content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Cool ($T_{\rm eff} \lesssim 10{,}000\,{\rm K}$) helium-atmosphere white dwarfs often have detectable traces of carbon in their atmospheres. Such carbon-polluted white dwarfs are known as DQ stars. Spectroscopically, DQ white dwarfs are recognized by the presence of C$_2$ molecular bands from the Swan transitions in the optical region. Given the relentless action of gravitational settling, heavy elements such as carbon are a priori not expected to be present in white dwarf atmospheres. But many physical processes can compete with gravitational settling. In the case of cool DQ stars, the presence of carbon is understood as the result of convective dredge-up from the deep interior \citep{pelletier1986,macdonald1998,althaus2005,dufour2005,koester2020}. Under reasonable modelling assumptions, the most recent evolutionary calculations agree remarkably well with the observed carbon abundance pattern of DQ stars \citep{bedard2022}.
The atmospheres of cool helium-rich white dwarfs are also frequently polluted by heavier elements, such as Ca, Mg and Fe. Those are known as DZ stars. The origin of heavy elements in DZ white dwarfs is now convincingly explained by the recent or ongoing accretion of rocky debris \citep[e.g.,][]{jura2014,farihi2016}. Many independent observations all converge to this explanation, including the abundance patterns observed in DZ atmospheres \citep{zuckerman2007,klein2010,gansicke2012,doyle2019,harrison2021}, the identification of circumstellar disks \citep{rocchetto2015,wilson2019,manser2020}, the discovery of planetary debris by the transit method \citep{vanderburg2015,vanderbosch2020,vanderbosch2021} and the detection of X-rays from the accretion process itself \citep{cunningham2022}.
Approximately 15\% of all helium-atmosphere white dwarfs are DZs \citep{mccleery2020,hollands2022}. Under the assumption that the DQ phenomenon is independent from the external accretion of rocky planetesimals, about 15\% of DQ white dwarfs should also have atmospheres polluted by elements heavier than carbon. But it is well established that DQZ/DZQ stars (white dwarfs showing both carbon features and absorption lines from heavier elements) are exceedingly rare. Only about 2\% of DQs show the presence of heavier elements in their spectra \citep{coutu2019,farihi2022}. To add to this mystery, the inferred external accretion rates of the few known DQZs are orders of magnitude smaller than those typical of DZ stars. To solve this long-standing conundrum, \cite{farihi2022} recently suggested that DQ stars are the product of a binary evolution that has altered their circumstellar environments in a way that prevents the pollution of the white dwarf.
In this Letter, we show that this hypothesis is not necessary. We demonstrate that the C$_2$ Swan bands of typical DQ white dwarfs are efficiently suppressed if the atmosphere is polluted by an amount of metals typical of that observed in DZ stars. This naturally explains both the paucity of DQZ white dwarfs and the systematically low accretion rates inferred for known DQZ stars. In Section~\ref{sec:suppress}, we present DQZ model atmosphere calculations that show how the Swan bands can disappear if the atmosphere is moderately polluted by heavy elements following the accretion of rocky debris. We then detail in Section~\ref{sec:physics} the physical mechanism that leads to this Swan bands suppression. In Section~\ref{sec:implications}, we compare our DQZ models to observations to demonstrate that the rarity of DQZ stars can naturally be explained by this process. Finally, our conclusions are stated in Section~\ref{sec:conclusion}.
\section{Metals can suppress the C$_2$ Swan bands}
\label{sec:suppress}
The possibility of Swan bands suppression in DQZ stars was recently studied by \cite{hollands2022}. Their Figure~3 shows how their models predict that the Swan bands disappear if a sufficient quantity of polluting metals is added to the atmosphere while keeping the other parameters constant \chg{(specifically, $T_{\rm eff}=8100\,{\rm K}$, $\log g=8.02$, $\log\,{\rm H/He}=-3.2$\footnote{In this work, $\log\,{\rm X/He}$ corresponds to the logarithm of the number abundance ratio, $\log_{10}n_{\rm \scriptscriptstyle X}/n_{\rm \scriptscriptstyle He}$.} and $\log\,{\rm C/He}=-4.5$)}. However, \cite{hollands2022} argue that this disappearance of the Swan bands cannot explain the extreme rarity of DQZ stars, since they find that the Swan bands are suppressed only at extreme levels of metal pollution. Therefore, they conclude that while the Swan bands can be sufficiently inhibited to become undetectable in some strongly polluted white dwarfs, they are generally too weakly suppressed to explain the dearth of DQZ white dwarfs.
\chg{In Figure~\ref{fig:demo_C45} we repeat the exercise of \cite{hollands2022} using our own state-of-the-art model atmospheres \citep{blouin2018,blouin2018b,blouin2019c}. We examine how the strength of the Swan bands changes for a normal-mass white dwarf with $T_{\rm eff} = 8000\,{\rm K}$ and $\log\,{\rm C/He}=-4.5$ as a function of the external pollution level $\log\,{\rm Ca/He}$\footnote{We only quote the Ca abundance, but all metals from N to Cu are included with chondritic abundance ratios with respect to Ca. This applies to all models presented in this work.} (no hydrogen is included). These atmospheric parameters are similar to those used in Figure~3 of \cite{hollands2022} but there are small differences since in that work the hydrogen and individual metal abundances were adjusted to the specific case of SDSS~J095645.12+591240.7. We qualitatively replicate their finding that a high metal pollution of $\log\,{\rm Ca/He} \gtrsim -8.5$ is required to strongly suppress the Swan bands at that temperature and carbon abundance.} This good agreement between independent calculations indicates that differences with respect to the constitutive physics of our atmosphere code and that used in \citealt{hollands2022} \citep{koester2010} are negligible in the context of this work.
\begin{figure}
\includegraphics[width=\columnwidth]{figs/demo_C45.pdf}
\caption{Synthetic spectra for a white dwarf with $T_{\rm eff}=8000\,$K, $\log g=8$, $\log\,{\rm C/He}=-4.5$ and varying amounts of polluting metals. No hydrogen trace was included. The blue regions indicate the location of the two strongest C$_2$ Swan bands.}
\label{fig:demo_C45}
\end{figure}
For their analysis, \cite{hollands2022} chose a fixed carbon abundance of $\log\,{\rm C/He}=-4.5$. This is not the ideal value to study the Swan bands suppression: DQ white dwarfs at 8000\,K typically have 10 times less carbon than that. This is shown in Figure~\ref{fig:DQscatter}, where the carbon abundances of a large sample of DQ white dwarfs are plotted as a function of $T_{\rm eff}$. At 8000\,K, $\log\,{\rm C/He}=-4.5$ \chg{(identified as a grey star)} lies a full one dex above the sequence on which the vast majority of DQs are found. While some stars are in this upper region, apparently forming a continuous sequence with the ``warm'' DQs found above $10{,}000\,$K \citep{coutu2019,koester2019}, they likely have a different origin than the classical DQs, possibly being the descendants of the Hot DQ white dwarfs. In short, the analysis presented in Figure~\ref{fig:demo_C45} of this work and in Figure~3 of \cite{hollands2022} is not representative of the typical DQ white dwarf.
\begin{figure}
\includegraphics[width=\columnwidth]{figs/DQscatter.pdf}
\caption{Photospheric carbon abundance of DQ white dwarfs as a function of effective temperature (data taken from \protect\citealt{coutu2019} and \protect\citealt{blouin2019d}). The vast majority of DQ white dwarfs follow a clear sequence in this plane as indicated by the dashed red line.}
\label{fig:DQscatter}
\end{figure}
We now repeat the exercise of \cite{hollands2022}, but this time using a carbon abundance of $\log\,{\rm C/He}=-5.5$, a representative composition for a 8000\,K DQ. Figure~\ref{fig:demo_C55} reveals that the weaker Swan bands are more easily erased from the spectrum when metals are added. The C$_2$ bands are very shallow even at the relatively low pollution level of $\log\,{\rm Ca/He}=-10.5$. This result is hardly surprising (if there is less carbon, then surely the Swan bands are weaker for any given metal pollution level), but it reopens the door to the idea that the dearth of DQZ white dwarfs could be explained by this simple effect. To test this idea, we need to compare our synthetic spectra to observations. This is the subject of Section~\ref{sec:implications}; for now, we turn to the identification of the physical mechanism responsible for the Swan bands suppression.
\begin{figure}
\includegraphics[width=\columnwidth]{figs/demo_C55.pdf}
\caption{Same as Figure~\ref{fig:demo_C45} but this time assuming $\log\,{\rm C/He}=-5.5$. Note the different vertical scale compared to Figure~\ref{fig:demo_C45}.}
\label{fig:demo_C55}
\end{figure}
\section{Physical explanation}
\label{sec:physics}
To identify the reason behind the disappearance of the C$_2$ bands following the addition of metals, we look in more details at the $\log\,{\rm Ca/He}=-11.5$ and $\log\,{\rm Ca/He}=-9.5$ models of Figure~\ref{fig:demo_C55}. The first thing to establish is at which Rosseland optical depths ($\tau_R$) the Swan bands are formed in the vertical stratification of the atmosphere. To answer this question, we have computed a series of model spectra where the Swan bands opacity was artificially ignored at certain optical depths (Figure~\ref{fig:taucontrib}). From this exercise, we conclude that at this effective temperature and carbon abundance, the Swan bands are mostly formed in the $0.001 \lesssim \tau_R \lesssim 0.03$ range.
Equipped with this information, we can now compare the $\log\,{\rm Ca/He}=-11.5$ and $\log\,{\rm Ca/He}=-9.5$ models to see how they differ. Figure~\ref{fig:kappa} compares their opacities as a function of $\tau_R$ at $\lambda=5160\,\textrm{\AA}$ (which corresponds to the wavelength where the Swan bands are strongest). It shows that the total opacity at a given $\tau_R$ (dashed lines) does not change much following the hundredfold increase in external metal pollution. In contrast, the C$_2$ Swan opacity (solid lines) decreases by a factor $\simeq 10$ in the region where the Swan bands are formed ($0.001 \lesssim \tau_R \lesssim 0.03$). Therefore, the suppression of the Swan bands seen in Figure~\ref{fig:demo_C55} is not caused by a masking effect from other opacity sources but rather by a decrease of the C$_2$ opacity itself.
\begin{figure}
\includegraphics[width=\columnwidth]{figs/kappa.pdf}
\caption{Opacity at $5160\,\textrm{\AA}$ as a function of Rosseland optical depth for DQZ white dwarfs with $T_{\rm eff}=8000\,$K, $\log g=8$, $\log\,{\rm C/He}=-5.5$ and different Ca/He values (black lines correspond to $\log\,{\rm Ca/He}=-11.5$ and red lines to $\log\,{\rm Ca/He}=-9.5$). The dashed lines show the total opacity while the solid lines show the opacity due to the Swan bands only. $5160\,\textrm{\AA}$ corresponds to the wavelength where the Swan bands are strongest. The grey region indicates where the Swan bands are formed.}
\label{fig:kappa}
\end{figure}
By inspecting our model structures, we found that this decline of the Swan opacity is directly caused by a tenfold decrease of the C$_2$ number density in the same region of the atmosphere (Figure~\ref{fig:C2abn}). This decrease is in turn largely explained by a $\simeq 2.5$x decrease of the total density between the $\log\,{\rm Ca/He}=-11.5$ and $\log\,{\rm Ca/He}=-9.5$ models in the relevant atmospheric layers ($0.001 \lesssim \tau_R \lesssim 0.03$, Figure~\ref{fig:structure}). The reason for this decline is well documented \citep{provencal2002,dufour2005,bergeron2019}. Metals act as electron donors that increase the total opacity of helium-dominated atmospheres mostly through an increase of the He$^-$ free-free opacity, which dominates the total atmospheric opacity in those stars \citep[][Figure~18]{saumon2022}. Due to this opacity increase, a given optical depth $\tau_R$ is attained higher up in the atmosphere, at a lower pressure.
To understand how this density decline affects the molecular carbon density, consider the C$_2$ dissociation equilibrium equation. Neglecting nonideal effects, it is given by
\begin{equation}
\frac{n_{\rm \scriptscriptstyle C}^2}{n_{{\rm \scriptscriptstyle C}_2}} = \frac{Q_{\rm \scriptscriptstyle C}^2 (T)}{Q_{{\rm \scriptscriptstyle C_2}} (T)} \left( \frac{ 2\pi m_{\rm \scriptscriptstyle C}^2 k_B T}{m_{{\rm \scriptscriptstyle C_2}} h^2}\right)^{3/2} e^{-D_0 / k_B T} \equiv f(T),
\label{eq:C2}
\end{equation}
where $n_i$ is a number density, $Q_i(T)$ is a partition function, $m_i$ is a mass, $k_B$ is the Boltzmann constant, $h$ is the Planck constant and $D_0=6.21\,{\rm eV}$ is the dissociation energy of the C$_2$ molecule. At a fixed temperature and given C/He, we have
\begin{equation}
n_{{\rm \scriptscriptstyle C}_2} \propto n_{\rm \scriptscriptstyle C}^2 \propto \rho^2,
\end{equation}
meaning that the 2.5x decrease in mass density noted above translates into a $\simeq 6$x reduction of the C$_2$ density.
In addition to this change in the density structure, the temperature profile is also affected (Figure~\ref{fig:structure}). This is a second-order but still important effect to explain the shift in molecular carbon abundances. In the region of interest ($0.001 \lesssim \tau_R \lesssim 0.03$), the temperature rises by $\simeq 400\,$K following the increase in metal pollution. This in turn translates into a twofold increase of the right-hand side of Equation~\eqref{eq:C2} (which we have plotted in Figure~\ref{fig:fT} for reference). Combining this effect with the density effect described in the previous paragraph we see how moderate changes to the density and temperature stratifications explain the tenfold decrease of the C$_2$ density in the Swan bands forming region of the atmosphere. We can conclude that the suppression of the Swan bands described in Section~\ref{sec:suppress} is due to changes to the atmospheric structure that tilt the scale in favour of dissociation in the C$_2$/C equilibrium equation.
\section{Implications for DQZ white dwarfs}
\label{sec:implications}
In the previous sections, we have seen how and why the accretion of metals on DQ white dwarfs can suppress their C$_2$ Swan bands. This suggests that DQ white dwarfs may transform directly into DZs (and not DQZs) following the accretion of rocky debris, which would naturally explain the apparent dearth of DQZ stars. To test this hypothesis, we turn to the Sloan Digital Sky Survey (SDSS) DZ sample of \cite{dufour2007}. This sample contains 72 DZ white dwarfs cooler than 9000\,K (a temperature range where the classical DQ sequence is well populated, see Figure~\ref{fig:DQscatter}) that have been recently reanalysed by \cite{coutu2019} using Gaia parallaxes and updated model atmospheres. Under the assumption that DQ white dwarfs descend from stars that hosted planetary systems similar to those hosted by the progenitors of non-DQ white dwarfs, we should expect that \chg{$11 \pm 3$} stars in this sample were DQs before accreting planetary material \citep[given that $\simeq 15$\% of cool helium-atmosphere white dwarfs are DQs,][]{mccleery2020}. \chg{The probability that none was a DQ is a priori $\lesssim 10^{-5}$.} The fact that Swan bands are detected in none of the 72~objects can mean two things: either the progenitors of DQs actually had different planetary system architectures after all \citep{farihi2022} or DQs generally transform directly into DZs after the accretion of metals. Of course, those two scenarios are not mutually exclusive, but if DQs do transform directly into DZs, then the need for the first scenario lessens considerably.
\begin{figure*}[ht]
\centering
\includegraphics[width=1.85\columnwidth]{figs/panel1.pdf}
\caption{Comparison between SDSS spectra (grey) and different model atmospheres (described in the text, see Section~\ref{sec:implications}) convolved at the SDSS resolution. Panels for the 66 remaining objects are available in Figure~\ref{fig:panelapp}.}
\label{fig:panel1}
\end{figure*}
To demonstrate that most DQs do transform directly into DZs, we have calculated model atmospheres for each of those 72~objects using the $T_{\rm eff}$, $\log g$, Ca/He and H/He values given in \cite{coutu2019}. In addition, we have computed models where we have boosted the carbon abundance so that C/He matches the expected carbon abundance for a DQ at that particular effective temperature (following the dashed red line of Figure~\ref{fig:DQscatter}). The resulting synthetic spectra are shown in Figure~\ref{fig:panel1}. For each object, we show four different models:
\begin{itemize}
\item The blue model is a reference model computed assuming the same parameters as those found by \cite{coutu2019}. It has the same parameters as those given in each panel except for C/He, which is fixed so that C/Ca is chondritic.
\item The orange model contains a trace of carbon typical of a DQ white dwarf at that temperature. It has the same parameters as those given in each panel.
\item The green model is identical to the orange model, except that no trace of hydrogen was included in the calculations. In contrast, in \cite{coutu2019} and in the orange models, H/He is adjusted to reproduce the Balmer lines or to match the visibility limit of H$\alpha$ if no Balmer line is detected (in that case, an asterisk is displayed next to the hydrogen abundance in Figure~\ref{fig:panel1}).
\item The red model (dashed line) is a DQ model with no polluting metals other than carbon. It has the same parameters as those given in each panel except that ${\rm H}/{\rm He}=0$ and ${\rm Ca}/{\rm He}=0$.
\end{itemize}
In all cases, a full model structure is recalculated: the synthetic spectra and underlying model atmospheres are fully consistent. The only parameters that were adjusted to the observations are the absolute scale and the slope of the synthetic spectra. All other parameters are fixed as described above. Finally, since \cite{coutu2019} adjusted Ca/He while keeping other abundance ratios constant (e.g., Mg/He), some objects with high Mg/Ca are not well represented by our models in the Mg$\,{\textsc{i}}$ 5168/5174/5185\,{\rm \AA} triplet region. In the context of this work, this is only a cosmetic issue that can safely be ignored.
The first thing to notice is that in most cases, the SDSS spectra have a sufficiently high signal-to-noise ratio (SNR) that Swan bands could be detected in DQ stars at those same $T_{\rm eff}$ and with typical carbon abundances (red dashed lines). In contrast, for the vast majority of objects, the same amount of carbon should not be detectable once we account for the presence of polluting metals (orange lines). \chg{For most stars, it would remain impossible to conclusively detect Swan bands even with observations with ${\rm SNR}=100$.} This conclusion still holds if we remove any hydrogen impurities from the model calculations (green lines). Hydrogen can further suppress the Swan bands by adding more free electrons to the atmosphere, which explains the difference between the orange and green models.
The key takeaway from Figure~\ref{fig:panel1} is that if, as expected, some objects in this sample are stars that have dredged up carbon from their deep interiors, they could not have been classified as DQZ because their Swan bands are too strongly suppressed to be detectable. The rarity of DQZ white dwarfs is not a surprise and is naturally explained by the well-known effect that electron donors have on the atmospheric structures of helium-dominated atmospheres. Figure~\ref{fig:panel1} also shows why it is unsurprising that the few known DQZ white dwarfs have very small accretion rates compared to the DZ population \citep{farihi2022}. The few objects where the Swan bands could have a chance of being detected with higher SNR observations are precisely those with the lowest levels of external pollution. \chg{For instance, we estimate that the Swan bands in the orange models of Figure~\ref{fig:panel1}/\ref{fig:panelapp} for the three most weakly polluted objects could be conclusively detected with an SNR $\gtrsim 40$ (WD~1005+030, $\log\,{\rm Ca/He}=-11.32$) and $\gtrsim 150$ (LSPM J1341+4253, $\log\,{\rm Ca/He}=-11.16$; USNO$-$B1.0 0937$-$00210798, $\log\,{\rm Ca/He}=-10.95$).}
So far, the hunt for more DQZs has been mostly focused on a search for metal lines (chiefly Ca$\,{\textsc{ii}}$ H \& K) in known DQ white dwarfs \citep{farihi2022}. By exclusively selecting known DQs, this survey strategy is effectively selecting objects that are very unlikely to be DQZs. In fact, since most known DQ white dwarfs have easily detectable Swan bands, they are objects that must have very low (if any) external pollution, otherwise they would not have strong Swan bands in the first place. A more promising survey strategy might be to look for very shallow Swan bands in known DZ white dwarfs with very low Ca/He, since a low external pollution implies a weaker Swan bands suppression.
\section{Conclusion}
\label{sec:conclusion}
Using state-of-the-art model atmospheres, we have shown that a classical DQ star with a typical carbon abundance directly transforms into a DZ white dwarf following the accretion of a typical amount of metals from planetary debris. The accreted material decreases the density of the atmosphere, which results in a smaller C$_2$ abundance and a strong suppression of the Swan bands. This naturally explains the observed paucity of DQZ stars as well as the very small accretion rates inferred for known DQZs.
Those findings nullify the main argument put forward by \cite{farihi2022} to support the idea that all DQ stars are the product of binary evolution. We cannot definitely rule out this scenario, but our conclusions at least lessen the need for this hypothesis. That being said, we recognize that there are still other properties of DQ stars that remain hard to explain (notably the observed deficit of unevolved companions in post-common envelope binaries) and each of those should be further scrutinized.
\begin{acknowledgements}
SB thanks Pierre Bergeron, Patrick Dufour and Antoine B\'edard for useful discussions. SB is a Banting Postdoctoral Fellow and a CITA National Fellow, supported by the Natural Sciences and Engineering Research Council of Canada (NSERC). This work has made use of the Montreal White Dwarf Database \citep{dufour2017}.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{IQP implementation}
\label{df:app:diffrac}
In this section, we introduce an IQP implementation to solve for \eqref{df:eq:disambiguation}.
We first mention that our alternative minimization scheme is not restricted to well-behaved problems, before motivating the introduction of the IQP algorithm in two different ways, and finally describing its implementation.
\subsection{Initialization of alternative minimization for non-well-behaved problem}
Before describing the IQP implementation to solve \eqref{df:eq:algo}, we would like to stress that, even for non-well-behaved partial labeling problems, it is possible to search for smart ways to initialize variables of the alternative minimization scheme.
For example, one could look at $z_i^{(0)} \in \cap_{j;x_j\in{\cal N}_{k_i}} S_j$, where ${\cal N}_k$ designs the $k$ nearest neighbors of $x_i$ in $(x_j)_{j\leq n}$, and $k_i$ is chosen such that this intersection is a singleton.
\subsection{Link with Diffrac and empirical risk minimization}
Our IQP algorithm is similar to an existing disambiguation algorithm known as the Diffrac algorithm \citep{Bach2007,Joulin2010}.\footnote{The Diffrac algorithm was first introduced for clustering, which is a classical approach to unsupervised learning.
In practice, it consists of changing the constraint set $C_n = \prod S_i$ by a set of the type $C_n = \argmax_{(y_i) \in \mathcal{Y}^n} \sum_{i,j=1}^n \ind{y_i\neq y_j}$ in \eqref{df:eq:disambiguation} and \eqref{df:eq:iqp}, meaning that $(y_i)$ should be disambiguated into different classes.}
This algorithm was derived by implicitly following empirical risk minimization of \eqref{df:eq:principle}.
This approach leads to algorithms written as
\[
(y_i) \in \argmin_{(y_i) \in C_n} \inf_{f\in{\cal F}} \sum_{i=1}^n \ell(f(x_i), y_i)
+ \lambda \Omega(f),
\]
for ${\cal F}$ a space of functions, and $\Omega:{\cal F} \to \mathbb{R}_+$ a measure of complexity.
Under some conditions, it is possible to simplify the dependency in $f$ \citep[{\em e.g.},][]{Xu2004,Bach2007}.
For example, if $\ell(y, z)$ can be written as $\norm{\phi(y) - \phi(z)}^2$ for a mapping $\phi:\mathcal{X}\to\mathcal{Y}$, {\em e.g.} the Kendall loss detailed in Section \ref{df:sec:ranking},\footnote{Since $\norm{\phi(y)}$ is constant.} and the search of $\phi(f):\mathcal{X}\to\phi(\mathcal{Y})$ is relaxed as a $g:\mathcal{X}\to{\cal H}$.
With $\Omega$ and ${\cal F}$ linked with kernel regression on the surrogate functional space $\mathcal{X}\to{\cal H}$, it is possible to solve the minimization with respect to $g$ as $g(x_i) = \sum_{j=1}^n \alpha_j(x_i)\phi(y_i)$, with $\alpha$ given by kernel ridge regression \citep{Ciliberto2016}, and to obtain a disambiguation algorithm written as
\[
\argmin_{y_i \in S_i} \sum_{i=1}^n\big\|\sum_{j=1}^n \alpha_j(x_i) \phi(y_j) - \phi(y_i)\big\|^2.
\]
This IQP is a special case of the one we will detail. As such, our IQP is a generalization of the Diffrac algorithm, and this paper provides, to our knowledge, {\em the first consistency result for Diffrac}.
\subsection{Link with another determinism measure}
While we have considered the measure of determinism given by \eqref{df:eq:principle}, we could have considered its quadratic variant
\[
\mu^\star \in \argmin_{\mu\vdash\nu} \inf_{f:\mathcal{X}\to\mathcal{Y}} \E_{X\sim\nu_\mathcal{X}}
\bracket{\E_{Y, Y'\sim\mu\vert_x}\bracket{\ell(Y, Y')}}.
\]
This corresponds to the right drawing of Figure \ref{df:fig:objective}.
We could arguably translate it experimentally as
\begin{equation}
\label{df:eq:iqp}
(\hat{y}_i) \in \argmin_{(y_i) \in C_n} \sum_{i,j=1}^n \alpha_i(x_j)
\ell(y_i, y_j),
\end{equation}
and still derive Theorem \ref{df:thm:convergence} when substituting \eqref{df:eq:disambiguation} by \eqref{df:eq:iqp}.
When the loss is a correlation loss $\ell(y, z) = -\phi(y)^\top\phi(z)$.
This leads to the quadratic problem
\[
(\hat{y}_i) \in \argmin_{(y_i) \in C_n} -\sum_{i,j=1}^n \alpha_i(x_j)
\phi(y_i)^\top \phi(y_j).
\]
\subsection{IQP implementation}
In order to make our implementation possible for any symmetric loss $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$, on a finite space $\mathcal{Y}$, we introduce the following decomposition.
\begin{proposition}[Quadratic decomposition]
When $\mathcal{Y}$ is finite, any proper symmetric loss $\ell$ admits a decomposition with two mappings $\phi:\mathcal{Y}\to\mathbb{R}^m$, $\psi:\mathcal{Y}\to\mathbb{R}^m$, for an $m\in\mathbb{N}$ and a $c\in\mathbb{R}$, reading
\begin{equation}
\label{df:eq:qloss}
\forall\, y, z\in \mathcal{Y}, \quad\ell(y, z) = \psi(y)^\top\psi(z) - \phi(y)^\top
\phi(z)
\qquad \text{with}\qquad \norm{\phi(y)} = \norm{\psi(y)} = c
\end{equation}
\end{proposition}
\small
\begin{proof}
Consider $\mathcal{Y} = {y_1, \cdot, y_m}$ and $L = (\ell(y_i, y_j))_{i,j\leq m} \in \mathbb{R}^{m\times m}$.
$L$ is a symmetric matrix, diagonalizable as
\(
L = \sum_{i=1}^m \lambda_i u_i\otimes u_i,
\)
with $(u_i)$ an orthonormal basis of $\mathbb{R}^m$, and $\lambda_i \in \mathbb{R}$ its eigenvalues.
We have, with $(e_i)$ the Cartesian basis of $\mathbb{R}^m$,
\[
\ell(y_j, y_k) = L_{jk} = \scap{e_j}{Le_k} = \sum_{i=1}^m (\lambda_i)_+
\scap{e_j}{u_i} \scap{e_k}{u_i}
- \sum_{i=1}^m (\lambda_i)_- \scap{e_j}{u_i} \scap{e_k}{u_i}.
\]
We build the decomposition
\[
\tilde{\psi}(y_k) = \paren{\sqrt{(\lambda_i)_+} \scap{e_k}{u_i}}_{i\leq m},
\qquad\text{and}\qquad
\tilde{\phi}(y_k) = \paren{\sqrt{(\lambda_i)_-} \scap{e_k}{u_i}}_{i\leq m}.
\]
It satisfies
\(
\ell(y_j, y_k) = \tilde\psi(y)^\top\tilde\psi(z) - \tilde\phi(y)^\top\tilde\phi(z).
\)
We only need to show that we can consider $\phi$ of constant norm.
For this, first consider $C = \max_i\abs{\lambda_i}$, we have
\(
\norm{\tilde{\psi}(y_k)}^2 = \sum_{i=1}^m (\lambda_i)_+ \scap{u_i}{e_k}^2
\leq C \sum_{i=1}^m \scap{u_i}{e_k}^2 = C\norm{e_k}^2 = C
\)
The last equalities being due to the fact that $(u_i)$ is orthonormal.
Now, introduce the correction vector $\xi:\mathcal{Y}\to\mathbb{R}^m$,
\(
\xi(y_i) = \sqrt{C - \norm{\tilde\psi(y)}^2} e_i.
\)
And consider $\phi = \binom{\tilde\phi}{\xi}$, $\psi = \binom{\tilde\psi}{\xi}$.
By construction, $\psi$ is of constant norm being equal to $C$ and that $\ell(y, z) = \psi(y)^T\psi(z) - \phi(y)^T\phi(z)$.
Finally, because $\ell(y, z) = 0$, we also have $\phi$ of constant norm.
\end{proof}
\normalsize
Using the decomposition \eqref{df:eq:qloss}, \eqref{df:eq:iqp} reads, with
$\textbf{y} = (y_i)$
\[
{\bf \hat y} \in \argmin_{{\bf y}\in C_n}\sum_{i=1}^n \alpha_i(x_j) \psi(y_i) \psi(y_j) - \sum_{i=1}^n \alpha_i(x_j) \phi(y_i) \phi(y_j).
\]
By defining the matrix $A = (\alpha_i(x_j))_{ij\leq n} \in \mathbb{R}^{n\times n}$, $\Psi({\bf y}) = (\psi(y_i))_{i\leq n} \in \mathbb{R}^{n\times m}$ and $\Phi({\bf y}) = (\phi(y_i))_{i\leq n} \in \mathbb{R}^{n\times m}$, we cast it as
\[
{\bf \hat y} \in \argmin_{{\bf y}\in C_n}\trace\paren{A\Psi({\bf y})\Psi({\bf y})^\top} - \trace\paren{A\Phi({\bf y})\Phi({\bf y})^\top}.
\]
\paragraph{Objective convexification.} As $\alpha_i(x_j)$ is a measure of similarity between $x_i$ and $x_j$, $A$ is usually symmetric positive definite, making this objective convex in $\Psi$ and concave in $\Phi$.
However, recalling \eqref{df:eq:qloss}, we have $\trace\Phi\Phi^\top = \trace{\Psi\Psi^\top} = n c$, therefore considering the spectral norm of $A$, we convexify the objective as
\[
{\bf \hat y} \in \argmin_{{\bf y}\in C_n}\trace\paren{(\norm{A}_*I +
A)\Psi({\bf y})\Psi({\bf y})^\top} + \trace\paren{(\norm{A}_*I - A)\Phi({\bf y})\Phi({\bf y})^\top}.
\]
Considering
\[
B = \paren{\begin{array}{cc} \norm{A}_*I + A & 0 \\ 0 & \norm{A}_*I - A \end{array}}
\qquad\text{and}\quad
\Xi({\bf y}) = \binom{\Psi({\bf y})}{\Phi({\bf y})},
\]
simplifies this objective as
\[
{\bf \hat y} \in \argmin_{{\bf y}\in C_n}\trace\paren{B\Xi({\bf y})\Xi({\bf y})^\top}.
\]
When parametrized by $\xi = \Xi({\bf y})$, this is an optimization problem with a convex quadratic objective and ``integer-like'' constraint $\xi \in \Xi(C_n)$, identifying to an integer quadratic program (IQP).
\paragraph{Relaxation.} IQP are known to be NP-hard, several tools exist in literature and optimization libraries implementing them.
The most classical approach consists in relaxing the integer constraint $\xi \in \Xi(C_n)$ into the convex constraint $\xi \in \hull(\Xi(C_n))$, solving the resulting convex quadratic program, and projecting back the solution toward an extreme of the convex set.
Arguably, our alternative minimization approach is a better grounded heuristic to solve our specific disambiguation problem.
\section{Example with graphical illustrations}
\label{df:app:example}
To ease the understanding of the disambiguation principle \eqref{df:eq:principle}, we provide a toy example with a graphical illustration, Figure~\ref{df:fig:objective}.
Since \eqref{df:eq:principle} decorrelates inputs, we will consider $\mathcal{X}$ to be a singleton, in order to remove the dependency to $\mathcal{X}$.
In the following, we consider $\mathcal{Y} = \brace{a, b, c}$, with the loss given by
\[
L = (\ell(y, z))_{y, z\in \mathcal{Y}} =
\paren{\begin{array}{ccc}
0 & 1 & 1 \\
1 & 0 & 2 \\
1 & 2 & 0
\end{array}}.
\]
This problem can be represented on a triangle through the embedding of probability measures reading $\xi:\prob{\mathcal{Y}}\to \mathbb{R}^3; \mu\to\mu(a)e_1 + \mu(b)e_2 + \mu(c)e_3$, and onto the triangle $\brace{z\in\mathbb{R}_+^3 \,\middle\vert\, z^\top 1 = 1}$. Note that $\xi$ can be extended from any signed measure of total mass normalized to one onto the plane $\brace{z\in\mathbb{R}^3 \,\middle\vert\, z^\top 1 = 1}$, as well as the drawings Figure \ref{df:fig:objective} can be extended onto the affine span of the represented triangles.
The objective \eqref{df:eq:principle} reads pointwise as $\prob{\mathcal{Y}}\to\mathbb{R}; \mu \to \min_{i\leq 3} e_i^\top L \xi(\mu)$, while its quadratic version reads $\prob{\mathcal{Y}}\to\mathcal{Y}; \mu\to \xi(\mu)^\top L \xi(\mu)$.
Note that while $L$ is not definite negative, one can check that the restriction of $\mathbb{R}^3 \to \mathbb{R}; z \to z^\top Lz$ to the definition domain $\brace{z\in\mathbb{R}^3\,\middle\vert\, z^\top 1 = 1}$ is concave, as suggested by the right drawing of Figure \ref{df:fig:objective}.
\begin{figure*}[t]
\centering
\input{disambiguation/images/simplex}
\caption{Exposition of a pointwise problem in the simplex $\prob\mathcal{Y}$, with $\mathcal{Y} = \brace{a, b, c}$ and a proper symmetric loss defined by $\ell(a, b) = \ell(a, c) = \ell(b, c) / 2$.
(Left) Representation of the decision regions $R_z = \brace{\mu\in\prob{\mathcal{Y}}\,\middle\vert\, z\in \argmin_{z'\in\mathcal{Y}} \E_{Y\sim\mu}[\ell(z, y)]}$ for $z\in\mathcal{Y}$.
(Middle Left) Representation of $R_\nu = \brace{\mu\in\prob{\mathcal{Y}} \,\middle\vert\, \mu\vdash\nu}$ for $\nu = (5\delta_{\brace{a, b, c}} + \delta_{\brace{c}} + \delta_{\brace{a, c}} + \delta_{\brace{b, c}}) / 8$.
(Middle Right) Level curves of the piecewise function $\prob{\mathcal{Y}}\to\mathbb{R}; \mu\to\min_{z\in\mathcal{Y}}\E_{Y\sim\mu}[\ell(z, Y)]$ corresponding to \eqref{df:eq:principle}.
(Right) Level curves of the quadratic function $\prob{\mathcal{Y}}\to\mathbb{R}; \mu\to \E_{Y, Y'\sim\mu}[\ell(Y, Y')]$. Our disambiguation \eqref{df:eq:principle} corresponds to minimizing the concave function represented in the middle right drawing on the convex domain represented in the middle left drawing.}
\label{df:fig:objective}
\end{figure*}
It should be noted that $(\ell, \nu)$ being a well-behaved partial labeling problem can be understood graphically, as having the intersection of the decision regions $\cap_{z\in S} R_z$ non-empty for any set $S$ in the support of $\nu$.
As such, it is easy to see that our toy problem is well-behaved for any distribution $\nu$.
Formally, to match Definition \ref{df:def:init}, we can define $\mu_{\brace{e}} = \delta_e$ for $e\in\brace{a, b, c}$ and
\[
\mu_{\brace{a,b}} = .5\delta_a + .5\delta_b,\quad
\mu_{\brace{a,c}} = .5\delta_b + .5\delta_c,\quad
\mu_{\brace{b, c}} = \delta_b + \delta_c - \delta_a,\quad
\mu_{\brace{a, b, c}} = .5\delta_b + .5\delta_c.
\]
Graphically $\xi(\mu_{\brace{a,b}})$ can be chosen as any points on the horizontal dashed line on the middle right drawing of Figure \ref{df:fig:objective} (similarly for $\xi\mu_{\brace{a, c}}$), while $\xi(\mu_{\brace{a, b, c}})$ has to be chosen has the intersection $.5e_2 + .5e_3$, and while $\xi(\mu_{\brace{b,c}})$ has to be chosen outside the simplex on the half-line leaving $.5e_2 + .5e_3$ supported by the perpendicular bisector of $[e_2, e_3]$ and not containing $e_1$.
\section{Experiments}
\label{df:app:experiments}
While our results are much more theoretical than experimental, out of principle, as well as for reproducibility, comparison and usage sake, we detail our experiments.
\subsection{Interval regression - Figure \ref{df:fig:ir}}
Figure \ref{df:fig:ir} corresponds to the regression setup consisting of learning $f^*:[0, 1]\to\mathbb{R}; x\to\sin(\omega x)$, with $\omega=10\approx 3\pi$.
The dataset represented on Figure \ref{df:fig:ir} is collected in the following way.
We sample $(x_i)_{i\leq n}$ with $n = 10$, uniformly at random on $\mathcal{X}=[0, 1]$, after fixing a random seed for reproducibility.
We collect $y_i = f(x_i)$.
We create $(s_i)$ by sampling $u_i$ uniformly on $[0,1]$, defining $r_i = r - \gamma \log(u_i)$, with $r=1$ and $\gamma = 3^{-1}$, sampling $c_i$ uniformly at random on $[0, r_i]$, and defining $s_i = y_i + \sign(y_i)\cdot c_i + [-r_i, r_i]$.
The corruption is skewed on purpose to showcase disambiguation instability of the baseline \eqref{df:eq:baseline} compared to our method.
We solve \eqref{df:eq:disambiguation} with alternative minimization, initialized by taking $y_i^{(0)}$ at the center of $s_i$, and stopping the minimization scheme when $\sum_{i\leq n}\vert y_i^{(t+1)} - y_i^{(t)}\vert < \epsilon$ for $\epsilon$ a stopping criterion fixed to $10^{-6}$.
For $x\in \mathcal{X}$, the inference \eqref{df:eq:estimate} and \eqref{df:eq:baseline} is done through grid search, considering, for $f_n(x)$, 1000 guesses dividing uniformly $[-6, 6] \subset \mathcal{Y} = \mathbb{R}$.
We consider weights $\alpha$ given by kernel ridge regression with Gaussian kernel, defined as
\[
\alpha(x) = (K+n\lambda I)^{-1}K_x \in \mathbb{R}^n, \quad
K = (k(x_i, x_j))_{i,j\leq n} \in \mathbb{R}^{n\times n},\quad
K_x = (k(x_i, x))_{i\leq n} \in \mathbb{R}^n,
\]
with $k(x, x') = \exp\paren{-\frac{\norm{x-x'}^2}{2\sigma^2}}$, and $\lambda$ a regularization parameter, and $\sigma$ a standard deviation parameter.
In our simulation, we fix $\sigma = .1$ based on simple considerations on the data, while we consider $\lambda \in [10^{-1}, 10^{-3}, 10^{-6}]$.
The evaluation of the mean square error between $f_n$ and $f^*$, which is equivalent to evaluating the risk with the regression loss $\ell(y, z) = \norm{y - z}^2$, is done by considering 200 points dividing uniformly $\mathcal{X} = [0,1]$ and evaluating $f_n$ and $f^*$ on it.
The best hyperparameter $\lambda$ is chosen by minimizing this error.
It leads to $\lambda = 10^{-1}$ for the baseline \eqref{df:eq:baseline}, and $\lambda = 10^{-6}$ for our algorithm \eqref{df:eq:disambiguation} and \eqref{df:eq:estimate}.
This difference in $\lambda$ is normal since both methods are not estimating the same surrogate quantities.
The fact that $\lambda$ is smaller for our algorithm is natural as our disambiguation objective \eqref{df:eq:disambiguation} already has a regularization effect on the solution.\footnote{%
Moreover, the analysis in \citet{Cabannes2020} suggests that the baseline is estimating a surrogate function in $\mathcal{X}\to 2^\mathbb{R}$, while our method is estimating a function in $\mathcal{X}\to\mathbb{R}$, which is a much smaller function space, hence needing less regularization.
However, those reflections are based on upper bounds, that might be suboptimal, which could invalidate those considerations.}
Note that we used the same weights $\alpha$ for \eqref{df:eq:disambiguation} and \eqref{df:eq:estimate}, which is suboptimal, but fair to the baseline, as, consequently, both methods have the same number of hyperparameters.
\subsection{Classification - Figure \ref{df:fig:cl}}
Figure \ref{df:fig:cl} corresponds to classification problems, based on real datasets from the LIBSVM datasets repository.
At the time of writing, the datasets are available at \url{https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/multiclass.html}.
We present results on the ``DNA'' and ``Svmguide2'' datasets, that both have 3 classes ($m=3$), and respectively have 4000 samples with 180 features ($n=4000$,$d=180$) and 391 samples with 20 features ($n = 391$, $d=20$).
In terms of {\em complexity}, when $\mathcal{Y} = \bbracket{1, m} = \brace{1, 2, \cdots, m}$, and weights based on kernel ridge regression with Gaussian kernel as described in the last paragraph the complexity of performing inference for \eqref{df:eq:estimate} and \eqref{df:eq:baseline} can be done in $O(nm)$ in time and $O(n+m)$ in space, where $n$ is the number of training samples \citep{Nowak2019,Cabannes2020}.
The disambiguation \eqref{df:eq:disambiguation} performed with alternative minimization is done in $O(cn^2m)$ in time and in $O(n(n+m))$ in space, with $c$ the number of steps in the alternative minimization scheme.
In practice, $c$ is really small, which can be understood since we are minimizing a concave function and each step leads to a guess on the border of the constraint domain.
Based on the dataset $(x_i, y_i)$, we create $(s_i)$ by sampling it accordingly to $\gamma \delta_{\brace{y_i}} + 1-\gamma \delta_{\brace{y, y_i}}$, with $y$ the most present labels in the dataset (indeed we choose the two datasets because they were not too big and presenting unequal labels proportion), and $\gamma\in[0,1]$ the corruption parameter represented in percentage on the $x$-axis of Figure \ref{df:fig:cl}.
This skewed corruption allows distinguishing methods and invalidates the simple approach consisting of averaging candidate (AC) in set to recover $y_i$ from $s_i$, which works well when data are {\em missing at random} \citep{Heitjan1991}.
We separate $(x_i, s_i)$ in 8 folds, consider $\sigma \in d\cdot[1, .1, .01]$, where $d$ is the dimension of $\mathcal{X}$, and $\lambda \in n^{-1/2}\cdot [1, 10^{-3}, 10^{-6}]$, where $n$ is the number of data. We tested the different hyperparameter setup and reported the best error for each corruption parameter on Figure \ref{df:fig:cl}.
Those errors are measured with the 0-1 loss, computed as averaged over the 8 folds, {\em i.e.} cross-validated, with standard deviation represented as error bars on the figure.
The best hyperparameter generally corresponds to $\sigma = .1$ and $\lambda = 10^{-3}$ when the corruption is small and $\sigma = 1$, $\lambda = 10^{-3}$ when the corruption is big.
Differences between cross-validated error and testing error were small, and we presented the first one out of simplicity.
In terms of {\em energy cost}, the experiments were run on a personal laptop that has two processors, each of them running 2.3 billion instructions per second.
During experiments, all the data were stored on the random access memory of 8 GB.
Experiments were run on Python, extensively relying on the NumPy library \citep{Harris2020}.
The heaviest computation is Figure \ref{df:fig:cl}.
Its total runtime, cross-validation included, was around 70 seconds.
This paper is the result of experimentation, we evaluate the total cost of our experimentation to be three orders of magnitude higher than the cost of reproducing the final computations presented on Figure \ref{df:fig:ir}, \ref{df:fig:cl} and \ref{df:fig:ss}.
The total computational energy cost is negligible.
\subsection{Semi-supervised learning - Figure \ref{df:fig:ss}}
\begin{figure*}[ht]
\centering
\includegraphics{disambiguation/images/rk/setting_2.pdf}
\includegraphics{disambiguation/images/rk/setting_1.pdf}
\caption{
Ranking setting.
We consider $\mathcal{X}$ an interval of $\mathbb{R}$, and $\mathcal{Y} = \mathfrak{S}_m$ with $m=4$ on the figure.
(Right) To create a ranking dataset, we sample randomly $m$ lines in $\mathbb{R}^2$, embedding a value, or equivalently a score, associated to each item as a function of the input $x$.
(Left) By ordering those lines, we create preferences between items as a function of $x$. On the figure, when $x$ is small, the ``red'' item is preferred over the ``orange'' item, itself preferred over the ``blue'' item, itself preferred over the ``green'' item. While when $x$ is big, ``green'' is preferred over ``blue'', preferred over ``orange'', preferred over ``red''.
We create a partial labeling dataset by sampling $(x_i) \in \mathcal{X}^n$, and providing only partial ordering that the $(y_i)$ follow.
For example, for a small $x$, we might only give the partial information that ``red'' is preferred over ``blue''.
}
\label{df:fig:rk_set}
\end{figure*}
\begin{figure*}[ht!]
\centering
\includegraphics{disambiguation/images/rk/reconstruction.pdf}
\caption{
Performance of our algorithm for ranking with partial ordering.
This figure is similar to Figure \ref{df:fig:cl}, but is based on the ranking problem illustrated on Figure \ref{df:fig:rk_set}.
For this figure, we consider $m = 10$, as it is arguably the limit where the LP relaxation provided by \citet{Cabannes2020} of the NP-hard minimum feedback arcset problem still performs well.
The corruption parameter corresponds to the proportion of coordinates lost in the Kendall embedding when creating $s_i$ from $y_i$.
Because the Kendall embedding satisfies transitivity constraints, a corruption smaller than 50\% is almost ineffective to remove any information.
In this figure, we observe a similar behavior for ranking to the one observed for classification on Figure \ref{df:fig:cl}, suggesting that those empirical findings are not spurious.
}
\label{df:fig:rk}
\end{figure*}
On Figure \ref{df:fig:ss}, we review a semi-supervised classification problem with $\mathcal{Y} = \bbracket{1, 4}$, $\mathcal{X} = [-4.5, 4.5]^2$, $\mu_\mathcal{X}$ only charging $\brace{x=(x_1, x_2)\in\mathbb{R}^2\,\middle\vert\, x_1^2 + x_2^2 \in \mathbb{N}^*}$ and the solution $f^*:\mathcal{X} \to \mathcal{Y}$ being defined almost everywhere as $f^*(x) = x_1^2 + x_2^2$.
We collect a dataset $(x_i, s_i)$, by sampling 2000 points $\theta_i$ uniformly at random on $[0, 1]$, as well as $r_i$ uniformly at random in $\bbracket{1, 4} = \brace{1, 2, 3, 4}$, before building $x_i = r_i\cdot (\cos(2\pi \theta_i), \sin(2\pi\theta_i)) \in \mathcal{X}$, and $s_i = \mathcal{Y}$.
We add four labeled points to this dataset $x_{2001} = (-2\sqrt{3}, 2)$ with $s_{2001} = \brace{4}$, $x_{2001} = (1, -2\sqrt{2})$ with $s_{2002} = \brace{3}$, $x_{2001} = (-\sqrt{3}, -1)$ with $s_{2003} = \brace{2}$ and $x_{2001} = (-1, 0)$ with $s_{2004} = \brace{1}$.
We designed the weights $\alpha$ in \eqref{df:eq:disambiguation} with $k$-nearest neighbors, with $k=20$, and solve this equation with a variant of alternative minimization, leading to the optimal solution $\tilde{y}_i = y_i^*$.
In order to be able to compute the baseline \eqref{df:eq:baseline}, we design weights $\alpha$ for the inference task based on Nadaraya-Watson estimators with Gaussian kernel, defined as $\alpha_i(x) = \exp\paren{\norm{x-x_i}^{2} / h}$, with $h = .08$. We solve the inference task on a grid of $\mathcal{X}$ composed of 2500 points, and artificially recreate the observations to make them neat and reduce the resulting PDF size.
Note that it is possible to design weights $\alpha$ that capture the cluster structure of the data, which, in this case, will lead to a nice behavior of the baseline as well as our algorithm.
Arguably, this experiment showcases a regularization property of our algorithm \eqref{df:eq:disambiguation}.
\subsection{Ranking with partial ordering}
To conclude this experiment section, we look at ranking with partial ordering.
We refer to Section \ref{df:sec:ranking} for a clear description of this instance of partial labeling.
We provide to the reader eager to use our method, an implementation of our algorithm, available online at \url{https://github.com/VivienCabannes/partial_labelling/}.
It is based on LP relaxation of the NP-hard minimum feedback arcset problem.
This relaxation was proven exact when $m \leq 6$ by \citet{Cabannes2020}.
The LP implementation relies on CPLEX \citep{Cplex}.
As complementary experiments, we will not provide much reproducibility details, those details would be really similar to the previous paragraphs, and the curious reader could run our code instead.
We present our ranking setup on Figure \ref{df:fig:rk_set} and our results on Figure \ref{df:fig:rk}.
\section{Proofs}
\paragraph{Mathematical assumptions.}
To make formal what should be seen as implicit assumptions heretofore, we consider $\mathcal{X}$ and $\mathcal{Y}$ Polish spaces, $\mathcal{Y}$ compact, $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ continuous, ${\cal H}$ a separable Hilbert space, $\phi$ measurable, and $\psi$ continuous.
We also assume that for $\nu_x$-almost every $x\in\mathcal{X}$, and any $\mu\vdash\nu$, that the pushforward measure $\phi_*\mu\vert_x$ has a second moment.
This is the sufficient setup in order to be able to formally define objects and solutions considered all along the paper.
\paragraph{Notations.}
Beside standard notations, we use $\card{\mathcal{Y}}$ to design the cardinality of $\mathcal{Y}$, and $2^\mathcal{Y}$ to design the set of subsets of $\mathcal{Y}$.
Regarding measures, we use $\mu_\mathcal{X}$ and $\mu\vert_x$ respectively the marginal over $\mathcal{X}$ and the conditional accordingly to $x$ of $\mu\in\prob{\mathcal{X}\times\mathcal{Y}}$.
We denote by $\mu^{\otimes n}$ the distribution of the random variable $(Z_1, \cdots, Z_n)$, where the $Z_i$ are sampled independently according to $\mu$.
For $A$ a Polish space, we consider $\prob{A}$ the set of Borel probability measures on this space.
For $\phi:\mathcal{Y}\to{\cal H}$ and $S\subset \mathcal{Y}$, we denote by $\phi(S)$ the set $\brace{\phi(y)\,\middle\vert\, y\in S}$.
For a family of sets $(S_i)$, we denote by $\prod S_i$ the Cartesian product $S_1\times S_2\times\cdots$, also defined as the set of points $(y_i)$ such that $y_i \in S_i$ for all index $i$, and by $\mathcal{Y}^n$ the Cartesian product $\prod_{i\leq n}\mathcal{Y}$.
Finally, for $E$ a subset of a vector space $E'$, $\hull E$ denotes the convex hull of $E$ and $\Span(E)$ its span.
\paragraph{Abuse of notations.}
For readability’ sake, we have abused notations.
For a signed measure $\mu$, we denote by $\E_{\mu}[X]$ the integral $\int x\mathop{}\!\mathrm{d}\mu(x)$, extending this notation usually reserved to probability measure.
More importantly, when considering $2^\mathcal{Y}$, we should actually restrict ourselves to the subspace ${\cal S} \subset 2^\mathcal{Y}$ of closed subsets of $\mathcal{Y}$, as ${\cal S}$ is a Polish space (metrizable by the Hausdorff distance) while $2^\mathcal{Y}$ is not always.
However, when $\mathcal{Y}$ is finite, those two spaces are equals, $2^\mathcal{Y} = {\cal S}$.
\subsection{Proof of Lemma \ref{df:lem:cal}}
\label{df:proof:cal}
From Lemma 3 in \citet{Cabannes2021b}, we pull the calibration inequality
\[
{\cal R}(f_n) - {\cal R}(f^*) \leq 2c_\psi
\E\bracket{\ind{\norm{g_n(X) - g^*(X)} > d(g^*(X), F)}\norm{g_n(X) - g^*(X)}}.
\]
Where $F$ is defined as the set of points $\xi \in \hull\phi(\mathcal{Y})$ leading to two decodings
\[
F = \brace{\xi \in \hull\phi(\mathcal{Y})\,\middle\vert\,
\card{\argmin_{z\in\mathcal{Y}}\scap{\psi(z)}{\xi}} > 1},
\]
and $d$ is defined as the extension of the norm distance to sets, for $\xi \in {\cal H}$
\[
d(\xi, F) = \inf_{\xi'\in F} \norm{\xi - \xi'}_{\cal H}.
\]
Using that $\norm{g_n(X) - g^*(X)} \leq \norm{g_n(X) - g_n^*(X)} + \norm{g_n^*(X) - g^*(X)}$ and that, if $a \leq b + c$,
\[
\ind{a > \delta} a \leq \ind{b + c > \delta} b + c \leq \ind{2\sup(b, c) >
\delta} 2 \sup{b, c} = 2 \sup_{e\in{b, c}} \ind{e > \delta} e
\leq 2 \ind{b>\delta} b + 2\ind{c > \delta} c.
\]
We get the refined inequality
\begin{align*}
& {\cal R}(f_n) - {\cal R}(g^*) \\&\quad\leq 4 c_\psi
\E\bracket{\ind{2\norm{g_n(X) - g_n^*(X)} > d(g^*(X), F)}\norm{g_n(X) -
g_n^*(X)}
+ \ind{2\norm{g_n^*(X) - g^*(X)} > d(g^*(X), F)}\norm{g_n^*(X) - g^*(X)}}.
\end{align*}
The first term is bounded by
\[
\E\bracket{\ind{2\norm{g_n(X) - g_n^*(X)} > d(g^*(X), F)}\norm{g_n(X) -
g_n^*(X)}} \leq \norm{g_n - g_n^*}_{L^1}.
\]
While for the second term, we proceed with
\begin{align*}
& \E\bracket{\ind{2\norm{g_n^*(X) - g^*(X)} > d(g^*(X), F)}\norm{g_n^*(X)
-g^*(X)}}
\\&\qquad\leq \norm{g_n^* - g^*}_{L^\infty} \Pbb_X\paren{2\norm{g_n^*(X) - g^*(X)} >
\inf_{x\in\supp\nu_\mathcal{X}} d(g^*(X), F)}.
\end{align*}
When weights are positive and sum to one, both $g_n^*(X)$ and $g^*(X)$ are averaging of $\phi(y)$ for $y\in\mathcal{Y}$, therefore
\[
\norm{g_n^* - g^*}_{L^\infty} \leq 2 c_\phi.
\]
The same is true when $\sum_{i\leq n}\abs{\alpha_i(x)} \leq 1$.
Finally, when the labels are a deterministic function of the input, $g^*(X) = \phi(f^*(X))$, and $d(g^*(X), F) \leq \sup_{y\in\mathcal{Y}} d(\phi(y), F)$.
Defining $\delta := \sup_{y\in\mathcal{Y}} d(\phi(y), F) /2$, and adding everything together leads to Lemma~\ref{df:lem:cal}.
\subsection{Implication of Assumptions \ref{df:ass:non-ambiguity} and
\ref{df:ass:lipschitz}}
\label{df:proof:ass}
Assume that Assumption \ref{df:ass:non-ambiguity} holds, consider $x\in\supp\nu_\mathcal{X}$, let us show that $f^*(x) = y_x$ and $\mu^*\vert_x = \delta_{y_x}$.
First of all, notice that \(\bigcap_{S;S\in\supp\nu\vert_x} = \brace{y_x} \); that $\delta_{y_x} \vdash \nu\vert_x$, as it corresponds to $\pi\vert_{x, S} = \delta_{y_x} \in \prob{S}$, for all $S$ in the support of $\nu\vert_{x}$; and that, because $\ell$ is well-behaved,
\[
\inf_{z\in\mathcal{Y}} \ell(z, y_x) = \ell(y_x, y_x) = 0.
\]
This infimum is only achieved for $z = y_x$, hence if we prove that $\mu^*\vert_x = \delta_{y_x}$, we directly have that $f^*(x) = y_x$.
Finally, suppose that $\mu\vert_x\vdash\nu\vert_x$ charges $y \neq y_x$.
Because $y$ does not belong to all sets charged by $\nu\vert_x$, $\mu\vert_x$ should charge another $y'\in\mathcal{Y}$, and therefore
\[
\inf_{z\in\mathcal{Y}} \E_{Y\sim\mu\vert_x}[\ell(z, y)] \geq
\inf_{z\in\mathcal{Y}}\mu\vert_x(y) \ell(z, y) +
\mu\vert_x(y') \ell(z, y') > 0.
\]
Which shows that $\mu^*\vert_x = \delta_{y_x}$.
We deduce that $g^*(x) = y_x$.
Now suppose that Assumption \ref{df:ass:lipschitz} holds too, and consider two $x, x' \in\supp\nu_\mathcal{X}$ belonging to two different classes $f(x) = y$ and $f(x') = y'$.
We have that $g^*(x) = \phi(y)$ and $g^*(x') =
\phi(y')$, therefore,
\[
d(x, x') \geq c^{-1} \norm{\phi(y) - \phi(y')}_{\cal H}.
\]
Define $h_2 = \inf_{y\neq y'} c^{-1} \norm{\phi(y) - \phi(y')}_{\cal H}$.
Let us now show that $h_2 > 0$.
When $\mathcal{Y}$ is finite, this infimum is a minimum, therefore, $h_2 = 0$, only if there exists a $y \neq y'$, such that $\phi(y) = \phi(y')$, which would implies that $\ell(\cdot, y) = \ell(\cdot,y')$ and therefore $\ell(y, y') = \ell(y, y)$ which is impossible when $\ell$ is proper.
\subsection{Proof of Theorem \ref{df:thm:convergence}}
\label{df:proof:convergence}
Reusing Lemma \ref{df:lem:cal}, we have
\[
{\cal E}(f_n) \leq 4c_\psi \E_{{\cal D}_n, X}\bracket{\norm{g_n^*(X) - g_n(X)}_{\cal H}} +
8c_\psi c_\phi \E_{{\cal D}_n, X}\bracket{\ind{\norm{g_n^*(X) - g^*(X)} > \delta}}.
\]
We will first prove that
\[
\E_{{\cal D}_n}\bracket{\ind{\norm{g_n^*(X) - g^*(X)} > \delta}}
\leq \exp\paren{-\frac{np}{8}}
\]
as long as $k < np / 2$.
The error between $g^*$ and $g_n$ relates to classical supervised learning of $g^*$ from samples $(X_i, Y_i) \sim \mu^*$.
We invite the reader who would like more insights on this fully supervised part of the proof to refer to the several monographs written on local averaging methods and, in particular, nearest neighbors, such as \citet{Biau2015}.
Because of class separation, we know that if $k$ points fall at distance at most $h$ of $x \in \supp\nu_\mathcal{X}$, $g_n^*(x) = k^{-1}\sum_{i; X_i\in{\cal N}(x)} \phi(y_i) = \phi(y_x) = g^*(x)$, where ${\cal N}(x)$ designs the $k$-nearest neighbors of $x$ in $(X_i)$.
Because the probability of falling at distance $h$ of $x$ for each $X_i$ is lower bounded by~$p$, we have that
\[
\Pbb_{{\cal D}_n}(g_n^*(x) \neq g^*(x)) \leq \Pbb(\text{Bernoulli}(n, p) < k).
\]
This can be upper bound by $\exp(- np / 8)$ as soon as $k < np/2$, based on Chernoff multiplicative bound \citep[see][for a reference]{Biau2015}, meaning
\[
\E_{{\cal D}_n, X} \bracket{\ind{\norm{g_n^*(X) - g^*(X)} \geq \delta}}
\leq \exp(- np / 8).
\]
For the disambiguation part in $\norm{g_n - g_n^*}_{L^1}$, we distinguish two types of datasets, the ones where for any input $X_i$ its $k$-neighbors at are distance at least $h$, ensuring that disambiguation can be done by clusters, and datasets that does not verify this property.
Consider the event
\[
\mathbb{D} = \brace{(X_i)_{i\leq n} \,\middle\vert\, \sup_{i} d(X_i, X_{(k)}(X_i)) < h}
\]
where $X_{(k)}(x)$ design the $k$-th nearest neighbor of $x$ in $(X_i)_{i\leq n}$.
We proceed with
\[
\E_{{\cal D}_n, X}\bracket{\norm{g_n^*(X) - g_n(X)}_{\cal H}}
\leq \sup_{X\in\mathcal{X}} \norm{g_n^* - g_n}_{\infty} \Pbb_{{\cal
D}_n}((X_i) \notin\mathbb{D})
+ \E_{{\cal D}_n, X}\bracket{\norm{g_n^*(X) - g_n(X)}_{\cal H}\,\middle\vert\, (X_i) \in \mathbb{D}},
\]
Which is based on $E[Z] = \Pbb(Z\in A)\E[Z\vert A] + \Pbb(Z\notin A)\E[Z\vert ^cA]$.
For the term corresponding to bad datasets, we can bound the disambiguation error with the maximum error.
Similarly to the derivation for Lemma \ref{df:lem:cal}, because $g_n^*(x)$ and $g_n^*(X)$, are averaging of $\phi(y)$, we have that
\[
\sup_{x\in\supp\nu_\mathcal{X}} \norm{g_n(x) - g_n^*(x)} \leq 2c_\phi.
\]
Indeed, we allow ourselves to pay the worst error on those datasets as their probability is really small, which can be proved based on the following derivation.
\begin{align*}
\Pbb_{{\cal D}_n}((X_i)_{i\leq n} \notin\mathbb{D})
& = \Pbb_{(X_i)}( \sup_{i} d(X_i, X_{(k)}(X_i)) \geq h)
= \Pbb_{(X_i)}\paren{\cup_{i\leq n}\brace{d(X_i, X_{(k)}(X_i)) \geq h}}
\\&\leq \sum_{i=1}^n \Pbb_{(X_i)}\paren{d(X_i, X_{(k)}(X_i)) \geq h}
= n \Pbb_{X, {\cal D}_{n-1}}\paren{d(X, X_{(k)}(X)) \geq h}.
\end{align*}
This last probability has already been work out when dealing with the fully supervised part, and was bounded as
\[
\Pbb_{X, {\cal D}_{n-1}}\paren{d(X, X_{(k)}(X)) \geq h} \leq
\exp\paren{-(n-1)p / 8}.
\]
as long as $k < (n-1)p / 2$.
Finally, we have
\[
\sup_{X\in\mathcal{X}} \norm{g_n^* - g_n}_{\infty}
\Pbb_{{\cal D}_n}((X_i)_{i\leq n} \notin\mathbb{D})
\leq 2c_\phi n \exp\paren{-(n-1)p / 8}.
\]
For the expectation term, corresponding to datasets ${\cal D}_n \in \mathbb{D}$ that cluster data accordingly to classes, we have to make sure that $\hat{y}_i = y_i^*$ is the only acceptable solution of \eqref{df:eq:disambiguation}, which is true as soon as the intersection of $S_j$, for $x_j$ the neighbors of $x_i$, only contained $y_i^*$.
To work out the disambiguation algorithm, notice that
\begin{align*}
\norm{g_n - g_n^*}_{L^1}
& =\int_\mathcal{X} \norm{\sum_{i=1}^n \alpha_i(x) \phi(\hat{y}_i) - \phi(y^*_i)} \mathop{}\!\mathrm{d}\nu_\mathcal{X}(x)
\leq \int_\mathcal{X} k^{-1}\sum_{i=1}^n \ind{X_i \in{\cal N}(x)} \norm{\phi(\hat{y}_i) - \phi(y^*_i)} \mathop{}\!\mathrm{d}\nu_\mathcal{X}(x)
\\ &= k^{-1}\sum_{i=1}^n \Pbb_X\paren{X_i \in{\cal N}(X)} \norm{\phi(\hat{y}_i) - \phi(y^*_i)}
\leq 2c_\phi k^{-1}\sum_{i=1}^n \Pbb_X\paren{X_i \in{\cal N}(X)} \ind{\phi(\hat{y}_i) \neq \phi(y^*_i)}.
\end{align*}
Finally we have, after proper conditioning, considering the variability in $S_i$ while fixing $X_i$ first,
\begin{align*}
& \E_{{\cal D}_n, X}\bracket{\norm{g_n^*(X) - g_n(X)}_{\cal H}\,\middle\vert\, (X_i) \in \mathbb{D}}
\\&\qquad= 2c_\phi k^{-1}\E_{(X_i)}\bracket{\sum_{i=1}^n \Pbb_X\paren{X_i \in{\cal
N}(X)} \E_{(S_i)}\bracket{\ind{\phi(\hat{y}_i) \neq
\phi(y^*_i)} \,\middle\vert\, (X_i)} \,\middle\vert\, (X_i) \in \mathbb{D}}
\\&\qquad= 2c_\phi k^{-1}\E_{(X_i), X}\bracket{\sum_{i=1}^n \ind{X_i \in{\cal
N}(X)} \Pbb_{(S_i)}\paren{\phi(\hat{y}_i) \neq
\phi(y^*_i) \,\middle\vert\, (X_i)} \,\middle\vert\, (X_i) \in \mathbb{D}}.
\end{align*}
We design $\mathbb{D}$ so that the $k$-th nearest neighbor of any input $X_i$ is at distance at most $h$ of $X_i$, meaning the because of class separation, $y_{x_i} \in S_j$ for any $X_j \in {\cal N}(X_i)$.
This mean that outputting $(\hat{y}_i) = (y^*_i)$ and $z_j = y_j$, will lead to an optimal error in \eqref{df:eq:disambiguation}.
Now suppose that there is another solution for \eqref{df:eq:disambiguation} such that $\hat{y}_i \neq y_i^*$, it should also achieve an optimal error, therefore it should verify $z_j = \hat{y}_j$ for all $j$ as well as $\hat{y}_j = \hat{y}_i$ for all $j$ such that $X_j$ is one of the $k$ nearest neighbors of $X_i$.
This implies that $\hat{y}_i \in \cap_{j;X_j \in {\cal N}(X_i)} S_j$, which happen with probability
\[
\Pbb_{(S_j)_{j;X_j\in{\cal N}(X_i)}}(\exists z \neq y_i, z \in \cap_j S_j)
\leq m \Pbb_{S_j}(z \in S_j)^k
\leq m \eta^k = m\exp(-k\abs{\log(\eta)}).
\]
With $m = \card{\mathcal{Y}}$ the number of elements in $\mathcal{Y}$.
We deduce that
\[
\Pbb_{(S_i)}\paren{\phi(\hat{y}_i) \neq \phi(y^*_i) \,\middle\vert\, (X_i)}
\leq m \exp(-k\abs{\log(\eta)}).
\]
And because $\sum_{i=1}^n \ind{X_i \in {\cal N}(X)} = k$, we conclude that
\[
\E_{{\cal D}_n, X}\bracket{\norm{g_n^*(X) - g_n(X)}_{\cal H}\,\middle\vert\, (X_i)
\in \mathbb{D}} \leq 2c_\phi m \exp(-k\abs{\log(\eta)}).
\]
Finally, adding everything together we get
\[
{\cal E}(f_n) \leq 8 c_\phi c_\psi \exp\paren{-\frac{np}{8}}
+ 8 c_\phi c_\psi n \exp\paren{-\frac{(n-1)p}{8}}
+ 8 c_\phi c_\psi m \exp\paren{-k\abs{\log(\eta)}}.
\]
as long as $k < (n-1) p /2$, which implies Theorem \ref{df:thm:convergence} as long as $n \geq 2$.
\begin{remark}[Other approaches]
While we have proceeded with analysis based on local averaging methods, other paths could be explored to prove convergence results of the algorithm provided \eqref{df:eq:disambiguation} and \eqref{df:eq:estimate}.
For example, one could prove Wasserstein convergence of $\sum_{i=1}^n\delta_{(x_i, \hat{y}_i)}$ toward $\sum_{i=1}^n \delta_{(x_i, \hat{y}^*_i)}$, together with some continuity of the learning algorithm as a function of those distributions.\footnote{The Wasserstein metric is useful to think in terms of distributions, which is natural when considering partial supervision that can be cast as a set of admissible fully supervised distributions.
This approach has been successfully followed by \citet{Perchet2015} to deal with partial monitoring in games.}
This analysis could be understood as tripartite:
\begin{itemize}
\item A disambiguation error, comparing $\hat{y}_i$ to $y_i^*$.
\item A stability / robustness measure of the algorithm to learn $f_n$ from data
when substituting $y_i^*$ by $\hat{y}_i$.
\item A consistency result regarding $f_n^*$ learned with $(x_i, y_i^*)$.
\end{itemize}
Our analysis followed a similar path, yet with the first two parts tackled jointly.
\end{remark}
\subsection{Proof of Proposition \ref{df:prop:init}}
\label{df:proof:init}
Under the non-ambiguity hypothesis (Assumption \ref{df:ass:non-ambiguity}), the solution of \eqref{df:eq:solution} is characterized pointwise by $f^*(x) = y_x$ for all $x\in\supp\nu_\mathcal{X}$.
Similarly, under Assumption \ref{df:ass:non-ambiguity}, we have the characterization $f^*(x) \in \cap_{S\in\supp\nu\vert_x} S$.
With the notation of Definition \ref{df:def:init}, since $f^*(x)$ minimizes $z\to\E_{Y\sim\mu_S}[\ell(z, Y)]$ for all $S\in\supp\nu\vert_x$, it also minimizes $z\to\E_{S\sim\nu\vert_x}\E_{Y\sim\mu_S} [\ell(z, Y)]$.
For the second part of the proposition, we use the structured prediction framework of \citet{Ciliberto2020}.
Define the signed measure $\mu^\circ$ defined as $\mu^\circ_\mathcal{X} := \nu_\mathcal{X}$ and $\mu^\circ\vert_x := \E_{S\sim\nu\vert_x}\E_{Y\sim\mu_S}[\delta_Y]$, and $f^\circ:\mathcal{X}\to\mathcal{Y}$ the solution $f^\circ \in \argmin_{f:\mathcal{X}\to\mathcal{Y}}\E_{(X, Y)\sim\mu^\circ}[\ell(f(X), Y)] = \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X, Y)\sim\nu}\bracket{\E_{Y\sim\mu_S}[\ell(f(X), Y)]}$.
The first part of the proposition tells us that $f^\circ = f^*$ under Assumption \ref{df:ass:non-ambiguity}.
The framework of \citet{Ciliberto2020}, tells us that $f^\circ$ is obtained after decoding \eqref{df:eq:decoding} of $g^\circ:\mathcal{X}\to{\cal H}$, and that if $g_n^\circ$ converges to $g^\circ$ with the $L^1$ norm, $f_n^\circ$ converges to $f^\circ$ in terms of the $\mu^\circ$-risk.
Under Assumption \ref{df:ass:non-ambiguity} and mild hypothesis on $\mu^\circ$, it is possible to prove that convergence in terms of the $\mu^\circ$-risk implies convergence in terms of the $\mu$-risk (for example through calibration inequality similar to Proposition 2 of \citet{Cabannes2020}).
\subsection{Ranking with partial ordering is a well-behaved problem}
\label{df:proof:ranking}
Here, we discuss building directly $\xi_S$ to initialize our alternative minimization scheme or considering $\mu_S$ given by the definition of well-behaved problem (Definition \ref{df:def:init}).
Since the existence of $\mu_S$ implying $\xi_S$ defined as $\E_{Y\sim\mu_S}[\phi(Y)]$, we will only study when $\xi_S$ can be cast as a $\mu_S$.
In ranking, we have that $\psi = -\phi$, which corresponds to ``correlation losses''.
In this setting, we have that $\Span(\phi(\mathcal{Y})) = \Span(\psi(\mathcal{Y}))$.
More generally, looking at a ``minimal'' representation of $\ell$, one can always assume the equality of those spans, as what happens on the orthogonal of the intersection of those spans, does not modify the scalar product $\phi(y)^\top\psi(z)$.
Similarly, $\xi_S$ can be restricted to $\Span(\psi(\mathcal{Y}))$, and therefore $\Span(\phi(\mathcal{Y}))$, which exactly the image by $\mu\to\E_{Y\sim\mu}[\phi(Y)]$ of the set of signed measures, showing the existence of a $\mu_S$ matching Definition \ref{df:def:init}.
\section{Introduction}
In many applications of machine learning, such as recommender systems, where an input $x$ characterizing a user should be matched with a target $y$ representing an ordering of a large number $m$ of items, accessing fully supervised data $(x, y)$ is not an option.
Instead, one should expect weak information on the target $y$, which could be a list of previously taken (if items are online courses), watched (if items are plays), {\em etc.}, items by a user characterized by the feature vector $x$.
This motivates {\em weakly supervised learning}, aiming at learning a mapping from inputs to targets in such a setting where tools from supervised learning can not be applied off-the-shelves.
Recent applications of weakly supervised learning showcase impressive results in solving complex tasks such as action retrieval on instructional videos \citep{Miech2019}, image semantic segmentation \citep{Papandreou2015}, salient object detection \citep{Wang2017}, 3D pose estimation \citep{Dabral2018}, text-to-speech synthesis \citep{Jia2018}, to name a few.
However, those applications of weakly supervised learning are usually based on clever heuristics, and theoretical foundations of learning from weakly supervised data are scarce, especially when compared to statistical learning literature on supervised learning \citep{Vapnik1995,Boucheron2005,Steinwart2008}.
We aim to provide a step in this direction.
In this paper, we focus on partial labeling, a popular instance of weak supervision, approached with a structured prediction point of view \cite{Ciliberto2020}.
We detail this setup in Section \ref{df:sec:setup}.
Our contributions are organized as follows.
\begin{itemize}
\item In Section \ref{df:sec:algo}, we introduce a disambiguation algorithm to retrieve fully supervised samples from weakly supervised ones, before applying off-the-shelf supervised learning algorithms to the completed dataset.
\item In Section \ref{df:sec:stat}, we prove exponential convergence rates of our algorithm, in terms of the fully supervised excess of risk, given classical learnability assumptions.
\item In Section \ref{df:sec:optim}, we explain why disambiguation algorithms are intrinsically non-convex, and provide guidelines based on well-grounded heuristics to implement our algorithm.
\end{itemize}
We end this paper with a review of literature in Section \ref{df:sec:litterature}, before showcasing the usefulness of our method on practical examples in Section \ref{df:sec:example}, and opening on perspectives in Section \ref{df:sec:opening}.
\section{Disambiguation of partial labeling}
\label{df:sec:setup}
In this section, we review the supervised learning setup, introduce the partial labeling problem along with a principle to tackle this instance of weak supervision.
Algorithms can be formalized as mapping an input $x$ to a desired output $y$, respectively belonging to an input space~$\mathcal{X}$ and an output space $\mathcal{Y}$.
Machine learning consists in automating the design of the mapping $f:\mathcal{X}\to\mathcal{Y}$, based on a joint distribution $\mu \in \prob{\mathcal{X}\times\mathcal{Y}}$ over input/output pairings $(x, y)$ and a loss function $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$, measuring the error cost of outputting $f(x)$ when one should have output $y$.
The optimal mapping is defined as satisfying
\begin{equation}
\label{df:eq:sol_fs}
f^* \in \argmin_{f:\mathcal{X}\to \mathcal{Y}} \E_{(X, Y) \sim \mu} \bracket{\ell(f(X), Y)}.
\end{equation}
In \emph{supervised learning}, it is assumed that one does not have access to the full distribution $\mu$, but only to independent samples $(X_i, Y_i)_{i\leq n} \sim \mu^{\otimes n}$.
In practice, accessing such samples means building a dataset of examples.
While input data $(x_i)$ are usually easily accessible, getting output pairings $(y_i)$ generally requires careful annotation, which is both time-consuming and expensive.
For example, in image classification, $(x_i)$ can be collected by scrapping images over the Internet.
Subsequently, a ``data labeler'' might be asked to recognize a rare feline $y_i$ on an image $x_i$.
While getting $y_i$ will be hard in this setting, recognizing that it is a feline and describing elements of color and shape is easy, and already helps to determine what outputs $f(x_i)$ are acceptable.
A second example is given when pooling a known population $(x_i)$ to get estimation of their political orientation $(y_i)$, one might get information from recent election of percentage of voters across the political landscape, leading to global constraints that $(y_i)$ should verify.
A supervision that gives information on $(y_i)_{i\leq n}$ without giving its precise value is called \emph{weak supervision}.
\emph{Partial labeling}, also known as ``superset learning'', is an instance of weak supervision, in which, for an input $x$, we do not access the precise label $y$ but only a set $s$ of potential labels, $y\in s\subset \mathcal{Y}$.
For example, on a caracal image $x$, one might not get the label ``caracal'' $y$, but the set $s$ ``feline'', containing all the labels $y$ corresponding to felines.
It is modelled through a distribution $\nu \in \prob{\mathcal{X}\times 2^\mathcal{Y}}$ over $\mathcal{X}\times 2^\mathcal{Y}$ generating samples $(X, S)$, which should be compatible with the fully supervised distribution $\mu \in \prob{\mathcal{X}\times\mathcal{Y}}$ as formalized by the following definition.
\begin{definition}[Compatibility, \citet{Cabannes2020}]
\label{df:def:compatibility}
A fully supervised distribution $\mu \in \prob{\mathcal{X}\times\mathcal{Y}}$ is \emph{compatible} with a weakly supervised distribution $\nu \in \prob{\mathcal{X}\times\mathcal{Y}}$, denoted by $\mu\vdash\nu$ if there exists an underlying distribution $\pi \in \prob{\mathcal{X}\times\mathcal{Y}\times 2^\mathcal{Y}}$, such that $\mu$, and $\nu$, are the respective marginal distributions of $\pi$ over $\mathcal{X}\times \mathcal{Y}$ and $\mathcal{X}\times 2^\mathcal{Y}$, and such that $y\in s$ for any tuple $(x, y, s)$ in the support of $\pi$ (or equivalently $\pi\vert_{s} \in \prob{s}$, with $\pi\vert_{s}$ denoting the conditional distribution of $\pi$ given $s$).
\end{definition}
This definition means that a weakly supervised sample $(X, S) \sim \nu$ can be thought as proceeding from a fully supervised sample $(X, Y) \sim \mu$ after losing information on $Y$ according to the sampling of $S\sim \pi\vert_{X, Y}$.
The goal of partial labeling is still to learn $f^*$ from \eqref{df:eq:sol_fs}, yet without accessing a fully supervised distribution $\mu \in \prob{\mathcal{X}\times\mathcal{Y}}$ but only the weakly supervised distribution $\nu \in \prob{\mathcal{X}\times 2^\mathcal{Y}}$.
As such, this is an ill-posed problem, since $\nu$ does not discriminate between all $\mu$ compatible with it.
Following {\em lex parsimoniae}, \citet{Cabannes2020} have suggested looking for $\mu$ such that the labels are the most deterministic function of the inputs, which they measure with a loss-based ``variance'', leading to the disambiguation
\begin{equation}
\label{df:eq:principle}
\mu^* \in \argmin_{\mu\vdash\nu} \inf_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X, Y)\sim\mu}\bracket{\ell(f(X), Y)},
\end{equation}
and to the definition of the optimal mapping $f^*:\mathcal{X}\to\mathcal{Y}$
\begin{equation}
\label{df:eq:solution}
f^* \in \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X, Y)\sim\mu^*}\bracket{\ell(f(X), Y)}.
\end{equation}
This principle is motivated by Theorem 1 of \citet{Cabannes2020} showing that $f^*$ in \eqref{df:eq:solution} is characterized by $f^* \in \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X, S)\sim\nu}\bracket{\inf_{y\in S} \ell(f(X), y)}$, matching a prior formulation based on infimum loss \citep{Cour2011,Luo2010,Hullermeier2014}.
In practice, it means that if $(S\vert X=x)$ has probability 50\% to be the set ``feline'' and 50\% the set ``orange with black stripes'', $(Y\vert X=x)$ should be considered as 100\% ``tiger'', rather than 20\% ``cat'', 30\% ``lion'' and 50\% ``orange car with black stripes'', which could also explain $(S\vert X=x)$.
Similarly to supervised learning, partial labeling consists in retrieving $f^*$ without accessing $\nu$ but only samples $(X_i, S_i)_{i\leq n} \sim \nu^{\otimes n}$.
\begin{remark}[Measure of determinism]
\label{df:rmk:determinism}
\eqref{df:eq:principle} is not the only variational way to push toward distribution where labels are deterministic function of the inputs.
For example, one could minimize entropy \citep[{\em e.g.},][]{Berthelot2019,Lienen2021}.
However, a loss-based principle is appreciable since the loss usually encodes structures of the output space \citep{Ciliberto2020}, which will allow sample and computational complexity of consequent algorithms to scale with an intrinsic dimension of the space rather than the real one, {\em e.g.}, $m$ rather than $m!$ when $\mathcal{Y} = \mathfrak{S}_m$ and $\ell$ is a suitable ranking loss \citep[see Section~\ref{df:sec:ranking} or ][]{Nowak2019}.
\end{remark}
\section{Learning algorithm}
\label{df:sec:algo}
In this section, given weakly supervised samples, we present a disambiguation algorithm to retrieve fully supervised samples based on an empirical expression of \eqref{df:eq:principle}, before learning a mapping from $\mathcal{X}$ to $\mathcal{Y}$ based on those fully supervised samples, according to \eqref{df:eq:solution}.
Given a partially labeled dataset ${\cal D}_n = (x_i, s_i)_{i\leq n}$, sampled accordingly to $\nu^{\otimes n}$, we retrieve fully supervised samples, based on the following empirical version of \eqref{df:eq:principle}, with $C_n = \prod_{i\leq n} s_i \subset \mathcal{Y}^n$
\begin{equation}
\label{df:eq:disambiguation}
(\hat{y}_i)_{i\leq n} \in \argmin_{(y_i)_{i\leq n} \in C_n} \inf_{(z_i)_{i\leq n} \in \mathcal{Y}^n} \sum_{i, j=1}^n \alpha_j(x_i) \ell(z_i, y_j),
\end{equation}
where $(\alpha_i(x))_{i\leq n}$ is a set of weights measuring how much one should base its prediction for $x$ on the observations made at $x_i$.
This formulation is motivated by the Bayes approximate rule proposed by \citet{Stone1977}, which can be seen as the approximation of $\mu$ by $n^{-1}\sum_{i,j=1}^n \alpha_j(x_i)\delta_{x_i} \otimes \delta_{y_j}$ in \eqref{df:eq:principle}.
Once fully supervised samples $(x_i, \hat{y_i})$ have been recollected, one can learn $f_n:\mathcal{X}\to\mathcal{Y}$, approximating $f^*$, with classical supervised learning techniques.
In this work, we will consider the structured prediction estimator introduced by \citet{Ciliberto2016}, defined as
\begin{equation}
\label{df:eq:estimate}
f_n(x) \in \argmin_{z\in \mathcal{Y}} \sum_{i=1}^n \alpha_i(x) \ell(z, \hat{y}_i).
\end{equation}
\paragraph{Weighting scheme $\alpha$.}
For the weighting scheme $\alpha$, several choices are appealing.
Laplacian diffusion is one of them as it incorporates a prior on low density separation to boost learning \citep{Zhu2003,Zhou2003,Bengio2006,Hein2007}.
Kernel ridge regression is another due to its theoretical guarantees \citep{Ciliberto2020}.
In the theoretical analysis, we will use nearest neighbors.
Assuming $\mathcal{X}$ is endowed with a distance $d$, and assuming, for readability’ sake, that ties to define nearest neighbors do not happen, it is defined as
\[
\alpha_i(x) = \left\{
\begin{array}{cc}
k^{-1} & \text{if}\quad \sum_{j=1}^n \ind{d(x, x_j) \leq d(x, x_i)} \leq k \\
0 & \text{otherwise},
\end{array}\right.
\]
where $k$ is a parameter fixing the number of neighbors.
Our analysis, leading to Theorem \ref{df:thm:convergence}, also holds for other local averaging methods such as partitioning or Nadaraya-Watson estimators.
\section{Consistency result}
\label{df:sec:stat}
In this section, we assume $\mathcal{Y}$ finite, and prove the convergence of $f_n$ toward $f^*$ as $n$, the number of samples, grows to infinity.
To derive such a consistency result, we introduce a surrogate problem that we relate to the risk through a calibration inequality.
We later assume that weights are given by nearest neighbors and review classical assumptions, that we work to derive exponential convergence rates.
In the following, we are interested in bounding the expected generalization error, defined as
\begin{equation}
\label{df:eq:excess}
{\cal E}(f_n) = \E_{{\cal D}_n} {\cal R}(f_n) - {\cal R}(f^*),
\end{equation}
where
\(
{\cal R}(f) = \E_{(X, Y)\sim\mu^*}\bracket{\ell(f(X), Y)},
\)
by a quantity that goes to zero, when $n$ goes to infinity.
This implies, under boundedness of $\ell$, convergence in probability (the randomness being inherited from ${\cal D}_n$) of ${\cal R}(f_n)$ toward $\inf_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}(f)$, which is referred as \emph{consistency} of the learning algorithm.\footnote{If $\E|X| < +\infty$ and $\E[\abs{X_n - X}]\to 0$, $X_n \to X$ in probability.}
We first introduce a few objects.
\paragraph{Disambiguation ground truth \texorpdfstring{$(y_i^*)$}{}.}
Introduce $\pi^* \in \prob{\mathcal{X}\times\mathcal{Y}\times 2^\mathcal{Y}}$ expressing the compatibility of $\mu^*$ and $\nu$ as in Definition \ref{df:def:compatibility}.
Given samples $(x_i, s_i)_{i\leq n}$ forming a dataset ${\cal D}_n$, we enrich this dataset by sampling $y_i^* \sim \pi^*\vert_{x_i, s_i}$, which build an underlying dataset $(x_i, y_i^*, s_i)$ sampled accordingly $(\pi^*)^{\otimes n}$.
Given ${\cal D}_n$, while {\em a priori}, $y_i^*$ are random variables, sampled accordingly to $\pi^*\vert_{x_i, s_i}$, because of the definition of $\mu^*$ \eqref{df:eq:principle}, under basic definition assumptions, they are actually deterministic, defined as $y_i^* = \argmin_{y\in s_i} \ell(f^*(x_i), y)$.
As such, they should be seen as ground truth for $\hat{y}_i$.
\paragraph{Surrogate estimates.}
The approximate Bayes rule was successfully analyzed recently through the prism of plug-in estimators by \citet{Ciliberto2020}.
While we will not cast our algorithm as a plug-in estimator, we will leverage this surrogate approach, introducing two mappings $\phi$ and $\psi$ from~$\mathcal{Y}$ to a Hilbert space ${\cal H}$ such that
\begin{equation}
\label{df:eq:loss}
\forall\, z, y \in \mathcal{Y}, \qquad \ell(z, y) = \scap{\psi(z)}{\phi(y)},
\end{equation}
Such mappings always exist when $\mathcal{Y}$ is finite, and have been used to encode ``problem structure'' defined by the loss $\ell$ \citep{Nowak2019}.
We introduce three surrogate quantities that will play a major role in the following analysis, they map $\mathcal{X}$ to ${\cal H}$ as
\begin{gather}
\label{df:eq:surrogate}
g^*(x) = \E_{\mu^*}\bracket{\phi(Y)\,\middle\vert\, X=x},\qquad
g_n(x) = \sum_{i=1}^n \alpha_i(x) \phi(\hat{y}_i),\nonumber\\
g_n^*(x) = \sum_{i=1}^n \alpha_i(x) \phi(y^*_i).
\end{gather}
It is known that $f^*$ and $f_n$ are retrieved from $g^*$ and $g_n$, through the decoding, retrieving $f:\mathcal{X}\to\mathcal{Y}$ from $g:\mathcal{X}\to{\cal H}$ as
\begin{equation}
\label{df:eq:decoding}
f(x) = \argmin_{z\in\mathcal{Y}} \scap{\psi(z)}{g(x)},
\end{equation}
which explains the wording of {\em plug-in} estimator \citep{Ciliberto2020}.
We now introduce a {\em calibration inequality}, that relates the error between $f_n$ and $f^*$ with surrogate error quantities.
\begin{lemma}[Calibration inequality]\label{df:lem:cal}
When $\mathcal{Y}$ is finite, and the labels are a deterministic function of the input, {\em i.e.}, when $\mu^*\vert_x$ is a Dirac for all $x\in\supp\nu_\mathcal{X}$, for any weighting scheme such that $\sum_{i=1}^n \abs{\alpha_i(x)} \leq 1$ for all $x\in\supp\nu_\mathcal{X}$,
\begin{align}
\label{df:eq:cal}
& {\cal R}(f_n) - {\cal R}(f^*) \leq 4c_\psi \norm{g_n^* - g_n}_{L^1} \hspace*{2cm}\nonumber \\
& \hspace*{1cm} \qquad+
8c_\psi c_\phi \Pbb_X\paren{\norm{g_n^*(X) - g^*(X)} > \delta},
\end{align}
with $c_\psi = \sup_{z\in\mathcal{Y}} \norm{\psi(z)}$, $c_\phi = \sup_{z\in\mathcal{Y}}\norm{\phi(y)}$, and $\delta$ a parameter that depend on the geometry of $\ell$ and its decomposition through $\phi$.
\end{lemma}
This lemma, proven in Appendix \ref{df:proof:cal}, separates a part reading in $\norm{g_n - g_n^*}$, due to the {\em disambiguation error} between $(\hat{y}_i)$ and $(y_i^*)$ together with the {\em stability} of the learning algorithm when substituting $(\hat{y}_i)$ for $(y_i^*)$, and a part in $\norm{g_n^* - g^*}$ due to the {\em consistency} of the fully supervised learning algorithm.
The expression of the first part relates to Theorem 7 in \citet{Ciliberto2020} while the second part relates to Theorem 6 in \citet{Cabannes2021b}.
\subsection{Classical learnability assumptions}
In the following, we suppose that the weights $\alpha$ are given by nearest neighbors, that $\mathcal{X}$ is a compact metric space endowed with a distance $d$, that $\mathcal{Y}$ is finite and that $\ell$ is proper in the sense that it strictly positive except on the diagonal of $\mathcal{Y}\times\mathcal{Y}$ diagonal where it is zero.
We now review classical assumptions to prove consistency.
First, assume that $\nu_\mathcal{X}$ is regular in the following sense.
\begin{assumption}[$\nu_\mathcal{X}$ well-behaved]\label{df:ass:mass}
Assume that $\nu_\mathcal{X}$ is such that there exists $h_1, c_\mu, q > 0$ satisfying, with ${\cal B}$ designing balls in $\mathcal{X}$,
\[
\forall\, x\in\supp\nu_\mathcal{X},\ \forall\, r < h_1,
\qquad \nu_\mathcal{X}({\cal B}(x, r)) > c_\mu r^q.
\]
\end{assumption}
Assumption \ref{df:ass:mass} is useful to make sure that neighbors in ${\cal D}_n$ are closed with respect to the distance $d$, it is usually derived by assuming that $\mathcal{X}$ is a subset of $\mathbb{R}^q$; that $\nu_\mathcal{X}$ has a density $p$ against the Lebesgue measure $\lambda$ with {\em minimal mass} $p_{\min}$ in the sense that for any $x\in\supp\nu_\mathcal{X}$, $p(x) > p_{\min}$; and that $\supp\nu_\mathcal{X}$ has regular boundary in the sense that $\lambda({\cal B}(x, r) \cap \supp\nu_\mathcal{X}) \geq c\lambda({\cal B}(x, r)$ for any $x\in\supp\nu_\mathcal{X}$ and $r < h$ \citep[{\em e.g.},][]{Audibert2007}.
We now switch to a classical assumption in partial labeling, allowing for population disambiguation.
\begin{assumption}[Non ambiguity, \citet{Cour2011}]
\label{df:ass:non-ambiguity}
Assume the existence of $\eta \in [0, 1)$, such that for any $x\in\supp\nu_\mathcal{X}$, there exists $y_x \in \mathcal{Y}$, such that $\Pbb_{\nu}\paren{y_x \in S \,\middle\vert\, X=x} = 1$, and
\[
\forall\, z\neq y_x,\quad
\Pbb_{\nu}\paren{z \in S \,\middle\vert\, X=x} \leq \eta.
\]
\end{assumption}
Assumption \ref{df:ass:non-ambiguity} states that when given the full distribution $\nu$, there is one, and only one, label that is coherent with every observable sets for a given input.
It is a classical assumption in literature about the learnability of the partial labeling problem \citep[{\em e.g.},][]{Liu2014}.
When $\ell$ is proper, this implies that $\mu^*\vert_x = \delta_{y_x}$, and $f^*(x) = y_x$.
Finally, we assume that $g^*$ is regular.
As we are considering local averaging method, we will use Lipschitz-continuity, which is classical in such a setting.\footnote{Its generalization through H\"older-continuity would work too.}
\begin{assumption}[Regularity of $g^*$]\label{df:ass:lipschitz}
Assume that there exists $c_g > 0$, such that for any $x, x'\in\mathcal{X}$, we have
\[
\norm{g^*(x) - g^*(x')}_{\cal H} \leq c_g d(x, x').
\]
\end{assumption}
It should be noted that regularity of $g^*$, Assumption \ref{df:ass:lipschitz}, together with determinism of $\mu^*\vert_x$ inherited from Assumption~\ref{df:ass:non-ambiguity} implies that classes $\mathcal{X}_y = \brace{x\,\middle\vert\, f^*(x) = y}$ are separated in $\mathcal{X}$, in the sense that there exists $h_2 > 0$, such that for any $y, y' \in \mathcal{Y}$ and $(x, x') \in \mathcal{X}_y \times \mathcal{X}_{y'}$, $d(x, x') > h_2$, which is a classical assumption to derive consistency of semi-supervised learning algorithm \citep[{\em e.g.},][]{Rigollet2007}.
We detailed those implications in Appendix~\ref{df:proof:ass}.
\subsection{Exponential convergence rates}
We are now ready to state our convergence result.
We introduce $h = \min(h_1, h_2)$ and $p = c_\mu h^q$, so that for any $x \in \supp\nu_\mathcal{X}$, $\nu_\mathcal{X}({\cal B}(x, h)) > p$.
\begin{theorem}[Exponential convergence rates]
\label{df:thm:convergence}
When the weights $\alpha$ are given by nearest neighbors, under Assumptions \ref{df:ass:mass}, \ref{df:ass:non-ambiguity} and \ref{df:ass:lipschitz}, the excess of risk in~\eqref{df:eq:excess} is bounded by
\begin{align}
\label{df:eq:convergence}
{\cal E}(f_n) & \leq 8c_\psi c_\phi (n+1)\exp\paren{-{\frac{np}{16}}} \hspace*{2cm} \nonumber \\
& \hspace*{1cm} \qquad + 8c_\psi c_\phi m \exp\paren{-k\abs{\log(\eta)}},
\end{align}
as soon as $k < np / 4$, with $m = \card{\mathcal{Y}}$.
By taking $k_n = k_0 n$, for $k_0 < p/4$, this implies exponential convergence rates ${\cal E}(f_n) = O(n\exp(-n))$.
\end{theorem}
\begin{proof}[Sketch for Theorem \ref{df:thm:convergence}]
In essence, based on Lemma \ref{df:lem:cal}, Theorem \ref{df:thm:convergence} can be understood as two folds.
\begin{itemize}
\item A fully supervised error between $g_n^*$ and $g^*$.
This error can be controlled in $\exp(-np)$ as the non-ambiguity assumption implies a hard Tsybakov margin condition, a setting in which {\em the fully supervised estimate $g_n^*$ is known to converge to the population solution $g^*$ with such rates} \citep{Cabannes2021b}.
\item A weakly disambiguation error, that is exponential too, since, based on Assumption \ref{df:ass:non-ambiguity}, disambiguating between $z \in \mathcal{Y}$ and $y_x$ from $k$ sets $S$ sampled accordingly to $\nu\vert_x$ can be done in $\eta^k$, and disambiguating between all $z\neq y_x$ and $y_x$ in $m\eta^k = m \exp(-k\abs{\log(\eta)})$.
\end{itemize}
Appendix~\ref{df:proof:convergence} provides details.
\end{proof}
Theorem \ref{df:thm:convergence} states that under a non-ambiguity assumption and a regularity assumption implying no-density separation, one can expect exponential convergence rates of $f_n$ learned with weakly supervised data to $f^*$ the solution of the fully supervised learning problem, measured with excess of fully supervised risk.
Because of exponential convergence rates, we could expect polynomial convergence rates for a broader class of problems that are approximated by problems satisfying assumptions of Theorem \ref{df:thm:convergence}.
{\em The derived rates in $n\exp(-n)$ should be compared with rates in $n^{-1/2}$ and $n^{-1/4}$}, respectively derived, {\em under the same assumptions}, by \citet{Cour2011,Cabannes2020}.
\subsection{Discussion on assumptions}
While we have retaken classical assumptions from literature, those assumptions are quite strong, which allows us, by understanding their strength, to derive exponential convergence rates.
Assumptions \ref{df:ass:mass} and \ref{df:ass:lipschitz} are classical in the nearest neighbor literature with full supervision.
If we were using (reproducing) kernel methods to define the weighting scheme $\alpha$, those assumptions would be mainly replaced with ``$g^*$ belonging to the RKHS''.
Assumption \ref{df:ass:non-ambiguity} is the strongest assumption in our view, that we will now discuss.
\paragraph{How to check it in practice ?}
First, for Assumption \ref{df:ass:non-ambiguity} to hold, the labels have to be a deterministic function of the inputs.
In other words, a zero error is achievable.
Finally, Assumption \ref{df:ass:non-ambiguity} is related to dataset collection.
If dealing with images, weak supervision could take the form of some information on shape, color, or texture, etc., Assumption \ref{df:ass:non-ambiguity} supposes that the weak information potentially given on a specific image~$x$ allows retrieving the unique label $y$ of the image ({\em e.g.}, a “pig” could be recognized from its shape and its color).
This is a reasonable assumption, if, for a given $x$, we ask at random a data labeler to provide us information on shape, color, or texture, etc.
However, it will not be the case, if for some reasons ({\em e.g.} the dataset is built from several weakly annotated datasets), in some regions of the input space, we only get shape information, and in other regions, we only get color information.
In particular, it is not verified for semi-supervised learning when the support of the unlabeled data distribution is not the same as the support of the labeled input data distribution.
\paragraph{How to relax it and what results to expect?}
Previous works used Assumption \ref{df:ass:non-ambiguity} to derive a calibration inequality between the infimum loss to the original loss \citep[{\em e.g.}, see Proposition 2 by][]{Cabannes2020}.
In contrast, we relate the surrogate and original problem through a refined calibration inequality \eqref{df:eq:cal}.
This technical progress allows us to derive exponential convergence rates similarly to the work of \citet{Cabannes2021b}.
Importantly, in comparison with previous work, our calibration inequality Lemma \ref{df:lem:cal} can easily be extended without the determinism assumption provided by Assumption \ref{df:ass:non-ambiguity}.
Essentially, in our work, Assumption \ref{df:ass:non-ambiguity} is used to simplify the study of $(\hat{y}_i)_{i\leq n}$ given by the disambiguation algorithm \eqref{df:eq:disambiguation}, and therefore the study of the disambiguation error in~\eqref{df:eq:cal}.
The study of $(\hat{y}_i)_{i\leq n}$ without Assumption \ref{df:ass:non-ambiguity} would require other tools than the one presented in this paper.
It could be studied in the realm of graphical model and message passing algorithm, or with Wasserstein distance and topological considerations on measures.
With much milder forms of Assumption \ref{df:ass:non-ambiguity}, we expect the rates to degrade smoothly with respect to a parameter defining the hardness of the problem, similarly to the works of \citet{Audibert2007,Cabannes2021b}.
\section{Optimization considerations}
\label{df:sec:optim}
In this section, we focus on implementations to solve~\eqref{df:eq:disambiguation}.
We explain why disambiguation objectives, such as~\eqref{df:eq:principle} are intrinsically non-convex and express a heuristic strategy to solve~\eqref{df:eq:disambiguation} besides non-convexity in classical well-behaved instances of partial labeling.
Note that we do not study implementations to solve~\eqref{df:eq:estimate} as this study has already been done by \citet{Nowak2019}.
We end this section by considering a practical example to make derivations more concrete.
\subsection{Non-convexity of disambiguation objectives}
For readability, suppose that $\mathcal{X}$ is a singleton, justifying to remove the dependency on the input in the following.
Consider $\nu \in \prob{2^\mathcal{Y}}$ a distribution modelling weak supervision.
While the domain $\brace{\mu\in\prob{\mathcal{Y}} \,\middle\vert\, \mu \vdash \nu}$ is convex, a disambiguation objective ${\cal E}:\prob{\mathcal{Y}} \to \mathbb{R}$ defining $\mu^* \in \argmin_{\mu\vdash\nu} {\cal E}(\mu)$, similarly to~\eqref{df:eq:principle}, that is minimized for deterministic distributions, which correspond to $\mu$ a Dirac, {\em i.e.}, minimized on vertices of its definition domain $\prob{\mathcal{Y}}$, can not be convex.
In other terms, any disambiguation objective that pushes toward distributions where targets are deterministic function of the input, as mentioned in Remark \ref{df:rmk:determinism}, can not be convex.
Indeed, smooth disambiguation objectives such as entropy and our piecewise linear loss-based principle~\eqref{df:eq:principle}, reading pointwise
\(
{\cal E}(\mu) = \inf_{z\in\mathcal{Y}} \E_{Y\sim\mu}[\ell(z, Y)],
\)
are concave.
Similarly, its quadratic variant
\(
{\cal E}'(\mu) = \E_{Y, Y'\sim\mu}[\ell(Y, Y')],
\)
is concave as soon as $(\ell(y, y'))_{y, y'\in\mathcal{Y}}$ is semi-definite negative.
We illustrate those considerations on a concrete example with graphical illustration in Appendix \ref{df:app:example}.
We should see how this translates on generic implementations to solve the empirical objective~\eqref{df:eq:disambiguation}.
\subsection{Generic implementation for~\eqref{df:eq:disambiguation}}
Depending on $\ell$ and on the type of observed set $(s_i)$, \eqref{df:eq:disambiguation}~might be easy to solve.
In the following, however, we will introduce optimization considerations to solve it in a generic structured prediction fashion.
To do so, we recall the decomposition of $\ell$~\eqref{df:eq:loss} and rewrite~\eqref{df:eq:disambiguation} as
\begin{equation*}
(\hat{y}_i)_{i\leq n} \in \argmin_{y_i \in C_n} \inf_{(z_i) \in \mathcal{Y}^n} \sum_{i,j=1}^n \alpha_j(x_i) \psi(z_i)^\top \phi(y_j).
\end{equation*}
Since, given $(y_j)$, the objective is linear in $\psi(z_j)$, the constraint $\psi(z_j) \in \psi(\mathcal{Y})$ can be relaxed with $\zeta_i \in \hull\psi(\mathcal{Y})$.\footnote{The minimization pushes toward extreme points of the definition domain.}
Similarly, with respect to $\phi(y_j)$, this objective is the infimum of linear functions, therefore is concave, and the constraint $\phi(y_j) \in \phi(s_j)$, could be relaxed with $\xi_i \in \hull\phi(s_j)$.
Hence, with ${\cal H}_0 = \hull\psi(\mathcal{Y})$ and $\Gamma_n = \prod_{j\leq n} \hull\phi(s_j)$, the optimization is cast as
\begin{equation}
\label{df:eq:algo}
(\hat{\xi}_i)_{i\leq n} \in \argmin_{(\xi_i) \in \Gamma_n} \inf_{(\zeta_i) \in{\cal H}_0^n} \sum_{i,j=1}^n \alpha_j(x_i) \zeta_i^\top \xi_j.
\end{equation}
Because of concavity, $(\hat{\xi}_i)$ will be an extreme point of $\Gamma_n$, that could be decoded into $\hat{y}_i = \phi^{-1}(\hat{\xi}_i)$.
However, it should be noted that if only interested in $f_n$ and not in the disambiguation $(\hat{y}_i)$, this decoding can be avoided, since~\eqref{df:eq:estimate} can be rewritten as $f_n(x) \in \argmin_{z\in\mathcal{Y}} \psi(z)^\top \sum_{i=1}^n \alpha_i(x) \hat{\xi}_i$.
\subsection{Alternative minimization with good initialization}
To solve~\eqref{df:eq:algo}, we suggest using an alternative minimization scheme.
The output of such a scheme is highly dependent to the variable initialization.
In the following, we introduce well-behaved problem, where $(\xi_i)_{i\leq n}$ can be initialized smartly, leading to an efficient implementation to solve~\eqref{df:eq:algo}.
\begin{definition}[Well-behaved partial labeling problem]
\label{df:def:init}
A partial labeling problem $(\ell, \nu)$ is said to be well-behaved if for any $s \in \supp\nu_{2^\mathcal{Y}}$, there exists a signed measure $\mu_s$ on $\mathcal{Y}$ such that the function from $\mathcal{Y}$ to $\mathbb{R}$ defined as $z\to \int_{\mathcal{Y}} \ell(z, y) \mathop{}\!\mathrm{d}\mu_s(y)$ is minimized for, and only for, $z \in s$.
\end{definition}
We provide a real-world example of a well-behaved problem in Section \ref{df:sec:ranking} as well as a synthetic example with graphical illustration in Appendix \ref{df:app:example}.
On those problems, we suggest solving~\eqref{df:eq:algo} by considering the initialization $\xi^{(0)}_i = \E_{Y\sim\mu_{s_i}}[\phi(Y)]$, and performing alternative minimization of~\eqref{df:eq:algo}, until attaining $\xi^{(\infty)}$ as the limit of the alternative minimization scheme (which exists since each step decreases the value of the objective in~\eqref{df:eq:algo} and there is a finite number of candidates for $(\xi_i)$).
It corresponds to a disambiguation guess $\tilde{y}_i = \phi^{-1}(\xi_i^{(\infty)})$.
Then we suggest learning $\hat{f}_n$ from $(x_i, \tilde{y}_i)$ based on~\eqref{df:eq:estimate}, and existing algorithmic tools for this problem \citep{Nowak2019}.
To assert the well-foundedness of this heuristic, we refer to the following proposition, proven in Appendix \ref{df:proof:init}.
\begin{proposition}
\label{df:prop:init}
Under the non-ambiguity hypothesis, Assumption \ref{df:ass:non-ambiguity}, the solution of~\eqref{df:eq:solution} is characterized by
\(
f^* \in \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X, S) \sim \nu}\bracket{\E_{Y\sim\mu_S}[\ell(f(X), Y)]}.
\)
Moreover, if the surrogate function $g_n^{\circ}:\mathcal{X}\to{\cal H}$ defined as $g_n^{\circ}(x) = \sum_{i=1}^n\alpha_i(x)\xi_{s_i}$, with $\xi_s = \E_{Y\sim\mu_s}[\phi(Y)]$, converges toward $g^\circ(x) = \E_{S\sim\nu\vert_x}[\xi_S]$ in $L^1$, $f_n^{\circ}$ defined through the decoding~\eqref{df:eq:decoding} converges in risk toward $f^*$.
\end{proposition}
Given that our algorithm scheme is initialized for $\xi_i^{(0)} = \xi_{s_i}$ and $\zeta_i^{(0)} = f_n^{\circ}(x_i)$ and stopped once having attained $\xi_i^{(\infty)}$ and $\zeta_i^{(\infty)} = \hat{f}_n(x_i)$, $\hat{f}_n$ is arguably better than $f_n^{\circ}$, which given consistency result exposed in Proposition \ref{df:prop:init}, is already good enough.
\begin{remark}[IQP implementation for~\eqref{df:eq:disambiguation}]
Other heuristics to solve~\eqref{df:eq:disambiguation} are conceivable.
For example, considering $z_i = y_i$ in this equation, we remark that the resulting problem is isomorphic to an integer quadratic program (IQP).
Similarly to integer linear programming, this problem can be approached with relaxation of the ``integer constraint'' to get a real-valued solution, before ``thresholding'' it to recover an integer solution.
This heuristic can be seen as a generalization of the Diffrac algorithm \citep{Bach2007,Joulin2010}.
We present it in details in Appendix~\ref{df:app:diffrac}.
\end{remark}
\begin{remark}[Link with EM, \citep{Dempster1977}]
Arguably, our alternative minimization scheme, optimizing respectively the targets $\xi_i = \phi(y_i)$ and the function estimates $\zeta_i = \psi(f_n(x_i))$ can be seen as the non-parametric version of the Expectation-Maximization algorithm, popular for parametric model \citep{Dempster1977}.
\end{remark}
\subsection{Application: ranking with partial ordering}
\label{df:sec:ranking}
Ranking is a problem consisting, for an input $x$ in an input space $\mathcal{X}$, to learn a total ordering $y$, belonging to $\mathcal{Y} = \mathfrak{S}_m$, modelling preference over $m$ items.
It is usually approached with the Kendall loss $\ell(y, z) = - \phi(y)^\top \phi(z)$, with $\phi(y) = (\sign\paren{y(i) - y(j)})_{i,j\leq m} \in \brace{-1,1}^{m^2}$ \citep{Kendall1938}.
Full supervision corresponds, for a given $x$, to be given a total ordering of the $m$ items.
This is usually not an option, but one could expect to be given partial ordering that $y$ should follow \citep{Cao2007,Hullermeier2008,Korba2018}.
Formally, this equates to the observation of some, but not all, coordinates $\phi(y)_i$ of the vector $\phi(y)$ for some $i\in I \subset \bbracket{1, m}^2$.
In this setting, $s\subset \mathcal{Y}$ is a set of total orderings that match the given partial ordering.
It can be represented by a vector $\xi_s \in {\cal H}$, that satisfies the partial ordering observation, $(\xi_s)_I = \phi(y)_I$, and that is agnostic on unobserved coordinates, $(\xi_s)_{^c I} = 0$.
This vector satisfies that $z\to\psi(z)^\top\xi_s$ is minimized for, and only for, $z\in s$.
Hence, it constitutes a good initialization for the alternative minimization scheme detailed above.
We provide details in Appendix \ref{df:proof:ranking}, where we also show that $\xi_s$ can be formally translated in a $\mu_s$ to match the Definition \ref{df:def:init}, proving that ranking with partial labeling is a well-behaved problem.
Many real world problems can be formalized as a ranking problem with partial ordering observations.
For example, $x$ could be a social network user, and the $m$ items could be posts of her connection that the network would like to order on her feed accordingly to her preferences.
One might be told that the user $x$ prefer posts from her close rather than from her distant connections, which translates formally as the constraint that for any $i$ corresponding to a post of a close connection and $j$ corresponding to a post of a distant connection, we have $\phi(y)_{ij} = 1$.
Nonetheless, designing non-parametric structured prediction models that scale well when the intrinsic dimension $m$ of the space $\mathcal{Y}$ is very large (such as the number of post on a social network) remains an open problem, that this paper does not tackle.
\section{Related work}
\label{df:sec:litterature}
Weakly supervised learning has been approached through parametric and non-parametric methods.
Parametric models are usually optimized through maximum likelihood \citep{Heitjan1991,Jin2002}.
\citet{Hullermeier2014} show that this approach, as formalized by \citet{Denoeux2013}, equates to disambiguating sets by averaging candidates, which was shown inconsistent by \citet{Cabannes2020} when data are {\em not missing at random}.
Among non-parametric models, \citet{Xu2004,Bach2007} developed an algorithm for clustering, that has been cast for weakly supervised learning problem \citep{Joulin2010,Alayrac2016}, leading to a disambiguation algorithm similar than ours, yet without consistency results.
More recently, half-way between theory and practice, \citet{Gong2018} derived an algorithm geared toward classification, based on a disambiguation objective, incorporating several heuristics, such as class separation, and Laplacian diffusion.
Those heuristics could be incorporated formally in our model.
The infimum loss principle has been considered by several authors, among them \citet{Cour2011,Luo2010,Hullermeier2014}.
It was recently analyzed through the prism of structured prediction by \citet{Cabannes2020}, leading to a consistent non-parametric algorithm that will constitute the baseline of our experimental comparison.
This principle is interesting as it does not assume knowledge on the corruption process $(S\vert Y)$ contrarily to the work of \citet{CidSueiro2014} or \citet{vanRooyen2018}.
The non-ambiguity assumption has been introduced by \citet{Cour2011} and is a classical assumption of learning with partial labeling \citep{Liu2014}.
Assumptions of Lipschitzness and minimal mass are classical assumptions to prove convergence of local averaging method \citep{Audibert2007,Biau2015}.
Those assumptions imply class separation in $\mathcal{X}$, which has been leverage in semi-supervised learning, justifying Laplacian regularization \citep{Rigollet2007,Zhu2003}.
Note that those assumptions might not hold on raw representation of the data, but with appropriate metrics, which could be learned through unsupervised \cite{Duda2000} or self-supervised learning \cite{Doersch2017}.
As such, the practitioner might consider weights $\alpha$ given by similarity metrics derived through such techniques, before computing the disambiguation~\eqref{df:eq:disambiguation} and learning $f_n$ from the recollected fully supervised dataset with deep learning.
\section{Experiments}
\label{df:sec:example}
In this section, we review a baseline, and experiments that showcase the usefulness of our algorithm -- which corresponds to ~\eqref{df:eq:disambiguation}~and~\eqref{df:eq:estimate}.
\paragraph{Baseline.}
We consider as a baseline the work of \citet{Cabannes2020}, which is a consistent structured prediction approach to partial labeling through the infimum loss.
It is arguably the state-of-the-art of partial labeling approached through structured prediction.
It follows the same loss-based variance disambiguation principle, yet in an implicit fashion, leading to the inference algorithm, $f_n:\mathcal{X}\to\mathcal{Y}$,
\begin{equation}
\label{df:eq:baseline}
f_n(x) \in \argmin_{z\in\mathcal{Y}} \inf_{(y_i) \in C_n} \sum_{i=1}^n \alpha_i(x) \ell(z, y_i).
\end{equation}
Statistically, exponential convergence rates similar to Theorem \ref{df:thm:convergence} could be derived.
Yet, as we will see, our algorithm outperforms this state-of-the-art baseline.
\paragraph{Disambiguation coherence - Interval regression.}
The baseline~\eqref{df:eq:baseline} implicitly requires disambiguating $(\hat{y}_i(x))$ differently for every $x\in\mathcal{X}$.
This is counterintuitive since $(y_i^*)$ does not depend on $x$.
It means that $(\hat{y}_i)$ could be equal to some $(\hat{y}_i^{(0)})$ on a subset $\mathcal{X}_0$ of $\mathcal{X}$, and to another $(\hat{y}_i^{(1)})$ on a disjoint subset $\mathcal{X}_1 \subset \mathcal{X}$, leading to irregularity of $f_n$ between $\mathcal{X}_0$ and $\mathcal{X}_1$.
We illustrate this graphically on Figure \ref{df:fig:ir}.
This figure showcases an interval regression problem, which corresponds to the regression setup ($\mathcal{Y} = \mathbb{R}$, $\ell(y, z) = \abs{y - z}^2$) of partial labeling, where one does not observe $y\in\mathbb{R}$ but an interval $s\subset \mathbb{R}$ containing $y$.
Among others this problem appears in physics \citep{Sheppard1897} and economy \citep{Tobin1958}.
\begin{figure*}[t]
\centering
\includegraphics[width=.45\linewidth]{disambiguation/images/ir/setting.pdf}
\includegraphics[width=.45\linewidth]{disambiguation/images/ir/reconstruction.pdf}
\caption{
Interval regression.
See Appendix \ref{df:app:experiments} for the exact reproducible experimental setup (Left) Setup.
The goal is to learn $f^*:\mathcal{X}\to\mathbb{R}$ represented by the dashed line, given samples $(x_i, s_i)$, where $(s_i)$ are intervals represented by the blue segments.
(Right) We compare the infimum loss (IL) baseline~\eqref{df:eq:baseline} shown in green, with our disambiguation framework (DF), \eqref{df:eq:disambiguation} and \eqref{df:eq:estimate}, shown in orange; with weights $\alpha$ given by kernel ridge regression.
(DF) retrieves $\hat{y_i}$ before learning a smooth $f_n$ based on $(x_i, \hat{y}_i)$, while (IL) implicitly retrieves $\hat{y}_i(x)$ differently for each input, leading to irregularity of the consequent estimator of $f^*$.
}
\label{df:fig:ir}
\end{figure*}
\paragraph{Computation attractiveness - Ranking.}
Computationally, the baseline requires to solve a disambiguation problem, recovering $(\hat{y_i}(x)) \in C_n$ for every $x\in\mathcal{X}$ for which we want to infer $f_n(x)$.
This is much more costly, than doing the disambiguation of $(\hat{y_i}) \in C_n$ once, and solving the supervised learning inference problem~\eqref{df:eq:estimate}, for every $x\in\mathcal{X}$ for which we want to infer $f_n(x)$.
To illustrate the computation attractiveness of our algorithm, consider the case of ranking, defined in Section \ref{df:sec:ranking}.
Fully supervised inference scheme~\eqref{df:eq:estimate} corresponds to solving a NP-hard problem, equivalent to the minimum feedback arcset problem \citep{Duchi2010}.
While disambiguation approaches with alternative minimization implied by~\eqref{df:eq:disambiguation} and~\eqref{df:eq:baseline} require to solve this NP-hard problem for each minimization step.
In other terms, the baseline ask to solve multiple NP-hard problem every time one wants to infer $f_n$ given by~\eqref{df:eq:baseline} on an input $x\in\mathcal{X}$.
Meanwhile, our disambiguation approach asks to solve multiple NP-hard problem upfront to solve~\eqref{df:eq:disambiguation}, yet only require to solve one NP-hard problem to infer $f_n$ given by~\eqref{df:eq:estimate} on an input $x\in\mathcal{X}$.
\paragraph{Better empirical results - Classification.}
Finally, we compare our algorithm, our baseline~\eqref{df:eq:baseline} and the baseline considered by \citet{Cabannes2020} on real datasets from the LIBSVM dataset \citep{Chang2011}.
Those datasets $(x_i, y_i)$ correspond to fully supervised classification problem.
In this setup, $\mathcal{Y} = \bbracket{1,m}$ for $m$ a number of classes, and $\ell(y, z) = \ind{y\neq z}$.
We ``corrupt'' labels in order to create a synthetic weak supervision datasets $(x_i, s_i)$.
We consider skewed corruption, in the sense that $(s_i)$ is generated by a probability such that $\sum_{z\in\mathcal{Y}} \Pbb_{S_i}(z\in S_i \vert y_i)$ depends on the value of $y_i$.
This corruption is parametrized by a parameter that related with the ambiguity parameter $\eta$ of Assumption \ref{df:ass:non-ambiguity}.
Results on Figure \ref{df:fig:cl} show that, in addition to having a lower computation cost, our algorithm performs better in practice than the state-of-the-art baseline.\footnote{All the code is available online - \url{https://github.com/VivienCabannes/partial_labelling/}.}
\begin{figure*}[t]
\centering
\includegraphics[width=.45\linewidth]{disambiguation/images/cl/dna.pdf}
\includegraphics[width=.45\linewidth]{disambiguation/images/cl/svmguide2.pdf}
\caption{Testing errors as function of the supervision corruption on real dataset corresponding to classification with partial labels.
We split fully supervised LIBSVM datasets into training and testing dataset.
We corrupt training data in order to get partial labels.
Corruption is managed through a parameter, represented by the $x$-axis, that relates to the ambiguity degree $\eta$ of Assumption \ref{df:ass:non-ambiguity}.
For each method (our algorithm (DF), the baseline (IL), and the baseline of the baseline (AC, consisting of averaging candidates $y_i$ in sets $S_i$)), we consider weights $\alpha$ given by kernel ridge regression with Gaussian kernel, for which we optimized hyperparameters with cross-validation on the training set.
We then learn an estimate $f_n$ that we evaluate on the testing set, represented by the $y$-axis, on which we have full supervision.
The figure shows the superiority of our method, that achieves error similar to baseline when full supervision ($x=0$) or no supervision ($x=100\%$) is given, but performs better when only in presence of partial supervision.
See Appendix \ref{df:app:experiments} for reproducibility specifications, where we also provide Figure \ref{df:fig:rk} showcasing similar empirical results in the case of ranking with partial ordering.
}
\label{df:fig:cl}
\end{figure*}
\paragraph{Beyond~\eqref{df:eq:principle} - Semi-supervised learning.}
The main limitation of~\eqref{df:eq:principle} is that it is a pointwise principle that decorrelates inputs, in the sense that the optimization of $\mu^*\vert_x$, for $x\in\mathcal{X}$, only depends on $\nu\vert_x$ and not on what is happening on $\mathcal{X}\setminus \brace{x}$.
As such, this principle failed to tackle semi-supervised learning, where $\nu\vert_x$ is equal to $\mu\vert_x$ (in the sense that $\pi\vert_{x, y} =\delta_{\brace{y}}$) for $x\in\mathcal{X}_l$ and is equal to $\delta_\mathcal{Y}$ for $x \in \mathcal{X}_u:=\mathcal{X}\setminus\mathcal{X}_l$.
In such a setting, for $x\in\mathcal{X}_u$, $\mu^*\vert_x$ can be set to any $\delta_y$ for $y\in\mathcal{Y}$.
Interestingly, in practice, while the baseline suffer the same limitation, for our algorithm, {\em weighting schemes have a regularization effect}, that contrasts with those considerations.
We illustrate it on Figure \ref{df:fig:ss}.
\begin{figure*}[ht]
\centering
\includegraphics[width=.3\linewidth]{disambiguation/images/ss/setting.pdf}
\includegraphics[width=.3\linewidth]{disambiguation/images/ss/df.pdf}
\includegraphics[width=.3\linewidth]{disambiguation/images/ss/il.pdf}
\caption{Semi-supervised learning, ``concentric circle'' instance with four classes (red, green, blue, yellow).
Reproducibility details provided in Appendix \ref{df:app:experiments}.
(Left) We represent points $x_i \in \mathcal{X} \subset \mathbb{R}^2$, there is many unlabeled points (represented by black dots and corresponding to $S_i = \mathcal{Y}$), and one labeled point for each class (represented in color, corresponding to $S_i = \brace{y_i}$).
(Middle) Reconstruction $f_n:\mathcal{X} \to \mathcal{Y}$ given by our algorithm~\eqref{df:eq:disambiguation} and \eqref{df:eq:estimate}.
Our algorithm succeeds to comprehend the concentric circle structure of the input distribution and clusters classes accordingly.
(Right) Reconstruction $f_n:\mathcal{X}\to\mathcal{Y}$ given by the baseline~\eqref{df:eq:baseline}.
The baseline performs as if only the four supervised data points where given.}
\label{df:fig:ss}
\end{figure*}
\section{Conclusion}
\label{df:sec:opening}
In this work, we have introduced a structured prediction algorithm~\eqref{df:eq:disambiguation} and \eqref{df:eq:estimate}, to tackle partial labeling.
We have derived exponential convergence rates for the nearest neighbors instance of this algorithm under classical learnability assumptions.
We provided optimization considerations to implement this algorithm in practice, and have successfully compared it with the state-of-the-art.
Several open problems offer prospective follow-up of this works:
\begin{itemize}
\item \emph{Semi-supervised learning and beyond.}
While we only proved convergence in situation where $\mu^*$ of~\eqref{df:eq:principle} is uniquely defined, therefore excluding semi-supervised learning, Figure \ref{df:fig:ss} suggests that our algorithm \eqref{df:eq:disambiguation} could be analyzed in a broader setting than the one considered in this paper.
Among others, we conjecture that the non-ambiguity assumption could be replaced by a cluster assumption \citep{Rigollet2007} together with a non-ambiguity assumption cluster-wise in Theorem~\ref{df:thm:convergence}.
\item \emph{Hard-coded weak supervision.}
Variational principles~\eqref{df:eq:principle} and \eqref{df:eq:solution} could be extended beyond partial labeling to any type of hard-coded weak supervision, which is when weak supervision can be cast as a set of hard constraint that $\mu$ should satisfy, formally written as a set of fully supervised distributions compatible with weak information.
Hard-coded weak supervision includes label proportion \citep{Quadrianto2009,Dulac2019}, but excludes supervision of the type ``80\% of the experts say this nose is broken, and 20\% say it is not''.
Providing a unifying framework for those problems would make an important step in the theoretical foundation of weakly supervised learning.
\item \emph{Missing input data.}
While weak supervision assumes that only $y$ is partially known, in many applications of machine learning, $x$ is also only partially known, especially when the feature vector $x$ is built from various source of information, leading to missing data.
While we only considered a principle to fill missing output information, similar principles could be formalized to fill missing input information.
\end{itemize}
\chapter{Disambiguation of weak supervision with exponential convergence rates}
\chapter{Disambiguation Framework}
The following is a reproduction of \cite{Cabannes2021}.
\input{disambiguation/abstract}
\input{disambiguation/core}
\begin{subappendices}
\chapter*{Appendix}
\addcontentsline{toc}{chapter}{Appendix}
\input{disambiguation/appendix/proof}
\input{disambiguation/appendix/diffrac}
\input{disambiguation/appendix/example}
\input{disambiguation/appendix/experiments}
\end{subappendices}
\section{Experiments}
\label{il:app:experiments}
\subsection{Classification}\label{il:app:classification}
Let us consider the classification setting of Section~\ref{il:sec:classification}.
The infimum loss reads $L(z, S) = \ind{z\notin S}$.
Given a weak distribution $\tau$, the infimum loss is therefore solving for
\[
f(x) \in \argmin_{z\in\mathcal{Y}} \E_{S\sim\tau\vert_x}\bracket{L(z, S)}
= \argmin_{z\in\mathcal{Y}} \E_{S\sim\tau\vert_x}\bracket{\ind{z\notin S}}
= \argmin_{z\in\mathcal{Y}} \PP_{S\sim\tau\vert_x}(z\notin S)
= \argmax_{z\in\mathcal{Y}} \PP_{S\sim\tau\vert_x}(z\in S).
\]
Given data, $(z_i, S_i)$ our estimator consists in approximating the conditional distributions $\tau\vert_x$ as
\[
\hat\tau\vert_x = \sum_{i=1}^n \alpha_i(x) \delta_{S_i},
\]
from which we deduce the inference formula, that we could also be derived from~\eqref{il:eq:algorithm},
\[
\hat{f}(x) \in \argmax_{z\in\mathcal{Y}} \sum_{i=1}^n \alpha_i(x) \ind{z\in S_i}
= \argmax_{z\in\mathcal{Y}} \sum_{i; z\in S_i} \alpha_i(x).
\]
\subsubsection{Complexity analysis}
The complexity of our algorithm~\eqref{il:eq:algorithm} can be split in two parts:
\begin{itemize}
\item a training part, where given $(x_i, S_i)$ we precompute quantities that will be useful at inference.
\item an inference part, where given a new $x$, we compute the corresponding prediction $\hat{f}(x)$.
\end{itemize}
In the following, we will review the time and space complexity of both parts.
We give this complexity in terms of $n$ the number of data and $m$ the number of items in $\mathcal{Y}$.
Results are summed up in Table~\ref{il:tab:cl:complexity}.
\begin{table}[t]
\caption{Complexity of our algorithm for classification.}
\label{il:tab:cl:complexity}
\vskip 0.3cm
\centering
{\small\sc
\begin{tabular}{lcc}
\toprule
Complexity & Time & Space \\
\midrule
Training & ${\cal O}(n^2(n+m))$ & ${\cal O}(n(n+m))$ \\
Inference & ${\cal O}(nm)$ & ${\cal O}(n+m)$ \\
\bottomrule
\end{tabular}
}
\end{table}
\paragraph{Training.}
Let us suppose that computing $L(y, S) = \ind{y\notin S}$ can be done in a constant cost that does not depend on $m$.
We first compute the following matrices in ${\cal O}(nm)$ and ${\cal O}(n^2)$ in time and space.
\[
L = (L(y, S_i))_{i\leq n, y\in\mathcal{Y}} \in \mathbb{R}^{n\times m},\qquad
K_\lambda = (k(x_i, x_j) + n\lambda\delta_{i=j})_{ij} \in \mathbb{R}^{n\times n}.
\]
We then solve the following, based on the {\tt \_gesv} routine of Lapack, in ${\cal O}(n^3 + n^2m)$ in time and ${\cal O}(n(n+m))$ in space \citep[see][for details]{Golub1996}
\[
\beta = K_\lambda^{-1}L \in \mathbb{R}^{n\times m}.
\]
\paragraph{Inference.}
At inference, we first compute in ${\cal O}(n)$ in both time and space
\[
v(x) = (k(x, x_i))_{i\leq n} \in \mathbb{R}^n.
\]
Then we do the following multiplication in ${\cal O}(nm)$ in time and ${\cal O}(m)$ in space,
\[
{\cal R}_{S, x} = v(x)^T \beta \in \mathbb{R}^m.
\]
Finally, we take the minimum of ${\cal R}_{S, x}(z)$ over $z$ in ${\cal O}(m)$ in time and ${\cal O}(1)$ in space.
\subsubsection{Baselines}
The average loss is really similar to the infimum loss, it reads
\[
L_{\textit{ac}}(z, S) = \frac{1}{\module{S}}\sum_{y\in S} \ell(z, y) = 1 -
\frac{\ind{z\in S}}{\module{S}} \simeq \frac{1}{\module{S}}\cdot\ind{z\notin S} =
\frac{1}{\module{S}} L(z, S).
\]
Following similar derivations to the one for the infimum loss, given a distribution $\tau$, one can show that the average loss is solving for
\[
f_{\textit{ac}}(x) \in \argmax_{z\in\mathcal{Y}} \sum_{S; z\in S}
\frac{1}{\module{S}}\tau\vert_x(S),
\]
which is consistent when $\tau$ is not ambiguous.
The difference with the infimum loss is due to the term in $\module{S}$.
It can be understood as an evidence weight, giving less importance to big sets that do not allow discriminating efficiently between candidates.
Given data $(x_i, S_i)$, it leads to the estimator
\[
\hat{f}_{\textit{ac}}(x) \in \argmin_{z\in\mathcal{Y}} \sum_{i; z\in S_i} \frac{\alpha_i(x)}{\module{S_i}}.
\]
The supremum loss is really conservative since
\[
L_{\textit{sp}}(z, S) = \sup_{y\in S} \ell(y, z) = \sup_{y\in S} \ind{y\neq z}
= \ind{S\neq\brace{z}}.
\]
It is solving for
\[
f(x) \in \argmax_{z\in\mathcal{Y}} \tau\vert_x(\brace{z}),
\]
which empirically correspond to discarding all the set with more than one element
\[
\hat{f}_{\textit{sp}}(x) \in \argmin_{z\in\mathcal{Y}} \sum_{i; S_i = \brace{z}} \alpha_i(x).
\]
Note that $\tau$ could be not ambiguous while charging no singleton, in this case, the supremum loss is not informative, as its risk is the same for any prediction.
\subsubsection{Corruptions on the \emph{LIBSVM} datasets}\label{il:sec:libsvm}
To illustrate the dynamic of our method versus the average baseline, we used {\em LIBSVM} datasets~\citep{Chang2011}, that we corrupted by artificially adding false class candidates to transform fully supervised pairs $(x, y)$ into weakly supervised ones $(x, S)$.
We experiment with two types of corruption processes.
\begin{itemize}
\item A uniform one, reading, with the $\mu$ of Definition~\ref{il:def:eligibility}, for $z\neq y$,
\[
\PP_{(Y, S)\sim\mu\vert_{\mathcal{Y}\times 2^\mathcal{Y}}}\paren{z\in S\,\middle\vert\, Y=y} = c.
\]
with $c$ a corruption parameter that we vary between zero and one.
In this case, the average loss and the infimum one works the same as shown on Figure~\ref{il:fig:cl:uniform}.
\item A skewed one, where we only corrupt the pair $(x, y)$ when $y$ is the most present class in the dataset.
More exactly, if $y$ is the most present class in the dataset, for $z\in\mathcal{Y}$, and $z'\neq z$, our corruption process reads
\[
\PP_{(Y, S)\sim\mu\vert_{\mathcal{Y}\times 2^\mathcal{Y}}}\paren{z'\in S\,\middle\vert\, Y=z} = c\cdot\ind{z=y}.
\]
In unbalanced dataset, such as the ``DNA'' and ``svmguide2'' datasets, where the most present class represents more than fifty percent of the labels as shown Table~\ref{il:tab:cl:libsvm}, this allows fooling the average loss as shown Figure~\ref{il:fig:libsvm}.
Indeed, this corruption was designed to fool the average loss since we knew of the evidence weight $\frac{1}{\module{S}}$ appearing in its solution.
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=.45\textwidth]{infimum/images/cl/segment.pdf}
\includegraphics[width=.45\textwidth]{infimum/images/cl/vowel.pdf}
\vspace*{-.3cm}
\caption{Classification.
Testing risks \eqref{il:eq:risk} achieved by {\em AC} and {\em IL} on the ``segment'' and ``vowel'' datasets from {\em LIBSVM} as a function of corruption parameter $c$, when the corruption is uniform, as described in Section~\ref{il:sec:libsvm}.}
\label{il:fig:cl:uniform}
\end{figure}
\begin{table}[t]
\caption{{\em LIBSVM} datasets characteristics, showing the number of data, of classes, of input features, and the proportion of the most present class when labels are unbalanced.}
\label{il:tab:cl:libsvm}
\vskip 0.3cm
\centering
{\small\sc
\begin{tabular}{lccccc}
\toprule
Dataset & Data ($n$) & Classes ($m$) & Features ($d$) & Balanced & Most present \\
\midrule
DNA & 2000 & 3 & 180 & $\times$ & 52.6\% \\
Svmguide2 & 391 & 3 & 20 & $\times$ & 56.5\% \\
Segment & 2310 & 7 & 19 & $\checkmark$ & - \\
Vowel & 528 & 11 & 10 & $\checkmark$ & - \\
\bottomrule
\end{tabular}
}
\end{table}
\subsubsection{Reproducibility specifications}
All experiments were run with {\em Python}, based on the {\em NumPy} library.
Randomness was controlled by instantiating the random seed of {\em NumPy} to $0$ before doing any computations.
Results of Figures~\ref{il:fig:libsvm} and \ref{il:fig:cl:uniform} were computed by using eight folds, and trying out several hyperparameters, before keeping the set of hyperparameters that hold the lowest mean error over the eight folds.
Because we used a Gaussian kernel, there were two hyperparameters, the Gaussian kernel parameter $\sigma$, and the regularization parameter $\lambda$.
We search for the best hyperparameters based on the heuristic
\[
\sigma = c_\sigma d,\qquad \lambda = c_\lambda n^{-1/2},
\]
where $d$ is the dimension of the input $\mathcal{X}$ (or the number of features), and where the Gaussian kernel reads
\[
k(x, x') = \exp\paren{-\frac{\norm{x-x'}^2}{2\sigma^2}}.
\]
We tried $c_\sigma \in \brace{10, 5, 1, .5, .1, .01}$ and $c_\lambda \in\brace{10^i\,\middle\vert\, i\in\bbracket{3, -3}}$.
\subsection{Ranking}
Consider the ranking setting of Section~\ref{il:sec:ranking}, where $\mathcal{Y} = \mathfrak{S}_m$, $\phi$ is the Kendall's embedding and the loss is equivalent to $\ell(z, y) = - \phi(y)^T\phi(z)$.
\subsubsection{Complexity analysis}
Given data $(x_i, S_i)$, our algorithm is solving at inference for
\[
f(x) \in \argmin_{z\in\mathcal{Y}}\inf_{y_i \in S_i} - \sum_{i=1}^n \alpha_i(x)
\phi(z)^T\phi(y_i)
= \argmax_{z\in\mathcal{Y}}\sup_{y_i \in S_i} \sum_{i=1}^n \alpha_i(x) \phi(z)^T\phi(y_i)
\]
We solved it through alternate minimization, by iteratively solving in $z$ for
\[
\phi(z)^{(t+1)} = \argmax_{\xi \in \phi(\mathcal{Y})} \scap{\xi}{ \sum_{i=1}^n \alpha_i(x) \phi(y_i)^{(t)}},
\]
and solving for each $y_i$ for
\[
\phi(y_i)^{(t+1)} = \argmax_{\xi\in\phi(S_i)} \alpha_i(x)\scap{\xi}{\phi(z)}.
\]
We initialize the problem with the coordinates of $\phi(y_i)$ put to 0 when not specified by the constraint $y_i \in S_i$.\footnote{%
Coordinates of the Kendall's embedding correspond to pairwise comparison between two items $j$ and $k$, so we put to 0 the coordinates for which we can not infer preferences from $S$
between items $j$ and $k$.
}
Those two problems are minimum feedback arc set problems, that are \textit{NP}-hard in $m$, meaning that one has to check for all potential solutions, and there is $m!$ of them, which is the cardinal of $\mathfrak{S}_m$.
We suggest solving them using an integer linear programming (ILP) formulation that we relax into linear programming as explained in Appendix~\ref{il:app:fas}.
All the problems in $y_i$ share the same objective, up to a change in sign, but different constraint $\xi\in\phi(S_i)$, such a setting is particularly suited for warm start on the dual simplex algorithm to solve efficiently one after the other the linear programs associated to each $y_i$.
To give numbers, at training time, we compute the inverse $K_\lambda^{-1}$ in ${\cal O}(n^3)$ in time and ${\cal O}(n^2)$ in space, and at inference we compute $\alpha(x) K_\lambda^{-1}v(x)$ in ${\cal O}(n^2)$ in time and ${\cal O}(n)$ in space, before solving iteratively $n$ \textit{NP}-hard problem in $m$ of complexity $n \textit{NP}(m)$, that cost $nm^2$ in space to represent using {\em Cplex}~\citep{Cplex}, if we allow our self $e$ iterations, the inference complexity is ${\cal O}(n^2 + e\,n\,\textit{NP}(m))$ in time and ${\cal O}(nm^2)$ in space.
\subsubsection{Baselines}
The supremum loss is really similar to the infimum loss, only changing an infimum by a supremum.
However, algorithmically, this change leads to solving for a local saddle point rather than solving for a local minimum.
While the latter are always defined, there might be instances where no saddle point exists.
In this case, the supremum optimization might stall without getting to any stable solution, and the user might consider stopping the optimization after a certain number of iteration and outputting the current state as a solution.
The average loss, despite its simple formulation, does not lead to an easy implementation either.
Indeed, when given a set $S$, the average loss is implicitly computing the center of this set $c(S)$, and replacing $L_{\textit{ac}}(z, S)$ by $\ell(z, c(S))$, more exactly
\[
L_{\textit{ac}}(z, S) \simeq - \frac{1}{\module{S}} \sum_{y\in S} \phi(z)^T\phi(y)
= -\phi(y)^T\paren{\frac{1}{\module{S}}\sum_{y\in S}\phi(y)}.
\]
To compute the center $\paren{\frac{1}{\module{S}}\sum_{y\in S}\phi(y)}$, we sample $c_k \sim {\cal N}(0, I_{m^2})$, solve the resulting minimum feedback arc set problem, with the constraint $y\in S$, and end up with solutions $\phi(y_k)$.
After removing duplicates, we estimate the average with the empirical one.
Note that this work is done at training, leading the average loss to have a quite good inference complexity in ${\cal O}(nm + \textit{NP}(m))$ in time.
\subsubsection{Synthetic example: ordering lines}
In the following, we explain our synthetic example of Section~\ref{il:sec:ranking}.
It corresponds of choosing $\mathcal{X} = [0, 1]$, choose $m$ a number of items, simulate $a, b \sim {\cal N}(0, I_m)$, compute scores $v_i(x) = ax + b$, and order items according to their scores as shown on Figure~\ref{il:fig:rk:setting}.
For Figure~\ref{il:fig:rk:corruption}, we chose $m=10$, as this is the biggest $m$ for which can rely on our minimum feedback arc set heuristic to recover the real minimum feedback arc set solution and therefore not to play a role in what our algorithm will output.
The corruption process was defined as losing coordinates in the Kendall's embedding, more exactly given a point $x\in\mathcal{X}$, we have a score $(v_i(x))_{i\leq m}$ and an ordering $y\in\mathcal{Y}$.
To create a skewed corruption, we first compute the normalized distance between scores as
\[
d_{ij} = \frac{\module{v_i - v_j}}{\max_{k,l} \module{v_k - v_l}} \in [0, 1]
\]
and remove the pairwise comparison for which $d_{ij} > c$, where $c$ is a
corruption parameter between 0 and 1, formally
\[
S = \brace{z \in \mathcal{Y}\,\middle\vert\, \forall\, (j,k) \in I,\ \phi(z)_{jk} = \phi(y)_{jk}},\qquad\text{where}\qquad
I = \brace{(j,k) \,\middle\vert\, d_{(j, k)} < c },
\]
Because of the transitivity constraint, when $c$ is small the comparison that we lost can be found back using transitivity between comparisons.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{infimum/images/rk/reconstruction.pdf}
\vspace*{-.3cm}
\caption{Reconstruction of the problem of Figure~\ref{il:fig:rk:setting}, given $n=50$ random points $(x_i, y_i)_{i\leq n}$, after losing at random fifty percent of the coordinates $(\phi(y_i))_{i\leq n}$, leading to sets $(S_i)_{i\leq n}$ of potential candidates.
Hyperparameter were chosen as $\sigma = 1$ for the Gaussian kernel and $\lambda = 10^{-3}n^{-1/2}$ for the regularization parameter.
The percentage of error in the reconstructed Kendall's embedding is 3\% for {\em IL}, 4\% for {\em AC} and 13\% for {\em SP}.
As for classification, with such a random corruption process, {\em AC} and {\em IL} show similar behaviors.}
\label{il:fig:ranking_reconstruction}
\end{figure}
\subsubsection{Reproducibility specification}
To get Figure~\ref{il:fig:rk:corruption}, we generate eight problems that correspond to ordering $m=10$ lines, seen as eight folds.
We only cross validated results with the same heuristics as in Appendix~\ref{il:app:classification}, yet, because computations were expensive we only tried $c_\sigma \in \brace{1, .5}$, and $c_\lambda \in \brace{10^3, 1, 10^{-3}}$.
Again, randomness was controlled by instantiating random seeds to 0.
Solving the linear program behind our minimum feedback arc set was done using {\em Cplex} \citep{Cplex}, which is the fastest linear program solver we are aware of.
\subsection{Multilabel}
Multilabel is another application of partial labeling that we did not mention in our experiment section in the core paper.
This omission was motivated by the fact that, under natural weak supervision, the three losses (infimum, average and supremum) are basically the same.
However, we will provide, now, an explanation of this problem and our algorithm to solve it.
Multilabel prediction consists in finding which are the relevant tags (possibly more than one) among $m$ potential tags.
In this case, one can represent $\mathcal{Y} = \brace{-1, 1}^m$, with $y_i = 1$ (resp.~$y_i = -1$), meaning that tag $i$ is relevant (resp.~not relevant).
The classical loss is the Hamming loss, which is the decoupled sum of errors for each label:
\[
\ell(y, z) = \sum_{i=1}^m \ind{y_i\neq z_i}.
\]
Natural weak supervision consists in mentioning only a few relevant or irrelevant tags.
This is the setting of~\citet{Yu2014}.
This leads to sets $S$ that are built from a set $P$ of relevant items, and a set $N$ of irrelevant items.
\[
S = \brace{y\in \mathcal{Y} \,\middle\vert\, \forall\,i\in P, y_i = 1, \forall\,i\in N, y_i = -1}.
\]
In this case, the infimum loss reads,
\[
L(z, S) = \sum_{i\in P} \ind{z_i = -1} + \sum_{i\in N} \ind{z_i = 1}.
\]
For such supervision, the infimum, the average and the supremum loss are intrinsically the same, they only differs by constants, due to the fact that for each unseen labels, the infimum loss pays $0$, the average loss $1/2$ and the supremum loss $1$.
When considering data $(x_i, S_i)_{i\leq n}$, where $(S_i)$ is built from $(N_i, P_i)$, our algorithm in~\eqref{il:eq:algorithm} reads $\hat{f}(x) = (\sign(\hat{f}_j(x)))_{j\leq m}$, based on the scores
\[
\hat{f}_j(x) = \sum_{i;j\in P_i} \alpha_i(x) - \sum_{i;j\in N_i} \alpha_i(x).
\]
\subsubsection{Tackling positive bias.}
In the precedent development, we implicitly assumed that the ratio between positive and negative labels given by the weak supervision reflects the one of the full distribution.
An assumption that is often violated in practice.
It is common that partial labeling only mentions a subset of the relevant tags ({\em i.e.}, $N = \emptyset$).
This case is ill-conditioned as always outputting all tags ($y = \textbf{1}$) will minimize the infimum loss.
To solve this problem, we can constrain the prediction space to the top-$k$ space $\mathcal{Y}_{k} = \brace{y\in\mathcal{Y}\,\middle\vert\, \sum_{i=1}^m \ind{y_j = 1} = k}$, which will lead to taking the top-$k$ over the score $(\hat{f}_j)_{j\leq m}$.
We can also break the loss symmetry and add a penalization with $\epsilon >0$,
\[
\ell_\epsilon(z, y) = \ell(z, y) + \epsilon \sum_{i=1}^m \ind{z_i=1}.
\]
In this case, the inference algorithm will threshold scores at $\epsilon$ rather than $0$.
\subsubsection{Complexity analysis}
The complexity analysis is similar to the one for classification.
At training, we compute $L = (\ind{j\in P_i} - \ind{j \in N_i})$, and we solve for $\beta = K_\lambda^{-1}L$ in $\mathbb{R}^{n\times m}$.
At testing, we compute $v(x)$ and $\beta^T v(x)$ in $\mathbb{R}^m$, before thresholding it or taking the top-$k$ in either ${\cal O}(m)$ or ${\cal O}(m\log(m))$.
As such, complexity reads similarly as for the classification case.
Yet notice that, for multilabeling, the dimension of $\mathcal{Y}$ is not $m$ but $2^m$, meaning we do not scale with $\#\mathcal{Y}$ but with the intrinsic dimension.
\begin{table}[t]
\caption{Complexity of our algorithm for multilabels.}
\label{il:tab:ml:complexity}
\vskip 0.3cm
\centering
{\small\sc
\begin{tabular}{lcc}
\toprule
Complexity & Time & Space \\
\midrule
Training & ${\cal O}(n^2(n+m))$ & ${\cal O}(n(n+m))$ \\
Inference & ${\cal O}(nm)$ & ${\cal O}(n+m)$ \\
Inference top-$k$ & ${\cal O}(nm + m\log(m))$ & ${\cal O}(n+m)$ \\
\bottomrule
\end{tabular}
}
\end{table}
\subsubsection{Corruptions on the \emph{MULAN} datasets}\label{il:sec:mulan}
When sets are given by few positive and negative tags, all losses are the same.
Yet, under other types of supervision, such as when the sets come as Hamming balls, defined by
\[
B(z, r) = \brace{y\in \mathcal{Y} \,\middle\vert\, \ell(z, y) \leq r},
\]
the methods will not behave the same.
We experiment on MULAN datasets provided by~\citet{Tsoumakas2011}.
Because supervision with Hamming balls does not lead to efficient implementation, we went for extensive grid search for the best solution, which reduces our ability to consider large $m$.
Among MULAN datasets, we went for the ``scene'' one, with $m=6$ tags, and $n=2407$ data.
When given a pair $(x, y)$, we add corruption on $y$, by first sampling a radius parameter $r\sim\uniform([0, c*(m+1)])$, with $c$ a corruption parameter.
We then sample, with replacement, $\floor{r}$ coordinates to modify to pass from $y$ to a center $c$.
We then consider the supervision $S = B(c, r)$.
For such randomness, somehow uniform corruption, the infimum loss works slightly better than the average loss that both outperform the supremum loss as shown on Figure~\ref{il:fig:mulan}.
\begin{figure}[t]
\centering
\includegraphics{infimum/images/ml/scene.pdf}
\vspace*{-.3cm}
\caption{Multilabeling.
Testing risks (from~\eqref{il:eq:risk}) achieved by {\em AC} and {\em IL} on the ``scene'' dataset from MULAN as a function of corruption parameter c, shown in procedure defined unit, when the supervision is given as Hamming balls, as described in Section~\ref{il:sec:mulan}.}
\label{il:fig:mulan}
\end{figure}
\subsubsection{Reproducibility specification}
To get Figure~\ref{il:fig:mulan}, we follow the same cross-validation scheme as for classification and ranking.
More exactly, we cross-validated over eight folds with the same heuristics for $\sigma$, the Gaussian kernel parameter, and $\lambda$, the regularization one, with $c_\sigma \in \brace{10, 5, 1, .5, .1, .01}$, and $c_\lambda \in \brace{10^i\,\middle\vert\, i\in\bbracket{-3, 3}}$.
\subsection{Partial regression}
Partial regression is the regression instance of partial labeling.
When supervision comes as intervals, it is known as interval regression, and known as censored regression when sets come as half-lines.
Note that for censored regression, neither the average, nor the supremum loss can be properly defined.
\subsubsection{Baselines}
Given a bounded set $S$, learning with the average loss corresponds to considering the center of this set, since, for $z\in\mathcal{Y}$, with $\lambda$ the Lebesgue measure
\begin{align*}
L_{\textit{ac}}(z, S) & = \frac{1}{\lambda(S)}\int_{S} \norm{z - y}^2 \lambda(\mathop{}\!\mathrm{d} y)
= \norm{z}^2 - 2\scap{z}{\frac{1}{\lambda(S)}\int_{S} y \lambda(\mathop{}\!\mathrm{d} y)} +
\frac{1}{\lambda(S)}\int_{S} \norm{y}^2 \lambda(\mathop{}\!\mathrm{d} y)
\\&= \norm{z - \frac{1}{\lambda(S)}\int_{S} y \lambda(\mathop{}\!\mathrm{d} y)}^2 + \frac{1}{\lambda(S)}\int_{S} \norm{y}^2 \lambda(\mathop{}\!\mathrm{d} y) - \norm{\frac{1}{\lambda(S)}\int_{S} y \lambda(\mathop{}\!\mathrm{d} y) }^2
= \norm{z - c(S)}^2 + C_S,
\end{align*}
where $c(S) = \frac{1}{\lambda(S)}\int_{S} y \lambda(\mathop{}\!\mathrm{d} y)$ is the center of $S$.
As such, the average loss is always convex.
As the supremum of convex function, the supremum loss is also convex.
\subsubsection{Reproducibility specification}
To compute Figure~\ref{il:fig:interval-regression}, for both {\em AC} and {\em IL}, we consider $\sigma$, the Gaussian kernel parameter, and $\lambda$, the regularization parameter, achieving the best risk when measured with the fully supervised distribution~\eqref{il:eq:risk}.
We tried over $\sigma \in \brace{1, .5, .1, .05, .01}$ and $\lambda \in \brace{10^3, 1, 10^{-3}}$.
Randomness was controlled by instantiating random seeds.
\subsection{Beyond}
Beyond the examples showcased previously, advances in dealing with weak supervision could be beneficial for several problems.
Supervision on \emph{image segmentation} problems usually comes as partial pixel annotation.
This problem is often tackled through conditional random fields~\citep{Verbeek2007}, making it a perfect mix between partial labeling and structured prediction.
\emph{Action retrieval} on instructional video, where partial supervision is retrieved from the audio track is another interesting application~\citep{Alayrac2018}.
\section{Minimum feedback arc set}
\label{il:app:fas}
\subsection{Formulation}
Consider a directed weighted graph with vertices $\bbracket{1, m}$ and edges $\brace{i\rightarrow j}$ with weights $(w_{ij})_{i,j\leq m} \in \mathbb{R}_+^{m^2}$.
The goal is to find a directed acyclic graph $G = (V, E)$ that maximizes the weights on remaining edges
\[
\argmax_{E} \sum_{i\rightarrow j \in E} w_{ij}.
\]
This directed acyclic graph can be seen as a preference graph, item $j$ being preferred over item $i$.
Since $w_{ij}$ are non-negative, the underlying ordering in $G$ is necessarily total, and therefore can be written based on a score function, that can be embedded in the permutation of $\bbracket{1, m}$, $\sigma \in \mathfrak{S}_m$, with $\sigma(j) > \sigma(i)$ meaning that $j$ is preferred over $i$.
Thus the problem reads equivalently
\begin{align*}
\argmax_{\sigma\in\mathfrak{S}_m}\sum_{i,j \leq m} w_{ij} \ind{\sigma(j) > \sigma(i)}
& = \argmax_{\sigma\in\mathfrak{S}_m}\sum_{i<j \leq m} c_{ij} \ind{\sigma(j) > \sigma(i)}
= \argmax_{\sigma\in\mathfrak{S}_m}\sum_{i<j \leq m} c_{ij} \sign\paren{\sigma(j) - \sigma(i)}
\\&= \argmin_{\sigma\in\mathfrak{S}_m}\sum_{i<j \leq m} c_{ij} \sign\paren{\sigma(i) - \sigma(j)}
= \argmin_{\sigma\in\mathfrak{S}_m}\sum_{i<j \leq m} c_{ij} \ind{\sigma(i) > \sigma(j)}
\end{align*}
with $c_{ij} = w_{ij} - w_{ji}$.
This last formulation is the one usually encountered for ranking algorithms in machine learning~\citep{Duchi2010}.
We are going to study in depth this problem under the formulation
\begin{equation}\label{il:eq:fas}
\argmin_{\sigma \in\mathfrak{S}_m} \sum_{i<j\leq m} c_{ij} \sign\paren{\sigma(i) - \sigma(j)}
\end{equation}
\subsection{Integer linear programming}
\begin{definition}[Kendall's embedding]\label{il:def:ken}
For $\sigma \in \mathfrak{S}_m$, define Kendall's embedding, with $m_e =m(m-1) / 2$,
\[
\phi(\sigma) = \sign\paren{\sigma(i) - \sigma(j)}_{i<j \leq m} \in \brace{-1, 1}^{m_e}.
\]
We associate it to the Kendall's polytope of order $m$, $\hull\paren{\phi(\mathfrak{S}_m)}$.
\end{definition}
The Kendall's embedding, Definition~\ref{il:def:ken}, cast the minimum feedback arcset problem~\eqref{il:eq:fas} as a linear program
\[
\minimize{\scap{c}{x}}{x\in \hull\paren{\phi(\mathfrak{S}_m)}.}
\]
Since the objective is linear, the solution is known to lie on a vertex of the constraint polytope, which is the set of Kendall's embeddings of permutations.
Yet, how to describe Kendall's polytope?
\begin{definition}[Transitivity polytope]\label{il:def:tr}
The transitivity polytope of order $m$ is defined in $\mathbb{R}^{m_e}$ as
\[
{\cal M} = \brace{x\in\mathbb{R}^{m_e} \,\middle\vert\, \forall\,i<k<j; -1 \leq x_{ij} + x_{jk} - x_{ik} \leq 1 }
\]
This polytope encodes the transitivity constraints of Kendall's embeddings, Definition~\ref{il:def:ken}.
\end{definition}
The transitivity polytope, Definition~\ref{il:def:hull}, will be used to approximate Kendall's polytope based on the following property.
\begin{proposition}[Relaxed polytope]\label{il:prop:rel}
The intersection between the transitivity polytope and the vertex of the hypercube is exactly the set of Kendall's embeddings of permutations.
Mathematically
\[
\phi(\mathfrak{S}_m) = {\cal M} \cap \brace{-1, 1}^{m_e}.
\]
\end{proposition}
\begin{proof}
First of all it is easy to show that $\phi(\mathfrak{S}_m) \subset \brace{-1, 1}^{m_e}$, and that, $\phi(\mathfrak{S}_m)\subset{\cal M}$.
Let's now consider $x \in {\cal M} \cap\brace{-1,1}^{m_e}$.
Let's associate to $x$ the symmetric embedding
\[
\tilde{x}_{ij} = \left\{
\begin{array}{ccc}
x_{ij} & \text{if} & i < j \\
0 & \text{if} & i = j \\
-x_{ji} & \text{if} & j < i
\end{array}
\right.
\]
Let's consider the permutation $\sigma$ resulting from the ordering of $\sum_{k}\tilde{x}_{ik}$
\[
\sigma^{-1}(1) = \argmin_{i \in\bbracket{1,m}} \sum_{k = 1}^m \tilde{x}_{ik}\qquad\text{and}\qquad
\sigma^{-1}(i) = \argmin_{i \in\bbracket{1,m}\setminus \sigma^{-1}\paren{\bbracket{1, i-1}} } \sum_{k=1}^m \tilde{x}_{ik}.
\]
Let's now show that $\phi(\sigma) = x$, or equivalently that $\tilde{\phi}(\sigma) = (\sign(\sigma(i) - \sigma(j)))_{i,j\leq m} = \tilde{x}$.
First, one can show that $\tilde{x}$ verify the transitivity constraints
\[
\forall\, i, j, k \leq m, \qquad -1 \leq \tilde{x}_{ij} + \tilde{x}_{jk} - \tilde{x}_{ik} \leq 1.
\]
This can be proven for any ordering of $i, j, k$ based on the fact that $x \in {\cal M}$.
For example, if $i < k < j$, we have
\[
\bracket{-1, 1} \ni x_{ik} + x_{kj} - x_{ij} = \tilde{x}_{ik} - \tilde{x}_{jk} - \tilde{x}_{ij}.
\]
which leads to
\[
\tilde{x}_{ij} + \tilde{x}_{jk} - \tilde{x}_{ik} \in -\bracket{-1, 1} = \bracket{-1, 1}.
\]
Now suppose, without loss of generality, that $\tilde{x}_{ij} = 1$ (if $\tilde{x}_{ij} = -1$, just consider $\tilde{x}_{ji} = 1$).
The transitivity constraints tell us that $\tilde{x}_{ik} \geq \tilde{x}_{jk}$ for all $k$, therefore
\[
\sum_{k\not\in\brace{i,j}} \tilde{x}_{ik} \geq \sum_{k\not\in\brace{i,j}} \tilde{x}_{jk},
\qquad\Rightarrow\qquad
\sum_{k=1}^{m} \tilde{x}_{ik} > \sum_{k=1}^{m} \tilde{x}_{jk}.
\qquad\Rightarrow\qquad
\sigma(i) > \sigma(j).
\]
This shows that $\tilde{\phi(\sigma)}_{ij} = 1 = \tilde{x}_{ij}$.
Thus, we have shown that $x \in \phi(\mathfrak{S}_m)$, which concludes the proof.
\end{proof}
\begin{figure}[t]
\centering
\includegraphics[width=.45\textwidth]{infimum/images/fas/ilp_sol.pdf}
\includegraphics[width=.45\textwidth]{infimum/images/fas/ilp_sol_zoom.pdf}
\vspace*{-.3cm}
\caption{Evaluating the percentage of exact solutions of the ILP relaxation as $m$ grows large.
Evaluation is done by choosing an objective $c\sim{\cal N}(0, I_{m_e})$, solving the ILP relaxation, Definition~\ref{il:def:hull}, and evaluating if the solution is in $\brace{-1, 1}^{m_e}$.
The experience is repeated several times to estimate how often, on average, the original solution of~\eqref{il:eq:fas} is returned by the ILP.
}
\label{il:fig:ilp_rel}
\end{figure}
\begin{definition}[ILP relaxation]\label{il:def:hull}
Based on Proposition~\ref{il:prop:rel}, we define the canonical polytope ${\cal C} = {\cal M} \cap \bracket{-1, 1}^{m_e}$, and relax the problem~\eqref{il:eq:fas} into
\[
\minimize{\scap{c}{x}}{x\in{\cal C}}
\]
As soon as the solution $x$ is in $\brace{-1, 1}^{m_e}$, Proposition~\ref{il:prop:rel} tells us that $x$ recover the exact minimum feedback arc set solution~\eqref{il:eq:fas}.
\end{definition}
\iftrue
In small dimensions, the canonical polytope ${\cal C}$ is the same as the Kendall's one, and the ILP relaxation gives the right solution.
Yet, as shown in Figure~\ref{il:fig:ilp_rel}, as soon as $m > 5$, there exists vertex in ${\cal C}$ that does not correspond to a permutation embedding.
For small dimensions, proving that ${\cal C}$ is exactly the Kendall's polytope is done with a simple drawing for $m=3$, using unimodularity of the transitivity constraint matrix is enough for $m=4$~\citep{Hoffman2010}.
The case $m=5$ is also provable, based on several tricks that we will not discuss here.
\else
\begin{proposition}
The ILP relaxation, Definition~\ref{il:def:hull}, always recovers the solution of~\eqref{il:eq:fas} if and only if $m \leq 5$.
\end{proposition}
\begin{proof}
For $m > 5$, computation done to draw Figure~\ref{il:fig:ilp_rel} gives counter-examples, and small dimension counter-examples can be cast in bigger ones as $m$ grows.
For $m=3$, one can draw the Kendall's polytope and the canonical one, they are the same.
For $m=4$, let us use basic knowledge on integer linear program and unimodularity~\citep{Hoffman2010}.
With $x = (x_{12}, x_{13}, x_{14}, x_{23}, x_{24}, x_{34})$, the transitivity constraints are built from
\[
A =
\paren{\begin{array}{cccccc}
1 & -1 & 0 & 1 & 0 & 0 \\
1 & 0 & -1 & 0 & 1 & 0 \\
0 & 1 & -1 & 0 & 0 & 1 \\
0 & 0 & 0 & 1 & -1 & 1 \\
\end{array}}
\]
More precisely, the entire constraints read $Ax \leq 1$, $-Ax\leq -1$, $x\leq 1$, $-x\leq 1$.
The matrix $A$, and the one that build the transitivity constraints for $m=5$, is totally unimodular, meaning that any invertible sub-square matrix of $A$ is invertible in $\mathbb{Z}$, or equivalently, any sub-square matrix has determinant -1, 0 or 1.
It is easy to show that $\tilde{A} = [A, -A, I, -I]^T$ is also totally unimodular.
Consider a vertex $x$ in ${\cal C}$, it can be defined from $m_e$ linearly independent constraint hyperplanes, it solves an equation of the type $\tilde{A}_{\vert I} x = b$, with $b\in\brace{-1, 1}^{m_e}$.
$\tilde{A}_{\vert I}$ being an invertible sub-square matrix of $\tilde{A}$, it is invertible in $\mathbb{Z}$, leading to $x = \tilde{A}_{\vert I}^{-1}b$ being integer.
Therefore, $x\in\brace{-1, 0, 1}^{m_e}$.
To show that $x\in\brace{-1, 1}^{m_e}$, let's consider $x' = (x + 1)/2$, it corresponds to the embedding $(\ind{\sigma(i) > \sigma(j)})_{ij}$.
Solving the original problem with $x'$ or with $x$ is the same.
The constraints for $x'$ reads $Ax' \leq 1$, $-Ax'\leq 0$, $x' \leq 1$, $-x'\leq 0$.
Therefore, for the same reason as $x$, $x'$ is integer, more exactly $x'\in\brace{0, 1}^{m_e}$, from which $x$ is deduced to be in $\brace{-1,1}^{m_e}$.
By showing that all vertices of ${\cal C}$ are in $\brace{-1, 1}^{m_e}$, using Proposition~\ref{il:prop:rel} we have shown that the ILP relaxation does give the exact solution for $m = 4$.
For $m=5$, one can not use unimodularity as the transitivity constraint matrix shows sub-matrices of determinant $\pm 2$, for example
\[
\tilde{A} = \begin{blockarray}{cccccc}
x_{45} & x_{34} & x_{25} & x_{13} & x_{12} \\
\begin{block}{(ccccc)c}
1 & 1 & 0 & 0 & 0 & x_{34} + x_{45} - x_{35} \\
1 & 0 & -1 & 0 & 0 & x_{24} + x_{45} - x_{25} \\
0 & 1 & 0 & 1 & 0 & x_{13} + x_{34} - x_{14} \\
0 & 0 & 1 & 0 & 1 & x_{12} + x_{25} - x_{15} \\
0 & 0 & 0 & -1 & 1 & x_{12} + x_{23} - x_{13} \\
\end{block}
\end{blockarray}
\]
This matrix is not invertible in $\mathbb{Z}$.
The constraints and coefficients corresponding are written around it.
Let's first prove that ${\cal C} \in \frac{1}{2}\mathbb{Z}$.
Mainly we used that all sub-matrix of dimension $7\times 7$ are not invertible, the one of dimension $6\times 6$ are of determinant -1, 1, 0, and the one of $5\times 5$ are the only ones with determinant sometimes being $\pm 2$.
This shows that the determinant of the matrix defining the constraint for a vertex is at most 2, by developing the determinant with respect to constraint of the type $x_{ij} = \pm 1$.
Which shows that ${\cal C} \in \frac{1}{2}\mathbb{Z}$.
With the same embedding trick, with $x' = (x+1)/2$, we show that $x\in\mathcal{Z}$.
We know that the only potential trouble is when the matrix defining the constraints has a sub-factor of determinant $\pm 2$, and all the other constraint are of the type $x_{ij} = \pm 1$, meaning that there is at least 5 coordinates among the 10 non-zero.
If we put those into the 5 transitivity constraints, if we put one of them in a coefficient that appear twice, then necessarily there will be one transitivity constraint that get two of its three coefficient fixed by the simple constraints, this constraint can so be replaced by a fixed one on the last coefficient, meaning that the vertex could also be defined by 4 transitivity constraint and 6 simple constraint, in this case it will be invertible.
Meaning that we should distribute the simple constraint among the coefficients that appears only once.
This links all the remaining coefficients between them in a really simple way, either they are all zeros, or all integers.
If they are all zeros, one can check that they are not on the border by showing for example that $\max_{y\in{\cal C}}\scap{x}{y}$ is not only achieve for $y=x$, indeed any completing of the partial order given by the simple constraint will work
\end{proof}
\fi
\begin{remark}[Low noise consistency]
Remark that the low-noise setting considered by~\citet{Duchi2010} correspond to having $\sign(c) = -\phi(y)$ for a $y\in\mathcal{Y}$, in this case our algorithm is consistent and does recover the best solution $z=y$.
\end{remark}
\subsection{Sorting heuristics}
When formatting and solving the integer linear program takes too much time, one can go for a simple sorting heuristic, mainly based on a heuristic to compare items two by two and using quick sorting.
A review of some heuristic with guarantees is provided by~\citet{Ailon2005},
Similar study when in presence of constraints on the resulting total order can be found in~\citet{vanZuylen2007}.
\section{Proofs}
\label{il:app:proof}
In the paper, we have implicitly considered $\mathcal{X}, \mathcal{Y}$ separable and completely metrizable topological spaces, {\em i.e.} Polish spaces, which allows considering probabilities.
Moreover, we assumed that $\mathcal{Y}$ is compact, to have minimizers well-defined.
The observation space was considered to be the set of closed subsets of $\mathcal{Y}$ endowed with the Hausdorff distance, ${\cal S} = {\operatorname{Cl}(\mathcal{Y}), d_H}$.
As such, ${\cal S}$ is also a Polish metric space, inheriting this property from $\mathcal{Y}$ \citep{Beer1993}.
In the following, we will show that the closeness of sets is important in order to switch from the minimum variability principle to the infimum loss.
In terms of notations, we use the simplex notation $\Prob{\cal A}$ to denote the space of Borel probability measures over the space ${\cal A}$.
In particular, $\Prob{\mathcal{X}\times\mathcal{Y}}$, $\Prob{\mathcal{X}\times{\cal S}}$ and $\Prob{\mathcal{X}\times\mathcal{Y}\times{\cal S}}$ are endowed with the weak-* topology and are Polish, inheriting the properties from original spaces \citep{Aliprantis2006}.
The fact that such spaces are Polish allows defining the conditional probabilities given $x \in \mathcal{X}$.
We will denote this conditional probability $\rho\vert_x$ when, for example, $\rho\in\Prob{\mathcal{X}\times\mathcal{Y}}$.
Finally, we will denote by $\rho_\mathcal{X}$ the marginal distributions of $\rho$ over $\mathcal{X}$.
Before diving into proofs, we would like to point out that many of our results are pointwise results.
At an intuitive level, we only leverage the structure of the loss on the output space and aggregate those results over $\mathcal{X}$.
\begin{remark}[Going pointwise]
The learning frameworks in~\eqref{il:eq:risk}, \eqref{il:eq:infimum-disambiguation} and \eqref{il:eq:set-risk} are {\em pointwise separable} as their solutions can be written as aggregation of pointwise solutions \citep{Devroye1996}.
More exactly, the partial labeling risk (and similarly the fully supervised one) can be expressed as
\[
{\cal R}_S(f) = \E_X\bracket{{\cal R}_{S,X}(f(X))},
\]
where the conditional risk reads,
\[
{\cal R}_{S, x}(z) = \E_{S\sim\tau\vert_x}\bracket{L(z, S)},
\]
with $\tau\vert_x$ the conditional distribution of $\paren{S\,\middle\vert\, X=x}$.
Thus, minimizing ${\cal R}_S$ globally for $f\in\mathcal{Y}^\mathcal{X}$ is equivalent to minimizing locally ${\cal R}_{S,x}$ for $f(x)$ for almost all $x$.
Similarly, for~\eqref{il:eq:infimum-disambiguation},
\[
{\cal E}(\rho) = \inf_{f:\mathcal{X}\to\mathcal{Y}}\E_{\rho}\bracket{\ell(f(X), Y)}
= \E_X\bracket{\inf_{z\in\mathcal{Y}} \E_{Y\sim\rho\vert_x}\bracket{\ell(z, Y) \,\middle\vert\, X=x}}.
\]
Therefore, studies on risk can be done pointwise on instances $(\ell, \rho\vert_x, \tau\vert_x)$, before integrating along $\mathcal{X}$.
Actually, Theorems \ref{il:thm:ambiguity}, \ref{il:thm:infimum-loss}, \ref{il:thm:non-ambiguity} and \ref{il:thm:calibration} are pointwise results.
\end{remark}
\subsection{Proof of Theorem~\ref{il:thm:ambiguity}}
\label{il:proof:ambiguity}
Here we want to prove that when $\tau$ is non-ambiguous, then it is possible to define an optimal $\rho^\star$ that is deterministic on $\mathcal{Y}$, and that this $\rho^\star$ is characterized by solving~\eqref{il:eq:infimum-disambiguation}.
\begin{lemma}\label{il:lem:ambiguity}
When $\tau$ is non-ambiguous, and there is one, and only one, deterministic distribution eligible for $\tau$.
More exactly, if we write, for any $x\in\mathcal{X}$ in the support of $\tau_\mathcal{X}$, based on Definition~\ref{il:def:non-ambiguity}, $S_x = \brace{y_x}$, then this deterministic distribution is characterized as $\rho\vert_x=\delta_{y_x}$ almost everywhere.
\end{lemma}
\begin{proof}
Let us consider a probability measure $\tau\in\Prob{\mathcal{X}\times{\cal S}}$.
We begin by working on the concept of eligibility.
Consider $\rho\in\Prob{\mathcal{X}\times\mathcal{Y}}$ eligible for $\tau$ and a suitable $\pi$ as defined in Definition~\ref{il:def:eligibility}.
First of all, the condition that, for $y\in S$, $\PP_\pi\paren{S\,\middle\vert\, Y=y} = 0$, can be stated formally in terms of measure as
\[
\pi(\brace{(x, y, S)\in\mathcal{X}\times\mathcal{Y}\times{\cal S}\,\middle\vert\, y\notin S}) = 0,
\]
from which we deduced that, for $y\in\mathcal{Y}$ and $x\in\mathcal{X}$,
\begin{align*}
\rho\vert_x(y) & = \pi\vert_x(\brace{y}\times{\cal S})
= \pi\vert_x(\brace{y} \times \brace{S \in{\cal S}\,\middle\vert\, y\in S})
\\&\leq \pi\vert_x(\mathcal{Y} \times \brace{S \in{\cal S}\,\middle\vert\, y\in S})
= \tau\vert_x(\brace{S\in{\cal S}\,\middle\vert\, y\in S}).
\end{align*}
It follows that when $\rho$ is deterministic, if we write $\rho\vert_x = \delta_{y_x}$, then we have
\(
\tau\vert_x(\brace{S\in{\cal S}\,\middle\vert\, y_x\in S}) = 1,
\)
which means that $y_x$ is in all sets that are in the support of $\tau\vert_x$, or that, using notations of Definition~\ref{il:def:non-ambiguity}, $y_x \in S_x$.
So far, we have proved that if there exists a deterministic distribution, $\rho\vert_x=\delta_{y_x}$, that is eligible for $\tau\vert_x$, we have $y_x\in S_x$.
Reciprocally, one can do the reverse derivations, to show that if $\rho\vert_x = \delta_{y_x}$, with $y_x \in S_x$, for all $x\in\mathcal{X}$, then $\rho$ is eligible for $\tau$
When $\tau$ is non-ambiguous, $S_x$ is a singleton and therefore, there could be only one deterministic eligible distribution for $\tau$, that is characterized in the lemma.
\end{proof}
Now we use the characterization of deterministic distribution through the minimization of the risk~\eqref{il:eq:risk}.
\begin{lemma}[Deterministic characterization]\label{il:lem:deterministic}
When $\mathcal{Y}$ is compact and $\ell$ proper, deterministic distribution are exactly characterized by minimum variability~\eqref{il:eq:infimum-disambiguation} as
\[
{\cal E}(\rho) = \inf_{f: \mathcal{X}\to\mathcal{Y}}\E_\rho\bracket{\ell(f(X), Y)} = 0.
\]
\end{lemma}
\begin{proof}
Let's consider $\rho\in\Prob{\mathcal{X}\times\mathcal{Y}}$, because $\mathcal{Y}$ is compact and $\ell$ continuous, we can consider $f_\rho$ a minimizer of ${\cal R}(f;\rho)$.
Let's now suppose that ${\cal R}(f_\rho; \rho) = 0$, since $\ell$ is non-negative, it means that almost everywhere
\[
\E_{Y\sim\rho\vert_x}\bracket{\ell(f_\rho(x), Y)} = 0.
\]
Suppose that $\rho\vert_x$ is not deterministic, then there is at least two points $y$ and $z$ in $\mathcal{Y}$ in its support, then, because $\ell$ is proper, we come to the absurd conclusion that
\[
\E_{Y\sim\rho\vert_x}\bracket{\ell(f_\rho(x), Y)}
\geq \rho\vert_x(y) \ell(f_\rho(x), y) + \rho\vert_x(z) \ell(f_\rho(x), z) > 0.
\]
So ${\cal R}(f_\rho; \rho) = 0$ implies that $\rho$ is deterministic.
Reciprocally, when $\rho$ is deterministic it is easy to show that the risk is minimized at zero.
\end{proof}
\subsection{Proof of Theorem~\ref{il:thm:infimum-loss}}
\label{il:proof:infimum-loss}
At a comprehensive level, Theorem~\ref{il:thm:infimum-loss} is composed of two parts:
\begin{itemize}
\item A double minimum switch, to take the minimum over $\rho$ before the minimum over $f$, and for which we need some compactness assumption to consider the joint minimum.
\item A minimum-expectation switch, to take the minimum over $\rho\vdash\tau$ as a minimum $y\in S$ before the expectation to compute the risk, and for which we need some measure properties.
\end{itemize}
We begin with the minimum-expectation switch.
To proceed with derivations, we need first to reformulate the concept of eligibility in Definition~\ref{il:def:eligibility} in terms of measures.
\begin{lemma}[Measure eligibility]\label{il:lem:eligibility}
Given a probability $\tau$ over $\mathcal{X} \times {\cal S}$, the space of probabilities over $\mathcal{X}\times\mathcal{Y}$ satisfying $\rho \vdash \tau$ is characterized by all probability measures of the form
\[
\rho(C) = \int_{\mathcal{X}\times\mathcal{Y}\times{\cal S}} \ind{C}(x, y)
\mathop{}\!\mathrm{d}\pi\vert_{x,S}(y) \mathop{}\!\mathrm{d}\tau(x, S),
\]
for any $C$ a closed subset of $\mathcal{X}\times\mathcal{Y}$, and where $\pi$ is a probability measure over $\mathcal{X}\times\mathcal{Y}\times{\cal S}$ that satisfies $\pi_{\mathcal{X}\times{\cal S}} = \tau$ and $\pi\vert_{x,S}(S) = 1$ for any $(x, S)$ in the support of $\tau$.
\end{lemma}
\begin{proof}
For any $\rho$ that is eligible for $\tau$ there exists a suitable $\pi$ on $\mathcal{X} \times \mathcal{Y} \times {\cal S}$ as specified by Definition~\ref{il:def:eligibility}.
Actually, the set of $\pi$ leading to an eligible $\rho:=\pi_{\mathcal{X}\times\mathcal{Y}}$ is characterized by satisfying $\pi_{\mathcal{X}\times{\cal S}} = \tau$ and
\[
\pi(\{(x,y,S) \in \mathcal{X}\times\mathcal{Y}\times{\cal S} ~|~ y \notin S\}) = 0.
\]
This last property can be reformulated with the complementary space as
\[
\pi(\{(x,y,S) \in \mathcal{X}\times\mathcal{Y}\times{\cal S} ~|~ y \in S\}) = 1,
\]
which equivalently reads, that for any $(x, S)$ in the support of $\tau$, we have
\[
\pi\vert_{x, S}(S) = \pi\vert_{x, S}(\brace{y\in\mathcal{Y}\,\middle\vert\, y\in S}) = 1.
\]
Finally, using the conditional decomposition we have that, for $C$ a closed subset of $\mathcal{X}\times\mathcal{Y}$
\[
\rho(C) = \pi_{\mathcal{X}\times\mathcal{Y}}(C)
= \int_{\mathcal{X}\times\mathcal{Y}\times{\cal S}} \ind{C}(x, y)\mathop{}\!\mathrm{d}\pi(x, y, S)
= \int_{\mathcal{X}\times\mathcal{Y}\times{\cal S}} \ind{C}(x, y)\mathop{}\!\mathrm{d}\pi\vert_{x, S}(y)
\mathop{}\!\mathrm{d}\pi_{\mathcal{X}\times{\cal S}}(x, S),
\]
which ends the proof since $\tau = \pi_{\mathcal{X}\times{\cal S}}$.
\end{proof}
We are now ready to state the minimum-expectation switch.
\begin{lemma}[Minimum-Expectation switch]\label{il:lem:eligibility-infimum}
For a probability measure $\tau\in\Prob{\mathcal{X}\times{\cal S}}$, and measurable functions $\ell \in \mathbb{R}^{\mathcal{Y}\times\mathcal{Y}}$ and $f \in \mathcal{Y}^\mathcal{X}$, the infimum of eligible expectations of $\ell$ is the expectation of the infimum of $f$ over $S$ where $S$ is distributed according to $\tau$.
Formally
\[
\inf_{\rho\vdash\tau}\E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)} =
\E_{(X,S)\sim\tau}\bracket{\inf_{y\in S}\ell(f(X), y)}.
\]
\end{lemma}
\begin{proof}
Before all, note that $(x, S) \to \inf_{y\in S} \ell(f(x), y)$ inherit measurability from $f$ allowing to consider such an expectation \citep[see Theorem 18.19 of][and references therein for details]{Aliprantis2006}.
Moreover, let us use Lemma~\ref{il:lem:eligibility} to reformulate the right hand side problem as
\[
\inf_{\rho\vdash\tau}\E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)} =
\inf_{\pi \in {\cal M}} \int_{\mathcal{X}\times\mathcal{Y}\times{\cal S}} \ell(f(x),
y)\mathop{}\!\mathrm{d}\pi_{x, S}(y)\mathop{}\!\mathrm{d}\tau(x, S).
\]
Where we denote by ${\cal M} \subset \Prob{\mathcal{X}\times\mathcal{Y}\times{\cal S}}$ the space of probability measures $\pi$ that satisfy the assumption of Lemma~\ref{il:lem:eligibility}.
We will now prove the equality by showing that both quantities bound the other one.
\paragraph{($\geq$).} To proceed with the first bound, notice that for
$x\in\mathcal{X}$ and $S\in{\cal S}$, when $\pi\vert_{x, S}\in\Prob{\mathcal{Y}}$ only charge $S$, {\em i.e.} if $\pi \in {\cal M}$, then
\[
\int_\mathcal{Y} \ell(f(x), y) \mathop{}\!\mathrm{d}\pi_{x, S}(y) \geq \inf_{y\in S} \ell(f(x), y).
\]
The first bound is then obtained by taking the expectation over $\tau$ of this pointwise property.
\paragraph{($\leq$).} For the second bound, we consider the function $Y \in \mathcal{Y}^{\mathcal{X}\times{\cal S}}$ define as
\[
Y(x, S) = \argmin_{y\in S} \ell(f(x), y).
\]
Such a function is well-defined since $S$ is compact due to the fact that $\mathcal{Y}$ is compact and ${\cal S}$ is the set of closed set.
However, in more general cases, one can consider a sequence that minimizes $\ell(f(x), y)$ rather than the argmin to show the same as what we are going to show.
Now, if we define $\pi^{(f)}$ with $\pi^{(f)}_{\mathcal{X}\times{\cal S}}:= \tau$ and $\pi^{(f)}\vert_{x, S} := \delta_{Y(x,S)}$, because $Y(x, S)$ is in $S$, we have that $\pi^{(f)}$ is in ${\cal M}$, so, for $x \in \mathcal{X}$ and $S\in{\cal S}$
\[
\inf_{\pi\in{\cal M}} \int_{\mathcal{Y}} \ell(f(x), y)\mathop{}\!\mathrm{d}\pi_{x, S}(y) \leq
\int_{\mathcal{Y}} \ell(f(x), y)\mathop{}\!\mathrm{d}\pi^{(f)}_{x, S}(y) = \ell(f(x), Y(x, S)) =
\inf_{y\in S} \ell(f(x), y).
\]
We end the proof by integrating this over $\tau$.
\end{proof}
Now, we will move on to the minimum switch.
First, we make sure that the infimum loss minimizer is well-defined.
\begin{lemma}[Infimum loss minimizer]
When $\mathcal{Y}$ is compact and the observed set are closed, there exists a measurable function $f_S \in \mathcal{Y}^\mathcal{X}$ that minimize the infimum loss risk
\[
{\cal R}_S(f_S) = \inf_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}_S(f), \qquad\text{where}\qquad
{\cal R}_S (f) = \int \min_{y \in S} \ell(f(x), y) \mathop{}\!\mathrm{d}\tau(x, S).
\]
The infimum on the right hand side is a minimum because $S$ is a closed subset of $\mathcal{Y}$ compact, and therefore, is compact.
\end{lemma}
\begin{proof}
First note that $d(y,y') = \sup_{z \in \mathcal{Y}} \module{\ell(z,y) - \ell(z,y')}$ is a metric on $\mathcal{Y}$ when $\ell$ is a proper loss: the triangular inequality holds trivially; when $y = y'$ then $d(y,y') = 0$; when $y\neq y'$, by properness we have $\ell(y,y) = 0$ and $d(y, y') \geq \ell(y,y') > 0$.
Moreover, note that $L(z,S) = \min_{y \in S} \ell(z,y)$ is continuous and $1$-Lipschitz with respect to the topology induced by the Hausdorff distance $d_H$ based on $d$, indeed given two sets $S,S' \in {\cal S}$
\begin{align*}
\module{L(z,S) - L(z,S')} & \leq \max\brace{
\max_{y \in S}\min_{y' \in S'} \module{\ell(z,y)-\ell(z,y')},~
\max_{y' \in S'}\min_{y \in S} \module{\ell(z,y)-\ell(z,y')}} \\
& \leq \max\brace{
\max_{y \in S}\min_{y' \in S'} d(y,y'),~
\max_{y' \in S'}\min_{y \in S} d(y,y')} = d_H(S,S').
\end{align*}
The result of existence of a measurable $f_S$ minimizing ${\cal R}_S(f) = \int L(f(x),S) d\tau(x, S)$ follows by the compactness of $\mathcal{Y}$, the continuity of $L(z,S)$ in the first variable with respect to the topology induced by $d$, in the second with respect to the topology induced by $d_H$ and measurability of $\tau\vert_x$ in $x$, via Berge maximum theorem \citep[see Theorem 18.19 of][and references therein]{Aliprantis2006}.
\end{proof}
We can state the minimum switch now.
\begin{lemma}[Minimum switch]\label{il:lem:switch}
When $\mathcal{Y}$ is compact, and observed sets are closed, solving the partial labeling through the minimum variability principle
\[
f^* \in \argmin_{f\in\mathcal{Y}^\mathcal{X}} \E_{\rho^\star}\bracket{\ell(f(X), Y)}, \qquad\text{with}\qquad
\rho^\star \in \argmin_{\rho\vdash\tau} \inf_{f\in\mathcal{Y}^\mathcal{X}} \E_{\rho}\bracket{\ell(f(X), Y)}.
\]
can be done jointly in $f$ and $\rho$, and rewritten as
\[
f^*\in\argmin_{f\in\mathcal{Y}^\mathcal{X}}\inf_{\rho\vdash\tau}\E_{\rho}\bracket{\ell(f(X), Y)}.
\]
\end{lemma}
\begin{proof}
When $(\rho^\star, f^*)$ is a minimizer of the top problem, it also minimizes the joint problem $(\rho, f) \to {\cal R}(f; \rho)$, and we can switch the infimum order.
The hard part is to show that when $f_S$ minimizes the bottom risk, the infimum over $\rho$ is indeed a minimum.
Indeed, we know from Lemma~\ref{il:lem:eligibility-infimum} that $f_S$ is characterized as a minimizer of the infimum risk ${\cal R}_S$, those are well-defined as shown in precedent lemma.
To $f_S$, we can associate $\rho_S := \pi^{(f)}$ as defined in the proof of Lemma~\ref{il:lem:eligibility-infimum}, which is due to the closeness of sets in ${\cal S}$ and the compactness of $\mathcal{Y}$.
Indeed, $(f_S, \rho_S)$ minimize jointly the objective ${\cal R}(f, \rho)$, so we have that
\[
\rho_S \in \argmin_{\rho\vdash\tau}\inf_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}(f;
\rho),\qquad\text{and}\qquad
f_S \in \argmin_{f:\mathcal{X}\to\mathcal{Y}}{\cal R}(f; \rho_S).
\]
From which we deduced that $\rho_S$ can be written as a $\rho^\star$ and $f_S$ as an $f^*$.
\end{proof}
\begin{remark}[A counter example when sets are not closed.]
The minimum switch relies on compactness assumption, which can be violated when the observed sets in ${\cal S}$ are not closed.
Let us consider the case where $\mathcal{Y} = \mathbb{R}$, $\ell=\ell_2$ is the mean square loss.
Consider the pointwise weak supervision
\[
\tau = \frac{1}{2} \delta_{\mathbb{Q}} + \frac{1}{2}\delta_{\sqrt{2}\mathbb{Q}},
\]
In this case, we have $\rho^\star = \delta_0$.
Yet, for any $z$, we do have ${\cal R}_{S, x}(z) = 0$ for any $z \in \mathbb{R}$.
For example, if $z=\sqrt{2}$, one can consider
\[
\rho_n = \frac{1}{2}\delta_{\sqrt{2}} + \frac{1}{2}\delta_{\frac{\floor{10^n\sqrt{2}}}{10^n}},
\]
to show that $z\in\argmin_{z\in\mathcal{Y}}\inf_{\rho\vdash\tau} {\cal R}(z, \rho)$.
As one can see this is counter example is based on the fact that $\brace{\rho\,\middle\vert\, \rho\vdash\tau}$ is not complete, so that there exists infimum of ${\cal R}_{x}(z, \rho)$ that are not minimum such as ${\cal R}_x(\sqrt{2}, \delta_{\sqrt{2}})$.
\end{remark}
\subsection{Proof of Theorem~\ref{il:thm:non-ambiguity}}
\label{il:proof:non-ambiguity}
If $\tau$ is not ambiguous, then, almost surely for $x\in\mathcal{X}$, if $y_x$ is the only element in $S_x$ of Definition~\ref{il:def:non-ambiguity}, we know that $\rho^\star\vert_x = \delta_{y_x}$, and consequently we derive $f^*(x)=y_x$, so for it to be consistent with $f_0$, we need that $f_0(x)=y_x$.
Moreover, because $\tau$ is a weakening of $\rho_0$, $\rho_0$ is eligible for $\tau$.
When $\rho_0$ is deterministic, we know from considerations in the proof of Lemma~\ref{il:lem:ambiguity}, that it is $\rho^\star$, the only deterministic distribution eligible for $\tau$.
Thus, in fact, the condition $S_x = \brace{f_0(x)}$ is implied by $\rho_0$ deterministic.
\subsection{Proof of Theorem~\ref{il:thm:calibration}}
\label{il:proof:calibration}
When $\tau$ is not ambiguous, we know from Theorem~\ref{il:thm:ambiguity}, that $\rho^\star$ is deterministic.
Let us write $\rho^\star\vert_x = \delta_{y_x}$, we have $f^*(x) = y_x$, and ${\cal R}_x(f^*) = 0$, moreover, because $y_x$ is in every $S$ in the support of $\tau\vert_S$, then ${\cal R}_{S, x}(f^*) = 0$.
Similarly to the bound given by~\citet{Cour2011} for the 0-1 loss, we have
\begin{align*}
{\cal R}_{S, x}(z) & = \E_{S\sim\tau\vert_x}[\inf_{z'\in S} \ell(z, z')]
= \sum_{S; z\notin S} \inf_{z'\notin S} \ell(z, z')\PP_{S\sim\tau\vert_x}(S)
\\&\geq \inf_{z'\neq z} \ell(z, z')\PP_{S\sim\tau\vert_x}(z\notin S)
\geq \inf_{z'\neq z} \ell(z, z')\eta,
\end{align*}
while ${\cal R}_x(z) = \ell(z, y)$, so we deduce locally
\begin{align*}
{\cal R}_{x}(z; \rho^\star\vert_x) - {\cal R}_{x}(f^*(x); \rho^\star\vert_x)
& \leq \frac{\ell(z, y)}{\inf_{z'\neq z} \ell(z, z')} \eta^{-1} \paren{{\cal R}_{S, x}(z) - {\cal R}_{S, x}(f^*(x))}
\\&\leq e^\nu\eta^{-1}\paren{{\cal R}_{S, x}(z) - {\cal R}_{S, x}(f^*(x))}.
\end{align*}
Integrating over $x$ this last equation gives us the bound in Theorem~\ref{il:thm:calibration}.
\subsection{Refined bound analysis of Theorem~\ref{il:thm:calibration}}
\label{il:discussion:refinement-C}
The constant $C$ that appears in Theorem~\ref{il:thm:calibration} is the result of controlling separately the corruption process and the discrepancy of the loss.
Indeed, they can be controlled together, leading to a better constant.
To relates the two risk ${\cal R}$ and ${\cal R}_S$, we will consider the pointwise setting $\tau\in\Prob{2^\mathcal{Y}}$ and $\rho_0\in\Prob{\mathcal{Y}}$ that satisfies $\rho_0\vdash\tau$, we will also consider a prediction $z\in\mathcal{Y}$.
\begin{proposition}[Bound refinement]\label{il:prop:refinement}
When $\mathcal{Y}$ is discrete and $\tau$ not ambiguous, the best $C$ that verifies~\eqref{il:eq:calibration-bound} in the pointwise setting $\tau\in\Prob{2^\mathcal{Y}}$ is the maximum of $\lambda^{-1}$, for $\lambda \in [0, 1]$ such that there exists a point $z\neq y$ and signed measured $\sigma$ that verify ${\cal R}(z;\sigma) = 0$ and such that $\sigma + \lambda \delta_y + (1-\lambda)\delta_z$ is a probability measure that is eligible for $\tau$.
\end{proposition}
\begin{proof}
First, let's extend our study to the space ${\cal M}_\mathcal{Y}$ of signed measure over $\mathcal{Y}$.
We extend the risk definition in~\eqref{il:eq:risk} to any signed measure $\mu\in {\cal M}_\mathcal{Y}$, with
\[
{\cal R}_x(z; \mu) = \int_{\mathcal{Y}} \ell(z, y) \mathop{}\!\mathrm{d}\mu(y).
\]
Note that the risk is a linear function of the distribution $\mu$.
Two spaces are going to be of particular interest: the one of measures of mass one ${\cal M}_{\mathcal{Y}, 1}$, and the one of measures of mass null ${\cal M}_{\mathcal{Y}, 0}$, where
\[
{\cal M}_{\mathcal{Y}, p} = \brace{\mu\in{\cal M} \,\middle\vert\, \mu(\mathcal{Y}) = p}.
\]
Let's now relates for a $\rho_0$, $\tau$ and $z$, the risk ${\cal R}_x(z; \rho_0)$ and ${\cal R}_{S,x}(z)$.
To do so, we introduce the space of signed measures of null mass, that could be said orthogonal to $(\ell(z, y))_{y\in\mathcal{Y}}$, formally
\[
D_z = \brace{\mu\in{\cal M}_{\mathcal{Y}, 0} \,\middle\vert\, {\cal R}_x(z; \mu) = 0}.
\]
There are two alternatives: (1) either ${\cal R}_x(z; \rho_0) = 0$, and so ${\cal R}_{S, x}(z) = 0$ too, and we have related the two risk; (2) either ${\cal R}_x(z, \rho_0) \neq 0$, and the space ${\cal M}_{\mathcal{Y}, 1}$ can be decomposed as
\[
{\cal M}_{\mathcal{Y}, 1} = D_z + \brace{\lambda \rho_0 + (1-\lambda)\delta_z \,\middle\vert\,
\lambda \in \mathbb{R}}.
\]
To prove it take $\mu \in {\cal M}_{\mathcal{Y}, 1}$, and use linearity of the risk after writing
\[
\mu = \lambda \rho_0 + (1-\lambda)\delta_z + \paren{\mu - (\lambda \rho_0 + (1-\lambda)\delta_z)},
\qquad\text{with}\qquad \lambda = \frac{{\cal R}_{x}(z, \mu)}{{\cal R}_{x}(z, \rho_0)}.
\]
For such a $\mu$, using the linearity of the risk, and the properness of the loss, if we denote by $d_z$ the part in $D_z$ of the last decomposition, we have
\[
{\cal R}_x(z; \mu) = \lambda {\cal R}_x(z; \rho_0) + (1-\lambda){\cal R}_x(z;
\delta_z) + {\cal R}_x(z; d_z) = \lambda {\cal R}_x(z; \rho_0)
\]
If we denote by $R_\tau = \brace{\rho\in\Prob{\mathcal{Y}}\,\middle\vert\, \rho\vdash \tau}$, we can conclude that
\[
\frac{{\cal R}_{S, x}(z)}{{\cal R}_x(z; \rho_0)} = \inf\brace{\lambda\,\middle\vert\,
(\lambda \rho_0 + (1-\lambda)\delta_z) \in R_\tau + D_z}.
\]
Finally, when $\tau$ is not ambiguous, we know that $\rho^\star$ is deterministic, and if $\rho_0$ is deterministic then $\rho_0 = \rho^\star$.
In this case, there exists a $y$ such that $\rho_0 = \delta_y$, and we can suppose this $y$ is different from $z$ otherwise ${\cal R}_x(z; \rho_0) = 0$.
In this case, we also have ${\cal R}_x(z^*) = {\cal R}_{S, x}(z^*) = 0$ with $z^* = y$, and thus the excess of risk to relates in~\eqref{il:eq:calibration-bound} is indeed the relation between the two risks.
\begin{figure}[t]
\centering
\input{infimum/drawings/simplex_calibration}
\vspace*{-.3cm}
\caption{Geometrical understanding of Proposition~\ref{il:prop:refinement}, showing the link between the infimum and the fully supervised risk.
The drawing is set in the affine span of the simplex ${\cal M}_{\mathcal{Y}, 1}$, where we identify $a$ with $\delta_a$.
The underlying instance $(\ell, \tau)$ is taken from Section~\ref{il:sec:inconsistency}, and can be linked to the setting of Proposition~\ref{il:prop:refinement} with $z=b$, $y=c$.
Are represented in the simplex the level curves of the function $\rho\rightarrow {\cal R}(z;\rho)$.
Based on this drawing, one can recover ${\cal R}_S(b) = {\cal R}(b)/4$, which is better than the bound given in Theorem~\ref{il:thm:calibration}.
}
\label{il:fig:simplex-calibration}
\end{figure}
%
\end{proof}
\begin{remark}[Proposition \ref{il:prop:refinement} as a variant of Thales theorem]
Proposition \ref{il:prop:refinement} can be seen as a variant of the Thales theorem.
Indeed, with the geometrical embedding $\pi$ of the simplex in $\mathbb{R}^\mathcal{Y}$, $\pi(\rho) = (\rho(y))_{y\in\mathcal{Y}}$, one can have, with $d$ the Euclidean distance
\[
\frac{{\cal R}_{S, x}(z)}{{\cal R}_x(z; \rho_0)} = \frac{d(\pi(\delta_z +
D_z), \pi(R_\tau))}{d(\pi(\delta_z + D_z), \pi(\rho_0))}.
\]
And conclude by using the following variant of Thales theorem, that can be derived from Figure~\ref{il:fig:proof-thales}:
For $x, y, z \in \mathbb{R}^d$, and $S \subset \mathbb{R}^d$, with $d$ the Euclidean distance, if $y \in S$, $d(z+x^\perp, S) = \gamma d(z+x^\perp, y) $, where
\[
\gamma = \min\brace{\module{\lambda}\,\middle\vert\, \lambda\in \mathbb{R},
(\lambda y + (1- \lambda) z + x^\perp) \cap S \neq \emptyset}.
\]
Moreover, notice that if $S$ is contained in the half-space that contains $y$ regarding the cut with the hyperplane $z + x^\perp$, $\lambda$ can be restricted to be in $[0,1]$.
\begin{figure}[t]
\centering
\input{infimum/drawings/thales}
\vspace*{-.3cm}
\caption{A variant of Thales theorem.}
\label{il:fig:proof-thales}
\end{figure}
%
\end{remark}
\begin{remark}[Active labeling]
When annotating data, as a partial labeler, you could ask yourself how to optimize your labeling.
For example, suppose that you want to poll a population to retrieve preferences among a set of presidential candidates.
Suppose that for a given polled person, you can only ask her to compare between four candidates.
Which candidates would you ask her to compare?
According to the questions you are asking, you will end up with different sets of potential weak distribution $\tau$.
If aware of the problem $\ell$ that your dataset is intended to tackle, and aware of a constant $C = C(\ell, \tau)$ that verifies~\eqref{il:eq:calibration-bound}, you might want to design your questions in order to maximize on average over potential $\tau$, the quantity $C(\ell,\tau)$.
An example where $\tau$ is not well-designed according to $\ell$ is given in Figure~\ref{il:fig:example-calibration}.
\begin{figure}[t]
\centering
\input{infimum/drawings/example_calibration.tex}
\vspace*{-.3cm}
\caption{Example of a bad link between $\tau$ and $\ell$.
Same representation as Figure~\ref{il:fig:simplex-calibration} with a different instance where $\tau = \frac{1}{2} \delta_{\brace{a, c}} + \frac{1}{2}\delta_{\brace{b, c}}$ and $\ell(b, a) = 0$, $\ell(b, c) = 1$.
In this example $C_\ell(\tau)=+\infty$, and the infimum loss is $0$ on $\mathcal{Y}$ and therefore not consistent.
Given the loss structure, partial labeling acquisition should focus on specifying sets that do not intersect $\brace{a, b}$.
Note that this instance violates the proper loss assumption, explaining its inconsistency.
}
\label{il:fig:example-calibration}
\end{figure}
%
\end{remark}
\subsection{Proof of Theorems~\ref{il:thm:consistency} and \ref{il:thm:learning-rates}}
\label{il:proof:consistency}
First, note that, since ${\cal R}_S(f)$ is characterized by ${\cal R}_S(f) = \mathbb{E}_{(x,S)\sim\tau} \min_{u \in S} \ell(f(x), u)$, then the problem
\[
f^* = \argmin_{f:\mathcal{X} \to \mathcal{Y}} {\cal R}_S(f) = \argmin_{f:\mathcal{X} \to \mathcal{Y}}
\E_{(x,S)\sim\tau}\bracket{\min_{y \in S} \ell(f(x), y)}.
\]
can be considered as an instance of structured prediction with loss $L(z,S) = \min_{y \in S} \ell(f(x), y)$.
The framework for structured prediction presented in \citet{Ciliberto2016}, and extended in \citet{Ciliberto2020}, provides consistency and learning rates in terms of the excess risk ${\cal R}_S(f_n) - {\cal R}_S(f^*)$ when $f^*$ is estimated via $f_n$ defined as in~\eqref{il:eq:algorithm} and when the structured loss $L$ admits the decomposition
\[
L(z,S) = \langle \psi(z), \phi(S) \rangle_{\cal H},
\]
for a separable Hilbert space ${\cal H}$ and two maps $\psi: \mathcal{Y} \to {\cal H}$ and $\phi: {\cal S} \to {\cal H}$.
Note that since $\mathcal{Y}$ is finite $L$ always admits the decomposition, indeed the cardinality of $\mathcal{Y}$ is finite, {\em i.e.}, $|\mathcal{Y}| < \infty$ and $|{\cal S}| = 2^{|\mathcal{Y}|}$.
Choose an ordering for the elements in $\mathcal{Y}$ and in ${\cal S}$ and denote them respectively $o_\mathcal{Y}:\mathbb{N} \to \mathcal{Y}$ and $o_{\cal S}:\mathbb{N} \to {\cal S}$.
Let $n_\mathcal{Y}: \mathcal{Y} \to \mathbb{N}$ be the inverse of $o_\mathcal{Y}$, {\em i.e.} $o_\mathcal{Y}(n_\mathcal{Y}(y)) = y$ and $n_\mathcal{Y}(o_\mathcal{Y}(i)) = i$ for $y \in \mathcal{Y}$ and $i \in {1,\dots,|\mathcal{Y}|}$, define analogously $n_{\cal S}$.
Now let ${\cal H} = \mathbb{R}^{|\mathcal{Y}|}$ and define the matrix $B \in \mathbb{R}^{|\mathcal{Y}| \times 2^{|\mathcal{Y}|}}$ with element $B_{i,j} = L(o_\mathcal{Y}(i), o_{\cal S}(j))$ for $i=1,\dots,|\mathcal{Y}|$ and $j=1,\dots,2^{|\mathcal{Y}|}$, then define
\[
\psi(z) = e^{|\mathcal{Y}|}_{n_\mathcal{Y}(z)}, \quad \phi(S) = B e^{2^{|\mathcal{Y}|}}_{n_{\cal S}(S)},
\]
where $e^k_i$ is the $i$-th element of the canonical basis of $\mathbb{R}^k$.
We have that
\[
\langle \psi(z), \phi(S) \rangle_{\cal H} = \langle e^{|\mathcal{Y}|}_{n_\mathcal{Y}(z)}, B
e^{2^{|\mathcal{Y}|}}_{n_{\cal S}(S)} \rangle_{R^{|\mathcal{Y}|}} = B_{n_\mathcal{Y}(z), n_{\cal S}(S)} =
L(i_\mathcal{Y}(n_\mathcal{Y}(z)), i_{\cal S}(n_{\cal S}(S))) = L(z,S),
\]
for any $z \in \mathcal{Y}, S \in {\cal S}$.
So we can apply Theorem 4 and 5 of \cite{Ciliberto2016} \citep[see also their extended forms in Theorem 4 and 5 of][]{Ciliberto2020}.
The last step is to connect the excess risk on ${\cal R}_S$ with the excess risk on ${\cal R}(f, \rho^\star)$, which is done by our comparison inequality in Theorem~\ref{il:thm:calibration}.
\begin{remark}[Illustrating the consistency in a discrete setting]
Suppose that $\tau_{\vert x}$ has been approximate, as a signed measure $\hat{\tau}_{\vert x} = \sum_{i=1}^n \alpha_i(x)\delta_{S_i}$.
After renormalization, one can represent it as a region $R_{\hat\tau_{\vert x}}$ in the affine span of $\Prob{\mathcal{Y}}$.
Retaking the settings of Section~\ref{il:sec:inconsistency}, suppose that
\[
\hat{\tau}(\brace{a,b}) = \frac{1}{2},\qquad\hat{\tau}(\brace{c}) = \frac{1}{2},\qquad
\hat{\tau}(\brace{a,c}) = \frac{1}{4},\qquad \hat{\tau}(\brace{a,b,c}) = -\frac{1}{4}.
\]
This corresponds to the region $R_{\hat\tau}$ represented in Figure~\ref{il:fig:consistency}.
It leads to a disambiguation $\hat{\rho}$ that minimizes ${\cal E}$~\eqref{il:eq:infimum-disambiguation}, inside this space as
\[
\hat{\rho}(a) = \frac{1}{2}, \qquad \hat{\rho}(b) = -\frac{1}{4}, \qquad \hat{\rho}(c) = \frac{3}{4},
\]
and to the right prediction $\hat{z}=c$, since $\hat{\rho}$ felt in the decision region $R_c$.
As the number of data augments, $R_{\hat\tau}$ converges toward $R_\tau$, so does $\hat{\rho}$ toward $\rho^\star$ and the risk ${\cal R}(\hat{f})$ toward its minimum.
\begin{figure}[t]
\centering
\input{infimum/drawings/consistency}
\vspace*{-.3cm}
\caption{Understanding convergence of the algorithm~\eqref{il:eq:algorithm}.
Our method is approximating $\tau$ as a signed measured $\hat\tau$, which leads to $R_{\hat{\tau}}$ in dark gray compared to the ground truth $R_{\tau}$ in light gray.
The disambiguation of $\hat\rho$ and $\rho^\star$ is done on those two domains with the same objective ${\cal E}$~\eqref{il:eq:infimum-disambiguation}, which level curves are represented with light lines.}
\label{il:fig:consistency}
\end{figure}
\end{remark}
\subsection{Understanding of the average and the supremum loss}
\label{il:app:other-losses}
For the average loss, if there is discrepancy in the loss $\nu > 0$, then there exists $a, b, c$ such that $\ell(b, c) = (1+\epsilon) \ell(a, b)$, for some $\epsilon > 0$.
In this case, one can recreate the example of Section~\ref{il:sec:inconsistency} by considering $\rho_0 = \rho^\star = \delta_c$ and
\[
\tau = \lambda \delta_{\brace{c}} + (1 - \lambda) \delta_{\brace{a, b, c}},
\qquad\text{with}\qquad
\lambda = \frac{1}{2}\frac{\epsilon}{3\ell(a, b) + \epsilon},
\]
to show the inconsistency of the average loss.
Similarly, supposing, without loss of generality that $\ell(a, c) \in \bracket{\ell(a, b), \ell(b, c)}$, the case where $\rho_0 = \rho^\star = \delta_b$ and
\[
\tau = \lambda \delta_{\brace{b}} + (1 - \lambda) \delta_{\brace{a, b, c}},
\qquad \text{with}\qquad
\lambda = \frac{1}{2} \min\paren{\frac{\epsilon}{1+\epsilon},
\frac{1+\epsilon - x}{2+\epsilon - x}},\qquad x=\frac{\ell(a,c)}{\ell(a, b)},
\]
will fail the supremum loss, which will recover $z^* = a$, instead of $z^* = b$.
\section{Introduction}
Fully supervised learning demands tight supervision of large amounts of data, a supervision that can be quite costly to acquire and constrains the scope of applications.
To overcome this bottleneck, the machine learning community is seeking to incorporate weaker sources of information in the learning framework.
In this paper, we address those limitations through partial labeling: {\em e.g.}, giving only partial ordering when learning user preferences over items, or providing the label ``flower'' for a picture of Arum Lilies, instead of spending a consequent amount of time to find the exact taxonomy.
Partial labeling has been studied in the context of classification~\citep{Cour2011,Nguyen2008}, multilabeling~\citep{Yu2014}, ranking~\citep{Hullermeier2008,Korba2018}, as well as segmentation~\citep{Verbeek2007,Papandreou2015}, however a generic framework is still missing.
Such a framework is a crucial step toward understanding how to learn from weaker sources of information, and widening the spectrum of machine learning beyond rigid applications of supervised learning.
Some interesting directions are provided by~\citet{CidSueiro2014,vanRooyen2018}, to recover the information lost in a corrupt acquisition of labels.
Yet, they assume that the corruption process is known, which is a strong requirement that we want to relax.
In this paper, we make the following contributions:
\begin{itemize}
\item We provide a principled framework to solve the problem of learning with partial labeling, via {\em structured prediction}.
This approach naturally leads to a variational framework built on the {\em infimum loss}.
\item We prove that the proposed framework is able to recover the original solution of the supervised learning problem under identifiability assumptions on the labeling process.
\item We derive an explicit algorithm which is easy to train and with strong theoretical guarantees.
In particular, we prove that it is consistent, and we provide generalization error rates.
\item Finally, we test our method against some simple baselines, on synthetic and real examples.
We show that for certain partial labeling scenarios with symmetries, our infimum loss performs similarly to a simple baseline.
However, in scenarios where the acquisition process of the labels is more adversarial in nature, the proposed algorithm performs consistently better.
\end{itemize}
\section{Partial labeling with infimum loss}\label{il:sec:partial-labeling} In this section, we introduce a statistical framework for partial labeling, and we show that it is characterized naturally in terms of risk minimization with the infimum loss.
First, let's recall some elements of fully supervised and weakly supervised learning.
{\em Fully supervised learning} consists in learning a function~${f\in\mathcal{Y}^\mathcal{X}}$ between an input space~$\mathcal{X}$ and an output space~$\mathcal{Y}$, given a joint distribution ${\rho \in \Prob{\mathcal{X} \times \mathcal{Y}}}$ on~${\mathcal{X} \times \mathcal{Y}}$, and a loss function ${\ell \in \mathbb{R}^{\mathcal{Y}\times\mathcal{Y}}}$, that minimizes the risk
\begin{equation}\label{il:eq:risk}
{\cal R}(f; \rho) = \E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)},
\end{equation}
given observations ${(x_i, y_i)_{i\leq n} \sim \rho^{\otimes n}}$.
We will assume that the loss~$\ell$ is proper, {\em i.e.} it is continuous non-negative and is zero on, and only on, the diagonal of~${\mathcal{Y} \times \mathcal{Y}}$, and strictly positive outside.
We will also assume that $\mathcal{Y}$ is compact.
In \emph{weakly supervised learning}, given~$(x_i)_{i\leq n}$, one does not have direct observations of~$(y_i)_{i\leq n}$ but weaker information.
The goal is still to recover the solution~${f \in \mathcal{Y}^\mathcal{X}}$ of the fully supervised problem~\eqref{il:eq:risk}.
In \emph{partial labeling}, also known as \emph{superset learning} or as \emph{learning with ambiguous labels}, which is an instance of weak supervision, information is cast as closed sets $(S_i)_{i\leq n}$ in ${\cal S}$, where ${\cal S}\subset 2^\mathcal{Y}$ is the space of closed subsets of $\mathcal{Y}$, containing the true labels ${(y_i \in S_i)}$.
In this paper, we model this scenario by considering a data distribution ${\tau \in \Prob{\mathcal{X} \times {\cal S}}}$, that generates the samples $(x_i, S_i)$.
We will denote $\tau$ as {\em weak distribution} to distinguish it from $\rho$.
Capturing the dependence on the original problem, $\tau$~must be compatible with $\rho$, a matching property that we formalize with the concept of eligibility.
\begin{definition}[Eligibility]\label{il:def:eligibility}
Given a probability measure $\tau$ on $\mathcal{X} \times {\cal S}$, a probability measure $\rho$ on $\mathcal{X} \times \mathcal{Y}$ is said to be eligible for $\tau$ (denoted by $\rho \vdash \tau$), if there exists a probability measure $\pi$ over $\mathcal{X} \times \mathcal{Y} \times {\cal S}$ such that $\rho$ is the marginal of $\pi$ over $\mathcal{X} \times \mathcal{Y}$, $\tau$ is the marginal of $\pi$ over $\mathcal{X} \times {\cal S}$, and, for $y\in\mathcal{Y}$ and $S\in{\cal S}$
\[
y\notin S \qquad \Rightarrow \qquad \PP_{\pi}\paren{S\,\middle\vert\,
Y=y} = 0.
\]
We will alternatively say that $\tau$ is a {\em weakening} of~$\rho$, or that $\rho$ and $\tau$ are {\em compatible}.
\end{definition}
\subsection{Disambiguation principle}\label{il:sec:disambiguation-principle}
According to the setting described above, the problem of partial labeling is completely defined by a loss and a weak distribution~$(\ell, \tau)$.
The goal is to recover the solution of the original supervised learning problem in~\eqref{il:eq:risk} assuming that the original distribution verifies $\rho\vdash\tau$.
Since more than one $\rho$ may be eligible for $\tau$, we would like to introduce a guiding principle to identify a $\rho^\star$ among them.
With this goal we define the concept of {\em non-ambiguity} for $\tau$, a setting in which a natural choice for $\rho^\star$ appears.
\begin{definition}[Non-ambiguity]\label{il:def:non-ambiguity}
For any $x \in \mathcal{X}$, denote by $\tau\vert_x$ the conditional probability of $\tau$ given $x$, and define the set $S_x$ as \[ S_x = \bigcap_{S \in \supp(\tau\vert_x)} S. \] The weak distribution $\tau$ is said {\em non-ambiguous} if, for every $x\in\mathcal{X}$, $S_x$ is a singleton.
Moreover, we say that $\tau$ is {\em strictly non-ambiguous} if it is non-ambiguous and there exists ${\eta\in (0,1)}$ such that, for all ${x \in \mathcal{X}}$ and ${z \notin S_x}$
\[\PP_{S \sim \tau\vert_x}(z \in S) \leq 1 - \eta.\]
\end{definition}
This concept is similar to the one by~\citet{Cour2011}, but more subtle because this quantity only depends on~$\tau$, and makes no assumption on the original distribution~$\rho$ describing the fully supervised process that we can not access.
In this sense, it is also more general.
When $\tau$ is non-ambiguous, we can write $S_x = \brace{y_x}$ for any $x$, where $y_x$ is the only element of $S_x$.
In this case it is natural to identify $\rho^\star$ as the one satisfying $\rho^\star\vert_x = \delta_{y_x}$.
Actually, such a $\rho^\star$ is characterized without $S_x$ as the only deterministic distribution that is eligible for $\tau$.
Because deterministic distributions are characterized as minimizing the minimum risk~\eqref{il:eq:risk}, we introduce the following {\em minimum variability principle} to disambiguate between all eligible $\rho$'s, and identify $\rho^\star$,
\begin{equation}\label{il:eq:infimum-disambiguation}
\rho^\star \in \argmin_{\rho\vdash \tau}{\cal E}(\rho),\qquad {\cal E}(\rho) = \inf_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}(f; \rho).
\end{equation}
The quantity ${\cal E}$ can be identified as a variance, since if $f_\rho$ is the minimizer of ${\cal R}(f; \rho)$, $f_\rho(x)$ can be seen as the mean of $\rho\vert_x$ and $\ell$ the natural distance in $\mathcal{Y}$.
Indeed, when $\ell = \ell_2$ is the mean square loss, this is exactly the case.
The principle above recovers exactly $\rho^\star\vert_x = \delta_{y_x}$, when $\tau$ is non-ambiguous, as stated by Theorem \ref{il:thm:ambiguity}, proven in Appendix \ref{il:proof:ambiguity}.
\begin{proposition}[Non-ambiguity determinism]\label{il:thm:ambiguity}
When $\tau$ is non-ambiguous, the solution $\rho^\star$ \eqref{il:eq:infimum-disambiguation} exists and satisfies that, for any $x\in\mathcal{X}$, $\rho^\star\vert_x = \delta_{y_x}$, where $y_x$ is the only element of $S_x$.
\end{proposition}
Proposition \ref{il:thm:ambiguity} provides a justification for the usage of the minimum variability principle.
Indeed, under non-ambiguity assumption, following this principle will allow us to build an algorithm that recovers the original fully supervised distribution.
Therefore, given samples $(x_i, S_i)$, it is of interest to test if $\tau$ is non-ambiguous.
Such tests should leverage other regularity hypotheses on $\tau$, which we will not address in this work.
Now, we characterize the minimum variability principle in terms of a variational optimization problem that we can tackle in Section \ref{il:sec:algorithm} via empirical risk minimization.
\subsection{Variational formulation via the infimum loss}
Given a partial labeling problem $(\ell, \tau)$, define the solutions based on the minimum variability principle as the functions minimizing the recovered risk
\begin{equation}\label{il:eq:fstar-of-rhostar}
f^* \in \argmin_{f:\mathcal{X} \to \mathcal{Y}} {\cal R}(f; \rho^\star).
\end{equation}
for $\rho^\star$ a distribution solving~\eqref{il:eq:infimum-disambiguation}.
As shown in Theorem \ref{il:thm:infimum-loss} below, proven in Appendix \ref{il:proof:infimum-loss}, the proposed disambiguation paradigm naturally leads to a variational framework involving the {\em infimum loss}.
\begin{theorem}[Infimum loss (\emph{IL})]\label{il:thm:infimum-loss}
The functions $f^*$ defined in~\eqref{il:eq:fstar-of-rhostar} are characterized as \[ f^* \in \argmin_{f:\mathcal{X} \to \mathcal{Y}} {\cal R}_S(f), \] where the risk ${\cal R}_S$ is defined as
\begin{equation}\label{il:eq:set-risk}
{\cal R}_S(f) = \E_{(X,S)\sim\tau}\bracket{L(f(X), S)},
\end{equation}
and $L$ is the {\em infimum loss}
\begin{equation}\label{il:eq:infimum-loss}
L(z, S) = \inf_{y\in S} \ell(z, y).
\end{equation}
\end{theorem}
The infimum loss, also known as the ambiguous loss \citep{Luo2010,Cour2011}, or as the optimistic superset loss \citep{Hullermeier2014}, captures the idea that, when given a set~$S$, this set contains the good label $y$ but also a lot of bad ones, that should not be taken into account when retrieving~$f$.
In other terms, $f$ should only match the best guess in $S$.
Indeed, if $\ell$ is seen as a distance, $L$ is its natural extension to sets.
\subsection{Recovery of the fully supervised solutions}
In this subsection, we investigate the setting where an original fully supervised learning problem $\rho_0$ has been weakened due to incomplete labeling, leading to a weak distribution~$\tau$.
The goal here is to understand under which conditions on $\tau$ and $\ell$ it is possible to recover the original fully supervised solution based on the infimum loss framework.
Denote $f_0$ the function minimizing ${\cal R}(f;\rho_0)$.
The theorem below, proven in Appendix \ref{il:proof:non-ambiguity}, shows that under non-ambiguity and deterministic conditions, it is possible to fully recover the function $f_0$ also from $\tau$.
\begin{theorem}[Supervision recovery]\label{il:thm:non-ambiguity}
For an instance $(\ell, \rho_0, \tau)$ of the weakened supervised problem, if we denote by $f_0$ the minimizer of~\eqref{il:eq:risk}, we have the under the conditions that (1)~$\tau$ is not ambiguous (2)~for all $x\in\mathcal{X}$, $S_x = \brace{f_0(x)}$; the infimum loss recovers the original fully supervised solution, {\em i.e.} the $f^*$ defined in~\eqref{il:eq:fstar-of-rhostar} verifies $f^* = f_0$.
Furthermore, when $\rho_0$ is deterministic and $\tau$ not ambiguous, the $\rho^\star$ defined in~\eqref{il:eq:infimum-disambiguation} verifies $\rho^\star = \rho_0$.
\end{theorem}
At a comprehensive level, this theorem states that under non-ambiguity of the partial labeling process, if the labels are a deterministic function of the inputs, the infimum loss framework makes it possible to recover the solution of the original fully supervised problem while only accessing weak labels.
In the next subsection, we will investigate which is the relation between the two problems when dealing with an estimator $f$ of $f^*$.
\subsection{Comparison inequality}\label{il:sec:calibration}
In the following, we want to characterize the error performed by ${\cal R}(f;\rho^\star)$ with respect to the error performed by ${\cal R}_S(f)$.
This will be useful since, in the next section, we will provide an estimator for $f^*$ based on structured prediction, that minimizes the risk ${\cal R}_S$.
First, we introduce a measure of discrepancy for the loss function.
\begin{definition}[Discrepancy of the loss $\ell$]
Given a loss function $\ell$, the {\em discrepancy degree} $\nu$ of $\ell$ is defined as
\[ \nu = \log\sup_{y, z'\neq z} \frac{\ell(z, y)}{\ell(z, z')}. \]
$\mathcal{Y}$ will be said discrete for $\ell$ when $\nu < +\infty$, which is always the case when $\mathcal{Y}$ is finite.
\end{definition}
Now we are ready to state the comparison inequality that generalizes a result on classification with the $0-1$ loss from \citet{Cour2011} to arbitrary losses and output spaces.
\begin{proposition}[Comparison inequality]\label{il:thm:calibration}
When $\mathcal{Y}$ is discrete and $\tau$ is strictly non-ambiguous for a given $\eta \in (0,1)$, then the following holds
\begin{equation}\label{il:eq:calibration-bound}
{\cal R}(f; \rho^\star) - {\cal R}(f^*;\rho^\star) \leq C({\cal R}_S(f) - {\cal R}_S(f^*)),
\end{equation}
for any measurable function $f \in \mathcal{Y}^\mathcal{X}$, where $C$ does not depend on $\tau, f$, and is defined as follows and always finite
\[ C = \eta^{-1} e^\nu.\]
\end{proposition}
When $\rho_0$ is deterministic, since we know from Theorem~\ref{il:thm:non-ambiguity} that $\rho^\star=\rho_0$, this theorem allows bounding the error made on the original fully supervised problem with the error measured with the infimum loss on the weakly supervised one.
Note that the constant presented above is the product of two independent terms, the first measuring the ambiguity of the weak distribution $\tau$, and the second measuring a form of discrepancy for the loss.
In the appendix, we provide a more refined bound for $C$, that is $C = C(\ell, \tau)$, that shows a more elaborated interaction between $\ell$ and $\tau$.
This may be interesting in situations where it is possible to control the labeling process and may suggest strategies to active partial labeling, with the goal of minimizing the costs of labeling while preserving the properties presented in this section and reducing the impact of the constant $C$ in the learning process.
An example is provided in the Appendix \ref{il:discussion:refinement-C}.
\section{Consistent algorithm for partial labeling}\label{il:sec:algorithm}
In this section, we provide an algorithmic approach based on structured prediction to solve the weak supervised learning problem expressed in terms of infimum loss from Theorem~\ref{il:thm:infimum-loss}.
From this viewpoint, we could consider different structured prediction frameworks as structured SVM \citep{Tsochantaridis2005}, conditional random fields \citep{Lafferty2001} or surrogate mean estimation \citep{Ciliberto2016}.
For example, \citet{Luo2010} used a margin maximization formulation in a structured SVM fashion, \citet{Hullermeier2015} went for nearest neighbors, and \citet{Cour2011} design a surrogate method specific to the 0-1 loss, for which they show consistency based on~\citet{Bartlett2006}.
In the following, we will use the structured prediction method of~\citet{Ciliberto2016,Nowak2019}, which allows us to derive an explicit estimator, easy to train and with strong theoretical properties, in particular, consistency and finite sample bounds for the generalization error.
The estimator is based on the pointwise characterization of $f^*$ as
\[
f^*(x) \in \argmin_{z\in\mathcal{Y}} \E_{S\sim \tau\vert_x}\bracket{\inf_{y\in S}\ell(z, y)},
\]
and weights $\alpha_i(x)$ that are trained on the dataset such that $\hat{\tau}_{\vert x} = \sum_{i=1}^n \alpha_i(x) \delta_{S_i}$ is a good approximation of $\tau\vert_x$.
Plugging this approximation in the precedent equation leads to our estimator, that is defined explicitly as follows
\begin{equation}\label{il:eq:algorithm}
f_n(x) \in \argmin_{z\in\mathcal{Y}} \inf_{y_i \in S_i} \sum_{i=1}^n \alpha_i(x) \ell(z, y_i).
\end{equation}
Among possible choices for $\alpha$, we will consider the following kernel ridge regression estimator to be learned at training time
\[
\alpha(x) = (K + n\lambda)^{-1}v(x),
\]
with $\lambda > 0$ a regularizer parameter and $K = (k(x_i, x_j))_{i,j} \in \mathbb{R}^{n\times n}, v(x) = (k(x, x_i))_{i} \in \mathbb{R}^n$ where $k\in \mathcal{X}\times\mathcal{X} \to \mathbb{R}$ is a positive-definite kernel \citep{Scholkopf2001} that defines a similarity function between input points ({\em e.g.}, if $\mathcal{X} = \mathbb{R}^d$ for some $d \in \mathbb{N}$ a commonly used kernel is the Gaussian kernel $k(x,x') = e^{-\|x-x'\|^2}$).
Other choices can be done to learn $\alpha$, beyond kernel methods, a particularly appealing one is harmonic functions, incorporating a prior on low density separation to boost learning \citep{Zhu2003,Zhou2003,Bengio2006}.
Here we use the kernel estimator since it allows deriving strong theoretical results, based on kernel conditional mean estimation \citep{Muandet2017}.
\subsection{Theoretical guarantees}
In this following, we want to prove that $f_n$ converges to $f^*$ as $n$ goes to infinity, and we want to quantify it with finite sample bounds.
The intuition behind this result is that as the number of data points tends toward infinity, $\hat{\tau}$ concentrates toward $\tau$, making our algorithm in~\eqref{il:eq:algorithm} converging to a minimizer of~\eqref{il:eq:set-risk} as explained more in detail in Appendix \ref{il:proof:consistency}.
\begin{theorem}[Consistency]\label{il:thm:consistency}
Let $\mathcal{Y}$ be finite and $\tau$ be a non-ambiguous probability.
Let $k$ be a bounded continuous universal kernel, {\em e.g.} the Gaussian kernel \citep[see][for details]{Micchelli2006}, and $f_n$ the estimator in~\eqref{il:eq:algorithm} trained on $n \in \mathbb{N}$ examples and with $\lambda = n^{-1/2}$.
Then, holds with probability~$1$
\[ \lim_{n \to \infty} {\cal R}(f_n; \rho^\star) = {\cal R}(f^*; \rho^\star).\]
\end{theorem}
In the next theorem, instead we want to quantify how fast $f_n$ converges to $f^*$ depending on the number of examples.
To obtain this result, we need a finer characterization of the infimum loss $L$ as:
\[
L(z, S) = \scap{\psi(z)}{\phi(S)},
\]
where ${\cal H}$ is a Hilbert space and $\psi: \mathcal{Y} \to {\cal H}, \phi: 2^{\mathcal{Y}} \to {\cal H}$ are suitable maps.
Such a decomposition always exists in finite case (as
for the infimum loss over $\mathcal{Y}$ finite) and many explicit examples for losses of interest are presented by~\citet{Nowak2019}.
We now introduce the conditional expectation of $\phi(S)$ given $x$, defined as
\[
\myfunction{g}{\mathcal{X}}{\cal H}{x}{\E_{\tau}\bracket{\phi(S)\,\middle\vert\, X = x}.}
\]
The idea behind the proof is that the distance between $f_n$ and $f$ is bounded by the distance of $g_n$ an estimator of $g$ that is implicitly computed via $\alpha$.
If $g$ has some form of regularity, {\em e.g.} $g \in {\cal G}$, with ${\cal G}$ the space of functions representable by the chosen kernel \citep[see][]{Scholkopf2001}, then it is possible to derive explicit rates, as stated in the following theorem.
\begin{theorem}[Convergence rates]\label{il:thm:learning-rates}
In the setting of Theorem~\ref{il:thm:consistency}, if $\tau$ is $\eta$-strictly non-ambiguous for $\eta\in(0, 1)$, and if $g \in {\cal G}$, then there exists a $\tilde{C}$, such that, for any $\delta \in (0, 1)$ and $n\in \mathbb{N}$, holds with probability at least $1 - \delta$,
\begin{equation}\label{il:eq:rates}
{\cal R}(f_n; \rho^\star) - {\cal R}(f^*; \rho^\star) \leq
\tilde{C} \log\paren{\frac{8}{\delta}}^2 n^{-1/4}.
\end{equation}
\end{theorem}
Those last two theorem are proven in Appendix \ref{il:proof:consistency} and combines the consistency and learning results for kernel ridge regression \citep{Caponnetto2007,Smale2007}, with a comparison inequality of~\citet{Ciliberto2016} which relates the excess risk of the structured prediction problem with the one of the surrogate loss ${\cal R}_S$, together with our Theorem \ref{il:thm:calibration}, which relates the error ${\cal R}$ to ${\cal R}_S$.
Those results make our algorithm the first algorithm for partial labeling that, to our knowledge, is applicable to a generic loss $\ell$ and has strong theoretical guarantees as consistency and learning rates.
In the next section we will compare with the state of the art and other variational principles.
\section{Previous works and baselines}
\label{il:sec:inconsistency}
Partial labeling was first approached through discriminative models, proposing to learn $\paren{Y \,\middle\vert\, X}$ among a family of parametrized distributions by maximizing the log likelihood based on expectation-maximization scheme~\citep{Jin2002}, eventually integrating knowledge on the partial labeling process~\citep{Grandvalet2002,Papandreou2015}.
In the meanwhile, some applications of clustering methods have involved special instances of partial labeling, like segmentation approached with spectral method~\citep{Weiss1999}, semi-supervision approached with max-margin~\citep{Xu2004}.
Also, initially geared toward clustering,~\citet{Bach2007} considered the infimum principle on the mean square loss, and this was generalized to weakly supervised problems~\citep{Joulin2010}.
The infimum loss as an objective to minimize when learning from partial labels was introduced by~\citet{Cour2011} for the classification instance and used by~\citet{Luo2010,Hullermeier2014} in generic cases.
Compared to those last two, we provide a framework that derives the use of infimum loss from first principles and from which we derive an explicit and easy to train algorithm with strong statistical guarantees, which were missing in previous work.
In the rest of the section, we will compare the infimum loss with other variational principles that have been considered in the literature, in particular the supremum loss~\citep{Guillaume2017} and the average loss~\citep{Denoeux2013}.
\paragraph{Average loss (\emph{AC}).}
A simple loss to deal with uncertainty is to average over all potential candidates, assuming $S$ discrete,
\[
L_{\textit{ac}}(z, S) = \frac{1}{\module{S}} \sum_{y\in S} \ell(z, y).
\]
It is equivalent to a fully supervised distribution $\rho_{\textit{ac}}$ by sampling $Y$ uniformly at random among $S$
\[ \rho_{\textit{ac}}(y) = \int_{{\cal S}} \frac{1}{\module{S}} \ind{y\in S} \mathop{}\!\mathrm{d}\tau(S).\]
This directly follows from the definition of $L_{\textit{ac}}$ and of the risk ${\cal R}(z; \rho_{\textit{ac}})$.
However, as soon as the loss $\ell$ has discrepancy, {\em i.e.} $\nu > 0$, the average loss will implicitly advantage some labels, which can lead to inconsistency, even in the deterministic not ambiguous setting of Theorem \ref{il:thm:calibration} (see Appendix \ref{il:app:other-losses} for more details).
\paragraph{Supremum loss (\emph{SP}).}
Another loss that has been considered is the supremum loss~\citep{Wald1945,Madry2018}, bounding from above the fully supervised risk in~\eqref{il:eq:risk}.
It is widely used in the context of robust risk minimization and reads
\[
R_{\textit{sp}}(f) = \sup_{\rho\vdash\tau} \E_{(X,Y)\sim\rho}\bracket{\ell(f(x), S)}.
\]
Similarly to the infimum loss in Theorem~\ref{il:thm:infimum-loss}, this risk can be written from the loss function
\[
L_{\textit{sp}}(z, S) = \sup_{y\in S} \ell(z, y).
\]
Yet, this adversarial approach is not consistent for partial labeling, even in the deterministic non-ambiguous setting of Theorem \ref{il:thm:calibration}, since it finds the solution that best agrees with {\em all} the elements in $S$ and not only the true one (see Appendix \ref{il:app:other-losses} for more details).
\subsection{Instance showcasing superiority of our method}
In the rest of this section, we consider a pointwise example to showcase the underlying dynamics of the different methods.
It is illustrated in Figure~\ref{il:fig:inconsistency}.
Consider $\mathcal{Y} = \brace{a, b, c}$ and a proper symmetric loss function such that $\ell(a, b) = \ell(a, c) = 1$, $\ell(b, c) = 2$.
The simplex $\Prob{\mathcal{Y}}$ is naturally split into decision regions, for $e\in\mathcal{Y}$,
\[
R_e = \brace{\rho\in\Prob{\mathcal{Y}} \,\middle\vert\, e\in\argmin_{z\in\mathcal{Y}}\E_{\rho}[\ell(z, Y)]}.
\]
Both {\em IL} and {\em AC} solutions can be understood geometrically by looking at where $\rho^\star$ and $\rho_{\textit{ac}}$ fall in the partition of the simplex $(R_e)_{e\in\mathcal{Y}}$.
Consider a fully supervised problem with distribution $\delta_c$, and a weakening $\tau$ of $\rho$ defined by $\tau(\brace{a, b, c}) = \frac{5}{8}$ and $\tau(\brace{c}) = \tau(\brace{a,c}) = \tau(\brace{b,c}) = \frac{1}{8}$.
This distribution can be represented on the simplex in terms of the region $R_\tau = \brace{\rho\in\Prob{\mathcal{Y}}\,\middle\vert\, \rho\vdash \tau}$.
Finding $\rho^\star$ correspond to minimizing the piecewise linear function $ {\cal E}(\rho)$~\eqref{il:eq:infimum-disambiguation} inside $R_\tau$.
On this example, it is minimized for $\rho^\star = \delta_c$, which we know from
Theorem~\ref{il:thm:calibration}.
Now note that if we use the average loss, it disambiguates $\rho$ as
\[
\rho_{\textit{ac}}(c) = \frac{11}{24} =
\frac{1}{3}\frac{5}{8} + \frac{1}{8} + 2\cdot\frac{1}{2}\frac{1}{8}, \quad
\rho_{\textit{ac}}(b) = \rho_{\textit{ac}}(a) = \frac{13}{48}.
\]
This distribution falls in the decision region of $a$, which is inconsistent with the real label $y=c$.
For the supremum loss, one can show, based on ${\cal R}_{\textit{sp}}(a) = \ell(a, c) = 1$, ${\cal R}_{\textit{sp}}(b) = \ell(b, c) = 2$ and ${\cal R}_{\textit{sp}}(c) = 3/2$, that the supremum loss is minimized for $z = a$, which is also inconsistent.
Instead, by using the infimum loss, we have $f^* = f_0 = c$, and moreover that $\rho^\star = \rho_0$ that is the optimal one.
\begin{figure}[t]
\centering
\input{infimum/drawings/inconsistency}
\vspace*{-.3cm}
\caption{Simplex $\Prob{\mathcal{Y}}$.
(Left) Decision frontiers.
(Middle left) Full and weak distributions.
(Middle right) Level curves of the piecewise linear objective ${\cal E}$~\eqref{il:eq:infimum-disambiguation}, to optimize when disambiguating $\tau$ into $\rho^\star$.
(Right) Disambiguation of \emph{AC} and \emph{IL}.}
\label{il:fig:inconsistency}
\end{figure}
\subsection{Algorithmic considerations for AC, SP}
The averaging candidates principle, approached with the framework of quadratic surrogates \citep{Ciliberto2016}, leads to the following algorithm
\begin{align*}
f_{\textit{ac}}(x) & \in \argmin_{z\in\mathcal{Y}}
\sum_{i=1}^n \alpha_{i}(x)\frac{1}{\module{S_i}}\sum_{y\in S_i} \ell(z, y)
\\ &= \argmin_{z\in\mathcal{Y}} \sum_{y\in\mathcal{Y}}
\paren{\sum_{i=1}^n \mathbf{1}_{y\in S_i}\frac{\alpha_i(x)}{ \module{S_i}}} \ell(z,y).
\end{align*}
This estimator is computationally attractive because the inference complexity is the same as the inference complexity of the original problem when approached with the same structured prediction estimator.
Therefore, one can directly reuse algorithms developed to solve the original inference problem~\citep{Nowak2019}.
Finally, with a similar approach to the one in Section \ref{il:sec:algorithm}, we can derive the following algorithm for the supremum loss
\[
f_{\textit{sp}}(x) \in \argmin_{z\in\mathcal{Y}}
\sup_{y_i \in S_i}\sum_{i=1}^n \alpha_i(x) \ell(z, y_i).
\]
In the next section, we will use the average candidates as baseline to compare with the algorithm proposed in this paper, as the supremum loss consistently performs worth, as it is not fitted for partial labeling.
\section{Applications and experiments}\label{il:sec:application}
In this section, we will apply~\eqref{il:eq:algorithm} to some synthetic and real datasets from different prediction problems and compared with the average estimator presented in the section above, used as a baseline.
Code is available online.\footnote{\url{https://github.com/VivienCabannes/partial_labelling}}
\subsection{Classification}\label{il:sec:classification}
Classification consists in recognizing the most relevant item among $m$ items.
The output space is isomorphic to the set of indices $\mathcal{Y}=\bbracket{1, m}$, and the usual loss function is the 0-1 loss
\[
\ell(z, y) = \ind{y\neq z}.
\]
It has already been widely studied with several approaches that are calibrated in non-ambiguous deterministic settings, notably by~\citet{Cour2011}.
The infimum loss reads $L(z, S) = \ind{z\notin S}$, and its risk in~\eqref{il:eq:set-risk} is minimized for
\[
f(x) \in \argmax_{z\in \mathcal{Y}} \PP\paren{z\in S \,\middle\vert\, X=x}.
\]
Based on data $(x_i, S_i)_{i\leq n}$, our estimator~\eqref{il:eq:algorithm} reads
\[
f_n(x) = \argmax_{z\in\mathcal{Y}} \sum_{i;z\in S_i} \alpha_i(x).
\]
For this instance, the supremum loss is really conservative, only learning from set that are singletons $L_{\textit{sp}}(z, S) = \ind{S\neq\brace{z}}$, while the average loss is similar to the infimum one, adding an evidence weight depending on the size of $S$, $L_{\textit{ac}}(z, S) \simeq \ind{z\notin S} / \module{S}$.
\begin{figure}[t]
\centering
\includegraphics[width=.45\textwidth]{infimum/images/cl/dna.pdf}
\includegraphics[width=.45\textwidth]{infimum/images/cl/svmguide2.pdf}
\vspace*{-.3cm}
\caption{Classification.
Testing risks (from~\eqref{il:eq:risk}) achieved by {\em AC} and {\em IL} on the ``DNA'' and ``svmguide2'' datasets from {\em LIBSVM} as a function of corruption parameter $c$, when the corruption is as follows: for $y$ being the most present labels of the dataset, and $z'\neq z$, $\PP\paren{z'\in S\,\middle\vert\, Y=z} = c\cdot \ind{z=y}$.
Plotted intervals show the standard deviation on eight-fold cross-validation.
Experiments were done with the Gaussian kernel.
See all experimental details in~\eqref{il:app:experiments}.}
\label{il:fig:libsvm}
\end{figure}
\paragraph{Real data experiment.} To compare {\em IL} and {\em AC}, we used {\em LIBSVM} datasets~\citep{Chang2011} on which we corrupted labels to simulate partial labeling.
When the corruption is uniform, the two methods perform the same.
Yet, when labels are unbalanced, such as in the ``DNA'' and ``svmguide2'' datasets, and we only corrupt the most frequent label $y\in\mathcal{Y}$, the infimum loss performs better as shown in Figure~\ref{il:fig:libsvm}.
\subsection{Ranking}\label{il:sec:ranking}
Ranking consists in ordering $m$ items based on an input~$x$ that is often the conjunction of a user $u$ and a query $q$, ($x=(u,q)$).
An ordering can be thought of as a permutation, that is, $\mathcal{Y}=\mathfrak{S}_m$.
While designing a loss for ranking is intrinsically linked to a voting system \citep{Arrow1950}, making it a fundamentally hard problem; \citet{Kemeny1959} suggested approaching it through pairwise disagreement, which is current machine learning standard~\citep{Duchi2010}, leading to the Kendall embedding
\[
\phi(y) = \paren{\sign\paren{y_i - y_j}}_{i<j \leq m},
\]
and the Kendall loss~\citep{Kendall1938}, with $C = m(m-1)/2$
\[
\ell(y, z) = C - \phi(y)^T\phi(z).
\]
Supervision often comes as partial order on items, {\em e.g.},
\[
S = \brace{y\in\mathfrak{S}_m \,\middle\vert\, y_i > y_j > y_k, y_l > y_m}.
\]
It corresponds to fixing some coordinates in the Kendall embedding.
In this setting, \emph{AC} and \emph{SP} are not consistent, as one can recreate a similar situation to the one in Section~\ref{il:sec:inconsistency}, considering $m=3$, $a = (1,2,3)$, $b=(2,1,3)$ and $c=(1,3,2)$ (permutations being represented with $(\sigma^{-1}(i))_{i\leq m}$), and supervision being most often $S = (1>3) = \brace{a,b,c}$ and sometimes $S = (1>3>2) = \brace{c}$.
\paragraph{Minimum feedback arc set.}
Dealing with Kendall's loss requires solving problem of the form,
\[
\argmin_{y \in S} \scap{c}{\phi(y)},
\]
for $c\in\mathbb{R}^{m^2}$, and constraints due to partial ordering encoded in $S\subset\mathcal{Y}$.
This problem is an instance of the constrained minimum feedback arc set problem.
We provide a simple heuristic to solve it in Appendix~\ref{il:app:fas}, which consists of approaching it as an integer linear program.
Such heuristics are analyzed and refined for analysis purposes by~\citet{Ailon2005,vanZuylen2007}.
\paragraph{Algorithm specification.}
At inference, the infimum loss requires solving:
\begin{equation}\tag{\ref{il:eq:algorithm}}
f_n(x) = \argmax_{z\in\mathcal{Y}} \sup_{(y_i) \in S_i} \sum_{i=1}^n \alpha_i(x) \scap{\phi(z)}{\phi(y_i)}.
\end{equation}
It can be approached with alternate minimization, initializing $\phi(y_i) \in \hull(\phi(S_i))$, by putting $0$ on unseen observed pairwise comparisons, then, iteratively, solving a minimum feedback arc set problem in $z$, then solving several minimum feedback arc set problems with the same objective, but different constraints in $(y_i)$.
This is done efficiently using warm start on the dual simplex algorithm.
\begin{figure}[t]
\centering
\includegraphics{infimum/images/rk/setting.pdf}
\vspace*{-.3cm}
\caption{Ranking, experimental setting.
Colors represent four different items to rank.
Each item is associated with a utility function of $x$ shown on the left figure.
From those scores, an ordering $y$ of the items is retrieved as represented on the right.}
\label{il:fig:rk:setting}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics{infimum/images/rk/corruption.pdf}
\vspace* {-.3cm}
\caption{Ranking, results.
Testing risks~(from~\eqref{il:eq:risk}) achieved by {\em AC} and {\em IL} as a function of corruption parameter $c$.
When $c=1$, both risks are similar at $0.5$.
The simulation setting is the same as in Figure~\ref{il:fig:libsvm}.
The error bars are defined as for Figure~\ref{il:fig:libsvm}, after cross-validation over eight folds.
{\em IL} clearly outperforms {\em AC}.}
\label{il:fig:rk:corruption}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=.45\textwidth]{infimum/images/ir/setting.pdf}
\includegraphics[width=.45\textwidth]{infimum/images/ir/reconstruction.pdf}
\vspace* {-.3cm}
\caption{Partial regression on $\mathbb{R}$.
In this setting we aim at recovering a signal $y(x)$ given upper and lower bounds on its amplitude, and in thirty percent of case, information on its phase, or equivalently in $\mathbb{R}$, its sign.
{\em IL} clearly outperforms the baseline.
Indeed, {\em AC} is a particularly ill-fitted method on such a problem, since it regresses on the barycenter of the resulting sets.}
\label{il:fig:interval-regression}
\end{figure*}
\paragraph{Synthetic experiments.}
Let us consider $\mathcal{X} = [0,1]$ embodying some input features.
Let $\{1,\dots,m\}$, $m \in \mathbb{N}$ be abstract items to order, each item being linked to a utility function $v_i \in \mathbb{R}^\mathcal{X}$, that characterizes the value of $i$ for $x$ as $v_i(x)$.
Labels $y(x)\in\mathcal{Y}$ are retrieved by sorting $(v_i(x))_{i\leq m}$.
To simulate a problem instance, we set $v_i$ as $v_i(x) = a_i\cdot x + b_i$, where $a_i$ and $b_i$ follow a standard normal distribution.
Such a setting is illustrated in Figure~\ref{il:fig:rk:setting}.
After sampling $x$ uniformly on $[0, 1]$ and retrieving the ordering $y$ based on scores, we simulate partial labeling by randomly losing pairwise comparisons.
The comparisons are formally defined as coordinates of the Kendall's embedding $(\phi(y)_{jk})_{jk\leq m}$.
To create non-symmetric perturbations we corrupt more often items whose scores differ a lot.
In other words, we suppose that the partial labeling focuses on pairs that are hard to discriminate.
The corruption is set upon a parameter $c\in[0,1]$.
In fact, for $m=10$, until $c=0.5$, our corruption is fruitless since it can most often be inverted based on transitivity constraints in ordering, while the problem becomes non-trivial with $c \geq 0.5$.
In the latter setting, {\em IL} clearly outperforms {\em AC} on Figure~\ref{il:fig:rk:corruption}.
\subsection{Partial regression}\label{il:sec:partial-regression} Partial regression is an example of non-discrete partial labeling problem, where $\mathcal{Y}=\mathbb{R}^m$ and the usual loss is the Euclidean distance
\[
\ell(y, z) = \norm{y - z}^2.
\]
This partial labeling problem consists of regression where observations are sets $S\subset \mathbb{R}^m$ that contain the true output $y$ instead of $y$.
Among others, it arises for example in economical models, where bounds are preferred over approximation when acquiring training labels~\citep{Tobin1958}.
As an example, we will illustrate how partial regression could appear for some phase problems arising with physical measurements.
Suppose a physicist wants to measure the law between a vector quantity $Y$ and some input parameters~$X$.
Suppose that, while she can record the input parameters $x$, her sensors do not exactly measure $y$ but render an interval in which the amplitude $\norm{y}$ lays and only occasionally render its phase $y / \norm{y}$, in a fashion that leads to a set of candidates $S$ for $y$.
The geometry over $\ell^2$ makes it a perfect example to showcase superiority of the infimum loss as illustrated in Figure~\ref{il:fig:interval-regression}.
In this figure, we consider ${\cal Y} = \mathbb{R}$ and suppose that $Y$ is a deterministic function of $X$ as shown by the dotted blue line signal.
If, for a given $x_i$, measurements only provides that $\module{y_i} \in [1, 2]$ without the sign of $y_i$, a situation where the phase is lost, this corresponds to the set $S_i = [-2, -1] \cup [1, 2]$, explaining the shape of observed sets that are symmetric around the origin.
Whenever the acquired data has no phase, which happens seventy percent of the time in our simulation, AC will target the set centers, explaining the green curve.
On the other hand, IL is aiming at passing by each set, which explains the orange curve, crossing all blue bars.
\section{Conclusions}
In this paper, we deal with the problem of weakly supervised learning, beyond standard regression and classification, focusing on the more general case of arbitrary loss functions and structured prediction.
We provide a principled framework to solve the problem of learning with partial labeling, from which a natural variational approach based on the infimum loss is derived.
We prove that under some identifiability assumptions on the labeling process the framework is able to recover the solution of the original supervised learning problem.
The resulting algorithm is easy to train and with strong theoretical guarantees.
In particular, we prove that it is consistent, and we provide generalization error rates.
Finally, the algorithm is tested on simulated and real datasets, showing that when the acquisition process of the labels is more adversarial in nature, the proposed algorithm performs consistently better than baselines.
This paper focuses on the problem of partial labeling, however the resulting mathematical framework is quite flexible in nature, and it is interesting to explore the possibility to extend it to tackle also other weakly supervised problems, as imprecise labels from non-experts~\citep{Dawid1979}, more general constraints over the set $(y_i)_{i\leq n}$ \citep{Quadrianto2009} or semi-supervision \citep{Chapelle2006}.
\chapter{Infimum Loss}
The following is a reproduction of \cite{Cabannes2020}.
\input{infimum/abstract}
\input{infimum/core}
\begin{subappendices}
\chapter*{Appendix}
\addcontentsline{toc}{chapter}{Appendix}
\input{infimum/appendix/proof}
\input{infimum/appendix/experiments}
\input{infimum/appendix/fas}
\end{subappendices}
\chapter{Forewords}
\section{Acknowledgement}
This thesis would not have been the same without the brilliant minds of Francis Bach and Alessandro Rudi, who have constantly impressed me by the speed at which they formed deep thoughts on any subject I would bring up in meetings.
I consider myself really fortunate to have worked under your direction.
Beside your sagacity, you have been a great source of inspiration as a person: always wise and caring, even when I was wondering about life in uncalled-for fashions, I was abusing academic freedom, or I was claiming false statements.
My second thoughts go for my colleagues at INRIA Paris.
I am particularly grateful to Alex Nowak for our many conversations around structured prediction, and to Yann Labb\'e for his continuous support.
My last year as a PhD student has been highly enjoyable, which correlates with the arrival of Quentin Le Lidec in my office.
Finally, I would like to warmly thank Guillaume Lecu\'e and Eyke H\"ullermeier for having accepted to review this manuscript, and especially considering how valuable your research hence your time is.
In particular, Eyke, you were the first person to give me feedback on my work, which I have experienced as an accolade that marked my entry in the research world.
Une thèse de doctorat représente un accomplissement académique certain.
Ma reconnaissance s'exprime également envers toutes les entités et les personnes extérieures qui m'ont permis d'y accéder.
Tout d'abord tous les professeurs qui ont su me transmettre leur enthousiasme, en particulier Xavier Lacroix et ses merveilleuses digressions.
Rétrospectivement, je m'étonne du dévouement et de la pédagogie remarquable d'un grand nombre d'entre eux, l'exemple le plus frappant étant celui d'Alain Camanès qui a su m'enseigner la rigueur nécessaire pour entrer dans le monde des sciences du supérieur.
Enfin, mon parcours n'aurait pas été le même sans la bienveillance d'un certain nombre de personnes qui ont cru en moi plus que moi-même n'y croyait, me laissant nombre de vifs souvenirs.
Remercions également les institutions qui font de leur mieux pour permettre ce genre d'entreprises, notamment en les finançant.
En particulier, cette thèse a été financée par le gouvernement français via le programme \guillemotleft~Investissements d'avenir~\guillemotright\ de l'Agence Nationale de la Recherche (référence ANR-19-P3IA-0001).
Francis et Alessandro sont également soutenus financièrement par l'European Research Council (bourses SEQUOIA 724063 et REAL 94790).
Cet accomplissement s'inscrit également dans une histoire plus intime.
Celle-ci s'écrit depuis bien avant ma naissance, et si je peux ici ressusciter les morts pour les remercier, je le ferai pour l'importance qu'ils ont porté à l'éveil, l'instruction et la curiosité, des valeurs qu'à leur suite, mon père et ma grand-mère maternelle ont tout particulièrement souhaite m'inculquer.
Je dois également à ma famille de précieux sentiments, nourris par le souvenir, qui me fournissent souvent les aspirations nécessaires pour repartir quand la difficulté se fait trop présente en moi.
J'espère qu'en retour, je participe un peu au bonheur et au développement de chacun.
Si ce court texte ne permet pas de remercier personnellement toutes les personnes qui m'importent, je n'oublierai ni mon papy qui m'a hébergée pendant ces trois dernières années, ni ma mère qui s'est occupée de moi si longtemps.
À la famille s'ajoutent mes amis.
J'aimerai les remercier tous, qui sont des éléments constitutifs de ma personne.
En particulier, je citerai Thomas Kerdreux pour sa magnanimité et Charles Arnal pour ses pertinents conseils.
Tous deux ont une vigueur intellectuelle et physique qui ne cesse de m'impressionner.
\subsection{Résumé}
Les mathématiques appliquées et le calcul nourrissent beaucoup d'espoirs à la suite des succès récents de l'apprentissage supervisé.
Dans l'industrie, beaucoup d'ingénieurs cherchent à remplacer leurs anciens paradigmes de pensée par l'apprentissage machine.
Étonnamment, ces ingénieurs passent plus de temps à collecter, annoter et nettoyer des données qu'à raffiner des modèles.
Ce phénomène motive la problématique de cette thèse: peut-on définir un cadre théorique plus général que l'apprentissage supervisé pour apprendre grâce à des données hétérogènes?
%
Cette question est abordée via le concept de supervision faible, faisant l'hypothèse que le problème que posent les données est leur annotation.
On modélise la supervision faible comme l'accès, pour une entrée donnée, non pas d'une sortie claire, mais d'un ensemble de sorties potentielles.
On plaide pour l'adoption d'une perspective \guillemotleft~optimiste~\guillemotright\ et l'apprentissage d'une fonction qui vérifie la plupart des observations.
Cette perspective nous permet de définir un principe pour lever l'ambiguïté des informations faibles.
On discute également de l'importance d'incorporer des techniques sans supervision d'appréhension des données d'entrée dans notre théorie, en particulier de compréhension de la variété sous-jacente via des techniques de diffusion, pour lesquelles on propose un algorithme réaliste afin d'éviter le fléau de la dimension, à l'inverse de ce qui existait jusqu'alors.
%
Enfin, nous nous attaquons à la question de collecte active d'informations faibles, définissant le problème de \guillemotleft~catalogage en ligne~\guillemotright, où un intendant doit acquérir une maximum d'informations fiables sur ses données sous une contrainte de budget.
Entre autres, nous tirons parti du fait que pour obtenir un gradient stochastique et effectuer une descente de gradient, il n'y a pas besoin de supervision totale.
\subsection{Abstract}
Applied mathematics and machine computations have raised a lot of hope since the recent success of supervised learning.
Many practitioners in industries have been trying to switch from their old paradigms to machine learning.
Interestingly, those data scientists spend more time scrapping, annotating and cleaning data than fine-tuning models.
This thesis is motivated by the following question: can we derive a more generic framework than the one of supervised learning in order to learn from clutter data?
This question is approached through the lens of weakly supervised learning, assuming that the bottleneck of data collection lies in annotation.
We model weak supervision as giving, rather than a unique target, a set of target candidates.
We argue that one should look for an ``optimistic'' function that matches most of the observations.
This allows us to derive a principle to disambiguate partial labels.
We also discuss the advantage to incorporate unsupervised learning techniques into our framework, in particular manifold regularization approached through diffusion techniques, for which we derived a new algorithm that scales better with input dimension then the baseline method.
Finally, we switch from passive to active weakly supervised learning, introducing the ``active labeling'' framework, in which a practitioner can query weak information about chosen data.
Among others, we leverage the fact that one does not need full information to access stochastic gradients and perform stochastic gradient descent.
\section{Summary of contributions}
\begin{enumerate}
\item The main contribution of this thesis is to provide a generic framework to deal with partial supervision.
Along this framework, we provide a consistent algorithm for any structured prediction problem \citep{Cabannes2020}; a principled algorithm to disambiguate weak information into full supervision which go beyond the partial supervision setting \citep{Cabannes2021}, as well as a statistically and computationally efficient way to incorporate Laplacian regularization \citep{Cabannes2021c}.
\item We provide a method to derive fast rates for any discrete output problem \citep{Cabannes2021b} based on least-squares surrogate. Although geared toward least-squares surrogate with Tikhonov regularization, those derivations can easily be extended to self-concordant losses based on \cite{MarteauFerey2019} and to any spectral filtering techniques based on \cite{Lin2020}.
We hope that this could be useful for statistical learning people, and we wish to see in the future similar results for other losses, as well as a better theory to compare surrogate problems for classification.
As a first step in this direction, we provide some derivations for the hinge loss \citep{Cabannes2022b}.
\item We provide a statistically and computationally efficient method to approach Laplacian regularization \citep{Cabannes2021c}.
Our method could be useful for a vast number of applications. Our low-rank approximation could help scale up exciting methods for sampling based on Langevin dynamics \citep{PillaudVivien2020}, Gaussian processes with derivatives \citep{Eriksson2018}, sparse models based on regularization with derivatives \citep{Rosasco2013}.
The fact that we ``kernelized'' Laplacian regularization, making it statistically superior to existing local averaging methods, is exciting when considering the impact of ``local'' methods such as diffusion maps \citep{Coifman2006}.
\item We introduce the active labeling problem, which unify several problems of searching for information through imprecise query, and might be useful to deal with privacy issues.
We provide a stochastic gradient descent method that does not require full supervision to tackle this active weakly supervised learning problem that we formalized in \cite{Cabannes2022}.
\end{enumerate}
\subsection{List of publications}
On supervised learning, reproduced in Part \ref{part:discrete}.
\begin{itemize}
\item \citep{Cabannes2021b} Vivien Cabannes, Alessandro Rudi, and Francis Bach. Fast rates in structured prediction. In {\em Conference on Learning Theory}, 2021.
\item \citep{Cabannes2022b} Vivien Cabannes and Stefano Vigogna. A case of exponential convergence rates for SVM. {\em In preparation} 2022.
\end{itemize}
\noindent
On partial supervision, reproduced in Parts \ref{part:weakly} and \ref{part:collection}.
\begin{itemize}
\item \citep{Cabannes2020} Vivien Cabannes, Alessandro Rudi, and Francis Bach. Structured prediction with partial labeling through the infimum loss. In {\em International Conference on Machine Learning}, 2020.
\item \citep{Cabannes2021c} Vivien Cabannes, Alessandro Rudi, and Francis Bach. Disambiguation of weak supervision with exponential convergence rates. In {\em International Conference on Machine Learning}, 2021.
\item \citep{Cabannes2021c} Vivien Cabannes, Loucas Pillaud-Vivien, Francis Bach, and Alessandro Rudi. Overcoming the curse of dimensionality with Laplacian regularization in semi-supervised learning. In {\em Neural Information Processing Systems}, 2021.
\item \citep{Cabannes2022} Vivien Cabannes, Francis Bach, Vianney Perchet, and Alessandro Rudi. Active labeling: streaming stochastic gradients. {\em In preparation} 2022.
\end{itemize}
\noindent
Some artistic reflections formatted for workshops (not reproduced).
\begin{itemize}
\item \citep{Cabannes2019} Vivien Cabannes, Thomas Kerdreux, Louis Thiry, and Tina \& Charly. Dialog on a canvas with a machine. In {\em NeurIPS workshop on Creativity}, 2019.
\item \citep{Cabannes2020b} Vivien Cabannes, Thomas Kerdreux, Louis Thiry. Diptychs of human and machine perception. {\em In NeurIPS workshop on Creativity}, 2020.
\end{itemize}
\noindent
Side works done for French reviews (not reproduced).
\begin{itemize}
\item Vivien Cabannes. Perquisition de données sur le cloud : les Etats-Unis en avance, l’Europe à la traîne. In {\em le Grand Continent}, 2019.
\item \citep{Cabannes2022c} Vivien Cabannes. Le futur du numérique sera-t-il incarné ? {\em Esprit}, 487:117-125, 2022.
\end{itemize}
\section{Research chronology}
This section traces back some of my research history, before summarizing the contributions of this thesis.
I wish it to be useful in order to understand the thought process behind this thesis and to ease its reading.
The thesis originated from the motivations of my supervisors to confront the interest of Francis for weak supervision with the least-squares surrogate method of Alessandro.
My interest was sparked by the fact that I wondered how to better incorporate fuzzy human knowledge into learning processes.
Many abstractions can be learned in a hierarchical fashion, {\em e.g.} first recognizing edges on input images, then aggregating edges into tires, doors, handles, and those last components into cars, fridges, bridges, before understanding the semantic of an image and outputting a wanted label.
Often, we have a good understanding of this hierarchical structure, but it is hard to formulate it clearly in order to guide the learning process.
In this thesis, we focus on weak supervision, which is understood as specifying crucial elements in the last layer of abstraction before labels, such as specifying some animal attributes that allow us to discriminate the animals that our algorithm should learn to recognize.
Other priors on the problem to solve are assumed to be incorporated into the model or into regularizers.
Note that priors are distinct from supervision in the sense that they are given once and for all, while supervision can be refined ({\em e.g.} samples $(X_i, Y_i)_{i\leq n}$ are a function of $n\in\mathbb{N}$ that can be made larger).
From a theoretical point of view, successes of machine learning are understood as successes of supervised learning, which are successes of statistics when given enough clean data.
Yet, in many problems, accessing tidy data is too costly, so many heuristics have been developed to deal with settings that do not completely fit in the supervised learning setting.
The goal of this thesis is to ground those heuristics into principles.
To make those principles clear and rigorous, this work finds its roots in the statistical learning theory.
In particular, we assume that there is a unique metric to quantify the quality of a learning algorithm, and that this metric is specified by a loss function.
The first task was to define the setup to work on.
Many directions could have been taken.
For example, supervision can come as experts specifying what is their guess for the label, and what is their levels of certitude.
To avoid dealing with fuzziness, we decided to assume that our supervision provides sure information that allows us to locate the label into a set of label candidates, assuming that errors on the supervision could be captured by the randomness of the label conditionally to the input.
This generic yet interesting framework turns out to have already been studied in the literature under the name of partial labeling as well as superset learning.
In a first paper \citep{Cabannes2020}, we described this framework, advocate for combining weak information in a parsimonious way and came up with some theoretical results to strengthen our point.
We put on top the framework of Alessandro to exhibit a consistent algorithm with decent convergence rates.
While we initially wanted to compare this principle with several heuristics and associated principles, we soon realized that this task was endless, and we reduced the experimental comparison to ``comparable'' principles.
Several indications point us toward a second paper \citep{Cabannes2021}.
First of all, the algorithm we came up with was dealing with a really big related surrogate problem, hard to understand, hiding in ugly constants in our convergence rates.
Fundamentally, we were dealing with functions from input to set of labels, and were paying the price of the combinatorial structure of power sets.
To stay in the space of functions from input to labels, we decided to explicitly retrieve labels from weak supervision before learning with a fully supervised dataset.
To come up with a clear empirical objective, we thought in terms of distributions rather than functions, leveraging the kernel mean embedding structure beyond Alessandro framework.
However, theoretical guarantees were much harder to obtain, missing generic tools to understand deviation between empirical principle and population principle when these principles are expressed on distributions.
This forces us to incorporate assumptions to ease the theoretical understanding.
Surprisingly, those assumptions were not that strong and led to much better convergence rates, and this was not only the case for our problem but for the whole framework of Alessandro.
We published this interesting result in a different paper \citep{Cabannes2021b}, and left open research questions on how far we can push those derivations.
We came back to those questions recently, extending the proof mechanism to support vector machines \citep{Cabannes2022b}.
In the first two papers, we built on combining weak information conditionally to a given input.
While relation between inputs were captured by the model of functions we learn, the model we used was not smartly leveraging the input distribution.
As such, we wanted to strengthen our principle to deal with settings akin to semi-supervised learning.
In particular, we wanted to incorporate Laplacian regularization into our framework.
Surprisingly, there were no good ``kernelized'' methods for Laplacian regularization but only ``local averaging'' methods.
Benefiting from the expertise of my supervisors on kernel methods, as well as exchanging with Loucas Pillaud-Vivien, who already realized this fact, we came up with a paper on the matter \citep{Cabannes2021c}.
Arguably, this gave a satisfying final piece to the theoretical framework built in this thesis in order to learn from partial supervision.
From there were several natural continuations.
Deepen the study of partial supervision by understanding and incorporating more heuristics into our frameworks, such as collaborative filtering.
Refine our algorithms with refinement of the structured prediction framework of Alessandro, for example, what can we say about conditional random fields and partial supervision? could we simplify the link between the structure of the supervision and the structure in ``structured prediction'', as we have implicitly done in the first two papers applications and algorithms?
Widen our scope to all weak supervision frameworks, or even all unclutter data problems ({\em e.g.} missing input data).
Ultimately, we decided to go for dataset annotation.
The ultimate goal of this work is to be useful to the practitioner hustling with data.
Indeed, our research was not motivated by the fact that many datasets come with weak supervision, but by the idea that weak supervision is easier to collect for the practitioner.
As a consequence, if we suppose that the practitioner is annotating data, we might help him to efficiently provide the most useful supervision.
This problem has deep links with active learning, sequential design, non i.i.d. statistics and search games; links that were attractive to me.
Building on our previous framework, we came up with a simple model of annotation cost and looked at techniques to minimize this cost for a given level of probably approximately correct bound.
In dimension one, our atomic questions are exactly the gradient of the $L^1$ loss; this pushed us to realize that one does not need full supervision to acquire unbiased stochastic gradients \citep{Cabannes2022}.
However, when dealing with highly structured prediction, naive stochastic gradient descent on continuous surrogates suffer from suboptimality issues.
Hence, in order to have a finer control on the samples we collected, we are currently switching to a bandit setting, for which, with the help of Vianney Perchet, we hope to come up with creative solutions.
\subsection*{Ethical considerations}
This work aims at advancing our understanding of weakly supervised learning.
Weakly supervised learning enrolls in the quest of an automated artificial intelligence, free from the need of human supervision.
Automation, which is at the basis of computer science \citep{Turing1950}, is known to increase productivity at a reduced human labor cost, and is associated with several political/societal issues.
\chapter{Learning Rates with Discrete Outputs}
Understanding the reliability of machine learning is crucial to deploy learned models in the wild.
It was also an important point of this thesis that focuses on algorithms for weakly supervised learning with appreciable theoretical guarantees.
This chapter elaborates on such guarantees and discusses some of our work \citep{Cabannes2021b,Cabannes2022b}, which is reproduced entirely in part \ref{part:discrete}.
For simplicity, it is set in the supervised learning formalization made in the precedent chapter.
In particular, we consider $\mathcal{X}\subset\mathbb{R}^d$ an input space, $\mathcal{Y}$ a discrete output space, $\rho$ a joint distribution on $\mathcal{X}\times\mathcal{Y}$, and $\ell$ a loss function.
Our goal is to minimize the risk ${\cal R}:\mathcal{Y}^\mathcal{X}\to\mathbb{R}$ defined as
\begin{equation}
{\cal R}(f) = \E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)}.
\end{equation}
\section{Statistical learning theory}
In this section, we discuss results offered by statistical learning theory to get insights on what can be learned from data.
In particular, we get more precise about classical derivations of generalization bounds.
\subsection{Insights from information theory}
Artificial intelligence designs a putative system created by humans that would display signs of intelligence, whatsoever this means.
Machine learning consists in implementing such a system with a machine that learns, {\em i.e.} finds in a sort of autonomous fashion, how to behave in order to produce a desired output given some input parameters.
Statistical learning consists in presenting the machine with a set of training examples that show desired outputs for different inputs in order for the machine to infer a rule to produce outputs from inputs.
Statistical learning theory borrows many tools from information theory.
Information theory is concerned with transmitting information.
Think that we want to describe a function $f:\mathcal{X}\to\mathcal{Y}$ to a friend, with only a finite number of information.
If we know that $f$ belongs to a finite set of functions ${\cal F}$, based on dichotomy, we only need $\log_2(\card{\cal F})$ bits to encode this function among all functions in ${\cal F}$.
Now, if ${\cal F}$ is continuous, we cannot encode all functions on a finite number of bits.
Yet, if we only want to unravel a signal $f\in{\cal F}$ up to a precision $\epsilon$, given a certain error metric\footnote{Here, the word ``metric'' is used in a prosaic fashion since $L$ might not be a distance.} $L$ (typically $L(\hat{f}, f) = \E\bracket{\ell(\hat{f}(X), Y) - \ell(f(X), Y)}$), we could do so with an encoding on less than $\ceil{\log_2({\cal N}(\epsilon))}$, with ${\cal N}(\epsilon)$ the minimum number of balls of size $\epsilon$ (with the metric $L$) to cover ${\cal F}$.
To quantify the complexity of the set ${\cal F}$, it is natural to look at the behavior of the number of bits we need to encode any function in ${\cal F}$ up to an error $\epsilon$ as $\epsilon$ get to zero.
As a consequence, it is useful to define a notion of size of ${\cal F}$ (with respect to the error metric $L$) as the superior limit
\[
C_{\cal F} = \limsup_{\epsilon\to 0} - \log_2({\cal N}(\epsilon)) / \log_2(\epsilon).
\]
Note the analogy with statistical mechanics, where signals are microscopic arrangements of unit elements, and the set ${\cal F}$ is the corresponding macroscopic system.
Assuming that all arrangements have the same probability to appear, the Boltzmann entropy is exactly, up to the Boltzmann constant, the logarithm in base two of the cardinality of ${\cal F}$.
This explains the wording Kolmogorov entropy for the logarithm in base two of the covering number ${\cal N}$.
To make the preceding paragraph more concrete, suppose that we want to transmit a function $f:[0,1]^d \to \mathbb{R}$ to our friend.
Typically, we would give to our friend $m$ bits of discrete information, plus some information on the function class ({\em e.g.} $f$ is a polynomial of order $d$, or, thinking in terms of Fourier transform, it has a low energy on high harmonics).
For example, assume that $f$ belongs to the Sobolev space $H^m(\mathop{}\!\mathrm{d} x)$, {\em i.e.} it is $m$ times differentiable, with derivatives up to order $m$ square integrable against the Lebesgue measure.
Suppose that our friend knows that $\norm{f}_{H^m} \leq 1$, and suppose that we want to minimize the $L^2$ error. Then, we can give $n$ binary information to localize this function in a covering of ${\cal F} = \brace{f\,\middle\vert\, \norm{f}_{H^m} \leq 1}$ with $2^n$ $L^2$-balls.\footnote{Note that such a covering of ${\cal F}$ can be taken as it is compact with respect to the $L^2$-topology - even though it is not compact with the $H^m$-topology, $H^m$ being infinite dimensional.}
At best, our friend will be able to localize the signal in an $L^2$-ball of radius $\epsilon(2^n)$, thus making at most an error $\epsilon(2^n)$ on the reconstructed signal, with
\(
\epsilon(n) = \inf \bracket{\epsilon \,\middle\vert\, {\cal N}(\epsilon) \leq n} \approx n^{-C_{\cal F}},
\)
and $C_{\cal F}$ the size of ${\cal F}$ with respect to the $L^2$-metric.
We refer the interested reader to the seminal paper of \cite{Kolmogorov1959} for precise quantification of space sizes.
\subsection{Vapnik-Chervonenkis theory}
Machine learning is concerned with inferring information.
In this setting, we want to learn a task $f:\mathcal{X}\to\mathcal{Y}$, and we do so by collecting examples of the solved instance of this task $(x_i, y_i = f(x_i))$.
In the current machine learning paradigm, we do not choose the bits of information to transmit, but we are given some points $(x_i, y_i)_{i\leq n}$ that are assumed independently sampled according to the distribution that matters to us.
Let us review the picture offered by the classical statistical learning theory for the sake of completeness.
Recall that we want to study the excess of risk. We have, with $f_{\cal F}^*$ and $f_n$ the respective minimizers of the population and empirical risks over ${\cal F}$,
\[
{\cal R}(f_n) - {\cal R}(f_{\cal F}^*) \leq
{\cal R}(f_n) - {\cal R}_{{\cal D}_n}(f_n) +
{\cal R}_{{\cal D}_n}(f^*_{\cal F}) - {\cal R}(f^*_{\cal F}).
\]
For any function $f\in{\cal F}$, we can use concentration inequality (such as Bernstein inequality) to control the difference between the empirical risk ${\cal R}_{{\cal D}_n} (f)$, which inherits its randomness from ${\cal D}_n$, and its average ${\cal R}(f)$.
This works well for the second part of the equation, due to $f^*_{\cal F}$.
Sadly, as $f_n$ depends on ${\cal D}_n$, we can not apply concentration inequality directly to control the deviation between its empirical and population risk.
The classical way to proceed is to find a uniform concentration bound over ${\cal F}$, which we can do by controlling the following random quantity
\[
\sup_{f\in{\cal F}} {\cal R}(f) - {\cal R}_{{\cal D}_n}(f).
\]
When ${\cal F}$ is finite, we can control this supremum with a union bound, joining together all individual concentration inequality.
Similarly, when ${\cal F}$ is continuous, we could do similar things with a well-specified $\epsilon$-cover of ${\cal F}$, which can be refined based on chaining techniques, to avoid redundancy of events when performing union bound for supremum.
In the statistical learning literature, it is classical to rather use a symmetrization trick to relate ${\cal R}(f) - {\cal R}_{{\cal D}_n}(f)$ to ${\cal R}_{{\cal D}'_n}(f) - {\cal R}_{{\cal D}_n}(f)$ for ${\cal D}'_n$ another dataset, and study the Rademacher complexity defined as $\E\sup_{f\in{\cal F}} n^{-1} \sum_{i=1}^n \epsilon_i f(X_i)$ for $\E$ the expectation taken over the $\epsilon_i$ defined as variables taking value one or minus one with probability one half.
In the case of binary classification with the 0-1 loss, {\em i.e.} $\mathcal{Y} = \brace{-1, 1}$, $\ell(y, y') = \ind{y\neq y'}$, Vapnik and Chervonenkis proposed slightly different derivations leading to the following bound.
\begin{equation}
\label{eq:vcbound}
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f_{\cal F}^*) \leq
c\sqrt{\frac{V_{\cal F}}{n}},
\end{equation}
for $c$ a universal constant, and $V_{\cal F}$ the so-called VC dimension of ${\cal F}$, that relates to the average number of points that the class ${\cal F}$ is able to shatter \citep{Vapnik1995}.
A natural question with generalization bounds is how well they quantify the behavior of the excess of risk ${\cal R}(f_n) - {\cal R}^*$.
Indeed, the estimation of the excess of risk given by~\eqref{eq:vcbound} is known to be optimal, in the sense that, for any class of functions ${\cal F}$, it is possible to find a specific distribution $\rho$ such that this bound is a lower bound up to a multiplicative constant.
This behavior is referred to as {\em minimax optimality}, meaning that it is not possible to find a better bound, given the class ${\cal F}$ of estimators considered.
\section{Surrogate methods}
In this section, we discuss the difficulty of learning discrete-valued functions, and the possibility to tackle the learning problem through surrogate problems that consist of learning continuous-valued functions.
\subsection{Practical limitations of VC theory}
While VC theory was a keystone in machine learning, it does not exhaust practitioner issues, as this theory does suffer from some limitations.
\paragraph{Approximation/estimation trade-off.}
First of all, VC theory controls the estimation error in~\eqref{eq:risk-dec}, that is the error in terms of population risk ${\cal R}$ between $f_n$ the minimizer of the empirical risk in a hypothesis class ${\cal F}$ and $f_{\cal F}^*$ the minimizer of the population risk in the class ${\cal F}$.
In particular, equation \eqref{eq:vcbound} does not tell anything about the approximation error, that is the gap between ${\cal R}(f_{\cal F}^*)$ and ${\cal R}^*$.
To control the approximation error, we need to make assumptions on how good our model ${\cal F}$ is to minimize the risk ${\cal R}$.
There is a trade-off between choosing a big model ${\cal F}$, so that the optimal risk ${\cal R}^*$ can be achieved by ${\cal R}(f_{\cal F}^*)$ inside the hypothesis class, and choosing a small model, so that its capacity, captured by $C_{\cal F}$ or ${\cal V}_{\cal F}$, is small.
\paragraph{Optimization issues.}
Finally, if we were to apply VC theory to learn from a continuous input space $\mathcal{X} \subset \mathbb{R}^d$, to a discrete output space $\mathcal{Y}$, we could consider a model ${\cal F}$ of functions from $\mathcal{X}$ to $\mathcal{Y}$.
Those functions could be characterized through their decision regions and decisions boundaries, defined for $y, z \in \mathcal{Y}$ as
\[
f^{-1}(y) = \brace{x\in\mathcal{X} \,\middle\vert\, f(x) = y} \subset \mathcal{X},\qquad
\Delta_{y, z}(f) = \overline{f^{-1}(y)} \cap \overline{f^{-1}(z)} \subset \mathcal{X},
\]
where the bar stands for the closure of the set.
As such, a model ${\cal F}$ could be defined as a collection of putative decision regions.
The region of disagreement between two functions $f$ and $g$ for a prediction $y\in \mathcal{Y}$ can be expressed through the symmetric difference ${f^{-1}(y) \triangle g^{-1}(y)}$.
When using the 0-1 loss, the excess of risk between $f_n$ and $f^*_{\cal F}$ is controlled by the measure of the union of disagreement regions, that is $\rho_\mathcal{X}(\cup_{y\in\mathcal{Y}} (f_n)^{-1}(y)\triangle (f_{\cal F}^*)^{-1}(y))$.
It is then possible to retake the VC theory, using the covering number of ${\cal F}$ with respect to this pseudo-distance \citep{Mammen1999}.
Sadly, even for a simple model ${\cal F}$, such as binary classification with half-plane decision regions, minimizing the empirical risk is NP-hard \citep{Arora1997}.
This echoes optimization issues in machine learning. While VC theory asks to consider the empirical risk minimizer $f_n$, it might be really hard to find it in practice, and one might have to settle with an estimate $\hat{f}$ of $f_n$.
This adds a third term to the risk decomposition that is an optimization error term
\[
{\cal R}(\hat{f}) - {\cal R}^* =
\underbrace{{\cal R}(\hat{f}) - {\cal R}(f_n)}_{\text{optimization error}}
\ + \underbrace{{\cal R}(f_n) - {\cal R}(f_{\text{lin}}^*)}_{\text{estimation error}}
\ + \underbrace{{\cal R}(f_{\text{lin}}^*) - {\cal R}^*}_{\text{approximation error}}.
\]
\subsection{Related continuous surrogate problems}
When learning with discrete outputs, direct risk minimization leads to many combinatorial difficulties that translate into computational intractability.
This is related to the difficulties of studying and optimizing discrete-valued functions.
One technique to overcome this issue consists in learning a discrete-valued function, by learning a surrogate continuous-valued function and thresholding its output in order to make it discrete.
Consider binary classification, that is $\mathcal{Y} = \brace{-1, 1}$ and $\ell(y, z) = \ind{y\neq z}$.
In this setting, the optimal predictor $f^*:\mathcal{X}\to\mathcal{Y}$ is defined as
\[
f^*(x) = \sign(g^*(x)), \qquad\text{where}\qquad
g^*(x) = \E_{(X, Y)\sim\rho}\bracket{Y \,\middle\vert\, X=x}.
\]
This suggests learning the discrete-valued function $f^*:\mathcal{X}\to\mathcal{Y}$ by learning the continuous-valued function $g^*:\mathcal{X}\to\mathbb{R}$.
To an estimate $g$ of $g^*$, we associate the estimate $f = \sign(g)$ of $f^*$.
One method to learn the conditional expectation $g^*$ is through its characterization with the least-squares error
\[
g^* \in \argmin_{g:\mathcal{X}\to\mathbb{R}} {\cal R}_S(g) := \E_{(X, Y)\sim\rho}\bracket{\abs{g(X) - Y}^2}.
\]
Given data $(X_i, Y_i)$, one can then consider a model of real-valued functions ${\cal G} \subset \mathbb{R}^\mathcal{X}$ and the empirical risk minimization
\[
\hat{g} \in \argmin_{g\in{\cal G}} \frac{1}{n}\sum_{i=1}^n \abs{g(X_i) - Y_i}^2.
\]
This estimate is cast as an estimate of $f^*$ through the decoding $\hat{f} = \sign\hat{g}$.
It is natural to wonder how a good estimate of $g^*$ translate into a good estimate of $f^*$.
This question is answered by calibration inequalities that relate the excess of risk on the surrogate problem with the excess on risk of the original problem \citep{Bartlett2006}.
For example, for the least-squares surrogate considered here, we have
\[
{\cal R}(\sign g) - {\cal R}(f^*) \leq \sqrt{{\cal R}_S(g) - {\cal R}_S(g^*)}.
\]
While we only present the least-squares surrogate for binary classification, other surrogates such as the hinge loss, leading to support vector machine, the logistic loss, leading to softmax regression, can be considered.
Similarly, surrogate methods can be extended beyond binary classification, for example, in multiclass problem with the $0-1$ loss, that is $\mathcal{Y} = \brace{1, \cdots, m}$ and $\ell(y, z) = \ind{y\neq z}$, we can consider $g^*:\mathcal{X}\to\mathbb{R}^\mathcal{Y}$, defined as
\[
g^*(x) = \paren{\Pbb\paren{Y=y\,\middle\vert\, X=x}}_{y\in\mathcal{Y}}
= \E\bracket{(\ind{Y=y})_{y\in\mathcal{Y}} \,\middle\vert\, X=x},
\]
and the decoding
\[
f^*(x) = \argmax_{y\in\mathcal{Y}} g^*_y(x).
\]
In such a setting, $g^*$ is often referred to as a {\em score} function.
\begin{remark}[The pros of predicting scores]
From a formal perspective, if we are only interested in the optimal mapping $f:\mathcal{X}\to\mathcal{Y}$, learning $g$ can be seen as a waste of resources.
In essence, this waste of resources is similar to the one when learning the full probability function $(p(y))_{y\in\mathcal{Y}}$ for some $p \in\prob{\mathcal{Y}}$ while we only care about the argmax $y^* \in \argmax_{y\in\mathcal{Y}} p(y)$.
In practice, however, it might be of interest to get an estimate of $g^*(x) = \Pbb\paren{Y=y\,\middle\vert\, X=x}$, as it might provide important information about how specified is $f^*(x)$ and how we could confidently discard other potential labels $y$ for the input $x$.
\end{remark}
We refer the interested reader to \cite{Nowak2021} for deeper reflections about learning with surrogate methods.
\section{Fast rates derivation}
In this section, we sketch roughly the underlying machinery beyond \citet{Cabannes2021b}.
While writing this section, we gave a fresh look at it to avoid redundancy with results already published and reproduced in Chapter~\ref{chap:fast}.
Confronting this machinery with SVM has led us to~\cite{Cabannes2022b} which is reproduced in Chapter~\ref{chap:svm}.
\paragraph{Regularized risk minimization.}
For simplicity, consider a surrogate problem consisting of learning a real-valued function; with a linear parametric model of functions ${\cal G} =\brace{g_\theta: x\to \scap{\theta}{\phi(x)}\,\middle\vert\, \theta\in\Theta}\subset \mathbb{R}^\mathcal{X}$, parametrized by some Hilbert space $\Theta$, and some features $\phi:\mathcal{X}\to\Theta$.
Consider the regularized empirical risk minimization
\[
\theta_n \in \argmin_{\theta} {\cal R}_{S, {\cal D}_n}(g_\theta) + \lambda_n \norm{\theta}^2.
\]
Here, $\lambda_n$ is a regularization parameter that goes to zeros as the number of samples $n$ goes to infinity and the risk of overfitting vanishes, and ${\cal R}_{S, {\cal D}_n}$ is the empirical surrogate risk computed from the dataset ${\cal D}_n$.
Eventually, we could also consider $\hat{\theta}$ an approximation of $\theta_n$ related to some optimization techniques.
Finally, it is useful to introduce the bias estimate
$
\theta_\lambda \in \argmin_{\theta} {\cal R}_S(g_\theta) + \lambda \norm{\theta}^2.
$
\paragraph{Classical versus new convergence study.}
Classically, to prove that $\hat{f}$, the decoding of $g_{\hat{\theta}}$, will minimize the original risk ${\cal R}$ as $n$ goes to infinity and to give rates of convergence, one can use capacity/estimation assumptions to state that $\hat{\theta}$ concentrate around $\theta_\lambda$, as well as source/approximation assumptions to state that ${\cal R}_S(g_{\theta_\lambda})$ will convergence to ${\cal R}_S(g^*)$ as $\lambda$ goes to zero.
Finally, using calibration inequalities, one can relate concentration in $\theta$ to concentration in ${\cal R}_S$, and then to concentration in ${\cal R}$.
The problem with this technique is that cascading calibration inequalities can lead to quite suboptimal rates, as shown by the work of \citet{Audibert2007}, which bypassed this procedure, and, under a simple well-thought margin condition, derived dramatically faster convergence rates than classical ones.
In our view, this work can be generalized by deriving directly calibration inequality that relates the original excess of risk to concentration in the parameter space.
The specificity of those new calibration inequalities is that they are not universal, but depends on some approximation hypothesis that quantifies how hard or easy it is to solve the original problem based on the parametrized surrogate problem.
The most well-known such hypothesis is the Tsybakov margin condition, but others were proposed \citep[for example by][]{Steinwart2007}.
\begin{figure}
\centering
\input{drawings/fast_rates}
\caption{Our new convergence analysis consists in relating natural concentration given by the surrogate method to the original excess of risk without passing by the surrogate excess of risk.
As the drawing shows, concentration in parameter space $\Theta$ can be cast as deviation on the original excess of risk.
Yet, this casting relation depends on the geometry of this picture, which itself depends on which surrogate is used, what is the function to learn, how the bias estimator approaches it, and how our empirical estimate concentrates around the bias estimator.
}
\label{fig:discrete_analysis}
\end{figure}
\paragraph{Special case of exponential convergence rates.}
To give more light on the previous paragraph, consider the binary classification problem. Suppose that there exists $\lambda$ such that $\sign(g_{\theta_\lambda}) = \sign(g^*)$.
Suppose also that there exists $\epsilon > 0$ such that $\norm{\theta - \theta_\lambda} \leq \epsilon$ implies that $\sign(g_{\theta}) = \sign(g_{\theta_\lambda})$.
With those two approximation hypothesis, we have the following calibration inequality
\[
{\cal R}(\sign(g_\theta)) - {\cal R}(f^*) \leq
\ind{\norm{\theta - \theta_\lambda} > \epsilon}.
\]
As a consequence,
\(
\E_{{\cal D}_n}\bracket{{\cal R}(\sign(g_{\theta_{{\cal D}_n}}))} - {\cal R}(f^*) \leq
\Pbb_{{\cal D}_n}\paren{\norm{\theta_{{\cal D}_n} - \theta_\lambda} > \epsilon}.
\)
Assuming that this last $\theta_{{\cal D}_n}$ concentrated around $\theta_\lambda$ with sub-Gaussian tails, that is
\(
\Pbb_{{\cal D}_n}\paren{\norm{\theta_{{\cal D}_n} - \theta_\lambda} >
\epsilon}
\leq \exp(-cn\epsilon^2),
\)
for a constant $c$, we get exponential convergence rates.
\section{Discussion}
In this section, we discuss the practical usefulness of generalization bounds, and the different paradigms to define and quantify learnability.
\paragraph{Practical bounds on expected error.}
In theory, generalization bounds provide guarantees on how much error we might expect when deploying a model learned from data.
In practice, those bounds are rather taken as indications that the learning methods are sound, but rates are not reported in order to get confidence on the learned models.
This might have to do with the fact that bounds often depend on parameters or constants that are hard to know in practice. As such, practitioners often prefer to derive error indications from test samples.
In this line, research on {\em conformal prediction} is trying to leverage test samples to provide useful confidence information on learned models.
\paragraph{Assumptions to quantify learnability.}
Given data, it is highly valuable to get an idea of what can be learned from it.
Of course, learnability depends on regularity assumptions of the problem at hand.
Under such assumptions, lower bounds given by minimax rates are supposed to give optimal baselines for learning rates, thus to quantify learnability.
In practice, those lower bounds are derived by considering the most degenerate function respecting those assumptions.
Sometimes a simple well-thought additional assumption can lead to much better bounds.
As such, is there a natural paradigm of assumptions on the function to learn to discuss its learnability?
For example, assumptions on the surrogate solution $g^*$ are not intrinsic to the original problem but depends on the surrogate problem that is considered.
On the contrary, for a fixed decoding $d$, one can quantify the regularity of $f$ through the most regular function $g$ such that $f = d(g)$ (think of the binary classification case with $d = \sign$), this will correspond to the regularity of the decision frontier, rather than the one of the conditional expectation $x\to\E\bracket{Y\,\middle\vert\, X=x}$.
We refer to the paragraph ``Beyond least-squares'' in Section \ref{lap:sec:structured_prediction} for a practical example of such considerations.
This echoes the difference between realizable-consistency and Fisher-consistency of surrogate methods \citep{Long2013}.
\begin{figure}[t]
\centering
\input{drawings/surrogate_analysis}
\caption{
The behavior described by generalization bound might be relevant only when accessing an indecently large number of samples, thus being of little use in practice.
}
\label{fig:sur_bounds}
\end{figure}
\paragraph{Non-asymptoticness of finite-sample bounds?}
There are two points to mention about probably approximately correct bounds.
First they are derived by considering worst-cases, and as we said in the previous paragraph, simple assumptions often allow refining those bounds by removing pathological cases.
Second they tend to describe the behavior of how risks decrease with the number of samples when accessing numerous samples. As such, while those bounds are non-asymptotic in the sense that they are valid for any number of samples $n$, the described regime might not take place without accessing an indecent number of data, thus being of little use in practice.
To get more concrete, consider a loss $\ell$ and a surrogate loss $L$.
Suppose that we want to bound the excess of risk ${\cal E}_\ell(\theta)$ on the original problem with the excess of risk ${\cal E}_L(\theta)$ on the surrogate problem, where $\theta$ is a learned parameter.
The goal of calibration theory is to find the best function $\xi$ such that for any $\theta$, ${\cal E}_\ell(\theta) \leq \xi({\cal E}_L(\theta))$.
In practice, people try to find $\xi$ concave (because the inequality can be derived for a single $x$ and propagated to the whole space thanks to Jensen inequality), and will look for it under the form $\xi(x) = cx^\alpha$ with $c > 0$ and $\alpha \in (0, 1]$, trying to get the highest $\alpha$ -- that is trying to optimize the behavior of $\xi$ close to zero.
This will lead to somewhat tight bounds only in the regime where ${\cal E}_L(\theta)$ is small, a regime which might not be reached without an indecent number of samples.
We illustrate the regime where constants matter more than exponents on Figure \ref{fig:sur_bounds}.
In practice, it would be more useful to take into consideration the number of samples first, before looking for a function $\xi$ that verifies ${\cal E}_\ell(\theta) \leq \xi({\cal E}_L(\theta))$ for any $\theta$, and minimizes $\max\xi([a, b])$ for $[a,b]$ the range of value that we expect the surrogate risk to take given our fixed number of samples.
\chapter{The Machine Learning Paradigm}
In 2021, the International Data Corporation estimates that by 2025, the yearly global spending on artificial intelligence (AI) will overrun \$204 billions \citep{IDC2021}.
To give an order of magnitude, this corresponds to the gross domestic product of countries such as Greece, New Zealand or Peru.\footnote{According to the World Bank Open Data.}
It means that if this business was concentrated in Peru, its 33 millions inhabitants would only produce goods and services related to AI.
Nowadays, the core engine behind AI is machine learning, that is the learning of rules and algorithms by computers.
In this chapter, we provide a subjective exposition of what machine learning is about.
It is written to be accessible for a large public beyond the machine learning community.
\section{From algorithmics to machine learning}
In this section, we present machine learning as the evolution of algorithmics.
In a prosaic fashion, an algorithm can be thought of as a cooking recipe.
It is a sequence of instructions that, given some inputs, produces an output.
For example, the inputs could be some apples, lard, eggs, salt and flour, the sequence of instructions be the one of an apple pie recipe, and the output be a delicious pie.
This can be formalized mathematically, an algorithm being understood as a sequence of instruction $I$, leading to a mapping $f:\mathcal{X}\to\mathcal{Y}$, which, given an input $x$ belonging to an input space $\mathcal{X}$, outputs $y = f(x)$ belonging to an output space $\mathcal{Y}$.
\begin{example}[Cooking recipe]
\label{ex:cooking}
$\mathcal{X}$ is a set of potential ingredients, $\mathcal{Y}$ is a set of potential pies.
For some ingredients $x$, {\em e.g.} $x = ($5 apples, 100g of lard, 200g of flour, 1 egg, 1 pinch of salt$)$, $y = f(x)$ describes the pie obtained by following the instructions $I$ of a recipe.
\end{example}
\begin{example}[Sorting algorithm]
\label{ex:sorting}
$\mathcal{X} = \mathbb{R}^n$ is the set of lists of size $n$, $\mathcal{Y} = \brace{(y_1, \cdots, y_n) \in \mathbb{R}^n \,\middle\vert\, y_1 \leq \cdots \leq y_n}$ is the set of sorted lists of size $n$.
Any sorting algorithm corresponds to $f:\mathcal{X}\to\mathcal{Y}$ that maps a list $x = (x_i)_{i\leq n}$ to its sorted version $y = (x_{\sigma(i)})_{i\leq n} \in \mathcal{Y}$ with $\sigma \in \mathfrak{S}_n$ a permutation.
\end{example}
\begin{example}[House pricing]
\label{ex:housing}
$\mathcal{X} = \mathbb{R}^2_+$ represents real estate properties with two features: the living space of the house, the outdoor space of the garden, both in meter square.
$\mathcal{Y} = \mathbb{R}_+$ represents the price of the house in US dollars.
The mapping $f$ is a model for house prices.
\end{example}
\begin{example}[Classifying cats and dogs]
\label{ex:image_classification}
$\mathcal{X} = [255]^{d_1\times d_2\times 3}$ is a set of images of cats and dogs. Those images are represented by $d_1 \times d_2$ pixels, with the three RGB channels.
$\mathcal{Y} = \brace{-1, 1}$ with $f(x)=1$ if $x$ is an image of cats, and $f(x)=-1$ if $x$ is an image of dogs.
\end{example}
The problematics arising in our examples are all slightly different and will be helpful to illustrate differences between algorithmics, statistics and machine learning.
\paragraph{Algorithmics.}
In classical computer science, algorithms are implemented on $m$ bits of memories. As such, the sequence of instructions can be written as $I = (I_t)_{0\leq t\leq T+1}$, with $I_0:\mathcal{X}\to\brace{0, 1}^m$ an encoding of $\mathcal{X}$ on $m$ bits, $I_{T+1}:\mathbb{R}^m \to \mathcal{Y}$ a decoding of $m$ bits to $\mathcal{Y}$, and, for $t\in [T]$, $I_t:\brace{0,1}^m\to\brace{0,1}^m$ a basic bitwise operation.
In problems such as sorting lists, we know exactly which $f$ we want to achieve, but we do not know how this mapping can be decomposed into a set of basic operations which could be implemented practically on a computer.
In those algorithmic problems, an algorithm is thought of as the sequence of instructions $I$,and its quality is discussed under the light of two notions.
\begin{itemize}
\item Correctness: does this sequence of instructions correspond to the desired mapping $f$? In other terms, can we guarantee the equality $f = I_{T+1}\circ I_{T} \circ \cdots \circ I_0$?
\item Complexity: how much basic operations $T$ and computer memory $m$ does it take to perform the full sequence of instruction $I$?
\end{itemize}
In the cooking example \ref{ex:cooking}, the first question can be rephrased as: does our recipe, given a specification of ingredients, always lead to the same pie?
The second question is related to the equipment and the time required by the recipe.
When it comes to sorting lists, example \ref{ex:sorting}, deriving and implementing a sequence of basic operations to solve this task is a reasonable idea.
For example, think of how you sort a deck of cards, write rules about it, and implement those rules in your favorite programming language that will convert it into a sequence of bitwise instructions for the computer.
When it comes to understanding images, deriving and implementing a sequence of basic operations to differentiate cats and dogs have not been proven a successful way to proceed.
Indeed, even for language translation, efforts, made by linguists to describe translation as a set of simple syntactic rules, have not led to efficient translation algorithms.
\paragraph{Learning optimal sequence of instructions.}
Rather than designing an algorithm from a set of humanly-thought rules, machine learning suggests designing an algorithm by quantifying a notion of optimality, or equivalently a notion of error, cost or risk ${\cal R}$, and looking among all potential sequences of instructions $I \in {\cal I}$, in a class ${\cal I}$ of potential sequences of instruction, for the one that minimize the risk ${\cal R}(I)$.
For example, in the cooking example, we could consider ${\cal I}$ as all potential combinations of basic instructions such as ``breaking eggs'', ``mixing some ingredients'', ``frying/baking a batter'' and so on.
Optimality, and the risk ${\cal R}$, could be designed by some experts testing the pie, or through molecular gastronomy considerations.
To find the best recipe, that is the best sequence of basic instructions, a naive approach consists in trying all potential combinations of basic instructions $I$, and quantifying the result quality through ${\cal R}(I)$, and considering the sequence $I$ that achieves the minimal risk.
Of course, there are zillions of ways to combine basic instructions, and baking zillions of pie, and having them tested by experts is not a viable option.
But on a computer that can perform millions of operations by seconds, the picture is quite different.
Applied to the list sorting problem, such a procedure leads to a completely different approach from the classic algorithmic one.
Roughly speaking, the machine learning paradigm is to try all potential algorithms and take the best one, where ``the best'' is quantified through the risk ${\cal R}$.
\section{The emergence of statistics}
In this section, we get more specific on what machine learning is today, that is, statistics in the era of powerful computers.
While machine learning could be seen as the art of automatically designing algorithms, nowadays, it is rather a new take on statistics in the era of powerful computers.
As such, it is less focused on having a ``good'' correct sequence of instruction $I$ to perform a task, but rather in finding a mapping $f$ that is optimal for a task, such as pricing houses or distinguishing cats and dogs.
In such a setting, the risk ${\cal R}:\mathcal{Y}^\mathcal{X}\to \mathbb{R}_+$ associates a score ${\cal R}(f)$, quantifying a notion of risk, cost or error, to a function $f:\mathcal{X}\to\mathcal{Y}$.
Ideally, we would like to retrieve an optimal mapping $f^*:\mathcal{X}\to\mathcal{Y}$ such that
\begin{equation}
{\cal R}(f^*) = {\cal R}^*, \qquad\text{where}\qquad
{\cal R}^* = \inf_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}(f).
\end{equation}
The problematics of machine learning are slightly different from the ones of algorithmics.
The quality of a procedure, that allows obtaining automatically a mapping $f:\mathcal{X}\to\mathcal{Y}$ to solve a specified task, is discussed under the light of three notions.
\begin{itemize}
\item Optimality: is the risk ${\cal R}(f)$ of the obtained mapping $f$ close to ${\cal R}^*$?
\item Inference complexity: given an input $x\in\mathcal{X}$, is it easy to compute $f(x)$?
\item Training complexity: is it computationally easy to obtain the mapping $f$?
\end{itemize}
The first two issues are the direct translation of the algorithmic issues of correctness and complexity, while the third one is specific to machine learning, and has strong links with the field of optimization.
\begin{remark}[Correctness and Optimality]
In classical algorithms, there is the idea that an algorithm is perfect/correct, that we know exactly what it is doing.
In machine learning, the idea of $f$ being correct is dropped by the notion of $f$ being optimal or quite optimal.
For critical applications such as flying planes, replacing the notion of correctness brought in by classical algorithms, by a notion of quasi-optimality is not innocuous.
Yet, once a mapping $f$ is retrieved by an artificial intelligence system, it can be tested for correctness through formal proof management systems, or tested for statistical quality through an extensive number of testings.
While replacing ``being correct'' by ``being statistically good'' can offend purists, note that, in medicine, drugs are more often approved based on statistical studies, rather than on a precise description of their mechanisms and a complete understanding of their interactions with the body.
\end{remark}
If we consider the cooking example \ref{ex:cooking}, we could translate those three issues into prosaic terms. Suppose that you have a way to design new recipes.
Optimality translates into ``are the new dishes you make good?''.
Inference complexity translates into ``are the recipes easy to reproduce?''.
Optimization complexity translates into ``is your procedure for designing new recipes difficult, is it time and money consuming?''.
\subsection{Supervised learning}
A simple idea to learn a mapping $f$ between inputs $x$ and outputs $y$ is to consider a set of $n$ examples $(x_i, y_i)_{i\leq n} \in (\mathcal{X}\times\mathcal{Y})^n$, where $f(x_i)$ should be $y_i$, and try to infer from those examples a general law between $x$ and $y$.
That is, for house pricing, example \ref{ex:housing}, consider $n$ houses, report their characteristics $(x_i)_{i\leq n}$ and their prices $(y_i)_{i\leq n}$, and try to find a law that links house characteristics $x$ to their prices $y$.
In particular, a real estate agent can try to fit a linear model $y = w^\top x$, for $w \in \mathbb{R}^d$ a weight vector, the house characteristics $x$ assumed to belong to $\mathcal{X}\subset\mathbb{R}^d$, and $y$ the price of the house.
It is usual to fit $w$ by minimizing the least-squares error
\[
\hat w \in \argmin_{w\in\mathbb{R}^d} \frac{1}{n} \sum_{i=1}^n \norm{w^\top x_i -
y_i}^2 = (\sum_{i=1}^n x_i x_i^\top )^{\dagger} \sum_{i=1}^n y_i x_i,
\]
where, for a vector $x\in\mathbb{R}^d = \mathbb{R}^{d\times 1}$, $x^\top \in \mathbb{R}^{1\times d}$ designs its transpose, and, for a matrix $A$, $A^\dagger$ its pseudo-inverse.
When needing to price a new house, the real estate agent could report its characteristics $x$ and price it at $y = \hat{w}^\top x$.
Such a model raises many questions.
Which features $x$ govern the price of a house?
Can the price of the house be deduced as a linear combination of those features?
How many houses $n$ should I consider as training examples $(x_i, y_i)_{i\leq n}$ to retrieve a $\hat w$ such that the value $\hat w^\top x$ does reflect well the market price of a new house characterized by $x$?
How much should we expect the transaction price of a house to deviate from our model of its market price?
Those questions are addressed by statisticians and machine learning practitioners.
They are respectively linked with features engineering, model selection, data collection, and generalization guarantees.
Perhaps the most well known result of statistics is the large law number.
In essence, it tells you that if you repeat the same experiment, what you see as the result of this experiment, reflects what you are likely to see if you repeat this experiment one more time.
For example, if you have rolled a die zillion times, and you have gotten six forty-five percent of the time, the next time you roll this die, you can expect to get a six with forty-five percent of chance.
In the past two centuries, statisticians have developed many concentration inequalities that quantify, given the number of experiments you have run so far, how much the statistics collected from those experiments are likely to reflect the outcome of future experiments.
This provides tools to give generalization guarantees on models learned from data.
But beyond those results are some axioms on the experiments you run.
In particular, statistics are grounded in probability theory, assuming that the experiment's outcomes are governed by some probabilistic process that is stable over time, allowing prediction of future outcomes from prior ones.
Typically, when talking about dice, we assume that you roll the same dice, in the same fashion, in the same environment.
Learning theory, and in particular supervised learning, builds on this idea of understanding observable phenomena by assuming an underlying probabilistic model.
This thesis is set in this framework. It assumes the existence of a joint probability distribution $\rho \in \prob{\mathcal{X}\times\mathcal{Y}}$ over $\mathcal{X}$ and $\mathcal{Y}$, that have generated the dataset ${\cal D}_n = (X_i, Y_i)_{i\leq n} \sim \rho^{\otimes n}$ in an {\em independent identically distributed} fashion.
Moreover, it assumes that the measure of error results from the integration of a pointwise measure of error $\ell(f(X), Y)$, between a prediction $f(X)$ made at the point $X$ and its observed label $Y$, for $\ell \in \mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ a loss function, in the sense that for $f:\mathcal{X}\to\mathcal{Y}$,
\begin{equation}
{\cal R}(f) = \E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)}.
\end{equation}
As one has no access to the distribution $\rho$ but only to samples $(X_i, Y_i)_{i\leq n}$, it is natural to substitute the risk with its empirical counterpart ${\cal R}_{{\cal D}_n}:\mathcal{Y}^\mathcal{X}\to \mathbb{R}$, defined as
\begin{equation}
{\cal R}_{{\cal D}_n}(f) = \frac{1}{n}\sum_{i=1}^n\bracket{\ell(f(X_i), Y_i)}.
\end{equation}
A popular approach to learn a mapping $f:\mathcal{X}\to\mathcal{Y}$ is to consider some model of functions ${\cal F} \subset \mathcal{Y}^\mathcal{X}$ and to perform empirical risk minimization over this class of functions, leading to the estimate
\[
f_n \in \argmin_{f\in{\cal F}} {\cal R}_{{\cal D}_n}(f).
\]
Yet, how optimal is this mapping $f_n$? In other terms, does it exhibit a low risk ${\cal R}(f)$?
In the machine learning community, this question refers to generalization properties of $f_n$: does what we have learned based on $n$ examples generalize to new situations characterized by new inputs?
Statistical learning theory provides some answers to this question.
\begin{remark}[Why probability?]
Arguably, we live in a deterministic world. Hence, if everything has a reason, why introduce randomness in our model?
Think about rolling dice, the face that pops up is a pure deterministic function of the landscape of the support the dice are rolled on, as well as the initial moment, impulsion and position of the dice, which themselves depend on the mood of the person rolling the dice.
But as we cannot model those factors, we suppose them to be random, leading to the random distribution on the rolling dice, at the end of the causal chain from mood to initial physical values to dice output.
The same applies to house price, many idiosyncratic factors come into play when a buying offer is made on a house.
When tracking those explanation variables is too difficult, statisticians deal with this fact by adding randomness into the model.
\end{remark}
\subsection{Generalization bounds}
In this subsection, we mimic typical derivations provided by statistical learning theory which are used in this thesis.
Recall the house pricing example \ref{ex:housing}.
Supervised learning consists in assuming the existence of some abstract distribution $\rho\in\mathcal{X}\times\mathcal{Y}$, and that pairs of house characteristics and their prices $(x_i, y_i)_{i\leq n}$ are actually sampled from this distribution $(x_i, y_i) = (X_i, Y_i) \sim \rho^{\otimes n}$ -- using capital letters to denote the randomness in the sampling.
Trying to predict house prices from a linear combination of house characteristics leads to the model of functions ${\cal F} = \brace{x\to w^\top x\,\middle\vert\, w\in\mathbb{R}^d}$, assuming $x\in\mathcal{X}\subset \mathbb{R}^d$.
The best linear model, based on the mean-squares error, is defined by $f^*_{\text{lin}}:x\to x^\top w^*$, with the weighting scheme
\[
w^* \in \argmin_{w\in\mathbb{R}^d} \E_{(X, Y)\sim\rho}\bracket{\norm{X^\top w - Y}^2}
= (\E_{X\sim\rho_\mathcal{X}}\bracket{XX^\top})^{\dagger} \E_{(X, Y)\sim\rho}\bracket{Y X}.
\]
Here $\rho_\mathcal{X}$ designs the marginal of $\rho$ over $\mathcal{X}$, that is, the distribution of $X$ according to $\rho$.
In practice, we do not have access to $\rho$ but to the samples $(X_i, Y_i)$.
As a consequence, we can replace $\rho$ by $\hat\rho = \frac{1}{n} \sum_{i=1}^n \delta_{(X_i, Y_i)}$, which leads to the empirical risk minimizer $f_n: x\to x^\top \hat{w}$ with
\[
\hat{w} \in \argmin_{w\in\mathbb{R}^d} \frac{1}{n}\sum_{i=1}^n\norm{X_i^\top w - Y_i}^2
= \paren{\frac{1}{n}\sum_{i=1}^n X_iX_i^\top}^{\dagger} \paren{\frac{1}{n}
\sum_{i=1}^n Y_i X_i}.
\]
As the quality of a function $f$ is measured through the risk ${\cal R}$, it is natural to measure the quality of $f_n$ in terms of the excess of risk ${\cal R}(f_n) - {\cal R}^*$, with respect to the minimum achievable risk ${\cal R}^*$.
This error can be split in the following way
\begin{equation}
\label{eq:risk-dec}
{\cal R}(f_n) - {\cal R}^* =
\underbrace{{\cal R}(f_n) - {\cal R}(f_{\text{lin}}^*)}_{\text{estimation error}}
\ + \underbrace{{\cal R}(f_{\text{lin}}^*) - {\cal R}^*}_{\text{approximation error}}.
\end{equation}
This split separates an error due to the need for more data so that $\hat{w}$ estimates $w^*$ correctly and a part due to the fact that the best linear predictor $f^*_{\text{lin}}$ might not be the best pricing model.
The decomposition with estimation and approximation errors, is sometimes called bias-variance decomposition.
The variance, or estimation error, is due to the randomness of the samples, while the bias, or approximation error, is due to the gap between the average (or best) estimate we expect and the solution.
The approximation error is due to the fact that our model might be inaccurate.
This error can be controlled by assuming some sort of density of ${\cal F}$ in $\mathcal{X}\to\mathcal{Y}$ regarding the topology inherited from ${\cal R}$, or by assuming that the best predictor $f^*$ does belong to our class of functions ${\cal F}$ (or is well approximated by it with respect to the topology associated with ${\cal R}$).
The estimation error is due to the fact that we have estimated a quantity defined through the distribution $\rho$ thanks to sample $(X_i, Y_i)_{i\leq n}$.
Similarly to rolling dice or tossing coins, when knowing $\rho$, statistics and probability can quantify, for a given number of training examples $n$, what is the distribution of ${\cal R}(f_n) - {\cal R}(f_{\text{lin}}^*)$.
This distribution should be seen as the pushforward of $\rho^{\otimes n}$ regarding the process that leads to the estimate $f_n$ from samples $(X_i, Y_i)_{i\leq n}$.
When not knowing $\rho$, this pushforward cannot be estimated, but with few hypotheses on $\rho$, such as control of extreme behaviors, or equivalently control of high moments, it is possible to control deviation of ${\cal R}(f_n)$.
This is exactly what provide concentration inequalities, such as Hoeffding or Bernstein inequalities, that allows upper-bounding the deviations between the empirical means $\frac{1}{n}\sum_{i=1}^n X_i X_i^\top$ and $\frac{1}{n}\sum_{i=1}^n Y_i X_i$ and their population counterpart $\E[XX^\top]$ and $\E[YX]$, and therefore the deviation between $\hat{w}$ and $w^*$.
With assumptions to control the approximation and estimation errors, it is possible to get convergence results of the type,
\[
\forall\, t > 0,\qquad \Pbb_{{\cal D}_n}({\cal R}(f_n) - {\cal R}^* > t)
\leq \exp(-\sigma^{-2} nt^2).
\]
This specific result tells you that if you have trained your model with $n$ independent identically distributed samples $(X_i, Y_i)_{i\leq n}$, the risk of the learned mapping $f_n$ is comparable to a random number that follows a Gaussian distribution ${\cal N}(0, \sigma / \sqrt{n})$, the randomness being due to the randomness of the samples.
Such a bound is called a generalization bound as it tells us what risk or error we can expect on generic examples, which is a measure of how well our predictor generalizes to new situations.
When an algorithm benefits from such a bound, we call it probably approximately correct \citep{Valient2013}, in the sense that with high probability (the randomness being due to the sample population), it approximately minimizes the risk.
It is worth noting that, in practice, we only have one realization of ${\cal D}_n$, and that this result does not tell us that the risk ${\cal R}(f_n)$ is small, it only tells us that we would have been really unlucky if it was not the case.
\subsection{Unsupervised learning}
In the preceding sections, we have presented machine learning as concerned with the learning of a mapping from an input space $\mathcal{X}$ to an output space $\mathcal{Y}$.
As such, it differs from statistics that are rather concerned with inferring properties of a distribution $\rho_\mathcal{X}$ from samples $(X_i)_{i\leq n}$.
Indeed, many papers published as machine learning papers are concerned with this setting, usually referred to as unsupervised learning.
There are two major motivations for unsupervised learning.
First, statistical problems such as density estimation, eventually done through maximum likelihood estimation, are instances of problems where we would like to infer properties of a distribution $\rho_\mathcal{X}$ through samples $(X_i)_{i\leq n}$.
The usage of a computer, and of optimization routines, is crucial to deal with a large number of data in this setting.
Indeed, many of the computational and statistical problematics arising from those problems are shared with standard supervised learning, justifying results on those problems to be published in machine learning conferences, and justifying the machine learning community to welcome statisticians.
Second, learning a mapping from inputs to outputs, highly benefits from features engineering, which might be done without access to the supervision $(Y_i)_{i\leq n}$.
From a machine learning perspective, the main goal of unsupervised learning is to discover structure beyond input data, assuming that this structure will help to find laws that correlate those inputs to potential outputs.
For example, when described as an array of pixels, images are understood as points in $\mathbb{R}^d$ for $d$ of order $10^9$.
Yet, we expect natural images to have a strong structure that forbid them to be any points in $\mathbb{R}^d$.
Indeed, we can expect natural images to be concentrated close to a low-dimensional manifold, potentially parametrized in $\mathbb{R}^m$ for $m$ much smaller than $d$, describing the many parameters of freedom of a natural image.
Finding such a small representation would ease ``downstream'' tasks, such as learning to recognize the content of an image.
Having been approached with many perspectives, unsupervised learning has been divided into many sub-problems, such as clustering, manifold learning, sparse dictionary learning, or self-supervised learning.
\section{A tour of problems}
In this section, we review diverse practical problems that are arguably understood as machine learning problems.
We review machine learning to create artificial intelligence with human-alike capability.
We showcase business motivations and driving forces beyond machine learning.
We explain how data science can fit into intelligent systems, and discuss how machine learning can help science and how its scope is growing behind statistical learning.
We do not aim for completeness, but for simple examples.
\subsection{Seeing, hearing and talking}
For non-experts of machine learning, probably the most important result of this last decade was the result achieved by deep learning trained on GPU on the ImageNet challenge of 2012 \citep{Krizhevsky2012}. In essence, for the first time, a computer was able to showcase human-like performance when recognizing simple objects in images.
Similarly, we can dictate to our electronic devices and have those devices writing down our thoughts as if they were secretaries.
Moreover, those devices can emit sounds articulated into words and sentences to answer a question we have eventually asked.
All as if machines were to see, hear and speak.
This echoes a long fantasy of machines bearing signs of human intelligence.\footnote{Indeed, some researchers use concepts of human psychology to describe their algorithms such as attention and saliency maps \citep{Cabannes2020b}.}
Interestingly, this ``artificial intelligence'' replicates tasks that many animals do automatically and instantaneously, concerned with processing simple information.
As such it is more concerned with intelligence as collecting pieces of information (such as ``military intelligence''), rather than the more deep thought and contemplative process captured by the French notion of {\em intelligence}.
Those recent successes are somewhat based on supervised learning, providing strong evidence of the usefulness of this theoretical framework, especially when compared to previous attempts to build computer vision, audio and natural language processing from so-called ``expert systems''.
Solving those problems has required machine learning to scale with millions of training examples $n$, as well as millions of input dimensions $d$, which has strongly pushed to advance calculations engineering \citep[see, \emph{e.g.},][]{Chowdhery2022}.
\subsection{Playing games}
Another highly publicized result of machine learning was the beating of all humans at Go, an abstract strategy board game known for the richness of all potential strategies \citep{Silver2018}.
Games such as Go or chess do not fit exactly the framework of supervised learning.
Those games are characterized by a position, or state, which mainly corresponds to the disposition of pieces on the game board, that is changed by moves, or action, made iteratively by two players.
The game is over when a terminal state is reached.
Each terminal state is associated with a game output, could it be a draw, a win of player one or a win of player two.
What needs to be learned is a strategy, or policy, that can be seen as a function that maps a game state to an action or move.
The goal is to come up with a winning strategy whatsoever the opponent strategy.
These problems can be approached with reinforcement learning \citep{Sutton2018}, whose goal is to navigate an environment over time in order to maximize some reward functions.
The wording ``reinforcement'' came from behavioral psychology and describes the fact that the strategy is learned by trials and errors.
In the case of games, the computer can simulate many games against itself before finding a winning strategy.
\subsection{Recommender systems}
\label{int:sec:coll_filt}
Far away from the fantasy of creating superhuman creatures, a consequent amount of machine learning development is the fruit of businesses trying to better identify potential customers and prospects of bestseller products.
A particularly hot topic is recommender systems, whose goal is to propose to a given customer characterized by some features $x$, the most adequate content $y$.
Recommender systems are now everywhere to suggest personalized content, could it be for browsing music and movies, or for targeted advertising.
In 2009, Netflix ran a competition to predict how many stars over five a user might rank a movie. Their goal was clear : have a system that finds the most relevant movies to recommend to the user. They offered \$1,000,000 to anyone coming up with the best performance on this prediction problem \citep{Bennett2007,Lohr2009}.
A specificity of this problem was the absence of features to describe movies or users. Among suggestions of many participants, the best solution was provided by "collaborative filtering". This technique consists in trying to factorize the note given to a movie by a user as a product $u^\top m$, where $u$ characterizes the user, and $m$ characterizes the movie with the smallest (in terms of number of coordinates) vectors $u$ and $m$.
The goal is to find few characteristics for each user that respond to the same number of few characteristics for each movie in order to explain given scores and to predict future scores -- think of $u = (\text{how much I like romantic movie}, \text{how much I like action movie})$ and of
$a = (\text{how much is it a romantic movie}, \text{how much is it an action movie})$.
Formally this technique is formalized by trying to factorize the table $S=(\text{score}(\text{user},\text{movie}))_{\text{user}, \text{movie}}$ as $S = U M$ on observed scores for $U$ and $M$ of smallest rank.
While the rank is not a continuous function, making this problem hard to solve, it is possible to relax it into a concave function thanks to the nuclear norm and get nicely behaving solutions.
\subsection{Intelligent infrastructure}
The machine learning of today is the science of data. The competency to manage massive amounts of data opens the way for massive systems of information.
Moreover, advances in sensors and measurement tools as well as robots allowing vast automation of tasks create optimistic perspectives in fields such as farming or medicine.
The big dreamer might fall for the internet of things governing hydrometry sensors, cameras recognizing aphids, automatic release of pesticides, analysis of the best companion plants and so on.
But without going that far, simply putting in place numerical tools to easily perform large scale cross-sectional and longitudinal studies would highly accelerate medicine -- recall debates on Hydroxychloroquine as a COVID treatment based on small studies such as \cite{Raoult2020}.
Arguably, the revolution promised by machine learning is not the revolution of machine learning per se, nor the creation of supra-humanoid, but the future deployment of massive active monitoring systems \citep{Jordan2019}.
To give a personal touch to this discussion, in 2019, with two friends, we created an ``artificial intelligent system'' to interact with artists during their creation process \citep{Cabannes2019}. In this project, much of our work was engineering to make sure that the process was innocuous for the artists. The machine learning algorithm running in the middle was only a part of the complete picture.
\subsection{Making sense of scientific data}
While we have described the application of machine learning for non-science purposes, in particular for marketing and agriculture, there is traction for machine learning as science for science.
Of course, supervised learning techniques can be applied to scientific problems, such as the protein folding problem \citep{Jumper2021}.
But machine learning techniques are also handy to make sense of data coming from massive scientific experiments such as the quest for the Higgs boson at the CERN \citep{Larkoski2020}, or the analysis of gravitational waves \citep{Gebhard2019}.
In those last two examples, the goal is to retrieve a signal that has caused some observations given a potential high-level of noise.
\subsection{Signal recovery and inverse problems}
Signal recovery from noisy observations is an active field of research with many appreciable results.
The goal of compression is to, given a signal $x$ that is heavy to describe, compress it into a signal $y=f(x)$ much lighter to describe, store, transmit, {\em etc.} such that given the knowledge of $f$ and a prior on $x$, one can retrieve the signal $x$ from the observation $y$.
Historically, this has been studied for telecommunications with several results coming from the Bell Labs, such as the Nyquist-Shannon theorem \citep{Shannon1949,Nyquist1928}: assuming $x$ to be a function with bounded frequencies, it can be reconstructed by interpolation \citep{Whittaker1915}.
More recently, \cite{Candes2006} and \cite{Donoho2006} looked at the transmission of a vector $x\in\mathbb{R}^n$ known to be sparse, {\em i.e.} having only $s<<n$ non-zero components, under the form $y = A x\in\mathbb{R}^m$ for a chosen $A\in\mathbb{R}^{m\times n}$.
They showed that by taking the row of $A$ sufficiently ``incoherent'' it is possible to recover $x$ with $m\approx 2s$.
Moreover, when the coefficients of $A$ are random, normally distributed, the ``incoherence'' property holds with high probability.
This allows to drastically reduce compression, as well as acquisition, of sparse signal, and had successful applications in medical imaging \citep{Lustig2008,Sidky2008}.
Think of the impact of compressed sensing when it permits reducing X-ray exposure during tomography for the same amount of reconstruction fidelity.
Signal recovery is not the primary focus of machine learning, but as the field is gaining traction, its tools and scope of applications are growing, and similarly to compressed sensing papers that find their place in machine learning journals and conferences, other imaging techniques might as well.
To conclude on those inverse problems beyond machine learning, let us mention the work of \cite{Fink2002}, which leverages wave reversibility and information preservation to design time-reversal mirrors in order to locate a wave source and recreate the emitted wave.
In the same vein, techniques such as sonar or medical ultrasonography are based on propagating waves into a medium and deduce a topography of the medium from the waves echo.
Among others, they have been used to locate oil in the ground, after creating a wave with dynamite.
Quite remarkably, recent research has shown those explosions to be a waste of resources as it is possible to leverage ambient noise of the Earth's crust in order to image it \citep{Garnier2016}.
\section{A tour of models}
In this section, we focus on the supervised learning settings, and we discuss classical models to learn functions.
Our classification into local averaging, reproducing kernel and deep learning methods is aligned with the class taught by Francis, on learning from first principles \citep{Bach2023}.
\subsection{Local averaging methods}
If you are a real estate agent pricing houses, and you encounter a new house and need to price it, a natural idea is to look for similar houses that have been sold recently, and infer a price from those similar examples.
This is exactly what captures local averaging methods.
Suppose that you have collected examples $(X_i, Y_i)_{i\leq n}$ with features $X_i$ characterizing a house indexed by $i \in [n]$ and its price $Y_i$.
Given a new house with characteristics $x$, you can look at the $k$-th nearest neighbors ${\cal N}_k(x)$ that contains $k$ elements in $[n]$ and such that for $i\in{\cal N}_k(x)$ and $j\in[n]\setminus{\cal N}_k(x)$, $d(x, X_i) \leq d(x, X_j)$ according to a distance $d:\mathcal{X}\times\mathcal{X}\to\mathbb{R}$.
This defines the nearest neighbors predictor $f_n:\mathcal{X}\to\mathcal{Y}$ as
\(
f_n(x) = \frac{1}{k}\sum_{i\in{\cal N}_k(x)} Y_i
\).
Eventually, you might want to give more importance to houses that are really similar to the one you are trying to price, and less to the $k$-th nearest neighbor.
To do so, you can introduce a system of weights $w_i=w_i^{(n)}:\mathcal{X}\to\mathbb{R}$ that are positive and sum to one, and refine $f_n$ as
\[
f_n(x) = \sum_{i=1}^n w_i(x) Y_i =
\sum_{i=1}^n w_i(x) f^*(X_i) + \sum_{i=1}^n w_i(x) \epsilon_i
\]
with $\epsilon_i = Y_i - f^*(X_i)$ linked with the noise of having observed $Y_i$ when we would have preferred to observe $f^*(X_i)$.
\paragraph{Universal consistency.}
While this pricing method sounds really sensible, the job of statisticians is to make formal statements to prove its soundness.
The most classical statement is to prove that when the number of examples goes to infinity, we retrieve the optimal pricing rule.
This property is known as {\em consistency}.
Consistency is usually defined formally through convergence in probability (although some prefer to define it with convergence almost surely) of the risk ${\cal R}(f_n)$, which is seen a random variable depending on the sampling of the dataset ${\cal D}_n = (X_i, Y_i)_{i\leq n}\sim \rho^{\otimes n}$, toward the infimum risk ${\cal R}^*$.
Methods that are consistent without assumptions on the solution $f^*$ are known as {\em universally consistent}.
Consistency of local averaging methods has first been derived by \cite{Stone1977}.
In essence, it consists in assuming that the noises $(\epsilon_i)_{i\leq n}$ are well-behaved, such that $\sum_{i=1}^n w_i(x) \epsilon_i$ cancels out as the number of samples $n$ goes to infinity, and assuming the weighting scheme concentrates locally around $x$ such that, for $\rho_\mathcal{X}$-almost all $x$, informally,
\[
\lim_{n\to+\infty}f_n(x) =
\lim_{n\to+\infty} \sum_{i=1}^n w_i(x)f^*(X_i)
= \lim_{r\downarrow 0} \frac{1}{\rho_\mathcal{X}(B(x, r))} \int_{B(x, r)}
f^*(t)\rho_\mathcal{X}(\mathop{}\!\mathrm{d} t) = f^*(x).
\]
The last inequality being due to Lebesgue differentiation theorem (the derivative of the integral of a function equals the original function), which is true when $f^*$ is bounded.
\paragraph{Curse of dimensionality.}
Once universal consistency has been derived, statisticians focus on asymptotic rates of convergence, as well as non-asymptotic generalization bounds.
Those results are more informative than universal consistency, similarly to the fact that concentration inequalities are arguably stronger than the central limit theorem which is stronger than the law of large numbers.
When providing non-asymptotic generalization bounds, we should specify a loss function.
When we output continuous values, $\mathcal{Y}=\mathbb{R}$, for algebraic reasons, people tend to use the mean-squares loss, defined as $\ell(z, y) = \norm{z-y}^2$.
In this setting, $f^*(x) = \E_\rho\bracket{Y\,\middle\vert\, X=x}$, and under the condition that $f^*$ is Lipschitz plus some mild condition on $\rho$ and classical assumptions on the weighting scheme \citep{Gyorfi2002}, it is possible to get convergence rates of the type
\begin{equation}
\label{par:eq:nn_rates}
{\cal R}(f_n) - {\cal R}(f^*) \leq c n^{-\frac{2}{d+2}},
\end{equation}
for $c$ a constant depending on the hardness of the problem, and $d$ the dimension of the input space $\mathcal{X}$.
The dependence in $d$ is explained by the fact that to cover the space $[0,1]^d$ with precision $\epsilon$ in order to have meaningful neighbors to predict the value of a Lipschitz function, one need $\epsilon^{-d}$ points.
Meaning that as the dimension increases, one will need exponentially more points in order to guarantee a given risk level, a phenomenon which is referred to as the {\em curse of dimensionality}.
The curse of dimensionality is troublesome for machine learning problems where the input dimension can be really large such as on images with millions of pixels.
\paragraph{Leveraging smoothness and higher-order information.}
It is possible to break the curse of dimensionality in~\eqref{par:eq:nn_rates} by leveraging additional assumed structure of $f^*$.
In particular, if $f^*$ admits smooth derivatives, one can use Taylor formula, and try to infer derivatives values by finite differences.
This will lead to a much more precise estimate $f_n$ of $f^*$.
In particular, if $f^*$ is $s+1$ times differentiable in every direction, when a zeroth order Taylor expansion makes a local error of order $\epsilon$, an $s$-th order expansion will make an error of order $\epsilon^s$.
As such, it is natural to hope for an estimator $f_n$ that achieve the following generalization bound,
\begin{equation}
\label{par:eq:smooth_nn_rates}
{\cal R}(f_n) - {\cal R}(f^*) \leq c n^{-\frac{2s}{d+2s}},
\end{equation}
This is exactly the rates that local polynomial methods are achieving.
We refer to \cite{Tsybakov2009} for additional details on the matter.
Note that higher-order derivatives need more points in the neighborhood of $x$ in order for finite difference methods to provide a good estimate of those derivatives.
As such, we might expect a transitory regime, before the number of samples $n$ gets sufficiently large, where there are not enough points in the neighborhood of $x$ to benefit from high-order smoothness.
\subsection{Reproducing kernel methods}
\label{int:krr}
Suppose that you are the same real agent, and that you encounter a new house that is really different from the one you have in your database, so that local averaging prediction is meaningless.
For example, suppose that the house is huge, but it is not close to any good schools, and while you have already sold big houses, or houses without good nearby schools, you have no records for big houses without good nearby schools.
Then a natural idea is to proceed with factor analysis, trying to figure how the price of a house is driven by its size, and by good nearby schools.
Simple factor analysis model supposes that characteristics $x \in \mathbb{R}^d$ combine linearly to form the output $y\in\mathbb{R}$.
This leads to the search of an estimate $f_n$ as a linear model $f_n(x) = w^\top x$, for $w\in\mathbb{R}^d$ to be determined in order to ensure that $f_n(X_i) \approx Y_i$ on your training dataset $(X_i, Y_i)_{i\leq n}$.
\paragraph{Packing features.}
Consider the case where $\mathcal{X}=\mathbb{R}$ and when plotting $(X_i, Y_i)_{i\leq n}$ seems to indicate that a quadratic relationship links $x$ to $y$ rather than a linear one, as illustrated on Figure \ref{par:fig:lin_reg}.
In this case, one can enrich the input $x$ with the features $\phi(x) = (1, x, x^2) \in \mathbb{R}^3$, and perform a linear regression with the data $(\phi(X_i), Y_i)_{i\leq n}$.
When it is hard to visualize the relationship between $x$ and $y$, one might be tempted to concatenate a lot of features in the vector $\phi(x)$ in order to make sure to be able to predict $y$ from $\phi(x)$.
However, such a technique is prone to learn spurious correlation between features and outputs of the training data, and not to generalize well to unseen I/O pairs.
This phenomenon is called overfitting.
\begin{figure}
\centering
\includegraphics[width=.3\textwidth]{drawings/set_reg.pdf}
\includegraphics[width=.3\textwidth]{drawings/lin_reg.pdf}
\includegraphics[width=.3\textwidth]{drawings/qua_reg.pdf}
\caption{Plotting $(X_i, Y_i)_{i\leq n}$ (left) indicates that a quadratic regression (right) is better suited to derive $y$ from $x$ than a linear regression (middle).
However a quadratic regression is nothing but a linear regression with features $(1, x, x^2)$.}
\label{par:fig:lin_reg}
\end{figure}
\paragraph{Avoiding overfitting.}
If we directly find the predictor $f_n(x) = w^\top\phi(x)$ by minimizing $\sum_i \ell(w^\top \phi(X_i), Y_i)$, our method will likely overfit the training data.
Overfitting is often the fruit of a linear combination of features $\omega^\top\phi(X) \approx 0$ for $\omega \in \mathbb{R}^d$ (assuming $\phi(X)\in\mathbb{R}^d$) that is noise spuriously correlated with $Y$ on the training data $(X_i, Y_i)_{i\leq n}$, leading to $w = C\cdot\omega$ for a big $C$.
As a consequence, a way to avoid overfitting is to constraint $w$ to have a small norm, or to look for the minimizer of the regularized objective $\frac{1}{n} \sum_i \ell(w^\top \phi(X_i), Y_i) + \lambda \Omega(w)$, for $\Omega:\mathbb{R}^d\to\mathbb{R}$ a regularizing function and $\lambda\in\mathbb{R}_+$ a regularizing parameter.
To ease this minimization, it is possible to use $\Omega(w) = \norm{w}^2_2$.
An interesting alternative is to use the $\ell_1$-norm $\Omega(w) = \norm{w}_1$ as it pushes $w$ to be sparse \citep{Tibshirani1996}, which makes the predictor $f_n(x) = w^\top\phi(x)$ relatively simple to interpret.
Sparsity is a rich concept \citep{Bach2012}, yet it was not a focus of this thesis.
In the following, we will consider $\Omega(w) = \norm{w}^2_2$.
\paragraph{Kernel methods.}
When minimizing $\sum_i \ell(w^\top\phi(X_i), Y_i) + n\lambda \norm{w}^2_2$, it is clear that $w^* \in \Span(\phi(X_i))_{i\leq n}$ as any part of $w$ supported on the orthogonal of the span of the $(\phi(X_i))_{i\leq n}$ will not change the value of any $w^\top \phi(X_i)$ but will higher $\norm{w}_2$, a fact that is referred as the representer theorem \citep{Scholkopf2001}.
Looking for $w\in\mathbb{R}^d$ under the form $w = \sum_{i} \alpha_i \phi(X_i)$ with $\alpha \in \mathbb{R}^n$ leads to the new problem $\sum_i \ell([K\alpha]_i, Y_i) + \lambda \alpha^\top K \alpha$, with $K = (\scap{\phi(X_i)}{\phi(X_j)})_{i, j\leq n} \in \mathbb{R}^{n\times n}$, $[K\alpha]_i$ the $i$-th entry of the vector $K\alpha \in\mathbb{R}^n$ and $f_n(x) = \sum_i \alpha_i \scap{\phi(X_i)}{\phi(x)}$.
As one can notice, everything can be expressed through the scalar product $k:\mathcal{X}\times\mathcal{X}\to\mathbb{R}_+; (x, x') = \scap{\phi(x)}{\phi(x')}$, forgetting about the features map $\phi:\mathcal{X}\to\mathbb{R}^d$.
As such, rather than making explicit the features, one can make explicit the kernel $k$, under the sole condition that for any $m\in\mathbb{N}$ and $(x_i)_{i\leq m}\in\mathcal{X}^m$, $k(x_i, x_j)_{i,j\leq m}$ is symmetric semidefinite positive.
In particular, when $\mathcal{X}$ is a collection of structured objects, it can be easier to describe how similar are two $x$ through $k$ rather than deriving features $\phi$.
When $\ell(z, y) = \norm{z - y}^2$, the empirical risk minimization admits a closed form solution
\[
f_n(x) = K_x^\top (K + n\lambda I)^{-1} \mathbf{Y},
\]
with $\mathbf{Y} = (Y_i)_{i\leq n} \in \mathbb{R}^n$ and $K_x = (k(X_i, x))_{i\leq n} \in \mathbb{R}^n$.
As $n$, the number of data, goes to infinity, this converges to
\(
f_\lambda(x) = \bar{K}_x (\bar{K} + \lambda I)^{-1} f^*
\)
where $\bar{K}$ operates on functions as $\bar{K}(f) = x\to \int k(x, x')f(x')\mathop{}\!\mathrm{d}\rho_\mathcal{X}(x')$ and $\bar{K}_x(f) = \bar{K}(f)(x)$.
And as $\lambda$ goes to zero, under mild reachability assumption, $f_\lambda$ converges to $f^*$ \citep{Caponnetto2007}.
An advantage of kernel methods is that it reduces the search of $w\in\mathbb{R}^d$ to the search of $\alpha\in\mathbb{R}^n$ which is valuable when $d >> n$.
This is known as the ``kernel trick''.
\paragraph{Reproducing kernel Hilbert space (RKHS).}
Kernel methods are backed-up by a rich mathematical concept, linked with well-behaved classes of functions.
It can be introduced in a completely different fashion, as it has been historically by \cite{Aronszajn1950}.
Consider a scalar product of functions mapping $\mathcal{X}$ to $\mathcal{Y}=\mathbb{R}$ and the class of functions ${\cal F} = \brace{f:\mathcal{X}\to\mathcal{Y}\,\middle\vert\, \norm{f} = \sqrt{\scap{f}{f}}<+\infty}$.
Suppose also that the evaluation maps $L_x:{\cal F}\to\mathbb{R}; f\to f(x)$ are bounded linear operators.
This implies the existence of representer $k_x\in{\cal F}$ such that the ``reproducing property'' $L_x(f) = \scap{k_x}{f}$ holds for any $x\in\mathcal{X}$ and $f\in{\cal F}$.
It can be shown that ${\cal F}$ is the completion of the linear combination of $(k_x)_{x\in\mathcal{X}}$, ${\cal F} = \overline{\Span(k_x)_{x\in\mathcal{X}}}^{\norm{\cdot}}$.
One inclusion is due to the fact that $k_x\in{\cal F}$ and that ${\cal F}$ is close by linear combination and completion; the other due to the fact that if $f\in{\cal F}\cap (k_x)_{x\in\mathcal{X}}^\perp$ than $f(x) = \scap{k_x}{f} = 0$.
Hence, there is a unique association between the kernel $k:\mathcal{X}\times\mathcal{X}\to\mathbb{R}; (x, x') \to \scap{k_x}{k_{x'}}$ and the so-called reproducing kernel Hilbert space ${\cal F}$ \citep{Mercer1909}.
\paragraph{Smoothness adaptability.}
Under the knowledge that the target function $f^*$ belongs to a RKHS ${\cal F}$, using the associated kernel $k$ generates good learning rates, akin to linear regression.
Moreover, when using the mean-squares error, one can guarantee universal consistency of the method described above, under the assumption that ${\cal F}$ is dense in $L^2(\rho_\mathcal{X})$, a property that is verified for many usual kernels \citep{Micchelli2006}.
But what about the rates in this case, do they deteriorate badly when $f^*\notin{\cal F}$?
How does kernel methods adapt to the regularity of the function $f^*$?
Indeed the operator $\bar{K}$ introduced above provides the answer.
In fact, it is possible to show that $L^2(\rho_\mathcal{X}) = \ima \bar{K}^0$ and ${\cal F} = \ima \bar{K}^{1/2}$ (we provide many details on this operator in sections \ref{rates:app:rkhs} and \ref{lap:app:operators}), and that by only adapting the regularization parameter $\lambda$ (which is usually done through cross-validation) the excess of risk for the least-squares error of the estimator introduced above is ${\cal O}(n^{-2q/(2q+1)})$ for the biggest $q\in(0,1)$ such that $f^*\in\ima \bar{K}^q$ \citep{Caponnetto2007}.
\paragraph{Transitory regimes.}
While research papers tend to showcase rates that are best in terms of exponents raising the number of samples, those rates are often not observed in practice before accessing a high number of samples.
As such, practitioners might prefer to write
\begin{equation}
\label{par:eq:kkr_rates}
{\cal R}(f_n) - {\cal R}(f^*) \leq \inf_{0\leq t\leq q} c_t n^{-\frac{2t}{2t+1}},
\end{equation}
for the biggest $q\in(0,1)$ such that $f^*\in\ima\bar{K}^q$, and some constants $(c_t)_{0\leq t\leq q}$.
The transitory regime corresponds to the case where the best bound is not achieved for $t=q$, and which is when, if the kernel $k$ is leveraging smoothness, points are too far apart to benefit from higher-order information.
\subsection{Neural networks and deep learning}
Most of the state-of-the-art algorithms derived through machine learning are currently based on deep learning.
Deep learning consists in long artificial neural networks.
Artificial neural networks are a cascade of linear operators and non-linearities that are optimized through gradient descent on empirical risks.
This model appeared in the middle of the 20\textsuperscript{th} century \citep{McCulloch1943,Rosenblatt1958}, and were advocated by a few researchers in the 90s \citep{Lecun1995}.
It rose to dominance when implemented on graphical processing units, achieving astonishing results on image recognition \citep{Krizhevsky2012}.
\paragraph{Cascade of linear operation.}
A neural network $f_n:\mathbb{R}^d\to\mathbb{R}^m$ with three hidden layers is parametrized with four matrices $W_1 \in \mathbb{R}^{h_1\times d}$, $W_2\in \mathbb{R}^{h_2\times h_1}$, $W_3\in \mathbb{R}^{h_3\times h_2}$, $W_4\in \mathbb{R}^{m\times h_3}$ and reads
\[
f_n(x) = W_4\sigma (W_3\sigma (W_2 \sigma(W_1x))),
\]
with $h_1, h_2, h_3 \in \mathbb{N}$ the number of neurons in the first, second and third hidden layers, $\sigma:\mathbb{R}\to\mathbb{R}$ a non-linearity, {\em e.g.} $\sigma(x) = \max(x, 0)$, and the convention $\sigma((x_1, \cdots, x_n)^\top) = (\sigma(x_1), \cdots, \sigma(x_n))^\top$.
We illustrate it in Figure~\ref{par:fig:nn_archi}.
Given a loss $\ell$, some data ${\cal D}_n = (X_i, Y_i)_{i\leq n}$, a regularizer term $\lambda\Omega(W_1, W_2, W_3, W_4)$ and the consequent regularized empirical risk ${\cal R}_{{\cal D}_n}$, the parameters $W_1, W_2, W_3, W_4$ can be optimized in order to minimize ${\cal R}_{{\cal D}_n}(f_n)$.
This minimization is usually done with gradient descent, or stochastic gradient descent \citep{Robbins1951}, which is better suited for large scale learning problems \citep{Bottou2007}.
Gradients are obtained thanks to the chain rule, and while the computation of $f(x)$ can be seen as forward propagation from $x$ to $W_1x$ to $W_2\sigma(W_1x)$ and so on, the chain rule naturally leads to {\em backward propagation} or {\em backpropagation}, which is the name used by researchers to refer to the derivation of gradients with neural networks.
Astonishingly, while such a procedure is supposed to stall into undesirable local minima, it has not constrained the recent successes of deep learning.
\begin{figure}
\centering
\input{drawings/nn}
\caption{Representation of a neural network with three hidden layers and parameters $(d, h_1, h_2, h_3, m) = (3, 5, 5, 4, 2)$, as a weighted directed acyclic graph.
The input $x = (x_{(i)}) \in \mathbb{R}^3$ is passed through the first set of edges, leading to the first layer with value $x_1 = (x_{1, (i)}) \in \mathbb{R}^5$ specified by $x_{1,(i)} = \sigma(\sum_{1\leq j\leq 3} W_{1, (i,j)} x_{(j)})$ based on the set of weights $W_1 = (W_{1, (i, j)}) \in \mathbb{R}^{5\times 3}$ and the non-linearity $\sigma:\mathbb{R}\to\mathbb{R}$, and so on until reaching the output $y \in \mathbb{R}^2$.
The wording ``deep learning'' refers to learning with neural networks that
have a really large number of hidden layers, which can lead to hundreds of
billions of parameters \citep[see {\em e.g.}][]{Brown2020}.
}
\label{par:fig:nn_archi}
\end{figure}
\paragraph{Hierarchy and convolution.}
In the last paragraph, we have described the most basic architecture of a neural network, only made of fully connected layers.
There are many architectures that refine this basic neural network, the most well known are recurrent neural networks \citep{Williams1986} that allow dealing with time series of different length and convolutional neural networks that were proven successful to understand images \citep{LeCun2015}.
We refer the interested reader to this last paper to get a more precise idea on what convolutional neural networks are.
In substance, a one dimensional convolutional filter replaces the linear mapping $\mathbb{R}^{h_i}\to\mathbb{R}^{h_o} x\to Wx$ with $W \in \mathbb{R}^{h_0\times h_i}$ by the linear mapping $\mathbb{R}^{h_i}\to\mathbb{R}^{h_i-p}; x \to w*x$ with $w\in \mathbb{R}^p$ and $(w*x)_i = \sum_{j=1}^p w_i x_{i+j}$.
This is useful when $x$ has a translation-invariant structure \citep{Fukushima1980}.
Rather than learning to recognize the same pattern at different places by tuning similarly different columns of $W$, it allows for parameter sharing.
As one filter allows for the recognition of one pattern, a convolutional layer is made of several filters, whose responses are concatenated along a new dimension in order to form the layer output.
After passing through a filter bank, the response is usually pruned by keeping only local maxima, an operation known as max-pooling.
This subsampling operation builds a hierarchical structure, allowing for further filters to learn broader patterns \citep{Behnke2003}.
Translation invariance and hierarchy are two important concepts \citep{Simon1962,Mallat1989} that have been used to design others ``hand-made'' methods that were used in the past to analyze images \citep{Lowe1999,Dalal2005}.
\paragraph{Plethora of architectures.}
While convolutional neural networks are the model of choice to work with images, they do not exhaust what deep learning is about.
Recently, transformer models \citep{Vaswani2017} have become the model of choice to deal with time series in natural language processing, with amazing realizations in language understanding \citep{Devlin2019} and text generation \citep{Brown2020}, and generative adversarial networks \citep{Goodfellow2014} introduced earlier have also led to impressive applications such as ``deepfakes'' \citep{Karras2019}, which are synthetic images that look real and might be used for malicious purposes.
The rapid spread of deep learning is partly due to its ease to implement through differential programming libraries that implement automatic differentiation and backpropagation when forward propagation is specified \citep{Abadi2016,Paszke2019}.
This allows practitioners to easily develop and test new ideas and architectures, even though training neural networks is known to be quite unstable and greedy in terms of time and energy \citep[see {\em e.g.}][]{Garcia2019}.
\paragraph{Statistical properties.}
While neural networks work better than local averaging and kernel methods, statisticians have not arrived at a consensus in order to describe theoretically why neural networks work so well.
In particular, it is currently hard to state generalization bounds akin~\eqref{par:eq:nn_rates}, \eqref{par:eq:smooth_nn_rates} and \eqref{par:eq:kkr_rates}.
Because of their wide-spread use, many are trying to derive a better understanding of those models.
Some are trying to come up with similar models based on concepts that are better understood such as wavelets \citep{Bruna2013}, kernels \citep{Mairal2014} or sparse coding \citep{Papyan2017}.
Others are looking at the properties of neural networks once the training has taken place \citep{Mahendran2015,Selvaraju2017}, notably discovering the possibility to fool them with imperceptible noise \citep{Szegedy2014,Ilyas2019}.
Finally, statisticians are trying to explain some of their characteristics with different tools, could it be some measure of complexity \citep{Bartlett2017,Zhang2021}, group theory and harmonic analysis \citep{Mallat2016}, statistical physics \citep{Spigler2019} or asymptotic limit \citep{Jacot2018,Chizat2019,Mei2019}.
\section{The practice}
In this section, we discuss the daily job of data scientists before confronting this reality with the prior theory in order to discuss the limitations of the supervised learning framework.
\subsection{The job of data scientists}
There are arguably four steps in a project of a data scientist when it comes to machine learning.
While we present those tasks in a downstream fashion, in practice, it is important to get a big picture and a rough sketch of each step first in order to ease the transitions between steps and avoid going back too often to prior steps.
All along the process, intuition can be found by reflecting on how the work would be accomplished manually, that is without learning perspectives.
\paragraph{1. Define the problem.}
First of all, data scientists should be clear about the task to solve, look at the current solution, and see how it can be improved with machine learning.
In order to prove the usefulness of the new method, they should define some clear measures of success, and benchmark simple baselines.
At this point comes the definition of the output space $\mathcal{Y}$ and of the loss $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ that should be defined in coherence with the precedent measures of success.
Once the target space $\mathcal{Y}$ is well-defined, the next step is to find a set of relevant features that defines the input space $\mathcal{X}$ and allows prediction of targets.
Of course, the features that can be accessed, hence the definition of $\mathcal{X}$, depend on the data one can get.
\paragraph{2. Get clean data.}
Once $\mathcal{X}$ and $\mathcal{Y}$ are well-defined, it is time to find a set of I/O examples $(X_i, Y_i)_{i\leq n}$.
This is often the most time-consuming task.
How many samples are needed regarding the statistical hardness of the task at hand? Where to scrap data? How to easily associate outputs to inputs?
Once enough samples have been collected, it is usual to set aside a part of those to form a test dataset, that would only be looked at before deployment to validate the working status of the built system.
The remaining data form the train set.
At this point, data is often very cluttered, implying a consequent amount of cleaning before feeding a machine learning model.
In particular, when collected from different sources, there might be some coordinates missing for various input samples.
For example when working with countries, it will be hard to get an indication of the number of people below poverty line in North Korea.
A decision should be made with those missing data, could it be filling with the feature average, or dropping completely the corresponding sample.
Similarly, the practitioners might have to deal with outliers or data sources that got stalled on wrong values.
\paragraph{3. Train a model.}
Once clean data have been collected, machine learning models are ready to be trained.
In order to help models, some features engineering might be useful, such as centering and normalizing the variance of the data, or applying more advanced transformations to get samples that are well-behaved, {\em e.g.} following a unit Gaussian distribution.
It is usual to combine different models together, when those models capture different relationships between inputs and outputs, in order to boost the overall performance, a procedure known as boosting \citep{Schapire1990}.
Hyperparameters are generally tuned with cross-validation.
Cross-validation consists in splitting the samples into a given number of folds, using all the folds but one to train the models and getting a measure of success on the last fold before averaging this measure over the different folds that can be retained for testing.
This associates a score to a method with a given set of hyperparameters.
Best hyperparameters are chosen as the ones leading to the highest score.
Practitioners should be careful not to try too many models or hyperparameters as the more they try, the more likely will be learned spurious correlations that only explain the training samples I/O relationship.
\paragraph{4. Deploy the model.}
Once a model has been learned to predict output from input, one should excavate the test dataset, and see if the method performs better than baselines.
If so, it is time to benefit from this new method and put it into production.
At this point, it is important to monitor the quality of new input data, to make sure that it is not corrupted, and that its distribution is similar to the distribution of training data.
There might also be an emphasis on engineering problems, such as managing databases and computing architectures.
For the curious reader, there are many good references to prepare future data scientists for their job \cite[see {\em e.g.}][]{Geron2017}.
\subsection{Framework limitations}
\paragraph{Loss design.}
In the formalization of supervised learning that we gave earlier, we assumed that the loss $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ is clearly defined and that the risk ${\cal R}$ to minimize is nothing but the average over samples of the error made by a predictor measured with this loss.
In some critical applications, such as medical one, one might prefer to build ${\cal R}_{{\cal D}_n}$ as the worst case of $(\ell(f(X_i), Y_i))$ or build ${\cal R}$ as some quantile of the pushforward distribution $\ell(f(\cdot), \cdot)_\# \rho$ rather than building them as their respective means.
Moreover, in practice, it might be hard to define clearly the loss $\ell$.
For example, it is the case when it comes to style transfer, that is taking two images and outputting an image with the content of the first one and the style of the second, or to super-resolution, that is enhancing the resolution of an image \citep{Johnson2016}.
The same is true with our cooking example, how to access the quality of a recipe derived by a computer? Should it be the most healthy, the most tasty or somewhat healthy and tasty? Can we derive such losses based on chemical considerations? Or should we only measure success based on people reviews, meaning high costs to evaluate the risk ${\cal R}$?
Sometimes, losses can be partially or totally designed by leveraging the structure of the problem.
For example, on problems where there is a desired invariance defined by a set $G$ of functions $g:\mathcal{X}\to\mathcal{X}$ such that we should have $f\circ g = f$, if this invariance is not built into the model, it can be built into the loss by replacing $\ell(f(X), Y)$ with $\sum_{g \in G}\ell(f(g(X)), Y)$, which is notably the idea of ``data augmentation''.
Another example is when we have a parametric model of the relation between $X$ and $Y$, $Y = g(X, \theta, \epsilon)$ for $g$ a known function, $\theta$ an unknown parameter to optimize and $\epsilon$ some random noise.
The parameter $\theta$ can be fitted by maximizing the likelihood of having observed ${\cal D}_n = (X_i, Y_i)_{i\leq n}$ under the model.
Assuming the samples are independent, this probability factorizes as the product of the probability of observing each $(X_i, Y_i)$.
As a consequence, maximizing the likelihood, or equivalently the log-likelihood, is similar to minimizing the risk ${\cal R}$ defined through the loss $\ell(X, Y, \theta) =-\log \Pbb\paren{Y\,\middle\vert\, X; \theta}$ with the last term designing the probability of observing $Y$ given the observation of $X$ under the model parametrized by $\theta$.
When the parametric model is an exponential family, this procedure is to be linked with generalized linear models \citep{Nelder1972}.
\paragraph{Independent identically distributed variables?}
Suppose you want to predict annual growth from indicators, such as the gross domestic product, regarding a country.
You might collect a dataset $(X_{c,t}, Y_{c,t})_{c\in C, t\in T}$ with many countries, indexed by a set $C$, and many years, indexed by a set $T$, features $X$ and corresponding growth $Y$, but this dataset will violate the independent identically distributed assumption of statistical learning.
For example, an oil crisis happening in year $t$ will have a repercussion on each $(Y_{c,t})_{c\in C}$, and if one try to add oil prices in the features $X$ to enforce the independence of the conditional variables $\paren{Y_{c,t}\,\middle\vert\, X_{c,t}}_{c\in C}$, than the $(X_{c,t})_{c\in C}$ will not be independent as this factor will be constant over countries.
Similarly, for a given country, indicators are arguably forming a non-stationary process that depends on politics in place, hence $(X_{c,t})_{t\in T}$ is not a family of independent random variables.
However, many data scientists might be tempted to still use the many resources provided by supervised learning, without trying to leverage the underlying processes.
More importantly, suppose that one trains a model on western countries, benefiting from higher quality data, before deploying the model to help some decision-making in developing countries.
There might be some important cofactors such as quantitative easing policies, or trust of populations in money, that change completely the picture when it comes to predicting growth from country indicators.
As such, the testing distribution is not the same as the training distribution, a phenomenon known as {\em covariate shift} \citep{Heckman1979,Shimodaira2000}.
Recently, this problem has come back to denounce the risk of automatic systems that reproduce human-biases \citep{Caliskan2017}, which has found an echo in the civil society \citep{Benjamin2019}.
Interestingly, while data scientists often try to artificially squeeze their problems into the framework of supervised learning, many researchers are trying to leverage specific settings, such as the fact that data might be coming from different datasets \citep{Arjovsky2019}.
\paragraph{Practice and theory regarding generalization.}
In theory, generalization error might be given by derivations allowing to get results such as~\eqref{par:eq:nn_rates}, \eqref{par:eq:smooth_nn_rates} and \eqref{par:eq:kkr_rates}.
In practice, those bounds might be hard to compute because of unknown constants, yet they provide precious insights on the number of samples to consider in order to achieve statistical significance.
For example, bounds reading $d/n$ suggests that the number of samples $n$ should be bigger than the number of dimensions $d$ of the input space $\mathcal{X}$, while bounds in $n^{-1/d}$ suggests that the number of samples should be exponentially bigger than the number of dimensions.
Anyhow, data scientists mostly compute generalization scores by computing the empirical risk of a predictor $f_n$ on a fresh test set of data.
According to this practice, researchers might study the benefits of the test set validation procedure.
For example, they may leverage the distribution of the empirical test error in order to transform a predictor $f_n:\mathcal{X}\to\mathcal{Y}$ into a confidence interval predictor $f_n:\mathcal{X}\to 2^\mathcal{Y}$ with a given probability to be valid \citep{Vovk2008}.
\chapter{Partially Supervised Learning}
Many data scientists spend more time scrapping, annotating and cleaning data than fine-tuning models.
This motivates the following question: can we derive a more generic framework than the one of supervised learning in order to learn from cluttered data?
In this thesis, this question is approached through the lens of weakly supervised learning, assuming that the bottleneck of data collection lies in data annotation.
This chapter summarizes our main contributions to weakly supervised learning.
It is based on our published work \cite{Cabannes2020,Cabannes2021,Cabannes2021c}, which is reproduced in part \ref{part:weakly}.
\section{Navigating frameworks of weakly supervised learning}
In this section, we give a brief introduction to weakly supervised learning.
Weakly supervised learning is concerned with the setting where it is easy to get $n$ inputs $(X_i)_{i\leq n}$, but it is hard to get the corresponding outputs $(Y_i)_{i\leq n}$, a setting which is relevant for many applications.
For example, when dealing with images, one can easily scrap zillions of images on the web, which provides many input data, but getting the corresponding outputs demands laborious annotations of data.
The situation is similar when it comes to medical applications, such as understanding drug interactions \citep{CTDPfizer2013} or recognizing cancers with radiography.
While one can access databases from hospitals to get many X-ray images, recognizing cancers on those requires the expertise of a radiologist, which is expensive to access.
Indeed, annotation costs can critically add up when one plans to use powerful models that require millions of data points to be trained.
As a counter example, notice that weakly supervised learning does not help when input data are features coming from heterogeneous sources leading to missing data.
We shall discern three types of weak supervision, the last two being not completely disjoint.
\paragraph{1. Global statistics on groups of inputs.}
This setting consists in accessing global information on bags of samples -- {\em e.g.} knowing that half of the $(Y_i)_{i\in I}$ are zeros for a given $I\subset[n]$.
Examples of global statistics supervision include multiple-instance learning \citep{Dietterich1997} and learning from label proportion \citep{Quadrianto2009}.
We refer to those articles for motivations, descriptions and applications of those two problems.
\paragraph{2. Weak classifiers.}
A second approach consists in assuming the access to many weak classifiers $\phi_j:\mathcal{X}\to\mathcal{Y}$ for $j\in[p]$ that weakly correlate with $f^*$.
Those classifiers might model labelers from a crowdsourcing platform, experts or noisy measurements.
More generally, they might be provided by side information, coming from {\em distant supervision} \citep{Mintz2009,Craven1999} or {\em transferred} from algorithms that have been designed for a related task \citep{Pan2010}.
Recently, \cite{Ratner2020} have implemented a software interface based on this approach.
\paragraph{3. Incomplete annotation.}
Finally, weak supervision might be understood as the access to partial knowledge on what $Y_i$ is for each $i\in[n]$.
In some instances, partial observation on $Y_i$ can be cast as a set of potential labels $S_i\subset\mathcal{Y}$ that are compatible with this partial observation, which is the setting of partial supervision \citep{Cour2011}.
Partial supervision is a generalization of semi-supervised learning, which has been the classical approach to overcome the bottleneck of data annotation \citep{Chapelle2006} and consists in learning from a set of labeled data, $S_i=\brace{Y_i}$ for $i \in L\subset[n]$, and of unlabeled data, $S_i = \mathcal{Y}$ for $i\in [n]\setminus L$.
Beyond those three settings, limitations that motivate weakly supervised learning might be tackled by leveraging human knowledge under the form of priors \citep{Mann2010} or of function architectures \citep[reviving old approaches of artificial intelligence such as][]{Muggleton1994}.
\section{The importance to create consensus}
In this section, we review our first approach to the problem of learning from partial supervision. We first formalize this setting before introducing a variational objective to minimize and providing guarantee regarding this approach.
\subsection{Partial supervision setting}
In this thesis, hence in the following, we focus on partial supervision, which is also known as {\em superset learning}.
In other terms, we assume the access to weak supervision for each label $y$ under the form of a set $s\subset\mathcal{Y}$ that contains the true labels $y \in s$.
To formalize this problem, we introduce the set ${\cal S}\subset 2^\mathcal{Y}$ of closed subsets of $\mathcal{Y}$ and the distribution $\tau\in\prob{\mathcal{X}\times{\cal S}}$ generating samples $(X_i, S_i)_{i\leq n} \sim \tau^{\otimes n}$.
The goal is still to minimize the risk
\begin{equation}
\label{int:eq:risk}
{\cal R}(f) = \E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)},
\end{equation}
for a known loss function $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ and a distribution on I/O pairs $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$.
In contrast with supervised learning, $\rho$ will never be accessed, nor any samples $(X_i, Y_i) \sim \rho$.
In order to motivate our theoretical setting, we now discuss simple instances of partial supervision.
\begin{example}[Classification with attributes]
Suppose that you want to learn fine-grained classes on images, {\em e.g.} to precisely distinguish ``caracals'' and ``domestic cats'', as well as ``hedgehog mushrooms'' and ``shiitakes''.
It will probably be hard for commoners hired on crowdsourcing platforms to label images with precise classes.
Yet, this commoner might easily label attributes, such as ``tufted ears'' characterizing ``caracals'' among ``felines'', or ``spines rather gills underside of the cap'' and ``tan irregular cap'' characterizing ``hedgehogs'' among ``mushrooms''.
Those partial labels can easily be cast as sets of labels: ``feline'' would be $\brace{\text{``lion'', ``panther'', \dots}}$ and ``tufted ears'' be $\brace{\text{``great horned owl'', ``Araucana chicken'', \dots}}$.
The key idea here is that it is much cheaper to get a set of many partial labels that is informative enough to learn $f^*$ than a set of few complete labels enabling the same learning.
\end{example}
\begin{example}[Ranking with partial ordering]
Consider a ranking problem where, given a user characterized by some features $x$, the goal is to learn their preference over $m$ flowers, defining the output $y\in\mathfrak{S}_m$ as an ordering of $m$ elements.
Once again, getting the full $y$ is a laborious task, as one might have a hard time to rank all $m$ flowers. In contrast, one might easily provide partial ordering information, such as ``I prefer roses to lilies'', or ``tulips are my favorite'', which can be cast as the set of total orderings that verify those partial orderings.
\end{example}
\begin{example}[Regression with censored data]
For many regression problems, it is not possible to get exact labels but only possible to access bins or censored labels $[a,b] \ni y$.
This could be due to the use of a measuring scale, or, if retaking example \ref{ex:housing}, due to some uncertainty regarding a price at which a house has been sold.
\end{example}
\subsection{Solution definition through the infimum loss}
Currently, the problem $(\ell, \tau)$ is ill-defined since we cannot define clearly the goal \eqref{int:eq:risk} from those two objects only -- $\ell$ being a loss function on pairs of outputs and $\tau$ a distribution on $(X, S)$.
One way to redefine a solution in this context is to keep the precedent variational point of view and define a loss $L:\mathcal{Y}\times 2^{\mathcal{Y}}\to\mathbb{R}$ that given a prediction $z\in \mathcal{Y}$ and an observation $S\subset \mathcal{Y}$, provides a compatibility or performance score.
Hence the new objective
\[
{\cal R}_S(f) = \E_{(X, S)\sim\tau}\bracket{L(f(X), S)}.
\]
Yet, how to derive the loss $L(z, S)$ for $z\in\mathcal{Y}$ and $S\subset\mathcal{Y}$ when the original task was specified by the loss $\ell(z, y)$ with $y\in S$?
Arguably there are three possibilities.
\begin{enumerate}
\item Bounding the original excess risk from above with the supremum loss $L(z, S) = \sup_{y\in S}\ell(z, y)$.
Suprema have been used in statistics for robustness purposes \citep{Wald1945}.
The idea here is that if we prevent ourselves against the worst possible candidates (in terms of error given a prediction $z$), we will prevent ourselves against whatever is the ground truth label in $S$.
Yet, as we will discuss later, this approach is too conservative in our setting.
\item Matching a bit of all the elements in the set $S$ with the average loss $L(z, S) = \frac{1}{\abs{S}} \sum_{y\in S} \ell(z, S)$.
This loss has the benefit of being agnostic on what should be the true $y\in S$ regardless of what the prediction $z$ is.
Under symmetry assumptions of the original loss, for example when $\ell$ is the 0-1 loss in classification, averaging candidates is a reasonable solution.
When the loss is not symmetric, it can insidiously bias the solution as we describe on Figure \ref{il:fig:inconsistency}.
In many cases, it is possible to correct for this asymmetry, which techniques implicitly implied by Proposition \ref{df:prop:init} and illustrated with the filling with zero techniques in section \ref{df:sec:ranking}.
\item Making sure that the prediction $z$ matches at least one element with the infimum loss $L(z, S) = \inf_{y\in S} \ell(z, S)$.
When $\ell$ is seen as a distance, $L$ is its natural extension to sets.
\cite{Hullermeier2014} has referred to it as the optimistic loss, since given a prediction $z$, it somehow disambiguates $y\in S$ by considering the best label in order to minimize $\ell(z, y)$ under the constraints $y\in S$.
\end{enumerate}
Let us take a step back and recall the commoner providing the set ``feline'' that is $S_1 = \brace{\text{``cat'', ``lion'', \dots}}$ and the set ``tufted ears'', that is $S_2 = \brace{\text{``owl'', ``Araucana chicken''}, \dots}$, when annotating the picture of a ``caracal''.
Naturally, we would like to output a prediction $z\in S_1\cap S_2 = \brace{\text{``caracal''}}$ of a feline with tufted ears.
In other terms, we want to create consensus between observations, which is what the infimum loss provides.
The supremum loss is outcast from the good solutions, since it might disambiguate $y_1 \in S_1$ as ``leopard'' and $y_2 \in S_2$ as ``owl'', as well as the average loss as we illustrated on Figure \ref{int:fig:il_ac}.
\begin{figure}
\centering
\input{drawings/il_ac}
\hspace{2cm}
\input{drawings/il_ac_bis}
\caption{Two different figures of partial supervision in $\mathcal{Y}=\mathbb{R}^2$ without input (or $\mathcal{X}$ being a singleton).
The loss is given by the mean square error, {\em i.e.} the usual Euclidean geometry.
On both figures, we represent $\mathcal{Y}$ with two sets, $S_1$ in light gray and $S_2$ in gray, and the estimates $z = \argmin_{z\in\mathcal{Y}} \brace{L(z, S_1) + L(z, S_2)}$ provided by the infimum loss ($z^i$) and the average loss ($z^a$).
The point $y_1^a$ represents the disambiguation of $S_1$ by the average loss.
On the left figure, although $S_1$ intersects $S_2$, only the infimum loss verifies $z^i \in S_1\cap S_2$.
}
\label{int:fig:il_ac}
\end{figure}
\subsection{Theoretical guarantees}
With the infimum loss, we switch to the original supervised learning problem \eqref{int:eq:risk} to the new problem defined through the risk
\begin{equation}
\label{int:eq:inf_risk}
{\cal R}_S(f) = \E_{(X, S)\sim\tau}\bracket{\inf_{y\in S} \ell(f(X), y)}.
\end{equation}
How are those two problems related?
In particular, we would like to make sure that solving the partially supervised problem does help to tackle the fully supervised one.
Following the formalism of the previous chapter, we can try to derive a calibration inequality that relates the excess of the original risk \eqref{int:eq:risk} to the excess of the ``surrogate'' risk \eqref{int:eq:inf_risk}.
To derive such a generic inequality, we need strong assumptions, which are usually given by ensuring non-ambiguity of the partial supervision (Definition \ref{il:def:non-ambiguity}), that is for each $x$, the sets in the support of the conditional distribution $(S\,\vert\,X=x)$ intersect into a singleton $\brace{y_x} = \brace{f^*(x)}$.
Such an inequality was first provided by \cite{Cour2011} in the case of classification with the 0-1 loss, and is generalized to generic problems by Proposition \ref{il:thm:calibration}.
This technique allows us to derive rates on a structured prediction algorithm learning from partially supervised data in Theorem \ref{il:thm:learning-rates}.
In practice, given samples $(X_i, S_i)$, the practitioner can use their favorite algorithm to translate this population principle into an empirical objective to be optimized in order to get an estimate $f_n$ of the risk minimizer.
\begin{remark}[A false case against partial supervision]
Partially supervised learning and the infimum loss are often criticized because of the strength of the non-ambiguity assumption.
This should rather be seen as a limitation of theoretical analysis, rather than a limitation of the framework.
Actually, a more careful analysis based on the idea provided in the previous chapter shows that the non-ambiguity assumption is akin to the Massart noise condition, and can be used to derive much faster rates without modifying algorithms as shown by Theorem \ref{df:thm:convergence}.
This suggests that there are many interesting behaviors to characterize beyond the non-ambiguity assumption.
\end{remark}
\section{Label disambiguation to complete supervision}
In this section, we characterize the solution provided by the infimum loss as a disambiguation strategy, that consists in retrieving full supervision first, before learning from it.
This leads to a different perspective in terms of learning and practical implementations.
\subsection{Expected testing distribution}
Let us roll back a little and think back on which solution to look for when in presence of partial supervision.
The ultimate goal is to find a mapping that minimizes the generalization error on the testing distribution $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$.
Yet which distribution $\rho$ should we expect given the loss $\ell$ and the distribution $\tau \in \prob{\mathcal{X}\times 2^\mathcal{Y}}$?
The only clearly defined constraints is that the distribution should be compatible with the weak information we have seen on the data.
The notion of compatibility can be formalized through the following definition, which we introduced in \cite{Cabannes2020}.
\begin{definition}[Compatibility]
\label{int:def:comp}
Two distributions $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$ and $\tau\in\prob{\mathcal{X}\times{\cal S}}$ are said to be {\em compatible} if there exists an {\em underlying distribution} of probability $\pi\in\prob{\mathcal{X}\times\mathcal{Y}\times{\cal S}}$ such that $\rho$ and $\tau$ are the respective marginals of $\pi$ according to $\mathcal{X}\times\mathcal{Y}$ and $\mathcal{X}\times{\cal S}$ and such that for any $(x, y, s) \in \supp\pi$, we have $y\in s$.
This compatibility property will be denoted by $\rho\vdash\tau$.
\end{definition}
To discriminate the expected $\rho$ among all the distributions compatible with $\tau$, we need a principle.
It is natural to look for a distribution that satisfies some properties.
As mentioned in the last section, we would like to go for a distribution that is ``consensual'', {\em i.e.} we would like to reduce the noise of the conditional $\paren{Y\,\middle\vert\, X}$.
In other terms, we would like those conditional distributions to be as deterministic as possible.
As a consequence, we define the expected testing distribution as
\[
\rho^\star \in \argmin_{\rho\vdash\tau}{\cal E}(\rho),
\]
for ${\cal E}:\prob{\mathcal{X}\times\mathcal{Y}}\to\mathbb{R}$ a measure of determinism.
A well-known measure of determinism is the entropy ${\cal E}(\rho):=\E[-\log(\rho(X, Y))]$.
This measure is independent of the loss $\ell$, which might be a desirable property if one would like to reuse the same supervised distribution for different tasks linked with different losses.
Yet, when there is a structure on $\mathcal{Y}$ such that its explicit dimension is much bigger than the intrinsic dimension of this space seen through the loss $\ell$, the entropy will scale with the explicit dimension.
For example, in ranking problems where the goal is to output an ordering over $m$ items, which is similar to a permutation in $\mathfrak{S}_m$, the output space is of cardinality $m!$, while the intrinsic dimension is $m$.
For those problems, it is wise to incorporate the loss in the objective, in particular, we suggest the following loss-based variance ${\cal E}(\rho) := \inf_{f:\mathcal{X}\to\mathcal{Y}} \E_\rho[\ell(f(X), Y)]$.
Once the distribution $\rho$ is recovered, a function $f$ can be learned in a supervised learning fashion
\begin{equation}
\label{eq:int:df}
f^* \in \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X, Y)\sim\rho^\star}[\ell(f(X), Y)], \qquad\text{with}\qquad \rho^\star = \argmin_{\rho\vdash\tau} \inf_{f:\mathcal{X}\to\mathcal{Y}} \E_\rho[\ell(f(X), Y)].
\end{equation}
Remarkably, this second solution to the partially supervised learning problem is exactly the same as the one provided by the infimum loss.
This is based on the fact that, under really mild definition assumptions
\[
\inf_{\rho\vdash\tau} \inf_{f:\mathcal{X}\to\mathcal{Y}} \E_{\rho}[\ell(f(X), Y)] = \inf_{f:\mathcal{X}\to\mathcal{Y}}\inf_{\rho\vdash\tau} \E_{\rho}[\ell(f(X), Y)] = \inf_{f:\mathcal{X}\to\mathcal{Y}} \E_{\tau}[\inf_{y\in S}\ell(f(X), y)].
\]
A nice property of this new characterization of the solution $f^*$ is that it goes beyond the partial labeling problem.
As a matter of fact, \eqref{eq:int:df} can be applied more generally to any weakly supervised learning paradigm where the weak information acquired on the problem can be translated into hard constraints, {\em e.g.} in the case of label proportion.
\subsection{Optimization issues}
Before going any further, let us point out that any objective ${\cal E}(\rho)$ that is minimized for deterministic distributions is intrinsically not-convex.
Consider the case where there is no input space (or $\mathcal{X}$ is a singleton), $\mathcal{Y}$ is discrete, and ${\cal E}$ is a function taking as argument elements of the simplex $\prob{\mathcal{Y}}$ and outputting a real-valued score.
If ${\cal E}$ is minimized for deterministic distributions, it is minimized on extremities of the simplex.
In other terms, ${\cal E}$ should really be thought of as a concave function, making its minimization a non-convex optimization problem.
We picture level lines of such an objective on Figure \ref{df:fig:objective}.
Regarding the constraints, it should be noted that $\brace{\rho\in\prob{\mathcal{X}\times\mathcal{Y}}\,\middle\vert\, \rho\vdash\tau}$ is a convex set.
So we are trying to minimize a concave function over a convex set.
In \cite{Cabannes2021}, we suggested two different heuristics to approach this minimization procedure.
The first one consists in starting with a good ``center point'' and finding an extremity of the definition domain after gradient descent.
Our notion of center takes into account the loss to avoid insidious bias (similarly to what we discussed the average loss).
The second one is based on the Diffrac algorithm \citep{Bach2007}, and consists in convexifying the objective by adding a quadratic term that is constant on extremities, perform gradient descent, and get a solution that is a convex combination of extreme points.
We refer to sections \ref{df:sec:optim} and \ref{df:app:diffrac} for further discussions on those techniques.
\subsection{Empirical objective}
Given data $(X_i, S_i)_{i\leq n}\sim \tau^{\otimes n}$, how to translate the disambiguation principle \eqref{eq:int:df} into an empirical objective allowing to disambiguate the sets $(S_i)_{i\leq n}$ into labels $(\hat{Y}_i)_{i\leq n}$?
To answer this question, we need to translate an objective on distribution into an objective on samples.
The usual translation of risk minimization into empirical risk minimization can be thought as the approximation $\rho \approx \hat{\rho} = \frac{1}{n} \sum_{i\leq n} \delta_{X_i} \otimes \delta_{Y_i}$ plus some constraints $f\in{\cal F}$ on the functions that we minimize.
As such, we could go for
\[
\min_{(y_i \in S_i)_{i\leq n}} \inf_{f\in{\cal F}} \sum_{i=1}^n \ell(f(X_i), y_i).
\]
This will be really hard to solve in practice.
The solution we suggested is based on kernel mean embedding \citep{Muandet2017}, which is a natural way to understand the surrogate approach of \cite{Ciliberto2016}.
It consists in incorporating directly the hypothesis on the solution of the problem in the estimated distribution with $\hat{\rho} \approx \frac{1}{n} \sum_{i,j\leq n} \delta_{X_i} \otimes \alpha_j(X_i)\delta_{Y_j}$, where $\alpha:\mathcal{X}\to\mathbb{R}^n$ is a weighting scheme that states how to diffuse information $(Y_i)_{i\leq n}$ seen at $(X_i)_{i\leq n}$ to any point $x\in\mathcal{X}$.
In particular, those weights might be used to estimate the conditional distribution $\paren{Y\vert X=x}$ as $\sum_{j\leq n} \alpha_j(x)\delta_{Y_j}$.
This leads to the following empirical objective
\begin{equation}
\label{eq:int:df_emp}
(\hat{y}_i)_{i\leq n} \in \argmin_{(y_i \in S_i)_{i\leq n}} \inf_{(z_i\in \mathcal{Y})_{i\leq n}} \sum_{i,j=1}^n \alpha_i(X_j)\ell(z_i, y_j),
\end{equation}
In the case of kernel ridge regression, those weights are given by
$\alpha(x) = (K + \lambda n)^{-1}K_x$ with the notations of section \ref{int:krr}.
But they could also be given by similarity metrics, or derived with more advanced unsupervised techniques such as the one we present in \cite{Cabannes2021c} as we detail in section \ref{lap:sec:extension}.
\subsection{Theoretical guarantees}
The convergence of an estimate $f_n$ learned on the dataset $(X_i, \hat{y}_i)_i$ with $(\hat{y}_i)_{i\leq n}$ recovered through \eqref{eq:int:df_emp} cannot easily be performed with the tools presented so far.
Indeed, it is hard to understand how the different disambiguation $(y_i \in S_i)$ interacts when studying \eqref{eq:int:df_emp}.
This relates to the non-concavity of this objective: a small change in the optimal $(z_i)_{i\leq n}$ can lead to completely different optimal $(y_i)_{i\leq n}$.
While there are probably good tools to think in terms of distribution and derive a consistent empirical estimate of \eqref{eq:int:df}, in \cite{Cabannes2021}, we used the fact that usual learnability assumptions for the partial labeling problem are actually quite strong, which allows us to prove impressively fast rates.
To derive consistency of algorithm learning with partially supervised data, it is classical to assume non-ambiguity of the distribution $\paren{S\,\middle\vert\, X=x}$ as well as some regularity of the function $\mathcal{X}\to\prob{\mathcal{Y}}; x\to\paren{Y\,\middle\vert\, X=x}$.
Those assumptions actually imply that the data are clustered by groups where labels are all equal.
When this class separation hypothesis holds true, one can leverage this structure by considering a weighting scheme based on nearest neighbors.
As we discussed in more details in section \ref{df:sec:stat}, this endows the estimate $f_n$ with exponential convergence rates, meaning that the risk ${\cal R}(f_n)$ converges toward the minimum risk exponentially fast as the number of samples $n$ goes to infinity.
This is the statement of Theorem \ref{df:thm:convergence}.
\subsection{Can collaborative filtering help us to deal with weak supervision?}
There is a natural link between completion of weak information and the collaborative filtering problem described in section \ref{int:sec:coll_filt}.
To make this link we will need to introduce more concepts.
Assume that the loss $\ell:\mathcal{Y}\to\mathcal{Y}\to\mathbb{R}$ can be written as $\ell(z, y) = -\scap{\psi(z)}{\phi(y)}$ with $\psi,\phi:\mathcal{Y}\to{\cal H}$ two embeddings into a Hilbert space ${\cal H}$, encoding the structure of the loss.
For example, the Kendall loss in ranking corresponds to the correlation measure $\phi = -\psi = (\ind{\sigma(i) > \sigma(j)})_{i,j\leq m}$ for $\sigma \in \mathcal{Y} = \mathfrak{S}_m$ a permutation.
We refer to \cite{Nowak2019} for more examples.
Assume also that the weak supervision correspond to observing $A_i\phi(Y_i)$ instead of $\phi(Y_i)$ for a known masking operator $A_i:{\cal H}\to{\cal H}$.
For example, observing a partial ordering that $\sigma$ should verify is equivalent to observing some coordinates of the vector $\phi(\sigma)$.
The disambiguation problem can be reduced to the matrix completion problem of retrieving $(\phi(y_i))_{i\leq n}$ from the observations $(A_i\phi(Y_i))_{i\leq n}$.
Our algorithm suggests solving the minimization problem
\[
\minimize{\norm{
\paren{\begin{array}{ccc}
\vert & & \vert \\
\psi(z_1) & \cdots & \psi(z_n) \\
\vert & & \vert \\
\end{array}}
\paren{\begin{array}{ccc}
& & \\
& \alpha_i(X_j) & \\
& & \\
\end{array} }
\paren{\begin{array}{ccc}
\vert & & \vert \\
\phi(y_1) & \cdots & \phi(y_n) \\
\vert & & \vert \\
\end{array}}^\top
}_F^2}{A_i\phi(y_i) = A_i\phi(Y_i).}
\]
In the case where $\psi = -\phi$ and there is no context variables, hence $\alpha_i(X_j) = 1$, the slightly different measure of determinism exposed in section \ref{df:app:diffrac} leads to the problem
\[
\minimize{-\norm{ \sum_{i=1}^n \phi(y_i)}^2}
{A_i\phi(y_i) = A_i\phi(Y_i).}
\]
For correlation losses, it is usual for $\norm{\phi(\cdot)}$ to be constant, so to minimize the last objective, one should try to align all $\phi(y_i)$ in one direction.
This differs from the collaborative filtering solution which completes the $(\phi(y_i))$ by minimizing the cardinality of the set $\brace{\phi(y_i)\,\middle\vert\, i\in[N]}$.
Formally, collaborative aims at solving the following problem
\[
\minimize{\operatorname{rank}
\paren{\begin{array}{ccc}
\vert & & \vert \\
\phi(y_1) & \cdots & \phi(y_n) \\
\vert & & \vert \\
\end{array}}
}{A_i\phi(y_i) = A_i\phi(Y_i),}
\]
which is practically implemented with the nuclear norm instead of the rank.
In essence, the infimum loss builds an understanding of $\paren{Y\,\middle\vert\, X}$ by collapsing all observations onto a single vector in $\phi(\prob{\mathcal{Y}})$, while collaborative filtering builds such an understanding by minimizing the number of vectors in $\phi(\mathcal{Y})$ to complete observations without discrepancy.
Bining observations into several groups makes sense when context variables do not allow for inputs to characterize outputs univocally.
Further investigations of the link between our work and collaborative filtering is a research direction that we left open.
We conjecture that this link might prove useful to incorporate context variables into collaborative filtering techniques.
\section{Leveraging input structure}
In this section, we discuss unsupervised learning techniques that could be incorporated in the frameworks described previously in order to leverage unlabeled input data.
\subsection{The case of semi-supervised learning}
The infimum loss and the disambiguation framework described previously are based on a pointwise principle that discriminates a distribution $\paren{Y\,\middle\vert\, X}$ from a weakly supervised distribution $\paren{S\,\middle\vert\, X}$.
Such a principle fails to provide any guideline for semi-supervised learning, which is a specific instance of partial supervision where $S$ is either a singleton or the full set $\mathcal{Y}$.
In particular for unsupervised parts of the input space, which correspond to points in the support of $\paren{X\,\middle\vert\, S=\mathcal{Y}}$ that do not belong to the support of $\paren{X\,\middle\vert\, \card S = 1}$, this principle does not provide any information on what $f^*(x)$ should be.
Eventually, the solution on those points could be inferred from the solution on the other points by looking for the simplest function completion with respect to some criterion quantifying how ``simple'' a function is.
For regression problems, if one searches to minimize the empirical risk in a Banach space of function, the norm of this Banach space provides a simple criterion for ``simplicity''.
More generally, in metric space of function, this norm can be replaced more generally as the distance between a neutral function ({\em e.g.} the null function in a vector space) and the function itself.
\subsection{Laplacian regularization}
The classical approach to solve semi-supervised learning regression problem is to choose the criterion of ``simplicity'' as the Dirichlet energy $f\to \E_{\rho_\mathcal{X}}[\norm{\nabla f(X)}^2]$.
This regularization is similar to the square norm of the Hilbert space of functions $W^{1,2}(\rho_\mathcal{X})$, which is the weighted Sobolev space of functions with weak derivative endowed with the probability measure $\rho_\mathcal{X}$.
Considering a square of the norm rather than the norm itself is classical in machine learning, making analysis as well as computations easier.
Measuring regularity through the probability measure $\rho_\mathcal{X}$ rather than the Lebesgue measure is crucial in order to leverage the unsupervised data.
This regularization has been introduced in the seminal paper of \cite{Zhu2003} which estimates the Dirichlet energy through finite difference methods akin to the Nadayara-Watson estimator.
This regularization formalized many intuitions.
For regression problems, it is natural to assume that the target function does not vary too much on densely populated input regions, or said otherwise, that its variations concentrate on regions without too much data.
For classification problems, this echoes the low-density separation hypothesis, assuming that decision frontiers, that is frontier between classes in the input space, lies on sparsely populated regions.
In \cite{Cabannes2021c}, we proposed a ``kernelized'' version of this estimation in order to bypass the curse of dimensionality under regularity assumptions, together with some clever low-rank factorization in order to reduce computation time to a decent amount.
It should be noted that Laplacian regularization has also a natural interpretation in terms of diffusion, which is captured by the Langevin dynamic, and explains the wording ``label propagation'' found in semi-supervised literature.
\subsection{Complete framework}
With Laplacian regularization, one can deepen the {\em lex parsimoniae} of the precedent sections when this rule does not discriminate a unique distribution by looking for the distribution that will minimize the Dirichlet energy of its (or a continuous surrogate) risk minimizer.
In particular, this approach is interesting as it could remove the usual non-ambiguity assumption usually made to define and study solutions of the partial supervision problems \citep{Cour2011,Luo2010,Liu2014,Cabannes2020,Cabannes2021}.
Yet, removing such an assumption will come at the price of adding assumptions on the surrogate problem we are solving that might be harder to verify in practice.
Finally, it should be noted that Laplacian regularization can be introduced as soon as during the preprocessing/feature engineering part of a practical machine learning pipeline.
In particular, it provides principled guidelines to retrieve and parametrize a small manifold on which the input data might lie, and might prove itself useful to strengthen self-supervision techniques.
At this point, it is less a theory than experiments that could confirm those intuitions.
\section{Active labeling}
Weakly supervised learning is concerned with retrieving a target function under weak observations.
It is often motivated by the bottleneck of annotating data.
Yet it fails to answer the most simple question practitioners might ask themselves: how to collect the most discriminative dataset to learn this function under some cost constraints.
Hence, the last part of this thesis is devoted to the ``active'' collection of weak supervision.
Active refers to the fact that we will iteratively and adaptively search for annotations, in contrast with passive settings in which annotations are given once and for all by an exterior mechanism.
The cost of dataset annotation depends on different task specifications.
Hence, a model of annotation cost can only describe a limited number of situations.
As well as we split our understanding of weakly supervised learning in three parts (group statistics, weak classifiers, and labels corruption), we could split dataset annotation with the same three categories.
In practice, it is frequent to annotate dataset by grouping input data according to a classifier output before identifying the most present class in each group and filtering mistakes, which allows annotating large chunks of data at once.
In a similar fashion, ImageNet was collected by searching for different categories on image search engines, before spotting outliers in a batch of images resulting from a single query \citep{ImageNet}.
The final part of this thesis touches upon the active variant of partially supervised learning \citep{Cabannes2022}.
In particular, we assume that given data $(X_i)_{i\leq n}$, one can query any information of the type $\ind{Y_{i_t}\in S_t}$ for $i_t \in [n]$ an index to choose and $S_t\in {\cal S}\subset 2^\mathcal{Y}$ a set to choose among certain subsets of labels.
We refer the reader to Part \ref{part:collection} for additional considerations on the matter.
\section{Consistency analysis}
\label{lap:app:consistency}
This section is devoted to the proof of Theorem \ref{lap:thm:consistency}.
The proof is based on~\eqref{lap:eq:filtering} and \eqref{lap:eq:filtering_zero}, and splits the error of $\norm{g_\rho - \hat{g}_p}_{L^2}^2$ into several components linked with how well we approximate $S^\star g_\rho$, and how well we approximate the eigenvalue decomposition $(\lambda_i, \theta_{i})$ of $(\Sigma, L)$.
\subsection{Sketch and understanding of the proof}
In this subsection, we explain how the proofs work for consistency theorems such as Theorem \ref{lap:thm:consistency}.
Let us define the mapping $G:{\cal H}\times {\cal C}\to L^2$ with $\cal C$ the set of pairs of self-adjoint operators on ${\cal H}$ that admit a generalized eigenvalue decomposition, as
\begin{equation}
G(\theta, (A, B)) = \sum_{i\in\mathbb{N}} \psi(\lambda_i) \scap{\theta}{\theta_i} S\theta_i
\qquad\text{with}\qquad (\lambda_i, \theta_i) \in \operatorname{GEVD}(A, B).
\end{equation}
$G(\theta, (A, B)) \in L^2$ corresponds to writing $\theta\in{\cal H}$ in the basis associated with the generalized eigenvalue decomposition (GEVD) of $(A, B)$.
\begin{proposition}
Under Assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation}, and with $\psi$ defined in Theorem \ref{lap:thm:consistency}
\[
g_\lambda = G(S^\star g_\rho, (\Sigma, L)), \qquad\text{and}\qquad
\hat{g}_p = G(\widehat{S^\star g_\rho}, (P\hat\Sigma P, P\hat LP + \mu P)),
\]
with $P$ the projection matrix from ${\cal H}$ to $\Span\brace{k_{X_i}}_{i\leq p}$.
\end{proposition}
\begin{proof}
This is a direct application of Assumptions~\ref{lap:ass:source}, \ref{lap:ass:approximation}, \eqref{lap:eq:filtering_zero} and Algorithm~\ref{lap:alg:imp}.
\end{proof}
The main point of the proof is to relate $g_\rho$ to $\hat{g}_p$.
To do so, we will use several functions in $L^2$ generated by $G$.
We detail our steps in Table \ref{lap:tab:step}.
Table \ref{lap:tab:step} gives a first answer to the two questions asked in the opening of Section \ref{lap:sec:statistics}.
The number of unlabeled data controls the convergence of the operators $(P\hat\Sigma P, P\hat LP+\mu P)$ toward $(\Sigma, L + \mu)$.
The number of labeled data controls the convergence of the vector $\widehat{S^\star g_\rho}$ toward $S^\star g_\rho$.
Priors on the structure of the problem, such as source and approximation conditions, control the convergence of the bias estimate $g_{\lambda, \mu}$ toward $g_\rho$.
Furthermore, a more precise study reveals that the concentration of operators are related to efficient dimension \citep{Caponnetto2007} and are accelerated by capacity assumptions on the functional space whose norm is $\norm{g} = \norm{({\cal L} + K^{-1})^{1/2}g}_{L^2}$, and that the concentration of the vector $\widehat{S^\star g_\rho}$ is accelerated by assumptions on moments of the variable $Y(I + \lambda {\cal L})^{-1}\delta_X$ (inheriting randomness from $(X, Y) \sim \rho$).
\begin{table}[h]
\caption{Steps in the consistency analysis}
\label{lap:tab:step}
\vskip 0.1in
\centering
\begin{tabular}{cccc}
\toprule
Estimate & Vector & Property of convergence & Basis \\
\midrule
& & & \\
$\hat{g}_p$ & $\widehat{S^\star g_\rho}$ & & $(P\hat\Sigma P, P\hat LP+\mu P)$ \\
& & & \\
& & Low-rank approx. \citep{Rudi2015} & $\downarrow$ \\
& & & \\
$\hat{g}$ & $\widehat{S^\star g_\rho}$ & & $(\hat\Sigma , \hat L+\mu )$ \\
& & & \\
& & Operator concentration \citep{Minsker2017} & $\downarrow$ \\
& & & \\
$g_{n_\ell}$ & $\widehat{S^\star g_\rho}$ & & $(\Sigma, L +\mu)$ \\
& & & \\
& & Vector concentration \citep{Yurinskii1970} & $\downarrow$ \\
& & & \\
$g_{\lambda, \mu}$ & $S^\star g_\rho$ & & $(\Sigma, L +\mu)$ \\
& & & \\
& & Source condition \citep{Lin2020} & $\downarrow$ \\
& & & \\
$g_{\lambda}$ & $S^\star g_\rho$ & & $(\Sigma,L)$ \\
& & & \\
$\downarrow$ & & Source condition \citep{Caponnetto2007} & \\
& & & \\
$g_\rho$ & & & \\
\bottomrule
\end{tabular}
\end{table}
\paragraph{Control of biases.}
We begin our study in a downward fashion regarding Table \ref{lap:tab:step}.
Indeed, for Tikhonov regularization~\eqref{lap:eq:laplacian_tikhonov}, we can show that for $q \in [0, 1]$,
\[
\norm{g_\lambda - g_\rho}_{L^2}
\leq \lambda^q \norm{{\cal L}^q g_\rho}_{L^2}.
\]
Meaning that if we have the source condition $g_\rho \in \ima {\cal L}^q$ (which is a condition on how fast $(\scap{g_\rho}{e_i})_{i\in\mathbb{N}}$ decreases compared to $(\lambda_i)_{i\in\mathbb{N}}$ for $(\lambda_i, e_i)$ the eigenvalue decomposition of ${\cal L}^{-1}$), the rates of convergence of this term when $n$ goes to infinity is controlled by the regularization scheme $\lambda_n^q$.
Similarly to the kernel-free bias above, for $(q_i) \in (0, 1)^\mathbb{N}$, we can have
\[
\norm{g_{\lambda, \mu} - g_\lambda}_{L^2}^2
\leq 2 \sum_{i\in\mathbb{N}} \lambda^{2q_i} \mu^{2q_i} \paren{\frac{\lambda_i}{\lambda+\lambda_i}}^2\abs{\scap{e_i}{g_\rho}}^2 \norm{K^{-q_i}e_i}_{L^2}^2.
\]
This shows explicitly the usefulness of controlling at the same time how $g_\rho$ is supported on the eigenspaces of ${\cal L}$ and how the eigenvectors are well approximated by the RKHS ${\cal H}$, which can be read in the value of $(q_i)$ such that all $e_i \in \ima K^{q_i}$.
\paragraph{Vector concentration.}
Let us now switch to concentration of $\widehat{S^\star g_\rho} = n_\ell^{-1}\sum_{i=1}^{n_\ell} Y_i k_{X_i}$ toward $S^\star g_\rho = \E_{(X, Y)\sim \rho}\bracket{Yk_X}$, it will allow controlling $\norm{g_{n_\ell} - g_{\lambda, \mu}}_{L^2}^2$ with the notations appearing in Table~\ref{lap:tab:step}.
Note how this convergence should be measured in terms of the reconstruction error
\[
\norm{\sum_{i\in\mathbb{N}} \psi(\lambda_{i,\mu})
\scap{S^\star g_\rho - \widehat{S^\star g_\rho}}{\theta_{i,\mu}}
S\theta_{i,\mu}}_{L^2}.
\]
This error might behave in a much better fashion than the $L^2$ error between $S S^\star g_\rho$ and $S\widehat{S^\star g_\rho}$.
In particular, on Figure \ref{lap:fig:intro}, we can consider $\psi(\lambda_{i,\mu})= 0$ for $i > 4$, and we might have $\scap{Yk_{X}}{\theta_{i,\mu}} = Y\ind{x\in C_i}$, for $i \leq 4$ and $(X, Y) \in \mathcal{X}\times\mathcal{Y}$, where $C_i$ is the $i$-th innermost circle.
In this setting, when all four $\paren{Y\,\middle\vert\, X\in C_i}$ are deterministic, we only need one labeled point per circle to clear the reconstruction error.
Based on concentration results in Hilbert space, when $\abs{Y}$ is bounded by a constant $c_\mathcal{Y}$, and $x\to k(x, x)$ by a constant $\kappa^2$, we have, with ${\cal D}_{n_\ell}\sim\rho^{\otimes {n_\ell}}$ the dataset generating the labeled data
\[
\E_{{\cal D}_{n_\ell}}\bracket{\norm{g_{n_\ell} - g_{\lambda, \mu}}^2_{L^2}}
\leq 2 \sigma_\ell^2 (\mu\lambda n_\ell)^{-1} + \frac{4}{9} c_\mathcal{Y}^2 \kappa^2 (\mu\lambda n_\ell^2)^{-1}.
\]
where $\sigma_\ell^2 \leq c_\mathcal{Y}^2 \trace(\Sigma)$ is a variance parameter to relate to the variance of $Y(I + \lambda {\cal L})^{-1} \delta_X$ (where the randomness is inherited from $(X, Y)\sim\rho$).
The fact that the need for labeled data depends on the variance of $(X, Y)$ after being diffused through ${\cal L}$ is coherent with the results obtained by \citet{Lelarge2019} in the specific case of a mixture of two Gaussian.
\paragraph{Basis concentration.}
We are left with the comparison of $g_{n_\ell}$, which is the filtering of $\widehat{S^\star g_\rho}$ with the operators $(\Sigma, L + \mu)$, and $\hat{g}_p$, which is the filtering of the same vector with the operators $(P\hat\Sigma P, P \hat{L} P + \mu P)$.
As the number of samples grows toward infinity, we know that $(P\hat\Sigma P, P \hat{L} P + \mu P)$ will converge in operator norm toward $(\Sigma, L + \mu)$.
Yet, how to leverage this property to quantify the convergence of $\hat{g}_p$ toward $g_{n_\ell}$?
Let us write $(\lambda_i, \theta_i) = \operatorname{GEVD}(\Sigma, L + \mu)$, and $(\lambda_i', \theta_i') = \operatorname{GEVD}(P\hat\Sigma P, P \hat{L} P + \mu P)$, we have
\[
\norm{\hat{g}_p - g_{n_\ell}}_{L^2} = \norm{\sum_{i\in\mathbb{N}}
\psi(\lambda_i)\scap{\theta_i}{\hat\theta_\rho} S\theta_i -
\psi(\lambda_i')\scap{\theta_i'}{\hat\theta_\rho} S\theta_i'}
\qquad\text{with}\qquad \hat\theta_\rho = \widehat{S^\star g_\rho}.
\]
The generic study of this quantity requires controlling eigenspaces one by one.
Note that we expect convergence of eigenspaces to depend on gaps between eigenvalues.
However, when considering Tikhonov regularization, this quantity can be written under a simpler form.
In particular, the concentration of operators is controlled, up to few leftovers, through the quantity
$\norm{(\Sigma + \lambda L + \mu\lambda)^{-1/2}((\Sigma - \hat\Sigma) +
\lambda(L - \hat{L}))(\Sigma + \lambda L + \mu\lambda)^{-1/2}}_{\op}$
where $\norm{\cdot}_{\op}$ designs the operator norm.
In this setting, the low-rank approximation is controlled through $\norm{(\Sigma + \lambda L)^{1/2}(I - P)}_{\op}$, and when $L \preceq c\Sigma^\alpha$, this term can be controlled by $\norm{\Sigma^{1/2} (I - P)}_{\op} + \lambda^{1/2}\norm{\Sigma^{1/2}(I-P)}^\alpha_{\op}$ which can be controlled based on the work of \citet{Rudi2015}.
\subsection{Risk decomposition}
In this subsection, we decompose the risk appearing in Theorem \ref{lap:thm:consistency}.
\subsubsection{Control of biases}
We begin by splitting the error $\norm{g_\rho - \hat{g}_p}_{L^2}$ between a bias term due to the regularization parameters and a variance term due to the data.
With the notation of Table \ref{lap:tab:step},
\begin{equation}
\norm{g_\rho - \hat{g}_p}_{L^2} \leq \norm{g_\rho - g_\lambda}_{L^2} +
\norm{g_\lambda - g_{\lambda, \mu}}_{L^2} + \norm{g_{\lambda, \mu} - \hat{g}_p}_{L^2}.
\end{equation}
We will control the first two terms here, and the last term in the following subsections.
\begin{proposition}[Bias in $\lambda$]
Under Assumption \ref{lap:ass:source}
\begin{equation}
\norm{g_\lambda - g_\rho}_{L^2} \leq \lambda \norm{{\cal L}g_\rho}_{L^2}.
\end{equation}
\end{proposition}
\begin{proof}
Based on the definition of $g_\lambda = (I + \lambda {\cal L})^{-1}g_\rho$, we have
\[
g_\lambda - g_\rho = ((I + \lambda {\cal L})^{-1} - I)g_\rho
= -\lambda {\cal L} g_\rho.
\]
Because $g_\rho$ is supported on the first eigenvectors of the Laplacian (Assumption \ref{lap:ass:source}), we have $g_\rho = \sum_{i=1}^r \scap{g_\rho}{e_i} e_i$, with $e_i$ the eigenvector of ${\cal L}$ appearing in~\eqref{lap:eq:evd}, and
\[
\norm{{\cal L}g_\rho}_{L^2}^2 = \norm{\sum_{i=1}^r \lambda_i^{-1}
\scap{g_\rho}{e_i} e_i}_{L^2}^2
= \sum_{i=1}^r \lambda_i^{-2} \scap{g_\rho}{e_i}^2 \leq
\lambda_r^{-2} \norm{g_\rho}_{L^2}^2 < +\infty.
\]
This ends the proof of this proposition.
\end{proof}
\begin{proposition}[Bias in $\mu$]
Under Assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation}, we have
\begin{equation}
\norm{g_{\lambda, \mu} - g_\lambda}_{L^2}^2
\leq \lambda\mu c_a^2\norm{g_\rho}^2_{L^2},\qquad\text{with}\qquad
c_a^2 = \sum_{i=1}^r \norm{K^{-1/2}e_i}_{L^2}^2 = \sum_{i=1}^r \norm{e_i}_{\cal H}^2.
\end{equation}
\end{proposition}
\begin{proof}
Before diving into the proof, recall that the RKHS norm penalization can be written as $\norm{g}_{\cal H} = \norm{K^{-1/2}g}_{L^2}$.
Using the fact that $A^{-1} - B^{-1} = A^{-1} (B - A) B^{-1}$, we have
\begin{align*}
g_{\lambda, \mu} - g_\lambda & =
((I + \lambda {\cal L} + \mu\lambda K^{-1})^{-1} - (I + \lambda {\cal L})^{-1}) g_\rho
\\&= - (I + \lambda {\cal L} + \mu\lambda K^{-1})^{-1} \lambda\mu K^{-1}(I + \lambda {\cal L})^{-1} g_\rho
\\&= - (\lambda\mu)^{1/2} (I + \lambda {\cal L} + \mu\lambda K^{-1})^{-1/2} (I + \lambda {\cal L} + \mu\lambda K^{-1})^{-1/2} \\&
\qquad\cdots\times (\lambda\mu K^{-1})^{1/2} K^{-1/2}(I + \lambda {\cal L})^{-1} g_\rho.
\end{align*}
As a consequence,
\begin{align*}
\norm{g_{\lambda, \mu} - g_\lambda}^2_{L^2}
& \leq \lambda \mu \norm{K^{-1/2}(I + \lambda {\cal L})^{-1} g_\rho}^2_{L^2},
\end{align*}
where we used the fact that $I + \lambda {\cal L} + \mu\lambda K^{-1} \succeq I$, so that $\norm{(I + \lambda {\cal L} + \mu\lambda K^{-1})^{-1/2}}_{\op} \leq 1$ (with $\norm{\cdot}_{\op}$ the operator norm), and that
\begin{align*}
\norm{(I + \lambda {\cal L} + \mu\lambda K^{-1})^{-1/2} (\lambda\mu K^{-1})^{1/2}}_{\op}^2
& =
\lambda\mu\norm{K^{-1/2}(I + \lambda {\cal L} + \mu\lambda K^{-1})^{-1} K^{-1/2}}_{\op}
\\&=
\lambda\mu\norm{(K + \lambda K^{1/2}{\cal L}K^{1/2} + \mu\lambda)^{-1}}_{\op}
\leq 1.
\end{align*}
We continue the proof with
\begin{align*}
\norm{K^{-1/2}(I + \lambda {\cal L})^{-1} g_\rho}
& = \norm{\sum_{i=1}^r \frac{\lambda_i}{\lambda + \lambda_i} \scap{g_\rho}{e_i} K^{-1/2}e_i}
\leq \sum_{i=1}^r \frac{\lambda_i}{\lambda + \lambda_i} \abs{\scap{g_\rho}{e_i}} \norm{K^{-1/2}e_i}
\\&\leq \sum_{i=1}^r \abs{\scap{g_\rho}{e_i}} \norm{K^{-1/2}e_i}_{L^2}
\leq \norm{g_\rho}_{L^2} \paren{\sum_{i\leq r} \norm{K^{-1/2}e_i}^2_{L^2}}^{1/2}.
\end{align*}
Putting all the pieces together ends the proof.
\end{proof}
\subsubsection{Vector concentration}
We are left with the study of the variance $\norm{\hat{g}_p - g_{\lambda,\mu}}$.
To ease derivations, we denote $C = \Sigma + \lambda L$, $\hat{C} = \hat\Sigma + \lambda \hat L$, $\theta_\rho = S^\star g_\rho$, $\hat\theta_\rho = \widehat{S^\star g_\rho}$ and $P$ the projection from ${\cal H}$ to $\Span\brace{k_{X_i}}_{i\leq p}$.
We have, for Tikhonov regularization
\begin{align*}
\norm{\hat{g}_p - g_{\lambda, \mu}}_{L^2}
& = \norm{S \paren{P(P\hat{C}P + \lambda\mu)^{-1} P \hat\theta_\rho - (C + \lambda\mu)^{-1}\theta_\rho}}_{L^2} \\
& = \norm{\Sigma^{1/2} \paren{P(P\hat{C}P + \lambda\mu)^{-1} P \hat\theta_\rho - (C + \lambda\mu)^{-1}\theta_\rho}}_{\cal H}.
\end{align*}
We begin by isolating the dependency to labeled data
\begin{equation}
\begin{split}
\norm{\hat{g}_p - g_{\lambda,\mu}}_{L^2}
&\leq \norm{\Sigma^{1/2} P(P\hat{C}P + \lambda\mu)^{-1} (\hat\theta_\rho - \theta_\rho)}_{\cal H}
\\&\qquad\cdots+ \norm{\Sigma^{1/2} \paren{P(P\hat{C}P + \lambda\mu)^{-1} \theta_\rho - (C+\lambda\mu)^{-1}\theta_\rho}}_{\cal H}.
\end{split}
\end{equation}
We will control the first term here, and the second term in the following subsection.
\begin{lemma}[Vector term]
When $\norm{(C + \lambda\mu)^{-1/2} (\hat{C} - C) (C+\lambda\mu)^{-1/2}}_{\op} \leq 1/2$, we have
\begin{equation}
\norm{\Sigma^{1/2} P(P\hat{C}P + \lambda\mu)^{-1} P (\hat\theta_\rho - \theta_\rho)}_{\cal H}
\leq 2 \norm{(C + \lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}.
\end{equation}
\end{lemma}
\begin{proof}
We begin with the splitting
\begin{align*}
\norm{\Sigma^{1/2} P(P\hat{C}P + \lambda\mu)^{-1} P (\hat\theta_\rho - \theta_\rho)}_{\cal H}
& \leq \norm{\Sigma^{1/2} P(P\hat{C}P + \lambda\mu)^{-1} P (C + \lambda\mu)^{1/2}}_{\op}
\\&\qquad\cdots \times\norm{(C + \lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}.
\end{align*}
The first term will concentrate toward a matrix smaller than identity, while the second term concentrates toward zero.
We can make those considerations more formal.
Following basic properties with the L\"owner order on operators, we have
\begin{align*}
& (C + \lambda\mu)^{-1/2} (C - \hat{C}) (C+\lambda\mu)^{-1/2} \preceq t
\\\Rightarrow\qquad
& \hat C \succeq (1 - t) C - t\lambda\mu
\\\Rightarrow\qquad
& P\hat CP \succeq (1 - t) PCP - t\lambda\mu P \succeq (1 - t) PCP - t\lambda\mu
\\\Rightarrow\qquad
& P\hat CP + \lambda\mu \succeq (1 - t) (PCP + \lambda\mu)
\\\Rightarrow\qquad
& (P\hat CP + \lambda\mu)^{-1} \preceq (1 - t)^{-1} (PCP + \lambda\mu)^{-1}
\\\Rightarrow\qquad
& (C+\lambda\mu)^{1/2}P(P\hat CP + \lambda\mu)^{-1}P(C+\lambda \mu)^{1/2}
\\&\qquad\preceq (1 - t)^{-1} (C+\lambda\mu)^{1/2}P(PCP + \lambda\mu)^{-1}P(C+\lambda\mu)^{1/2}
\preceq (1-t)^{-1},
\end{align*}
where we have used the fact that the last operator is a projection.
As a consequence, for any $t \in (0, 1)$, we have
\begin{align*}
& \norm{(C + \lambda\mu)^{-1/2} (\hat{C} - C) (C+\lambda\mu)^{-1/2}}_{\op} \leq t
\\\qquad\Rightarrow\qquad&
\norm{(C + \lambda \mu)^{1/2} P(P\hat{C}P + \lambda\mu)^{-1}(C+\lambda\mu)^{1/2}}_{\op}
\leq (1-t)^{-1}.
\\\qquad\Rightarrow\qquad&
\norm{\Sigma^{1/2} P(P\hat{C}P + \lambda\mu)^{-1}P (C+\lambda\mu)^{1/2}}_{\op}
\leq (1-t)^{-1}.
\end{align*}
Where the last implication follows from the fact that $C + \lambda \mu = \Sigma + \lambda L + \lambda\mu \succeq \Sigma$.
\end{proof}
\subsubsection{Basis concentration}
We are left with the study of the basis concentration with the number of unlabeled data.
\begin{lemma}[Basis term]
When $\norm{(C + \lambda\mu)^{-1/2} (\hat{C} - C) (C+\lambda\mu)^{-1/2}}_{\op} \leq 1/2$, we have
\begin{equation}
\begin{split}
&\norm{\Sigma^{1/2}(P(P\hat{C}P + \lambda\mu)^{-1} - (C + \lambda\mu)^{-1})\theta_\rho}_{\cal H}
\\&\qquad
\leq 3 \norm{C^{1/2} (I - P)}_{\op} \norm{g_\lambda}_{\cal H}
+ 2 \norm{(C + \lambda\mu)^{-1/2} (\hat C - C) (C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}.
\end{split}
\end{equation}
Notice that Assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation} imply $\norm{g_\lambda}_{\cal H} \leq c_a\norm{g_\rho}_{L^2} < +\infty$.
\end{lemma}
\begin{proof}
First of all, using that $A^{-1} - B^{-1} = A^{-1}(B - A)B^{-1}$, notice that
\begin{align*}
& \norm{\Sigma^{1/2}(P(P\hat{C}P + \lambda\mu)^{-1} - (C + \lambda\mu)^{-1})\theta_\rho}_{\cal H}
\\&
= \norm{\Sigma^{1/2}P(P\hat{C}P + \lambda\mu)^{-1}P(C - \hat{C}P )(C + \lambda\mu)^{-1}\theta_\rho
- \Sigma^{1/2}(I-P)(C + \lambda\mu)^{-1}\theta_\rho}_{\cal H}
\\&
\leq \norm{\Sigma^{1/2}P(P\hat{C}P + \lambda\mu)^{-1}P(\hat{C}P - C)(C + \lambda\mu)^{-1}\theta_\rho}_{\cal H}
+ \norm{\Sigma^{1/2}(I-P)(C + \lambda\mu)^{-1}\theta_\rho}_{\cal H}
\\&
\leq \norm{\Sigma^{1/2}P(P\hat{C}P + \lambda\mu)^{-1}P (C+\lambda\mu)^{1/2}}_{\op}
\norm{(C+\lambda\mu)^{-1/2}P(\hat{C}P - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}
\\&
\qquad\cdots+ \norm{\Sigma^{1/2}(I-P)}_{\op}\norm{(C + \lambda\mu)^{-1}\theta_\rho}_{\cal H}.
\end{align*}
Because $\Sigma \preceq \Sigma + \lambda L = C$, we have $\norm{\Sigma^{1/2}(I-P)}_{\op} \leq \norm{C^{1/2} (I-P)}_{\op}$, and we also have
\[
\norm{(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H} \leq
\norm{C^{-1}\theta_\rho}_{\cal H}
= \norm{K^{-1/2} S C^{-1}\theta_\rho}_{\cal H}
= \norm{K^{-1/2}g_\lambda}_{L^2} = \norm{g_\lambda}_{\cal H}.
\]
Recall, that, for any $t \in (0, 1)$, we have already shown that
\begin{align*}
& \norm{(C + \lambda\mu)^{-1/2} (\hat{C} - C) (C+\lambda\mu)^{-1/2}}_{\op} \leq t
\\\qquad\Rightarrow\qquad&
\norm{\Sigma^{1/2} P(P\hat{C}P + \lambda\mu)^{-1}P (C+\lambda\mu)^{1/2}}_{\op}
\leq (1-t)^{-1}.
\end{align*}
We are left with one last term to work on
\begin{align*}
\norm{(C+\lambda\mu)^{-1/2}P(\hat{C}P - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}
& \leq \norm{(C+\lambda\mu)^{-1/2}P(\hat{C} - C)P(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}
\\&\cdots+ \norm{(C+\lambda\mu)^{-1/2}C (I-P)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}.
\end{align*}
We control the first term with the fact for $A, B, C$ three self-adjoint operators and $x$ a vector we have
\[
\norm{APBPCx} = \norm{APBPC x\otimes x C PBPA}_{\op}^{1/2},
\]
and that
\begin{align*}
& PCx\otimes xCP \preceq Cx\otimes xC
\\\qquad\Rightarrow\qquad&
PBPC x\otimes xCPBP \preceq BPCx\otimes xCPB \preceq BCx\otimes xCB
\\\qquad\Rightarrow\qquad&
APBPC x\otimes xCPBPA \preceq ABCx\otimes xCBA,
\end{align*}
so that
\[
\norm{(C+\lambda\mu)^{-1/2}P(\hat{C} -
C)P(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}
\leq
\norm{(C+\lambda\mu)^{-1/2}(\hat{C} - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}.
\]
We control the second term with
\begin{align*}
& \norm{(C+\lambda\mu)^{-1/2}C (I-P)(C+\lambda\mu)^{-1}\theta_\rho}
\\&\qquad\leq \norm{(C+\lambda\mu)^{-1/2}C^{1/2}}\norm{C^{1/2}(I-P)}\norm{(C+\lambda\mu)^{-1}\theta_\rho}.
\end{align*}
Using that
\(
(C+\lambda\mu)^{-1/2}C^{1/2}\preceq I,
\)
we can add up everything to get the lemma.
For the part concerning $\norm{g_\lambda}_{\cal H}$, notice that
\begin{align*}
\norm{g_\lambda}_{\cal H} & = \norm{K^{-1/2}g_\lambda}_{L^2}
= \norm{ \sum_{i=1}^r \frac{\lambda_i}{\lambda_i +
\lambda}\scap{g_\rho}{e_i} K^{-1/2} e_i}_{L^2}
\leq \sum_{i=1}^r \abs{g_\rho}{e_i} \norm{K^{-1/2}e_i}
\\&\leq \norm{g_\rho}_{L^2} \paren{\sum_{i=1}^d \norm{K^{-1/2}e_i}^2_{L^2}}^{1/2}
= c_a \norm{g_\rho}_{L^2},
\end{align*}
with $c_a$ defined as before.
\end{proof}
\subsubsection{Conclusion}
Based on the last subsections, we have proved the following proposition.
\begin{proposition}[Risk decomposition]
When $\norm{(C+\lambda\mu)^{-1/2}(\hat{C} - C)(C+\lambda\mu)^{-1/2}} \leq 1/2$, Under the assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation},
\begin{equation}
\label{lap:eq:risk_dec}
\begin{split}
&\norm{\hat{g}_p - g_\rho}_{L^2}^2 \leq
4\lambda^2 \norm{{\cal L}g_\rho}_{L^2}^2
+ 4\lambda\mu c_a^2 \norm{g_\rho}_{L^2}^2
+ 8 \norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}^2
\\&\qquad\cdots+ 12c_a^2\norm{C^{1/2}(I - P)}_{\op}^2\norm{g_\rho}_{L^2}^2
+ 8\norm{(C+\lambda\mu)^{-1/2} (\hat{C} - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}^2.
\end{split}
\end{equation}
\end{proposition}
We are left with the quantification of the different convergences when the number of labeled and unlabeled data grows toward infinity.
We will quantify those convergences based on concentration inequalities.
\subsection{Probabilistic inequalities}
In this subsection, we bound each term appearing in~\eqref{lap:eq:risk_dec} based on concentration inequalities.
\subsubsection{Vector concentration}
The concentration of $\hat\theta_\rho = \widehat{S^\star g_\rho}$ toward $\theta_\rho = S^\star g_\rho$ is controlled through Bernstein inequality.
\begin{theorem}[Concentration in Hilbert space \citep{Yurinskii1970}]
\label{lap:thm:bernstein-vector}
Let denote by ${\cal A}$ a Hilbert space and by $(\xi_{i})$ a sequence of independent random vectors in ${\cal A}$ such that $\E[\xi_{i}] = 0$, that are bounded by a constant $M$, with finite variance $\sigma^2 = \E[\sum_{i=1}^{n}\norm{\xi_{i}}^2]$.
For any $t>0$,
\[
\Pbb(
\norm{\sum_{i=1}^{n} \xi_{i}} \geq t) \leq 2\exp\paren{-\frac{t^2}{2\sigma^2 +
2tM / 3}}.
\]
\end{theorem}
\begin{proposition}[Vector concentration]
When $\abs{Y}$ is bounded by a constant $c_\mathcal{Y}$, and $x\to k(x, x)$ by a constant $\kappa^2$, we have, with ${\cal D}_{n_\ell}\sim\rho^{\otimes {n_\ell}}$ the dataset generating the labeled data
\begin{equation}
\Pbb_{{\cal D}_{n_\ell}}\paren{\norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H} \geq t}
\leq
2\exp\paren{-\frac{n_\ell t^2}{2\sigma_\ell^2(\mu\lambda)^{-1} + 2t c_\mathcal{Y} (\mu\lambda)^{-1/2}\kappa / 3}},
\end{equation}
where $\sigma_\ell^2 \leq c_\mathcal{Y}^2 \trace(\Sigma)$ is a variance parameter to relate with the variance of $Y(I + \lambda {\cal L})^{-1} \delta_X$ (where the randomness is inherited from $(X, Y)\sim\rho$).
\end{proposition}
\begin{proof}
Recall that
\[
(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho) = (\Sigma + \lambda L +\lambda\mu)^{-1/2} (n_\ell^{-1}\sum_{i=1}^{n_\ell}Y_i k_{X_i} - \E_{\rho}[Yk_X])
\]
We want to apply Bernstein inequality to the vector $\xi_{i} = (\Sigma + \lambda L + \mu \lambda)^{-1/2}Y_ik_{X_i}$, after centering it.
Let us denote by $c_\mathcal{Y}$ a bound on $\abs{Y}$, $c_\mathcal{Y} \in \mathbb{R}$ exists since we have supposed $\rho$ of compact support.
We have
\begin{align*}
\sigma^2 & = \E[\sum_{i=1}^{n_\ell}\norm{\xi_{i} - \E[\xi_i]}^2]
= n_\ell \E[\norm{\xi - \E[\xi_i]}^2]
\leq n_\ell \E[\norm{\xi}^2]
\\&= n_\ell \E_{(X, Y)\sim\rho}\bracket{ Y^2 \scap{k_X}{(\Sigma + \lambda L + \mu \lambda)^{-1} k_{X}}}
\\&\leq n_\ell c_\mathcal{Y}^2 \E_{X\sim\rho_\mathcal{X}}\bracket{\scap{k_X}{(\Sigma + \lambda L + \mu \lambda)^{-1} k_{X}}}
\\&= n_\ell c_\mathcal{Y}^2 \trace\paren{(\Sigma + \lambda L + \mu \lambda)^{-1} \Sigma}
\\&\leq n_\ell c_\mathcal{Y}^2 \trace(\Sigma) \norm{(\Sigma + \lambda L + \mu \lambda)^{-1}}_{\op}
\leq n_\ell c_\mathcal{Y}^2 \trace(\Sigma) (\mu \lambda)^{-1}.
\end{align*}
Note that we have proceed with a generic upper bound, but we expect this variance, which is related to the variance of $Y(I + \lambda {\cal L} + \lambda \mu K^{-1})^{-1} \delta_X$ to be potentially much smaller -- if we remove the term in $P$ the vector concentration is the concentration of the vector
$S(S^*S + \lambda S^\star {\cal L}S +
\lambda\mu)^{-1}Yk_X \simeq
K^{1/2}(K + \lambda K^{1/2} {\cal L}K^{1/2} +
\lambda\mu)^{-1}K^{1/2}S^{-\star}Yk_X
= (I + \lambda L + \lambda\mu K^{-1})^{-1}
YS^{-\star}k_X \simeq (I + \lambda L + \lambda\mu K^{-1})^{-1}Y\delta_X$.
To capture this fact, we will write $\sigma^2 \leq n_\ell \sigma_\ell^2(\mu\lambda)^{-1}$, with $\sigma_\ell = c_\mathcal{Y} \trace(\Sigma)^{1/2}$ in our analysis, but potentially much smaller under refined hypothesis and in practice.
Similarly to the bound on the variance, we have
\[
\norm{\xi - \E[\xi]} \leq \norm{\xi}
= \norm{(\Sigma + \lambda L + \mu\lambda)^{-1/2} Y_i k_{X_i}}
\leq (\mu\lambda)^{-1/2} c_\mathcal{Y} \kappa,
\]
with $\kappa$ an upper bound on $k(x, x)^{1/2}$ for $x\in\supp\rho_\mathcal{X}$.
As a consequence, applying Bernstein concentration inequality, we get, for any $t>0$,
\[
\Pbb_{{\cal D}_{n_\ell}}(\norm{n_\ell^{-1}\sum_{i=1}^{n_\ell} \xi_{i}- E[\xi_i]} \geq t) \leq
2\exp\paren{-\frac{n_\ell t^2}{2\sigma_\ell^2(\mu\lambda)^{-1} + 2t c_\mathcal{Y} (\mu\lambda)^{-1/2}\kappa / 3}}.
\]
This ends the proof.
\end{proof}
\subsubsection{Operator concentration}
The convergence of $\hat{C}$ toward $C$ is controlled with Bernstein inequality for self-adjoint operators.
\begin{theorem}[Bernstein inequality for self-adjoint \citep{Minsker2017}]
Let ${\cal A}$ be a separable Hilbert space, and $(\xi_i)$ a sequence of independent random self-adjoint operators on ${\cal A}$.
Assume that $(\xi_i)$ are bounded by $M \in \mathbb{R}$, in the sense that almost everywhere $\norm{\xi}_{\op} < M$, and have a finite variance $\sigma^2 = \norm{\sum_{i=1}^n\E[\xi_i^2]}_{\op}$.
For any $t > 0$,
\begin{align*}
& \Pbb\paren{\norm{\sum_{i=1}^{n}(\xi_i - \E[\xi_i])}_{\op} > t}
\leq 2 \paren{1 + 6\ \frac{\sigma^2 + M t / 3}{t^2}} \frac{\trace\paren{\sum_{i=1}^n\E[\xi_i^2]}}{\norm{\sum_{i=1}^n\E[\xi_i^2]}_{\op}}
\exp\paren{-\frac{t^2}{2 \sigma^2 + 2 t M / 3}}.
\end{align*}
\end{theorem}
\begin{proposition}[Operator concentration]
When $x\to k(x, x)$ is bounded by $\kappa^2$, and $x \to \partial_{1,j}\partial_{2,j} k(x, x)$ is bounded by $\kappa_j^2$, we have
\begin{equation}
\begin{split}
&\Pbb_{{\cal D}_{n}}\paren{\norm{(C + \mu\lambda)^{-\frac{1}{2}}(C - \hat C)(C + \mu\lambda)^{-\frac{1}{2}}}_{\op} > 1/2}
\leq \paren{2 + 56\ \frac{\kappa^2 + \lambda\sum_{i=1}^d \kappa_j^2}{\lambda\mu n}}
\\&\qquad\cdots \times(1 + \lambda\mu \norm{C}_{\op}^{-1})
\frac{\kappa^2 + \lambda\sum_{j=1}^d \kappa_j^2}{\lambda\mu}
\exp\paren{-\frac{\lambda\mu n}{10\paren{\kappa^2 + \lambda\sum_{j=1}^d \kappa^2_j}}}.
\end{split}
\end{equation}
\end{proposition}
\begin{proof}
We want to apply the precedent concentration inequality to
\[
\xi_i = (\Sigma + \lambda L +\lambda\mu)^{-1/2}(k_{X_i}\otimes k_{X_i} +
\lambda\sum_{j=1}^d \partial_j k_{X_i}\otimes \partial_j k_{X_i})(\Sigma + \lambda L +\lambda\mu)^{-1/2},
\]
since we have, based on the fact that $C = \Sigma + \lambda L$ and that $\Sigma = \E[k_X\otimes k_X]$ and $L = \E[\sum_{j=1}^n\partial_j k_X \otimes \partial_j k_X]$,
\[
\norm{(C+\lambda\mu)^{-1/2}(\hat C - C)(C+\lambda\mu)^{-1/2}}_{\op}
= n^{-1}\norm{\sum_{i=1}^n \xi_i - \E[\xi_i]}_{\op}.
\]
We bound $\xi$ with
\begin{align*}
\norm{\xi}_{\op} & = \norm{(C + \mu\lambda)^{-\frac{1}{2}}
\paren{k_{X} \otimes k_{X} + \lambda\sum_{j=1}^d \partial_j k_X \otimes \partial_j k_X } (C + \mu\lambda)^{-\frac{1}{2}}}_{\op}
\\&\leq \trace\paren{(C + \mu\lambda)^{-\frac{1}{2}} \paren{k_X \otimes k_X + \lambda\sum_{j=1}^d \partial_j k_X \otimes \partial_j k_X } (C + \mu\lambda)^{-\frac{1}{2}}}
\\&= \trace\paren{(C + \mu\lambda)^{-\frac{1}{2}} k_X \otimes k_X (C + \mu\lambda)^{-\frac{1}{2}}}
\\&\qquad\cdots + \lambda \sum_{j=1}^d \trace\paren{(C + \mu\lambda)^{-\frac{1}{2}} \partial_j k_X \otimes \partial_j k_X (C + \mu\lambda)^{-\frac{1}{2}}}
\\&= \norm{(C + \mu\lambda)^{-\frac{1}{2}} k_{X}}_{\cal H}^2
+ \lambda\sum_{j=1}^d \norm{(C + \mu\lambda)^{-\frac{1}{2}} \partial_j k_X}_{\cal H}^2
\leq (\lambda\mu)^{-1} \paren{\kappa^2 + \lambda\sum_{j=1}^d \kappa^2_j}.
\end{align*}
With $\kappa^2$ an upper bound on the kernel $k$ and $\kappa_j^2$ an upper bound on $\partial_{1,j}\partial_{2,j} k$.
For the variance we have, using L\"owner order,
\begin{align*}
\E[\xi^2]
& \preceq \sup_{X\in\mathcal{X}} \norm{\xi(X)}_{\op} \E[\xi]
\preceq (\lambda\mu)^{-1}\paren{\kappa^2 + \lambda\sum_{j=1}^d \kappa^2_j} \E[\xi]
\\&= (\lambda\mu)^{-1}\paren{\kappa^2 + \lambda\sum_{j=1}^d \kappa^2_j} (C + \lambda)^{-1}C
\preceq (\lambda\mu)^{-1}\paren{\kappa^2 + \lambda\sum_{j=1}^d \kappa^2_j}.
\end{align*}
Therefore, we get for any $t>0$,
\begin{align*}
& \Pbb_{{\cal D}_{n}}\paren{\norm{(C + \mu\lambda)^{-\frac{1}{2}}(C - \hat C)(C + \mu\lambda)^{-\frac{1}{2}}}_{\op} > t}
\\&\qquad\leq 2\paren{1 + 6\ \frac{(\kappa^2 + \lambda \sum_{i=1}^d \kappa_j^2)(1 + t / 3)}{\lambda\mu n t^2}}\, \frac{\norm{C}_{\op} + \lambda\mu}{\norm{C}_{\op}}
\trace\paren{(C + \lambda)^{-1}C}
\\&\qquad\cdots\times\exp\paren{-\frac{nt^2}{2 (\lambda\mu)^{-1}\paren{\kappa^2 + \lambda\sum_{j=1}^d \kappa^2_j} (1 + t/3)}}.
\end{align*}
Remark that
\[
\trace\paren{(C + \mu\lambda)^{-1}C} \leq \norm{(C +
\mu\lambda)^{-1}}_{\op}\trace(C) \leq
(\lambda\mu)^{-1}(\kappa^2 + \lambda \sum_{j=1}^d \kappa_j^2).
\]
Taking $t = 1/2$ ends the lemma.
\end{proof}
\subsubsection{Basis concentration}
Similarly, we could control $\norm{(C+\lambda\mu)^{-1/2}(\hat C - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}$ by using concentration of self-adjoint, yet this will lead to laxer bounds, than using concentration on vectors.
\begin{proposition}[Basis concentration]
When $x\to k(x, x)$ is bounded by $\kappa^2$, $x \to \partial_{1,j}\partial_{2,j} k(x, x)$ is bounded by $\kappa_j^2$, with Assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation}, we have
\begin{equation}
\Pbb_{{\cal D}_{n}}\paren{\norm{(C + \mu\lambda)^{-\frac{1}{2}}(C - \hat C)(C + \mu\lambda)^{-1}\theta_\rho}_{\cal H} > t}
\leq 2\exp\paren{
- \frac{\mu\lambda nt^2}{2 c_1(c_1 + \lambda^{1/2}\mu^{1/2} t / 3)}
},
\end{equation}
with $c_1 = (\kappa^2 + \lambda\sum_{i=1}^d \kappa_j^2) c_a \norm{g_\rho}_{L^2}$.
\end{proposition}
\begin{proof}
We want to apply Bernstein concentration inequality to the vectors
\[
\xi_i = (C+\mu\lambda)^{-1/2} \paren{k_{X_i}\otimes k_{X_i} + \lambda \sum_{j=1}^d
\partial_j k_{X_i} \otimes \partial_j k_{X_i}} (C+\lambda\mu)^{-1}\theta_{\rho},
\]
since
\[
\norm{(C + \mu\lambda)^{-\frac{1}{2}}(C - \hat C)(C + \mu\lambda)^{-1}\theta_\rho}_{\cal H}
= n^{-1} \norm{\sum_{i=1}^n \xi_i - \E[\xi_i]}_{\cal H}.
\]
We bound $\xi$, reusing prior derivations, with
\begin{align*}
\norm{\xi_i}_{\cal H}
& = \norm{(C+\mu\lambda)^{-1/2} \paren{k_{X_i}\otimes k_{X_i} + \lambda \sum_{j=1}^d
\partial_j k_{X_i} \otimes \partial_j k_{X_i}} (C+\lambda\mu)^{-1}\theta_{\rho}}_{\cal H}
\\&\leq \norm{(C+\mu\lambda)^{-1/2}}_{\op} \norm{\paren{k_{X_i}\otimes k_{X_i} + \lambda \sum_{j=1}^d
\partial_j k_{X_i} \otimes \partial_j k_{X_i}}}_{\op} \norm{(C+\mu\lambda)^{-1}\theta_\rho}_{\cal H}.
\\&\leq (\mu\lambda)^{-1/2} (\kappa^2 + \lambda\sum_{i=1}^d \kappa_j^2) c_a \norm{g_\rho}_{L^2}.
\end{align*}
For the variance, we have, similarly to prior derivations,
\begin{align*}
\E[\norm{\xi}^2]
& \leq
\sup_{X\in\mathcal{X}}\norm{k_{X}\otimes k_{X} + \lambda \sum_{j=1}^d
\partial_j k_{X} \otimes \partial_j k_{X}}_{\op}\norm{(C+\lambda\mu)^{-1}\theta_{\rho}}^2
\\&\qquad\qquad\cdots \times \E\bracket{ \norm{(C+\mu\lambda)^{-1} \paren{k_{X}\otimes k_{X} + \lambda \sum_{j=1}^d
\partial_j k_{X_i} \otimes \partial_j k_{X}}_{\op}}}
\\&\leq
\paren{\kappa^2 + \lambda\sum_{i=1}^d \kappa_i^2} c_a^2 \norm{g_\rho}^2_{L^2}
\\&\qquad\qquad\cdots\E\bracket{\norm{(C+\mu\lambda)^{-1} k_{X}\otimes k_{X}}_{\op} + \lambda \sum_{j=1}^d
\norm{(C+\mu\lambda)^{-1}\partial_j k_{X_i} \otimes \partial_j k_{X}}_{\op}}
\\&=
\paren{\kappa^2 + \lambda\sum_{i=1}^d \kappa_i^2} c_a^2 \norm{g_\rho}^2_{L^2}
\trace\paren{(C+\mu\lambda)^{-1} C }
\\&\leq (\lambda\mu)^{-1}\paren{\kappa^2 + \lambda\sum_{i=1}^d \kappa_i^2}^2 c_a^2 \norm{g_\rho}^2_{L^2}.
\end{align*}
As a consequence, using Bernstein inequality,
\[
\Pbb\paren{n^{-1}\norm{\sum_{i=1}^n \xi_{i} - \E[\xi_i]} > t}
\leq 2\exp\paren{
- \frac{\mu\lambda nt^2}{2 c_1(c_1 + \lambda^{1/2}\mu^{1/2} t / 3)},
}
\]
with $c_1 = (\kappa^2 + \lambda\sum_{i=1}^d \kappa_j^2) c_a \norm{g_\rho}_{L^2}$.
Note that we have bounded naively the variable $\xi$ and its variance, but we have shown how appears $\sup_{X\in\mathcal{X}} \norm{(C+\lambda\mu)^{-1}k_X} + \lambda\sum_{i=1}^d \norm{(C+\lambda\mu)^{-1}\partial_i k_X}$ and $\trace((C+\lambda\mu)^{-1}C)$, which under interpolation and capacity assumptions could be controlled in a better fashion.
\end{proof}
\subsubsection{Low-rank approximation}
We now switch to Nystr\"om approximation.
\begin{proposition}[Low-rank approximation]
When $x\to k(x, x)$ is bounded by $\kappa^2$, for any $p \in \mathbb{N}$ and $t > 0$, we have
\begin{equation*}
\Pbb_{{\cal D}_p}\paren{\norm{(I-P)\Sigma^{1/2}}^2 > t}
\leq \paren{2 + \frac{116\kappa^2}{t p}} (2 + t \norm{\Sigma}^{-1}_{\op}) \frac{\kappa^2}{t}
\exp\paren{- \frac{p t }{10\kappa^2}},
\end{equation*}
\end{proposition}
\begin{proof}
Reusing Proposition 3 of \citet{Rudi2015}, for any $\gamma > 0$, we have, with $P$ the projection on $\Span\brace{k_{X_i}}_{i\leq p}$ and $\hat\Sigma = p^{-1}\sum_{i=1}^p k_{X_i} \otimes k_{X_i}$,
\[
\norm{(I-P)\Sigma^{1/2}}^2
\leq \gamma\norm{(\hat\Sigma + \gamma)^{-1/2}\Sigma^{1/2}}^2_{\op}
\leq \gamma \norm{\Sigma^{1/2} (\hat\Sigma + \gamma)^{-1} \Sigma^{1/2}}_{\op}.
\]
As a consequence, skipping derivations that can be retaken from our precedent proofs,
\begin{align*}
& \Pbb_{{\cal D}_p}\paren{\norm{(I-P)\Sigma^{1/2}}^2 > t}
\leq \inf_{\gamma > 0} \Pbb_{{\cal D}_p}
\paren{ \gamma \norm{\Sigma^{1/2} (\hat\Sigma + \gamma)^{-1} \Sigma^{1/2}}_{\op} > t}
\\&\qquad\leq \inf_{\gamma > 0} \Pbb_{{\cal D}_p}
\paren{\norm{(\Sigma + \gamma)^{-1/2}(\hat\Sigma - \Sigma)
(\Sigma + \gamma)^{-1/2}}_{\op} > (1 - \gamma t^{-1})}
\\&\qquad\leq \inf_{\gamma > 0}
\paren{2 + 56\ \frac{\kappa^2}{\gamma p}} (1 + \gamma \norm{\Sigma}^{-1}_{\op}) frac{\kappa^2}{\gamma}
\exp\paren{- \frac{p\gamma u^2}{2\kappa^2 (1 + u/ 3)}},
\end{align*}
with $u = (1 - \gamma t^{-1})$.
Taking $\gamma = t/2$, this term is simplified as
\[
\Pbb_{{\cal D}_p}\paren{\norm{(I-P)\Sigma^{1/2}}^2 > t}
\leq \paren{2 + 116\ \frac{\kappa^2}{t p}} (2 + t \norm{\Sigma}^{-1}_{\op}) \frac{\kappa^2}{t}
\exp\paren{- \frac{p t }{10\kappa^2}},
\]
which is the object of this proposition.
\end{proof}
\begin{lemma}
When $L \leq c_d \Sigma^a$, we have
\begin{equation}
\norm{(I-P)C^{1/2}}_{\op}^2 \leq \norm{(I-P)\Sigma^{1/2}}^2_{\op}
+ c_d\lambda\norm{(I-P)\Sigma^{1/2}}^{2a}_{\op}.
\end{equation}
\end{lemma}
\begin{proof}
This follows from the fact that
\begin{align*}
\norm{C^{1/2}(I - P)}^2_{\op} & = \norm{(I-P) C (I-P)}_{\op}
= \norm{(I-P) (\Sigma + \lambda L)(I-P)}_{\op}
\\&\leq \norm{(I-P)\Sigma(I-P)}_{\op}
+ \lambda \norm{(I-P)L(I-P)}_{\op}
\\&\leq \norm{(I-P)\Sigma(I-P)}_{\op}
+ \lambda c_d \norm{(I-P)\Sigma^a(I-P)}_{\op}
\\&= \norm{(I-P)\Sigma^{1/2}}_{\op}^2
+ \lambda c_d \norm{(I-P)\Sigma^{a/2}}_{\op}^2
\\&= \norm{(I-P)\Sigma^{1/2}}_{\op}^2
+ \lambda c_d \norm{(I-P)^{a}\Sigma^{a/2}}_{\op}^2
\\&\leq \norm{(I-P)\Sigma^{1/2}}_{\op}^2
+ \lambda c_d \norm{(I-P)\Sigma^{1/2}}_{\op}^{2a},
\end{align*}
where we used the fact that $(I - P)^a = (I-P)$ and that $\norm{A^s B^s} \leq \norm{AB}^s$ for $s \in [0, 1]$ and $A, B$ positive self-adjoint.
\end{proof}
\subsection{Averaged excess of risk - ending the proof}
Based on the precedent excess of risk decomposition, and precedent concentration inequalities, we have all the elements to derive convergence rates of our algorithm.
We will enunciate this convergence in terms of the averaged excess of risk of $\E_{{\cal D}_n}\bracket{\norm{\hat{g}_p - g_\rho}_{L^2}^2}$.
\begin{lemma}
Under Assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation},
\begin{equation}
\begin{split}
&\E_{{\cal D}_n}\bracket{\norm{\hat{g}_p - g_\rho}_{L^2}^2}
\leq 4c_\mathcal{Y}^2\Pbb\paren{\norm{(C+\lambda\mu)^{-1/2}(\hat{C} - C)(C+\lambda\mu)^{-1/2}} \leq 1/2}
\\&\qquad\cdots + 4\lambda^2 \norm{{\cal L}g_\rho}_{L^2}^2
+ 4\lambda\mu c_a^2 \norm{g_\rho}_{L^2}^2
\\&\qquad\cdots + 8 \E_{{\cal D}_n}\bracket{ \norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}^2}
+ 12c_a^2\norm{g_\rho}_{L^2}^2 \E_{{\cal D}_n}\bracket{ \norm{C^{1/2}(I - P)}_{\op}^2}
\\&\qquad\cdots+ 8\E_{{\cal D}_n}\bracket{\norm{(C+\lambda\mu)^{-1/2} (\hat{C} - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}^2}.
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
We proceed using the fact that $\E[X] = \E[X\,\vert\, ^cA]\Pbb(^cA) + \E[X\,\vert\, A]\Pbb(A) \leq \sup X \Pbb(^cA) + \E[X\, \vert\, A]\Pbb(A)$, with $A = \brace{{\cal D}_n \,\middle\vert\, \norm{(C+\lambda\mu)^{-1/2}(\hat{C} - C)(C+\lambda\mu)^{-1/2}} \leq 1/2}$,
\[
\E_{{\cal D}_n}\bracket{\norm{\hat{g}_p - g_\rho}_{L^2}^2}
\leq \sup_{{\cal D}_n} \norm{\hat{g}_p - g_\rho}^2\Pbb\paren{^c A}
+ \E_{{\cal D}_n}\bracket{\norm{\hat{g}_p - g_\rho}^2 \,\middle\vert\, A}\Pbb(A).
\]
When $Y$ is bounded by $c_\mathcal{Y}$, because $g_\rho$ is a convex combination of $Y$, we know that $\norm{g_\rho}_{L^2} \leq c_\mathcal{Y}$, as a consequence, we can clip $\hat{g}_p$ to $[-c_\mathcal{Y}, c_\mathcal{Y}]$, which will only improve the estimation of $g_\rho$, as a consequence, we can consider the clipping estimate for which we have
\(
\sup_{{\cal D}_n} \norm{\hat{g}_p - g_\rho}_{L^2}^2 \leq 4 c_\mathcal{Y}^2.
\)
Regarding the second part, we have already decomposed the risk under the event $A = \brace{{\cal D}_n \,\middle\vert\, \norm{(C+\lambda\mu)^{-1/2}(\hat{C} - C)(C+\lambda\mu)^{-1/2}} \leq 1/2}$.
As a consequence, we have
\begin{align*}
& \E_{{\cal D}_n}\bracket{\norm{\hat{g}_p - g_\rho}_{L^2}^2}
\leq 4c_\mathcal{Y}^2\Pbb(^cA)
+ 4\lambda^2 \norm{{\cal L}g_\rho}_{L^2}^2 \Pbb(A)
+ 4\lambda\mu c_a^2 \norm{g_\rho}_{L^2}^2 \Pbb(A)
\\&\cdots + 8 \E_{{\cal D}_n}\bracket{ \norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}^2 \,\middle\vert\, A}\Pbb(A)
\\&\cdots+ 12c_a^2\norm{g_\rho}_{L^2}^2 \E_{{\cal D}_n}\bracket{ \norm{C^{1/2}(I - P)}_{\op}^2 \,\middle\vert\, A}\Pbb(A)
\\&\cdots+ 8\E_{{\cal D}_n}\bracket{\norm{(C+\lambda\mu)^{-1/2} (\hat{C} - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}^2 \,\middle\vert\, A}\Pbb(A).
\end{align*}
To control the conditional expectation, we use that, when $X$ is positive
\[
\E\bracket{X\,\middle\vert\, A} P(A) = \E[X] - \E\bracket{X \,\middle\vert\, ^cA}\Pbb(^cA) \leq \E[X].
\]
This ends the proof.
\end{proof}
Based on deviation inequalities, we can control expectations based on the equality, for $X$ positive, $\E[X] = \int_0^{+\infty} \Pbb(X > t) \mathop{}\!\mathrm{d} t$.
\begin{lemma}
In the setting of the paper,
\begin{equation}
\E_{{\cal D}_n}\bracket{ \norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}^2}
\leq 8 \sigma_\ell^2 (n_\ell \mu\lambda)^{-1} +
8c_\mathcal{Y}^2\kappa^2(n_\ell^{2} \mu\lambda)^{-1}.
\end{equation}
\end{lemma}
\begin{proof}
First, recall that
\begin{align*}
\Pbb\paren{ \norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H} > t}
& \leq 2\exp\paren{-\frac{n_\ell t^2}{2\sigma_\ell^2 (\mu\lambda)^{-1} + 2 t c_\mathcal{Y}(\lambda\mu)^{-1/2} \kappa / 3}}
\\&
\leq 2\exp\paren{-\frac{n_\ell t^2}{2\max\paren{2\sigma_\ell^2 (\mu\lambda)^{-1}, 2 t c_\mathcal{Y}(\lambda\mu)^{-1/2} \kappa / 3}}}
\\&
\leq 2\exp\paren{-\frac{n_\ell \mu\lambda t^2}{4\sigma_\ell^2}}
+ 2\exp\paren{-\frac{3n_\ell \mu^{1/2}\lambda^{1/2} t}{4 c_\mathcal{Y} \kappa}}.
\end{align*}
As a consequence
\begin{align*}
& \E\bracket{ \norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}^2}
=
\int_0^{+\infty}\Pbb\paren{ \norm{(C+\lambda\mu)^{-1/2}(\hat\theta_\rho - \theta_\rho)}_{\cal H}^2 > t}\mathop{}\!\mathrm{d} t
\\&\qquad\leq 2 \int \exp\paren{-\frac{n_\ell \mu\lambda t}{4\sigma_\ell^2}}\mathop{}\!\mathrm{d} t
+ 2\int \exp\paren{-\frac{3n_\ell \mu^{1/2}\lambda^{1/2} t^{1/2}}{4 c_\mathcal{Y} \kappa}}\mathop{}\!\mathrm{d} t.
\\&\qquad = 8 \sigma_\ell^2 (n_\ell \mu\lambda)^{-1}
+ \frac{64c_\mathcal{Y}^2\kappa^2}{9}(n_\ell^{2} \mu\lambda)^{-1}.
\end{align*}
This is the result stated in the lemma.
\end{proof}
\begin{lemma}
In the setting of the paper,
\begin{equation}
\begin{split}
\E_{{\cal D}_n}\bracket{\norm{(C+\lambda\mu)^{-1/2} (\hat{C} - C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}^2}
&\leq 8(\kappa^2 + \lambda\partial \kappa^2)^2 c_a^2 \norm{g_\rho}_{L^2}^2
\\&\qquad\cdots\times\paren{(\mu\lambda n)^{-1} + (\mu \lambda n^2)^{-1}},
\end{split}
\end{equation}
with $\partial \kappa^2 = \sum_{i=1}^d \kappa_i^2$.
\end{lemma}
\begin{proof}
Let us denote by $A$ the quantity $\norm{(C+\lambda\mu)^{-1/2} (\hat{C} -
C)(C+\lambda\mu)^{-1}\theta_\rho}_{\cal H}$, and
$\partial \kappa^2 = \sum_{i=1}^d \kappa_j^2$.
Recall that
\begin{align*}
\Pbb\paren{A > t}
& \leq 2\exp\paren{-\frac{\mu\lambda n t^2}{2c_1(c_1 + \lambda^{1/2}\mu^{1/2} t/3)}}
\\&\leq 2\exp\paren{-\frac{\mu\lambda n t^2}{4c_1^2}}
+ 2\exp\paren{-\frac{3(\mu\lambda)^{1/2} n t}{4c_1}}.
\end{align*}
We conclude the proof similarly to the precedent lemma.
\end{proof}
\begin{lemma}
Under Assumption \ref{lap:ass:decay},
\begin{equation}
\begin{split}
\E_{{\cal D}_n}\bracket{\norm{C^{1/2}(I - P)}_{\op}^2}
&\leq
\paren{\frac{10\kappa^2\log(p)}{p} + \frac{10^a\kappa^{2a} c_d \lambda \log(p)^a}{p^a}}
\\&\cdots \times \paren{1 + \frac{2\kappa^2}{\norm{\Sigma}_{\op}\log(p)} \paren{1 +
\frac{6}{\log(p)}} \paren{\frac{1}{p} + \frac{1}{5\log(p)}}}.
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
Once again, this result comes from integration of the tail bound obtained on $\norm{C^{1/2}(I-P)}_{\op}^2$ through the one we have on $\norm{\Sigma^{1/2}(I - P)}_{\op}^2$ and the fact that $\norm{C^{1/2}(I - P)}_{\op}^2 \leq \norm{\Sigma^{1/2}(I - P)}^2_{\op} + c_d\lambda \norm{\Sigma^{1/2}(I - P)}^{2a}_{\op}$.
For any $a, b > 0$, we have
\begin{align*}
& \E_{{\cal D}_n}\bracket{\norm{\Sigma^{1/2}(I - P)}_{\op}^2}
= \int_0^\infty \Pbb_{{\cal D}_n}\paren{\norm{\Sigma^{1/2}(I - P)}_{\op}^2 > t} \mathop{}\!\mathrm{d} t
\\&\qquad\leq \int_0^\infty \min\brace{1, 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 + \frac{58\kappa^2}{t p}} \paren{1 + \frac{2\kappa^2}{t}}
\exp\paren{- \frac{p t }{10\kappa^2}}} \mathop{}\!\mathrm{d} t
\\&\qquad =
\frac{10\kappa^2 a}{p} \int_0^\infty \min\brace{1, 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 +
\frac{58}{10 au}} \paren{1 + \frac{p}{5 au}}
\exp\paren{- au }} \mathop{}\!\mathrm{d} u
\\&\qquad \leq
\frac{10\kappa^2 a}{p} \paren{b + \int_b^\infty 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 +
\frac{6}{au}} \paren{1 + \frac{p}{5 au}}
\exp\paren{- au } \mathop{}\!\mathrm{d} u}
\\&\qquad \leq
\frac{10\kappa^2}{p} \paren{ab + 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 +
\frac{6}{ab}} \paren{1 + \frac{p}{5 ab}}
\exp\paren{- ab }}.
\end{align*}
This last quantity is optimized for $ab = \log(p)$, which leads to the first part of the lemma.
Similarly,
\begin{align*}
& \E_{{\cal D}_n}\bracket{\norm{\Sigma^{1/2}(I - P)}_{\op}^{2a}}
= \int_0^\infty \Pbb_{{\cal D}_n}\paren{\norm{\Sigma^{1/2}(I - P)}_{\op}^{2a} > t} \mathop{}\!\mathrm{d} t
\\&\qquad = \int_0^\infty \Pbb_{{\cal D}_n}\paren{\norm{\Sigma^{1/2}(I - P)}_{\op}^{2} > t^{1/a}} \mathop{}\!\mathrm{d} t
\\&\qquad\leq \int_0^\infty \min\brace{1, 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 + \frac{58\kappa^2}{t^{1/a} p}} \paren{1 + \frac{2\kappa^2}{t^{1/a}}}
\exp\paren{- \frac{p t^{1/a} }{10\kappa^2}}} \mathop{}\!\mathrm{d} t
\\&\qquad =
\frac{10^a\kappa^{2a} a c^a}{p^a} \int_0^\infty \min\brace{u^{a-1}, 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 +
\frac{58}{10 cu}} \paren{1 + \frac{p}{5 cu}}
\frac{1}{u^{1-a}}\exp\paren{- cu }} \mathop{}\!\mathrm{d} u
\\&\qquad \leq
\frac{10^a\kappa^{2a} a c^a}{p^a} \paren{\frac{b^a}{a} + \int_b^\infty 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 +
\frac{6}{cu}} \paren{1 + \frac{p}{5 cu}}
\frac{1}{u^{1-a}}\exp\paren{- cu } \mathop{}\!\mathrm{d} u}
\\&\qquad \leq
\frac{10^a\kappa^{2a}}{p^a} \paren{(cb)^a + 2\kappa^2 \norm{\Sigma}^{-1}_{\op} \paren{1 +
\frac{6}{cb}} \paren{1 + \frac{p}{5 cb}}
\frac{1}{(cb)^{1-a}}\exp\paren{- cb }}.
\end{align*}
Once again this is optimized for $cb = \log(p)$.
\end{proof}
\begin{remark}[Leverage scores]
Out of simplicity, we only present a low rank approximation with random subsampling.
Yet, we can improve the result by considering subsampling based on leverage scores.
If we consider the Gaussian kernel, $Sk_x\in L^2$ can be thought of as a function that is a little bump around $x\in\mathcal{X}$.
In essence, subsampling based on leverage scores, consists in representing the solution on a subsampled sequence $(k_{X_i})_{i\in I}$ where the $X_i$ are far from one another so that the bump functions $(Sk_{X_i})$ can approximate a maximum of functions.
\citep{Rudi2015} shows that with leverage scores, we can take $p = (\mu\lambda)^{\gamma} \log(n)$, with $\gamma$ linked with the capacity of the RKHS linked with the kernel $k$.
\end{remark}
If we add all derivations, we have derived the following theorem.
\begin{theorem}
Under Assumptions \ref{lap:ass:source}, \ref{lap:ass:approximation} and \ref{lap:ass:decay},
\begin{equation}
\label{lap:eq:pre_final}
\begin{split}
&\E_{{\cal D}_n}\bracket{\norm{\hat{g} - g_\rho}_{L^2}^2}
\\&\qquad\leq 8c_\mathcal{Y}^2 \paren{1 + 28\ \frac{\kappa^2 + \lambda \partial\kappa^2}{\lambda\mu n}} (1 + \lambda\mu \norm{C}_{\op}^{-1})
\frac{\kappa^2 + \lambda\partial\kappa^2}{\lambda\mu}
\exp\paren{-\frac{\lambda\mu n}{10\paren{\kappa^2 + \lambda\partial\kappa^2}}}
\\&\qquad\cdots + 4\lambda^2 \norm{{\cal L}g_\rho}_{L^2}^2
+ 4\lambda\mu c_a^2 \norm{g_\rho}_{L^2}^2
+ 64 \sigma_\ell^2 (n_\ell \mu\lambda)^{-1} +
57 c_\mathcal{Y}^2\kappa^2(n_\ell^{2} \mu\lambda)^{-1}
\\&\qquad\cdots + 64(\kappa^2 + \lambda\partial \kappa^2)^2 c_a^2 \norm{g_\rho}_{L^2}^2 (\mu\lambda n)^{-1}
+ 57 (\kappa^2 + \lambda\partial \kappa^2)^2 c_a^2 \norm{g_\rho}_{L^2}^2 (\mu\lambda n^2)^{-1}
\\&\qquad\cdots +
12 c_a^2 \norm{g_\rho}^2_{L^2}\paren{\frac{10\kappa^2\log(p)}{p} + \frac{10^a\kappa^{2a} c_d \lambda \log(p)^a}{p^a}}
\\&\qquad\qquad\qquad\cdots \times
\paren{1 + \frac{2\kappa^2}{\norm{\Sigma}_{\op}\log(p)} \paren{1 +
\frac{6}{\log(p)}} \paren{\frac{1}{p} + \frac{1}{5\log(p)}}}.
\end{split}
\end{equation}
where $c_\mathcal{Y}$ is an upper bound on $Y$, $\kappa^2$ is an upper bound on $x \to k(x, x)$, $\partial\kappa^2 = \sum_{i=1}^d \kappa_i^2$ with $\kappa_i^2$ a bound on $x \to \partial_{1i}\partial_{2i} \partial k_{x_i}$, $c_d$ and $a$ the constants appearing in Assumption \ref{lap:ass:decay}, $c_a$ a constant such that $\norm{g}_{\cal H} \leq c_a \norm{g}_{L^2}$ and $\sigma_\ell^2 \leq c_\mathcal{Y}^2 \kappa^2$ a variance parameter linked with the variance of $Y(I + \lambda{\cal L})^{-1}\delta_X$.
\end{theorem}
Theorem \ref{lap:thm:consistency} is a corollary of this theorem.
\section{Spectral decomposition}
\label{lap:app:algebra}
In this section, we recall facts on spectral regularization, before proving Proposition \ref{lap:thm:decomposition} and extending it to the case $\mu = 0$.
\subsection{Generalized singular value with matrices}
\paragraph{Generalized singular value decomposition.}
Let $A \in \mathbb{R}^{m1\times n}$ and $B \in \mathbb{R}^{m_2\times n}$ be two matrices.
There exists $U \in \mathbb{R}^{m_1 \times m_1}$, $V\in \mathbb{R}^{m_2 \times m_2}$ two orthogonal matrices, $c \in \mathbb{R}^{m_1 \times r}$ and $s\in \mathbb{R}^{m_2 \times r}$ two $1$-diagonal matrices such that $c^\top c + s^\top s = I_r$, and $H\in \mathbb{R}^{n\times r}$ a non-singular matrix such that
\[
A = UcH^{-1}, \qquad B = VsH^{-1}.
\]
To be more precise $c$ is such that only entries $c_{ii} = \cos (\theta_i)$ for $i < \min(r, m_2)$ are non-zeros and $s$ such that only entries $s_{m_1-i, r-i} = \sin(\theta_{r-i})$ for $i < \min(r, m_1)$ are non-zeros, with $\theta_i \in [-\pi/2, \pi/2]$ some angle.
Here, $c$ stands for cosine, $s$ for sinus and $r$ for rank.
\paragraph{Link with generalized eigenvalue problem.}
As well as the singular value of $A$ is linked with the eigenvalue of $A^\top A$, the generalized singular value decomposition of $[A; B]$ is linked with the generalized eigenvalue problem linked with $(A^\top A, B^\top B)$.
Indeed, we have
\[
A^\top A = H^{-\top} c^\top c H^{-1}, \qquad B^\top B = H^{-\top} s^\top s H^{-1}.
\]
In particular, with $(e_i)$ the canonical basis of $\mathbb{R}^r$, and $h_i$ the $i$-th column of $H$, we get
\[
H^\top A^\top A h_i = \cos(\theta_i)^2 e_i = \tan(\theta_i)^{-2}
\sin(\theta_i)^2 e_i
= \tan(\theta_i)^{-2} H^\top B^\top B h_i.
\]
From which we deduce that, since $\ima A \cup \ima B \subset \ima H^\top$,
\[
A^\top A h_i = \tan(\theta_i)^{-2} B^\top B h_i,\qquad
h_j^\top B^\top B h_i = \sin(\theta_i)^2 \ind{i=j}.
\]
So if we denote by $f_i = \abs{\sin(\theta_i)}^{-1} h_i$ and $\lambda_i = \tan(\theta_i)^{-2}$, assuming $\lambda_i \neq 0$ for all $i \leq r$ (which corresponds to $\ker B \subset \ker A$), $(\lambda_i)_{i\leq r}, (f_i)_{i\leq r}$ provide the generalized eigenvalue decomposition of $(A^\top A, B^\top B)$ in the sense that
\[
A^\top A f_i = \lambda_i B^\top B f_i,\qquad
f_j^\top B^\top B f_i = \ind{i=j},\qquad
f_j^\top A^\top A f_i = \lambda_i \ind{i=j}.
\]
\subsection{Tikhonov regularization}
Define the Tikhonov regularization
\[
x_\lambda = \argmin_{x\in \mathbb{R}^n} \norm{Ax - b}^2 + \lambda \norm{Bx}^2.
\]
When this problem is well-defined, the solution is defined as
\[
x_\lambda = (A^\top A + \lambda B^\top B)^\dagger A^\top b.
\]
With the generalized singular value decomposition of $A$ and $B$, we have
\[
A^\top A + \lambda B^\top B = H^{-\top} \gamma_\lambda H^{-1},
\quad\text{with}\quad \gamma_\lambda = c^\top c + \lambda s^\top s.
\]
Using the fact that $A^\top b = H^{-\top} c^\top U^\top b$, we get
\[
x_\lambda = H \gamma_\lambda^{-1} c^\top U^\top b
= \paren{\sum_{i=1}^r \frac{\cos(\theta_i)}{\cos(\theta_i)^2 + \lambda \sin(\theta_i)^2} h_i \otimes u_i} b.
\]
Now, we would like to replace $c_{ii}$, $s_{ii}$, $h_i$ and $u_i$ with quantities that depend on $\lambda_i$, $f_i$ and $A$.
To do so recall that $AH = Uc$, therefore $\cos(\theta_i) u_i = Ah_i$, and recall that $h_i = \sin(\theta_i) f_i$ and $\lambda_i = \cos(\theta_i)^2 / \sin(\theta_i)^2$.
Inputting this equality in the last expression of $x_\lambda$ we get
\[
x_\lambda = \paren{\sum_{i=1}^r \frac{\sin(\theta_i)^2}{\cos(\theta_i)^2 +
\lambda \sin(\theta_i)^2} f_i \otimes Af_i} b
= \paren{\sum_{i=1}^r \frac{1}{\lambda_i +
\lambda} f_i \otimes Af_i} b.
\]
Finally,
\[
b_\lambda = A x_\lambda
= \sum_{i=1}^r \psi(\lambda_i) \scap{Af_i}{b} Af_i,
\qquad\text{where}\qquad
\psi(x) = \frac{1}{x+\lambda}.
\]
\subsection{Extension to operators}
To end the proof of Proposition \ref{lap:thm:decomposition}, we should prove that we can apply the generalized eigenvalue decomposition to operators.
We will only prove that it is possible for $(\Sigma, L+\mu)$ based on simple considerations.
\begin{proposition}
When $k$ is continuous and $\supp\rho_\mathcal{X}$ is bounded, $\Sigma$ is a compact operator.
\end{proposition}
\begin{proof}
We have $\Sigma = \E[k_X\otimes k_X]$ and $\norm{k_x} = k(x, x)$.
Since $k$ is continuous and $\supp\rho_\mathcal{X}$ is compact, for $x\in\supp\rho_\mathcal{X}$, $k(x, x)$ is bounded.
Hence, $\Sigma$ is a trace class and compact operator.
\end{proof}
\begin{proposition}
When $k$ is twice differentiable with continuous derivative, and $\supp\rho_\mathcal{X}$ is compact, $L$ is a compact operator.
As a consequence, $L$ has a compact spectrum, and has a pseudo-inverse that we will denote, with a slight abuse of notation, by $L^{-1}$.
\end{proposition}
\begin{proof}
The proof is similar to the one showing that $\Sigma$ is compact, based on the fact that $L = \sum_{i=1}^d \E[\partial_i k_X \otimes \partial_i k_X]$, and $\norm{\partial_i k_X}^2 = \partial_{1i}\partial_{2i} k(x, x)$.
\end{proof}
\begin{proposition}
When $\Sigma$ is compact, for all $\mu > 0$, $(L+\mu)^{-1/2}\Sigma (L+\mu)^{-1/2}$ is a compact operator.
\end{proposition}
\begin{proof}
The proof is straightforward
\[
\trace((L+\mu)^{-1/2} \Sigma (L+\mu)^{-1/2}) =
\trace(\Sigma (L+\mu)^{-1}) \leq
\norm{(L+\mu)^{-1}}_{\op} \trace(\Sigma)
\leq \mu^{-1} \trace(\Sigma) < +\infty.
\]
Therefore, the operator is trace class, hence compact.
\end{proof}
\begin{proposition}
For any $\mu > 0$, the generalized eigenvalue decomposition of $(\Sigma, L+\mu)$ as defined in Proposition \ref{lap:thm:decomposition} exists.
\end{proposition}
\begin{proof}
Using the spectral theorem, since $(L+\mu)^{-1/2}\Sigma (L+\mu)^{-1/2}$ is positive self-adjoint compact operator, there exists $(\xi_i)$ a basis of ${\cal H}$ and $(\lambda_i) \in \mathbb{R}_+$ a decreasing sequence (note that $\ker (L+\mu) = \ker \Sigma = \brace{0}$), such that
\[
(L+\mu)^{-1/2}\Sigma (L+\mu)^{-1/2} = \sum_{i\in\mathbb{N}} \lambda_i \xi_i \otimes \xi_i.
\]
Taking $\theta_i = (L + \mu)^{-1/2} \xi_i$, we get $\Sigma \theta_i = \lambda_i L\theta_i$.
Because $(\xi_i)$ generates ${\cal H}$, and $(L+\mu)^{-1/2}$ is bijective (since $L$ is compact, $(L+\mu)^{-1}$ is coercive), $((L+\mu)^{-1/2}\xi_i)$ generates ${\cal H}$.
\end{proof}
Proposition \ref{lap:thm:decomposition} follows from prior discussion on Tikhonov regularization extended to infinite summations.
\subsection{The case \texorpdfstring{$\mu = 0$}{}}
When $\mu = 0$, \eqref{lap:eq:filtering} should be seen as the rewriting of~\eqref{lap:eq:decomposition_zero} based on the RKHS ${\cal G} = \ima S$.
This can only be done when the eigenvectors of ${\cal L}$ appearing in~\eqref{lap:eq:evd} belongs to ${\cal G} = \ima S = \ima K^{1/2}$, which is exactly what Assumption \ref{lap:ass:approximation} provides.
In such a setting, we can find $(\theta_i) \in {\cal H}^\mathbb{N}$ to write $\lambda_i^{1/2} e_i = S\theta_i$ as soon as $\lambda_i \neq 0$ (write $M e_i = S\theta_i$ for $M$ an abstraction representing $+\infty$ when $\lambda_i = 0$, handling the potential fact that $\ker B \not\subset \ker A$), we get $\theta_iS^*S\theta_j = \lambda_i\ind{i=j}$, and $L \theta_i = \lambda_i^{-1}\Sigma\theta_i$, and we can extend Proposition \ref{lap:thm:decomposition} to the case $\mu = 0$, with
\begin{equation}
\label{lap:eq:filtering_zero}
g_\lambda = \sum_{i\in\mathbb{N}} \psi(\lambda_i) \scap{S^\star g_\rho}{\theta_i} S\theta_i,
\end{equation}
where we handle the null space of ${\cal L}$ with the equality $M\psi(M) = 1$, verified by $M$ our abstraction representing $+\infty$, so that $\psi(M) \scap{S^\star g_\rho}{\theta_i}S\theta_i = \scap{g_\rho}{e_i} e_i$ as soon as $\lambda_i = 0$.
\paragraph{Beyond Assumption \ref{lap:ass:approximation}.}
Assumption \ref{lap:ass:approximation} could be made generic by considering the biggest $(p_i)\in\mathbb{R}_+^\mathbb{N}$ such that $K^{-p_i} e_i$ belongs to $L^2$, and rewriting~\eqref{lap:eq:filtering_zero} under the form $g_\lambda = \sum_{i\in\mathbb{N}} \psi(\lambda_i) \scap{(S_0 K^{p_i})^\star g_\rho}{\theta_i} S_0 K^{p_i} \theta_i$, with $\theta_i = \lambda_i^{-1/2} S_0^{-1} K^{-p_i}\theta_i$ and $S_0 = K^{-1/2}S$ the isomorphism between ${\cal H}$ and $L^2$ (assuming that $S$ is dense in $L^2$).
Such an assumption would describe all situations from no assumption ($p_i = 0$ for all $i$), Assumption \ref{lap:ass:approximation} ($p_i = 1/2$ for all $i$) to even more optimistic assumptions ($p_i \geq 1$ for all $i$).
\section{Experiments}
\label{lap:sec:experiment}
\subsection{Low-rank approximation}
Cutting computation cost thanks to low-rank approximation, as we did by going from the naive exact empirical risk minimizer $\hat{g}$~\eqref{lap:eq:estimate} to the smart implementation $\hat{g}_p$ Algorithm \ref{lap:alg:imp}, is associated with a trade-off between computational versus statistical performance.
This trade-off can be studied theoretically thanks to Theorem \ref{lap:thm:consistency}, which shows that under mild assumptions, considering $p = n^{1/2}\log(n)$ does not lead to any loss in performance, in the sense that the convergence rates in $n$, the number of samples, are only changed by a constant factor.
We show on Figure \ref{lap:fig:cut_loss} that in the setting of Figure \ref{lap:fig:exp}, our low-rank approximation is not associated with a loss in performance.
Actually low-rank approximation can even be beneficial as it tends to lower the risk for overfitting as discussed by \citet{Rudi2015}.
\begin{figure*}[t]
\centering
\includegraphics{laplacian/images/classification_error_bis.pdf}
\includegraphics{laplacian/images/regression_error_bis.pdf}
\vskip -0.1in
\caption{
Cut in computation costs are not associated with a loss in performance.
The estimate $\hat{g}_p$ Algorithm \ref{lap:alg:imp} (in blue), based on low-rank approximation that cut computation cuts performs as well as the exact computation of $\hat{g}$~\eqref{lap:eq:estimate}.
(Left) Classification error in the setting of Figure \ref{lap:fig:exp}.
(Right) Regression error in the same setting.
The fact that the error of the graph-based method stalls around one, is due to the amplitude of the estimate being very small, which is coherent with behaviors described in \citep{Nadler2009}.
}
\label{lap:fig:cut_loss}
\end{figure*}
\subsection{Comparison with graph-based Laplacian}
One the main goal of this paper is to make people drop graph-based Laplacian methods and adopt our ``kernelized'' technique.
As such, we would like to discuss in more detail our comparison with graph-based Laplacian.
In particular, we will discuss how and why we choose the hyperparameters and the setting of Figure \ref{lap:fig:exp}.
\begin{figure*}[t]
\centering
\includegraphics{laplacian/images/comparison.pdf}
\vskip -0.1in
\caption{
Setting of Figure \ref{lap:fig:exp} with $n = 1000$.
(Left) Training set.
We represent a cut of $\mathcal{X} \subset \mathbb{R}^d$ according to the two first coordinates $\brace{(x_1, x_2) \,\middle\vert\, (x_1, x_2, \cdots, x_d)\in \mathcal{X}}$.
We have two Gaussian distributions with unit variance, one centered at $x = (0, 0, \cdots 0)$ and the other one centered at $x = (3, 0, \cdots, 0)$.
One of the Gaussian distributions is associated with the blue class, the other one with the orange class.
We consider $n = 1000$ unlabeled points, represented by small points, colored according to their classes, and $n_\ell = 100$ labeled points, represented in color with black edges.
(Middle) Reconstruction with our kernelized Laplacian methods.
Our method uncovers correctly the structure of the problem, and allows making a quite optimal reconstruction.
The optimal decision frontier is illustrated by the gray line $\partial X$.
(Right) Reconstruction with graph-Laplacian.
The graph-Laplacian diffuses information too far away from what it should, leading to many incorrect guesses.
}
\label{lap:fig:exp_setting}
\end{figure*}
The setting of Figure \ref{lap:fig:exp} is the one of Figure \ref{lap:fig:exp_setting}, we considered two Gaussian with unit variance and whose centers are at distance $\delta = 3$ of each other.
We chose Gaussian distributions as it is a well-understood setting.
We chose $\delta = 3$ so that there is a mild overlap between the two distributions.
For the bandwidth parameter, we considered $\sigma_n = Cn^{-\frac{1}{d+4}}\log(n)$ as this is known to be the optimal bandwidth for graph Laplacian \citep{Hein2007}.
We chose $C = 1$ as this leads to $\sigma_n$ of the order of $\delta$.
We chose $\lambda = 1$ to enforce Laplacian regularization and $\mu_n = 1/n$, as this is a classical regularization parameter in RKHS.
Furthermore, we did not cross-validate parameters in order to be fair with graph-Laplacian that do not have as many parameters as our kernel method.
We compute the error in a transductive setting, retaking the exact problem and algorithm of \citet{Zhu2003}.
We choose $d = 10$, as we know that this is a good dimension parameter in order to illustrate the curse of dimensionality phenomenon without needing too much data.\footnote{%
Note that our consistency result Theorem \ref{lap:thm:consistency} describes a convergence regime that applies to a vast class of problems.
Such a regime usually takes place after a certain number of data (depending on the value of the constant $C$).
Before entering this regime, describing the error of our algorithm would require more precise analysis specific to each problem instance, eventually involving tools from random matrix theory.
}
\subsection{Usefulness of Laplacian regularization}
It is natural to ask about the relevance of Laplacian regularization.
To give convergence results, we have used Assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation}, which imply that $g^*$ belongs to the RKHS ${\cal H}$, and we got convergence rates in $n_l^{1/2}$, which is not better than the rates we could get with pure kernel ridge regression.
In particular, our algorithm can be split between an unsupervised part that learn the penalty $\norm{{\cal L}^{1/2}g}_{L^2(\rho_\mathcal{X})}^2$ and a supervised part, that solve the problem of estimating $g_\lambda$ from few labels $(X_i, Y_i)$ given the penalty associated to ${\cal L}$.
But the same method can be used for pure kernel ridge regression: unsupervised data could be leveraged to learn the covariance matrix $\Sigma$ \eqref{lap:eq:block_op}, and supervised data could be used to get $\widehat{S^\star g_\rho}$ to converge toward $S^\star g_\rho$.
The same analysis would yield the same type of convergence rates.
Yet the parameter $\sigma_\ell$ appearing in Theorem \ref{lap:thm:consistency} would not be linked with the variance of $Y(I + \lambda{\cal L} + \lambda\mu K^{-1})^{-1}\delta_X$ but with the variance of $Y(I + \mu K^{-1})^{-1}\delta_X$.
This is a key fact, the geometry of the covariance operator $\Sigma$ is not supposed to be that relevant to the problem, while the one of $L$ is.
We illustrate this fact on Figure \ref{lap:fig:use}.
\begin{figure*}[t]
\centering
\includegraphics{laplacian/images/S.pdf}
\includegraphics{laplacian/images/L.pdf}
\vskip -0.2in
\caption{
Usefulness of Laplacian regularization.
We illustrate the reconstruction based on our spectral filtering techniques based on the sole use of the covariance matrix $\Sigma$ on the left, and on the sole use of the Laplacian matrix $L$ on the right.
We see that the covariance matrix does not capture the geometry of the problem, which contrasts with the use of Laplacian regularization.
}
\label{lap:fig:use}
\vskip -0.1in
\end{figure*}
\section{Extension to partially supervised learning}
\label{lap:sec:extension}
In this section, we first show how our work can be extended to generic semi-supervised learning problems, beyond real-valued regression.
This first extension is based on the least-square surrogate introduced by \citet{Ciliberto2020} for structured prediction problems.
We later show how our work can be extended to generic partially-supervised learning.
This second extension is based on the work of \citet{Cabannes2020}.
\subsection{Structured prediction and least-square surrogate}
\label{lap:sec:structured_prediction}
Until now, we have considered the least-square problem with $Y \in \mathbb{R}$.
Indeed, our work can be extended easily to a wide class of learning problems.
Consider $\mathcal{Y}$ an output space, $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ a loss function, and keep $\mathcal{X}\subset \mathbb{R}^d$ and $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$.
Suppose that we want to retrieve
\begin{equation}
\label{lap:eq:structured_prediction}
f^* = \argmin_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}(f),\qquad\text{with}\qquad
{\cal R}(f) = \E_{(X, Y)\sim \rho} \bracket{\ell(f(X), Y)}.
\end{equation}
\citet{Ciliberto2020} showed that
as soon as $\ell$ can be decomposed through two mappings $\phi:\mathcal{Y}\to{\cal H}_\mathcal{Y}$ and $\psi:\mathcal{Y}\to{\cal H}_\mathcal{Y}$ with ${\cal H}_\mathcal{Y}$ a Hilbert space as $\ell(y, z) = \scap{\phi(y)}{\psi(z)}_{{\cal H}_\mathcal{Y}}$, it is possible to leverage the least-square regression by considering the surrogate problem
\begin{equation}
\label{lap:eq:surrogate}
g^* \in \argmin_{g:\mathcal{X}\to{\cal H}_\mathcal{Y}}
\E_{(X, Y)\sim\rho}\bracket{\norm{g(X) - \phi(Y)}^2_{{\cal H}_\mathcal{Y}}}.
\end{equation}
This surrogate problem relates to the original one through the decoding $d$ that relates a surrogate estimate $g:\mathcal{X}\to{\cal H}_\mathcal{Y}$ to an estimate of the original problem $f:\mathcal{X}\to\mathcal{Y}$ as $f = d(g)$ defined through, for $x\in\supp\rho_\mathcal{X}$,
\begin{equation}
\label{lap:eq:decoding}
f(x) = \argmin_{z\in\mathcal{Y}} \scap{\psi(z)}{g(x)}_{{\cal H}_\mathcal{Y}}.
\end{equation}
In the real-valued regression case, presented previously, our estimates for $g_n$ can all be written as $g_n(x) = \sum_{i=1}^{n_\ell} \beta_i(x) Y_i$, where $\beta_i(x)$ is a function of the $(X_i)_{i\leq n}$, involving the kernel $k$ and its derivatives.
Those estimates can be cast to vector-valued regression by considering coordinates-wise regression,\footnote{%
To parametrize functions $g$ from $\mathcal{X}$ to ${\cal H}_\mathcal{Y}$, we can parametrize independently each coordinates $\scap{g}{e_i}_{{\cal H}_\mathcal{Y}}$, for $(e_i)$ a basis of ${\cal H}_\mathcal{Y}$, by the space ${\cal G}$ -- note that it is possible to generalize real-valued kernel to parametrize coordinates in a joint fashion \citep{Caponnetto2007}.
The coordinate-wise parametrization corresponds to the tensorization ${\cal H}' = {\cal H}_\mathcal{Y} \otimes {\cal H}$ and to the parametric space ${\cal G}' = \brace{x\to \Theta k_x \,\middle\vert\, \Theta \in {\cal H}'}$ of functions from $\mathcal{X}$ to ${\cal H}_\mathcal{Y}$.
${\cal G}'$ naturally inherits the Hilbertian structure of ${\cal H}'$, itself inherited from the structure of ${\cal H}$ and ${\cal H}_\mathcal{Y}$.
} which leads to $g_n(x) = \sum_{i=1}^{n_\ell}\beta_i(x) \phi(Y_i)$, and to the original estimates, for any $x\in\supp\rho_\mathcal{X}$,
\begin{equation}
\label{lap:eq:loss_trick}
f_n(x) \in \argmin_{z\in{\cal Z}} \sum_{i=1}^{n_\ell} \beta_i(x) \ell(z, Y_i).
\end{equation}
The behavior of $f_n$ being independent of the decomposition $(\phi, \psi)$ of $\ell$ was referred to as the loss trick.
In particular, \citet{Ciliberto2020} showed that convergence rates derived between $\norm{g_n - g^*}_{L^2}$ does not change if we consider $g:\mathcal{X}\to\mathbb{R}$ or $g:\mathcal{X}\to{\cal H}_\mathcal{Y}$ and that those rates can be cast directly as convergence rates between ${\cal R}(f_n)$ and ${\cal R}(f^*)$ with $f_n = d(g_n)$ defined by~\eqref{lap:eq:decoding}.
Moreover, when $\mathcal{Y}$ is a discrete output space, it is possible to get much better generalization bound on ${\cal R}(f_n) - {\cal R}^*$by introducing geometrical considerations regarding $g^*$ and decision frontier between classes \citep{Cabannes2021b}.
\begin{example}[Binary classification]
\label{lap:ex:binary}
This framework aims at generalizing well known surrogate considerations in the case of the binary classification.
Binary classification corresponds to $\mathcal{Y} = \brace{-1, 1}$, $\ell$ the $0-1$ loss.
In this setting, ${\cal H}_\mathcal{Y} = \mathbb{R}$, $\phi: \mathcal{Y}\to\mathbb{R}; y \to y$, and $\psi = -\phi$.
This definition verifies $\ell(y, z) = .5 - .5 \phi(y)^\top \phi(z) \simeq \phi(y)^\top \psi(z)$.
This corresponds to the usual least-square surrogate, which is \( {\cal R}_S(g) = \E[\norm{g(X) - Y}^2]\), \(g(x) = \E\bracket{Y\,\middle\vert\, X=x}\) and $f = \sign g$.
\end{example}
\paragraph{Beyond least-squares.}
Considering a least-square surrogate assumes that retrieving $g^*$ \eqref{lap:eq:surrogate} is the way to solve the original problem \eqref{lap:eq:structured_prediction} and that the low-density separation hypothesis can be expressed as Assumption \ref{lap:ass:source} being verified by $g^*$.
We would like to point out that the low-density separation could be expressed under a much weaker form, which is that there exists $g$ such that $f^* = d(g)$ \eqref{lap:eq:decoding} and $g$ verifies Assumption \ref{lap:ass:source}.
In particular, the cluster assumption \citep{Rigollet2007} could be understood as assuming that $g = \phi(f^*)$, the trivial embedding of $f^*$ in ${\cal H}_\mathcal{Y}$, is constant on clusters, which means that $g$ belongs to the kernel of the Laplacian operator ${\cal L}$.
Yet, $g^*:x\to\E[\phi(Y)\vert X=x]$, which depends on the labeling noise, could be really non-smooth, even under the cluster assumption.
Those considerations are related to an open problem in machine learning, which is that we do not know what is the best statistical way (and the best surrogate problem) to solve the fully supervised binary classification problem \citep[see {\em e.g.}][]{Zhang2020}.
However, many points introduced in the work could be retaken with other surrogate, could it be SVM (which leads to $g^* = \phi(f^*)$, with $g^*$ minimizing the Hinge loss), softmax regression (used in deep learning) or others.
\subsection{Partially supervised learning}
Partial supervision is a popular instance of weak supervision, which generalizes semi-supervised learning.
It has been known under the name of partial labeling \citep{Cour2011}, superset learning \citep{Liu2014}, as well as learning with partial label \citep{Grandvalet2002}, with partial annotation \citep{Lou2012}, with candidate labeling set \citep{Luo2010} or with multiple label \citep{Jin2002}.
It encompasses many problems such as ``classification with partial labels'' \citep{Nguyen2008,Cour2011}, ``multilabeling with missing labels'' \citep{Yu2014}, ``ranking with partial ordering'' \citep{Hullermeier2008}, ``regression with censored data'' \citep{Tobin1958}, ``segmentation with pixel annotation'' \citep{Verbeek2007,Papandreou2015}, as well as instances of ``action retrieval'', especially on instructional videos \citep{Alayrac2016,Miech2019}.
It consists, for a given input $x$, in not observing its label $y\in\mathcal{Y}$, but observing a set of potential labels $s\in 2^\mathcal{Y}$ that contains the labels ($y\in s$).
Typically, if $\mathcal{Y}$ is the space $\mathfrak{S}_m$ of orderings between $m$ items (\emph{e.g.} movies on a streaming website), for a given input $x$ (\emph{e.g.} some feature vectors characterizing a user) $s$ might be specified by a partial ordering that the true label $y$ should satisfy (\emph{e.g.} the user prefers romantic movies over action movies).
In this setting, it is natural to create consensus between the different sets giving information on $(y\vert x)$, which has been formalized mathematically by the infimum loss $(z,s)\in\mathcal{Y}\times 2^\mathcal{Y} \to \inf_{y\in s} \ell(z, y)\in\mathbb{R}$ for $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ a specified loss on the underlying fully supervised learning problem.
This leads, for $\tau \in \prob{\mathcal{X}\times 2^\mathcal{Y}}$ encoding the distribution generating samples $(X, S)$, to the formulation
\(
f^* \in {\cal F} = \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X,
S)\sim\tau}\bracket{\inf_{Y\in S}\ell(f(X), Y)}.
\)
To study this problem, a non-ambiguity assumption is usually made \citep{Cour2011,Luo2010,Liu2014,Cabannes2020,Cabannes2021}.
This is a very strong assumption to ensure that ${\cal F}$ is, in essence, a singleton.
Highly adequate to this setting, the Laplacian regularization allows relaxing this assumption, assuming that ${\cal F}$ can be big, but that we can discriminate between function in ${\cal F}$ by looking for the smoothest one in the sense defined by the Laplacian penalty.
Moreover, the loss trick \eqref{lap:eq:loss_trick} allows endowing, in an off-the-shelf fashion, the recent work of \citet{Cabannes2020,Cabannes2021} on the partial supervised learning problem with our considerations on Laplacian regularization.
\section{Central operators}
\label{lap:app:operators}
The paper makes an intensive use of operators.
This section aims at providing details and intuitions on those operators, in order to help the reader.
In particular, we discuss Assumptions \ref{lap:ass:source} and \ref{lap:ass:approximation}, and we prove the equality in~\eqref{lap:eq:block_op}.
\begin{table}[ht]
\caption{Notations}
\label{lap:tab:notations}
\vskip 0.1in
\centering
\begin{tabular}{cc}
\toprule
Symbol & Description \\
\midrule
$(X_i)_{i\leq n}$ & $n$ samples of input data \\
$(Y_i)_{i\leq n_l}$ & $n_l$ labels \\
$\rho$ & Distribution of $(X, Y)$ \\
$g_\rho$ & Function to learn \eqref{lap:eq:least_square} \\
$\lambda, \mu$ & Regularization parameters \\
$g_{\lambda}, g_{\lambda,\mu}$ & Biased estimates (\ref{lap:eq:laplacian_tikhonov}, \ref{lap:eq:tikhonov}) \\
$\hat{g}$ & Empirical estimate \eqref{lap:eq:estimate} \\
$\hat{g}_p$ & Empirical estimate with low-rank approximation (Algo. \ref{lap:alg:imp}) \\
${\cal H}$ & Reproducing kernel Hilbert space \\
$k$ & Reproducing kernel \\
$S$ & Embedding of ${\cal H}$ in $L^2$ \\
$S^\star$ & Adjoint of $S$, operating from $L^2$ to ${\cal H}$ \\
$\Sigma = S^\star S$ & Covariance operator on ${\cal H}$ \\
$K = SS^\star$ & Equivalent of $\Sigma$ on $L^2$ \\
${\cal L}$ & Diffusion operator (a.k.a. Laplacian) \\
$L = S^\star{\cal L}S$ & Restriction of the diffusion operator to ${\cal H}$ \\
$g$ & Generic element in $L^2$ \\
$\theta$ & Generic element in ${\cal H}$ \\
$\lambda_i$ & Generic eigenvalue \\
$e_i$ & Generic eigenvector in $L^2$ \\
\bottomrule
\end{tabular}
\end{table}
\subsection{The diffusion operator}
In this subsection, we extend on the diffusion operator, and recall its basic properties.
The diffusion operator is a well-known operator in the realm of partial differential equations.
Let us assume that $\rho_\mathcal{X}$ admits a smooth density $\rho_\mathcal{X}(dx) = p(x) dx$, say $p \in {\cal C}^2(\mathbb{R}^d)$ that cancels outside a domain $\Omega \subset \mathbb{R}^d$.
Then the diffusion operator ${\cal L}$ can be explicitly written, for $g$ twice differentiable, as
\[
{\cal L}g(x) = - \Delta g(x) + \frac{1}{p(x)} \scap{\nabla p(x)}{\nabla g(x)}.
\]
This follows from the fact that for $f$ once and $g$ twice differentiable, using Stokes theorem,
\begin{align*}
\scap{f}{{\cal L} g}_{L^2(\rho_\mathcal{X})}
& = \scap{\nabla f}{\nabla g}_{L^2(\rho_\mathcal{X})}
= \scap{\nabla f}{ p \nabla g}_{L^2(\mathop{}\!\mathrm{d} x)}
\\&= \int_\mathcal{X} \Div(f p \nabla g ) \mathop{}\!\mathrm{d} x -
\scap{f}{\Div(p\nabla g)}_{L^2(\mathop{}\!\mathrm{d} x)}
= -\scap{f}{\Div(p \nabla g)}_{L^2(\mathop{}\!\mathrm{d} x)}
\\&= -\scap{f}{p^{-1}\Div (p \nabla g)}_{L^2(\rho_\mathcal{X})}
= -\scap{f}{\Div \nabla g + p^{-1} (\nabla p). \nabla g}_{L^2(\rho_\mathcal{X})}.
\end{align*}
Note that when the distribution is uniform on $\Omega$, the diffusion operator is exactly the opposite of the usual Laplacian operator $\Delta$.
As for the Laplacian case, it can be shown that under mild assumption on $p$, whose smoothness properties directly translates to the smoothness properties of the boundary of $\Omega$, that the diffusion operator ${\cal L}$ has a compact resolvent (that is, for $\lambda \notin \spec({\cal L})$,
$({\cal L} + \lambda I)^{-1}$ is compact).
This is a standard result implied by a standard version of the famous Rellich-Kondrachov compactness embedding theorem: $H^2(\Omega)$ is compactly injected in $L^2(\Omega)$ whenever $\Omega$ is a bounded open with $C^2$-boundaries.
In such a setting, we can consider the eigenvalue decomposition of ${\cal L}^{-1}$, that is, $(\lambda_i, e_i) \in (\mathbb{R}_+ \times L^2)^\mathbb{N}$, with $(e_i)_{i\in\mathbb{N}}$ an orthonormal basis of $L^2$ and $(e_i)_{i\leq \dim\ker{\cal L}}$ generating the null space of ${\cal L}$, with the convention $\lambda_i = M$ for $i \leq \dim\ker {\cal L}$, with $M$ an abstraction representing $+\infty$, and $(\lambda_i)$ decreasing toward zero afterwards.
This decomposition reads
\begin{equation}
\label{lap:eq:L}
{\cal L}^{-1} = \sum_{i \in \mathbb{N}} \lambda_i e_i \otimes e_i.
\end{equation}
Note that the fact that all the $(\lambda_i)$ are positive, is due to the fact that ${\cal L}^{-1}$ is the inverse of a positive self-adjoint operator.
As a consequence, the diffusion operator has discrete spectrum, and can be written as
\begin{equation}
\label{lap:eq:evd}
{\cal L} = \sum_{i\in\mathbb{N}} \lambda_i^{-1} e_i \otimes e_i.
\end{equation}
In such a setting, the kernel-free Tikhonov regularization~\eqref{lap:eq:laplacian_tikhonov} reads
\begin{equation}
\label{lap:eq:decomposition_zero}
g_\lambda = \sum_{i\in\mathbb{N}} \psi(\lambda_i)\scap{g_\rho}{\lambda_i^{1/2} e_i}_{L^2}
\lambda_i^{1/2} e_i,
\end{equation}
with $\psi:x \to (x+\lambda)^{-1}$, and the convention $M\psi(M) = 1$.
\subsection{Regularity of the eigenvectors of the diffusion operator}
In this subsection, we extend on the regularity assumed in Assumption \ref{lap:ass:approximation}.
Introducing the kernel $k$ and its associated RKHS ${\cal H}$ is useful when the eigenvectors of ${\cal L}$ can be well approximated by functions in ${\cal H}$.
In applications, people tend to go for kernels that are translation-invariant, which implied that the RKHS ${\cal H}$ is made of smooth functions, could it be analytical functions (for the Gaussian kernel) or functions in $H^m$ (for Sobolev kernels).
As a consequence, we should investigate the regularity of those eigenvectors.
Indeed, if $\rho$ derives from a Gibbs potential, that is $\rho(\mathop{}\!\mathrm{d} x) = e^{-V(x)} \mathop{}\!\mathrm{d} x$, the eigenvectors of ${\cal L}$ can be shown to inherit from the smoothness of the potential $V$ \citep{PillaudVivien2020b}.
For example if $V$ belongs to $H^m$, and $H^m \subset {\cal H}$, we expect $(e_i)$ to belong to ${\cal H}$, thus verifying Assumption \ref{lap:ass:approximation}.
\paragraph{Counter-example and beyond Assumption \ref{lap:ass:approximation}.}
Note that if $\rho$ has several connected components of non-empty interiors, the null space of ${\cal L}$ is made of functions that are constants on each connected component of $\supp\rho_\mathcal{X}$.
Those functions are not analytic.
In such a setting, the Gaussian kernel is not sufficient for Assumption \ref{lap:ass:approximation} to hold, and one should favor kernel associated with richer functional space such as the Laplace kernel or the neural tangent kernel \citep{Chen2021}.
However, as illustrated by Figure \ref{lap:fig:unsupervised}, $e_i$ not belonging to ${\cal H}$ does not mean that $e_i$ can not be well approximated by ${\cal H}$.
Indeed, it is well known that the approximation power of ${\cal H}$ for $e_i$ can be measure in the biggest power $p$ such that $e_i \in \ima K^p$ \citep{Caponnetto2007}, where $K = SS^*$.
Assumption \ref{lap:ass:approximation} corresponds to $p=1/2$, but it should be seen as a specific instance of more generic approximation conditions.
\paragraph{Handling constants in RKHS.}
Finally, note also that many RKHS do not contain constant functions, and therefore might not contain the constant function $e_0$ (although we are only looking for equality in the support of $\rho_\mathcal{X}$), however this specific point with $e_0$ can easily be circumvented either by assuming that $g_\rho$ has zero mean, either by centering the covariance matrices $\Sigma$ and $\hat\Sigma$ \citep{PillaudVivien2020b}.
This relates with the usual technique for SVM consisting in adding an unpenalized bias \citep{Steinwart2008}.
\subsection{Low-density separation}
In this subsection, we discuss how Assumption \ref{lap:ass:source} relates to the idea of low-density separation.
\paragraph{Low-variation intuition.}
The low-density separation supposes that the variations of $g^*$ take place in region with low-density, so that $\norm{{\cal L}^{1/2}g^*} / \norm{g^*}$ is small.
As such, using Courant-Fischer principle, Assumption \ref{lap:ass:source} can be reformulated as $g^*$ belonging to the space
\[
\Span\brace{e_i}_{i\leq r} =
\argmin_{\substack{{\cal F}\subset L^2;\\ \dim{\cal F} = r}}
\max_{g\in{\cal F}} \frac{\norm{{\cal L}^{-1/2}g}_{L^2}^2}{\norm{g}_{L^2}^2}.
\]
In other terms, Assumption \ref{lap:ass:source} can be restated as $g^*$ belonging to a finite dimensional space that minimizes a measure of variation given by the Dirichlet energy.
To tell the story differently, suppose that we are in a classification setting, {\em i.e.} $Y\in\brace{-1, 1}$, and that the $\supp\rho_\mathcal{X}$ is connected.
Then we know that the null space of ${\cal L}$ is made of constant functions.
Then the first eigenvector $e_2$ of ${\cal L}$ is a function that is orthogonal to constants.
Hence, $e_2$ is a function that changes its sign and that is ``balanced'' in the sense that $\E[e_2] = 0$ -- {\em i.e.} if $e_2(x) = \E_{\mu}[Y\,\vert\, X=x]$ for some measure $\mu$, we have $\E_\mu[Y] = 0$, meaning that classes are ``balanced''.
Moreover, in order to minimize $\norm{{\cal L}^{1/2}e_2}$, the variations of $e_2$ should take place in low-density regions of $\mathcal{X}$.
\paragraph{Diffusion intuition.}
Finally, as ${\cal L}$ is a diffusion operator, we also have an interpretation of Assumption \ref{lap:ass:source} in terms of diffusion.
Consider $(\lambda_i, e_i)$ the eigenelements of~\eqref{lap:eq:evd}.
The diffusion of $g_\rho$ according the density $\rho_\mathcal{X}$ can be written as, for $t \in \mathbb{R}$,
\[
g_t = e^{-t {\cal L}} g_\rho = \sum_{i\in\mathbb{N}} e^{-t\lambda_i^{-1}} \scap{g_\rho}{e_i} e_i.
\]
This diffusion will cut off the high frequencies of $g_\rho$ that corresponds to $\scap{g_\rho}{e_i}$ for big $i$, and big $\lambda_i^{-1}$.
Indeed, the difference between the diffusion and the original $g_\rho$ can be measured as
\[
\norm{g_t - g_\rho}_{L^2}^2 = \sum_{i\in\mathbb{N}} (e^{-t\lambda_i^{-1}} - 1)^2\scap{g_\rho}{e_i}^2
= \sum_{i\in\mathbb{N}} t^2\lambda_i^{-2} \scap{g_\rho}{e_i}^2 + o(t^2\lambda_i^{-2}).
\]
Hence, assuming that $g_\rho$ is supported on a few of the first eigenvectors of ${\cal L}$, can be rephrased as saying that the diffusion of $g_\rho$ does not modify it too much.
\paragraph{The variance $\sigma_\ell$.}
Theorem \ref{lap:thm:consistency} shows that the need for labels depends on the variance parameter $\sigma_{\ell}^2$.
It is natural to wonder how this parameter relates to the low-density hypothesis.
As we discussed, this parameter is linked to the variance of $Z = Y(I + \lambda {\cal L})^{-1}\delta_X$.
We can separate the variability of this variable due to $X$ and the variability due to $Y$
\[
Z = Z_X + Z_Y, \qquad\text{with}\qquad
Z_X = (I + \lambda {\cal L})^{-1} g_\rho(X)\delta_X,\quad
Z_Y = (I + \lambda {\cal L})^{-1} (Y - \E[Y\,\vert\,X])\delta_X.
\]
As such we see that this variance depends on the structure of the density $\rho_\mathcal{X}$ with the variance of $(I + \lambda{\cal L})^{-1}\delta_X$, and the labeling noise with the variance of $(Y\,\vert\, X)$.
The low-density separation does not tell us anything about the level of noise in $Y$ or the diffusion structure linked with $\rho_\mathcal{X}$, but additional hypotheses could be made to characterize those.
\subsection{Kernel operators}
In this subsection, we define formally the operators $S$ and $\Sigma$.
We now turn toward operators linked with the Hilbert space ${\cal H}$.
Recall that for $k:\mathcal{X}\to\mathcal{X}\to\mathbb{R}$ a kernel, ${\cal H}$ is defined the closure of the span of the elements $k_x$ under the scalar product $\scap{k_x}{k_{x'}} = k(x, x')$.
In particular, $\norm{k_x}^2_{\cal H} = k(x, x)$.
${\cal H}$ parametrizes a vast class of functions in $\mathbb{R}^\mathcal{X}$ through the mapping
\[
\myfunction{S}{\cal H}{\mathbb{R}^\mathcal{X}}{\theta}{(\scap{k_x}{\theta})_{x\in\mathcal{X}}.}
\]
Under mild assumptions, $S$ maps ${\cal H}$ to a space of functions belonging to $L^2$.
\begin{proposition}
When $x\to k(x, x)$ belongs to $L^1(\rho_\mathcal{X})$, $S$ is a continuous mapping from ${\cal H}$ to $L^2(\rho_\mathcal{X})$.
This is particularly the case when $\rho_\mathcal{X}$ has compact support and $k$ is continuous.
\end{proposition}
\begin{proof}
Consider $\theta\in{\cal H}$, we have
\begin{align*}
\norm{S\theta}_{L^2}^2 & = \int_\mathcal{X} \scap{k_x}{\theta}^2\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
\leq \int_\mathcal{X} \scap{k_x}{\theta}^2_{\cal H}\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
\leq \int_\mathcal{X} \norm{k_x}^2_{\cal H} \norm{\theta}^2_{\cal H}\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
\\&= \norm{\theta}_{\cal H}^2 \int_\mathcal{X} k(x, x)\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
= \norm{\theta}^2_{\cal H}\norm{x\to k(x, x)}_{L^1}.
\end{align*}
Moreover, when $\rho_\mathcal{X}$ has compact support and $k$ is continuous, $k$ is bounded on the support of $\rho_\mathcal{X}$ therefore $x\to k(x, x)$ belongs to $L^1$.
\end{proof}
As a continuous operator from the Hilbert space ${\cal H}$ to the Hilbert space $L^2$, $S$ is naturally associated with many linear structures: in particular its adjoint $S^\star$, but also the self-adjoint operators $K = SS^\star$ and $\Sigma = S^\star S$.
\begin{proposition}
The adjoint of $S$ is defined as
\[
\myfunction{S^\star}{L^2}{\cal H}{g}{\int_\mathcal{X} g(x)k_x \rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
= \E_{X\sim\rho_\mathcal{X}}[g(X)k_X].}
\]
To $S$ is associated the kernel self-adjoint operator on $L^2$
\[
\myfunction{K := SS^\star}{L^2}{L^2}{g}{(x\to\int_\mathcal{X} k(x, x') g(x') \rho_\mathcal{X}(\mathop{}\!\mathrm{d} x')),}
\]
as well as the (non-centered) covariance on ${\cal H}$,
\(
\Sigma := S^\star S = \E_{X\sim\rho_\mathcal{X}}[k_X\otimes k_X].
\)
\end{proposition}
\begin{proof}
We shall prove the equality defining those operators.
Consider $\theta \in {\cal H}$ and $g \in L^2$, we have
\[
\scap{S^\star g}{\theta}_{\cal H}
= \scap{g}{S\theta}_{L^2}
= \E_{X\sim\rho_\mathcal{X}}[g(X) \scap{k_X}{\theta}_{\cal H}]
= \scap{\E_{X\sim\rho_\mathcal{X}}[g(X)k_X]}{\theta}_{\cal H}.
\]
We also have, for $x\in \mathcal{X}$,
\[
(SS^\star g)(x) = \scap{k_x}{\E_{X\sim\rho_\mathcal{X}}[g(X)k_X]}_{\cal H}
= \E_{X\sim\rho_\mathcal{X}}[g(X)\scap{k_x}{k_X}_{\cal H}]
= \E_{X\sim\rho_\mathcal{X}}[g(X)k(X, x)].
\]
Finally, we have
\[
S^\star S \theta = \E_{X\sim\rho_\mathcal{X}}[S\theta(X)k_X]
= \E_{X\sim\rho_\mathcal{X}}[\scap{\theta}{k_X}_{\cal H}k_X]
= \E_{X\sim\rho_\mathcal{X}}[k_X\otimes k_X]\theta.
\]
This provides the last of all the equalities stated above.
\end{proof}
\paragraph{The functional space \texorpdfstring{${\cal H}$}{}.}
In the main text, we have written everything in terms of $\theta$, highlighting the parametric nature of kernel methods.
This made it easier to dissociate the norm on functions derived from ${\cal H}$ and the one derived from $L^2$ or $H^1$.
In literature, people tend to keep everything in terms of functions $g_\theta = S\theta$ without even mentioning the dependency in $\theta$.
Such a setting consists in considering directly the RKHS ${\cal H}$ whose scalar product is defined for $g, g'\in (\ker K)^\perp$ by $\scap{g}{g'}_{\cal H} = \scap{g}{K^{-1}g'}_{L^2}$.
\subsection{Derivative operators}
In this subsection, we extend on derivatives in RKHS, and we formally define the operator $L$.
As well as evaluation maps can be represented in ${\cal H}$, under mild assumptions, derivative evaluation maps can benefit from such a property.
Indeed, for $g_\theta = S\theta$, $x \in \mathcal{X}$ and $u \in {\cal B}_\mathcal{X}(0, 1)$ a unit vector, we have
\[
\partial_u g_\theta(x) = \lim_{t\to 0} \frac{g_\theta(x+tu) - g_\theta(x)}{t}
= \lim_{t\to 0} \frac{\scap{\theta}{k_{x+tu}}_{\cal H} - \scap{g}{k_x}_{\cal H}}{t}
= \lim_{t\to 0} \scap{\theta}{\frac{k_{x+tu} - k_x}{t}}_{\cal H}
\]
As a linear combination of elements in ${\cal H}$, the difference quotient evaluation map $t^{-1}(k_{x+tu} - k_x)$ belongs to ${\cal H}$ and has a norm
\[
\norm{\frac{k_{x+tu} - k_x}{t}}^2_{\cal H}
= \frac{k(x+tu, x+tu) - 2k(x+tu, x) + k(x, x)}{t^2}.
\]
In order for the limit when $t$ goes to zero to belong to ${\cal H}$, we see the importance of $k$ to be twice differentiable.
This limit $\partial_u k_x$, whose existence is proven formally by \citet{Zhou2008}, provides a derivative evaluation map in the sense that
\[
\partial_u g_\theta(x) = \scap{\theta}{\partial_u k_x}_{\cal H}.
\]
From this equality, we derive that $\partial_{1i}k(x, x') = \scap{k_{x'}}{\partial_i k_x}$, and recursively that $\scap{\partial_i k_x}{\partial_j k_{x'}} = \partial_{1i}\partial_{2j} k(x, x')$.
Similarly to the operator $S$, we can introduce the operators $Z_i$ for $i\in\bracket{1, d}$, defined as
\[
\myfunction{Z_i}{\cal H}{\mathbb{R}^\mathcal{X}}{\theta}{(\scap{\partial_ik_x}{\theta}_{\cal H})_{i\leq d}}.
\]
Once again, under mild assumptions, $\ima Z_i$ inherit from a Hilbertian structure.
\begin{proposition}
When $x\to \partial_{1i} \partial_{2i} k(x, x)$ belongs to $L^1(\rho_\mathcal{X})$, $Z_i$ is a continuous mapping from ${\cal H}$ to $L^2(\rho_\mathcal{X})$.
This is particularly the case when $\rho_\mathcal{X}$ has compact support and $k$ is twice differentiable with continuous derivatives.
\end{proposition}
\begin{proof}
Consider $\theta\in{\cal H}$, similarly to before, we have
\begin{align*}
\norm{Z\theta}_{L^2}^2 & = \int_\mathcal{X} \scap{\partial_i k_x}{\theta}^2\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
\leq \norm{\theta}^2_{\cal H} \int_\mathcal{X} \norm{\partial_i k_x}^2_{\cal H}\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
= \norm{\theta}^2_{\cal H}\norm{x\to \partial_{1,i}\partial_{2,i} k(x, x)}_{L^1}.
\end{align*}
Moreover, when $\rho_\mathcal{X}$ has compact support and $\partial_{1,i}\partial_{2,i}k$ is continuous, $\partial_{1,i}\partial_{2,i}k$ is bounded on the support of $\rho_\mathcal{X}$ therefore $x\to\partial_{1,i}\partial_{2,i}k$ belongs to $L^1$.
\end{proof}
Among the linear operators that can be built from $Z_i$, in the theoretical part of this paper, we are mainly interested in $Z_i^\star Z_i$.
In the empirical part however, we might be interested in $Z_i Z_j^\star$ as well as $Z_iS^\star$ as it appears in Algorithm \ref{lap:alg:imp} (where the notation $Z_n$ there has to be understood as the empirical version of $Z = [Z_1; \cdots; Z_d]$).
\begin{proposition}
The Dirichlet energy on ${\cal H}$ can be represented through the operator
\[
S^\star {\cal L} S = \sum_{i=1}^d Z_i^\star Z_i
= \sum_{i=1}^d \E_{X\sim\rho_\mathcal{X}}[\partial_i k_X\otimes \partial_i k_X].
\]
\end{proposition}
\begin{proof}
Let $\theta \in {\cal H}$ and $g_\theta = S\theta$, we have
\begin{align*}
\scap{g_\theta}{{\cal L}g_\theta}_{L^2}
& = \scap{\theta}{S^\star{\cal L}S\theta}_{\cal H}
= \E_{X\sim\rho_\mathcal{X}}\bracket{\norm{\nabla g_\theta(X)}^2}
= \sum_{i=1}^d\E_{X\sim\rho_\mathcal{X}}\bracket{(\partial_i g_\theta(X))^2}
\\&= \sum_{i=1}^d\E_{X\sim\rho_\mathcal{X}}\bracket{\scap{\partial_i k_X}{\theta}^2_{\cal H}}
= \sum_{i=1}^d\norm{Z_i\theta}^2_{L^2}
= \scap{\theta}{\sum_{i=1}^d Z_i^\star Z_i \theta}_{\cal H}
\\&= \sum_{i=1}^d\E_{X\sim\rho_\mathcal{X}}\bracket{\scap{\theta}{(\partial_i k_X \otimes \partial_i k_X)\theta}_{\cal H}}
= \scap{\theta}{\sum_{i=1}^d\E_{X\sim\rho_\mathcal{X}}\bracket{\partial_i k_X \otimes \partial_i k_X}\theta}_{\cal H}.
\end{align*}
Since the three operators are self-adjoint, and they all represent the same quadratic form, they are equals.
\end{proof}
\subsection{Relation between \texorpdfstring{$\Sigma$}{} and \texorpdfstring{$L$}{}}
In this subsection, we discuss the relation between $\Sigma$ and $L$ and show that we can expect the existence of $a \in (1 - 2 / d, 1]$ and $c > 0$ such that $L \preceq c \Sigma^a$.
\paragraph{Informal capacity considerations.}
We want to compare $\Sigma$ and $L$, as $L \preceq c\Sigma^a$ with the biggest $a$ possible.
This depends on how fast the eigenvalues are decreasing, which is linked to the entropy numbers of those two compact operators.
Those entropy numbers are linked with the capacity of the functional spaces $\brace{g\in L^2\,\middle\vert\, \norm{K^{-1/2}g}_{L^2} < \infty}$ and $\brace{g\in L^2\,\middle\vert\, \norm{K^{-1/2}{\cal L}^{-1/2}g}_{L^2} < \infty}$.
The first space is the reproducing kernel Hilbert space linked with $k$, the second space is, roughly speaking, a space of functions whose integral belongs to the first space.
As such, if the first space is ${\cal H}^m$, the second is ${\cal H}^{m-1}$, and we can consider $a = (m - 1) / m$.
Because we are considering kernels, we have $m > d / 2$ (this to make sure that the evaluation functionals $L_X : f\to f(x)$ are continuous), so that $a > 1 - 2 / d$.
Without trying to make those ``algebraic'' considerations more formal, we will give an example on the torus.
\paragraph{Translation-invariant kernel and Fourier transform.}
Consider $L^2([0, 1]^d, \mathop{}\!\mathrm{d} x)$ the space of periodic functions in dimension $d$, square integrable against the Lebesgue measure on $[0, 1]^d$.
For simplicity, we will suppose that $\rho_\mathcal{X}$ is the Lebesgue measure on $[0, 1]^d$.
Consider a translation invariant kernel
\[
k(x, y) = q(x - y) \qquad \text{for } q:\mathbb{R}^d \to \mathbb{R} \text{ that is one periodic}.
\]
In this setting, the operator $K$, operating on $L^2$, is the convolution by $q$, that is
\[
\myfunction{K}{L^2}{L^2}{g}{q*g}, \qquad\text{hence}\qquad
\widehat{Kg} = \hat{q}\hat{g}.
\]
Where we have used the fact that convolutions can be represented by a product in the Fourier domain.
Note that, from B\"ochner theorem, we know that $k$ being positive definite implies that the Fourier transform of $q$ exists and is not negative.
Let us define the Fourier coefficient and the inverse Fourier transform as
\[
\forall\,\omega\in\mathbb{Z}^d, \quad
\hat{g}(\omega) = \int_{[0,1]^d} g(x)e^{-2i\pi \omega^\top x} \mathop{}\!\mathrm{d} x,
\quad\text{and}\quad
\forall\,x\in[0, 1]^d, \quad
g(x) = \sum_{\omega \in \mathbb{Z}^d} e^{2i\pi \omega^\top x} \hat{g}(\omega).
\]
$K$ being a convolution operator, it is diagonalizable with eigenelements $(\hat{q}(\omega), x\to e^{2i\pi \omega^\top x})_{\omega\in\mathbb{Z}^d}$.
From this, we can make explicit many of our abstract operators.
First of all, using Perceval's theorem,
\[
\norm{g}_{\cal H}^2 = \scap{g}{K^{-1}g}_{L^2}
= \sum_{\omega\in\mathbb{Z}^d} \frac{\abs{\hat{g}(\omega)}^2}{\hat{q}(\omega)}.
\]
Hence, we can parametrize ${\cal H}$ with $(\theta_\omega)_{\omega \in \mathbb{Z}^d} \in \mathbb{C}^{\mathbb{Z}^d}$ and the $\ell^2$-metric, where $\theta_\omega = \hat{g}(\omega) / \sqrt{\hat{q}(\omega)}$ and
\[
(S\theta)(x) = g_\theta(x) = \sum_{\omega\in\mathbb{Z}^d} e^{2i\pi\omega^\top x}\sqrt{\hat{q}(\omega)}\theta_\omega.
\]
Note that this is not the usual parametrization of ${\cal H}$ by elements $\theta\in{\cal H}$ as $(\mathbb{C}^{\mathbb{Z}^d}, \ell^2)$ is not a space of functions.
However, such a parametrization of ${\cal H}$ does not change any of the precedent algebraic considerations on the operators $S$, $\Sigma$, $K$, and $L$.
\paragraph{Diffusion operator and Fourier transform.}
As well as convolution operators are well represented in the Fourier domain, derivation operators are.
Indeed, when $g$ is regular, we have
\[
\norm{{\cal L}^{1/2} g}^2_{L^2} =
\norm{\nabla g}^2_{L^2} = \sum_{j=1}^d \norm{\partial_j g}^2_{L^2}
= \sum_{j=1}^d \sum_{\omega \in\mathbb{Z}^d} \omega_j^2 \abs{\hat{g}(\omega)}^2.
\]
As a consequence, using the expression of $S\theta$, we have
\[
\Sigma\theta = \sum_{\omega\in\mathbb{Z}^d} \hat{q}(\omega) \theta_\omega,
\qquad\text{while}\qquad
L\theta = \sum_{\omega\in\mathbb{Z}^d} \norm{\omega}_2^2 \hat{q}(\omega) \theta_\omega,
\quad\text{where}\quad \norm{\omega}_2^2 = \sum_{j=1}^d \omega_i^2.
\]
In this setting, the eigenelements of $\Sigma$ are $(\hat{q}(\omega), \delta_\omega)_{\omega\in\mathbb{Z}^d}$ while the one of $L$ are $(\norm{\omega}^2_2\hat{q}(\omega), \delta_\omega)_{\omega\in\mathbb{Z}^d}$.
\paragraph{Eigenvalue decay comparison.}
Hence, having $L \preceq c\Sigma^a$ is equivalent to having $\norm{\omega}_2^2\hat{q}(\omega) \leq c\hat{q}(\omega)^a$.
Now suppose that the decay of $\hat{q}$ is governed by
\[
c_1 (1 + \sigma^{-1}\norm{\omega}_2^2)^{-m} \leq \hat{q}(\omega)
\leq c_2 (1 + \sigma^{-1}\norm{\omega}_2^2)^{-m},
\]
for two constants $c_1, c_2 > 0$.
In particular, this is verified for Mat{\'e}rn kernels, corresponding to the fractional Sobolev space $H^m$, and for the Laplace kernel with $m = (d+1) / 2$, which reads $k(x, y) = \exp(-\sigma^{-1}\norm{x-y})$.
The Gaussian kernel could be seen as $m = +\infty$ as it has exponential decay.
With such a decay we have, assuming without restrictions that we are in one dimension
\begin{align*}
\omega^2\hat{q}(\omega) \leq c_2\omega^2 (1 + \sigma^{-1}\omega^2)^{-m}
\leq c_2\sigma(1 + \sigma^{-1}\omega^2)^{-(m-1)}
\leq c_1^{\frac{m}{m-1}}c_2\sigma \hat{q}(\omega)^{\frac{m-1}{m}}.
\end{align*}
In other terms, we can consider $c = c_1^{\frac{m}{m-1}}c_2 \sigma$ and $a = (m-1)/m$.
Assuming that $q$ is square-integrable, so is $\hat{q}$, which implies that $2m > d$.
As a consequence, we do have $a > 1 - 2 / d$.
Note that this reasoning could be extended to the case where $\rho_\mathcal{X}$ has a density against the Lebesgue measure, that is bounded above and below away from zero.
\section{Introduction}
In the last decade, machine learning has been able to tackle amazingly complex tasks, which was mainly allowed by computational power to train large learning models on large annotated datasets.
For instance, ImageNet is made of tens of millions of images, which have all been manually annotated by humans \citep{ImageNet}.
The greediness in data annotation of such a current learning paradigm is a major limitation.
In particular, when annotation of data demands in-depth expertise, relying on techniques that require zillions of labeled data is not viable.
This motivates several research streams to overcome the need for annotations, such as self-supervised learning for images or natural language processing \citep{Devlin2019}.
Aiming for generality, semi-supervised learning is the most classical one, assuming access to a vast amount of input data, but among which only a scarce percentage is labeled.
To leverage the presence of unlabeled data, most semi-supervised techniques assume a form of low-density separation hypothesis, as detailed in the recent review of \citet{Engelen2020}, and illustrated by state-of the-art models \citep{Berthelot2019,Verma2019}.
This hypothesis assumes that the function to learn from the data varies smoothly in highly populated regions of the input space, but might vary more strongly in sparsely populated areas, or that the decision frontiers between classes lie in regions with low-density.
In such a setting, it is natural to enforce constraints on the variations of the function to learn.
While semi-supervised learning is an important learning framework, it has not provided as many exciting realizations as one could have expected.
This might be related to the fact that it is classically approached through graph-based Laplacian, a technique that does not scale well with the dimension of the input space~\citep{Bengio2006}.
\paragraph{Paper organization.} In Section \ref{lap:sec:laplacian}, we motivate Laplacian regularization, and recall drawbacks of naive implementations.
These limitations are overcome in Section \ref{lap:sec:kernel} where we expose a theoretically principled path to derive well-behaved algorithms.
More precisely, we unveil a vast class of estimates based on spectral filtering.
We turn to implementation in Section~\ref{lap:sec:implementation} where we provide realistic guidelines to ensure scalability of the proposed algorithms.
Statistical properties of our estimators are stated in Section \ref{lap:sec:statistics}.
\paragraph{Contributions.} They are two folds.
({\em i}) Statistically, we explain that Laplacian regularization can be properly leveraged based on functional space considerations, and that those considerations can be turned into concrete implementations thanks to kernel methods.
As a result, we provide consistent estimators that exhibit fast convergence rates under a low density separation hypothesis, and that, in particular, do not suffer from the curse of dimensionality.
({\em ii}) Computationally, we avoid dealing with large matrices of derivatives by providing a low-rank approximation that allows dealing with $n^{\gamma}\log(n)\times n^{\gamma}\log(n)$ matrices, with a parameter $\gamma \in (0, 1]$ depending on the regularity of the problem, instead of $n(d+1)\times n(d+1)$ matrices, thus cutting down to ${\cal O}(\log(n)^2n^{1+2\gamma}d)$ the potential ${\cal O}(n^3d^3)$ training cost.
\paragraph{Related work.} Interplay between graph theory and machine learning were proven successful in the 2000s \citep{Smola2003}.
The seminal paper of \citet{Zhu2003} introduced graph-Laplacian as a transductive method in the context of semi-supervised learning.
A smoothing variant was proposed by \citep{Zhou2003}, which is coherent with the fact that enforcing constraints on labeled points leads to spikes \citep{Alaoui2016}.
Interestingly, graph Laplacian do converge to diffusion operators linked with the weighted Laplace Beltrami operator \citep{Hein2007,GarciaTrillos2019}.
However, these local diffusion methods are known to suffer from the curse of dimensionality \citep{Bengio2006}.
That is, local averaging methods are intuitive learning methods that have been used for more than half a century \citep{Fix1951}.
Yet, those methods do not scale well with the dimension of the input space \citep{Yang1999}.
This is related to the fact that to cover $[0,1]^d$, we need $\epsilon^{-d}$ balls of radius $\epsilon$.
Interestingly, if the function to learn is $m$ times differentiable with smooth partial derivatives, it is possible to leverage more information from function evaluations and overcome the curse of dimensionality when $m \gtrsim d$.
This property is related to covering numbers (a.k.a. capacity) of Sobolev spaces \citep{Kolmogorov1959}, and is leveraged by (reproducing) kernel methods \citep{Steinwart2008,Caponnetto2007}.
The crux of this paper is to apply this fact to Laplacian regularization techniques.
Note that derivatives with reproducing kernel methods in machine learning have already been considered in different settings by \citep{Zhou2008,Rosasco2013,Eriksson2018}.
\section{Laplacian regularization}
\label{lap:sec:laplacian}
In this section, we introduce the notations and concepts related to the semi-supervised learning regression problem, noting that most of our results extend to any convex loss beyond least-squares.
We motivate and describe Laplacian regularization that will allow us to leverage the low-density separation hypothesis.
We explain statistical drawbacks usually linked with Laplacian regularization, and discuss how to circumvent them.
In the following, we denote by $\mathcal{X} = \mathbb{R}^d$ the input space, $\mathcal{Y}=\mathbb{R}$ the output space, and by $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$ the joint distribution on $\mathcal{X}\times\mathcal{Y}$.
For simplicity, we assume that $\rho$ has compact support.
In the following, we denote by $\rho_\mathcal{X}$ the marginal of $\rho$ over $\mathcal{X}$, and by $\rho\vert_x$ the conditional distribution of $Y$ given $X=x$.
As usual, for $p \in \mathbb{N}^*$, $L^p(\mathbb{R}^d)$ is the space of functions $f$ such that $f^p$ is integrable.
Moreover, we define usual Sobolev spaces: for $s \in \mathbb{N}$, $W^{s,p}(\mathbb{R}^d)$ stands for the space of functions whose weak derivatives of order $s$-th are in $L^p(\mathbb{R}^d)$.
When $p=2$, they have a Hilbertian structure, and we denote, $H^{s}(\mathbb{R}^d) = W^{s,2}(\mathbb{R}^d)$ these Hilbertian spaces.
Ideally, we would like to retrieve the mapping $g^*:\mathcal{X}\to\mathcal{Y}$ defined as
\begin{equation}
\label{lap:eq:least_square}
g^* = \argmin_{g\in L^2(\rho_\mathcal{X})}\E_{(X, Y)\sim \rho}\bracket{\norm{g(X) - Y}^2}
= \argmin_{g\in L^2(\rho_\mathcal{X})}\norm{g - g_\rho}^2_{L^2(\rho_\mathcal{X})} = g_\rho,
\end{equation}
where $g_\rho:\mathcal{X}\to\mathcal{Y}$ is defined as $g_\rho(x) = \E\bracket{Y\,\middle\vert\, X=x}$.
In semi-supervised learning, we assume that we do not have access to $\rho$, but we have access to $n$ independent samples $(X_i)_{i\leq n} \sim\rho_\mathcal{X}^{\otimes n}$, among which we have $n_\ell$ labels $Y_i \sim \rho\vert_{X_i}$ for $i \leq n_\ell$, with $n_\ell$ potentially much smaller than $n$.
In other terms, we have $n_\ell$ supervised pairs $(X_i, Y_i)_{i\leq n_\ell}$, and $n-n_\ell$ unsupervised samples $(X_i)_{n_\ell < i \leq n}$.
While we restrict ourselves to real-valued regression for simplicity, our exposition indeed applies generically to partially supervised learning.
In particular, it can be used off-the-shelve to complement the approaches of \citep{Cabannes2020,Cabannes2021} as we detailed in Appendix \ref{lap:sec:extension}.
\subsection{Diffusion operator \texorpdfstring{${\cal L}$}{}}
In order to leverage unlabeled data, we will assume that $g^*$ varies smoothly on highly populated regions of $\mathcal{X}$, and might vary highly on low density regions.
For example, this is the case when data are clustered in well separated regions of space, and labels are constant on clusters.
This is captured by the fact that the Dirichlet energy
\begin{equation}
\label{lap:eq:diffusion_operator}
\int_{\mathcal{X}} \norm{\nabla g^*(x)}^2 \rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
= \E_{X\sim\rho_\mathcal{X}} \bracket{\norm{\nabla g^*(X)}^2}
=: \norm{{\cal L}^{1/2} g}_{L^2(\rho_\mathcal{X})}^2,
\end{equation}
is assumed to be small.
Because the quadratic functional \eqref{lap:eq:diffusion_operator} will play a crucial role in our exposition, we define ${\cal L}$ as the self-adjoint operator on $L^2(\rho_\mathcal{X})$, extending the operator on $H^1(\rho_\mathcal{X})$ representing this functional.
Under mild assumptions on $\rho_\mathcal{X}$, ${\cal L}^{-1}$ can be shown to be a compact operator, which we will assume in the following.
In essence, we will assume that if we have a lot of unlabeled data and $\norm{{\cal L}^{1/2} g}$ can be well approximated for any function $g$, then we do not need a lot of labeled data to estimate correctly $g^*$.
To illustrate this, at one extreme, if we know that $\norm{{\cal L}^{1/2} g^*} = 0$, then $g^*$ is known to be constant on each connected component of $\rho_\mathcal{X}$ so that, along with the knowledge of $\rho_\mathcal{X}$, only a few labeled points would be sufficient to recover perfectly $g^*$.
We illustrate those considerations on Figure \ref{lap:fig:intro}.
\begin{figure*}[t]
\centering
\includegraphics{laplacian/images/setting.pdf}
\includegraphics{laplacian/images/krr_0_1.pdf}
\includegraphics{laplacian/images/krr_1_n.pdf}
\vskip -0.2in
\caption{
Motivating example.
(Left) We suppose given $n = 2000$ points in $\mathcal{X}=\mathbb{R}^2$, represented as black dots, spanning $4$ concentric circles.
Among those points are $n_\ell = 4$ labeled points, with labels being either $1$ represented in red, and $-1$ represented in blue.
In this setting, it is natural to assume that $g^*$ should be constant on each circle, which can be encoded as $\norm{\nabla g^*} = 0$ on $\supp\rho_\mathcal{X}$.
(Middle) Kernel ridge regression estimate based on the labeled points with Gaussian kernel of bandwidth $\sigma = .2r$, $r$ being the radius of the innermost circle.
(Right) Laplacian regularization reconstruction.
The reconstruction is based on approximate empirical risk minimization with $p=n$, which ensures a computational complexity of $O(p^2 nd)$, instead of $O(n^3d^3)$ needed to recover the exact empirical risk minimizer \eqref{lap:eq:estimate}.
}
\label{lap:fig:intro}
\vskip -0.2in
\end{figure*}
\subsection{Drawbacks of naive Laplacian regularization}
Following the motivations presented previously, it is natural to consider the regularized objective and solution defined, for $\lambda > 0$, as
\begin{equation}
\label{lap:eq:laplacian_tikhonov}
\begin{split}
g_\lambda
&= \argmin_{g\in H^1(\rho_\mathcal{X})} \E_{(X,Y)\sim\rho}\bracket{\norm{g(X) - Y}^2} + \lambda
\E_{X\sim\rho_\mathcal{X}}\bracket{\norm{\nabla g(X)}^2_{\mathbb{R}^d}}
\\&= \argmin_{g\in H^1(\rho_\mathcal{X})} \norm{g - g_\rho}^2_{L^2(\rho_\mathcal{X})} + \lambda
\norm{{\cal L}^{1/2} g}^2_{L^2(\rho_\mathcal{X})} = (I+\lambda{\cal L})^{-1} g_\rho.
\end{split}
\end{equation}
This regularization has nice properties.
In particular, for small $\lambda$, it can be seen as a first order approximation of the heat equation solution $e^{-\lambda {\cal L}}g_\rho$, which represents the temperature profile at time $t=\lambda$, instantiated with the initial profile $g_\rho$, and with $\rho_\mathcal{X}$ modeling the thermal conductivity.
It also has interpretations in terms of random walk and Langevin diffusion \citep{PillaudVivien2020,Klus2020}.
In a word, $g_\lambda$ is the diffusion of $g_\rho$ with respect to the density $\rho_\mathcal{X}$, which relates to the idea of diffusing labeled data with respect to the intrinsic geometry of the data, which is the idea captured by \citep{Zhu2003}.
However, from a learning perspective, \eqref{lap:eq:laplacian_tikhonov}~is linked with the prior that $g^*$ belongs to $H^1(\rho_\mathcal{X})$, a prior that is not strong enough to overcome the curse of dimensionality as we saw in the related work section.
Moreover, assuming we have enough unsupervised data to suppose known $\rho_\mathcal{X}$, and therefore $\cal L$, \eqref{lap:eq:laplacian_tikhonov}~leads to the naive empirical estimate
\(
g_{\text{(naive)}} \in \argmin_{g:\mathcal{X}\to\mathbb{R}} \sum_{i=1}^{n_\ell} \norm{g(X_i) - Y_i}^2 + n_\ell\lambda \norm{{\cal L}^{1/2}g}^2.
\)
While the definition of $g_{\text{(naive)}}$ could seem like a great idea, in fact, such an estimate $g_{\text{(naive)}}$ is known to be mostly constant and spiking to interpolate the data $(X_i, Y_i)$ as soon as $d > 2$ \citep{Nadler2009}.
This is to be related with the capacity of the space associated with the pseudo-norm $\norm{{\cal L}^{1/2}g}$ in $L^2$.
This capacity, related to $H^1$, is too large for the Laplacian regularization term to constraint $g_{\text{(naive)}}$ in a meaningful way.
In other terms, we need to regularize with stronger penalties.
\subsection{Stronger regularization}
\label{lap:sec:kernel_free_reg}
In this subsection, we discuss techniques to overcome the issues encountered with $g_{\text{(naive)}}$.
Those techniques are based on functional space constraints or on spectral filtering techniques.
\paragraph{Functional spaces.} A solution to overcome the capacity issue of $H^1$ in $L^2$ is to constrain the estimate of $g^*$ to belong to a smaller functional space.
In the realm of graph Laplacian, \citep{Alaoui2016} proposed to solve this problem by considering the $r$-Laplacian regularization reading
\(
\Omega_{r} = \int_\mathcal{X} \norm{\nabla g(X)}^r \rho(\mathop{}\!\mathrm{d} x),
\)
with $r > d$.
In essence, this restricts $g$ to live in $W^{1,r}(\rho_\mathcal{X})$ for $r > d$, and allows avoiding spikes associated with $g_{\text{(naive)}}$.
However, considering high power of the gradient is likely to introduce instability (think that $d$ is the potentially colossal dimension of the input space), and from a learning perspective, the capacity of $W^{1,r}$, which compares to the one of $H^2$, is still too big.
In this paper, we will rather keep the diffusion operator ${\cal L}$, and add a second penalty to reduce the space in which we look for the solution.
With ${\cal G}$ a Hilbert space of functions, we could look for, with $\mu > 0$ a second regularization parameter
\begin{equation}
\label{lap:eq:tikhonov}
g_{\lambda, \mu} = \argmin_{g:{\cal G}\cap H^1(\rho_\mathcal{X})} \norm{g - g_\rho}^2_{L^2(\rho_\mathcal{X})}
+ \lambda \norm{{\cal L}^{1/2} g}^2_{L^2(\rho_\mathcal{X})} + \lambda\mu\norm{g}_{\cal G}^2.
\end{equation}
This formulation restricts $g_{\lambda, \mu}$ to belong both to ${H}^1(\rho_\mathcal{X})$ (thanks to the term in $\lambda$) and ${\cal G}$ (thanks to the term in $\mu$).
In particular the resulting space $H^1(\rho_\mathcal{X}) \cap {\cal G}$ to which $g_{\lambda, \mu}$ belongs, has a smaller capacity in $L^2$ than the one of ${\cal G}$ in $L^2$.
In practice, we do not have access to $\rho$ and $\rho_\mathcal{X}$ but to $(X_i, Y_i)_{i\leq n_\ell}$ and $(X_i)_{i\leq n}$, and we might consider the empirical estimator defined through empirical risk minimization
\begin{equation}
\label{lap:eq:estimate}
g_{n_\ell,n} = \argmin_{g\in{\cal G}} n_\ell^{-1}\sum_{i=1}^{n_\ell} \norm{g(X_i) - Y_i}^2
+ \lambda n^{-1} \sum_{i=1}^n\norm{\nabla g(X_i)}^2 + \lambda\mu\norm{g}^2_{\cal G}.
\end{equation}
For example, we could consider ${\cal G}$ to be the Sobolev space ${H}^m(\mathop{}\!\mathrm{d} x)$.
Note the difference between ${\cal G}$ linked with $\mathop{}\!\mathrm{d} x$, the Lebesgue measure, that is known, and ${\cal L}$ linked with $\rho_\mathcal{X}$, the marginal of $\rho$ over $\mathcal{X}$, that is not known.
In this setting, the regularization $\|{\cal L}^{1/2}g\|^2 + \mu\|g\|_{\cal G}^2$ reads $\int_\mathcal{X} \norm{D g(x)}^2 \rho_\mathcal{X}(\mathop{}\!\mathrm{d} x) + \mu \int_\mathcal{X} \sum_{\alpha=0}^m \norm{D^\alpha g(x)}^2 \mathop{}\!\mathrm{d} x$.
Because of the size of ${H}^m$ in $L^2$, this allows for efficient approximation of $g_{\lambda, \mu}$ based on empirical risk minimization.
In particular, if $n = +\infty$, we expect the minimizer \eqref{lap:eq:estimate} to converge toward $g_{\lambda, \mu}$ at rates in $L^2$ scaling similarly to $n_\ell^{-m/d}$ in $n_\ell$.
To complete the picture, depending on a prior on $g_\rho$, $g_{\lambda, \mu}$ might exhibit good convergence properties toward $g_\rho$ as $\lambda$ and $\mu$ go to zero.
This contrasts with the problem encountered with $g_{\text{(naive)}}$.
Those considerations are exactly what reproducing kernel Hilbert space will provide, additionally with a computationally friendly framework to perform the estimation.
Note that quantities similar to $g_{\lambda,\mu}$ were considered in \citep{Zhou2008,Rosasco2013}.
\paragraph*{Spectral filtering.}
Without looking for higher power-norm, \citep{Nadler2009} proposed to overcome the capacity issue by considering approximation of the operator ${\cal L}$ based on the graph-based technique provided by \citep{Belkin2003,Coifman2006} and to reduce the search of $g_{n_\ell}$ on the space spanned by the first few eigenvectors of the Laplacian.
In particular, on Figure \ref{lap:fig:intro}, $g^*$ could be searched in the null space of ${\cal L}$, that is, among functions that are constant on each connected component of $\supp\rho_\mathcal{X}$.
This technique exhibits two parts, the ``unsupervised'' estimation of ${\cal L}$ that will depend on the total number of data $n$, and the ``supervised'' search for $g_\rho$ on the first few eigenvectors of ${\cal L}$ that will depend on the number of labels~$n_\ell$.
While, at first sight, this technique seems to be completely different from Tikhonov regularization~\eqref{lap:eq:tikhonov}, it can be cast, along with gradient descent, into the same \emph{spectral filtering} framework \citep{Lin2020}.
This point of view enables the use of a wide range of techniques offered by spectral manipulations on the diffusion operator ${\cal L}$.
This paper is motivated by the fact that current well-grounded semi-supervised learning techniques are implemented based on graph-based Laplacian, which is a local averaging method that does not leverage smartly functional capacity.
In particular, as recalled earlier, graph-based Laplacian is known to suffer from the curse of dimensionality, in the sense that the convergence of the empirical estimator $\widehat{\cal L}$ toward the ${\cal L}$ exhibits a rate of convergence of order ${\cal O}(n^{-1/d})$ with $d$ the dimension of the input space ${\cal X}$ \citep{Hein2007}.
In this work, we will bypass this curse of dimensionality by looking for $g$ in a smooth universal reproducing kernel Hilbert space, which will lead to efficient empirical estimates.
\section{Spectral filtering with kernel Laplacian}
\label{lap:sec:kernel}
In this section, we approach Laplacian regularization from a functional analysis perspective.
We first introduce kernel methods and derivatives in reproducing kernel Hilbert space (RKHS).
We then translate the considerations provided in Section \ref{lap:sec:kernel_free_reg} in the realm of kernel methods.
\subsection{Kernel methods and derivatives evaluation maps}
In this subsection, we introduce kernel methods (see \citep{Aronszajn1950,Scholkopf2001,Steinwart2008} for more details).
Consider $({\cal H}, \scap{\cdot}{\cdot}_{\cal H})$ a reproducing kernel Hilbert space, that is a Hilbert space of functions from $\mathcal{X}$ to $\mathbb{R}$ such that the evaluation functionals $L_x:{\cal H}\to\mathbb{R};g\to g(x)$ are continuous linear forms for any $x\in\mathcal{X}$.
Such forms can be represented by $k_x \in {\cal H}$ such that, for any $g\in{\cal H}$, $L_x(g) = \scap{k_x}{g}_{\cal H}$.
A reproducing kernel Hilbert space can alternatively be defined from a symmetric positive semi-definite kernel $k:\mathcal{X}\to\mathcal{X}\to\mathbb{R}$, that is a function such that for any $n\in\mathbb{N}$ and $(x_i)_{i\leq n} \in \mathcal{X}^n$ the matrix $(k(x_i,x_j))_{i,j}$ is symmetric positive semi-definite, by building $(k_x)_{x\in\mathcal{X}}$ such that $k(x, x') = \scap{k_x}{k_{x'}}_{\cal H}$.
From a learning perspective, it is useful to use the evaluation maps to rewrite ${\cal H} = \brace{g_\theta:x\to\scap{k_x}{\theta}_{\cal H}\,\middle\vert\, \theta\in{\cal H}}$.
As such, kernel methods can be seen as ``linear models'' with features $k_x$, allowing to parametrize large spaces of functions \citep{Micchelli2006}.
In the following, we will differentiate $\theta$ seen as an element of ${\cal H}$ and $g_\theta$ seen as its embedding in $L^2$.
To make this distinction formal, we define the embedding $S:({\cal H}, \scap{\cdot}{\cdot}_{\cal H})\hookrightarrow (L^2(\rho_\mathcal{X}), \scap{\cdot}{\cdot}_{L2}); \theta\to g_\theta$, as well as its adjoint $S^\star:L^2(\rho_\mathcal{X})\to{\cal H}$.
Given a linear parametric model of functions $g_\theta(x) = \scap{\theta}{k_x}_{\cal H}$, it is possible to compute derivatives of $g_\theta$ based on derivatives of the feature vector -- think of ${\cal H} = \mathbb{R}^p$ and of $k_x = \phi(x)$ as a feature vector with $\phi:\mathbb{R}^d \to \mathbb{R}^p$.
For $\alpha \in \mathbb{N}^d$, with $\abs\alpha = \sum_{i\leq d} \alpha_i$, we have the following equality of partial derivatives, when $k$ is $2\abs{\alpha}$ times differentiable,
\[
D^\alpha g_\theta(x) = \scap{\theta}{D^\alpha k_x},\qquad\text{where}\qquad
D^\alpha = \frac{\partial^{\abs\alpha}}{(\partial x_1)^{\alpha_1} (\partial x_2)^{\alpha_2}\cdots (\partial x_d)^{\alpha_d}}.
\]
Here and $D^\alpha k_x$ has to be understood as the partial derivative of the mapping of $x\in\mathcal{X}$ to $k_x\in{\cal H}$, which can be shown to belong to ${\cal H}$ \citep{Zhou2008}.
In the following, we assume that $k$ is twice differentiable with continuous derivatives, and will make an extensive use of derivatives of the form $\partial_i k_x = \partial k_x / \partial x_i$ for $i \leq d$ and $x\in\mathcal{X}$.
Note that, as well as we can describe completely the Hilbertian geometry of the space $\Span\brace{k_x \,\middle\vert\, x\in\mathcal{X}}$ through $k(x, x') = \scap{k_x}{k_x'}$, for $x, x'\in\mathcal{X}$, we can describe the Hilbertian geometry of $\Span\brace{k_x \,\middle\vert\, x\in\mathcal{X}} + \Span\brace{\partial_i k_x\,\middle\vert\, x\in\mathcal{X}}$, through
\[
\partial_{1,i} k(x, x') = \scap{\partial_i k_x}{k_{x'}}_{\cal H},
\qquad\text{and}\qquad
\partial_{1,i}\partial_{2,j} k(x, x') = \scap{\partial_i k_x}{\partial_j
k_{x'}}_{\cal H},
\]
where $\partial_{1,i}$ denotes the partial derivative with respect to the $i$-th coordinates of the first variable.
This echoes the so-called ``representer theorems''.
\begin{example}[Gaussian kernel]
\label{lap:ex:rbf}
A classical kernel is the Gaussian kernel, also known as radial basis function, defined for $\sigma > 0$ as the following $k$, and satisfying, for $i\neq j$, the following equalities,
\begin{gather*}
k(x, x') = \exp\paren{-\frac{\norm{x-x'}^2}{2\sigma^2}},\qquad
\partial_{1,i}\partial_{2,j} k(x, y) = - \frac{(x_i - y_i)(x_j -
y_j)}{\sigma^4} k(x, y),
\\
\partial_{1,i} k(x, y) = -\frac{(x_i - y_i)}{\sigma^2} k(x, y),
\qquad
\partial_{1,i}\partial_{2,i} k(x, y) = \paren{\frac{1}{\sigma^2} -
\frac{(x_i - y_i)^2}{\sigma^4}} k(x, y),
\end{gather*}
where $x_i$ designs the $i$-th coordinates of the vector $x\in\mathcal{X}=\mathbb{R}^d$.
\end{example}
\subsection{Tikhonov, spectral filtering and dimensionality reduction}
Given the kernel $k$, its associated RKHS ${\cal H}$ and $S$ the embedding of ${\cal H}$ in $L^2$, we rewrite~\eqref{lap:eq:tikhonov} under its ``parametrized'' version
\begin{equation}
\tag{\ref{lap:eq:tikhonov}}
g_{\lambda, \mu} = S \, \argmin_{\theta\in{\cal H}} \brace{\norm{S\theta - g_\rho}^2_{L^2(\rho_\mathcal{X})}
+ \lambda \norm{{\cal L}^{1/2} S\theta}^2_{L^2(\rho_\mathcal{X})} + \lambda\mu\norm{\theta}_{\cal H}^2 }.
\end{equation}
Do not hesitate to refer to Table \ref{lap:tab:notations} to keep track of notations.
In the following, we will use that
\(
\norm{{\cal L}^{1/2} S\theta}^2_{L^2(\rho_\mathcal{X})} + \mu\norm{\theta}_{\cal
H}^2
= \norm{(S^\star {\cal L} S + \mu I)^{1/2} \theta}_{\cal H}^2.
\)
This equality explains why we consider $\mu\lambda$ instead of $\mu$ in the last term.
In the RKHS setting, the study of~\eqref{lap:eq:tikhonov} unveils the three operators $\Sigma$, $L$, and $I$ on ${\cal H}$, (indeed
\(
g_{\lambda, \mu} = S \, \argmin_{\theta\in{\cal H}} \brace{
\theta^\star(\Sigma + \lambda L + \lambda\mu)\theta - 2\theta^\star S^\star g_\rho }
\))
where $I$ is the identity, and, as we detail in Appendix \ref{lap:app:operators},
\begin{equation}
\label{lap:eq:block_op}
\Sigma = S^\star S = \E_{X\sim\rho_\mathcal{X}}\bracket{k_X \otimes k_X},
\qquad\text{and}\qquad
L = S^\star {\cal L}S = \E_{X\sim\rho_\mathcal{X}}
\bracket{\sum_{i=1}^d \partial_j k_X \otimes \partial_j k_X}.
\end{equation}
Regularization and spectral filtering have been well-studied in the inverse-problem literature.
In particular, the regularization~\eqref{lap:eq:tikhonov} is known to be linked with the generalized singular value decomposition of $[\Sigma; L+\mu I]$ (see, \emph{e.g.}, \cite{Edelman2020}), which is linked to the generalized eigenvalue decomposition of $(\Sigma, L+\mu I)$ \citep{Golub2013}.
We derive the following characterization of~\eqref{lap:eq:tikhonov}, whose proof is reported in Appendix \ref{lap:app:algebra}.
\begin{proposition}
\label{lap:thm:decomposition}
Let $(\lambda_{i,\mu})_{i\in\mathbb{N}} \in \mathbb{R}^\mathbb{N}, (\theta_{i,\mu})_{i\in\mathbb{N}} \in {\cal H}^\mathbb{N}$ be the generalized eigenvalue decomposition of the pair $(\Sigma, L + \mu I)$, that is $(\theta_{i,\mu})$ generating ${\cal H}$ and such that for any $i, j \in \mathbb{N}$, $\Sigma \theta_{i,\mu} = \lambda_{i,\mu} (L + \mu I) \theta_{i,\mu}$, and $\scap{\theta_{i,\mu}}{(L+\mu I) \theta_{j,\mu}} = \ind{i=j}$.
\eqref{lap:eq:tikhonov} can be rewritten as
\begin{equation}
\label{lap:eq:filtering}
g_{\lambda, \mu} =
\paren{\sum_{i\in\mathbb{N}} \psi(\lambda_{i,\mu}) S\theta_{i,\mu} \otimes S\theta_{i, \mu}} g_\rho =
\sum_{i\in\mathbb{N}} \psi(\lambda_{i,\mu}) \scap{S^\star g_\rho}{\theta_{i,\mu}} S\theta_{i,\mu},
\end{equation}
with $\psi:\mathbb{R}_+ \to \mathbb{R}; x\to (x+\lambda)^{-1}$.
\eqref{lap:eq:filtering}~should be seen as a specific instance of spectral filtering based on a filter function $\psi:\mathbb{R}_+\to\mathbb{R}$.
\end{proposition}
\begin{figure*}[t]
\centering
\includegraphics{laplacian/images/eigen1_f.pdf}
\includegraphics{laplacian/images/eigen4_f.pdf}
\includegraphics{laplacian/images/eigen8_f.pdf}
\vskip -0.2in
\caption{
Few of the first generalized eigenvectors of $(\hat\Sigma; \hat{L} + \mu I)$ (with $\mu = 1/n$).
The first four eigenvectors correspond to constant functions on each circle, as shown with $e_1$ and $e_4$.
The few eigenvectors after correspond to second harmonics localized on a single circle as shown with $e_8$.
}
\label{lap:fig:unsupervised}
\vskip -0.2in
\end{figure*}
Interestingly, the generalized eigenvalue decomposition of the pair $(\Sigma, L+\mu I)$ was already considered by \citet{PillaudVivien2020} to estimate the first eigenvalue of the Laplacian.
Moreover, \citet{PillaudVivien2020b} suggests leveraging this decomposition for dimensionality reduction based on the first eigenvectors of the Laplacian.
As well as \eqref{lap:eq:tikhonov}~contrasts with graph-based semi-supervised learning techniques, this dimensionality reduction technique contrasts with methods based on graph Laplacian provided by \citep{Belkin2003,Coifman2006}.
Remarkably, the semi-supervised learning algorithm that consists in using the unsupervised data to perform dimensionality reduction based on the Laplacian eigenvalue decomposition, before solving a small linear regression problem on the small resulting space, can be seen as a specific instance of spectral filtering, based on regularization by thresholding/cutting-off eigenvalue, which corresponds to $\psi:x\to x^{-1}\ind{x>\lambda}$ for a given threshold $\lambda > 0$ in~\eqref{lap:eq:filtering}.
\section{Implementation}
\label{lap:sec:implementation}
In this section, we discuss how to practically implement estimates for~\eqref{lap:eq:filtering} based on empirical data $(X_i, Y_i)_{i\leq n_\ell}$ and $(X_i)_{n_\ell < i \leq n}$.
We first review how we can approximate the integral operators of~\eqref{lap:eq:block_op} based on data.
We then discuss how to implement our methods practically on a computer.
We end this section by considering approximations that allow cutting down high computational costs associated with kernel methods involving derivatives.
\begin{algorithm}[H]
\caption{Empirical estimates based on spectral filtering.}
\KwData{$(X_i, Y_i)_{i\leq n_\ell}$, $(X_i)_{n_\ell < i\leq n}$, a kernel $k$, a filter $\psi$, a regularizer $\mu$}
\KwResult{$\hat{g}_p$ through $c \in \mathbb{R}^p$ defining $\hat{g}_p(x) = \sum_{i=1}^n c_i k(x, X_i) = k_x^\star T_ac$}
Compute $S_nT_a = (k(X_i, X_j))_{i\leq n, j\leq p} \in \mathbb{R}^{n\times p}$ in ${\cal O}(pn)$\\
Compute $Z_nT_a = (\partial_{1,j} k(X_l, X_i))_{(j\leq d, l\leq n), i\leq p} \in \mathbb{R}^{nd\times p}$ in ${\cal O}(pnd)$\\
Build $T_a^\star \hat\Sigma T_a = n^{-1}(S_nT_a)^\top(S_nT_a)$ in ${\cal O}(p^2n)$\\
Build $T_a^\star \hat{L} T_a = n^{-1}(Z_nT_a)^\top(Z_nT_a)$ in ${\cal
O}(p^2nd)$\footnote{
Building this matrix can be avoided by using the generalized singular value decomposition rather than the generalized eigenvector decomposition.
Implemented with Lapack, such a procedure will also require $O(p^2 nd)$ floating point operations, but with a smaller constant in the big $O$ \citep{Golub2013}.}\\
Build $T_a^\star T_a = (k(X_i, X_j))_{i,j\leq p} \in \mathbb{R}^{p\times p}$ in ${\cal O}(1)$ as a partial copy of $S_nT_a$\\
Get $(\lambda_{i, \mu}, u_{i,\mu})_{i\leq n}$ the generalized eigenelements of $(T_a^\star \hat\Sigma T_a, T_a^\star (\hat{L} + \mu I) T_a)$ in~${\cal O}(p^3)$\\
Get $b = T_a^\star \hat\theta = (n_\ell^{-1}\sum_{i=1}^{n_\ell} Y_i k(X_i, X_j))_{j\leq p} \in \mathbb{R}^p$ in ${\cal O}(pn_\ell)$\\
Return $c = \sum_{i=1}^n \psi(\lambda_i) u_i u_i^\top b \in \mathbb{R}^p$ in ${\cal O}(p^3)$.
\label{lap:alg:imp}
\end{algorithm}
\subsection{Integral operators' approximation}
The classical empirical risk minimization in~\eqref{lap:eq:estimate} can be understood as the plugging of the approximate distributions $\hat\rho = n_\ell^{-1}\sum_{i=1}^{n_\ell} \delta_{X_i} \otimes \delta_{Y_i}$ and $\hat\rho_\mathcal{X} = n^{-1} \sum_{i=1}^n \delta_{X_i}$ instead of $\rho$ and $\rho_\mathcal{X}$ in~\eqref{lap:eq:tikhonov}.
It can also be understood as the same replacement when dealing with integral operators, leading to the three following important quantities to rewrite~\eqref{lap:eq:filtering},
\begin{equation}
\label{lap:eq:approximate_operator}
\hat\Sigma := n^{-1}\sum_{i=1}^{n} k_{X_i}\otimes k_{X_i},\
\hat{L} := n^{-1}\sum_{i=1}^n \sum_{j=1}^d \partial_j k_{X_i}\otimes \partial_j k_{X_i},\
\hat\theta := \widehat{S^\star g_\rho} := n_\ell^{-1}\sum_{i=1}^{n_\ell} Y_i k_{X_i}.
\end{equation}
It should be noted that while considering $n$ in the definition of $\hat\Sigma$ is natural from the spectral filtering perspective, to make it formally equivalent with the empirical risk minimization \eqref{lap:eq:estimate}, it should be replaced by $n_\ell$.
\eqref{lap:eq:approximate_operator}~allows rewriting~\eqref{lap:eq:filtering} without relying on the knowledge of $\rho$, by considering $(\hat\lambda_{i,\mu}, \hat\theta_{i,\mu})$ the generalized eigenvalue decomposition of $(\hat\Sigma, \hat L)$ and considering
\begin{equation}
\hat g = \sum_{i\in\mathbb{N}} \psi(\hat\lambda_{i,\mu}) \scap{\widehat{S^\star g_\rho}}{\hat\theta_{i,\mu}} S\hat\theta_{i,\mu},
\end{equation}
We present the first eigenvectors (after plunging them in $L^2$ through $S$) of the generalized eigenvalue decomposition of $(\hat\Sigma, \hat{L} + \mu I)$ on Figure \ref{lap:fig:unsupervised}.
The first eigenvectors recover the null space of ${\cal L}$.
This explains clearly the behavior on the right of Figure \ref{lap:fig:intro}.
\subsection{Matrix representation and approximation of operators}
Currently, we are dealing with operators ($\hat\Sigma, \hat{L}$) and vectors (\emph{e.g.}, $\hat\theta$) in the Hilbert space ${\cal H}$.
It is natural to wonder how to represent this on a computer.
The answer is the object of representer theorems (see Theorem 1 of \citep{Zhou2008}), and consists in noticing that all the objects introduced are actually defined in, or operate on, ${\cal H}_n + {\cal H}_{n,\partial} \subset {\cal H}$, with ${\cal H}_n = \Span\brace{k_{X_i} \,\middle\vert\, i\leq n}$ and ${\cal H}_{n,\partial} = \Span\brace{\partial_j k_{X_i} \,\middle\vert\, i\leq n, j\leq d}$.
This subspace of ${\cal H}$ is of dimension at most $n(d+1)$ and if $T:\mathbb{R}^p \to {\cal H}_n + {\cal H}_{n,\partial}$ (with $p \leq n(d+1)$) parametrizes ${\cal H}_n + {\cal H}_{n,\partial}$, our problem can be cast in $\mathbb{R}^{p}$ by considering the $p\times p$ matrices $T^\star \hat\Sigma T$ and $T^\star (\hat{L} + \mu I)T$ instead of the operators $\hat\Sigma$ and $\hat{L} + \mu I$.
The canonical representation consists in taking $p = n(d+1)$ and considering for $c \in \mathbb{R}^{n(d+1)}$, the mapping $T_c c = \sum_{i=1}^n c_{i0} k_{X_i} + \sum_{j=1}^d c_{ij}\partial_j k_{X_i}$ \citep{Zhou2008,Rosasco2013}.
This exact implementation implies dealing and finding the generalized eigenvalue decomposition of $p\times p$ matrices with $p = n(d+1)$, which leads to computational costs in ${\cal O}(n^3d^3)$, which can be prohibitive.
Two solutions are known to cut down prohibitive computational costs of kernel methods.
Both methods consist in looking for a space that can be parametrized by $\mathbb{R}^p$ for a small $p$ and that approximates well the space ${\cal H}_n + {\cal H}_{n,\partial} \subset {\cal H}$.
The first solution is provided by random features \citep{Rahimi2007}.
It consists in approximating ${\cal H}$ with a space of small dimension $p\in\mathbb{N}$, linked with an explicit representation $\phi:\mathcal{X}\to\mathbb{R}^p$ that approximate $k(x, x') \simeq k_\phi(x, x') = \scap{\phi(x)}{\phi(x')}_{\mathbb{R}^p}$.
In theory, it substitutes the kernel $k$ by $k_\phi$.
In practice, all computations can be done with the explicit feature $\phi$.
\paragraph{Approximate solution.}
The second solution, which we are going to use in this work, consists in approximating ${\cal H}_n + {\cal H}_{n,\partial}$ by ${\cal H}_p = \Span\brace{k_{X_i}}_{i\leq p}$ for $p \leq n$.
This method echoes the celebrated Nystr\"om method \citep{Williams2000}, as well as the Rayleigh–Ritz method for Sturm–Liouville problems.
In essence, \citep{Rudi2015} shows that, when considering subsampling based on leverage score, $p=n^{\gamma}\log(n)$, with $\gamma \in (0,1]$ linked to the ``size'' of the RKHS and the regularity of the solution, is a good enough approximation, in the sense that it only downgrades the sample complexity by a constant factor.
In theory, we know that the space ${\cal H}_p$ will converge to ${\cal H} = \text{Closure}\Span\brace{k_x}_{x\in\supp\rho_\mathcal{X}}$ as $p$ goes to infinity.
In practice, it means considering the approximation mapping $T_a: \mathbb{R}^p \to {\cal H}; c \to \sum_{i=1}^p c_i k_{X_i}$, and dealing with the $p\times p$ matrices $T_a^\star \Sigma T_a$ and $T_a^\star L T_a$.
It should be noted that the computation of $T_a^\star L T_a$ requires to multiply a $p\times nd$ matrix by its transpose.
Overall, training this method can be done with ${\cal O}(p^2 n d)$ basic operations, and inference with this method can be done in ${\cal O}(p)$.
The saving cost of this approximate method is huge: without compromising the precision of our estimator, we went from $O(n^3 d^3)$ run time complexities to $O(\log(n)^2n^{1+2\gamma} d)$ computations, with $\gamma$ possibly very small.
Similarly, the memory cost went from $O(n^2 d^2)$ down to $O(nd + n^{2\gamma})$.\footnote{%
Our code is available online at \url{https://github.com/VivienCabannes/partial_labelling}.}
\section{Statistical analysis}
\label{lap:sec:statistics}
In this section, we are interested in quantifying the risk of the learned mapping $\hat{g}$.
We study it through the generalization bound, which consists in obtaining a bound on the averaged excess risk~$\E_{\textrm{data}}\|\hat{g} - g_\rho\|_{L^2}^2$.
In particular, we want to answer the following points.
\begin{enumerate}
\item How, and under which assumptions, Laplacian regularization boosts learning?
\item How the excess of risk relates to the number of labeled and unlabeled data?
\end{enumerate}
In terms of priors, we want to leverage a low-density separation hypothesis.
In particular, we can suppose that when diffusing $g_\rho$ with $e^{-t{\cal L}}$ we stay close to $g_\rho$, or that $g_\rho$ is supported on a finite dimensional space of functions on which $\norm{{\cal L}^{1/2}g}$ (which measures the variation of $g$) is small.
Both those assumptions can be made formal by assuming the $g_\rho$ is supported by the first eigenvectors of the diffusion operator ${\cal L}$.
\begin{assumption}[Source condition]
\label{lap:ass:source}
$g_\rho$ is supported on a finite dimensional space that is left stable by the diffusion operator ${\cal L}$.
In other terms, if $(e_i) \in (L^2)^\mathbb{N}$ are the eigenvectors of ${\cal L}$, there exists $r\in\mathbb{N}$, such that $g_\rho \in \Span\brace{e_i}_{i\leq r}$.
\end{assumption}
We will also assume that the diffusion operator ${\cal L}$ can be well approximated by the RKHS associated with $k$.
In practice, under mild assumptions, {\em c.f.} Appendix \ref{lap:app:operators}, the eigenvectors of the Laplacian are known to be regular, in particular to belong to $H^m$ for $m\in\mathbb{N}$ bigger than $d$.
As such, many classical kernels would verify the following assumption.
\begin{assumption}[Approximation condition]
\label{lap:ass:approximation}
The eigenvectors $(e_i)$ of ${\cal L}$ belong to the RKHS ${\cal H}$.
\end{assumption}
We add one technical assumptions regarding the eigenvalue decay of the operator $\Sigma$ compared to the operator $L$, with $\preceq$ denoting the L\"owner order ({\em i.e.}, for $A$ and $B$ symmetric, $A \preceq B$ if $B-A$ is positive semi-definite).
\begin{assumption}[Eigenvalue decay]
\label{lap:ass:decay}
There exists $a \in [0, 1]$ and $c > 0$ such that $L \preceq c \Sigma^\alpha$.
\end{assumption}
Note that, in our setting, $L$ is compact and bounded and Assumption \ref{lap:ass:decay} is always satisfied with $a = 0$.
For translation-invariant kernels, such as Gaussian or Laplace kernels, based on considerations linking eigenvalue decay of operators with functional space capacities \citep{Steinwart2008}, under mild assumptions, we can take $a > 1 - 2 / d$.
We discuss all assumptions in more detail in Appendix~\ref{lap:app:operators}.
To study the consistency of our algorithms, we can reuse the extensive literature on kernel ridge regression \citep{Caponnetto2007,Lin2020}.
This literature body provides an extensive picture on convergence rates relying on various filters and assumptions of capacity, a.k.a. effective dimension, and source conditions.
Our setting is slightly different and showcases two specificities:
({\em i}) the eigenelements $(\lambda_{i, \mu}, \theta_{i,\mu})$ are dependent of $\mu$;
({\em ii}) the low-rank approximation in Algorithm \ref{lap:alg:imp} is specific to settings with derivatives.
We end our exposition with the following convergence result, proven in Appendix \ref{lap:app:consistency}.
Note that the dependency of $p$ in $n$ can be improved based on subsampling techniques that leverage expressiveness of the different $(k_{X_i})$ \citep{Rudi2015}.
Moreover, universal consistency results could also be provided when the RKHS is dense in $H^1$, as well as convergence rates for other filters and laxer assumptions which we discuss in Appendix \ref{lap:app:consistency} (in particular, the source condition can be relaxed by considering the biggest $q\in(0, 1]$ such that $g\in\ima{\cal L}^q$).
\begin{figure}[t]
\centering
\includegraphics{laplacian/images/classification_error.pdf}
\includegraphics{laplacian/images/computation_time.pdf}
\vskip -0.1in
\caption{(Left) Comparison between our kernelized Laplacian method
(Tikhonov regularization version with $\lambda = 1$, $\mu_n = n^{-1}$, $p=50$) and graph-based Laplacian based on the same Gaussian kernel with bandwidth $\sigma_n = n^{-\frac{1}{d+4}}\log(n)$ as suggested by graph-based theoretical results \citep{Hein2007}.
We report classification error as a function of the number of samples $n$.
The error is averaged over 50 trials, with error bars representing standard deviations.
We fixed the ratio $n_\ell / n$ to one tenth, and generated the data according to two Gaussian in dimension $d=10$ with unit variance and whose centers are at distance $\delta = 3$ of each other (similar to the setting of \citep{Castelli1995,Lelarge2019}).
Our method discovers the structure of the data much faster than graph-based Laplacian (to get a 20\% error we need $40$ points, while graph-based needs $700$).
(Right) Time to perform training with graph-based Laplacian in orange, with Algorithm \ref{lap:alg:imp} in blue (with the specification of the left figure), and with the naive representation in $\mathbb{R}^{n(d+1)}$ of the empirical minimizer \eqref{lap:eq:estimate} in green.
When dealing with $1000$ points, our algorithm, as well as graph-based Laplacian, can be computed in about one tenth of a second on a 2 GHz processor, while the naive kernel implementation requires 10 seconds.
We show in Appendix \ref{lap:sec:experiment} that this cut in costs is not associated with a loss in performance.
}
\label{lap:fig:exp}
\vskip -0.15in
\end{figure}
\begin{theorem}[Convergence rates]
\label{lap:thm:consistency}
Under Assumptions \ref{lap:ass:source}, \ref{lap:ass:approximation} and \ref{lap:ass:decay}, for $n_\ell, n \in \mathbb{N}$, when considering the spectral filtering Algorithm \ref{lap:alg:imp} with $\psi_{\lambda}: x \to (x + \lambda)^{-1}$, there exists a constant $C$ independent of $n$, $n_\ell$, $\lambda$, $\mu$ and $p$ such that the estimate $\hat{g}_p$ defined in Algorithm \ref{lap:alg:imp} verifies
\begin{equation}
\E_{{\cal D}_n}\bracket{\norm{\hat{g}_p - g_\rho}_{L^2}^2} \leq
C \Big(\lambda^2 + \lambda\mu + \frac{\sigma^2_\ell n_\ell^{-1} + n_\ell^{-2} + n^{-1}}{\lambda\mu}
+ \frac{\log(p)}{p}
+ \lambda \frac{\log(p)^a}{p^a}\Big),
\end{equation}
with $\sigma_\ell^2$ is a variance parameter that relates to the variance of the variable $Y(I + \lambda {\cal L})^{-1}\delta_X$, inheriting its randomness from $(X, Y) \sim \rho$.
In particular, when the ratio $r = n_\ell / n$ is fixed, with the regularization scheme $\lambda_n = \lambda_0 n^{-1/4}$, $\mu_n = \mu_0 n^{-1/4}$, for any $\lambda_0 > 0$ and $\mu_0 > 0$, and the subsampling scheme $p_n = p_0 n^s \log(n)$ for any $p_0 > 0$ and with $s = \max(\sfrac{1}{2}, \sfrac{1}{4a})$, there exists a constant $C'$ independent of $n$ and $n_\ell$ such that the excess of risk verifies
\begin{equation}
\E_{{\cal D}_n}\bracket{\norm{\hat{g}_p - g_\rho}_{L^2}^2} \leq
C' (n^{-1/2} + \sigma_\ell^2 n_\ell^{-1/2}).
\end{equation}
\end{theorem}
Theorem \ref{lap:thm:consistency} answers the two questions asked at the beginning of this section.
In particular, it characterizes the dependency of the need for labeled data to a variance parameter linked with the diffusion of observations $(X_i, Y_i)$ based on the density $\rho_\mathcal{X}$ through the operator ${\cal L}$.
Intuitively, if the index set $I \subset \{1, 2, \cdots, n\}$ of data $(X_i)_{i\in I}$ we labeled does not change the profile of the diffusion solution $\hat{g}$, then we do not need that much labeled data -- as this is the case on Figure \ref{lap:fig:intro}.
Finally, Theorem~\ref{lap:thm:consistency} is remarkable in that it exhibits no dependency to the dimension of $\mathcal{X}$ in the power of $n$ and $n_\ell$.
This contrasts with graph-based Laplacian methods that do not scale well with the input space dimensionality \citep{Bengio2006,Hein2007}.
Indeed, Figure \ref{lap:fig:exp} shows the superiority of our method over graph-based Laplacian in dimension $d=10$, with a mixture of Gaussian.
We provide details as well as additional experiments in Appendix \ref{lap:sec:experiment}.
\section{Conclusion}
Diffusing information or enforcing regularity through penalties on derivatives are natural ideas to tackle many machine learning problems.
Those ideas can be captured informally with graph-based techniques and finite element differences, or captured more formally with the diffusion operator we introduced in this work.
This formalization allowed us to shed light on Laplacian regularization techniques based on statistical learning considerations.
In order to make our method usable in practice, we provided strong computational guidelines to cut down prohibitive costs associated with a naive implementation of our methods.
In particular, we were able to develop computationally efficient semi-supervised techniques that do not suffer from the curse of dimensionality.
This work paves the way to many extensions beyond semi-supervised learning.
For example, in Appendix \ref{lap:sec:extension}, we describe its usefulness to the partial supervised learning problem, where minimizing the Dirichlet energy provide a learning principle, in order to bypass the restrictive non-ambiguity assumption usually made in this setup \citep{Cour2011,Liu2014,Cabannes2020,Cabannes2021}.
Moreover, in the context of active learning, retaking the strategy of \citet{Karzand2020}, this energy provides a computationally-effective, theoretically-grounded, data-dependent score to select the next point to query.
As such, follow-ups would be of interest to see how this introductory theoretical paper makes its way into the world of concrete applications.
\chapter{Laplacian Regularization}
The following is a reproduction of \cite{Cabannes2021c}.
\input{laplacian/abstract}
\input{laplacian/core}
\begin{subappendices}
\chapter*{Appendix}
\addcontentsline{toc}{chapter}{Appendix}
\input{laplacian/appendix/extension}
\input{laplacian/appendix/experiments}
\input{laplacian/appendix/operators}
\input{laplacian/appendix/decomposition}
\input{laplacian/appendix/consistency}
\end{subappendices}
\part{Introduction}
\input{introduction/paradigm}
\input{introduction/discrete}
\input{introduction/partial}
\part{Considerations on Learning Theory}
\label{part:discrete}
\input{rates/main}
\input{svm/main}
\part{Learning with Partial Supervision}
\label{part:weakly}
\input{infimum/main}
\input{disambiguation/main}
\input{laplacian/main}
\part{Active Labeling}
\label{part:collection}
\input{sgd/main}
\chapter*{Conclusion}
\pagestyle{plain}
\addcontentsline{toc}{part}{Conclusion}
\input{introduction/conclusion}
\cleardoublepage
\phantomsection
\addcontentsline{toc}{chapter}{Bibliography}
\section{Kernel proofs}
\label{rates:app:rkhs}
In this section, we study $L^\infty$ convergence rates of the kernel ridge regression estimate.
We use the $L^2$-proof scheme of \citet{Caponnetto2007} with the remark of \citet{Ciliberto2016} to factorize the action of $K$ on $L^2(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}})$ through its action on $L^2(\mathcal{X}, \mathbb{R}, \rho_{\mathcal{X}})$.
We retake the work of \citet{PillaudVivien2018} to relax the source condition, and use \citet{Fischer2020} to cast in $L^\infty$ thanks to interpolation inequality.
While those results, leading to Lemma \ref{rates:lem:rkhs}, are not new, we present them entirely to provide the reader with self-contained materials.
\subsection{Construction of reproducing kernel Hilbert space (RKHS)}
\label{rates:app:rkhs-construction}
In the following, we suppose that $k$ is bounded by $\kappa^2$.
\paragraph{Vector-valued RKHS.}
To study the estimator $g_{n}$, it is useful to introduce the reproducing kernel Hilbert space ${\cal G}$ associated with $k$ and ${\cal H}$ \citep{Aronszajn1950}.
To define ${\cal G}$, define the atoms $k_{x}:{\cal H}\to{\cal G}$ and the scalar product, for $x, x' \in \mathcal{X}$ and $\xi, \xi' \in {\cal H}$ as
\[
\scap{k_{x}\xi}{k_{x'}\xi'}_{\cal G} = \scap{\xi}{\Gamma(x, x')\xi'} =
k(x, x')\scap{\xi}{\xi'}_{\cal H}.
\]
Where $\Gamma$ is the vector valued kernel inherited from $k$ as $\Gamma(x, x') = k(x, x') I_{\cal H}$ \citep{Schwartz1964}.
${\cal G}$ is defined as the closure, under the metric induced by this scalar product, of the span of the atoms $k_{x} \xi$ for $x\in\mathcal{X}$ and $\xi\in{\cal H}$.
Note that $k_{x}$ is linear, and continuous of norm $\norm{k_{x}}_{\op} = \sqrt{k(x, x)}$.
When $k(\cdot, x)$ is square integrable for all $x\in\supp\rho_{\mathcal{X}}$, ${\cal G}$ is homomorphic to a functional space in $L^2(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}})$ through the linear mapping $S$ that associates the atom $k_{x}\xi$ in ${\cal G}$ to the function $k(\cdot, x)\xi$ in $L^2$, defined formally as
\[
\myfunction{S}{\cal G}{L^2}{\gamma}{x\to k_{x}^{\star} \gamma.}
\]
While intrinsically similar, it is useful to distinguish between $\cal G$ and $\ima S \subset L^2$.
Note that $S$ is continuous, since on atom $k_x \xi$, $\norm{Sk_x\xi}_{L^2} \leq \norm{k_x(\cdot)}_{L^2} \norm{\xi}_{\cal H} \leq k(x,x)\norm{\xi}_{\cal H} = \norm{k_x\xi}_{\cal G}$.
The fact that $S$ is a bounded operator justifies the introduction of the following operators.
\paragraph{Central operators.}
In the following, we will make an extensive use of $S^{\star}:L^2(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}})\to{\cal G}$ the adjoint of $S$, defined as $S^{\star} g = \E_{\rho_{\mathcal{X}}}[k_{X}g(X)]$; the covariance operator $\Sigma:{\cal G}\to{\cal G}$, defined as $\Sigma := S^{\star} S = \E_{\rho_{\mathcal{X}}}[k_{X} k_{X}^{\star}]$; and its action on $L^2$, $K:L^2(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}}) \to L^2(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}})$, defined as $Kg := SS^{\star} g = \E_{X}[k(\cdot, X)g(X)]$.
Finally, we have defined the four central operators
\begin{equation}
\begin{array}{ll}
S\gamma = k_{(\cdot)}^{\star} \gamma, & S^{\star} g = \E_{\rho_{\mathcal{X}}}[k_{X}g(X)] \\
\Sigma := S^{\star} S = \E_{\rho_{\mathcal{X}}}[k_{X} k_{X}^{\star}],\qquad & Kg := SS^{\star} g = \E_{X}[k(\cdot, X)g(X)].
\end{array}
\end{equation}
It should be noted that this construction is usually avoided since, based on the fact that the Frobenius norm of $K$ behave like $\dim({\cal H})$, meaning that when $\cal H$ is infinite dimensional, $K$ is not a compact operator.
However, since we consider $\cal Z$ finite, we can always consider ${\cal H} = \mathbb{R}^{\card{\cal Z}}$ with $\phi(y) = (\ell(z, y)_{z\in\cal Z}$ and $\psi(z) = (\ind{z=z'})_{z'\in{\cal Z}}$, and moreover, we will see that a way can be worked out, even when $\cal H$ is infinite dimensional, which was already shown by \citet{Ciliberto2016}.
\subsubsection*{Relation between real-valued versus vector-valued RKHS.}
Usually convergence in RKHS is studied for real-valued functions.
We need convergence results for vector-valued functions.
As mentioned above, we only need the results for Euclidean space, however, we will do it for functions that are maps going into potentially infinite dimensional Hilbert space.
Indeed, this does not lead to major complications.
We provide here one way to get around this issue.
An alternative formal way to proceed can be found \citep{Ciliberto2016}.
\paragraph{Real-valued RKHS.}
We build the real-valued RKHS ${\cal G}_{\mathcal{X}}$ as the closure of the span of the atoms $\bar{k}_{x}$ for $x\in\mathcal{X}$, under the metric induced by the scalar product $\scap{\bar{k}_{x}}{\bar{k}_{x'}} = k(x, x')$.
Similarly, we build $\bar{S}$, $\bar{S}^{\star}$, $\bar\Sigma$ and $\bar{K}$.
We shall see that the action of $\Sigma$ on ${\cal G}$ can be factorized through its actions $\bar\Sigma$ on ${\cal G}_{\mathcal{X}}$.
\paragraph{Algebraic equivalences.}
Based on the fact that $\norm{\bar{k}_{x}}_{{\cal G}_{\mathcal{X}}} = \norm{k_{x}}_{\op} = \sqrt{k(x, x)}$, it is possible to build an isometry that match $\bar{k}_{x}$ in ${\cal G}_{\mathcal{X}}$ to $k_{x}$ in the space of continuous linear operator from ${\cal H}$ to ${\cal G}$.
With $(\bar{e}_{i})_{i\in \mathbb{N}}$ an orthogonal basis of ${\cal G}_{\mathcal{X}}$, and $(f_{j})_{j\in\mathbb{N}}$ an orthogonal basis of ${\cal H}$, we get an orthogonal basis $(e_{i}f_{j})_{i,j\in\mathbb{N}}$ of ${\cal G}$.
This is exactly the construction ${\cal G} = {\cal G}_{\mathcal{X}} \otimes {\cal H}$ of \citep{Ciliberto2016}.
Note that for $\mu_1, \mu_2$ two measures on $\mathcal{X}$, we can check that
\begin{align*}
\norm{\E_{X\sim\mu_1}[k_{X}k_{X}^{\star}]\E_{X_0\sim\mu_2}[k_{X_0}]}_{\op}^2
& = \norm{\E_{X\sim\mu_1}[\bar{k}_{X}\bar{k}_{X}^{\star}]\E_{X_0\sim\mu_2}[\bar{k}_{X_0}]}_{{\cal G}_{\mathcal{X}}}^2
\\&=\E_{X, X'\sim\mu_1; X, X'\sim\mu_2} [k(X_0, X)k(X, X')k(X', X'_0)].
\end{align*}
This explains that we will allow ourselves to write derivations of the type
\[
\norm{(\Sigma+\lambda)^{-\frac{1}{2}} k_{x} g_{n}(x)}_{\cal G} \leq
\norm{(\bar\Sigma+\lambda)^{-\frac{1}{2}} \bar{k}_{x}}_{{\cal G}_{\mathcal{X}}}
\norm{g_{n}(x)}_{\cal H}.
\]
Note also that for $g := \sum_{ij} c_{ij} e_{i}f_{j} \in {\cal G}$, with $\sum_{ij}c_{ij}^2 = 1$, $\bar{c}_{i} := (c_{ij})_{j\in\mathbb{N}} \in \ell^2$, $\bar{A}$ a self-adjoint operator on ${\cal G}_{\mathcal{X}}$ and $A$ its version on ${\cal G}$, we have
\begin{align*}
\norm{Ag}^2_{{\cal G}}
& = \sum_{ijk} c_{ij} c_{kj} \scap{\bar{A}\bar{e}_{i}}{\bar{A}\bar{e}_k}_{\cal G}
= \sum_{ij} \scap{\bar{c}_{i}}{\bar{c}_{j}}_{\ell^2} \scap{\bar{A}\bar{e}_{i}}{\bar{A}\bar{e}_k}_{{\cal G}_{\mathcal{X}}}
\\& \leq \sum_{ij}\norm{\bar{c}_{i}}_{\ell^2}\norm{\bar{c}_{j}}_{\ell^2} \scap{\bar{A}\bar{e}_{i}}{\bar{A}\bar{e}_k}_{{\cal G}_{\mathcal{X}}}
= \norm{\bar{A} \sum_{i} \norm{\bar{c}_{i}}_{\ell^2} \bar{e}_{i}}^2_{{\cal G}_{\mathcal{X}}}
\leq \norm{\bar{A}}_{\op}^2,
\end{align*}
which explains why we will consider derivations of the type
\[
\norm{(\Sigma+\lambda)^{\frac{1}{2}}(\hat\Sigma+\lambda)^{-1}(\Sigma+\lambda)^{\frac{1}{2}}}_{\op}
\leq \norm{(\bar\Sigma+\lambda)^{\frac{1}{2}}(\hat{\bar\Sigma}+\lambda)^{-1}(\bar\Sigma+\lambda)^{\frac{1}{2}}}_{\op}.
\]
Finally, notice that because of the same consideration, if $(\bar{u}_i)_{i\in\mathbb{N}} \in {\cal G}_\mathcal{X}^\mathbb{N}$ diagonalize $\bar{A}$, $(u_if_j)_{i,j\leq \mathbb{N}}\in{\cal G}^{\mathbb{N}\times\mathbb{N}}\simeq{\cal G}^\mathbb{N}$ diagonalize $A$ in ${\cal G}$.
This justifies the consideration of fractional operators $A^p$ for $p \in \mathbb{R}_+$, such as in Assumptions \ref{rates:ass:interpolation} and \ref{rates:ass:source}.
Based on this equivalence, we will forget the bar notations, we incite the careful and attentive reader to recover them.
\subsection{Estimate \texorpdfstring{$g_{n}$}{} as an empirical approximate projection on RKHS}
To obtain bounds like~\eqref{rates:eq:concentration}, it is sufficient to control the convergence of $g_{n}$ to $g^{*}$ in $L^\infty$.
Assumption \ref{rates:ass:interpolation} allow us to cast in $L^2$ the study of the convergence in $L^\infty$.
The convergence of $g_{n}$ toward $g^{*}$ can be split in two terms, a term expressing the convergence of $g_\lambda$ toward $g^{*}$ that is based on geometrical properties and a term expressing the convergence of $g_{n}$ toward $g_\lambda$, that is based on concentration inequalities in ${\cal G}$, such as the ones given by \citet{Pinelis1986,Minsker2017}.
For this last term, we need to characterize $g_{n}$ and $g_\lambda$ with the following lemma.
\begin{lemma}[Approximation of integral operators]
$g_{n}$ can be understood as the empirical approximation of $g_\lambda$ since
\[
g_{n} = S(\E_{\hat\rho}[k_{X} k_{X}^{\star}] + \lambda)^{-1}
\E_{\hat\rho}[k_{X}\phi(Y)],
\qquad
g_\lambda = S(\E_{\rho}[k_{X} k_{X}^{\star}] + \lambda)^{-1}
\E_{\rho}[k_{X}\phi(Y)],
\]
with $\hat\rho = \frac{1}{n} \sum_{i=1}^{n} \delta_{X_{i}} \otimes \delta_{Y_{i}}$,
\end{lemma}
\begin{proof}
Indeed, the expression of $g_{n}$ and its convergences toward $g^{*}$ will be understood thanks to the operator $S$ and its derivatives.
When $\ima S$ is closed in $L^2$, on can be defined the orthogonal projection of $g^{*}$ to $\ima S$, with the $L^2$ metric as $\pi_{\ima S}(g^{*}) = S(S^{\star} S)^\dagger S^{\star} g^{*}$.
When $\ima S$ is not closed, or equivalently when $\Sigma$ has positive eigenvalues converging to zero, one can define approximate orthogonal projection, through eigenvalue thresholding or Tikhonov regularization.
This last choice leads to the estimate
\[
g_\lambda = S(\Sigma + \lambda)^{-1}S^{\star} g^{*} = S (S^{\star} S + \lambda)^{-1} S^{\star} g^{*}
= SS^{\star} (SS^{\star} +\lambda)^{-1}g^{*} = K (K + \lambda)^{-1}g^{*}.
\]
Note that, because of the Bayes optimum characterization of $g^{*}$, $S^{\star} g^{*} = \E_{\rho}[k_{X} \phi(Y)]$.
This explains the characterization of $g_\lambda$.
Interestingly, the approximation of $\rho$ by $\hat\rho$ can be thought with the approximation of $L^2(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}})$ by $\ell^2({\cal H}^{n}) \simeq L^2(\mathcal{X}, {\cal H}, \hat\rho_{\mathcal{X}})$ where for $\Xi = (\xi_{i}), Z = (\zeta_{i}) \in {\cal H}^{n}$,
\[
\scap{\Xi}{Z}_{\ell^2} = \frac{1}{n}\sum_{i=1}^{n} \scap{\xi_{i}}{\zeta_{i}}_{\cal H},
\]
and with the empirical probability measure $\hat\rho = \frac{1}{n}\sum_{i=1}^{n} \delta_{x_{i}}\otimes \delta_{y_{i}}$.
We redefine the natural homomorphism of $\cal G$ into $\ell^2$ with
\[
\myfunction{\hat{S}}{\cal G}{\ell^2}{\gamma}{\paren{k_{x_{i}}^{\star} \gamma}_{i\leq n}.}
\]
We check that its adjoint is, for $\Xi\in{\cal H}^{n}$ and $\gamma\in{\cal G}$
\[
\scap{\hat{S}^{\star} \Xi}{\gamma}_{\cal G} = \scap{\Xi}{\hat S \gamma}_{\ell^2}
= \frac{1}{n} \sum_{i=1}^{n}\scap{\xi_{i}}{k_{x_{i}}^{\star}\gamma}_{\cal H}
= \scap{\frac{1}{n} \sum_{i=1}^{n} k_{x_{i}}\xi_{i}}{\gamma}_{\cal G}.
\]
Similarly, we define $\hat K:{\cal H}^{n}\to{\cal H}^{n}$ and $\hat\Sigma:{\cal G}\to{\cal G}$, with
\[
\hat K\Xi = \hat S \hat S^{\star} \Xi = \paren{\frac{1}{n} \sum_{i=1}^{n} k(x_{j}, x_{i})
\xi_{i}}_{j\leq n},\qquad
\hat\Sigma = \frac{1}{n}\sum_{i=1}^{n} k_{x_{i}} \otimes k_{x_{i}} = \E_{\hat\rho_{\mathcal{X}}}[k_{X}k_{X}^{\star}].
\]
Finally, we define $\hat{\Phi} = (\phi(y)_{i})_{i\leq n} \in {\cal H}^{n}$, so that
\[
\widehat{S^{\star} g^{*}} := \E_{\hat\rho}[\phi(Y)\cdot k_{X}] = \hat S^{\star} \hat\Phi.
\]
Finally, we can express $g_{n}$ as
\[
g_{n} = S(\hat\Sigma + \lambda)^{-1}\hat S^{\star} \hat\Phi
= S(\hat S^{\star} \hat S + \lambda)^{-1}\hat S^{\star} \hat\Phi
= S\hat S^{\star} (\hat S\hat S^{\star} + \lambda)^{-1} \hat\Phi
= S\hat S^{\star} (\hat K + \lambda)^{-1} \hat\Phi.
\]
This explains the equivalence between $g_{n}$ defined at the beginning of Section \ref{rates:sec:rkhs} and the $g_{n}$ expressed in the lemma, that will be used for derivations of theorems.
\end{proof}
\subsection{Linear algebra and equivalent assumptions to Assumptions \ref{rates:ass:capacity},
\ref{rates:ass:interpolation}}
To proceed with the study of the convergence of $g_{n}$ toward $g_\lambda$ in $L^2$, it is helpful to pass by ${\cal G}$.
To do so, we need to express Assumptions \ref{rates:ass:capacity} and \ref{rates:ass:interpolation} in ${\cal G}$, which we can do using the following linear algebra property.
\begin{lemma}[Linear algebra on compact operators]
There exist $(u_{i})_{i\in\mathbb{N}}$ an orthogonal basis of ${\cal G}_{\mathcal{X}}$, $(v_{i})_{i\in\mathbb{N}}$ an orthogonal basis of $L^2(\mathcal{X}, \mathbb{R}, \rho_{\mathcal{X}})$, and $(\lambda_{i})_{i\in \mathbb{N}}$ a decreasing sequence of positive real number such that
\begin{align}
S = \sum_{i\in\mathbb{N}} \lambda_{i}^{1/2}u_{i}v_{i}^{\star},\qquad
S^{\star} = \sum_{i\in\mathbb{N}} \lambda_{i}^{1/2}v_{i}u_{i}^{\star},\qquad
\Sigma = \sum_{i\in\mathbb{N}} \lambda_{i} u_{i}u_{i}^{\star},\qquad
K = \sum_{i\in\mathbb{N}} \lambda_{i}v_{i}v_{i}^{\star},
\end{align}
where the convergence of series has to be understood with the operator norms.
Moreover, we have that, if the kernel $k$ is bounded by $\kappa^2$,
\[
\sum_{i\in\mathbb{N}} \lambda_{i} \leq \kappa^2 < +\infty.
\]
Therefore, $K$ and $\Sigma$ are trace-class, and $S$ and $S^{\star}$ are Hilbert-Schmidt.
\end{lemma}
\begin{proof}
Notice that $\Sigma = \E_{X}[k_{X}\otimes k_{X}]$ and that $\norm{k_{x}\otimes k_{x}}_{\op({\cal G}_{\mathcal{X}})} = \norm{k_{x}}_{{\cal G}_{\mathcal{X}}} = k(x, x) \leq \kappa^2$.
Therefore, $\Sigma$ is a nuclear operator, so it is trace class and so it is compact.
The first point results from diagonalization of kernel operators, known as Mercer's Theorem \citep{Mercer1909,Steinwart2012}.
$\Sigma$ is a compact operator, therefore, the Spectral Theorem gives the existence of a sequence $(\lambda_{i}) \in \mathbb{R}^{\mathbb{N}}$ and an orthonormal basis $(u_{i}) \in {\cal G}_{\mathcal{X}}^{\mathbb{N}}$ of ${\cal G}_{\mathcal{X}}$ such that
\[
\Sigma = \sum_{i\in\mathbb{N}} \lambda_{i} u_{i} u_{i}^{\star},
\]
where the convergence has to be understood with the operator norm.
Because $\Sigma$ is of the form $S^{\star} S$, one can consider $(\lambda_{i})$ a decreasing sequence of positive eigenvalues.
Then, by defining, for all $i\in\mathbb{N}$ with $\lambda_{i} > 0$,
\[
v_{i} = \lambda_{i}^{-1/2} Su_{i}
\]
we check that $(v_{i})$ are orthonormal, and we complete them to form an orthonormal basis of $(L^2(\mathcal{X}, \mathbb{R}, \rho_{\mathcal{X}}))$.
Finally, we check that
\[
S = \sum_{i\in\mathbb{N}} \lambda_{i}^{1/2} v_{i}u_{i}^{\star},
\]
and that the other equalities hold too.
To check the second assertion, we use that $k_{x}k_{x}^{\star}$ is rank one when operating on ${\cal G}_{\mathcal{X}}$ and therefore
\begin{align*}
\trace{\Sigma} & = \trace\paren{\E_{X}[k_{X}k_{X}^{\star}]} = \E_{X}\bracket{\trace\paren{k_{X}k_{X}^{\star}}}
= \E_{X}\bracket{\norm{k_{X}k_{X}^{\star}}_{\op({\cal G}_{\mathcal{X}})}}
\\&= \E_{X}\bracket{\norm{k_{X}}_{{\cal G}_{\mathcal{X}}}} = \E_{X}[k(x, x)] \leq \kappa^2.
\end{align*}
This shows that $S$ and $S^{\star}$ are Hilbert-Schmidt operators and that $K$ is also trace class.
\end{proof}
This allows us to cast in ${\cal G}_{\mathcal{X}}$ the assumptions expressed in $L^2$.
\begin{lemma}[Equivalence of capacity condition]
For $\sigma \in (0, 1]$, it is equivalent to suppose that
\begin{itemize}
\item $\trace_{L^2(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}})}(K^\sigma) < +\infty$.
\item $\trace_{{\cal G}_{\mathcal{X}}}(\Sigma^\sigma) < +\infty$.
\item $\sum_{i\in\mathbb{N}} \lambda_{i}^\sigma < +\infty$.
\end{itemize}
\end{lemma}
In Assumption \ref{rates:ass:capacity}, the smaller $\sigma$, the faster the $\lambda_{i}$ decreases, the easier it will be to approximate $\Sigma$ based on approximation of $\rho$.
This appears explicitly in Theorem \ref{rates:thm:matrix}.
Indeed, for $\sigma=0$, the condition should be defined as $\Sigma$ of finite rank.
Note that when $k$ is bounded, we know that $\Sigma$ is trace class, and therefore, Assumption \ref{rates:ass:capacity} holds with $\sigma = 1$.
\begin{lemma}[Interpolation inequality in RKHS]
\label{rates:lem:interpolation}
Assumption \ref{rates:ass:interpolation} implies that
\begin{equation}
\forall\ \gamma\in{\cal G},\qquad
\norm{S\gamma}_{L^\infty} \leq c_p \norm{\Sigma^{\frac{1}{2}-p}\gamma}_{\cal G}.
\end{equation}
\end{lemma}
\begin{proof}
We begin by showing the property for $\gamma \in {\cal G}_{\mathcal{X}}$.
When $\gamma = \sum_{i\in\mathbb{N}} c_{i} v_{i}$ with $\sum_{i\in\mathbb{N}} c_{i}^2 < +\infty$, denote $g = \sum_{i\in\mathbb{N}} \lambda_{i}^{\frac{1}{2}-p}c_{i}u_{i}$, we have $g\in L^2$, therefore, using Assumption \ref{rates:ass:interpolation},
\begin{align*}
\norm{S\gamma}_{L^\infty}
= \norm{K^pg}_{L^\infty}
\leq c_p\norm{g}_{L^2}
= c_p\norm{\Sigma^{\frac{1}{2}-p}\gamma}_{{\cal G}_{\mathcal{X}}}.
\end{align*}
This ends the proof for ${\cal G}_{\mathcal{X}}$.
Note also that when the result of the Lemma holds, then Assumption \ref{rates:ass:interpolation} holds for any $g\in\ima_{L^2(\mathcal{X}, \mathbb{R},\rho_{\mathcal{X}})} K^{\frac{1}{2}-p}$.
Let us switch to ${\cal G}$ now.
Let $\gamma \in {\cal G}$, and denote $g = S\gamma$.
Suppose that $g$ achieve it maximum in $x_\infty$, define the direction $\xi = \sfrac{g(x_\infty)}{\norm{g(x_\infty)}_{\cal H}}$, and define $g_{\xi}: x\to \scap{g(x)}{\xi}_{\cal H} \in L^2(\mathcal{X}, \mathbb{R}, \rho_{\mathcal{X}})$, and $\gamma_{\xi} = \sum_{j\in\mathbb{N}} \scap{g_{\xi}}{v_{i}}_{L^2} u_{i} \in {\cal G}_{\mathcal{X}}$.
We have
\[
\norm{S\gamma}_{L^\infty} = \norm{S\gamma_{\xi}}_{L^\infty} \leq
c_p\norm{\Sigma^{\frac{1}{2}-p}\gamma_{\xi}}_{{\cal G}_{\mathcal{X}}}
\leq c_p\norm{\Sigma^{\frac{1}{2}-p}\gamma}_{{\cal G}}.
\]
When $g$ does not achieve its maximum, one can do a similar reasoning by considering a basis $(f_{i})_{i\in\mathbb{N}}$ of ${\cal H}$ and decomposition $\gamma$ on the basis $(u_{i}f_{j})_{i,j\in\mathbb{N}}$, before summing the directions.
\end{proof}
In Assumption \ref{rates:ass:interpolation}, the bigger $\sfrac{1}{2}-p$ the more we are able to control our problem in ${\cal G}$, the better.
Note that this reformulation of the interpolation inequality allows generalizing it for $p$ smaller than zero.
Note that when $k$ is bounded,
\(
\norm{(S\gamma)(x)}_{\cal H} = \norm{k_{X}^{\star} \gamma}_{\cal H}
\leq \norm{k_{X}}_{\op} \norm{\gamma}_{\cal G} = \sqrt{k(x, x)}\norm{\gamma}_{\cal G},
\)
hence Assumption \ref{rates:ass:interpolation} holds with $p=\sfrac{1}{2}$.
\subsection{Linear algebra with atoms \texorpdfstring{$k_{x}$}{} and useful inequalities}
From the study of the convergence of $g_{n}$ to $g_\lambda$ will emerge two quantities linked to eigenvalues of $\Sigma$ and the position of $k_{x}$ regarding eigenspaces, that are
\begin{equation}\label{rates:eq:eigen-quantity}
{\cal N}(\lambda) = \trace\paren{(\Sigma+\lambda)^{-1}\Sigma},\qquad
{\cal N}_\infty(\lambda) = \sup_{x\in\supp\rho_{\mathcal{X}}}\norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}.
\end{equation}
While those quantities could be bounded with brute force consideration, Assumptions \ref{rates:ass:capacity} and \ref{rates:ass:interpolation} will help to control them more subtly.
\begin{proposition}[Characterization of capacity condition]
The property $\sum_{i\in\mathbb{N}}\lambda_{i}^\sigma < +\infty$, can be rephrased in terms of eigenvalues of $\Sigma$ as the existence of $a_1 > 0$ such that, for all $i > 0$,
\begin{equation}
\lambda_{i} \leq a_1 (i+1)^{-\frac{1}{\sigma}}.
\end{equation}
\end{proposition}
\begin{proof}
Denote by $u_{i}$ and $S_{n}$ the respective quantities $\lambda_{i}^{\sigma}$ and $\sum_{i=1}^{n} u_{i}$.
Because $S_{n}$ converge, it is a Cauchy sequence, so there exists $N$ such that for any $p > q > N$, \( S_p - S_{q} = \sum_{i=q+1}^p u_{i} \leq 1.\)
In particular, considering $p=2q$, and because $(\lambda_{i})$ is decreasing, we have \( q u_{2q} \leq \sum_{i=q+1}^{2q} u_{i} \leq 1.\)
Therefore, we have that for all $i > 2N$, $u_{i} \leq 3(i+1)^{-1}$, considering $(a_1)^{\sigma} = 3 + \max_{i\leq 2N}\brace{(i+1)u_{i}}$, we get the desired result.
\end{proof}
\begin{proposition}[Characterization of ${\cal N}$]
When $\trace\paren{K^{\sigma}} < +\infty$, with $a_2 = \int_0^\infty \frac{a_1}{a_1 + t^{\frac{1}{\sigma}}}\mathop{}\!\mathrm{d} t$,
\begin{equation}
\forall\ \lambda > 0, \qquad {\cal N}(\lambda, r) \leq a_2 \lambda^{-\sigma}.
\end{equation}
\end{proposition}
\begin{proof}
Expressed with eigenvalues, we have
\[
{\cal N}(\lambda) = \trace\paren{(\Sigma + \lambda)^{-1}\Sigma} =
\sum_{i\in\mathbb{N}} \frac{\lambda_{i}}{\lambda_{i} + \lambda}.
\]
Using that $\lambda_{i} \leq a_1(i+1)^{-\frac{1}{\sigma}}$, that $x\to\frac{x}{x+a}$ is increasing with respect to $x$ for any $a> 0$ and the series-integral comparison, we get for $\sigma \in (0, 1]$
\begin{align*}
{\cal N}(\lambda) & \leq \sum_{i\in\mathbb{N}} \frac{a_1(i+1)^{-\frac{1}{\sigma}}}{a_1(i+1)^{-\sigma} + \lambda}
\leq \int_0^\infty \frac{a_1t^{-\frac{1}{\sigma}}}{a_1t^{-\frac{1}{\sigma}} + \lambda}\mathop{}\!\mathrm{d} t
= \int_0^\infty \frac{a_1}{a_1 + \lambda t^{\frac{1}{\sigma}}}\mathop{}\!\mathrm{d} t
\\&= \lambda^{-\sigma}\int_0^\infty \frac{a_1}{a_1 + (\lambda^\sigma t)^{\frac{1}{\sigma}}}\mathop{}\!\mathrm{d} (\lambda^\sigma t)
= a_2\lambda^{-\sigma},
\end{align*}
where we check the convergence of the integral.
\end{proof}
Indeed, Assumption \ref{rates:ass:interpolation} has a profound linear algebra meaning, it is a condition on $\rho_{\mathcal{X}}$-almost all the vector $k_{x} \in {\cal G}_{\mathcal{X}}$ not to be excessively supported on the eigenvector corresponding to small eigenvalue of $\Sigma$.
\begin{proposition}[Characterization of interpolation condition]
The interpolation Assumption \ref{rates:ass:interpolation} implies that, for all $i\in\mathbb{N}$
\begin{equation}
\sup_{x\in\rho_{\mathcal{X}}} \abs{\scap{k_{x}}{u_{i}}_{{\cal G}_{\mathcal{X}}}} \leq c_p\lambda_{i}^{\frac{1}{2}-p}.
\end{equation}
\end{proposition}
\begin{proof}
Consider the decomposition of $k_{x} \in {\cal G}_{\mathcal{X}}$ according to the eigenvectors of $\Sigma$, with $a_{i}(x) = \scap{k_{x}}{u_{i}}$.
The interpolation condition Assumption \ref{rates:ass:interpolation}, expressed in ${\cal G}_{\mathcal{X}}$ with Lemma \ref{rates:lem:interpolation}, leads to for any $\gamma_{\mathcal{X}} \in {\cal G}_{\mathcal{X}}$, and $S\gamma_{\mathcal{X}} :\mathcal{X}\to\mathbb{R}$,
\begin{align*}
\abs{(S\gamma_{\mathcal{X}})(x)}
= \abs{\scap{k_{x}}{\gamma_{\mathcal{X}}}_{{\cal G}_{\mathcal{X}}}}
\leq \norm{S\gamma_{\mathcal{X}}}_{L^\infty}
\leq c_p \norm{\Sigma^{\frac{1}{2}-p} \gamma_{\mathcal{X}}}_{{\cal G}_{\mathcal{X}}}
\end{align*}
This implies that
\[
\scap{k_{x}}{\gamma_{\mathcal{X}}}_{{\cal G}_{\mathcal{X}}}^2 = \paren{\sum_{i\in\mathbb{N}}
\scap{k_{x}}{u_{i}}\scap{\gamma_{\mathcal{X}}}{u_{i}}}^2 \leq c_p^2
\norm{\Sigma^{\frac{1}{2}-p} \gamma_{\mathcal{X}}}_{{\cal G}_{\mathcal{X}}}^2 = c_p^2
\sum_{i\in\mathbb{N}} \lambda_{i}^{1-2p} \scap{\gamma_{\mathcal{X}}}{u_{i}}^2.
\]
Taking $\gamma_{\mathcal{X}} = u_{i}$, we get that
\[
\abs{\scap{k_{x}}{u_{i}}} \leq c_p\lambda_{i}^{\frac{1}{2}-p}.
\]
This result relates the interpolation condition to the fact that $k_{x}$ is not excessively supported on the eigenvectors corresponding to vanishing eigenvalues of $\Sigma$.
\end{proof}
\begin{proposition}[Characterization of ${\cal N}_{\infty}(\lambda, r)$]
Under the interpolation condition, Assumption \ref{rates:ass:interpolation}, we have with $a_3 = c_p (2p)^{-p} (1-2p)^{\frac{1}{2}-p}$, or $a_3=c_p$ when $p=\sfrac{1}{2}$,
\begin{equation}
{\cal N}_{\infty}(\lambda) \leq a_3\lambda^{-p}.
\end{equation}
\end{proposition}
\begin{proof}
First of all, notice that
\begin{align*}
\norm{(\Sigma + \lambda)^{-\frac{1}{2}}k_{x}}_{{\cal G}_{\mathcal{X}}}
& = \sup_{\norm{\gamma_{\mathcal{X}}}_{\cal \mathcal{X}} = 1} \scap{\gamma_{\mathcal{X}}}{(\Sigma + \lambda)^{-\frac{1}{2}}k_{x}}_{{\cal G}_{\mathcal{X}}}
= \sup_{c; \sum_{i\in\mathbb{N}} c_{i}^2 = 1} \sum_{i\in\mathbb{N}} \frac{c_{i} \scap{k_{x}}{u_{i}}}{(\lambda + \lambda_{i})^{\frac{1}{2}}}
\\& \leq c_p \sup_{c;\sum_{i\in\mathbb{N} c_{i}^2=1}} \sum_{i\in\mathbb{N}} \frac{c_{i}\lambda_{i}^{\frac{1}{2}-p}}{(\lambda + \lambda_{i})^{\frac{1}{2}}}
\leq \sup_{t\in\mathbb{R}_+} c_p \frac{t^{\frac{1}{2}-p}}{(\lambda + t)^{\frac{1}{2}}}.
\end{align*}
When $p\in(0, \sfrac{1}{2})$, this last function is zero in zero and in infinity, therefore its maximum $t_0$ verifies, taking the derivative of its logarithm,
\[
\frac{\sfrac{1}{2}-p}{t_0} = \frac{1}{2(t_0+\lambda)} \quad\Rightarrow\quad
t_0 = \frac{(1-2p)\lambda}{2p} \quad\Rightarrow\quad
\sup_{t\in\mathbb{R}_+}\frac{t^{\frac{1}{2}-p}}{(\lambda + t)^{\frac{1}{2}}} =
(2p)^{-p} (1-2p)^{\frac{1}{2}-p}\lambda^{-p}.
\]
The cases $p\in\brace{0, 1}$ are easy to treat.
\end{proof}
In the previous analysis, one fact does not appear, it is that $\Sigma$ and $k_{x}$ are linked to one another, since $\Sigma = \E_{X}[k_{X}k_{X}^{\star}]$.
The following remark builds on it to relate $\cal N$ and ${\cal N}_\infty$.
\begin{remark}[Relation between interpolation and capacity condition]
The capacity and interpolation condition are related by the fact that it is unreasonable not to consider that
\(
p \leq \sfrac{\sigma}{2}.
\)
\end{remark}
\begin{proof}
Because $k_{x}k_{x}^{\star}$ is of rank one in ${\cal G}_{\mathcal{X}}$, we have
\begin{align*}
{\cal N}(\lambda) & = \trace\paren{(\Sigma + \lambda)^{-1}\Sigma}
= \E_{X}\bracket{\trace\paren{(\Sigma + \lambda)^{-1}k_{X}k_{X}^{\star}}}
= \E_{X}\bracket{\trace\paren{k_{X}^{\star}(\Sigma + \lambda)^{-1}k_{X}}}
\\&= \E_{X}\bracket{\norm{k_{X}^{\star}(\Sigma + \lambda)^{-1}k_{X}}_{\op}}
= \E_{X}\bracket{\norm{(\Sigma + \lambda)^{-\frac{1}{2}}k_{X}}^2_{{\cal G}_{\mathcal{X}}}}.
\end{align*}
So indeed, ${\cal N}(\lambda)$ is the expectation of the square $\norm{(\Sigma + \lambda)^{-\frac{1}{2}}k_{X}}_{{\cal G}_{\mathcal{X}}}$, when ${\cal N}_{\infty}(\lambda)$ is the supremum of this last quantity.
Therefore,
\[
{\cal N}(\lambda) \leq {\cal N}_{\infty}(\lambda)^2
\]
Supposing that the dependencies in $\lambda$ proved above are tight, we should have \( \sigma \geq 2p, \) which is the statement of this remark.
We refer the reader to Lemma 6.2 of \citet{Fischer2020} for more consideration to relates $\sigma$ and $p$ (reading $p$ and $\sfrac{\alpha}{2}$ with their notations)
\end{proof}
\subsection{Geometrical control of the residual \texorpdfstring{$\norm{g_\lambda - g^{*}}_{L^\infty}$}{}}
\label{rates:proof:krr-1}
The proof of the first assertion in Lemma \ref{rates:lem:rkhs} follows from, using Assumption \ref{rates:ass:source}, with $g_0 \in K^{-q}g^{*}$,
\begin{align*}
g_\lambda - g^{*} & = (K(K + \lambda)^{-1} - I)g^{*} = -\lambda (K+\lambda)^{-1} g^{*} = -\lambda (K + \lambda)^{-1}K^{q}g_0
\\&= -\lambda K^{p}(K + \lambda)^{-1}K^{q-p}g_0.
\end{align*}
Then using Assumption \ref{rates:ass:interpolation},
\begin{align*}
\norm{g^{*} - g_\lambda}_{\infty} & \leq c_p \lambda\norm{K^{q-p}(K + \lambda)^{-1}}_{\op}\norm{g_0}_{L^2}
\\&\leq c_p\lambda \norm{K(K + \lambda)^{-1}}^{q-p}_{\op} \norm{(K + \lambda)^{-1}}^{1+p-q}_{\op}\norm{g_0}_{L^2}
\\&\leq c_p\lambda 1^{q-p} \lambda^{-(1+p-q)}\norm{g_0}_{L^2}
= b_1\lambda^{q-p},
\end{align*}
where we have used that $\norm{K(K + \lambda)^{-1}}_{\op} = \sfrac{\norm{K}_{\op}}{(\norm{K}_{\op} + \lambda)} \leq 1$ and that $\norm{(K + \lambda)^{-1}} \leq \lambda^{-1}$.
\subsection{Convergence of \texorpdfstring{$\norm{g_{n} - g_\lambda}$}{} through concentration inequality}
\label{rates:proof:krr-2}
For the proof of the second assertion in Lemma \ref{rates:lem:rkhs}, we will put ourselves in ${\cal G}$.
For this, we define in ${\cal G}$
\begin{equation}
\gamma = \E_\rho[k_{X}\phi(Y)],\qquad
\gamma_\lambda = (\Sigma + \lambda)^{-1}\gamma,\qquad
\hat\gamma = \E_{\hat\rho}[k_{X}\phi(Y)],
\end{equation}
so that $g_\lambda = S\gamma_\lambda$, and $g_{n} = S(\hat\Sigma + \lambda)^{-1}\hat\gamma$.
\subsubsection{Decomposition into a matrix and a vector term}
We begin by expressing $g_{n} - g_\lambda$ in ${\cal G}$ with
\begin{align*}
g_{n} - g_\lambda
& = S\paren{(\hat\Sigma + \lambda)^{-1} \hat\gamma - (\Sigma + \lambda)^{-1} \gamma}
\\&= S\paren{(\hat\Sigma + \lambda)^{-1} (\hat\gamma - \gamma) + ((\hat\Sigma + \lambda)^{-1} - (\Sigma + \lambda)^{-1}) \gamma)}
\\&= S\paren{(\hat\Sigma + \lambda)^{-1} (\hat\gamma - \gamma) + (\hat\Sigma + \lambda)^{-1}(\Sigma - \hat\Sigma)(\Sigma + \lambda)^{-1} \gamma)}
\\&= S\paren{(\hat\Sigma + \lambda)^{-1} ((\hat\gamma - \hat\Sigma \gamma_\lambda) - (\gamma - \Sigma \gamma_\lambda))},
\end{align*}
where we have used that $A^{-1} - B^{-1} = A^{-1}(B - A)B^{-1}$.
Therefore, using the expression, Lemma \ref{rates:lem:interpolation}, of Assumption \ref{rates:ass:interpolation} in ${\cal G}$, we get
\begin{align*}
\norm{g_{n} - g_\lambda}_{L^\infty}
& \leq c_p\norm{\Sigma^{\frac{1}{2}-p} (\hat\Sigma + \lambda)^{-1} (\Sigma + \lambda)^{\frac{1}{2}+p}}_{\op} \times \cdots \\
& \qquad\qquad \norm{(\Sigma + \lambda)^{-(\frac{1}{2}+p)}((\hat\gamma - \hat\Sigma \gamma_\lambda) - (\gamma - \Sigma \gamma_\lambda))}_{\cal G}.
\end{align*}
On the one hand, we have concentration of matrix term toward $\Sigma^{\frac{1}{2}-p} (\Sigma + \lambda)^{-\paren{\frac{1}{2}-p}} \preceq I$.
On the other hand, we have concentration of the vector $\hat\gamma - \hat\Sigma \gamma_\lambda$ toward $\gamma - \Sigma \gamma_\lambda$.
Indeed, the concentration of the matrix term is hard to prove (it is only a conjecture), therefore we will go for another decomposition, that will result in similar rates when $p \geq 0$, that is
\begin{equation}
\begin{split}
&\norm{g_{n} - g_\lambda}_{L^\infty}
\leq c_p\norm{\Sigma^{\frac{1}{2}-p} (\Sigma + \lambda)^{-\frac{1}{2}}}_{\op} {\cal A}(\lambda){\cal B}(\lambda)\\
&{\cal A}(\lambda) = \norm{(\Sigma + \lambda)^{\frac{1}{2}}(\hat\Sigma + \lambda)^{-1} (\Sigma + \lambda)^{\frac{1}{2}}}_{\op}, \\
&{\cal B}(\lambda) = \norm{(\Sigma + \lambda)^{-\frac{1}{2}}((\hat\gamma - \hat\Sigma \gamma_\lambda) - (\gamma - \Sigma \gamma_\lambda))}_{\cal G}.
\end{split}
\end{equation}
Recall the definition of the following important quantity that are going to pop up from the analysis
\begin{equation}\tag{\ref{rates:eq:eigen-quantity}}
{\cal N}(\lambda) = \trace\paren{(\Sigma+\lambda)^{-1}\Sigma},\qquad
{\cal N}_\infty(\lambda) = \sup_{x\in\supp\rho_{\mathcal{X}}}\norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}.
\end{equation}
\subsubsection{Extra matrix term}
We control the extra matrix term with
\[
\norm{\Sigma^{\frac{1}{2}-p} (\Sigma + \lambda)^{-\frac{1}{2}}}_{\op}
= \norm{\Sigma^{\frac{1}{2}-p} (\Sigma +
\lambda)^{-\paren{\frac{1}{2}-p}}}_{\op} \norm{(\Sigma + \lambda)^{-p}}_{\op}
\leq \lambda^{-p}.
\]
Using that $\norm{(\Sigma + \lambda)^{-1}}_{\op} \leq \lambda^{-1}$ and that $\norm{(\Sigma + \lambda)^{-1}\Sigma}_{\op} \leq \sfrac{\norm{\Sigma}_{\op}}{(\norm{\Sigma}_{\op} + \lambda)} \leq 1$.
\subsubsection{Matrix concentration}
Let us make explicit the concentration in the matrix term with
\begin{align*}
(\Sigma + \lambda)^{\frac{1}{2}} (\hat\Sigma + \lambda)^{-1} (\Sigma + \lambda)^{\frac{1}{2}}
& = I +
(\Sigma + \lambda)^{\frac{1}{2}} \paren{(\hat\Sigma + \lambda)^{-1} - (\Sigma + \lambda)^{-1}} (\Sigma + \lambda)^{\frac{1}{2}}
\\&= I +
(\Sigma + \lambda)^{\frac{1}{2}} (\hat\Sigma + \lambda)^{-1} \paren{\Sigma - \hat\Sigma} (\Sigma + \lambda)^{-1} (\Sigma + \lambda)^{\frac{1}{2}}.
\end{align*}
From here, notice the following implications (that are actually equivalence)
\begin{align*}
\Sigma - \hat\Sigma \preceq t (\Sigma + \lambda)
& \quad\Rightarrow\quad \hat\Sigma + \lambda \succeq (1 - t)(\Sigma + \lambda)
\\&\quad\Rightarrow\quad (\hat\Sigma + \lambda)^{-1} \preceq (1 - t)^{-1}(\Sigma + \lambda)^{-1}.
\\&\quad\Rightarrow\quad (\hat\Sigma + \lambda)^{-1} - (\Sigma + \lambda)^{-1} \preceq t(1 - t)^{-1}(\Sigma + \lambda)^{-1}.
\\&\quad\Rightarrow\quad (\Sigma + \lambda)^{\frac{1}{2}}
\paren{(\hat\Sigma + \lambda)^{-1} - (\Sigma + \lambda)^{-1}} (\Sigma + \lambda)^{\frac{1}{2}}\preceq t(1-t)^{-1}
\\&\quad\Rightarrow\quad (\Sigma + \lambda)^{\frac{1}{2}}
(\hat\Sigma + \lambda)^{-1} (\Sigma + \lambda)^{\frac{1}{2}} \preceq (1-t)^{-1}.
\end{align*}
The probability of the event $\Sigma - \hat\Sigma \preceq t (\Sigma + \lambda)$, can be studied through the probability of the event $(\Sigma + \lambda)^{-\frac{1}{2}}(\Sigma - \hat\Sigma)(\Sigma + \lambda)^{-\frac{1}{2}} \preceq t$, which can be studied through concentration of self adjoint operators.
Finally, we have shown that
\begin{equation}
\norm{(\Sigma + \lambda)^{-\frac{1}{2}}(\Sigma - \hat\Sigma)(\Sigma + \lambda)^{-\frac{1}{2}}}_{\op} \leq t
\quad\Rightarrow\quad {\cal A}(\lambda)
\leq \frac{1}{1-t}.
\end{equation}
The best result that we are aware of, for covariance matrix inequality, is the extension to self-adjoint Hilbert-Schmidt operators provided by \citet{Minsker2017} in Section 3.2 of its concentration inequality on random matrices Theorem 3.1.
It can be formulated as the following.
\begin{theorem}[Concentration of self-adjoint operators \citep{Minsker2017}]
\label{rates:thm:matrix}
Let denote by $(\xi_{i})_{i\leq n}$ a sequence of independent self-adjoint operator acting on a separable Hilbert space ${\cal A}$, such that $\ker(\E[\xi_{i}]) = {\cal A}$, that are bounded by a constant $M \in \mathbb{R}$, in the sense $\norm{\xi_{i}}_{\op} \leq M$, with finite variance $\sigma^2 = \norm{\E\sum_{i=1}^{n} \xi_{i}^2}_{\op}$.
For any $t>0$ such that $6 t^2 \geq (\sigma^2 + \sfrac{Mt}{3})$,
\[
\Pbb\paren{\norm{\sum_{i=1}^{n}\xi_{i}}_{\op} > t} \leq 14\
r\paren{\sum_{i=1}^{n} \E\xi_{i}^2}
\exp\paren{-\frac{t^2}{2\sigma^2 + 2t M / 3}},
\]
with $r(\xi) = \sfrac{\trace{\xi}}{\norm{\xi}_{\op}}$.
\end{theorem}
Let us define $\xi$ that goes from $\mathcal{X}$ to the space of self-adjoint operator action on ${\cal G}_{\mathcal{X}}$ as
\begin{equation}
\xi(x) = (\Sigma + \lambda)^{-\frac{1}{2}} k_{x}k_{x}^{\star} (\Sigma + \lambda)^{-\frac{1}{2}}.
\end{equation}
We have that $(\Sigma + \lambda)^{-\frac{1}{2}}(\Sigma - \hat\Sigma)(\Sigma + \lambda)^{-\frac{1}{2}} = \E_\rho[\xi(X)] - \frac{1}{n}\sum_{i=1}^{n} \xi(x_{i})$.
To apply operator concentration, we need to bound $\xi$ and its variance.
\paragraph{Bound on $\xi$.}
To bound $\xi$ we proceed with, because $k_{x}k_{x}^{\star}$ is of rank one,
\begin{align*}
\norm{\xi(x)}_{\op} & = \norm{ (\Sigma + \lambda)^{-\frac{1}{2}} k_{x}k_{x}^{\star} (\Sigma + \lambda)^{-\frac{1}{2}}}_{\op}
= \trace\paren{ (\Sigma + \lambda)^{-\frac{1}{2}} k_{x}k_{x}^{\star} (\Sigma + \lambda)^{-\frac{1}{2}}}
\\&= \trace\paren{ k_{x}^{\star} (\Sigma + \lambda)^{-1} k_{x}}
= \norm{(\Sigma + \lambda)^{-\frac{1}{2}}k_{x}}_{{\cal G}_{\mathcal{X}}}^2
\leq {\cal N}_{\infty}(\lambda)^2.
\end{align*}
\paragraph{Variance of $\xi$.}
For the variance of $\xi$ we proceed by noticing that
\begin{align*}
\E\xi(X) & = \E_{X}(\Sigma + \lambda)^{-\frac{1}{2}} k_{X}k_{X}^{\star} (\Sigma + \lambda)^{-\frac{1}{2}}
= (\Sigma + \lambda)^{-\frac{1}{2}} \E_{X}\bracket{k_{X}k_{X}^{\star}} (\Sigma + \lambda)^{-\frac{1}{2}}
\\&= (\Sigma + \lambda)^{-\frac{1}{2}} \Sigma (\Sigma + \lambda)^{-\frac{1}{2}}
= (\Sigma + \lambda)^{-1} \Sigma.
\end{align*}
Hence,
\begin{align*}
\E\xi(X)^2 \preceq \sup_{x\in\mathcal{X}} \norm{\xi(x)}_{\op} \E[\xi(X)] \preceq {\cal N}_\infty(\lambda)^2(\Sigma + \lambda)^{-1} \Sigma.
\end{align*}
And as a consequence
\begin{equation*}
\norm{\E\xi(x)^2} \leq {\cal N}_\infty(\lambda)^2,
\end{equation*}
where we have used that $\norm{(\Sigma + \lambda)^{-1} \Sigma}_{\op} = \sfrac{\norm{\Sigma}_{\op}}{(\norm{\Sigma}_{\op} + \lambda)} \leq 1$.
\paragraph{Concentration bound on $\xi$.}
Using the self-adjoint concentration theorem, we get for any $t>0$, such that $6nt^2 \geq {\cal N}_{\infty}(\lambda)^2 (1 + \sfrac{t}{3})$,
\begin{align*}
\Pbb_{{\cal D}_{n}}\paren{\norm{\E_{\hat\rho}[\xi] - \E_\rho[\xi]}_{\op} > t}
\leq 14\, \frac{\norm{\Sigma}_{\op} + \lambda}{\norm{\Sigma}_{\op}}
{\cal N}(\lambda)
\exp\paren{-\frac{nt^2}{2{\cal N}_{\infty}(\lambda)^2 (1 + t/3)}}.
\end{align*}
Therefore, using the contraposition of the prior implication, we get
\begin{equation}
\Pbb_{{\cal D}_{n}}\paren{{\cal A}(\lambda) > \frac{1}{1-t}} \leq
14\, \frac{\norm{\Sigma}_{\op} + \lambda}{\norm{\Sigma}_{\op}}
{\cal N}(\lambda)
\exp\paren{-\frac{nt^2}{2{\cal N}_{\infty}(\lambda)^2 (1 + t/3)}}.
\end{equation}
\subsubsection{Decomposition of vector term in a variance and a bias term}
Let switch to the vector term, consider $\xi:\mathcal{X}\times\mathcal{Y}\to{\cal G}$, defined as
\begin{equation*}
\xi = (\Sigma + \lambda)^{-\frac{1}{2}}k_{x}(\phi(y) - k_{x}^{\star} \gamma_\lambda).
\end{equation*}
It allows expressing in simple form the vector term as
\[
{\cal B}(\lambda) =
\norm{\frac{1}{n}\sum_{i=1}^{n} \xi(X_{i}, Y_{i}) - \E_{(X, Y)\sim\rho}[\xi(X, Y)]}.
\]
We can study this term through concentration inequality in ${\cal G}$.
To proceed we will dissociate the variability due to $Y$ to the one due to $X$, recalling that $g_\lambda(x) = k_{x}^{\star}\gamma_\lambda$ and going for the following decomposition
\begin{equation}
\begin{split}
&\xi(x, y) = \xi_v(x, y) + \xi_b(x)\\
&\xi_v(x, y) = (\Sigma + \lambda)^{-\frac{1}{2}}k_{x}(\phi(y) - g^{*}(x)),\\
&\xi_b(x) = (\Sigma + \lambda)^{-\frac{1}{2}}k_{x}(g^{*}(x) - g_\lambda(x)),
\end{split}
\end{equation}
which corresponds to the decomposition
\begin{equation}
\begin{split}
&{\cal B}(\lambda) \leq {\cal B}_{v}(\lambda) + {\cal B}_{b}(\lambda)\\
&{\cal B}_v(\lambda) = \norm{\E_{\hat\rho}[\xi_v(X, Y)]-\E_{\rho}[\xi_v(X, Y)]}\\
&{\cal B}_b(\lambda) = \norm{\E_{\hat\rho}[\xi_b(X, Y)]-\E_{\rho}[\xi_b(X, Y)]}.
\end{split}
\end{equation}
The first term is due to the error because of having observed $\phi(y)$ rather than $g^{*}(x)$, often called ``variance'', and the second term is due to the aiming for $g_\lambda$ instead of $g^{*}$ often called ``bias''.
\subsubsection{Control of the variance}
To control the variance term, we will use the Bernstein inequality stated Theorem \ref{rates:thm:bernstein-vector-full}.
\paragraph{Bound on the moment of $\xi_v$.}
First of all notice that
\[
\norm{\xi_v(x, y)}_{\cal G} \leq \norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op} \norm{\phi(y) -
g^{*}(x)}_{\cal H}.
\]
Therefore, under Assumption \ref{rates:ass:moment}, for $m \geq 2$:
\begin{align*}
\E_{(X, Y)\sim \rho}\bracket{\norm{\xi_v(X, Y)}^m}
& \leq
\E_{X\sim\rho_{\mathcal{X}}}\bracket{\norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}^m
\E_{Y\sim\rho\vert_{X}}\bracket{\norm{\phi(y) -
g^{*}(x)}_{\cal H}^m}}
\\& \leq \frac{1}{2} m! \sigma^2 M^{m-2}\E_{X\sim\rho_{\mathcal{X}}}\bracket{\norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}^m}.
\end{align*}
We bound the last term with
\begin{align*}
\E_{X\sim\rho_{\mathcal{X}}}\bracket{\norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}^m}
& \leq \sup_{x\in\supp\rho_{\mathcal{X}}} \norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}^{m-2}
\E_{X\sim\rho_{\mathcal{X}}}\bracket{\norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}^2}
\\&= {\cal N}_{\infty}(\lambda)^{(m-2)} {\cal N}(\lambda).
\end{align*}
\paragraph{Concentration on $\xi_v$.}
Applying Theorem \ref{rates:thm:bernstein-vector-full}, we get, for any $t > 0$, that
\begin{equation}
\Pbb\paren{{\cal B}_v(\lambda) > t}
\leq 2\exp\paren{-\frac{nt^2}{2\sigma^2{\cal N}(\lambda) + 2 M{\cal N}_\infty(\lambda)t}}.
\end{equation}
\subsubsection{Control of the bias}
To control the bias, we recall a simpler version of Bernstein concentration inequality, that is a corollary of Theorem \ref{rates:thm:bernstein-vector-full}.
\begin{theorem}[Concentration in Hilbert space \citep{Pinelis1986}]
\label{rates:thm:bernstein-vector}
Let denote by ${\cal A}$ a Hilbert space and by $(\xi_{i})$ a sequence of independent random vectors on ${\cal A}$ such that $\E[\xi_{i}] = 0$, that are bounded by a constant $M$, with finite variance $\sigma^2 = \E[\sum_{i=1}^{n}\norm{\xi_{i}}^2]$.
For any $t>0$,
\[
\Pbb(\norm{\sum_{i=1}^{n} \xi_{i}} \geq t) \leq 2\exp\paren{-\frac{t^2}{2\sigma^2 +
2tM / 3}}.
\]
\end{theorem}
\paragraph{Bound on $\xi_b$.}
We have
\[
\norm{\xi_b(x)}_{\cal G} \leq \sup_{x\in\supp\rho_{\mathcal{X}}} \norm{(\Sigma+\lambda)^{-\frac{1}{2}}k_{x}}_{\op}
\norm{g_\lambda(x) - g^{*}(x)}_{\cal H}
\leq {\cal N}_{\infty}(\lambda)\norm{g_\lambda - g^{*}}_{\infty}.
\]
Therefore, with Appendix \ref{rates:proof:krr-1}, we get
\[
\norm{\xi_b(x)}_{\cal G} \leq b_1 \lambda^{q-p} {\cal N}_{\infty}(\lambda).
\]
\paragraph{Variance of $\xi_b$.}
For the variance we proceed with
\[
\norm{\xi_b(x)}_{\cal G}^2 \leq {\cal N}_{\infty}(\lambda)^2
\norm{g_\lambda(x) - g^{*}(x)}_{\cal H}^2.
\]
Therefore,
\[
\E[\norm{\xi_b(X)}^2] \leq {\cal N}_{\infty}(\lambda)^2 \norm{g_\lambda -
g^{*}}_{L^2}^2.
\]
Using the derivations made in Appendix \ref{rates:proof:krr-1}, we have, using that $q \leq 1$,
\begin{align*}
\norm{g_\lambda - g^{*}}_{L^2} & = \lambda \norm{(K+\lambda)^{-1}K^q g_0}_{L^2}
\leq \lambda \norm{(K+\lambda)^{-(1-q)}}_{\op}
\norm{(K+\lambda)^{-q}K^q}_{\op}\norm{g_0}_{L^2}
\\&\leq \lambda^q \norm{g_0}_{L^2}.
\end{align*}
\paragraph{Concentration on $\xi_b$.}
Adding everything together, we get
\begin{equation}
\Pbb\paren{{\cal B}_b(\lambda) > t}
\leq 2\exp\paren{-\frac{nt^2}{2
\paren{\lambda^{2q} {\cal N}_{\infty}(\lambda)^{2}\norm{g_0}^2_{L^2} + b_1\lambda^{q-p}{\cal N}_{\infty}(\lambda)t / 3}}}.
\end{equation}
Note that based on the bound on the variance, we would like ${\cal N}_{\infty}(\lambda)^2\lambda^{2q} \approx\lambda^{2(q-p)}$ to be smaller than ${\cal N}(\lambda) \approx \lambda^{-\sigma}$.
It is the case since $q > p$.
\subsubsection{Union bound}
To control $\norm{g_{n} - g_\lambda}_{L^\infty} \leq c_p\lambda^{-p} {\cal A}(\lambda) ({\cal B}_v(\lambda) + {\cal B}_b(\lambda))$, we need to perform a union bound on the control of ${\cal A}$ and the control of ${\cal B} := {\cal B}_v + {\cal B}_b$, we use that for any $t > 0$ and $0<s<1$, $c_p \lambda^{-p} {\cal A}{\cal B} > t$ implies ${\cal A} > \sfrac{1}{(1-s)}$ or ${\cal B} > \sfrac{(1-s)t \lambda^p}{c_p}$.
Similarly, ${\cal B}_v + {\cal B}_b > t$, implies that either ${\cal B}_v > \sfrac{t}{2}$, either ${\cal B}_b > \sfrac{t}{2}$.
Therefore, we have, the following inclusion of events (with respect to ${\cal D}_{n}$)
\[
\brace{\norm{g_{n} - g_\lambda}_{L^\infty} > t} \subset \brace{{\cal A} >
\frac{1}{1-s}}
\cup \brace{{\cal B}_v > \frac{(1-s)t\lambda^{p}}{2c_p}}
\cup \brace{{\cal B}_b > \frac{(1-s)t\lambda^{p}}{2c_p}}.
\]
In terms of probability this leads to
\begin{equation}
\Pbb_{{\cal D}_{n}}\paren{\norm{g_{n} - g_\lambda}_{L^\infty} > t} \leq
\Pbb_{{\cal D}_{n}}\paren{{\cal A} > \frac{1}{1-s}}
+ \Pbb_{{\cal D}_{n}}\paren{{\cal B} > \frac{(1-s)t\lambda^p}{c_p}}.
\end{equation}
Looking closer it is the term in ${\cal B}$ that will be the more problematic, therefore we would like $s$ to be small.
If we take $s$ to be a constant with respect to $t$, we will get something that behaves like $\Pbb(B > t\lambda^p)$, which is the best we can hope for (this also explain why we divide ${\cal B} > t$ in ${\cal B}_v > \sfrac{t}{2}$ or ${\cal B}_b > \sfrac{t}{2}$).
We will consider $s=\sfrac{1}{2}$.
We express concentration based on the expression of ${\cal N}$ and ${\cal N}_{\infty}$, assuming $\lambda \leq \norm{\Sigma}_{\op}$, and $n > a_3^2\lambda^{-2p}$
\[
\Pbb_{{\cal D}_{n}}({\cal A} > 2) \leq 28 a_2 \lambda^{-\sigma}\exp(-
\frac{n\lambda^{2p}}{10 a_3^2}).
\]
Similarly, we get, when $\lambda \leq 1$, using that $\lambda^{-\sigma} \geq 1$
\[
\Pbb_{{\cal D}_{n}}({\cal B}_v > \sfrac{t}{4}) \leq
2 \exp\paren{-\frac{n\lambda^\sigma t^2}{32\sigma^2 a_2 + 8Ma_3\lambda^{-p}t}}.
\]
For the bias term, we can proceed at a brutal bounding, based on the fact that for $\lambda \leq 1$, $\lambda^{q-p} \leq 1 \leq \lambda^{-\sigma}$, to get
\[
\Pbb_{{\cal D}_{n}}({\cal B}_b > \sfrac{t}{4}) \leq
2 \exp\paren{-\frac{n\lambda^\sigma t^2}{32 a_3^2\norm{g_0}_{L^2} +
8b_1a_3\lambda^{-p}t/3}}.
\]
With $b_4 = \max(32\sigma^2 a_2, 32 a_3^2\norm{g_0}_{L^2})$ and $b_5 = \max(8Ma_3, 8b_1a_3/3)$, we get the following union bound
\[
\Pbb_{{\cal D}_{n}}\paren{{\cal B} > \frac{t\lambda^p}{2}}
\leq 4 \exp\paren{-\frac{n\lambda^{2p+\sigma}t^2}{b_4 + b_5t}}.
\]
We proceed with the union bound on $\norm{g_{n} - g_\lambda}_{L^\infty}$ as
\[
\Pbb_{{\cal D}_{n}}(\norm{g_{n} - g_\lambda}_{L^\infty} > t)
\leq b_2 \lambda^{-\sigma}\exp(-b_3n\lambda^{2p})
+ 4 \exp\paren{-\frac{n\lambda^{2p+\sigma}t^2}{b_4 + b_5t}},
\]
with $b_2 = 28a_2$ and $b_3^{-1} = 10a_3^2$, as long as $b_3n > \lambda^{-2p}$, and $\lambda \leq \max(1, \norm{K}_{\op})$.
\subsubsection{Refinement of Lemma \ref{rates:lem:rkhs}}
Remark that the uniform control in Lemma \ref{rates:lem:rkhs} is more than we need, we only need control for each $x$ as described in Assumption \ref{rates:ass:concentration}.
Indeed, if $p(x)$ is such that there exists a constant $\tilde{c_p}$ (that does not depend on $x$ or $\lambda$), such that for any $i\in\mathbb{N}$
\[
\scap{k_{x}}{u_{i}}_{{\cal G}_{\mathcal{X}}} \leq \tilde{c_p}\lambda_{i}^{p(x)},
\]
then considering that
\[
g_{n}(x) - g_\lambda(x) = k_{x}^{\star}(\gamma_{n} - \gamma_\lambda)
= k_{x}^{\star} \paren{\Sigma+\lambda}^{-\frac{1}{2}} \paren{\Sigma+\lambda}^{\frac{1}{2}}\paren{\gamma_{n} - \gamma_\lambda},
\]
we can improve the results of Lemma \ref{rates:lem:rkhs} by replacing $p$ by $p(x)$.
While we considered $p = {\sup_{x\in\rho_{\mathcal{X}}} p(x)}$ as a consequence of our proof scheme, one can expect to end up with the $\E_{X}[\lambda^{p(X)}]$ instead of $\lambda^p$ when deriving the proof of Theorems \ref{rates:thm:krr-no-density} and \ref{rates:thm:krr-low-density} (for which one has to refine Theorem \ref{rates:thm:low-density} in order to integrate dependency of $L$ to $x$, similarly to what is done in Lemma \ref{rates:lem:ref-no-density}), which will lead to better rates.
Yet, because of the complexity of expressing a quantity of the type $\E_{X}[\phi(p(X))]$, for some function $\phi$, we decided not to present this improved version in the paper.
\subsection{Proof of Theorem \ref{rates:thm:krr-no-density}}
\label{rates:proof:krr-no-density}
Based on the proof of Theorem \ref{rates:thm:no-density}, we know that
\[
\E_{{\cal D}_{n}} {\cal R}(f_{n}) - {\cal R}(f^{*}) \leq \ell_\infty
\Pbb_{{\cal D}_{n}}\paren{\norm{g_{n} - g^{*}}_{\infty} > t_0}.
\]
Now we use that
\[
\Pbb_{{\cal D}_{n}}\paren{\norm{g_{n} - g^{*}}_{\infty} > t_0}
\leq \Pbb_{{\cal D}_{n}}\paren{\norm{g_{n} - g_\lambda}_{\infty} > t_0 - \norm{g_\lambda - g^{*}}_{\infty}}.
\]
The result follows from derivations in Appendix \ref{rates:proof:krr-2}, where we used that when $k$ is bounded, Assumptions \ref{rates:ass:capacity} and \ref{rates:ass:interpolation} are verified with $\sigma=1$ and $p=\sfrac{1}{2}$.
Note that we do not need the source assumption, since we can bound directly $\norm{g_\lambda - g^{*}}_{L^2} \leq \norm{g_\lambda - g^{*}}_{L^\infty} < t_0$ while retaking the proof in Appendix \ref{rates:proof:krr-2}.
Moreover, the results of this last proof holds under the condition $n \lambda b_3 > 1$, but, since $\E_{{\cal D}_{n}} {\cal R}(f_{n}) - {\cal R}(f^{*}) \leq \ell_\infty$, we can augment the constant $b_6$ so that the result in Theorem \ref{rates:thm:krr-no-density} still holds for any $n \in \mathbb{N}^{*}$.
\subsection{Proof of Theorem \ref{rates:thm:krr-low-density}}
\label{rates:proof:krr-low-density}
We can rephrase Lemma \ref{rates:lem:rkhs}, using a union bound
\begin{align*}
\Pbb_{{\cal D}_{n}}(\norm{g_{n} - g^{*}} > t)
& \leq \Pbb_{{\cal D}_{n}}(\norm{g_{n} - g_\lambda} > \sfrac{t}{2})
+ \Pbb_{{\cal D}_{n}}(\norm{g_\lambda - g^{*}} > \sfrac{t}{2})
\\& \leq b_2\lambda^{-\sigma}\exp\paren{-b_3n\lambda^{2p}}
+ 4\exp\paren{-\frac{n\lambda^{2p+\sigma}t^2}{4b_4 + 2b_5t}}
+ \ind{t\leq 2\lambda^{q-p}}.
\end{align*}
Using variant of Theorem \ref{rates:thm:low-density} presented in Appendix \ref{rates:app:ref-low-density}, we get
\begin{align*}
{\cal R}(f_{n}) - {\cal R}(f^{*})
& \leq \ell_\infty b_2 \lambda^{-\sigma} \exp\paren{-b_3n\lambda^{2p}}
+ 2c_{\psi}c_{\alpha} 2^{\alpha + 1} \lambda^{(q-p)(\alpha+1)}
\\&\qquad\qquad+ 2c_{\psi}c_{\alpha} c\paren{b_4^{\frac{\alpha+1}{2}} (n\lambda^{2p+\sigma})^{-\frac{\alpha+1}{2}} + b_5^{\alpha+1}(n\lambda^{2p+\sigma})^{-(\alpha + 1)}}.
\end{align*}
As long as $\lambda \leq \max(\norm{K}_{\op}, 1)$ and $n \geq (b_3\lambda^{2p})^{-1}$.
We optimize those rates with $\lambda = \lambda_0 n^{-\gamma}$, and $\gamma$ satisfying
\[
2\gamma(q-p) = 1 - \gamma(2p+\sigma) \qquad\Rightarrow\qquad
\gamma = (2q + \sigma)^{-1}.
\]
This leads to, for $n$ after a certain $N\in\mathbb{N}^{*}$
\begin{align*}
{\cal R}(f_{n}) - {\cal R}(f^{*})
& \leq \ell_\infty b_2 \lambda_0^{-\sigma} n^{\frac{\sigma}{2q+\sigma}} \exp\paren{-b_3n\lambda_0^{2p} n^{\frac{2(q-p) + \sigma}{2q+\sigma}}}
\\&\qquad + 2c_{\psi}c_{\alpha} 2^{\alpha + 1} \lambda_0^{(q-p)(\alpha+1)} n^{-\frac{(q-p)(\alpha+1)}{2q+\sigma}}
\\&\qquad + 2c_{\psi}c_{\alpha} c\paren{b_4^{\frac{\alpha+1}{2}} \lambda_0^{\frac{(2p+\sigma)\alpha+1}{2}} n^{-\frac{(q-p)(\alpha+1)}{2q+\sigma}} + b_5^{\alpha+1} \lambda_0^{(2p+\sigma)\alpha+1} n^{-\frac{2(q-p)(\alpha+1)}{2q+\sigma}}}
\\&\leq b_8 n^{-\frac{2(q-p)(\alpha+1)}{2q+\sigma}}.
\end{align*}
Since $\ell$ is bounded, ${\cal R}(f_{n}) - {\cal R}(f^{*}) \leq \ell_\infty$, and we can always higher $b_8$, in order to have the inequality for any $n\in\mathbb{N}^{*}$.
\section{Nearest neighbors}
\subsection{Usual assumptions to derive nearest neighbors convergence rates}
\label{rates:app:nn-assumptions}
Assumption \ref{rates:ass:lipschitz} can be seen as the backbone to control $\norm{g_n^*(x) - g^*(x)}$ in a uniform manner.
This assumption that relates the regularity of $g^*$ with the density of $\rho_{\mathcal{X}}$ has been historically approached in the following manner.
Assume that $g^*$ is $\beta'$-H\"older, that is, for any $x, x' \in \supp\rho_{\mathcal{X}}$
\[
\norm{g^*(x) - g^*(x')} \leq a_1 d(x, x')^{\beta'}.
\]
Suppose that $\mathcal{X} = \mathbb{R}^d$, that $\rho_{\mathcal{X}}$ is continuous against $\lambda$, the Lebesgue measure, with minimal mass in the sense that there exists a $p_{\min} > 0$ such that $\frac{\mathop{}\!\mathrm{d}\rho_{\mathcal{X}}}{\mathop{}\!\mathrm{d}\lambda}(\mathcal{X})$ does not intersect $(0, p_{\min})$, and that $\supp\rho_{\mathcal{X}}$ has regular boundaries in the sense that there exist $a_2, t_0 > 0$ such that for any $x\in\supp\rho_{\mathcal{X}}$ and $t \in (0, t_0)$
\[
\lambda\paren{{\cal B}(x, t)\cap\supp\rho_{\mathcal{X}}} \geq a_2\lambda\paren{{\cal B}(x, t)}.
\]
For example an orthant satisfies this property with $a_2 = 2^{-d}$ and $t_0 = \infty$, and ${\cal B}(0, 1)$ satisfies this property with $a_2 = \lambda\paren{{\cal B}(0, 1) \cap {\cal B}(1, 1)} / \lambda\paren{{\cal B}(0, 1)}$ and $t_0 = 1$.
In such a setting, we get
\[
\norm{g^*(x) - g^*(x')} \leq a_1 d(x, x')^{\beta'}
= a_1 \paren{\frac{\lambda({\cal B}(x, d(x, x')))}{\lambda({\cal B}(0, 1))}}^{\frac{\beta'}{d}}.
\]
When $d(x, x') < t_0$, we have
\[
\lambda({\cal B}(x, d(x,x'))) \leq a_2^{-1} \lambda\paren{{\cal B}(x, d(x,
x'))\cap \supp\rho_{\mathcal{X}}}
\leq a_2^{-1}p_{\min}^{-1} \rho_{\mathcal{X}}({\cal B}(x, d(x,x')))^\beta.
\]
This means that for any $x \in \supp\rho_{\mathcal{X}}$ and $x' \in {\cal B}(x, t_0)$ we have, with $\beta = \frac{\beta'}{d}$ and the constant $a_3= a_1 a_2^{-\beta} p_{\min}^{-\beta}\lambda({\cal B}(0,1))^{-\beta}$
\[
\norm{g^*(x) - g^*(x')} \leq a_3 \rho_{\mathcal{X}}({\cal B}(x, d(x, x')))^{\beta}.
\]
While, we actually do not need the bound to hold for $d(x, x') > t_0$ in the following proof, to check the veracity of our remark on Assumption \ref{rates:ass:lipschitz}, one can verify that under our assumptions on $\rho_{\mathcal{X}}$, $\supp\rho_{\mathcal{X}}$ is bounded, and therefore $g^*$ is too.
And if $g^*$ is bounded by $c_\phi$, by considering $a_3' = \max\paren{2c_\phi a_2^{-\beta}p_{\min}^{-\beta} t_0^{-\beta'}, a_3}$, this bound holds for any $x, x' \in \supp\rho_{\mathcal{X}}$.
\subsection{Proof of Lemma \ref{rates:lem:nn-concentration}}
\label{rates:proof:nn-concentration}
\paragraph{Control of the variance term.}
For $x\in\rho_{\mathcal{X}}$, the variance term can be written
\[
\norm{g_n(x) - g_n^*(x)} = \norm{\frac{1}{k}\sum_{i=1}^k \phi(Y_{(i)}) -
\E\bracket{Y_{(i)}\,\middle\vert\, X_{(i)}}}.
\]
Where the index $(i)$ is such that $X_{(i)}$ is the $i$-th nearest neighbor of $x$ in $(X_i)_{i\leq n}$.
Since, given $(X_i)_{i\leq n}$, the $(Y_i)_{i\neq n}$ are independent, distributed according to $\otimes_{i\leq n}\rho\vert_{X_i}$, we can use a concentration inequality to control it.
We recall Bernstein concentration inequality in such spaces, derived by \citet{Yurinskii1970}, we will use the formulation of Corollary 1 from \citet{Pinelis1986}.
\begin{theorem}[Concentration in Hilbert space \citep{Pinelis1986}]
\label{rates:thm:bernstein-vector-full}
Let denote by ${\cal A}$ a Hilbert space and by $(\xi_i)$ a sequence of independent random vectors on ${\cal A}$ such that $\E[\xi_i] = 0$, and that there exists $M, \sigma^2 > 0$ such that for any $m \geq 2$
\[
\sum_{i=1}^n \E\bracket{\norm{\xi_i}^m} \leq \frac{1}{2} m!\sigma^2 M^{m-2}.
\]
Then for any $t>0$
\[
\Pbb(\norm{\sum_{i=1}^n \xi_i} \geq t) \leq
2\exp\paren{-\frac{t^2}{2\sigma^2 + 2tM}}.
\]
\end{theorem}
This explains Assumption \ref{rates:ass:moment}, allowing, because there is only $k$ $\xi_i$ active in $\sum_{i=1}^n \alpha_i(x)\xi_i$, to get
\[
\Pbb_{{\cal D}_n}\paren{\norm{g_n(x) - g_n^*(x)} > t}
\leq 2\exp\paren{-\frac{kt^2}{2\sigma^2 + 2M t}}.
\]
\paragraph{Control of the bias term.}
Under the Modified Lipschitz condition, Assumption \ref{rates:ass:lipschitz},
\begin{align*}
\norm{g_n^*(x) - g_n(x)} & = \norm{\sum_{i=1}^n \alpha_i(x) \paren{g_n(x) - g^*(X_i)}}
\leq \sum_{i=1}^n \alpha_i(x) \norm{g_n(x) - g^*(X_i)}
\\&\leq c_{\beta} \sum_{i=1}^n \alpha_i(x) \rho_{\mathcal{X}}\paren{{\cal B}(x, d(x, X_i))}^\beta
\leq c_{\beta} \rho_{\mathcal{X}}\paren{{\cal B}(x, d(x, X_{k}(x)))}^\beta.
\end{align*}
When $\rho_{\mathcal{X}}$ is continuous, it follows from the probability integral transform (also known as universality of the uniform) that $\rho_{\mathcal{X}}\paren{{\cal B}(x, d(x, X_{k}(x)))}$ behaves like the $k$-th order statistics of a sample $(U_i)_{i\leq n}$ of $n$ uniform distributions on $[0,1]$.
Therefore, for any $s\in[0,1]$
\begin{align*}
\Pbb_{{\cal D}_n}\paren{\rho_{\mathcal{X}}\paren{{\cal B}(x, d(x, X_{k}(x)))} > s}
& = \Pbb\paren{\sum_{i=1}^n \ind{U_i < s} \leq k}.
\end{align*}
Recall the multiplicative Chernoff bound, stating that for $(Z_i)_{i\leq n}$ $n$ independent random variables in $\brace{0,1}$, if $Z = \sum_{i=1}^n Z_i$, and $\mu = \E[Z]$, for any $\delta > 0$
\[
\Pbb\paren{Z \leq (1-\delta)\mu} \leq \exp\paren{-\frac{\delta^2\mu}{2}}.
\]
Since, for $s \in [0, 1]$, $\E[\ind{U_i < s}] = \Pbb(U_i < s) = s$, we get, when $k \leq ns / 2$
\[
\Pbb\paren{\sum_{i=1}^n \ind{U_i < s} \leq k} \leq
\exp\paren{-\frac{\paren{ns - k}^2}{2 ns}}
\leq \exp\paren{-\frac{ns}{8}}.
\]
With $s = c_{\beta}^{-1}t^{\frac{1}{\beta}}$, we get
\[
\Pbb_{{\cal D}_n}\paren{\norm{g_n^*(x) - g_n(x)} > t}
\leq \exp\paren{-\frac{nt^{\frac{1}{\beta}}}{8c_{\beta}}}.
\]
Remark that when $g^*$ is $\beta'$ H\"older, we get the same result with $\rho_{\mathcal{X}}({\cal B}(x, t))$ instead of $t^{\frac{1}{\beta}}$ by considering $\ind{X_i \in {\cal B}(x, t)}$ instead of $\ind{U_i \leq t}$.
Note that there is a way to bound a Binomial distribution with a Gaussian for $t$ smaller than the mean of the binomial distribution, which would lead to a bound that holds for any $t>0$ \citep{Slud1977}.
\subsection{Proof of Theorem \ref{rates:thm:nn-no-density}}
\label{rates:proof:nn-no-density}
Using the proof of Theorem \ref{rates:thm:no-density}, we get
\[
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*)
\leq \ell_\infty\Pbb_{{\cal D}_n}\paren{\norm{g(x) - g_n^*(x)} > t_0}.
\]
Because $\norm{g_n(x) - g^*(x)} > t_0$ implies that either $\norm{g_n(x) - g_n^*(x)} > t_0/2$ or $\norm{g_n^*(x) - g^*(x)} > t_0/2$, we get using Lemma \ref{rates:lem:nn-concentration}
\[
\Pbb_{{\cal D}_n}\paren{\norm{g(x) - g_n^*(x)} > t_0}
\leq
2\exp\paren{-\frac{b_1kt_0^2}{4 + 2b_2t_0}}
+ \exp\paren{-2^{-\frac{1}{\beta}}b_3 nt_0^{\frac{1}{\beta}}}
+\ind{t_0 < \paren{k/2n}^\beta}.
\]
This explains the result of Theorem \ref{rates:thm:nn-no-density}.
\subsection{Proof of Theorem \ref{rates:thm:nn-low-density}}
\label{rates:proof:nn-low-density}
First of all for $t > 0$, and $x \in \supp\rho_{\mathcal{X}}$, because $\norm{g_n(x) - g^*(x)} >t$ implies that either $\norm{g_n(x) - g_n^*(x)} > t / 2$ or $\norm{g_n^*(x) - g^*(x)} > t / 2$, we have the inclusion of events:
\[
\brace{{\cal D}_n \,\middle\vert\, \norm{g_n(x) - g^*(x)} > t}
\subset \brace{{\cal D}_n \,\middle\vert\, \norm{g_n(x) - g_n^*(x)} > t/2}
\cup \brace{{\cal D}_n \,\middle\vert\, \norm{g_n(x) - g^*(x)} > t/2},
\]
which translates in terms of probability as
\begin{align*}
\Pbb_{{\cal D}_n}\paren{\norm{g_n(x) - g^*(x)} > 2t}
& \leq \Pbb_{{\cal D}_n}\paren{\norm{g_n(x) - g_n^*(x)} > t}
+\Pbb_{{\cal D}_n}\paren{ \norm{g_n^*(x) - g^*(x)} > t}.
\\&\leq 2\exp\paren{-\frac{b_1kt^2}{1 + b_2t}}
+ \exp\paren{-b_3 nt^{\frac{1}{\beta}}}
+ \ind{t < \paren{\frac{k}{2n}}^\beta}.
\end{align*}
Using the refinements of Theorem \ref{rates:thm:low-density} exposed in Appendix \ref{rates:app:ref-low-density}, we get that there exists a constant $c > 0$ that does not depend on $k$ or $n$ such that
\[
\E_{{\cal D}_n}{\cal R}{f_n} - {\cal R}(f^*)
\leq
c \paren{k^{-\frac{\alpha + 1}{2}} + n^{-\beta(\alpha + 1)} +
(nk^{-1})^{\beta(\alpha + 1)}}.
\]
We optimize this last quantity with respect to $k$ by taking $k = n^{\gamma}$, and choosing $\gamma$ such that $\gamma = 2(1-\gamma)\beta$ leading to $\gamma = 2\beta / (2\beta + 1)$ and to rates in $n$ to the power minus $\beta(\alpha + 1) / (2\beta + 1)$.
\subsection{Numerical experiments}
\begin{figure}[ht]
\centering
\includegraphics{rates/drawings/more_rates-eps-converted-to.pdf}
\caption{Supplement to Figure \ref{rates:fig:rates}.
We specify the error is evaluated on 100 points forming a regular partition of $\mathcal{X} = [-1, 1]$, and the expectation $\E_{{\cal D}_n}$ is approximated by considering 100 datasets.
The violet curve is cropped at $n\approx 10^5$, because the error was null afterwards with our evaluation parameters (only 100 points to evaluate the error), forbidding us to consider the logarithm of the excess of risk.}
\label{rates:fig:more_rates}
\end{figure}
Interestingly, on numerical simulations such as the one presented on Figure \ref{rates:fig:rates}, we observed two regimes.
A first regime where bound is meaningless because of constants being too big, and where the error decreases independently of the exponent expected through Theorem \ref{rates:thm:nn-low-density}, and a final regime where rates correspond to the bond given by the theorem.
Note that when $\alpha >> 1$, with our computation parameter, we do not get to really illustrate convergence rates, as this final regime get place for bigger $n$ than what we have considered ($n_{\max} = 10^6$), this being partly due to the constant $c_\beta$ in Assumption \ref{rates:ass:lipschitz} being big for the $g^*$ we considered.
Furthermore, note that, for example, if a problem satisfied Assumption \ref{rates:ass:no-density} with a tiny $t_0$, we expect that exponential convergence rates are only going to be observed for $n > N$, with $N$ huge, and for which the excess of risk is already minuscule.
\section{Fast rates}
\label{rates:sec:proof}
In the following, we consider $\mathcal{X}$ and $\mathcal{Y}$ to be Polish spaces, {\em i.e.}, separable completely metrizable topological spaces, in order to define the distribution $\rho$.
We also consider $\cal Z$ endowed with a topology that makes it compact, and that makes $z\to\E_{Y\sim\mu}\ell(z, Y)$ continuous for any $\mu\in\prob{\mathcal{Y}}$, in order to have minimizer well-defined.
For a Polish space ${\cal A}$, we denote by $\prob{\cal A}$ the simplex formed by the set of Borel probability measures on this space.
For $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$, we denote by $\rho\vert_{x}$ the conditional distribution of $Y$ given $x$, and by $\rho_{\mathcal{X}}$ the marginal distribution over $\mathcal{X}$.
We suppose ${\cal H}$ separable Hilbert and that the mapping $\phi$ is measurable in order to define the pushforward measure $\phi_*\rho\vert_{x}$.
We assume that, for $\rho_{\mathcal{X}}$-almost every $x$, $(\phi(Y)\vert X=x)$ has a second moment, in order to consider the conditional mean $g^*(x)$ as the solution of the well-defined problem consisting of minimizing $\norm{\xi - \phi(Y)}^2$ for $\xi \in {\cal H}$.
We consider $\psi$ to be continuous, in order to have the decoding problem well posed.
\subsection{Proof of Lemma \ref{rates:lem:cal}}
\label{rates:proof:cal}
With the notation of Lemma \ref{rates:lem:cal}, for $x\in\supp\rho_{\mathcal{X}}$
\begin{align*}
\E_{Y\sim\rho\vert_{x}} & \bracket{\ell(f_n(x), Y) - \ell(f^*(x), Y)}
= \scap{\psi(f_n(x)) - \psi(f^*(x))}{g^*(x)}_{\cal H}
\\& = \scap{\psi(f_n(x))}{g_n(x)}
+ \scap{\psi(f_n(x))}{g^*(x) - g_n(x)}
- \scap{\psi(f^*(x))}{g^*(x)}
\\& \leq \scap{\psi(f^*(x))}{g_n(x)}
+ \scap{\psi(f_n(x))}{g^*(x) - g_n(x)}
- \scap{\psi(f^*(x))}{g^*(x)}
\\& = \scap{\psi(f_n(x)) - \psi(f^*(x))}{g^*(x) - g_n(x)}
\\& \leq \norm{\psi(f_n(x)) - \psi(f^*(x))}_{\cal H}\norm{g^*(x) - g_n(x)}_{\cal H}
\\& \leq 2c_{\psi}\norm{g^*(x) - g_n(x)}_{\cal H},
\end{align*}
where the inequality $\scap{\psi(f_n(x))}{g_n(x)} \leq \scap{\psi(f^*(x))}{g_n(x)}$ is due to the fact that $f_n(x)$ minimizes the functional $z\to \scap{\psi(z)}{g_n(x)}$.
Integrating over $\mathcal{X}$ leads to the results in Lemma \ref{rates:lem:cal}.
\subsection{Proof of Lemma \ref{rates:lem:calibration}}
\label{rates:proof:calibration}
The first part of the lemma is a geometrical result stating that to go from two elements $\xi_1$ and $\xi_2$ in $\prob{\phi(\mathcal{Y})}$, leading to two different decoding, one has to pass by a point $\xi_{1/2} \in F$, where there is at least two possible decodings.
Let us make it clearer.
Consider $x\in\supp\rho_{\mathcal{X}}$ and suppose that $f_n(x) \neq f^*(x)$, define the path
\[
\myfunction{\zeta}{[0,1]}{\prob{\phi(\mathcal{Y})}}{\lambda}{\lambda g_n(x) +
(1-\lambda)g^*(x).}
\]
Consider $d:\prob{\phi(\mathcal{Y})}\to{\cal Z}$ the decoding function used to retrieve $f^*$ and $f_n$, from $g^*$ and $g_n$, satisfying $d(\xi) \in \argmin_{z\in{\cal Z}} \scap{\psi(z)}{\xi}$.
Consider the path $d\circ\zeta:[0, 1] \to {\cal Z}$, it goes from $\zeta(0) = f^*(x)$ to $\zeta(1) = f_n(x)$.
Consider $\lambda_{\infty}$ the supremum of $(d\circ\zeta)^{-1}(f^*(x))$.
We will show that $\zeta(\lambda_{\infty}) \in F$, this will lead to
\[
\norm{g_n(x) - g^*(x)} = \norm{g_n(x) - \zeta(\lambda_{\infty})}
+ \norm{\zeta(\lambda_{\infty}) - g^*(x)}
\geq \norm{\zeta(\lambda_{\infty}) - g^*(x)}
\geq d(g^*(x), F),
\]
and to Lemma \ref{rates:lem:calibration} by contraposition.
To show that $\zeta(\lambda_{\infty}) \in F$, we will show that \( f^*(x) \in \argmin_{z} \scap{\psi(z)}{\zeta(\lambda_{\infty})} \not\subset\brace{f^*(x)}.\)
By definition of the supremum, there exists a sequence $(\lambda_p)_{p\in\mathbb{N}}$ converging to $\lambda_{\infty}$ such that
\[
f^*(x) \in
\argmin_{z} \scap{\psi(z)}{\lambda_p g_n(x) + (1-\lambda_p)g^*(x)},
\]
meaning that for all $z\neq f^*(x)$
\[
\scap{\psi(f^*(x))}{\lambda_p g_n(x) + (1-\lambda_p)g^*(x)}
\leq \scap{\psi(z)}{\lambda_p g_n(x) + (1-\lambda_p)g^*(x)}.
\]
By continuity of the scalar product, it holds for $p=\infty$, which means $f^*(x) \in \argmin_z \scap{\psi(z)}{\zeta(\lambda_{\infty})}$.
Now, suppose that $\argmin_z\scap{\psi(z)}{\zeta(\lambda_{\infty})} = \brace{f^*(x)}$, this means that for all $z\neq f^*(x)$,
\[
\scap{\psi(f^*(x))}{\lambda_{\infty} g_n(x) + (1-\lambda_{\infty})g^*(x)}
< \scap{\psi(z)}{\lambda_{\infty} g_n(x) + (1-\lambda_{\infty})g^*(x)}.
\]
By continuity of this function according to $\lambda$, this means that this still holds for $\lambda_{\infty} + \epsilon_z$ for $\epsilon_z > 0$.
Taking $\epsilon = \inf_{z\in{\cal Z}} \epsilon_z$, it means that $\lambda_{\infty} + \epsilon \in (d\circ\zeta)^{-1}(f^*(x))$.
When ${\cal Z}$ is finite, $\epsilon > 0$, which contradicts the definition of $\lambda_{\infty}$.
Therefore, $\zeta(\lambda_{\infty}) \in F$.
The second part of Lemma \ref{rates:lem:calibration} follows from derivations in Appendix \ref{rates:proof:cal}.
\begin{remark}[Extension to discrete cases]\label{rates:rmk:extension-1}
Note that the same argument can be generalized to discrete problems -- which could be defined as ${\cal Z}$ endowed with a topology that makes $z\to\E_{Y\sim\mu}[\ell(z, Y)]$ continuous with respect to $z$, and ${\cal Z}\setminus\brace{z}$ locally compact for any $z\in\cal Z$ -- that are not degenerate, in the sense that $\rho_{\mathcal{X}}$ almost all $x\in\mathcal{X}$, there exists $t > 0$ such that the cardinality of the set defined as $\brace{z\,\middle\vert\,\E_{Y\sim\rho\vert_{x}}[\ell(z, Y)] - \inf_{z'\in\mathcal{Y}}\E_{Y\sim\rho\vert_{x}}[\ell(z', Y)] < t}$ if finite.
This holds for classification with infinite countable classes, but it does not for regression on the set of rational numbers.
\end{remark}
\begin{remark}[Extension to general cases]\label{rates:rmk:extension-2}
To remove the condition ${\cal Z}$ finite, one can change the definition of $d(g^*(x), F)$ to
\(
\inf_{\xi\in{\cal H}; \brace{f^*(x)} \neq \argmin\scap{\psi(z)}{\xi}} \norm{\xi - g^*(x)},
\)
in order to make Lemma \ref{rates:lem:calibration} hold for any ${\cal Z}$.
\end{remark}
\subsection{Equivalence between generalizations of the Tsybakov margin condition}
\label{rates:proof:def-low-density}
While we state the margin condition with $d(g^*(x), F)$, it could also be stated with $d(g^*(x), F\cap\hull(\phi(\mathcal{Y})))$ or with, which is the quantity considered by \citep{Nowak2019},
\[
\gamma(x) = \inf_{z\neq z^*} \E_{Y\sim \rho\vert_{x}}\ell(z, Y) -
\E_{Y\sim \rho\vert_{x}}\ell(z^*, Y)
= \inf_{z\neq z^*}\scap{\psi(z) - \psi(z^*)}{g^*(x)}.
\]
Indeed, when $\cal Z$ is finite and $\ell$ is proper in the sense that $\ell(\cdot, y) = \ell(\cdot, z)$ implies $z=y$, and that there is no $z$ that minimizes a linear combination of $(\ell(\cdot, y))_{y\in\mathcal{Y}}$ without minimizing a convex combination of the same family, we have the existence of two constants such that
\[
c\gamma(x) \leq d(g^*(x), F\cap\hull(\phi(\mathcal{Y}))) \leq d(g^*(x), F) \leq c'\gamma(x).
\]
\subsubsection{Mildness of our condition}
Let $z'$ be the argmin defining $\gamma$, geometric properties of the scalar product imply the existence of a $\xi \in (\psi(z') - \psi(z^*))^\perp$ such that
\[
\scap{\psi(z') - \psi(z^*)}{g^*(x)}
= \norm{\psi(z') - \psi(z^*)} \norm{g^*(x) - \xi}.
\]
Therefore,
\[
\scap{\psi(z') - \psi(z^*)}{g^*(x)}
\geq \min_{y, y'} \norm{\psi(y) - \psi(y')}
\norm{g^*(x) - \xi}.
\]
Note that, by definition of $\xi$, $\scap{\xi}{\psi(z')} = \scap{\xi}{\psi(z^*)}$.
If $\xi \in R_{z^*}$ then $\xi \in F$, otherwise $\xi \notin R_{z^*}$ and then, there exists a point between $\xi$ and $g^*(x)$ that belongs to the decision frontier (see Appendix \ref{rates:proof:calibration} for a proof - for which we need some regularity assumption such as $\cal Z$ finite).
In every case,
\[
\norm{g^*(x) - \xi} \geq d(g^*(x), F).
\]
This implies the existence of $c'$.
\subsubsection{Strength of our condition}
For any $g_n$ such that $f_n(x) = z$, we have
\begin{align*}
\scap{\psi(z) - \psi(z^*)}{g^*(x)}
& = \scap{\psi(z)}{g^*(x) - g_n(x)}
+ \scap{\psi(z)}{g_n(x)} - \scap{\psi(z^*)}{g^*(x)}
\\&\leq \scap{\psi(z)}{g^*(x) - g_n(x)}
+ \scap{\psi(z^*)}{g_n(x)} - \scap{\psi(z^*)}{g^*(x)}
\\&\leq 2 c_\psi\norm{g^*(x) - g_n(x)}.
\end{align*}
If we take the infimum on both sides we have
\[
d(g^*(x), F) = \inf_{g_n(x)\notin R_{f^*(x)}} \norm{g_n(x) - g^*(x)}
\geq \frac{1}{2c_\psi}\inf_{z\neq z^*} \scap{\psi(z) - \psi(z^*)}{g^*(x)},
\]
where the left equality is provided, when ${\cal Z}$ is finite, by a similar reasoning to the one in Appendix \ref{rates:proof:calibration}.
This implies the existence of $c$.
Note also that if the loss is proper in the sense that if $z$ minimizes $\scap{\psi(z)}{\xi}$ for a $\xi\in{\cal H}$, there exists a $\xi \in \hull{\phi(\mathcal{Y})}$ such that $z$ minimizes $\scap{\psi(z)}{\xi}$, we can consider $g_n(x) \in \hull(\phi(\mathcal{Y}))$, and therefore restrict $F$ to $F\cap\hull{\phi(\mathcal{Y})}$.
Finally, we have shown that, when $\cal Z$ finite and $\ell$ proper
\[
c\gamma(x) \leq d(g^*(x), F\cap\hull(\phi(\mathcal{Y}))) \leq d(g^*(x), F) \leq c'\gamma(x).
\]
This explains why we would consider $\gamma(x)$, $d(g^*(x), F\cap\hull(\phi(\mathcal{Y})))$ or $d(g^*(x), F)$ to define the margin condition, it will only change the value of constants in Assumptions \ref{rates:ass:no-density} and \ref{rates:ass:low-density}.
\subsection{Refinement of Theorem \ref{rates:thm:no-density}}
\label{rates:proof:no-density}
It is possible to refine Theorem \ref{rates:thm:no-density} to remove the condition that the loss $\ell$ is bounded.
In the following, we omit the dependency of $L_n$ and $M_n$ to $n$.
\begin{lemma}[Refinement of Theorem \ref{rates:thm:no-density}]
\label{rates:lem:ref-no-density}
Under refined calibration \eqref{rates:eq:calibration}, concentration, Assumption \ref{rates:ass:concentration}, and no-density separation, Assumption \ref{rates:ass:no-density}, the risk is controlled as
\[
\E_{{\cal D}_n} {\cal R}(f_n) - {\cal R}(f^*) \leq
4c_\psi L^{-1/2} \exp\paren{-\frac{t_0^2 L}{2}}^{1/2}
+ 4c_\psi ML^{-1} \exp\paren{- \frac{t_0 L}{2M}}.
\]
Note that it is not possible to derive a better bound only given~\eqref{rates:eq:concentration}, \eqref{rates:eq:calibration} and \eqref{rates:eq:no-density}.
Yet when $\ell$ is bounded by $\ell_{\infty}$, we have
\[
\E_{{\cal D}_n} {\cal R}(f_n) - {\cal R}(f^*) \leq \ell_{\infty}
\exp\paren{-\frac{Lt_0^2}{1 + Mt_0}}.
\]
\end{lemma}
\begin{proof}
Using the calibration inequality along with the no-density separation one, we get
\begin{align*}
{\cal R}(f_n) - {\cal R}(f^*)
& \leq 2c_\psi \E_{X}\bracket{\ind{\norm{g_n(X) - g^*(X)} \geq t_0} \norm{g_n(X) - g^*(X)}} \\
& = 2c_\psi \int_{t_0}^\infty \Pbb_{X}\paren{\norm{g_n(X) - g^*(X)} \geq t} \mathop{}\!\mathrm{d} t.
\end{align*}
Taking the expectation over ${\cal D}_n$ and using concentration inequality we have
\begin{align*}
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*)
& \leq 2c_\psi \int_{t_0}^\infty \Pbb_{X, {\cal D}_n}\paren{\norm{g_n(X) - g^*(X)} \geq t} \mathop{}\!\mathrm{d} t \\
& \leq 2c_\psi \int_{t_0}^\infty \exp\paren{-\frac{Lt^2}{1+Mt}} \mathop{}\!\mathrm{d} t.
\end{align*}
We only need to study the integral
\(
\int_{t_0}^\infty \exp\paren{-\frac{Lt^2}{1+Mt}} \mathop{}\!\mathrm{d} t.
\)
We first clean the dependency on $t$ inside the exponential using that
\[
\frac{1}{2}\paren{\exp(-Lt^2) + \exp\paren{-\frac{Lt}{M}}}
\leq \exp\paren{-\frac{Lt^2}{1 + Mt}}
\leq \exp\paren{-\frac{Lt^2}{2}} + \exp\paren{-\frac{Lt}{2M}}.
\]
We are left with the study of
\(
\int_{t_0}^\infty \exp(-At^p)\mathop{}\!\mathrm{d} t,
\)
for $p \in \brace{1, 2}$ and $A > 0$.
The case $p=1$, directly leads to
\(
A^{-1}\exp(-At_0),
\)
explaining the part in $L / M$.
The case $p=2$ is similar to the Gaussian integral, and can be handled with the following tricks
\begin{align*}
\int_{t_0}^\infty \exp(-At^2)\mathop{}\!\mathrm{d} t
& = \frac{1}{2} \int_{(-\infty, -t_0]\cup[t_0, \infty)} \exp(-At^2)\mathop{}\!\mathrm{d} t
\\&= \frac{1}{2}\paren{\int_{((-\infty, -t_0]\cup[t_0, \infty))^2} \exp(-A\norm{x}^2))
\mathop{}\!\mathrm{d} x}^{1/2}.
\end{align*}
This last integral corresponds to integrate the function $x\to\exp(-A\norm{x}^2)$ for $x\in\mathbb{R}^2$ on the domain $((-\infty, -t_0]\cup[t_0, \infty))^2$.
This function being positive and the domain being included in the domain $\brace{\norm{x} \geq t_0}$ and containing the domain $\brace{\norm{x} \geq \sqrt{2}t_0}$, we get
\begin{align*}
\int_{\brace{\norm{x}\geq \sqrt{2}t_0}}^\infty \exp(-A\norm{x}^2)\mathop{}\!\mathrm{d} x
\leq \paren{2\int_{t_0}^\infty \exp(-At^2)\mathop{}\!\mathrm{d} t}^2
\leq \int_{\brace{\norm{x}\geq t_0}}^\infty \exp(-A\norm{x}^2)\mathop{}\!\mathrm{d} x.
\end{align*}
Using polar coordinate we get
\[
\int_{\brace{\norm{x}\geq t_0}}^\infty \exp(-A\norm{x}^2)\mathop{}\!\mathrm{d} x
= 2\pi\int_{t_0}^\infty r\exp(-Ar^2)\mathop{}\!\mathrm{d} r
= \pi A^{-1} \exp(-At_0^2).
\]
Therefore,
\[
2^{-1}\sqrt{\pi} A^{-1/2} \exp(-A 2 t_0^2)^{1/2}
\leq \int_{t_0}^\infty \exp(-At^2)\mathop{}\!\mathrm{d} t
\leq 2^{-1}\sqrt{\pi} A^{-1/2} \exp(-A t_0^2)^{1/2}.
\]
This explains the rates in $L$.
\end{proof}
\subsection{Proof of Theorem \ref{rates:thm:low-density}}
\label{rates:proof:low-density}
Using the calibration and Bernstein inequalities we get, omitting the dependency of $L_n$ and $M_n$ to $n$,
\begin{align*}
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*)
& \leq 2c_{\psi}\E_{{\cal D}_n, X}\bracket{\ind{d(g_n(X), g^*(X)) \geq d(g^*(X), F)}\norm{g_n(X) - g^*(X)}}
\\& = 2c_{\psi}\int_{0}^\infty \Pbb_{{\cal D}_n, X}\paren{\ind{d(g_n(X), g^*(X)) \geq d(g^*(X), F)}\norm{g_n(X) - g^*(X)} \geq t} \mathop{}\!\mathrm{d} t
\\& = 2c_{\psi}\int_{0}^\infty \E_{X}\Pbb_{{\cal D}_n}\paren{\norm{g_n(X) - g^*(X)} \geq \max\brace{t, d(g^*(X), F)}} \mathop{}\!\mathrm{d} t
\\& \leq 2c_{\psi}\int_{0}^\infty \E_{X}\exp\paren{-\frac{L\max\brace{t, d(g^*(X), F)}^2}{1 + M \max\brace{t, d(g^*(X), F)}^2}} \mathop{}\!\mathrm{d} t
\\& = 2c_{\psi}\int_{0}^\infty \E_{X}\bracket{\ind{d(g^*(X), F) < t} \exp\paren{-\frac{Lt^2}{1 + Mt}}}\mathop{}\!\mathrm{d} t
\\&\qquad + 2c_{\psi} \int_0^\infty \E_{X}\bracket{\ind{d(g^*(X), F) \geq t}\exp\paren{-L\frac{d(g^*(X), F)^2}{1 + Md(g^*(X), F)}}}\mathop{}\!\mathrm{d} t
\\& = 2c_{\psi}\int_{0}^\infty \Pbb_{X}\paren{d(g^*(X), F) < t} \exp\paren{-\frac{Lt^2}{1 + Mt}}\mathop{}\!\mathrm{d} t
\\& \qquad + 2c_{\psi}\E_{X}\bracket{d(g^*(X), F) \exp\paren{-\frac{Ld(g^*(X), F)^2}{1 + Md(g^*(X), F)}}}.
\end{align*}
Let us begin by working on the first term.
We have, using the low-density separation hypothesis
\begin{align*}
\int_{0}^\infty \Pbb_{X}\paren{d(g^*(X), F) < t} \exp\paren{-\frac{Lt^2}{1 + Mt}}\mathop{}\!\mathrm{d} t
& \leq c_{\alpha} \int_{0}^\infty t^\alpha \exp(-\frac{Lt^2}{1+Mt})\mathop{}\!\mathrm{d} t.
\end{align*}
Recall the expression of the Gamma integral
\[
\int_0^\infty t^\alpha \exp(-Lt) \mathop{}\!\mathrm{d} t = \frac{\Gamma(\alpha+1)}{L^{\alpha+1}}
\qquad\text{and}\qquad
\int_0^\infty t^\alpha \exp(-Lt^2) \mathop{}\!\mathrm{d} t = \frac{\Gamma\paren{\frac{\alpha+1}{2}}}{2 L^{\frac{\alpha+1}{2}}}.
\]
Let us briefly talk about optimality.
Up to now, we have only used three inequalities: calibration, exponential concentration and low-density separation.
Therefore, when those inequalities hold as equalities, we get a lower bound of order on the excess of risk as
\begin{align*}
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*)
& \geq 2c_{\psi}c_{\alpha} \int_{0}^\infty t^\alpha \exp(-\frac{Lt^2}{1+Mt})\mathop{}\!\mathrm{d} t
\\&\geq 2c_{\psi}c_{\alpha} \int_{0}^\infty \frac{1}{2} t^\alpha \paren{\exp(-Lt^2) + \exp\paren{-\frac{Lt}{M}}}\mathop{}\!\mathrm{d} t
\\&= 2c_{\psi} c_{\alpha}\paren{\frac{\Gamma\paren{\frac{\alpha+1}{2}}}{4} L^{-\frac{\alpha+1}{2}}
+ \frac{\Gamma(\alpha+1)}{2} M^{\alpha+1}L^{-(\alpha + 1)}}.
\end{align*}
For the upper bound, using that $\exp(-a/1+b) \leq \exp(-a/2) + \exp(-a/2b)$, we get
\begin{align*}
\int_{0}^\infty t^\alpha \exp(-\frac{Lt^2}{1+Mt})\mathop{}\!\mathrm{d} t
& \leq \int_{0}^\infty t^\alpha \exp(-\frac{Lt^2}{2})\mathop{}\!\mathrm{d} t
+ \int_{0}^\infty t^\alpha \exp(-\frac{Lt}{2M})\mathop{}\!\mathrm{d} t
\\&= 2^{\frac{\alpha - 1}{2}} \Gamma\paren{\frac{\alpha+1}{2}} L^{-\frac{\alpha + 1}{2}}
+ 2^{\alpha+1} \Gamma(\alpha + 1) M^{\alpha + 1} L^{-(\alpha + 1)} .
\end{align*}
Let study the second term in the excess of risk inequality.
To enhance readability, write $\eta(X) = d(g^*(X), F)$.
We will first dissociate the two parts in the exponential with
\begin{align*}
\E_{X}\bracket{\eta(X) \exp\paren{-\frac{L\eta(X)^2}{1 + M\eta(X)}}}
\leq \E_{X}\bracket{\eta(X) \paren{\exp\paren{-\frac{L\eta(X)^2}{2}} + \exp\paren{-\frac{L\eta(X)}{2M}}}.}
\end{align*}
We are left with studying $\E[\eta(X) \exp(-A\eta(X)^p)]$, for $A > 0$ and $p\in \brace{1, 2}$.
The function $t\to~t\exp(-At^p)$ achieves its maximum in $t_0 = (pA)^{-1/p}$, increasing before and decreasing after.
Notice that the quantity
\begin{align*}
\Pbb(\eta(X) < t_0)\E_{X}\bracket{\eta(X) \exp(-A\eta(X)^p)\,\middle\vert\, \eta(X) < t_0}
& \leq c_{\alpha} t_0^{\alpha + 1} \exp(-A t_0^p)
\\&= c_{\alpha} p^{-\frac{\alpha + 1}{p}}\exp(-p^{-1/p}) A^{-\frac{\alpha + 1}{p}},
\end{align*}
is exactly of the same order as the control we had on the first term in the excess of risk decomposition.
This suggests considering the following decomposition
\begin{align*}
& \E_{X}\bracket{\eta(X)\exp(-A\eta(X)^p)}=
\Pbb(\eta(X) < t_0)\E_{X}\bracket{\eta(X) \exp(-A\eta(X)^p)\,\middle\vert\, \eta(X) < t_0}
\\&\qquad +\sum_{i=0}^\infty \Pbb(2^i t_0\leq \eta(X) < 2^{i+1}t_0)\E_{X}\bracket{\eta(X) \exp(-A\eta(X)^p)\,\middle\vert\, 2^it_0 \leq \eta(X) < 2^{i+1}t_0}
\\&\qquad \leq c_{\alpha} t_o^{\alpha+1} \exp(-At_0^p)
+ \sum_{i=0}^\infty c_{\alpha} 2^\alpha (2^i t_0)^{\alpha + 1} \exp(-At_0^p (2^i)^p)
\\&\qquad = c_{\alpha} t_o^{\alpha+1} \paren{\exp(-p^{-1/p})
+ \sum_{i=0}^\infty 2^\alpha 2^{i(\alpha + 1)} \exp(-p^{-1/p} 2^{ip})}.
\end{align*}
The convergence of the last series, ensures the existence of a constant $c$ such that
\begin{align*}
\E_{X}\bracket{\eta(X) \exp\paren{-\frac{L\eta(X)^2}{1 + M\eta(X)}}}
& \leq c \paren{\paren{\frac{L}{2M}}^{-(\alpha + 1)} + \paren{\frac{L}{2}}^{-\frac{\alpha + 1}{2}}}.
\end{align*}
Adding everything together, we get the existence of two constants $c', c''$, such that
\[
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*) \leq 2c_{\psi}c_{\alpha}\paren{c' M^{\alpha+1}
L^{-(\alpha+ 1)} + c'' L^{-\frac{\alpha + 1}{2}}}.
\]
This ends the proof by considering $c = \max\paren{c', c''}$.
\subsection{Refinement of Theorem \ref{rates:thm:low-density}}
\label{rates:app:ref-low-density}
Some convergence analyses lead to exponential inequalities that are not of Bernstein type, indeed, our result still holds in those settings, as mentioned by the following lemma.
In the following, we omit the dependency of $L_n$ and $M_n$ to $n$.
\begin{lemma}[Refinement of Theorem \ref{rates:thm:low-density}]
\label{rates:lem:ref-low-density}
Under the assumptions of Theorem \ref{rates:thm:low-density}, if the concentration is not given by Assumption \ref{rates:ass:concentration} but given, for some positive constants $(a_{i}, b_{i}, p_{i})_{i\leq m}$, by, for all $x\in\supp\rho_{\mathcal{X}}$ and $t > 0$,
\[
\Pbb_{{\cal D}_n}(\norm{g_n(x) - g^*(x)} > t) \leq \sum_{i=1}^n a_{i} \exp(-b_{i}t^{p_{i}}).
\]
Then the excess of risk is controlled by
\[
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*) \leq c\sum_{i=1}^n a_{i} b_{i}^{-\frac{\alpha + 1}{p_{i}}},
\]
for a constant $c$ that does not depend on $(a_{i}, b_{i})_{i\leq m}$.
\end{lemma}
\begin{proof}
First of all, remark that the proof of Theorem \ref{rates:thm:low-density} is linear in $\Pbb_{{\cal D}_n}(\norm{g_n(x) - g^*(x)} > t)$, therefore we only need to prove this lemma for $(a, b, p)$, for which we proceed as in Theorem \ref{rates:thm:low-density}
\begin{align*}
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*)
& \leq 2c_{\psi} \E_{{\cal D}_n, X}\bracket{\ind{\norm{g_n(X) - g(X)} \geq d(g(X), F)}\norm{g_n(X) - g(X)}}
\\& = 2c_{\psi}\int_{0}^\infty \Pbb_{{\cal D}_n, X}\paren{\ind{\norm{g_n(X) - g(X)} \geq d(g(X), F)}\norm{g_n(X) - g(X)} \geq t} \mathop{}\!\mathrm{d} t
\\& = 2c_{\psi}\int_{0}^\infty \E_{X}\Pbb_{{\cal D}_n}\paren{\norm{g_n(X) - g(X)} \geq \max\brace{t, d(g(X), F)}} \mathop{}\!\mathrm{d} t
\\& \leq 2c_{\psi}a \int_{0}^\infty \E_{X}\exp\paren{- b\max\brace{t, d(g(X), F)}^p} \mathop{}\!\mathrm{d} t
\\& = 2c_{\psi}a\int_{0}^\infty \E_{X}\bracket{\ind{d(g(X), F) < t} \exp\paren{-bt^p}}\mathop{}\!\mathrm{d} t
\\&\qquad\qquad + 2c_{\psi}a \int_0^\infty \E_{X}\bracket{\ind{d(g(X), F) \geq t}\exp\paren{-bd(g(X), F)^p}}\mathop{}\!\mathrm{d} t
\\& = 2c_{\psi}a\int_{0}^\infty \Pbb_{X}\paren{d(g(X), F) < t} \exp\paren{-bt^p}\mathop{}\!\mathrm{d} t
\\&\qquad\qquad+ 2c_{\psi}a\E_{X}\bracket{d(g(X), F) \exp\paren{-bd(g(X), F)^p}}.
\end{align*}
Let us begin by working on the first term.
We have, using the low-density separation hypothesis
\begin{align*}
& \int_{0}^\infty \Pbb_{X}\paren{d(g(X), F) < t} \exp\paren{-bt^2}\mathop{}\!\mathrm{d} t
\leq c_{\alpha} \int_{0}^\infty t^\beta \exp(-bt^p)\mathop{}\!\mathrm{d} t.
\\ &\qquad\qquad= b^{-\frac{1+\beta}{p}} c_{\alpha} \int_{0}^\infty (b^{1/p}t)^\beta \exp(-(b^{1/p}t)^p)\mathop{}\!\mathrm{d} (b^{1/p}t).
\\ &\qquad\qquad= b^{-\frac{1+\beta}{p}} c_{\alpha} \int_{0}^\infty t^\beta \exp(-t^p)\mathop{}\!\mathrm{d} t = c_{\alpha} \Gamma(\beta, p) b^{-\frac{1+\beta}{p}}.
\end{align*}
Let study the second term in the excess of risk inequality.
To enhance readability, write $\eta(X) = d(g(X), F)$.
We are left with studying $\E[\eta(X) \exp(-b\eta(X)^p)]$.
The function $t\to~t\exp(-bt^p)$ achieves its maximum in $t_0 = (pb)^{-1/p}$, increasing before and decreasing after.
Notice that the quantity
\begin{align*}
\Pbb(\eta(X) < t_0)\E_{X}\bracket{\eta(X) \exp(-b\eta(X)^p)\,\middle\vert\, \eta(X) < t_0}
& \leq c_{\alpha} t_0^{\beta + 1} \exp(-b t_0^p)
\\&= c_{\alpha} p^{-\frac{\beta + 1}{p}}\exp(-p^{-1/p}) b^{-\frac{\beta + 1}{p}},
\end{align*}
is exactly of the same order as the control we had on the first term in the excess of risk decomposition.
This suggests considering the following decomposition
\begin{align*}
& \E_{X}\bracket{\eta(X)\exp(-b\eta(X)^p)}
=
\Pbb(\eta(X) < t_0)\E_{X}\bracket{\eta(X) \exp(-b\eta(X)^p)\,\middle\vert\, \eta(X) < t_0}
\\&\quad\qquad +\sum_{i=0}^\infty \Pbb(2^i t_0\leq \eta(X) < 2^{i+1}t_0)\E_{X}\bracket{\eta(X) \exp(-b\eta(X)^p)\,\middle\vert\, 2^it_0 \leq \eta(X) < 2^{i+1}t_0}
\\&\qquad \leq c_{\alpha} t_o^{\beta+1} \exp(-bt_0^p)
+ \sum_{i=0}^\infty c_{\alpha} 2^\beta (2^i t_0)^{\beta + 1} \exp(-bt_0^p (2^i)^p)
\\&\qquad = c_{\alpha} t_o^{\beta+1} \paren{\exp(-p^{-1/p})
+ \sum_{i=0}^\infty 2^\beta 2^{i(\beta + 1)} \exp(-p^{-1/p} 2^{ip})}.
\end{align*}
The convergence of the last series ensures the existence of a constant $c'$ such that
\begin{align*}
\E_{X}\bracket{\eta(X) \exp(-b\eta(X)^p)}
\leq c' b^{-\frac{\beta+1}{p}}.
\end{align*}
Adding everything together ends the proof of this lemma.
Note that we have the same type of optimality as the one stated in Theorem \ref{rates:thm:low-density}.
\end{proof}
Because we use concentration inequalities for terms that are not necessarily centered, we usually get that~\eqref{rates:eq:concentration} only holds for $t > \epsilon_0$ where, typically $\epsilon_0 = \norm{\E_{{\cal D}_n}g_n(x) - g^*(x)}$, we can bypass this problem by adding in $\ind{t<\epsilon_0}$ in the probability, motivating the study leading to the following lemma.
\begin{lemma}[Handling bias in concentration inequality]
Under the assumptions of Theorem \ref{rates:thm:low-density}, if the concentration is not given by Assumption \ref{rates:ass:concentration} but given, for an $\epsilon_0 > 0$, by, for all $x\in\supp\rho_{\mathcal{X}}$ and $t > 0$,
\[
\Pbb_{{\cal D}_n}(\norm{g_n(x) - g^*(x)} > t) \leq \ind{t<\epsilon_0}.
\]
Then the excess of risk is controlled by
\[
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*) \leq 2c_{\psi}c_{\alpha} \epsilon_0^{\alpha + 1}.
\]
\end{lemma}
\begin{proof}
We retake the beginning of the proof of Theorem \ref{rates:thm:low-density}, and change its ending with
\begin{align*}
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*)
& \leq 2c_{\psi} \E_{{\cal D}_n, X}\bracket{\ind{\norm{g_n(X) - g(X)} \geq d(g(X), F)}\norm{g_n(X) - g(X)}}
\\& = 2c_{\psi}\int_{0}^\infty \Pbb_{{\cal D}_n, X}\paren{\ind{\norm{g_n(X) - g(X)} \geq d(g(X), F)}\norm{g_n(X) - g(X)} \geq t} \mathop{}\!\mathrm{d} t
\\& = 2c_{\psi}\int_{0}^\infty \E_{X}\Pbb_{{\cal D}_n}\paren{\norm{g_n(X) - g(X)} \geq \max\brace{t, d(g(X), F)}} \mathop{}\!\mathrm{d} t
\\& \leq 2c_{\psi} \int_{0}^\infty \E_{X} \ind{t<\epsilon_0}\ind{d(g(X), F) < \epsilon_0} \mathop{}\!\mathrm{d} t
= 2c_{\psi} \epsilon_0 \Pbb_{X}\paren{d(g(X), F) <\epsilon_0} \mathop{}\!\mathrm{d} t.
\end{align*}
This leads to the result after applying the $\alpha$-margin condition.
\end{proof}
\section{Introduction}
Machine learning is raising high hopes to tackle a wide variety of prediction problems, such as language translation, fraud detection, traffic routing, speech recognition, self-driving cars, DNA-binding proteins, {\em etc}.
Its framework is appreciated as it removes humans from the burden to come up with a set of precise rules to accomplish a complex task, such as recognizing a cat on an array of pixels.
Yet, it comes at a price, which is of forgetting about algorithm correctness, meaning that machine learning algorithms can make mistakes, {\em i.e.}, wrong predictions, which can have dramatic implications, {\em e.g.}, in medical applications.
This motivates work on generalization error bounds, quantifying how often one should expect errors.
Many of the problems discussed above are of discrete nature, in the sense that the number of potential outputs is finite, or infinite countable.
To learn such problems, a classical technique consists in defining a continuous surrogate problem, which is easier to solve, and such that:
\begin{itemize}
\item[(1)] an algorithm on the surrogate problem translates into an algorithm on the original problem;
\item[(2)] errors on the original problem are bounded by errors on the surrogate problem.
\end{itemize}
The first point refers to the concept of plug-in algorithms, while the second point to the notion of calibration inequalities.
For example, binary classification can be approached through regression by estimating the conditional expectation of the output $Y$ given an input $X$ \citep{Bartlett2006}.
On the one hand, continuous surrogates for discrete problems are interesting, as they benefit from functional analysis knowledge, when discrete problems are more combinatorial in nature.
On the other hand, continuous surrogates can be deceptive, as they are asking to solve for more than needed.
Considering the example of binary classification, where $Y\in\brace{-1, 1}$, one only has to predict the sign of the conditional expectation, rather than its precise value.
Interestingly, without modifying the continuous surrogate approach, this last remark can be leveraged in order to tighten generalization bounds derived through calibration inequalities \citep{Audibert2007}.
In this work, we extend those considerations, known in binary classification \citep[\emph{e.g.},][]{Koltchinskii2005, Chaudhuri2014}, to generic discrete supervised learning problems, and show how it can be applied to the kernel ridge regression algorithm introduced by \citet{Ciliberto2016}.
\subsection{Contributions}
Our contributions are organized in the following order.
\begin{itemize}
\item In Section \ref{rates:sec:structured}, we consider the general structured prediction from \citet{Ciliberto2020} based on Lipschitz-continuous losses and derive refined calibration inequalities to leverage the fact that learning a mapping into a discrete output space is easier than learning a mapping into a continuous space.
\item In Section \ref{rates:sec:margin}, we show how to exploit exponential concentration inequalities to turn them into fast rates under a condition generalizing the Tsybakov margin condition.
\item In Section \ref{rates:sec:nn}, we apply Section \ref{rates:sec:margin} to local averaging methods with the particular example of nearest neighbors.
This leads to extending the rates known for regression and classification to a wide variety of structured prediction problems, with rates that match minimax rates known in binary classification.
\item In Section \ref{rates:sec:rkhs}, we show how Section \ref{rates:sec:margin} can be applied to kernel ridge regression.
This allows us to improve rates known in $n^{-1/4}$ to arbitrarily fast rates depending on the hardness of the associated discrete problem.
\end{itemize}
\subsection{Related work}
\paragraph{Surrogate framework.}
The surrogate problem we will consider to tackle structured prediction finds its roots in the approximate Bayes rule proposed by \citet{Stone1977}, analyzed through the prism of mean estimation as suggested by \citet{Friedman1994} for classification, and analyzed by \citet{Ciliberto2020} in the wide context of structured prediction.
In particular, we will specify results on two classes of surrogate estimators: local averaging methods, or kernel ridge regression.
\paragraph{Local averaging methods.}
Neighborhood methods were first studied by \citet{Fix1951} for statistical testing through density estimation.
Similarly, Parzen–Rosenblatt window methods \citep{Parzen1962,Rosenblatt1956} were developed.
Those methods were cast in the context of regression as nearest neighbors \citep{Cover1957} and Nadayara-Watson estimators \citep{Watson1964,Nadaraya1964}.
\citet{Stone1977} was the first to derive consistency results for a large class of localized methods, among which are nearest neighbors and some window estimators \citep{Spiegelman1980,Devroye1980}.
Rates were then derived, with minimax optimality \citep{Stone1980,Yang1999}.
Several reviews can be found in the literature, such as \citet{Gyorfi2002,Tsybakov2009,Biau2015,Chen2018}.
\paragraph{Reproducing kernel ridge regression.}
The theory of real-valued reproducing kernel Hilbert spaces was formalized by \citet{Aronszajn1950}, before finding applications in machine learning \citep[\emph{e.g.},][]{Scholkopf2001}.
Minimax rates for kernel ridge regression were achieved by casting the empirical solution estimate as a result of integral operator approximation \citep{Smale2007,Caponnetto2007}, allowing to control convergence through concentration inequalities in Hilbert spaces \citep{Yurinskii1970,Pinelis1986} and on self-adjoint operator on Hilbert spaces \citep{Minsker2017}.
First derived in $L^{2}$-norm, rates were cast in $L^{\infty}$-norm through interpolation inequalities \citep[\emph{e.g.},][]{Fischer2020,Lin2020}.
\paragraph{Tsybakov margin condition.}
Learning a mapping into a discrete output space is indeed easier than learning a continuous mapping, as, for binary classification for example, one typically only has to predict the sign of $\E[Y\vert X]$ rather than its precise value.
As such, calibration inequalities that relate the error on a discrete structured prediction problem to an error on a smooth surrogate problem are often suboptimal.
This phenomenon was exploited for density discrimination, a problem consisting of testing if samples were drawn from one or the other of two potential distributions, by \citet{Mammen1999}, and for binary classification by \citet{Audibert2007}.
Those works introduce a parameter $\alpha\in[0, \infty)$ characterizing the hardness of the discrete problem, and leverage concentration inequalities to accelerate rates known for regression by a power $\alpha+1$ \citep{Audibert2007}, while rates plugged-in directly through calibration inequalities only present an acceleration by a power $\sfrac{2(\alpha+1)}{(\alpha + 2)}$ \citep[see, \emph{e.g.}][]{Boucheron2005,Bartlett2006,Bartlett2006b,Ervan2015,Nowak2019}.
\section{Structured prediction with surrogate control}
\label{rates:sec:structured}
In this section, we introduce the classical supervised learning problem, and a surrogate problem that consists of conditional mean estimation.
We recall a calibration inequality relating the original problem to the surrogate one.
We mention how empirical estimations of the conditional means usually deviate from the real means following a sub-exponential tail bound, similarly to bounds obtained through Bernstein inequality.
We end this section by providing refined surrogate control, that is the key toward ``superfast'' rates, that is, rates faster than $1/n$.
\subsection{Surrogate mean estimation}
Consider a classic supervised learning problem, where given an input space $\mathcal{X}$, an observation space~$\mathcal{Y}$, a prediction space $\cal Z$, a joint distribution $\rho \in \prob{\mathcal{X}\times\mathcal{Y}}$ and a loss function $\ell:{\cal Z}\times\mathcal{Y}\to\mathbb{R}_{+}$, one would like to retrieve $f^{*}:\mathcal{X}\to{\cal Z}$ minimizing the risk $\cal R$.
\[
f^{*} \in \argmin_{f:\mathcal{X}\to{\cal Z}} {\cal R}(f)
\qquad\text{with}\qquad
{\cal R}(f) = \E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)}.
\]
In practice, $\mathcal{X}$, $\mathcal{Y}$, ${\cal Z}$ and $\ell$ are givens of the problem, while $\rho$ is unknown, yet partially observed thanks to a dataset ${\cal D}_{n} = (X_{i}, Y_{i})_{i\leq n} \sim \rho^{\otimes n}$, with data $(X_{i}, Y_{i})$ sampled independently from $\rho$.
Note that in fully supervised learning, the observation space is the same as the prediction space $\mathcal{Y} = {\cal Z}$, yet we distinguish the two for our results to stand in more generic settings, such as instances of weak supervision \citep{Cabannes2020}.
In the following, we consider ${\cal Z}$ finite.
In several cases, solving the supervised learning problem can be done through solving a surrogate problem that is easier to handle.
\cite{Ciliberto2016} provides a setup that reduces a wide variety of structured prediction problems $(\ell, \rho)$ to a problem of mean estimation.
It works under the following assumption.
\begin{assumption}[Bilinear loss decomposition]\label{rates:ass:loss}
There exists a Hilbert space ${\cal H}$ and two mappings
$\psi:{\cal Z}\to{\cal H}$, $\phi:\mathcal{Y}\to{\cal H}$ such that
\[
\ell(z, y) = \scap{\psi(z)}{\phi(y)}.
\]
We will also assume that $\psi$ is bounded (in norm) by a constant $c_\psi$.
\end{assumption}
This assumption is not really restrictive \citep{Ciliberto2020}.
Among others, it works for any losses on finite spaces, usually with spaces $\cal H$ whose dimensionality is only polylogarithmic with respect to the cardinality of $\cal Z$ \citep{Nowak2019}.
Under Assumption \ref{rates:ass:loss}, solving the supervised learning problem can be done through estimating the surrogate conditional mean $g^{*}:\supp\rho_{\mathcal{X}}\to{\cal H}$, defined as
\begin{equation}
\label{rates:eq:gast}
g^{*}(x) = \E_{Y\sim\rho\vert_{x}}\bracket{\phi(Y)},
\end{equation}
where we denote $\rho\vert_{x}$ the conditional law of $\paren{Y\,\middle\vert\, X}$ under $(X, Y) \sim \rho$.
\begin{lemma}[\citet{Ciliberto2016}]\label{rates:lem:cal}
Given an estimate $g_{n}$ of $g^{*}$ in \eqref{rates:eq:gast}, consider the estimate $f_{n}:\mathcal{X}\to{\cal Z}$ of $f^{*}$, which is obtained from ``decoding'' $g_{n}$ as
\begin{equation}\label{rates:eq:decoding}
f_{n}(x) = \argmin_{z\in{\cal Z}} \scap{\psi(z)}{g_{n}(x)}.
\end{equation}
Then the excess risk is controlled through the surrogate error as
\begin{equation}
\label{rates:eq:L1}
{\cal R}(f_{n}) - {\cal R}(f^{*}) \leq 2c_\psi
\norm{g_{n} - g^{*}}_{L^{1}(\mathcal{X}, {\cal H}, \rho)}.
\end{equation}
\end{lemma}
Inequalities relating the original excess risk ${\cal R}(f_{n}) - {\cal R}(f^{*})$ with a measure of error on a surrogate problem are called \emph{calibration inequalities}.
They are useful when the measure of error between $g_{n}$ and $g^{*}$ is easier to control than the one between $f_{n}$ and $f^{*}$.
\begin{example}[Binary classification]\label{rates:ex:binary}
Binary classification corresponds to $\mathcal{Y}={\cal Z} = \brace{-1, 1}$ and $\ell(z, y) = \ind{z\neq y}$ (or equivalently $\ell(z, y) = 2\ind{z\neq y}-1$).
The classical surrogate consists of taking $\cal H = \mathbb{R}$, with $\phi=\textit{id}$ and $\psi=-\textit{id}$.
In this setting, we have $g^{*}(x) = \E_{\rho}[Y\vert X=x]$, and the decoding $f_{n}(x) := \sign g_{n}(x)$, for any $g_{n}(x) \in\cal H$.
In this case
\(
{\cal R}(f_{n}) - {\cal R}(f^{*})
= \E_{X}\bracket{\ind{f_{n}(X)\neq f^{*}(X)} \abs{g^{*}(X)}}
\leq 2\norm{g_{n} - g^{*}}_{L^{1}}
\leq 2\norm{g_{n} - g^{*}}_{L^{2}}.
\)
Note that in regression the excess risk reads as the square of the $L^{2}$ norm, explaining a loss of a power one half in convergence rates, when going from regression to classification \citep[\emph{e.g.}][]{Chen2018}.
\end{example}
Differences between an empirical estimate and its population version are generally handled through concentration inequalities.
In this work, we will leverage concentration on $\norm{g_{n}(x) - g(x)}$ that is uniform for $x \in \supp \rho_{\mathcal{X}}$, motivating the introduction of Assumption \ref{rates:ass:concentration}.
\begin{assumption}[Exponential concentration inequality]\label{rates:ass:concentration}
Suppose that for $n\in\mathbb{N}$, there exists two reals $L_n$ and $M_n$, such that the tails of $\norm{g_{n}(x) - g(x)}$ can be controlled for any $t > 0$ as
\begin{equation}
\label{rates:eq:concentration}
\sup_{x\in\supp\rho_{\mathcal{X}}} \Pbb_{{\cal D}_{n}}\paren{\norm{g_{n}(x) - g(x)} > t} \leq
\exp\paren{-\frac{L_n t^{2}}{1 + M_n t}}.
\end{equation}
\end{assumption}
Note that to satisfy Assumption \ref{rates:ass:concentration}, it is sufficient, yet \emph{not necessary}, to have a uniform control on $g_{n} - g^{*}$, {\emph i.e.}, a control on the tail of $\norm{g_{n} - g^{*}}_{L^{\infty}}$, since
\(
\sup_{x} \Pbb\paren{{A_{x} > t}}
\leq \Pbb\paren{\cup_{x} \brace{A_{x} > t}} = \Pbb\paren{\sup_{x} A_{x} > t},
\)
with $(A_{x})$ a family of random variables indexed by $x\in\mathcal{X}$.
Usually, in bounds like \eqref{rates:eq:concentration}, $M_n$ is a constant of the problem, while $L_n$ depends on the number of samples, therefore, we would like to give rates depending on $L_n$.
Typically, in Bernstein inequalities (see Theorem \ref{rates:thm:bernstein-vector} in Appendix), $L_n = n\sigma^{-2}$ with $\sigma^{2}$ a variance parameter and $M_n = c\sigma^{-2}$ with $c$ a constant of the problem that does not depend on $n$.
\subsection{Refined calibration}
While it is sufficient to control the excess risk through an $L^{1}$-norm control on $g$ from~\eqref{rates:eq:L1}, it is not always necessary.
In other words, the calibration bound in Lemma \ref{rates:lem:cal} is not always tight.
Indeed, we do not predict optimally, that is, $\brace{f_{n}(x) \neq f^{*}(x)}$ only if $g_{n}(x)$ and $g^{*}(x)$ do not lead to the same decoding $f_{n}(x)$ and $f^{*}(x)$.
When $\cal Z$ is finite, this is characterized by $g_{n}(x)$ and $g^{*}(x)$ not falling in the same region $R_z$ of ${\cal H}$, where
\[
R_z = \big\{\xi \in {\cal H} \big| z \in \argmin_{z' \in{\cal Z}}
\scap{\psi(z')}{\xi}\big\}.
\]
To ensure that $g_{n}(x)$ and $g^{*}(x)$ fall in the same region, one can ensure that $g_{n}(x)$ is closer to $g^{*}(x)$ than $g^{*}(x)$ is on the frontier of those regions.
Those frontiers are defined by points leading to at least two minimizers in~\eqref{rates:eq:decoding}:
\begin{align*}
F = \brace{\xi\in{\cal H} \,\middle\vert\,
\big|\argmin_{z\in{\cal Z}}\scap{\psi(z)}{\xi}\big| > 1}.
\end{align*}
The introduction of $F$ is motivated by the following geometric results.
\begin{lemma}[Refined surrogate control]\label{rates:lem:calibration}
When $\cal Z$ is finite, for any $x\in\supp\rho_{\mathcal{X}}$,
\[
\norm{g_{n}(x) - g^{*}(x)} < d(g^{*}(x), F) \qquad\Rightarrow\qquad f_{n}(x) = f^{*}(x),
\]
with $d$ the extension of the norm distance to sets as $d(g^{*}(x), F) = \inf_{\xi\in F} \norm{g^{*}(x) - \xi}$.
This result allows refining the calibration control from Lemma \ref{rates:lem:cal} as
\begin{equation}
\label{rates:eq:calibration}
{\cal R}(f_{n}) - {\cal R}(f^{*})
\leq 2c_{\psi} \E_{X} \bracket{ \ind{\norm{g_{n}(X) - g^{*}(X)} \geq d(g^{*}(X), F)}\norm{g_{n}(X) - g^{*}(X)}}.
\end{equation}
\end{lemma}
\begin{example}[Binary classification]\label{rates:ex:binary-2}
In binary classification (\emph{cf.} Example \ref{rates:ex:binary}), $F = \brace{0}$, and, for any $x\in\supp\rho_{\mathcal{X}}$, $d(g^{*}(x), F) = \abs{g^{*}(x)}$.
Lemma \ref{rates:lem:calibration} is based on the fact that $f^{*}(x) \neq f_{n}(x)$ implies that $\sign g^{*}(x) \neq \sign g_{n}(x)$ which itself implies that $\abs{g^{*}(x) - g_{n}(x)} = \abs{g^{*}(x)} + \abs{g_{n}(x)} \geq \abs{g^{*}(x)}$.
\end{example}
To leverage~\eqref{rates:eq:calibration}, we need to control $d(g^{*}(x), F)$ below and $\norm{g_{n}(x)- g^{*}(x)}$ above.
While upper bounds on $\norm{g_{n}(x)- g^{*}(x)}$ are assumed to have been derived through concentration inequalities, lower bounds on $d(g^{*}(x), F)$ will be assumed as a given parameter of the problem, see~\eqref{rates:eq:no-density}~and~\eqref{rates:eq:low-density}.
\begin{remark}[Scope of our work]
While we derived the refined calibration inequality~\eqref{rates:eq:calibration} for the surrogate conditional mean $g^*$ and the associated pointwise metric $\norm{\cdot}_{\cal H}$, similar inequality could be obtained for other types of surrogate methods.
This suggests that our work could be extended to any smooth surrogate such as the ones considered by \citet{Nowak2019b}, as well as Fenchel-Young losses \citep{Blondel2020}.
\end{remark}
\subsection{Geometric understanding}
In this subsection, we detail how to understand geometrically Lemma \ref{rates:lem:calibration}.
While the introduction of $\phi$ and $\psi$ could seem arbitrary, it can be thought in a more intrinsic manner by considering the embedding $\phi(y) = \delta_y$ belonging to the Banach space ${\cal H}$ of signed measured, $g^{*}(x) = \rho\vert_{x}$, with the bracket operator, for $\mu\in{\cal H}$ and $z\in{\cal Z}$, $\scap{z}{\mu} = \int_{\mathcal{Y}} \ell(z, y)\mu(\mathop{}\!\mathrm{d} y)$, and the distance between signed measures being $d(\mu_{1}, \mu_{2}) = \sup_{z\in{\cal Z}} \scap{z}{\mu_{1} - \mu_{2}}$.
Note that Lemma \ref{rates:lem:calibration} is a pointwise result, holding for any $x\in\mathcal{X}$, that is integrated over $\mathcal{X}$ afterwards.
Therefore, it is enough to consider $\mathcal{X} = \brace{x}$ and remove the dependency in $\mathcal{X}$ to understand it.
The simplex $\prob{\mathcal{Y}}$ naturally splits into the decision region $R_z$ for $z\in{\cal Z}$ as illustrated on Figure~\ref{rates:fig:separation}.
The main idea of Lemma \ref{rates:lem:calibration} is that one does not have to precisely estimate $g^*(x) = \rho\vert_x$ but only has to make sure that $g_n(x)$ falls in the same region on Figure~\ref{rates:fig:separation}.
\begin{figure}[ht]
\centering
\input{rates/drawings/frontier}
\caption{Illustration of Lemma \ref{rates:lem:calibration}.
Simplex $\prob{\mathcal{Y}}$, for $\mathcal{Y} = {\cal Z} = \brace{a, b, c}$ and $\ell$ a symmetric loss defined as $\ell(a, b) = \ell(a, c) = 1$ and $\ell(b, c) = 2$, while $\ell(z, z) = 0$.
This leads to the decision regions $R_z$ represented in colors.
Given $x\in\mathcal{X}$, if $g^*(x)$ corresponds to a distribution $\mu^* := \rho\vert_{x}$ falling in $R_a$, and if $g_n(x)$ represented by $\hat\mu$ falls closer to $\mu^*$ than the distance between $\mu^*$ and the decision frontier $F$ (represented by a circle on the right figure), then $\hat\mu$ is also in $R_a$, and therefore $f^*(x)=f_n(x)=a$.
}
\label{rates:fig:separation}
\vspace*{-.5cm}
\end{figure}
\section{Rate acceleration under margin condition}
\label{rates:sec:margin}
In this section, we introduce a condition that $g^{*}$ is not too often close to the decision frontier~$F$.
It generalizes the so-called ``Tsybakov margin condition'' known for classification.
Under this condition, we prove rates that generalize the results of \cite{Audibert2007} from binary classification to generic structured prediction problems, which opens the way to ``superfast'' rates in structured prediction.
\subsection{No density separation}
To get fast convergence rates, one has to make assumptions on the problem.
A classical assumption is that $g^{*}$ is smooth enough in order to get concentration bounds similar to Assumption \ref{rates:ass:concentration} when considering a specific class of estimates $g_{n}$.
In our decoding setting (Lemma \ref{rates:lem:cal}), learning is made easy when it is easy to estimate in which region $R_z$ the optimal $g^{*}$ will fall in.
This is in particular the case, when there is a margin $t_{0}>0$, for which, for no point $x \in \supp\rho_{\mathcal{X}}$, $g^{*}(x)$ falls at distance $t_{0}$ of the decision frontier $F$, motivating the following definition.
\begin{assumption}[No-density separation]\label{rates:ass:no-density}
A surrogate solution $g^{*}$ will be said to satisfy the \emph{no-density separation}, if there exists a $t_{0} > 0$, such that
\begin{equation}
\label{rates:eq:no-density}
\Pbb_{X}(d(g^{*}(X), F) < t_{0}) = 0.
\end{equation}
This condition is alternatively called the \emph{hard margin condition}, or sometimes ``Massart's noise condition'' for binary classification~\citep{Massart2006}.
\end{assumption}
\begin{figure}[ht]
\vspace*{-.5cm}
\centering
\input{rates/drawings/nodensity}
\caption{Illustration of Remark \ref{rates:rmk:separation}.
We represent two instances of binary classification (see Examples \ref{rates:ex:binary} and \ref{rates:ex:binary-2}).
On the left example, when $\rho_{\mathcal{X}}$ is such that there is no mass where the sign of $g^{*}$ changes, classes are separated in $\mathcal{X}$, yet the no-density separation is not verified.
On the right, classes are not separated in $\mathcal{X}$, but the problem satisfies the no-density separation as there is no $x$ such that $d(g^{*}(x), F) = \abs{g^{*}(x)} < t_{0}$.
Note that when $\rho_{\mathcal{X}}$ is uniform, the left problem satisfies a milder separation condition, introduced thereafter and called the $1$-low-density separation.}
\label{rates:fig:no-density}
\end{figure}
\begin{remark}[Separation in $\mathcal{Y}$ and separation in $\mathcal{X}$]\label{rates:rmk:separation}
It is important to realize that~\eqref{rates:eq:no-density} is a condition of separation in $\prob{\mathcal{Y}}$ that should hold for all $x\in\mathcal{X}$, but it does not state any separation between classes in $\mathcal{X}$ for $f^{*}:\mathcal{X}\to{\cal Z}$.
To visualize it, consider the classification problem where $\mathcal{X} = [-1,1]$, $\mathcal{Y} = {\cal Z} = \brace{-1, 1}$ and $\ell(z, y) = \ind{z\neq y}$.
\begin{itemize}
\item A situation where $\rho_{\mathcal{X}}$ is uniform on $\mathcal{X}$ and $\E[Y\vert X=x] = 2\cdot\ind{x\in p\mathbb{N} + \brace{a\,\middle\vert\, \abs{a} < p/4}} - 1$, for $p = 1/50$, satisfies separation in $\prob{\mathcal{Y}}$ \eqref{rates:eq:no-density}, but classes are not separated in $\mathcal{X}$.
\item A situation where $\rho$ is uniform on $[-1, -.5] \cup [.5, 1]$, with $\E[Y\vert X=x] = \sign(x) (1 - \abs{x})^{p}$, for $p>0$, satisfies a separation of classes in $\mathcal{X}$ but does not satisfy \eqref{rates:eq:no-density}.
\end{itemize}
%
Note that continuity of $g^{*}$ and the no-density separation in~\eqref{rates:eq:no-density} imply separation of classes in $\mathcal{X}$.
Note also that to get concentration inequality such as~\eqref{rates:eq:concentration}, one usually supposes that $g^{*}$ is smooth.
We refer the curious reader to Section 2.4 in \citet{Steinwart2007} for separation in $\mathcal{X}$.
\end{remark}
The introduction of Assumption \ref{rates:ass:no-density} is motivated by the following result.
\begin{theorem}[Rates under no-density separation]\label{rates:thm:no-density}
When $\ell$ is bounded by $\ell_{\infty}$ ({\em i.e.}, $\ell(z, y) \leq \ell_{\infty}$ for any $(z, y) \in {\cal Z} \times \mathcal{Y}$) and satisfies Assumption \ref{rates:ass:loss}, and $\mathcal{Z}$ is finite, under the no-density separation Assumption \ref{rates:ass:no-density}, and the concentration Assumption \ref{rates:ass:concentration}, the excess risk is controlled
\[
\E_{{\cal D}_{n}}{\cal R}(f_{n}) - {\cal R}(f^{*}) \leq \ell_{\infty} \exp\paren{-\frac{L_n t_0^2}{1 + M_n t_0}}.
\]
\end{theorem}
\begin{proof}
Because we make a mistake only when $d(g^{*}(x), F) \geq \norm{g_{n}(x) - g^{*}(x)}$, we make no mistake when $\norm{g_{n}(x) - g^{*}(x)} < t_{0}$; otherwise we can consider the worse error we are going to pay, that is~$\ell_{\infty}$, leading to
\[
{\cal R}(f_{n}) - {\cal R}(f^{*}) \leq \ell_{\infty} \Pbb_{X}(\norm{g_{n}(x) - g^{*}(x)} > t_{0}).
\]
Taking the expectation with respect to ${\cal D}_{n}$ and using the fact that \(\E_A\Pbb_B(Z) = \E_A\E_B[\ind{Z}] = \E_B\E_A[\ind{Z}] = \E_B\Pbb_A(Z)\), and plug-in the concentration inequality~\eqref{rates:eq:concentration}, we get the result.
\end{proof}
\begin{example}[Image classification]
In image classification, one can arguably assume that the class of an image is a deterministic function of this image.
With the 0-1 loss, it implies that the image classification problem verifies the no-density separation.
The same holds for any discrete problem where the label is a deterministic function of the input.
Based on Theorem \ref{rates:thm:no-density} and~\eqref{rates:eq:concentration} in which $M$ is generally a constant when $L$ is proportional to the number of data, it is reasonable to ask for exponential convergences rates on such problems.
\end{example}
\subsection{Low density separation}
While we presented the no-density separation first for readability, it is a strong assumption.
Recall our example, Remark \ref{rates:rmk:separation}, with $\E[Y\vert X=x] = \sign(x)(1 - \abs{x})^{p}$, only around $x=1$ and $x=-1$ is $d(g^{*}(x), F)$ not bounded away from zero.
While the neighborhood of those points should be studied carefully, the error on all other points $x \in [-1+t, 1-t]$ can be controlled with exponential rates.
The low-density separation, also known as the Tsybakov margin condition in binary classification, will allow a refined control to get fast rates in such a setting.
\begin{assumption}[Low-density separation]\label{rates:ass:low-density}
A surrogate solution $g^{*}$ is said to satisfy the \emph{low-density separation}, if there exists $c_{\alpha} > 0$, and $\alpha > 0$, such that for any $t > 0$
\begin{equation}
\label{rates:eq:low-density}
\Pbb_{X}(d(g^{*}(X), F) < t) \leq c_{\alpha} t^{\alpha}.
\end{equation}
This condition is alternatively called the \emph{margin condition}.
\end{assumption}
The low-density separation spans all situations from the hard margin condition, that can be seen as $\alpha = +\infty$, to situations without any margin assumption corresponding to $\alpha = 0$.
The coefficient~$\alpha$ is an intrinsic measure of the easiness of finding $f^{*}$ in the problem $(\ell, \rho)$.
For example, the setting described in the last paragraph corresponds to the case $\alpha = 1/p$.
We discuss the equivalence of Assumption \ref{rates:ass:low-density} to definitions appearing in the literature in Remark \ref{rates:rmk:low-density}.
\begin{theorem}[Optimal rates under low density separation]\label{rates:thm:low-density}
Under refined calibration in~\eqref{rates:eq:calibration}, concentration (Assumption \ref{rates:ass:concentration}), and low-density separation (Assumption \ref{rates:ass:low-density}), the risk is controlled as
\[
\E_{{\cal D}_{n}} {\cal R}(f_{n}) - {\cal R}(f^{*}) \leq
2c_{\psi} c_{\alpha} c \paren{M_n^{\alpha + 1}L_n^{-(\alpha + 1)} + L_n^{-\frac{\alpha + 1}{2}}},
\]
for $c$ a constant that only depends on $\alpha$, that can be expressed through the Gamma function evaluated in quantity depending on $\alpha$, meaning that when $\alpha$ is big, $c$ behaves like $\alpha^{\alpha}$.
Note that it is not possible to derive a better bound only given~\eqref{rates:eq:concentration}, \eqref{rates:eq:calibration} and \eqref{rates:eq:low-density}.
\end{theorem}
\begin{proof}[Sketch for Theorem \ref{rates:thm:low-density}, details in Appendix \ref{rates:proof:low-density}]
Based on the refined calibration inequality in~\eqref{rates:eq:calibration}, and using that $\E[X] = \int_{0}^{\infty} \Pbb(X > t)\mathop{}\!\mathrm{d} t$, it is possible to show that the expectation of the excess risk behave like
\[
\int_{0}^{\infty} \Pbb_{X}(d(g^{*}(x), F) < t) \sup_{x} \Pbb_{{\cal D}_{n}}(\norm{g_{n}(x)
- g^{*}(x)} > t)\mathop{}\!\mathrm{d} t.
\]
Based on Assumptions \ref{rates:ass:concentration} and \ref{rates:ass:low-density}, the integrand behaves like $t^{\alpha} \exp(- L_n t^{2} /(1+M_n t))$.
A change of variable and the study of the Gamma function leads to the result.
We provide all the details in Appendix \ref{rates:proof:low-density}.
Note that while we stated Theorem \ref{rates:thm:low-density} under an exponential inequality of Bernstein type (Assumption \ref{rates:ass:concentration}), similar theorems can be derived for any type of exponential concentration inequality, as stated in Lemma \ref{rates:lem:ref-low-density} in Appendix \ref{rates:app:ref-low-density}.
\end{proof}
Theorem \ref{rates:thm:low-density} is to put in perspective with the work of \citet{Nowak2019} which considers the same setup as ours, yet only succeeds to derive acceleration by a power $2(\alpha+1) / (\alpha + 2)$, while we got an acceleration by a power $(\alpha + 1) / 2$ as already mentioned in the related work section.
This gain will appear more clearly in Theorem \ref{rates:thm:krr-low-density}.
\begin{remark}[Independence to the decomposition of $\ell$]\label{rates:rmk:low-density}
While we have stated results based on the quantity $d(g^{*}(x), F)$, generalization of the Tsybakov margin condition has also been expressed through the quantity $\inf_{z\neq z^{*}} \E_{Y\sim \rho\vert_{x}}\ell(z, Y) - \E_{Y\sim \rho\vert_{x}}\ell(z^{*}, Y)$ instead of $d(g^{*}(x), F)$ \citep{Nowak2019}.
We show in Appendix \ref{rates:proof:def-low-density} that the two definitions of the margin condition are equivalent.
\end{remark}
\begin{remark}[Scope of our work]\label{rates:rmk:l2-pointwise}
Our work relies on pointwise exponential concentration inequalities (Assumption \ref{rates:ass:concentration}) which are specially designed to work well with the Tsybakov margin condition.
It is natural for localized averaging methods such as nearest neighbors, or for surrogate methods leading to $L^\infty$ concentration.
For surrogate methods leading to concentration of other quantities, it is possible to use similar tricks under different ``margin'' conditions ({\em e.g.} \citet{Steinwart2007} for a margin condition designed for the Hinge loss).
Note that $L^2$ concentration on $g_n$ toward $g^*$ (such as the one derived by \citet{MarteauFerey2019} for logistic regression) could also be turned into fast convergence of $f_n$ toward $f^*$, since, in essence, for points $x\in\mathcal{X}$ where $\rho(\mathop{}\!\mathrm{d} x)$ is high, the quantity $g^*(x) - g_n(x)$ will have a non-negligible contribution to $\norm{g^* - g_n}_{L^2}$ -- allowing to cast concentration in $L^2$ to concentration pointwise in $x$ -- and for points $x\in\mathcal{X}$ where $\rho(\mathop{}\!\mathrm{d} x)$ is negligible, it is acceptable to pay the worst error, since it will have a small contribution to the excess of risk.
Finally, note that it is also possible to let the right hand-side term in~\eqref{rates:eq:concentration} depends on $x$, and to modify Theorem \ref{rates:thm:no-density} with $L = \E[L(X)]$.
\end{remark}
\subsection{The importance of constants}
In this subsection, we discuss the importance of constants when providing learning rates.
Assumption \ref{rates:ass:no-density} corresponds to asking for $g^*(x)$ never to enter a neighborhood of $F$ defined through~$t_0$.
Similarly, when $\mathcal{X}$ is parametrized such that $\rho_{\mathcal{X}}$ is uniform, the parameter $\alpha$ in Assumption~\ref{rates:ass:low-density} corresponds to the speed at which $g^*(x)$ ``get through'' the decision frontier $F$.
In order to have a higher $\alpha$ and optimize the dependency in $n$ in the bound of Theorem \ref{rates:thm:low-density}, it is natural to think of infinitesimal perturbations of $g^*$ to make it cross the boundary orthogonally (or even jump over it and satisfy the no-density separation).
To give a precise example, in binary classification, let us artificially add smoothness to the function $g^*(x) = x^q$ when approaching zero.
Consider $g^*:[0, 1]\to[-1,1], x\to c^{q-p}x^p\ind{x<c} + x^q\ind{x\geq c}$, and $x$ uniform, and $p < q$.
In this setting, $\alpha$ can be taken anywhere in $[0, p^{-1})$.
Naively, we could ask for the biggest possible $\alpha$ in order to have the best dependency in $n$ in the learning rates given by Theorem \ref{rates:thm:low-density}.
While this approach will higher $\alpha$, it will also higher $c_\alpha$, compensating the gain one could expect from such a strategy.
Indeed, for $\alpha \in [0, p^{-1}]$, at best, we can take $c_\alpha = \ind{\alpha < q^{-1}} + c^{1-q\alpha}\ind{\alpha \geq q^{-1}}$.
This shows the importance of optimizing both $\alpha$ and $ c_\alpha c$ to minimize the lower bound appearing in Theorem \ref{rates:thm:low-density} when given a fixed number of sample $n$.
In a word, while we only give results that are optimized in $n$, when $n$ is fixed, better bounds could be given by optimizing parameters and constants simultaneously.
For example, when $\mathcal{X} = \mathbb{R}^d$ and $g^*$ belongs to the Sobolev space $H^m$ for all $m \in [0, m_*]$, and satisfies Assumption \ref{rates:ass:low-density} for all $\alpha \in [0, \alpha_*]$, we expect the best bound, that could be derived from our proof technique, to be of form
\[
\E_{{\cal D}_n}{\cal R}(f_n) - {\cal R}(f^*) \leq
\min_{m \leq m_*, \alpha < \alpha_*} \alpha^\alpha c_\alpha c_\psi\norm{g^*}_{H^m}
n^{-\frac{m(\alpha + 1)}{2 d}}.
\]
Yet, for simplicity, we will express those bounds as $b n^{-\frac{m_*(\alpha_* + 1)}{2d}}$, for $b$ a big constant.
\section{Application to nearest neighbors}
\label{rates:sec:nn}
In this section, we consider the Bayes approximate risk estimator proposed by \citet{Stone1977}, with weights given by nearest neighbors \citep{Cover1957}.
We prove, under regularity assumptions, concentration inequalities similar to~\eqref{rates:eq:concentration}, which allow us to derive exponential and polynomial rates.
Given samples $(X_{i},Y_{i}) \sim \rho^{\otimes n}$, $k \in \mathbb{N}$ and a metric $d$ on $\mathcal{X}$, the estimator is
\begin{equation}\label{rates:eq:nn}
\!\!\! g_{n}(x) = \sum_{i=1}^{n} \alpha_{i}(x) \phi(Y_{i}),\ \
\text{with}\ \
\alpha_{i}(x) = \left\{
\begin{array}{cl}
k^{-1} & \text{if}\quad \sum_{j=1}^{n} \ind{d(x, X_{j}) \leq d(x, X_{i})} < k \\
0 & \text{if}\quad \sum_{j=1}^{n} \ind{d(x, X_{j}) < d(x, X_{i})} \geq k \\
(pk)^{-1} & \text{else, with } p = \sum_{j=1}^{n} \ind{d(x, X_{j}) = d(x, X_{i})}.
\end{array}\right.
\end{equation}
To study the convergence of $g_{n}$, we introduce the noise free estimator $g_{n}^{*} = \sum_{i=1}^{n} \alpha_{i}(x)g^{*}(X_{i})$.
This allows separating the error due to the randomness of the labels $Y_{i} \sim \rho\vert_{X_{i}}$, and the error due to the difference between $g^{*}(x)$ and the averaging of $g^{*}$ on the neighbors of $x$ defining $g_{n}$.
To control the fist error, we need a bounded moment condition on $\phi(Y)$.
We reuse an assumption from \citet{Bernstein1924}, that is classic in machine learning \citep[\emph{e.g.},][]{Caponnetto2007,Lin2020}.
\begin{assumption}[Sub-exponential moment of $\rho\vert_{x}$]\label{rates:ass:moment}
Suppose that there exists $\sigma^{2}, M > 0$ such that for any $x\in\supp\rho_{\mathcal{X}}$, for any $m \geq 2$, we have
\[
\E_{Y\sim\rho\vert_{x}}\bracket{\norm{\phi(Y) - g^{*}(x)}^{m}}
\leq \frac{1}{2} m! \sigma^{2} M^{m-2}.
\]
\end{assumption}
\begin{example}[Moment bound on $\phi(Y)$]
Assumption \ref{rates:ass:moment} is a classical assumption that is notably satisfied when $\phi(Y)$ is bounded by $M$, with $\sigma^2$ its variance, or when $\paren{\phi(Y)\,\middle\vert\, X}$ is Gaussian with covariance bounded by a constant independent of $X$ \citep[see a proof of this standard result by][]{Fischer2020}.
\end{example}
To control the second error, we notice, for $x\in\supp\rho_{\mathcal{X}}$, that the quantity $\norm{g^{*}(x) - g_{n}^{*}(x)}$ behaves like $\sup_{x'\in{\cal B}(x, r)} \norm{g^{*}(x) - g^{*}(x')}$, with $r$ such that $\rho_{\mathcal{X}}({\cal B}(x, r)) \approx k/n$, such an $r$ modeling the distance between $x$ and its $k$-th neighbor.
This motivates the following assumption.
\begin{assumption}[Modified Lipschitz condition \citep{Chaudhuri2014}]\label{rates:ass:lipschitz}
$g^{*}$ is said to verify the $\beta$-Modified Lipschitz condition if there exists $c_{\beta} > 0$ such that for any $x, x' \in \supp\rho_{\mathcal{X}}$
\[
\norm{g^{*}(x) - g^{*}(x')} \leq c_{\beta} \rho_{\mathcal{X}}({\cal B}(x, d(x,x')))^{\beta},
\]
where $d$ is the distance on $\mathcal{X}$, and ${\cal B}(x, t)\subset \mathcal{X}$ the ball of center $x$ and radius $t$.
\end{assumption}
Typically, the $\beta$ that appears in Assumption \ref{rates:ass:lipschitz} is linked with the dimension of a subset of $\mathcal{X}$ containing most of the mass of $\rho_{\mathcal{X}}$ (see below).
This will slow the rates according to this dimension parameter, a property referred to as the curse of dimensionality.
\begin{example}[Classical assumptions]\label{rates:ex:nn}
When $\mathcal{X} = \mathbb{R}^{d}$, if $g$ is $\beta'$-H\"older continuous, and $\rho_{\mathcal{X}}$ is regular in the sense that, there exists a constant $c$ and $t^{*} > 0$ such that for $x\in\supp\rho_{\mathcal{X}}$ and any $t \in [0, t^{*}]$,
\(
\rho_{\mathcal{X}}({\cal B}(x, t)) \geq c \lambda({\cal B}(x, t)),
\)
with $\lambda$ the Lebesgue measure on $\mathcal{X}$, then $g$ satisfies the modified Lipschitz condition with $\beta = \beta' / d$.
The condition on $\rho_{\mathcal{X}}$ is usually split in a condition of minimal mass of $\rho_{\mathcal{X}}$, and a condition of regular boundaries of $\supp\rho_{\mathcal{X}}$ \citep[\emph{e.g.},][]{Audibert2007}.
We provide more details in Appendix \ref{rates:app:nn-assumptions}.
\end{example}
We now state convergence results, respectively proven in Appendices \ref{rates:proof:nn-concentration}, \ref{rates:proof:nn-no-density} and \ref{rates:proof:nn-low-density}, in which the constant values $b_1$ to $b_6$ appear explicitly.
Note that results provided by Lemma \ref{rates:lem:nn-concentration} are already known in the literature \citep{Gyorfi2002}, while Theorems \ref{rates:thm:nn-no-density} and \ref{rates:thm:nn-low-density} were only known in binary classification, but we generalize them to any discrete structured prediction problem.
It should be noted that rates in Theorem \ref{rates:thm:nn-low-density} match the minimax rates derived by \cite{Audibert2007} in the case of binary classification.
\begin{lemma}[Nearest neighbors concentration]\label{rates:lem:nn-concentration}
Under Assumptions \ref{rates:ass:moment} and \ref{rates:ass:lipschitz}, there exist constants $b_{1}, b_{2}, b_{3} > 0$, such that for any $x\in\supp\rho_{\mathcal{X}}$ and any $t > 0$,
\[
\Pbb_{{\cal D}_{n}}\paren{\norm{g_{n}(x) - g_{n}^{*}(x)} > t} \leq
2\exp\paren{-\frac{b_{1}kt^{2}}{1 + b_{2}t}}.
\]
And for $t > \paren{k/2n}^{\beta}$, when $\rho_{\mathcal{X}}$ is continuous\footnote{Note that this topological assumption ease derivations but is not fundamental for such non-asymptotic results.}
\[
\Pbb_{{\cal D}_{n}}\paren{\norm{g_{n}^{*}(x) - g^{*}(x)} > t} \leq
\exp\paren{-b_{3} nt^{\frac{1}{\beta}}}.
\]
\end{lemma}
\begin{theorem}[Nearest neighbors fast rates under no-density assumption]
\label{rates:thm:nn-no-density}
When $\ell$ is bounded by $\ell_{\infty}$, satisfies Assumption \ref{rates:ass:loss}, and $\mathcal{Z}$ is finite, under the no-density separation, Assumption \ref{rates:ass:no-density}, and Assumptions \ref{rates:ass:moment} and \ref{rates:ass:lipschitz}, there exist two constants $b_{4}, b_{5} > 0$ that do not depend on $n$ or $k$ such that for any $n \in \mathbb{N}^{*}$ and any $k$ such that $\paren{k/2n}^{\beta} < t_{0}$, we have
\begin{equation}
\E_{{\cal D}_{n}}{\cal R}(f_{n}) - {\cal R}(f^{*}) \leq
2\ell_{\infty}\exp(-b_{4}k) + \ell_{\infty}\exp(-b_{5}n).
\end{equation}
\end{theorem}
\begin{theorem}[Nearest neighbors fast rates under low-density assumption]
\label{rates:thm:nn-low-density}
When $\ell$ satisfies Assumption \ref{rates:ass:loss}, and ${\cal Z}$ is finite, under the low-density separation, Assumption \ref{rates:ass:low-density}, and Assumptions \ref{rates:ass:moment} and \ref{rates:ass:lipschitz}, considering the scheme $k_{n} = \floor{k_{0}n^{\frac{2\beta}{2\beta + 1}}}$, for any $k_{0} > 0$, there exists a constant $b_{6} > 0$ that does not depend on $n$ such that for any $n \in \mathbb{N}^{*}$,
\begin{equation}\label{rates:eq:nn_rates}
\E_{{\cal D}_{n}}{\cal R}(f_{n}) - {\cal R}(f^{*}) \leq
b_{6} n^{-\frac{\beta(\alpha + 1)}{2\beta + 1}}.
\end{equation}
\end{theorem}
\begin{remark}[Scope of our work]
The same type of argument works for other local averaging methods, such as Nadaraya-Watson \citep{Nadaraya1964,Watson1964}, local polynomials \citep{Cleveland1979,Audibert2007} or decision trees \citep{Breiman1984}.
\end{remark}
\begin{figure}[ht]
\vspace*{-.5cm}
\centering
\includegraphics{rates/drawings/rates-eps-converted-to.pdf}
\caption{Empirical convergence rates.
We consider binary classification, with $\mathcal{X} = [-1, 1]$, $g^{*}(x) = \sign(x)*\abs{x}^{\frac{1}{\alpha}}$, for $\alpha \in \brace{.1, 1}$ and $\rho_{\mathcal{X}}$ uniform.
We plot in solid the logarithm of the excess risk averaged over 100 trials against the logarithm of the number of samples for $n \in [10, 10^6]$, and plot in dashed the expected slope of those curves due to Theorem \ref{rates:thm:nn-low-density} (\emph{i.e.}, we fit the constant $C$ in the rate $Cn^{-\gamma}$ with $\gamma$ obtained from the bound in~\eqref{rates:eq:nn_rates}).}
\label{rates:fig:rates}
\end{figure}
\section{Application to reproducing kernel ridge regression}
\label{rates:sec:rkhs}
In this section, we consider the kernel ridge regression estimate $g_{n}$ of $g^{*}$ first proposed by \cite{Ciliberto2016}, and we prove, under regularity assumptions, uniform concentration inequalities similar to~\eqref{rates:eq:concentration}, which allow us to derive superfast rates at the end of the section.
Given a symmetric, positive semi-definite kernel $k:\mathcal{X}\times\mathcal{X}\to\mathbb{R}_+$, the kernel ridge regression estimation $g_{n}$ of $g^{*}$ is defined similarly to~\eqref{rates:eq:nn} yet with weights $\alpha(x) \in \mathbb{R}^{n}$ defined as
\[
\alpha(x) = (\hat K + \lambda)^{-1} \hat K_{x},\quad
\hat K = \paren{\frac{1}{n} k(X_{i}, X_{j})}_{i,j\leq n} \in \mathbb{R}^{n\times n},\
\hat K_{x} = \paren{\frac{1}{n} k(x, X_{i})}_{i\leq n} \in \mathbb{R}^{n}.
\]
To state regularity assumptions, we introduce a minimal setup linked to the reproducing kernel $k$.
To keep the exposition clear, we relegate technicalities in Appendix \ref{rates:app:rkhs}.
We define the operator $K$ operating on functions $f\in L^{2}(\mathcal{X}, {\cal H}, \rho_{\mathcal{X}})$ and $K_{\mathcal{X}}$ operating on $f\in L^{2}(\mathcal{X}, \mathbb{R}, \rho_{\mathcal{X}})$, both defined as
\[
(Kf)(x') = \int_{\mathcal{X}} k(x', x)f(x)\mathop{}\!\mathrm{d}\rho_{\mathcal{X}}(x).
\]
Inheriting from the symmetry and positive semi-definiteness of $k$, $K$ is self-adjoint with the spectrum in $\mathbb{R}_+$.
To study the convergence of $g_{n}$ to $g^{*}$, it is useful to introduce the approximate orthogonal projection on $\ima K^{\frac{1}{2}}$, defined for $\lambda > 0$ as
\[
g_\lambda = K (K+\lambda)^{-1} g^{*}.
\]
We introduce three assumptions linked with the regularity of the problem, referred to as the capacity condition, interpolation inequality and source condition.
Those are classical assumptions to prove uniform rates of the kernel ridge regression estimates.
They could be found, in particular, by \citet{Fischer2020} under the respective names of (EVD), (EMB) and (SRC), but also by \citet{PillaudVivien2018,Lin2020}.
Our assumptions differ in that they are expressed for vector-valued functions, which usually generate compactness issues \citep{Caponnetto2007}.
However, when ${\cal Z}$ is finite, ${\cal H}$ is finite dimensional, and $K$ can be shown to be a compact operator, which allows considering fractional power without definition issues.
\begin{assumption}[Capacity condition]\label{rates:ass:capacity}
Suppose $\trace\paren{K_{\mathcal{X}}^{\sigma}} < +\infty$ for
$\sigma \in [0, 1]$.
\end{assumption}
\begin{assumption}[Interpolation inequality]\label{rates:ass:interpolation}
Assume the existence of $p \in [0,\frac{1}{2}]$, $c_p > 0$ such that
\[
\forall\ g\in \paren{\ker K}^{\perp}, \qquad
\norm{K^{p}g}_{L^{\infty}} \leq c_p\norm{g}_{L^{2}}.
\]
\end{assumption}
\begin{assumption}[Source condition]\label{rates:ass:source} Suppose
$g^{*} \in \ima K^{q}$ for $q \in (p, 1]$.
\end{assumption}
When $q=\sfrac{1}{2}$, the source condition is often expressed as $g^{*}$ belonging to the reproducing kernel Hilbert space associated to the kernel $k$.
Note that when $k$ is bounded, Assumptions~\ref{rates:ass:capacity} and~\ref{rates:ass:interpolation} hold with $\sigma=1$ and $p=\sfrac{1}{2}$.
In those assumptions, for $p$ and $\sigma$ the smaller, and for $q$ the bigger, the faster the convergence rates will be.
\begin{example}[Classical assumptions]
For Assumption \ref{rates:ass:interpolation} to hold, minimal mass and regular support of $\rho$, similarly to Example \ref{rates:ex:nn}, are often assumed, as well as regularity of functions in $\ima K^{p}$, in coherence with Remark \ref{rates:rmk:l2-pointwise}.
For Assumption \ref{rates:ass:source} to hold, it is classical to assume regularity of $g^{*}$, matching the regularity of function spaces derived from the kernel $k$.
The value of $\sigma$ in Assumption \ref{rates:ass:capacity} often comes as a bonus of regularity assumptions on $\rho$ and specificity of the RKHS implied by $k$.
See Example 2 by \citet{PillaudVivien2018} and Section 4 by \citet{Fischer2020} as well as references therein for concrete examples.
\end{example}
We now state convergence results respectively proven in Appendices \ref{rates:proof:krr-1} and \ref{rates:proof:krr-2}, \ref{rates:proof:krr-no-density}, and \ref{rates:proof:krr-low-density}.
Lemma \ref{rates:lem:rkhs} is a generalization to vector-valued functions of kernel ridge regression uniform convergence rates known for real-valued function \citep[see][]{Fischer2020}.
Note that a similar result to Theorem \ref{rates:thm:krr-no-density} was provided for binary classification by \citet{Koltchinskii2005}, but we generalize exponential rates with kernel ridge regression to any discrete structured prediction problem.
Theorem \ref{rates:thm:krr-low-density} is new, even in the context of binary classification.
It states that, while, up to now, only rates in $n^{-1/4}$ were known for $f_{n}$ \citep{Ciliberto2020}, one can indeed hope for arbitrarily fast rates, depending on the hardness of the problem, read in the value of $\alpha\in[0, \infty)$.
\begin{lemma}[Reproducing kernel concentration]\label{rates:lem:rkhs}
Under Assumptions \ref{rates:ass:capacity}, \ref{rates:ass:interpolation} and \ref{rates:ass:source}, for any $\lambda > 0$,
\[
\norm{g_\lambda - g^{*}}_{L^{\infty}} \leq b_{1} \lambda^{q-p}.
\]
With $b_{1} = c_p\norm{K^{-q}g^{*}}_{L^{2}}$.
Moreover, when the kernel $k$ is bounded and under Assumption \ref{rates:ass:moment}, there exists three constants $b_{2}, b_{3}, b_{4}, b_{5} > 0$ that does not depend nor on $\lambda$ nor on $n$ such that
\[
\Pbb\paren{\norm{g_{n} - g_\lambda}_{\infty} > t} \leq
b_{2} \lambda^{-\sigma}\exp\paren{-b_{3}n\lambda^{2p}}
+ 4\exp\paren{-\frac{n\lambda^{2p + \sigma}t^{2}}
{b_{4}+b_{5}t}}.
\]
As long as $b_{3}n \geq \lambda^{-p}$, and $\lambda \leq \min\paren{\norm{K}_{\op}, 1}$.
\end{lemma}
\begin{theorem}[Kernel ridge regression fast rates under no-density assumption]
\label{rates:thm:krr-no-density}
When the loss $\ell$ is bounded, satisfies Assumption \ref{rates:ass:loss} and ${\cal Z}$ is finite, under the $t_{0}$-no-density separation condition, and Assumptions \ref{rates:ass:moment}, when $k$ is bounded, if $\lambda_{n} = \lambda$, for any $\lambda > 0$ such that $\norm{g^{*} - g_\lambda}_{L^{\infty}} < t_{0}$, then there exist two constants $b_{6}, b_{7} > 0$ such that, for any $n\in\mathbb{N}^{*}$,
\begin{equation}
\E_{{\cal D}_{n}} {\cal R}(f_{n}) - {\cal R}(f^{*}) \leq b_{6} \exp(-b_{7} n),
\end{equation}
with $f_{n}$ given by the kernel ridge regression surrogate estimate.
\end{theorem}
\begin{theorem}[Kernel ridge regression fast rates under low-density assumption]
\label{rates:thm:krr-low-density}
When $\ell$ satisfies Assumption \ref{rates:ass:loss}, is bounded and ${\cal Z}$ is finite, under the $\alpha$-low-density separation condition, and Assumptions \ref{rates:ass:moment}, \ref{rates:ass:capacity}, \ref{rates:ass:interpolation} and \ref{rates:ass:source}, if $\lambda_{n} = \lambda_{0} n^{-\frac{1}{2q + \sigma}}$, for any $\lambda_{0} > 0$, there exists $b_{8} > 0$, such that for any $n \in \mathbb{N}^{*}$,
\begin{equation}
\E_{{\cal D}_{n}} {\cal R}(f_{n}) - {\cal R}(f^{*}) \leq b_{8} n^{-\frac{(q-p)(1+\alpha)}{2q + \sigma}}.
\end{equation}
\end{theorem}
\section{Conclusion}
In this paper, we have shown how, for discrete problems, to leverage exponential concentration inequalities derived on continuous surrogate problems, in order to derive faster rates than rates directly obtained through calibration inequalities.
Those rates are arbitrarily fast, depending on a parameter characterizing the hardness of the discrete problem.
We have shown how this method directly applies to local averaging methods and to kernel ridge regression, which allows deriving ``superfast'' rates for any discrete structured prediction problem.
This opens the way to several follow-up, such as
\begin{itemize}
\item Application follow-up, consisting of tackling concrete problem instances, such as predicting properties of DNA-sequence \citep{Jaakkola2000}, \emph{e.g.}, gene mutations responsible for diseases, with well-designed kernels on DNA in order to higher the exponent appearing in Theorem \ref{rates:thm:krr-low-density}.
\item Computational follow-up, pushing our analysis further to understand how to design better algorithms on discrete problems.
For example, by adding a regularization pushing $g_{n}$ away from the decision frontier $F$, and adding a term in $\ind{\norm{g_{n}(x) - g^{*}(x)} > d(g_{n}(x), F)}$ in~\eqref{rates:eq:calibration} for the analysis.
\item Theoretical follow-up, to widen our analysis to other types of smooth surrogates, and to parametric methods, such as deep learning models, assuming that functions are parametrized by a parameter $\theta$, that some analysis gives concentration on $\theta_{n} - \theta^{*}$ similar to~\eqref{rates:eq:concentration} and that calibration inequalities relate the error on $\theta$ with the error between $f_{n}=f_{\theta_{n}}$ and $f^{*} = f_{\theta^{*}}$.
\end{itemize}
\chapter{Fast Rates for Structured Prediction}
\label{chap:fast}
The following is a reproduction of \cite{Cabannes2021b}.
\input{rates/abstract}
\input{rates/core}
\begin{subappendices}
\chapter*{Appendix}
\addcontentsline{toc}{chapter}{Appendix}
\input{rates/appendix/proof}
\input{rates/appendix/nearest}
\input{rates/appendix/kernel}
\end{subappendices}
\section{Median surrogate}
\label{sgd:proof:sur}
Let us begin this section by proving Proposition \ref{sgd:prop:sur}.
This result is actually the integration over $x\in\mathcal{X}$ of a pointwise result, so let us fix $x\in\mathcal{X}$.
Consider a probability distribution $p\in\prob{\mathcal{Y}}$ over $\mathcal{Y}$, and its median $\Theta^* \subset \mathbb{R}^\mathcal{Y}$ defined as the minimizer of ${\cal R}_S(\theta) = \E_p[\norm{\theta - e_Y}]$.
We will to prove that $\cup_{\theta\in\Theta^*}\argmax_{y\in\mathcal{Y}} \theta_y = \argmax_{y\in\mathcal{Y}} p(y)$.
Let us begin by the inclusion $\argmax_{y\in\mathcal{Y}} p(y) \subset \cup_{\theta\in\Theta^*}\argmax_{y\in\mathcal{Y}} \theta_y$.
To do so, consider $\theta\in\mathbb{R}^\mathcal{Y}$ and $\sigma \in \mathfrak{S}_\mathcal{Y}$ the transposition of two elements $y$ and $z$ in $\mathcal{Y}$.
Denote by $\theta_\sigma \in \mathbb{R}^\mathcal{Y}$, the vector such that $(\theta_\sigma)_{y'} = \theta_{\sigma(y')}$ for any $y'\in\mathcal{Y}$.
We have
\begin{align*}
{\cal R}_S(\theta) - {\cal R}_S(\theta_\sigma)
& = \sum_{y'\in \mathcal{Y}} p(y') \paren{\norm{\theta - e_{y'}} - \norm{\theta_\sigma - e_{y'}}}
\\&= \sum_{y'\in \mathcal{Y}} p(y') \paren{\sqrt{\sum_{z'\in\mathcal{Y}}\theta_{z'}^2 + (1 - \theta_{y'})^2 - \theta_{y'}^2} -
\sqrt{\sum_{z'\in\mathcal{Y}}\theta_{\sigma(z')}^2 + (1 - \theta_{\sigma(y')})^2 - \theta_{\sigma(y')}^2}}
\\&= (p(y) - p(z)) \paren{\sqrt{\sum_{z'\in\mathcal{Y}}\theta_{z'}^2 + 1 - 2\theta_{y}} -
\sqrt{\sum_{z'\in\mathcal{Y}}\theta_{z'}^2 + 1 - 2\theta_{z}}}.
\end{align*}
Because, for any $a \in \mathbb{R}_+$, the function $x \to \sqrt{a - 2x}$ is increasing, if $p(y) > p(z)$, then to minimize ${\cal R}$, we should make sure that $\theta_y \geq \theta_z$i.
As a consequence, because of symmetry, the modes of $p$ do correspond to argmax of $(\theta^*_y){y\in\mathcal{Y}}$ for some $\theta^*\in\Theta^*$.
Let us now prove the second inclusion.
To do so, suppose that $p(1) > p(2)$, and let us show that $\theta^*_1 > \theta^*_2$.
Let us parametrize $\theta_1 = a + \epsilon$ and $\theta_2 = a - \epsilon$ for a given $a$, and show that $\epsilon = 0$ is not optimal in order to minimize the risk ${\cal R}_S$ seen as a function of $\epsilon$.
To do so, we can use the Taylor expansion of $\sqrt{1 + x} = 1 + x / 2$.
Hence, with $A = \sum_{y > 2} (\theta^*_y)^2$, retaking the last derivations
\begin{align*}
{\cal R}_S(\epsilon)
& = p(1)\sqrt{(a+\epsilon)^2 + (a-\epsilon)^2 + A + 1 - 2(a + \epsilon)}
\\&\qquad+ p(2)\sqrt{(a+\epsilon)^2 + (a-\epsilon)^2 + A + 1 - 2(a - \epsilon)}
\\&\qquad+ \sum_{y>2} p(y)\sqrt{(a+\epsilon)^2 + (a-\epsilon)^2 + A + 1 - 2\theta^*_y}
\\&= p(1)\sqrt{2a^2+ 2\epsilon^2 + A + 1 - 2a - 2\epsilon}
\\&\qquad+ p(2)\sqrt{2a^2+2\epsilon^2 + A + 1 - 2a + 2\epsilon} + c + o(\epsilon)
\\&= \tilde{c} + \frac{\epsilon}{\sqrt{2a^2 + A + 1 - 2a}}(p(2) - p(1)) + o(\epsilon).
\end{align*}
This shows that taking $\theta_1^* = \theta_2^*$, that is $\epsilon=0$, is not optimal, hence we have the second inclusion, which ends the proof.
Note that we have proven a much stronger result, we have shown that $(\theta_y)$ and $p(y)$ are order in the exact same fashion (with respect to the strict comparison $p(y)>p(z)\Rightarrow \theta_y^* > \theta_z^*$ for any $\theta^*\in\Theta^*$).
\subsection{Discussion around the median surrogate.}
The median surrogate have some nice properties for a surrogate method, in particular it does not fully characterize the distribution $p(y)$ in the sense that there is no one-to-one mapping from $p$ to $\theta^*$.
For example, when $\mathcal{Y} = \brace{1, 2, 3}$ if
\(
p(y=e_1), p(y=e_2), p(y=e_3) \propto (1, 1, 2\cos(\pi / 6)),
\)
then the geometric median correspond to $\theta^* = e_3$.
This differs from smooth surrogates, such as logistic regression or least-squares, that implicitly learn the full distribution $p$, which should be seen as a waste of resources.
Non-smooth surrogates tend to exhibit faster rates of convergence (in terms of decrease of the original risk as a function of the number of samples) than smooth surrogates when rates are derived through calibration inequalities \citep{Nowak2021}.
It would be nice to derive generic calibration inequality for the median surrogate for multiclass, and see how to derive a median surrogate for more structured problems such as ranking problems.
\begin{figure}[ht]
\centering
\includegraphics{sgd/images/noisy_class}
\includegraphics{sgd/images/noisy_ls}
\includegraphics{sgd/images/noisy_median}
\includegraphics{sgd/images/noisy_ls_bis}
\includegraphics{sgd/images/noisy_median_bis}
\caption{
{\em Comparison of least-squares and absolute deviation with noise irregularity} for a classification problem specified by $\mathcal{X} = [0, 3]$, $\mathcal{Y} = \brace{-1, 1}$ with $X$ uniform on $[0, 1]\cup[2, 3]$ and $\eta(x) = \E\brace{Y\,\middle\vert\, X=x}$ specified on the left figure.
The optimal classifier, with respect to the zero-one loss, $f^*(x) = \sign \eta$ takes value one on $[0, 1]$ and value minus one on $[2, 3]$.
The regularized solution are defined as $\argmin_g \E[\norm{\scap{\phi(X)}{\theta} - Y}^p] + \lambda \norm{\theta}$ with $p=2$ for least-squares (middle), and $p=1$ for the median (right).
They can be translated into classifiers with the decoding $f = \sign g$.
In this figure, we choose $\phi$ implicitly through the Gaussian kernel $k(x, x') = \scap{\phi(x)}{\phi(x')} = \exp(-\norm{x-x'}^2 / 2\sigma^2)$ with $\sigma = .1$ which explains the frequency of the observed oscillations, and choose $\lambda = 10^{-6}$ (top) and $\lambda = 10^{-2}$ (bottom).
On the one hand, because the least-squares surrogate is trying to estimate $\eta$ it suffers from its lack of regularity, leading to Gibbs phenomena that restricts it to be a perfect classifier.
On the other hand, the absolute deviation is trying to approach the function $f^*$ itself, and does not suffer from its lack of regularity.
In this setting, if we approach the original classification problem by minimization of the surrogate empirical risks, and denote by $g_n$ this minimizer and $f_n =\sign g_n$ its decoding, $f_n$ obtained through median regression will converge exponentially fast toward $f^*$, while $f_n$ obtained through least-squares will never converge to the solution $f^*$.
}
\label{sgd:fig:med_ls}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics{sgd/images/simplex1}
\includegraphics{sgd/images/simplex2}
\includegraphics{sgd/images/simplexleg}
\caption{
{\em Comparison of least-squares and median surrogate without context.}
Consider a context-free classification problem that consists in estimating the mode of a distribution $p\in\prob{\mathcal{Y}}$, or equivalently the minimizer of the 0-1 loss.
Such a problem can be visualized on the simplex $\prob{m}$ where $\mathcal{Y} = \brace{y_1, \cdots, y_m} \simeq \brace{1, \cdots, m}$ is mapped to the canonical basis $\brace{e_i}_{i\in[m]} \in \mathbb{R}^m$.
The figure illustrates the case $m=3$.
The least-squares and median surrogate methods can be understood as working in this simplex, estimating a quantity $z \in \prob{\mathcal{Y}}$, before performing the decoding $y(z) = \argmax_y \scap{z}{e_y}$.
Such a decoding partitions the simplex in regions whose frontiers are represented in dashed blue on the figure.
The distribution $p$ is characterized on the simplex by $\eta = \E_{Y\sim p}[e_Y] = \argmin \E_{Y\sim p} [\norm{z - e_Y}^2]$.
This quantity $\eta$ is exactly the quantity estimated by the least-squares surrogate.
The median surrogate searches the minimizer $z^*$ of the quantity ${\cal E}(z) = \E_{Y\sim p}[\norm{z - e_Y}]$, whose level lines are represented in solid on the figure.
One of the main advantage of the median surrogate compared to the least-squares one is that $z^*$ is always farther away from the boundary frontier than $\eta$, meaning that for a similar estimation error on this quantity, the error on the decoding, which corresponds to an estimate of the mode of $p$, will be much smaller for the median surrogate.
The left figure represents the case $p = (1, 0, 0)$, the right figure the case $p =(.45, .35, .2)$.}
\label{sgd:fig:med_simplex}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics{sgd/images/simplex_query}
\caption{
{\em Query strategy based on regression surrogate.}
Retaking the simplex representation of Figure \ref{sgd:fig:med_simplex}, the query strategy for classification approached with least-squares surrogate or median surrogate consists in looking at the current surrogate estimate $z$ in the simplex $\prob{\mathcal{Y}}$, taking a random direction $u\in\mathbb{R}^\mathcal{Y}$ and querying $\sign(\scap{e_Y - z}{u})$.
We see that with three elements, when $Y$ is deterministic, the optimal query strategy consists in considering $s =\brace{y}$, while surrogate strategies, such as least-squares and median regression, that learn $z^* = e_y$, would only make such a query only two third of the time (which is the ratio of the solid angle of $[e_2, e_3]$ from $e_1$ divided by $\pi$).
This shows that those surrogate strategies do not fully leverage the specific structure of the output.
}
\label{sgd:fig:med_query}
\end{figure}
\section{Classification with a min-max game}
\label{sgd:proof:minmax}
In this section, we prove and extend on Proposition \ref{sgd:prop:minmax}.
First of all, let us consider the average loss, for $(v_y)\in\mathbb{R}^\mathcal{Y}$ summing to one
\[
\bar L(v, s) = 1 - \sum_{y\in s} v_y = \sum_{y\notin s} v_y.
\]
Consider now this loss conditioned on the observation $\ind{y\in s}$, we have plenty of characterizations of $L$,
\begin{align*}
L(v, s; \ind{y\in s} - \ind{y\notin s})
& = \ind{y\in s} \bar L(v, s) + \ind{y\notin s}\bar L(v, \mathcal{Y}\setminus s)
= \ind{y\in s}\sum_{y\notin s} v_y + \ind{y\notin s}\sum_{y\in S} v_y
\\&= \ind{y\in s} + (\ind{y\notin s} - \ind{y\in s})\sum_{y\in s} v_y
= \ind{y\notin s} + (\ind{y\in s} - \ind{y\notin s})\sum_{y\notin s} v_y
\\&= \frac{1}{2} - \frac{1}{2} (\ind{y\in s} - \ind{y\notin s})\paren{\sum_{y\in s} v_y - \sum_{y\notin s} v_y}
= \frac{1}{2} + \frac{1}{2} (\ind{y\in s} - \ind{y\notin s})\paren{1 - 2\sum_{y\in s} v_y}.
\end{align*}
Minimizing this loss or the loss $2L-1$ as defined in Proposition \ref{sgd:prop:minmax} is equivalent.
\subsection{Consistency}
Let us consider the loss as defined in this proposition, we have the characterization
\[
L(v, s; \ind{y\in s} - \ind{y\notin s}) = (\ind{y\in s} - \ind{y\notin s})\paren{\sum_{y\in s} v_y - \sum_{y\notin s} v_y}.
\]
Let us rewrite~\eqref{sgd:eq:minmax} based on this previous characterization of the loss, we have
\[
\E_Y[L(v, s, \ind{Y\in s} - \ind{Y\notin s})]
= -(\Pbb_Y(Y\in s) - \Pbb_Y(Y\notin s))\paren{\sum_{y\in s} v_y - \sum_{y\notin s} v_y}.
\]
Hence, without any context variable, the min-max game \eqref{sgd:eq:minmax} can be rewritten as
\begin{equation}
\min_{v\in\prob{\mathcal{Y}}} \max_{\mu\in\prob{\cal S}}
- \sum_{s\in{\cal S}} \mu_s (\Pbb_Y(Y\in s) - \Pbb_Y(Y\notin s))\paren{\sum_{y\in s} v_y - \sum_{y\notin s} v_y}.
\end{equation}
We will analyze this problem through the lens of a mix-actions zero-sum game.
We know from \cite{VonNeumann1944} that a solution to this min-max problem exists, and that one can switch the min-max to a max-min without modifying the value of the solution.
Let us denote by $(v^*,\mu^*)$ the argument of a solution.
To minimize the value of this game, the player $v$ should play such that
\[
\sign(\sum_{y\in s} v_y^* - \sum_{y\notin s} v_y^*)
= \sign (\Pbb(Y\in s) - \Pbb(Y\notin s))
= \sign (\sum_{y\in s} \Pbb(Y=y) - \sum_{y\notin s} \Pbb(Y=y)),
\]
which allows this player to ensure a negative value to the game.
Stated otherwise
\begin{equation}
\label{sgd:eq:loss_imp}
\forall\, s\in{\cal S}, \qquad
\Pbb(Y\in s) > \frac{1}{2}\quad\Rightarrow\quad\sum_{y\in s} v_y^* \geq \frac{1}{2}.
\end{equation}
As a consequence, if there exists any set such that $\Pbb(Y\in s) = 1/2$, the best strategy of player $\mu$ is to play only those sets to ensure the value zero, and any $v$ that satisfies~\eqref{sgd:eq:loss_imp} is optimal.
It should be noted that~\eqref{sgd:eq:loss_imp} does not generally imply that $(v_y)_{y\in\mathcal{Y}}$ has the same ordering as $(\Pbb(Y=y))_{y\in\mathcal{Y}}$.
When $\brace{y^*}\in{\cal S}$ and $\Pbb(Y=y^*) > 1/2$, if $v = \delta_{y^*}$, the prediction player is able to ensure a value of $\max_{s\in{\cal S}}-\abs{2\Pbb(Y\in s) - 1}$, which is maximized by the query player with $s = \brace{y^*} \cup s'$ for any $s'$ such that $\Pbb(Y\in s') = 0$. Other strategies for $v$ will only increase this value, hence $v^* = \delta_{y^*}$ which implies the first part of Proposition \ref{sgd:prop:minmax}.
\paragraph{A counter example.}
While we hope that the solution $(v^*, \mu^*)$ does characterize the original solution $y^*$, it should be noted that $v^*$ alone does not characterize $y^*$.
Indeed, it is even possible to have $v^*$ uniquely defined without having $y^* = \argmax_{y\in\mathcal{Y}} v^*_y$.
For example, consider the case where $\mathcal{Y} = \brace{1, 2, 3}$ and $(\Pbb(Y=i))_{i\in[3]} = (.4, .3, .3)$.
By symmetry, the player $\mu$ only has to play on ${\cal S} = \brace{\brace{1}, \brace{2}, \brace{3}}$, which leads to the min-max game
\[
\min_{v} \max_{\mu}
\paren{
\begin{array}{c}
\mu_{\brace 1} \\
\mu_{\brace 2} \\
\mu_{\brace 3} \\
\end{array}
}^\top
\paren{
\begin{array}{ccc}
.2 & -.2 & -.2 \\
-.4 & .4 & -.4 \\
-.4 & -.4 & .4 \\
\end{array}
}
\paren{
\begin{array}{c}
v_1 \\
v_2 \\
v_3 \\
\end{array}
}.
\]
The value of this game is $-.1$ and is achieved for $\mu^* = (.5, .25, .25)$, $v^* = (.25, .375, .375)$.
\subsection{Optimization procedure}
Let us rewrite the problem through the objective
\[
{\cal E}(g, \mu) = \E_{(X, y)\sim \rho}\E_{S\sim\mu(x)}[L(g(X), S, \ind{Y\in S} - \ind{Y\notin S})].
\]
We want to solve the min-max problem $\min_g\max_\mu{\cal E}(g, \mu)$.
This problem can be solved efficiently based on the vector field point of view of gradient descent \citep{Bubeck2015} if:
\begin{itemize}
\item we can parametrize the function $g:\mathcal{X}\to\prob\mathcal{Y}$ such that ${\cal E}$ is convex with respect to the parametrization of $g$;
\item we can access unbiased stochastic gradients of ${\cal E}$ with respect to $g$ that have a small second moment;
\item we can parametrize the function $\mu:\mathcal{X}\to\prob{\cal S}$ such that ${\cal E}$ is concave with respect to the parametrization of $\mu$;
\item we can access unbiased stochastic gradients of ${\cal E}$ with respect to $\mu$ that have a small second moment.
\end{itemize}
The first two points are no problems, $g$ can be parametrized with softmax regression, and since $L$ is linear with respect to the scores, it will keep the problem convex.
Moreover, to access a stochastic gradient of ${\cal E}$, one can sample $X_i\sim\rho_\mathcal{X}$ and $S_i\sim\mu(X_i)$ before querying $\ind{Y_i\in S_i}$ and computing the gradient of $L(g(X_i), S_i, \ind{Y_i\in S_i} - \ind{Y_i\notin S_i})$ with respect to the parametrization of $g$.
The third point is slightly harder to tackle.
Since ${\cal E}$ is linear with respect to $\mu$, one way to proceed is to find a linear parametrization of $\mu$.
In particular, one can take a family $(g_i)_{i\in[N]}$ of linearly independent functions from $\mathcal{X}$ to $\prob{\cal S}$ and search for $g$ under the form $\sum_{i\in[N]} c_i g_i$ for $(c_i)$ positive summing to one.
To build such a family, one can eventually use ``atom functions'' and simple operations such as symmetry with respect to $\mathcal{Y}$ and ${\cal S}$, rescaling, translation, rotations with respect to $\mathcal{X}$.
For example if $\mathcal{X}$ is a Banach space, one could define atom functions as, for $y_i\in\mathcal{Y}$
\[
g_i: x\to \frac{\norm{x}}{1+\norm{x}} \frac{1}{\card{\cal S}} \sum_{s\in{\cal S}} e_s + \frac{1}{1+\norm{x}} e_{\brace{y_i}}.
\]
Those functions could be rescaled and translated as $g_{\sigma, \tau, i}(x) = g_i(\sigma (x-\tau))$, in order to specify a family $(g_{\sigma, \tau, i})$ from few values for $\tau$ and $\sigma$.
The last point is the most difficult one.
Without context variables, and with no-parametrization for $\mu$, a naive unbiased gradient strategy for $\mu$ consists in asking random questions to update the full knowledge of $(\Pbb(Y\in s))_{s\in{\cal S}}$.
But such a strategy will be much worse than our median surrogate technique with queries $\ind{Y\in\brace{y}}$ for $y$ sampled uniformly at random in $\mathcal{Y}$.
Eventually, one should go for a biased gradient strategy, while making sure to update $\mu$ coherently to avoid getting stalled on bad estimates as a result of biases.
\section{Experimental details}
\label{sgd:app:experiments}
Our experiments are done in {\em Python}. We leverage the {\em C} implementation of high-level array instructions by \cite{Harris2020}, as well as the visualization library of \cite{Hunter2007}.
Randomness in experiments is controlled by choosing explicitly the seed of a pseudo-random number generator.
\begin{figure}[ht!]
\centering
\includegraphics{sgd/images/one_d_frame}
\caption{
{\em Streaming history of the active strategy} to reconstruct the signal in dashed blue in the same setting as Figure \ref{sgd:fig:exp_1}.
At any time $t$, a point $X_t$ is given to us, our current estimate of $\theta_t$ plotted in dashed orange gives us $z = f_{\theta_t}(X_t)$, and we query $\sign(Y_t - z)$.
Based on the answer to this query, we update $\theta_t$ to $\theta_{t+1}$ leading to the new estimate of the signal in solid orange.
In this figure, we see that it might be useful for the practitioners in a streaming setting to reduce the bandwidth of $\phi$ as they advance in time.}
\label{sgd:fig:exp_1_app}
\end{figure}
\subsection{Comparison with fully supervised SGD}
In this section, we investigate the difference between weakly and fully supervised SGD.
According to Theorem \ref{sgd:thm:sgd}, we only lost a constant factor of order $m^{3/2}$ in our rates compared to fully supervised (or plain) SGD.
This behavior can be checked by adding the plain SGD curve on Figure \ref{sgd:fig:exp_2}.
On the left side of Figure \ref{sgd:fig:plain_sgd}, we do observe that the risk of both Algorithm \ref{sgd:alg:sgd} and plain SGD decrease with same exponent with respect to number of iteration but with a different constant in front of the rates: that is we observe the same slopes on the logarithm scaled plot, but different intercepts.
Going one step further to check the tightness of our bound, one can plot the intercept, or the error achieved by both Algorithm \ref{sgd:alg:sgd} and plain SGD as a function of the output space dimension $m$.
The right side of Figure \ref{sgd:fig:plain_sgd} shows evidence that this error grows as $m^\epsilon$ for some $\epsilon \in [1, 3/2]$, which is coherent with our upper bound.
Similarly to Figure \ref{sgd:fig:exp_2}, this figure was computed after cross validation to find the best scaling of the step sizes for each dimension $m$.
\begin{figure}[t]
\centering
\includegraphics{sgd/images/plain_sgd}
\includegraphics{sgd/images/plain_m}
\caption{
{\em Comparison of generalization errors of weakly and fully supervised SGD} as a function of the annotation budget $T$ and output space dimension $m$.
The setting is similar to Figure \ref{sgd:fig:exp_2}.
We observe a transitory regime before convergence rates follows the behavior described by Theorem \ref{sgd:thm:sgd}.
The right side plots the error of both procedures after 10,000 iterations as a function of the output space dimension $m$ between 1 and 50.
The number of iteration ensures that, for all values of $m\in [50]$, the reported error is well characterized by our theory, in other terms that we have entered the regime described by Theorem \ref{sgd:thm:sgd}.}
\label{sgd:fig:plain_sgd}
\end{figure}
\subsection{Passive strategies for classification}
A simple passive strategy for classification based on median surrogate consists in using the active strategy with coordinates sampling, that is $u$ being uniform on $\brace{e_y}_{y\in\mathcal{Y}}$, where $(e_y)_{y\in\mathcal{Y}}$ is the canonical basis of $\mathbb{R}^\mathcal{Y}$ used to define the simplex $\prob{\mathcal{Y}}$ as the convex hull of this basis.
Querying $\ind{\scap{g_\theta(x) - e_y}{e_y} > 0}$ is formally equivalent to the query of $\ind{Y = y}$ when $g_\theta(x) \in \prob{\mathcal{Y}}$.
This is the baseline we plot on Figure \ref{sgd:fig:exp_2}.
\begin{figure}[ht]
\centering
\includegraphics{sgd/images/classification_tough}
\caption{
{\em Comparison with the infimum loss with better conditioned passive supervision} in a similar setting to Figure \ref{sgd:fig:exp_2} yet with $m=10$, $\epsilon=0$, that is $X$ uniform on $\mathcal{X}$, and $\gamma_0 = 7.5$ for the active strategy and $\gamma_0 = 15$ for the passive strategy.
We see no major differences between the active strategy based on the median surrogate and the passive strategy based on the median surrogate with the infimum loss.
Note that the standard deviation is sometimes bigger than the average of the excess of risk, explaining the dive of the dark area on this logarithmic-scaled plot.
}
\label{sgd:fig:exp_2_app}
\end{figure}
A more advanced passive baseline is provided by the infimum loss \citep{Cour2011,Cabannes2020}.
It consists in solving
\[
\argmin_{f:\mathcal{X}\to\mathcal{Y}}{\cal R}_I(f):=\E_{(X, Y)\sim\rho}\E_{S}\bracket{L(f(X), S, \ind{Y\in S})},
\]
where $S$ is a random subset of $\mathcal{Y}$ and $L$ is defined from the original loss $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ as, for $z\in\mathcal{Y}$, $s\subset \mathcal{Y}$ and $y\in\mathcal{Y}$,
\[
L(z, s, \ind{y\in s}) = \left\{\begin{array}{cl} \inf_{y'\in s} \ell(z, y') & \text{if } y\in s \\ \inf_{y'\notin s} \ell(z, y') & \text{otherwise.}\end{array}\right.
\]
Random subsets $S$ could be generated by making sure that the variable $(y \in S)_{y\in\mathcal{Y}}$ are independent balanced Bernoulli variables; and by removing the trivial sets $S = \emptyset$ and $S = \mathcal{Y}$ from the subsequent distribution.
In order to optimize this risk in practice, one can use a parametric model and a surrogate differentiable loss together with stochastic gradient descent on the empirical risk.
For classification with the 0-1 loss, we can reuse the surrogate introduced in Proposition \ref{sgd:prop:sur} and minimize, assuming that we always observed $\ind{Y_i\in S_i} = 1$ for simplicity,
\[
\hat{\cal R}_{I, S}(\theta) = \sum_{i=1}^n \inf_{y \in S_i} \norm{g_\theta(X_i) - e_y}.
\]
Stochastic gradients are then given by, assuming ties have no probability to happen,
\[
\nabla_\theta \inf_{y \in S_t} \norm{g_\theta(X_t) - e_y}
= \paren{\frac{g_\theta(X_t) - e_{y^*}}{\norm{g_\theta(X_t) - e_{y^*}}}}^\top D g_\theta(X_t)
\quad\text{with}\quad y^* := \argmax_{y\in S_t} \scap{g_\theta(X_t)}{e_y}.
\]
This gives a good passive baseline to compare our active strategy with.
In our experiments with the Gaussian kernel, see Figure \ref{sgd:fig:exp_2_app} for an example, we witness that this baseline is highly competitive.
Although we find that it is slightly harder to properly tune the step size for SGD, and that the need to compute an argmax for each gradient slows-down the computations.
\subsection{Real-world classification datasets}
\begin{figure}[ht]
\centering
\includegraphics{sgd/images/classification_usps}
\includegraphics{sgd/images/classification_pendigits}
\caption{
{\em Testing errors on two LIBSVM datasets} with a similar setting to Figure \ref{sgd:fig:exp_2_app}.
Those empirical errors are reported after averaging over 100 different splits of the datasets.
The step size parameter was optimized visually, which led to $\gamma_0=15$ for the active strategy on ``USPS'', $\gamma_0=60$ for the passive one, $\gamma_0=7.5$ for the active strategy on ``pen digits'', $\gamma_0=30$ for the passive one.
The dotted line represents ${\cal R} = 1 - m^{-1}$ which is the performance of a random model.
}
\label{sgd:fig:exp_libsvm}
\end{figure}
In Figure \ref{sgd:fig:exp_libsvm}, we compare the ``well-conditioned'' passive baseline with our active strategy on the real-world problems of LIBSVM \citep{Chang2011}.
We choose the ``USPS'' and ``pen digits'' datasets as they contain $m=10$ classes each with $n=7291$ and $n=7494$ samples respectively, with $d=50$ and $d=16$ features each.
We have chosen those datasets as they present enough classes that leads to many different sets $S$ to query, and they are made of the right number of samples to do some experiments on a laptop without the need for ``advanced'' computational techniques such as caching or low-rank approximation \citep{Meanti2020}.
On Figure \ref{sgd:fig:exp_libsvm}, we use the same linear model as for Figure \ref{sgd:fig:exp_2}, that is a Gaussian kernel.
We choose the bandwidth to be $\sigma = d/5$, and we normalize the features beforehand to make sure that they are all centered with unit variance.
We report error by taking two thirds of the samples for training and one third for testing, and averaging over one hundred different ways of splitting the datasets.
We observe that the active strategy leads to important gains on the ``USPS'' dataset, yet is not that useful for the ``pen digits'' dataset.
We have not dug in to understand those two different behaviors.
\subsection{Real-world regression dataset \& Nystr\"om method}
In this section, we provide two experiments on real-world datasets.
In order to deal with big regression datasets, it is useful to approximate the parameter space $\mathcal{Y}\otimes{\cal H}$ in Assumption \ref{sgd:ass:source} with a small dimensional space.
To do so, let us remark that given samples $(X_i)_{i\leq n} \in \mathcal{X}^n$ for $n\in\mathbb{N}$, we know that our estimate $f_{\theta_n}$ can be represented as
\[
f_{\theta_n}(\cdot) = \sum_{i\leq n}\sum_{j\leq m} a_{ij} \scap{\phi(x_i)}{\phi(\cdot)} e_j,
\]
for some $(a_{ij}) \in \mathbb{R}^{p\times m}$ and where $(e_j)_{j\leq m}$ is the canonical basis of $\mathcal{Y} = \mathbb{R}^m$.
For large datasets, that is when $n$ is large, it is smart to approximate this representation through the parameterization
\[
f_{a}(x) = \sum_{i\leq p}\sum_{j\leq m}a_{ij} k(x, x_i) e_j,
\]
where $p\leq n$ is the rank of our approximation, and $k$ is the kernel defined as $k(x, x') = \scap{\phi(x)}{\phi(x')}$.
Stated with words, we only use a small number $p$, instead of $n$, of vectors $\phi(x_i)$ to parameterize $f$.
This allows to only keep a matrix of size $p\times m$ in memory instead of $n\times m$, while not fundamentally changing the statistical guarantee of the method \citep{Rudi2015}.
In this setting, the stochastic gradients are specified from the fact that
\[
u^\top D_a f_a(x) = (u_j k(x, x_i))_{i,j} \in \mathbb{R}^{p\times m}.
\]
In other terms, in order to update the parameter $a$ with respect to the observation made at $(x, u)$, we check how much each coordinate of $a$ determines the value of $u^\top f_a(x)$.
In the following, we experiment with two real-world datasets.
In order to learn the relation between inputs and outputs, we use a Gaussian kernel after normalizing input features so that each of them has zero mean and unit variance.
To keep computational cost, we sample $p$ random (Nystr\"om) representers among the training inputs which are used to parameterize functions.
To avoid overfitting, we add a small regularization to the empirical objective. It reads $\lambda \norm{\theta}_{{\cal H}}^2$ with our notations and corresponds to the Hilbertian norm inherited from the reproducing kernel $k$ of the function $f_{\theta}$ \citep{Scholkopf2001}.
\begin{figure}[h]
\centering
\includegraphics{sgd/images/calcofi}
\includegraphics{sgd/images/real_temp}
\caption{
{\em Testing error on two real-world regression datasets.}
On both datasets, a single pass was made through the data in a chronological fashion, and errors were computed from the 26,453 most recent data samples for the ``Weather'' dataset, and from a random sample of 10,000 samples among the 155,140 most recent samples for the ``CalCOFI'' dataset.}
\label{sgd:fig:real_world}
\end{figure}
Our first experiment is based on the data collected by the California Cooperative Oceanic Fisheries Investigation between March 1949 and November 2016.
It consists of more than 800,000 seawater samples including measurements of nutriments (set aside in our experiments) together with pressure, temperature, salinity, water density, dynamic height (providing five input parameters), as well as dissolved oxygen, and oxygen saturation (the two outputs we would like to predict).
We assume that we can measure if any weighted sum of oxygen concentration and saturation is above a threshold by letting some population of bacteria evolves in the water sample and checking if it survives after a day.
If the measurements are done on the day of the sample collection, this setting exactly fits in the streaming active labeling framework.
After cleaning the dataset for missing values, the dataset contains 655,140 samples.
The ``CalCOFI'' dataset results are reported on the left of Figure \ref{sgd:fig:real_world}, parameters were chosen as $p=100$, $\sigma = 10$, $\lambda = 10^{-6}$ and $\gamma_0 = 1$.
For the passive strategy, random queries were chosen to follow a normal distribution with the same mean as the targets and one third of their standard deviation ({\em i.e.} we ask if the apparent temperature is lower than the usual one plus or minus a perturbation).
The plotted baseline corresponds to linear regression performed over the entire dataset.
It takes about 10,000 samples for our active strategy to be competitive with this baseline, and 200,000 samples for the passive one.
The second experiment makes use of data collected through the Dark Sky API (which is now part of Apple WeatherKit).
It is made of 96,454 weather summaries between 2006 and 2016 in the city of Szeged, Hungary.
Our task consists in computing the apparent temperature from real temperature, humidity, wind speed, wind bearing, visibility and pressure.
The apparent temperature is an index that searches to quantify the subjective feeling of heat that humans perceive, it is expressed on the same scale as real temperature.
One way to measure it would be to ask some humans if the outside is hotter or colder than a controlled room with a specific temperature and neutral meteorological conditions.
Once again, this exactly fits into our streaming active labeling setting.
The ``Weather'' dataset results are reported on the right of Figure \ref{sgd:fig:real_world}.
The baseline consists in predicting the apparent temperature as the real temperature.
We observe a transitory regime where the first 1,000 samples seem to be used to calibrate the weights $\alpha$.
During this regime, our estimate is too bad for the active strategy to make smarter queries than the ``random'' ones that have been calibrated on temperature statistics.
The main difference in the learning dynamic between the active and passive strategies is observed on the remaining 69,000 training samples.
The parameters were the same as the ``CalCOFI'' dataset but for $\gamma_0 = 10^{-2}$.
\section{Proofs of the statistical analysis}
In the following proofs, we assume $\mathcal{X}$ to be Polish and $\mathcal{Y} = \mathbb{R}^m$, so to define the joint probability $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$.
Moreover, we assume that $\E[\norm{Y}] < +\infty$ in order to define the risk of median regression.
We consider ${\cal H}$ to be a Hilbert space that is separable ({\em i.e.} only the origin is in all the neighborhood of the origin), and $\phi$ to be a measurable mapping from $\mathcal{X}$ to ${\cal H}$.
In terms of notations, we denote $\brace{1, 2, \cdots, n}$ by $[n]$ for any $n\in\mathbb{N}^*$, and by $(x_i)_{i\leq n}$ the family $(x_1, \cdots, x_n)$ for any sequence $(x_i)$.
The unit sphere in $\mathbb{R}^m$ is denoted by $\mathbb{S}^{m-1}$.
The symbol $\otimes$ denotes tensors, and is extended to product measures in the notation $\rho^{\otimes n} = \rho\times\rho\times\cdots\times\rho$.
We have used the isometry between trace-class linear mappings from ${\cal H}$ to $\mathcal{Y}$ and the tensor space $\mathcal{Y}\otimes{\cal H}$, which generalizes the matrix representation of linear map between two finite-dimensional vector spaces.
This space inherits from the Hilbertian structure of ${\cal H}$ and $\mathcal{Y}$ and we denote by $\norm{\cdot}$ the Hilbertian norm that generalizes the Frobenius norm on linear maps between Euclidean spaces.
\subsection{Upper bound for stochastic gradient descent}
\label{sgd:proof:sgd}
This subsection is devoted to the proof of Theorem \ref{sgd:thm:sgd}.
For simplicity, we will work with the rescaled step size $\gamma_t := c_2 \gamma(t)$ rather than the step size described in the main text $\gamma(t)$.
Convergence of stochastic gradient descent for non-smooth problems is a known result. For completeness, we reproduce and adapt a usual proof to our setting.
For $t\in\mathbb{N}$, let us introduce the random functions
\[
{\cal R}_t(\theta) = c_2^{-1}\abs{\scap{\theta\phi(X_t) - Y_t}{U_t}},
\qquad\text{where}\qquad
c_2 = \E_U[\abs{\scap{e_1}{U}}] = \E_U[\sign(\scap{e_1}{U}) \scap{e_1}{U}]
\]
for $(X_t, Y_t) \sim \rho$, $U_t$ uniform on the sphere $\mathbb{S}^{m-1} \subset \mathcal{Y}$.
Those random functions all average to ${\cal R}(\theta) = \E_{\rho} \E_U[c_2^{-1}\abs{\scap{\theta\phi(X) - Y}{U}}] = \E_\rho[\norm{\theta\phi(X) - Y}]$.
After a random initialization $\theta_0 \in\Theta$, the stochastic gradient update rule can be written for any $t \in \mathbb{N}$ as
\[
\theta_{t+1} = \theta_t - \gamma_t \nabla {\cal R}_t(\theta_t),
\]
where $\nabla {\cal R}_t$ denotes any sub-gradients of ${\cal R}_t$.
We can compute
\[
\nabla {\cal R}_t(\theta_t)
= c_2^{-1} \nabla \abs{\scap{\theta\phi(X_t) - Y_t}{U_t}}
= c_2^{-1} \sign\paren{\scap{\theta\phi(X_t) - Y_t}{U_t}} U_t\otimes \phi(X_t).
\]
This corresponds to the gradient written in Algorithm \ref{sgd:alg:sgd}.
Let us now express the recurrence relation on $\norm{\theta_{t+1} - \theta^*}$.
We have
\begin{align*}
\norm{\theta_{t+1} - \theta^*}^2
& = \norm{\theta_t - \gamma_t \nabla {\cal R}_t(\theta_t) - \theta^*}^2
\\&= \norm{\theta_t - \theta^*}^2 + \gamma_t^2\norm{\nabla {\cal R}_t(\theta_t)}^2 - 2\gamma_t\scap{\nabla {\cal R}_t(\theta_t)}{\theta_t - \theta^*}.
\end{align*}
Because ${\cal R}_t$ is convex, it is above its tangents
\[
{\cal R}_t(\theta^*) \geq {\cal R}_t(\theta_t) + \scap{\nabla{\cal R}_t(\theta_t)}{\theta^* - \theta_t}.
\]
Hence,
\[
\norm{\theta_{t+1} - \theta^*}^2
\leq \norm{\theta_t - \theta^*}^2 + \gamma_t^2\norm{\nabla {\cal R}_t(\theta_t)}^2 + 2\gamma_t ({\cal R}_t(\theta^*) - {\cal R}_t(\theta_t)).
\]
This allows bounding the excess of risk as
\[
2({\cal R}_t(\theta_t) - {\cal R}_t(\theta^*))
\leq \frac{1}{\gamma_t}(\norm{\theta_t - \theta^*}^2 - \norm{\theta_{t+1} - \theta^*}^2) + \gamma_t c_2^{-2}\norm{\phi(X_t)}^2.
\]
where we used the fact that $\norm{\nabla {\cal R}_t} = c_2^{-1}\norm{\phi(X_t)}$.
Let us multiply this inequality by $\eta_t > 0$ and sum from $t=0$ to $t=T-1$, we get
\begin{align*}
& 2(\sum_{t=0}^{T-1} \eta_t {\cal R}_t(\theta_t) - \sum_{t=0}^{T-1} \eta_t {\cal R}_t(\theta^*))
\leq \sum_{t=0}^{T-1} \frac{\eta_t}{\gamma_t}(\norm{\theta_t - \theta^*}^2 - \norm{\theta_{t+1} - \theta^*}^2) + \sum_{t=0}^{T-1} \eta_t\gamma_t c_2^{-2}\norm{\phi(X_t)}^2
\\&\qquad= \frac{\eta_0}{\gamma_0} \norm{\theta_0-\theta^*}^2 - \frac{\eta_{T-1}}{\gamma_{T-1}} \norm{\theta_T - \theta^*}^2 + \sum_{t=1}^{T-1} \paren{\frac{\eta_t}{\gamma_t} - \frac{\eta_{t-1}}{\gamma_{t-1}}}\norm{\theta_t - \theta^*}^2 + \sum_{t=0}^{T-1} \eta_t\gamma_t c_2^{-2}\norm{\phi(X_t)}^2.
\end{align*}
From here, there is several options to obtain a convergence result, either one assume $\norm{\theta_t - \theta^*}$ bounded and take $\eta_t\gamma_{t-1} \geq \eta_{t-1}\gamma_t$; or one take $\eta_t = \gamma_t$ but at the price of paying an extra $\log(T)$ factor in the bound; or one take $\gamma_t$ and $\eta_t$ independent of $t$.
Since we suppose the annotation budget given, we will choose $\gamma_t$ and $\eta_t$ independent of $t$, only depending on $T$.
\begin{align*}
& 2(\sum_{t=0}^{T-1} \eta {\cal R}_t(\theta_t) - \sum_{t=0}^{T-1} \eta {\cal R}_t(\theta^*))
\leq \frac{\eta}{\gamma} \norm{\theta_0-\theta^*}^2 + \sum_{t=0}^{T-1} \eta\gamma c_2^{-2}\norm{\phi(X_t)}^2.
\end{align*}
Let now take the expectation with respect to all the random variables, for the risk
\begin{align*}
\E_{(X_s, Y_s, U_s)_{s\leq t}}[{\cal R}_t(\theta_t)]
& = \E_{(X_s, Y_s, U_s)_{s\leq t}}\bracket{\E_{(X_t, Y_t)}\bracket{\E_{U_t}\bracket{{\cal R}_t(\theta_t)\,\middle\vert\, \theta_t}\,\middle\vert\, \theta_t}}
\\&= \E_{(X_s, Y_s, U_s)_{s\leq t}}\bracket{{\cal R}(\theta_t)}
= \E[{\cal R}(\theta_t)].
\end{align*}
For the variance, $\E[\norm{\phi(X_s)}^2] = \E[\norm{\phi(X)}^2] = \kappa^2$.
Let us fix $T$ and consider $\eta_t = 1/ T$, by Jensen we can bound the following averaging
\begin{align*}
2\paren{{\cal R}\paren{\sum_{t=0}^{T-1} \eta_t \theta_t} - {\cal R}(\theta^*)}
& \leq 2\paren{\sum_{t=0}^{T-1} \eta_t {\cal R}\paren{\theta_t} - {\cal R}(\theta^*)}
= 2\E\bracket{\sum_{t=0}^{T-1} \eta_t ({\cal R}_t\paren{\theta_t} - {\cal R}_t(\theta^*))}
\\&\qquad\leq \frac{1}{T\gamma}\norm{\theta_0-\theta^*}^2 + \gamma c_2^{-2}\kappa^2.
\end{align*}
Initializing $\theta_0$ to zero, we can optimize the resulting quantity to get the desired result.
\subsection{Upper bound for resampling strategy}
\label{sgd:proof:resampling}
For resampling strategies, the proof is built on classical statistical learning theory considerations.
Let us decompose the risk between estimation and optimization errors.
Recall the expression of the risk ${\cal R}$, the function taking as inputs measurable functions from $\mathcal{X}$ to $\mathcal{Y}$ and outputting a real number
\[
{\cal R}(f) = \E_{\rho}[\norm{f(X) - Y}].
\]
Let us denote by ${\cal F}$ the class of functions from $\mathcal{X}$ to $\mathcal{Y}$ we are going to work with.
Let $f_n$ be our estimate of $f^*$ which maps almost every $x\in\mathcal{X}$ to the geometric median of $\paren{Y\,\middle\vert\, X}$.
Denote by ${\cal R}_{{\cal D}_n}^*$ the best value that can be achieved by our class of functions to minimize the empirical average absolute deviation
\[
{\cal R}^*_{{\cal D}_n} = \inf_{f\in{\cal F}} {\cal R}_{{\cal D}_n}(f).
\]
Assumption \ref{sgd:ass:source} states that we have a well-specified model ${\cal F}$ to estimate the median, {\em i.e.} $f^* \in {\cal F}$.
Hence, the excess of risk can be decomposed as an estimation and an optimization error, without approximation error (it is not difficult to add an approximation error, but it will make the derivations longer and the convergence rates harder to parse for the reader).
Using the fact that ${\cal R}_{{\cal D}_n}(f^*) \geq {\cal R}_{{\cal D}_n}^*$ by definition of the infimum, we have
\begin{equation}
\label{sgd:eq:decomposition}
{\cal R}(f_n) - {\cal R}(f^*)
\leq \underbrace{{\cal R}(f_n) - {\cal R}_{{\cal D}_n}(f_n)
+ {\cal R}_{{\cal D}_n}(f^*) - {\cal R}(f^*)}_{\text{estimation error}}
+ \underbrace{{\cal R}_{{\cal D}_n}(f_n) - {\cal R}_{{\cal D}_n}^*}_{\text{optimization error}}.
\end{equation}
\paragraph{Estimation error.}
Let us begin by controlling the estimation error.
We have two terms in it.
${\cal R}_{{\cal D}_n}(f^*) - {\cal R}(f^*)$ can be controlled with a concentration inequality on the empirical average of $\norm{f^*(X) - Y}$ around its population mean.
Assuming sub-Gaussian moments of $Y$, it can be done with Bernstein inequality.
${\cal R}_{{\cal D}_n}(f_n) - {\cal R}(f_n)$ is harder to control as $f_n$ depends on ${\cal D}_n$, so we can not use the same technique.
The classical technique consists in going for the brutal uniform majoration,
\begin{equation}
\label{sgd:eq:sup_rade}
{\cal R} (f_n) - {\cal R}_{{\cal D}_n}(f_n) \leq
\sup_{f\in{\cal F}} \paren{{\cal R} (f) - {\cal R}_{{\cal D}_n}(f)},
\end{equation}
where ${\cal F}$ denotes the set of functions that $f_n$ could be in concordance with our algorithm.
While this bound could seem highly suboptimal, when the class of functions is well-behaved, we can indeed control the deviation ${\cal R}(f) - {\cal R}_{{\cal D}_n}(f)$ uniformly over this class without losing much (indeed for any class of functions, it is possible to build some really adversarial distribution $\rho$ so that this supremum behaves similarly to the concentration we are looking for \citep{Vapnik1995,Anthony1999}).
This is particularly the case for our model linked with Assumption \ref{sgd:ass:source}.
Expectations of supremum processes have been extensively studied, allowing to get satisfying upper bounds (note that when the $\norm{f(X) - Y}$ is bounded, deviation of the quantity of interest around its expectation can be controlled through McDiarmid inequality).
In the statistical learning literature, it is usual to proceed with Rademacher complexity.
\begin{lemma}[Uniform control of functions deviation with Rademacher complexity]
The expectation of the excess of risk can be bounded as
\begin{equation}
\label{sgd:eq:rademacher}
\frac{1}{2}\E_{{\cal D}_n} \bracket{\sup_{f\in{\cal F}} \paren{{\cal R} (f) - {\cal R}_{{\cal D}_n}(f)}}
\leq \mathfrak{R}_n({\cal F}, \ell, \rho) :=
\frac{1}{n} \E_{{\cal D}_n, (\sigma_i)}\bracket{ \sup_{f\in{\cal F}}\sigma_i \ell(f(X_i), Y_i)},
\end{equation}
where $(\sigma_i)_{i\leq n}$ is defined as a family of Bernoulli independent variables taking value one or minus one with equal probability, and $\mathfrak{R}_n({\cal F}, \ell, \rho)$ is called Rademacher complexity.
\end{lemma}
\begin{proof}
This results from the reduction to larger supremum and a symmetrization trick,
\begin{align*}
\E_{{\cal D}_n} \bracket{\sup_{f\in{\cal F}} \paren{{\cal R} (f) - {\cal R}_{{\cal D}_n}(f)}}
& = \E_{{\cal D}_n} \bracket{\sup_{f\in{\cal F}} \paren{\E_{{\cal D}_n'}{\cal R}_{{\cal D}_n'} (f) - {\cal R}_{{\cal D}_n}(f)}}
\\&\leq \E_{{\cal D}_n} \E_{{\cal D}_n'}\bracket{\sup_{f\in{\cal F}} \paren{{\cal R}_{{\cal D}_n'} (f) - {\cal R}_{{\cal D}_n}(f)}}
\\&= \E_{(X_i, Y_i), (X_i', Y_i')}\bracket{\sup_{f\in{\cal F}} \paren{\frac{1}{n} \sum_{i=1}^n \ell(f(X_i'), Y_i') - \ell(f(X_i), Y_i)}}
\\&= \E_{(X_i, Y_i), (X_i', Y_i'), (\sigma_i)}\bracket{\sup_{f\in{\cal F}} \paren{\frac{1}{n} \sum_{i=1}^n \sigma_i\paren{\ell(f(X_i'), Y_i') - \ell(f(X_i), Y_i)}}}
\\&\leq 2 \E_{(X_i, Y_i), (\sigma_i)}\bracket{\sup_{f\in{\cal F}} \paren{\frac{1}{n} \sum_{i=1}^n \sigma_i\paren{\ell(f(X_i), Y_i)}}},
\end{align*}
which ends the proof.
\end{proof}
In our case, we want to compute the Rademacher complexity for $\ell$ given by the norm of $\mathcal{Y}$, and ${\cal F} = \brace{x\to \theta \phi(x)\,\middle\vert\, \theta \in \mathcal{Y}\otimes{\cal H}, \norm{\theta} < M}$, for $M > 0$ a parameter to specify in order to make sure that $\norm{\theta^*} < M$, where the norm has to be understood as the $\ell^2$-product norm on $\mathcal{Y} \otimes{\cal H} \simeq {\cal H}^m$.
Working with linear models and Lipschitz losses is a well-known setting, allowing to derive directly the following bound.
\begin{lemma}[Rademacher complexity of linear models with Lipschitz losses]
The complexity of the linear class of vector-valued function ${\cal F} = \brace{x\to \theta \phi(x)\,\middle\vert\, \theta \in \mathcal{Y}\otimes{\cal H}, \norm{\theta} < M}$ is bounded as
\begin{equation}
\E_{(\sigma_i)}\bracket{\sup_{f\in{\cal F}}\paren{\frac{1}{n} \sum_{i=1}^n \sigma_i \norm{f(x_i) - y_i}}}
\leq M \kappa n^{-1/2}.
\end{equation}
\end{lemma}
\begin{proof}
This proposition is usually split in two.
First using the fact that the composition of a space of functions with a Lipschitz function does not increase the entropy of the subsequent space \citep{Vitushkin1954}.
Then bounding the Rademacher complexity of linear models.
We refer to \cite{Maurer2016} for a self-contained proof of this result (stated in its Section 4.3).
\end{proof}
Adding all the pieces together we have proven the following proposition, using the fact that the previous bound also applies to $\sup_{f\in{\cal F}} {\cal R}_{{\cal D}_n}(f) - {\cal R}(f)$ by symmetry, hence it can be used for the deviation of ${\cal R}_{{\cal D}_n}(f^*) - {\cal R}(f^*)$.
\begin{proposition}[Control of the estimation error]
Under Assumption \ref{sgd:ass:source}, with the model of computation ${\cal F} = \brace{x\in\mathcal{X}\to \theta\phi(x) \in \mathcal{Y} \,\middle\vert\, \norm{\theta} \leq M}$, the generalization error of $f_n$ is controlled by a term in $n^{-1/2}$ plus an optimization error on the empirical risk minimization
\begin{equation}
\E_{{\cal D}_n}\bracket{{\cal R}(f_n) - {\cal R}(f^*)}
\leq \frac{4 M\kappa}{n^{1/2}} + \E_{{\cal D}_n}\bracket{{\cal R}_{{\cal D}_n}(f_n) - {\cal R}_{{\cal D}_n}^*},
\end{equation}
as long as $f^* \in {\cal F}$.
\end{proposition}
Note that this result can be refined using regularized risk \citep{Sridharan2008}, which would be useful under richer (stronger or weaker) source assumptions \citep[\emph{e.g.},][]{Caponnetto2007}.
Such a refinement would allow switching from a constraint $\norm{\theta} < M$ to define ${\cal F}$ to a regularization parameter $\lambda\norm{\theta}^2$ added in the risk without restrictions on $\norm{\theta}$, which would be better aligned with the current practice of machine learning.
Under Assumption \ref{sgd:ass:source}, this will not fundamentally change the result.
The estimation error can be controlled with the derivation in Appendix \ref{sgd:proof:sgd}, where stochastic gradients correspond to random sampling of a coefficient $i_t\leq n$ plus the choice of a random $U_t$.
For the option without resampling, there exists an acceleration scheme specific to different losses in order to benefit from the strong convexity \citep[\emph{e.g.},][]{Bach2013}.
\subsection{Lower bound}
\label{sgd:proof:lower}
In this section, we prove Theorem \ref{sgd:thm:minmax_opt}.
Let us consider any algorithm ${\cal A}:\cup_{n\in\mathbb{N}}(\mathcal{X}\times\mathcal{Y})^n \to \Theta$ that matches a dataset ${\cal D}_n$ to an estimate $\theta_{{\cal D}_n}\in\Theta$.
Let us consider jointly a distribution $\rho$ and a parameter $\theta$ such that Assumption \ref{sgd:ass:source} holds, that is $f_\rho := \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_\rho[\ell(f(X), Y)] = f_\theta$.
We are interested in characterizing for each algorithm the worst excess of risk it can achieve with respect to an adversarial distribution.
The best worst performance that can be achieved by algorithms matching datasets to parameter can be written as
\begin{equation}
{\cal E} = \inf_{\cal A}\sup_{\theta\in\Theta, \rho\in\prob{\mathcal{X}\times\mathcal{Y}}; f_{\rho} = f_{\theta}} \E_{{\cal D}_n \sim \rho^{\otimes n}}\bracket{\E_{(X, Y)\sim\rho}\bracket{\ell(f_{{\cal A}({\cal D}_n)}(X), Y) - \ell(f_{\theta}(X), Y)}}.
\end{equation}
This provides a lower bound to upper bounds such as~\eqref{sgd:eq:thm} that can be derived for any algorithm.
There are many ways to get lower bounds on this quantity.
Ultimately, we want to quantify the best certainty one can have on an estimate $\theta$ based on some observations $(X_i, Y_i)_{i\leq n}$.
In particular, the algorithms ${\cal A}$ can be seen as rules to discriminate a model $\theta$ from observations ${\cal D}_n$ made under $\rho_\theta$, and where the error is measured through the excess of risk ${\cal R}(f_{\hat\theta}, \rho_\theta) - {\cal R}(f_\theta; \rho_\theta)$ where ${\cal R}(f; \rho) = \E_\rho[\ell(f(X), Y)]$ and $\rho_\theta$ is a distribution parametrized by $\theta$ such that $f_\theta = f_\rho$.
Let us first characterize the measure of error.
Surprisingly, when in presence of Gaussian noise or uniform noise, the excess of risk behaves like a quadratic metric between parameters.
\begin{lemma}[Quadratic behavior of the median regression excess of risk with Gaussian noise]
Consider the random variable $Y \sim {\cal N}(\mu, \sigma^2 I_m)$, denote by $\hat{\mu}$ an estimate of $\mu$, the excess of risk can be developed as
\begin{equation}
\E_{{\cal N}(\mu, \sigma^2 I_m)}[\norm{\hat{\mu} - Y} - \norm{\mu - Y}]
= \frac{c_4\norm{\hat{\mu} - \mu}^2}{\sigma}
+ o\paren{\frac{\norm{\hat{\mu} - \mu}^3}{\sigma^2}},
\end{equation}
where $c_4 = \Gamma(\frac{m+1}{2}) / (2\sqrt{2}\Gamma(\frac{m+2}{2})) \geq (m+2)^{-1/2} / 2$.
\end{lemma}
\begin{proof}
With this specific noise model, one can do the following derivations.
\[
\E_{{\cal N}(\mu, \sigma^2 I_m)}[\norm{\hat{\mu} - Y}]
= \E_{{\cal N}(0, I_m)} [\norm{\hat{\mu} - \mu - \sigma Y}]
= \sigma \E_{{\cal N}(0, I_m)} \bracket{\norm{\frac{\hat{\mu} - \mu}{\sigma} - Y}}.
\]
We recognize the mean of a non-central $\chi$-distribution of parameter $k=m$ and $\lambda = \norm{\frac{\hat{\mu} - \mu}{\sigma}}$.
It can be expressed through the generalized Laguerre functions, which allows us to get the following Taylor expansion
\begin{align*}
\E_{{\cal N}(\mu, \sigma^2 I_m)}[\norm{\hat{\mu} - Y}]
& = \frac{\sqrt{\pi}\sigma}{\sqrt{2}} L_{\frac{1}{2}}^{(\frac{m-2}{2})}\paren{-\frac{\norm{\hat{\mu} - \mu}^2}{2\sigma^2}}
\\&= \frac{\sqrt{\pi} \sigma}{\sqrt{2}} \paren{L_{\frac{1}{2}}^{(\frac{m-2}{2})}(0) + \frac{\norm{\hat{\mu} - \mu}^2}{2\sigma^2} L_{-\frac{1}{2}}^{(\frac{m}{2})}(0)}
+ o\paren{\frac{\norm{\hat{\mu} - \mu}^3}{\sigma^2}}.
\end{align*}
Hence, the following expression of the excess of risk,
\begin{align*}
\E_{{\cal N}(\mu, \sigma^2 I_m)}[\norm{\hat{\mu} - Y} - \norm{\mu - Y}]
& = \frac{\sqrt{\pi}\norm{\hat{\mu} - \mu}^2}{2\sqrt{2}\sigma} L_{-\frac{1}{2}}^{(\frac{m}{2})}(0)
+ o\paren{\frac{\norm{\hat{\mu} - \mu}^3}{\sigma^2}}
\\&= \frac{\Gamma(\frac{m+1}{2})\norm{\hat{\mu} - \mu}^2}{2\sqrt{2}\Gamma(\frac{m+2}{2})\sigma}
+ o\paren{\frac{\norm{\hat{\mu} - \mu}^3}{\sigma^2}}.
\end{align*}
Note that in dimension one, the calculation can be done explicitly by computing integrals with the error function.
\begin{align*}
\E_{{\cal N}(\mu, \sigma^2)}[\norm{\hat{\mu} - Y}]
& = \sigma \E_{{\cal N}(0, 1)} \bracket{Y - \frac{\hat{\mu} - \mu}{\sigma} + 2\ind{Y < \frac{\hat{\mu} - \mu}{\sigma}} \paren{\frac{\hat{\mu} - \mu}{\sigma} - Y}}
\\&= \mu - \hat{\mu} + 2(\hat{\mu} - \mu) \E_{{\cal N}(0, 1)}\bracket{\ind{Y < \frac{\hat{\mu} - \mu}{\sigma}}}
- 2\sigma \E_{{\cal N}(0, 1)}\bracket{Y\ind{Y < \frac{\hat{\mu} - \mu}{\sigma}}}
\\&= \mu - \hat{\mu} + 2(\hat{\mu} - \mu) \paren{\frac{1}{2} + \frac{1}{2}\operatorname{erf}\paren{\frac{\hat{\mu} - \mu}{\sqrt{2}\sigma}}}
- \frac{\sqrt{2}\sigma}{\sqrt{\pi}} \int_{-\infty}^{\frac{\hat{\mu} - \mu}{\sigma}} y e^{-\frac{y^2}{2}}\mathop{}\!\mathrm{d} y
\\&= (\hat{\mu} - \mu) \operatorname{erf}\paren{\frac{\hat{\mu} - \mu}{\sqrt{2}\sigma}}
- \frac{\sqrt{2}\sigma}{\sqrt{\pi}} e^{-\frac{(\hat{\mu} - \mu)^2}{2\sigma^2}},
\end{align*}
where we used the error function, which is the symmetric function defined for $x\in\mathbb{R}_+$ as
\[
\operatorname{erf}(x) = \frac{2}{\sqrt{\pi}} \int_0^x e^{-t^2} \mathop{}\!\mathrm{d} t
= \frac{2}{\sqrt{2 \pi}} \int_0^{\sqrt{2}x} e^{-\frac{u^2}{2}} \mathop{}\!\mathrm{d} u
= 2 \E_{{\cal N}(0, 1)}[\ind{0 \leq Y \leq \sqrt{2} x}].
\]
Developing those two functions in the Taylor series leads to the same quadratic behavior.
\end{proof}
Let us now add a context variable.
\begin{lemma}[Reduction to least-squares]
For $\mathcal{Y}=\mathbb{R}^m$, there exists a $\sigma_m > 0$, such that if $\phi$ is bounded by $\kappa$, and $f^*$ belongs to the class of functions ${\cal F} = \brace{x\to \theta\phi(x)\,\middle\vert\, \theta \in \mathcal{Y}\otimes{\cal H}, \norm{\theta}\leq M}$, and the conditional distribution are distributed as $\paren{Y\,\middle\vert\, X} \sim {\cal N}(f^*(x), \sigma^2 I_m)$, with $\sigma > 2M\kappa\sigma_m$,
\begin{equation}
\forall\, f\in{\cal F}, \qquad{\cal R}(f) - {\cal R}(f^*)
\geq \frac{c_4\norm{f - f^*}_{L^2(\rho_\mathcal{X})}^2}{2\sigma}.
\end{equation}
\end{lemma}
\begin{proof}
According to the precedent lemma, there exists $\sigma_m$ such that $\norm{\hat{\mu} - \mu}\sigma^{-1} \leq \sigma_m^{-1}$ leads to\footnote{This best value for $\sigma_m$ can be derived by studying the Laguerre function, which we will not do in this paper.}
\[
\E_{{\cal N}(\mu, \sigma^2 I_m)}[\norm{\hat{\mu} - Y} - \norm{\mu - Y}]
\geq \frac{c_4\norm{\hat{\mu} - \mu}^2}{2\sigma}.
\]
Let $f$ and $f^*\in{\cal F}$ be parametrized by $\theta$ and $\theta^*$.
For a given $x$, setting $\hat\mu = f_\theta(x) = \theta\phi(x)$ and $\mu = f_{\theta^*}(x)$, we get that, using the operator norm,
\[
\norm{\hat{\mu} - \mu} = \norm{(\theta - \theta^*)\phi(x)}
\leq \norm{\theta - \theta^*}_{\op}\norm{\phi(x)}
\leq \norm{\theta - \theta^*}\norm{\phi(x)}
\leq 2M\kappa.
\]
Hence, as soon as $2M\kappa \leq \sigma\sigma_m^{-1}$, we have that for almost all $x\in\mathcal{X}$
\[
\E_Y\bracket{\norm{f(X) - Y} - \norm{f^*(X) - Y}\,\middle\vert\, X=x}
\geq \frac{c_4\norm{f(X) - f^*(X)}^2}{2\sigma}.
\]
The result follows from integration over $\mathcal{X}$.
\end{proof}
We now have a characterization of the excess of risk that will allow us to reuse lower bounds for least-squares regression.
We will follow the exposition of \cite{Bach2023} that we reproduce and comment here for completeness.
It is based on the generalized Fano's method \citep{Ibragimov1977,Birge1983}.
Learnability over a class of functions depends on the size of this class of functions.
For least-squares regression with a Hilbert class of functions, the right notion of size is given by the Kolmogorov entropy.
Let us call $\epsilon$-packing of ${\cal F}$ with a metric $d$ any family $(f_i)_{i\leq N}\in {\cal F}^N$ such that $d(f_i, f_j) > \epsilon$.
The logarithm of the maximum cardinality of an $\epsilon$-packing defines the $\epsilon$-capacity of the class of functions ${\cal F}$.
We refer the interested reader to Theorem~6 in \cite{Kolmogorov1959} to make a link between the notions of capacity and entropy of a space.
To be perfectly rigorous, the least-squares error in not a norm on the space of $L^2$ functions, but we will call it a {\em quasi-distance} as it verifies symmetry, positive definiteness and the inequality $d(x, y) \leq K(d(x, z) + d(z, y))$ for $K \geq 1$.
Let us define an $\epsilon$-packing with respect to a quasi-distance similarly as before.
The $\epsilon$-capacity of a space ${\cal F}$ gives a lower bound on the number of information to transmit in order to recover a function in ${\cal F}$ up to precision $\epsilon$.
We will leverage this fact in order to show our lower bound.
Let us first reduce the problem to a statistical test.
\begin{lemma}[Reduction to statistical testing]
Let us consider a class of functions ${\cal F}$ and an $\epsilon$-packing $(f_i)_{i\leq N}$ of ${\cal F}$ with respect to a quasi-distance $d(\cdot, \cdot)$ verifying the triangular inequality up to a multiplicative factor $K$.
Then the minimax optimality of an algorithm ${\cal A}$ that takes as input the dataset ${\cal D}_n = (X_i, Y_i)_{i \leq n}$ and output a function in ${\cal F}$ can be related to the minimax optimality of an algorithm ${\cal C}$ that takes an input the dataset ${\cal D}_n$ and output an index $j\in[m]$ through
\begin{equation}
\inf_{\cal A} \sup_{\rho} \E_{{\cal D}_n\sim\rho^{\otimes n}}\bracket{d\paren{f_{{\cal A}({\cal D}_n)}, f_\rho}}
\geq \frac{\epsilon}{2K}
\inf_{\cal C} \sup_{i\in[N]} \Pbb_{{\cal D}_n\sim(\rho_i)^{\otimes n}}\paren{{{\cal C}({\cal D}_n)} \neq i},
\end{equation}
where the supremum over $\rho$ has to be understood as taken over all measures whose marginals can be written ${\cal N}(f^*(x), \sigma)$ for $\sigma$ bigger than a threshold $\sigma_m$ and $f^*\in{\cal F}$, and the supremum over $\rho_i$ taken over the same type of measures with $f^* \in (f_i)_{i\leq N}$.
\end{lemma}
\begin{proof}
Consider an algorithm ${\cal A}$ that takes as input a dataset ${\cal D}_n = (X_j, Y_j)_{j\leq n}$ and output a function $f\in{\cal F}$.
We would like to see ${\cal A}$ as deriving from a classification rule and relate the classification and regression errors.
The natural classification rule associated with the algorithm ${\cal A}$ can be defined through $\pi$ the projection from ${\cal F}$ to $[N]$ that minimizes $d\paren{f, f_{\pi(f)}}$.
The classification error and regression error made by $\pi\circ{\cal A}$ can be related thanks to the $\epsilon$-packing property.
For any index $j\in[N]$
\[
d\paren{f_{\pi\circ{\cal A}({\cal D}_n)}, f_j} \geq \epsilon \ind{\pi\circ{\cal A}({\cal D}_n) \neq j}.
\]
The error made by $f_{\cal A}({\cal D}_n)$ relates to the one made by $f_{\pi\circ{\cal A}({\cal D}_n)}$ thanks to the modified triangular inequality, using the definition of the projection
\[
d\paren{f_{\pi\circ{\cal A}({\cal D}_n)}, f_j}
\leq K \paren{d\paren{f_{\pi\circ{\cal A}({\cal D}_n)}, f_{{\cal A}({\cal D}_n)}}
+ d\paren{f_{{\cal A}({\cal D}_n)}, f_j}}
\leq 2 Kd\paren{f_{{\cal A}({\cal D}_n)}, f_j}.
\]
Finally,
\[
d\paren{f_{{\cal A}({\cal D}_n)}, f_{j}} \geq \frac{\epsilon}{2K} \ind{\pi\circ{\cal A}({\cal D}_n) \neq j}.
\]
Assuming that the data were generated by a $\rho_i$ and taking the expectation, the supremum over $\rho_i$ and the infimum over ${\cal A}$ leads to
\[
\inf_{\cal A}\sup_{\rho_i} \E_{{\cal D}_n\sim\rho_i^{\otimes n}}\bracket{d\paren{f_{{\cal A}({\cal D}_n)}, f_i}}
\geq \frac{\epsilon}{2K}\inf_{{\cal C}=\pi\circ{\cal A}}\sup_{(\rho_i)} \Pbb_{{\cal D}_n\sim\rho_i^{\otimes n}}\paren{{\cal C}({\cal D}_n) \neq i}.
\]
Because $\pi\circ{\cal A}$ are part of classification rules (indeed it parametrizes all the classification rules, simply consider ${\cal A}$ that matches a dataset to one of the functions $(f_i)_{i\leq N}$), and because the distributions $\rho_i$ are part of the distributions $\rho$ defined in the lemma, this last equation implies the stated result.
\end{proof}
One of the harshest inequalities in the last proof is due to the usage of the $\epsilon$-packing condition without considering error made by $d\paren{f_{\pi\circ{\cal A}({\cal D}_n)}, f_j}$ that might be much worse than $\epsilon$.
We will later add a condition on the $\epsilon$-packings to ensure that the $(f_i)$ are not too far from each other.
This will not be a major problem when considering small balls in big dimension spaces.
\subsubsection{Results from statistical testing}
In this section, we expand on lower bounds for statistical testing.
We refer the curious reader to \cite{Cover1991}.
We begin by relaxing the supremum by an average
\begin{align}
\inf_{\cal C} \sup_{i\in[N]} \Pbb_{{\cal D}_n\sim(\rho_i)^{\otimes n}}\paren{{{\cal C}({\cal D}_n)} \neq i}
& = \inf_{\cal C} \sup_{p\in\prob{N}} \sum_{i=1}^N p_i \Pbb_{{\cal D}_n\sim(\rho_i)^{\otimes n}}\paren{{{\cal C}({\cal D}_n)} \neq i}
\\&\geq \inf_{\cal C} \frac{1}{N}\sum_{i=1}^N \Pbb_{{\cal D}_n\sim(\rho_i)^{\otimes n}}\paren{{{\cal C}({\cal D}_n)} \neq i}.
\end{align}
The last quantity can be seen as the best measure of error that can be achieved by a decoder ${\cal C}$ of a signal $i\in[N]$ based on noisy observations ${\cal D}_n$ of the signal.
A lower bound on such a similar quantity is the object of Fano's inequality \citep{Fano1968}.
\begin{lemma}[Fano's inequality]
Let $(X, Y)$ be a couple of random variables in $\mathcal{X}\times\mathcal{Y}$ with $\mathcal{X}$, $\mathcal{Y}$ finite, and $\hat{X}:\mathcal{Y}\to\mathcal{X}$ be a classification rule.
Then, the error $e = e(X, Y) = \ind{X \neq \hat{X}(Y)}$ verifies
\[
H\paren{X\,\middle\vert\, Y} \leq H(e) + \Pbb(e) \log(\card{\mathcal{X}} - 1) \leq \log(2) + \Pbb(e)\log(\card\mathcal{X}).
\]
Where for $(X, Y)\in\prob{\mathcal{X}\times\mathcal{Y}}$, $H(X)$ and $H\paren{X\,\middle\vert\, Y}$ denotes the entropy and conditional entropy, defined as, with the convention $0\log 0 = 0$,
\begin{align*}
& H(X) = -\sum_{x\in\mathcal{X}} \Pbb(X=x) \log(\Pbb(X=x)),
\\ &H\paren{X\,\middle\vert\, Y} = -\sum_{x\in\mathcal{X},y\in\mathcal{Y}} \Pbb(X=x, Y=y) \log(\Pbb\paren{X=x\,\middle\vert\, Y=y}).
\end{align*}
\end{lemma}
\begin{proof}
This lemma is actually the result of two properties.
The first part of the proof is due to some manipulation of the entropy, consisting in showing that
\begin{equation}
\label{sgd:eq:proc_1}
H\paren{X\,\middle\vert\, \hat{X}(Y)} \leq H(e) + \Pbb(e)\log(\card\mathcal{X} - 1).
\end{equation}
Let us first recall the following additive property of entropy
\begin{align*}
H\paren{X, Y\,\middle\vert\, Z} & = -\sum_{x\in\mathcal{X}, y\in\mathcal{Y}, z\in{\cal Z}} \Pbb(X=x, Y=y, Z=z)\log(\Pbb\paren{X=x,Y=y\,\middle\vert\, Z=z})
\\&= -\sum_{x\in\mathcal{X}, y\in\mathcal{Y}, z\in{\cal Z}} \Pbb(X=x, Y=y, Z=z)\log(\Pbb\paren{Y=y\,\middle\vert\, X=x, Z=z})
\\&\qquad\qquad\qquad\qquad - \sum_{x\in\mathcal{X}, y\in\mathcal{Y}, z\in{\cal Z}} \Pbb(X=x, Y=y, Z=z)\log(\Pbb\paren{X=x\,\middle\vert\, Z=z})
\\&= H\paren{Y\,\middle\vert\, X, Z} + H\paren{X\,\middle\vert\, Z}.
\end{align*}
Using this chain rule, we get
\begin{align*}
H\paren{e, X\,\middle\vert\, \hat{X}} & = H\paren{e\,\middle\vert\, X, \hat{X}} + H\paren{X\,\middle\vert\,\hat{X}}
\\&= H\paren{X\,\middle\vert\, e, \hat{X}} + H\paren{e\,\middle\vert\,\hat{X}}
\end{align*}
Because $e$ is a function of $\hat{X}$ and $X$ one can check that $H\paren{e\,\middle\vert\, X, \hat{X}} = 0$,
\begin{align*}
H\paren{e\,\middle\vert\, X,\hat{X}}
& = -\sum_{e, X, \hat{X}} \Pbb(X, \hat{X}) \Pbb\paren{e\,\middle\vert\, X,\hat{X}} \log(\Pbb\paren{e\,\middle\vert\, X,\hat{X}})
\\&= -\sum_{e, X, \hat{X}} \Pbb(X, \hat{X}) \ind{e=\ind{X\neq\hat{X}}}\log(\ind{e=\ind{X\neq\hat{X}}})
= -\sum_{e, X,\hat{X}} \Pbb(X, \hat{X}) \cdot 0 = 0.
\end{align*}
Using Jensen inequality for the logarithm, we get
\begin{align*}
H\paren{X\,\middle\vert\, e, \hat{X}}
& = -\sum_{X, e, \hat{X}} \Pbb(X, e, \hat{X}) \log(\Pbb\paren{X\,\middle\vert\, e, \hat{X}})
\\&= -\sum_{x, x'} \Pbb(X=x, e=0, \hat{X}=x') \log(\Pbb\paren{X=x\,\middle\vert\, e=0, \hat{X}=x'})
\\&\qquad\qquad\qquad\qquad- \Pbb(X=x, e=1, \hat{X}=x') \log(\Pbb\paren{X=x\,\middle\vert\, e=1, \hat{X}=x'})
\\&= -\sum_{x, x'} \Pbb\paren{X = x, \hat{X}=x'}\ind{x=x'} \log(\ind{x=x'})
\\&\qquad\qquad\qquad\qquad- \Pbb(e=1)\ind{x\neq x'}\Pbb(X=x, \hat{X}=x') \log(\Pbb\paren{X=x\,\middle\vert\, \hat{X}=x'})
\\&= \Pbb(e=1)\sum_{x'} \Pbb(\hat{X}=x') \sum_{x\neq x'} \Pbb\paren{X=x\,\middle\vert\, \hat{X}=x'} \log\paren{\frac{1}{\Pbb\paren{X=x\,\middle\vert\, \hat{X}=x'}}}
\\&\leq \Pbb(e=1)\sum_{x'} \Pbb(\hat{X}=x') \log\paren{\sum_{x\neq x'} \Pbb\paren{X=x\,\middle\vert\, \hat{X}=x'} \frac{1}{\Pbb\paren{X=x\,\middle\vert\, \hat{X}=x'}}}
\\&= \Pbb(e=1)\log(\card{\mathcal{X}} - 1).
\end{align*}
Using that conditioning reduces the entropy, which follows again from Jensen inequality,
\begin{align*}
H(X) - H\paren{X\,\middle\vert\, Y}
& = \sum_{x, y} \Pbb(X=x,Y=y) \log\paren{\frac{\Pbb\paren{X=x\,\middle\vert\, Y=y}}{\Pbb\paren{X=x}}}
\\&= -\sum_{x, y} \Pbb(X=x,Y=y) \log\paren{\frac{\Pbb\paren{X=x}\Pbb\paren{Y=y}}{\Pbb\paren{X=x, Y=y}}}
\\&\geq -\log\paren{\sum_{x, y} \Pbb(X=x,Y=y) \frac{\Pbb\paren{X=x}\Pbb\paren{Y=y}}{\Pbb\paren{X=x, Y=y}}} = 0,
\end{align*}
we get
\[
H\paren{e \,\middle\vert\, \hat{X}} \leq H(e) \leq \log(2).
\]
Hence, we have proven that
\[
H\paren{X\,\middle\vert\, \hat{X}} \leq \Pbb(e=1)\log(\card\mathcal{X} - 1) + H(e).
\]
The rest of the proof follows from the so-called data processing inequality, that is
\begin{equation}
H\paren{X\,\middle\vert\, \hat{X}(Y)} \geq H\paren{X\,\middle\vert\, Y}.
\end{equation}
We will not derive it here, since it will not be used in the following.
\end{proof}
In our case, a slight modification of the proof of Fano's inequality leads to the following Proposition.
\begin{lemma}[Generalized Fano's method]
For any family of distributions $(\rho_i)_{i\leq N}$ on $\mathcal{X}\times\mathcal{Y}$ with $N\in\mathbb{N}^*$, any classification rule ${\cal C}:{\cal D}_n \to [N]$ cannot beat the following average lower bound
\begin{equation}
\inf_{\cal C} \frac{1}{N}\sum_{i=1}^N \Pbb_{{\cal D}_n \sim \rho_i^{\otimes n}}({\cal C}({\cal D}_n) \neq i) \log(N - 1) \geq \log(N) - \log(2) - \frac{n}{N^2} \sum_{i, j \in [N]} K\paren{\rho_i\,\middle\vert\vert\, \rho_j},
\end{equation}
where $K\paren{p\,\middle\vert\vert\, q}$ is the Kullback-Leibler divergence defined for any measure $p$ absolutely continuous with respect to a measure $q$ as
\[
K\paren{p\,\middle\vert\vert\, q} = \E_{X\sim q}\bracket{-\log\paren{\frac{\mathop{}\!\mathrm{d} p(X)}{\mathop{}\!\mathrm{d} q(X)}}}.
\]
\end{lemma}
\begin{proof}
Let us consider the joint variable $(X, Y)$ where $X$ is a uniform variable on $[N]$ and $\paren{Y\,\middle\vert\, X}$ is distributed according to $\rho_X^{\otimes n}$.
For any classification rule $\hat{X}:{\cal D}_n\to[N]$, using~\eqref{sgd:eq:proc_1} we get
\[
\frac{1}{N}\sum_{i=1}^N \Pbb_{{\cal D}_n\sim\rho_i^{\otimes n}}\paren{\hat{X}({\cal D}_n) \neq i} = \Pbb(\hat{X} \neq X)\log(N-1) \geq H\paren{X\,\middle\vert\, \hat{X}} - \log(2).
\]
We should work on $H\paren{X\,\middle\vert\, \hat{X}\,\middle\vert\, X}$ with similar ideas to the data processing inequality.
First of all, using the chain rule for entropy
\[
H\paren{X\,\middle\vert\, \hat{X}} = H(X, \hat{X}) - H(\hat{X}) = H(X) + (H(X, \hat{X}) - H(X) - H(\hat{X})) = \log(N) - I(X,\hat{X}),
\]
where $I$ is the mutual information defined as, for $X$ and $Z$ discrete
\begin{align*}
I(X, Z) & = H(X) + H(Z) - H(X, Z)
= \sum_{x, z} \Pbb\paren{X=x, Z=z} \log\paren{\frac{\Pbb(X=x, Z=z)}{\Pbb(X=x)\Pbb(Z=z)}}
\\&= \sum_x \Pbb(X=x) \sum_z \Pbb\paren{Z=z\,\middle\vert\, X=x} \log\paren{\frac{\Pbb\paren{Z=z\,\middle\vert\, X=x)}}{\Pbb(Z=z)}}.
\end{align*}
Similarly, one can define the mutual information for continuous variables.
In particular, we are interested in the case where $X$ is discrete and $Y$ is continuous, denote by $\mu_\mathcal{Y}$ the marginal of $(X, Y)$ over $Y$ and by $\mu\vert_x$ the conditional $\paren{Y\,\middle\vert\, X=x}$.
\[
I(X, Y) = \sum_{x}\Pbb(X=x)\int_{y} \mu\vert_x(\mathop{}\!\mathrm{d} y) \log\paren{\frac{\mu\vert_x(\mathop{}\!\mathrm{d} y)}{\mu(\mathop{}\!\mathrm{d} y)}}.
\]
Let us show the following version of the data processing inequality
\begin{equation}
I(X, \hat{X}(Y)) \leq I(X, Y).
\end{equation}
To do so, we will use the conditional independence of $X$ and $\hat{X}$ given $Y$, which leads to
\begin{align*}
\Pbb\paren{X=x\,\middle\vert\, \hat{X}=x'}
& = \int \Pbb\paren{X=x\,\middle\vert\, Y=\mathop{}\!\mathrm{d} y}\Pbb\paren{Y=\mathop{}\!\mathrm{d} y\,\middle\vert\, \hat{X}=z}
\\&= \int \frac{\Pbb(X=x)\mu\vert_x(\mathop{}\!\mathrm{d} y)}{\mu(\mathop{}\!\mathrm{d} y)}\Pbb\paren{Y=\mathop{}\!\mathrm{d} y\,\middle\vert\, \hat{X}=z}.
\end{align*}
Hence, using Jensen inequality,
\begin{align*}
I(X, \hat{X})
& = H(X) - H\paren{X\,\middle\vert\, \hat{X}}
\\&= H(X) + \sum_{z} \Pbb(\hat{X}=z) \sum_{x} \Pbb(X=x) \log(\Pbb\paren{X=x\,\middle\vert\, \hat{X}=z})
\\&= H(X) + \sum_{z} \Pbb(\hat{X}=z) \sum_{x} \Pbb(X=x) \log\paren{\int \frac{\Pbb(X=x)\mu\vert_x(\mathop{}\!\mathrm{d} y)}{\mu(\mathop{}\!\mathrm{d} y)}\Pbb\paren{Y=\mathop{}\!\mathrm{d} y\,\middle\vert\, \hat{X}=z}}
\\&\leq H(X) + \sum_{z} \Pbb(\hat{X}=z) \sum_{x} \Pbb(X=x)\int \Pbb\paren{Y=\mathop{}\!\mathrm{d} y\,\middle\vert\, \hat{X}=z} \log\paren{\frac{\Pbb(X=x)\mu\vert_x(\mathop{}\!\mathrm{d} y)}{\mu(\mathop{}\!\mathrm{d} y)}}
\\&= H(X) + \sum_{x} \Pbb(X=x) \int \mu(\mathop{}\!\mathrm{d} y) \log\paren{\frac{\Pbb(X=x)\mu\vert_x(\mathop{}\!\mathrm{d} y)}{\mu(\mathop{}\!\mathrm{d} y)}}
\\&=\sum_{x} \Pbb(X=x) \paren{\int \mu(\mathop{}\!\mathrm{d} y) \log\paren{\frac{\Pbb(X=x)\mu\vert_x(\mathop{}\!\mathrm{d} y)}{\mu(\mathop{}\!\mathrm{d} y)}} - \log(P(X=x)}
\\&=\sum_{x} \Pbb(X=x) \int \mu(\mathop{}\!\mathrm{d} y) \log\paren{\frac{\mu\vert_x(\mathop{}\!\mathrm{d} y)}{\mu(\mathop{}\!\mathrm{d} y)}}
\\&= I(X, Y).
\end{align*}
We continue by computing the value of $I(X, Y)$, by definition and using Jensen inequality, we get
\begin{align*}
I(X, Y) & = \frac{1}{N}\sum_{i\in[N]} \int_{{\cal D}_n\sim\rho_i^{\otimes n}} \rho_i^{\otimes n}(\mathop{}\!\mathrm{d} {\cal D}_n) \log\paren{\frac{\rho_i^{\otimes n}(\mathop{}\!\mathrm{d} {\cal D}_n)}{\frac{1}{N}\sum_{j\in[N]} \rho_j^{\otimes n}(\mathop{}\!\mathrm{d} {\cal D}_n)}}
\\&\leq \frac{1}{N}\sum_{i\in[N]} \int_{{\cal D}_n\sim\rho_i^{\otimes n}} \rho_i^{\otimes n}(\mathop{}\!\mathrm{d} {\cal D}_n) \frac{1}{N}\sum_{j\in[N]}\log\paren{\frac{\rho_i^{\otimes n}(\mathop{}\!\mathrm{d} {\cal D}_n)}{\rho_j^{\otimes n}(\mathop{}\!\mathrm{d} {\cal D}_n)}}
= \frac{1}{N^2}\sum_{i,j \in[N]} K\paren{\rho_i^{\otimes n}\,\middle\vert\vert\,\rho_j^{\otimes n}}.
\end{align*}
We conclude from the fact that for $p$ and $q$ two distributions on a space ${\cal Z}$, we have
\begin{align*}
K\paren{p^{\otimes n}\,\middle\vert\vert\, q^{\otimes n}}
& = \int_{{\cal Z}^n} -\log\paren{\frac{\mathop{}\!\mathrm{d} p^{\otimes n}(z_1,\cdots, z_n)}{\mathop{}\!\mathrm{d} q^{\otimes n}(z_1,\cdots, z_n)}} q^{\otimes n}(\mathop{}\!\mathrm{d} z_1,\cdots, \mathop{}\!\mathrm{d} z_n)
\\&= \int_{{\cal Z}^n} -\log\paren{\frac{\prod_{i\leq n}\mathop{}\!\mathrm{d} p(z_i)}{\prod_{i\leq n}\mathop{}\!\mathrm{d} q(z_i)}} q^{\otimes n}(\mathop{}\!\mathrm{d} z_1,\cdots, \mathop{}\!\mathrm{d} z_n)
\\&= \sum_{i\leq n}\int_{{\cal Z}^n} -\log\paren{\frac{\mathop{}\!\mathrm{d} p(z_i)}{\mathop{}\!\mathrm{d} q(z_i)}} q^{\otimes n}(\mathop{}\!\mathrm{d} z_1,\cdots, \mathop{}\!\mathrm{d} z_n)
\\&= \sum_{i\leq n}\int_{\cal Z} -\log\paren{\frac{\mathop{}\!\mathrm{d} p(z_i)}{\mathop{}\!\mathrm{d} q(z_i)}} q(\mathop{}\!\mathrm{d} z_i)
= n K\paren{p\,\middle\vert\vert\, q}.
\end{align*}
This explains the result.
\end{proof}
Let us assemble all the results proven thus far.
In order to reduce our excess risk to a quadratic metric, we have assumed that the conditional distribution $\rho_i\vert_x$ to be Gaussian noise.
In order to integrate this constraint into the precedent derivations, we leverage the following lemma.
\begin{lemma}[Kullback-Leibler divergence with Gaussian noise]
If $\rho_i$ and $\rho_j$ are two different distributions on $\mathcal{X}\times\mathcal{Y}$ such that there marginal over $\mathcal{X}$ are equal and the conditional distributions $\paren{Y\,\middle\vert\, X=x}$ are respectively equal to ${\cal N}(f_i(x), \sigma I_m)$ and ${\cal N}(f_j(x), \sigma I_m)$, then
\[
K\paren{\rho_i \,\middle\vert\vert\, \rho_j} = \frac{1}{2\sigma^2} \norm{f_i - f_j}^2_{L^2(\rho_\mathcal{X})}.
\]
\end{lemma}
\begin{proof}
We proceed with
\begin{align*}
K\paren{\rho_i \,\middle\vert\vert\, \rho_j}
& = \int_\mathcal{X} \E_{Y\sim{\cal N}(f_j(x), \sigma I_m)}\bracket{\frac{\norm{Y-f_i(x)}^2 - \norm{Y-f_j(x)}^2}{2\sigma^2}}\rho_j(\mathop{}\!\mathrm{d} x)
\\&= \int_\mathcal{X} \E_{Y\sim{\cal N}\paren{\frac{f_j(x) - f_i(x)}{\sqrt{2}\sigma}, I_m}}\bracket{\norm{Y}^2} - \E_{Y\sim{\cal N}(0, I_m)}\bracket{\norm{Y}^2}\rho_j(\mathop{}\!\mathrm{d} x)
\\&= \int_\mathcal{X} \paren{m + \frac{\norm{f_j(x) - f_i(x)}^2}{2\sigma^2} - m} \rho_j(\mathop{}\!\mathrm{d} x)
= \frac{\norm{f_j - f_i}^2_{L^2(\rho_\mathcal{X})}}{2\sigma^2},
\end{align*}
where we have used the fact that the mean of a non-central $\chi$-square variable of parameter $(m, \mu^2)$ is $m+\mu^2$.
One could also develop the first two squared norms and use the fact that for any vector $u\in\mathbb{R}^m$, $\E[\scap{Y-f_i(x)}{u}] = 0$ to get the result.
\end{proof}
Combining the different results leads to the following proposition.
\begin{lemma}
\label{sgd:prop:step_1}
Under Assumption \ref{sgd:ass:source} with ${\cal F} = \brace{x\in \mathcal{X}\to \theta\phi(x) \in \mathcal{Y}\,\middle\vert\, \norm{\theta} \leq M}$ and $\phi$ bounded by $\kappa$, for any family $(f_i)_{i\leq N_\epsilon} \in {\cal F}^N$ and any $\sigma > 2M\kappa \sigma_m$
\begin{align*}
& \inf_{\cal A}\sup_\rho \E_{{\cal D}_n\sim\rho^{\otimes n}}[{\cal R}(f_{{\cal A}({\cal D}_n)}; \rho)] - {\cal R}^*(\rho)
\\&\qquad\qquad\qquad\geq \frac{\min_{i,j\in[N]} \norm{f_i - f_j}_{L^2(\rho_\mathcal{X})}^2}{16(m+2)^{1/2}\sigma} \paren{1 - \frac{\log(2)}{\log(N)} - \frac{n\max_{i,j\in[N]} \norm{f_i - f_j}^2_{L^2(\rho_\mathcal{X})}}{2\sigma^2 \log(N)}},
\end{align*}
for any algorithm ${\cal A}$ that maps a dataset ${\cal D}_n \in (\mathcal{X}\times\mathcal{Y})^n$ to a parameter $\theta \in \Theta$.
\end{lemma}
\subsubsection{Covering number for linear model}
We are left with finding a good packing of the space induced by Assumption \ref{sgd:ass:source}.
To do so, we shall recall some property of reproducing kernel methods.
\begin{lemma}[Linear models are ellipsoids]
\label{sgd:lem:lin_ell}
For ${\cal H}$ a separable Hilbert space and $\phi:\mathcal{X}\to{\cal H}$ bounded, the class of functions ${\cal F} = \brace{x\in \mathcal{X}\to \theta\phi(x) \in \mathcal{Y}\,\middle\vert\, \norm{\theta} \leq M}$ can be characterized by
\begin{equation}
{\cal F} = \brace{f:\mathcal{X}\to\mathcal{Y}\,\middle\vert\, \norm{K^{-1/2}f}_{L^2(\rho_\mathcal{X})} \leq M},
\end{equation}
where $\rho_\mathcal{X}$ is any distribution on $\mathcal{X}$ and $K$ is the operator on $L^2(\rho_\mathcal{X})$ that map $f$ to
\[
Kf(x') = \int_{x\in\mathcal{X}} \scap{\phi(x)}{\phi(x')} f(x) \rho_\mathcal{X}(\mathop{}\!\mathrm{d} x),
\]
whose image is assumed to be dense in $L^2$.
\end{lemma}
\begin{proof}
This follows for isometry between elements in ${\cal H}$ and elements in $L^2$.
More precisely, let us define
\[\myfunction{S}{\mathcal{Y}\otimes{\cal H}}{L^2(\mathcal{X}, \mathcal{Y}, \rho_\mathcal{X})}{\theta}{x\to \theta\phi(x).}\]
The adjoint of $S$ is characterized by
\[\myfunction{S^*}{L^2(\mathcal{X}, \mathcal{Y}, \rho_\mathcal{X})}{\mathcal{Y}\otimes{\cal H}}{f}{\E[f(x)\otimes\phi(X)],}\]
which follows from the fact that for $\theta \in \mathcal{Y}\otimes{\cal H}$, $f\in L^2$ we have
\begin{align*}
\scap{\theta}{S^*f}_{\mathcal{Y}\otimes{\cal H}} = \scap{S\theta}{f}_{L^2} & = \sum_{i=1}^m \int_{\mathcal{X}} f_i(x) \scap{\theta_i}{\phi(x)}_{\cal H} \rho_\mathcal{X}(\mathop{}\!\mathrm{d} x)
\\&= \sum_{i=1}^m \scap{\theta_i}{\E[f_i(X) \phi(X)]}_{\cal H}
= \scap{\theta}{\E[f(X)\otimes\phi(X)]}_{\mathcal{Y}\otimes{\cal H}}.
\end{align*}
When $SS^*$ is compact and dense in $L^2$, we have
\[
\norm{\theta}_{\mathcal{Y}\otimes{\cal H}} = \norm{(SS^*)^{-1/2} S\theta}_{L^2(\rho_\mathcal{X})}.
\]
The compactness allows considering spectral decomposition hence fractional powers.
We continue by observing that $SS^* = K$, which follows from
\begin{align*}
(SS^*f)(x') = (S\E[f(X)\otimes\phi(X)])(x')
= \E[f(X)\otimes\phi(X)] \phi(x')
= \E[\scap{\phi(X)}{\phi(x')} f(X)].
\end{align*}
The compactness of $K$ derives from the fact that
\[
\norm{Kf(x')}^2 = \norm{\E[\scap{\phi(X)}{\phi(x')} f(X)]}^2
\leq \E[\norm{\scap{\phi(X)}{\phi(x')} f(X)}^2]
\leq \kappa^2 \norm{f}_{L^2}^2.
\]
Hence, $\norm{K}_{\text{op}} \leq \kappa^2$.
Indeed, it is not hard to prove that the trace of $K$ is bounded by $m\kappa^2$, hence $K$ is not only compact but trace-class.
\end{proof}
It should be noted that the condition on $K$ being dense in $L^2(\rho_\mathcal{X})$ is not restrictive, as indeed all the problem is only seen through the lens of $\phi$ and $\rho_\mathcal{X}$: one can replace $\mathcal{X}$ by $\supp\rho_\mathcal{X}$ and $L^2(\rho_\mathcal{X})$ by the closure of the range of $K$ in $L^2(\rho_\mathcal{X})$ without modifying nor the analysis, nor the original problem.
We should study packing in the ellipsoid ${\cal F} = \brace{f\in L^2\,\middle\vert\, \|K^{-1/2}f\|_{L^2(\rho_\mathcal{X})} \leq M}$.
It is useful to split the ellipsoid between a projection on a finite dimensional space that is isomorphic to the Euclidean space $\mathbb{R}^k$ and on a residual space $R$ where the energies $(\norm{f\vert_R}_{L^2(\rho_\mathcal{X})})_{f\in{\cal F}}^2$ are uniformly small.
We begin with the following packing lemma, sometimes referred to as Gilbert-Varshamov bound \citep{Gilbert1952,Varshamov1957} which corresponds to a more generic result in coding theory.
\begin{lemma}[$\ell_2^2$-packing of the hypercube]
For any $k \in \mathbb{N}^*$, there exists a $k/4$-packing of the hypercube $\brace{0, 1}^k$, with respect to Hamming distance, of cardinality $N = \exp(k/8)$.
\end{lemma}
\begin{proof}
Let us consider $\epsilon > 0$, and a maximal $\epsilon$-packing $(x_i)_{i\leq N}$ of the hypercube with respect to the distance $d(x, y) = \sum_{i\in[k]} \ind{x_i\neq y_i} = \norm{x-y}_1 = \norm{x-y}_2^2$.
By maximality, we have $\brace{0,1}^k \subset \cup_{i\in[N]}B_d(x_i,\epsilon)$, hence
\[
2^k \leq N \card{\brace{x\in\brace{0,1}^k\,\middle\vert\, \norm{x}_1 \leq \epsilon}}.
\]
This inequality can be rewritten with $Z$ a binomial variable of parameter $(k, \sfrac{1}{2})$ as
\(
1 \leq N \Pbb(Z \leq \epsilon).
\)
Using Hoeffding inequality \citep{Hoeffding1963}, when $\epsilon = k/4$ we get
\[
N^{-1} \leq \Pbb(Z \leq k/4) = \Pbb(Z - \E[Z] \leq k/4)
\leq \exp\paren{-\frac{2k^2}{4^2 k}} = \exp\paren{-k/8}.
\]
This is the desired result.
\end{proof}
\begin{lemma}[Packing of infinite-dimensional ellipsoids]
\label{sgd:lem:pack_ell}
Let ${\cal F}$ be the function in $L^2(\rho_\mathcal{X})$ such that $\norm{K^{-1/2}f}_{L^2(\rho_\mathcal{X})}\leq M$ for $K$ a compact operator and $M$ any positive number.
For any $k\in\mathbb{N}^*$, it is possible to find a family of $N \geq \exp(k/8)$ elements $(f_i)_{i\in[N]}$ in ${\cal F}$ such that for any $i\neq j$,
\begin{equation}
\frac{kM^2}{\sum_{i\leq k} \lambda_i^{-1}} \leq \norm{f_i - f_j}_{L^2(\rho_\mathcal{X})}^2 \leq \frac{4k M^2}{\sum_{i\leq k}\lambda_i^{-1}},
\end{equation}
where $(\lambda_i)_{i\in\mathbb{N}}$ are the ordered (with repetition) eigenvalues of $K$.
\end{lemma}
\begin{proof}
Let us denote by $(\lambda_i)_{i\in \mathbb{N}}$ the eigenvalues of $K$ and $(u_i)_{i\in\mathbb{N}}$ in $L^2$ the associated eigenvectors.
Consider $(a_s)_{s\in[N]}$ a $k$-packing of the hypercube $\brace{-1,1}^k$ for $N \geq \exp(k/8)$ with respect to the $\ell^2_2$ quasi-distance and define for any $a\in\brace{a_s}$
\[
f_a = \frac{M}{c} \sum_{s=1}^k a_i u_i,
\]
with $c^2 = \sum_{i=1}^k \lambda_i^{-1}$.
We verify that
\begin{align*}
& \norm{K^{-1/2} f_a}_{L^2}^2 = \frac{M^2}{c^2} \sum_{i=1}^k \lambda_i^{-1} = M^2.
\\&\norm{f_a - f_b}_{L^2}^2 = \frac{M^2}{c^2} \sum_{i=1}^k \abs{a_i - b_i}^2
= \frac{M^2}{c^2} \norm{a_i - b_i}_2^2
\in \frac{M^2}{c^2}\cdot [k, 4k].
\end{align*}
This is the object of the lemma.
\end{proof}
So far, we have proven the following lower bound.
\begin{lemma}
\label{sgd:prop:step_2}
Under Assumption \ref{sgd:ass:source} with ${\cal F} = \brace{x\in \mathcal{X}\to \theta\phi(x) \in \mathcal{Y}\,\middle\vert\, \norm{\theta} \leq M}$ and $\phi$ bounded by $\kappa$, for any family $(f_i)_{i\leq N_\epsilon} \in {\cal F}^N$ and any $\sigma > 2M\kappa \sigma_m$ and $km > 10$,
\[
\inf_{\cal A}\sup_\rho \E_{{\cal D}_n\sim\rho^{\otimes n}}[{\cal R}(f_{{\cal A}({\cal D}_n)}; \rho)] - {\cal R}^*(\rho)
\geq \frac{1}{128}\min\brace{\frac{M^2}{\sigma m^{1/2}\sum_{i\leq k} (k\lambda_i)^{-1}}, \frac{\sigma km^{1/2}}{32 n}},
\]
for any algorithm ${\cal A}$ that maps a dataset ${\cal D}_n \in (\mathcal{X}\times\mathcal{Y})^n$ to a parameter $\theta \in \Theta$, and where $(\lambda_i)$ are the ordered eigenvalue of the operator $K$ on $L^2(\mathcal{X},\mathbb{R},\rho_\mathcal{X})$ that maps any function $f$ to the function $Kf$ defines for $x'\in\mathcal{X}$ as
\[
(Kf)(x') = \int_{x\in\mathcal{X}} \scap{\phi(x)}{\phi(x')} f(x)\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x).
\]
In particular, when $\lambda_i = \kappa^2 i^{-a} / \zeta(\alpha)$, where $\zeta$ denotes the Riemann zeta function, we get the following bounds.
If we optimize with respect to $\sigma$, there exists $n_\alpha \in \mathbb{N}$ such that for any $n > n_\alpha$.
\begin{equation}
\inf_{\cal A}\sup_\rho \E_{{\cal D}_n\sim\rho^{\otimes n}}[{\cal R}(f_{{\cal A}({\cal D}_n)}; \rho)] - {\cal R}^*(\rho)
\geq \frac{M\kappa}{725\zeta(\alpha)^{1/2} n^{1/2}}.
\end{equation}
If we fix $\sigma = \beta M\kappa$ with $\beta \geq 2$, and we optimize with respect to $k$, there exists a constant $c_\beta$ and an integer $n_0$ such that for $n > n_0$ we have
\begin{equation}
\inf_{\cal A}\sup_\rho \E_{{\cal D}_n\sim\rho^{\otimes n}}[{\cal R}(f_{{\cal A}({\cal D}_n)}; \rho)] - {\cal R}^*(\rho)
\geq \frac{M\kappa c_\beta}{\zeta(\alpha)^{\frac{1}{1+\alpha}} n^{\frac{\alpha}{\alpha+1}}}.
\end{equation}
\end{lemma}
\begin{proof}
Reusing Lemma \ref{sgd:prop:step_1}, with the same notations, we have the lower bound in
\[
\frac{\min\norm{f_i - f_j}^2}{16\sigma(m+2)^{1/2}}\paren{1 - \frac{\log(2)}{\log(N)} - \frac{n\max\norm{f_i - f_j}^2}{2\sigma^2\log(N)}}.
\]
Let $K$ and $K_\mathcal{Y}$ be the self-adjoint operators on $L^2(\mathcal{X}, \mathbb{R}, \rho_\mathcal{X})$ and $L^2(\mathcal{X}, \mathcal{Y}, \rho_\mathcal{X})$ respectively, both defined through the formula
\[
(Kf)(x') = \int_{x\in\mathcal{X}} \scap{\phi(x)}{\phi(x')} f(x)\rho_\mathcal{X}(\mathop{}\!\mathrm{d} x).
\]
When $K$ is compact, it admits an eigenvalue decomposition $K = \sum_{i\in\mathbb{N}} \lambda_i u_i\otimes u_i$ where the equality as to be understood as the convergence of operator with respect to the operator norm based on the $L^2$-topology.
It follows from the product structure of $L^2(\mathcal{X}, \mathcal{Y}, \rho_\mathcal{X}) \simeq L^2(\mathcal{X},\mathbb{R},\rho_\mathcal{X})^m$ that $K_\mathcal{Y} = \sum_{i\in\mathbb{R}, j\in[m]} \sum_{i\in\mathbb{N}, j\in[m]} \lambda_i (e_i\otimes y_j) \otimes (e_i\otimes u_j)$ with $(e_j)$ the canonical basis of $\mathcal{Y} = \mathbb{R}^m$.
As a consequence, if $(\lambda_i)_{i\in\mathbb{N}}$ are the ordered eigenvalues of $K$ then $(\lambda_{\floor{i / m}})$ are the ordered eigenvalues of $K_\mathcal{Y}$.
Hence, with Lemmas \ref{sgd:lem:lin_ell} and \ref{sgd:lem:pack_ell}, it is possible to find $N=\exp(km/8)$ functions in ${\cal F}$ such that
\[
\frac{km M^2}{m\sum_{i\leq k} \lambda_i^{-1}} \leq \norm{f_i - f_j}_{L^2(\rho_\mathcal{X})}^2 \leq \frac{4km M^2}{m\sum_{i\leq k}\lambda_i^{-1}}.
\]
If we multiply those functions by $\eta\in[0,1]$ we get a lower bound in
\[
\frac{\eta^2 M^2}{16\sigma(m+2)^{1/2}\sum_{i\leq k}(k\lambda_i)^{-1}}\paren{1 - \frac{8\log(2)}{km} - \frac{16 M^2 n\eta^2}{\sigma^2 km \sum_{i\leq k} (k\lambda_i)^{-1}}}.
\]
Making sure that the last two terms are smaller than one fourth and one half respectively we get the following conditions on $k$ and $\eta$, with $\Lambda_k = \sum_{i\leq k} (k\lambda_i)^{-1}$,
\[
km \geq 32 \log(2),\qquad
32M^2 n \eta^2 \leq \sigma^2 km\Lambda_k.
\]
Using the fact that $\eta < 1$, the lower bound becomes
\[
\frac{M^2}{128\sigma m^{1/2}\Lambda_k} \min\brace{1, \frac{\sigma^2 km \Lambda_k}{32M^2 n}}
=\frac{1}{128}\min\brace{\frac{M^2}{\sigma m^{1/2}\Lambda_k}, \frac{\sigma km^{1/2}}{32 n}},
\]
as long as $km > 10$.
When $\lambda_i^{-1} = i^\alpha \zeta(\alpha) / \kappa^2$, since $\Lambda_k \leq \lambda_k^{-1}$, we simplify the last expression as
\[
\frac{1}{128}\min\brace{\frac{M^2 \kappa^2}{\sigma m^{1/2}k^\alpha \zeta(\alpha)}, \frac{\sigma km^{1/2}}{32 n}}.
\]
Optimizing with respect to $\sigma$ leads to
\[
\sigma^2 = \frac{32nM^2\kappa^2}{mk^{1+\alpha} \zeta(\alpha)} \geq 4M^2\kappa^2 \sigma_m.
\]
This gives
\[
n_{\alpha, m} = m\zeta(\alpha) \sigma_m^2 / 8.
\]
The dependency of $n_\alpha$ to $m$ can be removed since any problem with $\mathcal{Y}=\mathbb{R}^m$ can be cast as a problem in $\mathbb{R}^{m+1}$ by adding a spurious coordinate.
Taking $k=1$ and $m=10$ leads to the result stated in the lemma.
When $n < n_\alpha$, one can artificially multiply the bound by $n_\alpha^{1/2}$, since an optimal algorithm can not do better with fewer data.
After checking that one can take $\sigma_1 \geq 1$, this leads to a bound in
\[
\frac{M\kappa}{2048 n^{1/2}}.
\]
Optimizing with respect to $k$ leads to $k^{\alpha + 1} = 32 M^2 \kappa^2 n / (\sigma^2 m \zeta(\alpha))$ and a bound in
\[
\frac{(\sigma m^{1/2})^{\frac{\alpha-1}{\alpha+1}} (M\kappa)^{\frac{2}{\alpha+1}}}{128(32 n)^{\frac{\alpha}{\alpha+1}} \zeta(\alpha)^{\frac{1}{\alpha+1}}}.
\]
The condition $k > \min\brace{10m^{-1}, 1}$ and $\sigma \geq 2M\kappa\sigma_m$ translates into the condition
\[
4M^2\kappa^2\sigma_m^2 \leq \sigma^2 \leq \frac{32 M^2\kappa^2 n}{m\zeta(\alpha)} \min\brace{1, \frac{m^{1+\alpha}}{10^{1+\alpha}}}.
\]
We deduce that $\sigma_m = O(m^{-1/2})$, otherwise we would not respect the upper bound derived with Rademacher complexity (or have made a mistake somewhere).
Once again we can remove the dependency to $m$. Considering $\sigma =\beta M\kappa$ leads to the result stated in the lemma.
\end{proof}
\subsubsection{Controlling eigenvalues decay}
Based on Lemma \ref{sgd:prop:step_2}, in order to prove Theorem \ref{sgd:thm:minmax_opt}, we only need to show that there exists a mapping $\phi$, an input space $\mathcal{X}$ and a distribution $\rho_\mathcal{X}$ such that the integral operator $K$ introduced in the lemma verifies the assumption on its eigenvalues.
Notice that we show in the proof of Lemma \ref{sgd:prop:step_2} that the universal constant $c_3$ can be taken as $c_3 = 2^{-11}$.
To proceed, let us consider any infinite dimensional Hilbert space ${\cal H}$ with a basis $(e_i)_{i\in\mathbb{N}}$, $\mathcal{X}=\mathbb{N}$ and $\phi:\mathbb{N}\to{\cal H}; i\to \kappa e_i$.
For $a:\mathbb{N}\to\mathbb{R}$ we have
\[
(Ka)(i) = \sum_{j\in\mathbb{N}} \scap{\phi(i)}{\phi(j)} a(j) \rho(j) = \kappa^2 a(i)\rho(i).
\]
Hence, the eigenvalues of $K$ are $(\kappa^2\rho_\mathcal{X}(i))_{i\leq n}$.
It suffices to consider $\rho_\mathcal{X}(i) = i^{-\alpha} / \zeta(\alpha)$ to conclude.
The eigenvalue decay in $O(i^{-\alpha})$ can also be witnessed in many regression problems.
One way to build those cases is to turn a sequence of non-negative real values into a one-periodic function $h$ from $\mathbb{R}^d$ to $\mathbb{R}$ thanks to the inverse Fourier transform.
Using \cite{Bochner1933}, one can construct a map $\phi$ such that the convolution operator linked with $h$ corresponds to the operator $K$.
When $\rho$ is uniform on $[0,1]^d$, diagonalizing this convolution operator with the Fourier functions and using the property in Lemma \ref{sgd:lem:lin_ell} shows that the class of functions ${\cal F}$ are akin to Sobolev spaces.
Similar behavior can be proven when $\mathcal{X}=\mathbb{R}^d$ and $\rho_\mathcal{X}$ is absolutely continuous with respect to the Lebesgue measure and has bounded density \citep{Widom1963}.
We refer the curious reader to \cite{Scholkopf2001} or \cite{Bach2023} for details.
\section{Unbiased weakly supervised stochastic gradients}
\label{sgd:app:generic}
In this section, we provide a generic scheme to acquire unbiased weakly supervised stochastic gradients, as well as specifications of the formula given in the main text for least-squares and median regression.
\subsection{Generic implementation}
Suppose that $\Theta$ is finite dimensional, or that it can be approximated by a finite dimensional space without too much approximation error.
For example, in the realm of scalar-valued kernel methods, it is usual to consider either the random finite dimensional space $\Span\brace{\phi(x_i)}_{i\leq n}$ for $(x_i)$ the data points, or the finite dimension space linked to the first eigenspaces of the operator $\E[\phi(X)\otimes \phi(X)]$.
In the context of neural networks, the parameter space is always finite-dimensional.
Suppose also that, given $\theta$, we know an upper bound $M_\theta$ on the amplitude of $\nabla_\theta \ell(f_\theta(x), y)$, or that we know how to handle clipped gradients at amplitude $M_\theta$ for SGD.
Then, similarly to the least-squares method proposed in the main text, we can access weakly supervised gradient through the formula
\begin{equation*}
\nabla_\theta\ell(f_\theta(x), y) = \frac{2M_\theta(\card\Theta^2 + 4\card\Theta + 3)}{\pi^{3/2}} \E_{U \sim \uniform{B_\Theta},V\sim\uniform{[0, M_\theta]}}[\ind{y\in(z\to\scap{U}{\nabla_\theta \ell(f_\theta(x), z)})^{-1}([V, \infty))} U],
\end{equation*}
where $B_\Theta$ is the unit ball of $\Theta$.
This scheme is really generic, and we do not advocate for it in practice as one may hope to leverage specific structure of the loss function and the parametric model in a more efficient way.
This formula is rather a proof of concept to illustrate that our technique can be applied generically, and is not specific to least-squares or median regression.
\subsection{Specific implementations}
Let us prove the two formulas to get stochastic gradients for both least-squares and median regression.
We begin with median regression.
Consider $z\in\mathbb{S}^{m-1}$, and let us denote
\[
x = \E_U[\sign(\scap{z}{U}) U].
\]
The direction $x/\norm{x}\in\mathbb{S}^{m-1}$ is characterized by the argmax over the sphere of the linear form
\[
y\to \scap{\E_U[\sign(\scap{z}{U})U]}{y}
= \E_U[\sign(\scap{z}{U})\scap{U}{y}].
\]
This linear form has a unique maximizer on $\mathbb{S}^{m-1}$ and by invariance by symmetry over the axis $z$, this maximizer is aligned with $z$, hence $x = c_x\cdot z$.
We compute the amplitude with the formula, because $z$ is a unit vector
\[
c_x = \scap{x}{z} = \E_U[\sign(\scap{z}{U})\scap{U}{z}].
\]
By invariance by rotation of both the uniform distribution and the scalar product, $c_x$ is actually a constant, it is equal to its value $c_2 = c_{e_1}$.
The same type of reasoning applies for the least-squares case.
Consider $z \in\mathbb{R}^m$, and denote
\[
x = \E_{U, V}\bracket{\ind{\scap{z}{U} \geq V}\cdot U}.
\]
For the same reasons as before $x = c_x\cdot u$ for $u = z / \norm{z}$, and $c_x$ verifies
\begin{align*}
c_x
& = \scap{x}{u}
= \E_{U,V}[\ind{\scap{z}{U} \geq V}\scap{U}{u}]
= \E_U[\E_{V}[\ind{\scap{z}{U} \geq V}]\scap{U}{u}]
\\&= \E_{U}[\ind{\scap{z}{U} > 0} \frac{\scap{z}{U}}{M}\scap{U}{u}]
= \frac{\norm{z}}{M}\E_{U}[\ind{\scap{u}{U}>0} \scap{U}{u}^2].
\end{align*}
Hence,
\[
x = \frac{1}{M}\E_{U}[\ind{\scap{u}{U}>0} \scap{U}{u}^2] \cdot z = c_1 \cdot z.
\]
This explains the formula for least-squares.
\begin{lemma}[Constant for the uniform strategy]
Under the uniform distribution on the sphere
\begin{equation}
c_2 = \E_{u\sim\mathbb{S}^{m-1}} \bracket{\abs{\scap{u}{e_1}}}
= \frac{\sqrt{\pi}\,\Gamma(\frac{m-1}{2})}{m\,\Gamma(\frac{m}{2})} \geq \frac{\sqrt{2 \pi}}{m^{3/2}}.
\end{equation}
\end{lemma}
\begin{proof}
Let us compute $c_2 = \E_{u\sim\mathbb{S}^{m-1}} \bracket{\abs{\scap{u}{e_1}}}$.
This constant can be written explicitly as
\[
c_2 = \frac{\int_{x\in\mathbb{S}^{m-1}} \abs{x_1} \mathop{}\!\mathrm{d} x}{\int_{x\in\mathbb{S}^{m-1}} \mathop{}\!\mathrm{d} x}.
\]
Remark that for any function $f:\mathbb{R}\to\mathbb{R}$, we have
\[
\int_{\mathbb{S}^{m-1}} f(x_1) \mathop{}\!\mathrm{d} x
= \int_{x_1 \in [-1, 1]} f(x_1)\mathop{}\!\mathrm{d} x_1 \int_{\tilde{x}\in \sqrt{1-x_1^2}\cdot\mathbb{S}^{m-2}} \mathop{}\!\mathrm{d} \tilde{x}
= \int_{x_1\in[-1,1]} f(x_1) (1 - x_1^2)^{\frac{m-2}{2}} \mathop{}\!\mathrm{d} x_1 \int_{\tilde{x}\in\mathbb{S}^{m-2}} \mathop{}\!\mathrm{d} \tilde{x}.
\]
By denoting $S_{m}$ the surface of the $m$-sphere, the last integral is nothing but $S_{m-2}$.
By setting $f(x)=1$, we can retrieve by recurrence the expression of $S_m$.
In our case, $f(x)=\abs{x}$, so we compute, with $u = 1-x^2$
\[
\int_{x_1\in[-1,1]} \abs{x_1} (1 - x_1^2)^\frac{m-2}{2} \mathop{}\!\mathrm{d} x_1
=2\int_{x_1\in[0,1]} x_1 (1 - x_1^2)^\frac{m-2}{2} \mathop{}\!\mathrm{d} x_1
= \int_{u=0}^1 u^\frac{m-2}{2} \mathop{}\!\mathrm{d} u = \frac{1}{m}.
\]
This leads to
\[
c_2 = \frac{S_{m-2}}{m S_{m-1}} = \frac{\sqrt{\pi}\,\Gamma(\frac{m-1}{2})}{m\,\Gamma(\frac{m}{2})}.
\]
The ratio $S_{m-2} / S_{m-1}$ can be expressed with the integral corresponding to $f=1$, but it is common knowledge that $S_{m-1} = \sfrac{2\pi^{m/2}}{\Gamma(m/2)}$.
\end{proof}
\begin{lemma}[Constant for least-squares]
Under the uniform distributions on $[0, M]$ and the sphere
\begin{equation}
c_1 = \E_{y\sim[0, M]}\E_{u\sim\mathbb{S}^{m-1}} \bracket{\ind{\scap{u}{e_1} > v} \scap{u}{e_1}} = \frac{\pi^{3/2}}{M (m^2 + 4m + 3)}.
\end{equation}
\end{lemma}
\begin{proof}
Similarly to the previous case, this constant can be written explicitly as
\[
c_1 = \frac{1}{2}\frac{\int_{y\in[0,M]}\int_{x\in\mathbb{S}^{m-1}} \abs{x_1} \ind{\abs{x_1} > y} \mathop{}\!\mathrm{d} y\mathop{}\!\mathrm{d} x}{M\int_{x\in\mathbb{S}^{m-1}} \mathop{}\!\mathrm{d} x}
= \frac{\int_{x\in\mathbb{S}^{m-1}} x_1^2 \mathop{}\!\mathrm{d} x}{2M\int_{x\in\mathbb{S}^{m-1}} \mathop{}\!\mathrm{d} x}.
\]
We continue as before with
\[
\int_{x_1\in[-1,1]} \abs{x_1}^2 (1 - x_1^2)^\frac{m-2}{2} \mathop{}\!\mathrm{d} x_1
= 2\int_{x\in[0,1]} x^2 (1 - x^2)^\frac{m-2}{2} \mathop{}\!\mathrm{d} x
= \frac{2\pi \Gamma(\frac{m}{2})}{4\Gamma(\frac{m+3}{2})}.
\]
This leads to
\[
c_1 = \frac{\pi \Gamma(\frac{m}{2})}{4M\Gamma(\frac{m+3}{2})}\cdot\frac{\sqrt{\pi}\,\Gamma(\frac{m-1}{2})}{\Gamma(\frac{m}{2})}
= \frac{\pi^{3/2} \Gamma(\frac{m-1}{2})}{4M\Gamma(\frac{m+3}{2})}
= \frac{\pi^{3/2}}{M (m^2 + 4m + 3)}.
\]
This is the result stated in the lemma.
\end{proof}
\section{Introduction}
A large amount of the current hype around artificial intelligence was fueled by the recent successes of supervised learning.
Supervised learning consists in designing an algorithm that maps inputs to outputs by learning from a set of input/output examples.
When accessing many samples, and given enough computation power, this framework is able to tackle complex tasks.
Interestingly, many of the difficulties arising in practice do not emerge from choosing the right statistical model to solve the supervised learning problem, but from the problem of collecting and cleaning enough data \cite[see Chapters 1 and 2 of][for example]{Geron2017}.
Those difficulties are not disjoint from the current trends toward data privacy regulations \citep{GDPR}.
This fact motivates this work, where we focus on how to efficiently collect information to carry out the learning process.
In this paper, we formalize the ``active labeling'' problem for weak supervision, where the goal is to learn a target function by acquiring the most informative dataset given a restricted budget for annotation.
We focus explicitly on weak supervision that comes as a set of label candidates for each input, aiming to partially supervise input data in the most efficient way to guide a learning algorithm.
We also restrict our study to the streaming variant where, for each input, only a single partial information can be collected about its corresponding output.
The crux of this work is to leverage the fact that full supervision is not needed to acquire unbiased stochastic gradients, and perform stochastic gradient descent.
The following summarizes our contributions.
\begin{enumerate}
\item First, we introduce the ``active labeling'' problem, which is a relevant theoretical framework that encompasses many useful problems encountered by practitioners trying to annotate their data in the most efficient fashion, as well as its streaming variation, in order to deal with privacy preserving issues. This is the focus of Section~\ref{sgd:sec:fram}.
\item Then, in Section~\ref{sgd:sec:sgd}, we give a high-level framework to access unbiased stochastic gradients with weak information only. This provides a simple solution to the streaming ``active labeling'' problem.
\item Finally, we detail this framework for a robust regression task in Section~\ref{sgd:sec:median}, and provide an algorithm whose optimality is proved in Section~\ref{sgd:sec:stat}.
\end{enumerate}
As a proof of concept, we provide numerical simulations in Section~\ref{sgd:sec:exp}.
We conclude with a high-level discussion around our methods in Section~\ref{sgd:sec:discussion}.
\paragraph{Related work.}
Active query of information is relevant to many settings.
The most straightforward applications are searching games, such as Bar Kokhba or twenty questions \citep{Walsorth1882}.
We refer to \citet{Pelc2002} for an in-depth survey of such games, especially when liars introduce uncertainty, and their relations with coding on noisy channels.
But applications are much more diverse, {\em e.g.} for numerical simulation \citep{Chevalier2014}, database search \citep{Qarabaqi2014}, or shape recognition \citep{Geman1993}, to name a few.
In terms of motivations, many streams of research can be related to this problem, such as experimental design \citep{Chernoff1959}, statistical queries \citep{Kearns1998,Fotakis2021}, crowdsourcing \citep{Doan2011}, or aggregation methods in weak supervision \citep{Ratner2020}.
More precisely, ``active labeling''\footnote{Note that the wording ``active labeling'' has been more or less used as synonymous of ``active learning'' \cite[\emph{e.g.,}][]{Wang2014}. In contrast, we use ``active labeling'' to design ``active weakly supervised learning''.} consists in having several inputs and querying partial information on the labels. It is close to active learning \citep{Settles2010,Dasgupta2011,Hanneke2014}, where there are several inputs, but exact outputs are queried; and to active ranking \citep{Valiant1975,Ailon2011,Braverman2019}, where partial information is queried, but there is only one input.
The streaming variant introduces privacy preserving constraints, a problem that is usually tackled through the notion of differential privacy \citep{Dwork2006}.
In terms of formalization, we build on the partial supervision formalization of \cite{Cabannes2020}, which casts weak supervision as sets of label candidates and generalizes semi-supervised learning \citep{Chapelle2006}.
Finally, our sequential setting with a unique final reward is similar to combinatorial bandits in a pure-exploration setting \citep{Garivier2016,Fiez2019}.
\section{The ``active labeling'' problem}
\label{sgd:sec:fram}
Supervised learning is traditionally modeled in the following manner.
Consider $\mathcal{X}$ an input space, $\mathcal{Y}$ an output space, $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ a loss function, and $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$ a joint probability distribution.
The goal is to recover the function
\begin{equation}
\label{sgd:eq:obj}
f^*\in\argmin_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}(f) := \E_{(X, Y)\sim\rho}[\ell(f(X), Y)],
\end{equation}
yet, without accessing $\rho$, but a dataset of independent samples distributed according to $\rho$, ${\cal D}_n = (X_i, Y_i)_{i\leq n}\sim\rho^{\otimes n}$.
In practice, accessing data comes at a cost, and it is valuable to understand the cheapest way to collect a dataset allowing to discriminate $f^*$.
We shall suppose that the input data $(X_i)_{i\leq n}$ are easy to collect, yet that labeling those inputs to get outputs $(Y_i)_{i\leq n}$ demands a high amount of work.
For example, it is relatively easy to scrap the web or medical databases to access radiography images, but labeling them by asking radiologists to recognize tumors on zillions of radiographs will be both time-consuming and expensive.
As a consequence, \emph{we assume the $(X_i)_{i\leq n}$ given but the $(Y_i)_{i\leq n}$ unknown}.
As getting information on the labels comes at a cost ({\em e.g.}, paying a pool of label workers, or spending your own time), given a budget constraint, what information should we query on the labels?
To quantify this problem, we will assume that {\em we can sequentially and adaptively query $T$ information of the type $\ind{Y_{i_t} \in S_t}$, for any index $i_t\in\brace{1, \cdots, n}$ and any set of labels $S_t \subset \mathcal{Y}$ (belonging to a specified set of subsets of $\mathcal{Y}$).}
Here, $t \in \brace{1, \cdots, T}$ indexes the query sequence, and $T\in\mathbb{N}$ is a fixed budget.
The goal is to optimize the design of the sequence $(i_t, S_t)$ in order to get the best estimate of $f^*$ in terms of risk minimization~\eqref{sgd:eq:obj}.
In the following, we give some examples to make this setting more concrete.
\begin{example}[Classification with attributes]
Suppose that a labeler is asked to provide fine-grained classes on images \citep{Krause2016,Zheng2019}, such as the label ``caracal'' in Figure~\ref{sgd:fig:caracal}.
This would be difficult for many people. Yet, it is relatively easy to recognize that the image depicts a ``feline'' with ``tufted-ears'' and ``sandy color''.
As such, a labeler can give the weak information that $Y$ belongs to the set ``feline'', $S_1 = \brace{\text{``cat'', ``lion'', ``tiger''}, \dots}$, and the set ``tufted ears'', $S_2 = \brace{\text{``Great horned owl'', ``Aruacana chicken''}, \dots}$.
This is enough to recognize that $Y \in S_1 \cap S_2 = \brace{\text{``caracal''}}$.
The question $\ind{Y\in S_1}$, corresponds to asking if the image depicts a feline.
Literature on hierarchical classification and autonomic taxonomy construction provides interesting ideas for this problem \citep[{\em e.g.},][]{Cesa-Bianchi2006,Gangaputra2006}.
\end{example}
\begin{figure}[ht]
\centering
\includegraphics[width=.3\textwidth]{sgd/images/caracal.jpg}
\caption{Recognizing fine-grained classes is difficult, but recognizing attributes is easy.}
\label{sgd:fig:caracal}
\end{figure}
\begin{example}[Ranking with partial ordering]
Consider a problem where for a given input $x$, characterizing a user, we are asked to deduce their preferences over $m$ items.
Collecting such a label requires knowing the exact ordering of the $m$ items induced by a user. This might be hard to ask for.
Instead, one can easily ask the user which items they prefer in a collection of a few items. The user's answer will give weak information about the labels, which can be modeled as knowing $\ind{Y_i\in S} = 1$, for $S$ the set of total orderings that satisfy this partial ordering.
We refer the curious reader to active ranking and dueling bandits for additional contents \citep{Jamieson2011,Bengs2021}.
\end{example}
\begin{example}[Pricing a product]
Suppose that we want to sell a product to a consumer characterized by some features $x$, this consumer is ready to pay a price $y\in \mathbb{R}$ for this product.
We price it $f(x)\in\mathbb{R}$, and we observe $\ind{f(x) < y}$, that is if the consumer is willing to buy this product at this price tag or not \citep{Cesa-Bianchi2019,Liu2021}.
Although, in this setting, the goal is often to minimize the regret, which contrasts with our pure exploration setting.
\end{example}
As a counter-example, our assumptions are not set to deal with missing data, {\em i.e.} if some coordinates of some input feature vectors $X_i$ are missing \citep{Rubin1976}.
Typically, this happens when input data comes from different sources ({\em e.g.}, when trying to predict economic growth from country information that is self-reported).
\paragraph{Streaming variation.}
The special case of the active labeling problem we shall consider consists in its variant without resampling.
This corresponds to the online setting where one can only ask one question by sample, formally $i_t=t$.
This setting is particularly appealing for privacy concerns, in settings where the labels $(Y_i)$ contain sensitive information that should not be revealed totally.
For example, some people might be more comfortable giving a range over a salary rather than the exact value; or in the context of polling, one might not call back a previous respondent characterized by some features $X_i$ to ask them again about their preferences captured by $Y_i$.
Similarly, the streaming setting is relevant for web marketing, where inputs model new users visiting a website, queries model sets of advertisements chosen by an advertising company, and one observes potential clicks.
\section{Weak information as stochastic gradients}
\label{sgd:sec:sgd}
In this section, we discuss how stochastic gradients can be accessed through weak information.
Suppose that we model $f = f_\theta$ for some Hilbert space $\Theta\ni\theta$.
With some abuse of notations, let us denote $\ell(x, y, \theta) := \ell(f_\theta(x), y)$.
We aim to minimize
\(
{\cal R}(\theta) = \E_{(X, Y)}\bracket{\ell(X, Y, \theta)}.
\)
Assume that ${\cal R}$ is differentiable (or sub-differentiable) and denote its gradients by $\nabla_\theta {\cal R}$.
\begin{definition}[Stochastic gradient]
A stochastic gradient of ${\cal R}$ is any random function $G:\Theta \to \Theta$ such that
\(
\E[G(\theta)] = \nabla_\theta{\cal R}(\theta).
\)
Given some step size function $\gamma:\mathbb{N}\to\mathbb{R}^*$, a stochastic gradient descent (SGD) is a procedure, $(\theta_t) \in \Theta^\mathbb{N}$, initialized with some $\theta_0$ and updated as
\(
\theta_{t+1} = \theta_t - \gamma(t) G(\theta_t),
\)
where the realization of $G(\theta_t)$ given $\theta_t$ is independent of the previous realizations of $G(\theta_{s})$ given $\theta_s$.
\end{definition}
In supervised learning, SGD is usually performed with the stochastic gradients $\nabla_\theta \ell(X, Y, \theta)$.
More generally, stochastic gradients are given by
\begin{equation}
\label{sgd:eq:sgd_def}
G(\theta) = \ind{\nabla_\theta \ell(X, Y, \theta) \in T}\cdot \tau(T),
\end{equation}
for $\tau:{\cal T}\to\Theta$ with ${\cal T} \subset 2^\Theta$ a set of subsets of $\Theta$, and $T$ a random variable on ${\cal T}$, such that
\begin{equation}
\label{sgd:eq:condition}
\forall\, \theta \in\Theta,\quad \E_T[\ind{\theta\in T} \cdot \tau(T)] = \theta.
\end{equation}
Stated otherwise, if you have a way to image a vector $\theta$ from partial measurements $\ind{\theta\in T}$ such that you can reconstruct this vector in a linear fashion \eqref{sgd:eq:condition}, then it provides you a generic strategy to get an unbiased stochastic estimate of this vector from a partial measurement \eqref{sgd:eq:sgd_def}.
For $\psi:\mathcal{Y}\to\Theta$ a function from $\mathcal{Y}$ to $\Theta$ ({\em e.g.}, $\psi = \nabla_\theta(X, \cdot, \theta)$), a question $\ind{\psi(Y) \in T}$ translates into a question $\ind{Y\in S}$ for some set $S = \psi^{-1}(T) \subset \mathcal{Y}$, meaning that the stochastic gradient \eqref{sgd:eq:sgd_def} can be evaluated from a single query.
As a proof of concept, we derive a generic implementation for $T$ and $\tau$ in Appendix \ref{sgd:app:generic}.
This provides a generic SGD scheme to learn functions from weak queries when there are no constraints on the sets to query.
\begin{remark}[Cutting plane methods]
While we provide here a descent method, one could also develop cutting-plane/ellipsoid methods to localize $\theta^*$ according to weak information, which corresponds to the techniques developed for pricing by \cite{Cohen2020} and related literature.
\end{remark}
\section{Median regression}
\label{sgd:sec:median}
In this section, we focus on efficiently acquiring weak information providing stochastic gradients for regression problems.
In particular, we motivate and detail our methods for the absolute deviation loss.
Motivated by seminal works on censored data \citep{Tobin1958}, we shall suppose that {\em we query half-spaces}.
For an output $y\in\mathcal{Y}=\mathbb{R}^m$, and any hyper-plane $z + u^\perp \subset \mathbb{R}^m$ for $z \in \mathbb{R}^m$, $u \in \mathbb{S}^{m-1}$, we can ask a labeler to tell us which half-space $y$ belongs to.
Formally, {\em we access the quantity $\sign(\scap{y - z}{u})$ for a given unit cost}.
\paragraph{Least-squares.}
For regression problems, it is common to look at the mean square loss
\[
\ell(X, Y, \theta) = \norm{f_\theta(X) - Y}^2,\qquad
\nabla_\theta\ell(X, Y, \theta) = 2(f_\theta(X) - Y)^\top Df_\theta(X),
\]
where $Df_\theta(x)\in\mathcal{Y}\otimes\Theta$ denotes the Jacobian of $\theta \to f_\theta(x)$.
In rich parametric models, it is preferable to ask questions on $Y\in\mathcal{Y}$ rather than on gradients in $\Theta$ which is a potentially much bigger space.
If we assume that $Y$ and $f_\theta(X)$ are bounded in $\ell^2$-norm by $M \in \mathbb{R}_+$, we can adapt~\eqref{sgd:eq:sgd_def} and \eqref{sgd:eq:condition} through the fact that for any $z \in \mathcal{Y}$, such that $\norm{z} \leq 2M$, as proven in Appendix \ref{sgd:app:generic},
\[
\E_{U, V}\bracket{\ind{\scap{z}{U} \geq V}\cdot U}
= c_1\cdot z,\quad\text{where}\quad
c_1 = \E_{U, V}\bracket{\ind{\scap{e_1}{U} \geq V} \cdot\scap{e_1}{U}}
= \frac{\pi^{3/2}}{2M (m^2 + 4m + 3)},
\]
for $U$ uniform on the sphere $\mathbb{S}^{m-1}$ and $V$ uniform on $[0, 2M]$.
Applied to $z = f_\theta(X) - Y$, it designs an SGD procedure by querying information of the type
\(
\ind{\scap{Y}{U} < \scap{f_\theta(X)}{U} - V}.
\)
\paragraph{A case for median regression.}
Motivated by robustness purposes, we will rather expand on median regression.
In general, we would like to learn a function that, given an input, replicates the output of I/O samples generated by the joint probability $\rho$.
In many instances, $X$ does not characterize all the sources of variations of $Y$, {\em i.e.} input features are not rich enough to characterize a unique output, leading to randomness in the conditional distributions $(Y \vert X)$.
When many targets can be linked to a vector $x\in\mathcal{X}$, how to define a consensual $f(x)$?
For analytical reasons, statisticians tend to use the least-squares error which corresponds to asking for $f(x)$ to be the mean of the distribution $\paren{Y\vert X=x}$.
Yet, means are known to be too sensitive to rare but large outputs \citep[see \emph{e.g.},][]{Huber1981}, and cannot be defined as good and robust consensus in a world of heavy-tailed distributions.
This contrasts with the median, which, as a consequence, is often much more valuable to summarize a range of values.
For instance, median income is preferred over mean income as a population indicator \citep[see \emph{e.g.},][]{USCensus}.
\paragraph{Median regression.} The geometric median is variationally defined through the absolute deviation loss, leading to
\begin{equation}
\label{sgd:eq:median}
\ell(X, Y, \theta) = \norm{f_\theta(X) - Y},\qquad
\nabla_\theta \ell(X, Y, \theta) =
\paren{\frac{f_\theta(X) - Y}{\norm{f_\theta(X) - Y}}}^\top
Df_\theta(X).
\end{equation}
Similarly to the least-squares case, we can access weakly supervised stochastic gradients through the fact that for $z \in \mathbb{S}^{m-1}$, as shown in Appendix \ref{sgd:app:generic},
\begin{equation}
\label{sgd:eq:median_sgd}
\E_{U}\bracket{\sign\paren{\scap{z}{U}}\cdot U}
= c_2\cdot z,\quad\text{where}\quad
c_2 = \E_{U}\bracket{\sign\paren{\scap{e_1}{U}} \cdot\scap{e_1}{U}}
= \frac{\sqrt{\pi}\Gamma(\frac{m-1}{2})}{m \Gamma(\frac{m}{2})},
\end{equation}
where $U$ is uniformly drawn on the sphere $\mathbb{S}^{m-1}$, and $\Gamma$ is the gamma function.
This suggests Algorithm~\ref{sgd:alg:sgd}.
\begin{algorithm}[H]
\caption{Median regression with SGD.}
\KwData{A model $f_\theta$ for $\theta \in \Theta$, some data $(X_i)_{i\leq n}$, a labeling budget $T$, a step size rule $\gamma:\mathbb{N}\to\mathbb{R}_+$}
\KwResult{A learned parameter $\hat{\theta}$ and the predictive function
$\hat{f} = f_{\hat{\theta}}$.}
Initialize $\theta_0$.\\
\For{$t\gets 1$ \KwTo $T$}{
Sample $U_t$ uniformly on $\mathbb{S}^{m-1}$.\\
Query $\epsilon = \sign(\scap{Y_t-z}{U_t})$ for $z =
f_{\theta_{t-1}}(X_t)$.\\
Update the parameter
$\theta_{t} = \theta_{t-1} + \gamma(t) \epsilon\cdot U_t^\top(Df_{\theta_{t-1}}(X_t))$.
}
Output $\hat{\theta} = \theta_T$, or some average, {\em e.g.}, $\hat{\theta} = T^{-1}\sum_{t=1}^T \theta_t$.
\label{sgd:alg:sgd}
\end{algorithm}
\section{Statistical analysis}
\label{sgd:sec:stat}
In this section, we quantify the performance of Algorithm~\ref{sgd:alg:sgd} by proving optimal rates of convergence when the median regression problem is approached with (reproducing) kernels.
For simplicity, we will assume that $f^*$ can be parametrized by a linear model (potentially of infinite dimension).
\begin{assumption}
\label{sgd:ass:source}
Assume that the solution $f^*:\mathcal{X}\to\mathbb{R}^m$ of the median regression problem~\eqref{sgd:eq:obj} and \eqref{sgd:eq:median} can be parametrized by some separable Hilbert space ${\cal H}$, and a bounded feature map $\phi:\mathcal{X}\to{\cal H}$, such that, for any $i \in [m]$, there exists some $\theta_i^* \in \cal H$ such that
\(
\scap{f^*(\cdot)}{e_i}_{\mathcal{Y}} = \scap{\theta_i^*}{\phi(\cdot)}_{\cal H},
\)
where $(e_i)$ is the canonical basis of $\mathbb{R}^m$.
Written into matrix form, there exists $\theta^* \in \mathcal{Y}\otimes{\cal H}$, such that
\(
f^*(\cdot) = \theta^* \phi(\cdot).
\)
\end{assumption}
The curious reader can easily relax this assumption in the realm of reproducing kernel Hilbert spaces following the work of \citet{PillaudVivien2018}.
Under the linear model of Assumption~\ref{sgd:ass:source}, Algorithm~\ref{sgd:alg:sgd} is specified with
\(
u^\top Df_\theta(x) = u\otimes \phi(x).
\)
Note that rather than working with $\Theta = \mathcal{Y} \otimes {\cal H}$ which is potentially infinite-dimensional, empirical estimates can be represented in the finite-dimensional space $\mathcal{Y} \otimes \Span\brace{\phi(X_i)}_{i\leq n}$, and well approximated by small-dimensional spaces to ensure efficient computations \citep{Williams2000,Meanti2020}.
One of the key points of SGD is that gradient descent is so gradual that one can use noisy or stochastic gradients without loosing statistical guarantees while speeding up computations.
This is especially true when minimizing convex functions that are nor strongly-convex, {\em i.e.}, bounded below by a quadratic, nor smooth, {\em i.e.}, with Lipschitz-continuous gradient \citep[see, \emph{e.g.},][]{Bubeck2015}.
In particular, the following theorem, proven in Appendix~\ref{sgd:proof:sgd}, states that Algorithm~\ref{sgd:alg:sgd} minimizes the population risk at a speed at least proportional to $O(T^{-1/2})$.
\begin{theorem}[Convergence rates]
\label{sgd:thm:sgd}
Under Assumption~\ref{sgd:ass:source}, and under the knowledge of $\kappa$ and $M$ two real values such that $\E[\norm{\phi(X)}^2] \leq \kappa^2$ and $\norm{\theta^*} \leq M$, with a budget $T\in\mathbb{N}$, a constant step size $\gamma = \frac{M}{\kappa\sqrt{T}}$ and the average estimate $\hat{\theta} = \frac{1}{T}\sum_{t=0}^{T-1} \theta_t$, Algorithm~\ref{sgd:alg:sgd} leads to an estimate $f$ that suffers from an excess of risk
\begin{equation}
\label{sgd:eq:thm}
\E\bracket{{\cal R}\paren{f_{\hat{\theta}}}} - {\cal R}(f^*)
\leq \frac{2\kappa M}{c_2 \sqrt{T}}
\leq \kappa M m^{3/2} T^{-1/2},
\end{equation}
where the expectation is taken with respect to the randomness of $\hat{\theta}$ that depends on the dataset $(X_i, Y_i)$ as well as the questions $(i_t, S_t)_{t\leq T}$.
\end{theorem}
While we give here a result for a fixed step size, one could retake the extensive literature on SGD to prove similar results for decaying step sizes that do not require to know the labeling budget in advance ({\em e.g.} setting $\gamma(t) \propto t^{-1/2}$ at the expense of an extra term in $\log(T)$ in front of the rates), as well as different averaging strategies \citep[see \emph{e.g.},][]{Bach2023}.
In practice, one might not know {\em a priori} the parameter $M$ but could nonetheless find the right scaling for $\gamma$ based on cross-validation.
The rate in $O(T^{-1/2})$ applies more broadly to all the strategies described in Section~\ref{sgd:sec:sgd} as long as the loss $\ell$ and the parametric model $f_\theta$ ensure that ${\cal R}(\theta)$ is convex and Lipschitz-continuous.
Although the constants appearing in front of rates depend on the complexity to reconstruct the full gradient $\nabla_\theta \ell(f_\theta(X_i, Y_i))$ from the reconstruction scheme \eqref{sgd:eq:condition}.
Those constants correspond to the second moment of the stochastic gradient.
For example, for the least-squares technique described earlier one would have to replace $c_2$ by $c_1$ in~\eqref{sgd:eq:thm}.
Theorem~\ref{sgd:thm:minmax_opt}, proven in Appendix~\ref{sgd:proof:lower}, states that any algorithm that accesses a fully supervised learning dataset of size $T$ cannot beat the rates in $O(T^{-1/2})$, hence any algorithm that collects weaker information on $(Y_i)_{i\leq T}$ cannot display better rates than the ones verified by Algorithm~\ref{sgd:alg:sgd}. This proves minimax optimality of our algorithm up to constants.
\begin{theorem}[Minimax optimality]
\label{sgd:thm:minmax_opt}
Under Assumption~\ref{sgd:ass:source} and the knowledge of an upper bound on $\norm{\theta^*}\leq M$, assuming that $\phi$ is bounded by $\kappa$, there exists a universal constant $c_3$ such that for any algorithm~${\cal A}$ that takes as input ${\cal D}_T = (X_i, Y_i)_{i\leq T} \sim\rho^{\otimes T}$ for any $T\in\mathbb{N}$ and output a parameter $\theta$,
\begin{equation}
\sup_{\rho\in {\cal M}_{M}} \E_{{\cal D}_T\sim\rho^{\otimes T}}\bracket{{\cal R}(f_{{\cal A}({\cal D}_T; \rho)})} - {\cal R}(f_\rho; \rho) \geq c_3 M\kappa T^{-1/2}.
\end{equation}
The supremum over $\rho\in{\cal M}_{M}$ has to be understood as the supremum over all distributions $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$ such that the problem defined through the risk ${\cal R}(f; \rho) := \E_{\rho}[\ell(f(X), Y)]$ is minimized for $f_\rho$ that verifies Assumption~\ref{sgd:ass:source} with $\norm{\theta^*}$ bounded by a constant $M$.
\end{theorem}
The same theorem applies for least-squares with a different universal constant.
It should be noted that minimax lower bounds are in essence quantifying worst cases of a given class of problems.
In particular, to prove Theorem~\ref{sgd:thm:minmax_opt}, we consider distributions that lead to hard problems; more specifically, we assumed the variance of the conditional distribution $\paren{Y\,\middle\vert\, X}$ to be high.
The practitioner should keep in mind that it is possible to add additional structure on the solution, leverage active learning or semi-supervised strategy such as uncertainty sampling \citep{Nguyen2021}, or Laplacian regularization \citep{Zhu2003,Cabannes2021c}, and reduce the optimal rates of convergence.
To conclude this section, let us remark that most of our derivations could easily be refined for practitioners facing a slightly different cost model for annotation.
In particular, they might prefer to perform batches of annotations before updating $\theta$ rather than modifying the question strategy after each input annotation.
This would be similar to mini-batching in gradient descent.
Indeed, the dependency of our result on the annotation cost model and on Assumption~\ref{sgd:ass:source} should not be seen as a limitation but rather as a proof of concept.
\section{Numerical analysis}
\label{sgd:sec:exp}
In this section, we illustrate the differences between our active method versus a classical passive method, for regression and classification problems.
Extensive details are provided in Appendix \ref{sgd:app:experiments}.
Our code is available online at \url{anonymized-url}.
Let us begin with the regression problem that consists in estimating the function $f^*$ that maps $x\in[0,1]$ to $\sin(2\pi x) \in \mathbb{R}$.
Such a regular function, which belongs to any H\"older or Sobolev classes of functions, can be estimated with the Gaussian kernel, which would ensure Assumption \ref{sgd:ass:source}, and that corresponds to a feature map $\phi$ such that $k(x, x'):=\scap{\phi(x)}{\phi(x')} = \exp(-\abs{x-x'}/(2\sigma^2))$ for any bandwidth parameter $\sigma > 0$.\footnote{A noteworthy computational aspect of linear models, often refer as the ``kernel trick'', is that the features map $\phi$ does not need to be explicit, the knowledge of $k:\mathcal{X}\times\mathcal{X}\to\mathbb{R}$ being sufficient to compute all quantities of interest \citep{Scholkopf2001}. This ``trick'' can be applied to our algorithms.}
On Figure \ref{sgd:fig:exp_1}, we focus on estimating $f^*$ given data $(X_i)_{i\in [T]}$ that are uniform on $[0, 1]$ in the noiseless setting where $Y_i = f^*(X_i)$, based on the minimization of the absolute deviation loss.
The passive baseline consists in randomly choosing a threshold $U_i\sim{\cal N}(0, 1)$ and acquiring the observations $(\ind{Y_i > U_i})_{i\in [T]}$ that can be cast as the observation of the half-space $S_i = \brace{y\in\mathcal{Y}\,\middle\vert\, \ind{y > U_i} = \ind{Y_i > U_i}} =: s(Y_i, U_i)$.
In this noiseless setting, a good baseline to learn $f^*$ from the data $(X_i, S_i)$ is provided by the infimum loss characterization \citep[see][]{Cabannes2020}
\[
f^* = \argmin_{f:\mathcal{X}\to\mathcal{Y}} \E_{(X, S)}[\inf_{y\in S} \ell(f(X), y)],
\]
where the distribution over $X$ corresponds to the marginal of $\rho$ over $\mathcal{X}$, and the distribution over $\paren{S\,\middle\vert\, X=x}$ is the pushforward of $U\sim{\cal N}(0, 1)$ under $s(f^*(x), \cdot)$.
The left plot on Figure \ref{sgd:fig:exp_1} corresponds to an instance of SGD on such an objective based on the data $(X_i, S_i)$, while the right plot corresponds to Algorithm \ref{sgd:alg:sgd}.
We take the same hyperparameters for both plots, a bandwidth $\sigma=0.2$ and an SGD step size $\gamma =0.3$.
We refer the curious reader to Figure \ref{sgd:fig:exp_1_app} in Appendix \ref{sgd:app:experiments} for plots illustrating the streaming history, and to Figure \ref{sgd:fig:exp_libsvm} for ``real-world'' experiments.
\begin{figure}[ht]
\centering
\includegraphics{sgd/images/one_d_passive}
\includegraphics{sgd/images/one_d_active}
\caption{
{\em Visual comparison of active and passive strategies.}
Estimation in orange of the original signal $f^*$ in dashed blue based on median regression in a noiseless setting.
Any orange point $(x, u)\in\mathbb{R}^2$ corresponds to an observation made that $u$ is below $f^*(x)$, while a blue point corresponds to $u$ above $f^*(x)$.
The passive strategy corresponds to acquiring information based on $\paren{U\,\middle\vert\, x}$ following a normal distribution, while the active strategy corresponds to $\paren{u\,\middle\vert\, x} = f_{\theta}(x)$.
The active strategy reconstructs the signal much better given the budget of $T=30$ observations.
}
\label{sgd:fig:exp_1}
\end{figure}
To illustrate the versatility of our method, we approach a classification problem through the median surrogate technique presented in Proposition \ref{sgd:prop:sur}.
To do so, we consider the classification problem with $m\in\mathbb{N}$ classes, $\mathcal{X} = [0,1]$ and the conditional distribution $\paren{Y\,\middle\vert\, X}$ linearly interpolating between Dirac in $y_1$, $y_2$ and $y_3$ respectively for $x=0$, $x=1/2$ and $x=1$ and the uniform distribution for $x=1/4$ and $x=3/4$; and $X$ uniform on $\mathcal{X} \setminus ([1/4 - \epsilon, 1/4+\epsilon] \cup [3/4-\epsilon, 3/4+\epsilon])$.
\begin{figure}[ht]
\centering
\includegraphics{sgd/images/regression_error}
\includegraphics{sgd/images/classification_error}
\caption{
{\em Comparison of generalization errors of passive and active strategies} as a function of the annotation budget $T$.
This error is computed by averaging over 100 trials.
In solid is represented the average error, while the height of the dark area represents one standard deviation on each side.
In order to consider the streaming setting where $T$ is not known in advance, we consider the decreasing step size $\gamma(t) = \gamma_0 /\sqrt{t}$; and to smooth out the stochasticity due to random gradients, we consider the average estimate $\bar\theta_t = (\theta_1 + \cdots + \theta_t) / t$.
The left figure corresponds to the noiseless regression setting of Figure \ref{sgd:fig:exp_1}, with $\gamma_0 = 1$. We observe the convergence behavior in $O(T^{-1/2})$ of our active strategy.
The right setting corresponds to the classification problem setting described in the main text with $m=100$, $\epsilon = 1/20$, and approached with the median surrogate.
We observe the exponential convergence phenomenon described by \cite{Cabannes2021b}; its kicks in earlier for the active strategy.
The two plots are displayed with logarithmic scales on both axes.
}
\label{sgd:fig:exp_2}
\end{figure}
\section{Discussion}
\label{sgd:sec:discussion}
\subsection{Discrete output problems}
Learning problems with discrete output spaces are not as well understood as regression problems.
This is a consequence of the complexity of dealing with combinatorial structures in contrast with continuous metric spaces.
In particular, gradients are not defined for discrete output models.
The current state-of-the-art framework to deal with discrete output problems is to introduce a continuous surrogate problem whose solution can be decoded as a solution on the original problem \citep{Bartlett2006}.
For example, one could solve a classification task with a median regression surrogate problem, which is the object of the next proposition, proven in Appendix~\ref{sgd:proof:sur}.
\begin{proposition}[Consistency of median surrogate]
\label{sgd:prop:sur}
The classification setting where $\mathcal{Y}$ is a finite space, and $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ is the zero-one loss $\ell(y, z) = \ind{y\neq z}$ can be solved as a regression task through the simplex embedding of $\mathcal{Y}$ in $\mathbb{R}^\mathcal{Y}$ with the orthonormal basis $(e_y)_{y\in\mathcal{Y}}$.
More precisely, if $g^*:\mathcal{X}\to\mathbb{R}^\mathcal{Y}$ is the minimizer of the median surrogate risk ${\cal R}_S(g) = \E\bracket{\norm{g(X) - e_Y}}$, then $f^*:\mathcal{X}\to\mathcal{Y}$ defined as $f^*(x) = \argmax_{y\in\mathcal{Y}} g^*_y(x)$ minimizes the original risk ${\cal R}(f) = \E\bracket{\ell(f(X), Y)}$.
\end{proposition}
More generally, any discrete output problem can be solved by reusing the consistent least-squares surrogate of \cite{Ciliberto2020}.
Algorithm \ref{sgd:alg:sgd} can be adapted to the least-squares problem based on specifications at the beginning of Section \ref{sgd:sec:median}.
This allows using our method in an off-the-shelve fashion for all discrete output problems.
In this setting, Theorem \ref{sgd:thm:sgd} can be refined under margin conditions where our approach would exhibit exponential convergence rates as illustrated on Figure \ref{sgd:fig:exp_2}.
As a side note, while we are not aware of any generic theory encompassing the absolute-deviation surrogate of Proposition~\ref{sgd:prop:sur}, we showcase its superiority over least-squares on at least two types of problems on Figures \ref{sgd:fig:med_ls} and \ref{sgd:fig:med_simplex} in Appendix \ref{sgd:proof:sur}.
\subsection{Supervised learning baseline with resampling}
When resampling is allowed a simple baseline for the active labeling problem is provided by supervised learning.
In regression problems with the query of any half-space, a method that consists in annotating each $(Y_i)_{i\leq n(T, \epsilon)}$ up to precision $\epsilon$, before using any supervised learning method to learn $f$ from $(X_i, Y_i)_{i\leq n(T, \epsilon)}$ could acquire $n(T, \epsilon) \simeq T / m \log_2(\epsilon^{-1})$ data points with a dichotomic search along all directions, assuming $Y_i$ bounded or sub-Gaussian.
In terms of minimax rates, such a procedure cannot perform better than in $n(T,\epsilon)^{-1/2} + \epsilon$, the first term being due to the statistical limit in Theorem~\ref{sgd:thm:minmax_opt}, the second due to the incertitude $\epsilon$ on each $Y_i$ that transfers to the same level of incertitude on $f$.
Optimizing with respect to $\epsilon$ yields a bound in $O(T^{-{1/2}}\log(T)^{1/2})$.
Therefore, this not-so-naive baseline is only suboptimal by a factor $\log(T)^{1/2}$.
In the meanwhile, Algorithm \ref{sgd:alg:sgd} can be rewritten with resampling, as well as Theorem \ref{sgd:thm:sgd}, which we prove in Appendix \ref{sgd:proof:resampling}.
Hence, our technique will still achieve minimax optimality for the problem ``with resampling''.
In other terms, by deciding to acquire more imprecise information, our algorithm reduces annotation cost for a given level of generalization error (or equivalently reduces generalization error for a given annotation budget) by a factor $\log(T)^{1/2}$ when compared to this baseline.
The picture is slightly different for discrete-output problems.
If one can ask any question $s\in 2^\mathcal{Y}$ then with a dichotomic search, one can retrieve any label with $\log_2(m)$ questions.
Hence, to theoretically beat the fully supervised baseline with the SGD method described in Section~\ref{sgd:sec:sgd}, one would have to derive a gradient strategy \eqref{sgd:eq:sgd_def} with a small enough second moment ({\em e.g.}, for convex losses that are non-smooth nor strongly convex, the increase in the second moment compared to the usual stochastic gradients should be no greater than $\log_2(m)^{1/2}$).
How to best refine our technique to better take into account the discrete structure of the output space is an open question.
Introducing bias that does not modify convergence properties while reducing variance eventually thanks to importance sampling is a potential way to approach this problem.
A simpler idea would be to remember information of the type $Y_i \in s$ to restrict the questions asked in order to locate $f_{\theta_t}(X_i) - Y_i$ when performing stochastic gradient descent with resampling.
Combinatorial bandits might also provide helpful insights on the matter.
Ultimately, we would like to build an understanding of the whole distribution $\paren{Y\,\middle\vert\, X}$ and not only of $f^*(X)$ as we explore labels in order to refine this exploration.
\subsection{Min-max approaches}
Min-max approaches have been popularized for searching games and active learning, where one searches for the question that minimizes the size of the space where a potential guess could lie under the worst possible answer to that question.
A particularly well illustrative example is the solution of the Mastermind game proposed by \cite{Knuth1977}.
While our work leverages plain SGD, one could build on the vector field point-of-view of gradient descent \citep[see, \emph{e.g.},][]{Bubeck2015} to tackle min-max convex concave problems with similar guarantees.
In particular, we could design weakly supervised losses $L(f(x), s; \ind{y\in s})$ and min-max games where a prediction player aims at minimizing such a loss with respect to the prediction $f$, while the query player aims at maximizing it with respect to the question $s$, that is querying information that best elicit mistakes made by the prediction player.
For example, the dual norm characterization of the norm leads to the following min-max approach to the median regression
\[
\argmin_{f:\mathcal{X}\to\mathcal{Y}}{\cal R}(f) = \argmin_{f:\mathcal{X}\to\mathcal{Y}}\max_{U\in(\mathbb{S}^{m-1})^{\mathcal{X}\times\mathcal{Y}}} \E_{(X, Y)\sim\rho}\bracket{\scap{U(x, y)}{f(x) - y}}.
\]
Such min-max formulations would be of interest if they lead to improvement of computational and statistical efficiencies, similarly to the work of \cite{Babichev2019}.
For classification problems, the following proposition introduces such a game and suggests its suitability.
Its proof can be found in Appendix \ref{sgd:proof:minmax}.
\begin{proposition}
\label{sgd:prop:minmax}
Consider the classification problem of learning $f^*:\mathcal{X}\to\mathcal{Y}$ where $\mathcal{Y}$ is of finite cardinality, with the 0-1 loss $\ell(z, y) = \ind{z\neq y}$, minimizing the risk \eqref{sgd:eq:obj} under a distribution $\rho$ on $\mathcal{X}\times\mathcal{Y}$.
Introduce the surrogate score functions $g:\mathcal{X}\to\prob{\mathcal{Y}}; x\to v$ where $v = (v_y)_{y\in \mathcal{Y}}$ is a family of non-negative weights that sum to one, as well as the surrogate loss function $L:\prob{\mathcal{Y}}\times{\cal S}\times\brace{-1, 1}\to \mathbb{R}; (v, S, \epsilon) = \epsilon(1 - 2\sum_{y\in S} v_y)$, and the min-max game
\begin{equation}
\label{sgd:eq:minmax}
\min_{g:\mathcal{X}\to\prob{\mathcal{Y}}} \max_{\mu:\mathcal{X}\to\prob{\cal S}} \E_{(X,Y)\sim\rho} \E_{S\sim \mu(x)}\bracket{L(g(x), S; \ind{Y\in S}-\ind{Y\notin S})}.
\end{equation}
When ${\cal S}$ contains the singletons and with the low-noise condition that $\Pbb\paren{Y\neq f^*(x)\,\middle\vert\, X=x} < 1/2$ almost everywhere, then $f^*$ can be learned through the relation $f^*(x) = \argmin_{y\in\mathcal{Y}} g^*(x)_y$ for the unique minimizer $g^*$ of \eqref{sgd:eq:minmax}.
Moreover, the minimization of the empirical version of this objective with the stochastic gradient updates for saddle point problems provides a natural ``active labeling'' scheme to find this $g^*$.
\end{proposition}
\section{Conclusion}
We have introduced the ``active labeling'' problem, which corresponds to ``active partially supervised learning''.
We provided a solution to this problem based on stochastic gradient descent.
Although our method can be used for any discrete output problem, we detailed how it works for median regression, where we show that it optimizes the generalization error for a given annotation budget.
In a near future, we would like to focus on better exploiting the discrete structure of classification problems, eventually with resampling strategies.
Understanding more precisely the key issues in applications concerned with privacy, and studying how weak gradients might provide a good trade-off between learning efficiently and revealing too much information also provide interesting follow-ups.
Finally, regarding dataset annotation, exploring different paradigms of weakly supervised learning would lead to different active weakly supervised learning frameworks.
While this work is based on partial labeling, similar formalization could be made based on other weak supervision models, such as aggregation \citep[\emph{e.g.},][]{Ratner2020}, or group statistics \citep{Dietterich1997}.
In particular, annotating a huge dataset is often done by bagging inputs according to predicted labels and correcting errors that can be spotted on those bags of inputs \citep{ImageNet}.
We left for future work the study of variants of the ``active labeling'' problem that model those settings.
\chapter{Streaming Stochastic Gradients}
\label{chap:sgd}
The following is a reproduction of \cite{Cabannes2022}.
\input{sgd/abstract}
\input{sgd/core}
\begin{subappendices}
\chapter*{Appendix}
\addcontentsline{toc}{chapter}{Appendix}
\input{sgd/appendix/proofs}
\input{sgd/appendix/discussion}
\input{sgd/appendix/experiments}
\end{subappendices}
\section{Proofs}
\label{svm:app:proofs}
\paragraph{Notation.}
This paper makes use of the following standard notations.
We used the simplex notation $\prob{A}$ to denote the space of probability measure over a Polish space $A$. We also used the notation $\mathcal{Y}^\mathcal{X}$ to denote functions from $\mathcal{X}$ to $\mathcal{Y}$ (which can be seen as a sequence of elements in $\mathcal{Y}$ indexed by $\mathcal{X}$).
\paragraph{Equations.}
Some standard derivations in the fast rates literature were omitted in the main paper.
In particular, under Assumption \ref{svm:ass:margin}, we know that when $\abs{g(x) - \eta(x)} < \abs{\eta(x)}$ then $\sign g(x) = \sign \eta(x) = f^*(x)$, hence ${\cal R}(\sign g) - {\cal R}(f^*) \leq \Pbb_X(\sign g(X) \neq f^*(X)) \leq 1_{\norm{g - \eta}_{\infty} \geq \eta_0}$.
As a consequence,
$\E_{{\cal D}_n}[{\cal R}(\sign g_n)] - {\cal R}(f^*) \leq \Pbb_{{\cal D}_n}(\|g - \eta\|_{\infty} \geq \eta_0)$.
Similarly, the outline exposition follows from the fact that
\begin{align*}
\E_{{\cal D}_n}[{\cal R}(\sign g_{\theta_n})] - {\cal R}(f^*)
&= \E_{{\cal D}_n}[{\cal R}(\sign g_{\theta_n}) - {\cal R}(\sign g_{\theta^*})]
\leq \E_{{\cal D}_n}[1_{\sign g_{\theta_n} \neq (\sign g_{\theta^*})}]
\\&\leq \E_{{\cal D}_n}[1_{\norm{\theta_n - \theta^*}\geq \epsilon}]
\leq \Pbb_{{\cal D}_n}(\norm{\theta_n - \theta^*}\geq \epsilon)
\\&\leq \Pbb_{{\cal D}_n}({\cal R}_S(g_{\theta_n}) - {\cal R}(g_{\theta^*})\geq \epsilon).
\end{align*}
\subsection{Proof of Lemma \ref{svm:lem:l1}}
The first part follows by integration of a pointwise result.
Consider the function $h_p: \mathbb{R}\to\mathbb{R}; q \mapsto p(1-q)_+ + (1-p)(1+q)_+$, where $p\in(0, 1)$ represents $\Pbb(Y=1\vert X)$ and $q$ represents $g(x)$.
The function $h_p$ has a slope equal to $-p$ for $q < -1$, then slope $1-2p$ for $q\in(-1, 1)$, and $1-p$ for $q > 1$. Therefore, when $q_1, q_2 \in (-1, 1)$, we have
\[
h_p(q_2) - h_p(q_1) = (1-2p) (q_2 - q_1).
\]
Taking $p = \Pbb(Y=1\vert X)$, $q_2 = g_2(X)$ and $q_1 = g_1(X)$, we get $1-2p = -\E[Y\vert X] = -\eta(X)$. By integration, we obtain the claim.
From the previous slope considerations, it also follows that $h_p$ is minimized by $q=\sign(2p-1)$, meaning that one can take $g^*(X) = \sign\eta(X)$.
The second part follows from the fact that projecting on $[-1, 1]$ can only reduce the value of the hinge loss, that $\eta(x)(\pi(g)(x) - g^*(x))$ is always negative, and the reverse H\"older inequality:
\[
{\cal R}_S(g) - {\cal R}(g^*) \geq {\cal R}_S(\pi(g)) - {\cal R}(g^*) = \norm{\eta (\pi(g) - g^*)}_{1}
\geq \norm{\pi(g) - g^*}_{q} \norm{\abs{\eta}^{-1}}_{p}^{-1}.
\]
A H\"older inequality also holds for weak Lebesgue spaces \citep[see][Theorem 5.23]{Castillo2016}, whence
\[
{\cal R}_S(g) - {\cal R}(g^*) \geq \norm{\eta (\pi(g) - g^*)}_{1}
\geq \norm{\eta (\pi(g) - g^*)}_{1,\infty}
\geq \frac{1}{2} \norm{ |\eta|^{-1} }_{p,\infty}^{-1} \norm{ \pi(g) - g^* }_{\frac{p}{p+1},\infty}.
\]
This completes the proof.
\subsection{Proof of Lemma \ref{svm:lem2}}
Assume without restrictions that there exists $x \in \mathcal{X}_1$ such that $|g^*(x) - g(x)| \geq 1 $.
For any event $A = A(X)$, by the law of total probability we have
\[
\Pbb(A) =
\rho_\mathcal{X}(\mathcal{X}_1) \Pbb\paren{A\,\middle\vert\, X\in \mathcal{X}_1}
+ \rho_\mathcal{X}(\mathcal{X}_{-1}) \Pbb\paren{A\,\middle\vert\, X\in \mathcal{X}_{-1}}
\geq \rho_\mathcal{X}(\mathcal{X}_1) \Pbb\paren{A\,\middle\vert\, X\in \mathcal{X}_1}.
\]
Hence, since $g^*(\mathcal{X}_1) = \brace{1}$,
\begin{align*}
\norm{g - g^*}_{q, \infty}^q
= \sup_{t > 0} t^q \Pbb(\abs{g(X) - g^*(X)} > t)
\geq \sup_{t > 0} t^q\Pbb\paren{\abs{g(X) - 1} > t\,\middle\vert\, X\in\mathcal{X}_1}\rho_\mathcal{X}(\mathcal{X}_1).
\end{align*}
Using the triangular inequality, the $G$-Lipschitz continuity of $g$, and the definition of $x$, we have that, for any $x'\in\mathcal{X}$,
\[
|g(x') - 1| \geq |g(x) - 1| - |g(x') - g(x)| \geq 1 - Gd(x, x').
\]
As a consequence,
\[
\Pbb\paren{|g(X) - 1| > t\,\middle\vert\, X\in\mathcal{X}_1}
\geq \Pbb\paren{X\in{\cal B}\paren{x, \frac{1-t}{G}} \mid X\in\mathcal{X}_1}
= \frac{\rho_\mathcal{X}\paren{\mathcal{X}_1\cap {\cal B}\paren{x, \frac{1-t}{G}}}}{\rho_\mathcal{X}\paren{\mathcal{X}_1}}.
\]
Combined with the previous facts, we get
\[
\norm{g - g^*}_{q, \infty}^q
\geq \sup_{t > 0} t^q \rho_\mathcal{X}\paren{\mathcal{X}_1\cap {\cal B}\paren{x, \frac{1-t}{G}}}.
\]
Thanks to Assumption \ref{svm:ass:mass}, there exists $(c, r, d)$ such that \eqref{svm:eq:mass} holds for $\mathcal{X}_1$. Hence, when $G^{-1} < r$, we get the following lower bound:
\[
\norm{g - g^*}_{q, \infty}^q
\geq c G^{-d} \sup_{t \in [0, 1]} t^q (1 - t)^d
= c G^{-d} \frac{q^q d^d}{(d+q)^{d+q}}.
\]
This proves the statement in the lemma.
\subsection{Proof of Proposition \ref{svm:prop:mid}}
Suppose $ {\cal R}(\sign g) > {\cal R}(f^*) $.
Then, observing that $ \sign(\pi(t)) = \sign(t) $ for all $ t \in \mathbb{R} $, and taking $ g^* = f^* $, we know there must be $ x \in \supp \rho_\mathcal{X} $ such that $ \abs{ \pi(g(x)) - g^*(x) } \geq 1 $.
Hence, by Lemma \ref{svm:lem2}, we get $ \norm{ \pi(g) - g^* }_{q,\infty} \geq c_0 G^{-\frac{d}{q}} $, and therefore, by Lemma \ref{svm:lem:l1},
$
{\cal R}_S(g) - {\cal R}_S(g^*)
\geq 2^{-1} c_0 G^{-\frac{d}{q}} \norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} .
$
Thus, the proposition is proved.
\subsection{Proof of Theorem \ref{svm:thm}}
From Proposition \ref{svm:prop:mid} and Assumption \ref{svm:ass:source}, we get, with $\tilde{L} = \max\brace{M G_\phi \sigma^{-1}, r^{-1}}$ and $q = \sfrac{p}{p+1}$,
\begin{align*}
\E_{{\cal D}_n}[{\cal R}(\sign g_{{\cal D}_n})] - {\cal R}(f^*)
&\leq \Pbb_{{\cal D}_n} \paren{{\cal R}_S(\pi\circ g_{{\cal D}_n}) - {\cal R}_S(g^*) \geq 2^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 \tilde{L}^{-\frac{d}{q}}}
\\&\leq \Pbb_{{\cal D}_n} \paren{{\cal R}_S(\pi\circ g_{{\cal D}_n}) - {\cal R}_S(g_{M,\sigma}) \geq 4^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 \tilde{L}^{-\frac{d}{q}}}.
\end{align*}
To deal with this last quantity, we proceed by usign the fact that
\[
{\cal R}_{S,{\cal D}_n}(\pi\circ g_{{\cal D}_n}) \leq {\cal R}_{S,{\cal D}_n}(g_{{\cal D}_n}) \leq {\cal R}_{S,{\cal D}_n}(g_{M,\sigma}),
\]
where ${\cal R}_{S,{\cal D}_n}$ denotes the empirical surrogate risk,
to deduce that
\[
{\cal R}_S(\pi\circ g_{{\cal D}_n}) - {\cal R}_S(g_{M,\sigma})
\leq {\cal R}_S(\pi\circ g_{{\cal D}_n}) - {\cal R}_{S,{\cal D}_n}(\pi\circ g_{{\cal D}_n}) + {\cal R}_{S,{\cal D}_n}(g_{M,\sigma}) + {\cal R}_S(g_{M,\sigma}).
\]
Hence, we get the following union bound
\begin{align*}
\E_{{\cal D}_n}[{\cal R}(\sign g_{{\cal D}_n})] - {\cal R}(f^*)
&\leq \Pbb_{{\cal D}_n} \paren{{\cal R}_S(\pi\circ g_{{\cal D}_n}) - {\cal R}_{S, {\cal D}_n}(g_{{\cal D}_n}) \geq 8^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 \tilde{L}^{-\frac{d}{q}}}
\\&\qquad+ \Pbb_{{\cal D}_n} \paren{{\cal R}_S(\pi\circ g_{M,\sigma}) - {\cal R}_{S,{\cal D}_n}(g_{M,\sigma}) \geq 8^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 \tilde{L}^{-\frac{d}{q}}}.
\end{align*}
Regarding the first term, we can reuse the literature on Rademacher complexity for linear models on convex risks \citep{Bartlett2002}, which ensures that
\[
\E_{{\cal D}_n}\bracket{\sup_{g\in{\cal G}_{M,\sigma}}\abs{{\cal R}_S(\pi\circ g) - {\cal R}_{S, {\cal D}_n}(\pi\circ g)}} \leq M \norm{\phi}_{\infty} n^{-1/2}.
\]
Note that Assumption \ref{svm:ass:mass} implies that $\supp\rho_\mathcal{X}$ is compact, hence, if $\phi$ is Lipschitz-continuous, it is bounded on $\supp\rho_\mathcal{X}$.
This allows us to use McDiarmid inequality to get the same type of bound on the deviation of ${\cal R}_S(g_{{\cal D}_n})$ around its mean.
Let $H({\cal D}_n) = \sup_{g\in{\cal G}_{M,\sigma}} {\cal R}_S(\pi\circ g) - {\cal R}_{{\cal D}_n}(\pi\circ g)$.
Let us decompose ${\cal D}_n = ((x_1, y_1), \cdots, (x_n, y_n))$. We would like to show that if ${\cal D}_n'$ is equal to ${\cal D}_n$ for each datapoint but for $(x_i, y_i)$ that becomes $(x_i', y_i')$ then $H({\cal D}_n) - H({\cal D}_n')$ is bounded.
We have
\begin{align*}
H({\cal D}_n) - H({\cal D}_n')
&= \sup_{g\in{\cal G}_{M,\sigma}} {\cal R}_S(\pi\circ g) - {\cal R}_{S,{\cal D}_n}(\pi\circ g) - \sup_{g'\in{\cal G}_{M,\sigma}} {\cal R}_S(\pi\circ g') - {\cal R}_{S,{\cal D}_n'}(\pi\circ g')
\\&\leq \sup_{g\in{\cal G}_{M,\sigma}} {\cal R}_{S, {\cal D}_n}(\pi\circ g) - {\cal R}_{S,{\cal D}_n'}(\pi\circ g)
\\& = n^{-1}\sup_{g\in{\cal G}_{M, \sigma}} L(\pi\circ g(x_i'), y_i') - L(\pi\circ g(x_i), y_i)
\leq n^{-1}
\end{align*}
Using McDiarmid's inequality, we get
\[
\Pbb(H({\cal D}_n) - \E[H({\cal D}_n)] \geq t) \leq \exp(-2nt^2).
\]
In other terms, when adding the control we have on the expectation, we get
\begin{equation}
\Pbb_{{\cal D}_n}\paren{\sup_{g\in{\cal G}_{M,\sigma}} {\cal R}_S(\pi\circ g) - {\cal R}_{S, {\cal D}_n}(\pi\circ g) > t + M\norm{\phi}_{\infty} n^{-1/2}} \leq \exp\paren{-2nt^2}.
\end{equation}
When $8^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 \tilde{L}^{-\frac{d}{q}} \geq \norm{\phi}_{\infty} M n^{-1/2}$, this leads to
\begin{align*}
&\Pbb_{{\cal D}_n} \paren{{\cal R}_S(\pi\circ g_{{\cal D}_n}) - {\cal R}_{S, {\cal D}_n}(g_{{\cal D}_n}) \geq 8^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 \tilde{L}^{-\frac{d}{q}}}
\\&\leq \exp\paren{-\frac{n}{8} \paren{8^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 \tilde{L}^{-\frac{d}{q}} - M\norm{\phi}_{\infty} n^{-1/2}}^2}
\\&\leq \exp\paren{-\frac{c_0^2 \sigma^{\frac{2d(p+1)}{p}}}{512\norm{\abs{\eta}^{-1}}_{p,\infty}^2(MG_\phi)^{\frac{2d(p+1)}{p}}}\cdot n + \frac{c_0\norm{\phi}_{\infty}}{32\norm{\abs{\eta}^{-1}}_{p,\infty}M^{\frac{d(p+1)}{p}-1}G_\phi^{\frac{d(p+1)}{p}}} \cdot n^{-1/2} - \frac{M^2\norm{\phi}_{\infty}^2}{8}}.
\end{align*}
Regarding the second term, we can use classical concentration of ${\cal R}_{{\cal D}_n}(g_{M,\sigma})$ around its mean.
For example, using the fact that Assumption \ref{svm:ass:mass} implies that $\rho_\mathcal{X}$ is compact, and using the fact that $L$ and $g_{\sigma, M}$ are Lipschitz, we deduce that $L(g_{\sigma, M}, Y)$ is bounded, hence one can applies Hoeffding's inequality to get the same type of exponential control on this term.
The result follows form those concentration inequality and the fact that ${\cal R}$ is bounded by one and that any $\min(1, a \exp(-bn))$ for $a, b > 0$ and $n>1$ can be bounded by $\exp(-cn)$ for a $c>0$).
\section{Experimental details}
\label{svm:app:experiments}
In our experiments, we used the SVM implementation of \cite{Chang2011} through its {\em Scikit-learn} wrapper \citep{Pedregosa2011} in {\em Python}.
We used {\em Numpy} \citep{Harris2020} to reduce our work to high-level array instructions, and {\em Matplotlib} for visualization \citep{Hunter2007}.
Randomness in experiments was controlled with the random seed provided by {\em Numpy}, which we initialized at zero.
\begin{figure}[ht]
\centering
\includegraphics{svm/images/fig2_app}
\includegraphics{svm/images/fig2_krr}
\caption{
(Left) Similar setting as Figure \ref{svm:fig:fig1} but with $X$ uniform on $[-1, 1]$.
The behavior of the excess of risk is quite different without the separation in $\mathcal{X}$: no exponential convergence rate is kicking in after a thousand of samples.
(Right) Similar setting as Figure \ref{svm:fig:fig1}, using kernel ridge regression with the least-squares surrogate.
Exponential convergence rates are observed with a slight delay compared to the hinge loss, and are explained by the hard-margin condition \ref{svm:ass:margin}.
}
\label{svm:fig:fig1_app}
\end{figure}
Figures \ref{svm:fig:fig1} and \ref{svm:fig:fig1_app} are derived by averaging 100 trials of the following procedure.
We draw uniformly at random $n$ independent samples uniformly distributed on $\mathcal{X} \in \brace{[-1, 1], [-1, -.1]\cup [.1, 1]}$.
We draw randomly one output $y_i$ for each input $x_i$, according to $\eta(x_i)$.
We consider the Gaussian kernel $k(x, x') = \exp(-\norm{x-x'}^2 / 2\sigma^2)$ for $\sigma = .2$, and solve the empirical risk minimization associated to the hinge loss with the penalization $\lambda \norm{\theta}^2$ (rather than the hard constraint $\norm{\theta} < M$) for $\lambda = 10^{-4}$.
The generalization error is measured through the formula $\E[\norm{\eta(x)} \ind{f(x) \neq f^*(x)}]$, with an empirical approximation of this sum with the points $(x_i)_{i\leq n}$ chosen such that $\rho_{\mathcal{X}}([x_i, x_{i+1}]) = 1/n$ and $\rho_{\mathcal{X}}([x_n, +\infty)) = 1/n$, with $n = 10^4$ (which makes sure that the exponential behavior observed is not due to the lack of testing samples).
For each $x$, the height of each dark part corresponds to one standard deviation of the generalization error computed from the 100 trials, and the solid line corresponds to the empirical average.
The fact that the dark parts are not centered around the averages is due to the fact that we have drawn $\log$-plots but centered the interval for linearly-scaled plots.
Figure~\ref{svm:fig:fig2} is obtained by considering $\mathcal{X} = [0, 1]^2$ with uniform input distribution, the Gaussian kernel with $\sigma = .2$, and the penalty parameter $\lambda = 10^{-3}$ (instead of a hard constraint leading to a parameter $M$ as in the main text derivations).
We take $n=10^4 = 100^2$ points uniformly spread out on $\mathcal{X}$ (on the regular lattice $\frac{1}{\sqrt{n}}\cdot \mathbb{Z}^2 \cap \mathcal{X}$) to approximate $g_{\lambda, \sigma}$ with empirical risk minimization on this curated dataset.
We consider $\eta(x) = \pi_{[-1, 1]}(2 x_2 - .5\sin(2\pi x_1) - 1)$, and assign to each $x$ in the dataset a sample $(x, 1)$ weighted by $\Pbb\paren{Y=1\,\middle\vert\, X=x} = (\eta(x) - 1) / 2$, and a sample $(x, -1)$, weighted by $\Pbb\paren{Y=-1\,\middle\vert\, X=x}$.
The ``noiseless'' setting denotes the setting where $\paren{Y\,\middle\vert\, X}$ is deterministic, but with the same decision frontier between the classes $\mathcal{X}_1$ and $\mathcal{X}_{-1}$ characterized by $\brace{(x, .5 + .25\sin(2\pi x))\,\middle\vert\, x\in [0,1]}$.
Once we fit the support vector machine with this dataset, we test it with $n=2.5\cdot 10^5 = 500^2$ data points uniformly spread out on $\mathcal{X}$, and use {\em Matplotlib} to automatically draw level lines.
\begin{figure}[ht]
\centering
\includegraphics{svm/images/fig8_bis}
\includegraphics{svm/images/fig9_bis}
\includegraphics{svm/images/fig8_ter}
\includegraphics{svm/images/fig9_ter}
\caption{
Same setting as Figure \ref{svm:fig:fig3}, with $\sigma = .2$ and $\lambda = 10^{-6}$ (top), and with $\sigma = 1$ and $\lambda = 10^{-3}$ (bottom).
}
\label{svm:fig:fig3_app}
\end{figure}
Figures~\ref{svm:fig:fig3} and \ref{svm:fig:fig3_app} correspond to $\mathcal{X} = [0, 3]$ with the input distribution uniform on $[0,1]\cup[2,3]$.
Figure~\ref{svm:fig:fig3} is obtained with $\sigma=.1$ and $\lambda=10^{-6}$.
We derive it by considering $n=100$ points uniformly spread out on the domain of $\eta$, solving the equivalent curated empirical risk minimization, that approximates both
\begin{align}
g_{\lambda, \sigma} & = \argmin_{g:\mathcal{X}\to\mathbb{R}} \E_{\rho}[(0, 1- Y\scap{\theta}{\phi\paren{\frac{x}{\sigma}}})_+] + \lambda \norm{\theta}^2 ,
\\g_{(\text{LS})} &= \argmin_{g:\mathcal{X}\to\mathbb{R}} \E_{\rho}[\norm{\scap{\theta}{\phi\paren{\frac{x}{\sigma}}} - Y}^2] + \lambda \norm{\theta}^2.
\end{align}
The robustness of SVM might be understood from its geometrical definition: when trying to find the maximum separating margin, infinitesimal modifications that change the regularity properties of $\eta$ do not really matter.
The picture is different for the least-squares surrogate with kernel methods, where from few point evaluations, the system reconstructs a function by assuming regularity and inferring information on high-order derivatives.
This is similar to the Runge phenomenon with Hermite interpolation.
More precisely, the Gaussian kernel is linked to a space of functions with rapidly decreasing Fourier coefficients \citep[see, for example,][for a more precise link]{Bach2023}.
The function $\eta$ that needs to be approximated on Figure~\ref{svm:fig:fig3} is similar to the Heaviside function, whose Fourier coefficients are of the form $(\frac{1}{i\pi k})_{k\in\mathbb{N}^*}$ and do not decrease fast enough to be all reconstructed.
This leads to some high-frequency oscillations missing in the reconstruction as it appears on Figure~\ref{svm:fig:fig3}.
\section{Introduction}
To solve a problem with computer calculations, classical computer science consists in handcrafting a set of rules.
In contrast, machine learning is based on the collection of a vast amount of solved instances of this problem, and on the automatic tuning of an algorithm that maps inputs defining the problem to the desired outputs.
Denote by $x$ in a space $\mathcal{X}$ the inputs, by $y\in\mathcal{Y}$ the outputs, and by $f:\mathcal{X}\to\mathcal{Y}$ the input/output mappings.
To learn a mapping $f^*$, it is customary to introduce an explicit metric of error, and search for the function that minimizes it.
Define this metric through a loss $\ell:\mathcal{Y}\times\mathcal{Y}\to\mathbb{R}$ that quantifies how bad a prediction $f(x)$ is when we observe $y$.
If we assume the existence of a distribution over I/O pairs, $\rho\in\prob{\mathcal{X}\times\mathcal{Y}}$, that generates the instances of the problem we mean to solve, we aim to minimize the average loss value
\begin{equation}
\label{svm:eq:obj}
{\cal R}(f) = \E_{(X, Y)\sim\rho}\bracket{\ell(f(X), Y)}.
\end{equation}
In practice, this ``risk'' ${\cal R}$ can be evaluated approximately with samples ${\cal D}_n = (X_i, Y_i)_{i\leq n}$, collected by the machine learning scientist and assumed to have been drawn independently accordingly to $\rho$.
We shall focus on the binary classification problem where $\mathcal{Y} = \brace{-1, 1}$, and $\ell$ is the zero-one loss $\ell(y, z) = \ind{y\neq z}$.
In this setting, the risk ${\cal R}(f)$ captures the probability of mistakes of a classifier $f$, and its minimizer is characterized by
\begin{equation}
f^* = \argmin_{f:\mathcal{X}\to\mathcal{Y}} {\cal R}(f) = \sign \eta,
\qquad\text{where}\qquad
\eta(x) = \E\bracket{Y\,\middle\vert\, X=x}.
\end{equation}
Ideally, leveraging the dataset ${\cal D}_n$, we would like to find a mapping $f_{{\cal D}_n}:\mathcal{X}\to\mathcal{Y}$ that is close to be optimal, in the sense that the excess of risk
\(
{\cal E}(f_{{\cal D}_n}) = {\cal R}(f_{{\cal D}_n}) - {\cal R}(f^*)
\)
is as small as it could be.
Since this quantity is actually random, inheriting from the randomness of the samples, we will focus on controlling its average.\footnote{Note that one could also be interested in controlling its tail, but this is out-of-score of the current literature on ``super fast'' rates.}
In particular, we will show that, when our model is well-specified, as the number of samples grows, this average decays actually much faster than what usual statistical learning theory suggests.
We give a brief historical review of related literature before précising our contributions.
\subsection{Statistical learning theory}
The classical approach to minimize \eqref{svm:eq:obj} without the knowledge of $\rho$ but with the sole access to samples ${\cal D}_n \sim \rho^{\otimes n}$ is to restrict the search over functions in a class ${\cal F}\subset\mathcal{Y}^\mathcal{X}$, and look for an empirical risk minimizer
\begin{equation}
\label{svm:eq:erm}
f^*_{{\cal D}_n} \in \argmin_{f\in{\cal F}} {\cal R}_{{\cal D}_n},
\qquad\text{where}\qquad
{\cal R}_{{\cal D}_n}(f) = \frac{1}{n} \sum_{i=1}^n \ell(f(X_i), Y_i).
\end{equation}
If we denote by $f^*_{\cal F}$ the minimizer of ${\cal R}$ in ${\cal F}$, using the fact that ${\cal R}_{{\cal D}_n}(f^*_{\cal F}) \geq {\cal R}_{{\cal D}_n}(f^*_{{\cal D}_n})$, the excess of risk can be bounded as
\begin{equation}
\label{svm:eq:app_est}
{\cal R}(f_{{\cal D}_n}^*) - {\cal R}(f^*)\leq
\underbrace{2\sup_{f\in{\cal F}} \abs{{\cal R}(f) - {\cal R}_{{\cal D}_n}(f)}}_{\text{estimation error}}
+ \underbrace{{\cal R}(f^*_{\cal F}) - {\cal R}(f^*)}_{\text{approximation error}}.
\end{equation}
This bound can be seen as highly suboptimal because it bounds the deviation of a random function with the worst deviation in the function class.
However, for any class~${\cal F}$, there exists an ``adversarial'' distribution $\rho$ for which convergence rates (of the excess of risk toward zero as a function of the number of samples $n\in\mathbb{N}$) derived through this bound can not be improved beside lowering some multiplicative constants \citep{Vapnik1995}.
On the one hand, the estimation error can be controlled with general tools to bound the supremum of a random process \citep[\emph{e.g.},][]{Dudley1967}, and will decrease with a decrease in the size of the class ${\cal F}$.
On the other hand, the approximation error depends on assumptions of the problem, and the bigger the size of the class ${\cal F}$, the less restrictive it will be to assume that $f^*$ is not too different from $f_{\cal F}^*$.
Hence, there is a clear trade-off between controlling both errors, which should be balanced in order to optimize a bound on the full excess of risk.
\subsection{Surrogate methods}
In practice, due to the combinatorial nature of discrete-valued functions, finding the empirical risk minimizer \eqref{svm:eq:erm} is often an intractable problem \citep[\emph{e.g.},][]{Hoffgen1992,Arora1997}.
Therefore, people have approached the original problem with other perspectives.
A straightforward approach is given by \emph{plug-in classifiers}, \emph{i.e.}, classifiers of the form $\sign \hat\eta$, for $\hat\eta$ some estimator of $\eta$.
For example, such an estimator can be constructed as
\(
\hat\eta(x) = \sum_{i=1}^n \alpha_i(x) Y_i,
\)
for $\alpha_i(x)$ some weights that specify how much the observation $Y_i$ made at the point $X_i$ should diffuse to the point $x$ \citep[see][for an example]{Friedman1994}.
Another popular approach to solve classification problems is provided by \emph{support vector machines} (SVM), which were introduced from geometric considerations to maximize the margin between the classes $\brace{x\in\mathcal{X}\,\middle\vert\, f^*(x) = y}$ for $y \in\brace{-1, 1}$ \citep{Cortes1995}.
These two approaches can be conjointly understood as introducing a surrogate loss $L:\mathbb{R}\times\mathcal{Y}\to\mathbb{R}$ and looking for a continuous-valued function $g:\mathcal{X}\to\mathbb{R}$ that solves the surrogate problem
\begin{equation}
\label{svm:eq:decoding}
f = \sign g, \qquad
g^*\in \argmin_{g:\mathcal{X}\to\mathbb{R}} {\cal R}_S(g),
\qquad\text{where}\qquad {\cal R}_S(g) = \E_{(X,Y)\sim\rho}\bracket{L(g(X), Y)},
\end{equation}
where the notation $S$ stands for ``surrogate''.
To an estimate $g:\mathcal{X}\to\mathbb{R}$ of $g^*$ we associate an estimate $f:\mathcal{X}\to\mathcal{Y}$ of $f^*$ through the decoding step $f=\sign g$.
In particular, using the variational characterization of the mean, $\eta$ can be estimated through $L(z, y) = \abs{z - y}^2$.
Regarding SVM, they are related to the \emph{hinge loss} \cite[see, \emph{e.g.},][]{Steinwart2008}
\begin{equation}
\label{svm:eq:Hinge}
L(z, y) = \max\paren{0, 1 - zy},
\end{equation}
Surrogate methods benefit from their relative easiness to optimize and the quality of their practical results.
Arguably, they define the current state of the art in classification, softmax regression being particularly popular to train neural networks on classification tasks.
Surrogate methods were studied in depth by \cite{Bartlett2006}, who proposed a generic framework to relate the excess of the original risk to the excess of surrogate risk through an inequality of the type
\begin{equation}
\label{svm:eq:calibration}
{\cal R}(f) - {\cal R}(f^*) \leq \psi\paren{{\cal R}_S(g) - {\cal R}_S(g^*)},
\end{equation}
where $f = \sign g$
and $\psi$ is a concave function, uniquely defined from $L$ and verifying $\psi(0) = 0$.
The use of a concave function is motivated by Jensen inequality, allowing to integrate an inequality derived pointwise (conditionally on an input $x$).
\subsection{Exponential convergence rates}
On the one hand, calibration inequalities \eqref{svm:eq:calibration} are appealing, as they allow casting directly rates derived on the surrogate problem to rates on the original problem.
On the other hand, because~$\psi$ has to be concave, rates in $O(n^{-r})$ on the surrogate problem can not be cast as better rates on the original problem, corresponding to the optimal inequalities where $\psi(x) = cx$ for some $c > 0$,
Yet, one can find cases where the sign of $\eta$ can be estimated much faster than $\eta$ itself, even when this sign is estimated with surrogate methods.
In particular, \cite{Mammen1999} \citep[see also][]{Massart2006} introduced the following condition.
\begin{assumption}[Hard margin condition]
\label{svm:ass:margin}
The binary classification problem defined through the distribution $\rho$ is said to verify the (Tsybakov) hard margin condition if the conditional mean $\eta$ is bounded away from zero, i.e.,
\begin{equation}
\label{svm:eq:margin}
\exists\,\eta_0 > 0; \qquad \abs{\eta(X)} > \eta_0\qquad\textit{a.s.},
\end{equation}
where the notation a.s. stands for almost surely.
Equivalently, $\abs{\eta}^{-1} \in L^\infty(\rho_\mathcal{X})$.
\end{assumption}
Indeed, as shown in Appendix \ref{svm:app:proofs}, under Assumption \ref{svm:ass:margin}, leveraging sign equality, we get for any estimate $g_{{\cal D}_n}:\mathcal{X}\to\mathbb{R}$ computed from the dataset ${\cal D}_n$,
\[
\E_{{\cal D}_n}[{\cal R}(\sign g_{{\cal D}_n})] - {\cal R}(f^*)
\leq \Pbb_{{\cal D}_n}\paren{\norm{g_{{\cal D}_n} - \eta}_{L^\infty} > \eta_0}.
\]
As a consequence, an exponential concentration inequality on the $L^\infty$ distance between $g_{{\cal D}_n}$ and $\eta$ directly translates to exponential convergence rates on the average excess of risk.
In particular, estimation methods for~$\eta$ based on H\"older classes of functions, such as local polynomials, are known to be well-behaved with respect to the $L^\infty$ norm \citep[see, \emph{e.g.}, the construction of covering number by][]{Kolmogorov1959}.
This was leveraged by \cite{Audibert2007} in a seminal paper that shows how better rates can be achieved on the classification problem under Assumption \ref{svm:ass:margin} and a variety of weaker conditions (described later in Assumption \ref{svm:ass:weak_margin}).
Surprisingly, such an approach has remained somehow less popular than approaches based on calibration inequalities, and we are missing a framework to fully apprehend fast rates phenomena.
Some results were achieved by \cite{Koltchinskii2005}.
Recently, \cite{Cabannes2021b} showed that this result generalizes to any discrete output learning problem, and that approaches that naturally lead to concentration in $L^2$ could be turned into fast rates based on interpolation inequalities that relate the $L^2$ norm with the $L^\infty$ one (notably reusing the work of \cite{Fischer2020} on interpolation spaces).
Exploiting the work of \cite{MarteauFerey2019}, this can be generalized to any self-concordant loss (using self-concordance to reduce the problem to a least-squares problem); and, through the work of \cite{Lin2020}, to any spectral filtering technique (beyond Tikhonov regularization), such as stochastic gradient descent, which was actually shown earlier for binary classification by \cite{PillaudVivien2018b} and \cite{Nitanda2019}.
In the same stream of research, \cite{Vigogna2022} proposed a general framework to study exponential rates for smooth losses in multiclass classification beyond least-squares.
\subsection{Contribution}
The proofs of exponential convergence in the works quoted above are all based on the basic mechanism outlined in \cite{Audibert2007}.
Unfortunately, such a mechanism does not easily extend to the hinge loss.
Does this mean that support vector machines do not exhibit super fast rates, and thus they are inferior to other surrogate methods?
The practice seems to answer negatively.
In this paper, we give a firm theoretical answer to this question.
In particular, we show not only that support vector machines do achieve exponential rates, but also that they can do so even without assuming the hard margin condition.
Our main contribution is to introduce a general framework to prove exponential convergence rates, and show how this framework can be applied to the hinge loss while only considering classical assumptions.
\paragraph{Outline.}
Our general strategy is illustrated on Figure \ref{svm:fig:discrete_analysis} and consists in first finding a relation
\[
{\cal R}_S(g_{\theta}) - {\cal R}_S(g_{\theta^*}) \geq \norm{\theta - \theta^*},
\]
for some natural parameter $\theta$ in a Banach space $\Theta$ parametrizing a class of functions $g_\theta\in{\cal F}$, and then show that $\sign g_\theta = \sign g_{\theta^*}$ when $\norm{\theta - \theta^*}$ is small enough, that is,
\[
\exists\, \epsilon > 0; \qquad
\norm{\theta - \theta^*} \leq \epsilon \quad\Rightarrow\quad
\sign g_\theta = \sign g_{\theta^*}.
\]
Assuming that $\sign g_{\theta^*} = f^*$, we deduce that, as shown in Appendix \ref{svm:app:proofs},
\[
\E_{{\cal D}_n}[{\cal R}(\sign g_{\theta_n})] - {\cal R}(f^*)
\leq \Pbb_{{\cal D}_n}\paren{{\cal R}_S(g_{\theta_n}) - {\cal R}_S(g_{\theta^*}) \geq \epsilon},
\]
where $g_{\theta_n}$ is an estimate of $g_\theta$ based on the samples ${\cal D}_n$.
Finally, we conclude with an exponential concentration inequality that controls the deviation of the excess of risk based on classical statistical learning theory.
\begin{figure}
\centering
\input{svm/images/fast_rates}
\caption{Our convergence analysis consists in relating natural concentration given by surrogate methods to the original excess of risk without passing by the surrogate excess of risk.
As the drawing shows, concentration in parameter space $\Theta$ can be cast as deviation on the original excess of risk.
Yet, such a casting relation depends on the geometry of this picture, which itself depends on what surrogate is used, what is the function to learn, how a regularized estimator approached it, and how our empirical estimate concentrates around the regularized estimator.
Note that this figure illustrates an abstract mechanism that generalizes the simpler mechanism we use to derive exponential convergence rates.
}
\label{svm:fig:discrete_analysis}
\end{figure}
\section{Exponential convergence of SVM}
This section is devoted to the proof of exponential convergence rates for the hinge loss.
We shall fix the notation ${\cal R}_S$ as the surrogate risk associated with \eqref{svm:eq:Hinge}.
All the proofs are collected in Appendix \ref{svm:app:proofs}.
\subsection{Refined calibration for the hinge loss}
We start by introducing the classical weak margin condition \citep{Mammen1999}.
\begin{assumption}[Weak margin condition]
\label{svm:ass:weak_margin}
The binary classification problem defined through the distribution $\rho$ is said to verify the (Tsybakov) $p$-margin condition, with $p \in (0, \infty)$, if there exists a constant $c > 0$ such that
\begin{equation}
\Pbb_{\rho_\mathcal{X}}(0 < \abs{\eta(X)} < t) \leq c t^p,
\end{equation}
where the notation $\rho_\mathcal{X}$ denotes the marginal of $\rho$ over $\mathcal{X}$.
\end{assumption}
Assumption \ref{svm:ass:weak_margin} is equivalent to asking for the inverse of the conditional mean~$\abs{\eta}^{-1}$ (with the convention $0^{-1} = 0$) to belong to the Lorentz space $L^{p,\infty}(\rho_\mathcal{X})$ (also known as weak-$L^p$ space), which is the Banach space endowed with the norm (quasi-norm and quasi-Banach if $p < 1$)
\begin{equation}
\norm{f}_{p, \infty} = \sup_{t > 0} t\Pbb_{\rho_\mathcal{X}}(f(X) > t)^{\frac{1}{p}},
\end{equation}
where the $\rho_\mathcal{X}$ denotes the marginal of $\rho$ with respect to $\mathcal{X}$.
This definition can be extended to the case $p=\infty$ by setting $ L^{p,\infty}(\rho_\mathcal{X}) = L^\infty(\rho_\mathcal{X}) $, which characterizes the hard margin condition in Assumption \ref{svm:ass:margin}.
We will also use $\norm{\cdot}_p$, for $p\in[1,\infty]$, to denote the $L^p$-norm on $\mathcal{X}$ endowed with $\rho_\mathcal{X}$.
We now relate the excess of risk on the hinge loss to the deviation in these spaces.
\begin{lemma}[Weak-$L^q$ concentration due to the hinge loss]
\label{svm:lem:l1}
For any functions $g_1, g_2:\mathcal{X}\to[-1,1]$,
\begin{equation} \label{svm:lem1a}
{\cal R}_S(g_2) - {\cal R}_S(g_1) = \E_{\rho_\mathcal{X}}[ -\eta(X) (g_2(X) - g_1(X))].
\end{equation}
In particular, under Assumption \ref{svm:ass:margin}, for any $g:\mathcal{X}\to\mathbb{R}$,
\begin{equation} \label{svm:lem1b}
{\cal R}_S(g) - {\cal R}_S(g^*) \geq \norm{\abs{\eta}^{-1}}_{\infty}^{-1} \norm{\pi(g) - g^*}_{1},
\end{equation}
where $g^* = \sign \eta$ is a minimizer of ${\cal R}_S$ and $\pi$ is the projection of $\mathbb{R}$ on $[-1, 1]$, defined as mapping $t\in\mathbb{R}$ to $\pi(t) = \sign(t) \min\{|t|,1\}$.
Similarly, under Assumption \ref{svm:ass:weak_margin}, with $q = \sfrac{p}{p+1}$,
\begin{equation} \label{svm:lem1c}
{\cal R}_S(g) - {\cal R}_S(g^*)
\geq 2^{-1} \norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} \norm{\pi(g) - g^*}_{q,\infty}.
\end{equation}
\end{lemma}
Lemma \ref{svm:lem:l1} shows that we can set the minimizer $g^* = f^* \in \brace{-1, 1}^\mathcal{X}$.
This is a useful fact as it implies that the excess of the original risk is zero as soon as $\norm{g - g^*}_\infty < 1$.
In essence, the only piece missing in order to prove fast convergence rates is an interpolation inequality between $L^{q,\infty}$ and $L^\infty$.
In the following, we will leverage Lemma \ref{svm:lem:l1} more subtly by considering a class of functions ${\cal G}$ and assumptions on the distribution $\rho_\mathcal{X}$ such that, if an estimate $g\in{\cal G}$ has not the same sign almost everywhere as the estimand $g^*$, then $\norm{g-g^*}_{q, \infty}$ is bounded away from zero.
By contraposition, if $g\in{\cal G}$ presents a small excess of surrogate risk, then $\sign g = \sign g^*$.
When $\mathcal{X}$ is a metric space, one way to proceed is to assume that $g$ is Lipschitz-continuous, together with some minimal mass assumptions.
Let us begin with the minimal mass assumption.
We first need the following definition.
\begin{definition}[Well-behaved sets]
A set $U\subset\mathcal{X}$ is said to be well-behaved with respect to $\rho$ if there exist constants $c, r, d > 0$ such that, for any $x\in U$,
\begin{equation}
\label{svm:eq:mass}
\forall\,\epsilon \in [0, r]; \qquad \rho_\mathcal{X}(U\cap{\cal B}(x, \epsilon)) \geq c\epsilon^d,
\end{equation}
and ${\cal B}(x, \epsilon)$ the ball in $\mathcal{X}$ of center $x$ and radius $\epsilon$.
\end{definition}
The following examples show that the coefficient~$d$ that appears in \eqref{svm:eq:mass} results from the dimension of the ambient space, the regularity of singularities of the border of the set, and the decay of the density when approaching the frontier of the set.
\begin{example}
The set $[0, 1]^p$ is well-behaved with coefficients $r=1$, $d=p$ and $c=2^{-d}\operatorname{vol}(\mathbb{S}^{d-1})$ with respect to the Lebesgue measure in $\mathbb{R}^d$.
\end{example}
\begin{example}
The set $\brace{(x, y) \in \mathbb{R}^2\,\middle\vert\, x\in[0,1], y\in[0, x^{n-1}]}$ is well-behaved with coefficient $r=1$, $d=n$ and $c=n^{-1}$ with respect to the Lebesgue measure.
Reciprocally, the set $[0,1]$ is well-behaved with coefficient $r=1$, $d=n$ and $c=n^{-1}$ with respect to the measure whose density equals $p(x) = x^{n-1}$.
\end{example}
\begin{assumption}[Minimal mass assumption]
\label{svm:ass:mass}
The decision regions $\mathcal{X}_y = \brace{x\in\supp(\rho_\mathcal{X})\,\middle\vert\, f^*(x) = y}$ for $y \in \brace{-1, 1}$ are well-behaved.
\end{assumption}
Assumption \ref{svm:ass:mass} is a weakening of an assumption that is commonly found in the statistical learning literature.
More precisely, it is often assumed that $\rho$ is absolutely continuous according to the Lebesgue measure $\lambda$ on $\mathcal{X}$ (assumed to be a Euclidean space), that its density is bounded away from zero on its support, and that its support has smooth boundary, so that $\lambda(\supp\rho_\mathcal{X}\cap {\cal B}(x, \epsilon)) > c'\lambda({\cal B}(x, \epsilon))$ \citep[see the strong density assumption in][]{Audibert2007}.
The minimal mass requirement allows relating misclassification events to $L^{q,\infty}$ deviation.
\begin{lemma} \label{svm:lem2}
Under Assumption \ref{svm:ass:mass}, there exists a constant $c_0$ such that if $g$ is $G$-Lipschitz-continuous for $G > r^{-1}$, for any $q \in (0, 1]$
\begin{equation}
\exists\, x\in\supp\rho_\mathcal{X}; \abs{g(x) - g^*(x)} \geq 1 \quad\Rightarrow\quad \norm{g - g^*}_{q, \infty} \geq c_0 G^{-\frac{d}{q}}.
\end{equation}
\end{lemma}
Putting together Lemmas~\ref{svm:lem:l1} and~\ref{svm:lem2}, we obtain the following refined calibration.
\begin{proposition}
\label{svm:prop:mid}
Under Assumptions \ref{svm:ass:weak_margin} and \ref{svm:ass:mass}, if $g$ is $G$-Lipschitz-continuous with $G>r^{-1}$, we have
\begin{equation}
\label{svm:eq:mid}
{\cal R}_S(g) - {\cal R}_S(g^*) \leq 2^{-1}\norm{\abs{\eta}^{-1}}_{p,\infty}^{-1} c_0 G^{-\frac{d(p+1)}{p}} \qquad\Rightarrow\qquad {\cal R}(\sign g) = {\cal R}(f^*).
\end{equation}
\end{proposition}
\subsection{Trade-off between estimation and approximation errors}
We are now left with the research of $g_{{\cal D}_n}$ inside a class of Lipschitz-continuous functions such that ${\cal R}_S(g_{{\cal D}_n}) - {\cal R}_S(g^*)$ is sub-Gaussian (its randomness being inherited from the dataset ${\cal D}_n$ from which $g_{{\cal D}_n}$ is built).
To do so, let us consider a linear class of functions
\begin{equation}
\label{svm:eq:f_k}
{\cal G}_{M, \sigma} = \brace{x\mapsto\scap{\theta}{\phi(\sigma^{-1} x)}\,\middle\vert\, \theta \in {\cal H}, \norm{\theta}_{\cal H} \leq M} ,
\end{equation}
where ${\cal H}$ is a separable Hilbert space, $\phi:\mathcal{X}\to{\cal H}$ is a $G_\phi$-Lipschitz-continuous mapping, and $\sigma > 0$ is a scaling (or bandwidth) parameter.
Such a class of functions can be entirely described from the kernel $k(x, x') = \scap{\phi(x)}{\phi(x')}$ \citep[see][for a primer on kernel methods]{Scholkopf2001}.
An example for ${\cal G}$ is given by the Gaussian kernel, a.k.a. radial basis function, $k(x, x') = \exp(-\norm{x - x'}^2)$.
Using Cauchy-Schwarz, it is easy to show that any function in ${\cal G}_{M,\sigma}$ is $MG_\phi \sigma^{-1}$-Lipschitz-continuous.
In order to find a function $g_{{\cal D}_n}$ that is likely to minimize ${\cal R}_S$ without accessing the distribution $\rho$, but only i.i.d. samples ${\cal D}_n = (X_i, Y_i)_{i\leq n} \sim \rho^{\otimes n}$, it is classical to consider the empirical risk minimizer
\begin{equation}
\label{svm:eq:sur_erm}
g_{{\cal D}_n} \in \argmin_{g\in{\cal G}_{M, \sigma}} \frac{1}{n} \sum_{i=1}^n L(g(X_i), Y_i).
\end{equation}
This problem is convex with respect to $\theta$ parametrizing $g \in {\cal G}_{M,\sigma}$, and is easily optimized with duality.
We refer the curious reader to the extensive literature on SVM \citep[see][for books on the matter]{Cristianini2000,Scholkopf2001,Steinwart2008}.
In order to show that ${\cal R}_S(g_{{\cal D}_n})$ is close to ${\cal R}_S(g^*)$, one can apply classical results from statistical learning theory, and in particular \eqref{svm:eq:app_est}.
The estimation error can be bounded using the extensive literature on Rademacher complexity for linear classes of functions on Lipschitz-continuous losses \citep{Bartlett2002}.
To bound the approximation error, one needs to make additional assumptions on the problem.
We refer to \cite{Steinwart2007,Blaschzyk2018} for advanced considerations on the matter.
In view of our calibration result \eqref{svm:eq:mid}, the following additional assumption suffices to prove exponential convergence of SVM.
\begin{assumption}[Source condition]
\label{svm:ass:source}
The classification problem verifies the $p$-margin condition \ref{svm:ass:weak_margin}, and there exist $M, \sigma$ and a function $g\in{\cal G}_{M, \sigma}$ such that ${\cal R}_S(g) - {\cal R}_S(g^*) \leq 4^{-1}\|\abs{\eta}^{-1}\|_{p, \infty}^{-1} c_0 M^{-r} G_\phi^{-r} \sigma^r$ with $r = d(p+1) / p$.
\end{assumption}
It should be noted that Assumption \ref{svm:ass:source}, together with Assumption \ref{svm:ass:mass}, implies that the decision frontier $\overline{\mathcal{X}_{-1}} \cap \overline{\mathcal{X}_1}$ (the bar notation corresponding to space closure), inherits from the regularity of $g$, since it is included in the set $\brace{x\in\mathcal{X}\,\middle\vert\, g(x) = 0}$.
In particular, if ${\cal G}_{M,\sigma}$ is included in ${\cal C}^m$, this frontier would be in ${\cal C}^m$.
Hence, for Assumptions \ref{svm:ass:source} and \ref{svm:ass:mass} to hold, the boundary frontier should match the regularity implicitly defined by ${\cal G}_{M, \sigma}$.
We are finally ready to state our main result, establishing exponential convergence rates for SVM.
\begin{theorem}[Exponential convergence rates for SVM]
\label{svm:thm}
Under Assumptions \ref{svm:ass:weak_margin}, \ref{svm:ass:mass} and \ref{svm:ass:source}, there exists a constant $c > 0$ such that the empirical minimizer $g_{{\cal D}_n}$ defined by \eqref{svm:eq:sur_erm} verifies
\begin{equation}
\E_{{\cal D}_n}{\cal R}(\sign g_{{\cal D}_n}) - {\cal R}(f^*)
\leq \exp(-cn).
\end{equation}
\end{theorem}
\begin{figure}[t]
\centering
\includegraphics{svm/images/fig2}
\includegraphics{svm/images/fig1}
\caption{
SVM generalization error as a function of the number of samples (left) for a problem where $X$ is uniform on $[-1, .-1] \cup [.1, 1]$ and $\eta(x) = \sign(x) \abs{x}^p$ (right). We observe exponential convergence rates.
}
\label{svm:fig:fig1}
\end{figure}
\subsection{Relaxing assumptions}
Exponential convergence rates rely on strong assumptions in order to set the approximation error to zero.
In particular, it is customary to assume that the surrogate function to learn lies in the model we have chosen, that is, in our notation, $g^* \in {\cal G}_{M, \sigma}$.
In our case this would be a strong assumption, since $g^*$ is piecewise linear while ${\cal G}_{M, \sigma}$ is a smooth space of functions.
It turns out that the assumption $g^*\in{\cal G}_{M, \sigma}$ is not necessary, and what we actually need is a sufficiently small risk.
How small is enough is quantified by the statement of Proposition \ref{svm:prop:mid}.
From a qualitative point of view, the requirement of Assumption \ref{svm:ass:source} is natural (surrogate risk must be smaller than $\epsilon$), with the technical part being just a quantification of the needed behavior (bound on such $\epsilon$).
To deepen the study of the approximation error, one could leverage the following geometrical characterization of the risk of misclassification.
For $f:\mathcal{X}\to\brace{-1, 1}$, we have
\begin{equation}
\label{svm:eq:pre_source}
{\cal R}(f) - {\cal R}(f^*) = \E[\abs{\eta(X)} \ind{f(x) \neq f^*(X)}] \leq
\Pbb(f(X)\neq f^*(X)) = \rho_\mathcal{X}\paren{f^{-1}(\{1\})\,\triangle\, \mathcal{X}_1},
\end{equation}
where $\triangle$ denotes the symmetric difference of sets, {\em i.e.} $A \triangle B = (A \cup B) \setminus (A \cap B)$.
In particular, if the function class ${\cal G}_{M,\sigma}$ is rich enough, and the classes $\mathcal{X}_1$ and $\mathcal{X}_{-1}$ are separated, in the sense that the distance between any two points in each set is bounded away from zero,\footnote{This property is sometimes referred to as the \emph{cluster assumption} \citep{Rigollet2007} under which even weakly supervised learning techniques may exhibit exponential convergence rates \citep[\emph{e.g.}][]{Cabannes2021}. In terms of practical applications, this assumption says that no one can continuously modify an input to go from a region of the space linked with one class to a region linked with another class without going through inputs that will never exist.
This is typically true for well-curated image datasets such as CIFAR10: one can not continuously transform an image of a truck into an image of a horse without going through images that will never appear in the CIFAR10 dataset \citep{Krizhevsky2009}.} then Assumption \ref{svm:ass:source} holds, and the minimizer $g_{{\cal G}_{M,\sigma}}$ of the surrogate risk in ${\cal G}_{M,\sigma}$ verifies
\begin{equation}
\label{svm:eq:source}
\rho_\mathcal{X}\paren{(\sign g_{{\cal G}_{M,\sigma}})^{-1}(\{1\})\,\triangle\, \mathcal{X}_1} \leq \psi(M, \sigma),
\end{equation}
for $\psi$ a function that vanishes for sufficiently large $M$ and small $\sigma$.
On the one hand, one could control the approximation error by assuming or deriving inequalities akin to \eqref{svm:eq:pre_source} and \eqref{svm:eq:source}, with different profiles of $\psi$.
We conjecture that this can be done by assuming margin conditions that are well adapted to the geometric nature of SVM, such as the one proposed by \cite{Steinwart2007} \citep[see also][]{Gentile1999,Cristianini2000}.
On the other hand, the estimation error can be controlled by extending the ideas presented in this paper to study the worst value of the estimation error ${\cal R}(\sign g_{{\cal D}_n}) - {\cal R}(\sign g_{{\cal G}_{M,\sigma}})$ under the knowledge of ${\cal R}_S(g_{{\cal D}_n}) - {\cal R}_S(g_{{\cal G}_{M,\sigma}})$.
Fitting $M$ and $\sigma$ to trade estimation and approximation error, such derivations would open the way to fast polynomial rates under less restrictive assumptions.
\begin{figure}[t]
\centering
\includegraphics{svm/images/fig4}
\includegraphics{svm/images/fig5}
\includegraphics{svm/images/fig6}
\includegraphics{svm/images/fig4_bis}
\includegraphics{svm/images/fig5_bis}
\includegraphics{svm/images/fig6_bis}
\caption{
Study of the level lines of $g_{\lambda, \sigma}$ when $\eta^{-1}(0) \in {\cal C}^\infty$ (top) and $\eta^{-1}(0)\in{\cal C}^0\setminus {\cal C}^1$ (bottom).
The function $g^*$ takes values $-1$ below the optimal decision frontier plotted in red and $+1$ above, independently of the noise.
We observe that the bias error ${\cal R}(\sign g_{\lambda,\sigma}) - {\cal R}(f^*)$, which is bounded by the volume between the level lines $\brace{x\in\mathcal{X}\,\middle\vert\, g_{\lambda,\sigma}(x) = 0}$ and $\brace{x\in\mathcal{X}\,\middle\vert\, \eta(x) = 0}$ (plotted in red), depends on both the regularity of the latter, and on the noise level.
Here, $\sigma$ is taken to be of the order of 15\% of the diameter of the domain, which explains the regularity of the observed level lines.
The noiseless cases on the right correspond to the situations where $\E[Y\vert X] = \sign \eta(X)$ for $\eta$ plotted on the left.
}
\label{svm:fig:fig2}
\end{figure}
\section{Numerical analysis}
In this section, we provide experiments to illustrate and validate our theoretical findings.
In order to be inline with the current practice of machine learning, instead of considering the hard constraint $\norm{\theta} \leq M$ when minimizing a risk functional, we add a penalty $\lambda \norm{\theta}^2$ to the risk to be minimized.
Going from a constrained to a penalized framework does not change the nature of the statistical analysis, and one might loosely think of $\lambda$ as $1 / M$ \citep[see, for example,][]{Bach2023}.
All experiments are made with the Gaussian kernel.
Precise details of the different settings are provided in Appendix \ref{svm:app:experiments}.
First, we observe that the regime described in this paper kicks in when the error is already pretty small.
On many real-world problems, we do not expect the generalization error as a function of the number of data used for training to exhibit a clear exponential behavior until an unusually big number of samples is used.
This fact is illustrated on Figure~\ref{svm:fig:fig1}, where for hard problems, the exponential behaviors still do not kick in after a thousand of samples.
Second, this paper shows that, in order to get exponential convergence rates for SVM, one needs the minimizer $g_{M,\sigma}$ of the surrogate risk over the selected class of functions to be a perfect classifier, {\em i.e.} its sign equals the sign of $g^*$.
While this is not constraining under the cluster assumption, we inspect divergences from this condition on Figure~\ref{svm:fig:fig2}.
We observe that, even if $g^*$ does not depend on the noise, $g_{M,\sigma}$ does.
We also observe that the regularity of the decision boundary $\brace{x\in \mathcal{X}\,\middle\vert\, \eta(x) = 0}$ should match the regularity defined implicitly by the kernel $k$ and the scale parameter $\sigma$.
\begin{figure}[t]
\centering
\includegraphics{svm/images/fig7}
\includegraphics{svm/images/fig8}
\includegraphics{svm/images/fig9}
\caption{
Comparison of the regularized risk ({\em i.e.} ${\cal R}_S(g_\theta) + \lambda\norm{\theta}^2$) minimizer for the hinge loss surrogate (middle) and the least-squares surrogate (right), when $\eta$ is not regular (left).
In this setting, the hinge loss is minimized for $g = \sign(\eta)$, which can be chosen regular, while the least-squares loss is minimized for $g = \eta$, which can not be chosen regular.
The reconstruction is made with $\sigma$ about 3\% of the domain diameter, and $\lambda$ relatively small.
We assume no density in the middle of the domain, explaining the absence of definition of $\eta$ and the dashed lines of the right figures.
The oscillation on the later figures is related to the Gibbs phenomenon \citep{Wilbraham1848}.
This phenomenon prevents the regularized least-squares solution from being a perfect classifier.
}
\label{svm:fig:fig3}
\end{figure}
Experimental comparisons of different classification approaches have been done by many people, and our goal is not to showcase the superiority of the SVM over least-squares, which might be considered as general wisdom that led to the golden age of SVM in the pre-deep-learning area \citep[see][for example]{Joachims1998}.
In comparison with previous works based on calibration inequalities \citep{Rosasco2004,Steinwart2007b}, our analysis proves the robustness of SVM to noise far away from the decision boundary, in the sense that one does not need $\eta$ to be bounded away from zero.
This is a distinctive aspect of SVM compared to smooth surrogate methods \citep{Nowak2019b}, such as softmax regression, that implicitly estimate conditional probabilities and whose performance depends on the regularity of $\eta$.
We illustrate this fact graphically on Figure~\ref{svm:fig:fig3}.
\section{Limitations}
\paragraph{Are surrogate methods only a proxy for classification?}
From a theoretical perspective, if we are only interested in the optimal mapping $f:\mathcal{X}\to\mathcal{Y}$, learning surrogate quantities can be seen as a waste of resources.
In essence, this waste of resources is similar to the one occurring when we learn the full probability function $(p(y))_{y\in\mathcal{Y}}$ for some probability distribution $p$ on $\mathcal{Y}$, while we only care about its mode.
Yet, in practice, what we call a ``surrogate'' problem might actually be a problem of prime interest when we do not only want to predict $f^*(x)$, but we would also like to know how much we can confidently discard other potential outputs for an input $x$.
Furthermore, assuming that a problem is exactly defined through an ``original'' loss that defines a clear and unique measure of error can be questioned when some practitioners evaluate methods with several metrics of performance \cite[\emph{e.g.}][]{Chowdhery2022}.
\paragraph{Do PAC-bounds provide confidence levels?}
Since the parameters in Assumptions \ref{svm:ass:weak_margin} and \ref{svm:ass:mass} are hard to estimate in practice, it would be difficult to directly plug our bounds into a practical problem to derive confidence levels on how much error one might expect when deploying a model in production.
Less ambitiously, we see theorems akin to Theorem \ref{svm:thm} as providing theoretical indications that a learning method or a set of hyperparameters is sound.
This is a generic downfall of probably approximately correct (PAC) generalization bounds \citep{Valient2013}, which might explain why practitioners often prefer to derive error indications from test samples \citep[see, \emph{e.g.},][]{Geron2017}.
Along this line, research on conformal prediction provides interesting considerations to obtain useful confidence information from test samples \citep{Vovk2008}.
Finally, all these statistical methods to get confidence intervals assume representative (if not i.i.d.) data, an assumption sometimes hard to meet in practice, which is a problem that has found echoes in the civil society \citep[\emph{e.g.}][]{Benjamin2019}.
\section{Conclusion}
In this work, we were keen to illustrate a simple mechanism to get exponential convergence rates for a loss that is quite popular and whose understanding can not be easily reduced to existing work.
In particular, we show that the hard margin condition is not crucial in order to derive exponential convergence rates for the SVM.
This provides a crucial step to better understand convergence rates on classification problems.
An extension to generic discrete output problems could be made by considering polyhedral losses, and deriving variants of Lemma~\ref{svm:lem:l1} \citep[see][for calibration inequalities for such losses]{Frongillo2021}.
An important follow-up would be to provide a more global picture of fast polynomial rates for SVM under relaxations of Assumptions \ref{svm:ass:mass} and \ref{svm:ass:source}.
Finally, \cite{Chizat2020} have made a link between two-layer wide neural networks in the interpolation regime (which implies Assumption \ref{svm:ass:margin} with $\eta_0 = 1$) and max-margin classifiers over specific linear classes of functions.
As a consequence, we could directly plug in our analysis to prove exponential convergence rates for those small neural networks in this noiseless setting.
Studying rates, constants and hyperparameter tuning in this setting would be of particular interest if it was to provide practical guidelines to deep learning practitioners in the spirit of \cite{Yang2021}.
\chapter{Exponential Convergence Rates for SVM}
\label{chap:svm}
The following is a reproduction of \cite{Cabannes2022b}.
\input{svm/abstract}
\input{svm/core}
\begin{subappendices}
\chapter*{Appendix}
\addcontentsline{toc}{chapter}{Appendix}
\input{svm/appendix}
\end{subappendices}
|
\section{Introduction}\label{sect:intro}
Accurate radiative-transfer (RT) calculations remain a key challenge
in the study of astronomical transients such as supernovae. While
advances in computational capabilities and theoretical understanding
have allowed great progress in the ability to simulate radiation
transport, the large number of physical processes involved, in
particular opacity from thousands of atomic transitions with a mixed
absorptive/scattering character, prohibit comprehensive 3D
calculations based on first principles. Several physical
approximations of different forms, in particular different treatments
of the significant deviations from local thermodynamic equilibrium
(LTE) are employed by different RT codes to calculate the properties
of the gas and of the radiation field. Approximate treatment of atomic
physics is also required due to the partially calibrated atomic data.
The back-reaction of radiation on the hydrodynamics provides an
additional challenge, requiring the solution of hydrodynamic equations
coupled to the RT solution. In many cases however, in particular for
Type Ia supernovae (SNe~Ia) at phases beyond several days which we
will focus on here, the radiation carries a negligible fraction of the
energy and the ejecta are freely expanding homologously. The RT
problem in these cases is decoupled from the hydrodynamics problem,
the latter providing the initial ejecta profiles (henceforth `ejecta
models'). The ejecta profiles include density, composition, and
initial temperature as a function of position. The initial time (of
order 1 day) is such that on the one hand it is much larger than the
explosion time scale (of order 1 second) so that the expansion is
nearly homologous and on the other hand sufficiently early such that
radiation has hardly diffused across the ejecta and the only evolution
is due to adiabatic expansion and radioactive decay.
Comparisons of the results of different RT codes for the same ejecta
models play an important role for estimating the accuracy of different
approximations and for validating new codes. The number and
sophistication of RT codes has significantly developed in recent
years, increasing the need for diverse benchmark ejecta models that
will allow detailed and careful comparisons. In this paper we describe
the first results of a collaborative effort of ten groups around the
world that are developing existing RT codes (in alphabetical order:
ARTIS, CMFGEN, CRAB, KEPLER, SEDONA, STELLA, SUMO, SuperNu, TARDIS,
and URILIGHT; see Sect.~\ref{sect:codes} for descriptions and
references) to create a systematic code-comparison framework. As a
first important step, all groups agreed on a set of four test model
ejecta and standardized output formats. Each group calculated the
resulting radiative display with their respective codes for the same
ejecta models and shared the results in a new public electronic
repository on
GitHub\footnote{\url{https://github.com/sn-rad-trans}}. We did not
attempt to agree on a specific setup for each code, nor did we
synchronize atomic data between codes.
The structure of the paper is as follows. In Sect.~\ref{sect:models},
the benchmark models are described. These include two simplistic
sub-Chandrasekhar-mass (sub-\ensuremath{M_{\rm Ch}}) toy models with profiles that are
defined analytically and two more realistic models which result from
hydrodynamical simulations of the \ensuremath{M_{\rm Ch}}\ delayed-detonation
scenario. All the models considered here are spherically symmetric
(1D). We give short descriptions of the RT codes that were employed in
this first comparison in Sect.~\ref{sect:codes}, with emphasis on the
main physical approximations that are used in each. In
Sect.~\ref{sect:outputs} we describe the publicly-available repository
of results. In particular, descriptions are provided of the output
files and of python codes that are included for reading them. In
Sect.~\ref{sect:results}, several example comparisons of the results
of the different codes are provided in order to illustrate the
available outputs. We voluntarily make no attempt to analyse the
sources of discrepancies. While the comparisons focus on observable
aspects of the emission, comparison to observations and conclusions
regarding the implications for the applicability of the codes are
intentionally not addressed in order to keep the focus of the paper on
the description of the comparison. Finally, we outline future plans
for this comparison project in Sect.~\ref{sect:conclusions}.
\section{Test models}\label{sect:models}
The code-comparison test suite consists of four SN~Ia\ models. Two are
sub-\ensuremath{M_{\rm Ch}}\ models with analytic density and composition profiles
(Sect.~\ref{sect:toymodels}; ``toy'' models), and the remaining two
are \ensuremath{M_{\rm Ch}}\ models resulting from hydrodynamical simulations
(Sect.~\ref{sect:ddcmodels}; DDC models). The models were set up or
selected based on their \ensuremath{^{56}\rm{Ni}}\ yield, so as to have two models
corresponding to `normal' SNe~Ia\ (toy06 and DDC10 with
$\sim$0.6\,\ensuremath{\rm{M}_{\odot}}\ of \ensuremath{^{56}\rm{Ni}}) and two low-luminosity models (toy01 and
DDC25 with $\sim$0.1\,\ensuremath{\rm{M}_{\odot}}\ of \ensuremath{^{56}\rm{Ni}}). We present the toy and DDC
models in turn in the following sections, and summarize their
properties in Table~\ref{tab:models}. The density profiles at a
reference time of 1\,d past explosion and initial composition profiles
are shown in Figs.~\ref{fig:dens} and \ref{fig:composition},
respectively.
\begin{table*}
\footnotesize
\caption{Summary of ejecta conditions. The yields for representative species corresponds to the start of our simulations in our model set (2\,d past explosion for the toy models and $\sim$1\,d past explosion for the DDC models). The $^{56}$Ni mass is given prior to any decay.}
\label{tab:models}
\begin{center}
\begin{tabular}{lccccccccc}
\hline
Model & $M_{\rm ej}$ & $E_{\rm kin}$ & M($^{56}$Ni)$_{t=0}$ & M(Fe) & M(Ca) & M(S) & M(Si) & M(O) & M(C) \\
& (\ensuremath{\rm{M}_{\odot}}) & ($10^{51}$\,erg) & (\ensuremath{\rm{M}_{\odot}}) & (\ensuremath{\rm{M}_{\odot}}) & (\ensuremath{\rm{M}_{\odot}}) & (\ensuremath{\rm{M}_{\odot}}) & (\ensuremath{\rm{M}_{\odot}}) & (\ensuremath{\rm{M}_{\odot}}) & (\ensuremath{\rm{M}_{\odot}}) \\
\hline
toy06 & 1.00 & 1.00 & 0.600 & 0.001 & 0.040 & 0.140 & 0.220 & 0.000 & 0.000 \\
toy01 & 1.00 & 1.00 & 0.100 & 0.000 & 0.090 & 0.315 & 0.495 & 0.000 & 0.000 \\
DDC10 & 1.42 & 1.51 & 0.620 & 0.112 & 0.041 & 0.166 & 0.257 & 0.101 & 0.002 \\
DDC25 & 1.39 & 1.18 & 0.115 & 0.098 & 0.024 & 0.237 & 0.478 & 0.282 & 0.022 \\
\hline
\end{tabular}
\end{center}
\end{table*}
\subsection{Toy models}\label{sect:toymodels}
\begin{figure}
\includegraphics[width=\hsize]{fig1.pdf}
\caption{Density profiles of our model set at an adopted time of 1\,d past explosion.
}
\label{fig:dens}
\end{figure}
\begin{figure*}
\includegraphics[width=0.5\hsize]{fig2a.pdf}
\includegraphics[width=0.5\hsize]{fig2b.pdf}
\includegraphics[width=0.5\hsize]{fig2c.pdf}
\includegraphics[width=0.5\hsize]{fig2d.pdf}
\caption{Composition profiles at the start of our simulations in our
model set (2\,d past explosion for the toy models and $\sim$1\,d
past explosion for the DDC models). For the toy models this
represents the full set of species present (\ensuremath{^{56}\rm{Ni}}\ and decay
products \ensuremath{^{56}\rm{Co}}\ and \ensuremath{^{56}\rm{Fe}}, as well as IMEs: Ca S, Si), while for
the DDC models only a subset of species is shown for
illustration. The \ensuremath{^{56}\rm{Ni}}\ mass fraction is given at the time of
explosion. We also show the total iron-group element (IGE) mass fraction (from Sc to Ni;
dashed line) and the total IME mass fraction (from Na to Ca; dotted
line). The total IGE mass fraction coincides with the $\ensuremath{^{56}\rm{Ni}}_0$
line in the toy models and is not shown for sake of clarity.
}
\label{fig:composition}
\end{figure*}
The toy models were generated using the \verb+mk_snia_toy_model.py+
script (see Sect.~\ref{sect:outputs}) using the \verb+--highni+ (for
the normal SN~Ia\ model) and \verb+--lowni+ (for the low-luminosity
SN~Ia\ model) options. Both models have a total mass
$M_\mathrm{ej}=1.0$\,\ensuremath{\rm{M}_{\odot}}, a kinetic energy
$E_\mathrm{kin}=10^{51}$\,erg, and are calculated at the time
$t_f=2$\,d past explosion. The models have an exponential density
profile \citep[e.g.][]{Jefferey1999},
\begin{equation}\label{eq:rhoexp}
\rho = \rho_\mathrm{c}e^{-v/v_e},
\end{equation}
where
\begin{equation}\label{eq:ve}
v_e = \sqrt{\frac{E_\mathrm{kin}}{6M_\mathrm{ej}}} \approx 2895\mathrm{\,\ensuremath{\rm{km\,s}^{-1}}}
\end{equation}
is the e-folding velocity, and
\begin{equation}\label{eq:rhoc0}
\rho_\mathrm{c} = \frac{M_\mathrm{ej}}{8\pi v_e^3 t_f^3} \approx 6.32 \times 10^{-10}\mathrm{\,\ensuremath{\rm{g\,cm}^{-3}}}
\end{equation}
is the central density at the chosen time.
The ejecta consist of $n$ spherical shells indexed by
$i=1,\ldots,n$. We use a uniform velocity grid with width $\Delta v =
50$\,\ensuremath{\rm{km\,s}^{-1}}. Each shell then has an inner and outer velocity coordinate
given by: $v_{0,i} = (i-1) \Delta v$ and $v_{1,i} = v_{0,i} + \Delta
v$.
As in \cite{Jefferey1999} we define the dimensionless radial
coordinate $z = v / v_e$ which we use to compute the mass of each
shell as:
\begin{equation}\label{eq:mshell}
M_i = \int \rho(z)\ \mathrm{d}V = \rho_\mathrm{c} \int e^{-z}\ \mathrm{d}V,
\end{equation}
\noindent
where the integral runs from $z_{0,i}$ to $z_{1,i}$ and the volume
element is $dV = 4\pi r^2 dr = 4\pi v_e^3 t_f^3 z^2 dz$, where we have
assumed a homologously expanding ejecta ($r = vt = v_e z t$).
Likewise, each shell volume $V_i$ is computed from the inner and outer
radii ($r_{i,\{0,1\}} = v_{i,\{0,1\}} t_f$), which results in the mean
density of each shell:
\begin{equation}\label{eq:rhoave}
\rho_i = \frac{M_i}{V_i} = 3 \rho_c \frac{
e^{-z_{0,i}}(z_{0,i}^2 + 2z_{0,i} + 2) -
e^{-z_{1,i}}(z_{1,i}^2 + 2z_{1,i} + 2)}
{z_{1,i}^3 - z_{0,i}^3}.
\end{equation}
\noindent
Starting from the central shell, we keep adding successive shells
until the total mass is within 0.01\% of the required ejecta mass.
For the 1.0\,\ensuremath{\rm{M}_{\odot}}\ toy models considered here this results in 807
shells, where the last shell has an outer velocity of 40\,350\,\ensuremath{\rm{km\,s}^{-1}}.
We note that not all codes used this velocity grid; in particular the
number of shells and maximum velocity vary significantly among codes
(see Sect.~\ref{sect:codes}).
We assume the ejecta consist of a central region composed of
\ensuremath{^{56}\rm{Ni}}\ (and its decay products \ensuremath{^{56}\rm{Co}}\ and \ensuremath{^{56}\rm{Fe}}) and an outer region
composed of the intermediate-mass elements (IMEs) Ca, S, and Si with
constant mass fractions throughout the layer: $X(\mathrm{Ca})=0.1$,
$X(\mathrm{S})=0.35$, and $X(\mathrm{Si})=0.55$. These values were
chosen based on the delayed-detonation model DDC10 (see
Sect.~\ref{sect:ddcmodels} below). Our `normal' toy model consists of
0.6\,\ensuremath{\rm{M}_{\odot}}\ of \ensuremath{^{56}\rm{Ni}}\ and 0.4\,\ensuremath{\rm{M}_{\odot}}\ of IMEs, while our
low-luminosity toy model consists of 0.1\,\ensuremath{\rm{M}_{\odot}}\ of \ensuremath{^{56}\rm{Ni}}\ and
0.9\,\ensuremath{\rm{M}_{\odot}}\ of IMEs. The \ensuremath{^{56}\rm{Ni}}\ and IME composition profiles are
smoothly connected using an analytic function (here a cosine bell)
over a mass interval $\Delta M_\mathrm{trans}$ (set to 0.2\,\ensuremath{\rm{M}_{\odot}}\ for
the low-luminosity model and 0.4\,\ensuremath{\rm{M}_{\odot}}\ for the regular model). At a
given mass coordinate $m$, the \ensuremath{^{56}\rm{Ni}}\ mass fraction is set to:
\begin{equation}
X_{\ensuremath{^{56}\rm{Ni}}}(m) =
\begin{cases}
1 & \text{for $m < m_1$}, \\
1 - f_\mathrm{cos}(m) & \text{for $m_1 \le m \le m_2$}, \\
0 & \text{for $m > m_2$}.
\end{cases}
\end{equation}
\noindent
where $m_1 = M(\ensuremath{^{56}\rm{Ni}}) - \Delta $$M_\mathrm{trans}/2$,
$m_2 = M(\ensuremath{^{56}\rm{Ni}}) + \Delta $$M_\mathrm{trans}/2$, and
\begin{equation}
f_\mathrm{cos}(m) = \dfrac{1}{2} \left\{
1 - \cos \left[ \left( \dfrac{m - m_2}{\Delta M_\mathrm{trans}} \right) \pi \right]
\right\}.
\end{equation}
\noindent
The IME mass fraction is then simply set to $X_\mathrm{IME}(m) = 1 -
X_{\ensuremath{^{56}\rm{Ni}}}(m)$. Our toy models thus consist of only six chemical
species/isotopes (\ensuremath{^{56}\rm{Ni}}, \ensuremath{^{56}\rm{Co}}, \ensuremath{^{56}\rm{Fe}}, $^{40}$Ca, $^{32}$S,
$^{28}$Si): this was a voluntary choice so as to keep the model as
simple as possible while retaining the defining characteristics of a
SN~Ia.
The initial temperature profile at $t_f=2$\,d is determined by solving
the first law of thermodynamics assuming a radiation-dominated gas,
local energy deposition from \ensuremath{^{56}\rm{Ni}}\ decay, and no diffusion (i.e. the
temperature in each zone is solved independently from adjacent
zones). Given these assumptions, the temperature at $t_f$ can be
determined analytically by noting that the time-weighted internal
energy, $tE(t)$, equals the time-integrated time-weighted decay energy
deposition rate, $\int{t Q(t) \mathrm{d}t}$, as noted by
\cite{Katz2013}:
\begin{equation}\label{eq:temp}
T_i = \left( \frac{\rho_i X_i(\mathrm{\ensuremath{^{56}\rm{Ni}}})_0 \int_0^{t_f} tq_{\rm Ni}(t) \mathrm{d}t}{a ~ t_f} \right) ^{1/4},
\end{equation}
\noindent
where $X_i(\ensuremath{^{56}\rm{Ni}})_0$ is the \ensuremath{^{56}\rm{Ni}}\ mass fraction at $t\approx0$ in the
i'th cell, $a$ is the radiation constant and $q_{\rm Ni}(t)$ is the
energy release rate per unit mass (ignoring neutrinos) of the
\ensuremath{^{56}\rm{Ni}}$\rightarrow$\ensuremath{^{56}\rm{Co}}$\rightarrow$\ensuremath{^{56}\rm{Fe}}\ decay chain. In this
formulation we ignore the time-weighted internal energy shortly after
explosion, $E(t_0) t_0$.
It is clear from Eq.~\ref{eq:temp} that the temperature is predicted
to be zero in zones devoid of \ensuremath{^{56}\rm{Ni}}\ ($\gtrsim 12000$\,\ensuremath{\rm{km\,s}^{-1}}). We thus
impose a constant temperature floor in these zones of 5000\,K. The
resulting excess internal energy is quickly radiated away since these
layers have a relatively low optical depth, such that the impact on
the longer-term radiative display is negligible.
\subsection{Delayed-detonation models}\label{sect:ddcmodels}
In addition to the two toy models above we consider two
\ensuremath{M_{\rm Ch}}\ delayed-detonation models resulting from hydrodynamical
simulations. These were chosen from the DDC model grid presented in
\cite{Blondin2013}, in order to closely match the \ensuremath{^{56}\rm{Ni}}\ yields of the
toy models: Our low-luminosity model DDC25 yields $\sim
0.12$\,\ensuremath{\rm{M}_{\odot}}\ of \ensuremath{^{56}\rm{Ni}}\ (cf. 0.1\,\ensuremath{\rm{M}_{\odot}}\ for the toy01 model), and our
`normal' DDC10 model yields $\sim 0.52$\,\ensuremath{\rm{M}_{\odot}}\ of
\ensuremath{^{56}\rm{Ni}}\ (cf. 0.6\,\ensuremath{\rm{M}_{\odot}}\ for the toy06 model). We refer the reader to
\cite{Blondin2013} for a detailed description of the DDC models.
The outputs of the hydrodynamical modelling correspond to 30-60\,s past
explosion, by which time the ejecta have reached a state of homologous
expansion. We applied a small radial mixing to the hydrodynamical
input with a characteristic velocity width $\Delta v_{\rm
mix}=400$\,\ensuremath{\rm{km\,s}^{-1}}\ to smooth sharp variations in composition. The
ejecta were then evolved to 0.5\,d past explosion by solving the
energy equation given by the first law of thermodynamics, assuming the
plasma is radiation dominated and neglecting diffusion. Apart from the
two-step \ensuremath{^{56}\rm{Ni}}$\rightarrow$\ensuremath{^{56}\rm{Co}}$\rightarrow$\ensuremath{^{56}\rm{Fe}}\ decay chain, we
also treat eight additional two-step decay chains associated with
$^{37}$K, $^{44}$Ti, $^{48}$Cr, $^{49}$Cr, $^{51}$Mn, $^{52}$Fe,
$^{55}$Co, $^{57}$Ni, and a further six one-step decay chains
associated with $^{41}$Ar, $^{42}$K, $^{43}$K, $^{43}$Sc, $^{47}$Sc,
$^{61}$Co \citep[see][]{Dessart2014}.
The ejecta at 0.5\,d are then remapped onto the 1D, non-LTE,
radiative-transfer code CMFGEN of \cite{Hillier2012_cmfgen} and
evolved until $\sim 1$\,d past explosion (0.976\,d for the DDC10 model
and 1.3\,d for the DDC25 model; see Sect.~\ref{sect:cmfgen} for
details), at which point the ejecta serve as initial conditions for
the other radiative-transfer codes presented in
Sect.~\ref{sect:codes}.
\section{Radiative-transfer codes}\label{sect:codes}
\begin{table*}
\scriptsize
\caption{Physical ingredients and approximations used in each code for the test models in this paper.}\label{tab:codephys}
\begin{tabular}{lccccccccc}
\hline
Code & RT Method & Homologous & $\gamma$-ray & Non-Thermal & Excitation & Ionization & Radiation & Line Opacity & Thermalization \\
& & Expansion & Deposition & Deposition & & & Field $J_\nu$ & $\kappa_\nu$ & Parameter $\epsilon$ \\
(1) & (2) & (3) & (4) & (5) & (6) & (7) & (8) & (9) & (10) \\
\hline
ARTIS & MC & yes & MC & $\cdots$ & LTE($T_R$) & approx. $dn/dt=0$ & scaled LTE($T_R$) & Sobolev & $\cdots$ \\
ARTIS nebular & MC & yes & MC & Spencer-Fano & $dn/dt=0$ & $dn/dt=0$ & $dJ/dt$ & Sobolev & $\cdots$ \\
CMFGEN & RTE-CMF & yes & MC & Spencer-Fano & $dn/dt$ & $dn/dt$ & $dJ/dt$ & $\kappa_\nu$ & $\cdots$ \\
CRAB & RH-1G & no & grey & Kozma/Fransson & LTE($T_R$) & LTE($T_R$) & $dJ/dt$ & Expansion & 0.9 \\
KEPLER & FLD & no & grey & $\cdots$ & $\cdots$ & LTE($T_e$) & LTE($T_e$) & $\kappa=\mathrm{const.}$ & $\cdots$ \\
SEDONA & MC & yes & MC & $\cdots$ & LTE($T_e$) & LTE($T_e$) & $dJ/dt$ & Expansion & 0.8 or 1.0 \\
SUMO & MC & yes & grey & Spencer-Fano & $dn/dt=0$ & $dn/dt=0$ & $dJ/dt=0$ & Sobolev & $\cdots$ \\
STELLA & RH-MG & no & grey & $\cdots$ & LTE($T_e$) & LTE($T_e$) & $dJ/dt$ & Expansion & 0.9 \\
SuperNu & MC & yes & MC grey & $\cdots$ & LTE($T_e$) & LTE($T_e$) & $dJ/dt$ & $\kappa_\nu$ & 1.0 \\
TARDIS & MC & yes & $\cdots$ & $\cdots$ & scaled LTE($T_R$) & scaled LTE($T_R$) & scaled LTE($T_R$) & Sobolev & $\cdots$ \\
URILIGHT & MC & yes & MC & $\cdots$ & LTE($T_e$) & LTE($T_e$) & $dJ/dt$ & Expansion & 0.8 \\
\hline
\end{tabular}
\flushleft
\textbf{Notes:}
\textit{Column headings:}
(1) Code name.
(2) Numerical method used to solve the
radiative-transfer equation: FLD=Flux Limited Diffusion, MC=Monte
Carlo, RH-1G = one-group (grey) radiation hydrodynamics, RH-MG =
multi-group radiation hydrodynamics, RTE-CMF= Radiation Transfer
Equation Co-Moving Frame.
(3) The ejecta are assumed to be in homologous expansion ($v=rt$) in radiative-transfer codes. This is not the case for radiation-hydrodynamics codes (CRAB, KEPLER, STELLA).
(4) Treatment of $\gamma$-ray energy deposition.
(5) Non-thermal heating, excitation and ionization rates are calculated through a solution of the Spencer-Fano equation \citep{Spencer1954} or read in from tabulated values \citep{Kozma1992}.
(6) Solution method for the atomic level populations. LTE($T_X$) refers to a solution of the Boltzmann excitation formula setting the temperature to that of the electrons ($T_e$) or the radiation field ($T_R$). An approximate non-LTE treatment of excitation scales the Boltzmann occupation numbers by the dilution factor $W$ (cf. dilute-LTE treatment in TARDIS; Sect.~\ref{sect:tardis}). A non-LTE treatment requires the solution of the rate equations, either including time dependence ($dn/dt$) or assuming steady-state (statistical equilibrium, $dn/dt=0$).
(7) Treatment of ionization. Here LTE($T_X$) refers to a solution of the Saha-Boltzmann equation, which can be scaled for an approximate non-LTE treatment (cf. nebular approximation in TARDIS; Sect.~\ref{sect:tardis}). The non-LTE solution results from the solution of the rate equations, either including time dependence ($dn/dt$) or assuming steady-state ($dn/dt=0$).
(8) The radiation field can be computed via a solution of the radiative-transfer equation (possibly assuming steady-state, $dJ/dt=0$) or by following the propagation of photon packets in Monte Carlo codes. Alternatively, LTE treatments assume a Planckian radiation field (black body $B_\nu$) at a reference temperature $T_X$, possible scaled by the dilution factor $W$.
(9) Treatment of line opacity. This can be explicitly line by line, taking into account overlap in the co-moving frame ($\kappa_\nu$) or using the Sobolev approximation. Other treatments involve the use of an approximate frequency-dependent 'expansion' opacity, or assuming a constant value (e.g. KEPLER; Sect.~\ref{sect:kepler}).
(10) Global value of the thermalization parameter $\epsilon$, which sets the probability that a photon absorbed in a given transition is re-emitted in a different transition (see, e.g. URILIGHT; Sect.~\ref{sect:urilight}).
\end{table*}
In the following subsections each group provides a brief description
of their code, highlighting the specific setup used in the
calculations for this paper. Each code subsection follows a similar structure: brief description of the code (and whether it assumes a homologous velocity law); treatment of $\gamma$-ray energy deposition and non-thermal effects; computation of the temperature structure; treatment of excitation and ionization; evaluation of the radiation field; sources of opacity and atomic data; resolution (spatial and frequency) and typical run time. Table~\ref{tab:codephys} summarizes the physical ingredients and approximations used in each code for the test models considered here.
\subsection{ARTIS}\label{sect:artis}
ARTIS\footnote{Source code available at
\url{https://github.com/artis-mcrt/artis}} is a Monte Carlo radiative
transfer code that uses the indivisible energy packet method of
\cite{Lucy2002}. The code was originally described by \cite{Sim2007}
and \cite{Kromer2009}, with later improvements presented by
\cite{Bulla2015} and \cite{Shingles2020}. The code is
three-dimensional and follows the time-evolution of the radiation
field and state of the gas. It assumes a strictly homologous velocity
law.
Injection of energy into the ejecta is calculated by following the
deposition of $\gamma$-ray packets that are injected in accordance
with the radioactive decays of $^{56}$Ni and $^{56}$Co, following
\cite{Lucy2005} (additional decay channels have been included in the
studies of specific models). The simulated $\gamma$-ray transport is
non-grey and takes into account Compton scattering, photoelectric
absorption and pair-creation opacities. In our standard runs, the code
does not include the effects of excitation or ionization by
non-thermal particles. However, \cite{Shingles2020} presented updates
to the code that include a Spencer-Fano treatment of non-thermal
ionization, as required for late-phase modelling. Results obtained
with this improved version (\texttt{artisnebular}) are included for
late phases for the models in this study.
The electron temperature in each grid zone is estimated by balancing
of heating and cooling rates (accounting for $\gamma$-ray and positron
deposition, bound-bound, bound-free and free-free processes). In its
regular mode of operation (\texttt{artis}), the code uses an
approximate non-LTE treatment to estimate the ionization state in the
ejecta (based on Monte Carlo photoionization estimators; see
\citealt{Kromer2009}) and an LTE treatment of excitation is
adopted. This approach has been used in most of our published studies,
and is also used in most of the \texttt{artis} calculations presented
here. However, this method has limitations that become increasingly
important at later phases (e.g. it neglects non-thermal
heating/ionization and tends to overestimate the plasma temperature at
low densities due to incomplete treatment of cooling by forbidden
lines). To improve these issues, \cite{Shingles2020} presented updates
to the code that include a full non-LTE population solver (together
with the Spencer-Fano solver mentioned above). Results obtained with
this improved version (\texttt{artisnebular}) are included for
late-phase calculations here.
Monte Carlo estimators are used to track the radiation field in each
grid cell. In general we use volume-based estimators (see
\citealt{Lucy1999} or \citealt{Noebauer2019}) to extract radiation
field dependent quantities from the flight histories of our Monte
Carlo quanta. In its standard mode of operation, \texttt{artis} uses
detailed Monte Carlo estimators to obtain photoionization rates from
the radiation field, but relies on an estimated scaling for excited
state photoionization and on a dilute black-body radiation field model
for bound-bound excitation (see \citealt{Kromer2009} for details). The
improved \texttt{artisnebular} version, however, uses a more detailed,
frequency binned representation of the radiation field (see
\cite{Shingles2020} for details). The code has the capacity to iterate
on each time step with the aim of achieving consistency between the
radiation field estimates and the packet transport in each step. In
practise, however, we find that this iteration is not required and we
therefore generally simply use the radiation field quantities
extracted from the previous time step to estimate the radiative rates
that are needed for the current step.
In simulating ultraviolet to infra-red photon transport, the code
accounts for electron scattering, bound-bound, bound-free and
free-free processes. Bound-free and bound-bound processes are treated
using the \emph{Macro Atom} approach of \cite{Lucy2002, Lucy2003} and
adopting the Sobolev approximation for line opacity. The code does not
use an expansion opacity (or similar) but treats line opacity based on
a frequency-ordered list of transitions treated in the Sobolev limit
(i.e. no line overlap is taken into account).
In our simulations, atomic data is primarily drawn from the Kurucz
atomic line lists (see \citealt{Kromer2009}) -- in Appendix A.1 we
summarise the ions and numbers of levels and lines that we include.
The photon transport is carried out on a 3D Cartesian grid that
co-expands with the ejecta. The \texttt{artis} simulations included
here were carried out on a 100$^3$ grid. The resolution therefore
corresponds to around 500 -- 1000 km s$^{-1}$, depending on the
model. The simulations are typically run on 1000 compute cores for one
to two days.
\subsection{CMFGEN}\label{sect:cmfgen}
CMFGEN is a 1D, non-LTE, time-dependent, radiative-transfer code that
solves the transfer equation in the co-moving frame of spherical
outflows. Details about the code, techniques, and approximations can
be found in
\cite{Hillier1998_cmfgen,Hillier2003_cmfgen,Hillier2012_cmfgen_HSW}
and (for SN calculations) in
\cite{Hillier2012_cmfgen}\footnote{CMFGEN, with documentation, is
available at \url{www.pitt.edu/~hillier}}. The velocity law for the
outflow is in general monotonic (but see, e.g. \citealt{Dessart2015}
in the case of interacting supernovae) and is assumed here to be
homologous (such that $\partial v/\partial r = v/r$).
In the present calculations, non-local energy deposition from
radioactive decay is treated using a Monte-Carlo approach for
$\gamma$-ray transport \citep{Hillier2012_cmfgen}. Non-thermal
processes associated with the high-energy electrons produced by
Compton scattering and photoelectric absorption of these $\gamma$ rays
are accounted for through a solution of the Spencer-Fano equation
\citep{Spencer1954,Li2012}.
The temperature structure is constrained through the energy equation
that has the form:
\begin{equation}
\rho {De \over Dt} - {P \over \rho} {D\rho \over Dt}= 4\pi \int d\nu
(\chi_\nu J_\nu - \eta_\nu) + \dot \epsilon_{\rm decay} \,,
\label{eq_energy}
\end{equation}
where $D \over Dt$ is the Lagrangian derivative, \noindent $e$ is the
internal energy per unit mass, $P$ is the gas pressure, and $\dot
\epsilon_{\rm decay}$ is the energy absorbed per second per unit
volume (see \citealt{Hillier2012_cmfgen} for further details). We
verify the accuracy of the solution by examining a global energy
constraint (equivalent to conservation of flux in a static atmosphere;
see \citealt{Hillier2012_cmfgen} for details), and an equation
describing energy conservation as applied to the heating and cooling
of free electrons. These two equations are related to the above
Eq.~\ref{eq_energy} via the transfer equation, and the rate equations,
respectively \citep[e.g.][]{Hillier2003_cmfgen}. Because of model
assumptions (e.g. the use of super-levels) these two equations are not
satisfied exactly, but the errors (typical at the 1\% level or
smaller) are too small to affect the validity of the models.
Processes contributing to electron heating and cooling include
bound-free ionization and recombination, collisional ionization and
recombination, collisional excitation and de-excitation, free-free
emission, Auger ionization, charge exchange reactions (primarily with
H\ts {\sc i}\ and He\ts {\sc i}, and hence negligible in SNe~Ia\ ejecta), as well as
the net cooling from non-thermal processes and heating by radioactive
decay.
Atom and ion level populations are determined through a solution of
the time-dependent rate equations, coupled to the above energy
equation and the 0$^{\rm th}$ and 1$^{\rm st}$ moments of the
radiative-transfer equation (see below). We consider the following
processes: bound–bound processes, bound–free processes, collisional
ionization and recombination, collisional excitation and
de-excitation, Auger ionization
\citep{Hillier1987,Hillier1998_cmfgen}, and non-thermal excitation and
ionization \citep{Li2012}. We additionally consider further processes
involving H and He (two-photon decay, charge-exchange reactions, and
Rayleigh scattering), although these are negligible here (and
completely absent from the toy models which contain no H or He). To
ease the solution of the rate equations, atomic levels are grouped
into super levels (see \citealt{Hillier1998_cmfgen} for details).
The frequency-dependent mean intensity $J_\nu$ is obtained via a
solution of the time-dependent transfer equation in the co-moving
frame, to first order in $v/c$. More specifically, we solve the
0$^{\rm th}$ and 1$^{\rm st}$ moment equations, which are closed using
so-called Eddington factors $f_\nu=K_\nu/J_\nu$, where $K_\nu$ is the
second moment of the specific intensity (related to the radiation
pressure). The Eddington factors are obtained from a formal solution
of the time-independent transfer equation.
We consider the following sources of opacity: electron scattering,
bound-free (including photoionization from excited states),
bound-bound\footnote{In these calculations we assume a Doppler profile
with a constant effective Doppler width (including both thermal and
turbulent velocities) of 50\,\ensuremath{\rm{km\,s}^{-1}}. In more general supernovae
modelling, the effective Doppler width is varied to test its effect on
model results.}, free-free, and Auger ionization. As noted earlier,
Rayleigh scattering and two-photon processes (for H and He only) are
also part of the global opacity budget but are negligible here.
A description of the sources of atomic data can be found in the
Appendix (Sect.~\ref{sect:atomdata_cmfgen}). The number of levels
(both super-levels and full levels) and corresponding number of
bound-bound transitions are given in
Tables~\ref{tab:cmfgen_atoms_toy06}-\ref{tab:cmfgen_atoms_ddc_reducednicofe4}.
We ignore weak transitions with a $gf$ value\footnote{This value is
the product of the statistical weight $g$ of the lower level and the
oscillator strength $f$ of an atomic transition.} below some cutoff,
typically set to $10^{-4}$. For the toy models the following ions
were included: Si\ts {\sc ii}--{\sc iv}, S\ts {\sc ii}--{\sc iv}, Ca\ts {\sc ii}--{\sc iv},
Fe\ts {\sc i}--{\sc v}, Co\ts {\sc ii}--{\sc vii}, and Ni\ts {\sc ii}--{\sc v}. For the
delayed-detonation DDC models the following ions were included:
C\ts {\sc i}--{\sc iv}, O\ts {\sc i}--{\sc iv}, Ne\ts {\sc i}--{\sc iii}, Na\ts {\sc i},
Mg\ts {\sc ii}--{\sc iii}, Al\ts {\sc ii}--{\sc iii}, Si\ts {\sc ii}--{\sc iv}, S\ts {\sc ii}--{\sc
iv}, Ar\ts {\sc i}--{\sc iii}, Ca\ts {\sc ii}--{\sc iv}, Sc\ts {\sc ii}--{\sc iii},
Ti\ts {\sc ii}--{\sc iii}, Cr\ts {\sc ii}--{\sc iv}, Mn\ts {\sc ii}--{\sc iii}, Fe\ts {\sc i}--{\sc
vii}, Co\ts {\sc ii}--{\sc vii}, and Ni\ts {\sc ii}--{\sc vii} (we also include the
ground states of Cl\ts {\sc iv}, K\ts {\sc iii}, and V\ts {\sc i}\ for the sole purpose of
tracking changes in abundance of radioactive isotopes). For all the
aforementioned ions, we also consider ionizations to and
recombinations from the ground state of the next ionization stage
(e.g. \ion{Ni}{8} in the case of Ni). As time proceeds and the
temperature in the spectrum-formation region drops, the highest
ionization stages have a negligible impact on the RT solution. When
this occurs, smaller model atoms are used for these ions, or the ions
are removed altogether from the atomic model set.
The toy models were remapped onto a coarser spatial grid with 100
depth points. No remapping was performed for the DDC models. Typical
wall-clock run times are of the order of 24h per time step on a single
computing node with 8-12 CPUs, thus taking 2-3 months to complete for
a sequence covering the first 200\,d or so past explosion.
\subsection{CRAB}\label{sect:crab}
CRAB is a 1D, implicit, Lagrangian, radiation hydrodynamics code developed
to study the light curves during supernova outbursts and the corresponding
outflows evolved from hydrostatic state up to homologous expansion
\citep{Utrobin2004}.
Non-local energy deposition of $\gamma$ rays from radioactive decay is
determined by solving the corresponding one-group $\gamma$-ray transport
with the approximation of an effective absorption opacity of 0.06
$Y_\mathrm{e}$\,cm$^2$\,g$^{-1}$. Positrons are assumed to deposit
their energy locally. This energy deposition produces non-thermal
ionization and heating whose rates are taken from \cite{Kozma1992}.
The radiation hydrodynamic equations include a time-dependent energy
equation which is based on the first law of thermodynamics and determines
the gas temperature structure. In the outer, transparent and semitransparent
layers of the SN ejecta, the local energy balance is in control of a net
between heating and cooling processes, while, in the inner, optically thick
layers, it is determined by the diffusion of equilibrium radiation.
The code has two options for the treatment of atom and ion level
populations. In option A the elements H, He, C, N, O, Ne, Na, Mg, Si, S,
Ar, Ca, Fe, and the negative hydrogen ion H$^{-}$ are included in the
non-LTE ionization balance. All elements but H are treated with 3
ionization stages. The ionization balance is controlled by the
following elementary processes: photoionization and radiative
recombination, electron ionization and three-particle recombination,
and non-thermal ionization. In option B an LTE treatment of ionization
and excitation for elements from the option A list or all elements
from H to Zn modelled with an arbitrary number of ionization stages
is used. Atomic and ionic level populations are determined by the
Boltzmann formulae and the Saha equations for an element mixture
with the local electron or non-equilibrium radiation temperature.
The time-dependent radiation transfer is treated in an one-group (grey)
approximation in the outer, transparent and semitransparent layers of
the SN envelope, while, in the inner, optically thick layers, where
thermalization of radiation takes place and LTE applies, the diffusion
of equilibrium radiation is described in the approximation of radiative
heat conduction. The bolometric luminosity of the SN is calculated by
including retardation and limb-darkening effects.
In the inner, optically thick layers the Rosseland mean opacity is
evaluated for the local electron temperature, while in the outer,
transparent and semitransparent layers non-LTE effects are taken into
account when determining the mean opacities and the thermal emission
coefficient. The mean opacities include processes of photoionization,
free-free absorption, Thomson scattering on free electrons, and
Rayleigh scattering on neutral hydrogen. The contribution of lines to
the expansion opacity is evaluated by the generalized formula of
\citet{Castor1975} or by the formula of \citet{Blinnikov1996_opac}
using the Sobolev approximation for line opacity. The expansion opacity
in an expanding medium is treated with a thermalization parameter set
to 0.9 as recommended by \cite{Kozyreva2020} to model SNe~Ia.
The sources of atomic data for processes in continuum can be found in
\citep{Utrobin2004}. Oscillator strengths of lines are taken from the Kurucz
line database\footnote{\url{http://kurucz.harvard.edu/cdroms.html}}
containing nearly 530\,000 lines. Energy level data are from the
atomic spectra database of the National Institute of Standards and
Technology.
The zoning of model toy06 with 808 zones is adequate to model a light curve.
For this model a typical runtime is of the order of 5 hours on one CPU for
the first 140\,d.
\subsection{KEPLER}\label{sect:kepler}
KEPLER is a one-dimensional, implicit Lagrangian hydrodynamics code
with appropriate physics to study massive stellar evolution and
supernovae \citep{Weaver1978,Woosley2002}. The radiation transport is
flux-limited diffusion using a single temperature for the matter and
radiation.
An important difference between KEPLER and some of the other codes
used here is that KEPLER does not assume a coasting configuration.
Especially at early times when the matter is very optically thick,
energy input from $^{56}$Ni and $^{56}$Co decay will both heat the
matter and accelerate it. The correction to the kinetic energy is
small for the cases studied here, but the integral of decay energy
that is deposited will be less than the integral of the emitted light
(minus adiabatic losses).
Our approach to $\gamma$-ray energy deposition is discussed in the
appendix of \citet{Ensman1988}. Since the early 90's, a value of
$\kappa_{\gamma}$ = 0.054 cm$^{-2}$ g$^{-1}$ has been used for the
global opacity parameter in KEPLER's leakage scheme to model
SNe~Ia\ and that is the value used here. A better value can be
obtained by comparing to Monte Carlo calculations for a given class of
model, i.e. SN IIP, SN Ib, etc. The value 0.054 cm$^2$ g$^{-1}$ is
larger than the standard local grey opacity $(0.06 \pm 0.01) Y_e$
cm$^2$ g$^{-1}$ \citep[e.g.][]{Swartz1995}, where $Y_e \approx 0.5$,
since it is used to calculate the effective column depth from the
outer edge of a spherical zone to the surface. For all points except
the geometric centre of the explosion, the angle averaged column depth
would be greater than along this radial line. The averaging is thus
approximated by taking a larger opacity. In reality, this number would
vary with the radial distribution of $^{56}$Ni and would be smaller if
all the $^{56}$Ni were at the centre.
The temperature structure is computed by solving the hydrodynamics
equations including the effects of expansion and acceleration with
energy input by radioactive decay and transported by radiative
diffusion. Because of the way KEPLER handles flux-limited transport using a
single temperature for the radiation and matter, the pressure in the
outermost zones can be overestimated. This can lead to a small (of
order 10\%) overestimate of the conversion of radiation to kinetic
energy in those zones at late times. To alleviate this problem, the
luminosity was taken to be the maximum of the value at the
(electron-scattering) photosphere and the zone that included 95\% of
the mass. The former dominated the light curve until after peak.
An important parameter of the calculation is the {\sl
atomic} opacity used to transport thermal radiation. In KEPLER the
total `optical' opacity consists of two parts: a) electron scattering,
which is calculated using a full Saha solve for all ionization stages
of all 19 species present; and b) a constant additive opacity,
$\kappa_a$, taken to reflect the effect of Doppler broadened
lines. The electron-scattering opacity is temperature, density, and
composition dependent and thus varies with location and time. The
additive opacity is a constant everywhere for all time. Traditionally,
we have used a value of 0.1 cm$^2$ g$^{-1}$ in our studies of Type Ia
light curves, but the best value will depend on supernova type. A
comparison of SN Ib models calculated using CMFGEN and KEPLER
\citep{Ertl2020} suggested a value of 0.03 cm$^2$ g$^{-1}$ and we
regard that as a lower limit for the average. The actual value, of
course, varies with location and time in a complicated way. Here we
adopt $\kappa_a = 0.1$ cm$^2$ g$^{-1}$ for the results presented in
this paper.
KEPLER carries a nuclear network of 19 species \citep{Weaver1978}
which do not correspond identically with all the species in the
initial models provided. Care was taken to translate the $^{56}$Ni
and and $^{56}$Co abundances given to the zero age $^{56}$Ni and
stable iron mass fractions which are what is used by KEPLER for energy
generation. The species $^{12}$C and $^{16}$O were mapped without
change. Other species like $^{20}$Ne, $^{28}$Si, etc. were slightly
augmented where necessary by adding in nearby odd-Z abundances. That
is, $^{20}$Ne included $^{20}$Ne and $^{23}$Na, $^{24}$Mg included
$^{24}$Mg and $^{27}$Al, etc. Since KEPLER does not compute spectra,
this should have negligible consequence.
The four models were mapped into KEPLER. Total mass, kinetic energy,
and $^{56}$Ni mass were preserved to 0.05\%. Because KEPLER is not a
special relativistic code, zones with velocities greater than $0.1c$
were trimmed from the input. High velocity zones would also cause
difficulty with editing the luminosity at late times if the light
crossing time ceased to be negligible. Removing this high velocity
material decreased the kinetic energy of all models by about 2\%. This
should have negligible effect on the light curve.
The zoning of the DDC models was by traditional KEPLER standards
quite coarse. After trimming the high velocity zones, only 80 zones
remained. The zoning of the toy models was better with roughly 600
zones remaining. Rezoning was not carried out. For a light curve with
no nuclear burning, the zoning is adequate. All calculations took at
most a few minutes on a laptop.
\subsection{SEDONA}\label{sect:sedona}
SEDONA is a time-dependent, multi-frequency Monte Carlo radiative
transfer code originally developed to study supernova light curves,
spectra, and polarization \citep{Kasen2006}.
For this comparison study, the gas properties were tracked on 1D
spherical Lagrangian zones and are evolved under the assumption of
homologous expansion.
The radioactive decays of \ensuremath{^{56}\rm{Ni}}\ and \ensuremath{^{56}\rm{Co}}\ were tracked and used to
source $\gamma$-ray packets. For the DDC10 model, radioactive decays
of $^{48}$Cr and $^{48}$V were also included in addition to the
\ensuremath{^{56}\rm{Ni}}\ decay chain. More detailed radioactive decay networks can be
implemented in SEDONA, but only the previously mentioned radioactive
decay processes were used in this comparison study.
The $\gamma$-ray packets were transported using the Monte Carlo
procedure subject to simplified treatments of bound-free absorption
and Compton down-scattering. The $\gamma$-ray interactions were
treated as heating terms that entered into the thermal heating
balance that sourced a population of thermal photon packets.
The temperature in each shell was computed assuming radiative
equilibrium. In more detail, the LTE assumption let us set the gas
emissivity to the Planck function times the frequency-dependent line
opacity. Then, the temperature was iteratively adjusted until the
frequency-integrated emissivity equals the total radiative energy,
from both $\gamma$-rays and re-emitted photon packets absorbed over
the previous particle propagation step, as measured in the co-moving
frame of the fluid.
For most of its published applications, SEDONA assumes LTE in order to
compute opacities, although non-LTE capabilities have been developed
and applied recently \citep{Roth2016,Shen2021}. For this comparison
study, only the LTE opacity mode was used in order to compute the
excitation and ionization states of the gas. More precisely, this
means that ionization fractions for each element were computed by
simultaneously solving the Saha equation using a local temperature,
and the charge conservation equation across all elements and isotopes
present. Meanwhile, the bound-electron level populations within each
ionization stage were set by the Boltzmann factors given by the local
temperature.
Photon packets were transported in three dimensions as a direct
solution to the time-dependent radiative transfer equation. For all
interactions with the gas, the packets were mapped to the 1D
Lagrangian zones.
Once the level populations were computed, the bound-bound opacity was
computed using the expansion opacity formalism, as described by
\citet{Eastman1993} and \citet{Kasen2006}. For the toy01 and toy06
models, only the bound-bound opacities and electron scattering
opacity, using the Thomson cross section, were included. For DDC10,
simplified bound-free and free-free opacities were also included in
addition the previously mentioned opacities, although they did not
have a noticeable effect during the early stages of the explosion that
we wish to compare using these calculations.
A thermalization parameter $\epsilon$ of 1.0 was used for the toy01
and toy06 models. This means that all photon packets (other than
$\gamma$-ray packets) that were absorbed were immediately thermalized,
so that each absorption event was followed by re-emission of a photon
packet with a frequency sampled from the thermal emissivity. For the
DDC10 model, $\epsilon$ was set to 0.8, so in that case 20\% of
absorbed photon packets (not including the $\gamma$-ray packets) were
coherently scattered instead of having their frequency re-sampled.
The atomic data used for the bound-bound transitions was taken from
Kurucz CD 1. This is a larger set of atomic data than CD 23. The
details are described in Appendix~\ref{sect:atomdata_sedona}
For this comparison study, the co-moving frequency grid for the
thermalized photons (i.e. not the $\gamma$-rays) used 17664 bins, with
equal logarithmic spacing, ranging from $10^{14}$ Hz to $2 \times
10^{16}$ Hz. The output spectra used 1044 frequency bins over the
range $1.1 \times 10^{14}$ Hz to $2 \times 10^{16}$ Hz
($\sim150$\,\AA\ to $\sim 2.7$\,$\mu$m). The time steps for the
homologous expansion and radioactive decay began at approximately a
few hours at the start of the calculation, and grew to no longer than
1 day, with a maximum time step growth rate of 10\%. One million
$\gamma$-ray packets were emitted at each time-step. As these $\gamma$
rays deposited their energy, their packets were discarded, while one
thermally sourced photon packet was emitted for each discarded
$\gamma$-ray packet. With all of the settings described above, a
calculation run to 100 days post-explosion required about 10 CPU
hours, and could be efficiently performed in parallel across several
hundred processors, reducing the elapsed wall time to less than one
hour.
\subsection{STELLA}\label{sect:stella}
The multi-group radiation hydrodynamics code STELLA
\citep{Blinnikov1993,Blinnikov1998,Blinnikov2006} is capable of
computing the evolution of the radiation field coupled to the
hydrodynamics, as well as the bolometric light curve, spectral energy
distribution and resulting broad-band magnitudes and colours. Hence,
STELLA does not require condition of homologous expansion and is
capable to treat shock propagation and any dynamical processes
happening in the SN ejecta.
Energy deposition from \ensuremath{^{56}\rm{Ni}}\ and \ensuremath{^{56}\rm{Co}}\ radioactive decay is treated
in a one-group diffusion approximation with an absorption opacity of
0.05 $Y_\mathrm{e}$\,cm$^{\,2}$\,g$^{\,-1}$ according to
\cite{Swartz1995}. The energy of positrons are thermalized locally.
The ionization and level populations of a limited set of species (H,
He, C, N, O, Ne, Na, Mg, Al, Si, S, Ar, Ca, stable Fe, stable Co, and
stable Ni), is treated in the LTE approximation.
Radiation is not treated in equilibrium with the matter. The colour
temperature is estimated as a black body temperature via the
least-squares method.
The opacity includes photoionization, free–free absorption, and
electron scattering processes assuming LTE for the plasma, and line
interactions. Radioactive \ensuremath{^{56}\rm{Ni}}\ and \ensuremath{^{56}\rm{Co}}\ contribute to the stable
Fe when the line opacity is calculated. The expansion opacity is
calculated according to \citep{Eastman1993}. The thermalization
parameter for the line opacity treatment is set to 0.9
\citep{Kozyreva2020}. The line opacity is calculated using a data base
of 153\,441 spectral lines partly from \citet{Kurucz1995},
\cite{Verner1995} and \cite{Verner1996_res}.
The spectral energy distribution is computed in the wavelength range
from 1\,\AA{} to 50,000\,\AA{}. The frequency range is divided into
129 bins with equal logarithmic spacing, in which the radiative
transfer equations are solved at every time step. For the toy06 model
a higher-resolution simulation with 629 frequency bins was run, which
is the version used in the following sections for this model. The
overall opacity might be slightly underestimated in the simulations
with 629 frequency bins compared to simulations involving 129 bins,
because the expansion opacity is calculated for the lines in the given
bin and is not extended to another bin even if the velocity gradient
is very high. The final (pseudo-)bolometric light curve is obtained by
integrating the spectra over frequency. The typical run time on a
single processor was a few hours at most, keeping the original spatial
resolution of the input models.
\subsection{SUMO}\label{sect:sumo}
SUMO \citep{Jerkstrandthesis2011, Jerkstrand2011,Jerkstrand2012} is a
homologous non-LTE code with radiative transfer. It is specialized to
calculate spectra and light curves in the post-peak phases of the
supernova. The code is written in Fortran and is parallelized with
MPI.
Gamma-ray transfer is done by ray-tracing using a grey opacity of
$0.06 Y_e$\,cm$^2$\,g$^{-1}$. Positrons can be transferred (using an
effective opacity 8.5 times the $\gamma$-ray one) but were here
assumed locally trapped. The cascade of non-thermal electrons
following the scattering of gamma rays and positrons are solved for
with the Spencer-Fano method \citep{Kozma1992}.
Zone temperatures are solved from the first law of thermodynamics
considering heating by non-thermal processes and photoionization, and
cooling by net collisional line excitation, free-free emission, and
recombination. The temperature is solved either in steady state
(heating = cooling) \citep{Jerkstrand2012} or time-dependently
\citep{Pognan2022}. Here, the steady-state mode was used.
The rate equations are solved by considering spontaneous radiative
decay with Sobolev escape probabilities (assuming homology) modified
to include continuum processes and strong line overlaps, thermal and
non-thermal collisions (excitations and ionizations),
photoexcitation/de-excitation, photoionization, and
recombination. Options exist to include also charge transfer processes
and molecular chemistry \citep{Liljegren2020}, but these were not
considered here. The rate equations can be solved either in steady
state (inflow = outflow) \citep{Jerkstrand2012} or time-dependently
\citep{Pognan2022}. Here, the steady-state mode was used.
Radiative transfer is calculated with a Monte Carlo simulation that is
iterated with the solvers for temperature and
ionization/excitation. During the transfer, photon packets can
experience electron scattering, bound-free and free-free absorption,
and line absorption. The line transfer is resolved line by line (with
Sobolev formalism) rather than using an expansion opacity
formalism. Photoexcitation is either fully coupled to the non-LTE
solutions (for low and mid-lying levels), or decoupled (for high-lying
levels), instead giving a fluorescence cascade on-the-spot. The
radiation field is computed in steady state.
The atomic data comes from a variety of sources, mostly described in
\cite{Jerkstrand2011} and \cite{Jerkstrand2012}. Model atoms have
LS-states resolved, i.e. do not use super levels. In models computed
up to 2022 (including the ones here), ions up to and including doubly
ionized were included although higher ions are now being included
\citep{Pognan2022}. For the toy models computed here the model atoms
are Fe\ts {\sc i}\ (496 levels), Fe\ts {\sc ii}\ (578 levels), Fe\ts {\sc iii}\ (600
levels), Ni\ts {\sc i}\ (136 levels), Ni\ts {\sc ii}\ (500 levels), Ni\ts {\sc iii}\ (8
levels), Co\ts {\sc i}\ (317 levels), Co\ts {\sc ii}\ (108 levels), Co\ts {\sc iii}\ (306
levels), Si\ts {\sc i}\ (494 levels), Si\ts {\sc ii}\ (77 levels), Si\ts {\sc iii}\ (2
levels), S\ts {\sc i}\ (123 levels), S\ts {\sc ii}\ (5 levels), S\ts {\sc iii}\ (6 levels),
Ca\ts {\sc i}\ (198 levels), Ca\ts {\sc ii}\ (69 levels), and Ca\ts {\sc iii}\ (1
level)\footnote{For Ni\ts {\sc iii}\, Si\ts {\sc iii}, S\ts {\sc ii}\ and Ca\ts {\sc iii}\ the low
number of levels is sufficient as the next state energies are at
52\,152, 52\,853, 79\,395 and 203\,373 cm$^{-1}$, respectively.}.
For the runs here, the ejecta were resampled to $\Delta v =
500$\,\ensuremath{\rm{km\,s}^{-1}}, and truncated at $v = 30000$\,\ensuremath{\rm{km\,s}^{-1}}. The radiative transfer
was computed on a wavelength grid from 400\,\AA\ to 25000\,\AA, with a
logarithmic resolution $\mathrm{d}\lambda/\lambda = 10^{-3}$. Typical
wall-clock runtimes for a single epoch is a few hours on a typical 128
core setup.
\subsection{SuperNu}\label{sect:supernu}
SuperNu is a multi-group LTE radiative transfer code that employs
Implicit Monte Carlo (IMC) and Discrete Diffusion Monte Carlo (DDMC)
\citep[][]{Wollaeger2013,Wollaeger2014}. IMC solves the thermal
radiative transfer equations semi-implicitly by treating some
absorption and emission as instantaneous effective scattering
\citep[see e.g. ][]{Fleck1971}. DDMC optimizes IMC in optically thick
regions of space \citep[][]{Densmore2012} and ranges of wavelength
\citep[][]{Abdikamalov2012,Densmore2012} by replacing many low
mean-free-path scattering events with single leakage events. SuperNu
can apply IMC and DDMC in both static and homologous,
semi-relativistically expanding atmospheres. The code has been
verified by analytic and semi-analytic radiative transfer tests
\citep[][]{Wollaeger2013} and on the W7 model of SNe Ia
\citep[][]{Nomoto1984,Wollaeger2014}.
For the $\gamma$-ray transfer, SuperNu employs a constant absorption
opacity of 0.06 $Y_e$~cm$^2$\,g$^{-1}$, where $Y_e$ is ejecta gas
electron fraction, following the prescription
of~\citet[][]{Swartz1995}. The $\gamma$-ray packets in SuperNu are
not directly converted to optical packets, but instead are used to
tally the total $\gamma$-ray energy deposition per spatial cell. The
deposition energy values are then added to the thermal source for
optical packets.
The ejecta gas temperature is calculated using the standard IMC
semi-implicit linearization~\citep[][]{Fleck1971} of the comoving
internal energy equation~\citep[][]{Wollaeger2013}: internal energy is
recast to gas temperature using the standard relation $\partial e =
c_v\partial T$, where $e$ is internal energy, $c_v$ is heat capacity
at constant volume, and $T$ is gas temperature. Due to the IMC time
linearization, energy deposited from gamma-rays and (locally) from
beta particles appears simultaneously in both the comoving internal
energy equation and the radiation equation.
Ionization and excitation are treated with Saha-Boltzmann statistics
evaluated at the gas density and temperature, both at the beginning of
the time step. The multi-element system is solved iteratively by
converging the free electron number. The resulting population
densities are then used to calculate opacities. The radiation field
is represented by fully time-dependent Monte Carlo packets, which are
sourced from LTE emissivity. The radiation field is not constrained
to be in equilibrium with the gas, so in general the system is
``two-temperature''.
Opacity is discretized into groups via direct integration over
co-moving wavelength (but see \citealt{Fontes2020} for a study with
SuperNu using expansion opacity). Opacity in SuperNu includes
free-free \citep{Sutherland1998}, and bound-free
\citep{Verner1995,Verner1996} processes, as well as the bound-bound
opacities from the Kurucz line
lists\footnote{\url{http://kurucz.harvard.edu/atoms.html}}, and the
standard elastic Thomson scattering opacity. Weak lines in the Kurucz
data set are omitted from the SuperNu line list where the opacity is
dominated by stronger lines. The total number of available lines for
the present simulations is 591\,288. This list was motivated by
studies using the PHOENIX code in the work of \cite{vanRossum2012},
using the full line list (>10$^{7}$ lines) as a benchmark. For the
simulations in this work, opacity is computed in 1000 logarithmic
wavelength groups from 100 to 32000\,\AA.
Each simulation presented for SuperNu uses 4,194,304 source packets
per time step, with maximum active packet populations between 100 and
200 million. The wavelength group structure that the packets are
tracked through is the same as the opacity group structure (though for
the simulations here, the wavelength bounds for the flux tallies are
1000 and 32,000\,\AA). The DDC and toy model groups use 115 and 202
velocity space cells, respectively. For the time grid, the DDC and
toy model groups use 200 logarithmically increasing time steps out to
day 80, from about day 1 (DDC models) or day 2 (toy models)
post-explosion.
SuperNu has MPI+OpenMP parallelization. The SuperNu simulations of
the toy01, toy06, DDC10, and DDC25 models presented here cost 190,
320, 398, and 400 core-hours, respectively, each using 16 MPI ranks
and 8 OpenMP threads per rank.
\subsection{TARDIS}\label{sect:tardis}
TARDIS is an open-source steady-state 1D radiative transfer code that
uses indivisible energy packets as its transport quanta following the
methods in
\citet{Abbot1985,lucy1993,Mazzali1993,Lucy2002,Lucy2003,Lucy2005}. \citet{kerzendorf2014}
describes the initial version of the code which was primarily used to
model SNe Ia. Subsequently, the code has been significantly enhanced
to include a non-thermal approximation treatment for helium
\citep{boyle2017} and the continuum processes and relativity
treatments required for hydrogen-rich SNe \citep{vogl2019}. TARDIS
has been continuously enhanced since then \citep[see e.g. ][similarly
gamma-ray energy deposition is a new module of TARDIS, but has not
been used for the models in this work]{kerzendorf2022}. Full
documentation and an extended physics walk-through for TARDIS can be
found online\footnote{\url{https://tardis-sn.github.io/tardis}}.
\textsc{tardis} assumes a steady-state homologously expanding
supernova envelope and injects Monte Carlo packets, randomly sampled
from a given distribution (by default a black-body), from an inner
boundary. The code supports bound-bound, bound-free, free-free, and
Thomson opacities with several redistribution schemes from simple
scattering to a macro-atom \citep{Lucy2002, Lucy2003}. Summary packet
statistics are used to estimate radiation field quantities
(temperature, dilution factors, mean intensities, heating- and
photo-ionization rates, line source functions). The estimated
quantities are used to calculate the ionization/excitation populations
in steady state with a choice between LTE, and several formulations of
non-LTE (nebular approximation of \citealt{Abbot1985} for ionization,
dilute-LTE and full non-LTE for excitation). TARDIS then calculates
Sobolev optical depths for line interaction, and opacities for
continuum processes. These values are then used in the subsequent
Monte Carlo step, which produces summary statistics for updating the
opacities. The other convergence criterion is a match of the
integrated packet output luminosity and the requested luminosity,
which is achieved through iterative adjustments of the inner
temperature.
To handle atomic data, the TARDIS collaboration has developed an
additional package named \textsc{Carsus} \citep[available at
\url{https://github.com/tardis-sn/carsus};][]{Passaro2019_4062427}
with documentation available
online \footnote{\url{https://tardis-sn.github.io/carsus}}.
\textsc{Carsus} can read atomic data (masses, ionization energies,
levels, lines, photoionization cross sections, and collisional cross
sections) from NIST \citep{kramida_nist_1999}, Chianti
\citep{Dere_2019, Dere_1997}, CMFGEN \citep{Hillier1998_cmfgen,
Dessart2010}, and Kurucz \citep{Kurucz2009_ATD}. For the models in
this paper, we used Kurucz CD 23 as the source of atomic data (see
Appendix~\ref{sect:atomdata_tardis} for the full description).
The setup files for TARDIS that are used in this work are available
online\footnote{\url{https://github.com/tardis-sn/tardis-setups/tree/master/2022/sn_radtrans_compare}}
using an atomic data set from Kurucz CD 23. In the setup used for the
code comparison work, we run TARDIS in a mode that self-consistently
finds a temperature stratification given an inner boundary velocity
and output luminosity. We used the mean bolometric luminosity of the
other comparison codes for our output luminosity. The structure of the
toy01 and toy06 models was set at 520 shells between 9000 and 35000 km
s$^{-1}$. The DDC models were truncated to 40 shells between 9000 and
35000 km s$^{-1}$. The inner boundary velocity is found by iterating
until the dilution factor is close to 0.5 in the innermost zone.
Table~\ref{tab:tardis_lum_vel} shows the inner boundary velocity and
requested output luminosity for the different epochs.
\begin{table}
\footnotesize
\caption{Velocity and temperature at the inner boundary given the requested output luminosity for the TARDIS calculation of the toy06 model.}
\label{tab:tardis_lum_vel}
\begin{tabular}{cccc}
\hline
\multicolumn{1}{c}{Time} & $\log_{10} L_\mathrm{req}$ & \multicolumn{1}{c}{Velocity at IB} & \multicolumn{1}{c}{Temperature at IB} \\
\multicolumn{1}{c}{(days)} & (erg\,s$^{-1}$) & \multicolumn{1}{c}{(\ensuremath{\rm{km\,s}^{-1}})} & \multicolumn{1}{c}{(K)} \\
\hline
5 & 42.48 & 20500 & 8805 \\
10 & 42.95 & 17000 & 8870 \\
15 & 43.04 & 10000 & 11525 \\
20 & 43.00 & 5500 & 15513 \\
\hline
\end{tabular}
\flushleft
\textbf{Notes:} $L_\mathrm{req}$ = requested output luminosity; IB = inner boundary.
\end{table}
We used the \texttt{nebular} approximation for ionization and the
\texttt{dilute-lte} approximation for excitation. We used
$5\times10^5$ packets for estimating our radiation field and 30
iterations for convergence. In the final iteration, we estimated the
source functions with $5\times10^5$ packets and then used the formal
integral to synthesize a spectrum. For the comparison, we use line
opacities and Thomson opacities with the \texttt{macroatom}
interaction scheme. The resolution of the output spectra was uniform
from 500 to 20000 \AA\ with 2 \AA\ bin width. The models were run on
one CPU with runtimes less than one hour.
\subsection{URILIGHT}\label{sect:urilight}
URILIGHT is a time dependent Monte-Carlo code written in Fortran 90 by
Yoni Elbaz based on the approximations that are used in SEDONA (see
Sect.~\ref{sect:sedona} above, \citealt{Kasen2006} and references
therein), in particular assuming homologous expansion. A detailed
description of this program and previous comparisons to other
published radiative-transfer codes for several benchmark problems are
presented in \cite{Wygoda2019}. The code is publicly available and can
be downloaded from
\url{https://www.dropbox.com/sh/kyg1z1xwi0298ru/AAAqzUMbr6AkoVfkSVIYChTLa?dl=0}.
Energy deposition resulting from the decay of radioactive isotopes is
calculated by a Monte-Carlo solution of the $\gamma$-ray transport,
for which interaction with matter is included through Compton
scattering and photoelectric absorption. For the calculations in this
work, only $^{56}$Ni and $^{56}$Co decay was included.
The temperature is iteratively solved for in each cell by requiring
that the total emissivity equals the total absorbed energy.
LTE is assumed for calculating the ionization and excitation states:
ionization is obtained by solving the Saha equation, and excitation
levels are set by the Boltzmann-distribution.
Opacities include bound-bound and free-free absorption and Thomson
scattering off free electrons. The atomic line data for the
bound-bound transitions, which constitutes the main and most important
source of opacity in SNe~Ia, are taken from the extended set of lines
by Kurucz\footnote{CDs 1 and 23 from
\url{http://kurucz.harvard.edu/cdroms.html}}. Following a bound-bound
interaction, most photons (a fraction $\epsilon$, which is a global
parameter of the simulation; see \citealt{Kasen2006}) are thermalized
and re-emitted at a different wavelength with a distribution set by
the emissivity. The rest of the photons (fraction $1-\epsilon$) are
re-emitted at the same wavelength (within the line width). As in
\cite{Wygoda2019}, we have used $\epsilon=0.8$ in the runs performed
here.
The runs here use 162 spatial cells for the toy01 and toy06 models
and 115 spatial cells for the DDC10 and DDC25 models. All models are
run with 128 time steps logarithmically spaced between 2 and 210
days, and a uniform spectrum resolution of 10\,\AA\ between 100 and
30000\,\AA. Each run typically took 30 hours on a single core, or
approximately one hour when parallelized.
\begin{table*}
\footnotesize
\caption{Code outputs and computed models.}\label{tab:codedescr}
\begin{tabular}{lcccccccccccc}
\hline
Code & RT & Bolometric & Spectrum & Early & Nebular & \multicolumn{4}{c}{Computed Models} \\
Name & Method & Flux & or SED & Times & Times & toy06 & toy01 & DDC10 & DDC25 \\
\hline
ARTIS & MC & calculated & \cmark & \cmark & \cmark & \cmark & \xmark & \cmark & \xmark \\
CMFGEN & RTE-CMF & calculated & \cmark & \cmark & \cmark & \cmark & \cmark & \cmark & \cmark \\
CRAB & RH-1G & calculated & \xmark & \cmark & \xmark & \cmark & \xmark & \xmark & \xmark \\
KEPLER & FLD & calculated & \xmark & \cmark & \xmark & \cmark & \cmark & \cmark & \cmark \\
SEDONA & MC & calculated & \cmark & \cmark & \xmark & \cmark & \cmark & \cmark & \xmark \\
STELLA & RH-MG & calculated & \cmark & \cmark & \xmark & \cmark & \cmark & \cmark & \cmark \\
SUMO & MC & calculated & \cmark & \xmark & \cmark & \cmark & \cmark & \xmark & \xmark \\
SuperNu & MC & calculated & \cmark & \cmark & \xmark & \cmark & \cmark & \cmark & \cmark \\
TARDIS & MC & input & \cmark & \cmark & \xmark & \cmark & \cmark & \cmark & \cmark \\
URILIGHT & MC & calculated & \cmark & \cmark & \xmark & \cmark & \cmark & \cmark & \cmark \\
\hline
\end{tabular}
\flushleft
\textbf{Notes:} RT Method gives the numerical method used to solve the radiative-transfer equation (see Table~\ref{tab:codephys}).
\end{table*}
\section{Data repository of test models and standardized outputs}\label{sect:outputs}
The ejecta models and output files from the RT simulations of the
different codes are provided in a new data repository, which is
publicly available and can be accessed at
\url{https://github.com/sn-rad-trans/data1}.
Descriptions of the files that are available in the repository are
provided below, including ejecta model files
(Sect.~\ref{sect:outputs:model_formats}), output files
(Sect.~\ref{sect:outputs:result_formats}) and python codes that were
used to create the analytic toy model ejecta and codes for reading the
output files (Sect.~\ref{sect:outputs:codes}).
\subsection{Ejecta model files}\label{sect:outputs:model_formats}
As described in Sect.~\ref{sect:models}, the code comparison is
performed using four SN Ia models (main parameters in Table
\ref{tab:models}). The RT input files that were distributed among the
groups are provided in the repository, including two toy model files
\texttt{snia\_toy01.dat} and \texttt{snia\_toy06.dat} and the two
delayed-detonation model files \texttt{DDC10\_0.976d} and
\texttt{DDC25\_1.300d}.
\subsubsection{Toy model files}
The two toy model files, \texttt{snia\_toy01.dat} and
\texttt{snia\_toy06.dat}, represent a snapshot of the ejecta at 2.0
days since explosion and include 807 shells (rows) with the following
21 columns:
\begin{enumerate}[label=(\arabic*), align=left]
\item Shell index (1-807)
\item Velocity at shell centre (\ensuremath{\rm{km\,s}^{-1}})
\item Shell mass (\ensuremath{\rm{M}_{\odot}})
\item Lagrangian mass coordinate at the outer shell boundary (\ensuremath{\rm{M}_{\odot}})
\item Pre-decayed ($t=0$) stable IGE mass fraction (0 for all shells)
\item Pre-decayed ($t=0$) $^{56}$Ni mass fraction
\item IME mass fraction (of which 10\% is Ca, 35\% is S and 55\% is Si by mass)
\item Ti mass fraction (0 for all shells)
\item Unburnt C+O mass fraction (0 for all shells)
\item Radius at shell centre (cm) = velocity at shell centre $\times$ 2 days (homologous expansion)
\item Mean density over shell (\ensuremath{\rm{g\,cm}^{-3}}), \textit{not} density at shell centre
\item Temperature (K)
\item[(13)-(21)] Mass fractions of $^{56}$Ni, $^{56}$Co, $^{56}$Fe, Ca, S, Si, O, and C at 2 days since explosion
\end{enumerate}
\subsubsection{Delayed-detonation model files}
The two delayed-detonation model files \texttt{DDC10\_0.976d} and
\texttt{DDC25\_1.300d} represent a snapshot of the ejecta at 0.976
days and 1.3 days, respectively, with 115 shells with the following 50
columns:
\begin{enumerate}[label=(\arabic*), align=left]
\item Velocity at shell centre (\ensuremath{\rm{km\,s}^{-1}})
\item Radius at shell centre (cm) = velocity at shell centre
$\times$ the time since explosion
\item Shell volume (cm$^3$)
\item Density at zone centre (\ensuremath{\rm{g\,cm}^{-3}}), \textit{not} mean density
over shell
\item Shell mass estimate (g) = shell volume $\times$ density at
zone centre
\item Temperature (K)
\item[(7)-(26)] Elemental mass fractions at snapshot time of C, O,
Ne, Na, Mg, Al, Si, S, Cl, Ar, K, Ca, Sc, Ti, V, Cr, Mn, Fe, Co,
and Ni
\item[(27)-(50)] isotopic mass fractions of the following
radioactive nuclei at snapshot time: $^{56}$Ni, $^{56}$Co,
$^{57}$Ni, $^{57}$Co, $^{48}$Cr, $^{48}$V, $^{49}$Cr, $^{49}$V,
$^{51}$Mn, $^{51}$Cr, $^{55}$Co, $^{55}$Fe, $^{37}$K, $^{37}$Ar,
$^{52}$Fe, $^{52}$Mn, $^{44}$Ti, $^{44}$Sc, $^{41}$Ar,
$^{42}$Ar, $^{42}$K, $^{43}$K, $^{47}$Sc, and $^{61}$Co
\end{enumerate}
\subsection{RT output files}\label{sect:outputs:result_formats}
For each simulation by a specific group, six output file types are
generated for each of the ejecta models.
\subsubsection{Output file names}\label{sect:outputs:output_file_names}
The name of each file has the following structure:
\begin{verbatim}
<output type>_<model>_<code name>.txt
\end{verbatim}
\noindent
where
\begin{itemize}
\item \verb+<output type>+ represents the type of output
(described below) and can take one of six values:
\texttt{lbol\_edep}, \texttt{edep}, \texttt{phys},
\texttt{ionfrac\_<element>} (where \texttt{<element>} is the
name of a given element, e.g. \texttt{ca} for calcium),
\texttt{spectra}, and \texttt{wsynphot\_mags},\\
\item \verb+<model>+ represents one of the models and can take one
of four values: \texttt{toy06}, \texttt{toy01}, \texttt{ddc10},
and \texttt{ddc25},\\
\item \verb+<code name>+ represents the code name with an optional
additional descriptor (useful to distinguish between different
code settings when applied to a given model) and can take one of
12 values: \texttt{artis}, \texttt{artisnebular},
\texttt{cmfgen}, \texttt{crab}, \texttt{kepler},
\texttt{sedona}, \texttt{stella}, \texttt{stella\_fr600} (for
the STELLA runs for the toy06 models that use 629 frequency bins
instead of the default 129), \texttt{supernu}, \texttt{sumo},
\texttt{tardis}, and \texttt{urilight}.
\end{itemize}
In principle there are 12 code names $\times$ 4 ejecta models $\times$
5 output files (excluding the \texttt{ionfrac\_<element>} files) = 240
files and an additional 12 code names $\times$ (2 toy models $\times$
6 elements + 2 DDC models $\times$ 20 elements) = 624
\texttt{ionfrac\_<element>} files. This results in a total of 864
files, while in practice not all files are available for various
reasons: a) a code was not applied to a given model (it was agreed
that all groups should at least compute the toy06 model, but the other
three were considered optional); b) a given code cannot produce the
specified output (e.g. SEDs for grey codes); c) a given code does not
provide the desired output quantities by default (i.e. modification of
the source code would be necessary). Table~\ref{tab:codedescr}
summarizes the outputs and computed models for each code.
\subsubsection{Six output file types}\label{sect:outputs:output_file_types}
The six types of output files include:
\begin{enumerate}
\item \textbf{Pseudo-bolometric (UVOIR) luminosity and global energy deposition as function of time}\\
\textbf{File name:} \verb+lbol_edep_<model>_<code name>.txt+\\
\textbf{Header}: after some optional comment lines, the following
two lines give the number of epochs (\texttt{NTIMES}, 100 in this
example) and the column headings:
\begin{verbatim}
#<optional comment lines>
#NTIMES: 100
#time[d] Lbol[erg/s] Edep[erg/s]
\end{verbatim}
\vspace*{-1em}
\textbf{Contents}: \texttt{NTIMES} rows with the following three columns:
\begin{enumerate}[label=(\arabic*), align=left]
\item Time since explosion in days
\item Pseudo-bolometric (UVOIR) luminosity in erg s$^{-1}$
\item Global energy deposition by $\gamma$ rays and positrons in erg s$^{-1}$
\end{enumerate}
\vspace*{1em}
\item \textbf{Energy deposition}\\
\textbf{File name:} \verb+edep_<model>_<code name>.txt+\\
\textbf{Header}: after some optional comment lines, the following
four lines give the number of epochs (\texttt{NTIMES}, 100 in this
example), the number of cells (\texttt{NVEL}, 200 in this
example), the list of epochs in days with all \texttt{NTIMES}
values (the `...' should correspond to actual values), and finally
the column headings (here the `...' can be used as is):
\begin{verbatim}
#<optional comment lines>
#NTIMES: 100
#NVEL: 200
#TIMES[d]: 2.0 3.0 ... 100.0
#vel_mid[km/s] Edep_t0[erg/s/cm^3]
Edep_t1[erg/s/cm^3] ... Edep_tn[erg/s/cm^3]
\end{verbatim}
\vspace*{-1em}
\textbf{Contents}: \texttt{NVEL} rows with the following \texttt{NTIMES}+1 columns (101 in this example):
\begin{enumerate}[label=(\arabic*), align=left]
\item Velocity at the centre of each cell in \ensuremath{\rm{km\,s}^{-1}}
\item[(2)-(101)] Total $\gamma$-ray + positron energy deposition
rate at each of the \texttt{NTIMES} epochs within the
corresponding cell in erg s$^{-1}$ cm$^{-3}$
\end{enumerate}
\vspace*{1em}
\item \textbf{Physical conditions}\\
\textbf{File name:} \verb+phys_<model>_<code name>.txt+\\
\textbf{Header}: after some optional comment lines, the following
three lines give the number of epochs (\texttt{NTIMES}, 100 in
this example), the list of epochs with all \texttt{NTIMES} values
(the `...' should correspond to actual values), and finally one
empty comment line:
\begin{verbatim}
#<optional comment lines>
#NTIMES: 100
#TIMES[d]: 2.0 3.0 ... 100.0
#
\end{verbatim}
\vspace*{-1em}
\textbf{Contents}: \texttt{NTIMES} blocks (one for each epoch),
each containing a block header with three lines giving the epoch
(in days, 2.0\,d in this example), the number of cells saved for
this epoch (\texttt{NVEL}, 200 in this example), and finally the
column headings:
\begin{verbatim}
#TIME: 2.0
#NVEL: 200
#vel_mid[km/s] temp[K] rho[gcc] ne[/cm^3]
natom[/cm^3]
\end{verbatim}
\vspace*{-1em}
Each block content consists of \texttt{NVEL} rows with the following five columns:
\begin{enumerate}[label=(\arabic*), align=left]
\item Velocity at the centre of each cell in \ensuremath{\rm{km\,s}^{-1}}
\item Temperature in K
\item Density in \ensuremath{\rm{g\,cm}^{-3}}
\item Free electron density in cm$^{-3}$
\item Total atom density in cm$^{-3}$
\end{enumerate}
\vspace*{1em}
\item \textbf{Ionization fraction (one file per element)}\\
\textbf{File name:}\\
\verb+ionfrac_<element>_<model>_<code name>.txt+\\
\textbf{Header}: after some optional comment lines, the following
four lines give the number of epochs (\texttt{NTIMES}, 100 in this
example), the number of ionization stages (\texttt{NSTAGES}, 6 in
this example, starting at neutral and up to $\texttt{NSTAGES}-1$
times ionized), the list of epochs with all \texttt{NTIMES} values
(the `...' should correspond to actual values), and finally one
empty comment line:
\begin{verbatim}
#<optional comment lines>
#NTIMES: 100
#NSTAGES: 6
#TIMES[d]: 2.0 3.0 ... 100.0
#
\end{verbatim}
\vspace*{-1em}
\textbf{Contents}: \texttt{NTIMES} blocks (one for each epoch),
each containing a block header with three lines giving the epoch
(in days, 2.0\,d in this example), the number of cells saved for
this epoch (\texttt{NVEL}, 200 in this example), and finally the
column headings (where we consider the element Fe in this
example):
\begin{verbatim}
#TIME: 2.0
#NVEL: 200
#vel_mid[km/s] fe0 fe1 fe2 fe3 fe4 fe5
\end{verbatim}
\vspace*{-1em}
Each block content consists of \texttt{NVEL} rows with the
following \texttt{NSTAGES}+1 columns (7 in this example):
\begin{enumerate}[label=(\arabic*), align=left]
\item Velocity at the centre of each cell in \ensuremath{\rm{km\,s}^{-1}}
\item[(2)-(7)] Fraction of ions (dimensionless) in the
corresponding cell (\texttt{fe0} = Fe\ts {\sc i}, \texttt{fe1} =
Fe\ts {\sc ii}\ etc.). The sum of the fractions in each of the
\texttt{NVEL} rows is expected to be unity.
\end{enumerate}
\vspace*{1em}
\item \textbf{Spectral sequence}\\
\textbf{File name:} \verb+spectra_<model>_<code name>.txt+\\
\textbf{Header}: after some optional comment lines, the following
four lines give the number of epochs (\texttt{NTIMES}, 100 in this
example), the number of wavelengths (\texttt{NWAVE}, 2000 in this
example), the list of epochs in days with all \texttt{NTIMES}
values (the `...' should correspond to actual values), and finally
the column headings (here the `...' can be used as is):
\begin{verbatim}
#<optional comment lines>
#NTIMES: 100
#NWAVE: 2000
#TIMES[d]: 2.0 3.0 ... 100.0
#wavelength[Ang] flux_t0[erg/s/Ang]
flux_t1[erg/s/Ang] ... flux_tn[erg/s/Ang]
\end{verbatim}
\vspace*{-1em}
\textbf{Contents}: \texttt{NWAVE} rows with the following
\texttt{NTIMES}+1 columns (101 in this example):
\begin{enumerate}[label=(\arabic*), align=left]
\item Wavelength in \AA
\item[(2)-(101)] Fluxes at each of the \texttt{NTIMES} epochs at
the corresponding wavelength in erg s$^{-1}$ \AA$^{-1}$
\end{enumerate}
\vspace*{1em}
\item \textbf{Synthetic photometry}\\
\textbf{File name:} \verb+wsynphot_mags_<model>_<code name>.txt+\\
\textbf{Header}: after some optional comment lines, the following
three lines give the number of epochs (\texttt{NTIMES}, 100 in
this example), the number of photometric bands (\texttt{NBANDS}, 8
in this example), and finally the column headings (giving each
band name, $UBVRIJHK$ in our example):
\begin{verbatim}
#<optional comment lines>
#NTIMES: 100
#NBANDS: 8
#time[d] U B V R I J H K
\end{verbatim}
\vspace*{-1em}
\textbf{Contents}: \texttt{NTIMES} rows with the following \texttt{NBANDS}+1 columns (9 in our example):
\begin{enumerate}[label=(\arabic*), align=left]
\item Time since explosion in days
\item[(2)-(9)] Absolute magnitudes in each band at the corresponding time
\end{enumerate}
\end{enumerate}
\subsection{Useful python codes}\label{sect:outputs:codes}
All the figures presented in this paper are automatically generated
and accessible in a python notebook \verb+all_plots.ipynb+ which is
available in the data repository. A separate python notebook
\verb+photometry.ipynb+ in the \verb+code-comparison1/+ directory is
used to generate the synthetic photometry (\verb+wsynphot_mags+ files)
from the \verb+spectra+ files on the fly. Three further useful python
codes can be used to:
\begin{enumerate}
\item \textbf{Generate the toy models}\\
\textbf{Location:}\\ \verb+data1/input_models/mk_snia_toy_model.py+\\
\textbf{Calling syntax:}
\begin{verbatim}
python mk_snia_toy_model.py --highni
python mk_snia_toy_model.py --lowni
\end{verbatim}
\vspace*{-1em}
This code generates the toy models used in this paper, as
explained in Sect.~\ref{sect:toymodels}. The upper command
generates the toy06 model (\verb+snia_toy06.dat+), while the lower
line generates the toy01 model (\verb+snia_toy01.dat+). Comments
inside the code provide information on how to create new similar
test models with different parameters.
\vspace*{1em}
\item \textbf{Read the input files}\\
\textbf{Location:} \verb+data1/input_models/read_inputs.py+\\
\textbf{Calling syntax:}
\begin{verbatim}
python read_inputs.py
\end{verbatim}
\vspace*{-1em}
This code reads the input files of the test models and creates the
following files corresponding to Figs.~\ref{fig:dens} and
\ref{fig:composition} as well as Table~\ref{tab:models}:
\verb+density_profile.pdf+ (Fig.~\ref{fig:dens}),
\verb+composition_profile_<model>.pdf+ (where \texttt{<model>} is
one of \texttt{toy06}, \texttt{toy01}, \texttt{ddc10}, and
\texttt{ddc25}; Fig.~\ref{fig:composition}), and
\texttt{models\_summary.tex} (Table~\ref{tab:models}). The code
also includes the functions \texttt{read\_snia\_toy\_model()} and
\texttt{read\_ddc\_model()} to read the files into python
variables.
By default the code is expected to be executed from within the
\verb+data1/input_models/+ directory, although the path to the
\texttt{data1/} directory can be specified using the
\verb+--path2data+ option.
\vspace*{1em}
\item \textbf{Read the output files}\\
\textbf{Location:} \verb+data1/read_outputs.py+\\
\textbf{Calling syntax:}
\begin{verbatim}
python read_outputs.py file1.txt file2.txt
python read_outputs.py /path/to/file*.txt
\end{verbatim}
\vspace*{-1em}
This code reads the six output file types (see
\ref{sect:outputs:output_file_types}) and produces corresponding
plots. The code also includes the functions:
\texttt{read\_lbol\_edep()}, \texttt{read\_spectra()},
\texttt{read\_edep()}, \texttt{read\_phys()},
\texttt{read\_ionfrac()}, and \texttt{read\_mags()} to read the
output files into python variables.
The code can be executed from any directory since the full path of
each file can be specified (wildcards are also accepted). When
uploading new output files to the data repository users are
required to ensure they match the expected format exactly. This
can be achieved using the \verb+--checkformat+ option (and
optionally the \verb+--noplot+ option to disable the plotting
functionality when checking a large number of files).
\end{enumerate}
\section{Example results}\label{sect:results}
In this section we provide example results that are extracted from the
outputs of the different simulations. The purpose of this section is
to illustrate the contents of the data repository and comparisons that
can be made using it. No comparison to observations or in-depth
investigations of the sources of differences in results are made
here. We note that while some groups provided a few sets of results
for different physical approximations that appear in the repository,
only one set of results is shown for each code in this set of
examples. One exception is ARTIS, for which both the regular
(\texttt{artis}) and nebular (\texttt{artisnebular}) versions are
included for each model, in order to present results for this code
from early to late times.
The results are provided in the following subsections. In
Sect. \ref{sect:bol}, bolometric properties of the emission are shown
for simulations of the toy06 model, including the pseudo-bolometric
(UVOIR) light curve and the energy deposition rate as a function of
time. Rise times and peak luminosities are provided for the toy06 and
toy01 models. In Sect. \ref{sect:tion}, the resulting profiles of
temperature and mean ionization state of cobalt are shown at selected
times. In Sect. \ref{sect:multiband}, multi-band light curves and
colour curves are shown for model toy06. $B$-band rise times are
provided for all models where available. Example spectra at selected
early and late times are shown in Sect. \ref{sect:spec}. Throughout
the section each figure provides a legend with the colours or symbols
associated with each code. A uniform coding scheme is used, such that
a given code can be systematically identified in all the figures.
\subsection{Bolometric (UVOIR) evolution}\label{sect:bol}
\begin{figure*}
\centering
\includegraphics[width=0.85\hsize]{fig3.pdf}
\caption{Pseudo-bolometric (UVOIR) light curves for model
toy06. The inset zooms in on the maximum-light epoch (the
estimated time of maximum light is indicated with a `+' sign).
}
\label{fig:lbol}
\end{figure*}
\begin{figure}
\includegraphics[width=\hsize]{fig4.pdf}
\caption{Peak pseudo-bolometric (UVOIR) luminosities vs. rise times
(from explosion to peak) for all four test models. The data points
corresponding to the CMFGEN, CRAB, and SuperNu calculations of the
toy06 model are difficult to distinguish.
}
\label{fig:bolpeak}
\end{figure}
In this subsection bolometric properties are shown based on the
\verb+lbol_edep+ output files. In Fig.~\ref{fig:lbol}, the UVOIR light
curves for model toy06 are shown for the first 100 days after
explosion for the different codes. The rise times and peak
luminosities are shown in Fig.~\ref{fig:bolpeak} for all four test
models. The peak luminosities were obtained by a parabolic fit to
light curve around maximum light.
In Fig.~\ref{fig:edep}, the total energy deposition from $\gamma$ rays
and positrons based on the same files is shown as a function of time
for the first 200 days after explosion for the toy06 model. In order
to highlight the differences among codes, all results are normalized
to the same analytic approximation for the deposition:
\begin{align}\label{eq:Qdepnorm}
Q_{\rm dep, norm}\equiv&\frac{M_{\rm Ni}}{M_{\odot}}\left(0.97\left[1 - e^{-(40d/t)^2}\right] + 0.03\right)\times\cr
&\left(6.45~e^{-t/8.8d}~+~1.45~e^{-t/111.3d}~\right)~10^{43}~\rm erg~s^{-1}
\end{align}
\noindent
where $M_{\rm Ni} = 0.6$\,\ensuremath{\rm{M}_{\odot}}\ is the total (undecayed) $^{56}$Ni
mass in the toy06 model\footnote{The prefactor in parenthesis in the
first line of Eq. \eqref{eq:Qdepnorm}, is equivalent to a simplistic
approximation in which a fraction of 0.03 of the energy is emitted in
positron kinetic energy which is immediately deposited in the ejecta
while the rest is emitted in gamma-rays with a purely absorptive
optical depth given by $\tau_{\gamma}=t_0^2/t^2$ with a gamma-ray
escape time \citep[e.g.][]{Jefferey1999} of $t_0=40$d.}.
\begin{figure}
\includegraphics[width=\hsize]{fig5.pdf}
\caption{Total energy deposition rate from $\gamma$ rays and
positrons for model toy06, normalized to the analytic function
given by Eq. \ref{eq:Qdepnorm} (the normalization allows
differences to be seen more clearly).
}
\label{fig:edep}
\end{figure}
In Fig.~\ref{fig:bolratios}, two physical diagnostics of the relation
between the energy deposition and the bolometric light curves are
shown for the toy06 model (both expected to approach unity at late
times): the instantaneous ratio of the energy deposition rate and the
luminosity (left) and the ratio of cumulative time-weighted integrals
of the energy deposition rate and luminosity (right).
\begin{figure*}
\includegraphics[width=0.5\hsize]{fig6a.pdf}
\includegraphics[width=0.5\hsize]{fig6b.pdf}
\caption{
Left: Instantaneous ratio of the luminosity and the total energy
deposition rate for model toy06.
Right: Ratio of the time-weighted
integrals of the luminosity and energy deposition rate for model
toy06 (here we do not show the results for the SUMO code since the
calculation starts at 75\,d past explosion). Both ratios are
expected to reach unity at late times.
}
\label{fig:bolratios}
\end{figure*}
\subsection{Temperature and ionization profiles}\label{sect:tion}
In Fig.~\ref{fig:temp_and_ionization} the thermodynamic structure of
the ejecta as a function of velocity is shown based on the \verb+phys+
and \verb+ionfrac_co+ output files for model toy06. Three different
times of particular interest are shown: 15 days past explosion (upper
panels), close to the peak of the light curve, 40 days past explosion
(middle panels), close to the break in the $B-V$ colour curve, and 200
days past explosion, during the nebular phase (lower panels). For each
of these times, two profiles are shown, the (gas) temperature profile (left
panels), and the mean ionization level of Co, defined as
$\sum\nolimits_{i=0} i\cdot f_i$, where $f_i$ is the fraction of Co
ions ionized $i$ times (right panels). Co and Fe dominate the opacity
at these times and the ionization profiles of Fe are very similar to
those of Co (not shown here).
\begin{figure*}
\includegraphics[width=0.5\hsize]{fig7a.pdf}
\includegraphics[width=0.5\hsize]{fig7b.pdf}
\includegraphics[width=0.5\hsize]{fig7c.pdf}
\includegraphics[width=0.5\hsize]{fig7d.pdf}
\includegraphics[width=0.5\hsize]{fig7e.pdf}
\includegraphics[width=0.5\hsize]{fig7f.pdf}
\caption{Ejecta (gas) temperature and ionization profiles as a
function of velocity for the toy06 model. Left column: Temperature
at 15 days (top), 40 days (middle), and 200 days (bottom) past
explosion. Right column: Mean ionization of cobalt at 15 days
(top), 40 days (middle), and 200 days (bottom) past explosion. We
restrict the abscissa range of the ionization plots to $v \le
12000$\,\ensuremath{\rm{km\,s}^{-1}}\ since the Co abundance drops to 0 at larger
velocities.
}
\label{fig:temp_and_ionization}
\end{figure*}
\subsection{Multi-band light curves and colours}\label{sect:multiband}
In this section, multi-band properties are shown for the different
codes for different models. The photometry is extracted from the
spectra reported in the \verb+spectra+ output files using the
\texttt{wsynphot}
package\footnote{\url{https://github.com/starkit/wsynphot}} (using the
Vega calibration spectrum
\verb+alpha_lyr_stis_003.fits+\footnote{\url{https://archive.stsci.edu/hlsps/reference-atlases/cdbs/calspec/alpha_lyr_stis_003.fits}}
and a third-order spline interpolation). In Fig.~\ref{fig:lcfilt}, the
$UBVRIJHK$ light curves are shown for model toy06 up to 100 days after
explosion. The $B$-band rise times (from explosion to peak) are
extracted by fitting a high-order polynomial around maximum light and
shown for all models in Fig.~\ref{fig:trise}. The $B-V$ and $V-R$
colour curves for model toy06 are shown in Fig.~\ref{fig:color}.
\begin{figure*}
\includegraphics[width=\hsize]{fig8.pdf}
\caption{Multi-band ($UBVRIJHK$) light curves for the toy06
model. Note the different ordinate range for the $U$ band.
}
\label{fig:lcfilt}
\end{figure*}
\begin{figure}
\includegraphics[width=\hsize]{fig9.pdf}
\caption{Peak magnitude vs. rise time in the $B$ band for all four test models.
}
\label{fig:trise}
\end{figure}
\begin{figure*}
\includegraphics[width=0.5\hsize]{fig10a.pdf}
\includegraphics[width=0.5\hsize]{fig10b.pdf}
\caption{$B-V$ and $V-R$ colour evolution for the toy06 model.
}
\label{fig:color}
\end{figure*}
\subsection{Spectroscopic evolution}\label{sect:spec}
Spectra obtained for the toy06 model at selected times are shown in
Figs.~\ref{fig:earlyspec} and \ref{fig:latespec} based on the
\verb+spectra+ output files. The wavelength range is restricted to
3000-10000\,\AA\ to allow direct comparison. This does not represent
the full range available in the output files, which differs between
different codes. In Fig.~\ref{fig:earlyspec}, early-time spectra (5
days after explosion and around peak luminosity) are shown while in
Fig.~\ref{fig:latespec} we show spectra at later times (50 days and
200 days after explosion). Note that nebular times require specialized
treatment in non-LTE which is only implemented in the ARTIS, CMFGEN,
and SUMO codes for this paper\footnote{SEDONA also has non-LTE
capabilities but these were not used here.}.
\begin{figure*}
\includegraphics[width=0.5\hsize]{fig11a.pdf}
\includegraphics[width=0.5\hsize]{fig11b.pdf}
\caption{Spectra at 5\,d past explosion and at the time of peak UVOIR
luminosity (the caption indicates the time of the spectrum computed
closest to peak) for the toy06 model.
}
\label{fig:earlyspec}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.5\hsize]{fig12a.pdf}
\includegraphics[width=0.5\hsize]{fig12b.pdf}
\caption{Spectra at 50\,d and 200\,d past explosion for the toy06
model. The colour coding for the 50\,d spectra is the same as in
Fig.~\ref{fig:earlyspec} and is not repeated here for sake of
clarity.
}
\label{fig:latespec}
\end{figure*}
\section{Discussion and outlook}\label{sect:conclusions}
In this paper we present an online public data repository that
includes test models for radiative-transfer (RT) calculations of
emission from supernovae and standardized simulation outputs by ten
different groups that allows direct comparison. Python scripts that
generate the analytic toy models and read the different formats are
provided as well.
The main purpose of the repository is to allow studies of the
different physical approximations involved in the different codes and
to assess the robustness of different predictions of radiation
transfer. In addition, the repository can be used for finding bugs in
the codes or provide checks in the development of new codes.
We plan to extend the set of test models to include other supernova
types, as well as multi-dimensional models. In addition, we intend to
produce more specialised test cases for which exact solutions can be
found and agreed upon, which can provide benchmarks for RT code
development. We also aim to include models with standardized atomic
data sets that will allow the study of the effects of atomic physics
on the emission for the different approximations.
Other groups that are developing RT codes are encouraged to add their
results to the repository in the standardized format. Information for
this purpose is provided in the \texttt{README.md} file in the
repository.
\begin{acknowledgements}
The results presented in this paper are based on work performed before Feb 24\textsuperscript{th} 2022.
We thank the Schwartz/Reisman Institute for Theoretical Physics at the Weizmann institute of science and the Max Plank Institute for Astrophysics for hosting the workshops during 2018-2019 which lead to this collaboration.
This work was supported by the `Programme National de Physique Stellaire' (PNPS) of CNRS/INSU co-funded by CEA and CNES.
This research was supported by the Excellence Cluster ORIGINS which is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy EXC-2094-390783311.
SB acknowledges support from the ESO Scientific Visitor Programme in Garching.
CC acknowledges support by the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme under grant agreement No. 759253.
AJ acknowledges funding
from the European Research Council (ERC) under the European Union’s
Horizon 2020 Research and Innovation Program, ERC Starting Grant
803189 – SUPERSPEC.
Lawrence Livermore National Laboratory is operated by Lawrence Livermore National Security, LLC, for the U.S. Department of Energy, National Nuclear Security Administration under Contract DE-AC52-07NA27344.
AF and LS acknowledge support by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (ERC Advanced Grant KILONOVA No. 885281).
The reported study was funded by the Russian Scientific Foundation (RSF),
project number 19-12-00229, and the Russian Foundation for Basic Research
(RFBR) and the Deutsche Forschungsgemeinschaft (DFG), project number 21-52-12032.
SAS acknowledges funding from the UKRI STFC Grant ST/T000198/1. Part of this work was performed using the Cambridge Service for Data Driven Discovery (CSD3), part of which is operated by the University of Cambridge Research Computing on behalf of the STFC DiRAC HPC Facility (www.dirac.ac.uk). The DiRAC component of CSD3 was funded by BEIS capital funding via STFC capital grants ST/P002307/1 and ST/R002452/1 and STFC operations grant ST/R00689X/1. DiRAC is part of the National e-Infrastructure.
M.W. acknowledges support from the NASA Future Investigators in NASA Earth and Space Science and Technology grant (80NSSC21K1849) and support from the Thomas J. Moore Fellowship at New York University.
Los Alamos National Laboratory is operated by Triad National Security, LLC, for the National Nuclear Security Administration of U.S. Department of Energy (Contract No. 89233218CNA000001). RTW also acknowledges Daniel van Rossum for the development of SuperNu that enabled this work.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\noindent \textbf{Background.} The accurate prediction of biological features from genomic data is paramount for precision medicine, sustainable agriculture and climate change research.
Yet some characteristics of big biological data render most out-of-the-box methodologies inaccurate or inapplicable.
In particular, here we focus on 1) the high complexity (and sometimes impossibility) to align genomic sequences for certain organisms, and 2) the difficulty to obtain labeled samples to use in supervised learning.
First, while the growing interest of the biological community in machine-learning methods is undeniable, many existing methods \cite[e.g.]{Agarwal2019-pr, Pei2018-ei, wnuk2018predicting, greenside2018discovering, zhang2016deepsplice, Nguyen2020} need DNA or RNA alignments as input.
Fast evolving organisms like virus or bacteria -- especially when big sample sizes are needed -- muddle the alignment process with their heterogeneity and genomic diversity to the point that aligning the sequences becomes extremely challenging or impossible \cite{ren2018alignment, thompson2011comprehensive}.
Second, supervised learning models are the most accurate options for phenotype prediction. However, these methods rely on large samples of labeled data. In reality, many biological experiments produce partially observed labels given the money and time constraints to phenotype organisms or strains. One example that motivated our work is the case of mycovirus. Mycoviruses infect fungi and can be used as biocontrol of crop pests given that they induce hypovirulence on plant fungal pathogens. Not all mycoviruses cause hypovirulence in the fungal host, and thus, a standard mycovirus dataset will be partially labeled with some labeled sequences that have been tested \textit{in planta} for their hypovirulence potential, and many more unlabeled sequences that have never been tested in laboratory, and thus, have unknown hypovirulence potential.
Semi-supervised methods \cite{doostparast2018graph, bhardwaj2010genome} allow for partially observed labels, but existing methods require that we observe labels for all classes -- albeit in smaller frequencies than the unobserved labels. This condition is not met in the mycovirus dataset, for example, where negative controls (class of mycoviruses that do not cause host hypovirulence) are not available as observed labels only correspond to mycoviruses tested in the lab for their hypovirulence-induced potential. Thus, the use of most semi-supervised methods is prohibited for this maximally imbalanced dataset.
\noindent \textbf{Main contributions.}
We present BioKlustering (\url{https://bioklustering.wid.wisc.edu}) a user-friendly open-source web app to cluster unaligned genomic sequences based on maximally imbalanced partially observed labels.
Unlike most semi-supervised learning methods, our web app does not require all classes to be observed or has any requirements on class balance. Also, unlike most machine-learning methods, our web app does not require large sample sizes. While designed for a semi-supervised setup, our web app can also be used when no labels are measured (unsupervised case).
One of the main purposes of our web app is to allow easy and fast clustering of sequences that could inform future biological experiments. Indeed, we see our web app as a hypothesis-generating tool that will allow biological users, for example, to identify clusters of mycovirus sequences that could be either negative controls (from the unobserved class of mycovirus that do not serve as biocontrol and that have never been tested \textit{in planta}) or new mycovirus sequences that could serve as biocontrol for further experimentation.
We highlight that even when our work is visualization-driven, not inferential, we believe that it fills a important gap for biological scientists who need to identify similar sequences for experiments when sequence alignment is prohibited and when \textit{in vitro} or \textit{in planta} phenotyping of all labels is not possible.
\section{Methods}
\noindent \textbf{Input data.}
Sequences are input as FASTA files and are internally converted to kmer counts. For example, if a sequence is ``ACTGG", then its 3-mers are [``ACT", ``CTG", ``TGG"]. The user can select the length of the kmer and sequences can be aligned or unaligned.
Labels are input as an optional separate csv file. The matching is done by assuming that the rows in both files are in the same order. The labels file should assign a value of $-1$ to the sequences of unknown label.
\noindent \textbf{Models.} We implement three unsupervised clustering methods which are extended to the case of partially observed labels by internally optimizing the kmer length and other parameters until we reach maximum consensus with the observed labels.
\noindent \textit{K-means clustering.} The algorithm \cite{macqueen1967some} groups the data into $K$ clusters each formed around a centroid. Each data point is assigned to the nearest centroid, and these centroids are formed by minimizing the squared Euclidian distances within each cluster. The mean shift algorithm is used to identity locations of high density within the kmer space of the data, and then the unsupervised k-means model is run with these locations as the initial centroid coordinates.
Parameters of this model include the minimum and maximum kmer length, the random seed, and the number of clusters. Since the mean-shift algorithm does not allow an input number of clusters, when the number of clusters predicted by mean-shift exceeds the number of clusters requested, Algorithm \ref{algo:assignClustersKM} (Appendix) is used to assign output labels.
The unsupervised k-means model can incorporate known labels to create an effective semi-supervised model. The mean shift algorithm is again used to identity locations of high density within the kmer space of the data, and then the unsupervised k-means model is run with these locations as the initial centroid coordinates. Known labels are then compared against predicted labels and these clusters are reassigned into groups that minimize the prediction error.
\noindent \textit{Gaussian Mixture Model.} The algorithm fits a probabilistic model \cite{Buitinck2013-dn} that estimates the underlying multiple Gaussian distributions that best describe the clusters in the data. Starting with a random initialization model parameters, the Expectation-Maximization algorithm runs iteratively until convergence.
Parameters of this model include the minimum and maximum kmer length, the random seed, the number of classes, and the covariance type (which determines the shape of the clusters).
The semi-supervised GMM differs from a unsupervised one in the initialization and parameter selection process. A semi-supervised GMM starts with initializing means of the known data by their actual means and the rest with random numbers. A parameter selection process is implemented to compare models with different parameter combinations (kmer length and four covariance types), and the model with high accuracy calculated from known labels will be returned.
Finally, for the semi-supervised case only, we use Algorithm \ref{algo:assignClusters} (Appendix) to assign output labels from running the GMM to the input labels.
\noindent \textit{Spectral clustering.}
This algorithm \cite{pentney2005spectral} exploits the potential of eigenvalues of the matrix derived from the input data. In spectral clustering, the input data will be treated as a graph, and each sequence will represent a vertex in the graph. Then, the vertices in the graph will be partitioned based on their similarities.
Parameters of this model are the minimum and maximum kmer length, the random seed, the number of classes, and the manner in which we assign labels when the input is projected to a lower dimension space (denoted label assignment option).
The semi-supervised version of spectral clustering will find the optimal parameter values (kmer length and label assignment option) that maximizes the predictive accuracy of the observed labels.
We again use Algorithm \ref{algo:assignClusters} (Appendix) to assign output labels from running the spectral clustering algorithm to the input labels.
For all three algorithms, the predicted labels and embedded kmer data can be used to visualize the clusters using principle component analysis (PCA) or t-distributed stochastic neighbor embedding (t-SNE).
\section{Validation with \textit{Influenza} data}
We test our three semi-supervised algorithms with \textit{Influenza} data from two classes: bats and cats \cite{virus-database}. We run four settings: totally unsupervised, semi-supervised with 50\% of observed labels (both classes observed), semi-supervised with 10\% observed labels (both classes observed) and semi-supervised with 10\% observed labels (only one class observed). Table \ref{tab:pred} shows the prediction accuracy of the three learning algorithms in BioKlustering. See the Supplementary Material for clustering figures and more information about the data.
\setlength{\tabcolsep}{3pt}
\begin{Table}
\centering
\begin{tabular}{l|ccc}
& k-means & GMM & Spectral \\
\hline
Unsupervised & 71.3 & 75.3 & 75.8 \\
Semi-supervised (50\%) & 82.6 & 93.8 & 87.1 \\
Semi-supervised (10\%) & 73.6 & 90.4 & 75.3 \\
Semi-supervised (10\% 0's) & 71.3 & 51.1 & 70.2
\end{tabular}
\captionof{table}{Prediction accuracy (\%) of the three learning algorithms in BioKlustering on the \textit{Influenza} unaligned genome data for two classes (bats and cats) with respect to all the true labels.}
\label{tab:pred}
\end{Table}
\section{Web app interface}
BioKlustering is an open-source web application developed with Django, a Python-based framework (Figure \ref{webapp}). The users will be able to predict clusters in genomic data by 1) uploading FASTA files with (aligned/unaligned) genome sequences, 2) selecting a clustering algorithm, and 3) choosing in its corresponding parameters.
BioKlustering provides semi-supervised and unsupervised options depending on the presence of known labels. For each algorithm, the web app presents a description to aid users in the choice of parameters. The selected algorithm will generate prediction results which include an interactive plot (built with Plotly Dashboard \cite{plotly}) and a table with predicted labels. Users can download a static version of the plot, the table, and the parameter information in a zip file which can also be sent via email.
\noindent \textbf{Open-source code.} All of our code is open source in the following GitHub repository \url{https://github.com/solislemuslab/bioklustering}.
\noindent \textbf{Acknowledgements.} This work was supported by the Department of Energy [DE-SC0021016 to CSL]. We thank Dr. Aurelie Rakotondrafara and Helena Jaramillo Mesa for the motivation on the mycovirus data. We also want to thank Ben Huebner from the WID IT team who helped with the public deployment of the web app. Finally, we acknowledge the work in \cite{Hotaling2020} which helped us improve the scientific writing of this manuscript.
\bibliographystyle{unsrt}
|
\section{Introduction}
High-fidelity single-photon and multi-photon states are of central importance to emerging quantum technologies such as quantum cryptography \cite{pirandolaAdvancesQuantumCryptography2020}, quantum metrology \cite{barbieriOpticalQuantumMetrology2022}, quantum imaging \cite{genoveseRealApplicationsQuantum2016}, quantum spectroscopy \cite{mukamelRoadmapQuantumLight2020}, and quantum information processing \cite{flaminiPhotonicQuantumInformation2018}.
Spontaneous Parametric Downconversion (SPDC) has been a mainstay in experiments exploring these applications, being used to produce heralded single photons \cite{eisamanInvitedReviewArticle2011}, entangled photon pairs \cite{edamatsuEntangledPhotonsGeneration2007} and multi-photon entangled states \cite{panMultiphotonEntanglementInterferometry2012}.
Since SPDC produces photons in pairs, the most common approach in the latter case is to employ multiple SPDC photon-pair sources and to combine the pairs using entanglement-swapping \cite{zeilingerThreeParticleEntanglementsTwo1997}.
An alternative method is to combine SPDC sources in a cascade, using photons generated from an initial SPDC process as a pump for a second, as shown in Fig. \ref{fig:CSPDC}(a).
Cascaded SPDC (CSPDC) has been used to produce three-photon entangled states in the polarization \cite{chaissonPhasestableSourceHighquality2022} and energy-time \cite{agneObservationGenuineThreePhoton2017} degrees of freedom.
It can also serve to produce heralded single photons \cite{poitrasProposalLownoiseHeralded2018} and heralded Bell pairs \cite{hamelDirectGenerationThreephoton2014}.
Additionally, CSPDC has been used to implement photon precertification \cite{meyer-scottCertifyingPresencePhotonic2016}, which could allow for Bell inequality measurements free of the detection loophole even with transmission over lossy channels \cite{cabelloLoopholeFreeBellTest2012}, providing a path to long-distance device-independent quantum communication.
Currently, one of the main challenges with CSPDC experiments is the efficiency of the process; measured triplet rates have so far been limited to between a few detected triplets per hour \cite{hubelDirectGenerationPhoton2010,krapickOnchipGenerationPhotontriplet2016} and a few hundred per hour \cite{hamelDirectGenerationThreephoton2014}.
In order to improve these rates and make CSPDC attractive for practical applications, the most critical bottleneck is the second stage of downconversion, as any increase to this efficiency will directly enhance the measured rates.
In contrast, improving the efficiency of the first stage requires faster detectors to detect the higher rate of photons from the first downconversion.
Increasing the pump power has the same limitation.
Moreover, in the context of photon precertification, the generation of photons in the first stage should be considered independent, with the success probability of precertification directly linked to the efficiency of the second state.
It is therefore imperative, in order to enable more advanced experiments, to explore ways to increase the efficiency of the second stage.
One straightforward approach would be to use materials with stronger nonlinearities \cite{guoParametricDownconversionPhotonpair2017,stantonEfficientSecondHarmonic2020}.
Another promising technology is thin-film nonlinear devices which use stronger confinement to enhance conversion efficiency.
However, these devices have so far reached downconversion efficiencies on the order of $10^{-8}$ \cite{zhaoHighQualityEntangled2020,xinSpectrallySeparablePhotonpair2022}, whereas commercial PPLN waveguides such as the one used in this work already achieve efficiencies of over $10^{-6}$\cite{hubelDirectGenerationPhoton2010}.
Another option could be to place the second SPDC stage inside a cavity, which is a commonly used approach to enhance SPDC efficiency \cite{slatteryBackgroundReviewCavityEnhanced2019}.
However, the benefit of such an approach would be limited if the acceptance bandwidth of this cavity is smaller than that of the photons produced in the first stage.
While such a cavity could be an attractive option if the first stage consists of an atomic source \cite{dingHybridcascadedGenerationTripartite2015}, or of an SPDC source which is also placed in its own cavity with both cavities locked together, it is not immediately implementable if the first stage is a regular SPDC process.
In this work, we propose a different approach which aims to get the benefits of cavity enhancement without the typical bandwidth constraints of time-independent cavities.
One way around these limitations is through the use of time-dependent components \cite{mannNonreciprocalCavitiesTime2019,cardeaArbitrarilyHighTime2020}.
Here, we make use of the fact that arrival times of photons from the first downconversion are known in order to capture them inside a switchable cavity consisting of a delay loop and fast switch \cite{pittmanSinglePhotonsPseudodemand2002,kanedaTimemultiplexedHeraldedSinglephoton2015}.
Photons that fail to convert on the first pass through the second non-linear crystal are therefore provided with multiple additional chances of downconversion, thus amplifying the production of triplets.
To validate the concept, we develop a theoretical model to investigate the expected amplification from this method and provide experimental results from a proof of concept device.
\section{Theory}
Our scheme, shown in Fig. \ref{fig:CSPDC}(b), consists of a CSPDC device with an active loop controlled by a fast switch.
This optical switch should be open when a photon in mode (0) is expected to arrive.
The timing can be determined either using the detection of its partner photon in mode (1) or, in the case of a pulsed laser, with a periodic signal synchronized with the pump.
It is important that the switch is fast enough to be in a closed state before the photon completes the loop.
Otherwise, the photon will escape during its second pass instead of being trapped.
Without the optical switch, or by keeping it in its ``open'' state, the expected rate of produced triplets is given by
\begin{align}
T_0 = R_1 P_\mathrm{SPDC},
\end{align}
\noindent where $R_1$ is the rate of photons in mode (1) and $P_\mathrm{SPDC}$ is the probability of downconversion at SPDC 2.
With the addition of the switch, the photon in mode (0) is trapped in the loop and gets multiple passes through SPDC 2.
The contribution to the overall triplet rate from the k'th pass is given by
\begin{align}
T_k=\eta R_1 P_\mathrm{SPDC}(1- P_\mathrm{SPDC})^{k}\alpha^{k},
\end{align}
\noindent where $\eta$ is the efficiency of the switch and $\alpha$ is the loop transmission efficiency defined as the transmission probability after one full loop. We define the amplification factor, $A$, as the ratio of the sum of contributions from each $T_k$ term with the un-amplified rate:
\begin{equation}
A = \eta\sum_{k=0}^{\infty} \bigg[\alpha \big(1-P_\mathrm{SPDC}\big)\bigg]^{k}.
\label{Amp1}
\end{equation}
\begin{figure}[ht]
\centering
\begin{minipage}{0.48\textwidth}
\centering
\includegraphics[width=\linewidth]{TheoryDiagram_noAMP.pdf}
\end{minipage}
\begin{minipage}{0.48\textwidth}
\centering
\includegraphics[width=\linewidth]{TheoryDiagram_alt2.pdf}
\end{minipage}
\caption{Comparison of traditional CSPDC and our novel amplified CSPDC scheme. (a) A regular CSPDC scheme for producing time-correlated photon triplets. (b) An enhanced CSPDC scheme for producing time-correlated photon triplets with amplification by an active loop. The first pair is created at SPDC 1. The fast switch is shown to be in the closed position. With the right timing, triggered by a heralded photon pair or a pulsed laser, the switch is opened just before photon (0) arrives. Photon (0) is given a chance at downconversion at SPDC 2, then a Dichroic Mirror (DM) reflects it back into the loop. Until another photon pair arrives from SPDC 1, this photon will be given many chances to interact with SPDC 2 to create a second pair and thus a triplet.}\label{fig:CSPDC}
\end{figure}
\noindent Eq. (\ref{Amp1}) is a geometric series and the expected amplification from this scheme is thus
\begin{align}
A = \frac{1}{1-\alpha(1-P_\mathrm{SPDC})}.
\end{align}
This first model is valid as long as the switch is never activated while a photon is still in the loop.
For a more accurate calculation, we must take into account the effect of the switch being triggered while another photon is present.
If the switch is fast enough, it could be possible to have more than one photon inside the loop at a time.
However, this would lead to complications in discriminating triplets from one another.
Instead, we assume that the switch is always opened long enough such that its activation for a new photon will eject any photon already inside the loop.
Here, we consider the specific case of a continuous pump with triggers to the switch coming from detections of photon (1).
Since SPDC is a random process, we can use a Poisson distribution, $\mathrm{Pois}(x,\lambda)=\frac{{\lambda}^x \mathrm{e}^{-\lambda}}{x!}$, at $x=0$ to calculate the probability of no photons arriving in mode (1) within the loop time, $t_L$
\begin{align}
&\mathrm{Pois}(x=0,\lambda=R_1t_L)=\exp({-R_1t_L}),
\label{pois0}
\end{align}
\noindent where $\lambda$ is the expected number of photons in mode (0) in the time interval $t_L$. This gives an amplification of
\begin{align}
A=\eta\sum_{k=0}^{\infty} \Bigg[\frac{\alpha \big(1-P_\mathrm{SPDC}\big)}{\exp{(R_1t_L)}}\Bigg]^{k}.
\label{pois1}
\end{align}
Typically, the single-photon detectors have a certain deadtime $\tau$ when no detections can be made, therefore no new triggers are registered during this time.
We define the quantity $j=\lfloor t_L/ \tau \rfloor$ which is a floor function that gives the integer number of passes in which the photon that is trapped in the loop cannot be affected by any new photon pairs.
The contribution from the first $j$ passes will follow Eq. (\ref{Amp1}), whereas for passes beyond the value of $j$, the Poisson term must be taken into account as in Eq. (\ref{pois1}).
The more complete model for the amplification is thus given by
\begin{align}
A = \eta\sum_{k=0}^{j} \bigg[\alpha \Big(1-P_\mathrm{SPDC}\Big)\bigg]^{k}
+ \eta\sum_{k=j+1}^{\infty} \Bigg[\frac{\alpha \big(1-P_\mathrm{SPDC}\big)}{\exp{(R_1t_L)}}\Bigg]^{k}.
\label{ampFull}
\end{align}
As can be seen in Fig. \ref{fig:ampVSalpha}, at low values of alpha the amplification is modest and there is almost no dependence on the SPDC 1 rate.
This is because most photons in the loop will be lost before getting enough passes to be converted or will be ejected by a subsequent photon from SPDC 1.
However, at higher values of loop efficiency the amplification can be significant: for a heralding rate of around 1 million counts per second, which is reasonable for a commercially available single-photon avalanche diode, a loop efficiency of 0.93 is sufficient to provide an order of magnitude amplification.
For higher count rates, ejection of photons from the loop becomes significant, so a shorter loop time would become beneficial.
\begin{figure}[H]
\centering
\includegraphics[scale=0.55]{ampVSalpha_theory.pdf}
\caption{Dependence of the amplification factor $A$ (Eq. (\ref{ampFull})) on the loop coupling efficiency $\alpha$, for different mode (1) photon rates. The parameters used in these predictions, which are similar to those in our experiment, are $\eta=0.94$, $P_\mathrm{SPDC}=1.5 \e{-6}$, $\tau=45\si{\nano\second}$ and $t_L=23\si{\nano\second}$. As the mode (1) photon rate increases, the theoretical maximum amplification is reduced. This is due to new photons arriving and ejecting the already present photon from the loop, reducing the number of passes that each photon has in the second SPDC crystal.}
\label{fig:ampVSalpha}
\end{figure}
\section{Experiment}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.95]{ExperimentDiagram_Paper.pdf}
\caption{A $405\si{\nano\meter}$ pump beam, with polarization set by a half-wave plate (L/2), is incident on a PPKTP crystal. Type-2 SPDC occurs at the PPKTP crystal which produces two orthogonally polarized photon beams of wavelengths $849\si{\nano\meter}$ and $775\si{\nano\meter}$. These two overlapping beams are separated using a Polarizing Beam Splitter (PBS), followed by appropriate band-pass filters for removing the remaining pump beam. The $775\si{\nano\meter}$ photon is sent to a second stage of SPDC for the creation of a second time-correlated pair. Once the polarization has been modified by the Pockels Cell, the photons are coupled into a polarization maintaining (PM) fiber coupled to a PPLN waveguide. The output of the waveguide is coupled back into free space and is incident on a Dichroic Mirror (DM) which reflects the $775\si{\nano\meter}$ light back into the loop and transmits the $1550\si{\nano\meter}$ light towards the 50:50 Beamsplitter (BS) and then to the detectors. The signals from the 3 detectors are sent to a logic correlation unit for triplet detection. The half-waveplate (HWP) in the loop is used to control the loop coupling efficiency in order to test the model.}
\label{fig:experimentDiagram}
\end{figure}
To demonstrate the validity of the enhancement model, we implement the scheme experimentally as shown in Fig. \ref{fig:experimentDiagram}.
The cascade starts with a $405\si{\nano\meter}$ continuous-wave (CW) laser (Toptica Topmode) pump in a Periodically Poled Potassium Triphosphate (PPKTP) crystal (Raicol) which produces photon pairs at $775\si{\nano\meter}$ and $849\si{\nano\meter}$.
The $849\si{\nano\meter}$ photons are detected first, while the $775\si{\nano\meter}$ photons pass through a fiber optic delay line before being sent to the fast switch.
To implement the fast switch, a Pockels Cell (PC) is combined with a Polarizing Beam Splitter (PBS).
The PC is tuned as a half-wave plate so that the loop is open at high voltage and closed when the voltage is off.
After the PBS and before passing through the PC, the $775\si{\nano\meter}$ photons are horizontally polarized.
The trigger signal from the Single Photon Avalanche Detectors (SPADs) (Excelitas SPCM-AQ4C) detecting $849\si{\nano\meter}$ photons switches on the PC and the polarization of the $775\si{\nano\meter}$ beam is changed from horizontally polarized to vertically polarized.
This is the required pump polarization for type-0 SPDC at the second nonlinear crystal, a PPLN waveguide (HC Photonics).
The timing of the PC is set to allow sufficient time for it to reach half-wave voltage before the $775\si{\nano\meter}$ photon arrives.
The $775\si{\nano\meter}$ photons which do not downconvert within the nonlinear crystal are reflected by a DM towards the PBS.
Since these photons are now vertically polarized, they are reflected by the PBS and sent to the PC.
The time in which the PC is on after the photon enters the loop is sufficiently low such that by the time the vertically polarized $775\si{\nano\meter}$ photons reach the PC, it is turned off.
This ensures that the polarization is not changed a second time and the photons can stay in the loop until they are downconverted or absorbed.
Photon pairs produced inside the PPLN waveguide are separated probabilistically before being detected by superconducting nanowire single photon detectors (SNSPDs) (Photon Spot).
The timestamps of each triplet detection are recorded using a time tagging unit with $156.25\si{\pico\second}$ time resolution (UQDevices Logic 16), with subsequent events being labeled as a triplet if detections are recorded at all three detectors with a maximum delay between each detection of $95.9375\si{\nano\second}$.
\section{Results}
With a $4.3\si{\milli\watt}$ pump incident on the PPKTP crystal, $330,000$ coincidences per second, within a coincidence window of $1\si{\nano\second}$, are detected between the $775\si{\nano\meter}$ and $849\si{\nano\meter}$ beams at the SPADs.
Compared to a half-waveplate, the PC converts the polarization of $\eta_{PC}=94\pm3\%$ of the $775\si{\nano\meter}$ photons.
This loss of efficiency is mainly due to photons arriving within $100\si{\nano\second}$ of a previous detection, which are ignored by the PC driver to protect the high voltage circuit.
The PPLN waveguide used for cascaded SPDC has a downconversion efficiency of $P_\mathrm{SPDC}\approx10^{-6}$.
The combined detection and coupling efficiency for the $1550\si{\nano\meter}$ photons after the beamsplitter is $0.156$, measured as the ratio of coincidence and single photon detections.
Fig. \ref{fig:3Dhist} shows the time distribution of all photon triplets measured during 105 hours with the loop coupling efficiency at it's maximum value of 18\%.
Most of the triplets lie on the $t_3-t_2=0$ line, since the $1550\si{\nano\meter}$ photons are created simultaneously and travel equal distances.
From left to right, the first and largest peak in the histogram represents photons that downconverted on the first pass through the PPLN waveguide (SPDC 2).
The second peak corresponds to triplets generated after one additional pass through the PPLN waveguide, and so on.
The time difference between the two consecutive peaks is the time it takes the $775$ photon to traverse the loop once, $23\si{\nano\second}$.
Most accidentals in our experiment, visible as triplets on the fixed $t_3-t_2=0$ line but not within these main peaks, are therefore caused by a real $1550\si{\nano\meter}$ pair being detected along with an uncorrelated $849\si{\nano\meter}$ photon.
\begin{figure}[ht]
\centering
\includegraphics[scale=0.35]{3DHist.pdf}
\caption{A 3D histogram of the triplet counts for our maximum achievable loop coupling efficiency, $\alpha_0$. The $t_2-t_1$ axis is the time difference between a $849\si{\nano\meter}$ trigger photon at $D_1$ and a $1550\si{\nano\meter}$ photon at $D_2$. The $t_3-t_2$ axis is the time difference between $1550\si{\nano\meter}$ detections at $D_2$ and $D_3$.}
\label{fig:3Dhist}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[scale=0.50]{ampVSalpha_errorPropagation.pdf}
\caption{Comparison of measured amplification with model prediction. Each experimental data point corresponds to a 35 hour measurement, except for the $\alpha=0.18$ data point which was measured for 105 hours. The error bars for amplification are calculated assuming poissonian statistics on the number of counts in each peak. The uncertainty in $\alpha$ are estimated based on observed variations in fiber coupling efficiencies in the setup. The uncertainty band for the model prediction come from uncertainties in the values of $\eta_{PC}$ and $R_1$, given by $\eta_{PC}=0.94\pm0.03$ and $R_1=(1.52\pm0.03)\e{6}$ counts/sec. }
\label{fig:results}
\end{figure}
Based on this measurement, we find an average triplet rate of $48.9 \pm 0.7$ triplets/hour.
As a comparison, we also perform an un-amplified measurement for 17 hours by turning off the Pockels cell and using a half-wave plate to perform a polarization rotation.
This measurement produced $40.6 \pm 1.5$ triplets/hour.
We therefore infer an amplification of $20.5 \pm 4.8 \% $ from the active switching, which is consistent with the expected amplification.
In order to further verify the model, we repeat these measurements for various values of the loop coupling efficiency ($\alpha$) using a half-wave plate placed before the PBS in the loop.
From this comparison, rather than using separate runs performed with the Pockels cell turned off to calculate the observed amplification, we instead infer it by comparing the relative size of the peaks in the timing histograms.
The amplification is calculated as :
\begin{equation}
A_\mathrm{exp}=\eta_\mathrm{PC}\bigg(1+\frac{P_2 + P_3 + P_4}{P_1}\bigg)
\end{equation}
where $P_i$ is the triplet count on the $i$th peak.
We include $\eta_\mathrm{PC}=0.94\pm0.03$ in the calculation in order to quantify whether any advantage coming from the additional peaks is enough to compensate for the PC not being as efficient as using only a half-wave plate to set the polarization.
The results of these measurements are shown in Fig. \ref{fig:results} along with the predictions of the model.
For low values of $\alpha$, the amplification is initially less than 1, since the gain from multiple loops is not enough to compensate for the imperfect switching of the Pockels cell.
Then, once $\alpha > 0.07$, we see a gain and the switchable cavity becomes beneficial.
\section{Discussion}
While the experimental results are in agreement with our model, further improvements are required for the switchable cavity to provide a more appreciable improvement to cascaded downconversion.
In order to reach a regime where the enhancement justifies the added complexity, it will be necessary to increase the loop coupling efficiency.
Coupling efficiencies of over $85\%$ into single mode fibers have been reached out of nonlinear crystal waveguides \cite{urenEfficientConditionalPreparation2004,zhangDeviceindependentQuantumKey2022}, so a similar coupling efficiency should be achievable from a waveguide to itself.
With a comparable loop coupling efficiency, our model predicts that an amplification factor of 6 should be accessible with current technology.
Even higher loop efficiency would likely be possible with bulk nonlinear crystals to have a fully free-space cavity, in which case the limiting factors would likely become losses from the optical components.
However, it is unclear if any gains from reusing photons would be enough in this case to compensate for the loss in downconversion efficiency caused by moving away from waveguide devices.
Alternatively, the scheme may benefit from being implemented entirely as an integrated photonic circuit.
The production of photon triplets from cascaded downconversion has already been demonstrated to be possible on an integrated chip \cite{krapickOnchipGenerationPhotontriplet2016}.
The Pockels effect has also been shown to be viable on integrated devices \cite{chmielakPockelsEffectBased2011,abelLargePockelsEffect2019}.
It may therefore be possible to integrate this setup, including the second downconversion stage and the switchable cavity, on a photonic chip which could reduce coupling losses between the fast switch and the downconversion stage.
Another advantage of employing integrated circuits is the possibility of smaller loop lengths.
This is beneficial at high loop coupling efficiency, since in this regime it becomes necessary to allow for a large number of passes to happen before the switch is reactivated.
With such increases to the loop coupling efficiency, reusing photons with a switchable cavity will become a viable approach to increase the production rates of photon triplets from CSPDC by more than an order of magnitude.
This will make it attractive to optical quantum computing schemes which rely on either three-photon entangled states \cite{gimeno-segoviaThreePhotonGreenbergerHorneZeilingerStates2015} or heralded Bell pairs \cite{browneResourceEfficientLinearOptical2005}
as a resource.
It could also, if accompanied with high detection efficiency of the photons from the second downconversion, help improve the performance of photon pre-certification, paving the way towards long-distance device independent quantum communication.
\section{Acknowledgments}
We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada, the Canada Foundation for Innovation, Canada Research Chairs and the New Brunswick Innovation Foundation. The authors thank Z. Chaisson for the data collection software.
|
\section{Introduction}
\begin{figure*}
\centering
\includegraphics[width=.95\textwidth]{Figures/Schematic_NRSECorr_sep12-eps-converted-to.pdf}
\caption{ \label{figSchematic} Schematic of the corrected neutron resonance spin echo beamline. The neutron is polarized at the far left (Pol.) before traveling through the first arm, scattering from the sample (Sam.) at angle $\theta$, echoing through the second arm, and entering the analyzer (Ana.) at the far right; the detector is not shown. The correction magnets are labeled ``CM'' and the rf flippers ``RF''. The white space between the CMs is at zero field. The scattering plane for the example neutron path shown is the $x$-$y$ plane.
For the first arm, $L_n$ for $n=1,2,3$ is the distance from the beam-defining slit to the center of the corresponding correction magnet, while for the second arm, $L_n$ for $n = 4,5,6$ is the distance from the sample to each correction magnet center. The distance from the optical axis to the point that the neutron passes through the $n^{\mathrm{th}}$ correction magnet is defined as $y_n$. The distance between rf flippers in both arms is $L_{\mathrm{RF}}$.
}
\end{figure*}
Neutron Resonance Spin Echo (NRSE) is a modification of the Neutron Spin Echo (NSE) technique which replaces the static-field precession coils with radio-frequency (rf) spin-flippers. \cite{Golub1987} The underlying principle of both types of \textit{echo} measurements is that the instrument will measure the change in velocity, and hence the change in energy, of a neutron scattered from a sample.
Currently, most large neutron sources use static-field NSE instruments for high-energy-resolution measurements of slow dynamics.
In order to be competitive with existing NSE instruments, NRSE would need to achieve a Fourier time (also called the spin echo time) of about one hundred nanoseconds. \cite{Farago2015,Ohl2012}
The Fourier time $\tau$ is given by
\begin{equation}
\tau = \frac{2 m^2}{h^2} L_{\mathrm{RF}} f \lambda^3
\label{EqnFouriertime}
\end{equation}
where $L_{\mathrm{RF}}$ is the distance between the rf flippers in each arm, $f$ is the rf flipper (linear) frequency, $\lambda$ is the neutron wavelength, $m$ is the neutron mass, and $h$ is Planck's constant.\cite{Keller2002}
State-of-the-art rf flippers are already capable of producing a high performance NRSE instrument. As an example, suppose an NRSE beamline uses recently developed transverse rf flippers. \cite{Li2020} A beamline with those rf flippers operating at 4 MHz, with a 2 meter separation between the rf flippers, and with a standard NSE wavelength of 1 nm would have a Fourier time of 100 nanoseconds, which is comparable to modern NSE beamlines.\cite{Golub1988}
However, due to the long wavelength requirements for both NSE and NRSE, the neutron flux is often low and the relevant samples scatter weakly. Therefore, measurements are only possible by having a large spatial and angular beam size, which leads to aberrations. In conventional NSE, one aberration source is due to the variation in the static field strength across the beam due to the field profile created by a solenoid geometry.
In NRSE, the rf flippers are separated by zero field regions, so this aberration is not present. A second type of aberration arises from scattering from a sample. The sample is placed in the center of the two symmetric arms as shown in Fig. \ref{figSchematic}. If the neutron scatters with some non-zero momentum transfer, then the path length through the second arm will not be the same as the path length through the first arm, so the neutron will spend a different amount of time in the two arms.
Because NRSE instruments measure the velocity change of a neutron by measuring its Larmor phase $\Phi = 4 \pi f t$, where $t$ is the time it takes for the neutron to travel between the rf flippers, a difference in time is measured as a change in the neutron velocity. An uncorrected echo measurement will then conflate a change in scattering angle with a change in energy. The time can be written in terms of the neutron path length between the rf flippers as
\begin{equation}
t = \frac{L_{\mathrm{RF}}}{v \cos\theta},
\end{equation}
where $v$ is the neutron velocity and $\theta$ is the scattering angle in the scattering plane (see Fig. \ref{figSchematic}).\cite{Keller2002}
Thus, for our example NRSE beamline, a neutron scattering at 1 degree would be out of Larmor phase by more than 2000 degrees compared to the unscattered neutrons if there were no correction.
This aberration will be present for elastic, quasielastic, and inelastic neutron scattering.
Clearly, an NRSE instrument must have a method for correcting this geometric contribution to the Larmor phase. In NSE instruments, Fresnel coils with longitudinal-fields (i.e. the field is orientated along the optical axis) are used for an analogous correction as well as correcting the aberration from the static field variation. \cite{Ohl2005,Ohl2012,Farago2015} However, transverse-field rf flippers (i.e. the rf flipper's static-field is perpendicular to the optical axis) have been constructed for NRSE measurements, \cite{Li2020, Endo2019} which require a transverse-field correction magnet to improve the polarization.
In this paper, we present an analytical solution of the magnetic field profile needed to correct the path length aberrations and the design of a suitable prototype correction magnet. The Larmor phase that a neutron acquires traveling through our prototype NRSE correction magnet varies quadratically with the distance from the optical axis and is radially symmetric. We simulate an NRSE beamline with the correction magnets and experimentally measure the spatial dependence of the Larmor phase change through the device. We demonstrate that transverse-field NRSE beamlines can be corrected with transverse static-field magnets, and therefore have the potential to be competitive with NSE beamlines.
\section{Analytical Solution} \label{sec:ana}
The two arms of the NRSE instrument will be corrected independently; we will discuss the correction for the second arm first.
The necessary magnitude of the correction is proportional to the path~length~difference~$\Delta L_{\mathrm{RF}}$ between the two pairs of rf flippers, which is given by
\begin{equation} \label{pathlength}
\Delta L_{\mathrm{RF}} = L_{\mathrm{RF}} \left(\frac{1}{\cos\theta} - 1\right) \approx \frac{1}{2}L_{\mathrm{RF}}\theta^2,
\end{equation}
where $\theta$ is the small scattering angle shown in Fig. \ref{figSchematic}, defined relative to the optical axis.
This difference in path length will cause a delay in time, and thus a difference in the Larmor phase $\Phi$.
The difference in Larmor phase $\Delta \Phi$ between the unscattered and scattered beam for idealized rf flippers in the NRSE configuration is\cite{Keller2002}
\begin{equation} \label{deltaLinital}
\Delta \Phi = \frac{4 \pi f}{v} \Delta L_{\mathrm{RF}} \approx 2 \pi f \frac{\lambda m}{h} L_{\mathrm{RF}} \theta^2.
\end{equation}
The Larmor phase has already been defined as $\Phi = 4 \pi f t$ for an NRSE instrument, but it can also be defined for NSE instruments in terms of the magnetic field integral: $\Phi = (\gamma/v) \int ds \, B$, where $\mathrm{FI}_s = \int ds \, B$ the field integral experienced by the neutron traveling along the path $s$ and $\gamma \approx -1.832 \times 10^8$ rad/(T$\cdot$s) is the neutron's gyromagnetic ratio.
In NSE, the neutron magnetic moment rotates (precesses) in the plane perpendicular to an applied static field and the Larmor phase measures the amount of this rotation relative to some fixed direction in the lab. In NRSE, the rf field rotates in the plane perpendicular to a static field, and the Larmor phase measures that angle between the neutron magnetic moment and the rf field.
Hence, rf and static field effects on the neutron can be added together, as has been exploited recently.\cite{Jochum2020b}
To correct the phase difference, we must design a correction scheme consisting of static magnetic fields that generates a Larmor phase proportional to $\theta^2$, with the proportionality coefficient $\chi$ being
\begin{equation} \label{CC}
\chi = 2 \pi \frac{f}{\gamma} L_{\mathrm{RF}}.
\end{equation}
Notice that the required correction is independent of wavelength. With this correction, the Larmor phase of all diverging
neutrons will be corrected as if they traveled the same effective path length, namely the distance $L_{\mathrm{RF}}$.
Unfortunately, it is not obvious how to design a static magnetic field profile that would generate a purely $\theta^2$-dependent field integral term for a finite-sized beam.
However, we can generate such a term for a finite-sized beam with three correction magnets consisting of transverse static fields in which the magnitude of the field integral of a neutron traveling through the devices varies quadratically as a function of transverse position, with the center being the minimum and increasing radially outward. For scattering from a point-like sample, one can show that only two devices of this type are needed.
This design is a two-dimensional extension to the original solution proposed by Monkenbusch. \cite{Monkenbusch1999}
For simplicity, we only look at the aberrations in one dimension (transverse $y$ direction), but the following argument can be easily generalized to the entire two-dimensional plane perpendicular to the optical axis. The neutrons may be scattered from any position on the sample at a distance of $y_{\mathrm{sam}}$ from the optical axis into any angle $\theta$ defined relative to the optical axis.
If the sample were point-like, then the scattering angle $\theta$ would be defined just by the $y$ distance from the optical axis and the $\theta^2$ aberration would be known simply from the $x$ and $y$ position in the second arm. With a finite-size sample, the scattering position $y_{\mathrm{sam}}$ will add to the $y$ position from the scattering angle, so one device at a specific point along the beamline is not sufficient to correct the $\theta^2$ aberration.
To lowest order in scattering angle, the field integral per amp of a single prototype correction magnet is
\begin{equation} \label{correctiondeviceFIgeneral}
\mathrm{FI}_n = a_n + b_n y_n^2 + c_n y_n \theta,
\end{equation}
where $y_n$ is the distance between the neutron's path at the $n^{\mathrm{th}}$ correction magnet and the optical axis (see Fig. \ref{figSchematic}). There is no linear term in $y$ because the device is left-right symmetric; similarly, there is no linear term in $z$ due to the top-bottom symmetry. Each term in the field integral is proportional to the applied current, and $a_n$ has units of T$\cdot$m, $b_n$ has units of T/m, and $c_n$ has units of T/rad. Here $b_n$ is the term that corrects the path-length aberrations while the $a_n$ and $c_n$ terms appear because of the particular correction magnet geometry that we have chosen; higher order terms were found to have a negligible contribution to the field integral.
The transverse position that the scattered neutron passes through each correction magnet is given by
\begin{equation} \label{ydefinition}
y_n = y_{\mathrm{sam}}+ L_n \theta,
\end{equation}
where $\theta$ is assumed to be small and $L_n$ is the distance from the sample to the center of the $n$\textsuperscript{th} correction magnet, as shown in Fig. \ref{figSchematic}.
Combining Eqns. \eqref{correctiondeviceFIgeneral} and \eqref{ydefinition} for each device, we find the total field integral per amp $\mathrm{FI_T}$ experienced by a neutron arriving at the analyzer due to the three correction magnets to be
\begin{equation} \label{eqnLarmorPhaseCC}
\begin{aligned}
\mathrm{FI_T} = \sum_{n \in \{4,5,6\}} \big[& a_n + b_n y_{\mathrm{sam}}^2 + (c_n + 2 b_n L_n) y_{\mathrm{sam}} \theta \\
&+ L_n (c_n + b_n L_n) \theta^2 \big].
\end{aligned}
\end{equation}
The goal of the correction scheme is to have the coefficient of the $\theta^2$ term equal to Eqn. \eqref{CC} while having all other terms zero. Doing so, the series of correction magnets would correct the path-length aberration in the second arm of the instrument regardless of the $y_{\mathrm{sam}}$ position and without introducing any net field integral. These requirements on Eqn. \eqref{eqnLarmorPhaseCC} can be rewritten into several conditions:
\begin{gather*} \label{CoefficientCancellingEqns}
a_4 + a_5 + a_6 = b_4 + b_5 + b_6 = c_4 + c_5 + c_6 = 0 \\
b_4 L_4 + b_5 L_5 + b_6 L_6 = 0 \\
\chi = L_4 (c_4 + b_4 L_4) + L_5 (c_5 + b_5 L_5) + L_6 (c_6 + b_6 L_6).
\end{gather*}
Notice that if the sum of the currents through the three correction magnets is zero, then the first line will be satisfied.
Ignoring the $a_n$ terms for now, we solve this system of equations, obtaining
\begin{subequations} \label{SolutionEnc}
\begin{align}
b_4 &= \frac{\chi + c_4(L_5 - L_4) + c_6(L_5 - L_6)}{(L_4 - L_5)(L_4 - L_6)} \\
b_5 &= -\frac{\chi + c_4(L_5 - L_4) + c_6(L_5 - L_6)}{(L_4 - L_5)(L_5 - L_6)} \\
b_6 &= \frac{\chi + c_4(L_5 - L_4) + c_6(L_5 - L_6)}{(L_4 - L_6)(L_5 - L_6)} \\
c_5 &= - (c_4 + c_6),
\end{align}
\end{subequations}
with $c_4$ and $c_6$ being free parameters.
From this set of solutions, it is apparent that if we choose CM5, the fifth correction magnet, to be equidistant from CM4 and CM6 (such that $|L_4 - L_5| = |L_5 - L_6| = \delta L$) and also $c_4 = c_6$, then the angle-dependent $c_n$ terms cancel out, leaving $b_4 = b_6 = \chi/[2(\delta L)^2]$ and $b_5 = -\chi/(\delta L)^2$. Therefore, we can obtain our desired field integral by putting the same field in the first and last device and a field twice as large in the opposite direction in the middle device. We call this choice of fields the $(1,-2,1)$ configuration. In this configuration, the constant $a_n$ terms will also cancel out.
Next, we determine the magnitude required for $b_n$ for a realistic NRSE beamline. Plugging in Eqn. \eqref{CC} to the $b_n$ terms in Eqn. \eqref{SolutionEnc}, we see that
\begin{gather}
b_4 = \frac{\pi f L_{\mathrm{RF}}}{\gamma (\delta L)^2} \\
b_5 = -2 b_4, \quad b_6 = b_4. \nonumber
\end{gather}
To estimate the required field in the correction magnet, let $L_{\mathrm{RF}} = 2$ m, $\delta L = 1$ m, and $f = 4 $ MHz. Plugging in the numbers, we find $b_4 \approx -140$ mT/m. As we will show below, this value is attainable with our correction magnet.
The above discussion has only considered the second arm; now we look at the first arm.
If the initial beam is well-collimated (i.e. all neutrons in the first arm travel parallel to the optical axis), then no correction elements are required in the first arm even if there are correction magnets in the second arm. However, in practice, neutrons in a real instrument have some divergence angle relative to the optical axis.
This initial beam divergence will lead to a variation in path length between neutrons propagating at different angles in the first arm, similar to the scattering term for the second arm.
Therefore, we must correct for this variation in the first arm with another three correction magnets, as shown in Fig. \ref{figSchematic}. They must be in the $(-1,2,-1)$ configuration as the static magnetic field in both rf flippers is in the opposite direction relative the static fields in the rf flippers in the second arm.
Without this additional correction, we do not obtain the best possible improvement to the polarization.
The correction for the divergence angle of the neutron in the first arm will not require any changes to the correction magnet set-up in the second arm because the correction for the second arm is independent of angle or $y_{\mathrm{sam}}$ position.
With all six correction magnets installed, the Larmor phase, and hence Fourier time, of neutrons along any path in either arm will be corrected to the Larmor phase and Fourier time of a neutron traveling parallel to the optical axis.
\section{Development of the Correction Magnet} \label{sec:design}
\begin{figure}
\includegraphics[width=.95\linewidth]{Figures/CorrDevelopment_sep20-eps-converted-to.pdf}
\caption{\label{figCorrDevelopment} (a) CAD model of the correction magnet. The light blue surfaces are the high-temperature superconducting (HTS) films (front film not shown), the purple is the low-carbon steel chevron, the tan are the HTS coils, and the brown is the low-carbon steel flux return and pole pieces. (b and c) Simulation of fields at $y = 0$ and $z = 0$ with the coil current set to 10 amps, which have the same colorbar legend. The beam traveled from left to right. The HTS films are highlighted in white. (d) Contour plot of simulated field integral through the correction device at 10 amps for a neutron originating from a point source at $y = z = 0$ and $x = -20$ m. (e) Slices of the simulated field integrals in (d) through the origin.}
\end{figure}
To implement the analytical solution, we designed and constructed a vertical-field correction magnet following the drawing of Fig. \ref{figCorrDevelopment}(a). The bottom, top, and sides of the coils are enclosed by a magnetic flux return made of low-carbon steel (alloy 1018). High-temperature superconducting (HTS) films, gold-coated 350 nm thick YBCO on 0.5 mm thick sapphire substrate, were placed on the front and sides of the coils with another film placed 38 mm after the coils, outside of the magnetic circuit. These HTS films act as magnetic field screens due to the Meissner effect which creates sharp boundaries between field regions, as shown in Fig. \ref{figCorrDevelopment}(b,c).
Thus the magnet can be thought of in two parts: a contained region where the coils sit and an open region before the back film. HTS wire was wound around hollow low-carbon steel pole pieces and topped with ``chevrons'', low-carbon steel plates with v-shaped cutouts. The opening of the v is at the front of the device. The thickness of the chevrons was 3.2 mm, leaving a separation between chevrons of 50 mm. The angle of the chevron was 60 degrees, and the space from the coils to the rear HTS film was 38 mm. It was already well-known that a dipole magnet without a HTS film constraining the magnetic flux will create a magnetic field with a quadratic $z$-dependence, as correcting for this was one of the initial advantages of adding a HTS film. \cite{Wang2014}
The magnetic field in this device was simulated using the Siemens MagNet $\copyright$ software, which includes the material properties in its solutions via the finite-element method. The HTS films were simulated as perfect diamagnets preventing any perpendicular magnetic flux. We note that the explicit field profile in the device is arbitrary as long as the resulting field integral is quadratic across the device and the field direction does not change too quickly.
A useful feature of this design is that the $y$ and $z$ components of the field integral may be tuned independently. The $z$ component, shown in Fig. \ref{figCorrDevelopment}(b) is largely dependent on the distance to the back film while the $y$ component, shown in Fig. \ref{figCorrDevelopment}(c) is largely dependent on the chevron angle. As shown in \ref{figCorrDevelopment}(b), the quadratic behavior of the $z$ component comes from the bowing field lines protruding around the back of the coils due to the displaced back film.
A numerical solution to the field integral for any starting position $(y,z)$ and angle through the device was found by extracting the MagNet solution for the field at each point (mesh size 4 mm) and integrating the field along any chosen path. The simulated field integral through the coils at 10 amps for a neutron traveling from a far-away point source at $y= z =0$ and $x$ = -20 m is shown in Fig. \ref{figCorrDevelopment}(d). The difference in field integral between the center and edges is about 0.01 mT$\cdot$m, which approximately is the necessary correction value for an NRSE beamline with an rf flipper frequency of 1 MHz and 1 meter between the correction magnets. The field integral through the center is about 1.69 mT$\cdot$m.
\section{McStas Simulations of an NRSE Instrument} \label{sec:mcstas}
\FloatBarrier
Using McStas, a Monte Carlo neutron ray-tracing software package,\cite{Willendrup_Lefmann_2020, Willendrup_Lefmann_2021} we simulated an NRSE beamline with these correction magnets installed. The polarizer, rf flippers, analyzer, and detector were taken to be 100$\%$ efficient, while the correction magnet component was built using numerically simulated magnetic field data extracted from the MagNet simulations. The following model was used for the rf flipper at resonance:
\begin{equation}
\Phi_f = 2 \pi f (2t_i + \Delta t) -\Phi_i,
\end{equation}
where $\Phi_f$ is the final Larmor phase after exiting the rf flipper, $f$ the rf frequency (2 MHz for these simulations), $t_i$ the time at which the neutron enters the flipper, $\Delta t$ the time spent inside the flipper of thickness 15 cm, and $\Phi_i$ the initial Larmor phase when entering the flipper.
We used a modified version of the ``SANS\_spheres2'' sample, a default McStas sample that emulates elastically scattering hard spheres in a dilute solution. The sample parameters were chosen to prevent incoherent scattering or transmission without scattering, so the component acted like an idealized elastic scatterer with a maximum momentum transfer of 0.09 nm$^{-1}$, which corresponds to a maximum scattering angle of 0.7 degrees. The sample was 3 cm by 3 cm in transverse size with negligible thickness. The aperture diameter and the neutron wavelength were 2 cm and $0.8$ nm $ \pm 1\%$, respectively.
The separation between the rf flippers in each arm was taken to be 2.3 m, and the distance between the correction magnets 1 meter. The total distance from the source to the two-dimensional detector was 5.2 meters. With these parameters, the effective initial beam divergence is about 0.6 degrees.
Using the magnetic field data extracted from MagNet simulations, we determined that our prototype correction magnet had the following field integral per amp expansion across the correction magnet:
\begin{equation} \label{Eqn:fullFIE}
\mathrm{FI} = a + b_y y^2 + b_z z^2 + c_y y \theta + c_z z \psi,
\end{equation}
where $\psi$ is the vertical neutron divergence angle and the values of the fitted coefficients are given in Tab. \ref{tab:FIE CC coefficients} below.
Higher order terms were found to have a negligible contribution.
\renewcommand{\arraystretch}{2}
\begin{table}[h]
\centering
\newcolumntype{R}{>{\centering\arraybackslash}X}
\begin{tabularx}{.99\linewidth}{R|R|R|R|R}
$a \left(\frac{\mathrm{mT}\cdot \mathrm{m}}{\mathrm{A}}\right)$ &
$b_y \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{m}}\right)$ &
$b_z \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{m}}\right)$ &
$c_y \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{rad}}\right)$ &
$c_z \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{rad}}\right)$ \\
\hline
0.169 & 8.78 & 8.97 & -0.525 & 0.986
\end{tabularx}
\renewcommand{\arraystretch}{1}
\caption{\label{tab:FIE CC coefficients}
Values of the field integral expansion coefficients in Eqn. \eqref{Eqn:fullFIE}. The longitudinal length of the magnet is 14 cm.}
\end{table}
\begin{figure}[!b]
\centering
\includegraphics[width=.95\linewidth]{Figures/Correction_coil_mcstas_tinyinset.png}
\caption{\label{figNRSEmcstas} Simulated McStas polarization for an NRSE beamline with all six correction magnets on (red curve), only the final three correction magnets on (green curve), only the first three correction magnets on (orange curve), and no correction magnets on (blue curve).
The inset compares the polarization for all correction magnets on (red curve) to all correction magnets off and sample removed (purple curve) which shows that the polarization drops to about 0.9998 at the edge of the detector when both arms are corrected.}
\end{figure}
The NRSE beamline was simulated in McStas both with and without the correction magnets.
A plot comparing the simulated echo polarizations vs. radial position on the detector is shown in Fig. \ref{figNRSEmcstas}.
These simulations confirm that the inclusion of the correction elements greatly increases the polarization, especially for larger scattering angles which correspond to the edges of the detector.
Correcting only the second arm of the beamline improves the polarization for the larger scattering angles, although the polarization at the center of the detector is worsened compared to the uncorrected simulation due to the initial neutron divergence angle. However, if the initial beam divergence is large (e.g., about 1 degree or more for our specific simulation parameters), then both of the single arm correction schemes show very little improvement in the polarization, so both arms must be corrected. The alignment of the correction magnets is also important, with more precision required for higher Fourier times. For the simulation parameters used above, all correction magnets must be aligned within approximately $\pm 0.5$ mm in both the $y$ and $z$ directions.
\section{Experiment Results} \label{sec:experiment}
\begin{figure}[b]
\centering
\includegraphics[width=.95\linewidth]{Figures/beamline_diagram_sep21.png}
\caption{\label{figBeamline} Schematic of the experimental test of the correction magnet (CM). The beam travels from left to right. Precession occurred inside both the CM and guide field (GF).}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=.95\textwidth]{Figures/2d_analysis_10amp_sep12-eps-converted-to.pdf}
\caption{\label{figCorrAnalysis} Data with a current of -10 amps in the NRSE correction magnet. (a) The intensity vs. position recorded by the Anger camera when the guide field coil had a current of 1.11 amps. (b) A cosine fit of the intensity vs. guide field coil current for several pixels along the line $z=0$. (c) The phase and (d) phase error extracted from the cosine fit shown for each pixel. (e) The quadratic fit of the phase data to Eqn. \eqref{Quadratic2d} and (f) the quadratic fit subtracted from the phase data. }
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{Figures/quadratic_fig_sep2-eps-converted-to.pdf}
\caption{\label{figCorrAnalysis1d} (a) Horizontal and (b) vertical slices through the magnetic center for different currents in the correction magnet. Data are fit to a parabola.}
\end{figure}
A measurement of the field integral through the prototype correction magnet was performed on the cold-neutron, polarized test beamline CG4B at the High Flux Isotope Reactor (HFIR) at Oak Ridge National Laboratory (ORNL). As shown in Fig. \ref{figBeamline}, the correction magnet was installed in a vacuum chamber in front of a guide field magnet that also generated a field in the vertical, $z$-direction. The guide field magnet had a front and back HTS film, with the front film also serving as the back film of the correction magnet. The vertical guide field magnitude was designed to be spatially uniform, as the neutron will continue to precess in it.
S-benders served as the neutron polarizer and analyzer. A horizontal guide field outside of the correction magnet and the non-adiabatic field transition through the HTS film induced precession inside both the correction magnet and the spatially uniform guide field. Precession was stopped by another horizontal guide field after the back HTS film of the guide field. The beam size was determined by a square 1 by 1 cm slit located 1 meter in front of the correction magnet and a square 2.5 by 2.5 cm slit at the end of the guide field coil. The wavelength was 0.55 nm with a FWHM wavelength spread of less than 1$\%$.
The variation in the Larmor phase across the beam was measured across the two-dimensional detector. The detector was an Anger camera with 1.8 mm pixel size, as shown in Fig. \ref{figCorrAnalysis}(a). \cite{Riedel2015, Cao2018} From the detector image in Fig. \ref{figCorrAnalysis}(a), one can directly see an approximately ``bullseye'' shaped signal, suggesting a radial dependence of the Larmor phase.
The Larmor phase was measured by setting a current in the correction magnet and scanning the precessing guide field between 1 and 1.12 amps. This current range varies the phase of neutrons passing through the magnet by about $2 \pi$, as shown in Fig. \ref{figCorrAnalysis}(b). The difference in the phase of the curves shown in Fig. \ref{figCorrAnalysis}(b) shows the different Larmor phase acquired by neutrons traveling through the correction magnet. The intensity recorded in each pixel varies greatly due to the spatial non-uniformity in the CG4B beam intensity as well as the non-uniform detector efficiency.
The intensity $N$ as a function of pixel $(y,z)$ was fit to
\begin{equation}
N(y,z) = \alpha + \beta \cos[\phi + f_g(I - I_0)],
\end{equation}
where $\alpha$ and $\beta$ are fitting parameters, $\phi$ is the Larmor phase from the correction magnet, $f_g$ is the frequency of the oscillation in the polarization due to the Larmor phase produced by the guide field, $I$ is the current in the guide field, and $I_0$ is the current at the start of the scan (1 amp in this case).
The polarization of the signal is defined as $\beta/\alpha$. We fit the relative phase compared to the center which we set as zero.
The phase data for the phase $\phi$ were fit to the following two-dimensional quadratic function:
\begin{equation} \label{Quadratic2d}
\phi(y,z) = \phi_2 [(y - y_0)^2 + \epsilon(z-z_0)^2] + \phi_0,
\end{equation}
where $\phi_2$ and $\phi_0$ are fitting parameters and $\epsilon$ the eccentricity term which allows for a difference in the $y$ and $z$ correction terms, and $(y_0, z_0)$ the beam center. The fit is shown in Fig. \ref{figCorrAnalysis}(e). Subtracting the quadratic fit from the phase data gives the accessible corrected beam size, shown in Fig. \ref{figCorrAnalysis}(f) to be about 2 centimeters. MagNet simulations show that the eccentricity term can be tuned to unity by varying the chevron angle and back film separation distance.
There are several unexpected features in the data. Most notably, the center of the beam is not the same as the center of the quadratic fit, which we call the magnetic center. This discrepancy is possibly due to a misalignment of the beam mask and the correction magnet. Additionally, the positive $y$ side of the magnet has a larger discrepancy in the data-fit compared to the negative side. While these features are surprising, the most likely source for the off-center signal is due to misalignment between the beam apertures and the correction magnet, and the non-homogeneous signal is possibly due to a magnetic inhomogeneity in the soft-iron used in the pole and chevron pieces.
An additional feature is the non-radial dependence (i.e. the diamond shape of the fitted phase) of the data at large $(y,z)$. This feature can partially account for where the quadratic fit fails to match the data in Fig. \ref{figCorrAnalysis}(f). It can also be seen in the MagNet simulations of the field integral, suggesting that it is a result of the chevron design. A more sophisticated pole piece shape may be required to adjust the field integral into the proper quadratic shape for larger beam sizes.
In order to compare different currents in the correction magnet, we fit a vertical and horizontal slice through the magnetic center to a parabola as displayed in Fig. \ref{figCorrAnalysis1d}. The offset in $y$ of 3 mm remains approximately constant for all currents, which is consistent with the conclusion of misalignment between the correction magnet and the beam. The quadratic coefficient divided by the current should be the same for all currents if the field is generated solely by the current in the correction magnets. However the phase change at 15 amps is only 2.6 times the variation at 5 amps. This difference is possibly due to hysteresis effects and domain formation in the soft-iron inside of the correction magnet. It may also be due to the coupling of the field in the correction magnet to the external guide fields, although MagNet simulations show very little coupling.
\section{Discussion and Conclusion} \label{sec:dis}
This correction technique is for transverse-field NRSE instruments, while Fresnel coils may be installed for longitudinal NRSE.
An advantage of this device compared to Fresnel coils is the small amount of neutron-absorbing or scattering material in the beam. There is a 100 nm film of gold coating a 350 nm film of YBCO on a 0.5 mm sapphire substrate. With a 1 nm neutron wavelength, 12 of these films will have a transmission of $\sim 95\%$. Fresnel coils add at least 2 cm of aluminum wire which has a transmission of $\sim 86\%$ for 1 nm neutron wavelength. The exact amount of material for a Fresnel coil depends on the required current, so reaching a higher Fourier time generally produces more background scattering. However, the Fresnel coils have a long history of being successfully used to correct for NSE and have been built to accommodate much larger beam sizes.
One of the reasons longitudinal NRSE is preferred for the Reseda instrument at FRM-II is the historical difficulty in correcting transverse NRSE path length aberrations. \cite{Franz2019} If this correction technique can reach the same performance as Fresnel coils, the choice of longitudinal or transverse-NRSE will be more complicated: transverse rf flippers offer the opportunity to have a higher effective frequency in ``bootstrap'' mode, \cite{Li2020} while longitudinal rf flippers have a proven history of high performance.\cite{Franz2019}
Another method of correcting divergent neutrons has recently been installed at VIN-ROSE in JPARC, which addresses the same problem by adding elliptical mirrors to each arm so that all neutron paths will be the same length. \cite{Endo2019} To our knowledge, this correction magnet has not yet been used to correct for scattering from different points in the sample.
We have demonstrated a theory, simulation, and experiment of correcting aberrations caused from path deviations in a transverse NRSE beamline.
Simulation shows that an arrangement of six correction magnets maintains a high polarization even for 3 cm beam sizes and large rf flipper frequencies.
Experimental tests of the prototype magnet show that shaping the pole pieces and separating the coil from the back HTS film allow for the independent control of the $y$ and $z$ parameters of a quadratic field integral.
The most pressing improvements to future designs are more careful alignment, more magnetically-uniform material for pole pieces, and acceptance of larger beam sizes.
With these improvements, this type of correction magnet is ready to benefit transverse NRSE beamlines.
\section{Acknowledgements}
The authors would like to thank Lowell Crow, Georg Ehlers, Fumiaki Funama, and Steven Parnell for useful discussions. CAD drawings were made by Jak Doskow and machining was done by the Indiana University Physics machine shop: John Frye, Danny Clark, Darren Nevitt, and Todd Sampson. We thank Matthew Loyd for assistance with the Anger camera.
This research used resources at the High Flux Isotope Reactor, a DOE Office of Science User Facility operated by the Oak Ridge National Laboratory. F. Li would also like to acknowledge the support from DOE Early Career Research Program Award (KC0402010), under Contract No. DE-AC05- 00OR22725.
The work reported here was funded by the Department of Energy STTR program under grants DE-SC0021482 and DE-SC0018453.
A number of the authors acknowledge support from the US Department of Commerce through co- operative agreement number 70NANB15H259.
\bibliographystyle{apsrev4-2.bst}
\section{Introduction}
\begin{figure*}
\centering
\includegraphics[width=.95\textwidth]{Figures/Schematic_NRSECorr_sep12-eps-converted-to.pdf}
\caption{ \label{figSchematic} Schematic of the corrected neutron resonance spin echo beamline. The neutron is polarized at the far left (Pol.) before traveling through the first arm, scattering from the sample (Sam.) at angle $\theta$, echoing through the second arm, and entering the analyzer (Ana.) at the far right; the detector is not shown. The correction magnets are labeled ``CM'' and the rf flippers ``RF''. The white space between the CMs is at zero field. The scattering plane for the example neutron path shown is the $x$-$y$ plane.
For the first arm, $L_n$ for $n=1,2,3$ is the distance from the beam-defining slit to the center of the corresponding correction magnet, while for the second arm, $L_n$ for $n = 4,5,6$ is the distance from the sample to each correction magnet center. The distance from the optical axis to the point that the neutron passes through the $n^{\mathrm{th}}$ correction magnet is defined as $y_n$. The distance between rf flippers in both arms is $L_{\mathrm{RF}}$.
}
\end{figure*}
Neutron Resonance Spin Echo (NRSE) is a modification of the Neutron Spin Echo (NSE) technique which replaces the static-field precession coils with radio-frequency (rf) spin-flippers. \cite{Golub1987} The underlying principle of both types of \textit{echo} measurements is that the instrument will measure the change in velocity, and hence the change in energy, of a neutron scattered from a sample.
Currently, most large neutron sources use static-field NSE instruments for high-energy-resolution measurements of slow dynamics.
In order to be competitive with existing NSE instruments, NRSE would need to achieve a Fourier time (also called the spin echo time) of about one hundred nanoseconds. \cite{Farago2015,Ohl2012}
The Fourier time $\tau$ is given by
\begin{equation}
\tau = \frac{2 m^2}{h^2} L_{\mathrm{RF}} f \lambda^3
\label{EqnFouriertime}
\end{equation}
where $L_{\mathrm{RF}}$ is the distance between the rf flippers in each arm, $f$ is the rf flipper (linear) frequency, $\lambda$ is the neutron wavelength, $m$ is the neutron mass, and $h$ is Planck's constant.\cite{Keller2002}
State-of-the-art rf flippers are already capable of producing a high performance NRSE instrument. As an example, suppose an NRSE beamline uses recently developed transverse rf flippers. \cite{Li2020} A beamline with those rf flippers operating at 4 MHz, with a 2 meter separation between the rf flippers, and with a standard NSE wavelength of 1 nm would have a Fourier time of 100 nanoseconds, which is comparable to modern NSE beamlines.\cite{Golub1988}
However, due to the long wavelength requirements for both NSE and NRSE, the neutron flux is often low and the relevant samples scatter weakly. Therefore, measurements are only possible by having a large spatial and angular beam size, which leads to aberrations. In conventional NSE, one aberration source is due to the variation in the static field strength across the beam due to the field profile created by a solenoid geometry.
In NRSE, the rf flippers are separated by zero field regions, so this aberration is not present. A second type of aberration arises from scattering from a sample. The sample is placed in the center of the two symmetric arms as shown in Fig. \ref{figSchematic}. If the neutron scatters with some non-zero momentum transfer, then the path length through the second arm will not be the same as the path length through the first arm, so the neutron will spend a different amount of time in the two arms.
Because NRSE instruments measure the velocity change of a neutron by measuring its Larmor phase $\Phi = 4 \pi f t$, where $t$ is the time it takes for the neutron to travel between the rf flippers, a difference in time is measured as a change in the neutron velocity. An uncorrected echo measurement will then conflate a change in scattering angle with a change in energy. The time can be written in terms of the neutron path length between the rf flippers as
\begin{equation}
t = \frac{L_{\mathrm{RF}}}{v \cos\theta},
\end{equation}
where $v$ is the neutron velocity and $\theta$ is the scattering angle in the scattering plane (see Fig. \ref{figSchematic}).\cite{Keller2002}
Thus, for our example NRSE beamline, a neutron scattering at 1 degree would be out of Larmor phase by more than 2000 degrees compared to the unscattered neutrons if there were no correction.
This aberration will be present for elastic, quasielastic, and inelastic neutron scattering.
Clearly, an NRSE instrument must have a method for correcting this geometric contribution to the Larmor phase. In NSE instruments, Fresnel coils with longitudinal-fields (i.e. the field is orientated along the optical axis) are used for an analogous correction as well as correcting the aberration from the static field variation. \cite{Ohl2005,Ohl2012,Farago2015} However, transverse-field rf flippers (i.e. the rf flipper's static-field is perpendicular to the optical axis) have been constructed for NRSE measurements, \cite{Li2020, Endo2019} which require a transverse-field correction magnet to improve the polarization.
In this paper, we present an analytical solution of the magnetic field profile needed to correct the path length aberrations and the design of a suitable prototype correction magnet. The Larmor phase that a neutron acquires traveling through our prototype NRSE correction magnet varies quadratically with the distance from the optical axis and is radially symmetric. We simulate an NRSE beamline with the correction magnets and experimentally measure the spatial dependence of the Larmor phase change through the device. We demonstrate that transverse-field NRSE beamlines can be corrected with transverse static-field magnets, and therefore have the potential to be competitive with NSE beamlines.
\section{Analytical Solution} \label{sec:ana}
The two arms of the NRSE instrument will be corrected independently; we will discuss the correction for the second arm first.
The necessary magnitude of the correction is proportional to the path~length~difference~$\Delta L_{\mathrm{RF}}$ between the two pairs of rf flippers, which is given by
\begin{equation} \label{pathlength}
\Delta L_{\mathrm{RF}} = L_{\mathrm{RF}} \left(\frac{1}{\cos\theta} - 1\right) \approx \frac{1}{2}L_{\mathrm{RF}}\theta^2,
\end{equation}
where $\theta$ is the small scattering angle shown in Fig. \ref{figSchematic}, defined relative to the optical axis.
This difference in path length will cause a delay in time, and thus a difference in the Larmor phase $\Phi$.
The difference in Larmor phase $\Delta \Phi$ between the unscattered and scattered beam for idealized rf flippers in the NRSE configuration is\cite{Keller2002}
\begin{equation} \label{deltaLinital}
\Delta \Phi = \frac{4 \pi f}{v} \Delta L_{\mathrm{RF}} \approx 2 \pi f \frac{\lambda m}{h} L_{\mathrm{RF}} \theta^2.
\end{equation}
The Larmor phase has already been defined as $\Phi = 4 \pi f t$ for an NRSE instrument, but it can also be defined for NSE instruments in terms of the magnetic field integral: $\Phi = (\gamma/v) \int ds \, B$, where $\mathrm{FI}_s = \int ds \, B$ the field integral experienced by the neutron traveling along the path $s$ and $\gamma \approx -1.832 \times 10^8$ rad/(T$\cdot$s) is the neutron's gyromagnetic ratio.
In NSE, the neutron magnetic moment rotates (precesses) in the plane perpendicular to an applied static field and the Larmor phase measures the amount of this rotation relative to some fixed direction in the lab. In NRSE, the rf field rotates in the plane perpendicular to a static field, and the Larmor phase measures that angle between the neutron magnetic moment and the rf field.
Hence, rf and static field effects on the neutron can be added together, as has been exploited recently.\cite{Jochum2020b}
To correct the phase difference, we must design a correction scheme consisting of static magnetic fields that generates a Larmor phase proportional to $\theta^2$, with the proportionality coefficient $\chi$ being
\begin{equation} \label{CC}
\chi = 2 \pi \frac{f}{\gamma} L_{\mathrm{RF}}.
\end{equation}
Notice that the required correction is independent of wavelength. With this correction, the Larmor phase of all diverging
neutrons will be corrected as if they traveled the same effective path length, namely the distance $L_{\mathrm{RF}}$.
Unfortunately, it is not obvious how to design a static magnetic field profile that would generate a purely $\theta^2$-dependent field integral term for a finite-sized beam.
However, we can generate such a term for a finite-sized beam with three correction magnets consisting of transverse static fields in which the magnitude of the field integral of a neutron traveling through the devices varies quadratically as a function of transverse position, with the center being the minimum and increasing radially outward. For scattering from a point-like sample, one can show that only two devices of this type are needed.
This design is a two-dimensional extension to the original solution proposed by Monkenbusch. \cite{Monkenbusch1999}
For simplicity, we only look at the aberrations in one dimension (transverse $y$ direction), but the following argument can be easily generalized to the entire two-dimensional plane perpendicular to the optical axis. The neutrons may be scattered from any position on the sample at a distance of $y_{\mathrm{sam}}$ from the optical axis into any angle $\theta$ defined relative to the optical axis.
If the sample were point-like, then the scattering angle $\theta$ would be defined just by the $y$ distance from the optical axis and the $\theta^2$ aberration would be known simply from the $x$ and $y$ position in the second arm. With a finite-size sample, the scattering position $y_{\mathrm{sam}}$ will add to the $y$ position from the scattering angle, so one device at a specific point along the beamline is not sufficient to correct the $\theta^2$ aberration.
To lowest order in scattering angle, the field integral per amp of a single prototype correction magnet is
\begin{equation} \label{correctiondeviceFIgeneral}
\mathrm{FI}_n = a_n + b_n y_n^2 + c_n y_n \theta,
\end{equation}
where $y_n$ is the distance between the neutron's path at the $n^{\mathrm{th}}$ correction magnet and the optical axis (see Fig. \ref{figSchematic}). There is no linear term in $y$ because the device is left-right symmetric; similarly, there is no linear term in $z$ due to the top-bottom symmetry. Each term in the field integral is proportional to the applied current, and $a_n$ has units of T$\cdot$m, $b_n$ has units of T/m, and $c_n$ has units of T/rad. Here $b_n$ is the term that corrects the path-length aberrations while the $a_n$ and $c_n$ terms appear because of the particular correction magnet geometry that we have chosen; higher order terms were found to have a negligible contribution to the field integral.
The transverse position that the scattered neutron passes through each correction magnet is given by
\begin{equation} \label{ydefinition}
y_n = y_{\mathrm{sam}}+ L_n \theta,
\end{equation}
where $\theta$ is assumed to be small and $L_n$ is the distance from the sample to the center of the $n$\textsuperscript{th} correction magnet, as shown in Fig. \ref{figSchematic}.
Combining Eqns. \eqref{correctiondeviceFIgeneral} and \eqref{ydefinition} for each device, we find the total field integral per amp $\mathrm{FI_T}$ experienced by a neutron arriving at the analyzer due to the three correction magnets to be
\begin{equation} \label{eqnLarmorPhaseCC}
\begin{aligned}
\mathrm{FI_T} = \sum_{n \in \{4,5,6\}} \big[& a_n + b_n y_{\mathrm{sam}}^2 + (c_n + 2 b_n L_n) y_{\mathrm{sam}} \theta \\
&+ L_n (c_n + b_n L_n) \theta^2 \big].
\end{aligned}
\end{equation}
The goal of the correction scheme is to have the coefficient of the $\theta^2$ term equal to Eqn. \eqref{CC} while having all other terms zero. Doing so, the series of correction magnets would correct the path-length aberration in the second arm of the instrument regardless of the $y_{\mathrm{sam}}$ position and without introducing any net field integral. These requirements on Eqn. \eqref{eqnLarmorPhaseCC} can be rewritten into several conditions:
\begin{gather*} \label{CoefficientCancellingEqns}
a_4 + a_5 + a_6 = b_4 + b_5 + b_6 = c_4 + c_5 + c_6 = 0 \\
b_4 L_4 + b_5 L_5 + b_6 L_6 = 0 \\
\chi = L_4 (c_4 + b_4 L_4) + L_5 (c_5 + b_5 L_5) + L_6 (c_6 + b_6 L_6).
\end{gather*}
Notice that if the sum of the currents through the three correction magnets is zero, then the first line will be satisfied.
Ignoring the $a_n$ terms for now, we solve this system of equations, obtaining
\begin{subequations} \label{SolutionEnc}
\begin{align}
b_4 &= \frac{\chi + c_4(L_5 - L_4) + c_6(L_5 - L_6)}{(L_4 - L_5)(L_4 - L_6)} \\
b_5 &= -\frac{\chi + c_4(L_5 - L_4) + c_6(L_5 - L_6)}{(L_4 - L_5)(L_5 - L_6)} \\
b_6 &= \frac{\chi + c_4(L_5 - L_4) + c_6(L_5 - L_6)}{(L_4 - L_6)(L_5 - L_6)} \\
c_5 &= - (c_4 + c_6),
\end{align}
\end{subequations}
with $c_4$ and $c_6$ being free parameters.
From this set of solutions, it is apparent that if we choose CM5, the fifth correction magnet, to be equidistant from CM4 and CM6 (such that $|L_4 - L_5| = |L_5 - L_6| = \delta L$) and also $c_4 = c_6$, then the angle-dependent $c_n$ terms cancel out, leaving $b_4 = b_6 = \chi/[2(\delta L)^2]$ and $b_5 = -\chi/(\delta L)^2$. Therefore, we can obtain our desired field integral by putting the same field in the first and last device and a field twice as large in the opposite direction in the middle device. We call this choice of fields the $(1,-2,1)$ configuration. In this configuration, the constant $a_n$ terms will also cancel out.
Next, we determine the magnitude required for $b_n$ for a realistic NRSE beamline. Plugging in Eqn. \eqref{CC} to the $b_n$ terms in Eqn. \eqref{SolutionEnc}, we see that
\begin{gather}
b_4 = \frac{\pi f L_{\mathrm{RF}}}{\gamma (\delta L)^2} \\
b_5 = -2 b_4, \quad b_6 = b_4. \nonumber
\end{gather}
To estimate the required field in the correction magnet, let $L_{\mathrm{RF}} = 2$ m, $\delta L = 1$ m, and $f = 4 $ MHz. Plugging in the numbers, we find $b_4 \approx -140$ mT/m. As we will show below, this value is attainable with our correction magnet.
The above discussion has only considered the second arm; now we look at the first arm.
If the initial beam is well-collimated (i.e. all neutrons in the first arm travel parallel to the optical axis), then no correction elements are required in the first arm even if there are correction magnets in the second arm. However, in practice, neutrons in a real instrument have some divergence angle relative to the optical axis.
This initial beam divergence will lead to a variation in path length between neutrons propagating at different angles in the first arm, similar to the scattering term for the second arm.
Therefore, we must correct for this variation in the first arm with another three correction magnets, as shown in Fig. \ref{figSchematic}. They must be in the $(-1,2,-1)$ configuration as the static magnetic field in both rf flippers is in the opposite direction relative the static fields in the rf flippers in the second arm.
Without this additional correction, we do not obtain the best possible improvement to the polarization.
The correction for the divergence angle of the neutron in the first arm will not require any changes to the correction magnet set-up in the second arm because the correction for the second arm is independent of angle or $y_{\mathrm{sam}}$ position.
With all six correction magnets installed, the Larmor phase, and hence Fourier time, of neutrons along any path in either arm will be corrected to the Larmor phase and Fourier time of a neutron traveling parallel to the optical axis.
\section{Development of the Correction Magnet} \label{sec:design}
\begin{figure}
\includegraphics[width=.95\linewidth]{Figures/CorrDevelopment_sep20-eps-converted-to.pdf}
\caption{\label{figCorrDevelopment} (a) CAD model of the correction magnet. The light blue surfaces are the high-temperature superconducting (HTS) films (front film not shown), the purple is the low-carbon steel chevron, the tan are the HTS coils, and the brown is the low-carbon steel flux return and pole pieces. (b and c) Simulation of fields at $y = 0$ and $z = 0$ with the coil current set to 10 amps, which have the same colorbar legend. The beam traveled from left to right. The HTS films are highlighted in white. (d) Contour plot of simulated field integral through the correction device at 10 amps for a neutron originating from a point source at $y = z = 0$ and $x = -20$ m. (e) Slices of the simulated field integrals in (d) through the origin.}
\end{figure}
To implement the analytical solution, we designed and constructed a vertical-field correction magnet following the drawing of Fig. \ref{figCorrDevelopment}(a). The bottom, top, and sides of the coils are enclosed by a magnetic flux return made of low-carbon steel (alloy 1018). High-temperature superconducting (HTS) films, gold-coated 350 nm thick YBCO on 0.5 mm thick sapphire substrate, were placed on the front and sides of the coils with another film placed 38 mm after the coils, outside of the magnetic circuit. These HTS films act as magnetic field screens due to the Meissner effect which creates sharp boundaries between field regions, as shown in Fig. \ref{figCorrDevelopment}(b,c).
Thus the magnet can be thought of in two parts: a contained region where the coils sit and an open region before the back film. HTS wire was wound around hollow low-carbon steel pole pieces and topped with ``chevrons'', low-carbon steel plates with v-shaped cutouts. The opening of the v is at the front of the device. The thickness of the chevrons was 3.2 mm, leaving a separation between chevrons of 50 mm. The angle of the chevron was 60 degrees, and the space from the coils to the rear HTS film was 38 mm. It was already well-known that a dipole magnet without a HTS film constraining the magnetic flux will create a magnetic field with a quadratic $z$-dependence, as correcting for this was one of the initial advantages of adding a HTS film. \cite{Wang2014}
The magnetic field in this device was simulated using the Siemens MagNet $\copyright$ software, which includes the material properties in its solutions via the finite-element method. The HTS films were simulated as perfect diamagnets preventing any perpendicular magnetic flux. We note that the explicit field profile in the device is arbitrary as long as the resulting field integral is quadratic across the device and the field direction does not change too quickly.
A useful feature of this design is that the $y$ and $z$ components of the field integral may be tuned independently. The $z$ component, shown in Fig. \ref{figCorrDevelopment}(b) is largely dependent on the distance to the back film while the $y$ component, shown in Fig. \ref{figCorrDevelopment}(c) is largely dependent on the chevron angle. As shown in \ref{figCorrDevelopment}(b), the quadratic behavior of the $z$ component comes from the bowing field lines protruding around the back of the coils due to the displaced back film.
A numerical solution to the field integral for any starting position $(y,z)$ and angle through the device was found by extracting the MagNet solution for the field at each point (mesh size 4 mm) and integrating the field along any chosen path. The simulated field integral through the coils at 10 amps for a neutron traveling from a far-away point source at $y= z =0$ and $x$ = -20 m is shown in Fig. \ref{figCorrDevelopment}(d). The difference in field integral between the center and edges is about 0.01 mT$\cdot$m, which approximately is the necessary correction value for an NRSE beamline with an rf flipper frequency of 1 MHz and 1 meter between the correction magnets. The field integral through the center is about 1.69 mT$\cdot$m.
\section{McStas Simulations of an NRSE Instrument} \label{sec:mcstas}
\FloatBarrier
Using McStas, a Monte Carlo neutron ray-tracing software package,\cite{Willendrup_Lefmann_2020, Willendrup_Lefmann_2021} we simulated an NRSE beamline with these correction magnets installed. The polarizer, rf flippers, analyzer, and detector were taken to be 100$\%$ efficient, while the correction magnet component was built using numerically simulated magnetic field data extracted from the MagNet simulations. The following model was used for the rf flipper at resonance:
\begin{equation}
\Phi_f = 2 \pi f (2t_i + \Delta t) -\Phi_i,
\end{equation}
where $\Phi_f$ is the final Larmor phase after exiting the rf flipper, $f$ the rf frequency (2 MHz for these simulations), $t_i$ the time at which the neutron enters the flipper, $\Delta t$ the time spent inside the flipper of thickness 15 cm, and $\Phi_i$ the initial Larmor phase when entering the flipper.
We used a modified version of the ``SANS\_spheres2'' sample, a default McStas sample that emulates elastically scattering hard spheres in a dilute solution. The sample parameters were chosen to prevent incoherent scattering or transmission without scattering, so the component acted like an idealized elastic scatterer with a maximum momentum transfer of 0.09 nm$^{-1}$, which corresponds to a maximum scattering angle of 0.7 degrees. The sample was 3 cm by 3 cm in transverse size with negligible thickness. The aperture diameter and the neutron wavelength were 2 cm and $0.8$ nm $ \pm 1\%$, respectively.
The separation between the rf flippers in each arm was taken to be 2.3 m, and the distance between the correction magnets 1 meter. The total distance from the source to the two-dimensional detector was 5.2 meters. With these parameters, the effective initial beam divergence is about 0.6 degrees.
Using the magnetic field data extracted from MagNet simulations, we determined that our prototype correction magnet had the following field integral per amp expansion across the correction magnet:
\begin{equation} \label{Eqn:fullFIE}
\mathrm{FI} = a + b_y y^2 + b_z z^2 + c_y y \theta + c_z z \psi,
\end{equation}
where $\psi$ is the vertical neutron divergence angle and the values of the fitted coefficients are given in Tab. \ref{tab:FIE CC coefficients} below.
Higher order terms were found to have a negligible contribution.
\renewcommand{\arraystretch}{2}
\begin{table}[h]
\centering
\newcolumntype{R}{>{\centering\arraybackslash}X}
\begin{tabularx}{.99\linewidth}{R|R|R|R|R}
$a \left(\frac{\mathrm{mT}\cdot \mathrm{m}}{\mathrm{A}}\right)$ &
$b_y \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{m}}\right)$ &
$b_z \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{m}}\right)$ &
$c_y \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{rad}}\right)$ &
$c_z \left(\frac{\mathrm{mT}}{\mathrm{A}\cdot\mathrm{rad}}\right)$ \\
\hline
0.169 & 8.78 & 8.97 & -0.525 & 0.986
\end{tabularx}
\renewcommand{\arraystretch}{1}
\caption{\label{tab:FIE CC coefficients}
Values of the field integral expansion coefficients in Eqn. \eqref{Eqn:fullFIE}. The longitudinal length of the magnet is 14 cm.}
\end{table}
\begin{figure}[!b]
\centering
\includegraphics[width=.95\linewidth]{Figures/Correction_coil_mcstas_tinyinset.png}
\caption{\label{figNRSEmcstas} Simulated McStas polarization for an NRSE beamline with all six correction magnets on (red curve), only the final three correction magnets on (green curve), only the first three correction magnets on (orange curve), and no correction magnets on (blue curve).
The inset compares the polarization for all correction magnets on (red curve) to all correction magnets off and sample removed (purple curve) which shows that the polarization drops to about 0.9998 at the edge of the detector when both arms are corrected.}
\end{figure}
The NRSE beamline was simulated in McStas both with and without the correction magnets.
A plot comparing the simulated echo polarizations vs. radial position on the detector is shown in Fig. \ref{figNRSEmcstas}.
These simulations confirm that the inclusion of the correction elements greatly increases the polarization, especially for larger scattering angles which correspond to the edges of the detector.
Correcting only the second arm of the beamline improves the polarization for the larger scattering angles, although the polarization at the center of the detector is worsened compared to the uncorrected simulation due to the initial neutron divergence angle. However, if the initial beam divergence is large (e.g., about 1 degree or more for our specific simulation parameters), then both of the single arm correction schemes show very little improvement in the polarization, so both arms must be corrected. The alignment of the correction magnets is also important, with more precision required for higher Fourier times. For the simulation parameters used above, all correction magnets must be aligned within approximately $\pm 0.5$ mm in both the $y$ and $z$ directions.
\section{Experiment Results} \label{sec:experiment}
\begin{figure}[b]
\centering
\includegraphics[width=.95\linewidth]{Figures/beamline_diagram_sep21.png}
\caption{\label{figBeamline} Schematic of the experimental test of the correction magnet (CM). The beam travels from left to right. Precession occurred inside both the CM and guide field (GF).}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=.95\textwidth]{Figures/2d_analysis_10amp_sep12-eps-converted-to.pdf}
\caption{\label{figCorrAnalysis} Data with a current of -10 amps in the NRSE correction magnet. (a) The intensity vs. position recorded by the Anger camera when the guide field coil had a current of 1.11 amps. (b) A cosine fit of the intensity vs. guide field coil current for several pixels along the line $z=0$. (c) The phase and (d) phase error extracted from the cosine fit shown for each pixel. (e) The quadratic fit of the phase data to Eqn. \eqref{Quadratic2d} and (f) the quadratic fit subtracted from the phase data. }
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=.95\linewidth]{Figures/quadratic_fig_sep2-eps-converted-to.pdf}
\caption{\label{figCorrAnalysis1d} (a) Horizontal and (b) vertical slices through the magnetic center for different currents in the correction magnet. Data are fit to a parabola.}
\end{figure}
A measurement of the field integral through the prototype correction magnet was performed on the cold-neutron, polarized test beamline CG4B at the High Flux Isotope Reactor (HFIR) at Oak Ridge National Laboratory (ORNL). As shown in Fig. \ref{figBeamline}, the correction magnet was installed in a vacuum chamber in front of a guide field magnet that also generated a field in the vertical, $z$-direction. The guide field magnet had a front and back HTS film, with the front film also serving as the back film of the correction magnet. The vertical guide field magnitude was designed to be spatially uniform, as the neutron will continue to precess in it.
S-benders served as the neutron polarizer and analyzer. A horizontal guide field outside of the correction magnet and the non-adiabatic field transition through the HTS film induced precession inside both the correction magnet and the spatially uniform guide field. Precession was stopped by another horizontal guide field after the back HTS film of the guide field. The beam size was determined by a square 1 by 1 cm slit located 1 meter in front of the correction magnet and a square 2.5 by 2.5 cm slit at the end of the guide field coil. The wavelength was 0.55 nm with a FWHM wavelength spread of less than 1$\%$.
The variation in the Larmor phase across the beam was measured across the two-dimensional detector. The detector was an Anger camera with 1.8 mm pixel size, as shown in Fig. \ref{figCorrAnalysis}(a). \cite{Riedel2015, Cao2018} From the detector image in Fig. \ref{figCorrAnalysis}(a), one can directly see an approximately ``bullseye'' shaped signal, suggesting a radial dependence of the Larmor phase.
The Larmor phase was measured by setting a current in the correction magnet and scanning the precessing guide field between 1 and 1.12 amps. This current range varies the phase of neutrons passing through the magnet by about $2 \pi$, as shown in Fig. \ref{figCorrAnalysis}(b). The difference in the phase of the curves shown in Fig. \ref{figCorrAnalysis}(b) shows the different Larmor phase acquired by neutrons traveling through the correction magnet. The intensity recorded in each pixel varies greatly due to the spatial non-uniformity in the CG4B beam intensity as well as the non-uniform detector efficiency.
The intensity $N$ as a function of pixel $(y,z)$ was fit to
\begin{equation}
N(y,z) = \alpha + \beta \cos[\phi + f_g(I - I_0)],
\end{equation}
where $\alpha$ and $\beta$ are fitting parameters, $\phi$ is the Larmor phase from the correction magnet, $f_g$ is the frequency of the oscillation in the polarization due to the Larmor phase produced by the guide field, $I$ is the current in the guide field, and $I_0$ is the current at the start of the scan (1 amp in this case).
The polarization of the signal is defined as $\beta/\alpha$. We fit the relative phase compared to the center which we set as zero.
The phase data for the phase $\phi$ were fit to the following two-dimensional quadratic function:
\begin{equation} \label{Quadratic2d}
\phi(y,z) = \phi_2 [(y - y_0)^2 + \epsilon(z-z_0)^2] + \phi_0,
\end{equation}
where $\phi_2$ and $\phi_0$ are fitting parameters and $\epsilon$ the eccentricity term which allows for a difference in the $y$ and $z$ correction terms, and $(y_0, z_0)$ the beam center. The fit is shown in Fig. \ref{figCorrAnalysis}(e). Subtracting the quadratic fit from the phase data gives the accessible corrected beam size, shown in Fig. \ref{figCorrAnalysis}(f) to be about 2 centimeters. MagNet simulations show that the eccentricity term can be tuned to unity by varying the chevron angle and back film separation distance.
There are several unexpected features in the data. Most notably, the center of the beam is not the same as the center of the quadratic fit, which we call the magnetic center. This discrepancy is possibly due to a misalignment of the beam mask and the correction magnet. Additionally, the positive $y$ side of the magnet has a larger discrepancy in the data-fit compared to the negative side. While these features are surprising, the most likely source for the off-center signal is due to misalignment between the beam apertures and the correction magnet, and the non-homogeneous signal is possibly due to a magnetic inhomogeneity in the soft-iron used in the pole and chevron pieces.
An additional feature is the non-radial dependence (i.e. the diamond shape of the fitted phase) of the data at large $(y,z)$. This feature can partially account for where the quadratic fit fails to match the data in Fig. \ref{figCorrAnalysis}(f). It can also be seen in the MagNet simulations of the field integral, suggesting that it is a result of the chevron design. A more sophisticated pole piece shape may be required to adjust the field integral into the proper quadratic shape for larger beam sizes.
In order to compare different currents in the correction magnet, we fit a vertical and horizontal slice through the magnetic center to a parabola as displayed in Fig. \ref{figCorrAnalysis1d}. The offset in $y$ of 3 mm remains approximately constant for all currents, which is consistent with the conclusion of misalignment between the correction magnet and the beam. The quadratic coefficient divided by the current should be the same for all currents if the field is generated solely by the current in the correction magnets. However the phase change at 15 amps is only 2.6 times the variation at 5 amps. This difference is possibly due to hysteresis effects and domain formation in the soft-iron inside of the correction magnet. It may also be due to the coupling of the field in the correction magnet to the external guide fields, although MagNet simulations show very little coupling.
\section{Discussion and Conclusion} \label{sec:dis}
This correction technique is for transverse-field NRSE instruments, while Fresnel coils may be installed for longitudinal NRSE.
An advantage of this device compared to Fresnel coils is the small amount of neutron-absorbing or scattering material in the beam. There is a 100 nm film of gold coating a 350 nm film of YBCO on a 0.5 mm sapphire substrate. With a 1 nm neutron wavelength, 12 of these films will have a transmission of $\sim 95\%$. Fresnel coils add at least 2 cm of aluminum wire which has a transmission of $\sim 86\%$ for 1 nm neutron wavelength. The exact amount of material for a Fresnel coil depends on the required current, so reaching a higher Fourier time generally produces more background scattering. However, the Fresnel coils have a long history of being successfully used to correct for NSE and have been built to accommodate much larger beam sizes.
One of the reasons longitudinal NRSE is preferred for the Reseda instrument at FRM-II is the historical difficulty in correcting transverse NRSE path length aberrations. \cite{Franz2019} If this correction technique can reach the same performance as Fresnel coils, the choice of longitudinal or transverse-NRSE will be more complicated: transverse rf flippers offer the opportunity to have a higher effective frequency in ``bootstrap'' mode, \cite{Li2020} while longitudinal rf flippers have a proven history of high performance.\cite{Franz2019}
Another method of correcting divergent neutrons has recently been installed at VIN-ROSE in JPARC, which addresses the same problem by adding elliptical mirrors to each arm so that all neutron paths will be the same length. \cite{Endo2019} To our knowledge, this correction magnet has not yet been used to correct for scattering from different points in the sample.
We have demonstrated a theory, simulation, and experiment of correcting aberrations caused from path deviations in a transverse NRSE beamline.
Simulation shows that an arrangement of six correction magnets maintains a high polarization even for 3 cm beam sizes and large rf flipper frequencies.
Experimental tests of the prototype magnet show that shaping the pole pieces and separating the coil from the back HTS film allow for the independent control of the $y$ and $z$ parameters of a quadratic field integral.
The most pressing improvements to future designs are more careful alignment, more magnetically-uniform material for pole pieces, and acceptance of larger beam sizes.
With these improvements, this type of correction magnet is ready to benefit transverse NRSE beamlines.
\section{Acknowledgements}
The authors would like to thank Lowell Crow, Georg Ehlers, Fumiaki Funama, and Steven Parnell for useful discussions. CAD drawings were made by Jak Doskow and machining was done by the Indiana University Physics machine shop: John Frye, Danny Clark, Darren Nevitt, and Todd Sampson. We thank Matthew Loyd for assistance with the Anger camera.
This research used resources at the High Flux Isotope Reactor, a DOE Office of Science User Facility operated by the Oak Ridge National Laboratory. F. Li would also like to acknowledge the support from DOE Early Career Research Program Award (KC0402010), under Contract No. DE-AC05- 00OR22725.
The work reported here was funded by the Department of Energy STTR program under grants DE-SC0021482 and DE-SC0018453.
A number of the authors acknowledge support from the US Department of Commerce through co- operative agreement number 70NANB15H259.
\bibliographystyle{apsrev4-2.bst}
|
\section{Introduction}
This paper is centered on the area of \emph{distributed graph algorithms} and provides new methods and tools for developing improved \emph{deterministic} distributed algorithms.
It has been a central, well-known, and well-studied theme in this area that, for many of the graph problems of interest, known randomized algorithms outperform their deterministic counterparts. Concretely, the randomized variants have been much faster and/or achieved better output properties, e.g., approximation factors. As a prominent example, for several of the key problems of interest---including maximal independent set, maximal matching, $\Delta+1$ vertex coloring---we have known $O(\log n)$ round randomized algorithms since the 1986 work of Luby~\cite{luby86}. In contrast, developing even $\mathrm{poly}(\log n)$-time deterministic algorithms for many of these problems remained open for nearly four decades. See for instance the 2013 book of Barenboim and Elkin~\cite{barenboimelkin_book} which lists numerous such open questions. Only very recently, $\mathrm{poly}(\log n)$-time deterministic algorithms for these problems were developed~\cite{rozhonghaffari20, GGR20, chang2021strong, GhaffariK21, elkin2022deterministic}. However, currently, these deterministic algorithms are still quite far from their randomized counterparts.
In this paper, we focus on two of the most central tools in developing deterministic algorithms for local graph problems, namely \emph{network decompositions} and \emph{hitting sets}, and we present significantly improved deterministic distributed constructions of these tools. From a technical perspective, our novelty is in developing new randomized algorithms for these tools in such a way that we can analyze the algorithm by assuming only pairwise independence in the randomness it uses. We then describe how one can leverage this to derandomize the algorithms, i.e., to transform the randomized algorithm into an efficient deterministic algorithm. We next review the model and then state our contributions in the context of the recent progress.
\paragraph{Model.} We work with the standard distributed message-passing model for graph algorithms~\cite{peleg00}. The network is abstracted as an $n$-node graph $G=(V, E)$ where each node $v\in V$ corresponds to one processor in the network. Communications take place in synchronous rounds. Per round, each processor/node can send an $O(\log n)$-bit message to each of its neighbors in $G$. This model is called $\mathsf{CONGEST}\,$. The relaxed variant of the model where we allow unbounded message sizes is called $\mathsf{LOCAL}\,$. At the end of the round, each processor/node performs some computations on the data that it holds, before we proceed to the next communication round.
A graph problem in this model is captured as follows: Initially, the network topology is not known to the nodes of the graph, except that each node $v\in V$ knows its own unique $O(\log n)$-bit identifier and perhaps some of the global parameters of the network, e.g., the number $n$ of nodes in the network or a suitably tight upper bound on it. At the end of the computation, each node should know its own part of the output, e.g., in the graph coloring problem, each node should know its own color. When we discuss a particular graph problem, we will specify what part of the output should be known by each node.
\subsection{Network Decomposition}
Perhaps the most central object in the study of deterministic distributed algorithms for local graph problems has been the concept of \emph{network decomposition}, which was introduced by Awerbuch, Luby, Goldberg, and Plotkin~\cite{awerbuch89}. We next define this concept and explain its usefulness. Then, we discuss its existence and randomized distributed constructions. Afterward, we review the deterministic distributed constructions, especially the recent breakthroughs, and state our contributions.
Generally, the vertices of any $n$-node network can be colored using $O(\log n)$ colors such that in the subgraph induced by each color, each connected component has diameter $O(\log n)$. We call this an $O(\log n)$-color $O(\log n)$-diameter network decomposition (or sometimes $O(\log n)$-color $O(\log n)$-strong-diameter network decomposition, to contrast it with a weaker variant which we discuss later). This decomposition enables us to think of the entire graph as a collection of $O(\log n)$ node-disjoint graphs, each of which has a small $O(\log n)$-diameter per component; the latter facilitates distributed coordination and computation in the component.
As a prototypical example, given such a network decomposition, one easily gets an $O(\log^2 n)$-round deterministic algorithm for maximal independent set in the $\mathsf{LOCAL}\,$ model: we process the color classes one by one, and per color, in each $O(\log n)$-diameter component, we add to the output a maximal independent set of the nodes of the component that do not have a neighbor in the independent sets computed in the previous colors. Each color is processed in $O(\log n)$ rounds, as that is the component diameter, and thus the overall process takes $O(\log^2 n)$ rounds. See \cite{rozhonghaffari20, ghaffari2017complexity, ghaffari2018derandomizing} for how network decomposition leads to a general derandomization method in the $\mathsf{LOCAL}\,$ model, which transforms any $\mathrm{poly}(\log n)$-time randomized algorithm for any locally checkable problem~\cite{naor95} (roughly speaking, problems in which any proposed solution can be checked deterministically in $\mathrm{poly}(\log n)$-time, e.g., coloring, maximal independent set, maximal matching) into a $\mathrm{poly}(\log n)$-time deterministic algorithm.
The existence of such a $O(\log n)$-color $O(\log n)$-diameter network decomposition follows by a simple ball-growing process~\cite{Awerbuch-Peleg1990}.
We build the colors one by one, and each time, we color at least half of the remaining nodes with the next color. For one color $i$, start from an arbitrary node and grow its ball hop by hop, so long as the size is increasing by at least a $2$ factor per hop. This stops in at most $O(\log n)$ hops. Once stopped, color the inside of the ball with the current color $i$, and remove the boundary nodes, deferring them to the next colors. If we continue doing this from nodes that remain in the graph, in the end, at least half of the nodes of the graph (which remained after colors $1$ to $i-1$) are colored in this color $i$, each carved ball has diameter $O(\log n)$, and different balls are non-adjacent as we remove their boundaries.
Linial and Saks~\cite{linial92} gave a randomized distributed algorithm that computes almost such a network decomposition in $O(\log^2 n)$ rounds of the $\mathsf{CONGEST}\,$ model. The only weakness was in the diameter guarantee: the vertices of each color are partitioned into non-adjacent clusters so that per cluster, every two vertices of this cluster have a distance of at most $O(\log n)$ in the original graph. This is what we call $O(\log n)$ weak-diameter. In contrast, if the distance was measured in the subgraph induced by the nodes of this color, it is called a \emph{strong-diameter}. A $O(\log^2 n)$-round $\mathsf{CONGEST}\,$-model randomized algorithm for $O(\log n)$-color $O(\log n)$-strong-diameter network decomposition was provided much later, by Elkin and Neiman~\cite{elkin16_decomp}, building on a parallel algorithm of Miller, Peng, and Xu~\cite{miller2013parallel}.
In contrast, even after significant recent breakthroughs, deterministic constructions for network decomposition are still far from achieving similar measures, and this suboptimality spreads to essentially all applications of network decomposition in deterministic algorithms. The original work of Awerbuch et al.~\cite{awerbuch89} gave a $T$-round deterministic $\mathsf{LOCAL}\,$ algorithm for $c$-color and $d$-strong-diameter network decomposition where $c=d=T=2^{O(\sqrt{\log n\log\log n})}$. All these bounds were improved to $c=d=T=2^{O(\sqrt{\log n})}$ by Panconesi and Srinivasan~\cite{panconesi-srinivasan}. A transformation of Awerbuch et al.~\cite{awerbuch96} in the $\mathsf{LOCAL}\,$ model can transform these into a $O(\log n)$-color $O(\log n)$-strong-diameter network decomposition, but the time complexity remains $2^{O(\sqrt{\log n})}$ and this remained the state of the art for over nearly three decades.
Rozho\v{n} and Ghaffari~\cite{rozhonghaffari20} gave the first $\mathrm{poly}(\log n)$ time deterministic network decomposition with $\mathrm{poly}(\log n)$ parameters. Concretely, their algorithm computes a $O(\log n)$-color $O(\log^3 n)$-weak-diameter network decomposition in $O(\log^8 n)$ rounds of the $\mathsf{CONGEST}\,$ model. The construction was improved to a $O(\log n)$-color $O(\log^2 n)$-weak-diameter network decomposition in $O(\log^5 n)$ rounds of the $\mathsf{CONGEST}\,$ model, by Grunau, Ghaffari, and Rozho\v{n}~\cite{GGR20}. Both of these constructions were limited to only a weak-diameter guarantee. If one moves to the relaxed $\mathsf{LOCAL}\,$ model with unbounded message sizes, then by combining these with a known transformation of Awerbuch et al.~\cite{awerbuch96}, one gets $O(\log n)$-color $O(\log n)$-strong-diameter network decompositions, in a time complexity that is slower by a few logarithmic factors. However, such a transformation was not known for the $\mathsf{CONGEST}\,$ model, until a recent work of Chang and Ghaffari~\cite{chang2021strong}. They gave a $\mathsf{CONGEST}\,$-model reduction, which can transform the weak-diameter construction algorithm of Grunau et al.~\cite{GGR20} into a strong-diameter one, sacrificing some extra logarithmic factors. Concretely, they achieved a $O(\log n)$-color $O(\log^2 n)$-strong-diameter decomposition in $O(\log^{11} n)$ rounds. The time complexity of decomposition with these parameters was improved very recently by Elkin et al.~\cite{elkin2022deterministic}, obtaining a $O(\log n)$-color $O(\log^2 n)$-strong-diameter decomposition in $O(\log^5 n)$ rounds.
However, all these constructions are still far from building the arguably right object, i.e., an $O(\log n)$-color $O(\log n)$-strong-diameter decomposition, in the $\mathsf{CONGEST}\,$ model. This was true even if we significantly relax the time complexity, and as mentioned before, this sub-optimality spreads to all applications.
\paragraph{Our contribution.} In this paper, we present a novel deterministic construction of network decomposition which builds \emph{almost the right object}, achieving an $O(\log n)$-color $O(\log n \cdot \log\log\log n)$-strong-diameter decomposition, in $\mathrm{poly}(\log n)$ rounds. We note that all previous construction techniques seem to require diameter at least $\Omega(\log^2 n)$; see \cite{chang2021strong} for an informal discussion on this. Our algorithm breaks this barrier and reaches diameter $O(\log n \cdot \log\log\log n)$. The key novelty is in designing a new randomized algorithm that can be analyzed using only pairwise independence. We can thus derandomize this algorithm efficiently by using previously known network decompositions in a black-box manner, and in $\mathrm{poly}(\log n)$ time.
Furthermore, if we want faster algorithms, by a black-box combination of our new construction with the technically-independent recent work of Faour, Ghaffari, Grunau, Kuhn, and Rozho\v{n}~\cite{Faour2022} on locally derandomizing pairwise-analyzed randomized algorithms (roughly speaking, their approach works by a specialized weighted defective coloring, instead of using network decompositions), our construction becomes much faster than all the previous constructions, and therefore provides the new state-of-the-art:
\begin{theorem}\label{thm:NetDecomp} There is a deterministic algorithm that, in any $n$-node network, computes an $O(\log n)$-color $O(\log n \cdot \log\log\log n)$-strong-diameter decomposition in $\widetilde{O}(\log^3 n)$ rounds\footnote{We use the notation $\widetilde{O}(f(x)) = O(f(x) \cdot \mathrm{poly}\log f(x))$. } of the $\mathsf{CONGEST}\,$ model. The algorithm performs $\widetilde{O}(m)$ computations in total, where $m$ denotes the number of edges.
\end{theorem}
\subsection{Hitting Set}
While network decomposition is a generic tool for derandomization in the $\mathsf{LOCAL}\,$ model, and also a key tool for derandomization in the $\mathsf{CONGEST}\,$ model with extensive applications, a more basic tool that captures the usage of randomness in a range of distributed algorithms is \emph{hitting set}, as we describe next.
\paragraph{The Hitting Set Problem (basic case).} Given a collection of \lq\lq large\rq\rq\, sets in a ground set of elements, randomness gives us a very simple way of selecting a ``small\rq\rq\, portion of the elements such that we have at least one member of each set. The most basic variant is this: consider a bipartite graph $G=(A\sqcup B, E)$ where each node on one side $A$ has degree at least $k$. By using randomness, we can easily define a small subset $B'\subseteq B$ which, with high probability, has size $O(|B|\log n/k)$ and hits/dominates $A$, that is, each node $a\in A$ has a neighbor in $B'$. For that, simply include each element of $B$ in $B'$ with probability $p=O(\log n/k)$. This randomized selection in fact works in zero rounds. Finding such a small subset $B'$ in a deterministic manner is a key challenge in designing efficient deterministic distributed algorithms for many problems. For instance, Ghaffari and Kuhn~\cite{ghaffari2018congest-derandomizing} pointed out that this is the only use of randomness in some classic randomized algorithms for the construction of spanners and approximations of set cover. Indeed, a variant of this hitting set problem is a key ingredient even in our construction of the network decompositions mentioned above.
\paragraph{The Hitting Set Problem (general case).} Generalizing the problem allows us to capture a much wider range of applications. In some applications, we need to consider different sizes of the sets. Furthermore, we may not need to hit all sets, but instead, we would like to minimize the number, or the total cost, of those not hit. Following the bipartite graph terminology mentioned above, suppose each node $a\in A$ has a cost $c_a$, and its degree is denoted by $d_a$. Randomized selection with probability $p$ picks a subset $B'\subseteq B$ of size $p|B|$, in expectation, where the total cost of $A$-nodes that do not have a neighbor in $B'$ is $\sum_{a\in A} c_a (1-p)^{\deg(a)}$, in expectation. As a side comment, we note that in all applications that we are aware of, we may assume that $c_a\in [1, \mathrm{poly}(n)]$. Because of this, essentially without loss of generality, we can assume that for each node $a\in A$ we have $\deg(a)\leq O(\frac{1}{p} \cdot \log n)$. This is because the total expected cost of higher degree nodes is $1/\mathrm{poly}(n)$, which is negligible.
As an instructive example application, by defining $c_a:=\deg(a)$, we get that the total number of edges incident on $A$-nodes that are not hit is at most $O(|A|/p)$, in expectation. This particular guarantee is the sole application of randomness in some distributed constructions, e.g., the celebrated spanner construction of Baswana and Sen~\cite{baswana2007simple}.
\paragraph{Prior deterministic distributed algorithms for hitting set.} There are two known distributed constructions for hitting set~\cite{ghaffari2018congest-derandomizing, bezdrighin2022deterministic}, as we review next. Both of these algorithms are based on showing that a small collection of random bits are sufficient for the randomized algorithm and then using the conditional expectation method to derandomize this. However, both algorithms are computationally inefficient and use superpolynomial-time computations.
Ghaffari and Kuhn~\cite{ghaffari2018congest-derandomizing} observed that $O(\log n)$-wise independence is sufficient for the randomized algorithm in the basic hitting set problem, and thus $O(\log^2 n)$ bits of randomness are sufficient for the algorithm. Then, given a network decomposition with $c$ colors and strong diameter $d$, we can derandomize these bits one by one in a total of $O(cd \log^2 n)$ rounds, by processing the color classes one by one and fixing the bits in each color class in $O(d)$ time. However, the resulting algorithm is not computationally efficient: Each node has to perform $n^{O(\log n)}$-time local computations to calculate the conditional probabilities needed in the method of conditional expectation.
Parter and Yogev~\cite{parter2018congested} pointed out that one can replace the $O(\log n)$-wise independence with a pseudorandomness generator for read-once DNFs and this reduces the number of bits to $O(\log n (\log\log n)^3)$--this was presented in a different context of spanners in the congested clique model of distributed computing. More recently, Bezdrighin et al.~\cite{bezdrighin2022deterministic} further reduced that bound to $O(\log n \log\log n)$, by applying a pseudorandomness generator of Gopalan and Yeudayoff~\cite{gopalan2020concentration}, which is particularly designed for hitting events. This decreased the round complexity slightly to $O(cd\log n\log\log n)$. However, the conditional probability computations still remain quite inefficient: they are $n^{O(\log\log n)}$-time, which is still super-polynomial.
\paragraph{Our contribution.}
Instead of viewing the randomized hitting set algorithm as a one-shot process, we turn it into a more gradual procedure. Concretely, we show that one can turn the natural randomized algorithm for the general hitting set problem into a number of randomized algorithms (bounded by $O(\log n)$), in such a way that pairwise independence is sufficient for analyzing each step. Because of this, we can derandomize each step separately (using an overall potential function that ensures that the result after derandomizing all steps has the same guarantees as discussed above for the randomized algorithm). Thanks to this, in contrast to the prior algorithms~\cite{ghaffari2018congest-derandomizing, bezdrighin2022deterministic} which required super-polynomial computations, our algorithm uses only $\widetilde{O}(m)$ computations, summed up over the entire graph, where $m$ denotes the number of edges. Hence, our distributed algorithm directly provides a near-linear time low-depth deterministic parallel algorithm for the hitting set problem.
\begin{theorem} [Informal] There is a deterministic distributed algorithm that in $\mathrm{poly}(\log n)$ rounds and using $\widetilde{O}(m)$ total computations solves the generalized hitting set problem. That is, in the bipartite formulation mentioned above, the selected subset $B'$ has size $O(p \cdot |B|)$ and the total weight of nodes of $A$ not hit by $B'$ is $O(\sum_{a\in A} c_a (1-p)^{\deg(a)})$.
\end{theorem}
We present the formal version of this theorem in \Cref{thm:hittingset-congest} for the $\mathsf{CONGEST}\,$ model of distributed computing, and in \Cref{thm:hittingset-pram} for the $\mathsf{PRAM}$ model of parallel computation.
\paragraph{Applications of hitting set.} This efficiently derandomized hitting set has significant applications for a number of graph problems of interest. In this paper, as two examples, we discuss spanners and distance oracles. In the case of spanners, this deterministic hitting set leads to the first deterministic spanner algorithm with the best-known stretch-size trade-off, polylogarithmic round complexity, that has near-linear time computations. The best previously known deterministic constructions required superpolynomial computations~\cite{bezdrighin2022deterministic} (and in the \cite{ghaffari2018congest-derandomizing} case, had extra logarithmic factors in size). The formal statements are as follows, and the proofs are presented in \Cref{subsec:spanners}.
\begin{corollary} (\textbf{general stretch spanners, unweighted and weighted})
There is deterministic distributed algorithm that, in $\mathrm{poly}(\log n)$ rounds of the $\mathsf{CONGEST}\,$ model and with total computations $\widetilde{O}(m)$, for any integer $k\geq 1$, computes a $(2k-1)$-spanner with $O(nk + n^{1 + 1/k} \log k)$ and $O(nk + n^{1 + 1/k} k)$ edges for unweighted and weighted graphs, respectively.
\end{corollary}
\begin{corollary} (\textbf{ultra-sparse spanners})
There is deterministic distributed algorithm that, in $\mathrm{poly}(\log n)$ rounds of the $\mathsf{CONGEST}\,$ model and with total computations $\widetilde{O}(m)$, computes a spanner with size $(1+o(1)) n$ and with stretch $\log n \cdot 2^{O(\log^* n)}$ in weighted graphs.
\end{corollary}
\medskip
By slight generalizations of our hitting set, we also obtain an efficient parallel derandomization of approximate distance oracles constructions:
\begin{corollary} (\textbf{approximate distance oracle})
Given an undirected weighted graph $G = (V,E)$, a set of sources $S \subseteq V$ with $s=|S|$, and stretch and error parameters $k$ and $\varepsilon > 0$, there is a deterministic algorithm that solves the source-restricted distance oracle problem with $\widetilde{O}_{\varepsilon}(ms^{1/k})$ work and $\widetilde{O}_{\varepsilon}(\mathrm{poly}(\log n))$ depth in the $\mathsf{PRAM}\,$ model. The data structure has size $O(nk s^{1/k})$ and for each query $(u,v)$, the oracle can return a value $q$ in $O(k)$ time that satisfies
\begin{equation*}
d(u,v) \leq q \leq (2k-1)(1 + \varepsilon)d(u,v).
\end{equation*}
\end{corollary}
The proof is presented in \Cref{subsec:distance-oracles}. The corresponding centralized randomized construction was presented in the celebrated work of Thorup and Zwick~\cite{thorup2005approximate}. A centralized derandomization was given by Roddity, Thorup, and Zwick~\cite{roditty2005deterministic} but that approach does not appear to be applicable in parallel/distributed settings of computation.
\section{Preliminaries}
\label{sec:preliminaries}
We use standard graph theoretic notation throughout the paper.
All graphs are undirected and unweighted.
For a graph $G = (V, E)$, we use $d_G$ or just $d$ to denote the distance metric induced by its edges. For sets of nodes $U,W \subseteq V(G)$, we generalize $d$ by $d(U, W) = \min_{u \in U, w \in W} d(u,w)$.
\paragraph{Clustering}
Given a graph $G$, its cluster $C$ is simply a subset of nodes of $V(G)$.
The strong-diameter $\textrm{diam}(C)$ of a cluster $C$ is defined as $\textrm{diam}(C) = \max_{u,v \in C} d_{G[C]}(u,v)$.
We note that there is a related notion of \emph{weak-diameter} of a cluster $C$ which is defined as the smallest $D$ such that $\forall u,v \in C : d_G(u,v) \le D$. That is, $C$ can even be disconnected, but there has to be a short path between any two nodes if we are allowed to use all nodes of $G$, not just nodes of $C$.
Although a cluster is simply a subset of $V(G)$, during the construction of a clustering, we keep its \emph{center} node $v \in C$ and often we work with an arbitrary breadth first search tree of $C$ from $v$.
The basic object we construct in this paper is \emph{separated clusterings}, which we formally define next.
\begin{definition}[$s$-separated clustering]
\label{def:clustering}
Given an input graph $G$, a \emph{clustering} $\mathcal{C}$ is a collection of disjoint \emph{clusters} $C_1, \dots, C_t$, such that for each $i$ we have $C_i \subseteq V(G)$. We say that the clustering has \emph{(strong-)diameter} $D$ whenever the diameter of each graph $G[C_i]$, $1 \le i \le t$, is at most $D$. We say that the clustering is $s$-separated if for every $1 \le i < j \le t$ we have $d_G(C_i, C_j) \ge s$. We sometime refer to this by saying that the clustering has separation $s$.
\end{definition}
We will also need the following non-standard notion of $s$-hop degree of a cluster defined as follows.
\begin{definition}[$s$-hop degree]
\label{def:shop_degree}
Let $\mathcal{C}$ be some clustering and $C \in \mathcal{C}$ be a cluster with a fixed spanning tree $T_C$ rooted at $r \in C$.
The $s$-hop degree of $C$ in $\mathcal{C}$ is the minimum number $d$ such that for each $u \in C$ and the unique path $P_u$ from $u$ to $r$ in $T_C$ the following holds: The number of different clusters $C' \in \mathcal{C}$ such that $d(P_u, C') \le s$ is at most $d$.
\end{definition}
The $s$-hop degree of a clustering $\mathcal{C}$ is the maximum $s$-hop degree over all clusters $C \in \mathcal{C}$.
\section{Improved Network Decomposition, Outline}
\label{sec:outline}
To prove \Cref{thm:NetDecomp}, our core result is captured by the following low-diameter clustering statement, which clusters at least half of the vertices. \Cref{thm:NetDecomp} follows directly by repeating this clustering for $O(\log n)$ iterations, each time in the graph induced by the nodes that remain unclustered in the previous iterations.
\begin{theorem}\label{thm:Clustering-main} There is a deterministic $\mathsf{CONGEST}\,$ algorithm that runs in $\widetilde{O}(\log^2 n)$ rounds and computes a clustering of at least $\frac{n}{2}$ nodes, with strong diameter $O(\log n \cdot \log\log\log n)$, and separation $2$.
\end{theorem}
There are three ingredients in proving \Cref{thm:Clustering-main}, as we discuss next:
\paragraph{(A) Low-Degree Clustering.} The most important ingredient, captured by \Cref{thm:low_degree_main} and proven in \Cref{sec:low_degree}, is a clustering that manages to cluster half of the vertices but in which we have relaxed the separation/non-adjacency requirement of the clustering. Instead, we want each cluster to have $\text{s}$-hop degree of at most $\lceil 100 \log \log (n) \rceil$. See \cref{def:shop_degree} for the definition.
For this ingredient, we present a randomized algorithm with pairwise analysis and then we derandomize it.
\begin{restatable*}{theorem}{lowdegreemain}
\label{thm:low_degree_main}
Let $\text{s} \geq 2$ be arbitrary. There exists a deterministic $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log^2(n))$ rounds which computes a clustering $\mathcal{C}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log(n))$,
\item $\text{s}$-hop degree of at most $\lceil 100 \log \log (n) \rceil$, and
\item the number of clustered nodes is at least $n/2$.
\end{enumerate}
\end{restatable*}
\paragraph{(B) From Low-Degree to Isolation.} The second ingredient, captured by \Cref{thm:subsampling_main} and proven in \Cref{sec:LowDegtoIsolation} is able to receive the clustering algorithm of (A) and turn it into a true clustering with separation $s$, but at the expense of reducing the number of clustered nodes by an $O(\log\log n)$ factor. For this ingredient as well, we first present a simple randomized algorithm with pairwise analysis, and then we derandomize it.
\begin{restatable*}{theorem}{subsamplingmain}
\label{thm:subsampling_main}
Assume we are given a clustering $\mathcal{C}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log (n))$ and
\item $\text{s}$-hop degree of at most $\lceil 100 \log \log (n) \rceil$.
\end{enumerate}
There exists a deterministic $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log^2(n))$ rounds which computes a clustering $\mathcal{C}^{out}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log (n))$,
\item separation of $\text{s}$ and
\item the number of clustered nodes is $\frac{|\mathcal{C}|}{1000 \log \log (n)}$.
\end{enumerate}
\end{restatable*}
\paragraph{(C) Improving Fraction of Clustering Nodes.} The third and last ingredient, captured by \Cref{thm:clusteringmorenodesmain} and proven in \Cref{sec:clusteringmorenodes}, receives the clustering algorithm of part (B) with a suitably high separation parameter (which is at least logarithmically related to the fraction of nodes clustered) and transforms it into a clustering of at least half of the nodes, at the expense of reducing the separation to simply $2$. This ingredient is a deterministic reduction and needs no derandomization and explains the final logarithm in the guarantees of \cref{thm:Clustering-main} (the first two logarithms are coming already from \cref{thm:low_degree_main}).
\begin{restatable*}{theorem}{clusteringmorenodesmain}
\label{thm:clusteringmorenodesmain}
Let $x \geq 2$ be arbitrary.
Assume there exists a deterministic $\mathsf{CONGEST}\,$ algorithm $\mathcal{A}$ running in $R$ rounds which computes a clustering $\mathcal{C}$ with
\begin{enumerate}
\item strong diameter $O(x \log n)$,
\item separation $10 \cdot x$ and
\item clustering at least $\frac{n}{2^x}$ nodes.
\end{enumerate}
Then, there exists a deterministic $\mathsf{CONGEST}\,$ algorithm $\mathcal{A}'$ running in $O(2^x (R + x \log n))$ rounds which computes a clustering $\mathcal{C}'$ with
\begin{enumerate}
\item strong diameter $O(x \log n)$,
\item separation $2$ and
\item clustering at least $\frac{n}{2}$ nodes.
\end{enumerate}
\end{restatable*}
\vspace{1em}
Having all three ingredients \cref{thm:low_degree_main,thm:subsampling_main,thm:clusteringmorenodesmain}, we simply put them all together to prove \cref{thm:Clustering-main}.
\begin{proof}[Proof of \Cref{thm:Clustering-main}]
Let $x=\lceil \log(2000 \log\log n)\rceil$. First, from \Cref{thm:low_degree_main}, we get a clustering of $n/2$ nodes with strong diameter $O(\log n\cdot \log\log\log n)$ and $10x$-hop degree at most $\lceil 100 \log \log (n) \rceil$, in $\widetilde{O}(\log^2 n)$ rounds. Feeding this clustering algorithm to \Cref{thm:subsampling_main} produces a clustering algorithm that clusters $\frac{n}{2000 \log\log n}$ nodes with strong diameter $O(\log n\cdot \log\log\log n)$ and separation $10 x$, in $\widetilde{O}(\log^2 n)$ rounds. Hence, this clustering can be put as input for \Cref{thm:clusteringmorenodesmain}, which as a result gives a clustering of at least $n/2$ nodes with strong diameter $O(\log n \cdot \log\log\log n)$, and separation $2$, in $\widetilde{O}(\log^2 n)$ rounds.
\end{proof}
\section{Low-Degree Clustering}
\label{sec:low_degree}
This section is devoted to proving the following theorem discussed in \cref{sec:outline}.
\lowdegreemain
\paragraph{Intuition Behind the Proof of \Cref{thm:low_degree_main}.}
In this paragraph, we give a brief intuition behind the proof of \cref{thm:low_degree_main}.
Our clustering algorithm can be viewed as derandomization of the randomized clustering algorithm of Miller, Peng, and Xu~\cite{miller2013parallel} (MPX). This is an algorithm that can cluster $n/2$ nodes with strong diameter $O(s \log n)$ such that the $s$-hop degree of the constructed clustering is in fact $1$, or in other words, the clustering is $s$-separated.
In the MPX algorithm, we simply run a breadth first search from all nodes of $V(G)$ at once, but every node starts the search only after a random delay computed as follows. Every node $v \in V(G)$ starts with the delay $\mathrm{del}(v) = O(s \log n)$. Next, each node starts flipping a coin and each time it comes up heads, it decreases its delay by $5s$. If it comes up tails, it stops the process. That is, the delays come from an exponential distribution; even more precisely, each node gets a head start coming from an exponential distribution, we talk about delays and add $O(s \log n)$ to make all numbers positive with high probability.
The guarantees of the MPX algorithm stem from the following observation. Let $u \in V(G)$ be arbitrary and let $\mathrm{wait}(u)$ be the first time $u$ is reached by above breadth first search with delays. Let $\mathrm{frontier}^{2s}(u)$ be the number of nodes $v \in V(G)$ such that $\mathrm{del}(v) + d(v,u) \le \mathrm{wait}(u) + 2s$. That is, $\mathrm{frontier}^{2s}(u)$ contains nodes who can reach $u$ after at most $2s$ additional steps after $u$ is reached for the first time.
We claim that with positive constant probability $\mathrm{frontier}^{2s}(u) = 1$, i.e., after the first node reaches $u$, it takes at least $2s$ additional steps until the next node reaches $u$.
To see this, replace each node $v \in V(G)$ by a runner on a real line who starts at position $d(u,v) + O(s \log n)$ (and may move toward left, as we soon discuss). Then, the exponential distribution that defines the delays corresponds to each runner flipping her coin until it comes up tails. For each heads, the runner runs distance $5s$ to the left.
We now let the runners flip the coins one by one. When a runner $r_j$ is flipping her coin, we consider the leftmost runner $r'_j$ out of the runners $r_1, \dots, r_{j-1}$ that already flipped their coins.
We observe that if $r_j$ at some point reaches a position at most $5s$ to the right from $r'_j$, we also have that $r_j$ runs to the distance $5s$ to the left of $r'_j$ with positive constant probability.
{\bf Derandomization: } Let us now explain the intuitive reason why we lose a factor of $O(\log \log n)$ in \cref{thm:low_degree_main}. Our derandomized algorithm simulates the coin flipping procedure step by step, for $O(\log n)$ steps, until every runner finally flips a tail and finishes.
In contrast to the previous simple algorithm, we now have to track our progress after every step.
So, our analysis is a derandomization of the following, different, randomized analysis of the same running process.
In this new randomized analysis, in each step $i$ and for each node $u$, we consider, very informally speaking, the event that the coin of all the runners that are currently at distance at most $2s$ from the leading runner comes up tail, where $t$ is a parameter we compute later.
The probability of this event is $2^{-t}$.
This means that the probability of this bad event happening in one of the $O(\log n)$ steps is at most $O(\log n) \cdot 2^{-t}$.
Choosing $t = O(\log \log n)$ makes this probability constant. Going back to the analysis of MPX, we get that at least half of nodes $u$ have $|\mathrm{frontier}^{2s}(u)| = O(\log \log n)$.
Although this new randomized analysis loses a factor of $O(\log \log n)$, we can derandomize it in this section by setting up suitable potentials and derandomizing the coin flips of each step. To do so, we in fact simulate one fully-independent coin flip of each node in $O(\log\log n)$ steps where in each step we only use pairwise-independent random bits.
The rest of the section is structured as follows. In \cref{lem:pairwise_del_clustering}, we show how computing suitable delays gives rise to the final clustering. This step is simple and does not rely on derandomization. \cref{thm:low_degree_delay_main} then shows how to compute the node delays that simulate the MPX analysis as discussed above.
To sample even one ``coin flip'' of MPX, we need to invoke $O(\log\log n)$ times the local derandomization lemma of \cite{Faour2022}. One call of this lemma corresponds to \cref{thm:deterministic-goodSet-selection}.
\paragraph{Basic Definitions}
To prove \Cref{thm:low_degree_main}, we first need to define the notions of delay, waiting time, and a frontier:
\begin{definition}[delay $\mathrm{del}$, waiting time $\mathrm{wait}_\mathrm{del}(u)$, and frontier $\mathrm{frontier}^{D}_{\mathrm{del}}(u)$]
\label{def:low_degree_del}
A delay function $\mathrm{del}$ is a function assigning each node $u \in V$ a \emph{delay} $\mathrm{del}(u) \in \{0,1,\ldots,O(\text{s} \log (n))\}$.
The waiting time of a node $u \in V$, with respect to a delay function $\mathrm{del}$, is defined as
%
\[\mathrm{wait}_{\mathrm{del}}(u) = \min_{v \in V} \left( \mathrm{del}(v) + d(v,u)\right).\]
The intuition behind $\mathrm{wait}_{\mathrm{del}}(u)$ is as follows: Assume that each node $v$ starts sending out a token at time $\mathrm{del}(v)$. Then, $\mathrm{wait}(u)$ is the time it takes until $u$ receives the first token.
Furthermore, for every parameter $D \geq 0$, the \emph{frontier of width $D$} of a node $u \in V$, with respect to a delay function $\mathrm{del}$, is defined as
\[\mathrm{frontier}^{D}_\mathrm{del}(u) = \{v \in V \colon \mathrm{del}(v) + d(v,u) \leq \mathrm{wait}_{\mathrm{del}}(u) + D\}.\] Informally, $\mathrm{frontier}^D(u)$ contains each node $v$ whose token arrives at $u$ at most $D$ time units after $u$ receives the first token.
\end{definition}
\smallskip
\paragraph{Clustering from given delays.} The delay of each vertex is computed by a procedure provided in \Cref{alg:low_degree_delay_alg}. Before discussing that, we first explain how each delay function $\mathrm{del}$, along with a separation parameter $s$, give rise to a clustering $\mathcal{C}^{\mathrm{del}}$: The clustering $\mathcal{C}^{\mathrm{del}}$ clusters all the nodes that have a small frontier of width $2 \text{s}$.
In particular, each node $u \in V(G)$ satisfying $|\mathrm{frontier}^{2\text{s}}(u)| \leq \lceil 100 \log \log(n)\rceil$ is included in some cluster of $\mathcal{C}^{\mathrm{del}}$.
More concretely, each clustered node $u$ gets clustered to the cluster corresponding to the node with the smallest identifier in the set $\mathrm{frontier}^0(u)$. In other words, $u$ gets clustered with the cluster of the minimizer of $\mathrm{wait}(u)$, where we use the smallest identifier to break ties. In the following text, we denote this node by $c_u$. See \cref{fig:mpx} for an illustration of this clustering.
\begin{figure}
\centering
\includegraphics[width = .6\textwidth]{img/mpx.eps}
\caption{The figure shows the run of the clustering algorithm constructing $\mathcal{C}^{\mathrm{del}}$.
The algorithm can be seen as starting a breadth first search from a single node $\sigma$ connected to every node $u \in V(G)$ with an edge of length $\mathrm{del}(u)$ (the $\mathsf{CONGEST}\,$ implementation of the algorithm does not need to simulate any such node $\sigma$).
The value $\mathrm{wait}(u)$ is the time until the search reaches the node $u$. The node that reaches $u$ the first is denote $c_u$.
Moreover, we cluster only nodes such that the size of their frontier of width $2s$ is at most $O(\log\log n)$. For example, the node $v$ is not clustered because after it is reached by the first node $c_v$, it is reached by $\Omega(\log \log n)$ other nodes in the following $2s$ steps.
It can be seen that for any $w$ on the path from $c_u$ to $u$, we have $\mathrm{frontier}^{2s}(w) \subseteq \mathrm{frontier}^{2s}(u)$, hence the constructed clusters are connected.
}
\label{fig:mpx}
\end{figure}
\begin{lemma}
\label{lem:pairwise_del_clustering}
Let $\mathrm{del}$ be a delay function and $\mathcal{C}^{\mathrm{del}}$ the corresponding clustering, as described above. Then, the clustering $\mathcal{C}^{\mathrm{del}}$ has
\begin{enumerate}
\item strong diameter $O(\text{s} \log n)$,
\item $\text{s}$-hop degree at most $\lceil 100 \log \log(n)\rceil$ and
\item the set of clustered nodes is equal to $|\{u \in V \colon \mathrm{frontier}^{2\text{s}}(u)| \leq \lceil 100 \log \log(n)\rceil\}|$.
\end{enumerate}
Moreover, the clustering $\mathcal{C}^{\mathrm{del}}$ can be computed in $O(\text{s} \log n \log\log n)$ $\mathsf{CONGEST}\,$ rounds.
\end{lemma}
To prove \Cref{lem:pairwise_del_clustering}, we first observe that the frontiers have the following property.
\begin{claim}
\label{cl:frontier}
Let $w$ be any node on a shortest path from $u$ to $c_u$ and let $D \ge 0$. Then, we have (I) $\mathrm{frontier}^D(w) \subseteq \mathrm{frontier}^D(u)$, and (II) $c_w = c_u$.
\end{claim}
\begin{proof}
First, we prove (I) $\mathrm{frontier}^D(w) \subseteq \mathrm{frontier}^D(u)$. Consider any $v \in \mathrm{frontier}^D(w)$. We prove that $v \in \mathrm{frontier}^D(u)$. Since $v \in \mathrm{frontier}^D(w)$, we have
$$\mathrm{del}(v) + d(v, w) \le \mathrm{del}(c_u) + d(c_u, w) + D.$$
Since $w$ lies on a shortest path from $c_u$ to $u$, we can add $d(w, u)$ to both sides of the equation to conclude that
$$\mathrm{del}(v) + d(v, w) + d(w, u) \le \mathrm{del}(c_u) + d(c_u, u) + D = \mathrm{wait}(u) + D,$$
and thus we have
$$\mathrm{del}(v) + d(v, u) \le \mathrm{wait}(u) + D.$$
Hence, we have $v \in \mathrm{frontier}^D(u)$ and (I) is proven.
Next, we prove (II) $c_w = c_u$. In view of the above proof of (I), it suffices to show that $c_u \in \mathrm{frontier}^0(w)$.
%
To prove $c_u \in \mathrm{frontier}^0(w)$, we use the fact that $c_u \in \mathrm{frontier}^0(u)$ and write
\[\mathrm{del}(c_u) + d(c_u,u) = \mathrm{wait}(u) \leq \mathrm{wait}(w) + d(w,u)\]
%
Subtracting $d(w, u)$ from both sides of the equation and using that $w$ lies on a shortest path from $u$ to $c_u$ gives
\[\mathrm{del}(c_u) + d(c_u,w) \leq \mathrm{wait}(w).\]
%
Thus $c_u \in \mathrm{frontier}^0(w)$ and we are done.
\end{proof}
\medskip
Having \Cref{cl:frontier}, we now go back to present a proof of \cref{lem:pairwise_del_clustering}.
\begin{proof}[Proof of \Cref{lem:pairwise_del_clustering}]
We start with the first property. Let $u$ be an arbitrary clustered node and recall that $c_u$ is its cluster center.
As $c_u \in \mathrm{frontier}^0(u)$, we have $d(c_u,u) = \mathrm{wait}(u)$.
Moreover,
%
\[\mathrm{wait}(u) = \min_{v \in V} \mathrm{del}(v) + d(v,u) \leq \mathrm{del}(u) + d(u,u) = \mathrm{del}(u) = O(\text{s} \log n).\]
%
Hence, we have $d(c_u,u) = O(\text{s} \log n)$.
Moreover, \cref{cl:frontier} gives that all nodes on a shortest path from $c_u$ to $u$ are also clustered to $c_u$, implying that the diameter of the cluster is $O(s \log n)$.
Next, we prove the second property.
Consider an arbitrary clustered node $w$. We first show that for an arbitrary clustered node $y$ with $d(w,y) \leq \text{s}$, it holds that $c_y \in \mathrm{frontier}^{2\text{s}}(w)$.
%
To see this, we first use the definition of $c_w$ to write
\begin{align*}
\mathrm{del}(c_y) + d(c_y,y)
\leq \mathrm{del}(c_w) + d(c_w,y)
\end{align*}
%
On one hand, we can use triangle inequality to lower bound the left-hand side by
%
\begin{align*}
\mathrm{del}(c_y) + d(c_y,y) \ge \mathrm{del}(c_y) + d(c_y, w) - d(w,y)
\end{align*}
On the other hand, we can use triangle inequality to upper bound the right hand side by
\begin{align*}
\mathrm{del}(c_w) + d(c_w,y)
\le \mathrm{del}(c_w) + d(c_w, w) + d(w, y).
\end{align*}
%
Putting the two bounds together, we conclude that
\begin{align*}
\mathrm{del}(c_y) + d(c_y, w) \le \mathrm{del}(c_w) + d(c_w, w) + 2d(w,y) \le \mathrm{wait}(w) + 2s,
\end{align*}
where we used our assumption $d(w,y) \le 2s$. Thus, $c_y \in \mathrm{frontier}^{2\text{s}}(w)$.
Now, let $u$ be an arbitrary clustered node and $P_u$ the unique path between $u$ and $c_u$ in the tree associated with the cluster. Furthermore, let $C \in \mathcal{C}^{\mathrm{del}}$ be a cluster with $d(P_u,C) \leq \text{s}$. Then there exists $w \in P_u$ and $y \in C$ with $d(w,y) \leq \text{s}$ and the discussion above implies $c_y \in \mathrm{frontier}^{2\text{s}} (w)$.
We now use \cref{cl:frontier}, which implies that $\mathrm{frontier}^{2\text{s}} (w) \subseteq \mathrm{frontier}^{2\text{s}}(u)$.
Hence, for each cluster $C$ with $d(P_u,C) \leq \text{s}$, the corresponding cluster center is contained in $\mathrm{frontier}^{2\text{s}}(u)$. As $u$ is clustered, we know that $|\mathrm{frontier}^{2\text{s}}(u)| \leq \lceil100 \log \log n\rceil$. Therefore, the $\text{s}$-hop degree of $\mathcal{C}^{\mathrm{del}}$ is at most $\lceil100 \log \log (n)\rceil$.
The third property follows directly from the definition.
To finish the proof, we need to show that the algorithm can be implemented in $O(s \log n \cdot \log\log n)$ rounds.
To see this, note that we can compute for each node $u$ whether $|\mathrm{frontier}^{2s}(u)| \le \lceil 100 \log \log(n)\rceil$ or not, as follows: We run a variant of breadth first search that takes into account the delays, where each node $v$ starts sending out a BFS token at time $\mathrm{del}(v)$. Recall that in classical breadth first search, after a node $u$ is reached for the first time by a token (sent from $c_u$), it broadcasts this token to all its neighbors and then it does not redirect any other tokens sent to it.
In our version of the search, each node stops redirecting only after at least $\lceil 100 \log \log(n)\rceil$ tokens arrived (we do not take into account tokens that have already arrived earlier) or after it counts $2s$ steps from the arrival of the first token.
It can be seen that this algorithm can be implemented in the desired number of rounds. Moreover, every node $u$ learns the value of $|\mathrm{frontier}^{2s}(u)|$ whenever the value is at most $\lceil 100 \log \log(n)\rceil$ and otherwise, it learns the value is larger than this threshold.
\end{proof}
In view of \cref{lem:pairwise_del_clustering}, to prove the randomized variant of \Cref{thm:low_degree_main} with pairwise analysis, it suffices to show that \cref{alg:low_degree_delay_alg} stated next computes a delay function $\mathrm{del}$ such that the expected number of nodes $u$ with $|\mathrm{frontier}^{2\text{s}}(u)| \leq \lceil 100 \log \log(n)\rceil$ is at least $n/2$. We later discuss how this is derandomized.
\subsection{Computing Delays}
This subsection is dedicated to proving the following theorem that asserts that we can compute a suitable delay function that can be plugged in \cref{lem:pairwise_del_clustering} that constructs a clustering from it.
\begin{theorem}
\label{thm:low_degree_delay_main}
\cref{alg:low_degree_delay_alg} runs in $\widetilde{O}(\text{s} \log^2(n))$ $\mathsf{CONGEST}\,$ rounds and computes a delay function $\mathrm{del}$ that satisfies
\begin{align}
\label{eq:houkacka}
|\{u \in V \colon |\mathrm{frontier}_{\mathrm{del}}^{2s}(u)| \leq \lceil 100 \log \log (n) \rceil\}| \geq n/2.
\end{align}
\end{theorem}
\begin{restatable}{algorithm}{delays}
\caption{Computing Delay Function $\mathrm{del}$}
\label{alg:low_degree_delay_alg}
Input: A parameter $s$, an algorithm $\mathcal{A}_{i,j}$ computing a \emph{good set} from \cref{def:low_degree_del_good_set} in $\widetilde{O}(s \log n)$ rounds\\
Output: A delay function $\mathrm{del}$ from \cref{def:low_degree_del} satisfying \cref{eq:houkacka}
\begin{algorithmic}[1]
\Procedure{Delays}{}
\State $V_0^ { \textrm{active}} \leftarrow V$
\State $R \leftarrow \lfloor 2 \log(n) \rfloor$
\State $k \leftarrow \lceil 100 \log \log (n) \rceil$
\State $\forall u \in V \colon \mathrm{del}_0(u) \leftarrow 5 \text{s} R$
\For{$i \leftarrow 1,2, \ldots, R$}
\State $W_{i,0} \leftarrow \emptyset$
\For{$j \leftarrow 1,2,\ldots,k$}
\State $S_{i,j} \leftarrow \mathcal{A}_{i,j}(\mathrm{del}_{i-1}, W_{i,j-1})$ \Comment{$S_{i,j} \subseteq V_{i-1}^ { \textrm{active}}$}
\State $W_{i,j} \leftarrow W_{i,j-1} \cup S_{i,j}$
\EndFor
\State $V_i^{ \textrm{active}} \leftarrow W_{i,k}$
\For{$\forall u \in V$}
\If{$u \in V^\textrm{active}_i$}
\State $\mathrm{del}_i(u) \leftarrow \mathrm{del}_{i-1}(u) - 5s$
\Else
\State $\mathrm{del}_i(u) \leftarrow \mathrm{del}_{i-1}(u)$
\EndIf
\EndFor
\EndFor
\State $\mathrm{del} \leftarrow \mathrm{del}_R$
\State \Return $\mathrm{del}$
\EndProcedure
\end{algorithmic}
\end{restatable}
\paragraph{Intuitive Description of \cref{alg:low_degree_delay_alg}.}
The algorithm runs in $R = \lfloor 2 \log(n)\rfloor$ phases and each phase consists of $k = \lceil 100 \log \log n\rceil$ iterations.
In iteration $j$ of phase $i$, algorithm $\mathcal{A}_{i,j}$ is a deterministic algorithm which computes a good set $\mathcal{S}_{i,j} \subseteq V^{\textrm{active}}_{i-1}$ as defined later in \cref{def:low_degree_del_good_set}. The algorithm description of $\mathcal{A}_{i,j}$ is deferred to \cref{sec:low_degree_local_derandomization}. The high-level intuition is that $\mathcal{A}_{i,j}$ derandomizes the randomized process which obtains $S_{i,j}$ from $V^{\textrm{active}}_{i-1}$ by including each vertex with probability $\frac{1}{4k}$, pairwise independently.
Repeating this pairwise independent sampling process $k$ times then simulates including each vertex from $V^{\textrm{active}}_{i-1}$ to $V^{\textrm{active}}_{i}$ with positive probability.
The derandomization of the pairwise independent process is done efficiently using a novel local derandomization procedure introduced in \cite{Faour2022} which essentially allows to efficiently derandomize algorithms that only rely on pairwise analysis.
Throughout the algorithm, each node is assigned a delay. At the beginning, each node $u$ is assigned a delay of $\mathrm{del}_0(u) = 5 \text{s} R = O(\text{s} \log n)$. In each subsequent phase, for each node $u$, we have two possibilities: if $u \in V^{\textrm{active}}_i$, the delay of node $u$ is decreased by $5 \text{s}$, i.e., $\mathrm{del}_i(u) = \mathrm{del}_{i-1}(u) - 5 \text{s}$ if $u \in V^{\textrm{active}}_i$; if $u \notin V^{\textrm{active}}_i$, then its delay stays the same, i.e., $\mathrm{del}_i(u) = \mathrm{del}_{i-1}(u)$ if $u \notin V^{\textrm{active}}_i$.
For every $u \in V$, we define the shorthand $\mathrm{wait}_i(u) = \mathrm{wait}_{\mathrm{del}_i}(u)$ and for every $D \geq 0$, we define $\mathrm{frontier}_i^D(u) = \mathrm{frontier}_{\mathrm{del}_i}^D(u)$.
\paragraph{Communication Primitives.}
%
For the deterministic algorithm $\mathcal{A}_{i,j}$ which computes the set $S_{i,j}$, it is important that each node $u$ can efficiently compute the set $\textrm{alive}_{i-1}(u)$ and $\textrm{dead}_{i-1}(u)$ that are defined next.
Let us give a brief intuition behind the definition. In the ``runner intuition'' from the beginning of the section, we want to know in every step all runners that are currently at distance at most $2s$ after the front runner. For these runners, we want to ensure that not all of them stop running in one step. In the reality of the distributed $\mathsf{CONGEST}\,$ model, we however cannot compute even the size of $\mathrm{frontier}^{2s}(u)$.
Fortunately, for our purposes if the number of ``runners'' that are distance at most $2s$ from the front runner is larger than $O(\log\log n)$, it roughly speaking suffices to work with the first $O(\log \log n)$ runners in the analysis. This is formalized by the following definition of alive and dead nodes (dead nodes are runners that stopped flipping coins).
\begin{restatable}{definition}{alivedead}[$\textrm{alive}_i(u)$/$\textrm{dead}_i(u)$]
For every vertex $u \in V$ and $i \in \{0,1,\ldots,R\}$, let $\textrm{dead}_i(u) \subseteq \mathrm{frontier}^{2 \text{s}}_i(u) \setminus V_i^{\textrm{active}}$ be an arbitrary subset of size $\min(k, |\mathrm{frontier}^{2 \text{s}}_i(u) \setminus V_i^{\textrm{active}}|)$ and $\textrm{alive}_i(u) \subseteq \mathrm{frontier}^{2 \text{s}}_i(u) \cap V_i^{\textrm{active}}$ be an arbitrary subset of size $\min(k - |\textrm{dead}_i(u)|, |\mathrm{frontier}^{2 \text{s}}_i(u) \cap V_i^{\textrm{active}}|)$.
\end{restatable}
Note that $|\textrm{alive}_i(u)| + |\textrm{dead}_i(u)| \leq \min(k, |\mathrm{frontier}^{2 \text{s}}_i(u)|)$ and $|\textrm{dead}_i(u)| \leq |\textrm{alive}_{i-1}(u)| + |\textrm{dead}_{i-1}(u)|$.
For each node $v \in V$, let $M_{i-1}(v) = \{u \in V \colon v \in \textrm{alive}_{i-1}(u)\}$. Then, we need some simultaneous and efficient communication, that allows each $v\in V$ to broadcast a message to all nodes in $M_{i-1}(v)$, and for $v$ to receive an aggregate of messages prepared for $v$ in nodes $M_{i-1}(v)$.
\begin{lemma}
\label{lem:low_degree_delays_communication}
Suppose that we are at the beginning of some phase $i \in [R]$. Given delay function $\mathrm{del}_{i-1}$, and given the set $V_{i-1}^{\textrm{active}}$, there exists a $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log n)$ rounds which computes for each node $u \in V$ the sets $\textrm{alive}_{i-1}(u)$ and $\textrm{dead}_i(u)$. Moreover, let $M_{i-1}(v) = \{u \in V \colon v \in \textrm{alive}_{i-1}(u)\}.$ Then, there exists an $\widetilde{O}(\text{s} \log n)$ round $\mathsf{CONGEST}\,$ algorithm that allows each node $v$ to send one $O(\log n)$-bit message that is delivered to all nodes in $M_{i-1}(v)$. Similarly, there also exists an $\widetilde{O}(\text{s} \log n)$ round $\mathsf{CONGEST}\,$ algorithm that given $O(\log n)$-bit messages prepared at nodes in $M_{i-1}(v)$ specific for node $v$, it allows node $v$ to receive an aggregation of these messages, e.g., the summation of the values, in $\widetilde{O}(\text{s} \log n)$ rounds.
\end{lemma}
\begin{proof}[Proof of \Cref{lem:low_degree_delays_communication}] We run a variant of breadth first search (BFS) that takes into account the delays, and runs in $\widetilde{O}(s\log n)$ rounds: Each node $v$ starts sending out a BFS token at time $\mathrm{del}_{i-1}(v)$, where the token also includes the information whether $v\in V_{i}^{\textrm{active}}$ or not. During the entire process, each node $u$ forwards per time step at most $k= \lceil100\log\log n \rceil$ BFS tokens, breaking ties in favoring of tokens coming from nodes $v\notin V_{i}^{\textrm{active}}$. That is, all tokens that arrive at the same time step are forwarded in the next time step, except that the node forwards at most $k$ tokens in this time step, and moreover, the node first includes all tokens from nodes $v\notin V_{i}^{\textrm{active}}$ (up to $k$) before including tokens from nodes $v\in V_{i}^{\textrm{active}}$. Since per time step each node forwards at most $k$ tokens, each time step can be implemented in at most $k$ rounds of the $\mathsf{CONGEST}\,$ model. Furthermore, node $u$ starts counting time from the moment that it received the very first token (while forwarding any received tokens, up to $k$ per time step), and after $2s$ time steps have passed, node $u$ does not forward any other tokens.
Let us think of one tree for each node $v$, which is rooted at $v$ and includes all nodes $u$ that have received the token of node $v$. Every node $u$ receives the tokens of all nodes in $\mathrm{frontier}_{i-1}^{2s}(u)$, if there are at most $k$ of them. If there are more than $k$, node $u$ learns at least $k$ of them, with the following guarantee: The set of learned tokens includes all tokens from $v\notin V_{i}^{\textrm{active}}$, up to $k$ (if there were more).
Hence, given the received tokens, each node $u$ can form $\textrm{dead}_{i-1}(u) \subseteq \mathrm{frontier}^{2 \text{s}}_{i-1}(u) \setminus V_{i-1}^{\textrm{active}}$, which is subset of size $\min(k, |\mathrm{frontier}^{2 \text{s}}_{i-1}(u) \setminus V_{i-1}^{\textrm{active}}|)$. Furthermore, node $u$ can form $\textrm{alive}_{i-1}(u) \subseteq \mathrm{frontier}^{2 \text{s}}_{i-1}(u) \cap V_{i-1}^{\textrm{active}}$, which is a subset of size $\min(k - |\textrm{dead}_{i-1}(u)|, |\mathrm{frontier}^{2 \text{s}}_{i-1}(u) \cap V_{i-1}^{\textrm{active}}|)$.
By repeating the same communication, each node $v$ is able to send one message which is delivered to all nodes $M_{i-1}(v) = \{u \in V \colon v \in \textrm{alive}_{i-1}(u)\}$, all simultaneously in $\widetilde{O}(s\log n)$ rounds. Moreover, by repeating the same communication pattern but in the reverse direction of time, we can do an aggregation along each tree, again all simultaneously in $\widetilde{O}(s\log n)$ rounds, allowing each node $v$ to receive an aggregation of the messages prepared for $v$ in nodes $M_{i-1}(v)$.
\end{proof}
\paragraph{Potential Functions}
In this paragraph, we define an outer potential $\Phi_i$ for every phase $i$ and an inner potential $\phi_{i,j}$ for every
iteration $j$ within phase $i$. The inner potential satisfies that if $\phi_{i,j-1} \leq \phi_{i,j}$ in each iteration $j$, then $\Phi_i \leq \Phi_{i-1} + n$. The outer potential satisfies that $\Phi_0 = 2n$ and if $\Phi_R \leq 10n \log(n)$, then $|\{u \in V \colon |\mathrm{frontier}^{2 \text{s}}(u)|\}| \geq \frac{9n}{10}$.
\begin{definition}[Outer Potential]
\label{def:low_degree_delays_outer_potential}
For every $i \in \{0,1,\ldots,R\}$, the outer potential of a node $u$ after phase $i$ is defined as
\[\Phi_i(u) = e^{\frac{|\textrm{dead}_i(u)|}{10}}.\]
The outer potential after phase $i$ is defined as
\[\Phi_i = \sum_{u \in V} \Phi_i(u) + 2^i|V^{\textrm{active}}_i|.\]
\end{definition}
Here, "after phase $0$" should be read as "the beginning of phase $1$".
\cref{alg:low_degree_delay_alg} will make sure that the outer potential is sufficiently small. A small outer potential after phase $i$ implies on one hand that there are not too many nodes $u$ for which $|\textrm{dead}_i(u)|$ is large and on the other hand ensures that there are not too many nodes in $V^{\textrm{active}}_i$, i.e., $|V^{\textrm{active}}_i| \lesssim\frac{n}{2^i}$.
The following lemma captures the usefulness of the outer potential.
\begin{lemma}[Outer Potential Lemma]
\label{lem:low_degree_delays_outer_potential_lemma}
We have $\Phi_0 \leq 2n$. Moreover, if $\Phi_R \leq 10 n \log(n)$, then $|\{u \in V \colon |\mathrm{frontier}_{\mathrm{del}}^{2\text{s}}(u)| \leq 100 \log \log (n)\}| \geq \frac{9n}{10}$.
\end{lemma}
\begin{proof}
First, note that $\Phi_R \geq 2^R|V_R^ { \textrm{active}}| > 10n \log(n)|V_R^{ \textrm{active}}|$.
As we assume that $\Phi_R \leq 10 n \log(n)$, this directly implies $V_R^ { \textrm{active}} = \emptyset$.
In particular, every $u \in V$ with $|\mathrm{frontier}^{2\text{s}}_{\mathrm{del}}(u)| > 100 \log \log (n)$ contributes
%
\[\Phi_R(u) = e^{\frac{|\textrm{dead}_R(u)|}{10}}\geq e^{\frac{\min(k,|\mathrm{frontier}_{\mathrm{del}}^{2\text{s}}(u)|}{10})} \geq 100 \log(n) \]
%
to the potential. Hence, there can be at most $\Phi_R/(100 \log(n)) \leq n/10$ such nodes and therefore $\{u \in V \colon |\mathrm{frontier}_{\mathrm{del}}^{2\text{s}}(u)| \leq 100 \log \log (n)\} \geq \frac{9n}{10}$, as desired.
\end{proof}
\begin{restatable}{definition}{pessimisticprobability}[Pessimistic Estimator Probability $p_{i,j}(u)$]
For $i \in [R]$ and $j \in \{0,1,\ldots,k\}$, the pessimistic estimator probability of a node $u$ after iteration $j$ within phase $i$ is defined as
\[p_{i,j}(u) = I(\textrm{alive}_{i-1}(u) \cap W_{i,j} = \emptyset) \cdot \left(1 - \frac{|\textrm{alive}_{i-1}(u)|}{10k}\right)^{k-j}.\]
\end{restatable}
Here, "after iteration $0$", should be read as "the beginning of iteration $1$".
Let us briefly elaborate on the definition of $p_{i,j}(u)$. Assume that we would compute $S_{i,j}$ by sampling each vertex in $V^{\textrm{active}}_{i-1}$ with probability $\frac{1}{4k}$ pairwise independently.
By a simple pairwise analysis, one can show that this would imply $\Pr[S_{i,j} \cap \textrm{alive}_{i-1}(u) \neq \emptyset] \geq \frac{|\textrm{alive}_{i-1}(u)|}{10k}$. Hence, if we are currently at the beginning of iteration $j$ within phase $i$ just prior to sampling the set $S_{i,j}$, then $p_{i,j-1}(u)$ is an upper bound on the probability that no node in $\textrm{alive}_{i-1}(u)$ is contained in $ V^{\textrm{active}}_i$ (which one should think of as a bad event).
\begin{restatable}{definition}{innerpotential}[Inner Potential]
\label{def:low_degree_del_inner}
The inner potential of a node $u$ after iteration $j$ of phase $i$ is defined as
\[\phi_{i,j}(u) = p_{i,j}(u) e^{\frac{|\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)|}{10}}.\]
The inner potential after iteration $j$ of phase $i$ is defined as
\[\phi_{i,j} = \sum_{u \in V} \phi_{i,j}(u) + |W_{i,j}|2^i + \frac{k-j}{k}2^{i-1} |V_{i-1}^{\textrm{active}}|.\]
\end{restatable}
Again, assume for a moment that we would compute $S_{i,j}$ by sampling each vertex in $V^{\textrm{active}}_{i-1}$ with probability $\frac{1}{4k}$ pairwise independently. Assume we are at the beginning of iteration $j$ within phase $i$ just prior to sampling $S_{i,j}$. Then, using the fact that $\mathbb{E}[p_{i,j}(u)] \leq p_{i,j-1}(u)$, one directly gets that $\mathbb{E}[\phi_{i,j}(u)] \leq \phi_{i,j}(u)$ and it also follows that $\mathbb{E}[\phi_{i,j}] \leq \phi_{i,j-1}$. Moreover, one can also show that $\mathbb{E}[\Phi_i(u)] \leq \phi_{i,j-1}(u) + 1$ and $\mathbb{E}[\Phi_i] \leq \phi_{i,j-1} + n$.
In more detail, if at least one node in $\textrm{alive}_{i-1}(u)$ is contained in $V^{\textrm{active}}_i$, one can show that this implies $\textrm{dead}_i(u) = \emptyset$ and therefore $\Phi_i(u) = 1$. On the other hand, in the previous discussion we mentioned that with probability at most $p_{i,j-1}(u)$ no node in $\textrm{alive}_{i-1}(u)$ is included in $V^{\textrm{active}}_i$, and as $|\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)| \leq |\textrm{dead}_i(u)|$, we have $\ p_{i,j-1}(u)\Phi_i(u) \leq \phi_{i,j-1}$.
\begin{lemma}[Inner Potential Lemma]
\label{lem:low_degree_delays_inner}
For $i \in [R]$ and $j \in \{0,1,\ldots,k\}$,
Assume that in every iteration $j$ of phase $i$, $S_{i,j}$ is computed in such a way that $\phi_{i,j} \leq \phi_{i,j-1}$. Then, $\Phi_i \leq \Phi_{i-1} + n$ and $\Phi_R \leq 4n \log(n)$.
\end{lemma}
\begin{proof}
For each node $u \in V$, we have
\[\phi_{i,0}(u) = p_{i,0}(u)e^{\frac{|\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)|}{10}} = \left( 1 - \frac{|\textrm{alive}_{i-1}(u)|}{10k}\right)^{k-0} e^{\frac{|\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)|}{10}} \leq e^{\frac{|\textrm{dead}_{i-1}(u)|}{10}} = \Phi_{i-1}(u).\]
Therefore,
\[\phi_{i,0} = \sum_{u \in V} \phi_{i,0}(u) + |W_{i,0}|2^i + \frac{k-0}{k}2^{i-1} |V^{\textrm{active}}_{i-1}| \leq \sum_{u \in V} \Phi_{i-1}(u) + 2^{i-1}|V^{\textrm{active}}_{i-1}| = \Phi_{i-1}.\]
Consider an arbitrary $u \in V$. Next, we show that
\[e^{\frac{|\textrm{dead}_i(u)|}{10k}}=: \Phi_i(u) \leq \phi_{i,k}(u) + 1 = I(\textrm{alive}_{i-1}(u) \cap V_i^\textrm{active} = \emptyset)e^{\frac{|\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)|}{10k}} + 1.\]
It is easy to verify that the inequality is satisfied if $\textrm{alive}_{i-1}(u) \cap V_i^\textrm{active} = \emptyset$, as $|\textrm{dead}_i(u)| \leq |\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)|$.
Therefore, it remains to consider the case that there exists at least one node $v \in \textrm{alive}_{i-1}(u) \cap V_i^{\textrm{active}}$. The existence of such a node $v$ implies
%
\[\mathrm{wait}_i(u) \leq \mathrm{del}_i(v) + d(v,u) = \mathrm{del}_{i-1}(v) + d(v,u) - 5 \text{s} \leq \mathrm{wait}_{i-1}(u) + 2 \text{s} - 5 \text{s} = \mathrm{wait}_{i-1}(u) - 3 \text{s}.\]
%
For every node $w \notin V^{\textrm{active}}_i$, we have $ \mathrm{del}_i(w) = \mathrm{del}_{i-1}(w)$ and therefore
%
\[ \mathrm{del}_i(w) + d(w,u) = \mathrm{del}_{i-1}(w) + d(w,u) \geq \mathrm{wait}_{i-1}(u) \geq \mathrm{wait}_i(u) + 3 \text{s} > \mathrm{wait}_i(u) + 2 \text{s}\]
%
and thus $w \notin \mathrm{frontier}_i^{2 \text{s}}(u)$.
Hence, $\textrm{dead}_i(u) = \emptyset$ and the inequality is satisfied.
Therefore,
\[\phi_{i,k} = \sum_{u \in V} \phi_{i,k}(u) + |W_{i,k}|2^i + \frac{k-k}{k}2^{i-1} |V^{\textrm{active}}_{i-1}| \leq \left(\sum_{u \in V} \Phi_i(u) - 1 \right) + |V^{\textrm{active}}_i|2^i = \Phi_i - n.\]
A simple induction implies $\phi_{i,k} \leq \phi_{i,0}$.
Therefore,
\[\Phi_i \leq \phi_{i,k} + n \leq \phi_{i,0} + n = \Phi_{i-1} + n.\]
As $\Phi_0 \leq 2n$ according to \cref{lem:low_degree_delays_outer_potential_lemma}, a simple induction implies
\[\Phi_R \leq \Phi_0 + Rn \leq (2+R)n \leq 4n \log n.\]
\end{proof}
\paragraph{Good Set $S_{i,j}$:}
We are now going to define the good set of nodes $S_{i,j}$. Note that this is the part of \cref{alg:low_degree_delay_alg} whose definition we postponed.
\begin{definition}[Good Set $S_{i,j}$]
\label{def:low_degree_del_good_set}
For a set $S_{i,j} \subseteq V^{\textrm{active}}_{i-1}$ and $u \in V$, let
\[Y_{i,j}(u) = 1 - |\textrm{alive}_{i-1}(u) \cap S_{i,j}| + \binom{|\textrm{alive}_{i-1}(u) \cap S_{i,j}|}{2}.\]
We refer to the set $S_{i,j}$ as good if
\[ \sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_{i-1}(u)|/(10k))} + |S_{i,j}|\cdot 2^i \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}|.\]
\end{definition}
\begin{lemma}
\label{lem:low_degree_del_potential_nonicreasing}
If $S_{i,j}$ is a good set, then $\phi_{i,j} \leq \phi_{i,j-1}$.
\end{lemma}
\begin{proof}
For each $u \in V$, we have
\begin{align*}
\frac{I(\textrm{alive}_{i-1}(u) \cap S_{i,j} = \emptyset)p_{i,j-1}(u)}{1 - (|\textrm{alive}_{i-1}(u)|/(10k))} &=
\frac{I(\textrm{alive}_{i-1}(u) \cap S_{i,j} = \emptyset)I(\textrm{alive}_{i-1}(u) \cap W_{i,j-1} = \emptyset) \cdot \left(1 - \frac{|\textrm{alive}_{i-1}(u)|}{10k}\right)^{k-(j-1)}}{1 - (|\textrm{alive}_{i-1}(u)|/(10k))} \\
&= I(\textrm{alive}_{i-1}(u) \cap W_{i,j} = \emptyset) \left(1 - \frac{|\textrm{alive}_{i-1}(u)|}{10k}\right)^{k-j} \\
&= p_{i,j}(u).
\end{align*}
It also holds that $I(\textrm{alive}_{i-1}(u) \cap S_{i,j} = \emptyset) \leq Y_{i,j}(u)$. Therefore,
\begin{align*}
Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_{i-1}(u)|/(10k))} &\geq I(\textrm{alive}_{i-1}(u) \cap S_{i,j} = \emptyset)\frac{p_{i,j-1}(u)e^{\frac{|\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)|}{10}}}{1 - (|\textrm{alive}_{i-1}(u)|/10k)} \\
&= p_{i,j}(u) e^{\frac{|\textrm{dead}_{i-1}(u)| + |\textrm{alive}_{i-1}(u)|}{10}} \\
&= \phi_{i,j}(u).
\end{align*}
Thus, we get
\begin{align*}
\phi_{i,j} &= \sum_{u \in V} \phi_{i,j}(u) + |W_{i,j}|2^i + \frac{k-j}{k}2^{i-1}|V_{i-1}^{\textrm{active}}| \\
&\leq \sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + (|S_{i,j}| + |W_{i,j-1}|)2^i + \frac{k-j}{k}2^i|V^{\textrm{active}}_{i-1}| \\
&\leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}| + |W_{i,j-1}|2^i + \frac{k-j}{k}2^{i-1}|V^{\textrm{active}}_{i-1}| \\
&= \phi_{i,j-1}.
\end{align*}
\end{proof}
We now combine all the pieces to prove the main theorem of this subsection.
\paragraph{Proof of \cref{thm:low_degree_delay_main}}
We assume that in iteration $j$ of phase $i$, $\mathcal{A}_{i,j}$ computes a good set $S_{i,j}$. Therefore, \cref{lem:low_degree_del_potential_nonicreasing} implies that $\phi_{i,j} \leq \phi_{i,j-1}$. According to \cref{lem:low_degree_delays_inner}, this implies that $\Phi_R \leq 4n \log(n)$. Therefore, \cref{lem:low_degree_delays_outer_potential_lemma} implies that $|\{u \in V \colon |\mathrm{frontier}^{2 \text{s}}_{\mathrm{del}}(u)| \leq \lceil 100 \log \log (n)\rceil\}| \geq n/2$, as desired. It remains to discuss the $\mathsf{CONGEST}\,$ round complexity.
\cref{alg:low_degree_delay_alg} has $R \cdot k = \widetilde{O}(\log n)$ iterations in total. In iteration $i$ of phase $j$, algorithm $\mathcal{A}_{i,j}$ runs in $\widetilde{O}(\text{s} \log n)$ $\mathsf{CONGEST}\,$ rounds. Hence, the overall $\mathsf{CONGEST}\,$ complexity of \cref{alg:low_degree_delay_alg} is $\widetilde{O}(\text{s} \log^2 n)$.
\paragraph{Global Derandomization}
Here, we informally sketch a variant of \cref{alg:low_degree_delay_alg} which performs a global derandomization using the method of conditional expectation. A more formal discussion of this approach, though in a different context, is discussed in \cref{subsec:hittingset-impl} where we derandomize our algorithm for the hitting set problem in the $\mathsf{CONGEST}\,$ model. See in particular \cref{thm:hittingset-congest}.
\begin{definition}[Good Random Set $S_{i,j}$ (In Expectation)]
\label{def:low_degree_del_random_good_set}
For a set $S_{i,j} \subseteq V^{\textrm{active}}_{i-1}$ and $u \in V$, let
\[Y_{i,j}(u) = 1 - |\textrm{alive}_{i-1}(u) \cap S_{i,j}| + \binom{|\textrm{alive}_{i-1}(u) \cap S_{i,j}|}{2}.\]
%
We refer to a randomly computed subset $S_{i,j} \subseteq V^{\textrm{active}}_{i-1}$ as good in expectation if
%
\[ \mathbb{E} \left[\sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_{i-1}(u)|/(10k))} + |S_{i,j}|\cdot 2^i \right] \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}|.\]
\end{definition}
Note that we can recover \cref{def:low_degree_del_good_set} if we drop the expectation. Assume we choose $S_{i,j}$ by including each node in $V^{\textrm{active}}_{i-1}$ with probability $\frac{1}{4k}$, pairwise independently. One can show that the resulting set $S_{i,j}$ is good in expectation. Moreover, the pairwise distribution over the random set $S_{i,j}$ can be realized with a random seed length of $\widetilde{O}(\log n)$ using the construction of~\cite{roditty2005deterministic, luby1993removing} that is described in \Cref{subsec:hittingset-impl}.
The goal is now to fix the random seed one by one in such a way that the resulting deterministic set $S_{i,j}$ is a good set.
For the following discussion, let $X = \sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_{i-1}(u)|/(10k))} + |S_{i,j}|\cdot 2^i$.
The method of conditional expectation works by fixing the bits of the random seed one by one, each time fixing the $i$-th bit in such a way that
%
\[\mathbb{E}[X|\text{first $i$ bits are fixed to $b_0,\ldots,b_i$}] \leq \mathbb{E}[X|\text{first $i-1$ bits are fixed to $b_0,\ldots,b_{i-1}$}].\]
%
In particular, this ensures that
%
\[\mathbb{E}[X|\text{all bits are fixed}] \leq \mathbb{E}[X] \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}|\]
%
and hence the corresponding deterministic set $S_{i,j}$ is indeed a good set.
%
To find such a bit $b_i$, it suffices to compute two things:
\begin{itemize}
\item $\mathbb{E}[X|\text{first $i$ bits are fixed to $b_0,\ldots,b_{i-1},0$}]$, and
\item $\mathbb{E}[X|\text{first $i$ bits are fixed to $b_0,\ldots,b_{i-1}$}]$.
\end{itemize}
It is possible to decompose $X$ into $X = \sum_{u \in V} X_u$ such that each node $u$, when given $b_0,b_1,\ldots,b_i$, $\textrm{alive}_{i-1}(u)$ and $\phi_{i-1,j}(u)$, can efficiently compute $\mathbb{E}[X_u|b_0,b_1,\ldots,b_i]$, without any further communication. This in turn allows us to compute $\mathbb{E}[X|b_0,b_1,\ldots,b_i]$ in $O(D)$ rounds.
Hence, given that every node knows $\textrm{alive}_{i-1}(u)$ and $\phi_{i-1,j}(u)$, one can find a good set in $\widetilde{O}(D \log n)$ $\mathsf{CONGEST}\,$ rounds, where $D$ denotes the diameter of the network. Hence, computing $\textrm{alive}_{i-1}(u)$ and $\phi_{i-1,j}(u)$ can be done in $\widetilde{O}(\text{s} \log n)$ rounds at the beginning of phase $i$. Moreover, $\textrm{alive}_{i-1}(u)$ and $\phi_{i-1,j}(u)$ can be computed at the beginning of phase $i$ in $\widetilde{O}(\text{s} \log n)$ rounds according to \cref{lem:low_degree_delays_communication}. Hence, the overall resulting run-time of this variant of \cref{alg:low_degree_delay_alg} is $\widetilde{O}((D + \text{s}) \log^2(n)) = \widetilde{O}(D \log^2 n)$. This is the complexity for the setting where we have a low-diameter global tree of depth $D$. One can replace this by a standard application of network decomposition to reduce the round complexity to $\mathrm{poly}(\log n)$. In particular, given a $c$-color $d$-diameter network decomposition of $G^{O(s)}$, we can use independent randomness for the nodes of different colors, and for each color, we can perform the gathering and bit fixing in $(s+d)\mathrm{poly}(\log n)$ rounds. Hence, we can perform the same derandomization in $(s+d)\mathrm{poly}(\log n)$ rounds. With the algorithm of \cite{elkin2022deterministic} that computes a $O(\log n)$-color $\mathrm{poly}(\log n)$-strong-diameter network decomposition in $s \mathrm{poly}(\log n)$ rounds~\cite{elkin2022deterministic}, this becomes a complexity of $s \mathrm{poly}(\log n)$ rounds overall for the whole derandomization procedure. Please see the proof of \Cref{thm:hittingset-pram} where we perform such a global derandomization via network decomposition for the hitting set problem and provide more of the lower-order details. Instead of diving into those details here, in this section, we focus on the local derandomization which leads to a faster round complexity, as discussed in the next subsection.
\subsection{Algorithm $\mathcal{A}_{i,j}$ via Local Derandomization}
\label{sec:low_degree_local_derandomization}
This subsection is dedicated to providing the description of $\mathcal{A}_{i,j}$, that is proving \Cref{thm:deterministic-goodSet-selection} stated below. We note that this is the final missing piece in the proof of \cref{thm:low_degree_main}.
\begin{theorem}\label{thm:deterministic-goodSet-selection}
For every iteration $j$ of phase $i$, there exists a $\mathsf{CONGEST}\,$ algorithm $\mathcal{A}_{i,j}$ which computes a good set $S_{i,j} \subseteq V^{\textrm{active}}_{i-1}$ in $\widetilde{O}(\text{s} \log n)$ rounds.
\end{theorem}
The algorithm $\mathcal{A}_{i,j}$ makes use of the local rounding framework of Faour et al.~\cite{Faour2022} to compute a good set $S_{i,j}$. Their rounding framework works via computing a particular weighted defected coloring of the vertices, which allows the vertices of the same color to round their values simultaneously, with a limited loss in some objective functions that can be written as summation of functions each of which depend on only two nearby nodes. Next, we provide a related definition and then state their black-box local rounding lemma.
\begin{definition}[long-range d2-Multigraph]\label{def:long-range-d2multigraph}
A long-range \emph{d2-multigraph} is a multigraph $H=(V_H,E_H)$ that is simulated on top of an underlying communication graph $G=(V,E)$ by a distributed message-passing algorithm on $G$. The nodes of $H$ are a subset of the nodes of $G$, i.e., $V_H\subseteq V$. The edge set $E_H$ consists of two kinds of edges, \emph{physical edges} and \emph{virtual edges}. Physical edges in $E_H$ are edges between direct neighbors in $G$. For each physical edge in $e\in E_H$ with $V(e)=\set{u,v}$, both nodes $u$ and $v$ know about $e$. Virtual edges in $E_H$ are edges between two nodes $u,v\in V_H$, and for each such virtual edge, there is a manager node $w$ which knows about this edge.
We next describe the assumed communication primitives. Let $M(v)$ be the set of nodes $w$ who manage virtual edges that include $v$. We assume $T$-round primitives that provide the following: (1) each node $v$ can send one $O(\log n)$-bit message that is delivered to all nodes in $M(v)$ in $T$ rounds; (2) given $O(\log n)$-bit messages prepared at nodes $M(v)$ specific for node $v$, node $v$ can receive an aggregation of these messages, e.g., the summation of the values, in $T$ rounds.
\end{definition}
\begin{definition} (Pairwise Utility and Cost Functions) Let $H=(V_H,E_H)$ be a long-range d2-multigraph of an underlying communication graph $G=(V,E)$. For any label assignment $\vec{x}: V_H \rightarrow \Sigma$, a pairwise utility function is defined as $\sum_{u \ in V_H} \mathbf{u}(u, \vec{x}) + \sum_{e \in E_H} \mathbf{u}(e, \vec{x})$, where for a vertex $u$, the function $\mathbf{u}(u, \vec{x})$ is an arbitrary function that depends only on the label of $u$, and for each edge $e=\{u, v\}$, the function $\mathbf{u}(e, \vec{x})$ is an arbitrary function that depends only on the labels of $v$ and $u$. These functions can be different for different vertices $u$ and also for different edges $e$. A pairwise cost function is defined similarly. For a probabilistic/fractional assignment of labels to vertices $V_{H}$, where vertex $v$ assumes each label in $\Sigma$ with a given probability, the utility and costs are defined as the expected values of the utility and cost functions, if we randomly draw integral labels for the vertices from their corresponding distributions (and independently, though of course each term in the summation depends only on the labels of two vertices and thus pairwise independence suffices).
\end{definition}
\begin{lemma}\label{lemma:long-range-d2rounding}[Faour et al.~\cite{Faour2022}]
Let $H=(V_H,E_H)$ be a long-range d2-multigraph of an underlying communication graph $G=(V,E)$ of maximum degree $\Delta$, where the communication primitives have round complexity $T$. Assume that $H$ is equipped with pairwise utility and cost functions $\mathbf{u}(\cdot)$ and $\mathbf{c}(\cdot)$ (with label set $\Sigma$) and with a fractional label assignment $\lambda$. Further assume that the given rounding instance is polynomially bounded in a parameter $q \leq n$. Then for every constant $c>0$ and every $\varepsilon,\mu>\max\set{1/q^c, 2^{-c\sqrt{\log n}}}$, if $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)>\mu\mathbf{u}(\lambda)$, there is a deterministic $\mathsf{CONGEST}\,$ algorithm on $G$ to compute an integral label assignment $\ell$ for which $\mathbf{u}(\ell)-\mathbf{c}(\ell)\geq (1-\varepsilon)\cdot\big(\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\big)$ and such that the round complexity of the algorithm is
\[
T \cdot O\left(\frac{\log^2 q}{\varepsilon\cdot\mu}\cdot\left(\frac{|\Sigma| \log(q\Delta)}{\log n} + \log\log q \right)+\log q\cdot\log^* n \right).
\]
\end{lemma}
\paragraph{Our Local Derandomization.} In the following, for each node $u \in V$, we define $c_u = \frac{\phi_{i,j-1}(u)}{1- (|\textrm{alive}_{i-1} (u)|/(10k))}$.
The labeling space is whether each node in $V^{\textrm{active}}_{i-1}$ is contained in $S_{i,j}$ or not, i.e., each node in $V^{\textrm{active}}_{i-1}$ takes simply one of two possible labels $\Sigma=\{0,1\}$ where $1$ indicates that the node is in $S_{i,j}$. For a given label assignment $\vec{x} \in \{0,1\}^{V^{\textrm{active}}_{i-1}}$, we define the utility function
\[\mathbf{u}(\vec{x}) = \sum_{u \in V} c_u \sum_{v \in \textrm{alive}_{i-1}(u)} x_v + \frac{2^{i-1}}{k}|V_{i-1}^\textrm{active}| = \sum_{v \in V} \left( \sum_{u \in M_i(v)} c_u \right) x_v + \frac{2^{i-1}}{k}|V_{i-1}^{\textrm{active}}|,\]
and the cost function
\[\mathbf{c}(\vec{x}) = \sum_{u \in V} c_u \sum_{v \neq v' \in \textrm{alive}_{i-1}(u)} x_v x_{v'} + \sum_{v \in V^{\textrm{active}}_{i-1}} 2^i x_v.\]
If the label assignment is relaxed to be a fractional assignment $\vec{x} \in [0,1]^{V^{\textrm{active}}_{i-1}}$, where intuitively now $x_v$ is the probability of $v$ being contained in $S_{i,j}$, the same definitions apply for the utility and cost of this fractional assignment.
Note that the utility function is simply a summation of functions, each of which depends on the label of one vertex. Hence, it directly fits the rounding framework.
To capture the cost function as a summation of costs over edges, we next define an auxiliary multi-graph $H$ as follows: For each node $u \in V$ and every $v \neq v' \in \textrm{alive}_i(u)$, we add an auxiliary edge between $v$ and $v'$, with a cost function which is equal to $c_u$ when both $v$ and $v'$ are marked, and zero otherwise.
Note that $H$ is a long-range d2-Multigraph where the communication primitives have round complexity $\widetilde{O}(\text{s} \log n)$ as provided by \cref{lem:low_degree_delays_communication}.
We next argue that the natural fractional assignment where $x_v = \frac{1}{4k}$ for each $v\in V^{\textrm{active}}_{i-1}$ satisfies the conditions of \Cref{lemma:long-range-d2rounding}. First, note that these fractional assignments are clearly polynomially bounded in $q$ for $q=k=O(\log\log n)$. Next, we discuss that, for the given fractional assignment, utility minus cost is at least a constant factor of utility.
\begin{claim}
Let $\vec{x} \in [0,1]^{V^{\textrm{active}}_{i-1}}$ with $x_v = \frac{1}{4k}$ for every $v \in V^{\textrm{active}}_{i-1}$. Then, $\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}) \geq \mathbf{u}(\vec{x})/2$.
\end{claim}
\begin{proof}
We have
\begin{align*}
\mathbf{u}(\vec{x}) &= \sum_{u \in V} c_u \sum_{v\in \textrm{alive}_{i-1}(u)} x_v + \frac{2^{i-1}}{k}|V_{i-1}^{\textrm{active}}| \\
&\geq 2 \left(\sum_{u \in V} c_u \sum_{v,v' \in \textrm{alive}_{i-1}(u)} x_v x_{v'} + \sum_{v \in V^{\textrm{active}}_{i-1}} 2^i \frac{1}{4k} \right) \\
&\geq 2 \mathbf{c}(\vec{x}).
\end{align*}
and therefore indeed $\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}) \geq \mathbf{u}(\vec{x})/2$.
\end{proof}
Hence, we can apply \Cref{lemma:long-range-d2rounding} on these fractional assignments with $\mu=1/2$ and $\varepsilon=0.1$, which runs in $\widetilde{O}((\log\log \log n)^2)$ iterations of calling the communication primitives, each taking $\widetilde{O}(s\log^2 n)$ rounds. Hence, the entire procedure runs in $\widetilde{O}(s\log^2 n)$ rounds. As a result of applying
\Cref{lemma:long-range-d2rounding} with these parameters, we get an integral label assignment
$\vec{y} \in \{0,1\}^{V^\textrm{active}_{i-1}}$ which satisfies $\mathbf{u}(\vec{y}) - \mathbf{c}(\vec{y}) \geq 0.9 (\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}))$. We can then conclude
%
\begin{align*}
\mathbf{u}(\vec{y}) - \mathbf{c}(\vec{y}) &\geq 0.9 (\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x})) \\
&\geq0.9 \left( \sum_{u \in V} c_u \frac{|\textrm{alive}_{i-1}(u)|}{4k} + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}| - \left( \sum_{u \in V} c_u \frac{|\textrm{alive}_{i-1}(u)|}{16 k} + \frac{2^{i-2}}{k}|V^{\textrm{active}}_{i-1}|\right) \right)\\
&\geq \sum_{u \in V} c_u \frac{|\textrm{alive}_{i-1}(u)|}{10k}.
\end{align*}
%
This integral label assignment directly gives us $S_{i,j}$. In particular, let $S_{i,j} = \{v \in V^{\textrm{active}}_{i-1} \colon y_v = 1\}$. Note that
%
\begin{align*}
\mathbf{u}(\vec{y}) - \mathbf{c}(\vec{y}) = \sum_{u \in V} c_u \left( |\textrm{alive}_{i-1}(u) \cap S_{i,j}| - \binom{|\textrm{alive}_{i-1}(u) \cap S_{i,j}|}{2} \right) + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}| - 2^i|S_{i,j}|,
\end{align*}
%
and therefore
%
\begin{align*}
\sum_{u \in V} Y_{i,j}(u)\frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_{i-1}(u)|/(10k))} + |S_{i,j}| \cdot 2^i
&= \sum_{u \in V} c_u - \mathbf{u}(\vec{y}) + \mathbf{c}(\vec{y}) + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}| \\
&\leq \sum_{u \in V} c_u - \sum_{u \in V} c_u \frac{|\textrm{alive}_{i-1}(u)|}{10k} + \frac{2^{i-1}}{k}|V^\textrm{active}_{i-1}| \\
&\leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^{i-1}}{k}|V^{\textrm{active}}_{i-1}|, \end{align*}
which shows that $S_{i, j}$ is indeed a good set according to \Cref{def:low_degree_del_good_set}. This completes the description of our locally derandomized construction of good sets $S_{i, j}$, hence completing the proof of \Cref{thm:deterministic-goodSet-selection}.
\section{From Low-Degree Clusters to Isolated Clusters}
\label{sec:LowDegtoIsolation}
\subsamplingmain
Similar as in \cref{sec:low_degree}, we could get the same guarantees with a $\mathsf{CONGEST}\,$ algorithm with round complexity $O(s \mathrm{poly}(\log n))$ by performing a global derandomization with the help of a previously computed network decomposition.
\begin{proof}[Proof of \Cref{thm:subsampling_main}]
The clustering $\mathcal{C}^{out}$ is computed in two steps.
In the first step, we use the local rounding procedure to compute a clustering $\mathcal{C}'$ which one obtains from $\mathcal{C}$ by only keeping some of the clusters in $\mathcal{C}$ (any such cluster is kept in its entirety). Intuitively, the local rounding procedure derandomizes the random process which would include each cluster $C$ from $\mathcal{C}$ in the clustering $\mathcal{C}'$ with probability $\frac{1}{2k}$, $k = \lceil 100 \log \log n\rceil$, pairwise independently.
%
Given the clustering $\mathcal{C}'$, we keep each node $u \in \mathcal{C}'$ clustered in $\mathcal{C}^{out}$ if and only if the $\text{s}$-hop degree of $u$ in $\mathcal{C}'$ is $1$.
Note that given $\mathcal{C}'$, the output clustering $\mathcal{C}^{out}$ can be computed in $\widetilde{O}(\text{s} \log n)$ $\mathsf{CONGEST}\,$ rounds.
First, we discuss the first property, i.e., the strong diameter of the output clustering. The fact that the clustering $\mathcal{C}$ has strong diameter $O(\text{s} \log (n))$ directly implies that the clustering $\mathcal{C}'$ also has strong diameter $O(\text{s} \log(n))$, simply because each cluster of $\mathcal{C}'$ is exactly one of the clusters of $\mathcal{C}$. We next argue that $\mathcal{C}^{out}$ also has strong diameter $O(\text{s} \log n)$. Let $u$ be a node clustered in $\mathcal{C}'$ and $P_u$ the unique path from $u$ to its center in the tree associated with its cluster. Then, it directly follows from the definition that for every $w \in P_u$, the $\text{s}$-hop degree of $w$ in $\mathcal{C}'$ is at most the $\text{s}$-hop degree of $u$ in $\mathcal{C}'$. Therefore $u$ being clustered in $\mathcal{C}^{out}$ implies that $w$ is also clustered in $\mathcal{C}^{out}$. Hence, we conclude that $\mathcal{C}^{out}$ indeed has strong diameter $O(\text{s} \log n)$.
Next, we discuss the second property: The clustering $\mathcal{C}^{out}$ is $\text{s}$-hop separated. This directly follows from the fact that by definition every clustered node has a $\text{s}$-hop degree of $1$.
Finally, To prove \Cref{thm:subsampling_main}, the only remaining thing is to prove the third property, i.e., that we compute $\mathcal{C}'$ in such a way that $\mathcal{C}^{out}$ clusters at least $\frac{|\mathcal{C}|}{1000 \log \log (n)}$ nodes. The rest of this proof is dedicated to this property.
For each cluster $C \in \mathcal{C}$, we let $center(C)$ denote the cluster center of $C$ and define $Centers = \{center(C) \colon C \in \mathcal{C}\}$ as the set of cluster centers of $\mathcal{C}$.
%
Moreover, for each $u$ clustered in $\mathcal{C}$, recall that $c_u$ is the cluster center of the cluster of $u$ and let $P_u$ be the unique $u$-$c_u$ path in the tree associated with this cluster $C_u$. Now, let
%
\[S_u = \{C \in \mathcal{C} \colon d(P_u,C) \leq \text{s}\}.\]
%
Note that the size of $S_u$ is equal to the $\text{s}$-hop degree of $u$, which by assumption is at most $k$.
The labeling space is for each cluster center whether its cluster is contained in $\mathcal{C}'$ or not, i.e., each node in $Centers$ takes simply one of two possible labels $\{0,1\}$ where $1$ indicates that the corresponding cluster is in $\mathcal{C}'$. For a given label assignment $\vec{x} \in \{0,1\}^{Centers}$, we define
%
\[\mathbf{u}(\vec{x}) = \sum_{C \in \mathcal{C}} |C|x_{center(C)}\]
%
and
\[\mathbf{c}(\vec{x}) = \sum_{u \in V \colon \text{$u$ is clustered in $\mathcal{C}$}}\sum_{C \in S_u \setminus {C_u}} x_{c_u}x_{center(C)}.\]
If the label assignment is relaxed to be a fractional assignment $\vec{x} \in [0,1]^{Centers}$, where intuitively now $x_v$ is the probability of $v$'s cluster being contained in $\mathcal{C}'$, the same definitions apply for the utility and cost of this fractional assignment.
The utility function is simply a summation of functions, each of which depends on the label of one vertex in $Centers$. Hence, it directly fits the rounding framework.
To capture the cost function as a summation of costs over edges, we next define an auxiliary multi-graph $H$ as follows: For each node $u $ clustered in $\mathcal{C}$ and every $C_1 \neq C_2 \in S_u$, we add an auxiliary edge between $center(C_1)$ and $center(C_2)$ with a cost function which is equal to $1$ when both $C_1$ and $C_2$ are contained in $\mathcal{C}'$, and zero otherwise.
Note that $H$ is a long-range d2-Multigraph, according to \Cref{def:long-range-d2multigraph}. The communication primitives can be implemented in $\widetilde{O}(\text{s} \log n)$ rounds according to the lemma below.
\begin{lemma}
\label{lem:subsampling_delays_communication}
Let $\mathcal{C}$ be the input clustering of \cref{thm:subsampling_main}. There exists a $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log n)$ rounds which computes for each node $u \in V$ the sets $\{center(C) \colon C \in S_u\}$. Moreover, for each $v \in Centers$, let $M(v) = \{u \in V \colon C_v \in S_u\}.$ Then, there exists an $\widetilde{O}(\text{s} \log n)$ round $\mathsf{CONGEST}\,$ algorithm that allows each node $v$ to send one $O(\log n)$-bit message that is delivered to all nodes in $M(v)$. Similarly, there also exists an $\widetilde{O}(\text{s} \log n)$ round $\mathsf{CONGEST}\,$ algorithm that given $O(\log n)$-bit messages prepared at nodes in $M(v)$ specific for node $v$, it allows node $v$ to receive an aggregation of these messages, e.g., the summation of the values, in $\widetilde{O}(\text{s} \log n)$ rounds.
\end{lemma}
\begin{proof}[Proof of \Cref{lem:subsampling_delays_communication}]
The proof follows roughly along the lines of the proof of \cref{lem:low_degree_delays_communication}. First, we run the following variant of breadth first search: At the beginning, each node clustered in $\mathcal{C}$ has a token which is equal to the identifier of its cluster.
Now, in each of the $\text{s}$ iterations, each node that has received at most $k = \lceil 100 \log \log n \rceil$ identifiers in the previous iteration forwards all the identifiers it has received to its neighbors. If a node has received more than $k$ identifiers, it selects $k$ of them to forward.
The first phase can be implemented in $O(k \text{s})$ $\mathsf{CONGEST}\,$ rounds.
It directly follows from the fact that the $\text{s}$-hop degree of $\mathcal{C}$ is at most $k$ that after the first phase each node $w$ learns the identifiers of all cluster centers such that the corresponding cluster $C$ satisfies $d(w,C) \leq \text{s}$. The next phase propagates this information up in the cluster tree, from the root toward the leaves, such that each descendant of $w$---i.e., any node whose cluster path to the root passes through $w$---learns about all those cluster centers as well.
The second phase consists of $O(\text{s} \log n)$ iterations. In each iteration, each
clustered node sends all the identifiers it learned about so far to each of its children in the corresponding cluster tree.
It again follows from the fact that the $\text{s}$-hop degree of $\mathcal{C}$ is at most $k$ that each of the $O(\text{s} \log n)$ iterations in the second phase can be implemented in $O(k)$ $\mathsf{CONGEST}\,$ rounds. Hence, the overall $\mathsf{CONGEST}\,$ runtime is $\widetilde{O}(\text{s} \log n)$.
For each $v \in Centers$, let $M(v) = \{u \in V \colon C_v \in S_u\}.$ By repeating the above communication, we have a $\widetilde{O}(\text{s} \log n)$-round procedure that delivers one message from each node $v$ to all nodes $M(v)$. By reversing the same communication in time, we can also provide the opposite direction: if each node in $M(v)$ starts with a message for $v$, then in $\widetilde{O}(\text{s} \log n)$ rounds, we can aggregate these messages and deliver the aggregate to $v$, simultaneously for all $v$.
\end{proof}
\begin{claim}
Let $k= \lceil 100 \log \log (n) \rceil$ and $\vec{x} \in [0,1]^{Centers}$ with $x_v = \frac{1}{2k}$ for every $v \in Centers$. Note that this fractional label assignment is polynomially bounded in $q=k=O(\log\log n)$. Furthermore, we have $\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}) \geq \mathbf{u}(\vec{x})/2$.
\end{claim}
\begin{proof}
We have
%
\[\mathbf{u}(\vec{x}) = \sum_{C \in \mathcal{C}} |C|x_{center(C)} = \sum_{C \in \mathcal{C}} \frac{|C|}{2k} = \frac{|\mathcal{C}|}{2k},\]
%
and
%
\[\mathbf{c}(\vec{x}) = \sum_{u \in V \colon \text{$u$ is clustered in $\mathcal{C}$}}\sum_{C \in S_u \setminus \{C_u\}} x_{c_u}x_{center(C)} \leq \sum_{u \in V \colon \text{$u$ is clustered in $\mathcal{C}$}} \frac{1}{4k} \frac{|S_u|}{k} \leq \frac{|\mathcal{C}|}{4k}.\]
%
Therefore, indeed $\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}) \geq \mathbf{u}(\vec{x})/2$.
\end{proof}
We now invoking the rounding of \Cref{lemma:long-range-d2rounding} with parameters $\mu=0.5$, $\varepsilon=0.5$, and $q=k=O(\log\log n)$ on the fractional label assignment of $\vec{x} \in [0,1]^{Centers}$ where $x_v = \frac{1}{2k}$ for every $v \in Centers$. The procedure runs in $\widetilde{O}(s\log n)$ rounds. As output, we get an integral label assignment $\vec{y} \in \{0,1\}^{Centers}$ which satisfies
%
\[\mathbf{u}(\vec{y}) - \mathbf{c}(\vec{y}) \geq 0.5 (\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x})) \geq \frac{|\mathcal{C}|}{8k}.\]
Let $C' = \{C \in \mathcal{C} \colon y_{center(C)} = 1\}$. Note that for every $u \in \mathcal{C}$,
%
\[I(\text{$u$ is clustered in $\mathcal{C}^{out}$}) \geq y_{c_u} - \sum_{C \in S_u \setminus \{C_u\}} y_{c_u} y_{center(C)}.\]
%
%
Therefore,
%
\begin{align*}
|\mathcal{C}^{out}|&\geq \sum_{u \in V \colon \text{$u$ is clustered in $\mathcal{C}$}} I(\text{$u$ is clustered in $\mathcal{C}^{out}$}) \\
&\geq \sum_{u \in V \colon \text{$u$ is clustered in $\mathcal{C}$}} \left(y_{c_u} - \sum_{C \in S_u \setminus \{C_u\}} y_{c_u} y_{center(C)}\right)\\
&= \mathbf{u}(\vec{y}) - \mathbf{c}(\vec{y}) \\
&\geq \frac{|\mathcal{C}|}{8k}
\end{align*}
%
and therefore $\mathcal{C}^{out}$ clusters enough vertices to prove \Cref{thm:subsampling_main}.
\end{proof}
\section{Clustering More Nodes}
\label{sec:clusteringmorenodes}
In this section, we prove the following result, which says that once we have access to a clustering algorithm that clusters a nontrivial proportion of nodes with sufficient separation, we can turn it into an algorithm that clusters a constant proportion of nodes. We are paying for this with a slight decrease in the separation guarantees.
\clusteringmorenodesmain
It follows from the analysis of \cref{alg:clustering_with_more_vertices} and its subroutine \cref{alg:expanding}.
To understand the pseudocode of the algorithms, we note that for a set of nodes $C \subseteq V(G)$ and $D \in \mathbb{N}_0$, we define $$C^{\leq D} = \{v \in V \colon d(C,v) \leq D\}.$$
Moreover, we say that a cluster $C$ is \emph{good} in \cref{alg:expanding} if $\textrm{cut}(C) < +\infty$. Otherwise, $C$ is \emph{bad}.
\begin{algorithm}
\caption{Making a clustering algorithm cluster half of the nodes}
\label{alg:clustering_with_more_vertices}
\begin{algorithmic}[1]
\Procedure{\textsc{ClusterHalfNodes($G$)}}{}
\State $\mathcal{C}_0 = \emptyset$
\State $N = \lceil 4 \cdot 2^x\rceil$
\For{$i = 1,2 \ldots, N$}
\State $G_i = G\left[V \setminus \left( \bigcup_{C \in \mathcal{C}_{i-1}} C \right)^{\le 1} \right]$
\State $\mathcal{C} \leftarrow \mathcal{A}(G_i)$
\State $\hat{\fC}_i \leftarrow \textsc{Expand}(G_i, \mathcal{C})$
\State $\mathcal{C}_i = \mathcal{C}_{i-1} \cup \hat{\mathcal{C}}_i$
\EndFor
\Return $\mathcal{C}_N$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{Expanding an input clustering}
\label{alg:expanding}
{\bf Input:} A graph $G$ and its $10x$-separated clustering $\mathcal{C}$\\
{\bf Output:} An expanded clustering $\hat{\fC}$ with small boundary
\begin{algorithmic}[1]
\Procedure{\textsc{Expand($G, \mathcal{C}$)}}{}
\For{$C \in \mathcal{C}$}
\State Define $\textrm{cut}(C) = \min \left\{0 \le i \le 3x \colon |C^{\leq i+1}| \leq 1.5 |C^{\leq i}| \right\}$ and $\textrm{cut}(C) = +\infty$ if no such $i$ exists.
\State If $\textrm{cut}(C) < +\infty$, define $\textrm{expand}(C) = C ^{\textrm{cut}(C)}$
\EndFor
\State \Return $\hat{\fC} = \{\textrm{expand}(C) : C \in \mathcal{C}, \textrm{cut}(C) < +\infty\}$
\EndProcedure
\end{algorithmic}
\end{algorithm}
We start by analyzing \cref{alg:expanding} in the following lemma. Importantly, the fourth condition for $\hat{\fC}$ in the statement below states that the total number of unclustered vertices neighboring one of the clusters in $\hat{\mathcal{C}}$ is at most half the total number of clustered vertices. This is the reason why we can
\begin{lemma}
\label{lem:clustering_transformed}
Let $x \geq 2$ be arbitrary and $\mathcal{C}$ a clustering with
\begin{enumerate}
\item strong diameter $O(x \log n)$,
\item separation $10 x$ and
\item clustering at least $\frac{n}{2^x}$ nodes.
\end{enumerate}
Then, $\hat{\mathcal{C}}$ constructed in \cref{alg:expanding} is a clustering with
\begin{enumerate}
\item strong diameter $O(x \log n)$,
\item separation $4 x$,
\item clustering at least $0.5\frac{n}{2^x}$ nodes and
\item $\left|\left(\bigcup_{C \in \hat{\fC}} C\right)^{\leq 1}\right| \leq 1.5 \left|\left(\bigcup_{C \in \hat{\fC}} C\right)^{\leq 1}\right|$.
\end{enumerate}
Moreover, the algorithm can be implemented in $O(x \log n)$ $\mathsf{CONGEST}\,$ rounds.
\end{lemma}
\begin{proof}
The first property follows from the fact that for a set $S$ and $D \in \mathbb{N}_0$, $\textrm{diam}(S^{\leq D}) \leq \textrm{diam}(S) + 2D$. Hence, for a good cluster $C$,
\[\textrm{diam}(\textrm{expand}(C)) \leq \textrm{diam}(C) + 2\textrm{cut}(C) = O(x \log n).\]
To prove the second property, let $C_1 \neq C_2 \in \mathcal{C}$ be two arbitrary good clusters. For $i \in \{1,2\}$, let $u_i \in \textrm{expand}(C_i)$ be arbitrary. By triangle inequality, we have:
\[d(u_1,u_2) \geq d(C_1,C_2) - d(C_1,u_1) - d(C_2,u_2) \geq 10 x - 2 \cdot 3x \geq 4x.\]
To prove the third property, it suffices to show that at most $0.5 \frac{n}{2^x}$ of the nodes are contained in bad clusters. For a bad cluster $C$, a simple induction implies $|C^{\leq 3x}| \geq 1.5^{3x} |C| \geq 2 \cdot 2^x |C|$. Therefore,
\[\sum_{C \in \mathcal{C}, \text{$C$ is a bad cluster}} |C| \leq \frac{1}{2^{x+1}} \sum_{C \in \mathcal{C}, \text{$C$ is a bad cluster}} |C^{\leq 3x}| \leq \frac{n}{2^{x+1}}, \]
where the last inequality follows from the fact that for two clusters $C_1 \neq C_2 \in \mathcal{C}$, $C_1^{\leq 3x} \cap C_2^{\leq 3x} = \emptyset$.
To prove the fourth property we write
\begin{align*}
\left|\left(\bigcup_{C \in \hat{\mathcal{C}}} C\right)^{\leq 1}\right| &\leq \sum_{\hat{C} \in \hat{\mathcal{C}}} |\hat{C}^{\leq 1}| \\
&= \sum_{C \in \mathcal{C} \colon \text{$C$ is a good cluster}} |(C^{\leq \textrm{cut}(C)})^{\leq 1}| \\
&= \sum_{C \in \mathcal{C} \colon \text{$C$ is a good cluster}} |C^{\leq \textrm{cut}(C) + 1}| \\
&\leq 1.5 \sum_{C \in \mathcal{C} \colon \text{$C$ is a good cluster}} |C^{\leq \textrm{cut}(C)}| \\
&= 1.5 \left|\left(\bigcup_{C \in \hat{\fC}} C\right)^{\leq 1}\right|.
\end{align*}
It remains to discuss the $\mathsf{CONGEST}\,$ computation. Since we have for any $C_1, C_2 \in \mathcal{C}$ that $C_1^{\le 3x} \cap C_2^{\le 3x} = \emptyset$, each cluster $C \in \mathcal{C}$ can compute the values of $C_1^{\le 0}, C_1^{\le 1}, \dots, C_1^{\le 3x}$ by running one breadth first search from $C_1$ up to distance of $3x$.
\end{proof}
We are now ready to prove \cref{thm:clusteringmorenodesmain}.
\begin{proof}[Proof of \cref{thm:clusteringmorenodesmain}]
We show that the algorithm satisfies the following invariants for $i \in \{0,1,\ldots,N\}$:
\begin{enumerate}
\item $\mathcal{C}_i$ is $2$-separated
\item $|V(\mathcal{C}_i)| \geq n \cdot \min(0.5,\frac{i}{8 \cdot 2^x})$
\item $|V(\mathcal{C}_i^{\leq 1})| \leq 1.5 |V(\mathcal{C}_i)| $
\end{enumerate}
The base case $i = 0$ trivially holds. Now, consider an arbitrary $i \in [N]$ and assume that the invariant is satisfied for $i-1$. To check the first invariant, let $C_1 \neq C_2 \in \mathcal{C}_i$ be arbitrary. If $C_1,C_2 \in \mathcal{C}_{i-1}$, then it follows by induction that $d(C_1,C_2) \geq 2$. If $C_1,C_2 \in \hat{\mathcal{C}}_i$, then it follows from \cref{lem:clustering_transformed} that $d_{G_i}(C_1,C_2) \geq 2$ which also directly implies $d_G(C_1,C_2) \geq 2$. It remains to consider the case that one cluster, let's say $C_1$, is in $\mathcal{C}_{i-1}$ and $C_2$ is in $\hat{\mathcal{C}}_i$.
We have
\[C_2 \subseteq V(G_i) = V \setminus V(\mathcal{C}_{i-1}^{\leq 1}) \subseteq V \setminus C_1^{\leq 1}\]
and therefore $d(C_1,C_2) \geq 2$, as desired.
Next, we show that the second invariant is preserved. If $|V(\mathcal{C}_{i-1})| \geq n/2$, then there is nothing to show. Otherwise, we have
\[|V(G_i)| \geq n - |V(\mathcal{C}_{i-1}^{\leq 1})| \geq n - 1.5 |V(\mathcal{C}_{i-1})| \geq n - 1.5\frac{n}{2} = \frac{n}{4}.\]
Therefore, according to \cref{lem:clustering_transformed}, $\hat{C}_i$ clusters at least $0.5 \frac{(n/4)}{2^x} = \frac{n}{8 \cdot 2^x}$ vertices, which together with $|V(\mathcal{C}_{i-1})| \geq n \cdot \min(0.5,\frac{i-1}{8 \cdot 2^x})$ directly implies $|V(\mathcal{C}_i)| \geq n \cdot \min(0.5,\frac{i}{8 \cdot 2^x})$.
It remains to verify the third property. According to \cref{lem:clustering_transformed}, we have
\[|V(\hat{\mathcal{C}}^{\leq 1}_i) \setminus V(\mathcal{C}^{\leq 1}_{i-1})| \leq 1.5 |V(\hat{\mathcal{C}}_i)|.\]
Therefore,
\[|V(\mathcal{C}^{\leq 1}_i)| = |V(C_{i-1}^{\leq 1})| + |V(\hat{\mathcal{C}}^{\leq 1}_i) \setminus V(\mathcal{C}^{\leq 1}_{i-1})| \leq 1.5 |V(\mathcal{C}_{i-1})| + 1.5 |V(\hat{\mathcal{C}}_i)| = 1.5 |V(\mathcal{C}_i)|.\]
This finishes the proof that the invariants are satisfied throughout the algorithm. Hence, $\mathcal{C}_N$ is a $2$-separated clustering that clusters at least half of the vertices. Moreover, it directly follows from the strong diameter guarantee of \cref{lem:clustering_transformed} that $\mathcal{C}_N$ has strong diameter $O(x \log n)$. Finally, as $\hat{\mathcal{A}}$ has a round complexity of $O(x \log n)$, it follows that $\mathcal{C}_N$ is computed in $O(2^x(R + x \log n))$ $\mathsf{CONGEST}\,$ rounds. This concludes the proof of \cref{lem:pairwise_del_clustering}.
\end{proof}
\iffalse
\section{Low-Diameter Clustering With Pairwise Analysis}
\label{sec:pairwise}
This section is devoted to proving the following theorem.
\begin{theorem}
\label{thm:pairwise}
Let $\text{s} \geq 2$ be arbitrary.
There exists a randomized $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log n)$ rounds requiring access only to pairwise independent random variables that outputs a clustering $\mathcal{C}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log (n))$,
\item separation of $\text{s}$, and
\item the expected number of clustered nodes, i.e., $\mathbb{E} \left[ \left| \bigcup_{C \in \mathcal{C}} C\right|\right]$, is at least $\Omega(n /(\log \log (n)))$.
\end{enumerate}
\end{theorem}
In the subsequent sections, we will strengthen \cref{thm:pairwise} in two ways.
First, in ..., we derandomize the algorithm of \cref{thm:pairwise} using the method of conditional expectation. This derandomization method requires the computation of global aggregates and therefore the resulting round complexity of O() has a dependency in terms of the diameter of the graph.
Second, in ..., we derandomize the algorithm of \cref{thm:pairwise} using a novel local derandomization technique developed in a concurrent paper, resulting in a deterministic algorithm with round complexity O().
Second, one shortcoming of \cref{thm:pairwise} is that the (expected) number of clustered nodes is $\Omega(n/\log \log (n))$, while existentially it is possible to satisfy the first two guarantees and at the same time clustering half of the nodes.
In \cref{sec:clustering_deterministic_transformation}, we show that we can compute a clustering with strong diameter $O(\log n \log \log \log n)$ clustering half of the nodes. This clustering is computed by invoking the algorithm of the deterministic version of \cref{thm:pairwise} $O(\log \log n)$ times, each time with separation parameter $\text{s} = O(\log \log \log n)$.
The algorithm proving \cref{thm:pairwise} consists of two parts. The first part computes a clustering $\mathcal{C}$ with strong diameter $O(\text{s} \log n)$ which, in expectation, clusters at least half of the nodes. However, the clustering only satisfies that the $\text{s}$-hop degree, defined below, is $O(\log \log n)$. This relaxes the separation property in the sense that a (non-empty) clustering is $\text{s}$-separated if and only if its $\text{s}$-hop degree is $1$.
In the second part of the algorithm, we make the clusters $\text{s}$-separated, at the cost of clustering only $\Omega(1/\log\log n)$ fraction of nodes. We define the notion of $s$-hop degree next.
\begin{definition}[$\text{s}$-hop degree of a clustering $\mathcal{C}$]
Let $\text{s} \geq 2$ be arbitrary.
For a clustered node $u$, let $P_u$ denote the path between $u$ and its cluster center in the tree associated with its cluster. The $\text{s}$-hop degree of $u$ is defined as the number of different clusters $C \in \mathcal{C}$ with $d(P_u,C) \leq \text{s}$.
The $\text{s}$-hop degree of a clustering $\mathcal{C}$ is defined as the largest $\text{s}$-hop degree of any node $u \in C \in \mathcal{C}$.
\end{definition}
\cref{sec:pairwise_low_degree} is dedicated to proving the following theorem.
\begin{theorem}
\label{thm:pairwise_low_degree}
Let $\text{s} \geq 2$ be arbitrary. There exists a randomized $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log(n))$ rounds requiring only access to pairwise independent random variables and which computes a clustering $\mathcal{C}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log(n))$,
\item $\text{s}$-hop degree of at most $10 \log \log (n)$, and
\item the expected number of clustered nodes is at least $n/2$.
\end{enumerate}
\end{theorem}
Given such a clustering, a simple subsampling idea can be used to compute a $\text{s}$-separated clustering at the expense that the expected number of clustered nodes goes down by an $O(\log \log n)$ factor. In particular, in \cref{sec:pairwise_subsampling} we prove the following theorem.
\begin{restatable}{theorem}{subsampling}
\label{thm:pairwise_subsampling}
Assume we are given a clustering $\mathcal{C}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log (n))$ and
\item $\text{s}$-hop degree of at most $10 \log \log (n)$.
\end{enumerate}
There exists a randomized $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log(n))$ rounds requiring only access to pairwise independent random variables which computes a clustering $\mathcal{C}^{out}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log (n))$,
\item separation of $\text{s}$ and
\item the expected number of clustered nodes is at least $\frac{|\mathcal{C}|}{1000 \log \log (n)}$.
\end{enumerate}
\end{restatable}
Note that \cref{thm:pairwise} directly follows from \cref{thm:pairwise_low_degree} together with \cref{thm:pairwise_subsampling}.
\subsection{Computing a Low-Degree Clustering}
\label{sec:pairwise_low_degree}
This section is dedicated to the proof of \cref{thm:pairwise_low_degree}.
The algorithm proving \cref{thm:pairwise_low_degree} relies on the notion of a delay function $\mathrm{del}$ which assigns each vertex a value from the set $\{0,1,2,\ldots O(\text{s} \log(n))\}$.
\begin{definition}[Delay Function]
Given a graph $G$, a delay function $\mathrm{del}$ is a function assigning each node $u \in V(G)$ a its \emph{delay} $\mathrm{del}(u) \in \{0,1,\ldots,O(\text{s} \log (n))\}$.
\end{definition}
The delay of each vertex is computed by a simple subsampling procedure requiring only pairwise independence (\cref{alg:pairwise_delays}).
Before discussing and analyzing \cref{alg:pairwise_delays}, we first explain how each delay function $\mathrm{del}$ gives rise to a clustering $\mathcal{C}^{\mathrm{del}}$ with strong diameter $O(\text{s} \log n)$ and $\text{s}$-hop degree $O(\log \log n)$.
To this end, we need to introduce additional notation. Given a delay function $\mathrm{del}$, we define for every vertex $u \in V$ its \mathrm{wait} function $\mathrm{wait}(u)$ as
\[\mathrm{wait}(u) = \min_{v \in V} \mathrm{del}(v) + d(v,u).\]
The intuition behind $\mathrm{wait}(u)$ is as follows: Assume that each node $v$ starts sending out a token at time $\mathrm{del}(v)$. Then, $\mathrm{wait}(u)$ is the time it takes until $u$ receives the first token.
Finally, for every parameter $D \ge 0$ and any node $u \in V(G)$ we define its \emph{frontier of width $D$} as
\[\mathrm{frontier}^D(u) = \{v \in V \colon \mathrm{del}(v) + d(v,u) \leq \mathrm{wait}(u) + D\}.\]
Informally speaking, $\mathrm{frontier}^D(u)$ contains each node $v$ whose token arrives at $u$ at most $D$ time units after $u$ receives the first token.
We are now ready to define a clustering $\mathcal{C}^{\mathrm{del}}$ based on a delay function $\mathrm{del}$ and a separation parameter $s$.
The clustering $\mathcal{C}^{\mathrm{del}}$ clusters all the nodes which have a small frontier of width $2 \text{s}$.
In particular, each node $u \in V(G)$ satisfying $|\mathrm{frontier}^{2\text{s}}(u)| \leq 100 \log \log(n)$ is included in some cluster of $\mathcal{C}^{\mathrm{del}}$.
More concretely, each clustered node $u$ gets clustered to the cluster corresponding to the node with the smallest identifier in the set $\mathrm{frontier}^0(u)$. In other words, it gets clustered to the minimizer of $\mathrm{wait}(u)$ with the smallest identifier. In the following text, we denote this node by $c_u$ (see \cref{fig:mpx}).
\begin{figure}
\centering
\includegraphics[width = .6\textwidth]{img/mpx.eps}
\caption{The figure shows the run of the clustering algorithm constructing $\mathcal{C}^{\mathrm{del}}$.
The algorithm can be seen as starting a breadth first search from a single node $\sigma$ connected to every node $u \in V(G)$ with an edge of length $\mathrm{del}(u)$ (the $\mathsf{CONGEST}\,$ implementation of the algorithm does not need to simulate any such node $\sigma$).
The value $\mathrm{wait}(u)$ is the time until the search reaches the node $u$. The node that reaches $u$ the first is denote $c_u$.
Moreover, we cluster only nodes such that the size of their frontier of width $2s$ is at most $O(\log\log n)$. For example, the node $v$ is not clustered because after it is reached by the first node $c_v$, it is reached by $\Omega(\log \log n)$ other nodes in the following $2s$ steps.
It can be seen that for any $w$ on the path from $c_u$ to $u$, we have $\mathrm{frontier}^{2s}(w) \subseteq \mathrm{frontier}^{2s}(u)$, hence the constructed clusters are connected.
}
\label{fig:mpx}
\end{figure}
\begin{lemma}
\label{lem:pairwise_del_clustering}
Let $\mathrm{del}$ be a delay function and $\mathcal{C}^{\mathrm{del}}$ the corresponding clustering described above. Then, $\mathcal{C}^{\mathrm{del}}$ has
\begin{enumerate}
\item strong diameter $O(\text{s} \log n)$,
\item $\text{s}$-hop degree at most $100 \log \log (n)$ and
\item the total number of clustered nodes is equal to $|\{u \in V \colon \mathrm{frontier}^{2\text{s}}(u)| \leq 100 \log \log(n)\}|$.
\end{enumerate}
Moreover, the clustering $\mathcal{C}^{\mathrm{del}}$ can be computed in $O(\text{s} \log n \log\log n)$ $\mathsf{CONGEST}\,$ rounds.
\end{lemma}
To prove the lemma, we first observe that the frontiers have the following property.
\begin{claim}
\label{cl:frontier}
Let $w$ be any node on a shortest path from $u$ to $c_u$ and let $D \ge 0$. Then,
\[
\mathrm{frontier}^D(w) \subseteq \mathrm{frontier}^D(u).
\]
Moreover, $c_w = c_u$.
\end{claim}
\begin{proof}
Fix any such $w$ and consider any $v \in \mathrm{frontier}^D(w)$. The definition of $v$ implies that
$$\mathrm{del}(v) + d(v, w) \le \mathrm{del}(c_u) + d(c_u, w) + D.$$
Since $w$ lies on a shortest path from $c_u$ to $u$, we can add $d(w, u)$ to both sides of the equation to conclude that
$$\mathrm{del}(v) + d(v, w) + d(w, u) \le \mathrm{del}(c_u) + d(c_u, u) + D = \mathrm{wait}(u) + D$$
thus
$$\mathrm{del}(v) + d(v, u) \le \mathrm{wait}(u) + D.$$
Hence, we have $v \in \mathrm{frontier}^D(u)$ and we are getting $\mathrm{frontier}^D(w) \subseteq \mathrm{frontier}^D(u)$ as needed.
We continue with the second part of the statement. In view of the above proof that implies that $\mathrm{frontier}^0(w) \subseteq \mathrm{frontier}^0(u)$, it suffices to show that $c_u \in \mathrm{frontier}^0(w)$.
To see this, we use the fact that $c_u \in \mathrm{frontier}^0(u)$ to write
\[\mathrm{del}(c_u) + d(c_u,u) = \mathrm{wait}(u) \leq \mathrm{wait}(w) + d(w,u)\]
Subtracting $d(w, u)$ from both sides of the equation and using that $w$ lies on a shortest path from $u$ to $c_u$ gives
\[\mathrm{del}(c_u) + d(c_u,w) \leq \mathrm{wait}(w).\]
Thus $c_u \in \mathrm{frontier}^0(w)$ and we are done.
\end{proof}
We continue with the proof of \cref{lem:pairwise_del_clustering}.
\begin{proof}
We start with the first property. Let $u$ be an arbitrary clustered node and recall that $c_u$ is its cluster center.
As $c_u \in \mathrm{frontier}^0(u)$, we have $d(c_u,u) = \mathrm{wait}(u)$.
Moreover,
\[\mathrm{wait}(u) = \min_{v \in V} \mathrm{del}(v) + d(v,u) \leq \mathrm{del}(u) + d(u,u) = \mathrm{del}(u) = O(\text{s} \log n).\]
Hence, we have $d(c_u,u) = O(\text{s} \log n)$.
Moreover, \cref{cl:frontier} gives that all nodes on a shortest path from $c_u$ to $u$ are also clustered to $c_u$, implying that the diameter of the cluster is $O(s \log n)$.
Next, we prove the second property.
Consider an arbitrary clustered node $w$. We first show that for an arbitrary clustered node $y$ with $d(w,y) \leq \text{s}$, it holds that $c_y \in \mathrm{frontier}^{2\text{s}}(w)$.
To see this, we first use the definition of $c_w$ to write
\begin{align*}
\mathrm{del}(c_y) + d(c_y,y)
\leq \mathrm{del}(c_w) + d(c_w,y)
\end{align*}
On one hand, we can use triangle inequality to lower bound the left-hand side by
\begin{align*}
\mathrm{del}(c_y) + d(c_y,y) \ge \mathrm{del}(c_y) + d(c_y, w) - d(w,y)
\end{align*}
On the other hand, we can use triangle inequality to upper bound the right hand side by
\begin{align*}
\mathrm{del}(c_w) + d(c_w,y)
\le \mathrm{del}(c_w) + d(c_w, w) + d(w, y)
\end{align*}
Putting the two bounds together, we conclude that
\begin{align*}
\mathrm{del}(c_y) + d(c_y, w) \le \mathrm{del}(c_w) + d(c_w, w) + 2d(w,y) \le \mathrm{wait}(w) + 2s
\end{align*}
where we used our assumption $d(w,y) \le 2s$. Above inequality is in other words saying that $c_y \in \mathrm{frontier}^{2\text{s}}(w)$ as we wanted to prove.
Now, let $u$ be an arbitrary clustered node and $P_u$ the unique path between $u$ and $c_u$ in the tree associated with the cluster. Furthermore, let $C \in \mathcal{C}^{\mathrm{del}}$ be a cluster with $d(P_u,C) \leq \text{s}$. Then there exists $w \in P_u$ and $y \in C$ with $d(w,y) \leq \text{s}$ and the discussion above implies $c_y \in \mathrm{frontier}^{2\text{s}} (w)$.
We now use \cref{cl:frontier} that says that $\mathrm{frontier}^{2\text{s}} (w) \subseteq \mathrm{frontier}^{2\text{s}}(u)$.
Hence, for each cluster $C$ with $d(P_u,C) \leq \text{s}$, the corresponding cluster center is contained in $\mathrm{frontier}^{2\text{s}}(u)$. As $u$ is clustered, $|\mathrm{frontier}^{2\text{s}}(u)| \leq 100 \log \log n$, and therefore the $\text{s}$-hop degree of $\mathcal{C}^{\mathrm{del}}$ is at most $100 \log \log (n)$.
The third property follows directly from the definition.
To finish the proof, we need to show that the algorithm can be implemented in $O(s \log n \cdot \log\log n)$ rounds.
To see this, note that we can compute for each node $u$ whether $|\mathrm{frontier}^{2s}(u)| \le 100\log\log n$ as follows: We run a variant of breadth first search that takes into account the delays.
Moreover, recall that in classical breadth first search, after a node is $u$ is reached for the first time by a packet sent from $c_u$, it broadcasts it to all its neighbors and then it does not redirect any other packets sent to it.
In our version of the search, each node stops redirecting only after at least $100\log\log n$ packets arrived (we do not take into account packets that have already arrived earlier) or after it counts $2s$ steps from the arrival of the first packet.
It can be seen that this algorithm can be implemented in the desired number of rounds. Moreover, every node $u$ learns the value of $|\mathrm{frontier}^{2s}(u)|$ whenever the value is at most $100\log\log n$ and it learns the value is larger than this threshold otherwise.
\end{proof}
In view of \cref{lem:pairwise_del_clustering}, it suffices to show that \cref{alg:pairwise_delays} stated next computes a delay function $\mathrm{del}$ such that the expected number of nodes $u$ with $|\mathrm{frontier}^{2\text{s}}(u)| \leq 100 \log \log (n)$ is at least $n/2$.
\begin{restatable}{algorithm}{delays}
\caption{Computes Delay Function $\mathrm{del}$}
\label{alg:pairwise_delays}
\begin{algorithmic}[1]
\Procedure{Delays}{}
\State $V_0^ { \textrm{active}} = V$
\State $R = \lfloor 2 \log(n) \rfloor$
\State $k = \lceil 100 \log \log (n) \rceil$
\For{$i = 1,2, \ldots, R$}
\State $W_{i,0} = \emptyset$
\For{$j = 1,2,\ldots,k$}
\State $W_{i,j} = W_{i,j-1} \cup S_{i,j} = \mathcal{A}_{i,j}(V_0^ { \textrm{active}}, V_1^ { \textrm{active}}, \ldots, V_{i-1}^ { \textrm{active}}, W_{i,j-1})$ \Comment{In this section, we obtain the set $S_{i,j}$ by including each vertex with probability $\frac{1}{2k}$, assuming only pairwise independence.}
\EndFor
\State $V_{i+1}^{ \textrm{active}} = W_{i,k}$
\EndFor
\State For every $v \in V, i_v = \max \{i \in \{0,1,\ldots,R\} \colon v \in V_i^ { \textrm{active}}\}$
\State For every $v \in V$, $\mathrm{del}(v) = 5\text{s} (R - i_v)$ \\
\Return $\mathrm{del}$
\EndProcedure
\end{algorithmic}
\end{restatable}
The algorithm runs in $R = \lfloor 2 \log(n)\rfloor$ phases.
In the beginning, all of the nodes are active and in each phase, roughly half of the active nodes stop being active. Hence, after the last phase, all nodes are inactive, with high probability.
In the end, each node gets a delay of $\mathrm{del}(v) = 5 \text{s} (R - i_v) = O(\text{s} \log(n))$, where $i_v$ is the last phase where $v$ was still active.
For the sake of analysis, we define a delay function $\mathrm{del}_i$ for every phase $i \in \{0,1,2\ldots,R\}$ by defining
\[\mathrm{del}_i(v) = 5\text{s}(R - \min(i,i_v)).\]
Note that the delay of each node $v$ decreases by $5 \text{s}$ in each phase until $v$ stops being active, in which case the delay remains unchanged until the end. We also have $\mathrm{del}(v) = \mathrm{del}_R(v)$.
Similar as above, we now define for every vertex $u \in V$,
\[\mathrm{wait}_i(u) = \min_{v \in V} \mathrm{del}_i(v) + d(v,u)\]
and for every $D \in \mathbb{N}_0$,
\[\\frontierD_i(u) = \left\{v \in V \colon \mathrm{del}_i(v) + d(v,u) \leq \mathrm{wait}_i(u) + D \right\}.\]
\begin{restatable}{definition}{alivedead}[$\textrm{alive}_i(u)$/$\textrm{dead}_i(u)$]
For every vertex $u \in V$ and $i \in \{0,1,\ldots,R\}$, let $\textrm{dead}_i(u) \subseteq \mathrm{frontier}^{2 \text{s}}_i(u) \setminus V_i^{\textrm{active}}$ be an arbitrary subset of size $\min(k, |\mathrm{frontier}^{2 \text{s}}_i(u) \setminus V_i^{\textrm{active}}|)$ and $\textrm{alive}_i(u) \subseteq \mathrm{frontier}^{2 \text{s}}_i(u) \setminus V_i^{\textrm{active}}$ be an arbitrary subset of size $\min(k - |\textrm{dead}_i(u)|, |\mathrm{frontier}^{2 \text{s}}_i(u) \cap dead_i(u)|)$.
\end{restatable}
Recall that our goal is to show that the expected number of nodes in the set $\{u \in V \colon |\mathrm{frontier}^{2\text{s}}(u)| \leq 100 \log \log (n)\}$ is at least $n/2$. To that end, we define a suitable potential function $\Phi_i$ for every phase $i$.
\begin{definition}[Outer Potential]
\[\Phi_i(u) = e^{\frac{|\textrm{dead}_i(u)|}{10}}\]
\[\Phi_i = \sum_{u \in V} \Phi_i(u) + 2^i|V^{\textrm{active}}_i|.\]
\end{definition}
The following lemma states that the outer potential is at most $2n$ at the beginning and if the outer potential is sufficiently small at the end, then the corresponding clustering clusters a large fraction of the vertices.
\begin{lemma}[Outer Potential Lemma]
\label{lem:pairwise_outer_potential_lemma}
We have $\Phi_0 \leq 2n$. Moreover, if $\Phi_R \leq 10 n \log(n)$, then $\{u \in V \colon |\mathrm{frontier}^{2\text{s}}(u)| \leq 100 \log \log (n)\} \geq \frac{9n}{10}$.
\end{lemma}
\begin{proof}
First, note that $\Phi_R \geq 2^R|V_R^ { \textrm{active}}| > 10n \log(n)|V_R^{ \textrm{active}}|$.
As we assume that $\Phi_R \leq 10 n \log(n)$, this directly implies $V_R^ { \textrm{active}} = \emptyset$.
In particular, every $u \in V$ with $|\mathrm{frontier}^{2\text{s}}(u)| > 100 \log \log (n)$ contributes
\[\Phi_R(u) = e^{\frac{|\textrm{dead}_R(u)|}{10}}\geq e^{\frac{\min(k,|\mathrm{frontier}^{2\text{s}}(u)|}{10}} \geq 100 \log(n) \]
to the potential. Hence, there can be at most $\Phi_R/(100 \log(n)) \leq n/10$ such nodes and therefore $\{u \in V \colon |\mathrm{frontier}^{2\text{s}}(u)| \leq 100 \log \log (n)\} \geq \frac{9n}{10}$, as desired.
\end{proof}
\begin{restatable}{definition}{pessimisticprobability}[Pessimistic Estimator Probability $p_{i,j}(u)$]
\[p_{i,j}(u) = I(\textrm{alive}_i(u) \cap W_{i,j} = \emptyset) \cdot \left(1 - \frac{|\textrm{alive}_i(u)|}{10k}\right)^{k-j}\]
\end{restatable}
\todo{Briefly mention the intuition behind it}
\begin{restatable}{definition}{innerpotential}[Inner Potential]
Let .
\[\phi_{i,j}(u) = p_{i,j}(u) e^{\frac{|\textrm{dead}_i(u)| + |\textrm{alive}_i(u)|}{10}}\]
and
\[\phi_{i,j} = \sum_{u \in V} \phi_{i,j}(u) + |W_{i,j}|2^{i+1} + \frac{k-j}{k}2^i |V_i^{\textrm{active}}|.\]
\end{restatable}
\begin{restatable}{lemma}{Sijalg}
\label{lem:pairwise_Sij_alg}
Assume that for every $i \in [R], j \in [k]$, algorithm $\mathcal{A}_{i,j}$ is a deterministic algorithm which in $R$ rounds computes a set $S_{i,j}$ such that for
\[Y_{i,j}(u) = 1 - |\textrm{alive}_i(u) \cap S_{i,j}| + \binom{|\textrm{alive}_i(u) \cap S_{i,j}|}{2},\]
it holds that
\[ \sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + |S_{i,j}|\cdot 2^{i+1} \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i|.\]
Then, \cref{alg:pairwise_delays} runs in $O(R \log(n))$ rounds and computes a delay function $\mathrm{del}$ satisfying $|\mathcal{C}^{\mathrm{del}}| \geq \frac{n}{2}$.
Suppose $\mathcal{A}_{ij}$ is a randomized algorithm that in $R$ rounds computes a set $S_{i,j}$ such that
\[ \mathbb{E}[\sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + |S_{i,j}|\cdot 2^{i+1} ]\leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i|\]
Then, \cref{alg:pairwise_delays} runs in $O(R \log(n))$ rounds and computes a delay function $\mathrm{del}$ satisfying $\mathbb{E}[|C^{\mathrm{del}}| \geq \frac{n}{2}]$.
\end{restatable}
\begin{proof}
Consider an arbitrary node $u \in V$. Note that $I(\textrm{alive}_i(u) \cap S_{i,j} = \emptyset) \leq Y_{i,j}(u)$. Therefore,
\begin{align*}
&Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} \\
&\geq I(\textrm{alive}_i(u) \cap S_{i,j} = \emptyset)\frac{p_{i,j-1}(u)e^{\frac{|\textrm{dead}_i(u)| + |\textrm{alive}_i(u)|}{10}}}{1 - (|alive_i(u)|/10k)} \\
&= I(\textrm{alive}_i(u) \cap S_{i,j} = \emptyset)\frac{I(\textrm{alive}_i(u) \cap W_{i,j-1} = \emptyset)\left( 1 - \frac{|\textrm{alive}_i(u)}{10k}\right)^{k-(j-1)}1.1^{|\textrm{dead}_i(u)| + |\textrm{alive}_i(u)|}}{1 - (|\textrm{alive}_i(u)|/10k)}\\
&= I(\textrm{alive}_i(u) \cap W_{i,j} = \emptyset)\left( 1 - \frac{|\textrm{alive}_i(u)|}{10k}\right)^{k-j}1.1^{|\textrm{dead}_i(u)| + |\textrm{alive}_i(u)|} \\
&= \phi_{i,j}(u).
\end{align*}
Therefore,
\begin{align*}
\phi_{i,j} &= \sum_{u \in V} \phi_{i,j}(u) + |W_{i,j}|2^{i+1} + \frac{k-j}{k}2^i|V_i^{\textrm{active}}| \\
&\leq \sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + (|S_{i,j}| + |W_{i,j-1}|)2^{i+1} + \frac{k-j}{k}2^i|V^{\textrm{active}}_i| \\
&\leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i| + |W_{i,j-1}|2^{i+1} + \frac{k-j}{k}2^i|V^{\textrm{active}}_i| \\
&= \phi_{i,j-1}.
\end{align*}
As $\phi_{i,j} \leq \phi_{i,j-1}$, \cref{lem:inner_potential} implies that $\Phi_R \leq 4n \log n$. Therefore, \cref{lem:pairwise_outer_potential_lemma} gives that $\{u \in V \colon |\mathrm{frontier}^{2s}(u)| \leq 100 \log \log (n)\} \geq \frac{9n}{10}$. Hence, according to \cref{lem:pairwise_del_clustering}, $\mathcal{C}^{\mathrm{del}}$ clusters at least $\frac{9n}{10} \geq \frac{n}{2}$ nodes, as needed.
In the randomized case, the same set of calculations gives $\mathbb{E}[\phi_{i,j}] \leq \phi_{i,j-1}$ and therefore \cref{lem:inner_potential} implies that $\mathbb{E}[\Phi_R] \leq 4n \log(n)$. Hence, a Markov bound implies that $Pr[\Phi_R \leq 10n \log(n)] \geq 1 - \frac{4}{10} = \frac{3}{5}$. Hence, according to \cref{lem:pairwise_del_clustering},$\mathcal{C}^{\mathrm{del}}$ clusters at least $\frac{9n}{10}$ nodes with probability at least $\frac{3}{5}$ and hence the expected number of clustered nodes is at least $\frac{n}{2}$.
The $\mathsf{CONGEST}\,$ round complexity of $O(R \log n \log \log n)$ directly follows from the fact that for each $i \in [R]$ and $j \in [k]$, algorithm $\mathcal{A}_{i,j}$ is invoked exactly once.
\end{proof}
Hence, it suffices to show that the final potential is small in expectation. This directly follows from the following lemma.
\begin{restatable}{lemma}{pairwisesampling}[Pairwise Sampling Lemma]
Let $i \in [N]$ and $j \in [k]$ be arbitrary.
\label{lem:pairwise_analysis_sampling_lemma}
Let $\{X(v) \colon v \in V^{\textrm{active}}_i\}$ be a family of pairwise independent $0/1$-random variables with $Pr[X(v) = 1] = p$ for some $p \in [\frac{1}{4k},\frac{1}{2k}]$ for every $v \in V^{\textrm{active}}_i$. If one defines $S_{i,j} = \{v \in V^{\textrm{active}}_i \colon X(v) = 1\}$, then
\[\mathbb{E} \left[\sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + |S_{i,j}|\cdot 2^{i+1} \right] \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i|.\]
\end{restatable}
\begin{proof}
Let $u \in V$ be arbitrary. We have
\begin{align*}
\mathbb{E}[Y_{i,j}(u)] &= \mathbb{E} \left[ 1 - |\textrm{alive}_i(u) \cap S_{i,j}| + \binom{|\textrm{alive}_i(u) \cap S_{i,j}|}{2} \right] \\
&= \mathbb{E}[1 - \sum_{v \in \textrm{alive}_i(u)} X(v) + \sum_{v \neq v' \in \textrm{alive}_i(U)} X(v)X(v')] \\
&= 1 - \sum_{v \in \textrm{alive}_i(u)} \mathbb{E}[X(v)] + \sum_{v \neq v' \in \textrm{alive}_i(U)} \mathbb{E}[X(v)]\mathbb{E}[X(v')] \\
&\leq 1 - p|\textrm{alive}_i(u)| + p^2|\textrm{alive}_i(u)|^2 \\
&\leq 1 - (p/2)|\textrm{alive}_i(u)| \\
&\leq 1 - \frac{|\textrm{alive}_i(u)|}{10k}.
\end{align*}
Therefore,
\begin{align*}
\mathbb{E} \left[\sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + |S_{i,j}|\cdot 2^{i+1} \right] &= \sum_{u \in V} \mathbb{E}[Y_{i,j}(u)] \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + \mathbb{E}[|S_{i,j}|]\cdot 2^{i+1} \\
&\leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i|.
\end{align*}
\end{proof}
We now have all the pieces together to prove \cref{thm:pairwise_low_degree}.
\begin{proof}[Proof of \cref{thm:pairwise_low_degree}]
\cref{lem:pairwise_analysis_sampling_lemma} implies that for every $i \in [R], j \in [k]$, there exists a $0$-round randomized $\mathsf{CONGEST}\,$ algorithm only using pairwise independence which computes a set $S_{i,j}$ such that
\[\mathbb{E} \left[\sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + |S_{i,j}|\cdot 2^{i+1} \right] \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i|.\]
Hence, according to
\cref{lem:pairwise_Sij_alg}, there exists a $0$-round $\mathsf{CONGEST}\,$ algorithm only using pairwise independence which computes a delay function $\mathrm{del}$ such that the expected number of clustered nodes in the clustering $\mathcal{C}^{\mathrm{del}}$ is at least $n/2$.
Thus, \cref{thm:pairwise_low_degree} now directly follows from applying \cref{lem:pairwise_del_clustering}.
\end{proof}
\paragraph{Global Derandomization}
In \cref{sec:pairwise_low_degree}, we have proven the following lemma.
\pairwisesampling*
Next, we explain how we can use \cref{lem:pairwise_analysis_sampling_lemma} together with the method of conditional expectation to prove the following Lemma.
\begin{lemma}[Global Derandomization]
Let $D$ be the diameter of the input graph.
For every $i \in [R]$, $j \in [k]$, there exists a deterministic algorithm $\mathcal{A}_{i,j}$ which in $\widetilde{O}((\text{s} + D) \log n)$ rounds computes a set $S_{i,j}$ such that
\[\sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + |S_{i,j}| \cdot 2^{i+1} \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i|.\]
\end{lemma}
\begin{proof}
Let $\{X(v) \colon v \in V_i^{\textrm{active}}\}$ be a set of $0/1$-random variables following some distribution $Q$ and let $S_{i,j} = \{v \in V^{\textrm{active}}_i \colon X(v) = 1\}$.
For each $u \in V$, we define
\[T_Q(u) = \mathbb{E} \left[Y_{i,j}(u)\frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + I(u \in S_{i,j}) \cdot 2^{i+1}\right]\]
and $T_Q = \sum_{u \in V} T_Q(u)$.
The proof follows along the exact same lines as in \cref{subsec:hittingset-impl} where also the method of conditional expectation was used. We will therefore only discuss the necessary changes.
Let $p \in [\frac{1}{4k},\frac{1}{2k}]$ such that $p = 2^{-\ell}$ for some $\ell \in \mathbb{N}$.
The pairwise distribution of \cref{lem:pairwise_analysis_sampling_lemma} can be realized with a random seed length of $B = O(\log 1/p \cdot \log n) = \widetilde{O}(\log n)$ by using the pairwise distribution used in~\cite{luby1993removing, berger1989efficient}.
Now, algorithm $\mathcal{A}_{i,j}$ runs in $B$ rounds. At the beginning of round $i$, the first $i-1$ bits of the random seed are fixed and we deon and in round $i$ the $i$-th bit of the random seed is fixed in such a way .
\end{proof}
\paragraph{Local Derandomization}
\begin{lemma}[Global Derandomization]
\label{lem:global_derandomization}
For every $i \in [R]$, $j \in [k]$, there exists a deterministic algorithm $\mathcal{A}_{i,j}$ which in $\widetilde{O}(\text{s} \log^2 n)$ rounds computes a set $S_{i,j}$ such that
\[\sum_{u \in V} Y_{i,j}(u) \frac{\phi_{i,j-1}(u)}{1 - (|\textrm{alive}_i(u)|/10k)} + |S_{i,j}| \cdot 2^{i+1} \leq \sum_{u \in V} \phi_{i,j-1}(u) + \frac{2^i}{k}|V^{\textrm{active}}_i|.\]
\end{lemma}
\subsection{Derandomizing \cref{sec:pairwise_subsampling}}
\begin{restatable}{theorem}{subsampling}
\label{thm:pairwise_subsampling}
Assume we are given a clustering $\mathcal{C}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log^2 (n))$ and
\item $\text{s}$-hop degree of at most $10 \log \log (n)$.
\end{enumerate}
There exists a deterministic $\mathsf{CONGEST}\,$ algorithm running in $\widetilde{O}(\text{s} \log(n))$ rounds which computes a clustering $\mathcal{C}^{out}$ with
\begin{enumerate}
\item strong diameter $O(\text{s} \log (n))$,
\item separation of $\text{s}$ and
\item the number of clustered nodes is at least $\frac{|\mathcal{C}|}{1000 \log \log (n)}$.
\end{enumerate}
Let $D$ be the diameter of the input graph. There also exists a deterministic $\mathsf{CONGEST}\,$ algorithm with the same guarantees running in $\widetilde{O}((D + \text{s}) \log(n))$ rounds.
\end{restatable}
\paragraph{Global Derandomization}
\paragraph{Local Derandomization}
\subsection{Global Derandomization}
\subsection{Local Derandomization}
\fi
\section{Hitting Set}
\label{sec:hittingset}
In this section, first, we introduce a variant of the hitting set problem. Next, we propose a simple randomized algorithm for this problem using only pairwise independence. In the end, we describe an efficient distributed/parallel derandomization of our randomized algorithm.
\subsection{Problem Definition}
Consider a collection $\mathcal{S} = \{S_1, \dots, S_N\}$ of $N$ subsets from the universe $\{1,\dots,n\}$ and let $w_i \geq 0$ be the weight that is assigned to $S_i$. We say a subset $H \subseteq [n]$ hits $S_i$ if $H \cap S_i \neq \emptyset$. Our goal is to find a small $H$ with a small cost. Cost of $H$ is total weights of $S_i$ that are not hit by $H$, i.e., $\sum_{i:S_i \cap H = \emptyset} w_i$. For a random subset $H$ that includes each element with probability $p$ independently, the expected size of $H$ is $\mathbb{E}[|H|] = np$ and its expected cost is
\begin{equation*}
\sum_{i=1}^{N} w_i (1-p)^{|S_i|} \approx \sum_{i=1}^{N} w_i e^{-|S_i|p} = \tau^p_{\mathcal{S}}.
\end{equation*}
For example, suppose the regular case where $|S_i| = \Delta$. For $p = 10\log N / \Delta$, a random subset hits all sets with high probability $1 - 1/\mathrm{poly}(N)$ and for $p = 1 / \Delta$, constant fraction of sets are hit. Two important examples for weights is when $w_i = 1$ and $w_i = |S_i|$. In the former, we simply count the number of not hit sets. The latter indeed appears in our applications for constructing spanners and distance oracles (see \Cref{sec:hittingset-app}). There, we get penalized for each not hit set by its size.
In many applications, the expected size and cost of a random subset are enough. The challenge is to find a subset deterministically. Based on this, we formulate the following problem where we combine our two objectives in one potential function.
\label{subsec:hittingset-define}
\begin{definition}[Hitting Set Problem]
\label{def:hittingset}
Given a collection $\mathcal{S} = \{S_1, \dots, S_N\}$ of $N$ subsets from the universe $\{1,\dots,n\}$, an integer weight $w_i \geq 0$ for each $S_i$, and a sampling parameter $p \in (0,1)$, find a subset $H$ that minimizes the potential function
\begin{equation}
\label{eq:hittingset-potential}
\Phi^{p}_{\mathcal{S}}(H) = \frac{\sum_{i=1}^{N} w_i \cdot \mathbb{1}[H\cap S_i = \emptyset]}{\tau^p_{\mathcal{S}}} + \frac{|H|}{np}.
\end{equation}
\end{definition}
So if $\Phi^{p}_{\mathcal{S}}(H) = O(1)$, then $H$ has size $O(np)$ and its cost is $O(\tau^{p}_{\mathcal{S}})$. Our goal is to find such a set with constant potential function deterministically and efficiently. In the rest, we assume that $N \geq n$ as we can add dummy sets with zero costs. We also assume that $p \leq 1/2$ to ensure that $1 - p = e^{-\Theta(p)}$. Note that the case $p \geq 1/2$ is trivial since we tolerate constant deviation from a random subset and for $p \geq 1/2$, the expected size of a random subset is at least $n/2$. So our hitting can include all the $n$ elements.
\paragraph{Hitting in Ordered Sets.}
There are applications where $H$ is partially penalized even if we hit $S_i$. The amount of cost depends on which element of $S_i$ is being hit. In \Cref{sec:hittingset-app}, we encounter a particular instance of this generalization which is described in the following.
For each $S_i$, there is no weight but there is an order $\pi_i(\cdot)$ on its elements where $\pi_i(j)$ denotes the $j$-th element of $S_i$ for $j=1,\dots,|S_i|$. Then, $H$ has to pay $k-1$ for $S_i$ if $\pi_i(k) \in H$ and
\begin{equation*}
H \cap \{\pi_i(1), \dots, \pi_i(k-1)\} = \emptyset.
\end{equation*}
If $H$ does not hit $S_i$ at all, it has to pay $|S_i|$. Cost of $H$ is the sum of the expenses incurred by each $S_i$. We call this problem \textit{hitting ordered set}. With this definition, the expected cost for a random $H$ is
\begin{equation*}
\sum_{i=1}^{N} \sum_{j=1}^{|S_i|} (1-p)^j
\end{equation*}
The hitting ordered set problem is related to the original setting of \Cref{def:hittingset} in the following sense.
\begin{lemma}
\label{lem:hittingset-ordered-reduction}
Given an instance $\mathcal{I}_1$ of the hitting ordered set problem with $N$ sets $S_1, \dots, S_N \subseteq [n]$, we can construct an instance $\mathcal{I}_2$ of the original hitting set problem (see \Cref{def:hittingset}) with $O(N \log n)$ sets in $O(\sum_{i=1}^{N} |S_i|)$ time such that the following holds: For any $H \subseteq [n]$, if $c_1$ is the cost of $H$ in $\mathcal{I}_1$ and $c_2$ is the cost of $H$ in $\mathcal{I}_2$, then $c_1 \leq c_2 \leq 3c_1$.
\end{lemma}
\begin{proof}
To construct $\mathcal{I}_2$, for each $S_i$ in $\mathcal{I}_1$, we add $O(\log n)$ sets to $\mathcal{I}_2$. Suppose $2^\ell \leq |S_i| < 2^{\ell + 1}$. For $j\in [\ell]$, let $S_i^{j} = \{\pi_i(1), \dots, \pi_i(2^j)\}$ and let $S_i^{\ell+1} = S_i$. This completes the construction of sets of $\mathcal{I}_2$. Weight of $S_i^j$ in $\mathcal{I}_2$ is its size $|S_i^j|$.
Consider a subset $H \subseteq [n]$ and let $k$ be the minimum index that $\pi_i(k) \in H$. Suppose $k$ is $|S_i|+1$ if there is no such index. So $H$ has to pay $k-1$ in $\mathcal{I}_1$. In $\mathcal{I}_2$, it has to pay $\sum_{j:|S_i^j| < k} |S_i^j|$ which lies in the range $[(k-1), 3(k-1)]$ and concludes the proof.
\end{proof}
\subsection{Iterative Sampling}
\label{subsec:hittingset-alg}
The goal of this section is to find $H$ with $\Phi^{p}_{\mathcal{S}} = O(1)$ for the hitting set problem \Cref{def:hittingset}. Let $\Delta = \max_{i \in [N]} |S_i|$. Our algorithm has $T = \lceil 8p\Delta\rceil$ iterations. We start with a randomized algorithm and then we derandomize it. For $t = 1,\dots, T$, let $\mathcal{P}^t$ be a pairwise-independent distribution over $n$ binary random variables $X^t_1, \dots, X^t_n \in \{0,1\}$ with bias $q = 4p/T$. That is:
\begin{align*}
\forall i \in [n], \forall b\in \{0,1\},&\quad \Pr[X^t_i = b] = q^b (1-q)^{(1-b)},\\
\forall i,j \in [n], i \neq j,\forall b,b' \in \{0,1\},&\quad \Pr[X^t_i = b, X^t_j = b'] = q^{b+b'} (1-q)^{2-(b+b')}.
\end{align*}
Let the random subset $G^t$ be $\{i \in [n] \mid X^t_i = 1\}$. We replace $G^t$s one by one with an explicit set $H^t$. The final output of the algorithm is $H = \cup_{t=1}^T H^t$. Suppose we are in iteration $t$. Our goal is to find $H^t$. Let
\begin{equation}
\label{eq:pessimistic-def}
Y^t_i = \sum_{j \in S_i} X^t_j - \sum_{j \in S_i} \sum_{k \in S_i: j < k} X^t_j X^t_k.
\end{equation}
If $G^t$ does not hit $S_i$, then $Y^t_i = 0$. Otherwise, $Y^t_i \leq 1$ (because $a \leq \binom{a}{2}+1$ for all positive integers $a$). So $1 - Y^t_i$ is always greater than or equal to $\mathbb{1}[G^t \cap S_i = \emptyset]$ and is a pessimistic estimator for the event that $G^t$ does not hit $S_i$. We have the following upper bound on $\mathbb{E}[1 - Y^t_i]$.
\begin{lemma}
\label{lem:hittingset-pessimistic-bound}
$\mathbb{E}[1-Y^t_i] \leq 1 - 3|S_i|p/T \leq e^{-|S_i|p/T}.$
\end{lemma}
\begin{proof}
Note that:
\begin{equation*}
\mathbb{E}[Y^t_i] = |S_i|q - \binom{|S_i|}{2}q^2 \geq |S_i|q - |S_i|^2q^2/2 \geq 3|S_i|q/4 = 3|S_i|p/T
\end{equation*}
where in the last inequality we use $q = 4p/T \leq 1/2\Delta \leq 1/2|S_i|$.
\end{proof}
For a subset $G \subseteq [n]$, we define the function $f^t(G)$ as
\begin{equation*}
f^t(G) = \frac{\sum_{i: S_i \cap (H^1 \cup \dots \cup H^{t-1}) = \emptyset} (1 - Y_i) \cdot w_i e^{-|S_i|(T - t)p/T}}{\tau^p_{\mathcal{S}}} + \frac{\sum_{i=1}^{n} X_i + \sum_{j=1}^{t-1} |H^j| + 4n(T - t)p/T}{4np}
\end{equation*}
where $X_i = \mathbb{1}[i \in G]$ and $Y_i$ is defined from $X_1, \dots, X_n$ similar to \Cref{eq:pessimistic-def}.
\begin{lemma}
\label{lem:hittingset-first-iteration}
$\mathbb{E}[f^1(G^1)] \leq 2.$
\end{lemma}
\begin{proof}
Note that $\mathbb{E}[\sum_{i=1}^{n} X^1_i] = nq = 4np/T$ and from \Cref{lem:hittingset-pessimistic-bound}, we have $\mathbb{E}[1 - Y_i^1] \leq e^{-|S_i|p/T}$. Plugging these two bounds completes the proof.
\end{proof}
\begin{lemma}
\label{lemma:hittingset-invariant}
For $t \geq 2$, we have:
\begin{equation*}
\mathbb{E}[f^t(G^t)] \leq f^{t-1}(H^{t-1}).
\end{equation*}
\end{lemma}
\begin{proof}
Consider a subset $S_i$. If one of $H^1, \dots, H^{t-2}$ hits $S_i$, then the contribution of $S_i$ to the both sides of the inequality is zero. Otherwise, if $H^{t-1}$ hits $S_i$, the contribution of $S_i$ to $\mathbb{E}[f^t(G^t)]$ is zero. Note that it may contribute a non-zero amount into the RHS since we use pessimistic estimator $1 - Y_i$. The only remaining case is when $S_i$ is not hit in any of the first $t-1$ iterations. Then, the contribution of $S_i$ to the LHS is \begin{equation*}
\mathbb{E}[1 - Y^t_i] \cdot w_i e^{-|S_i|(T - t)p/T} \leq w_i e^{-|S_i|(T - t + 1)p / T}
\end{equation*}
where we use \Cref{lem:hittingset-pessimistic-bound}. On the other hand, the contribution of $S_i$ to the RHS is exactly $w_i e^{-|S_i|(T - t + 1)p / T}$. So the contribution of each $S_i$ to the LHS is less than or equal to its contribution to the RHS. Since $\mathbb{E}[|G_t|] = nq = 4np/T$, the second term that controls the size in $f^t(\cdot)$ and $f^{t-1}(\cdot)$ are equal which completes the proof.
\end{proof}
\begin{theorem}
\label{thm:hittingset-sampling}
If $f^t(H^t) \leq \mathbb{E}[f^t(G^t)]$ for all $t = 1, \dots, T$, then
\begin{equation*}
\Phi^p_{\mathcal{S}}(H = H^1 \cup \dots \cup H^T) \leq 2.
\end{equation*}
\end{theorem}
\begin{proof}
From \Cref{lem:hittingset-first-iteration} and \Cref{lemma:hittingset-invariant}, we get that $f^T(H^T) \leq 2$. Comparing $f^T(H^T)$ and $\Phi^p_{\mathcal{S}}(H)$ term by term, we can easily see that $f^T(H^T) \geq \Phi^p_{\mathcal{S}}(H)$.
\end{proof}
If $\Delta \gg 1/p$, then the number of iterations can be quite large. However, we are mostly interested in the regime where the number of iterations is logarithmic. We can achieve this as stated in the following.
\begin{corollary}
\label{cor:hittingset-sampling-logn}
Let $\mathcal{S}^+ = \{\text{the first $10 \log N/p$ elements of $S_i$} \mid |S_i| \geq 10 \log N / p\}$ and $\mathcal{S}^{-} = \mathcal{S} \setminus \mathcal{S}^+$. Run the algorithm twice: once for $\mathcal{S}^-$ with the same set of weights as before and once on $\mathcal{S}^+$ by setting all weights to $N^2$. Let the output of these two runs be $H^-$ and $H^+$. Then:
\begin{equation*}
\Phi^p_{\mathcal{S}}(H = H^- \cup H^+) \leq 4.
\end{equation*}
Each run takes at most $O(\log N)$ iterations. Moreover, all sets in $\mathcal{S}$ with size at least $10\log N/p$ are hit by $H$.
\end{corollary}
\subsection{Implementation}
\label{subsec:hittingset-impl}
The remaining piece of \Cref{thm:hittingset-sampling} is to find $H^t$ such that $f^t(H^t) \leq \mathbb{E}[f^t(G^t)]$. We first start with the construction of a suitable pairwise distribution.
\paragraph{Construction of Pairwise Independent Distribution.} From the algorithm of the previous section, we need a pairwise distribution $\mathcal{P}$ on $n$ binary random variables $X_1, \dots, X_n \in \{0,1\}$ with bias $q$. Assume that $q = 2^{-\ell}$ for some $\ell \in \mathbb{N}$ and $n$ is a positive integer of the form $n = 2^{m}-1$ for $m \in \mathbb{N}$. We use the pairwise distribution that is used in~\cite{luby1993removing, berger1989efficient} which has a random seed of length $\ell m = O(\log 1/p \cdot \log n)$. Let us quickly recall the construction. We first assign an $\ell$-bit label $L_i$ to each $X_i$. Then, we set $X_i$ to one if and only if all the $\ell$ bits of $L_i$ is one. To construct the labels, we decompose the random seed $R$ into $\ell$ groups each containing $m$ bits as follows:
\begin{equation*}
R = r^0_0\dots r^0_{m-1}r^1_0\dots r^1_{m-1}\dots r^{\ell-1}_{0}r^{\ell-1}_{m-1}
\end{equation*}
The $j$-th group $r^{j}_0\dots r^{j}_{m-1}$ is for constructing the $j$-bit of $L_i$s. To define $L_i(j)$ (the $j$-th bit of $L_i$), we use the bit representation of $i$. Suppose $i = \sum_{k=0}^{m-1} b_k 2^k$. Then:
\begin{equation*}
L_i(j) = b_0 r^{j}_0 \oplus \dots \oplus b_{m-1} r^{j}_{m-1}
\end{equation*}
This completes the construction. In the course of derandomization, we fix the random seed bit by bit. Suppose we fix the first $B$ bits of $R$ to $b_0,\dots,b_{B-1} \in \{0,1\}$. This gives us a new distribution $\mathcal{Q}$. The following result by Berger, Rompel, and Shor~\cite{berger1989efficient} is an important tool to achieve work-efficient derandomization.
\begin{lemma}[\cite{berger1989efficient}, Section 3.2]
\label{lem:berger-etal}
For any given subset $A \subseteq [n]$, we can compute
\begin{equation*}
\sum_{i \in A} \mathbb{E}_{\mathcal{Q}}[X_i],\quad
\sum_{i \in A} \sum_{j \in A} \mathbb{E}_{\mathcal{Q}}[X_i X_j]
\end{equation*}
with $O(|A|)$ processors and in $O(\log n)$ depth in the $\mathsf{PRAM}\,$ model. In particular, we can compute these two quantities in $O(|A| \log n)$ time in the standard model.
\end{lemma}
\paragraph{Bit Fixing.}
Suppose we are in iteration $t$ and we want to find $H^t$ such that $f^t(H^t) \leq \mathbb{E}[f^t(G^t)]$. Suppose $\mathcal{P}^t$ is $\mathcal{P}$ as described above. If $q$ is not a power of two (which is needed for the pairwise construction), replace it with a power of two in the range $[q,2q)$. We can observe that for any $H$
\begin{equation}
\label{eq:enlarge-q}
\Phi^{2p}_{\mathcal{S}}(H) \geq \Phi^{p}_{\mathcal{S}}(H)/2
\end{equation}
So with this replacement, we lose at most a two factor in the final bound for the potential function. Now, we start to fix the bits of the random seed of $\mathcal{P}^t$. Suppose we already fixed the first $B$ bits of the random seed $R$ by $b_0,\dots,b_{B-1}$. Let $e_x = \mathbb{E}[f(G^{t+1}) \mid R(0)=b_0,\dots,R(B-1)=b_{B-1},R(B)=x]$ for $x \in \{0,1\}$. If $e_0 \leq e_1$, then we fix $b_B$ to zero. Otherwise, we fix it to one. Suppose all the $\ell m$ bits are fixed and suppose that the random variable $X_i$ is $v_i \in \{0,1\}$ when we set the random seed to $b_0 \dots b_{\ell m - 1}$. Then, we set $H^t$ to $\{ i \in [n] \mid v_i = 1\}$. We can easily observe that $f^t(H^t) \leq \mathbb{E}[f^{t}(G^t)]$.
\paragraph{$\mathsf{PRAM}\,$ Model.} We have all the ingredients for implementing the algorithm in the $\mathsf{PRAM}\,$ model. This leads to the following theorem.
\begin{theorem}
\label{thm:hittingset-pram}
There is a deterministic algorithm that solves the hitting set problem by finding a subset $H$ with $\Phi^{p}_{\mathcal{S}}(H) \leq 4$ and with $\widetilde{O}(\sum_{i=1}^{N} |S_i|)$ work and
\begin{equation*}
O(\lceil p\Delta\rceil \cdot \log 1/p \cdot \log^2 n)
\end{equation*}
depth in the $\mathsf{PRAM}\,$ model. Moreover, there is a deterministic algorithm that finds a subset $H$ with $\Phi^{p}_{\mathcal{S}}(H) \leq 8$ and such that $H$ hits all $S_i$s with size greater than $10 \log N / p$. This algorithm runs with $\widetilde{O}(\sum_{i=1}^{N} |S_i|)$ work and
\begin{equation*}
O(\log N \cdot \log 1/p \cdot \log^2 n)
\end{equation*}
depth in the $\mathsf{PRAM}\,$ model.
\end{theorem}
\begin{proof}
The first algorithm is based on \Cref{thm:hittingset-sampling} and the second algorithm is based on \Cref{cor:hittingset-sampling-logn}. In those two algorithms, the potential function is upper bounded by $2$ and $4$. Here, we can only guarantee $4$ and $8$. This is because $q$, the sampling probability of one iteration, may not be a power of two. As discussed before (see \Cref{eq:enlarge-q}), we can handle this issue by paying an extra factor two in the approximation factor. In one iteration, we have $O(\log 1/p \cdot \log n)$ bit fixing. For each bit, we need to compute two conditional expectation which takes $O(\log n)$ depth and $O(\log n \cdot \sum_{i=1}^{N} |S_i|)$ work using \Cref{lem:berger-etal}. Multiplying the number of iterations gives us the claimed bounds.
\end{proof}
\paragraph{$\mathsf{CONGEST}\,$ Model.}
First, let us describe how the hitting set problem is represented in the distributed model. Consider an $(N+n)$-node bipartite network $G = (A \sqcup B, E)$ where $A = [N]$ and $B = [n]$. A node $i \in A$ represents set $S_i$ and a node $j \in B$ represents element $j \in [n]$. There is an edge between $i \in A$ and $j \in B$ if and only of $j \in S_i$. We assume that $p$, $n$, and $\tau_{\mathcal{S}}^p$ (or an upper bound of it) is known to all nodes.
To simulate global decision making, we use $3$-separated network decomposition. We need to execute the following operation fast: For an arbitrary color $j$, let $C_1, \dots, C_d$ be the set of clusters with color $j$ in the given $3$-separated network decomposition. Suppose that each node $v$ in $C_1 \cup \dots \cup C_d$ knows a value $a_v$. For each cluster $C_i$, we want to broadcast the value $\sum_{v \in C_i} a_v$ to all nodes in $C_i$. We denote the round complexity of executing this operation for all clusters $C_1, \dots, C_d$ by $T^{\mathrm{agg}}_{\mathrm{ND}}$.
\begin{theorem}
\label{thm:hittingset-congest}
Given a $Q$-color $3$-separated network decomposition with aggregation time $T^{\mathrm{agg}}_{\mathrm{ND}}$ (as described above), there is a deterministic algorithm that solves the hitting set problem by finding a subset $H$ with $\Phi_{\mathcal{S}}^{p}(H) \leq 4$ in
\begin{equation*}
O(\lceil p\Delta \rceil \cdot Q \cdot \log 1/p \cdot \log n \cdot T_{\mathrm{ND}}^{\mathrm{agg}})
\end{equation*}
rounds of the $\mathsf{CONGEST}\,$ model. Moreover, there is a deterministic algorithm that finds a subset $H$ with $\Phi_{\mathcal{S}}^{p}(H) \leq 8$ and such that $H$ hits all $S_i$s with size greater than $10 \log N / p$. This algorithm runs in
\begin{equation*}
O(\log N \cdot Q \cdot \log 1/p \cdot \log n \cdot T_{\mathrm{ND}}^{\mathrm{agg}})
\end{equation*}
rounds of the $\mathsf{CONGEST}\,$ model.
\end{theorem}
\begin{proof}
We want to derandomize iteration $t$. In contrast to the $\mathsf{PRAM}\,$ model \Cref{thm:hittingset-pram}, in the $\mathsf{CONGEST}\,$ model, we do not have global communication and so we cannot decide which bit should be fixed in a straightforward way. However, we can simulate such global decision-making with network decomposition paying an extra factor $Q$ in the round complexity. For each cluster $C$, we independently draw a sample from the pairwise-independent distribution $\mathcal{P}$ with bias $q$. Recall that the input graph is a bipartite graph $G = (A \sqcup B, E)$. These samples assign a binary value to each node of $B$. Observe that the assigned values are also pairwise independent since the product of pairwise independent distributions is pairwise-independent. Now, to derandomize, we go through the colors one by one. Suppose we are working on color $j \in [Q]$ with $d$ clusters $C_1, \dots, C_d$. Moreover, suppose the first $b$ bits of random seeds of $C_1, \dots, C_d$ are fixed. We fix the $(b+1)$-th bit. Let us emphasize that each cluster has its own random seed and different clusters may fix the $(b+1)$-th bit differently. Consider cluster $C_i$ and a node $v \in A$ that is either in $C_i$ or is in the boundary of $C_i$ (i.e., $v$ is not in $C_i$ but has a neighbor in $C_i$). So this node represents a set $S_v$ in the corresponding hitting set problem. We assign two values $a^0_v$ and $a^1_v$ to $v$ where $a^x_v$ corresponds to the case when we fix the $(b+1)$-bit of the random seed of $C_i$ to $x$. Note that each neighbor of $v$ represents an element of $S_v$. If $v$ has a neighbor in the clusters with color $\{1,\dots,j-1\}$ that is already decided to be in $H$ (our final hitting set), then we set $a^x_v$ to zero. So suppose this is not the case and let $d$ be the number of neighbors of $v$ that are not in $C_i$ and are in a cluster with color in $\{j+1, \dots, Q\}$. Then, we set $a_v^x$ to $$\frac{(1 + \binom{d}{2}q^2 - dq) \cdot F_v^x \cdot w_v e^{-|S_i|(T - t) p /T}}{\tau_{\mathcal{S}}^{p}}$$ where $F_v^b$ is
\begin{equation*}
F_v^x = \mathbb{E}[1 + \sum_{u \in C_i \cap B: u\in S_v} \sum_{u \in C_i \cap B: u\in S_v \wedge u < u'} X_u X_{u'} - \sum_{u \in C_i \cap B: u\in S_v} X_u \mid \text{first $b$ bits and $(b+1)$-th bit is $x$}]
\end{equation*}
where $X_u$ represents the indicator random variable of element $u$. Note that the given network decomposition is $3$-separated and so all the boundaries of $C_1, \dots, C_d$ are disjoint. So $v$ can compute $F_v^b$ in $\widetilde{O}(|S_v|)$ according to \Cref{lem:berger-etal}. Also, note that that this gives us the contribution of $S_v$ to
\begin{equation*}
\mathbb{E}[f^{t}(\cdot) \mid \text{first $b$ bits and $(b+1)$-th bit is $x$}]
\end{equation*}
Next, for each element $u \in A_i$, set $a_u^x$ to
\begin{equation*}
\frac{\mathbb{E}[X_u \mid \text{first $b$ bits and $(b+1)$-th bit is $x$}]}{4np}.
\end{equation*}
In the end, for each cluster $C_i$, we compute two values $e_i^x$ for $x \in \{0,1\}$ which is $$\sum_{v \in C_i \cup (\partial(C_i) \cap B)} a_v^x$$ where $\partial(C_i)$ denotes the boundary of $C_i$. We broadcast $e^b_i$ to each nodes in $C_i$. This can be done in $O(T_{\mathrm{ND}}^{\mathrm{agg}})$ rounds for all $C_i$s simultaneously. Next, nodes of $C_i$ set the $(b+1)$-bit of the random seed to zero if $e^0_i \leq e^0_i$ and set it to one otherwise. This completes the bit fixing.
There are $T$ sampling iterations (if we apply \Cref{thm:hittingset-sampling}, $T = \lceil p\Delta\rceil$, and if we apply \Cref{cor:hittingset-sampling-logn}, $T = O(\log N)$), $Q$ colors, and $O(\log 1/p \cdot \log n)$ bits to fix for each color. Multiplying these numbers gives us the number of bit fixing. Taking into account that fixing each bit takes $O(T_{\mathrm{ND}}^{\mathrm{agg}})$ rounds of the $\mathsf{CONGEST}\,$ model concludes the proof.
\end{proof}
\begin{corollary}
\label{cor:hittingset-congest-logn}
There is a deterministic algorithm that solves the hitting set problem by finding a subset $H\subseteq [n]$ with $\Phi^p_{\mathcal{S}}(H) = O(1)$ in $\mathrm{poly}(\log n)$ rounds of the $\mathsf{CONGEST}\,$ model and with total computations $\widetilde{O}(m)$.
\end{corollary}
\begin{proof}
There is a work-efficient deterministic algorithm for finding a $3$-separated $O(\log n)$-color network decomposition in $\textrm{polylog}(n)$ rounds and with $T_{\mathrm{ND}}^{\mathrm{agg}} = \textrm{polylog}(n)$(see Theorem 2.12 of Rozho\v{n} and Ghaffari~\cite{rozhonghaffari20}). Plugging this bound in \Cref{thm:hittingset-congest} concludes the proof.
\end{proof}
\section{Applications of Hitting Set}
In this section, we discuss two applications of the hitting set problem. One is the distributed construction of multiplicative spanners and the other is the parallel construction of distance oracles. Let us quickly define these notions. A subgraph $H = (V,E') \subseteq G = (V,E)$ is an $\alpha$-spanner of $G$ if for all pairs of nodes $u, v \in V$, we have:
\begin{equation*}
d_G(u,v) \leq d_H(u,v) \leq \alpha \cdot d_G(u,v).
\end{equation*}
A distance oracle is a data structure that accepts a pair of nodes $(u,v)$ as a query and returns their distance in $G$. In \Cref{subsec:distance-oracles}, we discuss \textit{source-restricted approximate distance oracle} in which $s$ nodes of $G$ are marked as source and it is guaranteed that $u$ is always a source. The term ``approximate'' allows the oracle to return an approximation of $d_G(u,v)$ rather than its exact value.
\label{sec:hittingset-app}
\subsection{Spanners}
\label{subsec:spanners}
\iffalse
In the paper~\cite{bezdrighin2022deterministic}, Bezdrighin et al. show
the following results on the deterministic construction of spanners in the $\mathsf{CONGEST}\,$ model are presented:
\begin{itemize}
\item A $\textrm{polylog}(n)$ rounds algorithm for constructing $(2k-1)$-spanner with $O(nk + n^{1 + 1/k} \log k)$ and $O(nk + n^{1 + 1/k} k)$ edges for unweighted and weighted graphs, respectively.
\item A $\textrm{polylog}(n)$ rounds algorithm for constructing $O(\log n \cdot 2^{\log^* n})$-spanner and $O(\log n \cdot 4^{\log^* n})$-spanner with $O(n)$ edges for unweighted and weighted graphs, respectively.
\end{itemize}
While all of these algorithms have poly-logarithmic round complexity, the total amount of computations by nodes is slightly super-polynomial, i.e., $O(n^{\log \log n})$. Keeping the $\textrm{polylog}(n)$ round complexity, here we improve the amount of computations to near-linear time, i.e., $\widetilde{O}(|E(G)|)$.
\fi
\begin{theorem}
There is deterministic algorithm in $\mathrm{poly}(\log n)$ rounds of the $\mathsf{CONGEST}\,$ model and with total computations $\widetilde{O}(m)$ that finds a $(2k-1)$-spanner with $O(nk + n^{1 + 1/k} \log k)$ and $O(nk + n^{1 + 1/k} k)$ edges for unweighted and weighted graphs, respectively.
\end{theorem}
\begin{proof}
We derandomize Baswana-Sen algorithm~\cite{baswana2007simple}. Let us quickly recall this algorithm. It consists of $k$ steps. The input of step $i$ is a clustering denoted by $\mathcal{C}_i$. Each cluster has a center node known to all of its members. The input of the first step is the trivial clustering: there are $n$ clusters each containing a single node. During one step, we sample some of the clusters, and then based on that sampling, some nodes stay in their clusters, some get unclustered, and some join other clusters. After this, the current step $i$ terminates, and the new clustering $\mathcal{C}_{i+1}$ is passed to the next step. Here is what we do in step $i$ for $i \leq k-1$ (we discuss the last step, $i$ equals $k$, later):
\begin{enumerate}
\item Each cluster of $\mathcal{C}_i$ is sampled with probability $p = n^{-1/k}$.
\item A node that is in a sampled cluster, stays put in its own cluster.
\item For a node $v$ in an unsampled cluster, let $C_1, \dots, C_d$ be the set of clusters containing at least one neighbor of $v$. Let $e_i = \{u_i \in C_i, v\}$ be an edge with the minimum weight between $v$ and one of the nodes in $C_i$. If there are several edges with the minimum weight, $v$ selects one of them arbitrarily. Let $w_i$ be the weight of $e_i$. Without loss of generality, suppose $w_1 \leq \dots \leq w_d$. If all of $C_1, \dots, C_d$ are unsampled, $v$ adds all edges $e_1, \dots, e_d$ to the output spanner and gets unclustered. Otherwise, let $j$ be the minimum index for which $C_j$ is sampled. Then, $v$ adds $e_1, \dots, e_{j}$ to the output spanner and joins the sampled cluster $C_j$. Note that all such $v$ runs this step simultaneously.
\end{enumerate}
In the last step, we do the exact same thing except that we sample no cluster (each cluster is sampled with probability zero rather than $n^{-1/k}$).
The output of Baswana-Sen is always a $(2k-1)$-spanner and only the size of the output depends on the randomness. From the algorithm description, you can see that the only randomized part of the Baswana-Sen algorithm is the sampling of clusters. Our goal is to find the set of sampled clusters of each step deterministically. If we have the following properties on the set of sampled clusters, then we can guarantee the claimed bounds on the size of the output spanner (see~\cite{bezdrighin2022deterministic}, Lemma 3.3):
\begin{enumerate}[(a)]
\item For each $i$, the number of clusters in $\mathcal{C}_i$ is at most $n^{1-(i-1)/k}$.
\item The number of edges added to the output spanner is bounded as follows: For the unweighted case, the total number of edges added by nodes with at least $\gamma_1 n^{1/k} \log k$ neighboring clusters for a large enough constant $\gamma_1 > 0$ is at most $O(n^{1 + 1/k}/k)$. For the weighted case, all nodes add at most $O(n^{1 + 1/k})$ edges to the output.
\item A node that is clustered in $\mathcal{C}_i$, remains clustered if it has at least $\gamma_2 n^{1/k} \log n$ neighbouring clusters for a large enough constant $\gamma_2 > 0$.
\end{enumerate}
We can frame these properties as a hitting set problem. To avoid cluttering the notation, we refer to the universe size in the corresponding hitting set problem of step $i$ by $n^h_i$ and its number of sets by $N^h_i$. In step $i$, we have the following hitting set problem: There is an element in the universe for each cluster in $\mathcal{C}_i$. So $n^h_i = |\mathcal{C}_i| \leq n$. For each clustered node $v$ in $\mathcal{C}_i$, there is a set $S_v$ containing all of its neighboring clusters. So $N^h_i \leq n$. The parameter $p$ for the hitting set problem is set to the sampling probability of Baswana-Sen divided by a large enough constant $\gamma_3 > 0$, i.e., $p = n^{-1/k}/\gamma_3$ (note that the last step is already deterministic and no derandomization is needed there). For unweighted graphs, we set the weight of $S_v$ to its size $w_v = |S_v|$. For weighted graphs, we consider the hitting ordered set problem as discussed in \Cref{lem:hittingset-ordered-reduction}. For each clustered node $v$, we assign the order $\pi_v(\cdot)$ on $S_v$. Suppose that the neighboring clusters of $v$ are $C_1, \dots, C_d$ and the minimum weight of an edge between $C_i$ and $v$ is $w_i$. Then $C_i$ comes before $C_j$ in $\pi_v(\cdot)$ if $w_i < w_j$ or $w_i = w_j$ and $i < j$.
With straightforward calculations, we can see that all the three required properties are satisfied if we solve the presented hitting set problem with \Cref{cor:hittingset-congest-logn} (for the hitting ordered set problem, we first use the reduction \Cref{lem:hittingset-ordered-reduction}).
We have $k \leq \log n$ steps in total. As described above, each step can be derandomized by solving a hitting set problem. So the total round complexity is $\mathrm{poly}(\log n)$ by applying \Cref{cor:hittingset-congest-logn}. One issue here is that each element in the defined hitting set problem corresponds to a cluster. This issue can be handled by contracting each cluster to a node and using the fact that the network decomposition of~\cite{rozhonghaffari20} also works on contracted graphs. This slows down the round complexity only by a factor $k = O(\log n)$ as each cluster has diameter $k$.
\end{proof}
\begin{theorem}
For any $\varepsilon > 0$, there is deterministic distributed algorithm in $\mathrm{poly}(\log n) / \varepsilon$ rounds of the $\mathsf{CONGEST}\,$ model and with total computations $\widetilde{O}(m)$ that finds a spanner with size $n(1 + \varepsilon)$ and with stretch $O(\log n \cdot 2^{\log^* n} / \varepsilon)$ and $O(\log n \cdot 4^{\log^* n} / \varepsilon)$ stretch for unweighted and weighted graphs, respectively.
\end{theorem}
\begin{proof}
We derandomize the algorithm of Pettie~\cite{pettie2010distributed} to get a spanner with $O(n)$ edges and with stretch $O(\log n \cdot 2^{\log^* n})$ and $O(\log n \cdot 4^{\log^*} n)$ for unweighted and weighted graphs, respectively. Pettie's algorithm is combining $O(\log^* n)$ application of Baswana-Sen back to back and the hitting set problem we encounter in Pettie's algorithm, is exactly the same as the Baswna-Sen. So we do not repeat this here. We refer interested readers to Theorem 1.5 of ~\cite{bezdrighin2022deterministic} where the full algorithm and a slower derandomized version of it is discussed. Let us note that the original algorithm of Pettie only works for unweighted graphs, but with a simple modification which is proposed in~\cite{bezdrighin2022deterministic}, it can work on weighted graphs as well. To reduce the number of edges from $O(n)$ to $n(1 + \varepsilon)$, we apply the deterministic reduction of ~\cite{bezdrighin2022deterministic}, Theorem 1.2.
\end{proof}
\subsection{Approximate Distance Oracles}
\label{subsec:distance-oracles}
This section is devoted to the parallel implementation of the approximate distance oracle by Roditty, Thorup, and Zwick~\cite{roditty2005deterministic}. There, given a weighted graph $G = (V,E)$, a stretch parameter $k$, and a set of $s$ sources $S \subseteq V$, they deterministically construct a data structure of size $O(kns^{1/k})$ and in $\widetilde{O}(ms^{1/k})$ time. For a query $(u,v)$, the data structure can compute a value $q$ such that
\begin{equation*}
d(u,v) \leq q \leq (2k-1) d(u,v)
\end{equation*}
in $O(k)$ time. See \Cref{alg:distance-oracle} for their algorithm for constructing the data structure and \Cref{alg:distance-oracle-query} for how they evaluate a query.
\begin{algorithm}
\caption{Approximate Distance Oracle~\cite{roditty2005deterministic}}
\label{alg:distance-oracle}
\begin{algorithmic}[1]
\Procedure{DistOracle}{G, k}
\State $A_0 = S, A_k = \emptyset$.
\State $\ell = 10s^{1/k} \log n$.
\For{$i=1,\dots,k-1$}
\State For each $v \in V$, find $p_i(v) \in A_{i-1}$ such that $d(p_i(v), v) = d(A_{i-1}, v).$
\State For every $v \in V$, compute $N_{i-1}(v)$ which is the set of $\ell$ closest nodes to $v$ in $A_{i-1}$.
\State Find a set $A_i \subseteq A_{i-1}$ such that:
\Statex $\quad \quad \quad \quad$ (a) $|A_i| \leq s^{1-i/k}$.
\Statex $\quad \quad \quad \quad$ (b) $A_i$ hits $N_{i-1}(v)$ for all $v \in V.$
\Statex $\quad \quad \quad \quad$ (c) $\sum_{v \in V} |\{w \in A_{i-1} - A_{i} \mid d(w,v) < d(A_i, v)\}| = O(ns^{1/k}).$
\EndFor
\State For each $v \in V$, compute $p_{k-1}(v)$.
\State For every $v \in V$, set $B(v) = A_{k-1}$.
\For{$i=0,\dots,k-2$}
\State For every $v \in V$, set $B(v) = B(v) \cup \{w \in N_i(v) \mid d(w,v) < d(A_{i+1}, v)\}$.
\EndFor
\State For each $v \in V$, create a hash table $H(v)$ with an entry $(v,d(v,w))$ for each $w \in B(v)$.
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{Evaluating a query~\cite{roditty2005deterministic}}
\label{alg:distance-oracle-query}
\begin{algorithmic}[1]
\Procedure{Query}{$u \in S$,$v$}
\State $w = u$, $i = 0$.
\While{$w \not \in B(v)$}
\State $i = i + 1$.
\State $(u,v) \leftarrow (v,u)$.
\State $w \leftarrow p_i(u)$
\EndWhile
\Return $d(w,u) + d(w,v)$
\EndProcedure
\end{algorithmic}
\end{algorithm}
\begin{theorem}
\label{thm:distance-oracle}
Given an undirected weighted graph $G = (V,E)$, a set of $s$ sources $S \subseteq V$, stretch parameter $k$, and error $\varepsilon > 0$, there is a deterministic algorithm that solves the source-restricted distance oracle problem with $\widetilde{O}_{\varepsilon}(ms^{1/k})$ work and $\widetilde{O}_{\varepsilon}(\mathrm{poly}(\log n))$ depth in the $\mathsf{PRAM}\,$ model. The data structure has size $O(nks^{1/k})$ and for each query $(u,v)$, the oracle can return a value $q$ in $O(k)$ time that satisfies
\begin{equation*}
d(u,v) \leq q \leq (2k-1)(1 + \varepsilon)d(u,v).
\end{equation*}
\end{theorem}
\begin{proof}
It is enough to provide a parallel algorithm with $\widetilde{O}_{\varepsilon}(\mathrm{poly}(\log n))$ depth for computing $A_i$, $N_i(\cdot)$, and the hash table. This gives us all the ingredients we need to run the algorithm.
Note that finding a suitable $A_i$ in \Cref{alg:distance-oracle} is just an instance of hitting ordered set problem and we can apply \Cref{lem:hittingset-ordered-reduction} and \Cref{cor:hittingset-sampling-logn}. The universe is $A_{i-1}$ and for each $v \in V$, we want to hit the set $N_{i-1}(v)$. We also need to determine $\pi_{i,v}(\cdot)$. An element $w$ comes before $w'$ in this order if $d(w,v) < d(w',v)$. If the distances are equal, we break the tie based on the identifier of $w$ and $w'$. If we set the sampling probability to $p = s^{-1/k}/\gamma$ for a large enough constant $\gamma > 0$ (indeed $\gamma = 24$ is enough), then we can compute a suitable $A_i$ satisfying all the three required properties with $\widetilde{O}(m)$ work and $O_{\varepsilon}(\mathrm{poly}(\log n))$ depth in the $\mathsf{PRAM}\,$ model using \Cref{thm:hittingset-pram} and the reduction \Cref{lem:hittingset-ordered-reduction}.
In~\cite{roditty2005deterministic}, they compute $N_i(\cdot)$ by running $\ell$ instances of Single Source Shortest Path problem (SSSP). There is no known parallel algorithm for SSSP with poly-logarithmic depth. However, recently, Rozhoň et al.~\cite{rozhovn2022undirected} proposed a work-efficient algorithm for computing $(1+\varepsilon)$-approximation of SSSP with poly-logarithmic depth. We can replace the exact computation with an approximation, losing $(1+\varepsilon)$ in the final stretch guarantee.
For computing the hash tables, we can apply the construction of Alon and Naor~\cite{alon1996derandomization}. There, they provide a deterministic hash table of $t$ elements into $O(t)$ space with read access of $O(1)$ time. While they did not discuss the parallel implementation of their construction, their algorithm can be implemented in $\mathrm{poly}(\log n)$ depth in a straightforward way. Their approach is derandomizing a randomized hash function using the method of conditional expectation on epsilon-biased spaces. They define a potential function (see section 3.1. of~\cite{alon1996derandomization}) which is a simple aggregation and can be parallelized. We do not discuss the full details as the implementation is straightforward.
\end{proof}
\newpage
\section*{Acknowledgments}
M.G., C.G., S.I., and V.R. were supported in part by the European Research Council (ERC) under the European Unions Horizon 2020 research and innovation program (grant agreement No.~853109) and the Swiss National Science Foundation (project grant 200021\_184735). B.H. was supported in part by NSF grants CCF-1814603, CCF-1910588, NSF CAREER award CCF-1750808, a Sloan Research Fellowship, funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement 949272), and the Swiss National Science Foundation (project grant 200021\_184735).
\bibliographystyle{alpha}
|
\section{Introduction}
At the heart of modern light emitting diodes (LEDs) operating in the blue to violet spectral region are (In,Ga)N/GaN multi-quantum well (MQW) systems~\cite{Hump2008}. While the efficiency of these LEDs is and can be very high, further efficiency gains will still directly reduce the cost of operating such LEDs. Moreover, extending efficient operation of (In,Ga)N-based LEDs into the green to red spectral range is a topic of current research interest~\cite{QuLiACSP2019,ScBiPhsStatSol2013, deMaPecPRL2016,DuBaJAP2020,HwHaAPE2014}. To achieve all this, understanding the fundamental electronic, optical and transport properties of (In,Ga)N-based MQW systems is of central importance to guide design of (In,Ga)N-based LEDs with new and improved capabilities. While experimental and theoretical studies on the electronic~\cite{CaSc2011,ChODo2021,TaCaRSC2016} and optical~\cite{JoTeAPL2017,DaScJAP2016} properties of such systems have already revealed that these properties are significantly impacted by alloy fluctuation induced carrier localization effects, the impact of alloy disorder on the carrier transport has only been targeted recently~\cite{ODoLu2021,MiOD2021_JAP,OdFaOQE2022,LyLhPhysRevApplied2022,BrMa2015,LiPi2017}. Here, studies are ranging from fully atomistic quantum mechanical approaches~\cite{ODoLu2021} up to modified continuum-based models~\cite{MiOD2021_JAP,OdFaOQE2022,LyLhPhysRevApplied2022,BrMa2015,LiPi2017}.
Recently, we have developed a three-dimensional (3-D) multiscale simulation framework that connects atomistic tight-binding theory with a modified, quantum corrected drift-diffusion (DD) solver. The framework has been employed to investigate uni-polar carrier transport and it was found that alloy fluctuations result in an increase in carrier transport of electrons (in an $n$-$i$-$n$ system)~\cite{MiOD2021_JAP}, but decreases transport in the case of holes (in a $p$-$i$-$p$ systems)~\cite{OdFaOQE2022}. In the present work we extend this scheme to investigate the active region of (In,Ga)N-based MQW LED structures (thus $p$-$i$-$n$ systems) and study how carriers distribute across the active region. In general, understanding the carrier distribution can help to guide maximizing the efficiency in an LED, since ideally the carriers shall be distributed evenly across the entire MQW region so that all QWs will contribute to emission~\cite{ZhZhJAP2011}.
However, previous experimental studies on carrier distribution in (In,Ga)N/GaN MQW systems have indicated that mainly the well closest to the $p$-doped contact side contributes to the light emission process~\cite{BGaPhysStatSol_2011,AuGrAPL2008,LiRyAPL2008,ZhZhJAP2011}. These samples were specifically designed to gain insight into the carrier distribution inside the active region of an LED.
Overall, this has been attributed to a sequential filling of the QWs, resulting in a high hole density only in the $p$-side QW. To establish accurate carrier transport models the trends found in the experimental studies of Refs.~\cite{BGaPhysStatSol_2011,AuGrAPL2008,LiRyAPL2008,ZhZhJAP2011} need to be captured. Previous theoretical studies have reproduced the experimentally observed behaviour, however this required (i) treating bound carriers in a quantum mechanical picture, (ii) softening of the QW barrier interface to account for tunneling effects, (iii) distinguishing between continuum and bound carriers in the carrier transport model (multi-population model), and (iv) allowing for scattering between the different populations~\cite{RoWiSPIE2017}. But, the impact of alloy disorder is basically neglected in this advanced but also complex carrier transport model.
In this paper we show that when employing our quantum corrected 3-D simulation framework that accounts for random alloy fluctuations, the experimentally observed trends are captured, without introducing for instance a multi-population scheme. This highlights that our developed solver presents an ideal starting point for future device design studies.
To highlight clearly the impact that random alloy fluctuations have on the carrier distribution in the active region of an (In,Ga)N-based LED, we use as a reference point a virtual crystal approximation (VCA) which effectively can be described by a 1-D model. The benefit of this is twofold. Firstly, this enables us to compare directly the outcomes of our quantum corrected model with results from 1-D commercial software simulations;
commercial software packages often employ a standard Schr\"odinger-Poisson-DD solver, which is numerically very costly and therefore unfeasible in large 3-D transport simulations. This motivates the need for an alternative implementation of quantum corrections.
Secondly, and building on this benchmark, alloy fluctuations can be included in the calculations, revealing clearly their impact on the results.
Our studies show, and when using the same input parameter set, only the model accounting for random alloy fluctuations produces trends that are consistent with the experimental data. The widely employed VCA yields results that are in contradiction with the experimental data, thus indicating that radiative recombination stems mainly from the well \emph{furthest away} from the $p$-side. Overall, this highlights (i) that alloy fluctuations are essential to achieve an accurate description of the carrier transport and (ii) have to be taken into account when theoretically guiding the design of energy efficient III-N light emitters.
The paper is organized as follows: in Section~\ref{sec:TheoryStructs} we outline the model structure used for calculations and briefly summarise some of the literature experimental data from Ref.~\cite{BGaPhysStatSol_2011}. The theoretical framework which we use is summarized in Section~\ref{sec:Theory}. Our results are discussed in Section~\ref{sec:Results}. Finally Section~\ref{sec:conclusion} presents our conclusions.
\section{Model MQW structures and literature experimental findings}\label{sec:TheoryStructs}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{1.pdf
\caption{Conduction and valence band edges (black) along with the quasi-Fermi energies for electrons and holes (grey) in an (In,Ga)N/GaN multi-quantum well system described in virtual crystal approximation. The band edge profile and the quasi Fermi levels are shown at a current density of $50\text{ A/cm}^2$. The leftmost (In,Ga)N quantum well contains 12.5\% indium while the other two (In,Ga)N wells (centre and right) contain 10\% indium.}
\label{fig:BEs_1DVCA}
\end{figure}
To investigate the carrier distribution in (In,Ga)N/GaN MQW systems we proceed similar to experimental studies in the literature~\cite{BGaPhysStatSol_2011,LiRyAPL2008} and target MQW systems where one of the wells in the MQW stack has a slightly higher In content compared to the remaining wells.
In our case, we study MQW systems with three (In,Ga)N/GaN wells. Here two are In$_{0.1}$Ga$_{0.9}$N (``shallow'') wells and one is an In$_{0.125}$Ga$_{0.875}$N (``deep'') QW. These QWs are 3 nm wide and separated by 5 nm GaN barriers. The band edge profile of such a system along the transport ($c$-) direction, using a VCA, is shown in Fig.~\ref{fig:BEs_1DVCA} at a current density of 50 A/cm$^2$.
\input{2
In the following, we investigate the carrier transport properties in two settings: (i) on an atomistic level accounting for random alloy fluctuations and (ii) in the frame of a VCA thus neglecting alloy fluctuations. In the latter VCA case (ii), at a given $z$-position (along the \emph{c}-direction), there is no variation in material properties within the growth plane ($c$-plane). This assumption is also made in the widely used 1-D transport simulations on (In,Ga)N MQWs.
\begin{figure*}[t]
\centering
\begin{tabular}{|c|c|c|}
\hline
\large \textbf{(a)} Valence band edge \normalsize & \large \textbf{(b)} Conduction band edge \normalsize & \large \textbf{(c)} Radiative recombination \normalsize \\
\hline
\includegraphics[width=0.3\textwidth]{3a.pdf} &
\includegraphics[width=0.3\textwidth]{3b.pdf}
\includegraphics[width=0.3\textwidth]{3c.pdf} \\%{Figures/recombslice.pdf} \\
\hline
\end{tabular}
\caption{Profile of (a) valence band edge energy, (b) conduction band edge energy, and (c) radiative recombination rate in the growth plane ($c$-plane) of an In$_{0.1}$Ga$_{0.9}$N quantum well; the current density is $50\text{ A/cm}^2$ in all depicted figures. The slice displayed is the through the center well. The data are shown in all cases on a linear scale.}
\label{fig:BEs_2D}
\end{figure*}
To study the carrier distribution in MQW systems using the simulation settings (i) and (ii), we follow again the experimental approach e.g. presented in Ref.~\cite{BGaPhysStatSol_2011} and
the deep QW is moved from the $n$-side (position 1 ($n$-side) in Fig.~\ref{fig:schematic_of_QWs}) to the $p$-side (position 3 ($p$-side) in Fig.~\ref{fig:schematic_of_QWs}). In the case of the random alloy structures, the same microscopic configuration is kept for each well and only the ordering is changed.
For each of these systems the ratio of radiative recombination from the shallow wells to the deep well is calculated using
\begin{equation}
\varrho = \frac{\mathcal{R}^{RAD}_{\Omega_S}}
{\mathcal{R}^{RAD}_{\Omega_D}}
\label{eq:RecombRatio}
\end{equation}
where
\begin{equation}
\mathcal{R}_{\Omega_i}^{RAD} = \int_{\Omega_i} R^{RAD}(\mathbf{r})dV\,\, ,
\end{equation}
is the total radiative recombination from the region $\Omega_i$.
Here, $\Omega_D$ is the region containing the deep QW, $\Omega_S$ is the region containing the shallow wells (as there are two shallow QWs this is the union of the two shallow QW regions). The radiative recombination rate at position $\mathbf{r}$, $R_{RAD}(\mathbf{r})$, is discussed in further detail in section~\ref{sec:TheoryDD}.
Since we are studying a system with three QWs, an even distribution of carriers across the MQWs would result in a ratio of $\varrho = 2$.
Previous experimental work on a similar system by Galler \textit{et. al.}~\cite{BGaPhysStatSol_2011} found that \emph{$\varrho$ was small} (i.e. emission is dominated by the deep QW) \emph{only when the deep well was closest to the $p$-doped side of the MQW system} (thus position 3 ($p$-side) in Fig.\ref{fig:schematic_of_QWs}). The authors conclude that holes are responsible for this behavior, and argue that they are mainly found in the $p$-side QW and not in wells further away from the $p$-side. As a consequence, the overall emission from the (In,Ga)N/GaN MQW system is dominated by the emission from this well closest to the $p$-doped region. In line with Ref.~\cite{BGaPhysStatSol_2011}, we calculate $\varrho$ at a current density of 50 A/cm$^2$, which allows us to compare the here predicted trends with the trends found in the experimental studies. The theoretical framework employed to gain insight into $\varrho$ is discussed in the following section.
\section{Theoretical framework}\label{sec:Theory}
\label{sec:theory}
In this section, we introduce the underlying (microscopic) theory of our multiscale simulations. We start in Section~\ref{sec:TheoryTB} with the electronic structure model, an atomistic tight-binding (TB) model, and discuss the drift-diffusion approach in Section~\ref{sec:TheoryDD}. Since all these ingredients have been discussed in detail in Refs.~\cite{MiOD2021_JAP,OdFaOQE2022}, we here give only a brief summary.
\subsection{Tight-binding energy landscape}\label{sec:TheoryTB}
In order to model the electronic structure of the above described (In,Ga)N MQW systems on an atomistic level, we employ a nearest-neighbour, $sp^3$ TB model~\cite{ScCa2015}. Here, local strain and polarization effects are included using a valence force field model and local polarization theory, respectively~\cite{CaSc2013local}.
To connect the TB model to a DD solver we proceed as follows. Firstly, the TB model is used to extract a potential energy landscape describing the MQW region of the device using an atomistic framework. To do so, at each atomic site in the three dimensional (3-D) supercell, a local TB Hamiltonian is constructed from the full TB Hamiltonian~\cite{ChODo2021}.
Subsequently, only the local TB Hamiltonian is diagonalized, yielding the conduction and valence band edge energy at each lattice site. These band edge energies include now already effects arising from alloy fluctuations and connected fluctuations in strain and built-in polarization field. The obtained 3-D confining energy landscape, after employing a Gaussian softening, forms the basis for our DD calculations.
In previous studies we have investigated and discussed in detail the influence of the Gaussian softening on transport calculations for electrons and holes~\cite{MiOD2021_JAP,OdFaOQE2022}. Here, we choose a Gaussian broadening on the order of the GaN lattice constant, $\sigma_{c,v} = a^\text{GaN} = 0.3189$ nm, in all calculations.
This value is large enough to average over a number of neighboring sites, while also small enough to retain fluctuations in the energy landscape.
To obtain an accurate description of carrier transport in (In,Ga)N-based LEDs, the DD equations, which will be discussed below, are often coupled with solving the Schr\"odinger equation to account for quantum corrections.
Such a Schr\"odinger-Poisson solver is widely available in commercial software packages. However, in these packages it is largely restricted to 1-D simulations, since the extension to a 3-D system is computationally basically unfeasible. Instead of solving the large eigenvalue problem connected to evaluating the Schr\"odinger equation, we have implemented quantum corrections via localization landscape theory (LLT)~\cite{FiPi2017}; this approach is numerically much more efficient and gives results similar to a full self-consistent Schr\"odinger-Poisson solver, as we will also discuss below.
From LLT we extract an effective confining potential for the conduction and valence band edge starting from the TB energy landscape.
An example of the resulting quantum corrected energy landscape is given in Fig.~\ref{fig:BEs_2D}~(a) and (b). Here, in-plane band edge profiles for a single atomic plane through an In$_{0.1}$Ga$_{0.9}$N QW, after LLT has been applied, are shown.
As Fig.~\ref{fig:BEs_2D} (a) reveals, the fluctuations in the valence band edge energy due to alloy fluctuations are of the order of 100 meV. In combination with the high effective hole mass, these fluctuations are large enough to give rise to strong carrier localization effects as seen in other studies already~\cite{WaGo2011,ScCa2015,DiPeJJAP2019}.
We therefore expect that, especially for holes, the inclusion of random alloy fluctuations in the simulation will impact the carrier distribution. Consequently recombination rates are also expected to be noticeably influenced.
The variation in the conduction band edge energy is significantly smaller (order of 30 meV), as can be seen in Figure~\ref{fig:BEs_2D} (b). Since the effective electron mass is much lower in comparison with the holes, electron wave functions are less strongly perturbed by alloy fluctuations.
The impact that these fluctuations in the band edge energies have on the radiative recombination is also seen in Fig.~\ref{fig:BEs_2D}~(c); the radiative recombination is calculated with \texttt{ddfermi} as will be described in section~\ref{sec:TheoryDD}. The correlation between the valence band edge maxima and regions of high radiative recombination can be clearly identified; similar spatial profiles can be seen for non-radiative (Auger) recombination (not shown).
In order to highlight the impact of random alloy fluctuations on carrier transport and the distribution of carriers across a MQW system, we compare our atomistic calculations with the outcome of a VCA. In the latter case a homogeneous effective crystal is constructed where material properties are chosen to be interpolated properties of the binaries InN and GaN within the QW region. Here a linear, composition weighted interpolation scheme is employed. A bandgap bowing of $-2.0$ eV is used, consistent with the underlying atomistic TB model~\cite{CaSc2013local}. The VCA description, without any Gaussian broadening, is similar to commercially available packages. However, and in contrast to commercial software packages, quantum corrections via LLT can also be taken into account in our VCA simulations, following the approach used for the random alloy case.
\subsection{Device simulation}\label{sec:TheoryDD}
Having outlined above the generation of the energy landscape of the active region, e.g. the (In,Ga)N/GaN MQW system, a full device mesh, including the $n$- and $p$-doped regions, needs to be constructed on which the DD equations are solved. To achieve this we proceed as follows and
divide the device mesh into two regions: an atomistic and a macroscopic one.
The atomistic region is used to describe the MQW region and has as many grid points as atoms in the system. These points contain information about the conduction and valence band edge energies calculated from TB, as discussed above.
In order to capture the effects of carrier localization in the calculations, the in-plane dimensions of our 3-D simulation cell should be larger than the localization length of the holes, given that electrons are less strongly affected by alloy fluctuations~\cite{TaCaRSC2016}. In our atomistic calculations we use a system with in-plane dimensions of $12.8 \times 11.0\text{ nm}^2$. This is large enough to see the effects of hole localization as the in-plane hole localization length for
In$_{0.1}$Ga$_{0.9}$N QWs is of the order of 1 nm~\cite{TaDa2020}. The in-plane dimensions can be seen in Fig.~\ref{fig:BEs_2D}~(a)~and~(b) where the in-plane valence and conduction band edges of an In$_{0.1}$Ga$_{0.9}$N QW are shown. In case of the VCA, given that there are no variations in material properties (band edge energies) within the growth plane ($c$-plane), a much smaller in-plane area is sufficient ($1.3 \times 1.1\text{ nm}^2$), which reduces the numerical effort.
LLT is solved on this (finite element) mesh using a finite element method (FEM).
The DD calculations are carried out employing a Voronoi finite volume method (FVM)~\cite{Farrell2017}. Therefore, the generated FEM mesh must be transferred to an appropriate FVM mesh.
Every point from the FEM mesh is included on the FVM mesh, as well as extra points required to produce a boundary-conforming Delaunay tetrahedral mesh; conduction and valence band data are then interpolated onto these additional nodes.
This mesh is then embedded within a macroscopic device mesh which contains information about the $p$- and $n$-doped regions. In our atomistic transport studies here, we focus on systems without an (Al,Ga)N electron blocking layer (EBL). In principle, an atomistic mesh resolution would be required for the EBL, given the alloy fluctuations in (Al,Ga)N. However, and as we will discuss below, an (Al,Ga)N EBL is of secondary importance for the questions targeted in the present study. Therefore, outside active MQW region of the system, pure GaN is assumed. Thus, the conduction and valence band edge values are position independent (except for changes due an applied bias). The absence of strongly fluctuating band edges in the macroscopic mesh region allows us to use a sparse mesh and scale the simulation to a full device.
The mesh is created using \texttt{TetGen}~\cite{Si15ACM} and the interpolation is handled via \texttt{WIAS-pdelib}~\cite{pdelib}. More details on the mesh generation can be found in Ref.~\cite{MiOD2021_JAP}.
Equipped with knowledge about the mesh generation, we turn now to the DD simulations. To do so we build on the van Roosbroeck system of equations~\cite{VanRoosbroeck1950}:
\begin{subequations}
\begin{equation}
-\nabla \cdot (\varepsilon_s(\mathbf{r})\nabla\psi(\mathbf{r})) = q(p(\mathbf{r})-n(\mathbf{r})+C(\mathbf{r}))\,\, ,
\label{eq:Poisson}
\end{equation}
\begin{equation}
\nabla\cdot \mathbf{j}_n(\mathbf{r}) = qR(\mathbf{r})\,\, ,
\label{eq:current_continuity_elec}
\end{equation}
\begin{equation}
\nabla\cdot \mathbf{j}_p(\mathbf{r}) = -qR(\mathbf{r})\,\, ,
\label{eq:current_continuity_hole}
\end{equation}
\begin{equation}
\mathbf{j}_n(\mathbf{r}) = -q \mu_n n(\mathbf{r}) \nabla\varphi_n(\mathbf{r})\,\, ,
\label{eq:current_elec}
\end{equation}
\begin{equation}
\mathbf{j}_p(\mathbf{r}) = -q \mu_p p(\mathbf{r}) \nabla\varphi_p(\mathbf{r})\,\, .
\label{eq:current_hole}
\end{equation}
\label{eq:VanRoosbroeck}
\end{subequations}
In the above equations, $q$ is the elementary charge, \mbox{$\varepsilon_s(\mathbf{r}) = \varepsilon_0\varepsilon_r (\mathbf{r})$} is the dielectric permittivity, $\psi(\mathbf{r})$ is the device electrostatic potential, $p(\mathbf{r})$ and $n(\mathbf{r})$ are the hole and electron densities, $C(\mathbf{r}) = N_D^+(\mathbf{r}) - N_A^+(\mathbf{r})$ is the net activated dopant density, $\mathbf{j}_n(\mathbf{r})$, $\mathbf{j}_p(\mathbf{r})$, $\varphi_n(\mathbf{r})$ and $\varphi_p(\mathbf{r})$ are the electron and hole current densities and the respective quasi-Fermi potentials. The total recombination rate is denoted by $R$.
The carrier densities are related to the band edge energies, $E^\text{dd}_c(\mathbf{r})$ and $E^\text{dd}_v(\mathbf{r})$, and quasi-Fermi potentials, $\varphi_n(\mathbf{r})$ and $\varphi_p(\mathbf{r})$, via the state equations~\cite{Farrell2017}
\begin{subequations}
\begin{equation}
n(\mathbf{r}) = N_c \mathcal{F}\Bigg( \frac{q(\psi(\mathbf{r})-\varphi_n(\mathbf{r}))-E_c^{dd}(\mathbf{r})}{k_BT}\Bigg)\, \, ,
\end{equation}
\begin{equation}
p(\mathbf{r}) = N_v \mathcal{F}\Bigg( \frac{E_v^{dd}(\mathbf{r}) - q(\psi(\mathbf{r})-\varphi_p(\mathbf{r}))}{k_BT} \Bigg)\,\, .
\end{equation}
\end{subequations}
Here, $N_c$ and $N_v$ are the effective density of states for the conduction and valence band, respectively. For the distribution function, $\mathcal{F}$, we use Fermi-Dirac statistics, and $k_B$ is the Boltzmann constant. A temperature of $T = 300$ K has been used in all calculations. It is to note that the valence band, $E_v^{dd}(\mathbf{r})$, and conduction band edge energy, $E_c^{dd}(\mathbf{r})$, can be described either by a VCA, a VCA including quantum corrections via LLT, or an atomistic random alloy calculation including LLT-based quantum corrections.
The total recombination rate in Eqs.~(\ref{eq:current_continuity_elec}) and~(\ref{eq:current_continuity_hole}), $R$, is calculated using the ABC model~\cite{KaOQE2015,PiPhysStatSol2010}. Here, $R$ is the sum of (defect related) Shockley-Read-Hall, $R^{SRH}$, radiative, $R^{RAD}$ and (non-radiative) Auger recombination rate, $R^{AUG}$.
The SRH rate is obtained from:
\begin{equation}
R^{SRH}(\mathbf{r}) = \frac{r(n,p)}{\tau_p\big(n(\mathbf{r}) + n_i(\mathbf{r}) \big) + \tau_n\big(p(\mathbf{r})+n_i(\mathbf{r}) \big)}\,\, ,
\label{eq:R_SRH}
\end{equation}
the radiative part via
\begin{equation}
R^{RAD}(\mathbf{r}) = B_0 r(n,p)\,\, ,
\label{eq:R_RAD}
\end{equation}
and the Auger rate is calculated as
\begin{equation}
R^{AUG}(\mathbf{r}) = ( C_n n(\mathbf{r}) + C_p p(\mathbf{r}) ) r(n,p)\,\, .
\label{eq:R_AUG}
\end{equation}
In Eqs.~(\ref{eq:R_SRH}) to~(\ref{eq:R_AUG})
$$r(n,p) = n(\mathbf{r})p(\mathbf{r}) - n_i^2(\mathbf{r})$$
and
$$
n_i^2(\mathbf{r}) = n(\mathbf{r})p(\mathbf{r})\exp\bigg(\frac{q\varphi_n - q\varphi_p}{k_BT}\bigg).
$$
The above equations require further input, namely the radiative recombination coefficient $B_0$, the Auger recombination coefficients $C_p$ and $C_n$ as well as the the SRH lifetimes $\tau_p$ and $\tau_n$. All these parameters will in principle carry a composition dependence~\cite{McKiPRB2022,McTa2020,KiQiNJOP2013}. Furthermore, $B_0$, $C_p$ and $C_n$ will also be carrier density dependent~\cite{DaYoPRA2019,DaNaECS2019,JoTeAPL2017}.
We follow here the widely made assumption that these coefficients are constant across the InGaN MQW region~\cite{LiPi2017,LyLhPhysRevApplied2022}. In the following we take a weighted average of parameters calculated in Ref.~\cite{McKiPRB2022} for an electron and hole density of $3.8\times 10^{18}$ cm$^{-3}$, which is a good approximation for the average carrier densities in the QWs at a current density of 50 A/cm$^2$.
As our active region consists of two In$_{0.1}$Ga$_{0.9}$N QWs and one In$_{0.125}$Ga$_{0.875}$N QW we evaluate the different recombination coefficients as follows:
\begin{equation}
R_i^{\text{eff}} = \frac{2\times(R_i^{10\%}) + 0.5\times(R_i^{15\%} + R_i^{10\%})}{3}\,\, .
\end{equation}
Here, $R_i \in \{B_0,C_n,C_p\}$ are the radiative recombination, electron-electron-hole and hole-hole-electron Auger recombination coefficients, respectively. As there are no values for an In$_{0.125}$Ga$_{0.875}$N QW in Ref.~\cite{McKiPRB2022}, a linear average of the coefficients in In$_{0.1}$Ga$_{0.9}$N and In$_{0.15}$Ga$_{0.85}$N wells has been used. A summary of the material parameters employed in all simulations is given in Table~\ref{tab:MatParams}.
\begin{table}[t]
\caption{Material parameters used in the different regions of the simulation supercell. Parameters denoted with $\dagger$ are taken from~\cite{LiPi2017}; parameters denoted with $\ddagger$ are derived from~\cite{McKiPRB2022} as described in the main text.}
\label{tab:MatParams}
\centering
\begin{tabular}{|lr|c||r|r|r|}
\hline
\multicolumn{3}{|c||}{Parameter} & \multicolumn{3}{c|}{Value in each region} \\
\hline
\multicolumn{1}{|c}{\textbf{Name}} & &
\multicolumn{1}{|c||}{\textbf{Units}} &
\multicolumn{1}{c}{\textbf{$p$-GaN}} &
\multicolumn{1}{|c}{\textbf{$i$-InGaN}} &
\multicolumn{1}{|c|}{\textbf{$n$-GaN}} \\
\hline
Doping &&cm$^{-3}$ & $5\times 10^{18}$ & $1\times 10^{16}$ & $5\times 10^{18}$ \\
$\mu_h$&$^\dag$ &cm$^2$/Vs & 5 & 10 & 23 \\
$\mu_e$&$^\dag$ &cm$^2$/Vs & 32 & 300 & 200 \\
$\tau_p$&$^\dag$ &s & $10$ & $1\times10^{-7}$ & $7\times10^{-10}$ \\
$\tau_n$&$^\dag$ &s & $6\times10^{-10}$ & $1\times10^{-7}$ & $10$ \\
$B_0$&$^\ddagger$ &cm$^3$/s& $2.8\times10^{-11}$ & $2.8\times10^{-11}$ & $2.8\times10^{-11}$ \\
$C_p$&$^\ddagger$ &cm$^6$/s & $5.7\times10^{-30}$ & $5.7\times10^{-30}$ & $5.7\times10^{-30}$ \\
$C_n$&$^\ddagger$ &cm$^6$/s & $1\times10^{-31}$ & $1\times10^{-31}$ & $1\times10^{-31}$ \\
\hline
\end{tabular}
\end{table}
The numerical approximation of the van Roosbroeck system is implemented (in 3-D) in \texttt{ddfermi}~\cite{ddfermi}. As already mentioned above, we employ the FVM and the current is discretized using the SEDAN (excess chemical potential) approach
\cite{Yu1988,Cances2020,Abdel2020b}, which yields a thermodynamically consistent flux approximation in the sense of Ref.~\cite{Farrell2017}.
To simulate the devices under study, we also used the commercial software \NN{~\cite{nextnano}}, which relies on the simulation of a self-consistent Schr\"odinger-Poisson-DD system.
In this work we use \NN{ }to simulate the carrier transport in the above discussed MQW systems within a 1-D approximation.
In \NN{ }we utilize the same parameter set as in the \texttt{ddfermi} simulations. Therefore, the obtained results can be directly compared to our 3-D VCA model. When including quantum corrections in \texttt{ddfermi}, LLT is used. In \NN{ } a self-consistent Schr\"odinger-Poisson-DD calculation is performed where a $\mathbf{k}\cdot\mathbf{p}$ Hamiltonian
is used to calculate eigenstates across the full simulation domain.
Following the \texttt{ddfermi} set up, in \NN{ }we employ also a 1-band model for the calculation of the electron and hole densities.
\section{Results}\label{sec:Results}
\begin{figure*}[t]
\centering
\begin{tabular}{|c|c|}
\hline
\large{\textbf{(a)} \NN{}} (1-D) & \large{\textbf{(b)} \texttt{ddfermi}} (3-D)\\
\hline
\text{ }\vspace{-5pt} & \text{ } \\
\includegraphics[width=0.3\textwidth]{4a.pdf}
\includegraphics[width=0.3\textwidth]{4b.pdf} \\%{Figures/latex_figures/Ratio_tex_DD.pdf}\\
\hline
\end{tabular}
\caption{Ratio of radiative recombination $\varrho$, Eq.~(\ref{eq:RecombRatio}), from the shallow wells (In$_{0.1}$Ga$_{0.9}$N) to recombination from the deep well (In$_{0.1}$Ga$_{0.9}$N) calculated as a function of the position of the deep well in the multi-quantum well stack. Here $\varrho$ is evaluated using (a) \NN{ }excluding (purple) and including (green) quantum corrections via a self-consistent Schr\"odinger-Poisson-drift diffusion solver; results are shown when excluding (solid, filled circles) and including (dotted, open circles) an Al$_{0.15}$Ga$_{0.85}$N blocking layer, and (b) \texttt{ddfermi} excluding (purple), including (green) quantum corrections via localization landscape theory (LLT) using a virtual crystal approximation (VCA) and a random alloy calculation including LLT-based quantum corrections (blue); these calculations neglect the AlGaN blocking layer.}
\label{fig:recombratio}
\end{figure*}
In this section we present the results of our study on the carrier distribution in the above described (In,Ga)N/GaN MQW systems. To understand the impact of the alloy microstructure on the carrier distribution, in Section~\ref{sec:ResultsEBL} we start with standard 1-D calculations building on the commercial software package \NN{~\cite{nextnano}}. We use this entirely continuum-based description of the QWs also to determine the impact (i) of an EBL and (ii) a self-consistent Schrodinger-Poisson-DD treatment on the transport properties.
Moreover, and as already mentioned above, (ii) can also be used as a benchmark for our 3-D \texttt{ddfermi} solver. In Section~\ref{sec:ResultsFluctuations} we then proceed to investigate the influence of random alloy fluctuations on the carrier distribution in the (In,Ga)N/GaN MQW stack.
\subsection{Continuum-based simulations of the carrier transport in (In,Ga)N-based LEDs}\label{sec:ResultsEBL}
\begin{figure*}
\centering
\begin{tabular}{|c|c|c|}
\hline
\large \textbf{(a) VCA} \normalsize & \large \textbf{(b) VCA + LLT} \normalsize & \large \textbf{(c) Random alloy + LLT} \normalsize \\
\hline
\text{ } \vspace{-5pt} & & \\
\includegraphics[width=0.3\textwidth]{5a.pdf}
\includegraphics[width=0.3\textwidth]{5b.pdf}
\includegraphics[width=0.3\textwidth]{5c.pdf}\\%{Figures/latex_figures/histogram_RA.pdf} \\
\hline
\end{tabular}
\caption{
Contribution of each quantum well ($n$-side; centre; $p$-side) in the (In,Ga)N multi-quantum well system to the total radiative recombination $\mathcal{R}^{RAD}_{\Omega_i}$ for $i\in \{n\text{-side},\text{center},p\text{-side}\}$ as a percentage of the total radiative recombination from all 3 quantum wells for (a) virtual crystal approximation (VCA), (b) virtual crystal approximation with quantum corrections included via localization landscape theory (VCA + LLT) and (c) a random alloy calculation including localization landscape theory based quantum corrections (Random alloy + LLT). That data are shown as a function of the position of the deep quantum well ($x$-axis). Each bar contains the percentage recombination from the $n$-side quantum well (purple), the center quantum well (green) and the $p$-side quantum well (blue). Labelling is consistent with that introduced in Fig.~\ref{fig:schematic_of_QWs}.
}
\label{fig:percentagecontribution}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{tabular}{|c||c||c|}
\hline
\large \textbf{(a)} $n$-side \normalsize & \large \textbf{(b)} center \normalsize & \large \textbf{(c)} $p$-side \normalsize \\
\hline
\hline
\multicolumn{3}{|c|}{\large \textbf{(i) VCA}} \\
\hline
\includegraphics[width=0.3\textwidth]{6ia.pdf}
\includegraphics[width=0.3\textwidth]{6ib.pdf}
\includegraphics[width=0.3\textwidth]{6ic.pdf} \\%{Figures/rhoRecomb_VCAQW3_log.pdf} \\
\hline
\hline
\multicolumn{3}{|c|}{\textbf{\large (ii) VCA + LLT}}\\
\hline
\includegraphics[width=0.3\textwidth]{6iia.pdf}
\includegraphics[width=0.3\textwidth]{6iib.pdf}
\includegraphics[width=0.3\textwidth]{6iic.pdf} \\%{Figures/rhoRecomb_VCALLTQW3_log.pdf} \\
\hline
\hline
\multicolumn{3}{|c|}{\large \textbf{(iii) Random alloy + LLT}} \\
\hline
\includegraphics[width=0.3\textwidth]{6iiia.pdf}
\includegraphics[width=0.3\textwidth]{6iiib.pdf}
\includegraphics[width=0.3\textwidth]{6iiic.pdf} \\%{Figures/rhoRecomb_RAQW3_log.pdf} \\
\hline
\end{tabular}
\caption{Hole density (black, solid), electron density (black, dashed), and radiative recombination rate (red, solid) averaged over each atomic plane along the transport direction. Results from calculations building on (i) a virtual crystal approximation (top), (ii) a virtual crystal including quantum corrections via localization landscape theory (LLT) (center) and a (iii) random alloy description including LLT-based quantum corrections (bottom); the deep well is located at (a) the $n$-side (left), (b) the center (middle) and (c) the $p$-side (right). The data are shown on a log scale.}
\label{fig:outofplaneDensities_log}
\end{figure*}
To examine the impact of random alloy fluctuations on the carrier distribution in an (In,Ga)N/GaN MQW stack, we start with a `standard' 1-D simulation approach that is widely applied in the literature.
In a first step we begin with \NN{ } calculations and as outlined above, compare the results to our \texttt{ddfermi} data.
\subsubsection{\NN{} simulations}\label{subsubsec:NNresults}
To study how the presence of an EBL affects the ratio of radiative recombination $\varrho$, Eq.~(\ref{eq:RecombRatio}), the systems outlined in Section~\ref{sec:TheoryStructs} are simulated with and without a 20 nm Al$_{0.15}$Ga$_{0.85}$N EBL using \NN{}. The EBL is separated from the $p$-side QW (position 3 ($p$-side) in Fig.~\ref{fig:schematic_of_QWs}) by a 10 nm GaN barrier. Similar settings for an (Al,Ga)N EBL have been used in previous studies~\cite{LiPi2017}.
The \NN{ } calculated ratio of radiative recombination $\varrho$,
when varying the position of the deep QW in the MQW stack, are depicted in Fig.~\ref{fig:recombratio}~(a). Turning first to the data without quantum corrections, we find that in the case of the employed 1-D VCA-like continuum-based description, \emph{$\varrho$ is small when the deep QW is at the $n$-side (position 1 ($n$-side) Fig.~\ref{fig:schematic_of_QWs}) and larger when the deep well is at the p-side (position 3 ($p$-side) Fig.~\ref{fig:schematic_of_QWs})}. Thus, the 1-D model \emph{predicts the opposite trend} when compared to experiment~\cite{BGaPhysStatSol_2011}.
This trend is only slightly changed when including quantum corrections via a self-consistent Schr\"odinger-Poisson-DD model. In this case a much weaker dependence of the results on the position of the deep QW in the MQW stack is observed. However, even when including quantum corrections, the \NN{} results for $\varrho$ are not reflecting the experimentally observed behavior (see discussion above). Figure~\ref{fig:recombratio}~(a) reveals also that qualitatively the results do not depend on the presence of the EBL, indicating that for the structures considered, this feature of an LED is of secondary importance for the aims of this work.
\subsubsection{\texttt{ddfermi} simulations}
Since we are also able to use the atomistic framework in a VCA setting, we compare our \texttt{ddfermi} results, cf. Fig.~\ref{fig:recombratio}~(b) (purple), with those from \NN{,} cf. Fig.~\ref{fig:recombratio}~(a) (purple, solid). We focus on structures which neglect the EBL as we have found above that it does not impact results in a VCA. In both \NN{ }and \texttt{ddfermi} a similar trend is found: the deep QW dominates recombination only when it is located at the $n$-side. This is illustrated further in Fig.~\ref{fig:percentagecontribution}~(a), which displays the contribution (in percent) to the radiative recombination rate from each QW (colors) in the MQW stack. The data are shown as a function of position of the deep QW in the MQW system.
This confirms that it is always the QW which is closest to the $n$-doped side (position 1) that dominates the recombination process;
the $n$-side QW contributes $\approx 95\%$ when the deep QW is at position 1, $\approx 70\%$ when the deep QW is at position 2 or 3. Again, we stress that this is the \emph{opposite} trend to the experimental findings in Ref.~\cite{BGaPhysStatSol_2011}.
To shed more light on this result, the upper row in Figure~\ref{fig:outofplaneDensities_log} depicts the average hole (black, solid), electron (black, dashed) and radiative recombination (red) rate along the \textit{c}-axis when the deep QW (In$_{0.125}$Ga$_{0.875}$N well) is (a) closest to the $n$-side (position 1), (b) in the centre of the MQW stack (position 2) and (c) closest to the $p$-side (position 3). Focusing on the VCA data, Figs.~\ref{fig:outofplaneDensities_log}~(i)~(a-c), we see the cause of the dominant recombination from the $n$-side QW: the hole density is always high in this region, independent of which well is closest to the $n$-side. In particular, the $p$-side QW fails to capture holes effectively and consistently has the lowest hole density.
We note that a similar behavior is also found in the \NN{} calculations discussed in Sec.~\ref{subsubsec:NNresults}
Given that our VCA \texttt{ddfermi} approach and \NN{ } treat (In,Ga)N as a homogeneous alloy that can be described by averaged material parameters which do not vary throughout the wells (no alloy fluctuations included), it allows us also to compare the implemented methods for quantum corrections in DD simulations. Here, as discussed above, \NN{} builds on the widely used Schr\"odinger-Poisson-DD model while \texttt{ddfermi} utilizes the recently developed LLT method. It has been discussed and shown in the literature that the LLT method can produce results in good agreement with the solution of the Schr\"odinger equation in the case of a 1-D effective mass approximation~\cite{FiPi2017,ChKe2020}.
Looking at Fig.~\ref{fig:recombratio}~(a) (green, solid) and Fig.~\ref{fig:recombratio}~(b) (green) we see that the results from our in-house developed \texttt{ddfermi}-based 3-D model, which employs LLT (3-D, \texttt{ddfermi}), are very similar to the standard self-consistent 1-D Schr\"odinger-Poisson-DD calculation underlying \NN{}. This gives confidence that our LLT treatment is providing a comparable description of the quantum corrections in the system.
Overall, Fig.~\ref{fig:recombratio}~(a) reveals that when including quantum corrections in the VCA calculations, the position of the deep QW has little impact on the ratio of the relative radiative recombination, $\varrho$.
From Fig.~\ref{fig:percentagecontribution}~(b) one can also gain more insight into this behavior and how quantum corrections impact the carrier distribution in the MQW stack. In the \emph{absence} of quantum corrections but utilizing a VCA, Fig.~\ref{fig:percentagecontribution}~(a), the well closest to the $n$-side dominates the relative radiative recombination ratio $\varrho$ independent of the position of the deep well in MQW systems. When \emph{including} quantum corrections this situation is now changed: the deep QW is now the dominant emitter independent of its position in the MQW stack.
This behavior becomes clear when analyzing the electron and hole densities as a function of the position of the deep well in the (In,Ga)N/GaN MQWs, as shown in Fig.~\ref{fig:outofplaneDensities_log}~(ii). Looking at the electron densities first, we find that electrons preferentially occupy the well closest to the $p$-side. This effect is enhanced when the deep QW is closest to the $p$-side (cf. Fig.~\ref{fig:outofplaneDensities_log}~(ii)~(c)).
In our previous study on uni-polar electron transport~\cite{MiOD2021_JAP}, we have already seen that including quantum corrections leads to a softening of the potential barrier at the QW barrier interfaces. This in turn can lead to an increased electron current at a fixed bias point, when compared to a VCA system without LLT treatment, and thus the electrons can more easily `overshoot' the wells in the MQW system. As a consequence, a lower electron density in the well closest to the $n$-side is observed. Turning to the hole density, the situation is different. Here, we find that holes preferentially populate the well closest to the $n$-side. Only when the deep QW is closest to the $p$-side, the hole density in this well is noticeably increased.
However, when comparing the distribution of holes in the MQW as a function of the position of deep well in absence (Fig.~\ref{fig:outofplaneDensities_log}~(i)) and presence (Fig.~\ref{fig:outofplaneDensities_log}~(ii)) of quantum corrections, the results are not very different. This indicates that quantum corrections, at least when employing a VCA, are of secondary importance for the hole distribution. This finding is consistent with our previous results on uni-polar hole transport~\cite{OdFaOQE2022}, where we have discussed that due to the high effective hole mass and the small valence band offset, quantum corrections have a smaller impact on the hole transport when compared to electrons. As a consequence, the distribution of holes follows a similar pattern to that of the VCA where quantum corrections are neglected.
Finally, when looking at the ratio of radiative recombination $\varrho$, it is important to note that this quantity is not only determined by having both large electron and hole densities in the same well but also by their spatial overlap. As one can infer from Fig.~\ref{fig:outofplaneDensities_log}~(ii)~(a-c), the largest radiative recombination rate is always observed in the deepest well. This indicates also that the spatial overlap of electron and hole densities largest in the deep QW regardless of its position across the MQW system.
We stress again that even when including quantum corrections in the VCA calculations, the resulting trend in $\varrho$ is not reflecting the trend observed in experimental studies~\cite{BGaPhysStatSol_2011}.
\subsection{Impact of a random alloy fluctuations on the carrier transport in (In,Ga)N/GaN MQWs} \label{sec:ResultsFluctuations}
In the last step, we move away from the VCA description of the system and include, in addition to quantum corrections, also random alloy fluctuations in the calculations. Figure~\ref{fig:recombratio}~(b) (blue) shows that, and this time in line with the experimental results by Galler \emph{et. al}~\cite{BGaPhysStatSol_2011}, the deep QW only contributes significantly to the radiative recombination when it is \emph{closest to the $p$-side} (position 3). In fact, when including random alloy fluctuations in the calculations, the well closest to the $p$-side always has the largest contribution to total radiative recombination, as can be seen in Fig.~\ref{fig:percentagecontribution}~(c).
To understand this behavior, Fig.~\ref{fig:outofplaneDensities_log}~(iii) depicts the electron and hole densities in the different wells as a function of the position of the deep well in the MQW systems. Looking at the electron density first, in comparison to the VCA calculations both including and excluding quantum corrections, random alloy fluctuations lead a reduction in electron density at the $n$-side QW.
As discussed above and previously, quantum corrections can lead to increased electron transport, and including alloy fluctuations adds further to this effect due to the softening of the barrier at the well interfaces~\cite{MiOD2021_JAP}.
As a consequence, the electrons can more easily `overshoot' the wells in the MQWs, which can also be seen in the increased electron density beyond the $p$-side QW when alloy fluctuations and quantum corrections are included.
However, in comparison to the VCA result including quantum corrections, the electron density in the $p$-side well is only slightly affected by alloy fluctuations.
In contrast, hole densities in the $p$-side QW are more dramatically changed by alloy fluctuations. As Figs.~\ref{fig:outofplaneDensities_log}~(ii) and~\ref{fig:outofplaneDensities_log}~(iii) show, in comparison to the VCA description, alloy fluctuations lead to an increase in the hole density in the \emph{$p$-side QW} (position 3) even when the deep QW is closest to the $n$-side (position 1) or in the centre (position 2) of the MQW system. While the smoothing of the well barrier interface can increase hole transport, as in the case of electrons, there are now also alloy disorder induced localization effects to contend with. As discussed in our previous work, these localization effects are \emph{detrimental} to hole transport~\cite{OdFaOQE2022} and result in an increased hole density in the $p$-side QW. As a consequence, the well closest to the $p$-side dominates radiative recombination
We note that there is still a reasonable hole density present in the $n$-side QW (Fig.~\ref{fig:outofplaneDensities_log}~(iii)~(a-c)). In general, the distribution of carriers will also depend on the GaN barrier width and a 5 nm barrier is narrow enough to allow for some hole transport across the MQW~\cite{LiRyAPL2008}; a similar dependence of hole transmission on the barrier width has been seen in previous non-equilibrium Green's function studies~\cite{ODoLu2021}. Thus we expect that increasing the barrier width will mainly lead to a reduction of the hole density in the well furthest away from the $p$-side, but should to a lesser extent affect the hole density in the well closest to the $p$-side. Therefore, even for a larger barrier width than the here considered 5 nm, we expect that the recombination will still be dominated by the $p$-side QW.
We note that based on the VCA results we did not consider the EBL in the atomistic calculations. In general the EBL needs to be treated with an atomistic resolution. Previous studies of (Al,Ga)N barriers in uni-polar device settings have found that the impact of these barriers is lower than what is expected from a 1-D simulation for both electrons~\cite{BrMiSST2017} and holes~\cite{QwMoAPL2020}. Thus given that our VCA calculations show that the presence of the EBL is of secondary importance for our study, we expect a similar conclusion when treating the EBL with alloy fluctuations. Therefore, it is unlikely that the EBL impacts the here presented result, however this question may be targeted in future studies.
\section{Conclusions}\label{sec:conclusion}
In this work we apply a 3-D quantum corrected multiscale simulation framework to gain insight into the impact of random alloy fluctuations on the electron and hole distribution across the active region of an (In,Ga)N/GaN LED.
To study the spatial distribution of carriers we have followed literature experimental studies~\cite{BGaPhysStatSol_2011} and analyzed the radiative recombination ratio in a multi-quantum well system, where one of the wells in the system has a higher indium content (deeper well) and its position is varied within the stack.
The here considered MQW systems are not only of interest for a comparison with experiment, they provide also the ideal opportunity to benchmark and validate results from our in-house developed 3-D multiscale simulation framework against commercially available software packages. To do so we treat the QWs in a virtual crystal approximation (VCA), to mimic the 1-D simulation widely used in the literature for (In,Ga)N QWs and implemented in the commercial software package \NN{}. In addition, this study allows us also to compare the different schemes to account for quantum corrections (localization landscape theory vs. Schr\"odinger-Poisson-DD simulations) in the simulations. Overall, this analysis showed very good agreement between results obtained from our in-house software and \NN{} (without random alloy fluctuation).
Equipped with this benchmarked model, our analysis reveals that including (random) alloy fluctuations in the calculations is vital for reproducing trends seen in experiment. More specifically, when using the widely employed virtual crystal approximation (VCA), the hole density in the well closest to the $p$-doped region of the device is significantly reduced compared to our atomistic random alloy calculation. As a consequence, and in contrast to the experiment, in VCA the well closest to the $p$-side contributes very little to the radiative recombination process, an effect that can be reduced by accounting for quantum corrections.
While this leads to enhanced radiative recombination from the well closest to the $p$-side, at least when this well is the deep well, it still does not reflect the trends observed in the experimental studies. However, when including random alloy fluctuations and quantum corrections in our 3-D simulation framework, these effects lead to an increase in the hole density in the well closest to the $p$-side. Consequently, this well dominates the radiative recombination process in line with the experimental data. We note that in addition to quantum corrections and alloy fluctuations no further ingredients are required (e.g. multi-population model) to explain the experimentally observed trends.
Therefore, our calculations highlight that alloy fluctuations are a key ingredient in simulations guiding the design of III-N based devices. Thus, the here developed model presents an ideal starting point for future calculations.
\begin{acknowledgements}
This work received funding from the Sustainable Energy Authority of Ireland and the Science Foundation Ireland (Nos. 17/CDA/4789, 12/RC/2276 P2 and 21/FFP-A/9014) and the Deutsche Forschungsgemeinschaft (DFG) under Germany’s Excellence Strategy EXC2046: MATH+, project AA2-15, the Leibniz competition 2020 as well as the Labex CEMPI (ANR-11-LABX-0007-01).
\end{acknowledgements}
|
\section{Introduction}
Quantum theory is described well via Feynman path integral.
In Feynman path integral, the action in the integrand is considered to be path-dependent,
while the measure is usually supposed to be path-independent.
However, we could consider a theory such that not only the action but also the measure is path-dependent.
This is the complex action theory (CAT)
whose action is complex at a fundamental level but expected to look real effectively\cite{Bled2006}.
The CAT provides us with
falsifiable predictions\cite{Bled2006,Nielsen:2008cm,Nielsen:2007ak,Nielsen:2005ub}.
Deeper understanding via the CAT have been tried even for the Higgs mass\cite{Nielsen:2007mj},
quantum mechanical philosophy\cite{newer1,Vaxjo2009,newer2},
some fine-tuning problems\cite{Nielsen2010qq,degenerate},
black holes\cite{Nielsen2009hq},
de Broglie-Bohm particles and a cut-off in loop diagrams\cite{Bled2010B},
a mechanism to obtain Hermitian Hamiltonians\cite{Nagao:2010xu},
the complex coordinate formalism\cite{Nagao:2011za},
and the momentum relation\cite{Nagao:2011is,Nagao:2013eda}.
There are two types for the CAT.
One is the future-not-included theory, where
only the past state $| A(T_A) \rangle$ at the initial time $T_A$ is given
and the time integration is performed over the period from $T_A$ to
a reference time $t$.
The other is the future-included one, where not only the past state $| A(T_A) \rangle$
but also the future state $| B(T_B) \rangle$ at the final time $T_B$ is given,
and the time integration is performed over the whole period from $T_A$ to $T_B$.
We elucidated various interesting properties of the future-not-included CAT\cite{Nagao:2013eda}.
In Ref.\cite{Nagao:2017ecx} we argued that, if a theory is described
with a complex action, then such a theory is suggested to be the future-included theory,
rather than the future-not-included one, as long as we respect objectivity.
Even so, the future-not-included CAT itself still remains to be a fascinating theory,
and a good playground to study various intriguing aspects of the CAT.
In the future-included theory,
the normalized matrix element
$\langle \hat{\mathcal O} \rangle^{BA}
\equiv \frac{ \langle B(t) | \hat{\mathcal O} | A(t) \rangle }{ \langle B(t) | A(t) \rangle }$
of an operator $\hat{\mathcal O}$ is expected to have a role of
an expectation value\cite{Bled2006}.\footnote{The normalized matrix element $\langle \hat{\mathcal O} \rangle^{BA}$ is called
the weak value\cite{AAV,generalized_two-state_vector_formalism} and has been studied intensively with a different philosophy
in the real action theory (RAT). For details, see Ref.\cite{review_wv} and references therein.}
Indeed, if we regard it so,
we can obtain nice properties such as the Heisenberg equation, Ehrenfest's theorem,
and a conserved probability current density\cite{Nagao:2012mj,Nagao:2012ye}.
However, $\langle \hat{\mathcal O} \rangle^{BA}$ is generically complex even
for Hermitian $\hat{\mathcal O}$, even though
any observables are real.
To resolve this problem, in Refs.\cite{Nagao:2015bya,Nagao:2017cpl},
we proposed a theorem that states that,
provided that an operator $\hat{\mathcal O}$ is $Q$-Hermitian, i.e.,
Hermitian with regard to a modified inner product $I_Q$ that makes a given non-normal
Hamiltonian\footnote{The Hamiltonian $\hat{H}$ is
generically non-normal, so it is
not restricted to the class of PT-symmetric non-Hermitian Hamiltonians that were
studied in
Refs.\cite{Bender:1998ke,Bender:1998gh,Mostafazadeh_CPT_ip_2002,Mostafazadeh_CPT_ip_2003,Bender:2011ke}.} normal
by using an appropriately chosen Hermitian operator $Q$,
the normalized matrix element defined with $I_Q$ becomes real and
time-develops under a $Q$-Hermitian Hamiltonian for the past and future states selected
such that the absolute value of the transition amplitude defined with $I_Q$ from the past state
to the future state is maximized.
We call this way of thinking the maximization principle.
We proved the theorem in the case of non-normal
Hamiltonians $\hat{H}$\cite{Nagao:2015bya}\footnote{The proof is based on the existence of imaginary parts of the eigenvalues of $\hat{H}$, so it cannot be applicable to the RAT case.
The maximization principle is reviewed in Refs.\cite{Nagao:2017book,Nagao:2017ztx}.} and
in the RAT\cite{Nagao:2017cpl}.
In addition, imposing a periodic condition, we studied the periodic CAT and
proposed a variant type of the maximization principle,
by which the period could be determined\cite{Nagao:2022rap}.
Maximization principle is based on the natural way of thinking and looks promising.
Behind the principle, the automatic hermiticity mechanism\cite{Nagao:2010xu} has a key role.
In the CAT the imaginary parts of the eigenvalues $\lambda_i$
of a given non-normal Hamiltonian $\hat{H}$ are supposed to be bounded from above for
the Feynman path integral $\int e^{\frac{i}{\hbar}S} {\mathcal D} \text{path}$ to converge.
Then we can imagine that some $\text{Im} \lambda_i$ take the maximal value $B$, and
denote the corresponding subset of $\{ i \}$ as $A$.
After a long time development, only the subset $A$ contributes most significantly,
and on the subset a $Q$-Hermitian Hamiltonian effectively emerges .
This is the automatic hermiticity mechanism that we proposed and studied explicitly
for pure states time-developing forward\cite{Nagao:2010xu}.
In Ref.\cite{Nagao:2012mj}, utilizing it for pure states time-developing forward and backward,
we showed that the normalized matrix element of ${\cal O}$
at the present time $t$
in the future-included theory for large $T_B-t$ and large $t- T_A$ corresponds to
the expectation value of ${\cal O}$ in the future-not-included theory
defined with a modified inner product $I_{Q'}$ for large $t- T_A$.
This study strongly suggests that the future-included CAT is not excluded phenomenologically,
even though it looks very exotic.
Thus the automatic hermiticity mechanism has an essential role for the CAT to be viable,
but so far we have studied it only for pure statues, not for mixed states.
Then it would be natural to have a question: how does it work for mixed states?
Even though mixed states can always be expressed by pure states
defined in a larger system that includes the mixed states in its subsystem,
it is interesting and worthwhile to study how mixed states are defined and how they
behave in the CAT.
Especially, it is intriguing to study the automatic hermiticity mechanism for mixed states
in the CAT, because the emergence of a Hermitian Hamiltonian is crucially important for the CAT
to be sensible, and also because mixed states are generic quantum states along with pure states.
We need to introduce density matrices to describe mixed states in the CAT.
In the future-not-included CAT, there is only
one class of state vectors
time-developing forward from the past,
while in the future-included CAT there are two classes of state vectors
time-developing not only forward from the past but also backward from the future.
Hence it would be much more non-trivial to define density matrices
and see the emergence of hermiticity for them in the future-included CAT
rather than the future-not-included one.
Therefore, in this manuscript, after reviewing the modified inner product $I_Q$ and automatic hermiticity mechanism for pure states,
we first define density matrices to describe mixed states
and study the emergence of hermiticity for them in the future-not-included CAT.
Next, we investigate
a couple of candidates for density matrices in the future-included CAT,
and introduce a ``skew density matrix'' composed of both ensembles of the future and past states
such that the trace of the product of it and an operator ${\cal O}$ becomes
a normalized matrix element of ${\cal O}$.
Furthermore, we argue that the skew density matrix defined with $I_Q$ at the present time $t$
for large $T_B-t$ and large $t-T_A$ approximately corresponds to
another type of density matrix composed of only an ensemble of the past state
and defined with another inner product $I_{Q_J}$ for large $t-T_A$.
Finally we summarize the study in this manuscript and discuss the outlook of our theory.
\section{Modified inner product and the automatic hermiticity mechanism for pure states}
In this section we briefly review the modified inner product $I_Q$ and
the automatic hermiticity mechanism
for pure states by following refs.\cite{Nagao:2010xu, Nagao:2011za}.
The eigenstates of a given non-normal Hamiltonian $\hat{H}$,
$| \lambda_i \rangle (i=1,2,\dots)$
obeying $\hat{H} | \lambda_i \rangle = \lambda_i | \lambda_i \rangle$, are not orthogonal to
each other in the usual inner product $I$.
Let us introduce a modified inner product
$I_Q$\cite{Nagao:2010xu,Nagao:2011za}\footnote{Similar inner products are also studied
in Refs.\cite{Geyer,Mostafazadeh_CPT_ip_2002,Mostafazadeh_CPT_ip_2003}.} such that
$| \lambda_i \rangle (i=1,2,\dots)$ become orthogonal to each other
with regard to it, i.e.,
for arbitrary kets $|u \rangle$ and $|v \rangle$, $I_Q(|u \rangle , |v \rangle) \equiv \langle u |_Q v \rangle
\equiv \langle u | Q | v \rangle$,
where $Q$ is a Hermitian operator that obeys $\langle \lambda_i |_Q \lambda_j \rangle = \delta_{ij}$.
Using the diagonalizing operator $P=(| \lambda_1 \rangle , | \lambda_2 \rangle , \ldots)$
such that $P^{-1}\hat{H}P=D=\text{diag}(\lambda_1, \lambda_2, \dots)$,
we choose $Q=(P^\dag)^{-1} P^{-1}$.
Next we define the $Q$-Hermitian conjugate $\dag^Q$ of
an operator $A$ by $\langle \psi_2 |_Q A | \psi_1 \rangle^* \equiv \langle \psi_1 |_Q A^{\dag^Q} | \psi_2 \rangle$, so $A^{\dag^Q} \equiv Q^{-1} A^\dag Q$.
If $A^{\dag^Q} = A$, we call $A$ $Q$-Hermitian.
We also introduce $\dag^Q$ for kets and bras as
$| \lambda \rangle^{\dag^Q} \equiv \langle \lambda |_Q $ and
$\left(\langle \lambda |_Q \right)^{\dag^Q} \equiv | \lambda \rangle$.
Then, since $P^{-1}=P^{\dag^Q}$, $\hat{H}$ is $Q$-normal, $[\hat{H}, \hat{H}^{\dag^Q} ] =0$.
We can decompose $\hat{H}$ as
$\hat{H}=\hat{H}_{Qh} + \hat{H}_{Qa}$,
where $\hat{H}_{Qh}= \frac{\hat{H} + \hat{H}^{\dag^Q} }{2}$ and
$\hat{H}_{Qa} = \frac{\hat{H} - \hat{H}^{\dag^Q} }{2}$ are
$Q$-Hermitian and anti-$Q$-Hermitian parts of $\hat{H}$, respectively.
Let us consider a state\footnote{We consider a state with an index $i$ just for our later convenience. }
$|A_i (t) \rangle$, which obeys the Schr\"{o}dinger equation
\begin{equation}
i \hbar \frac{d}{dt} | A_i (t) \rangle = \hat{H} | A_i (t) \rangle. \label{Schr_for_Ai}
\end{equation}
We introduce a normalized state and an expectation value of an operator ${\cal O}$ by
$| A_i (t) \rangle_{N}
\equiv \frac{1}{\sqrt{ \langle A_i (t) |_Q ~{A_i}(t) \rangle} } | {A_i}(t) \rangle$,
$\langle \hat{\cal O} \rangle_Q^{A_i A_i} (t)
\equiv {}_{N} \langle A_i (t) |_Q {\cal O} | A_i (t) \rangle_{N}$.
They obey
$i\hbar \frac{\partial}{ \partial t} | A_i (t) \rangle_{N}
= \hat{H}_{Qh} | A_i (t) \rangle_{N}
+ \hat{\Delta} \left( \hat{H}_{Qa} ; | A_i (t) \rangle_{N} \right)
| A_i (t) \rangle_{N}$,
$\frac{d}{dt}\langle \hat{\cal O} \rangle_Q^{A_i A_i} (t)
=
-\frac{i}{\hbar} \langle \left[ \hat{\cal O} , \hat{H}_{Qh} \right] \rangle_Q^{A_i A_i} (t)
-\frac{i}{\hbar} \langle \left\{ \hat{\cal O}, \hat{\Delta} \left( \hat{H}_{Qa} ; | A_i (t) \rangle_{N} \right) \right\} \rangle_Q^{A_i A_i} (t)$,
where we have introduced
\begin{equation}
\hat{\Delta} \left( \hat{H}_{Qa} ; | A_i (t) \rangle_{N} \right)
\equiv \hat{H}_{Qa} -{}_{N} \langle A_i (t) |_Q \hat{H}_{Qa} | A_i (t) \rangle_{N} . \label{DeltahatA_i}
\end{equation}
It seems that, in the classical limit,
since $\hat{\Delta} \left( \hat{H}_{Qa} ; | A_i (t) \rangle_{N} \right)$ is suppressed,
the system time-develops by a $Q$-Hermitian Hamiltonian, and Ehrenfest's theorem holds.
This property is intriguing, but we will see the emergence of the $Q$-hermiticity
even before considering the classical limit via the automatic hermiticity mechanism, which we explain below.
Expanding $| A_i (t) \rangle$ as
$| A_i (t) \rangle = \sum_j a_j^{(i)} (t) | \lambda_j \rangle$
and introducing $| A'_i (t) \rangle = P^{-1} | A_i (t) \rangle= \sum_j a_j^{(i)} (t) | e_j \rangle$,
which obeys $i \hbar \frac{d}{dt} | A'_i(t) \rangle =D | A'_i(t) \rangle$,
we obtain
$| A_i(t) \rangle = P e^{- \frac{i}{\hbar} D (t-t_0)} | A'_i(t_0) \rangle
= \sum_j a_j^{(i)} (t_0)
e^{ \frac{1}{\hbar} \left( \text{Im} \lambda_j - i \text{Re} \lambda_j \right) (t-t_0)}
| \lambda_j \rangle$.
Now we assume that the anti-$Q$-Hermitian part of $\hat{H}$ is bounded from above
for the Feynman path integral $\int e^{\frac{i}{\hbar}S} {\cal D}\text{path}$ to converge.
Based on this assumption
we can imagine that some $\text{Im} \lambda_j$
take the maximal value $B$, and denote the corresponding subset of $\{ j \}$ as $A$.
After a long time has passed, i.e., for large $t-t_0$,
the states with $\text{Im} \lambda_j |_{j \in A}$
contribute most in the sum.
Let us define a diagonalized Hamiltonian $\tilde{D}_{R}$ by
\begin{equation}
\langle e_j | \tilde{D}_{R} | e_k \rangle \equiv
\left\{
\begin{array}{cc}
\langle e_j | D_R | e_k \rangle =\delta_{jk} \text{Re} \lambda_j & \text{for} \quad j \in A , \\
0 &\text{for} \quad j \not\in A , \\
\end{array}
\right. \label{DRtilder}
\end{equation}
and introduce $\hat{H}_{\text{eff}} \equiv P \tilde{D}_{R} P^{-1}$.
Since $(\tilde{D}_{R})^{\dag} = \tilde{D}_{R}$, $\hat{H}_{\text{eff}}$ is $Q$-Hermitian,
$\hat{H}_{\text{eff}} ^{\dag^Q} =\hat{H}_{\text{eff}}$, and
satisfies $\hat{H}_{\text{eff}} | \lambda_i \rangle = \text{Re} \lambda_i | \lambda_i \rangle$.
Then $| A_i (t) \rangle$ is evaluated as
$| A_i (t) \rangle
\simeq e^{ \frac{1}{\hbar} B (t-t_0)}
\sum_{j \in A} a_j^{(i)} (t_0)$
$e^{-\frac{i}{\hbar} {\text Re} \lambda_j (t-t_0)} | \lambda_j \rangle
=e^{ \frac{1}{\hbar} B (t-t_0)} e^{-\frac{i}{\hbar} \hat{H}_{\text{eff}} (t-t_0)}
| \tilde{A}_i (t_0) \rangle = | \tilde{A}_i (t) \rangle$,
where we have introduced $| \tilde{A}_i (t) \rangle \equiv \sum_{j \in A} a_j^{(i)} (t)| \lambda_j \rangle$.
Since the factor $e^{ \frac{1}{\hbar} B (t-t_0)}$
is dropped out for the normalized state $| A_i(t) \rangle_N$,
we have effectively obtained a $Q$-Hermitian Hamiltonian $\hat{H}_{\text{eff}}$
after the long time development.
In fact, the normalized state
$| A_i (t) \rangle_{N}
\simeq
| \tilde{A}_i (t) \rangle_{N}
\equiv
\frac{1}{\sqrt{ \langle \tilde{A}_i (t) |_Q ~\tilde{A}_i(t) \rangle} } | \tilde{A}_i (t) \rangle$
and the expectation value of an operator ${\cal O}$,
$\langle \hat{\cal O} \rangle_Q^{A_i A_i} (t)
\simeq
\langle \hat{\cal O} \rangle_Q^{\tilde{A}_i \tilde{A}_i} (t)
\equiv
{}_{N} \langle \tilde{A}_i (t) |_Q {\cal O} | \tilde{A}_i (t)
\rangle_{N}$,
obey
\begin{eqnarray}
&&i\hbar \frac{\partial}{ \partial t} | \tilde{A}_i (t) \rangle_{N} = \hat{H}_{\text{eff}} | \tilde{A}_i (t) \rangle_{N}, \label{Schreq_for_AitildeketN} \\
&&
\frac{d}{dt} \langle \hat{\cal O} \rangle_Q^{\tilde{A}_i} (t)
=
-\frac{i}{\hbar} \langle \left[ \hat{\cal O}, \hat{H}_{\text{eff}} \right]
\rangle_Q^{\tilde{A}_i \tilde{A}_i} (t).
\end{eqnarray}
Thus we have seen for pure states that
the $Q$-hermitian Hamiltonian $\hat{H}_{\text{eff}}$ emerges.
\section{Density matrices for mixed states in the future-not-included CAT}
In this section we define density matrices to describe mixed states
and study the automatic hermiticity mechanism for them in the future-not-included CAT.
For a given ensemble\footnote{We note that each $| A_i (t) \rangle$ does not need to be orthogonal to each other, and that the number of elements does not have to match the order of the Hilbert space. }
$\left\{ | A_i (t) \rangle \right\}$,
each of which obeys Eq.(\ref{Schr_for_Ai}),
let us consider a mixed state that is composed of $| A_i (t) \rangle_{N}$
with the probability $q_i$ for each index $i$ ($q_i \ge 0$, $\sum_i q_i=1$).
We define the density matrix\footnote{When the density matrix is composed of only one component,
$\hat{\rho}_Q^{A,\text{mixed}}(t)=| A_1(t) \rangle_N {}_N \langle A_1(t) |_Q$,
it describes a pure state,
and satisfies $\hat{\rho}_Q^{A,\text{mixed}}(t)^2 = \hat{\rho}_Q^{A,\text{mixed}}(t)$ and
$\tr\left(\hat{\rho}_Q^{A,\text{mixed}}(t)\right)=1$.}
and expectation value of an operator $\hat{\cal O}$ for it by
\begin{eqnarray}
\hat{\rho}_Q^{AA,\text{mixed}}(t)
&\equiv& \sum_i q_i | A_i (t) \rangle_N {}_N \langle A_i (t) |_Q
\equiv \sum_i q_i \hat{\rho}_Q^{A_i A_i}(t) ,\label{rhoQAmixed} \\
\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{AA,\text{mixed}}} (t)
&\equiv&
\tr\left( \hat{\rho}_Q^{AA,\text{mixed}}(t) \hat{\cal O} \right)
\equiv \sum_i q_i \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{A_i A_i}} (t)
=\sum_i q_i \langle \hat{\cal O} \rangle_Q^{A_i A_i} (t) ,
\end{eqnarray}
where $\hat{\rho}_Q^{A_i A_i} (t)$ obeys
$\hat{\rho}_Q^{A_i A_i}(t)^2 = \hat{\rho}_Q^{A_i A_i}(t)$
and $\tr\left(\hat{\rho}_Q^{A_i A_i} (t) \right)=1$.
So $\tr\left(\hat{\rho}_Q^{A A,\text{mixed}}(t) \right)=1$.
We note that $\hat{\rho}_Q^{A A,\text{mixed}}(t)$ and
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{AA,\text{mixed}}} (t)$
are $Q$-Hermitian and real for $Q$-Hermitian $\hat{\cal O}$, respectively.
They time-develop as follows:
$\frac{d}{dt} \hat{\rho}_Q^{AA,\text{mixed}} (t)
=-\frac{i}{\hbar} \left[ \hat{H}_{Qh}, \hat{\rho}_Q^{AA,\text{mixed}} (t) \right]
-\frac{i}{\hbar} \sum_i q_i
\left\{ \hat{\Delta} \left( \hat{H}_{Qa} ; | A_i (t) \rangle_{N} \right),
\hat{\rho}_Q^{A_i A_i} \right\} (t)$,
$\frac{d}{dt} \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{AA,\text{mixed}}} (t)
=
-\frac{i}{\hbar} \langle \left[ \hat{\cal O} , \hat{H}_{Qh} \right] \rangle_{\hat{\rho}_Q^{AA,\text{mixed}}} (t)$
$-\frac{i}{\hbar} \sum_i q_i$
$\langle \left\{ \hat{\cal O}, \hat{\Delta} \left( \hat{H}_{Qa} ; | A_i (t) \rangle_{N} \right) \right\} \rangle_{\hat{\rho}_Q^{A_i A_i}} (t)$.
It is interesting that, in the classical limit,
since $\hat{\Delta} \left( \hat{H}_{Qa} ; A_i (t) \right)$ given in Eq.(\ref{DeltahatA_i})
is suppressed,
$\hat{\rho}_Q^{AA,\text{mixed}}(t)$
obeys Liouville equation with a $Q$-Hermitian Hamilonian
and Ehrenfest's theorem holds.
Now, let us consider the long time development.
Then, since $| A_i (t) \rangle_{N} \simeq | \tilde{A}_i (t) \rangle_{N}$
obeys Eq.(\ref{Schreq_for_AitildeketN}),
we obtain the following relations
for $\hat{\rho}_Q^{AA,\text{mixed}} (t) \simeq \hat{\rho}_Q^{\tilde{A} \tilde{A}, \text{mixed}} (t)$,
$\hat{\rho}_Q^{A_i A_i} (t) \simeq \hat{\rho}_Q^{\tilde{A}_i \tilde{A}_i}(t)$,
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{AA,\text{mixed}}} (t)
\simeq \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}} (t)$,
and $\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{A_i A_i}} (t) \simeq \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{A}_i \tilde{A}_i}}(t)$:
\begin{eqnarray}
&&
\hat{\rho}_Q^{\tilde{A} \tilde{A}, \text{mixed}} (t)
\equiv \sum_{i} q_i \hat{\rho}_Q^{\tilde{A}_i \tilde{A}_i}(t)
=
\hat{U}_{\text{eff}} (t-T_A)
\hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}(T_A)
\hat{U}_{\text{eff}} (t-T_A)^{\dag^Q} , \label{rhohatAtildemixedQt} \\
&&
\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}} (t)
\equiv
\tr\left( \hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}(t) \hat{\cal O} \right)
\equiv \sum_i q_i \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{A}_i \tilde{A}_i}} (t)
=\sum_i q_i \langle \hat{\cal O} \rangle_Q^{\tilde{A}_i \tilde{A}_i} (t) , \\
&&
\frac{d}{dt} \hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}(t)
=
-\frac{i}{\hbar} \left[ \hat{H}_{\text{eff}} , \hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}(t) \right] , \\
&&
\frac{d}{dt} \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}} (t)
=
-\frac{i}{\hbar} \langle \left[ \hat{\cal O}, \hat{H}_{\text{eff}} \right]
\rangle_{\hat{\rho}_Q^{\tilde{A} \tilde{A},\text{mixed}}} (t) ,
\end{eqnarray}
where $\hat{U}_{\text{eff}} (t-T_A) \equiv e^{-\frac{i}{\hbar} \hat{H}_{\text{eff}} (t-T_A)}$
is ``$Q$-Unitary'',
i.e., $U_{\text{eff}} (t-T_A)^{\dag^Q}=U_{\text{eff}} (t-T_A)^{-1}$.
We find that $\hat{\rho}_Q^{\tilde{A}\tilde{A}, \text{mixed}} (t)$
obeys von Neumann equation with
the $Q$-Hermitian Hamiltonian $\hat{H}_{\text{eff}}$ and Ehrenfest's theorem holds.
Thus we have confirmed that the automatic hermiticity mechanism works for mixed states as well as
for pure states in the future-not-included CAT.
\section{Density matrices for mixed states in the future-included CAT}
In this section we attempt to introduce density matrices to describe mixed states
and study their properties in the future-included CAT.
In addition we investigate the automatic hermiticity mechanism for the mixed states.
The future-included theory is described not only by
the state vector $| A_i (t) \rangle$ that time-develops forward from the initial time $T_A$
according to
the Schr\"{o}dinger equation (\ref{Schr_for_Ai}) but also by
the other one\footnote{We adopt a convention with an index $i$ for the state,
since it will be convenient later.}
$| B_i (t) \rangle$ that time-develops backward from the final time $T_B$ according to
the other Schr\"{o}dinger equation\footnote{In Ref.\cite{Nagao:2015bya}, we adopted the modified inner product $I_Q$
for all quantities in the future-included CAT\cite{Bled2006,Nagao:2012mj,Nagao:2012ye}. We follow the formalism in this manuscript.}:
$i \hbar \frac{d}{dt} | B_i (t) \rangle = {\hat{H}}^{\dag^Q} | B _i (t) \rangle
\Leftrightarrow -i \hbar \frac{d}{dt} \langle B_i (t) |_Q = \langle B_i (t) |_Q \hat{H}$.
The states $| A_i (t) \rangle$ and $| B_i (t) \rangle$ are normalized by
$\langle A_i (T_A) |_Q A_i (T_A) \rangle = \langle B_i (T_B) |_Q B_i (T_B) \rangle = 1$.
The normalized matrix element\footnote{In the case of $Q=1$, this corresponds to the weak value\cite{AAV, review_wv} that is well known in the RAT.}
\begin{equation}
\langle \hat{\mathcal O} \rangle_Q^{B_i A_i} (t)
\equiv \frac{ \langle B_i(t) |_Q \hat{\mathcal O} | A_i(t) \rangle }{ \langle B_i(t) |_Q A_i(t) \rangle }
\label{OQB_iA_i}
\end{equation}
is a good candidate for an expectation value of an operator ${\cal O}$ in the future-included CAT,
because, if it is viewed as such,
then we can obtain the Heisenberg equation, Ehrenfest's theorem,
and a conserved probability current density\cite{Nagao:2012mj,Nagao:2012ye}.
In the future-included CAT,
let us consider the other
ensemble\footnote{Just as each $| A_i (t) \rangle$ does not have to be orthogonal to each other,
neither does each $| B_i (t) \rangle$.} $\left\{ | B_i (t) \rangle \right\}$
besides the ensemble $\left\{ | A_i (t) \rangle \right\}$.
Now we have a simple question: what kind of mixed states can be considered in
the future-included theory?
One possible candidate would be the same type of mixed states as we considered
in the previous section.
Since $| A_i (t) \rangle$ time-develops according to Eq.(\ref{Schr_for_Ai})
in the same way as before, let us consider
the same mixed state described by the density matrix
$\hat{\rho}_Q^{AA,\text{mixed}}(t)$ defined in Eq.(\ref{rhoQAmixed}) for $| A_i (t) \rangle$,
and consider similar ones for $| B_i (t) \rangle$.
Let us introduce a normalized state
and an expectation value of an operator ${\cal O}$ for it by $| B_i (t) \rangle_{N}
\equiv \frac{1}{\sqrt{ \langle B_i (t) |_Q ~B_i (t) \rangle} } | B_i (t) \rangle$ and
$\langle \hat{\cal O} \rangle_Q^{B_i B_i} (t)
\equiv {}_{N} \langle B_i (t) |_Q {\cal O} | B_i (t) \rangle_{N}$,
which time-develop as
$i\hbar \frac{\partial}{ \partial t} | B_i (t) \rangle_{N}
= \hat{H}_{Qh} | B_i (t) \rangle_{N}
- \hat{\Delta} \left( \hat{H}_{Qa} ; | B_i (t) \rangle_{N} \right) | B_i (t) \rangle_{N}$,
$\frac{d}{dt}\langle \hat{\cal O} \rangle_Q^{B_i B_i} (t)
= -\frac{i}{\hbar} \langle \left[ \hat{\cal O} , \hat{H}_{Qh} \right] \rangle_Q^{B_i B_i} (t)
+\frac{i}{\hbar} \langle \left\{ \hat{\cal O}, \hat{\Delta} \left( \hat{H}_{Qa} ; | B_i (t) \rangle_{N} \right) \right\} \rangle_Q^{B_i B_i} (t)$,
where
$\hat{\Delta} \left( \hat{H}_{Qa} ; | B_i (t) \rangle_{N} \right)$ is defined like Eq.(\ref{DeltahatA_i}).
Next let us consider
a mixed state that is given by $| B_i (t) \rangle_{N}$
with the probability $r_i$ for each index $i$ ($r_i \ge 0$, $\sum_i r_i=1$).
We define the density matrix to describe the mixed state
and the expectation value of ${\cal O}$ for it by
$\hat{\rho}_Q^{BB,\text{mixed}}(t)
\equiv \sum_i r_i | B_i (t) \rangle_N {}_N \langle B_i (t) |_Q
\equiv \sum_i r_i \hat{\rho}_Q^{B_i B_i}(t)$,
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BB,\text{mixed}}} (t)
\equiv
\tr\left( \hat{\rho}_Q^{BB,\text{mixed}}(t) \hat{\cal O} \right)
\equiv \sum_i r_i \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{B_i B_i}} (t)
=\sum_i r_i \langle \hat{\cal O} \rangle_Q^{B_i B_i} (t)$,
where $\hat{\rho}_Q^{B_i B_i}(t)$ obeys $\hat{\rho}_Q^{B_i B_i}(t)^2 = \hat{\rho}_Q^{B_i B_i}(t)$
and $\tr \left(\hat{\rho}_Q^{B_i B_i} (t) \right)=1$, so
$\tr\left(\hat{\rho}_Q^{B B,\text{mixed}}(t) \right)=1$.
We note that $\hat{\rho}_Q^{B B,\text{mixed}}(t)$ and
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BB,\text{mixed}}} (t)$
are $Q$-Hermitian and
real for $Q$-Hermitian $\hat{\cal O}$, respectively.
They time-develop as follows:
$\frac{d}{dt} \hat{\rho}_Q^{BB,\text{mixed}} (t)
=-\frac{i}{\hbar} \left[ \hat{H}_{Qh}, \hat{\rho}_Q^{BB,\text{mixed}} (t) \right]
+\frac{i}{\hbar} \sum_i r_i $
$\left\{ \hat{\Delta} \left( \hat{H}_{Qa} ;
| B_i (t) \rangle_{N} \right),
\hat{\rho}_Q^{B_i B_i} \right\} (t)$,
$\frac{d}{dt} \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BB,\text{mixed}}} (t)
=
-\frac{i}{\hbar} \langle \left[ \hat{\cal O} , \hat{H}_{Qh} \right] \rangle_{\hat{\rho}_Q^{BB,\text{mixed}}} (t)
+\frac{i}{\hbar} \sum_i r_i
\langle \left\{ \hat{\cal O}, \hat{\Delta} \left( \hat{H}_{Qa} ; | B_i (t) \rangle_{N} \right) \right\} \rangle_{\hat{\rho}_Q^{B_i B_i}} (t)$,
which are almost the same as those for $\hat{\rho}_Q^{AA,\text{mixed}}(t)$.
The only difference is that the sign in front of $\hat{H}_{Qa}$ is opposite.
Therefore, if we use the automatic hermiticity mechanism
for $| B_i (t) \rangle = \sum_j b_j^{(i)} (t) | \lambda_j \rangle$, then obtaining
$| B_i(t) \rangle
\simeq
e^{ \frac{1}{\hbar} B (T_B-t)} e^{-\frac{i}{\hbar} \hat{H}_{\text{eff}} (t-T_B)}
| \tilde{B}_i (T_B) \rangle
=\sum_{j \in A} b_j^{(i)} (t) | \lambda_j \rangle
\equiv | \tilde{B}_i (t) \rangle$ and
$| B_i (t) \rangle_{N}
\simeq \frac{1}{\sqrt{ \langle \tilde{B}_i (t) |_Q ~\tilde{B}_i(t) \rangle} } | \tilde{B}_i (t) \rangle
\equiv | \tilde{B}_i (t) \rangle_{N}$ for large $T_B -t$,
we find that the various relations for
$\hat{\rho}_Q^{BB,\text{mixed}}(t) \simeq \hat{\rho}_Q^{\tilde{B}\tilde{B},\text{mixed}}(t)$
become the same as those for $\hat{\rho}_Q^{\tilde{A}\tilde{A},\text{mixed}}(t)$.
Thus we have seen that the the automatic hermiticity mechanism works for mixed states
described by the density matrices
$\hat{\rho}_Q^{AA,\text{mixed}} (t) \simeq \hat{\rho}_Q^{\tilde{A} \tilde{A}, \text{mixed}} (t)$ and
$\hat{\rho}_Q^{BB,\text{mixed}} (t) \simeq \hat{\rho}_Q^{\tilde{B} \tilde{B}, \text{mixed}} (t)$, and
that,
via the mechanism, both of the density matrices nicely obey von Neumann equation
with the effectively obtained $Q$-Hermitian Hamiltonian $\hat{H}_{\text{eff}}$.
In addition,
$\hat{\rho}_Q^{A_i A_i} (t)$ and $\hat{\rho}_Q^{B_i B_i} (t)$
have real meanings as density matrices of $| A_i (t) \rangle_N$
and $| B_i (t) \rangle_N$.
However, neither
$\tr\left( \hat{\rho}_Q^{A_i A_i} (t) \hat{\cal O} \right)
={}_{N} \langle A_i (t) |_Q \hat{\cal O} | A_i (t) \rangle_{N}$
nor
$\tr\left( \hat{\rho}_Q^{B_i B_i} (t) \hat{\cal O} \right)
={}_{N} \langle B_i (t) |_Q \hat{\cal O} | B_i (t) \rangle_{N}$
matches the normalized matrix element $\langle \hat{\mathcal O} \rangle_Q^{B_i A_i} (t)$
given in Eq.(\ref{OQB_iA_i}).
In the future-included CAT, we have a philosophy such that
it is not ${}_{N} \langle A_i (t) |_Q \hat{\cal O} | A_i (t) \rangle_{N}$ nor
${}_{N} \langle B_i (t) |_Q \hat{\cal O} | B_i (t) \rangle_{N}$
but $\langle \hat{\mathcal O} \rangle_Q^{B_i A_i} (t)$ that has a role of an expectation value of
$\hat{\mathcal O}$.
Therefore, $\hat{\rho}_Q^{A_i A_i} (t)$ and $\hat{\rho}_Q^{B_i B_i} (t)$ might
not be good density matrices in this sense.
Then what should we adopt as a density matrix
in the future-included CAT if we wish to respect the philosophy?
We are now motivated to consider the other kind of density matrix
such that
the trace of the product of each component with an index $i$ and $\hat{\cal O}$
corresponds to $\langle \hat{\mathcal O} \rangle_Q^{B_i A_i} (t)$.
Introducing
$| A_i (t) \rangle_M \equiv
\frac{| A_i (t) \rangle }{ \sqrt{ \langle B_i (t) |_Q A_i (t) \rangle} }$ and
$| B_i (t) \rangle_M \equiv
\frac{| B_i (t) \rangle }{ \sqrt{ \langle A_i (t) |_Q B_i (t) \rangle} }$,
which obey
$i \hbar \frac{d}{dt} | A_i (t) \rangle_M = \hat{H} | A _i (t) \rangle_M$,
$i \hbar \frac{d}{dt} | B_i (t) \rangle_M = {\hat{H}}^{\dag^Q} | B _i (t) \rangle_M$,
and
${}_M\langle B_i (t) |_Q A _i (t) \rangle_M = 1$,
let us define the following ``skew density matrix'' $\hat{\rho}_Q^{BA,\text{mixed}}(t)$
and ``expectation value'' of $\hat{\cal O}$ for it\footnote{$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}} (t)$ for $Q=1$ corresponds to the weak value for a generalized state
introduced in Ref.~\cite{generalized_two-state_vector_formalism}.
We note that $\tr\left( \hat{\rho}_Q^{B_i A_i} (t) \hat{\cal O} \right)$ can be expressed as
$\tr\left( \hat{\rho}_Q^{B_i A_i} (t) \hat{\cal O} \right) =
\frac{\tr \left( \hat{\rho}_Q^{B_i B_i} (t) \hat{\cal O} \hat{\rho}_Q^{A_i A_i} (t) \right)}{
\tr \left( \hat{\rho}_Q^{B_i B_i} (t) \hat{\rho}_Q^{A_i A_i} (t) \right)}$, where
the expression on the right-hand side for $Q=1$ is introduced in Refs.\cite{WuMolmera,TamateNakanishiKitano}.
} by
\begin{eqnarray}
\hat{\rho}_Q^{BA,\text{mixed}}(t)
&\equiv& \sum_i s_i | A_i (t) \rangle_M {}_M\langle B_i (t) |_Q
\equiv \sum_i s_i \hat{\rho}_Q^{B_i A_i} (t) , \\
\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}} (t)
&\equiv&
\tr\left( \hat{\rho}_Q^{BA,\text{mixed}}(t) \hat{\cal O} \right)
\equiv\sum_i s_i \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{B_i A_i}} (t)
=\sum_i s_i \langle \hat{\cal O} \rangle_Q^{B_i A_i} (t) ,
\end{eqnarray}
where the weight $s_i$ for each $\hat{\rho}_Q^{B_i A_i} (t)$ obeys $s_i \ge 0$, $\sum_i s_i=1$,
and $\hat{\rho}_Q^{B_i A_i} (t)$ obeys
$\tr\left( \hat{\rho}_Q^{B_i A_i} (t) \right) =1$,
$\left( \hat{\rho}_Q^{B_i A_i} (t) \right)^2 = \hat{\rho}_Q^{B_i A_i} (t)$.
So $\tr\left(\hat{\rho}_Q^{BA,\text{mixed}} (t) \right)=1$.
In addition, $\hat{\rho}_Q^{BA,\text{mixed}}(t)$ can be expressed as
$\hat{\rho}_Q^{BA,\text{mixed}}(t)=\hat{U} (t-t_r) \hat{\rho}_Q^{BA,\text{mixed}} (t_r) \hat{U} (t-t_r) ^{-1}$, where
$\hat{U} (t-t_r) \equiv e^{-\frac{i}{\hbar} \hat{H} (t-t_r)}$ is neither Unitary nor $Q$-Unitary,
and $t_r$ is a reference time.
They time-develop as follows:
$\frac{d}{dt} \hat{\rho}_Q^{BA,\text{mixed}}(t)
=
-\frac{i}{\hbar} \left[ \hat{H}, \hat{\rho}_Q^{BA,\text{mixed}}(t) \right]$,
$\frac{d}{dt} \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}} (t)
=
-\frac{i}{\hbar} \langle \left[ \hat{\cal O} , \hat{H} \right] \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}} (t)$,
which show that $\hat{\rho}_Q^{BA,\text{mixed}} (t)$ obeys
von Neumann equation and Ehrenfest's theorem holds as they are.
These properties are quite in contrast to those of $\hat{\rho}_Q^{AA,\text{mixed}} (t)$
and $\hat{\rho}_Q^{BB,\text{mixed}} (t)$.
If we consider the long time development, then for
$| A_i (t) \rangle_M
\simeq | \tilde{A}_i (t) \rangle_M
\equiv \frac{| \tilde{A}_i (t) \rangle }{ \sqrt{ \langle \tilde{B}_i (t) |_Q \tilde{A}_i (t) \rangle} }$ and
$| B_i (t) \rangle_M
\simeq | \tilde{B}_i (t) \rangle_M
\equiv \frac{| \tilde{B}_i (t) \rangle }{ \sqrt{ \langle \tilde{A}_i (t) |_Q \tilde{B}_i (t) \rangle} }$,
we find that $\hat{\rho}_Q^{BA,\text{mixed}}(t) \simeq \hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}(t)
\equiv \sum_i s_i | \tilde{A}_i (t) \rangle_M {}_M\langle \tilde{B}_i (t) |_Q
\equiv \sum_i s_i \hat{\rho}_Q^{\tilde{B}_i \tilde{A}_i}(t)$
and
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}} (t)
\simeq \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}} (t)
\equiv \tr\left(\hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}(t) \hat{\cal O} \right)$
$\equiv\sum_i s_i \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{B}_i \tilde{A}_i}} (t)
=\sum_i s_i \langle \hat{\cal O} \rangle_Q^{\tilde{B}_i \tilde{A}_i} (t)$
time-develop with an effectively obtained $Q$-Hermitian Hamiltonian $\hat{H}_{\text{eff}}$
as follows:
$\frac{d}{dt} \hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}(t)
= -\frac{i}{\hbar} \left[ \hat{H}_{\text{eff}} , \hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}(t) \right]$,
$\frac{d}{dt} \langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}} (t)
= -\frac{i}{\hbar} \langle \left[ \hat{\cal O}, \hat{H}_{\text{eff}} \right] \rangle_{\hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}} (t)$.
However,
$\hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}(t)$ and
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{\tilde{B} \tilde{A},\text{mixed}}} (t)$
are neither $Q$-Hermitian nor real for $Q$-Hermitian $\hat{\cal O}$, respectively,
because $| \tilde{A}_i (t) \rangle_M$ and $| \tilde{B}_i (t) \rangle_M$ are different states.
This is quite in contrast to the cases for $\hat{\rho}_Q^{AA,\text{mixed}} (t)$ and
$\hat{\rho}_Q^{BB,\text{mixed}} (t)$,
where only either $| A_i (t) \rangle_N$ or $| B_i (t) \rangle_N$ is used, respectively.
To resolve this problem, we will consider it in another way.
\section{Hermiticity and reality for $\hat{\rho}_Q^{BA,\text{mixed}}(t)$ and $\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}}(t)$}
In Ref.\cite{Nagao:2012mj}, utilizing the automatic hermiticity mechanism
for pure states time-developing forward and backward,
we obtained the following correspondence:
\begin{equation}
\langle {\cal O} \rangle^{BA} ~\text{for large $T_B-t$ and large $t- T_A$}
\simeq
\langle {\cal O} \rangle_{Q'}^{AA} ~\text{for large $t- T_A$},
\end{equation}
based on the Schr\"{o}dinger equations (\ref{Schr_for_Ai})
and $i\hbar \frac{d}{dt} | B(t) \rangle = H^\dag| B(t) \rangle$,
where $\langle {\cal O} \rangle^{BA}=\frac{\langle B(t) | {\cal O} | A(t) \rangle}{\langle B(t) | A(t) \rangle}$ is a matrix element of an operator ${\cal O}$ defined with a usual inner product ($Q=1$)
in the future-included theory,
while $\langle {\cal O} \rangle_{Q'}^{AA}=\frac{\langle A(t) |_{Q'} {\cal O} | A(t) \rangle}{\langle A(t) |_{Q'} A(t) \rangle}$ is a usual expectation value of ${\cal O}$
defined with a modified inner product $I_{Q'}$ in the future-not-included theory.
We showed this correspondence by improving the method used in Ref.\cite{Bled2006},
which first multiplies $\langle {\cal O} \rangle^{BA}$ by
$1=\frac{ \langle A(t) | B(t) \rangle }{\langle A(t) | B(t) \rangle}$ and then
evaluates $| B(t) \rangle \langle B(t) |$.
This correspondence strongly suggests that, the future-included CAT is not excluded phenomenologically,
even though it looks very exotic.
Utilizing this method, let us estimate $\langle {\cal O} \rangle_Q^{BA}$ and
$\hat{\rho}_Q^{BA}(t)$,
based on the Schr\"{o}dinger equations (\ref{Schr_for_Ai})
and $i\hbar \frac{d}{dt} | B(t) \rangle = H^{\dag^Q} | B(t) \rangle$.
Respecting the inner product $I_Q$ for all quantities,
let us multiply $\langle {\cal O} \rangle_Q^{BA} (t)$ by
$1=\frac{ \langle A(t) |_{Q} B(t) \rangle }{\langle A(t) |_{Q} B(t) \rangle}$, instead of
$1=\frac{ \langle A(t) | B(t) \rangle }{\langle A(t) | B(t) \rangle}$.
Then $\langle {\cal O} \rangle_Q^{BA}$ is rewritten as
$\langle {\cal O} \rangle_Q^{BA} (t)
= \frac{ \langle A(t) |_{Q} B(t) \rangle \langle B(t) |_Q {\cal O} | A(t) \rangle}{\langle A(t) |_{Q} B(t) \rangle \langle B(t) |_Q A(t) \rangle}$.
The expansion of $| B(T_B) \rangle$ used in Ref.\cite{Nagao:2012mj},
$| B(T_B) \rangle = \sum_i b_i | \lambda_i \rangle_B$ in terms of
the eigenstate of $\hat{H}^\dag$, $|\lambda \rangle_B=Q|\lambda \rangle$,
is found to produce too many $Q$, so it does not seem to be
an appropriate choice in the present study.
Hence we adopt another expansion:
$| B(T_B) \rangle = \sum_i c_i | \lambda_i \rangle = \sum_i J(\lambda_i)^* | \lambda_i \rangle$,
where $J(\lambda_i)$ is a function of $\lambda_i$.
Then $| B(t) \rangle \langle B(t) |_Q$ is evaluated as follows:
\begin{eqnarray}
| B(t) \rangle \langle B(t) |_Q
&=&
e^{-\frac{i}{\hbar} \hat{H}^{\dag^Q} (t - T_B)} | B(T_B) \rangle
\langle B(T_B) |_Q e^{\frac{i}{\hbar} \hat{H} (t - T_B) } \nonumber \\
&=&
\sum_{i,j} c_i c_j^* e^{\frac{i}{\hbar} \text{Re}(\lambda_j - \lambda_i) (t - T_B)}
e^{ \frac{1}{\hbar} \text{Im}(\lambda_j + \lambda_i) (T_B - t)}
| \lambda_i \rangle \langle \lambda_j |_Q \nonumber \\
&\simeq&
\frac{\int_{t-\Delta t}^{t+ \Delta t} | B(t) \rangle \langle B(t) |_Q dt }{ \int_{t-\Delta t}^{t+ \Delta t} dt }
\simeq
\sum_i | c_i |^2 e^{ \frac{2}{\hbar}\text{Im} \lambda_i (T_B - t)} | \lambda_i \rangle ~\langle \lambda_i |_Q
\nonumber \\
&\simeq& e^{ \frac{2}{\hbar}B (T_B - t)} Q_4 \quad \text{for large $T_B - t$} ,
\end{eqnarray}
where in the third line we have smeared the present time $t$ a little bit,
and the off-diagonal elements wash to $0$.
In the last line we have used the automatic hermiticity mechanism
for large $T_B - t$, and introduced
$Q_4 \equiv \sum_{i \in A} | c_i |^2 | \lambda_i \rangle \langle \lambda_i |_Q$,
which is expressed as follows:
$Q_4
=J(\hat{H}_{\text{eff}} + i B \Lambda_A )^{\dag^Q}
\Lambda_A
J(\hat{H}_{\text{eff}} + iB \Lambda_A )
=Q^{-1} \tilde{J}(\hat{H}_{\text{eff}} )^\dag Q \tilde{J}(\hat{H}_{\text{eff}} )
\equiv Q^{-1} Q_{\tilde{J}}$.
Here, supposing that $\text{Re} \lambda_i$ are not degenerate,
we have introduced
$\Lambda_A \equiv \sum_{i \in A} | \lambda_i \rangle \langle \lambda_i |_Q$,
a function $\tilde{J}$
such that $\tilde{J}(\text{Re} \lambda_i ) \equiv J(\text{Re} \lambda_i + iB) = c_i^*$ for $i \in A$,
and $Q_{\tilde{J}} \equiv \tilde{J}(\hat{H}_{\text{eff}} )^\dag Q \tilde{J}(\hat{H}_{\text{eff}} )$.
Now we use the automatic hermiticity mechanism for large $t-T_A$.
Then, since $| A(t) \rangle \equiv \sum_i a_i(t) | \lambda_i \rangle $ behaves as
$| \tilde A(t) \rangle \equiv \sum_{i \in A} a_i(t) | \lambda_i \rangle$,
we obtain
$\langle {\cal O} \rangle_Q^{BA}
\simeq
\frac{ \langle \tilde{A} (t) |_{Q_{\tilde{J}}} {\cal O} | \tilde{A}(t) \rangle }
{ \langle \tilde{A}(t) |_{Q_{\tilde{J}}} \tilde{A}(t) \rangle }
\equiv \langle {\cal O} \rangle_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}
\quad \text{for large $T_B - t$ and large $t-T_A$}$.
Next, let us consider the expectation value in the future-not-included theory:
$\langle {\cal O} \rangle_{Q_J}^{AA}
\equiv
\frac{ \langle A(t) |_{Q_J} {\cal O} | A(t) \rangle }{ \langle A(t) |_{Q_J} A(t) \rangle }$,
where $Q_J \equiv J(\hat{H})^\dag Q J(\hat{H}) =({P_{J^{-1}} }^{-1} )^\dag {P_{J^{-1}} }^{-1}$,
and $P_{J^{-1}} \equiv J(\hat{H})^{-1} P$ diagonalizes $\hat{H}$:
$(P_{J^{-1}})^{-1} \hat{H} P_{J^{-1}} = P^{-1} \hat{H} P = D$.
We introduce
$| \lambda_i \rangle^{J^{-1}} \equiv J(\hat{H})^{-1} | \lambda_i \rangle$,
so that $| \lambda_i \rangle^{J^{-1}}$ is $Q_J$-orthogonal, i.e.,
$I_{Q_J} ( | \lambda_i \rangle^{J^{-1}} , | \lambda_j \rangle^{J^{-1}} )
\equiv {}^{J^{-1}}\langle \lambda_i | Q_J | \lambda_j \rangle^{J^{-1}}
=\delta_{ij}$.
We use the automatic Hermiticity mechanism for large $t-T_A$.
Then, $| A(t) \rangle$ behaves as
$| \tilde A(t) \rangle = \sum_{i \in A} a_i(t) | \lambda_i \rangle$,
and $Q_J$ is estimated as follows:
$Q_J
\simeq J(\hat{H}_{\text{eff}} + iB \Lambda_A )^\dag Q J(\hat{H}_{\text{eff}} + iB \Lambda_A )
=\tilde{J}(\hat{H}_{\text{eff}} )^\dag Q \tilde{J}(\hat{H}_{\text{eff}} )
= Q_{\tilde{J}}$.
Then the expectation value in the future-not-included theory is expressed as
$\langle {\cal O} \rangle_{Q_J}^{AA}
\simeq \frac{ \langle \tilde{A}(t) |_{Q_{\tilde{J}}} {\cal O} | \tilde{A}(t) \rangle }
{ \langle \tilde{A}(t) |_{Q_{\tilde{J}}} \tilde{A}(t) \rangle }
=\langle {\cal O} \rangle_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}} \quad \text{for large $t-T_A$}$.
Thus we have obtained the following correspondence:
\begin{equation}
\langle {\cal O} \rangle_Q^{B A} ~\text{for large $T_B-t$ and large $t-T_A$}
\quad
\simeq
\quad
\langle {\cal O} \rangle_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}
\quad
\simeq
\quad
\langle {\cal O} \rangle_{Q_J}^{AA} ~\text{for large $t-T_A$}.
\end{equation}
This correspondence suggests that
the future-included theory is not excluded, although it looks very exotic.
We note that $\langle {\cal O} \rangle_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}$ is real
for $Q_{\tilde{J}}$-Hermitian ${\cal O}$, and time-develops
according to the $Q_{\tilde{J}}$-Hermitian Hamiltonian $\hat{H}_{\text{eff}}$.
We can apply this correspondence to each $i$-component
$\langle \hat{\cal O} \rangle_Q^{B_i A_i} (t)$
of $\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}} (t)$.
Next let us evaluate the skew density matrix
$\hat{\rho}_Q^{B A} (t)=\frac{| A (t) \rangle \langle B (t) |_Q }{\langle B (t) |_Q A (t) \rangle }$
by multiplying it by $1=\frac{ \langle A(t) |_{Q} B(t) \rangle }{\langle A(t) |_{Q} B(t) \rangle}$.
For $\hat{\rho}_Q^{B A} (t)
=\frac{| A (t) \rangle \langle A(t) |_{Q} B(t) \rangle \langle B (t) |_Q }{\langle A(t) |_{Q} B(t) \rangle \langle B (t) |_Q A (t) \rangle }$,
we make use of the above evaluation of $| B(t) \rangle \langle B(t) |_Q$.
Then we obtain the correspondence:
\begin{equation}
\hat{\rho}_Q^{B A} (t) ~\text{for large $T_B-t$ and large $t-T_A$}
\quad
\simeq
\quad
\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}} (t)
\quad
\simeq \quad
\hat{\rho}_{Q_J}^{AA} (t)~\text{for large $t-T_A$} ,
\end{equation}
where
$\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}(t)
\equiv
\frac{| \tilde{A} (t) \rangle \langle \tilde{A}(t) |_{Q_{\tilde{J}}} }{\langle \tilde{A}(t) |_{Q_{\tilde{J}}} \tilde{A} (t) \rangle }
=\hat{U}_{\text{eff}} (t-t_r)
\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}(t_r)
\hat{U}_{\text{eff}} (t-t_r)^{\dag^{Q_{\tilde{J}}}}$.
Here $t_r$ is a reference time, and $\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}(t)$ obeys
$\tr\left( \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}} \right) =1$ and
$\left( \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}} \right)^2 = \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}$.
$\hat{U}_{\text{eff}} (t-t_r) = e^{-\frac{i}{\hbar} \hat{H}_{\text{eff}} (t-t_r)}$ is $Q_{\tilde{J}}$-Unitary,
and $\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}\tilde{A}}(t)$ is $Q_{\tilde{J}}$-Hermitian.
We can apply this correspondence to each $i$-component
$\hat{\rho}_Q^{B_i A_i} (t)$ of $\hat{\rho}_Q^{BA,\text{mixed}}(t)$.
Therefore, though our skew density matrix $\hat{\rho}_Q^{B_i A_i} (t)$
is not $Q$-Hermitian by its definition,
after a long time development it results in a usual expression of density matrix
$\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}_i \tilde{A}_i} (t)$
that is $Q_{\tilde{J}}$-Hermitian.
Application to $\hat{\rho}_Q^{BA,\text{mixed}}(t) = \sum_i s_i \hat{\rho}_Q^{B_i A_i} (t)$
is rather straightforward and we easily see that it time-develops similarly.
Indeed, applying this correspondence to each component
$\hat{\rho}_Q^{B_i A_i} (t)$ of $\hat{\rho}_Q^{BA,\text{mixed}}(t)$,
we find that the expectation value of ${\cal O}$ for $\hat{\rho}_Q^{B_i A_i} (t)$,
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{B_i A_i}} (t)$, is expressed
for large $T_B-t$ and large $t-T_A$ as
\begin{eqnarray}
&&\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{B_i A_i}} (t)
=
\tr\left( \hat{\rho}_Q^{B_i A_i} (t) \hat{\cal O} \right)
\simeq
\tr\left( \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}_i \tilde{A}_i} (t) \hat{\cal O} \right)
\equiv
\langle \hat{\cal O} \rangle_{\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}_i \tilde{A}_i}} (t)
=\langle {\cal O} \rangle_{Q_{\tilde{J}}}^{\tilde{A}_i \tilde{A}_i} (t) ,
\end{eqnarray}
which is real for $Q_{\tilde{J}}$-Hermitian ${\cal O}$.
Finally,
$\hat{\rho}_Q^{BA,\text{mixed}}(t) \simeq \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}(t)
= \sum_i s_i \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}_i \tilde{A}_i}(t)$
and
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}} (t)
\simeq \langle \hat{\cal O} \rangle_{\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}} (t)
=\sum_i s_i \langle \hat{\cal O} \rangle_{\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A}_i \tilde{A}_i}} (t)$
time-develop according to
\begin{eqnarray}
&&
\frac{d}{dt} \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}(t)
=
-\frac{i}{\hbar} \left[ \hat{H}_{\text{eff}} , \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}(t) \right] , \\
&&
\frac{d}{dt} \langle \hat{\cal O} \rangle_{\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}} (t)
=
-\frac{i}{\hbar} \langle \left[ \hat{\cal O}, \hat{H}_{\text{eff}} \right]
\rangle_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}} (t) ,
\end{eqnarray}
which show that
$\hat{\rho}_Q^{BA,\text{mixed}}(t) \simeq \hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}(t)$
obeys von Neumann equation with
the $Q_{\tilde{J}}$-Hermitian Hamiltonian $\hat{H}_{\text{eff}}$ and Ehrenfest's theorem holds.
We note that $\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}(t)$ is
$Q_{\tilde{J}}$-Hermitian, and
$\langle \hat{\cal O} \rangle_{\hat{\rho}_{Q_{\tilde{J}}}^{\tilde{A} \tilde{A},\text{mixed}}} (t)$
is real for $Q_{\tilde{J}}$-Hermitian ${\cal O}$.
Thus we have seen that the problem with $\hat{\rho}_Q^{BA,\text{mixed}}(t)$ and
$\langle \hat{\cal O} \rangle_{\hat{\rho}_Q^{BA,\text{mixed}}}(t)$
mentioned at the end of the previous section can be effectively resolved
by considering the long time development for large $T_B-t$ and large $t-T_A$.
\vspace*{0.5cm}
\section{Discussion}
In this manuscript, we first briefly reviewed the modified inner product $I_Q$ that
makes a given non-normal Hamiltonian normal with regard to it,
and the automatic hermiticity mechanism\cite{Nagao:2010xu,Nagao:2011za,Nagao:2012mj}, which
we previously proposed and studied
for pure states in the CAT.
Next, in the case of the future-not-included CAT, we defined
a density matrix $\hat{\rho}_Q^{AA,\text{mixed}}(t)$ to describe a mixed state and
an expectation value of an operator ${\cal O}$ for it,
and studied their properties.
Then we found that, in the classical limit,
since the effect of the anti-$Q$-Hermitian part of a given non-normal Hamiltonian $\hat{H}$ is suppressed,
the density matrix obeys Liouville equation with a $Q$-Hermitian Hamilonian
and Ehrenfest's theorem holds.
In addition, we showed that, if we consider a long time development,
then eigenvectors having the largest imaginary part of the eigenvalues of $\hat{H}$ dominate most,
and on the subspace spanned by such eigenvectors,
$Q$-Hermitian Hamiltonian effectively emerges, and the expectation value of
${\cal O}$ becomes real for $Q$-Hermitian ${\cal O}$.
Thus we confirmed that the automatic hermiticity mechanism works for mixed states
in the future-not-included theory.
On the other hand, the situation becomes more complicated and non-trivial in the future-included theory.
It is because in the future-not-included theory there is only one class of ensembles of state vectors,
$\left\{ | A_i (t) \rangle \right\}$,
that time-develops forward from the past, while in the future-included theory there are two
classes of
ensembles of state vectors, $\left\{ | A_i (t) \rangle \right\}$ and $\left\{ | B_i (t) \rangle \right\}$,
that time-develop
forward from the initial time $T_A$ and backward from the final time $T_B$, respectively.
So it seems that there are at least a couple of candidates for density matrices in the future-included theory.
As the first candidate, we investigated a pair of density matrices
$\hat{\rho}_Q^{AA,\text{mixed}}(t)$ and $\hat{\rho}_Q^{BB,\text{mixed}}(t)$,
which are composed of only either
$\left\{ | A_i (t) \rangle \right\}$ or $\left\{ | B_i (t) \rangle \right\}$, and argued that,
though the pair has nice properties,
it has common disadvantage in the future-included theory.
In general, the trace of the product of
a density matrix and an operator ${\cal O}$
has to match an expectation value of ${\cal O}$, but it is not the case for this pair,
because it is the matrix element of ${\cal O}$, $\langle {\cal O} \rangle_Q^{BA}$,
that is expected to work as an expectation value of ${\cal O}$ in the future-included theory.
To resolve this problem, we introduced a ``skew density matrix'' $\hat{\rho}_Q^{BA,\text{mixed}}(t)$,
which is composed of both
$\left\{ | A_i (t) \rangle \right\}$ and $\left\{ | B_i (t) \rangle \right\}$.
It has a nice property such that the trace of the product of it
and an operator ${\cal O}$
matches the matrix element $\langle {\cal O} \rangle_Q^{BA}(t)$.
In addition,
utilizing the automatic hermiticity mechanism, we showed that
the skew density matrix $\hat{\rho}_Q^{BA,\text{mixed}}(t)$
and matrix element
$\langle {\cal O} \rangle_Q^{BA}(t)$ defined with an inner product $I_Q$
in the future-included theory for large $T_B-t$ and large $t-T_A$ approximately correspond to
another density matrix $\hat{\rho}_{Q_J}^{AA,\text{mixed}}(t)$ and an expectation value
$\langle {\cal O} \rangle_{Q_J}^{AA}(t)$ defined with another inner product $I_{Q_J}$
in the future-not-included theory for large $t-T_A$.
Therefore, even though the skew density matrix is not a density matrix in a usual sense,
it can effectively work as if it were a usual density matrix.
Thus we argued that it is the skew density matrix that is expected to have a role of a density matrix
in the future-included theory.
In addition, we confirmed that the automatic hermiticity mechanism works for mixed states
in the future-included theory.
Now density matrices have been implemented in the CAT.
What should we study by using the density matrices?
It would be interesting to investigate the classical dynamics of the CAT in phase space.
For this purpose, we need to study further in detail
the harmonic oscillator model that we previously formulated by introducing
the two-basis formalism\cite{Nagao:2019dew}.
Also, it would be intriguing to evaluate von Neumann entropy in the CAT.
Furthermore, density matrices are necessary tools if we wish to deal with
quantum measurement appropriately.
We expect that the density matrices would be convenient for various
analyses in the CAT, and would like to report such studies
in the future.
\ack
This work was supported by JSPS KAKENHI Grant Number JP21K03381, and accomplished
during K.N.'s sabbatical stay in Copenhagen.
He would like to thank the members and visitors of NBI
for their kind hospitality and Klara Pavicic for her various kind arrangements and
consideration during his visits to Copenhagen.
H.B.N. is grateful to NBI for allowing him to work there as emeritus.
Furthermore, the authors would like to express their gratitude to David Gross
for stimulating discussion when they attended the Bohr-100 workshop.
\let\doi\relax
|
\section{Introduction}\label{s:intro}
Consider an insoluble dye in an incompressible fluid.
Stirring the fluid typically causes filamentation, stretching blobs of die into fine tendrils.
Diffusion, on the other hand, efficiently damps these small scales, and the combination of these two effects results in \emph{enhanced dissipation} -- the tendency of passive scalars to diffuse faster than in the absence of stirring.
This phenomenon has been extensively studied in many contexts, and various authors have established a link between mixing and dissipation enhancement~\cite{ConstantinKiselevEA08,Zlatos10,FengIyer19,CotiZelatiDelgadinoEA20}, studied dissipation enhancement in more general situations~\cite{Seis20,AlbrittonBeekieEA21,NobiliPottel22} and studied it extensively for shear flows~\cite{Taylor53,BedrossianCotiZelati17,Wei19,GallayZelati21,ColomboCotiZelatiEA21}.
Enhanced dissipation has also been used to suppress non-linear effects arising in certain situations~\cite{FannjiangKiselevEA06,KiselevXu16,FengFengEA20,IyerXuEA21}, and is a subject of active study.
The purpose of this work is to quantify dissipation enhancement for cellular flows, thus providing simple and explicit examples of flows with arbitrarily large dissipation enhancement.
Cellular flows arise as a model problem where ambient fluid velocity is a periodic array of opposing vortices.
They have been extensively studied in the context of fluid dynamics, homogenization and as random perturbations of dynamical systems~\cite{Childress79,ChildressSoward89,FannjiangPapanicolaou94,Koralov04,NovikovPapanicolaouEA05,DolgopyatKoralov08,Bakhtin11,HairerIyerEA18}.
We will use probabilistic techniques to estimate the \emph{mixing time} of a diffusion whose drift is a cellular flow.
We then estimate the dissipation enhancement in terms of the mixing time.
The bounds we obtain are significantly better than the bounds previously obtained in~\cite{IyerXuEA21}, and (up to a logarithmic factor) they agree with the optimal heuristic bounds.
\section{Main Result}
We will study the concentration of a dye, denoted by $\phi$, as a passive scalar, evolving according to the advection diffusion equation
\begin{equation}\label{e:phiEq}
\partial_t \phi - (u \cdot \grad) \phi - \frac{\kappa}{2} \lap \phi = 0 \qquad \text{in } (0, \infty) \times \T^d\,.
\end{equation}
Here $-u$ represents the velocity field of the ambient fluid, and $\kappa/2 > 0$ is the molecular diffusivity.
We restrict our attention to the periodic $d$-dimensional torus $\T^d$ with side length~$1$, and we will normalize the initial concentration, $\phi_0$, so that
\begin{equation*}
\int_{\T^d} \phi_0(x) \, dx = 0\,.
\end{equation*}
As time evolves, the dye spreads uniformly across the torus and $\phi(\cdot, t) \to 0$ as~$t \to \infty$.
One measure of convergence rate that will interest us is the~\emph{dissipation time}: the time required for solutions to~\eqref{e:phiEq} to lose a constant fraction of their initial energy (see for instance~\cite{FannjiangNonnenmacherEA04,ConstantinKiselevEA08,FengIyer19}).
Explicitly, \emph{dissipation time}, denoted by $t_{\!\diss} = t_{\!\diss}(\kappa, u)$ is defined by
\begin{equation}\label{e:tdisDefPDE}
t_{\!\diss} \defeq \inf\set[\Big]{t \geq 0 \st \norm{\phi(s + t)}_{L^2} \leq \frac{1}{2} \norm{\phi(s)}_{L^2}\,~ \text{for all } s \geq 0,~ \phi(s) \in \dot L^2}\,.
\end{equation}
Here $\dot L^2$ denotes the space of all mean-zero, square integrable functions on the torus~$\T^2$.
The Poincar\'e inequality and the fact that $u$ is divergence free immediately imply
\begin{equation}\label{e:poincare}
t_{\!\diss}(\kappa, u) \leq \frac{1}{4 \pi^2 \kappa}\,.
\end{equation}
However, this is only an upper bound, and the dissipation time may in fact be much smaller than $O(1/\kappa)$.
When this occurs (i.e.\ when $t_{\!\diss}(u, \kappa) \leq o(1/\kappa)$) it is known as \emph{enhanced dissipation}.
Intuitively, enhanced dissipation when the stirring velocity field generates small scales (e.g. through filamentation), which are then damped much faster by the diffusion.
Seminal work of Constantin et\ al.~\cite{ConstantinKiselevEA08} provides a spectral characterization of (time independent) velocity fields for which $t_{\!\diss} = o(1/\kappa)$.
More explicit, improved bounds were recently obtained in terms of the mixing rate of~$u$.
For instance, if~$u$ is exponentially mixing then one can show $t_{\!\diss} \leq O(\abs{\ln \kappa}^2)$ (see for instance~\cite{FengIyer19,Feng19,CotiZelatiDelgadinoEA20}).
In the context of applications, various authors have shown that sufficiently enhanced dissipation can be used to quench reactions, stop phase separation and prevent singularity formation (see for instance~\cite{FannjiangKiselevEA06,KiselevXu16,FengFengEA20,IyerXuEA21,FengMazzucato22,FengShiEA22}).
Thus finding simple and explicit examples of flows which sufficiently enhance dissipation (i.e.\ make~$t_{\!\diss}$ arbitrarily small) are useful for many applications.
While such flows can be found by rescaling velocity fields with strong enough mixing properties (see for instance~\cites{FengFengEA20,IyerXuEA21}), examples of mixing velocity fields on the torus are notoriously hard to construct.
The main goal of this work is to provide a simple and explicit family of velocity fields for which~$t_{\!\diss}$ can be made arbitrarily small.
The family of flows we construct are two dimensional cellular flows.
These arise frequently in fluid dynamics as flows around strong arrays of opposing vortices and have been extensively studied~\cite{Childress79,RhinesYoung83,ChildressSoward89,FannjiangPapanicolaou94,Heinze03,NovikovPapanicolaouEA05,Koralov04}.
Given $\epsilon > 0$, consider the cellular flow~$v$ defined by
\begin{equation}\label{e:v}
v \defeq \grad^\perp (\xi H)
= \begin{pmatrix} - \partial_2 (\xi H) \\ \phantom-\partial_1 (\xi H) \end{pmatrix}
\,,
\qquad\text{where }
H(x) \defeq
\sin\paren[\Big]{ \frac{2 \pi x_1}{\epsilon} }
\sin\paren[\Big]{ \frac{2 \pi x_2}{\epsilon} }\, ,
\end{equation}
and $\xi$ is a smooth periodic cutoff function function such that
\begin{equation*}
\xi(x) = \begin{cases}
1 & \abs{H(x)} \leq \frac{1}{4}\,,\\
0 & \abs{H(x)} \geq \frac{1}{2}\,.
\end{cases}
\end{equation*}
\begin{figure}[htb]
\centering
\includegraphics[width=.5\textwidth]{images/cflow.pdf}
\caption{%
Stream lines of the cellular flow defined in equation~\eqref{e:v}.
The flow is only non-zero in the shaded region.
}
\label{f:cflow}
\end{figure}
This flow has cell size~$O(\epsilon)$, and its stream lines are shown in Figure~\ref{f:cflow}.
Our main result chooses $u = A v$ for $A \gg 1$, and estimates the \emph{mixing time} explicitly in terms of the flow amplitude~$A$, cell size~$\epsilon$ and diffusivity~$\kappa$ as follows.
\begin{theorem}\label{t:main}
If $u = A v$ with $v$ defined in~\eqref{e:v}, then there exists a finite constant $C$, independent of $\epsilon$, $A$, and $\kappa$, such that
\begin{equation}\label{e:tdisBound}
t_{\!\diss} \leq 3 t_\mix
\leq C \Bigl( \dfrac{\varepsilon^2}{\kappa} + \dfrac{\abs{\ln \delta}^2}{A \varepsilon^2} \Bigr) \,,
\qquad\text{where }
\delta = \sqrt{\frac{\kappa}{A}}\,.
\end{equation}
\end{theorem}
Here $t_\mix = t_\mix(u, \kappa)$ is the \emph{mixing time}, a notion used to measure the convergence rate of Markov processes~\cite{LevinPeresEA09,MontenegroTetali06}.
Roughly speaking, the \emph{mixing time} is the minimum amount of time required for the fundamental solution of~\eqref{e:phiEq} to be $L^1$-close to the invariant distribution, which in our case is the uniform distribution.
That is, if $\rho(x, s; y, t)$ is the fundamental solution of~\eqref{e:phiEq}, the mixing time is defined by
\begin{equation}\label{e:tmixDef}
t_\mix \defeq \inf\set[\Big]{ t \geq 0 \st \sup_{x \in \T^d,\; s \geq 0} \int_{\T^d} \abs{ \rho(x, s; y, s+t) - 1} \, dy < \frac{1}{2}\,,~\forall s\geq 0 }\,.
\end{equation}
The mixing time and dissipation time are related to each other: the dissipation time is bounded by three times the mixing time.
The mixing time can also be bounded by the dissipation time, up to a logarithmic factor.
This is a general result and is not specific to cellular flows.
\begin{proposition}\label{p:tMixTDis}
Let $u \in L^\infty( [0, \infty); W^{1, \infty}(\T^d) )$ be a divergence free vector field, and let $t_\mix = t_\mix(u, \kappa)$, $t_{\!\diss} = t_{\!\diss}(u, \kappa)$ denote the mixing time and dissipation time respectively.
There exists a dimensional constant $C = C(d) < \infty$, independent of $u$ and $\kappa$ such that for all sufficiently small~$\kappa > 0$ we have
\begin{equation}\label{e:disMix}
t_{\!\diss}
\leq 3 t_\mix
\leq C t_{\!\diss} \ln \paren[\Big]{ 1 + \frac{1}{\kappa t_{\!\diss}} }\,.
\end{equation}
\end{proposition}
\begin{remark}
By rescaling we see that on a torus with side length~$\ell$, the above becomes
\begin{gather}\label{e:disMixEll}\tag{\ref*{e:disMix}$'$}
t_{\!\diss}
\leq 3 t_\mix
\leq C t_{\!\diss} \ln \paren[\Big]{ 1 + \frac{\ell^2}{\kappa t_{\!\diss}} }\,.
\end{gather}
for some dimensional constant $C=C(d)$ that is independent of~$\ell$.
\end{remark}
We are presently unaware whether or not the logarithmic factor is necessary.
We prove Proposition~\ref{p:tMixTDis} in Section~\ref{s:tdismix}, below.
\smallskip
We now compare Theorem~\ref{t:main} to previously available bounds.
Previous work of the first author, Xu and Zlato\v s~\cite{IyerXuEA21} has already shown that the dissipation time of a sufficiently strong and fine cellular flow can be made arbitrarily small.
The estimates in~\cite{IyerXuEA21}, however, are neither explicit nor optimal!
In particular, Theorem 1.3 in~\cite{IyerXuEA21} only asserts the existence of sufficiently strong and fine cellular flows with arbitrarily small~$t_{\!\diss}$, without providing a quantitative bound.
A more explicit bound is provided in~\cite[Remark 6.6]{IyerXuEA21} which yields a sub-optimal bound of the form $t_{\!\diss} \leq C \log(A/\kappa) A^{-1/64} \kappa^{-1}$ after rescaling.
This is much weaker than~\eqref{e:tdisBound}.
We next compare Theorem~\ref{t:main} to the well known homogenization results that estimate the \emph{effective diffusivity}.
Recall standard results (see for instance~\cites{BensoussanLionsEA78,PavliotisStuart08}) show that the long time behavior of~\eqref{e:phiEq} is effectively that of the purely diffusive equation
\begin{equation*}
\partial_t \bar \phi - \frac{1}{2} D_\textit{eff} \lap \bar \phi = 0\,,
\end{equation*}
with an enhanced diffusion coefficient $D_\textit{eff}$, known as the \emph{effective diffusivity}.
The effective diffusivity of cellular flows has been extensively studied~\cites{Childress79,FannjiangPapanicolaou94,Koralov04} and is known to asymptotically be
\begin{equation*}
D_\textit{eff} \approx O\paren{ \epsilon^2 \sqrt{\kappa A} }\,.
\end{equation*}
Given this one would expect from~\eqref{e:poincare} that
\begin{equation*}
t_{\!\diss} \leq \frac{1}{4 \pi^2 D_\textit{eff}} = O\paren[\Big]{\frac{1}{\epsilon^2 \sqrt{\kappa A}} }
\end{equation*}
For the above to be valid, however, one would need $t_{\!\diss} \geq O(\epsilon^2 / \kappa)$, the time the dye takes to diffuse through one cell.
This heuristic leads one to choose $\epsilon = O(\kappa / A)^{1/2}$, leading to the bound $t_{\!\diss} \leq O(1/\kappa)$.
Not only is this much weaker than~\eqref{e:tdisBound}, but it also provides no significant improvement over~\eqref{e:poincare}.
Thus the mixing time bound in Theorem~\ref{t:main} captures an effect on time scales \emph{much smaller} than the time scales at which standard homogenization results are applicable.
\medskip
We now discuss optimizing the choice of parameters~$\epsilon, A$ in Theorem~\ref{t:main}.
\begin{remark}
One strategy is to simply minimize the right hand side of~\eqref{e:tdisBound} in the cell size~$\epsilon$, and choose
\begin{equation*}
\epsilon = \sqrt{ \delta \abs{\ln \delta} }
= \frac{1}{\sqrt{2}} \paren[\Big]{\frac{\kappa}{A}}^{1/4} \paren[\Big]{ \ln\paren[\Big]{\frac{A}{\kappa}} }^{1/2}\,.
\end{equation*}
This choice of~$\epsilon$ leads to
\begin{equation*}
t_\mix \leq \frac{C \delta \abs{\ln \delta} }{\kappa}
= \frac{C \abs{\ln \delta} }{\sqrt{\kappa A} }\,.
\end{equation*}
For a fixed~$\kappa$ the right hand side vanishes as $A \to \infty$, providing a simple family of explicit flows with arbitrarily small dissipation time.
\end{remark}
\begin{remark}[Fixed energy]
Another strategy to optimize the choice of parameters in~\eqref{e:tdisBound} is to minimize the right hand side while holding the \emph{energy} $\norm{u}_{L^2}^2 = O(A^2 / \epsilon^2)$ constant.
Physically the energy~$\norm{u}_{L^2}^2$ is proportional to the total kinetic energy of the ambient fluid.
In this case we choose
\begin{equation*}
\epsilon
\approx \paren[\Big]{ \frac{ \kappa \abs{\ln \delta}^2}{\norm{u}_{L^2} }}^{1/5}
= \frac{ \abs{\ln \delta}^{2/5}}{\mathrm{Pe}^{1/5}} \,.
\end{equation*}
Here $\mathrm{Pe} = A / (\epsilon \kappa)$ is the \emph{P\'eclet number} --- a non-dimensional ratio measuring the relative strength of the convection and diffusion.
This choice of cell size results in
\begin{equation}\label{e:fixedEnergy}
t_\mix
\leq \frac{C \abs{\ln \delta}^{4/5}}{\kappa^{3/5} \norm{u}_{L^2}^{2/5}}
= \frac{C \abs{\ln \delta}^{4/5}}{\kappa \mathrm{Pe}^{2/5}}
\end{equation}
For any fixed molecular diffusivity~$\kappa > 0$, the right hand side vanishes as $\norm{u}_{L^2}^2 \to \infty$, and~\eqref{e:fixedEnergy} provides the optimal asymptotic behavior of the mixing time of cellular flows with high energy.
\end{remark}
\begin{remark}[Fixed power]
Finally we choose the parameters~$\epsilon, A$ to minimize the right hand side of~\eqref{e:tdisBound} while holding the \emph{enstrophy}~$\norm{\grad u}^2 = O(A^2 / \epsilon^4)$ constant.
Physically the enstrophy~$\norm{\grad u}_{L^2}^2$ is proportional to the power dissipated by the ambient fluid (see~\cite{LinThiffeaultEA11}).
In this case we choose
\begin{equation}\label{e:fixedPower}
\epsilon \approx \paren[\Big]{ \frac{\abs{\ln \delta}^2 \kappa }{\norm{\grad u} } }^{1/6}\,,
\qquad\text{which implies}\qquad
t_\mix \leq \frac{C \abs{\ln \delta}^{2/3} }{\kappa^{2/3} \norm{\grad u}^{1/3} }\,.
\end{equation}
For any fixed molecular diffusivity~$\kappa > 0$, the right hand side vanishes as $\norm{u}_{L^2}^2 \to \infty$, and~\eqref{e:fixedPower} provides the optimal asymptotic behavior of the mixing time of cellular flows with high enstrophy.
\end{remark}
We now describe the main idea behind our proof.
The Ito diffusion associated to~\eqref{e:phiEq} is defined by the SDE
\begin{equation}\label{eq:itodef}
d X_t = A v(X_t) \, dt + \sqrt{\kappa} \, d B_t\,,
\end{equation}
on the $2$-dimensional torus $\T^2$.
Here $B$ is a standard $2$-dimensional Brownian motion.
Since $\dv v = 0$, the invariant measure of~$X$ is the Lebesgue measure on the torus.
We will estimate the mixing time of~$X$ by constructing a successful coupling (see for instance~\cite{LevinPeresEA09,MontenegroTetali06}).
To do this, we first project the process to one cell (of side length $\epsilon$, where we enforce $1/\varepsilon$ to be integer), and couple the projections.
The drift doesn't help at this stage and it takes on average $O(\epsilon^2/\kappa)$ time to couple using the diffusion.
Next, following the Freidlin Wentzell point of view~\cites{FreidlinWentzell93,FreidlinWentzell94,IyerNovikov16,HairerIyerEA18} we view the dynamics of~$X$ as a random walk on a lattice of $O(1/\epsilon^2)$ cells.
It is known that the expected coupling time of a discrete random walk on such lattice is bounded above by $O(1/\varepsilon^2)$, and that roughly takes one step of a random walk every time it crosses a \emph{boundary layer} of thickness $\epsilon \delta$.
Thus we can construct a successful coupling after the number of boundary layer crossings is at least $O(1/\epsilon^2)$.
Crossings of the boundary layer happen through the diffusion alone, and estimating their frequency yields Theorem~\ref{t:main}.
We carry out the details in Section~\ref{sec:main}, below.
Before delving into the details we make three remarks:
First, the extra logarithmic factor $\abs{\ln \delta}$ in~\eqref{e:tdisBound} arises due to the logarithmic slow down of Hamiltonian systems as they approach hyperbolic critical points (all cell corners, in our case).
Second, the smooth cutoff~$\xi$ in~\eqref{e:v} is used to initiate the coupling of the projected processes in a time that is independent of~$A$.
Third, the explicit formula for~$H$ in~\eqref{e:v} is used to construct a simple coupling in subsequent steps using symmetry.
While the logarithmic factor~$\abs{\ln \delta}$ is unavoidable, both the smooth cutoff~$\xi$ and the explicit formula for~$H$ are mainly used to simplify technicalities in the proof.
\subsection*{Plan of this paper}
In the next section (Section~\ref{sec:main}) we prove Theorem~\ref{t:main}, modulo several technical lemmas bounding certain hitting times.
In Section~\ref{s:tdismix} we prove Proposition~\ref{p:tMixTDis}, relating the dissipation time and mixing time for general incompressible flows.
In Section~\ref{s:cellCoupling} we prove an $O(\epsilon^2 / \kappa)$ bound on the coupling time when the process is projected to a torus of side length~$\epsilon$.
Finally in Section~\ref{s:phase2} we prove the remaining lemmas stated in Section~\ref{sec:main} by counting boundary layer crossings.
\iffalse
\bigskip\hrule\bigskip
\begingroup\color{red}
The position of microscopic tracer particles diffusing in an ambient fluid are governed by the SDE
Let $\rho(x,s; y, t)$ denote the transition density of the process~$X$.
Since $u$ is divergence free, the invariant measure of~$X$
The concentration of the dye
Consider the associated It\^o diffusion
on the torus $\T^2$, where $B$ is a standard, two dimensional Brownian motion.
Define its \emph{mixing time} by
\begin{equation}
t_\textit{mix}(\alpha) = \inf \left\{ t \geq 0 \mid \norm{\mu_{X_t} - \pi}_\textit{TV} \le \alpha \norm{\mu_{X_0} - \pi}_\textit{TV} \right\}.
\end{equation}
Then it has the following connection to the dissipation.
It is easy to see that the dissipation time is bounded by $3$ times the mixing time.
It turns out that the mixing time can also be estimated by the dissipation time, up to a logarithmic factor.
These are both a general results which are not specific to cellular flows and we present them in Appendix~\ref{s:tdismix}.
\begin{remark}\color{red}
Can alternately choose the normal cellular flow, but turn the flow off for the first $\epsilon^2/\kappa$ time units.
But need some care to make this work, as we need the processes completely coupled before turning the flow on.
\end{remark}
\begin{remark}[Choosing the cell size~$\varepsilon$ optimally]
By AM-GM inequality, when $\varepsilon^4 = \delta^2 \abs{\ln \delta}$, we have
\begin{equation}
t_\mix \le O \left( \frac{\abs{\ln \delta}^{3/2}}{\sqrt{\kappa A}} \right).
\end{equation}
As the map $t \mapsto t \abs{\ln (\kappa t)}$ monotonically decreases when $t$ is small, we have
\begin{align*}
t_{\!\diss} & \le O \left[ \frac{\abs{\ln \delta}^{3/2}}{\sqrt{\kappa A}} \abs{\ln \left( \kappa \cdot \frac{\abs{\ln \delta}^{3/2}}{\sqrt{\kappa A}} \right)} \right] \\
& \le O \left[ \frac{\abs{\ln \delta}^{3/2}}{\sqrt{\kappa A}} \abs{\ln (\delta \abs{\ln \delta}^{3/2})} \right] \\
& \le O \left( \frac{\abs{\ln \delta}^{5/2}}{\sqrt{\kappa A}} \right).
\end{align*}
\end{remark}
\medskip
\subsection*{Plan of this Paper}
\endgroup
\fi
\section{Proof of the Mixing Time Bound (Theorem~\ref{t:main})}\label{sec:main}
The goal of this section is to prove Theorem~\ref{t:main}.
In light of Proposition~\ref{p:tMixTDis}, we only need to bound the mixing time.
We will do this by a \emph{coupling construction}.
To fix notation, we will subsequently assume~$X$ and~$\tilde X$ are solutions of the SDEs
\begin{align}
\label{e:sdeX}
d X_t &= A v(X_t) dt + \sqrt{\kappa} \, d B_t \,,
\\
\label{e:sdeXtilde}
d \tilde{X}_t &= A v(\tilde{X}_t) dt + \sqrt{\kappa} \, d \tilde{B}_t \,.
\end{align}
with initial data
\begin{equation*}
X_0 = x\,,
\quad
\tilde X_0 = \tilde x
\qquad
\P^{(x, \tilde x)} \text{-almost surely}\,.
\end{equation*}
Here $B$ and $\tilde{B}_t$ are both 2D Brownian motions.
We will choose~$\tilde B$ in terms of $B$ in a manner that ensures a suitable bound on the \emph{coupling time}.
Recall the coupling time
\begin{equation*}
\tau_\cpl \defeq \inf \set{ t \geq 0 \st X_t = \tilde X_t }\,,
\end{equation*}
is the first time $X$ and $\tilde X$ meet, and standard results (see for instance~\cite[Ch.~5]{LevinPeresEA09}) guarantee
\begin{equation}\label{e:tMixTCouple}
t_\mix \leq C \sup_{(x, \tilde x) \in \T^2 \times \T^2} \E^{(x, \tilde x)} \tau_\cpl\,.
\end{equation}
Thus our task is now to choose the Brownian motion $\tilde B$ and bound $\E\tau_\cpl$.
The construction of~$\tilde B$ can be described quickly, however, the bound on~$\E\tau_\cpl$ requires several technical lemmas.
For clarity of presentation we will describe the construction of~$\tilde B$ below, and momentarily postpone the lemmas bounding the coupling time.
\begin{proof}[Proof of Theorem~\ref{t:main}]
The coupling construction is divided into several stages, which we describe individually.
\restartsteps
\step[Coupling projections]
Observe first that the drift~$v$ is periodic with period~$\epsilon$, and thus both~$X$ and~$\tilde X$ can be viewed as diffusions on a torus with side length~$\epsilon$.
Let $\T^2_\epsilon = [0, \epsilon)^2$ be the two dimensional torus with side length~$\epsilon$, and let $\Pi_\epsilon \colon \T^2 \to \T^2_\epsilon$ be the projection defined by
\begin{equation*}
\Pi_\epsilon(x_1, x_2) = \paren[\big]{ x_1 \pmod \epsilon,~x_2 \pmod \epsilon}\,.
\end{equation*}
We will subsequently assume that $1/\epsilon \in \N$, so the above projection is well defined.
(We also clarify that $\T^2$ above denotes the standard two dimensional torus with side length~$1$.)
Consider the projected diffusions
\begin{equation}\label{e:Ydef}
Y = \Pi_\epsilon X\,, \qquad \tilde Y = \Pi_\epsilon \tilde X\,,
\end{equation}
on the torus $\T^2_\epsilon$.
Since the drift~$v$ is divergence free one can use PDE methods to show that the mixing time of $Y$ is bounded by $O(\epsilon^2/ \kappa)$.
This, however, is not sufficient for our purposes as we need a coupling between~$Y$ and $\tilde Y$ for subsequent steps, and we need the coupling time to be bounded independent of~$A$.
We will couple~$Y$ and~$\tilde Y$ by waiting until they enter the central region of cells where $u = 0$.
In this region~$Y$ and~$\tilde Y$ are simply Brownian motions, and we can couple them by reflection (see for instance~\cite{LindvallRogers86}), in time~$\tau_\cpl^Y$ that is bounded independent of~$A$.
Explicitly, we will show (Lemma~\ref{l:cellCoupling}, below) that
\begin{equation}\label{e:ycouple}
\E^{(x, \tilde x)} \tau_\cpl^Y \leq \frac{C \epsilon^2}{\kappa} \,,
\end{equation}
for some finite constant $C$.
Here, and subsequently, we will assume that the constant $C$ is independent of the parameters $\epsilon$, $A$, $\kappa$, the initial data $x, \tilde x$, and may increase from line to line.
\step[Moving to vertical cell boundaries]
By the Markov property, we may now restart time and assume that at time $0$ we have $\Pi_\epsilon X_0 = \Pi_\epsilon \tilde X_0$.
In this step we will now choose $B = \tilde B$ and wait until $X$ and $\tilde X$ hit a vertical cell boundary.
That is, we set
\begin{equation}\label{e:sigmavDef}
\sigma_v \defeq \inf \set[\Big]{ t \geq 0 \st X^1_t \in \frac{\epsilon}{2} \Z}\,,
\qquad
\tilde{\sigma}_v \defeq \inf \set[\Big]{ t \geq 0 \st \tilde X^1_t \in \frac{\epsilon}{2} \Z}\,,
\end{equation}
where $X^1, \tilde X^1$ denote the first coordinates process of $X$, and $\tilde X$ respectively.
Periodicity of~$v$ will ensure~$\sigma_v = \tilde \sigma_v$, and we will show (Lemma~\ref{l:vhit}, below) that
\begin{equation}\label{e:vhit}
\E^x \sigma_v \leq \frac{C \epsilon^2 }{\kappa}
\end{equation}
\begin{figure}[htb]
\centering
\includegraphics[width=.7\textwidth]{images/coupling.pdf}
\caption{%
Sample trajectories illustrating the coupling in steps 2 and~3.
Here $X_0 = (0.75, 0.25)$, $\tilde X_0 = (0.25, 0.25)$, and the trajectory of~$X$ is shown in blue.
Until $\tilde X$ hits a vertical cell boundary the trajectory of~$\tilde X$ (shown in green) is simply a shift of the trajectory of~$X$.
After this time, the trajectory of~$\tilde X$ (shown in red) is a mirror image of the trajectory of~$X$ until they hit the same vertical line ($x = 0.5$ in this case).
}
\label{f:traject}
\end{figure}
\step[Vertical Coupling]
By the Markov property again, we restart time and assume $\Pi_\epsilon X_0 = \Pi_\epsilon \tilde X_0$, and $X^1_0, \tilde X^1_0 \in \frac{\epsilon}{2} \Z$.
We will now choose $\tilde B^1 = -B^1$ and $\tilde B^2 = B^2$, and wait until time $\tau_v$ defined by
\begin{equation}\label{e:tauvDef}
\tau_v \defeq \inf\set{t \geq 0 \st X^1_t = \tilde X^1_t}\,.
\end{equation}
Note that by symmetry of $v$ we will have%
\footnote{We clarify here that $(\Pi_\epsilon X_t)^2$ refers to the second coordinate of $\Pi_\epsilon X_t$.}
$(\Pi_\epsilon X_t)^2 = (\Pi_\epsilon \tilde X_t)^2$ for all $t \leq \tau_v$, and thus at time $\tau_v$ we will have~$\Pi_\epsilon X_t = \Pi_\epsilon \tilde X_t$.
(See Figure~\ref{f:traject}, below, for an illustration of trajectories of~$X$ and~$\tilde X$ under this choice of noise.)
We will show (Lemma~\ref{l:vcpl}, below) that
\begin{equation}\label{e:vcpl}
\E \tau_v \leq C \left( \dfrac{\varepsilon^2 }{\kappa} + \dfrac{\abs{\ln \delta}^2}{A \varepsilon^2} \right) \,.
\end{equation}
The proof of~\eqref{e:vcpl} requires a estimates on the number of times the flow crosses the boundary layer; this is technical, but has been well studied by numerous authors and the proofs can be readily adapted to our situation.
\step[Horizontal hitting and coupling]
At this point we have arranged for $\Pi_\epsilon X = \Pi_\epsilon \tilde X$, and $X^1 = \tilde X^1 \in \frac{\epsilon}{2} \Z$.
As usual, we restart time and assume that the above happens at time $0$.
We will now repeat steps 2 and 3 in the horizontal direction:
First choose $\tilde B = B$ until $X^2, \tilde X^2 \in \frac{\epsilon}{2} \Z$, then choose $\tilde B^1 = B^1$, $\tilde B^2 = - B^2$, and then wait until $X^2 = \tilde X^2$.
The time taken for each of these steps is bounded in Lemmas~\ref{l:hhit} and~\ref{l:hcpl}, below.
The symmetry of $v$ will ensure that when $X^2 = \tilde X^2$, we will also have $X^1 = \tilde X^1$, thus giving a successful coupling of $X, \tilde X$.
\smallskip
Using Chebychev's inequality, the above guarantees us a coupling of $X, \tilde X$ with probability at least $1/16$ in time at most twice the expected value of the stopping times in each of the above steps.
Thus using the Markov property and Lemmas~\ref{l:cellCoupling}--\ref{l:hcpl}, below, we obtain a successful coupling with the coupling time bounded by
\begin{equation}\label{e:couplingBound}
\E^{(x, \tilde x)} \tau_\textit{cpl} \leq C \paren[\Big]{ \dfrac{\varepsilon^2 }{\kappa} + \dfrac{\abs{\ln \delta}^2}{A \varepsilon^2} }\,.
\end{equation}
Using~\eqref{e:tMixTCouple}, concludes the proof.
\end{proof}
It remains to bound the stopping times in each of the above steps.
For clarity of presentation we state each bound as a Lemma below, and prove the lemmas in subsequent sections.
\begin{lemma}[Coupling of projections]\label{l:cellCoupling}
There exists a Brownian motion~$\tilde{B}$ such that $(Y, \tilde{Y})$ is a coupling of $Y$ (on the torus $\T^2_\epsilon$), and the coupling time satisfies~\eqref{e:ycouple}.
\end{lemma}
\begin{lemma}[Vertical boundary hitting time]\label{l:vhit}
Suppose $\Pi_\epsilon X_0 = \Pi_\epsilon \tilde X_0$.
Choose $\tilde B = B$, and let~$\sigma_v$ and $\tilde \sigma_v$ (equation~\eqref{e:sigmavDef}) be the first hitting times of $X$ and $\tilde X$ to the vertical cell boundaries, respectively.
Then $\sigma_v = \tilde \sigma_v$ and equation~\eqref{e:vhit} holds.
\end{lemma}
\begin{lemma}[Vertical coupling]\label{l:vcpl}
Suppose $\Pi_\epsilon X_0 = \Pi_\epsilon \tilde X_0$, and $X_0^1 \in \frac{\epsilon}{2} \Z$.
Let $\tilde B = (-B^1, B^2)$, and let~$\tau_v$ (equation~\eqref{e:tauvDef}) be the first time first time $X$ and $\tilde X$ are on the same vertical line.
Then~$(\Pi_\epsilon X_t)^2 = (\Pi_\epsilon \tilde X_t)^2$ for all $t \leq \tau_v$, the expected value of~$\tau_v$ is bounded by~\eqref{e:vcpl}.
\end{lemma}
\begin{lemma}[Horizontal boundary hitting time]\label{l:hhit}
Suppose that $\Pi_\epsilon X_0 = \Pi_\epsilon \tilde X_0$, and $X_0^1 = \tilde X^1_0 \in \frac{\epsilon}{2} \Z$.
Choose $\tilde B = B$, and let
\begin{equation*}
\sigma_h \defeq \inf \set[\Big]{ t \geq 0 \st X^2_t \in \frac{\epsilon}{2} \Z}\,,
\qquad
\tilde \sigma_h \defeq \inf \set[\Big]{ t \geq 0 \st \tilde X^2_t \in \frac{\epsilon}{2} \Z}\,,
\end{equation*}
be the first hitting time to the horizontal cell boundaries.
Then
\begin{equation*}
\sigma_h = \tilde \sigma_h\,,
\qquad
X^1_{\sigma_h} = \tilde X^1_{\tilde \sigma_h}\,,
\qquad\text{and}\qquad
\E \sigma_h
= \E \tilde \sigma_h
\leq \frac{C \epsilon^2}{\kappa}\,.
\end{equation*}
\end{lemma}
\begin{lemma}[Horizontal coupling]\label{l:hcpl}
Suppose $\Pi_\epsilon X_0 = \Pi_\epsilon \tilde X_0$, $X_0^1 = \tilde X^1_0$, and $X^2_0 \in \frac{\epsilon}{2} \Z$.
Choose $\tilde B = (B^1, -B^2)$ and let
\begin{equation*}
\tau_h = \inf \set[\Big]{t \geq 0 \st X^2_t = \tilde X^2_t} \,,
\end{equation*}
be the first time $X$ and $\tilde X$ are on the same horizontal line.
Then
\begin{equation*}
\tau_h = \tilde \tau_h\,,
\qquad
X_{\tau_h} = \tilde X_{\tau_h}\,,
\qquad\text{and}\qquad
\E \tau_h \le \dfrac{C \abs{\ln \delta}^2}{A \varepsilon^2} \,.
\end{equation*}
\end{lemma}
Each of these lemmas will be proved in subsequent sections.
\section{Relationship between the dissipation time and mixing time}\label{s:tdismix}
In this section we prove Proposition~\ref{p:tMixTDis} which relates the mixing time and the dissipation time of general incompressible flows.
Throughout this section we will assume~$u \in L^\infty( [0, \infty); W^{1, \infty}(\T^d) )$ is a divergence free vector field, and let~$X$ be the (time inhomogeneous) Markov process on $\T^d$ defined by the SDE
\begin{equation}\label{e:sdeXu}
dX_t = u(X_t) \, dt + \sqrt{\kappa} \, dB_t\,.
\end{equation}
Here $B$ is a standard $d$-dimensional Brownian motion on the torus.
Let $\rho(x, s; y, t)$ be the transition density of~$X$.
By the Kolmogorov equations, we know that~$\rho$ is the fundamental solution to~\eqref{e:phiEq}, and thus the mixing time of~$X$ is given by~\eqref{e:tmixDef}.
Using the Kolmogorov equations again, the dissipation time (defined in~\eqref{e:tdisDefPDE}) can be equivalently defined by
\begin{equation}\label{e:tdisDef}
t_{\!\diss} \defeq \inf\set[\Big]{ t \geq 0 \st \sup_{x \in \T^d,\; s \geq 0} \norm{ \E^{(\cdot,s)} \vartheta(X_{s+t}) }_{L^2} < \frac{1}{2} \norm{\vartheta}_{L^2} \,,~\forall \vartheta\in \dot L^2(\T^d) }\,.
\end{equation}
Recall $\dot L^2(\T^d)$ is the set of all mean zero $L^2$ functions on the torus $\T^d$, and $\E^{(x,s)}$ denotes the expected value under the probability measure $\P^{(x,s)}$ under which $X_s = x$ almost surely.
The constant~$1/2$ in~\eqref{e:tdisDefPDE}, \eqref{e:tmixDef} and~\eqref{e:tdisDef} is chosen for convenience.
Replacing it by any constant that is strictly smaller than~$1$ will only change $t_\mix$ and~$t_{\!\diss}$ by a constant factor that is independent of~$u$, $\kappa$ and $d$.
The first inequality in~\eqref{e:tmixDef} can be proved elementarily, and we do that first.
\begin{lemma}\label{l:disMix}
The dissipation time and mixing time satisfy the inequality
\begin{equation*}
t_{\!\diss}
\leq 3 t_\mix\,.
\end{equation*}
\end{lemma}
\begin{proof}
For simplicity, and without loss of generality, we will assume that $s = 0$ in both~\eqref{e:tmixDef} and~\eqref{e:tdisDef}.
Let $\theta_t(x) \defeq \E^{(x,0)} \theta_0(X_t) = \E^x \theta_0(X_t)$ for some $\theta_0$ in $\dot L^2$.
Since $\theta_0$ is mean~$0$, we note
\begin{equation*}
\theta_t(x)
= \int_{\T^d} \rho(x, 0; y, t) \theta_0(y) \, dy
= \int_{\T^d} (\rho(x, 0; y, t) - 1) \theta_0(y) \, dy \,.
\end{equation*}
and hence
\begin{align}\label{e:thetaL21}
\nonumber
\norm{\theta_t}_{L^2}^2
&\leq \paren[\Big]{
\int_{\T^d \times \T^d} \abs{ \rho(x, 0; y, t) - 1 } \, dy \, dx
}
\cdot
\\
&\qquad\qquad
\paren[\Big]{
\int_{\T^d \times \T^d } \theta_0(y)^2 (\rho(x, 0; y, t) + 1) \, dx \, dy
}\,.
\end{align}
Since the Lebesgue measure is invariant, we note
\begin{equation*}
\int_{\T^d} \rho(x, 0; y, t) \, dx = 1\,,
\qquad\text{for every }y \in \T^d\,,
\end{equation*}
and hence~\eqref{e:thetaL21} implies
\begin{equation}\label{e:thetaL2}
\norm{\theta_t}_{L^2}^2 \leq 2 \norm{\theta_0}_{L^2}^2 \paren[\Big]{
\sup_{x \in \T^d} \int_{\T^d} \abs{ \rho(x, 0; y, t) - 1 } \, dy
}\,.
\end{equation}
The Chapman Kolmogorov equations and invariance of the Lebesgue measure immediately imply
\begin{equation}\label{e:ntmix}
\sup_x \int_{\T^d} \abs{\rho(x, 0; y, n t_\mix ) - 1} \leq \frac{1}{2^n}\,,
\end{equation}
for any natural number~$n \in \N$.
Choosing $t = 3 t_\mix$ and using~\eqref{e:thetaL2}, we see that~\eqref{e:ntmix} immediately implies $\norm{\theta_{3t_\mix}}_{L^2}^2 \leq \frac{1}{4} \norm{\theta_0}_{L^2}^2$, which finishes the proof.
\end{proof}
The proof of the second inequality in~\eqref{e:disMix} follows from Proposition~4.2 in~\cite{IyerXuEA21}, which provides an~$L^1$ to $L^\infty$.
We reproduce this here for convenience, and then go on to prove the second inequality in~\eqref{e:disMix}.
\iffalse
\begin{proposition}\label{p:tStarPQ}
For any $p, q \in [1, \infty]$ there exists a constant $C = C(d, p, q)$ such that for all sufficiently small $\kappa$ we have
\begin{equation*}
t^*_p \leq C_{p,q} t^*_q \ln \paren[\Big]{ 1 + \frac{1}{\kappa t^*_q} }\,.
\end{equation*}
\end{proposition}
\fi
\begin{lemma}[Proposition 4.2 in~\cite{IyerXuEA21}]\label{l:L1Linf}
There exists a constant $C = C(d)$, independent of $u$, such that for all $\vartheta \in \dot L^2$, and all sufficiently small~$\kappa > 0$ we have
\begin{equation}\label{e:L1Linf}
\norm{\E^{(x, s)} \vartheta(X_{s+t}) }_{L^\infty} \leq \frac{1}{2} \norm{\vartheta}_{L^1}\,,
\quad\text{for all }
s \geq 0\,,~
t \geq C t_{\!\diss} \ln \paren[\Big]{1 + \frac{1}{\kappa t_{\!\diss}} } \,.
\end{equation}
\end{lemma}
\begin{remark}\label{r:L1LinfN}
Since $\norm{\vartheta}_{L^1} \leq \norm{\vartheta}_{L^\infty}$, we can iterate~\eqref{e:L1Linf} to yield
\begin{equation*
\norm{\E^{(x, s)} \vartheta(X_{s+t}) }_{L^\infty} \leq 2^{-n} \norm{\vartheta}_{L^1}\,,
\quad\text{for all }
s \geq 0\,,~
t \geq n C t_{\!\diss} \ln \paren[\Big]{1 + \frac{1}{\kappa t_{\!\diss}} } \,.
\end{equation*}
\end{remark}
\begin{proof}
For simplicity, and without loss of generality we assume $s = 0$.
Using well known drift independent estimates (see for instance Lemma~5.6 in~\cite{ConstantinKiselevEA08}, Lemmas 3.1, 3.3 in~\cite{FannjiangKiselevEA06}, and Lemma 5.4 in~\cite{Zlatos10}) we know
\begin{equation*}
\norm{\theta_{t + 2t_{\!\diss}}}_{L^\infty}
\leq \frac{c}{(\kappa t_{\!\diss})^{\frac{d}{4}}} \norm{\theta_{t + t_{\!\diss}}}_{L^2} \,,
\qquad\text{and}\qquad
\norm{\theta_{t_{\!\diss}}}_{L^2} \leq \frac{c}{(\kappa t_{\!\diss})^{\frac{d}{4}} } \norm{\theta_0}_{L^1}\,,
\end{equation*}
for some dimensional constant $c = c(d)$.
Now iterating~\eqref{e:tdisDef} we see
\begin{equation*}
\norm{\theta_{t + t_{\!\diss}}}_{L^2}
\leq 2^{- \floor{t/t_{\!\diss}} } \norm{\theta_{t_{\!\diss}}}_{L^2}
\leq 2^{1 - t/t_{\!\diss} } \norm{\theta_{t_{\!\diss}}}_{L^2} \,,
\end{equation*}
and hence
\begin{equation*}
\norm{\theta_{t + 2t^*_p}}_{L^\infty}
\leq \frac{c}{(\kappa t_{\!\diss})^{\frac{d}{4}}} \norm{\theta_{t + t_{\!\diss}}}_{L^2}
\leq \frac{c 2^{1 - t/t_{\!\diss}}}{(\kappa t_{\!\diss})^{\frac{d}{4}}} \norm{\theta_{t_{\!\diss}}}_{L^2}
\leq \frac{c^2 2^{1 - t/t_{\!\diss}}}{(\kappa t_{\!\diss})^{\frac{d}{2}}} \norm{\theta_0}_{L^1}
\end{equation*}
Thus if we choose $t \geq C t_{\!\diss} \ln ( 1 + 1/(\kappa t_{\!\diss}) )$ for some sufficiently large constant $C = C(p, d)$, we obtain
\begin{equation*}
\norm{\theta_{t + 2t_{\!\diss}}}_{L^\infty} \leq \frac{1}{2} \norm{\theta_0}_{L^1}\,.
\end{equation*}
This finishes the proof of Lemma~\ref{l:L1Linf}.
\end{proof}
We can now prove the second inequality in~\eqref{e:disMix}.
\begin{lemma}\label{l:mixDis}
There exists a dimensional constant $C = C(d)$, independent of~$u$ and~$\kappa$ such that
\begin{equation*}
t_\mix
\leq C t_{\!\diss} \ln \paren[\Big]{ 1 + \frac{1}{\kappa t_{\!\diss}} }\,.
\end{equation*}
\end{lemma}
\begin{proof}
For simplicity, and without loss of generality we assume $s = 0$.
Choose large enough so that~\eqref{e:L1Linf} holds.
By standard regularity theory, we know that for any $\epsilon > 0$, the density $\rho(x, 0; y, \epsilon)$ is integrable in~$y$.
Since $\rho \geq 0$, we note
\begin{equation*}
\int_{\T^d} \abs{ \rho(x, 0; y, \epsilon) - 1 } \, dy
\leq \int_{\T^d} \paren{ \rho(x, 0; y, \epsilon) + 1 } \, dy
= 2\,.
\end{equation*}
Let $C$ be the constant from~\eqref{e:L1Linf} and choose
\begin{equation*}
t = 2 C t_{\!\diss} \ln \paren[\Big]{ 1 + \frac{1}{\kappa t_{\!\diss}} } \,.
\end{equation*}
Iterating Lemma~\ref{l:L1Linf} (as in Remark~\ref{r:L1LinfN}), we note that for every $x \in \T^d$
\begin{align*}
\norm{\rho(x, 0; y, 2t + \epsilon) - 1}_{L^1(y)}
&\leq \norm{\rho(x, 0; y, 2t + \epsilon) - 1}_{L^\infty(y)}
\\
&\leq \frac{1}{4} \norm{\rho(x, 0; y, \epsilon) - 1}_{L^1(y)}
\leq \frac{1}{2} \,.
\end{align*}
This shows $t_\mix \leq t$, finishing the proof.
\end{proof}
The proof of Proposition~\ref{p:tMixTDis} follows immediately from Lemmas~\ref{l:disMix} and~\ref{l:mixDis}.
\section{Coupling of Projections (Proof of Lemma~\ref{l:cellCoupling})}\label{s:cellCoupling}
In this section we prove Lemma~\ref{l:cellCoupling} showing that the projected processes $Y$, $\tilde Y$ (defined in~\eqref{e:Ydef}) will couple in time $O(\epsilon^2 / \kappa)$ in expectation.
Coupling of diffusions have been studied by many authors, dating back to Lindvall and Rogers~\cite{LindvallRogers86}.
In their original work, Lindvall and Rogers~\cite{LindvallRogers86} provide a method to couple diffusions in $\R^d$ by ``reflecting'' the noise.
Unfortunately, if we use their methods directly the bound we obtain on the coupling time will depend on the Lipschitz constant of the drift; in our case, this is $O(A/\varepsilon)$ which is unbounded.
It is for this reason that we modify the cellular flows using the cutoff function~$\xi$.
With the cutoff, we have a central region in each square where there is no drift.
Once $Y, \tilde Y$ enter this region, they can be successfully coupled by reflection.
To carry out the details of the above plan, define
\begin{equation*}
Q = [0, \epsilon/2]^2\,,
\quad
U = Q \cap \set{H > 1/2}\,,
\quad
U' = Q \cap \set{H > h_0}\,,
\end{equation*}
for some ${h_0 \in (3/4, 1)}$ that is independent of $\epsilon$, $A$, $\kappa$ and will be chosen shortly.
We will run $Y$ and $\tilde Y$ independently until they both enter~$U'$, and then reflect the noise until they couple.
To estimate the time taken by each of these steps we use the following results.
\begin{lemma}\label{l:crudeMixTimeBd}
Let $u \in L^\infty( [0, \infty); W^{1, \infty}(\T^d)$ be a general (not necessarily cellular) divergence free drift, and consider the SDE~\eqref{e:sdeXu} on the $d$-dimensional torus $\T^d$.
The mixing time of~$X$ is bounded by
\begin{equation}\label{e:tmixX}
t_\mix(X) \leq \frac{C}{\kappa}\,,
\end{equation}
for some dimensional constant~$C = C(d)$ that is independent of~$u$ and~$\kappa$.
\end{lemma}
\begin{remark}\label{r:crudeMixTimeBdRescaled}
By rescaling, on a torus with side length~$\ell$, the bound~\eqref{e:tmixX} becomes
\begin{gather}\label{e:tmixXprime}
\tag{\ref*{e:tmixX}$'$}
t_\mix(X) \leq \frac{C \ell^2}{\kappa}\,,
\end{gather}
for some dimensional constant~$C$ that is independent of~$\ell$,~$u$ and~$\kappa$.
\end{remark}
\begin{remark}
We believe that in this generality there exists a coupling for which $\E \tau_\cpl \leq C / \kappa$, however we are presently unable to produce such a coupling.
\end{remark}
\begin{lemma}\label{l:mix1cell}
Let $\tilde B$ be a Brownian motion that is independent of~$B$.
There exists a time $t_1 \leq O(\varepsilon^2 /\kappa )$ such that for all $t \ge t_1$, we have
\begin{equation}\label{e:mix1cell}
\inf_{y, \tilde y \in \T^2_\varepsilon} \P^{(y, \tilde y)} \paren{Y_t,~ \tilde Y_t \in U'} \ge \frac{\abs{U'}^2}{4\epsilon^4}.
\end{equation}
Here~$\abs{U'} = \leb(U')$ denotes the Lebesgue measure of~$U'$.
\end{lemma}
\begin{lemma}\label{l:coupleBeforeExit}
Choose the Brownian motion~$\tilde B$ to be the Brownian motion~$B$ reflected about the line perpendicular to $y - \tilde y$.
Explicitly, choose $\tilde B = M B$, where
\begin{equation*}
M = I - 2 \hat n \hat n^T\,,
\qquad\text{and}\qquad
\hat n = \frac{y - \tilde y}{\abs{y - \tilde y}}\,.
\end{equation*}
There exists a time $t_2 \leq O(\epsilon^2/\kappa)$ and a constant $c > 0$ such that
\begin{equation*}
\inf_{y, \tilde y \in U'} \P^{(y, \tilde y)} \paren{ \tau_\cpl^Y \leq t_2 } \geq c \,.
\end{equation*}
\end{lemma}
Momentarily postponing the proofs of Lemmas~\ref{l:mix1cell} and~\ref{l:coupleBeforeExit}, we prove Lemma~\ref{l:cellCoupling}.
\begin{proof}[Proof of Lemma~\ref{l:cellCoupling}]
Choose~$t_1$ according to Lemma~\ref{l:mix1cell} and run $Y$ and $\tilde Y$ independently until time $t_1$.
Lemma~\ref{l:mix1cell} guarantees that at time~$t_1$ we have~\eqref{e:mix1cell}.
(Note that $\abs{U'} = O(\varepsilon^2)$, and so $\frac{\abs{U'}^2}{4 \varepsilon^4} = O(1)$.)
Now choose~$t_2$ and~$\tilde B$ according to Lemma~\ref{l:coupleBeforeExit}.
This construction will guarantee
\begin{equation*}
\inf_{y, \tilde y \in U'} \P^{(y, \tilde y)} \paren{ \tau_\cpl^Y \geq (t_1 + t_2) } \le 1 - c'\,,
\quad\text{where}\quad
c' \defeq \frac{c \abs{U'}^2}{4 \epsilon^4} > 0\,,
\end{equation*}
and~$c$ is the constant in Lemma~\ref{l:coupleBeforeExit}.
In the event that $\tau_\cpl^Y > t_1 + t_2$, we simply repeat the above two steps.
The Markov property will guarantee
\begin{equation*}
\inf_{y, \tilde y \in U'} \P^{(y, \tilde y)} (\tau_\cpl^Y \geq n (t_1 + t_2) ) \le (1-c')^n\,.
\end{equation*}
Thus, for any $y, \tilde y \in U'$ we see
\begin{align*}
\E^{(y, \tilde y)} \tau_\cpl^Y
&= \int_0^\infty \P^{(y, \tilde y)} ( \tau_\cpl^Y \geq t ) \, dt
\leq (t_1 + t_2) \sum_{n = 0}^\infty \P^{(y, \tilde y)} ( \tau_\cpl^Y \geq n(t_1 + t_2) )
\\
&\leq (t_1 + t_2) \sum_{n = 0}^\infty (1 - c')^n
\leq \frac{t_1 + t_2}{c'}\,,
\end{align*}
concluding the proof.
\end{proof}
It remains to prove Lemmas~\ref{l:crudeMixTimeBd}, \ref{l:mix1cell} and~\ref{l:coupleBeforeExit}.
\begin{proof}[Proof of Lemma~\ref{l:crudeMixTimeBd}]
Let~$\phi$ solve~\eqref{e:phiEq}.
Multiplying by~$\phi$, using the fact that $\dv u = 0$ and the Poincar\'e inequality shows
\begin{equation*}
\norm{\phi(s + t)}_{L^2} \leq e^{\lambda_1 \kappa t} \norm{\phi(s)}_{L^2}\,,
\end{equation*}
where~$\lambda_1$ is the first non-zero eigenvalue of the Laplacian on $\T^d$.
This immediately implies~$t_{\!\diss}(X) \leq 1 / (\lambda_1 \kappa)$, and using Proposition~\ref{p:tMixTDis} concludes the proof.
\end{proof}
\begin{proof}[Proof of Lemma \ref{l:mix1cell}]
Using Lemma~\ref{l:crudeMixTimeBd} and rescaling (see Remark~\ref{r:crudeMixTimeBdRescaled}), we see
\begin{equation*}
t_\mix(Y) \leq \frac{C \epsilon^2}{\kappa}\,.
\end{equation*}
Now choose $N = \log_2\paren{ 2\epsilon^2/\abs{U'} }$, and note that $N$ is independent of~$\epsilon$.
For every $t \geq N t_\mix(Y)$ and $y \in \T^2_\epsilon$, we have
\begin{equation*}
\abs[\Big]{
\P^y (Y_t \in U')
- \frac{\abs{U'}}{\epsilon^2}
}
\leq 2^{-N} \frac{\abs{U'}}{2\epsilon^2}
\quad\text{and hence}\quad
\P^y (Y_t \in U')
\geq \frac{\abs{U'}}{2\epsilon^2}\,,
\end{equation*}
Since $\tilde Y$ is independent of~$Y$ and satisfies the same bound we obtain~\eqref{e:mix1cell} as claimed.
\end{proof}
\begin{proof}[Proof of Lemma \ref{l:coupleBeforeExit}]
Notice that as long as $Y$, $\tilde Y$ remain in~$U$, they are simply rescaled standard Brownian motions.
Let $\ell_b$ be perpendicular bisector of the line segment joining~$y$ and $\tilde y$, and $\tau_\ell$ be the hitting time of~$Y$ to $\ell_b$.
The choice of~$\tilde B$ ensures that if $Y, \tilde Y$ remain inside~$U$ until they $\ell_b$, then they couple at time~$\tau_\ell$.
In order to estimate the hitting time to~$\ell_b$ before exiting~$U$, let $R = \abs{y - \tilde y} / 2$ be the distance of $y$ to $\ell_b$.
Let $K$ be the square with center $y$, side length $2R$, and one pair of sides parallel to $\ell_b$.
Note that if $h_0$ is sufficiently closed to $1$, this square lies entirely in $U$.
Let $\tau_K$ be the exit time of $Y$ from~$K$, and note that
\begin{equation*}
\P^{y, \tilde y} \paren{ \tau_\cpl^Y \leq t }
\geq \P^y \paren{ \tau_K \leq t, ~ Y_{\tau_K} \in \ell_b }
= \frac{1}{4} \P^y \paren{ \tau_K \leq t } \,.
\end{equation*}
The last equality followed by symmetry, as at time $\tau_K$ it is equally likely that $Y_{\tau_K}$ belongs to any of the four sides of~$K$.
The last term on the right can be bounded by Chebyshev's inequality, and the fact that the expected exit time of Brownian motion from a square is known.
Namely,
\begin{equation*}
\P^y(\tau_K \leq t)
\geq 1 - \frac{\E^y \tau_K}{t}
\geq 1 - \frac{R^2}{\kappa t}
\geq 1 - \frac{\epsilon^2}{\kappa t}
\geq \frac{1}{2}\,,
\end{equation*}
provided $t \geq 2 \epsilon^2 / \kappa$.
Choosing $t_2 = 2 \epsilon^2 / \kappa$ concludes the proof.
\end{proof}
\iffalse
The following auxilliary lemma deals with the hitting time in the proof of Lemma~\ref{l:coupleBeforeExit}.
\begin{lemma}
\label{l:BMinSquare}
Consider a 2-D coordinate system. Four points $(-r, 0), \, (-r, 2r), \, (r, 2r)$, and $(r, 0)$ form a square, where $r > 0$.
Consider a scaled Brownian motion $Z_t$ given by
\begin{equation*}
dZ_t = \sqrt{\kappa} \, dB_t \quad Z_0 = (0, r) \,.
\end{equation*}
Let $\tau_K = \inf \{t \ge 0 \mid Z_t \notin K\}$. Then
\begin{equation}
\label{e:tauK}
\E \tau_K \le \frac{r^2}{2 \kappa} \,,
\end{equation}
and
\begin{equation}
\label{e:exitFromBisec}
\P [Z_{\tau_K} \in (-r, r) \times \{0\}] = \frac{1}{4} \,.
\end{equation}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{l:BMinSquare}]
If we look at a circle centered at $(0, r)$ with radius $\sqrt{2} r$, the expected exit time of $Z$ from that circle is bounded above by $\frac{(\sqrt{2} r)^2}{2 \kappa} = \frac{r^2}{\kappa}$ (see~\cite{Oksendal03}).
Since the square $K$ is totally contained in that circle, the exit time $\tau_K$ should be bounded by $\frac{r^2}{\kappa}$ as well, which proves (\ref{e:tauK}).
Notice that $Z_0$ is the center of the square $K$. By symmetry of Brownian motion, $Z_{\tau_K}$ settles on each side with equal probability.
That gives (\ref{e:exitFromBisec}).
\end{proof}
\fi
\section{Synchronization and Reflection}\label{s:phase2}
In this section we prove Lemmas~\ref{l:vcpl}--\ref{l:hcpl}.
In order to prove these lemmas we will bounds on the boundary layer crossing time.
These have been studied previously by various authors (see for instance~\cite{Koralov04,FreidlinWentzell12,IyerNovikov16,HairerIyerEA18}), and the version we quote here can be obtained by a direct rescaling of those in~\cite{IyerNovikov16}.
Define the \emph{boundary layer} by $\mathcal{B}_\delta$ by
\begin{equation}\label{e:Bdelta}
\mathcal{B}_\delta \defeq \{\abs{H} < \delta\} \subset \T^2\,,
\end{equation}
where we recall from~\eqref{e:tdisBound} that~$\delta = \sqrt{\kappa / A}$.
The middle of the boundary layer is the level set~$\set{H = 0}$, and is known as the \emph{separatrix}.
We will now study repeated exits from the boundary layer, followed by returns to the separatrix.
Define the sequences of stopping times~$\sigma_n$ and~$\tau_n$ inductively by starting with $\tau_0 = 0$.
For $n \geq 1$, define
\begin{align}
\label{def:sigma-n}
\sigma_n &= \inf \{t \geq \tau_{n-1} \st X_t \notin \mathcal{B}_\delta\}
\\
\label{def:tau-n}
\tau_n &= \inf \{t \geq \sigma_n \st H(X_t) = 0\}\,.
\end{align}
That is,~$\sigma_n$ is the first exit from the boundary layer~$\mathcal{B}_\delta$ after time~$\tau_n$, and $\tau_n$ is the first return to the separatrix after time~$\sigma_n$.
At time~$\tau_n$ we must have either~$X^1 \in \frac{\epsilon}{2} \Z$, or~$X^2 \in \frac{\epsilon}{2} \Z$.
We now separate the times when $X^1 \in \frac{\epsilon}{2} \Z$, and when~$X^2 \in \frac{\epsilon}{2} \Z$.
Given~$i \in \set{1, 2}$, let~$\tau^i_0 = 0$ and inductively define
\begin{equation*}
\tau_n^i
= \inf \set[\Big]{\tau_k > \tau_{n-1}^i \st X^i_{\tau_k} \in \frac{\varepsilon}{2} \Z }\,.
\end{equation*}
We claim that up to a logarithmic factor, the chance that $\tau_n^i \leq t$ is comparable to the number of crossings of a standard Brownian motion over an interval of size~$\epsilon / \sqrt{A}$.
This is the first lemma we state.
\begin{lemma}\label{l:nthhits}
There exists a constant $c > 0$ such that, for $n \in \N$, $i \in \{1, 2\}$, we have
\begin{equation}
\label{eq:nthhits}
\inf_{\abs{H(x)} < \delta} \P^x \paren{\tau^i_n \le t} \geq \left( 1 - \dfrac{c n \varepsilon \delta \abs{\ln \delta}}{\sqrt{\kappa t}} \right)^+.
\end{equation}
\end{lemma}
This lemma is simply a rescaling of Lemma~2.2 in~\cite{IyerNovikov16}, and we refer the reader there for the proof.
We remark, however, that the proof in~\cite{IyerNovikov16} uses PDE techniques from~\cite{ChildressSoward89,FannjiangPapanicolaou94,NovikovPapanicolaouEA05,IyerKomorowskiEA14}.
Lemma~\ref{l:nthhits} can also be proved directly using probabilistic techniques, and we refer the reader to~\cite{FreidlinWentzell12,Koralov04,DolgopyatKoralov08,HairerIyerEA18} for related crossing estimates.
In order to apply Lemma~\ref{l:nthhits}, we need the process~$X$ to enter the boundary layer~$\mathcal{B}_\delta$.
This happens in time at most $O(\epsilon^2 \abs{\ln \delta} / \kappa )$, and is the content of our next lemma.
\begin{lemma}\label{l:exit-int}
Let $\sigma_e$ be the first hitting time of $X$ to the level set $\set{H = \delta}$ (i.e.\ $\sigma_e = \inf \{t \geq 0 \mid H(X_t) = \delta\}$).
Then
\begin{equation}
\sup_{x \in \T^2} \E^{x} \sigma_e \le \frac{C \varepsilon^2}{\kappa} \,.
\end{equation}
\end{lemma}
\begin{proof}
We first project to the torus of side length~$\epsilon$, and note that $\sigma_e = \inf\set{ t \geq 0 \st H(Y_t) = \delta }$.
Note $\set{ H > \delta }$ contains two connected components, each occupying an area of at most $1/4$ of the torus $\T^2_\epsilon$.
For any $x \in \T^2_\epsilon$, let $U$ be the connected component of $\set{H > \delta}$ that contains~$x$.
Thus, for any $t \geq t_\mix(Y)$ we know
\begin{equation*}
\abs[\Big]{ \P^{x} (X_t \in U^c) - \frac{\abs{U^c}}{\epsilon^2} } \leq \frac{1}{4} \,,
\quad\text{and hence}\quad
\P^{x_0} ( X_t \in U) \geq \frac{1}{2}\,.
\end{equation*}
By continuity of trajectories we note that the event $\set{\sigma_e \le t} \supseteq \set{ X_t \in U^c}$, and so $\P^{x_0} ( \sigma_e < t ) \geq 1/2$.
In the event that $\sigma_e > t$, we use the Markov property and repeat the above argument to yield
\begin{equation*}
\E^{x} \sigma_e \leq 2 t_\mix(Y)\,.
\end{equation*}
By~\eqref{e:tmixXprime} with $\ell = \epsilon$ we know $t_\mix(Y) \leq C \epsilon^2 / \kappa$, concluding the proof.
\end{proof}
\subsection{Proofs of the hitting time estimates (Lemmas~\ref{l:vhit} and~\ref{l:hhit})}\label{sb:hit}
Then we may estimate the first hit at vertical boundary lines.
\begin{proof}[Proof of Lemma~\ref{l:vhit}]
Notice that, periodicity of~$v$ and the synchronous choice~$\tilde B = B$, implies $\sigma_v = \tilde \sigma_v$.
Thus we only have to prove~\eqref{e:vhit}.
Without loss of generality assume $(\Pi_\epsilon X_0)^1 \in [0, \epsilon/2)$.
(We clarify that $(\Pi_\epsilon X_0)^1$ refers to the first coordinate of $\Pi_\epsilon X_0$.)
If $(\Pi_\epsilon X_0)^1 = 0$, then $\sigma_v = 0$, and there is nothing to prove, and thus we may assume $(\Pi_\epsilon X_0)^1 \in (0, \epsilon/2)$.
Let $V \subseteq \T^2_\epsilon$ be the set of all points $y$ such that $y_1 \in [\epsilon/2, \epsilon]$, and note that $V$ occupies half the area of $\T^2_\epsilon$.
Thus for any $t \geq 2t_\mix(Y)$ we see
\begin{equation*}
\abs[\Big]{ \P ( Y_t \in V ) - \frac{1}{2} } \leq \frac{1}{4}
\quad\text{and hence}\quad
\P ( Y_t \in V ) \geq \frac{1}{4}\,.
\end{equation*}
By continuity of trajectories, $\set{ \sigma_v \leq t } \supseteq \set{ Y_t \in V }$, and so $\P ( \sigma_v \le t ) \geq 1/4$.
If $\sigma_v > t$, then we use the Markov property and repeat the above argument to show
\begin{equation*}
\E \sigma_v \leq 8 t_\mix(Y) \leq \frac{C \epsilon^2 }{\kappa} \,,
\end{equation*}
as desired.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{l:hhit}]
The proof is identical to the proof of Lemma \ref{l:vhit}.
\end{proof}
\subsection{Coupling time estimates (Lemmas~\ref{l:vcpl} and~\ref{l:hcpl})}\label{sb:cpl}
We now turn our attention to Lemma~\ref{l:vcpl}.
Note first that by definition $v$ is~$\epsilon$ periodic and
\begin{align*}
v_1(-x_1, x_2) &= -v_1(x_1, x_2)\,,
&
v_2(-x_1, x_2) &= v_2(x_1, x_2)\,,
\\
v_1(x_1, -x_2) &= v_1(x_1, x_2)\,,
&
v_2(x_1, -x_2) &= -v_2(x_1, x_2)\,,
\\
v( x_1 + \tfrac{\epsilon}{2}, x_2 ) &= -v(x_1, x_2)\,,
&
v( x_1, x_2 + \tfrac{\epsilon}{2} ) &= -v(x_1, x_2)\,.
\end{align*}
As a result choosing~$\tilde B = (-B_1, B_2)$ and the assumptions $\Pi_\epsilon X_0 = \Pi_\epsilon \tilde X$ and $X_0^1, \tilde X_0^1 \in \frac{\varepsilon}{2} \Z$ imply
\begin{equation}\label{e:symmetry}
X_t^1 = -\tilde X_t^1 \pmod{\tfrac{\epsilon}{2}}\,,
\quad\text{and}\quad
X_t^2 = \tilde X_t^2 \pmod{\epsilon}\,.
\end{equation}
Let
\begin{equation*}
\ell_1 \defeq \set[\Big]{ \frac{X^1_0 + \tilde X^1_0}{2} }\times \T^1 \subseteq \T^2
\end{equation*}
be the vertical line half way between $X_0$ and $\tilde X_0$.
Note $(X^1_0 + \tilde X^1_0) / 2 \in \frac{\varepsilon}{2} \Z$ and so~$\ell_1$ is contained in the separatrix.
By~\eqref{e:symmetry} we see that~$\tau_v$ is exactly the hitting time of~$X$ to $\ell_1$.
Thus we may now ignore~$\tilde X$ and simply estimate the hitting time of~$X$ to~$\ell_1$.
Note that $X_{\tau^1_n} \in (\frac{\epsilon}{2} \Z) \times \R $ for all $n$ and behaves like a random walk on the collection of vertical lines~$(\frac{\epsilon}{2} \Z) \times \T^1 \subseteq \T^2$.
There are $2/\epsilon$ such vertical lines in the torus $\T^2$, and so we expect that after $O(1/\epsilon^2)$ steps of this random walk, $X_{\tau^1_n}$ will land in our desired line segment~$\ell_1$.
This is our next result.
\begin{lemma}\label{l:tauvN-hit-ell1}
Note that $\tau_v = \inf \{t \ge 0 \st X_{t} \in \ell_1\}$.
There exists $p_0 > 0$, and a constant $C_1$, independent of $A, \varepsilon, \kappa$, such that, for $n = C_1 / \varepsilon^2$, and $x \in \T^2$ such that $x_1 \in \frac{\varepsilon}{2} \Z$,
\begin{equation*}
\P^x \paren{ \tau_v \le \tau^1_n }
\geq p_0\,.
\end{equation*}
\end{lemma}
Postponing the proof of Lemma~\ref{l:tauvN-hit-ell1} to Section~\ref{s:pullback}, we prove Lemma~\ref{l:vcpl}.
\begin{proof}[Proof of Lemma~\ref{l:vcpl}]
As explained above, $\tau_v$ is the hitting time of~$X$ to the bisector~$\ell_1$.
Using Lemmas~\ref{l:nthhits} and~\ref{l:tauvN-hit-ell1} we see that
\begin{equation}
\label{e:froml6.1}
\P (\tau_n^1 \leq t_1) \geq 1 - \frac{p_0}{2}\,,
\quad
\text{where }
t_1 = \frac{4c^2 C_1^2 \abs{\ln \delta}^2 }{p_0^2 \epsilon^2 A}
\,,
\quad
n = \frac{C_1}{\epsilon^2}\,.
\end{equation}
Here $c$ is the constant from equation~\eqref{eq:nthhits}, and $p_0$,~$C_1$ are constants from Lemma~\ref{l:tauvN-hit-ell1}.
With Lemma~\ref{l:tauvN-hit-ell1}, we also see that
\begin{equation}
\label{e:froml6.3}
\P^x (\tau_v \le \tau^1_n) \ge p_0 \,.
\end{equation}
Combining~\eqref{e:froml6.1} and~\eqref{e:froml6.3} gives
\begin{equation*}
\P^x \paren[\big]{\tau_v \le \tau^1_n \leq t_1 } \geq \frac{p_0}{2} \,,
\end{equation*}
which implies $\P ( \tau_v \leq t_1 ) \geq \frac{p_0}{2}$.
Using the Markov property and iterating this implies
\begin{equation*}
\P (\tau_v > k t_1) \leq \paren[\Big]{1 - \frac{p_0}{2} }^k\,,
\quad\text{and hence }
\E \tau_v \leq \frac{2 t_1}{p_0}\,,
\end{equation*}
finishing the proof.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{l:hcpl}]
The proof is identical to the proof of Lemma \ref{l:vcpl}.
Note that at times when $X^2_t = \tilde X^2_t$, we actually have $X_t = \tilde X_t$ and hence $X_{\tau_h} = \tilde X_{\tau_h}$.
\end{proof}
\subsection{The hitting time to the bisector (Lemma~\ref{l:tauvN-hit-ell1})}\label{s:pullback}
In order to prove Lemma~\ref{l:tauvN-hit-ell1} we will lift trajectories of $X$ from the torus $\T^2$ to the covering space $\R^2$.
For clarity, we will denote the lifted process by $\hat X$.
Define the family of lines
\begin{equation*}
\hat \ell_1 = \set[\Big]{ x \in \R^2 \st x_1 = n + \frac{n_0\epsilon }{2}\,, n \in \Z }\,,
\end{equation*}
where $n_0 \in \Z$ is chosen such that
\begin{equation*}
\ell_1 = \set[\Big]{ x \in \T^2 \st x_1 = \frac{n_0 \epsilon }{2} }\,.
\end{equation*}
Note that the event of $X$ hitting $\ell_1$ on $\T^2$ is exactly the same as the event of $\hat X$ hitting $\hat \ell_1$ on $\R^2$.
Moreover, if $\hat X$ travels a horizontal distance of at least $1$, then it must pass through one of the lines in~$\hat \ell_1$.
We will use this to estimate $\P(\tau_v \leq \tau_n^1)$.
\begin{lemma}\label{l:pullback}
Suppose $\hat X$ satisfies the SDE~\eqref{eq:itodef} in~$\R^2$, with $\hat X_0 = \hat x \in \R^2$ such that $\hat x_1 = 0$.
There exist constants $C_1$, $p_0 > 0$, independent of $A, \varepsilon, \kappa$, such that, for $n = C_1/\varepsilon^2$ we have
\begin{equation}\label{e:XhatTaun}
\P^{\hat x} \paren{ \abs{\hat X_{\tau^1_n}^1} > 1 } \ge p_0 \,.
\end{equation}
\end{lemma}
\begin{proof}
Let $S_n = \hat X^1_{\tau^1_n}$, and observe that by symmetry of $v$ we must have $\E^{\hat x} S_n = 0$.
If $\E^{\hat x} S_n^2 \geq 1$ we note
\begin{equation*}
\paren[\Big]{ \E^{\hat x} S_n^2 - 1 }^2
\leq \paren[\Big]{ \E^{\hat x} S_n^2 \one_{\set{ \abs{ S_n} \geq 1 }} }^2
\leq \E^{\hat x} S_n^4 \P^{\hat x} \paren{ \abs{S_n} \geq 1 }
\end{equation*}
and hence
\begin{equation}\label{e:petrovBound}
\P^{\hat x} \paren{ \abs{S_n} \geq 1 } \ge \frac{(\E^{\hat x} S_n^2 - 1)^2}{\E^{\hat x} S_n^4}\,.
\end{equation}
whenever $\var(S_n) > 1$.
To use~\eqref{e:petrovBound}, we need to show~$\E^{\hat x} S_n^2 \geq 1$, and find a suitable upper bound for~$\E^{\hat x} S_n^4$.
For the first part we note~\cite{IyerNovikov16} shows that the variance of $S_n$ is comparable to that of a random walk with steps of size~$\epsilon$.
That is, we know
\begin{equation}\label{e:varBound}
\E^{\hat x} S_n^2 \ge c_1 n \varepsilon^2 \,,
\end{equation}
for some constant $c_1 > 0$, that is independent of $\epsilon$, $A$ and $\kappa$.
Thus choosing $n = 2 / (c_1 \varepsilon^2)$ will guarantee $\E^{\hat x} S_n^2 \geq 1$.
For the second part we need to find an upper bound for $\E S_n^4$.
For simplicity, let $\xi_m = \hat X^1_{\tau^1_{m+1}} - \hat X^1_{\tau^1_m}$, with $\tau^1_0 = 0$, so that $S_n = \xi_0 + \cdots + \xi_{n-1}$.
Notice
\begin{equation*}
\E^x S_n^4
= \sum_{m=0}^{n-1} \E^x \abs{\xi_m}^4
+ 6 \sum_{m' = 1}^{n-1} \sum_{m = 0}^{m'-1} \E^x \abs{\xi_m}^2 \abs{\xi_{m'}}^2 \,,
\end{equation*}
since the cross terms vanish by symmetry.
From Lemma~2.1 in~\cite{IyerNovikov16}, we know that
\begin{equation*}
\E^{\hat x} \abs{\xi_m}^2 \le c_2 \varepsilon^2 \,,
\end{equation*}
for some finite constant $c_2$ that is independent of~$\epsilon$, $A$ and $\kappa$.
The same proof (Section~5 in~\cite{IyerNovikov16}) also shows that
\begin{equation*}
\E^{\hat x} \abs{\xi_m}^4 \le c_2 \varepsilon^4 \,.
\end{equation*}
Moreover, for $m < m'$, by tower property,
\begin{equation*}
\E^{\hat x} \paren{ \abs{\xi_m}^2 \abs{\xi_{m'}}^2 }
= \E^{\hat x} \paren[\Big]{ \abs{\xi_m}^2 \E^{\hat X_{\tau^1_{m+1}}} \abs{\xi_{m'}}^2 }
\le \E^x \left[ \abs{\xi_m}^2 c_2 \varepsilon^2 \right] \le (c_2 \varepsilon^2)^2 \,.
\end{equation*}
Thus
\begin{equation}
\label{e:4mnt}
\E^{\hat x} S_n^4 \le c_2 n \varepsilon^4 + c_2^2 n^2 \varepsilon^4 \le c n^2 \varepsilon^4 \,,
\end{equation}
where $c = 2 c_2 (1 + c_2)$.
Combining~\eqref{e:petrovBound}, \eqref{e:varBound} and~\eqref{e:4mnt} we see
\begin{equation}
\P^{\hat x} \paren{ \abs{S_n} > 1 } \ge \frac{(c_1 n \varepsilon^2 - 1)^2}{c n^2 \varepsilon^4} \,.
\end{equation}
Choosing $n = C_1 / \varepsilon^2$ for some large constant $C_1$, we obtain~\eqref{e:XhatTaun} as desired.
\end{proof}
Using this, we prove Lemma~\ref{l:tauvN-hit-ell1}.
\begin{proof}[Proof of Lemma~\ref{l:tauvN-hit-ell1}]
Note that $x_1 \in \frac{\varepsilon}{2} \Z$.
Using symmetry and periodicity, we may, without loss of generality, assume $x_1 = 0$.
Lifting the process~$X$ to $\R^2$, we recall that when $\abs{\hat X^1_t} \geq 1$, the trajectory of $\hat X$ must have passed through one of the lines in~$\hat \ell_1$.
This implies
\begin{equation*}
\P^x\paren[\big]{ \tau_v \leq \tau^1_n }
\geq \P^{\hat x} \paren[\big]{ \abs{\hat X^1_{\tau_n^1}} \geq 1 }\,,
\end{equation*}
and applying Lemma~\ref{l:pullback} concludes the proof.
\iffalse
Now, pull back the whole setting to $\R^2$.
The diffusion starts on the vertical axis and satisfies the SDE~\eqref{eq:itodef} on $\R^2$.
By Lemma~\ref{l:pullback}, we know there exist constants $C_1$ and $p_0$ such that, for $n = C_1 / \varepsilon^2$,
\begin{equation}
\label{e:frompullback}
\P^x \left[ \abs{X^1_{\tau^1_n}} > 1 \right] \ge p_0 \,.
\end{equation}
When $\abs{X^1_{\tau^1_n}} > 1$, there exists some $t \le \tau^1_n$ such that $X_t^1 \in \{x_1 = a\} \cup \{x_1 = a - 1\}$, due to continuity of trajectories.
If we push forward to $\T^2$, this is exactly equivalent to $X_t^1 \in \{x_1 = a\}$, {i.e.}, $X_t \in \ell_1$.
So such $t$ corresponds to $\tau_v$ on this event.
This along with~\eqref{e:frompullback} implies
\begin{equation*}
\P^x \big[\tau_v \le \tau^1_n\big] \ge p_0 \,.
\end{equation*}
\fi
\end{proof}
\bibliographystyle{halpha-abbrv}
|
\section{Introduction}
\subsection{Related works}
Avoidance games belong to the class of positional games, that were introduced by Hales and Jewett in 1963~\cite{Hales1963} and popularised by Erd\H os and Selfridge in 1973~\cite{Erdos1973}. In this class of games, the board is a hypergraph and two players alternately select a vertex of the hypergraph that has not been claimed before. Winning conditions depend on the convention and are related to the hyperedges.
{\sc Tic-Tac-Toe} and {\sc Hex} are two famous examples of positional games. To learn more about positional games, we refer the reader to the recent survey of Hefetz {\em et al.}~\cite{Hefetz2014}.
Among positional games, a natural dichotomy exists: on the one hand, there are games in which players seek to build a structure, and on the other hand, there are games in which players want to avoid a structure. The former set contains both Maker-Maker and Maker-Breaker conventions, in which the hyperedges are winning sets, and the player either want to fill a winning set (Maker role), or to play at least once in each of them (Breaker role). The latter contains Avoider-Avoider and Avoider-Enforcer conventions, that can be seen as the misère version of the former. In these games, the hyperedges are losing sets, and the players either want not to fill one losing set (Avoider role), or to force their opponent to fill one of them (Enforcer role).
When positional games were introduced, the focus was on Maker-Breaker games, {\em i.e.} games in which one player, Maker, aims to fill a hyperedge, and the second one, Breaker, wants to prevent it by claiming at least one vertex in each hyperedge. This convention is the most popular, and several games were studied according to this convention. In particular, the survey of Beck~\cite{Beck2008} presents several results obtained for Maker-Breaker games. The field of Maker-Breaker games is still well investigated today, and some Maker-Breaker games were introduced recently~\cite{Duchene2020, nenadov2016}.
The first Avoider-Avoider game was introduced in 1968 with the game of SIM and is presented in~\cite{simmons1968}, but the first study of the complexity of Avoidance games was done by Schaefer in 1978~\cite{Schaefer1978}. Avoider-Enforcer games were introduced later by Lu in 1991~\cite{Lu1991,Lu1995} under the name of Antimaker-Antibreaker games and corresponds to the misère version of Maker-Breaker games. The standard name of this convention, Avoider-Enforcer, was popularised by Hefetz and different co-authors in 2007~\cite{hefetz2007-4, hefetz2007-2, hefetz2007-1, hefetz2007-3}. In this game, Enforcer wins if at some point during the game, Avoider has played all the vertices of a hyperedge, otherwise Avoider wins.
Even if most of the studies of positional games are focused on Maker-Breaker games, Avoider-Enforcer games have become more and more relevant: the famous Ramsey game were introduced in Avoider-Enforcer convention by Beck in 2002~\cite{beck2002} as a generalisation of SIM. As it was done in the Maker-Breaker convention, some games on graphs were introduced in Avoider-Enforcer or Avoider-Avoider conventions, where the loosing sets correspond to some structure in the graph, see~\cite{anuradha2008, barat2009, Grzesik2013, hefetz2010}.
In terms of complexity, an overview of the field is proposed by Demaine~\cite{demaine2001}. In positional games, as they are perfect information games, one player always has a winning strategy (or both players can ensure a draw). The natural decision problem related to games is thus: does the first player have a winning strategy ?
This problem was quickly proven to be {\sf PSPACE}-complete for Maker-Breaker games by Schaefer in 1978~\cite{Schaefer1978} even restricted to $11$-uniform hypergraphs (i.e. hypergraphs in which all hyperedges have size $11$). This bound was recently improved by Rahman and Watson in 2021~\cite{Rahman2021}, proving that the problem is still {\sf PSPACE}-complete if the hypergraph is $6$-uniform. These two proofs are very technical and a simpler proof of the {\sf PSPACE}-completeness was provided by Byskov in 2004~\cite{byskov2004}, proving in the same time that Maker-Maker games are also {\sf PSPACE}-complete. The complexity of Maker-Breaker games is still studied today, as Galliot {\em et al.}~\cite{galliot2022} have proven that the winner of a $3$-uniform Maker-Breaker game can be computed in polynomial time, but the gap between the complexity of $6$-uniform hypergraphs and $3$-uniform hypergraphs remains to be closed.
Despite the fact that Avoidance games were introduced at the same time as Maker-Breaker games, only partial results on complexity are known: determining the winner in Avoider-Avoider games, was proven to be {\sf PSPACE}-complete by Slany in 2002~\cite{slany2002} for endgames, i.e. games in which some vertices are already attributed to the players, but there is no result yet in the general case. Concerning Avoider-Enforcer, Bonnet {\em et al.} in 2017~\cite{Bonnet2017} mentioned that the complexity of this problem is still open, when they proved that short games, i.e. games in which a player has only few moves to make, are co-W[1]-hard, with the number of moves taken as a parameter. The best known result today is due to Miltzow and Stojakovi{\'c} in 2022~\cite{Miltzow2022} that states the {\sf NP}-hardness of this decision problem and conjectures its {\sf PSPACE}-completeness.
\subsection{Presentation of the results}
The Avoider-Enforcer game is played as follows: given a hypergraph $H$, two players, called {\em Avoider} and {\em Enforcer}, alternately claim an unclaimed vertex of $H$ with Avoider starting. The game ends when all the vertices have been claimed. If Avoider has claimed all the vertices of a hyperedge, Enforcer wins. Otherwise, Avoider wins. The related decision problem is the following one.
\begin{problem}{\sc Avoider-Enforcer}
\noindent Input: A hypergraph $H$.
\smallskip
\noindent Output: True if and only if Avoider has a winning strategy in the Avoider-Enforcer game on $H$.
\end{problem}
This paper will focus on the proof of the following result:
\begin{restatable}{theorem}{maintheorem}\label{main theorem}
The {\sc Avoider-Enforcer} problem is {\sf PSPACE}-complete, even when the entry is restricted to hypergraphs with hyperedges of size at most $6$.
\end{restatable}
Our proof of \cref{main theorem} follows a similar idea to the proof of Rahman and Watson~\cite{Rahman2021} and the proof of Schaefer \cite{Schaefer1978}, by constructing some hyperedges forcing the order of the moves. Contrary to Maker-Breaker games, in Avoider-Enforcer convention, there is no vertex in which the players are urged to play, as in general, players do not want to move in avoidance games. The key idea of this reduction is to create some structures in which playing first is a losing move. In the provided construction, at any moment of the game, only few moves are not losing moves. Thus, we can control the vertices played by the two players.
The proof provided for {\sf PSPACE}-completeness of Avoider-Enforcer games, enables us to state the following corollary for Avoider-Avoider games that will also be proven later:
\begin{problem}{\sc Avoider-Avoider}
\noindent Input: A hypergraph $H$.
\smallskip
\noindent Output: True if and only if the second player has a winning strategy in the Avoider-Avoider game on $H$.
\end{problem}
\begin{restatable}{corollary}{avoideravoider}\label{Avoider-Avoider}
The {\sc Avoider-Avoider} problem is {\sf PSPACE}-complete, even when the entry is restricted to $7$-uniform hypergraphs.
\end{restatable}
This paper is organised as follows. In section 2, we introduce two lemmas that will be used in the main proof of the results. In particular, we show that pairing strategies that are often used in Maker-Breaker conventions can also be applied to Avoider-Enforcer games.
Section 3 describes the reduction used to prove the {\sf PSPACE}-completeness and define an order on the move that we call the {\em legitimate order}. We also show in this section that the proof holds if both players follow the legitimate order.
In section 4, we show that if a player does not follow the legitimate order then it cannot be a disadvantage to the other player, completing the proof of the Theorem~\ref{main theorem}.
Finally, in section 5, we use Theorem~\ref{main theorem} to prove Corollary~\ref{Avoider-Avoider} and to show that the Avoider-Enforcer versions of the Domination game and of the $H$-game are {\sf PSPACE}-complete.
\section{Preliminaries}
In Maker-Breaker games, some moves appear to be better than others. For instance, if a vertex is in all the hyperedges, it is always an optimal move to play it for both players. We present here a similar result to prove that some moves are better than others.
\begin{lemma} \label{lemma:included moves}
Let $H$ be a hypergraph, and $u,v$ two vertices of $H$ such that, for every hyperedge $e$ containing $u$, $e$ also contains $v$. If a player has a winning strategy, then this player has a winning strategy in which he never plays $v$ while $u$ is unclaimed.
\end{lemma}
\begin{proof}
Let $H$ be a hypergraphs and let $u,v$ be two vertices such that, for every hyperedge $e$ containing $u$, $e$ also contains $v$. Let $\mathcal{S}_1$ be a winning strategy for Avoider (Enforcer resp.). We define the strategy $\mathcal{S}_2$ as follows:
\begin{itemize}
\item If $\mathcal{S}_1$ claims a vertex $w$ in $V\setminus \{v\}$, $\mathcal{S}_2$ claims $w$.
\item If $\mathcal{S}_1$ claims $v$ while $u$ is unclaimed, claim $u$ and consider the strategy obtained if $v$ was claimed in $\mathcal{S}_1$. If at some point the other player plays $v$ continue as if he played $u$ in $\mathcal{S}_1$, and, if the strategy of $\mathcal{S}_1$ requires that you play $u$, play on $v$ and continue following $\mathcal{S}_1$.
\end{itemize}
Consider a sequence of moves played against $\mathcal{S}_2$. Consider the same sequence played against $\mathcal{S}_1$ by exchanging the roles of $u$ and $v$. Note $V_1$ the vertices claimed by Avoider (Enforcer resp.) in $\mathcal{S}_1$, and $V_2$ the vertices claimed by Avoider (Enforcer resp.) in $\mathcal{S}_2$ according to this sequence. Note that if $V_1 \neq V_2$, then $V_1 \setminus V_2 = \{v\}$ and $V_2 \setminus V_1 = \{u\}$.
As for every hyperedge $e$ such that $u \in e$, we have $v \in e$, in particular, we have for any hyperedge $e$, $|e \cap V_1| \ge |e \cap V_2|$. Thus, in this sequence, as $\mathcal{S}_1$ was a strategy for Avoider (Enforcer resp.) that do not claim all the vertices of a hyperedge (that do not claim a vertex in each hyperedge resp.), $\mathcal{S}_2$ neither. As this result holds for any sequence of moves, $\mathcal{S}_2$ is a winning strategy.
\end{proof}
The second tool we introduce here is {\it pairing strategies} in Avoider-Enforcer games. In Maker-Breaker convention, these strategies are often described by using the fact that a player can claim at least one vertex in each pair of vertices. Here in Avoider-Enforcer convention, the main idea of pairing strategies is that it is always possible to force the opponent to play at least once in each pair.
In this section, we will refer to the players by Alice and Bob, as the strategy can be applied both by Avoider and by Enforcer.
\begin{lemma}\label{pairing strategy}
Let $H = (V,E)$ be a hypergraph. Suppose that Alice plays last in $H$, i.e. if the game is played until all the vertices have been claimed, Alice will claim the last one.
Let $(a_1, b_1), \dots, (a_n, b_n)$ be pairwise disjoint pairs of vertices, and let $v \not \in \underset{i=1}{\overset{n}{\bigcup}} \{a_i,b_i\}$.
Alice has a strategy which forces Bob to play at least one vertex in each pair $(a_i,b_i)$. Bob has a strategy which forces Alice to play $v$ and at least one vertex in each pair $(a_i,b_i)$.
\end{lemma}
A strategy satisfying the hypothesis of Lemma~\ref{pairing strategy} will be called a {\em pairing strategy}.
\begin{proof}
We prove by induction that if Alice plays last on $H$ she can force Bob to play at least once in each pair of unplayed vertices $(a_i,b_i)$ :
\begin{itemize}
\item If $|V| \in \{0,1\}$, then there can be no pair of distinct vertices $(a_i,b_i)$ and the strategy is trivially achieved.
\item If $|V| = 2p+1$ with $p \geq 0$, then, since Alice plays last on $H$, it is Alice's turn to play and there is at least one vertex $x$ in $V$ that is in no pair and Alice can play it. Then, by induction hypothesis, Bob will play one vertex in each pair $(a_i, b_i)$ in $V\setminus \{x\}$.
\item If $|V| = 2p$, with $p \geq 1$, then it is Bob's turn to play. If Bob plays a vertex $x$ that is in no pair $(a_i,b_i)$, then, by induction hypothesis, Alice can force Bob to play once in each pair $(a_i, b_i)$ in $V\setminus \{x\}$. If Bob plays on a vertex $a_i$ or $b_i$, then Alice can play in the other vertex of the pair. Bob played in the pair $(a_i, b_i)$ and by induction hypothesis Alice can force Bob to play once in each pair $(a_j, b_j)_{j \neq i}$ in $V \setminus \{a_i,b_i\}$.
\end{itemize}
We prove by induction that if Alice plays last on $H$, Bob can force Alice to play at least once in each pair of unplayed vertices $(a_i,b_i)$, and to play on the vertex $v$ :
\begin{itemize}
\item If $|V| = 0$, the situation cannot exist as there can be no vertex $v$.
\item If $|V| = 1$, it must be that $V = \{v\}$ and that there is no pair of vertices $(a_i,b_i)$ and since it is the last move, it is Alice's turn, and she has to play in $v$. Thus, she played on $v$ and in each pair.
\item If $|V| = 2p$ with $p \geq 1$, then, since Alice plays last on $H$, it is Bob's turn to play and there is at least one vertex $x$ in $V\setminus \{v\}$ that is in no pair and Bob can play it. Then, by induction hypothesis, Alice will play on $v$ and on one vertex in each pair $(a_i, b_i)$ in $V\setminus \{x\}$.
\item If $|V| = 2p+1$, with $p \geq 1$, then it is Alice's turn to play. If Alice plays a vertex $x$ that is neither $v$ neither in a pair $(a_i,b_i)$, then by induction hypothesis Bob can force Alice to play one $v$ and at lea st once in each pair $(a_i, b_i)$ in $V\setminus \{x\}$. If Alice plays on a vertex $a_i$ or $b_i$ then Bob can play in the other vertex of the pair. Alice played in the pair $(a_i, b_i)$ and by induction hypothesis Bob can force Alice to play on $v$ and once in each pair $(a_j, b_j)_{j \neq i}$ in $V \setminus \{a_i,b_i\}$. If Alice plays on $v$ then either there is no pair $(a_i,b_i)$ and then Alice played on $v$ and in each pair, or there is at least one pair $(a_i,b_i)$. Bob can then play on $a_i$, by induction hypothesis Bob can force Alice to play on $b_i$ and in each pair $(a_j, b_j)_{j \neq i}$. Since Alice played on $v$, and will be forced to play on $b_i$ and in each pair $(a_j, b_j)_{j \neq i}$ the result holds.
\end{itemize}
\end{proof}
\section{Proof of the main theorem}
In this section, we begin the proof of Theorem~\ref{main theorem} by describing the reduction from $3$-QBF, introducing an order on the move of Avoider and Enforcer called the {\em legitimate order} and giving a sketch of the general proof.
\maintheorem*
The first step of the proof is to prove that this game is in {\sf PSPACE}.
\begin{lemma}\label{inPspapce}
The {\sc Avoider-Enforcer} problem is in {\sf PSPACE}.
\end{lemma}
\begin{proof}
Let $H = (V,E)$ be a hypergraph. As the players are not allowed to play an already claimed vertex, any game ends after at most $|V|$ moves. Therefore, according to Lemma~2.2 of Schaefer~\cite{Schaefer1978}, as the game has a polynomial length and a polynomial number of moves, its winner can be computed with polynomial space.
\end{proof}
\subsection{Construction of the hypergraph}\label{subsection construction}
We reduce the problem {\sc 3-QBF} to an {\sc Avoider-Enforcer} game. This problem has been proven {\sf PSPACE}-complete by Stockemeyer and Meyer \cite{stockmeyer1973}, and we use the gaming version of this problem as it was formulated by Rahman and Watson~\cite{Rahman2021}. The game is played on a quantified formula $\varphi$ of the form $\forall X_1 \exists X_2 \dots \forall X_{2n-1} \exists X_{2n} \psi$, with $\psi$ a $3$-SAT formula. Alternately, two players, namely Falsifier and Satisfier, chose valuation for the variables, Falsifier for the odd variables (quantified with a $\forall$) and Satisfier for the even ones (quantified with a $\exists$). When all the variables have a valuation Satisfier wins if $\psi$ is satisfied, otherwise, Falsifier win.
\begin{problem}{3-QBF}
\noindent Input: A $3$-SAT quantified formula $\varphi$ of the form $\forall X_1 \exists X_2 \dots \forall X_{2n-1} \exists X_{2n} \psi$.
\smallskip
\noindent Output: True if and only if Satisfier has a winning strategy in the 3-QBF game on $\phi$
\end{problem}
A round in a $3$-QBF formula corresponds to the steps $i$ during which Falsifier gives a valuation to $X_{2i-1}$ and then Satisfier gives a valuation to $X_{2i}$. In this reduction, any round corresponds to ten vertices and eight hyperedges. Four of the ten vertices are $\{x_{2i-1}, \overline{x_{2i-1}}, x_{2i}, \overline{x_{2i}}\}$, and the six others are $u_{6i-5}, u_{6i-4}, u_{6i-3}, u_{6i-2}, u_{6i-1}, u_{6i}$. The eight hyperedges are constructed as follows:
\vspace{-0.4cm}
\begin{align*}
A_{2i} &= \hspace{.1cm} ( x_{2i}, \overline{x_{2i}}, u_{6i+1}, u_{6i+3} ) &\\
C^+_{6i} &= \hspace{.1cm} ( u_{6i}, u_{6_i+1}, u_{6i+3}, x_{2i} ) &\\
C^-_{6i} &= \hspace{.1cm} ( u_{6i}, u_{6_i+1}, u_{6i+3}, \overline{x_{2i}} ) &\\
C^+_{6i-2} &= \hspace{.1cm} ( u_{6i-2}, u_{6_i-1}, u_{6i+1}, x_{2i} ) &\\
C^-_{6i-2} &= \hspace{.1cm} ( u_{6i-2}, u_{6_i-1}, u_{6i+1}, \overline{x_{2i}} ) &\\
B_{2i-1} &= \hspace{.1cm} ( x_{2i-1}, \overline{x_{2i-1}}, u_{6i-1} ) &\\
C^+_{6i-4} &= \hspace{.1cm} ( u_{6i-4}, u_{6_i-3}, u_{6i-1}, x_{2i-1} ) &\\
C^-_{6i-4} &= \hspace{.1cm} ( u_{6i-4}, u_{6_i-3}, u_{6i-1}, \overline{x_{2i-1}} ) &
\end{align*}
\medskip
\noindent Moreover, for each clause $F_j = l_{j_1} \vee l_{j_2} \vee l_{j_3} \in \psi$ where the vertices $l_{j_1}, l_{j_2}$ and $l_{j_3}$ are literals either positive or negative, we add a hyperedge
$D_j$ containing six vertices. For $k = 1,2,3$, if $l_{j_k}$ is a positive variable $X_i$, then $x_i$ is in $D_j$, if $l_{j_k}$ is a negative one $\neg X_i$, then $\overline{x_i}$ is in $D_j$. If $j_k$ is odd, then $u_{6j_k -1}$ is in $D_j$, if $j_k$ is even, then $u_{6 j_k +1}$ is in $D_j$.
\medskip
\noindent Finally, the CNF game $\phi$ is reduced to the hypergraph $H = (V,E)$ with
\vspace{-0.25cm}
$$ V = \hspace{.1cm} \left \{ \{x_i\}_{1 \le i \le 2n} \cup \{ \overline{x_i}\}_{1 \le i \le 2n} \cup \{ u_j \}_{1 \le j \le 6n} \right \}$$ $$ E = \hspace{.1cm} \left \{ \{A_{2i}\}_{1 \le i \le n} \cup \{C^+_{2i}\}_{1 \le i \le 3n} \cup \{C^-_{2i}\}_{1 \le i \le 3n} \cup \{B_{2i-1}\}_{1 \le i \le n} \cup \{D_j\}_{1 \le j \le m} \right \}
$$
With this construction, we say that Avoider and Enforcer follow a {\em legitimate order} if during each round $i$ the moves are made in the following way, with each round played in increasing order:
\noindent {\bf Legitimate order during round $i$}
\begin{enumerate}
\setlength\itemsep{-.25em}
\item Avoider starts and plays $u_{6i-5}$.
\item Enforcer plays $u_{6i-4}$.
\item Avoider plays $u_{6i-3}$.
\item Enforcer plays one of $x_{2i-1}$ or $\overline{x_{2i-1}}$.
\item Avoider plays the remaining vertex in $(x_{2i-1}, \overline{x_{2i-1}})$
\item Enforcer plays $u_{6i-2}$.
\item Avoider plays $u_{6i-1}$.
\item Enforcer plays $u_{6i}$.
\item Avoider plays one of $x_{2i}$ or $\overline{x_{2i}}$.
\item Enforcer plays the remaining vertex in $(x_{2i}, \overline{x_{2i}})$.
\end{enumerate}
\subsection{Sketch of the proof}
To prove that, with our construction, Avoider wins the {\sc Avoider-Enforcer} game, if and only if Satisfier wins the QBF game, we first prove that this statement is true if the order of the move is legitimate, as the moves will give the valuation obtained in QBF. To force the players to play in the legitimate order, the main idea of the construction is that players want to play some vertices as late as possible. Therefore, we prove that it is always optimal to respect the legitimate order of the moves. We introduce the three following lemmas that will be proved in the next section.
\begin{restatable}{lemma}{lemmaorder}
\label{lem:order_resp}
When the game is restricted to the legitimate order, Avoider has a winning strategy in the Avoider-Enforcer game on $H$ if and only if Satisfier has a winning strategy for the {\sc 3-QBF} game on $\varphi$.
\end{restatable}
\begin{restatable}{lemma}{falsifierwin}
\label{lem:false_win}
If Enforcer has a winning strategy in $\varphi$ when the legitimate order is respected by the two players, then he has a winning strategy in $H$.
\end{restatable}
\begin{restatable}{lemma}{satisfierwin}
\label{lem: satis win}
If Avoider has a winning strategy in $\varphi$ when the legitimate order is respected by the two players, then she has a winning strategy in $H$.
\end{restatable}
We first admit these lemmas and we prove \cref{main theorem}.
\begin{proof}
Let $\varphi$ be a boolean formula. Consider the hypergraph $H$ obtained from $\varphi$ by following the construction of \cref{subsection construction}. This construction has polynomial size, and the hypergraph $H$ has all its hyperedges of size at most $6$. According to \cref{lem:order_resp}, when the order is respected, if Satisfier (Falsifier resp.) has a winning strategy in $\varphi$, Avoider (Enforcer resp.) has a winning strategy in $H$. Thus, according to \cref{lem: satis win} ( \cref{lem:false_win} resp.), if Avoider (Enforcer resp.) has a winning strategy on $H$ when the legitimate order is respected, she (he resp.) has one in general in $H$. Thus, Satisfier wins on $\varphi$ if and only if, Avoider wins on $H$. Therefore, the {\sc Avoider-Enforcer problem} is {\sf PSPACE}-complete.
\end{proof}
\subsection{Game in legitimate order} \label{order respected}
In this section, we suppose that both players follow a legitimate order of moves.
If the order of moves is legitimate, the only choices available for Avoider and Enforcer are on the vertices $x_i$ and $\overline{x_i}$. For each $1 \le i \le 2n$, Avoider plays one of $x_i, \overline{x_i}$ and Enforcer the other. Therefore, if both Avoider and Enforcer plays one vertex in $\{x_i, \overline{x_i}\}$, we define the {\em underlying valuation} given to $\psi$ as the following one:
\begin{align*}
X_i = \left\{
\begin{array}{ll}
True & \mbox{if Avoider has played $\overline{x_i}$ and Enforcer has played $x_i$} \\
False & \mbox{if Avoider has played $x_i$ and Enforcer has played $\overline{x_i}$}
\end{array}
\right.
\end{align*}
\lemmaorder*
\begin{proof}
Consider a game played on $H$ for which both Avoider and Enforcer respected the legitimate order through the whole game.
\begin{claim}
Avoider won the game on $H$ if and only if the formula $\psi$ is satisfied by the underlying valuation of the $X_i$s.
\end{claim}
\begin{claimproof}
Since the legitimate order is respected, Enforcer played all the vertices $u_{2i}$ and thus played at least once in all the hyperedges $C^+_{2i}$ and $C^-_{2i}$. Moreover, for each pair of variables $(x_i,\overline{x_i})$, Enforcer played one of the vertices of the pair and so played at least one vertex in all the hyperedges $A_i$ and $B_i$.
Thus, the only hyperedges that could possibly be fully played by Avoider are the hyperedges $D_j$.
Since, in the legitimate order, Avoider plays on all the vertices $u_{2i+1}$, a hyperedge $D_j$ is fully played by Avoider if and only if she played on all the vertices $x(l_k)$ for $l_k \in F_j$. If this is the case, then this means that the formula $\psi$ is not satisfied by the underlying valuation because the clause $F_j$ has all its literals assigned to False. On the contrary, if the formula $\psi$ is satisfied by the underlying valuation, then, for all clause $F_j$, at least one of its literal is assigned to True and so Enforcer played at least once in each hyperedge $D_j$.
Therefore, Avoider won the game on $H$ if and only if $\psi$ is satisfied.
\end{claimproof}
Suppose Satisfier has a winning strategy $\mathcal{S}$ on $\varphi$. We define a strategy for Avoider as follows: Whenever Avoider has to play a vertex $x_{2k}$ or $\overline{x_{2k}}$, Avoider considers the underlying valuation given to the $X_i$s with $i < 2k$. Then, if Satisfier had put $X_{2k}$ to True, she plays $\overline{x_{2k}}$. Otherwise, she plays $x_{2k}$. With this strategy, at the end of the game, the underlying valuation of the variables of $H$ will be equivalent to the valuation given by the game that Satisfier played on $\varphi$. Since Satisfier has a winning strategy on $\varphi$, the underlying valuation satisfies $\psi$ and so Avoider win the game.
Similarly, if Falsifier has a winning strategy, Enforcer can follow the strategy in such a way that at the end the valuation of variables in the game played by Falsifier correspond to the underlying valuation in $H$. Since Falsifier has a winning strategy, Enforcer wins the game on $H$.
\end{proof}
\section{Proofs of Lemma~\ref{lem:false_win} and Lemma~\ref{lem: satis win}}
The first part of our constructions showed that, if the legitimate order is respected, Avoider wins if and only if $\varphi$ is satisfied by the underlying valuation provided by the moves. We now prove that, if a player has a winning strategy, he has one that respects the legitimate order of the moves. We introduce here sets of variables in which players do not want to play. These sets will be the main point of the proofs of \cref{lem:false_win} and \cref{lem: satis win}.
For $i = 1$ to $4n$,
we define the set of vertices $S_i$ as $S_{4n} = \{u_{6n},x_{2n},\overline{x_{2n}}\}$ and for $i < 4n$:
\begin{itemize}
\item if $i=4k$, $S_i = \{u_{6k},x_{2k},\overline{x_{2k}},u_{6k+1}\} \cup S_{i+1} $
\item if $i=4k-1$, $S_i = \{u_{6k-2},u_{6k-1}\} \cup S_{i+1}$
\item if $i=4k-2$, $S_i = \{x_{2k-1},\overline{x_{2k-1}}\} \cup S_{i+1}$
\item if $i=4k-3$, $S_i = \{u_{6k-4},u_{6k-3}\} \cup S_{i+1}$
\end{itemize}
\subsection{Proof of Lemma~\ref{lem:false_win}}
\falsifierwin*
\begin{proof}
Suppose Falsifier has a winning strategy in $\varphi$. Consider a strategy for Enforcer in which he plays according to the legitimate order until Avoider does not. If Avoider respects the order until all the vertices are played, by Lemma~\ref{lem:order_resp}, Enforcer wins. Otherwise, the proof of the following claim provides a winning strategy to Enforcer.
\begin{claim}
If, during the game, Avoider plays in a set $S_i$ in which Enforcer has not played yet, then, after this move, Enforcer has a strategy to win the game.
\end{claim}
\begin{claimproof}
The proof is by induction on $i$.
First, notice that each $S_i$ has an odd number of vertices and, as the total number of vertices in $H$ is $10n$, there is also an odd number of vertices outside $S_i$. Therefore, if Avoider plays first in an $S_i$, Enforcer answers by playing an arbitrary vertex that is not in $S_i$ and considers an arbitrary pairing outside $S_i$, which exists as there is an even number of vertices outside $S_i$ after his move. This way, Avoider has to be the next player to play in $S_i$.
{\bf Base cases:}
\begin{itemize}
\item Case $i= 4n$: If Avoider plays first is $S_{4n}$, by pairing the two other vertices in $S_{4n}$, by using Lemma~\ref{pairing strategy}, Enforcer can force Avoider to play another vertex in $S_{4n}$. Hence, as $(u_{6n},x_{2n})$, $(u_{6n},\overline{x_{2n}})$ and $(x_{2n}, \overline{x_{2n}})$ are three hyperedges, Avoider will claim the two vertices of one of them and then lose.
\item Case $i=4n-1$: As shown previously, Enforcer has as strategy such that Avoider is the next player to play in $S_{4n-1}$. If Avoider has played at least one of its two first moves in $S_{4n}$, she has lost by the case $i=4n$. Otherwise, she has played exactly $u_{6n-2}$ and $u_{6n-1}$. In this case, Enforcer plays on $u_{6n}$ and pairs $x_{2n}$ and $\overline{x_{2n}}$ and by Lemma~\ref{pairing strategy} he forces Avoider to claim all the vertices of $C^+_{6n-2}$ or $C^-_{6n-2}$.
\end{itemize}
{\bf Induction steps:}
Suppose that the first time Avoider does not respect the order of the move, she plays in a set $S_i$ for $i \le 4n-2$. If the second move of Avoider in $S_i$ is in $S_{i+1}$, Enforcer wins by induction hypothesis. Thus, we can suppose that Avoider has played two vertices in $S_i \setminus S_{i+1}$. Moreover, as Enforcer has arbitrarily paired the vertices outside $S_i$, we describe here the strategy in $S_i$, and Enforcer plays according to the pairing outside $S_i$. This strategy ensure that the moves in $S_i$ alternate between both players.
\begin{itemize}
\item Case $i= 4k$: Avoider has played twice in $\{u_{6k}, x_{2k}, \overline{x_{2k}}, u_{6k+1}\}$. At least one of $\{u_{6k}, x_{2k}, \overline{x_{2k}}\}$ is available. Enforcer plays it. Avoider has to play a third vertex in this quadruple, otherwise, she plays first in $S_{i+1}$ and loses by induction, and necessarily one of the three vertices she has played is $u_{6k+1}$. Enforcer plays $u_{6k+2}$. Avoider either plays first in $S_{i+2}$ and loses by induction hypothesis, or plays $u_{6k+3}$. At this moment, Avoider has played on the vertices $u_{6k+1}$ and $u_{6k+3}$, and two of the vertices of $\{u_{6k}, x_{2k}, \overline{x_{2k}}\}$. So she has completed one of the hyperedge $C^+_{6k} = (u_{6k}, u_{6k+1}, u_{6k+3}, x_{2k})$, $C^-_{6k} = (u_{6k}, u_{6k+1}, u_{6k+3}, \overline{x_{2k}})$ or $A_{2k} = (x_{2k}, \overline{x_{2k}}, u_{6k+1}, u_{6k+3}) $.
\item Case $i=4k-1$: Avoider has played $u_{6k-2}$ and $u_{6k-1}$. Enforcer plays $u_{6k}$. Avoider has to play on vertex in $\{x_{2k}, \overline{x_{2k}}, u_{6k+1}\}$, otherwise she plays first in $S_{i+2}$ and loses by induction. Enforcer plays either $x_{2k}$ or $\overline{x_{2k}}$, as at least one of them is available. If Avoider plays a vertex in $S_{i+2}$ she loses by induction. So she has to play the last vertex available in $S_i \setminus S_{i+1}$. With this strategy, Avoider has necessarily played $u_{6k+1}$ and one of $x_{2k}$ and $\overline{x_{2k}}$. Thus, she has played all the vertices of either $C^+_{6k-2} = (u_{6k-2}, u_{6k-1}, u_{6k+1}, x_{2k})$ or $C^-_{6k-2} = (u_{6k-2}, u_{6k-1}, u_{6k+1}, \overline{x_{2k}})$.
\item Case $i=4k-2$: Avoider has played $x_{2k-1}$ and $\overline{x_{2k-1}}$. Enforcer plays $u_{6k-2}$. Either Avoider plays first in $S_{i+2}$ and loses by induction, or she plays $u_{6k-1}$, the last available vertex in $S_{i+1}$ and loses by having played all the vertices in $B_{2k-1} = (x_{2k-1}, \overline{x_{2k-1}}, u_{6k-1})$.
\item Case $i=4k-3$: Avoider has played $u_{6k-4}$ and $u_{6k-3}$. Enforcer plays $x_{2k-1}$. Avoider has to play $\overline{x_{2k-1}}$, otherwise she plays first in $S_{i+2}$ and loses by induction. Then Enforcer plays $u_{6k-2}$. If Avoider plays in $S_{i+3}$ she loses by induction. The last vertex available in $S_i \setminus S_{i+3}$ is $u_{6k-1}$, and if Avoider plays it, she loses by playing all the vertices $C^-_{6k-4} = (u_{6k-4}, u_{6k-3}, u_{6k-1}, \overline{x_{2k-1}})$.
\end{itemize}
By applying this induction, at any moment of the game, if Avoider plays first in a set $S_i$, she looses.
\end{claimproof}
Finally, if Enforcer has played according to the legitimate order, at any moment of the game, Avoider has to play in a set $S_i$ in which Avoider has already played. Therefore, she has to respect the order of the moves. The only moment when she can change this order is by playing $u_{6k+1}$ instead of one of the vertices $x_{2k}, \overline{x_{2k}}$. But if she does so, Enforcer can play one of them, for instance $x_{2k}$, and Avoider will be forced to play $\overline{x_{2k}}$. If this happens, everything happens as if Avoider has played $\overline{x_{2k}}$ first and $u_{6k+1}$ after. The strategy can then continue as if the order has been respected, as this does not change the order in which the valuation of the variables is chosen in $\varphi$, and Falsifier had a winning strategy if Satisfier had put $X_{2k}$ to True by hypothesis.
To conclude, if Falsifier has a winning strategy in $\varphi$, Enforcer has a winning strategy in $H$.
\end{proof}
\subsection{Proof of Lemma~\ref{lem: satis win}}
In this section, we prove that if Satisfier has a winning strategy in $\varphi$, Avoider has one in $H$, even if Enforcer does not respect the order. The main idea of the strategy is to respect the order, and if Enforcer does not respect the order, Avoider has a pairing strategy to force enforcer to claim some vertices $x_i$, or the odd $u_j$ that follows it. By construction, any hyperedge containing a vertex $x_i$ contains also the next odd vertex $u_i$ in the order, and this will prove that when Enforcer do not respect the order, it will benefit Avoider.
\satisfierwin*
\begin{proof}
If Satisfier has a winning strategy $\mathcal{S}$ in $\varphi$, and if Enforcer respects the order, Avoider has a winning strategy according to Lemma~\ref{lem:order_resp}.
While Enforcer respects the legitimate order, Avoider also respects it. Suppose that at any moment of the game, Enforcer does not respect the legitimate order. Denote by $y_A$ the vertex he would have played according to the legitimate order, and by $y_E$ the vertex he has played instead. First, note that, according to Lemma~\ref{lemma:included moves}, we can suppose that $y_E$ is not a vertex $u_j$ with $j$ odd. Indeed, for each vertex $u_{2i+1}$, the hyperedges that contain the previous vertex in the legitimate order (if this vertex is a vertex $x_j$ or $\overline{x_j}$ this is true for any of them), also contain $u_{2i+1}$. If $y_A$ is a vertex $x_{2i}$ or $\overline{x_{2i}}$, Avoider pairs it with $u_{6i+1}$ and continues as if Enforcer should have played $u_{6i+2}$. As any hyperedge containing $x_{2i}$ or $\overline{x_{2i}}$, also contains $u_{6i+1}$ which is the next vertex Avoider should have played according to the legitimate order, it benefits her if Enforcer finally plays $u_{6i+1}$ according to Lemma~\ref{lemma:included moves}.
Denote by $k$ the smallest integer such that $y_E \notin S_k$, and by $k'$ the largest integer such that $y_A \in S_{k'}$. Note that all the vertices outside $S_{k'}$ have already been played or are paired, and that $S_{k'} \setminus S_k$ is then the set of vertices perturbed by the move of Enforcer. We consider $k = 4n+1$ if $y_E \in S_{4n}$. As all the set $S_i$ have an odd number of variables, Avoider knows that the number of remaining vertices outside $S_k$ is odd. (as an even number of moves have been done in it). According to $\mathcal{S}$, Satisfier has a winning strategy starting with the value already given to the variables $x_i$ and $\overline{x_i}$ that are outside $S_{k'}$. Avoider considers than arbitrary moves for Falsifier in $\mathcal{S}$ and the corresponding moves for Satisfier until all the vertices $x_i$s and $\overline{x_i}$s in $S_k$ are played in $\mathcal{S}$. According to these moves, we will denote by $x^E_j$ the vertex among $(x_j, \overline{x_j})$ played by Enforcer and by $x^A_j$ the vertex played by Avoider, such that their underlying valuation is the one desired by $\mathcal{S}$.
Avoider will then play a strategy different in $S_k$ and outside $S_k$:
\begin{itemize}
\item In $H\setminus S_{k}$, Avoider plays $y_A$, the vertex that Enforcer should have played according to the order, then plays according to a pairing strategy, that is presented in the next paragraph.
\item In $S_k$, Avoider considers the strategy she would have played if all the vertices outside $S_k$ were played according to the legitimate order, with the valuation she considered in $\mathcal{S}$.
\end{itemize}
The pairing we define is the following one: $(u_{6i-4}, x_{2i-1}^A)$, $(u_{6i-3}, u_{6i-6})$, $(x_{2i-1}^E, u_{6i-1})$, $(u_{6i-2}, x_{2i}^A)$, $(u_{6i+1}, x_{2i}^E)$. This pairing starts at the vertex $y_A$ and we consider only pairs containing a vertex outside $S_k$. Note that, by construction, exactly one vertex of this pairing is already played, and exactly one is in $S_k$. Therefore, to make the pairing contain only vertices not played and outside $S_k$, some modifications are done. These modifications are presented in Figure~\ref{fig:changes matching}. By applying Lemma~\ref{pairing strategy}, Avoider can ensure that Enforcer plays at least one in each of these pairs.
\begin{figure}
\centering
\begin{tabular}{ |c|c| } \hline
$y_A$ & changes \\
\hline
$u_{6i-4}$ & $u_{6i-3} \longleftrightarrow x_{2i-1}^A$ \\
$x_{2i-1}$ or $\overline{x_{2i-1}}$ & $x^*_{2i-1} \longleftrightarrow u_{6i-1}$ \\
$u_{6i-2}$ & $u_{6i-1} \longleftrightarrow x_{2i}^A$ \\
$u_{6i}$ & $u_{6i+3} \longleftrightarrow x_{2i}^A$ \\ \hline
\end{tabular}
\begin{tabular}{ |c|c| } \hline
$y_E$& changes \\ \hline
$u_{6i-4}$ & no changes \\
$x_{2i-1}$ or $\overline{x_{2i-1}}$ & $x_{2i-1}^* \longleftrightarrow u_{6i-4}$ \\
$u_{6i-2}$ & no changes \\
$u_{6i}$ & no changes\\
$x_{2i}$ or $\overline{x_{2i}}$ & $x_{2i}^* \longleftrightarrow u_{6i+1}$, $u_{6i-2} \longleftrightarrow u_{6i}$ \\\hline
\end{tabular} \hfill
\caption{Changes of the matching. ($x_k^*$ refers to the variable between $x_k$ and $\overline{x_k}$ that has not been played ) }
\label{fig:changes matching}
\end{figure}
\begin{claim}\label{strat ABC}
The pairing strategy ensures that Enforcer plays at least once in each hyperedge $A_i$, $B_i$ or $C_i$ containing at all their vertices in $S_{k'}$ and at least one outside $S_k$.
\end{claim}
\begin{claimproof}
First, if the hyperedge contain no vertex whose pairing has been modified because of their appurtenance to $y_A$ or $y_E$, it contains two paired vertices. We show in bold text the paired vertices:
\vspace{-.4cm}
\begin{align*}\label{pairing Avoider}
A_{2i} &= \hspace{.1cm} ( x^A_{2i}, {\bf x^E_{2i}}, {\bf u_{6i+1} }, u_{6i+3} ) &\\
C^A_{6i} &= \hspace{.1cm} ( {\bf u_{6i}}, u_{6_i+1}, {\bf u_{6i+3}}, x^A_{2i} ) &\\
C^E_{6i} &= \hspace{.1cm} ( u_{6i}, {\bf u_{6_i+1}}, u_{6i+3}, {\bf x^E_{2i}} ) &\\
C^A_{6i-2} &= \hspace{.1cm} ( {\bf u_{6i-2}}, u_{6_i-1}, u_{6i+1}, {\bf x^A_{2i}} ) &\\
C^E_{6i-2} &= \hspace{.1cm} ( u_{6i-2}, u_{6_i-1}, {\bf u_{6i+1}}, {\bf x^E_{2i}} ) &\\
B_{2i-1} &= \hspace{.1cm} ( x^A_{2i-1}, {\bf x^E_{2i-1}}, {\bf u_{6i-1}} ) &\\
C^A_{6i-4} &= \hspace{.1cm} ( {\bf u_{6i-4}}, u_{6_i-3}, u_{6i-1}, {\bf x^A_{2i-1}} ) &\\
C^E_{6i-4} &= \hspace{.1cm} ( u_{6i-4}, u_{6_i-3}, {\bf u_{6i-1}}, {\bf x^E_{2i-1}} ) &
\end{align*}
\smallskip
For the first hyperedges of the matching, there are two paired vertices:
\begin{itemize}
\item If $y_A = u_{6i-4}$, only the hyperedges $C^E_{6i-4}$ and $C^A_{6i-4}$ are concerned by the changes. In the former $x^E_{2i-1}$ is paired with $u_{6i-1}$, in the latter $x^A_{2i-1}$ is paired with $u_{6i-3}$.
\item If $y_A\in \{ x_{2i-1}, \overline{x_{2i-1} } \}$, the only hyperedges concerned by the change is $B_{2i-1}$. In it, the other vertex in $ \{ x_{2i-1}, \overline{x_{2i-1} } \}$ is paired with $u_{6i-1}$
\item If $y_A = u_{6i-2}$, only the hyperedges $C^E_{6i-2}$ and $C^A_{6i-2}$ are concerned by the changes. In the former $x^E_{2i}$ is paired with $u_{6i+1}$, in the latter $x^A_{2i}$ is paired with $u_{6i-1}$.
\item If $y_A = u_{6i}$, only the hyperedges $C^E_{6i}$ and $C^A_{6i}$ are concerned by the changes. In the former $x^E_{2i}$ is paired with $u_{6i+1}$, in the latter $x^A_{2i}$ is paired with $u_{6i+3}$.
\end{itemize}
For the last hyperedges that contains vertices of the matching, the following happens:
\begin{itemize}
\item If $y_E = u_{6i-4}$, the pairing stops at $u_{6i-3}$. The only two hyperedges that contains at least one vertex in $S_k$ and one vertex outside $S_k$ are $C^+_{6i-4}$ and $C^-_{6i-4}$, in which Enforcer has played $y_E$.
\item If $y_E = x_{2i -1}$ or $\overline{x_{2i-1}}$, the pairing stops after the second vertex in $\{x_{2i-1}, \overline{x_{2i-1}}\}$. The only one hyperedge containing at least one vertex in $S_k$ and one outside $S_k$ is $B_{2i-1}$, in which Enforcer has already played $y_E$.
\item If $y_E = u_{6i-2}$, the pairing stops at $u_{6i-1}$. The only two hyperedges that contains at least one vertex in $S_k$ and one vertex outside $S_k$ are $C^+_{6i-2}$ and $C^-_{6i-2}$, in which Enforcer has played $y_E$.
\item If $y_E = u_{6i}$, the pairing stops at $u_{6i+1}$. The three hyperedges that contains both vertices in $S_k$ and vertices outside $S_k$ are $C^+_{6i}$, $C^-_{6i}$ and $A_{2i}$. In $C^+_{6i}$, $C^-_{6i}$, Enforcer has played $y_E$, and in $A_{2i}$, Enforcer will play one of $x^E_{2i}$ or $u_{6i+1}$ as these two vertices are paired together.
\item If $y_E = x_{2i}$ or $\overline{x_{2i}}$, the pairing stops at $u_{6i+1}$. The three hyperedges that contains vertices inside $S_k$ and outside $S_k$ are $C^+_{6i}$, $C^-_{6i}$ and $A_{2i}$. As the second vertex in $\{x_{2i}, \overline{x_{2i}}\}$ is paired with $u_{6i+1}$, either Enforcer has played both $x_{2i}$ and $\overline{x_{2i}}$, and any of these three hyperedges contain at least one of them; or Enforcer has played $u_{6i+1}$ which is in these three hyperedges.
\end{itemize}
If the pairing stops because it goes until the end (i.e. $k = 4n +1$), one vertex is not paired. According to Lemma~\ref{pairing strategy}, as Enforcer plays the last move in $H$, Avoider can force him to play it and still play once in each pair of the pairing.
Finally, in any hyperedge $A_i, B_i$ or $C_i$ containing at least one vertex of the matching, Enforcer has played at least one vertex.
\end{claimproof}
Now, we can prove that the strategy we defined for Avoider is a winning strategy. In all the hyperedges $A_i, B_i$ or $C_i$, Enforcer played at least once: If Enforcer has respected the order until vertices of this hyperedge are reached, he has to play in it, otherwise by Claim~\ref{strat ABC}, Avoider can force Enforcer to play in it as this hyperedge is considered in a set of hyperedges in which Enforcer has not respected the order, as the only hyperedges in which Enforcer is not the first player to claim a vertex according to the order are the $A_{2i}$, and in them, the vertex $x_{2i}^E$ is always paired with $u_{6i+1}$.
In the hyperedges $D_j$, as the strategy defined by $\mathcal{S}$ is a winning strategy, Satisfier has a strategy to force at least one literal $l_i$ to be True in $F_j$. By construction, if when the vertex $x(l_i)$ has to be played, the order was respected, Enforcer has played it. If it has not, then Avoider has paired it with the vertex $u(l_i)$. In both cases, Enforcer has played in $D_j$.
\end{proof}
\section{Applications}
\subsection{From 6-hypergraphs to 6-uniform hypergraphs}
The construction provided in Section~\ref{subsection construction} provided a hypergraphs in which all hyperedges have size at most six. We prove here that we can suppose that all hyperedges have size six without changing the outcome.
A hypergraph $H = (V,E)$ is a {\it $k$-hypergraph} if each edge $e\in E$ has size at most $k$. It is said to be {\it $k$-uniform} if each edge $e \in E$ has size exactly $k$.
\begin{lemma}\label{to6uniform}
Let $H = (V,E)$ be a $k$-hypergraph. Let $m = \underset{e \in E}{\min} |e|$. If $m <k$, there exists a $k$-hypergraph $H'= (V',E')$ where $\underset{e \in E}{\min} |e| = m+1$, and having $|E'|\le 2|E|$ and $|V'| \le |V| + 2$ such that Avoider has a winning strategy in Avoider-Enforcer on $H$ if and only if she has one in $H'$.
\end{lemma}
\begin{proof}
Let $H = (V,E)$ be a $k$-hypergraph. Let $m = \underset{e \in E}{\min} |e|$. We define $H' = (V', E')$ as follows. We start from $H' = H$. We add two vertices $\{a_1, a_2\}$ in $V'$. For each edge $e \in E$, we do the following:
\begin{itemize}
\item If $|e| > m$, we keep $e$ in $E'$.
\item If $|e| = m$, we consider $e_1 = e \cup \{a_1\}$ and $e_2 = e \cup \{a_2\}$. We then replace $e$ in $E'$ by $e_1$ and $e_2$.
\end{itemize}
We have $|V'|= |V| +2$, $|E'| \le 2|E|$ and $\underset{e \in E'}{\min} |e| = m+1$.
Now, if Avoider (Enforcer resp.) had a winning strategy $\mathcal{S}$ in $E$, we can define a strategy $\mathcal{S}'$ in $E'$ as follows:
\begin{itemize}
\item If the opponent plays a vertex in $V$, or if it is the first move of the player, play as in $\mathcal{S}$.
\item If the opponent plays a vertex in $\{a_1,a_2\}$, or if there is no vertex in $V$ available, play an available vertex in $\{a_1,a_2\}$.
\end{itemize}
Following this strategy, Avoider (Enforcer resp.) has played exactly the same vertices in $H'$ as he (she resp.) would have played in $H$ according to $\mathcal{S}$ with the addition of exactly one of $\{a_1,a_2\}$.
Therefore, if Avoider had a winning strategy in $H$, then for each $e \in E$, there exists one vertex $v \in e$, that Enforcer has played. This vertex is also in $e$ if $e \in E'$ or in $e_1$ and $e_2$ if $e_1$ and $e_2$ were added to $E'$ when we considered $e$. So this strategy ensure Avoider that Enforcer has played one vertex in $e$, and so it is a winning strategy for Avoider.
If Enforcer had a winning strategy in $H$, following this strategy, there exists an edge $e \in E$ in which Avoider has coloured all the vertices. If $|e| \ge m+1$, Avoider has also coloured all the vertices of $e \in E'$, so Enforcer has won. If $|e| = m$, as the strategy $\mathcal{S}'$ forces Avoider to play at least one of $\{a_1, a_2\}$, suppose without loss of generality that she has played $a_1$. Then, she has played $a_1$ and all the vertices of $e$, so she has filled the edge $e_1$. Therefore, this strategy is a winning strategy for Enforcer.
Finally, $H'$ has the same outcome as $H$ and $\underset{e \in E'}{\min} |e| = m+1$.
\end{proof}
\begin{corollary}\label{corollary6-unif}
Avoider-Enforcer is {\sf PSPACE}-complete even on $6$-uniform hypergraphs
\end{corollary}
Several games have been proven to be {\sf PSPACE}-complete in the Maker-Breaker convention, thanks to the proofs of Schaefer or of Rahman and Watson. Due to the similarities between the two convention, some reductions may be adapted to prove that these games are {\sf PSPACE}-complete in the Avoider-Enforcer convention. In particular, we prove in this section that Avoider-Avoider games are {\sf PSPACE}-complete, and we show that the Domination Game and the $H$-Game are {\sf PSPACE}-complete in Avoider-Enforcer convention.
\subsection{Avoider-Avoider games}
We prove here that Avoider-Avoider games are {\sf PSPACE}-complete.
\avoideravoider*
\begin{proof}
Consider the construction provided in the proof of \cref{corollary6-unif}. Consider $H'$ the hypergraph obtained by adding a vertex $v_0$ in $H$ and adding it in all the hyperedges of $H$. Note that in $H'$, all the hyperedges have size seven. By Lemma~\ref{lemma:included moves}, both player have an optimal strategy in which $v_0$ will be played last, and as the graph has an odd number of vertices, the first player will play it. Therefore, the second player cannot fill a hyperedge and plays as Enforcer would in the Avoider-Enforcer game. By applying the same strategy as in Avoider-Enforcer, If Avoider wins in Avoider-Enforcer, the game ends by a draw, otherwise the second player wins.
\end{proof}
\subsection{ Particular Avoider-Enforcer games}
\subsubsection{Avoider-Enforcer Domination game}
The Maker-Breaker Domination game was introduced by Duchêne {\em et al.} in 2020~\cite{Duchene2020} and follows the study of Domination games on graphs, which were investigated since 2002~\cite{Alon2002, Brevsar2010}. In Maker-Breaker, two players, namely Dominator and Staller alternately claim an unclaimed vertex of the graph, Dominator wins if he manages to take all the vertices from a dominating set. Otherwise, Staller wins. They proved that determining whether Dominator or Staller has a winning strategy is {\sf PSPACE}-complete using a reduction from Maker-Breaker. The Avoider-Enforcer Domination game can be similarly defined, with Staller winning if Dominator claims a dominating set and Dominator winning otherwise. We prove here that determining the winner of the Avoider-Enforcer Domination Game is {\sf PSPACE}-complete. Note that the proof is very similar to the reduction from Maker-Breaker games to Maker-Breaker Domination game.
\begin{problem}{\sc Avoider-Enforcer Domination Game}
Input: A graph $G$
Output: True if and only if Avoider wins the Avoider-Enforcer Domination Game on $G$.
\end{problem}
\begin{restatable}{theorem}{Domination}
Avoider-Enforcer Domination game is {\sf PSPACE}-complete
\end{restatable}
\begin{proof}
The proof of {\sf PSPACE}-completeness in Avoider-Enforcer is roughly the same as in Maker-Breaker.
First, Avoider-Enforcer Domination game is in {\sf PSPACE}, as the number of moves in a game is the number of vertices, and as determining if a set is a dominating set or not, the game is in {\sf PSPACE}.
Let $H = (V_H, E_H)$ be a hypergraph. Without loss of a generality, we can suppose that each vertex is in at least one hyperedge. We construct the following graph $G = (V,E)$ as follows:
\begin{itemize}
\item For each vertex $u_i$ in $V_H$, we add a vertex $v_i$ in $V$.
\item For each hyperedge $C$ in $E_H$, we add two vertices $v_C^1$ and $v_C^2$ in $V$.
\item If a vertex $u_i$ of $V_H$ belongs to a hyperedge $C$ of $E_H$, we add the edges $v_iv_C^1$ and $v_iv_C^2$ to $E$.
\end{itemize}
Note that the graph created here is bipartite
Suppose Avoider (Enforcer resp.) has a winning strategy $\mathcal{S}$ in $H$. We define a strategy $\mathcal{S'}$ for Bob (Alice resp.) in $G$ as follows.
\begin{itemize}
\item If Avoider (Enforcer resp.) plays the first move in $H$, play first a vertex $v_i$ such that $u_i$ is the first vertex played in $\mathcal{S}$.
Then:
\item If the opponent plays a vertex $v_i$, play a vertex $v_j$ such that $u_j$ is the answer to the vertex $u_i$ in $\mathcal{S}$.
\item If a player plays a vertex $v_C^k$ for $k \in \{1,2\}$, plays the vertex $v_C^{k'}$ for $k'\neq k \in \{1,2\}$.
\end{itemize}
Now if Avoider had a winning strategy in $H$, by applying the strategy $\mathcal{S}'$, for any vertex $v_C^i$, Bob has not played all the $v_j$s adjacent to it. Therefore, Alice has played one of them and all the $v_C^i$s are dominated. Moreover, all the vertex $u_j$s are in at least one edge $C$ of $H$. So $v_j$ is dominated by Alice has played one of $(v_C^1, v_C^2)$. So Bob has won.
If Enforcer had a winning strategy in $H$, by applying the strategy $\mathcal{S}'$, Bob knows that there exists a pair of vertices $(v_C^1, v_C^2)$, such that Alice has played all the $v_j$s adjacent to them. As Bob has played exactly one of them, he has not played the second one. Therefore, he does not dominate it and has won.
\end{proof}
In Figure~\ref{fig:Domination} we provide an example of reduction. Note that by connecting all the vertices $v_i$, they will form a clique, and this gives the proof Avoider-Enforcer Domination game on split graphs. So the Avoider-Enforcer game is {\sf PSPACE}-complete on bipartite and split graphs, as the Maker-Breaker Domination game for roughly the same construction.
\begin{figure}[ht]
\centering
\scalebox{1}{
\begin{tikzpicture}
\node[noeud] (x1) at (0,0){};
\node[noeud] (x2) at (1.25,0){};
\node[noeud] (x3) at (2.5,0){};
\node[noeud] (x4) at (3.75,0){};
\node[above] at (x1) {$v_1$};
\node[above] at (x2) {$v_2$};
\node[above] at (x3) {$v_3$};
\node[above] at (x4) {$v_4$};
\node[noeud] (c1) at (-0.8,-3){};
\node[noeud] (c1b) at (0.1,-3){};
\node[noeud] (c2) at (1.5,-3){};
\node[noeud] (c2b) at (2.4,-3){};
\node[noeud] (c3) at (3.8,-3){};
\node[noeud] (c3b) at (4.7,-3){};
\node[below=1pt] at (c1){$v_A^1$};
\node[below=1pt] at (c1b) {$v_A^2$};
\node[below=1pt] at (c2) {$v_B^1$};
\node[below=1pt] at (c2b) {$v_B^2$};
\node[below=1pt] at (c3) {$v_C^1$};
\node[below=1pt] at (c3b) {$v_C^2$};
\draw (x1) -- (c1);
\draw (x1) -- (c1b);
\draw (x1) -- (c2);
\draw (x1) -- (c2b);
\draw (x2) -- (c1);
\draw (x2) -- (c1b);
\draw (x2) -- (c3);
\draw (x2) -- (c3b);
\draw (x3) -- (c3);
\draw (x3) -- (c3b);
\draw (x4) -- (c2);
\draw (x4) -- (c2b);
\draw (x4) -- (c3);
\draw (x4) -- (c3b);
\node[noeud] (u1) at (-5.5,-2.25){};
\node[noeud] (u2) at (-5.5,-0.75){};
\node[noeud] (u3) at (-4,-0.75){};
\node[noeud] (u4) at (-4,-2.25){};
\node (u1') at (-5.3,-2.25){};
\node (u2') at (-5.3,-0.75){};
\node (u3') at (-3.8,-0.75){};
\node (u4') at (-3.8,-2.25){};
\node[right] at (u1){$u_1$};
\node[right] at (u2){$u_2$};
\node[right] at (u3){$u_3$};
\node[right] at (u4){$u_4$};
\draw \convexpath{u2',u3',u4'}{0.5cm};
\draw \convexpath{u1',u2'}{0.45cm};
\draw \convexpath{u1',u4'}{0.4cm};
\node[left] at (-5.7, -1.4){$A$};
\node[right] at (-3.4, -1){$B$};
\node at (-4.5, -2.9){$C$};
\end{tikzpicture}
}
\caption{Reduction from {\sc Avoider-Enforcer}.}
\label{fig:Domination}
\end{figure}
\subsubsection{Vertex H-game}
The vertex $H$-Game has been introduced by Kronenberg, Mond and Naor in \cite{Kronenberg2019} on random graphs. It is presented in several conventions, but we will focus here on the Avoider-Enforcer one. The game is played as follows:
Let $H$ be a graph. Avoider and Enforcer play on the vertex set of another graph $G$. Alternately, Avoider and Enforcer claim an unclaimed vertex of $G$. Avoider wins if the set of vertices she has claimed do not contain $H$ as a subgraph (not necessarily induced). Otherwise, Enforcer wins.
We prove here that determining the winner of the Avoider-Enforcer $H$-Game is a {\sf PSPACE}-complete problem for several graphs $H$ .
\begin{problem}{\sc Avoider-Enforcer $H$-Game}
\noindent Input: A graph $G$
\noindent Output: True if and only if Avoider wins the Avoider-Enforcer $H$-Game played on $G$.
\end{problem}
We prove that the vertex $H$-game is {\sf PSPACE}-complete for several graphs $H$. To define these graphs, we first need to define some graphs and operations:
\begin{itemize}
\item We will denote by $I_k$ the graph being an independent set of size $k$, i.e. containing $k$ vertices and no edge.
\item If $G$ and $H$ are two graphs, we denote by $G\bowtie H$ their join, i.e. if $G = (V_G, E_G)$ and $H = (V_H, E_H)$, we have $G \bowtie H = (V,E)$ with $V = V_G \cup V_H$ and $E = E_G \cup E_H \cup \{(v_G,v_H)| v_G\in V_G, v_H \in V_H\}$.
\item If $G$ and $H$ are two graphs, we denote by $G\boxtimes H$ their strong product, i.e. if $G = (V_G, E_G)$ and $H = (V_H, E_H)$, we have $G \boxtimes H = (V,E)$ with $V = \{ x_{u,v}| u \in V_G, v \in V_H \}$ and $E = \{ (x_{u_1,v_1}, x_{u_2,v_2}) | \left (u_1 = u_2 \mbox{ or } (u_1,u_2) \in E_G \right )$ and $\left ( v_1=v_2 \mbox{ or } (v_1,v_2) \in E_H \right ) \}$.
\end{itemize}
Remark that for any graph $G$, $G \boxtimes P_2$ (where $P_2$ design the path of length $2$) is obtained by taking two copies of $G$ and connecting each vertex to its copy and its copy's neighbours.
We prove here that determining the winner of the Avoider-Enforcer $H$-Game is a {\sf PSPACE}-complete problem for several graphs $H$ .
\begin{theorem}
Let $H_0$ be a graph containing at least one edge or at least $6$ vertices, and let $k\ge 6$. Consider $H = I_k \bowtie H_0 $. The {\sc Avoider-Enforcer $H$-game} problem is {\sf PSPACE}-complete.
\end{theorem}
Note that complete bipartite graphs $K_{n,m}$, with $n,m \ge 6$, are of this type. Indeed, $K_{n,m} = I_k \bowtie H_0$ for $H_0 = I_m$ and $k= n$.
\begin{proof}
First, the $H$-game is in {\sf PSPACE}. Indeed, as it is a positional game, if $G = (V,E)$ is a graph, the game ends after at most $|V|$ moves. After that, determining whether a graph $H$ is a subgraph of a graph $G$ can be done in polynomial space.
We do our reduction from {\sc Avoider-Enforcer} on $6$-uniform hypergraphs which is {\sf PSPACE}-complete by \cref{corollary6-unif}. Let $H_0$ be a graph containing at least one edge or at least $6$ vertices, and let $k \ge 6$. Let $H = I_k \bowtie H_0$. To avoid confusion while describing the strategies in the two games, we will call the players of the $H$-Game Alice and Bob, with Alice avoiding creating a subgraph $H$ and Bob forcing her to create one.
Let $H' = (V', E')$ be a $6$-uniform hypergraph. Let $H_0'$ be the strong product $H_0 \boxtimes P_2$.
We build $G = (V,E)$ an instance of $H$-Game as follows:
\begin{itemize}
\item {\bf Step 1}: For any vertex $v_i' \in V'$, we add a vertex $v_i \in V$.
\item {\bf Step 2}: For any edge $C \in E'$, we add $2(k-6)$ vertices $v^C_1, \dots, v^C_{2(k-6)}$ (note that if $k=6$ these vertices do not exist).
\item {\bf Step 3}: For any edge $C \in E'$, we add a copy $H^C_0$ of the graph $H_0'$ in $G$, and we connect any vertex of $H^C_0$ to all the vertices $v_i$ such that $v_i' \in C$ and to all the vertices $v^C_j$ for $1 \le j \le 2k-6$.
\end{itemize}
\begin{claim}
If Avoider has a winning strategy in $H'$, Alice has a winning strategy in $G$.
\end{claim}
\begin{claimproof}
Suppose Avoider has a winning strategy $\mathcal{S}'$ in $H'$. We define Alice's strategy $\mathcal{S}$ in $G$ as follows:
\begin{itemize}
\item She starts by playing the vertex $v_i$, corresponding to the vertex $v'_i$ that Avoider would have played in $H'$ according to $\mathcal{S}'$.
\item If Bob plays a vertex $v_i$, she answers with the vertex $v_{j}$ corresponding to the vertex $v_j'$ that Avoider would have played by $\mathcal{S}'$ in $H'$ if Enforcer has played $v_i'$.
\item In $H^C_0$, as it is a strong product $H_0 \boxtimes P_2$, Alice considers the pairing between any vertex and its copy in the strong product. If Bob plays one of them, she plays the second one.
\item For any edge $C$ in $H'$, Alice considers the set of vertices $v^C_j$ for $1 \le j \le 2k-6$. As there is an even number of them, if Bob plays one of them, she can also play one of them.
\end{itemize}
At the end of the game, by the matching strategy, for each clause $C$ in $H'$, Alice will have played exactly a copy of $H_0$ in each $H^C_0$, exactly $k-6$ vertices among the $v^C_j$ with $1 \le j \le 2(k-6)$, and the vertices $v_i$ corresponding to the $v_i'$ that Avoider would have played according to $\mathcal{S}'$ in $H'$.
Now, consider any copy $H_1$ of $H$ in $G$. Suppose that Alice has played all the vertices of $H_1$.
Suppose $H_0$ has at least one edge. We first prove that $H_1$ cannot contain two vertices $v_C$ and $v_{C'}$ for $C \neq C'$ that have been created by the Steps~$2$ or Step~$3$ of our construction. Suppose it does. Consider a decomposition of $H_1 = I_k \bowtie H^1_0$. By construction, the $v_C$s and the $v_{C'}$s are not adjacent. Therefore, as these two components are fully connected one to the other, they must either be both in $I_k$ or both in $H^1_0$.
\noindent{\bf Case 1:} $H_0$ has at least one edge.
$H_1$ also has one edge $e = (u_1,u_2)$. As $I_k$ is a stable set, we have $u_1,u_2 \in H^1_0$. Now, as $u_1$ can only be adjacent to vertices $v_i$s created by the vertices of $H'$(Step~1), $v_C$ and $v_{C}'$ cannot be both adjacent to $u_1$. Therefore, as we supposed $C \neq C'$, and as only vertices created during Step~1 can be adjacent to both $v_C$ and $v_{C'}$, any vertex in $I_k$ must be a vertex $v_i$. Which is not possible otherwise, Alice would have play $k \ge 6$ vertices $v_i$ adjacent to a same $v_C$, which means that, according to $\mathcal{S}'$ she would have played $k \ge 6$ vertices in the same hyperedge in $H'$, which contradicts the fact that $\mathcal{S}'$ was a winning strategy for Avoider in $H'$.
Now, as all the vertices of $H_1$ are either $v_i$s or were created by considering the same hyperedge $C$, and $|H_1| = |H_0| + k $, by the pairing, we know that exactly $6$ of them are $v_i$s. As there are no edges between the $v_i$s, they must all be on the same side of the join, and therefore, they are all connected to a same vertex. By construction, this is only the case if these six vertices were in a same hyperedge of $H'$ which contradicts that $\mathcal{S}'$ was a winning strategy for Avoider in $H'$.
\noindent{\bf Case 2:} $H_0$ has no edges and has $k'\ge 6$ vertices.
This means that $H$ can be written $I_k \bowtie I_{k'}$ for $k,k' \ge 6$ (note that $H$ is a complete bipartite graph). Once again, consider two vertices $v_C$ and $v_{C'}$ for $C \neq C'$ in $H_1$. As they cannot be adjacent, they must be both in $I_k$ or both in $I_{k'}$. Thus, $v_C$ and $v_{C'}$ have $\min(k,k') \ge 6$ common neighbours. This implies that, if $C \neq C'$, at least one of their common neighbour is not a vertex $v_i$ created during Step~1, otherwise Alice would have played six vertices in the same hyperedge of $H'$. This is not possible by construction. So once again, $H_1$ cannot contain $v_C$ and $v_{C'}$ created from different hyperedges from $H'$. Now, if Alice has played all the vertices of $H_1$, by construction as $|H_1| = k+k'$, and as her pairing strategy ensure her to play $k-6$ vertices created during step~2 and $k'$ created during step~3, necessarily, she has played six vertices $v_i$ creating during step~1. As there are no edges between these six vertices, they must all be in the same independent set $I_k$ or $I_{k'}$. Thus, they have a common neighbour. This common neighbour must then be a vertex $v_C$ creating during step~3 as only them are connected to the $v_i$s. Finally, these six vertices corresponds to six vertices $v_i'$s that are in the same hyperedge $C$ of $H'$. Once again, this contradicts the fact that $\mathcal{S}'$ was a winning strategy for Avoider in $H'$.
\end{claimproof}
\begin{claim}
If Enforcer has a winning strategy $\mathcal{S}'$ in $H'$, Bob has a winning strategy in $G$.
\end{claim}
\begin{claimproof}
Let $\mathcal{S}'$ be a winning strategy for Enforcer in $H'$. We consider a strategy $\mathcal{S}$ for Bob in $G$ as follows:
\begin{itemize}
\item If Alice plays a vertex $v_i$, he answers with the vertex $v_j$ that corresponds to the vertex $v_j'$ that Enforcer would have played in response to $v_i'$ in $\mathcal{S}'$.
\item In $H^C_0$, as it is a strong product $H_0 \boxtimes P_2$, Bob considers the pairing between any vertex and its copy in the strong product. If Alice plays one of them, he plays the second one.
\item For any edge $C \in H'$, Bob pairs the vertices $v^C_j$ for $1 \le j \le 2k-6$. If Alice plays one of them, he plays one of them too.
\item If at a certain moment of the game, it is Bob's turn and the remaining vertices are all in some $H^C_0$ or vertices $v^C_j$s, he applies the pairing strategy, so that Alice plays once in any pair of the matching by Lemma~\ref{pairing strategy}.
\end{itemize}
Consider the graph at the end of the game. As $\mathcal{S}$ was a winning strategy for Enforcer in $H'$, there exists a hyperedge $C \in H'$ in which Avoider has played the six vertices. Up to a renaming of the vertices, denote by $v_1', \dots, v_6'$ be these six vertices. Alice has then played $v_1, \dots, v_6$ in $G$. According to the pairing strategy, Bob knows that Alice will play exactly $k-6$ vertices from the $v^C_j$, denote them $v_7, \dots, v_k$, and exactly one copy $H_1$ of $H_0$ from the vertices of $H^C_0$. Now, by construction, the vertices $v_1, \dots v_k$ are a stable set and all the edges exist between any $v_i$ ($1 \le i \le k$) and any vertex $v$ of $H_1$. Thus, the subgraph formed by these vertices, which were all played by Alice, is isomorphic to $I_k \bowtie H_0 = H$. Thus, Bob has won.
\end{claimproof}
Finally, the $H$-Game played on $G$ is won by Alice if and only if the Avoider-Enforcer game played on $H'$ is won by Avoider, and determining the winner of the $H$-Game is {\sf PSPACE}-complete.
\end{proof}
Theses two games are two examples of games in which the proof of {\sf PSPACE}-completeness in Avoider-Enforcer convention is similar to the one in Maker-Breaker convention, in several other games the Maker-Breaker complexity proof can be adapted to an Avoider-Enforcer one.
\section*{Acknowledgments}
We want to thank Eric Duchêne and Aline Parreau for their help in the redaction of this paper
|
\section{Introduction}
Increasingly complex studies of transcriptome dynamics can be carried out now using high-throughput sequencing of reverse-transcribed RNA molecules. Such a procedure is typically called RNA-sequencing (RNA-seq). Studying RNA-seq data helps researchers gain a deeper understanding of how changes in transcriptional activity reflect various cell types and contribute to phenotypic differences. One of the ways to gain new insights from RNA-seq data is to identify groups (clusters) of co-expressed genes. This can help researchers target genes involved in similar biological processes, thus helping in the design of new pharmaceutical drugs, or finding genes that are candidates for co-regulation. Identification of clusters of co-expressed genes also helps characterize biological functions for orphan genes.
By now, a number of algorithms has been proposed for clustering RNA-seq data. Note that these data have a number of characteristics that makes modeling them rather difficult. First of all, they tend to be highly skewed and have a large dynamic range. Second, they typically demonstrate positive correlation between the gene length and read counts. Third, these data are almost always overdispersed (i.e. their variance is larger than their mean).
Most of the model-based clustering methods used for RNA-seq data typically view each cluster as represented by a distinct distribution, while the entire dataset is modeled as a finite mixture of these distributions. One of the advantages of this approach is that it allows a researcher to assess the appropriate number of clusters, the distance between clusters, and test hypotheses about these quantities. In the existing literature, most of these models tend to be parametric; in other words, it is assumed that distributions of individual biological samples are modeled as belonging to a particular distribution. The choice of distribution ranges from Poisson \cite{rau2015co} to negative binomial \cite{si2014model} to complicated distributions built specifically for the given task, such as multivariate Poisson-lognormal \cite{silva2019multivariate,subedi2020parsimonious}. The choice that has to be made here is a rather difficult one as there are few ready-made goodness of fit tests for most multivariate discrete distributions. As a practical concern, many practitioners may find it difficult to make such a choice in practice. Thus, an approach that avoids making such a choice altogether will be beneficial.
In this manuscript, we suggest an alternative way of handling the uncertainty in choosing a distribution for biological samples. Instead of trying to make a choice between various parametric families, we suggest viewing these distributions as ``just" probability density functions that do not belong to any such family. The nonparametric approach to multivariate mixture modeling and clustering has been studied in statistics for some time; several methods have been proposed to fit such models and establish the structure of relevant clusters \cite{benaglia2009like, levine2011maximum, chauveau2015semi}. However, their use in bioinformatics in general has been very limited (see e.g. \cite{anchang2014ccast}); to the best of our knowledge, they have not been used to cluster the data at all.
In this work, we suggest the use of an algorithm that can fit a general multivariate nonparametric mixture model with conditionally independent marginals in the RNA-seq data context. This method is a so-called npMSL (the nonparametric Maximum Smoothed Likelihood) method that was originally proposed in \cite{levine2011maximum}. The corresponding algorithm is an MM (Maximization-Minorization) algorithm that possesses the monotonicity property, similarly to the EM algorithm, and is guaranteed to converge.
The remainder of this article is organized as follows. In the Methods section, we describe the model used in detail, introduce the necessary notation, and define the algorithm that is to be used to perform the clustering task. Here, we also describe the model selection procedure that we use to select an appropriate number of clusters. In the Real Data Analysis section, we describe in detail two datasets that we use to illustrate our approach. The results are illustrated using both a simple visualization method and the results of a GO functional enrichment analysis. Finally, some concluding remarks are provided in the Discussion section.
\section{Methods}
Let $Y_{ij}$ be the random variable corresponding to the digital gene expression measure for a biological entity $i$ ($i=1,\ldots,n)$ of condition $j$ $(j=1,\ldots, d).$ We denote the corresponding observed value $y_{ij}.$ This setting implies that the data $\vv y$ is the $n\times d$ matrix of the digital gene expression for all observations and variables. Also, $\vv y_{i}$ is the $d-$ dimensional vector of digital gene expression for all variables of observation $i.$
\subsection{Nonparametric mixture model with conditionally independent measurements}
Historically, it has been common to use parametric mixture models to cluster RNA-seq data (see e.g.\cite{si2014model,rau2015co,rau2018transformation}). To the best of our knowledge, nonparametric approach to model-based clustering of RNA-seq data has not been tried before. At the same time, it has been used in many other statistical application areas e.g. developmental psychology, hydrology, and others \cite{chauveau2015semi, shen2018mm}. This approach assumes that the density functions of clusters do not come from a particular parametric family (e.g. Gaussian, Poisson etc.). More specifically,
the data are assumed to come from $m$ distinct clusters with the density of the $k$th cluster, $k=1,\ldots,m$ being $f_{k}.$ It is further assumed that each one of these densities $f_{k}$ is equal with probability $1$ to the product of its marginal densities:
\begin{equation}\label{marg_independence}
f_{k}({\vv y_i})=\prod _{j=1}^{d}f_{kj}(y_{ij})
\end{equation}
with $\vv y_i=(y_{i1},\ldots,y_{id})^{'}.$ Taking a fully nonparametric approach with regard to the $f_{kj}$, we may therefore express the density function of any observation $\vv Y_i$ according to a nonparametric finite mixture model as
\begin{equation}\label{model}
\vv Y_i \ \sim \
g_{\boldsymbol\theta}({\vv y}_{i})\ =\ \sum_{k=1}^{m}\pi_{k}\prod_{j=1}^{d}f_{kj}(y_{ij}),
\end{equation}
where $\boldsymbol\pi=(\pi_{1},\ldots,\pi_{m})$ must satisfy
\begin{equation}\label{lambdaconstraints}
\sum_{k=1}^m \pi_k=1 \quad\mbox{and each $\pi_k\ge 0$.}
\end{equation}
Here, we assume $\vv Y_i = (Y_{i1}, \ldots, Y_{id})^{'}$ and we let
$\boldsymbol\theta$ denote the vector of parameters to be estimated,
including the mixing proportions $\pi_{1},\ldots,\pi_{m}$ and the
univariate densities $f_{kj}$. For convenience, we will also use the notation $\vv f$ for a vector of all the marginal densities $\{f_{kj}\},$ $j=1,\ldots,d,$ $k=1,\ldots,m.$ Furthermore, throughout this article, $k$ and $j$ always denote the component and coordinate indices, respectively;
thus, $1\le j\le d$ and $1\le k \le m.$ Thus, the overall population of the sample size $n$ is distributed according to
\begin{equation}\label{prod_marg}
f(\vv y_i; m,\boldsymbol\theta)=\prod_{i=1}^{n}\sum_{k=1}^{m}\pi_{k}\prod_{j=1}^{d}f_{kj}(y_{ij}).
\end{equation}
The conditional independence may seem rather limiting at first sight. However, it may be thought of as a simplification of the commonly used repeated measures random effects model. In such a model, one usually assumes that the multivariate observations on an individual are independent, conditional on the identity of the individual in question. Here, the individual-level effects are replaced by the entity (gene)- level effects. Note also that the conditional independence assumption has been used on a number of occasions when modeling RNA-seq data using parametric mixtures e.g. \cite{rau2015co,si2014model}. On the other hand, no individual marginal density $f_{jk}$ is assumed to have come from a family of densities indexed by a finite-dimensional parameter vector, such as Gaussian, Student, etc. Such a nonparametric approach represents a substantial generalization compared to approaches typically used in bioinformatics literature.
\subsection{Inference}
We are going to start with some needed notation. First, let $Z_{ik}\in \{0,1\}$ be a Bernoulli random variable indicating that an individual $i$ comes from the component $k.$ Because each individual comes from exactly one component, it implies that $\sum_{k=1}^{m}Z_{ik}=1.$ Therefore, the complete data would be the set of all $(\vv x_i,\vv Z_i),$ where $1\le i \le n$. This suggests that the entire data can be viewed as consisting of observable and unobservable parts. Therefore, an EM-type algorithm seems to be an appropriate option for estimating of parameters of the model \eqref{model}.
\cite{levine2011maximum} introduced an algorithm that minimizes a smoothed loglikelihood function of the data produced by the model \eqref{model}. This algorithm has a provable monotonicity property. We only give a brief description of this algorithm; for detailed discussion of its monotonicity property see \cite{levine2011maximum}. For the purpose of this discussion, some additional notation is needed.
Let $K(\cdot)$ be a kernel density function on the real line. With a slight abuse of notation, let us
define the product kernel function in the $d$-dimensional space as $K({\vv u})=\prod_{j=1}^{d}K(u_{j})$ and its rescaled version $K_{h}({\vv u})=h^{-d}\prod_{j=1}^{d}K(h^{-1}u_{j})$ for a positive parameter $h$ that is commonly called the bandwidth. Furthermore, we smooth a function $f$ using the following smoothing operation ${\mathcal S}f(\vv y)=\int K_h(\vv y-\vv u)f(\vv u)\,d\vv u.$ The same smoothing operation can be applied to an $m$-dimensional vector of functions by defining ${\mathcal S}\vv f = ({\mathcal S}f_1, \ldots, {\mathcal S}f_m)^{'}.$ We also define a nonlinear smoothing operation ${\mathcal N}$ as
\[
{\mathcal N} f(\vv y)\ =\ \exp\left\{ ( {\mathcal S} \log f )(\vv y) \right\}
\ =\ \exp \int K_h(\vv y- \vv u) \log f(\vv u) \, d\vv u.
\]
To simplify notation, we introduce the finite mixture operator ${\mathcal M}_{\boldsymbol\pi}\vv f(\vv y):=\sum_{k=1}^{m}\pi_{k}f_{k}({\vv y}),$ whence we obtain ${\mathcal M}_{\boldsymbol\pi}\vv f(\vv y) =g_{\boldsymbol\theta}(\vv y).$ Also, we denote ${\mathcal M}_{\boldsymbol\pi} {\mathcal N}\vv f(\vv y):=\sum_{k=1}^{m}\pi_{k}{\mathcal N}f_{k}({\vv y}).$ With this notation in mind, we define the following algorithm. Given initial values $(\vv f^0, \boldsymbol\pi^0),$ iterate the following three steps for $t=0,1,\ldots:$
\begin{itemize}
\item{\bf E-step:\ }
Define, for each $i$ and $k$,
\begin{equation}\label{wij}
w_{ik}^t
= \frac{\pi^t_k {\mathcal N} f_k^t (\vv y_i) } {{\mathcal M}_{\boldsymbol\pi^t}
{\mathcal N}\vv f^t (\vv y_i) }
= \frac{\pi^t_k {\mathcal N} f_k^t (\vv y_i) } {\sum_{a=1}^m
\pi_a^{t} {\mathcal N}f_{a}^t (\vv y_i) }.
\end{equation}
\item
{\bf M-step, part 1:\ }
Set
\begin{equation}\label{lambda}
\pi_k^{t+1} = \frac{1}{n}\sum_{i=1}^n w_{ik}^t
\end{equation}
for $k=1, \ldots, m$.
\item
{\bf M-step, part 2:\ }
For each $j$ and $k$, let
\begin{align}\label{densest}
&f_{kj}^{t+1}(u) = \frac { \sum_{i=1}^n w_{ik}^t
K_h\left(u-y_{ij}\right) }
{\sum_{i=1}^n w_{ik}^t }
\\
&=\frac {1}{nh\pi_{k}^{t+1}} \sum_{i=1}^n w_{ik}^t
K\left(\frac{u-y_{ij}}{h}\right)\nonumber.
\end{align}
\end{itemize}
Let us define the following functional of $\boldsymbol\theta$ (and, implicitly, $g$):
\begin{equation}\label{elldefn}
\ell(\boldsymbol\theta)=\int g({\vv y})\log\frac{g({\vv y})}{[{\mathcal M}_{\boldsymbol\pi}{\mathcal N}\vv f]({\vv y})}
\,d{\vv y}.
\end{equation}
This functional represents conceptually a smoothed negative log-likelihood. Then, \cite{levine2011maximum} shows that the value of this functional decreases at each step of the introduced algorithm. This algorithm will be referred to as npMSL (nonparametric Maximum Smoothed Likelihood) algorithm.
The npMSL algorithm can also be generalized to a model where there are blocks of coordinates that are identically distributed (in addition to being conditionally independent). If we let $b_{j}$ be the block index of the $j$th coordinate, where $1\le b_{j}\le L$ and $L$ is the total number of such blocks, then the model \eqref{model} is modified as
\[
g_{\boldsymbol\theta}({\vv y}_{i})\ =\ \sum_{k=1}^{m}\pi_{k}\prod_{j=1}^{d}f_{kb_{j}}(y_{ij}).
\]
If all the blocks have the size $1,$ we are back to the original model \eqref{model}. The nonlinear smoothing operator ${\mathcal N}f_{k}=\prod_{j=1}^{d}{\mathcal N}f_{kb_{j}}$ and definitions of ${\mathcal M}_{\boldsymbol\pi}\vv f$ and ${\mathcal M}_{\boldsymbol\pi}{\mathcal N}\vv f$ remain unchanged. The only element of the algorithm that actually needs an update is the density estimation step \eqref{densest}. For the $k$th component and block $l\in \{1,\ldots,L\},$ we now have
\[
f_{kl}^{t+1}(u)=\frac{\sum_{j=1}^{d}\sum_{i=1}^{n}w_{ik}^{t}I_{b_{j}=l}K_{h}(u-y_{ij})}{\sum_{j=1}^{d}\sum_{i=1}^{n}w_{ik}^{t}I_{b_{j}=l}}
\]
where $I_{b_{j}=l}$ is an indicator function of the event $b_{j}=l$.
The block version of the npMSL algorithm is the one that we have used in our study. In other words, when applying the npMSL method, we assume that replicates of the same condition represent a block of identically distributed coordinates.
\subsection{Bandwidth and kernel function selection issues}
Of the two issues - kernel and bandwidth selection - mentioned in the heading above, the first is a simpler of the two. There seems to be a general consensus in the literature on the density estimation that the choice of the kernel function does not matter much, at least in terms of efficiency of resulting estimators; see e.g. \cite{scott2015multivariate} for a general discussion of this issue.
On the contrary, sensible choice of the bandwidth $h$ is a challenging problem. The form of the npMSL algorithm introduced here assumes that $h$ is the same for each component and coordinate, or block. It is straightforward to introduce component- and block-specific bandwidths $h_{jl}$. Note also that individual component densities are not observed in the mixture setting. This fact complicates selection of the bandwidth in the mixture setting.
From the practical viewpoint, we found out that selecting a constant bandwidth according to the so-called Silverman's rule of thumb \cite{silverman2018density}, p. 48, works reasonably well. This method suggests choosing the bandwidth as
\begin{equation}\label{Silverman}
h=0.9(nd)^{-1/5}\min\left\{SD,\frac{IQR}{1.34}\right\}
\end{equation}
where $SD$ is the standard deviation, $IQR$ the interquartile range, and $nd$ is the size of the entire dataset.
Note that this is a rather crude method in the nonparametric mixture
setting. It is possible that it may result in under- or oversmoothing. First, pooling all of the data implies that one treats all of the mixture components as though they are from the same distribution. This can lead to an inflation of the bandwidth, especially if the mixture components’
centers are well-separated. This is true because, in such a case, the variability of the pooled
dataset will be larger than that of the individual components. Similarly, if
the vector coordinates are not identically distributed within each component/block,
the bandwidth could be biased upward for the same reason.
Yet operating in the opposite direction is the fact that the expression $nd$ in the formula defining the bandwidth above is an overestimate of the ``true" sample size. One can think of the ``true" sample size from each component being approximately equal to $\lambda_{k}nd$.
The arguments above show first of all that it would be useful to know
something about the mixture structure in order to select a bandwidth. This
suggests an iterative procedure in which the value of $h$ is modified, and
the algorithm reapplied, after the output from the algorithm is obtained. This, however, is going to result in the violation of monotonicity property of npMSL algorithm. For a more detailed discussion of this topic, see \cite{benaglia2011bandwidth}. As the above suggests, a careful exploration of the bandwidth selection question is a research topic unto itself. Thus, to make our application of npMSL algorithm to the analysis of RNA-seq data simpler, we are only using the constant bandwidth value selected according to \eqref{Silverman}.
\subsection{Selecting the number of clusters}\label{Model_Selec}
The npMSL algorithm assumes that the number of clusters is known in advance. This is almost never true when working with biological data. Unlike the case of parametric mixture models, there are few if any approaches to determining the number of clusters in a nonparametric mixture model. Some preliminary results in this direction have been obtained in \cite{kwon2021estimation}. We suggest the following approach to this problem. Recall that our nonparametric approach to clustering implies, as a first step, fitting a repeated measures model \eqref{model}. In light of this, it seems reasonable to fit first a {\it Gaussian} repeated measures mixture model over a range of possible number of clusters. To do this, we use the function repnormmix.sel of the mixtools package \cite{benaglia2009mixtools}. For every choice of the number of clusters, the initial cluster means were generated from a multivariate normal distribution of the correct dimensionality with the diagonal covariance matrix. We used the default settings of the function repnormmix.sel whereby the mean is determined by a normal distribution according to a binning method done on the data while the reciprocal of the variance has random standard exponential entries also according to a binning method done on the data. The initial probability weights were generated from a uniform Dirichlet distribution. For more details see the package description at \cite{benaglia2009mixtools}. The procedure used generated values of four model selection criteria - Akaike’s information criterion (AIC), Schwartz’s Bayesian information criterion (BIC), Bozdogan’s consistent AIC (CAIC), and Integrated Completed Likelihood (ICL) in order to choose the optimal number of clusters \cite{bozdogan1987model,biernacki2000assessing}. In cases where different criteria suggested different choices for the number of clusters in a model, we used the choice suggested by at least two (out of the four) criteria. We have not experienced a situation where all four of the criteria suggested different choices for the number of clusters.
\section{Real data analysis}
In the following, we illustrate the use of npMSL algorithm using two real RNA-seq datasets. Note that it is not possible to compare the co-expression results obtained using these two methods to a ``true" clustering of the data, as, in general, such a classification does not exist. In order to identify whether the co-expressed genes seem to be implicated in similar biological processes, we conduct functional enrichment analysis of gene ontology (GO) terms for the clusters identified by the suggested methods. The data that we are using are a mouse RNA-seq dataset consisting of lung, kidney, liver, and small intestine tissues \cite{Tsoucas:2019tx} and a prostate cancer cell-line RNA-seq dataset \cite{liu2019PCa}. The data are written in the matrix form where each row corresponds to a gene and each column to an experimental condition. The row names are the ENSEMBL gene names for each gene (ENSEMBL is a genome database project that is a scientific project of the European Bioinformatics Institute). Each row constitutes a digital gene expression of a particular gene across a set of cell lines (together with replicates) in our case. The goal is to cluster digital gene expression profiles in order to discover networks of co-expressed genes.
Unless explicitly stated to the contrary, both datasets are normalized first. The normalization procedure that we use is commonly called FPKM (Fragments Per Kilobase of transcript per Million mapped reads). Let us denote the result of this procedure $Y_i$ for the $i$th gene. Then, it is defined as
\[
Y_i = \frac{X_i}{N*s_i}*10^9
\]
where $N$ is the total number of reads sequenced, $s_i$ is the length of gene i, and $X_i$ is the number of counts for the $i$th gene. This type of normalization is commonly used for visualization and clustering. It is necessary if we want to be able to perform within sample comparisons (“gene A is expressed higher or lower than gene B”) because it is, effectively, a procedure that normalizes for gene length. Such a normalization is in order because, the longer the gene’s length is, the more fragments (“reads”) we sequence from that gene.
The data sets have also been filtered using a cutoff of 1.5 CPM (counts per million), in which gene read counts are divided by the sum of read counts for a given sample and multiplied by a million. The CPM of the $i$th gene, denoted as $CPM_i$ is defined as
\[
CPM_i = \frac{X_i}{N}*10^6
\]
where $N$ is the total number of reads sequenced for a given sample and $X_i$ is the number of counts for the $i$th gene.
\subsection{Mouse tissue dataset}
The first dataset that we focus on is a dataset in which bulk RNA-seq was performed to profile the gene expression profiles in lung, kidney, liver, and small intestine tissues from six- to ten-week-old male C57BL/6J mice. As gene expression data is highly tissue-specific, data such as these can allow for the identification of tissue-specific expression modules. Two biological replicates are present for each tissue type and separate mice were used for each replicate. Observations in these datasets are commonly referred to as ``counts" where a count is the number of reads that align to a particular feature(gene). Filtering at the level of $1.5$ CPM as described above resulted in a file containing $116,512$ genes. These data are available through the Gene Expression Omnibus (GEO) repository through accession number GSE124419.
It has been also noted earlier that the RNA-seq data often have a very large dynamic range and tend to be heavily skewed. This often represents a significant problem when modeling these data. Therefore, before modeling the data, we applied a logarithmic transform to it as a first step.
To select an appropriate number of clusters, we fit several Gaussian repeated measures models over a range of cluster numbers from $1$ to $20,$ as described earlier (in more details) in the Section \eqref{Model_Selec}. The comparison of the four model selection criteria suggests using the majority vote criterion that $17$ is an optimal number of clusters for the npMSL method. In this case, the procedure was not sensitive to the choice of starting values.
As is well known, visualizing results of a co-expression analysis for RNA-seq data can be rather complicated. This is due to the extremely large dynamic range of digital gene expression and the fact that the more highly expressed genes tend to exhibit much higher variability than weakly expressed genes. The most appropriate manner in which the results of a co-expression analysis for RNA-seq data should be displayed is still an open research question. In this manuscript, we follow the visualization approach that is conceptually similar to what has been suggested first in \cite{rau2015co}. In this approach, bar widths correspond to the estimated proportions for the corresponding cluster $\hat\pi_{k}.$ The proportion of reads that is attributed to each cell line in each cluster is represented by the corresponding colored segment within each bar. More specifically, let $y_{ijq}$ be the read of the $i$th gene for the $q$th replicate of the $j$th condition where $j=1,\ldots,J$, $i=1,\ldots,I_{k}$, $q=1,\ldots,3$ and $I_{k}$ is the number of genes in the $k$th cluster. Then, the height of the vertical bin that corresponds to the $j$th condition in $k$th cluster is
\[
\lambda_{jk}=\frac{\sum_{i=1}^{i_{k}}\sum_{q=1}^{3}y_{ijq}}{\sum_{k=1}^{K}\sum_{i=1}^{i_{k}}\sum_{q=1}^{3}y_{ijq}}.
\]
This approach to visualization allows us to assess the relative size of clusters with different level of expression in different tissues. For example, in different clusters rather easily. For example, one can see that clusters that contain mostly genes that are strongly expressed in small intestine and lungs tend to be larger than those that contain mostly genes that are expressed in liver or kidneys. The visualization results are presented in the Figure \eqref{fig:Mouse}.
\begin{figure}
\centering
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{npMSL_lambda_plot_mouse}
\caption{Cluster behavior for the mouse tissue dataset:npMSL}
\label{fig:npMSL_mouse}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{Poisson_lambda_plot_mouse}
\caption{Cluster behavior for the mouse tissue dataset:Poisson method}
\label{fig:Poisson_mouse}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{Logit_Transformation_lambda_plot_mouse}
\caption{Cluster behavior for the prostate cancer cell line dataset:Transformation Method} \label{fig:Transformation_mouse}
\end{subfigure}
\caption{Three cluster visualization plots for the mouse tissue dataset}
\label{fig:Mouse}
\end{figure}
In order to determine the biological relevance of our clustering results, we performed gene ontology (GO) enrichment analyses on genes in clusters. This analysis identifies enrichment of biological processes amongst genes in clusters, thus investigating whether genes that our methods identify as co-expressed encode proteins that perform similar biological functions. Initially, methods were applied to the mouse dataset to identify co-expression clusters. Out of the $17$ clusters identified by the npMSL method, $16$ of them ($94.1$\%) show significant enrichment of biological processes. The clusters identified are associated with highly specific biological processes. For example, cluster $2$ involves genes that function in water homeostasis (GO:0030104). Cluster $12$ is clearly associated with processes involved in the adaptive immune response (GO:0002250), and also includes specific terms such as regulation of T cell activation (GO:0042110) and B cell activation (GO:0042113). Cluster $16$ on the other hand involves GO terms related to the innate immune response, such as the innate immune response in mucosa (GO:0002227) and the defense response to Gram-positive bacterium (GO:0050830). For this dataset, all three methods produce similar results: the Poisson method produces $19$ clusters that are all associated with statistically enriched GO terms while the logit transformation method results in the lowest performance, with $22$ out of $25$ clusters ($88.0$\%) enriched for biological processes.
\subsection{Human prostate cancer cell line dataset}
Again, we begin with the description of the dataset. We have four cell lines used that consist of samples that are sensitive to chemotherapeutic agents (C4-2 cells and LNCaP cells) as well as those that are resistant (MR49F and C4-2B cells). More specifically, a line of C4-2 cells is a subline of LNCaP cells (described below). Second, C4-2B cells are enzalutamide-resistant cells derived from C4-2 cells. Third, LNCaP (Lymph Node Carcinoma of the Prostate) cell line is the line that has been established from a metastatic lesion of human prostatic adenocarcinoma. Finally, the MR49F cell line consists of enzalutamide-resistant cells derived from LNCaP cells. The resistant cell lines are of interest because these cells do not respond well to treatment. Also, each of the cell lines has three replicates. Filtering at the level of $1.5$ CPM as described above results in a file containing sequencing data for $16,247$ genes. As before, we also applied the log transform to the data as a first step in our data analysis.
Note that it is unlikely that replicates in a cell line dataset are independent. If so, this dataset may violate a conditional independence model assumption that underlies the npMSL method. Due to this, an attempt to run our method on such a dataset may be viewed as a test of the algorithm's robustness to the violation of the conditional independence assumption. Indeed, we discovered that the algorithm tends not to converge if the dataset is used ``as is" even when multiple choices of starting values are considered. Therefore, we decided to treat zeros as missing observations and use a simple imputation procedure, which turned out to be useful. More specifically, every zero was substituted with a number generated from a uniform distribution on an interval $[0,A]$ where $A$ was the smallest count observed in the entire dataset. We would like to note here that the practice of treating zeros as missing observations, although not very common in the analysis of RNA-seq data, is quite widespread when analyzing the scRNA-seq (Single Cell RNA-seq) data \cite{hou2020systematic}.
As for the previous dataset, we fit several Gaussian repeated measures models over a range of cluster numbers from $1$ to $20$. Once again, the comparison of the four model selection criteria suggests, using the majority vote criterion, that $17$ is an optimal number of clusters for the npMSL method. In this case, the procedure was not sensitive to the choice of starting values.
To see if a solution with a larger number of clusters may also be possible, we also tried to run this procedure over the range of $1$ to $30$ clusters. In this case, selection of a starting point turns out to be a difficult problem since many choices result in singularities. To avoid this problem, we ran a small-EM type initialization procedure using the repnormmixmodel.sel procedure from the R mixtools package \cite{benaglia2009mixtools}. The concept of small-EM type initialization procedure is described in detail in Section 2.3.2 of \cite{rau2015co}; see also \cite{biernacki2003choosing} for additional information. In our case, the only two differences were the use of the repeated measures model and the use of only $5$ iterations of the EM algorithm. With this initialization procedure, we found a set of starting values that did not result in a singularity; however, the suggested solution had only $11$ clusters that were not well separated. In the final analysis, we decided to keep the $17$ cluster solution that we described earlier as an optimal one.
For visualization purposes, we followed the same approach as for the previous dataset. The results are given in the Figure \eqref{fig:Cell_Line}.
\begin{figure}
\centering
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{npMSL_lambda_plot}
\caption{Cluster behavior for the prostate cancer cell line dataset:npMSL}
\label{fig:npMSL}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{Poisson_lambda_plot}
\caption{Cluster behavior for the prostate cancer cell line dataset:Poisson method}
\label{fig:Poisson}
\end{subfigure}
\hfill
\begin{subfigure}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{Arcsin_Transformation_lambda_plot}
\caption{Cluster behavior for the prostate cancer cell line dataset:Transformation Method} \label{fig:Transformation}
\end{subfigure}
\caption{Three cluster visualization plots for the prostate cancer cell line dataset}
\label{fig:Cell_Line}
\end{figure}
For comparison purposes, we compare clustering results obtained using the npMSL method with two other methods commonly used for model-based clustering of RNA-seq data. The first of these methods is the Poisson method of \cite{rau2015co} while the second one is the transformation-based method of \cite{rau2018transformation}. We believe that the Poisson method is an appropriate comparison benchmark since it, alongside our methods, assumes conditional independence of marginals. Our method simply takes this assumption one step further and does not impose any specific distributional assumption on the marginal distributions. Note that the Poisson distribution only models the integer-valued data. Thus, when applying the Poisson method, we have used the raw data counts instead of the FPKM normalized dataset. We used the so-called slope heuristics method to select the optimal number of clusters in this case; our approach resulted in the choice of $18$ clusters when running the slope heuristics approach over a range from $1$ to $35$ clusters. The slope heuristics method has been suggested in \cite{birge2001gaussian,birge2007minimal} while some of the practical suggestions concerning its use can be found in \cite{baudry2012slope}.
Another method that we used for comparison purposes has been the transformation method of \cite{rau2018transformation}. This method is based on the use of data transformations in conjunction with Gaussian mixture models. It also uses a penalized model selection criterion to select the number of clusters present in the data. Compared to the Poisson method, it allows for modeling of per-cluster correlation among biological samples. To analyze our dataset, we used this method with an arcsin transform. The number of clusters was selected using the so-called ICL (Integrated Complete Likelihood) method \cite{biernacki2000assessing} which is a type of the penalized selection criterion method. Using this criterion over a possible range of $1$ to $40$ clusters, we found that $2$5 seemed to be the optimal number in this case.
All three clustering methods identify fewer biologically relevant clusters when applied to this cell line dataset compared to the mouse dataset. Furthermore, many of the enriched biological processes identified are at a higher level and are thus less specific as compared to those identified in the mouse dataset. Overall, the npMSL method identifies more biologically relevant clusters than either the Poisson or the arcSin transformation methods. Out of $17$ clusters identified by the npMSL method, $14$ ($82.4$\% of clusters) show enrichment of biological process GO terms. For example, cluster $1$ is associated with terms related to translation initiation (GO:0006413) such as cotranslational protein targeting to membrane (GO:0006613) and SRP-dependent cotranslational protein targeting to membrane (GO:0006614). Cluster $6$ is associated with non-coding RNA processing (GO:0034470). Cluster $10$ is heavily enriched for mRNA catabolic processes(GO:0006402) and cluster $14$ is involved in tRNA processing (GO:0008033) including tRNA modifications (GO:0006400) and tRNA metabolic processes (GO:0006399). In comparison, the arcSin transformation method resulted in $12$ out of $35$ clusters ($34.2$\% of the clusters) showing a significant enrichment of GO terms and the clusters derived from the Poisson method results in only $5$ out of $18$ clusters ($27.8$\% of clusters) with enrichment of GO terms.
\section{Discussion}
In this manuscript, we proposed a possible method of discovering gene co-expression networks in RNA seq data. The suggested method entails the use of a rigorous framework for parameter estimation, based on MM (Maximization-Minorization) algorithms. The model we use is distinct from most models routinely used in the clustering of digital gene expression profiles since it is a nonparametric one. More specifically, we assume that, conditional on knowing the cluster an observations has been generated from, biological samples (corresponding to marginal distributions) are independent. Moreover, each marginal distribution is not assumed to belong to any predetermined distributional family.
Conceptually, our proposal is closest to that of \cite{rau2015co} that also uses model with conditionally independent biological samples but imposes a Poisson restriction on marginal distributions. The Poisson assumption may not be very realistic in practice since it does not account for the overdispersion routinely observed in the RNA-seq data. The attempts to handle the overdispersion problem from the parametric viewpoint typically concentrate on proposing specific distribution to handle it such as e.g. negative binomial in \cite{si2014model} or the multivariate Poisson-lognormal in \cite{subedi2020parsimonious}. By comparison, our method is a more general one since it imposes no specific assumptions on the marginal distributions of individual biological samples. Moreover, fitting per-gene dispersion parameters using the experimental data with multiple conditions is typically rather difficult due to a small number of replicates available in such datasets. Our proposed method avoids this problem altogether by avoiding the parametric framework for marginal distributions. The use of nonparametric multivariate clustering in bioinformatics has so far been extremely limited \cite{anchang2014ccast}; to the best of our knowledge, these models have not been used for the discovery of gene co-expression networks before.
The algorithm that we use to fit the suggested model has been described earlier in the statistical literature \cite{levine2011maximum} and its use illustrated in various simulation settings. Moreover, there does not seem to be a general agreement on how to simulate the data resembling the true RNA-seq data in an optimal way. Due to this, we do not include a simulation study in this manuscript. Instead, we simply demonstrate the behavior of our method on two real datasets: a mouse tissue dataset and a human prostate cancer cell line dataset. The results are compared to those obtained by applying the Poisson method of \cite{rau2015co} and a transformation method of \cite{rau2018transformation} to those same datasets. In both cases, the npMSL method seems to identify a number of biologically meaningful clusters that is at least comparable to the number produced by other methods; in the human prostate cancer cell-line dataset case, it outperforms both alternative methods. For both datasets, the performance of the methods has also been compared using the adjusted Rand index (ARI) \cite{hubert1985comparing}.
The results are summarized in Tables \eqref{Table1} and \eqref{Table2}. The most salient feature of this summaries is that the clustering produced by the nonparametric npMSL method is very different from those produced by either Poisson method or a transformation-based method. In other words, these seem to represent different solutions of a problem. We would like to note here that in the only case we are aware of where a researcher tried to applied a nonparametric method to clustering of the RNA-seq data the conclusion was a rather similar one - the clustering produced has been quite different from those produced by parametric methods \cite{mazo2019constraining}.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
& npMSL & Poisson & Transformation \\
\hline
npMSL & 1 & 0.123015052
& 0.089651457
\\
\hline
Poisson &0.123015052
&1 & 0.22480933
\\
\hline
Transformation &0.089651457
& 0.22480933
&1 \\
\hline
\end{tabular}
\caption{ARI values for the mouse tissue dataset}
\label{Table1}
\end{table}
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
& npMSL & Poisson & Transformation \\
\hline
npMSL & 1 &0.013246073
& 0.026826513
\\
\hline
Poisson &0.013246073
&1 & 0.110776115
\\
\hline
Transformation &0.026826513
&0.110776115
&1 \\
\hline
\end{tabular}
\caption{ARI values for the human prostate cancer cell line dataset}
\label{Table2}
\end{table}
The suggested model does not take into account possible dependence between biological samples because it enforces the conditional independence between them. The natural next step in this line of research is to try to lift this assumption while preserving the general nonparametric nature of marginal distributions in our proposed model. One way this can be done is by considering explicit dependence structures between biological samples modeled using the so-called copula functions \cite{mazo2019constraining}. A useful direction of the future research will be the introduction of specific multivariate copulas such as e.g. Gaussian copula to model the dependence between biological samples in RNA-seq data. Note that this approach allows a researcher to model any type of dependence and not just correlation which is often claimed to be the benefit of using certain so-called hidden layer approaches \cite{silva2019multivariate}. Our research in this direction is ongoing.
\section{Funding}
This work was supported in part by funds from the Purdue University Center for Cancer Research[P30CA082709]; the IU Comprehensive Cancer Center [P30ca082709]; and the Walther Cancer Foundation.
\bibliographystyle{unsrt}
|
\section{Introduction}
Researchers interested in determining the features and dynamics of a specific environment have access to vast quantities of digital data from which to derive answers to their inquiries. The amount of written news content available on the internet alone provides a plethora of knowledge, context, and documentation that may be utilized to solve social-scientific issues. Even at the computational level, sifting through this amount of data and highlighting conversations and patterns of special interest is difficult, let alone for individual scholars to explore. Topic models are Bayesian statistical models that have proven their accuracy in many applications \cite{textmining}. Given a large corpus, these models permit the extraction of topics that structure the texts, and the topics themselves can be simplified to a list of keywords. Dirichlet-multinomial regression (DMR) topic model \cite{mimno2012topic} with document features such as author, references, dates, etc. can enhance the performance of such topic models.
\section{Related Works}
The present study aims to measure how the robustness of terms, derived from observed topics occurring in the corpus manually, may evolve temporally and correspond with known events described in the corpus' component papers.
In comparison to a study of current relevant methodologies, this analysis borrows and improves on previous methods to provide three key strengths in a novel approach to solving the stated problem: Capturing meaningful language use beyond the level of a word, the use of predefined topics to empower framework-based analysis and considering the temporal change in n-gram usage across a corpus of documents over a time period of interest. Kherwa and Bansal \cite{Kherwangram}, though using a topic modeling approach, similarly harness the semantic power of n-grams; they cite the difference between the words ``New'' and ``York'' occurring separately and ``New York'' as a bigram. A term-frequency measurement similar to Don et al. \cite{59394777347e41faa5b6431040808dfb} is applied, but scaled to the level of a corpus, as opposed to single documents with distinct models. Ahmed, Traore, and Saad \cite{Ahmed2017DetectionOO} note an epistemological limitation in their analysis: Identification of terms and trends of interest in text data typically requires some prior knowledge of what is to be found.
Bamler et al. \cite{bamler2017dynamic} presented a language model for probabilistic temporal text data which tracks the semantic evolution of individual words over time. The model uses Word2Vec \cite{mikolov2013efficient} to create embeddings, which are connected through inference algorithms that allow training the model jointly over all time periods. Word embedding trends are more interpretable and lead to higher predictive likelihoods than competing methods that are based on static models trained separately on time slices. This method is used to evaluate the change in the semantics of each word over time.
Topic modeling is useful for discovering latent topics that can later be associated with known trends or events. To attribute text in a document to a specific topic, LDA \cite{10.5555/944919.944937} is utilised. It creates a topic per document model and a words per topic model using Dirichlet distributions as the modelling framework. However, The approach in the present study employs a novel analysis, i.e hypothesizing that given a known set of topics, emergent trends or events can be discovered by tracking the temporal change between terms. Insights synthesized from the aforementioned research and development of a novel approach are used to test this hypothesis.
\section{Background}
When deciding what algorithms to use to model trends, the first consideration was towards the final output that needs to reach the end-user so they can gain insights. In the present study, the idea was to create a framework that would not only detect how much a topic evolves, but also in which way it is evolving semantically. In addition to the idea of giving the users flexibility in choosing their topic that is the ``\textit{umbrella}'' term, the users can decide the \textit{sub-topics} within the umbrella term as well. Thus, word embeddings and \textit{k}-means clustering became the immediate choices as the semantic changes in a word could be modeled over time and also for generating topics.
\subsection{Word2Vec Context}
Word2Vec \cite{mikolov2013distributed} can utilize either of two model architectures to produce a distributed representation of words: continuous \textit{bag-of-words} (CBoW) or continuous \textit{skip-gram}. In the continuous bag-of-words architecture, the model predicts the current word from a window of surrounding context words (window size). The order of context words does not influence prediction (bag-of-words assumption). In the continuous skip-gram architecture, the model uses the current word to predict the surrounding window of context words. The skip-gram architecture weighs nearby context words more heavily than most distant context words. The need for word prediction when the goal is to generate embeddings comes from the nature of language which is unlabeled. To overcome this, an artificial task is created for the neural network: predict a targeted word from its context and vice versa in a skip-gram. Although the inputs and outputs of the single layer neural network are not the objective, the weights that feed into the final softmax layer will be the eventual embeddings that convey the input word's semantic meaning. The softmax, multinomial distribution can be defined as follows:
\[P( w_j | w_i ) = \frac{exp({v'_{w_j}}^{T}*v_{w_i})}{\sum_{j'=1}^{V} exp({v'_{w_j'}}^{T}*v_{w_i})} \]
Where the probability of the target word w$_{j}$ given the surrounding word(s) w$_{i}$ is maximized. Also note that v$_{w}$ and v'$_{w}$ are two representations of the word ``w''. v$_{w}$ comes from rows of the input hidden weight matrix, and v'$_{w}$ comes from the columns of the hidden output matrix. Using gradient descent, the embeddings can be optimized without having a labeled dataset \cite{word2vecparameterrong}. Since this model is trained in an online setting, the goal is to take a small step mediated by the ``learning rate'' to minimize the distance between the current vectors for w$_{j}$ and w$_{i}$, thereby increasing the probability P(w$_{j}$ $|$ w$_{i}$). By repeating this process over the entire corpus, the vectors for words that habitually co-occur tend to be nudged closer and closer together. By gradually lowering the learning rate, this process converges towards some final state of the vectors. By the Distributional Hypothesis, words with similar distributional properties tend to share aspects of semantic meaning \cite{Firth1957}. For example, sentences in the corpus such as ``I like to play X with my friends,'' where X is the target word 'w$_{j}$' may be names of sports or activities that are semantically related.
\subsection{\textit{k}-means Clustering Context}
\textit{k}-means clustering is used to give the end-user the flexibility of choosing the subset of information that they may want to target. The user picks a single word topic, which then is converted to a vector, which in-turn is used to find the 100 words closest to it angle-wise with cosine similarity. \textit{k}-means clustering makes this simple because it essentially finds the center of mass (centroid) of a topic without having the rigid constraints of a Latent Dirichlet Allocation (LDA) \cite{mimno2012topic}. The \textit{k}-means clustering when mathematically depicted below computes the objective/squared error function that minimizes the distance from centroid to all other data points in a given cluster (\(m_n\) to all the \(x_n\) word topics).
\[J = \sum_{n=1}^{N}\sum_{k=1}^{K} r_{nk} * {||x_n - m_n||}^2\]
\section{Methodology}
The idea behind the present study being that, in an environment where comparable corpora (belonging to the same topic or sub-topic) are generated in a controlled and steady manner—such as a fashion magazine or financial newspaper—the more a word/topic is talked about, the more that term will be closely related to the umbrella field. Such as in a sports magazine; as the Olympics come closer, track and field will be much more closely related to the umbrella term of sports itself.
To demonstrate the effectiveness of the aforementioned framework, two case studies were performed comparing the relation between various themes of trends related to specific and arbitrary events. To implement the above mentioned ideology, the two main steps that were to be taken into account are the preparation of data by carving out the corpora on the basis of time (the time period of comparison being 'months' in present study), hence creating corresponding Word2Vec models for each slice, as well as formulating word trends over all the created models.
\subsection{Dataset}
The \textit{20 Newsgroups dataset} encompasses about 21,000 newsgroup posts or documents on 20 different topics or subtopics (such as atheism, religion, basketball, space etc). Some of the newsgroups are very closely related to each other (e.g. "PC hardware" vs "mac hardware"), while others are highly unrelated (e.g "for sale" vs "Christianity"), making this dataset the optimal choice for the research in this study. The acquired data is split in two subsets: one for training (or development) and the other one for testing (or for performance evaluation). The split between the train and test set is based upon messages posted before and after a specific date.
\subsection{Data Prepossessing}
Utilizing the 20 Newsgroups dataset, the corpora was loaded up, put into ascending order by date, then first split into smaller splices of data on the basis of years (2016, 2017, 2018 etc.) and further split into their corresponding months. Each corpus being a collection of articles, gives them a formal and structural quality. Therefore, these data points can have a variety of hidden meaning that might be conveyed through semantic analysis.
The standard data pre-processing steps were performed e.g., keyword tokenization, stopwords and removal of punctuation characters on the dataset. Lemmatization was also performed to get more consistent results. After the data was made adequately usable it was saved to be used later in the Processing step.
Before moving onto the next step, \textit{k}-means clustering was done to cross-examine if the data moved as it was hypothesized as. For example, in accordance with expectations, in June 2020, during the first wave of the COVID-19 pandemic, the embeddings of ``Unofficial'' and ``CDC'' (Centers for Disease Control) were at their farthest (Fig: \ref{fig:1a}), but in the following months, they converged. (Fig: \ref{fig:1b})
\begin{figure}
\begin{subfigure}{0.48\textwidth}
\captionsetup{justification=centering}
\includegraphics[width=\linewidth]{A.png}
\caption{Long distance between the words in Word2Vec model for June 2020} \label{fig:1a}
\end{subfigure}%
\hspace*{\fill}
\begin{subfigure}{0.48\textwidth}
\captionsetup{justification=centering}
\includegraphics[width=\linewidth]{B.png}
\caption{Short distance between the words in Word2Vec model for December 2020} \label{fig:1b}
\end{subfigure}%
\caption{Comparing semantic similarity between 'CDC' and 'Unofficial' over News Data in 2020} \label{fig:1}
\end{figure}
\subsection{Process}
In the present study, a Continuous Bag of Words Model was trained with a \textbf{worker\_size} of \textbf{5,000}. Training a \textit{GenSim} skip-gram Word2Vec model and tracking the temporal movements of said clusters (refer to Fig. II). As the clusters for each time period are calculated differently, the relative distance between vectors change with accordance to their positioning in clusters over different language models. This very change in distance is an indication of semantic drift, how two separate topics might be more closely related or be less related with each time-step.
The average \textbf{vector\_size} of large text corpora is around 300. In the present study large input corpora was converted into smaller slices, \textbf{vector\_size} is set to 100. We use a window size of 5 to get the best semantic relation between any two terms. \textbf{min\_count} is kept at 10 as well to ignore all words with low usage, such as names of specific officials or names of athletes, only focusing on words with high frequency. After the model for each slice was trained and created, the proceeding step was to measure the trend of words that may or may not be related to the overarching theme of the datasets.
In the present study, the likeness between a pair of topics is measured by their cosine similarity, i.e. a number ranging from 0 to 1, where a value close to 0 means that the words are highly correlated (interchangeable) to each other and 1 means that they are never used in the same context. The cosine value is the metric to define the absolute similarity between two words. The maximum cosine value (least correlated) in a pair is highlighted by red whereas the minimum cosine value (most correlated) in a pair is highlighted by green. If a term is not present in the model, its value denoted by '0.0' should be ignored, and is highlighted by blue.
The first instance of the model (refer to Figure \ref{fig:graph1} and Table \ref{tab:tab1}) shows the trend of a base term with respect to a single relative term. It is the trend of the term ``CDC'' over 2 years, considering the months of January, June and December of 2020 and 2021. The present study considers the base term to be ``unofficial'' to emphasize how the status quo sees the Center of Disease Control. A smaller value of the cosine distance between these two terms suggests that the mainstream considers the organization to be insignificant. In contrast, higher the value, higher will be the people's trust in the organization.
\begin{table}[H]
\centering
\captionsetup{justification=centering}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{7}{|c|}{Base Term : \textbf{Unofficial}}\\
\hline
Relative Term & \multicolumn{6}{c|}{Timeline (2020 - 2021)}\\
\hline
Months & Jan'20 & Jun'20 & Dec'20 & Jan'21 & Jun'21 & Dec'21\\
\hline
\hline
CDC & \cellcolor{blue!25}00.00 & \cellcolor{red!25}0.262 & 0.123 & 0.116 & \cellcolor{green!25}0.109 & 0.112\\
\hline
\end{tabular}
\caption{Cosine Similarity between topics over time on 20 Newsgroups Dataset. Maximum cosine value is red whereas the Minimum cosine value is green.}
\label{tab:tab1}
\end{table}
\begin{figure}[H]
\centering
\captionsetup{justification=centering}
\includegraphics[width=1\textwidth]{1_example.png}
\caption{Trend of the terms ``CDC'' through 2020-21 with the relative term ``Unofficial''}
\label{fig:graph1}
\end{figure}
As it can be deduced from the data (refer to Figure \ref{fig:graph1} and Table \ref{tab:tab1}), the term ``CDC'' starts being considered from June 2020, corresponding with the sharp increase in COVID-19 cases. A comparatively large cosine distance can be observed between ``CDC'' and ``Unofficial'', conveying that the organization was held in high regard as an entity to look for guidance and information in such a situation. As time progresses to June and December, the public's trust in the ``Media'' and government organizations deteriorate with increasing deaths and growth in the anti-vaccination agenda as well as an increased support from pro-vaccination groups \cite{10.1371/journal.pone.0247642}. Leading to the variance in cosine distance ($\sim$50\%), hence showcasing the growing distrust in the organization.
\subsection{Case Study : COVID-19 Trends}
In this instance, the model created in the previous example is further extended and built upon on, where now instead of equating a single relative term to the base term, multiple relative terms are considered. Therefore, not only analyzing their trends with respect to the base term but with themselves as well. Now, considering the base term as ``Trust'' while examining the association of this term with relative terms such as ``CDC'', ``Fauci'' and ``Experts''.
\begin{table}[ht]
\centering
\captionsetup{justification=centering}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{7}{|c|}{Base Term : \textbf{Trust}}\\
\hline
Relative Term & \multicolumn{6}{c|}{Timeline (2020 - 2021)}\\
\hline
Months & Jan'20 & Jun'20 & Dec'20 & Jan'21 & Jun'21 & Dec'21\\
\hline
\hline
CDC & \cellcolor{blue!25}00.00 & \cellcolor{red!25}0.413 & 0.132 & 0.128 & \cellcolor{green!25}0.103 & 0.114\\
\hline
Fauci & \cellcolor{blue!25}00.00 & \cellcolor{red!25}0.521 & 0.187 & \cellcolor{green!25}0.111 & 0.139 & 0.153\\
\hline
Experts & 0.280 & \cellcolor{red!25}0.591 & 0.133 & 0.126 & 0.146 & \cellcolor{green!25}0.109\\
\hline
\end{tabular}
\caption{Cosine Similarity between COVID-19 topics over time in 2020-21. The maximum Cosine Value is red whereas the minimum is green.}
\label{tab:tab2}
\end{table}
\begin{figure}[H]
\centering
\captionsetup{justification=centering}
\includegraphics[width=1\textwidth]{2_cs_COVID.png}
\caption{Trend of the terms ``CDC'', ``Fauci'' and ``experts'' through 2020-21 with the relative term ``Trust''}
\label{fig:graph2}
\end{figure}
The following case study is a perfect example to signify the discrepancies observed by quantizing such arbitrary concepts such as trust and formalism of an organization. Even though the terms ``Unofficial'' (refer to Figure \ref{fig:graph1} and Table \ref{tab:tab1}) and ``Trust'' (refer to Figure \ref{fig:graph2} and Table \ref{tab:tab2}) may seem different semantically, both of them produce similar trends when in relation to ``CDC''. This may be attributed to the absence of a general consensus.
An important fact that can be observed in both cases is the large difference in the values of the terms between June 2020 to December 2020. This corresponds to a greater shift in the political landscape as well as the public opinion. Where one school of thought went from considering CDC from competent to inept, in contrast, another may have started considering the said organization as more trustworthy. But, the COVID-19 case study also displays that all the relative terms (that are closely correlated to each-other) such as CDC, Dr. Anthony Fauci (Chief Medical Advisor to the President of the United States) and Experts all show similar trends.
\subsection{Case Study : The Olympics Trends}
While the previous example showed an exception to these trends, a notable demonstration of the inference power of such trends can be observed by taking the example of the Olympic games. As these games are periodic, occurring every 4 years, have very specific sub-topics (certain sports that are conducted at the Olympics) and do not have differentiating opinions about them, they are a perfect fit for the present study.
The base term is set as ``Olympics'' in the models and the relative terms were kept as pivotal sporting events, ``Track'', ``Tennis'', ``Gymnastics'' and ``Race'', as well as certain sports that are not a part of the Olympics (and hence not closely related) such as American Football as the relative term ``Football''.
These trends were drawn over 3 years (2016, 2017, 2021) where only 2017 was a non-Olympic year. Out of each of these years, 5 months were considered (January, March, July, September and December) to draw out these trends from.
\subsubsection{2016 Trends}
\begin{table}[H]
\centering
\captionsetup{justification=centering}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{6}{|c|}{Base Term : \textbf{Olympics}}\\
\hline
Relative Term & \multicolumn{5}{c|}{Timeline (2016)}\\
\hline
Months & Jan'16 & Mar'16 & Jul'16 & Sep'16 & Dec'16\\
\hline
\hline
Track & \cellcolor{red!25}0.368 & 0.355 & \cellcolor{green!25}0.115 & 0.215 & 0.309 \\
\hline
Football & 0.632 & \cellcolor{red!25}0.658 & \cellcolor{green!25}0.589 & 0.633 & 0.641\\
\hline
Tennis & \cellcolor{red!25}0.563 & 0.418 & \cellcolor{green!25}0.234 & 0.380 & 0.518\\
\hline
Gymnastics & 0.433 & \cellcolor{red!25}0.456 & \cellcolor{green!25}0.218 & 0.373 & 0.439\\
\hline
Race & \cellcolor{red!25}0.878 & 0.675 & \cellcolor{green!25}0.268 & 0.480 & 0.780\\
\hline
\end{tabular}
\caption{Cosine Similarity between Sports topics over time in 2016. Maximum cosine value is red whereas the Minimum cosine value is green.}
\label{tab:tab3}
\end{table}
\begin{figure}[H]
\centering
\captionsetup{justification=centering}
\includegraphics[width=1\textwidth]{3_cs_olympics_2016.png}
\caption{Trend of the terms ``track'', ``football'', ``tennis'', ``gymnastics'' and ``race'' through 2016 with the relative term ``Olympics''}
\label{fig:graph3}
\end{figure}
Using the data present in sports articles and analyzing the trends of such sports as track, tennis, gymnastics (all relating to and conducted at the Olympics) similar trends are shown. Whereas, American Football (not related to the Olympics) unlike the rest of the relative terms remains a nearly constant and static trend.
Moreover, the 2016 Summer Olympics, (also known as Rio 2016) was held from \nth{5} to \nth{21} August 2016 in Rio de Janeiro, Brazil, with preliminary events in some sports beginning on \nth{3} August. This sharply coincides with the results, where all relative terms related to Olympics show the highest convergence in the month of July 2016.
\subsubsection{2017 Trends}
\begin{table}[H]
\centering
\captionsetup{justification=centering}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{6}{|c|}{Base Term : \textbf{Olympics}}\\
\hline
Relative Term & \multicolumn{5}{c|}{Timeline (2017)}\\
\hline
Months & Jan'17 & Mar'17 & Jul'17 & Sep'17 & Dec'17\\
\hline
\hline
Track & 0.344 & 0.361 & \cellcolor{green!25}0.337 & \cellcolor{red!25}0.388 & 0.362 \\
\hline
Football & \cellcolor{green!25}0.622 & 0.673 & \cellcolor{red!25}0.681 & 0.641 & 0.679\\
\hline
Tennis & \cellcolor{red!25}0.587 & 0.568 & 0.543 & 0.562 & \cellcolor{green!25}0.418\\
\hline
Gymnastics & 0.454 & 0.456 & \cellcolor{red!25}0.531 & \cellcolor{green!25}0.357 & 0.459\\
\hline
Race & \cellcolor{green!25}0.678 & \cellcolor{red!25}0.875 & 0.743 & 0.756 & 0.866\\
\hline
\end{tabular}
\caption{Cosine Similarity between Sports topics over time in 2017. The Maximum cosine value is red whereas the Minimum cosine value is green.}
\label{tab:tab4}
\end{table}
\begin{figure}[H]
\centering
\captionsetup{justification=centering}
\includegraphics[width=1\textwidth]{4_cs_olympics_2017.png}
\caption{Trend of the terms ``track'', ``football'', ``tennis'', ``gymnastics'' and ``race'' through 2017 with the relative term ``Olympics''}
\label{fig:graph4}
\end{figure}
Now, using the same data from sports articles and other miscellaneous sources in 2017, trends are generated using the same relative and base term. All trends (refer to Table \ref{tab:tab4}) still show variation over time but do not show the same convergence as observed in the sports trends for 2016 (refer to Table \ref{tab:tab5}). Instead, all trends follow stable and consistent paths, exhibiting minimal variation.
An important contextual distinction that can be discerned is the fact that 2017 is a year in which Olympics are not held in. The crest and trough of the trends may very well show some particular event that relates to the sport, hence affecting its relative term. However, as there are no specific entities common between these, no particular trends are observed.
\subsubsection{2021 Trends}
\begin{table}[H]
\centering
\captionsetup{justification=centering}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{6}{|c|}{Base Term : \textbf{Olympics}}\\
\hline
Relative Term & \multicolumn{5}{c|}{Timeline (2021)}\\
\hline
Months & Jan'21 & Mar'21 & Jul'21 & Sep'21 & Dec'21\\
\hline
\hline
Track & \cellcolor{red!25}0.353 & 0.328 & \cellcolor{green!25}0.109 & 0.263 & 0.312 \\
\hline
Football & 0.656 & \cellcolor{red!25}0.681 & \cellcolor{green!25}0.597 & 0.624 & 0.638\\
\hline
Tennis & \cellcolor{red!25}0.573 & 0.410 & 0.244 & \cellcolor{green!25}0.210 & 0.318\\
\hline
Gymnastics & 0.212 & 0.349 & \cellcolor{green!25}0.218 & 0.398 & \cellcolor{red!25}0.441\\
\hline
Race & \cellcolor{red!25}0.900 & 0.576 & \cellcolor{green!25}0.254 & 0.503 & 0.823\\
\hline
\end{tabular}
\caption{Cosine Similarity between Sports topics over time in 2021. The Maximum cosine value is red whereas the Minimum cosine value is green.}
\label{tab:tab5}
\end{table}
\begin{figure}[H]
\centering
\captionsetup{justification=centering}
\includegraphics[width=1\textwidth]{5_cs_olympics_2021.png}
\caption{Trend of the terms ``track'', ``football'', ``tennis'', ``gymnastics'' and ``race'' through 2021 with the relative term ``Olympics''}
\label{fig:graph5}
\end{figure}
Finally, we see a great deal of similarity between the trends of 2021 (refer to Table \ref{tab:tab5}) and 2016 (refer to Table \ref{tab:tab3}). This shows that the trends generated in the present study are reproducible and show periodicity, where the trends of 2017 are visibly differentiable from 2016 and 2021.
The 2020 Summer Olympics, (also known as Tokyo 2020), was held from \nth{23} July to \nth{8} August 2021 in Tokyo, Japan, with some preliminary events that began on \nth{21} July, 2021. Similar to 2016, this sharply coincides with the results, where all relative terms related to Olympics show the highest convergence in the month of July 2016.
\newpage
\section{Results and Discussion}
The present study has discussed the limitations of using stochastic methods to try to understand how language and the topics within it are transformed over time. Thus, introducing a method of using Word2Vec to capture semantic meaning to draw out changes in the meaning and use of words over time, and therefore understanding and visualizing the trends related to them, has been helpful. We have presented two case studies to substantiate our hypothesis.
The model shows how articles pertaining to sports show multiple trends of certain terms with respect to sports itself. But, it is none the more apparent with competitive games that are and are not related to the Olympics. Sports like track \& field, racing, tennis, gymnastics show a trend of higher correlation with sports during the months around June, compared to American Football, (refer to Figure \ref{fig:graph3}) which is not an Olympic sport, to show no such trend and remaining mostly non-divergent. Whereas track \& field, arguably the sport most dependent on the Olympics, shows the highest deviation compared to other sports such as tennis, which has other important events all year around.
Another case study relates to terms associated with healthcare during the COVID-19 pandemic, where we consider the trend of terms with respect to the dimension of ``Trust'' and ``Unofficial''. The first cases of COVID were reported in China in the first weeks of January 2020. This shows the center of disease control, ``CDC'' with a comparatively low correlation to trust (refer to Figure \ref{fig:graph2}) as well as unofficial (refer to Figure \ref{fig:graph1}). Germani et al \cite{10.1371/journal.pone.0247642}, in their behavioural analysis research of anti-vaccination rhetoric on social media, depict a lot of similarity with the trends in the present study.
\section{Conclusion \& Future Work}
The present study presents a new method to measure the temporal trend of words by capturing the change in their context. The semantic movement of a target topical word is tracked and compared against a base term with the cosine similarity of the word vectors taken over time, measuring how topical keywords can evolve. This approach goes beyond the current explored methods which only track single keyword occurrences over time. Hence, this approach also provides a better means to visualize the trends of topics over time.
As a standalone tool, this framework can assist in understanding the evolutionary aspects language undergoes as words can gain and lose meaning over time. This is especially true in the social networks where the primary mode of communication is text, and discourse occurs quite frequently and rapidly. In the presented two case studies, the domains of healthcare and sports with a news dataset were explored, both exhibiting visible discernment.
Combined with automated topic extraction methods, such as LDA \cite{10.5555/944919.944937} and Top2Vec \cite{angelov2020top2vec} , topic discovery could be enhanced with topical evolution as an end-to-end framework for developing deeper insights into a corpus. This includes inter-topic trends as words can be part of a topic in one time period, and then be a part of another.
The current approach utilizes multiple models, each corresponding to a specified time period. Training a corpus on a single model preserves the basis space of the embeddings as to minimize the correlation errors, and in our case, between paired topics. The hope is to develop a single model that jointly trains the embeddings with respect to the time aspects of the corpus could solve this issue. However, as the intention was to develop a practical analytical tool since time-based learning is limited in current language models, careful construction of the dataset is a suitable solution. Whether it be by joining all the month datasets together and tagging each word with its month, or by creating intermediate datasets between each month by virtue of an an overlapping window.
Recently, more robust deep learning architectures that yield improved embeddings due to bidirectional learning (e.g. BERT \cite{devlin2018bert} or ELMo \cite{sarzynska2021detecting}). As these language models can learn different usages and interpretations of a token, our framework would be less prone to errors in the presence of linguistic polysemy.
\newpage
\printbibliography
\end{document}
|
\section{Supplemental Material}
\label{SM}
\noindent Here we present some details of derivations of the results obtained for the one-sided case
and also show that in the limit $H \to 1/2$ the expressions (6), (7), (9) and (11) converge to the Wiener result in Eq. (1).
\renewcommand{\theequation}{S\arabic{equation}}
\setcounter{equation}{0}
\section{Two-sided fBm}
\subsection{Sub-diffusion, $0 < H < 1/2$. The limit of $H\to 1/2$}
\noindent
In order to take the limit $H\to 1/2$, we take advantage of the identity
\begin{align}
\frac{1}{|t_1 - t_2|^{2H}} = \frac{1}{(1-2H)} \frac{d}{dt_1} \frac{{\rm sgn}(t_1 - t_2)}{|t_1 - t_2|^{2H-1}} \,,
\end{align}
which permits us to formally rewrite the kernel in Eq. (3) as
\begin{align}
\frac{{\rm cot}(\pi H)}{2 \pi H} \frac{1}{|t_1 - t_2|^{2H}} = \frac{{\rm cot}(\pi H)}{2 \pi H(1-2H)} \frac{d}{dt_1} \frac{{\rm sgn}(t_1 - t_2)}{|t_1 - t_2|^{2H-1}} \,.
\end{align}
Taking the limit $H \to 1/2$ in the both sides of the latter equality
and noticing that
\begin{align}
\lim_{H \to 1/2} \frac{{\rm cot}(\pi H)}{(1- 2H)} = \frac{\pi}{2} \,,
\end{align}
we get
\begin{align}
\lim_{H \to 1/2} \frac{{\rm cot}(\pi H)}{2 \pi H} \frac{1}{|t_1 - t_2|^{2H}} &= \lim_{H \to 1/2} \frac{{\rm cot}(\pi H)}{2 \pi H(1-2H)} \frac{d}{dt_1} \frac{{\rm sgn}(t_1 - t_2)}{|t_1 - t_2|^{2H-1}} \\
&= \frac{1}{2} \frac{d}{dt_1} {\rm sgn}(t_1 - t_2) = \delta(t_1 - t_2) \,,
\end{align}
which yields the Wiener expression (1).
\subsection{Super-diffusion, $1/2 < H < 1$. The limit of $H\to 1/2$}
We turn to the limit $H \to 1/2$ directly in Eq. (22) in the main text to get
\begin{align}
\lim_{H \to 1/2}\mathcal{Z}(\tau)= \mathcal{Z}_{H=1/2}(\tau) = \frac{1}{2} \frac{d^2}{d \tau^2} {\rm sgn}(\tau) = \frac{d}{d \tau} \delta(\tau)
\end{align}
The action $S$ written in terms of the derivative of the fractional Gaussian noise involves the kernel function $\mathcal{Q}$, which is given by a triple integral of $\mathcal{Z}_{H=1/2}(\tau)$ [see Eq. (23)]. Consequently, the action has the form
\begin{align}
S=\frac{1}{2} \int^{\infty}_{-\infty} \int^{\infty}_{-\infty} dt_1 dt_2 \ddot{x}(t_1) \ddot{x}(t_2) \int^{t_1}_{-\infty} \int^{t_2}_{-\infty} d\tau_1 d\tau_2 \delta(\tau_1 - \tau_2) \,.
\end{align}
Integrating the latter expression by parts, we arrive at the Wiener's result in Eq. (1).
\section{One-sided fBm}
\noindent Here we present brief derivations of our expressions (7) and (11) of the main text.
\subsection{Sub-diffusion, $0 < H < 1/2$}
For the one-sided sub-diffusive fBm the inverse kernel $\mathcal{C}(t_1,t_2)$ is not a difference kernel, and Eq.~(13) gives way to the equation
\begin{eqnarray}
\int_{0}^{\infty} d t_1 \,\mathcal{C}(t_1,t_2) \frac{d}{dt_1} \left[|t_1-t_3|^{2H-1} \text{sgn}
(t_1-t_3)\right] = \frac{1}{H}\,\delta(t_2-t_3)\,.&&
\end{eqnarray}
Integrating by part, we arrive at
\begin{equation}\label{inteqn2a}
\int_{0}^{\infty} d t_1 \,\frac{\text{sgn} (t_1-t_3)}{|t_1-t_3|^{1-2H}} \,\mathcal{D}(t_1,t_2)
=-\frac{1}{H}\,\delta(t_2-t_3),
\end{equation}
where $\mathcal{D}(t_1,t_2) = (\partial/\partial t_1) \,\mathcal{C}(t_1,t_2)$.
The solution can be found in Ref. [47]. Getting rid of the $t_1$-derivative, we
obtain $\mathcal{C}(t_1,t_2)$ in the following three alternative (but equivalent) forms
\begin{align}\label{q1}
\mathcal{C}(t_1,t_2) &= \frac{(t_1 t_2)^{-H} \sqrt{z'}}{H(1-2H) \sin(\pi H) \Gamma(2H)
\Gamma^2\left(\frac{1}{2}-H\right)} \,_2F_1\left(1,H+\frac{1}{2};\frac{3}{2}-H;z'\right) \,, \quad z' = \frac{{\rm min}(t_1,t_2)}{{\rm max}(t_1,t_2)} \\
&= \frac{{\rm cot}(\pi H) (t_1 t_2)^{-H} \sqrt{z}}{2 \pi H (1-2H) B(1/2-H,2H)} \,_2F_1\left(\frac{1}{2},1;\frac{3}{2}-H;z\right) \,, \quad z = \frac{4 t_1 t_2}{(t_1 + t_2)^2} \\
& = \frac{{\rm cot}(\pi H)}{2 \sqrt{\pi} \Gamma(1/2-H) \Gamma(1+H)} \frac{B_z(1/2-H,H)}{|t_1 - t_2|^{2H}} \,,
\end{align}
where $B(a,b)$ and $B_z(a,b)$ are the complete and incomplete beta-functions, respectively, and
$_2F_1(\dots)$ is the hypergeometric function.
Recalling next the definition of the regularized incomplete beta-function [see Eq. (8)], we obtain our result in Eq. (7).
\vspace{0.25 cm}
\textbf{Limit $H \to 1/2$.} To take the limit $H \to 1/2$ in Eq. (7) in the main text it is expedient to use
the representations of the kernel $\mathcal{C}(t_1,t_2)$ given in the second line in Eq. (\ref{q1}). We formally rewrite the Gauss hypergeometric function entering this representation as
\begin{align}
\,_2F_1\left(\frac{1}{2},1;\frac{3}{2}-H;z\right) = \frac{\Gamma(H) \Gamma(3/2-H)}{\sqrt{\pi} (1-z)^{H} z^{1/2-H}} - \frac{(1/2-H)}{H} \,_2F_1\left(\frac{1}{2},1;1+H;1 - z\right)
\end{align}
such that, after some algebra, the kernel $\mathcal{C}(t_1,t_2) $ can be cast into the form
\begin{align}
\mathcal{C}(t_1,t_2) = \frac{{\rm cot}(\pi H)}{2 \pi H} \frac{1}{|t_1 - t_2|^{2 H}} - \frac{{\rm cot}(\pi H)}{2 \pi H^2 B(1/2-H,2H)} \frac{(t_1 t_2)^{1/2-H}}{(t_1 + t_2)} \,_2F_1\left(\frac{1}{2},1;1+H;\left(\frac{t_1-t_2}{t_1+t_2}\right)^2\right)
\end{align}
Further on, we observe that the numerical $H$-dependent amplitude in the second term in the latter expression vanishes in the limit $H \to 1/2$ as
\begin{align}
\frac{{\rm cot}(\pi H)}{2 \pi H^2 B(1/2-H,2H)} \simeq 2 \left(\frac{1}{2} - H\right)^2 \,,
\end{align}
which signifies that this term does not contribute in this limit. On the contrary, as demonstrated above, the first term converges to the delta-function which ensures that the action in Eq. (7) in the main text converges to the Wiener's result in Eq. (1).
\subsection{Super-diffusion, $1/2 < H < 1$}
Here we present a derivation of Eq. (11) and also check the limit of $H\to 1/2$. In particular, this derivation
highlights the reason why the representation of the action in terms of $\ddot{x}$ is
advantageous.
We start the derivation by representing the action in terms of the first derivative of the fBm, that is in terms of the fGn:
\begin{align}
\label{q2}
S = \frac{1}{2} \int^{\infty}_0 \int^{\infty}_0 dt_1 dt_2 \mathcal{C}(t_1,t_2) \dot{x}(t_1) \dot{x}(t_2) \,.
\end{align}
The inverse kernel $\mathcal{C}(t_1,t_2)$ obeys the integral equation
\begin{align}
\label{q3}
\int^{\infty}_0 \frac{dt_1 \, \mathcal{C}(t_1,t_2)}{|t_1 - t_3|^{2-2H}} = \frac{\delta(t_2 - t_3)}{H (2H-1)} \,.
\end{align}
The explicit solution of this equation can be found in Ref. \cite{Lundgren}. After straightforward transformations, it reads:
\begin{align}
\label{q4}
\mathcal{C}(t_1,t_2) &= \frac{1}{2 H \sin(\pi H) \Gamma(2H) \Gamma^2(3/2-H)} \frac{d^2}{dt_1 dt_2} \int_0^{{\rm min}(t_1,t_2)} \frac{d\tau}{(t_1 - \tau)^{H-1/2} (t_2 - \tau)^{H-1/2}} \\
&= \frac{1}{2 H \sin(\pi H) \Gamma(2H) \Gamma^2(3/2-H)} \frac{d^2}{dt_1 dt_2} |t_1 - t_2|^{2 - 2H} B_{z'}(3/2-H,2H-2) \,, \quad z'= \frac{{\rm min}(t_1,t_2)}{{\rm max}(t_1,t_2)} \,.
\end{align}
We observe that the function $\mathcal{C}(t_1,t_2)$, defined in the right-hand-side of Eq. (\ref{q4}), contains a non-integrable singularity at $t_1 = t_2$, which shows why the representation of the action in Eq. (\ref{q2}) in terms of the first derivatives $\dot{x}(t)$ is problematic. To get a regular result, we integrate Eq. (\ref{q2}) by part,
\textit{i.e.} express it in terms of the second derivatives $\ddot{x}(t_1)$ and $\ddot{x}(t_2)$. Then, using the definition of the regularized incomplete beta-function in Eq. (8), we arrive at the final result in Eq. (11).
\vspace{0.25 cm}
\textbf{Limit $H \to 1/2$.} The limit $H \to 1/2$ can be conveniently taken in the expression given in the first line of Eq. (\ref{q4}). This yields
\begin{align}
\lim_{H \to 1/2} \mathcal{C}(t_1,t_2) = \frac{d^2}{dt_1 dt_2} {\rm min}(t_1,t_2) = \delta(t_1 - t_2) \,,
\end{align}
and we recover the Wiener's expression in Eq. (1).
\end{widetext}
|
\section{Introduction}
Probabilistic representations are at the heart of robotics~\cite{thrun2005probabilistic} and allow for a wide range of applications, including SLAM and visual odometry, by providing formal methods to account for uncertain sensor data when measuring and representing the environment. 3D scene representations using multilayer perceptron neural networks (MLP), have witnessed overwhelming interest from neighboring scientific communities during the last few years, notably because of the impressive level of realism achieved in representing scenes in this way. However, these methods do not explicitly account for uncertainties in their representations. Sensors measuring real-world quantities inherently contain some uncertainty (such as cameras, depth sensors, etc.), and this should be accounted for in the case where a robot interacts with its environment. Subsequently, the aim of this article is to investigate the problem of incorporating uncertainty into neural scene representations and to show its importance in terms of geometric consistency and increased map quality.
In robotics, 3D scene representations have often been modeled in various ways, including voxels~\cite{newcombe2011dtam, whelan2013vo} and key-frame graphs~\cite{meilland2011vslam, meilland2013unifying}, which aim to model scene surfaces. In this case, different primitives can be considered, for example, point clouds \cite{keller2013Realtime, wiles19synsin}, meshes \cite{nimier19mitsuba}, or implicit surfaces \cite{curless96avolumetric, takikawa2021nglod}. In ADOP (Approximate Differentiable One-Pixel Point Rendering) \cite{ruckert2021adop}, for instance, 3D point clouds are used to store learnable features that are interpreted by an MLP (Multilayer Perceptron) to produce view-dependent effects. Furthermore, Deferred Neural Rendering \cite{thies19DNR} represent surfaces as meshes, proposing to use the so-called neural textures as feature maps for the meshes. In addition to that, \cite{zhang2021learning} uses a signed distance field and a surface light field to represent scenes, taking advantage of a depth estimator network to supervise its geometry branch. Surface-based representation only models the information present on the surface of the scene elements, allowing this method to limit its memory footprint. However, this comes at the cost of not being able to accurately model volumetric matter \cite{tewari2021advances}.
Volumetric scene representations are more general, as they permit to model solid elements such as surfaces along with volumetric matter as smoke. Historically, voxel grids were the first volumetric approaches to be developed \cite{connolly84octree, chien88octree}. They offer a well-defined formalism for modeling 3D scenes. However, such approaches suffer from important memory footprints, which can become prohibitively large when the image resolution is high \cite{tewari2021advances}.
The past years have witnessed the development of a new kind of approach for volumetric scene representation that combines methods from \textit{machine learning} with those from \textit{computer vision} and \textit{computer graphics}, known as neural scene representation. An example of such approaches is Neural Volumes \cite{lombardi19NV}, which proposes to train a 3D CNN (Convolutional Neural Network) to reconstruct single scenes from multiview images. Another kind of approach that allowed to produce photorealistic images using classical image rendering is NeRF (Neural Radiance Fields) \cite{mildenhall2020nerf}. This approach is able to take into account specularity effects, while being able to model high-frequency image elements due to the use of positional encoding. It also exhibits interesting compression properties, which come at the cost of intensive computations, as several seconds are required to render a single high-resolution image on a typical desktop GPU. Consequently, different methods were proposed to overcome this limitation \cite{yu2021plenoctrees, yu2021pixelnerf} and improve the performance of the approach \cite{zhang2020nerf, mueller2022instant}. In addition, other approaches focused on the use of depth information to improve the modeled geometry and the quality of the rendered image, such as \cite{wei2021nerfingmvs, neff2021donerf, dey2022mip}.
The aforementioned approaches related to NeRF make the assumption that the camera poses for the training images are known and rely on SFM (Structure From Motion) methods such as COLMAP \cite{schoenberger2016mvs,schoenberger2016sfm} to produce the desired poses in the case of real data. However, some approaches have been proposed to estimate camera poses along with neural network parameters \cite{SCNeRF2021,wang2021nerfmm,lin2021barf}. Therefore, self-calibrating neural radiation fields \cite{SCNeRF2021} proposes to additionally estimate intrinsic, extrinsic, and distortion camera parameters; while BARF \cite{lin2021barf} proposes to apply an annealing schedule to each component of the positional encoding for coarse-to-fine trajectory estimation.
Other works aimed at using neural radiance fields in robotics have recently been developed. For instance, \cite{sucar2021imap} and \cite{Zhu2022nice} proposed SLAM systems that use NeRF-based map representation. Other approaches such as \cite{yen2020inerf} aimed to solve the inverse problem of estimating the pose of a query image using a trained NeRF. Furthermore, \cite{adam2022ral} used the output of NeRF to build a trajectory planning algorithm, and~\cite{yen2022nerfsupervision} used this kind of approach to learn view invariant object descriptors. These approaches, however, didn't take explicitly model the uncertainty distribution, which does not allow them to best exploit the available information.
In this work, the objective is integrate such an uncertainty distribution into the learning process of NeRF, which is beneficial for various robotics applications such as SLAM. Therefore, the learned distribution of depths along view rays is supervised using the available geometric information, such as point cloud estimates or depth measurements. This supervision during the learning process of neural radiance fields, in addition to being an inductive bias towards the correct scene geometry, allows a direct integration of uncertainty; resulting in an increased rendering quality and better geometry modeling. The proposed method is hence described in Section~\ref{sec:method_description}, followed by a presentation of the validation results (Section~\ref{sec:results}); in addition to that conclusions and perspectives are presented in Section~\ref{sec:conclusion}.
\section{Method Description} \label{sec:method_description}
\subsection{Preliminaries}
This section presents a recap of the approach proposed in~\cite{mildenhall2020nerf}, which is considered in this article. NeRF represents the scene as a continuous volumetric field using an MLP. Its inputs are the position $\mathbf{x}_k$ and view direction $\mathbf{d}_k$ of a 3D point in space $\mathbf{p}_k$. Its outputs are the corresponding color $\mathbf{c}_k$ and density $\tau_k$:
\begin{equation}
[\mathbf{c}_k, \tau_k] = \mathrm{MLP} \left( \gamma( \mathbf{x}_k, \mathbf{d}_k); \theta \right).
\end{equation}
Here $\gamma(.)$ is an encoding function that maps positions and directions to a high-dimensional sine-cosine space:
\begin{equation}
\gamma(\mathbf{x}) = [\mathrm{sin}(\mathbf{x}), \mathrm{cos}(\mathbf{x}), ..., \mathrm{sin}(2^{L-1} \mathbf{x}), \mathrm{cos}(2^{L-1} \mathbf{x})]^T,
\end{equation}
with $L$ being a hyperparameter. The images are rendered by casting a ray $\mathbf{r}(t) = \mathbf{o} + t \mathbf{d}$ from the camera's center of projection $\mathbf{o}$ through each image pixel, with $t$ representing the distance along the ray. Therefore, the sample distances $t_{k} \in \mathbf{t}$ are drawn according to a sampling strategy along each ray between the near ($t_{n}$) and far ($t_{f}$) bounds ($\mathbf{t}$ is the vector of the samples). These resulting sampling distances are then used to compute the 3D positions of the input points, while the input direction is defined by the ray direction. After that, this set of sample points (position and direction) is inputted into the neural net to predict the corresponding colors and densities. The pixel values are finally computed using an approximate volume rendering integral \cite{Max1995OpticalMF}:
\begin{equation}
\mathbf{C(r; t, \theta)} = \sum_{k} w_{k} \mathbf{c}_{k},
\end{equation}
where
\begin{equation}
w_{k} = T_k \Big( 1 - \mathrm{exp} \big( -\tau_{k}(t_{k+1} - t_{k}) \big) \Big) \mathbf{c}_{k},
\end{equation}
and
\begin{equation}
T_k = \mathrm{exp}\left(-\sum_{k'<k} \tau_{k'}(t_{k'+1} - t_{k'})\right).
\end{equation}
To increase the sampling efficiency, NeRF trains a coarse and fine MLP, through minimization of the following loss:
\begin{equation}\label{eq:loss_photo}
\mathcal{L} = \sum_{\mathbf{r} \in R} \left( ||\mathbf{C}_{coarse} - \mathbf{C}^{*}_{coarse}||_{2}^{2} + ||\mathbf{C}_{fine} - \mathbf{C}^{*}_{fine}||_{2}^{2} \right).
\end{equation}
The parameters of $\mathbf{C}$ were omitted for readability. $\mathbf{C}$ and $\mathbf{C}^*$ are, respectively, the predicted and target pixel colors for the coarse and fine networks. $R$ represents the set of rays for training and $||.||_{2}$ is the L2 norm.
\subsection{Probabilistic Learning and Training}
This work proposes to leverage uncertainty information (e.g. uncertainty along viewing rays) to train neural radiance fields in a probabilistic manner. NeRF-based approaches tend to fail to produce consistent novel views when training data are overly uncertain. This is the case of vanilla NeRF which assumes absolute uncertainty along each viewing ray. NeRF overcomes this absolute uncertainty by using a large number of training views which allows to improve geometric consistency. One way to circumvent this drawback is to use additional probabilistic information during model training.
In this paper, the case of depth uncertainty along the rays is considered because it is a primary source of uncertainty in NeRF architectures. Knowledge of depth uncertainty can be obtained from uncertain depth sensor measurements, uncertain correspondences between images, or even pose uncertainty. However, the uncertainty along rays is not explicitly modeled by the classic approach, which instead relies on the training process to recover it. As will be shown, such a strategy is indeed non-optimal, as it induces a need for additional data (views) to learn the scene representation. For example, consider Figure~\ref{fig:failure_example}, which shows an example of a rendered image using DONeRF \cite{neff2021donerf} for two cases (training with a subset of the training dataset and with the full training dataset).
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{data/do_nerf_illustration.pdf}
\caption{(left) image rendered using DONeRF trained with 60 images, (right) the same image rendered using DONeRF trained with 210 images. The method requires the full training set of images to learn a correct scene model.}
\label{fig:failure_example}
\end{figure}
The proposition of this work is to integrate uncertainty into the training process and demonstrate its effectiveness through the example of depth uncertainty. This is achieved by fitting a depth PDF (Probability Density Function) obtained from the NeRF-based model $\mathfrak{D}(r)$ to a target ray PDF $\mathfrak{D}^*(r)$ obtained by various means. The former is computed by normalizing the weights $w_k$ along each ray \cite{mildenhall2020nerf} as in Equation (\ref{eq:depth_pdf_est}), while the latter is computed using a parametric PDF. The target PDF can be computed using available geometric information; such as depth sensor measurements $\mathbf{D}_i$, image correspondence/3D point uncertainty, camera pose uncertainty, and/or SFM algorithms (e.g. COLMAP~\cite{schoenberger2016sfm}). Figure~\ref{fig:depth_density_illustration} shows an illustration of the proposed representation of depth uncertainty. In the rest of this article, the target uncertainty will be modeled by a Gaussian PDF, centered around a depth value.
\begin{equation} \label{eq:depth_pdf_est}
\mathfrak{D}(r) = \frac{1}{\sum_{j} w_j}[w_0,..., w_N] \: / \: j = {0, 1, ..., N}.
\end{equation}
$N$ and $[w_0,..., w_N]$ represent, respectively, the number of samples and the sample vector along the considered ray. One of the advantages regarding this approach is its ability to constrain depth uncertainty (inherent to depth measurements or 3D points) through a target uncertainty distribution. It is therefore possible to supervise the learned depth in accordance with the uncertainty sources (such as the uncertainty on the 3D point position or the camera position).
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{data/depth_density.pdf}
\caption{The proposed approach aims to supervise the learning of the depth PDF obtained from the network $\mathcal{D}(r)$ using a target PDF $\mathcal{D}^*(r)$. Left, the initial PDF in green has high uncertainty and right, after training the learnt PDF matches the target uncertainty.}
\label{fig:depth_density_illustration}
\end{figure}
Equation (\ref{eq:color_depth_density_loss}) shows the proposed loss function, which takes into account depth uncertainty within the training process.
\begin{equation} \label{eq:color_depth_density_loss}
\mathcal{L} = K_{color} \, \sum_{r \in R} \mathcal{L}_{color} + K_{density} \, \mathcal{L}_{density} + K_{depth} \, \mathcal{L}_{depth},
\end{equation}
with
\begin{equation}
\mathcal{L}_{color} = ||\mathbf{C}_{coarse} - \mathbf{C}^{*}_{coarse}||_{1} + ||\mathbf{C}_{fine} - \mathbf{C}^{*}_{fine}||_{1},
\end{equation}
\begin{equation}
\mathcal{L}_{density} = ||\mathfrak{D}_{coarse} - \mathfrak{D}^*_{coarse}||_{1} + ||\mathfrak{D}_{fine} - \mathfrak{D}^*_{fine}||_{1},
\end{equation}
and
\begin{equation}
\mathcal{L}_{depth} = ||\mathbf{D}_i - \mathbf{D}_i^*||_{1}.
\end{equation}
The depth PDF parameters were omitted for readability. $||.||_1$ is the L1 norm. This was preferred over the L2 norm used by NeRF due to the higher error values it produces when the error is smaller than 1 (which is the case for the depth densities). Moreover, it exhibits increased robustness toward outliers. $K_{color}$, $K_{density}$, and $K_{depth}$ are the gains applied to the terms. The depth $\mathbf{D_i}$ is computed as in Equation (\ref{eq:depth}), while $\mathbf{D_i}^*$ is a reference depth that can be obtained from a depth measurement or a point cloud as shown in Figure \ref{fig:3d_point_pnerf}. It is worth noting that the Gaussian PDF represented in Figure \ref{fig:3d_point_pnerf} takes into account the uncertainty of the 3D point in addition to the one of the camera pose along the $z$ axis.
\begin{equation} \label{eq:depth}
\mathbf{D}_i = \sum_{k} w_k \mathbf{t}_k.
\end{equation}
The depth term in this loss function aims to bias the learning process towards the most probable depth (the mean value), while taking into account its variance through the density
term.
\begin{figure}[h]
\centering
\includegraphics[width=0.32\textwidth]{data/3d_point_pnerf.png}
\caption{Target depth PDF $\mathcal{D}^*(r)$ obtained from the uncertain estimate of a 3D point. In this configuration, $\mathcal{D}^*(r)$ allows to take into account the pose uncertainty of the camera along the $z$ axis.}
\label{fig:3d_point_pnerf}
\end{figure}
\section{Results} \label{sec:results}
In this section, the effect of the different cost function terms is studied first, showing the benefit of the proposed loss. This is followed by a comparison with state-of-the-art methods on both synthetic and real data. The models considered here were trained on an Nvidia GeForce RTX 3090 GPU with 24 Gb RAM. In addition to that, the implementation of Pytorch Lightning of NeRF\footnote {https://github.com/kwea123/nerf\_pl/} was used to develop the proposed approach. Noting that the target Gaussian considered in these experiments is a unit Gaussian PDF centered around the target depth.
\subsection{Ablation study}
This part presents a comparison between NeRF models trained with the different terms of the loss described in Equation (\ref{eq:color_depth_density_loss}). The first model is trained with pure photometric loss (Equation (\ref{eq:loss_photo})) as in the original NeRF. The loss of the second model, for its part, contains only the color and density terms (modeling the uncertainty). The loss of the third model contains all the terms of Equation (\ref{eq:color_depth_density_loss}). Training is done using a stereo image pair that was generated from Blender using the Sponza virtual scene\footnote {https://www.intel.com/content/www/us/en/developer/topic-technology/graphics-research/samples.html}. In addition, two additional stereo pairs were generated for validation and testing. The objective here is to provide insights about the influence of the loss terms on the quality of the learnt representation.
Table \ref{table:ablation_quantitative} shows the PSNR values (Peak Signal-to-Noise Ratio) obtained with the different loss terms. It can be seen that considering only the photometric term provides the lowest PSNR value, as the model cannot learn a geometrically consistent scene representation (Figure \ref{fig:ablation_qualitative}). Integrating depth uncertainty through the addition of the density term to the loss allows to synthesize images with better quality. This is notably due to the fact that the learned scene geometry is consistent, allowing for a better generalization. However, sharp elements are not correctly learned, as shown in Figure \ref{fig:ablation_qualitative}. The term related to depth uncertainty, in fact, allows the model to learn a distribution of the depth corresponding to the image pixels, inducing a smoothing of the predicted depth. Therefore, the edges are poorly modeled. This aspect is compensated by adding to the depth term the loss, which forces the model to predict depth values that tend towards mean depth.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{data/ablation_study_qualitative.pdf}
\caption{Qualitative results for the ablation study. The introduction of depth information allows to better model the scene. Adding only the depth uncertainty term is not sufficient to allow the model to learn sharp scene elements, while biasing it towards the mean depth allows for a better scene modeling.}
\label{fig:ablation_qualitative}
\end{figure}
\begin{table}
\centering
\begin{tabular}{c | c}
\hline
Terms & PSNR $\uparrow$ \\
\hline
$\mathcal{L}_{photo}$ & 14.44 \\
$\mathcal{L}_{photo + density}$ & 20.66 \\
$\mathcal{L}_{photo + density + depth}$ & \textbf{20.85} \\
\end{tabular}
\caption{PSNR values for the considered loss functions. Adding the depth uncertainty term to the loss allows for a better modeling of the scene, but does not allow to model sharp elements. Adding the depth term allows to achieve a better modeling of sharp scene elements.}
\label{table:ablation_quantitative}
\end{table}
\subsection{Validation on Synthetic Data}
The dataset provided by \cite{neff2021donerf} was considered to validate the proposed approach. More specifically, the \textit{Classroom} and \textit{Barbershop} sequences that represent indoor synthetic scenes were considered as the proposed approach aims at modeling scenes with finite depth. The metrics considered here are the PSNR and the SSIM (Structural SIMilarity) that allow to measure the level of corruption of the rendered images and the perceptual similarity to the test images, respectively. Note that the model trained with the proposed loss function is referred to as P-NeRF, which renders $200\times 200$ images to speed up training. The models were trained with ADAM \cite{kingma2014adam} for $200$ epochs with a batch size of $2048$ and a learning rate of $0.0005$. The total number of evaluations is $256$ per iteration. In addition, the checkpoints used for the test are those with the lowest PSNR at the test time.
P-NeRF is compared with the DONeRF approach \cite{neff2021donerf} that uses an oracle depth network to produce a sampling region for the NeRF model to render $400\times 400$ images. The publicly available code\footnote {https://github.com/facebookresearch/DONERF} was used for this evaluation. Furthermore, the models were trained with $20$, $60$, $140$, and full sequence lengths. This validation allowed to study the effect of the size of the training set on the generalization.
Tables \ref{table:synthetic_data_partial_quantitative_barbershop} and \ref{table:synthetic_data_partial_quantitative_classroom} show the PSNR and SSIM values on the complete test sets of the Barbershop and Classroom sequences, for both P-NeRF and DONeRF using different training set sizes (20, 60, and 140 images). It can be seen that P-NeRF renders better image quality at test time compared to DONeRF, when the training set contains 20, 60, and 140 randomly selected images. This supports the claim that the integration of the depth information in a probabilistic manner allows the proposed P-NeRF to learn a better model of the scene, resulting in better-rendering images.
\begin{table}
\centering
\begin{tabular}{c | c | c | c}
\hline
Training Images & Method & PSNR $\uparrow$ & SSIM $\uparrow$ \\
\Xhline{1pt}
\multirow{2}{5em}{20 Images} & DONeRF & 26.75 & 0.87 \\
& P-NeRF & \textbf{27.23} & 0.87 \\
\Xhline{0.7pt}
\multirow{2}{5em}{60 Images} & DONeRF & 27.68 & 0.87 \\
& P-NeRF & \textbf{31.30} & \textbf{0.93} \\
\Xhline{0.7pt}
\multirow{2}{5em}{140 Images} & DONeRF & 29.07 & 0.89 \\
& P-NeRF & \textbf{32.92} & \textbf{0.95} \\
\Xhline{0.7pt}
\end{tabular}
\caption{Quantitative results for the DONeRF and the proposed P-NeRF using different training sets from the Barbershop sequence. The proposed approach permits to predict images with a better quality than the DONeRF, when the training set is limited.}
\label{table:synthetic_data_partial_quantitative_barbershop}
\end{table}
\begin{table}
\centering
\begin{tabular}{c | c | c | c}
\hline
Training Images & Method & PSNR $\uparrow$ & SSIM $\uparrow$ \\
\Xhline{1pt}
\multirow{2}{5em}{20 Images} & DONeRF & 28.83 & 0.90 \\
& P-NeRF & \textbf{29.85} & 0.90 \\
\Xhline{0.7pt}
\multirow{2}{5em}{60 Images} & DONeRF & 29.83 & 0.88 \\
& P-NeRF & \textbf{31.88} & \textbf{0.92} \\
\Xhline{0.7pt}
\multirow{2}{5em}{140 Images} & DONeRF & 34.53 & 0.94 \\
& P-NeRF & \textbf{35.54} & \textbf{0.97} \\
\Xhline{0.7pt}
\end{tabular}
\caption{Quantitative results for the DONeRF and the proposed P-NeRF using different training sets from the Classroom sequence. The proposed approach permits to predict images with a better quality than the DONeRF, when the training set is limited.}
\label{table:synthetic_data_partial_quantitative_classroom}
\end{table}
In this part of the validation, the full training sets (210 views) are considered, and the proposed P-NeRF is compared with methods from state-of-the-art. The results of this comparison are shown in Table \ref{table:synthetic_data_full_quantitative}, where the PSNR values of the baseline methods were taken from Table 2 of \cite{neff2021donerf}. Therefore, the SSIM metric is not considered here as it was not presented in \cite{neff2021donerf}. The considered state-of-the-art baseline methods are listed below:
\begin{itemize}
\item \textbf{DONeRF} \cite{neff2021donerf}: the variant (in terms of sampling strategy and use of ground truth depth) that provides the best image quality in each case is considered here.
\item \textbf{NeRF} \cite{mildenhall2020nerf}: the results shown here were obtained using a NeRF PyTorch implementation (https://github.com/yenchenlin/nerf-pytorch) with a total of $256$ network evaluations per iteration. The mention (log + warp) refers to the sampling strategy proposed by \cite{neff2021donerf} that was applied to the original NeRF.
\item \textbf{NeX (Neural Basis Expansion)} \cite{Wizadwongsa2021NeX}: the implementation provided by the authors (https://github.com/nex-mpi/nex-code/) was considered.
\item \textbf{LLFF (Local Light Field Fusion)} \cite{mildenhall2019llff}: the open source code available at (https://github.com/Fyusion/LLFF) was used for this evaluation.
\end{itemize}
It is possible to see that considering the full training set induces an increase in the quality of the images rendered by DONeRF and P-NeRF. The level of quality reached by the proposed approach, however, is the highest, as all the baseline methods produce lower image quality at test time. This result shows that P-NeRF is better able to generalize to novel view rendering.
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{data/qualitative_result_synthetic.png}
\caption{Example images rendered using P-NeRF and DONeRF. It is possible to notice that more details are present in the images rendered by P-NeRF, and that scene elements (chair, lamp) are dropped by DO-NeRF while being modeled by P-NeRF.}
\label{fig:synthetic_data_full_quantitative}
\end{figure}
\begin{table}
\centering
\begin{tabular}{c | c | c | c}
\hline
Method & Barbershop & Classroom & Mean \\
\Xhline{1.0pt}
P-NeRF & \textbf{34.05} & \textbf{36.74} & \textbf{35.40} \\
DONeRF & 32.80 & 36.27 & 34.54 \\
NeRF & 33.63 & 34.02 & 33.83\\
NeRF (log + warp) & 33.60 & 35.19 & 34.40 \\
NeX & 22.98 & 30.34 & 26.66\\
LLFF & 24.13 & 24.87 & 24.50\\
\hline
\end{tabular}
\caption{Values of the PSNR for the considered methods trained on the full training sets (210 images). The values of the baseline methods are obtained from Table 2 in \cite{neff2021donerf}. The proposed approach predicts images with a better quality than the DONeRF, when the training set is limited.}
\label{table:synthetic_data_full_quantitative}
\end{table}
The result of a qualitative comparison between P-NeRF and DONeRF is presented in Figure \ref{fig:synthetic_data_full_quantitative}. It shows that the proposed P-NeRF is capable of learning a consistent representation of the scenes, permitting to render images with a better quality in comparison with DONeRF. This difference is especially apparent when the training data are limited (20 training views); as this is the case for the Barbershop sequence, where P-NeRF is able to learn to represent scene elements that are dropped by DONeRF.
The results of the validation on synthetic data show that the proposed P-NeRF outperforms the baseline state-of-art methods. This is notably due to the fact that P-NeRF learns a consistent scene representation that permits to achieve a better generalization, and this even when the training data is limited.
\subsection{Validation on Real Data}
The results presented in this part are related to the real dataset provided by \cite{mildenhall2020nerf}. In order to learn the scenes models, the poses of the training frames and a sparse point cloud are first computed using COLMAP \cite{schoenberger2016sfm}. The sparse set of 3D points is then used to integrate depth information into the training process in the following way. The pixels corresponding to the sparse set of estimated 3D points are obtained through projection. Therefore, it is possible to consider for each of these pixels the loss described in Equation (\ref{eq:color_depth_density_loss}), where the corresponding target depth PDF is an approximation of the intersection between the ray passing by the pixel and the ellipsoid representing the uncertainty of the projected 3D point (Figure \ref{fig:3d_point_pnerf}). This allows to account for the uncertainty of the pose that propagates into the camera's $z$ axis. For its part, the corresponding depth is the distance between the camera and the 3D point. In addition to that, pixels with high reprojection error are down-weighted similarly to \cite{kangle2021dsnerf} in order to favor the most precise depth estimates during training. For the remaining set of pixels, only the photometric term is considered.
The comparison is made between the proposed P-NeRF and DS-NeRF \cite{kangle2021dsnerf}. The methodology of \cite{kangle2021dsnerf} was followed to select the test images with indexes {$0$, $8$, $16$, ...} until the end of the sequences. The training set for its part contains $10$ randomly selected images. The number of network evaluations per iteration is 128.
Table \ref{table:quantitatve_real} shows the PSNR and SSIM values reached in the test set for both P-NeRF and DS-NeRF. The proposed method produces better image quality on the test set for both the Fern and Fortress sequences and has the best mean metrics in the test set. This difference in the quality of the rendered images results from the fact that the proposed approach is capable of learning the details of the scene, which are not modeled by DS-NeRF, as shown in Figure \ref{fig:qualitative_real_data}.
\begin{table}[h]
\centering
\begin{tabular}{ c | c | c | c | c}
\hline
\multirow{2}{5em}{Sequence} & \multicolumn{2}{ c | }{DS-NeRF} & \multicolumn{2}{ c }{P-NeRF} \\
\cline{2-5}
& PSNR $\uparrow$ & SSIM $\uparrow$ & PSNR $\uparrow$ & SSIM $\uparrow$ \\
\Xhline{1pt}
Fern & 24.27 & 0.74 & \textbf{25.53} & \textbf{0.80} \\
Fortress & 23.48 & 0.62 & \textbf{24.13} & \textbf{0.70} \\
Flower & \textbf{23.08} & \textbf{0.68} & 22.45 & 0.66 \\
\hline
\end{tabular}
\caption{Quantitative results for the DS-NeRF and the proposed P-NeRF for different training sets. The proposed approach produces better image quality at test time.}
\label{table:quantitatve_real}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{data/ds_nerf_comparison.pdf}
\caption{Rendered images and depth maps using DS-NeRF \cite{kangle2021dsnerf} and P-NeRF with a training set of 10 images. The proposed approach is able to learn more details from the scene.}
\label{fig:qualitative_real_data}
\end{figure}
\section{Conclusions and Perspectives} \label{sec:conclusion}
This paper has proposed a novel method to account for uncertainty in the learning process of neural scene representations and has subsequently made a first step in providing a probabilistic representation that is fundamental for robotics applications. A case study was made based on depth uncertainty to supervise the depth PDF obtained from the NeRF using the available depth information to learn a geometrically consistent scene representation. This explicit introduction of uncertainty supervision into the learning process allowed to take into account different sources of depth uncertainty (depth maps, point clouds, and camera pose). As shown in the results, such probabilistic learning allows to enhance the learning process and permits to reach better image rendering quality compared to state-of-the-art methods, which also use depth information.
One perspective of this approach is to consider different uncertainty model sources in the learning process in order to further improve the quality of the scene representation. For example, the proposed approach could potentially account for probabilistic sources of information such as uncertain stereo matching, uncertain depth prediction, uncertain semantic segmentation, camera extrinsic and intrinsic uncertainty, etc. Further works will involve adapting real-time NeRF approaches~\cite{mueller2022instant} to take into account this probabilistic model for application to a real-world robotic navigation task.
\section*{ACKNOWLEDGMENT}
This work was funded by the EU H2020 MEMEX research project under grant
agreement No. 870743. Parts of the experiments that led to this article were performed using HPC resources from GENCI-IDRIS (Grant 2022-011012578).
\bibliographystyle{plain}
|
\section{Introduction}
Lepton flavor conservation stands out in the Standard Model (SM)
among all other symmetries
because it is not associated with any underlying conserved current.
Lepton flavor violation (LFV) in the charged sector is predicted
by many new physics (NP) models.
The small but finite mass of the neutrinos in the SM allow charged LFV
in neutrino-less two-body decays,
e.g., $\tau^- \to \ell_i^- \gamma$ decays~(charge conjugate modes are implied throughout the text, unless otherwise specified),
where $\ell_i^-$ ($i$ = 1, 2) denotes light charged leptons ($e^-$, $\mu^-$).
However, such decays are suppressed
by a factor of $({m_\nu^2}/{m_W^2})^2$~\cite{Lee:1977tib},
which produces experimentally unreachable rates of the order of $10^{-54}$.
For neutrino-less three-body decays,
e.g., $\tau^- \to \ell_i^- \ell_j^+ \ell_a^- $ decays
(where $a~=~i~\mathrm{or}~j$, and~$i$ may or may not be equal to $j$),
two conflicting predictions existed in the literature:
one of the order of $10^{-55}$~\cite{Petcov:1976ff}
and another of the order of $10^{-14}$~\cite{Pham:1998fq}.
Recently, the~contributions due to finite neutrino masses for such decays
were re-scrutinized and found to be in the range of $[10^{-56},10^{-54}]$~\cite{Hernandez-Tome:2018fbq,Blackstone:2019njl},
thereby laying to rest the claim that such decays could be
of the order of $10^{-14}$ in the SM.
Thus, any observation of charged LFV is an unambiguous signature of~NP.
{{{A discovery}}} of LFV in the charged sector
may {{{provide deeper insight into}}} several unsolved mysteries
such as the origin of the dark sector, large baryon number asymmetry,
number of flavor generations and extra dimensions.
Many NP models, such as low-scale seesaw models~\cite{Cvetic:2002jy},
supersymmetric standard models~\cite{Ellis:1999uq, Ellis:2002fe, Dedes:2002rh, Brignole:2003iv, Masiero:2002jn, Hisano:2009ae,Fukuyama:2003hn},
little Higgs models~\cite{Choudhury:2006sq, Blanke:2007db},
leptoquark models~\cite{Davidson:1993qk},
non-universal $Z^\prime$ models~\cite{Yue:2002ja},
and extended Higgs models~\mbox{\cite{Akeroyd:2009nu, Harnik:2012pb, Celis:2013xja, Omura:2015nja, Goudelis:2011un}},
predict LFV in $\tau$ decays at the level of $10^{-10}$--$10^{-8}$
which are just below the current experimental bounds.
Predictions for two-body and three-body neutrino-less $\tau$ decays
from some of these
NP models~\cite{Cvetic:2002jy,Ellis:1999uq,Ellis:2002fe,Dedes:2002rh,Brignole:2003iv,Masiero:2002jn,Fukuyama:2003hn,Yue:2002ja}
are tabulated in Table~\ref{taulfv_predictions}.
\begin{table}[H]
\setlength{\tabcolsep}{8.15mm}
\caption{The branching fractions (${\cal{B}}$) for $\tau^- \to \ell^- \gamma$ and $\tau^- \to \ell^- \ell^+ \ell^-$ decays in some NP~models.}
\label{taulfv_predictions}
\begin{tabular}{lcc}
\noalign{\hrule height 1pt}
& {\boldmath{$\BR(\taulg)$}} & {\boldmath{$\BR(\taulll)$}}\\\midrule
SM + seesaw~\cite{Cvetic:2002jy} & $10^{-9}$ & $10^{-10}$ \\
SUSY + Higgs~\cite{Dedes:2002rh,Brignole:2003iv} & $10^{-10}$ & $10^{-8}$ \\
SUSY + SO(10)~\cite{Masiero:2002jn,Fukuyama:2003hn} & $10^{-8}$ & $10^{-10}$ \\
Non-universal Z$^\prime$~\cite{Yue:2002ja} & $10^{-9}$ & $10^{-8}$ \\ \noalign{\hrule height 1pt}
\end{tabular}
\end{table}
Since $\tau$ is the only lepton that can decay hadronically,
many neutrino-less final states are allowed via LFV processes in NP models
within the observable parameter space~\cite{Black:2002wh}.
Example of such processes include final states containing
a light lepton and a meson
e.g., $\tau^- \to \ell^- M^0$ (where $M^0 = \pi^0, K_S^0, \eta, \rho^0, \omega,
K^{\star0}, \bar{K}^{\star0}, \eta^\prime, a_0, f_0, \phi$),
or a light lepton and two mesons: $\tau^- \to \ell^- h^+h^-, \;
\ell^+ h^- h^-$, \; $\ell^- h^0 h^0$
(where $h^\pm = \pi^\pm/K^\pm, h^0 = \pi^0/K_S^0$).
Searches for all possible LFV processes in decays of $\tau$ are necessary
because there are strong correlations
between the expected rates of the different channels in various models.
For example, in~some supersymmetric seesaw models~\cite{Babu:2002et, Sher:2002ew},
the relative rates of
${\cal{B}}(\tau^\pm \to \mu^\pm \gamma)$ :
${\cal{B}}(\tau^\pm \to \mu^\pm \mu^+\mu^-)$ :
${\cal{B}}(\tau^\pm \to \mu^\pm \eta)$
are predicted to have specific ratios,
depending on the model parameters.
In the unconstrained minimal supersymmetric model,
which includes various correlations between the $\tau$ and $\mu$ LFV rates,
the LFV branching fractions of the $\tau$ lepton in some decay channels
can be as high as $10^{-8}$~\cite{Brignole:2004ah, Goto:2007ee}),
while still respecting the strong experimental bounds
on the LFV decays of the $\mu$ lepton~\cite{MEG:2016leq,SINDRUM:1987nra}.
Thus, it is critical to probe all possible LFV modes of the $\tau$ lepton,
because any excess in a single channel will not provide sufficient information
to nail down the underlying LFV mechanism or even to identify an underlying~theory.
More exotic decay modes, such as $\tau^-\to\pi^+\ell^-\ell^-$ and $K^+\ell^-\ell^-$,
accompanied by a violation of the lepton number (LNV),
are predicted at the level of $10^{-10}$--$10^{-8}$ in several scenarios
beyond the SM~\cite{LopezCastro:2012udb}.
Several of these decay modes are expected to have branching ratios
close to existing experimental limits in NP models,
e.g., heavy Dirac \mbox{neutrinos~\cite{Gonzalez-Garcia:1991brm, Ilakovac:1999md}},
supersymmetric processes~\cite{Sher:2002ew, Arganda:2008jj},
flavor-changing $Z^\prime$ exchanges with non-universal couplings~\cite{Li:2005rr},
etc., to name a few.
Wrong-sign $\tau^-\to\ell_i^+\ell_j^-\ell_j^-$ decays are very intriguing
because they are expected at rates only one order of magnitude below the present bounds
in some NP models, e.g.,~the Littlest Higgs model
with T-parity realizing an inverse seesaw~\cite{Pacheco:2021djh}.
Most models for baryogenesis, a~hypothetical physical process
based on different descriptions of the interaction between the fundamental particles
that took place during the early universe
producing the observed matter--antimatter asymmetry,
require baryon number violation (BNV),
which in charged lepton decays automatically implies LNV and LFV{{~\cite{LHCb:2013fsr}}}.
Angular momentum conservation requires the difference of
net baryon number (B) and lepton number (L) to be equal to either 0 or 2.
Although the SM conserves this difference,
the symmetry group for the sum of baryon number and lepton number
can be associated with an anomalous current.
A set of models predicts baryogenesis that conserves B-L
but includes instanton induced B+L violating currents~\cite{Fuentes-Martin:2014fxa}.
In a large class of models~\cite{Hou:2005iu},
BNV in $\tau$ decay modes containing baryons in the final state, for~example,
$\tau^- \to \pi^- \Lambda, \pi^- \bar{\Lambda}, K^- \Lambda, K^- \bar{\Lambda},
\bar{p} \gamma, \bar{p} \ell_i \bar{\ell}_j ~\mathrm{and}~ p \ell_i \ell_j$,
are predicted at observable rates in the large $\tau$ data set
that the Belle II detector will record over the coming~years.
\section{Belle II Experiment at~SuperKEKB}
The most restrictive limits on LFV in $\tau$ decays at the level of $10^{-8}$ have been obtained by the first generation of the B-Factory experiments, Belle and \babar,
where a big data sample of $\tau$' s was generated
thanks to large and similar values of the production cross-sections of
$B-$ mesons and $\tau-$ pairs around the $\Upsilon(4S)$ resonance
at the level of a nanobarn ($nb$)~\cite{Banerjee:2007is}.
Belle and \babar\ experiments collected approximately
one attobarn-inverse ($ab^{-1}$) and half an $ab^{-1}$ of $e^-e^+$ annihilation data,
respectively.
The next generation of the B-Factory experiment, Belle II,
is expected to collect 50 $ab^{-1}$ of data over the next decade~\cite{Belle-II:2022cgf}.
Such a huge data sample corresponding to $10^{11}$ single $\tau$-decays
would lower the limits on LFV in $\tau$ decays by one or two orders of~magnitude.
\subsection{Luminosity Upgrade of~SuperKEKB}
The asymmetric beam energy $e^-e^+$ collider, SuperKEKB,
is an upgrade of the KEKB accelerator facility in Tsukuba, Japan,
and has a circumference of about 3~km.
The main components of the SuperKEKB collider complex
are a 7~GeV electron ring known as the high-energy ring (HER),
a 4~GeV positron ring known as the low-energy ring (LER),
and an injector linear accelerator with a 1.1~GeV positron damping ring~\cite{Akai:2018mbz}.
The HER and the LER have four straight sections named Tsukuba, Oho, Fuji, and~Nikko,
with the interaction point in the straight section of Tsukuba,
where the Belle II detector is~located.
The target integrated luminosity of 50 $ab^{-1}$
to be collected by the Belle II experiment
will be achieved by increasing the instantaneous luminosity by a factor of 30.
Two major upgrades account for this increase:
a modest two-fold increase in the beam currents,
and a fifteen-fold reduction of the vertical beta function
at the interaction point $(\beta^\star_y)$ from 5.9~mm to 0.4~mm,
according to the “nano-beam” scheme described~below.
Compared to KEKB, the~asymmetry between the beam energies for the HER/LER beams
were reduced from 8.0/3.5~GeV to 7.0/4.0~GeV,
which reduces the beam loss due to Touschek scattering.
This also improves the solid-angle acceptance of the experiment,
which helps to analyze events with large missing energy.
Additionally, the~effects of synchrotron radiation as a result of higher currents
are mitigated. Since synchrotron radiation is proportional to the product of beam current
and the fourth power of beam energy,
the HER at SuperKEKB emits $(7/8)^4 = 59\%$
as much synchrotron radiation per unit of beam current
compared to KEKB~\cite{Wu:2015gta}.
This facilitates the SuperKEKB collider to operate at a beam current
twice the value of the~KEKB.
The very high luminosity environment of SuperKEKB required
significant upgrades of the injection beams with high current and low emittance.
The upgraded accelerator complex houses a new electron-injection gun
and a new target for positron production.
A new damping ring was installed for injection of the positron beam
with low-emittance, as~well as for improving simultaneous top-up injections
needed for the high luminosity upgrade.
The upgrade also features completely redesigned lattices for the LER and HER,
replacement of short dipoles with longer ones in the LER,
a new Titanium Nitride coated beam pipe with antechambers
to suppress the electron-cloud effect, a~modified RF system,
and a completely redesigned interaction region~\cite{Suetsugu:2016twp}.
The design of the beam parameters at SuperKEKB~\cite{Ohnishi:2013fma}
follows the “nano-beam” and the “crab-waist” schemes,
which were originally proposed for the SuperB-Factory in Italy~\cite{SuperB:2007lel}.
Accordingly, the~transverse sizes of the beam bunches
in the horizontal plane ($\sigma_x$) are squeezed to have very small values
and made to collide at a larger horizontal crossing angle
($2\phi_x$) $=$ $83~{\rm{mrad}}$ at Belle II, instead of =22~{\rm{mrad}} at Belle.
Thus, the~effective size of the overlap region ($\tilde{\sigma}_z$)
is much shorter than what it would have been in the case of a normal head-on collision,
which is given by the longitudinal size of the beam bunches
in the horizontal plane ($\sigma_z$)~\cite{Ohnishi:2021ei}.
The vertical beta function at the interaction point (IP) is constrained
due to the hour-glass effect as:
$\beta^\star_y > \tilde{\sigma}_z = \frac{\sigma^\star_x}{\sin \phi_x} =
\frac{\sigma_z}{\Phi}$, where $\Phi$ is the large Piwinski angle.
With $\sigma_z$ of the order of 6~mm, the~$\beta^\star_y$ is thus squeezed down
to about $400~{\upmu}{\rm{m}}$, which is much shorter than the real bunch length.
In addition to increasing the luminosity,
a reduction of the interaction region of the colliding beams
restricts the vertex position along the beam axis,
thus providing an additional benefit
of more precise estimation of the primary vertex,
which helps in the reconstruction of the complete event topology
during physics~analysis.
The instantaneous luminosity in an $e^-e^+$ collider is
$${\cal{L}} = \frac{N_{e^{-}} N_{e^{+}} f}{4\pi \sigma^\star_x \sigma^\star_y} R_{\cal{L}},$$
where $N_{e^{-}}$ is the number of electrons per bunch,
$N_{e^{+}}$ is the number of positrons per bunch,
$f$ is the collision frequency of the bunch,
$\sigma^\star_x$ and $\sigma^\star_y$ are the transverse beam-profile sizes at the IP,
and $R_{\cal{L}}$ is the luminosity-reduction factor (of the order of unity)
due to the finite beam-crossing angle.
In terms of beam currents $I_{\pm} = N_{e^{\pm}} e f$, the~luminosity becomes
$${\cal{L}} = \frac{N_{e^{\mp}}I_{e^{\pm}}}{e} \frac{1}{4\pi} \frac{ R_{\cal{L}}}{\sigma^\star_x \sigma^\star_y},$$
where $e$ is the charge of the electron. Each beam affects the stability of the other,
which can be characterized by the beam--beam tune shift parameters given by
$$\xi_{(x,y)}^{\pm} =
\frac{r_e}{2\pi\gamma_{\pm}}
\frac{N_{e^{\mp}} \beta^\star_{(x,y)}}{\sigma^{\star}_{(x,y)} \big(\sigma^{\star}_{x} + \sigma^{\star}_{y}\big)}
R_{{\xi}_{(x,y)}},$$
where $r_e$ is the classical radius of the electron, $\gamma_{\pm}$ is the relativistic gamma factor of $e^{-}(e^{+})$ beams,
and $R_{{\xi}_{(x,y)}}$ is the geometric reduction factor (also of the order of unity) due to the hour-glass effect.
Putting all these factors altogether,
we arrive at the following expression for instantaneous luminosity:
$${\cal{L}} = \frac{\gamma_{\pm}}{2er_e}
\Big( 1 + \frac{\sigma^\star_y}{\sigma^\star_x}\Big)
\frac{I_{e^{\pm}}\xi_{y}^{\pm}}{\beta^\star_y}
\Big( \frac{R_{\cal{L}}}{R_{{\xi}_{y}^\pm}} \Big),$$
where the design parameters for beam--beam tune shifts are
$\xi_{(x,y)}$ $=$ $(0.0012,~0.0807)$ for HER and $=$ $(0.0028,~0.0881)$ for LER~\cite{Ohnishi:2013fma}.
The horizontal/vertical beam sizes at the IP are reduced from
$\sigma_x^\star/\sigma_y^\star$ = 170~$\upmu$m/940~{nm} for HER and
=147~$\upmu$m/940~{nm} for LER at Belle to
=10.7~$\upmu$m/62~{nm} for HER and
=10.1~$\upmu$m/48~{nm} for LER at Belle II, respectively~\cite{Akai:2018mbz}.
The beam currents for Belle were 1.19~{\rm{A}}
and 1.64~{\rm{A}} for HER and LER, respectively,
compared to the design values of 2.6~{\rm{A}}
and 3.6~{\rm{A}} for HER and LER, respectively,
at Belle II~\cite{Akai:2018mbz}.
This allows one to improve upon the value of instantaneous luminosity
from \linebreak ${\cal{L}} = 2.1 \times 10^{34}{~\rm{{cm^2}{s^{-1}}}}$ at Belle to
$6.5 \times 10^{35}{~\rm{{cm^2}{s^{-1}}}}$ at Belle II~\cite{Belle-II:2022cgf}.
\subsection{Detector Upgrade of Belle~II}
From the IP outward, the~main components of the Belle II detector are
vertexing and tracking detectors, particle identification systems,
calorimeter and muon chambers,
as shown in Figure~\ref{belle2det}~\cite{Belle-II:2010dht}.
The tracking detectors consist of an inner Silicon PiXel Detector (PXD),
a Silicon Vertex Detector (SVD) and a Central Drift Chamber (CDC).
Two dedicated particle identification systems are
the Time-Of-Propagation (TOP) detector in the barrel region and
the Aerogel Ring-Imaging CHerenkov detector (ARICH) in the forward endcap.
These are surrounded by an Electromagnetic CaLorimeter (ECL) and
a superconducting solenoid providing a homogeneous magnetic field of 1.5 T.
A $K_L^0$ and Muon detector (KLM) is the largest
and outermost part of the Belle II~detector.
\begin{figure}[H]
\centering
\includegraphics*[width=\textwidth,height=.3\textheight]{figure1.jpg}
\caption{Schematic view of the Belle II detector~\cite{Belle-II:2010dht}. The self-annotated figure is based on overview available from SuperKEKB/Belle II public page:
\url{https://www.belle2.org/project/super_kekb_and_belle_ii} (accessed on 11 April 2022). Picture copyright and credit: KEK.}
\label{belle2det}
\end{figure}
Some upgrades of the Belle II detector~\cite{Belle-II:2010dht,Belle-II:2018jsg} over the Belle detector~\cite{Belle:2000cnh} are:
\begin{itemize}
\item Vertexing:\\
In Belle, the~beam pipe was at 15~mm~\cite{Abe:2004ma},
the innermost layer of
a 4-layer silicon vertex detector~\cite{Kibayashi:2006vj}
was at 20~mm
and the outermost layer of the vertex detector was at a radius of 88~mm.
In Belle II, the~beam pipe is at 10~mm,
the inner two layers of the PXD{{, consisting of silicon pixels,}} are closer to the IP at 14~mm and 22~mm,
respectively, and~the outermost layer of the four layers of the SVD{{, consisting of silicon strips,}}
goes to a larger radius of 140~mm.
{{The PXD is}} based
on the Depleted Field Effect Transistor (DEPFET) technology,
which allows for thin sensors with 50~$\upmu$m thickness.
The readout of the new silicon strip detector is based on the APV25 chip,
which has a much shorter shaping time to accommodate for higher background rates
in Belle II than the VAITA chip-based readout used at Belle.
As a result of these upgrades, considerably better performance
is expected in Belle II than Belle.
For example, the~vertex resolution at Belle II is improved by
the excellent spatial resolution of the two innermost pixel detector layers.
\item Tracking:\\
The large volume CDC at Belle II,
with 56 layers organized in 9 super-layers,
has smaller drift cells than in Belle.
CDC starts just outside the expanded silicon strip detector,
and extends to a larger radius of 1130~mm in Belle II
as compared to 880~mm in Belle.
The measured spatial resolution of the CDC is about 100~$\upmu$m,
while the relative precision of the $dE/dx$ measurement
for particles with an incident angle of $90^\circ$ is around $12\%$.
The angular resolution achieved between tracks is $\sim$4.5~mrad.
The efficiency to reconstruct $K_S^0 \to \pi^-\pi^+$ decays in Belle II
is also improved because the silicon strip detector occupies a larger volume.
\item Particle Identification:\\
Belle II has two completely new, more compact particle identification devices
of the Cherenkov imaging type:
TOP in the barrel and ARICH in the endcap regions.
Both detectors are equipped with very fast read-out electronics,
leading to very good kaon versus pion separation in the kinematic limits
of the experiment.
The two Cherenkov detectors are designed to differentiate between
$K$ and $\pi$ particles over the entire momentum range,
and also differentiates among $\pi$, $\mu$, and~$e$ below 1~GeV/c.
\item Calorimetry:\\
The ECL is made of CsI(Tl) scintillation crystals of size 6 cm $\times$ 6 cm each
with high light output, a~short radiation length, and~good mechanical properties,
covering the range of 12$^\circ$ $<\theta<$ 155$^\circ$ in the polar angle,
e.g.,~90\% of solid angle coverage in the center-of-mass system.
The ECL is divided into two parts: the barrel and the endcap.
While the barrel part consists of 6624 crystals,
the endcap part consists of 2112 crystals.
The new electronics of the ECL are of the wave-form-sampling type,
which has particular relevance in missing-energy studies
by reducing the noise due to pile up considerably.
The ECL is able to detect neutral particles in a wide energy range,
from 20~MeV up to 4~GeV,
with a high resolution of $\sigma_E/E$ $=$ 4\% at 100~MeV,
and angular resolution of 13~mrad (3~mrad) at low (high) energies.
This gives a mass resolution for reconstructing $\pi^0\to\gamma\gamma$
of about 4.5~MeV/c$^2$~\cite{BaBar:2014omp}.
\item $K_L^0$ and Muon Detection:\\
The $K_L^0$ and muon detector (KLM) at Belle was based
on glass-electrode resistive plate chambers (RPC).
Since larger backgrounds are expected in the high luminosity environment
at Belle II, the~upgraded KLM system consists of RPC
only in some parts of the barrel.
The two innermost layers in the barrel and the entire endcap section
of KLM at Belle II consist of layers of scintillator strips
with wavelength shifting fibers,
read out by silicon photomultiplier (SiPMs) as light sensors~\cite{Balagura:2005gh}.
Although the high neutron background can cause damage to the SiPMs,
the upgraded KLM has been demonstrated to operate reliably during irradiation tests by
appropriately setting the discrimination thresholds.
\end{itemize}
\subsection{Daq Upgrade of Belle~II}
The new data acquisition (DAQ) system~\cite{Yamada:2015xjy}
meets the requirements of considerably higher event rates at Belle II.
It consists of a Level One (L1)~\cite{Iwasaki:2011za} and High Level Trigger (HLT).
The L1 trigger has a latency of 5~$\upmu$s and
a maximum trigger output rate of 30~kHz,
limited by the read-in rate of the DAQ.
The HLT must suppress online event rates to 10~kHz for offline storage
using complete reconstruction with all available information from the entire detector.
To enable readout from high-speed data transmission,
a peripheral component interconnect express based readout module (PCIe40)
with high data throughput of up to 100 Gigabytes/s was adopted
for the upgrade of the Belle II DAQ system~\cite{Zhou:2020qed}.
The trigger system at Belle II achieves almost
100~$\mathrm{\%}$ trigger efficiency for $\Upsilon(4S)\rightarrow B\bar{B}$ events
and nearly high efficiency for other physics processes of interest,
e.g., $\tau$-pair~events.
\section{Search~Strategies}
\unskip
\subsection{Event~Topology}
\label{EvtTopo}
B-Factories typically operate at center-of-mass energies
around the $\Upsilon(4S)$ resonance, e.g.,~10.58~GeV.
Tau-pair production via $e^-e^+$ annihilation in this energy regime
leads to cleanly separated event topology
associated with the decay of each $\tau$ lepton,
and are well simulated by state-of-the-art event generators:
{\texttt{KK2F}}~\cite{Jadach:1999vf,Ward:2002qq,Arbuzov:2020coe},
{\texttt{Tauola}}~\cite{Jadach:1993hs,Chrzaszcz:2016fte} and
{\texttt{Photos}}~\mbox{\cite{Barberio:1993qi,Davidson:2010ew}}.
Searches for LFV in $\tau$ decays in B factories exploit these event characteristics,
assuming that only one of the two $\tau$'s produced in the $e^-e^+\to\tau^-\tau^+$ process
could have decayed in this rare mode, and~the other $\tau$ decays
via the allowed SM processes. By~dividing the event into a pair of hemispheres
perpendicular to the thrust axis~\cite{Brandt:1964sa,Farhi:1977sg}
in the center-of-mass frame, $\tau$ decay products can thus be identified
as coming from the signal-side and the tag-side,
corresponding to decay via LFV and the SM decays of the $\tau$ lepton, respectively.
\subsection{Signal Characteristics}
The characteristic feature of $\tau$ decays via LFV is that the final state
does not contain $\nu_\tau$.
Thus, there is no missing momentum associated with the signal-side,
and the kinematics of the signal $\tau$ lepton can be completely
reconstructed from measurements of the final state particles.
Simulation studies for more than a hundred possible decays via LFV
that can be searched with such signal characteristics
for each sign of the $\tau$ lepton, are possible
with recent updates of the {\texttt{Tauola}}
event generator~\cite{Chrzaszcz:2016fte,Antropov:2019ald,Banerjee:2021rtn},
which have been seamlessly integrated into the software of the Belle II~experiment.
A very interesting feature of $\tau$-pair production in $e^-e^+$ annihilation
is that the energy of each $\tau$ lepton is known to be exactly
half of the center-of-mass (CM) energy of the collision,
except for corrections due to initial and final state radiations.
Therefore, the~uncertainty of the energy of the $\tau$ lepton
is independent of the performance of the detector,
and is known from the beam energy spread of SuperKEKB
to be approximately 5~MeV~\cite{Belle-II:2018jsg}.
As a first example of the signal mode, let us consider $\tau^- \to \ell^- \gamma$ decays,
which are predicted with rates just lower than the current experimental bounds
in the widest variety of NP models
and are hence regarded as “golden modes” in searches for LFV.
The total energy in the CM frame of the $\tau$ decay products
in the signal-side is $E_{\ell\gamma}^{\mathrm{CM}} = \sqrt{s}/2$,
and the invariant mass of the $\ell\gamma$ pair can be calculated as
$m_{\ell\gamma} = 2 \sqrt{p_\ell E_\gamma} \sin(\frac{\theta}{2})$,
where $p_\ell$ is the magnitude of the three-momentum of the lepton,
$E_\gamma$ is the energy of the photon,
and $\theta$ is the opening angle between them.
The invariant mass is ideal as a discriminating variable,
because its resolution is given by~\cite{Fabre:1992vpa}:\vspace{-6pt}
$$\frac{\Delta m_{\ell\gamma}}{m_{\ell\gamma}}
= \frac{1}{2} \sqrt{ \Big(\frac{\Delta p_\ell}{p_\ell}\Big)^2
+ \Big(\frac{\Delta E_\gamma}{E_\gamma}\Big)^2
+ \Big(\frac{\Delta \theta}{\tan \frac{\theta}{2}}\Big)^2
}
$$
which simultaneously combines all the available experimental precision
on the measured energy/momentum from the calorimeter/tracking systems
with the measured uncertainties on the position measurements of the
observable final state decays products.
The resolution of this kinematic variable is further improved
by considering the beam-energy-constrained mass,
$M_{\mathrm{bc}}$, given as:\vspace{-6pt}
$$
M_{\mathrm{bc}}= \sqrt{(E_{\mathrm{beam}}^{\mathrm{CM}})^{2} - |\vec{p}_{\ell\gamma}^{\mathrm{~CM}}|^{2}},
$$
\noindent where $E_{\mathrm{beam}}^{\mathrm{CM}}=\sqrt{s}/2$
and $\vec{p}_{\ell\gamma}^{\mathrm{~CM}}$ is the sum of the lepton and photon momenta
in the CM frame, because~the resolution of $E_{\mathrm{beam}}^{\mathrm{CM}}$
comes from the accelerator instead of the~detector.
The beam-energy-constrained $\tau$ mass,
labeled somewhat differently as $\mathrm{m}_{\mathrm EC}$ for the \babar\ search~\cite{BaBar:2009hkt},
is typically obtained from a kinematic fit that constrains the CM energy of the $\tau$
to be $\sqrt{s}/2$. Its resolution was further improved in the \babar\ search
by assigning the origin of the photon candidate to the point of the closest approach
of the signal lepton track to the $e^-e^+$ collision axis.
Figure~\ref{fig:minv-mec-mg} shows a comparison study
using a simulated sample of signal $\tau^- \to \mu^- \gamma$ events,
where the resolution of the invariant mass is crudely estimated
to be 18.2~MeV/c$^{2}$ from a single Gaussian fit to the peak of the distribution,
while that of $\mathrm{m}_{\mathrm EC}$ is 12.1~MeV/c$^{2}$,
which improves further to 8.3~MeV/c$^{2}$ after the vertex~constraint.
\begin{figure}[H]
\includegraphics*[width=.3\textwidth,height=0.222\textheight]{figure2a.jpg}
\includegraphics*[width=.3\textwidth,height=0.222\textheight]{figure2b.jpg}
\includegraphics*[width=.3\textwidth,height=0.222\textheight]{figure2c.jpg}
\vspace{-6pt}
\caption{Invariant mass (\textbf{left}), $\mathrm{m}_{\mathrm EC}$ without vertex constraint (\textbf{middle})
and $\mathrm{m}_{\mathrm EC}$ with vertex constraint (\textbf{right})
obtained from simulated samples during analysis strategy development studies
for $\tau^- \to \mu^- \gamma$ search at the \babar\ experiment~\cite{BaBar:2009hkt}.}
\label{fig:minv-mec-mg}
\end{figure}
The most distinguishing feature of signal events is obtained
by considering the characteristic mass of the decay products of the LFV $\tau$ decays
along with the normalized difference in their energy
from half the center-of-mass energy
in $e^-e^+$ annihilation,
so that the search can be uniformly performed
at energies other than the $\Upsilon(4S)$ peak,
to take advantage of the larger luminosity including all the recorded data:
$$\Delta E/\sqrt{s} = ( E_{\ell\gamma}^{\mathrm{CM}} - \sqrt{s}/2 )/\sqrt{s}.$$
The signal events are clustered around
$M_{\mathrm{bc}} \sim m_{\tau}$ and $\Delta E/\sqrt{s} \sim 0$
in the two-dimensional plots of $\Delta E$ vs. $M_{\mathrm{bc}}$,
as shown in Figures~\ref{fig:belle_taulg} and~\ref{fig:babar_taulg}
for $\tau^- \to e^- \gamma$ (left) and $\tau^- \to \mu^- \gamma$ (right) searches
at Belle~\cite{Belle:2021ysv} and \babar~\cite{BaBar:2009hkt} experiments, respectively,
where the variable $m_{EC}$ refers to the beam-energy-constrained mass in the latter,
as mentioned~earlier.
\begin{figure}[H]
\includegraphics*[width=.49\textwidth]{figure3a.jpg}
\includegraphics*[width=.49\textwidth]{figure3b.jpg}
\caption{Two-dimensional distributions of $\Delta E/\sqrt{s}$ versus $M_\mathrm{bc}$
for Belle searches~\cite{Belle:2021ysv}
for $\tau^- \to e^- \gamma$ (\textbf{left}) and $\tau^- \to \mu^- \gamma$ (\textbf{right}) decays.
Black points are data, blue squares are $\tau^- \to \ell^- \gamma$ signal MC events,
and magenta ellipses show the signal region~($\pm 2\sigma$ region).
This figure has been reprinted with permission from Ref.~\cite{Belle:2021ysv}.
}
\label{fig:belle_taulg}
\end{figure}
\unskip
\begin{figure}[H]
\includegraphics*[width=.49\textwidth]{figure4a.jpg}
\includegraphics*[width=.49\textwidth]{figure4b.jpg}
\caption{
Two-dimensional distributions in the $\mathrm{m}_{\mathrm EC}$ vs. $\Delta {\rm{E}}$ plane for the \babar\
searches~\cite{BaBar:2009hkt}
for $\tau^- \to e^- \gamma$ (\textbf{left}) and $\tau^- \to \mu^- \gamma$ (\textbf{right}) decays.
Data are shown as dots and contours containing 90\% (50\%) of signal MC events
are shown as light-yellow (dark-green) shaded regions,
along with the $2\sigma$ contours shown as black ellipses.
This figure has been reprinted with permission from Ref.~\cite{BaBar:2009hkt}.
}
\label{fig:babar_taulg}
\end{figure}
All analyses are developed in a blind manner,
e.g., optimizing the event selection before
looking at the data events inside the signal region,
to avoid experimental bias in the search for LFV $\tau$ decays.
The search sensitivity can be optimized to give the smallest expected upper limit
in the background-only hypothesis inside a $2\sigma$ ellipse,
for example, amongst other possible choices.
A typical $2\sigma$ signal region is defined as the following elliptical regions:
\begin{eqnarray*} \label{eq:sr}
\frac{(M_{\mathrm{bc}} - \mu_{M_{\mathrm{bc}}})^{2}}{(2\sigma_{M_{\mathrm{bc}}})^{2}}
&+&
\frac{(\Delta E/\sqrt{s} - \mu_{\Delta E /\sqrt{s}})^{2}}{(2\sigma_{\Delta E/\sqrt{s}})^{2}} < 1.0, \\ \nonumber
\\ \nonumber
\sigma_{M_{\mathrm{bc}}} &=& 0.5(\sigma_{M_{\mathrm{bc}}}^{\mathrm{high}} + \sigma_{M_{\mathrm{bc}}}^{\mathrm{low}}), \\ \nonumber
\\ \nonumber
\sigma_{\Delta E/\sqrt{s}} &=& 0.5(\sigma_{\Delta E/\sqrt{s}}^{\mathrm{high}} + \sigma_{\Delta E/\sqrt{s}}^{\mathrm{low}}). \nonumber
\end{eqnarray*}
Here, $\sigma^{\mathrm{high/low}}_{M_{\mathrm{bc}}}$ and $\sigma^{\mathrm{high/low}}_{\Delta E/\sqrt{s}}$
are the widths on the higher$/$lower side of the peak obtained by fitting the signal distribution to an asymmetric Gaussian function~\cite{Belle:2007qih}.
\textls[-25]{For the Belle search~\cite{Belle:2021ysv}, the~resolutions are
$\sigma_{M_{\mathrm{bc}}}^{\mathrm{high/low}} = 11.55\pm0.27/10.59\pm0.19$ MeV/c$^{2}$} and
$\sigma_{\Delta E/ \sqrt{s}}^{\mathrm{high/low}} = (6.1\pm0.7)/(4.4\pm0.3)\times10^{-3}$ for $\tau^- \to e^- \gamma$ events;
and $\sigma_{M_{\mathrm{bc}}}^{\mathrm{high/low}} = 11.08\pm0.08/7.46\pm0.23$ MeV/c$^{2}$ and
$\sigma_{\Delta E/ \sqrt{s}}^{\mathrm{high/low}} = (5.6\pm0.4)/(4.2\pm0.2)\times10^{-3}$ for $\tau^- \to \mu^- \gamma$ events.
The mean values are
$\langle {M_{\mathrm{bc}}} \rangle = 1.79$ MeV/c$^{2}$
and $\langle {\Delta E/\sqrt{s}} \rangle = -1.0\times10^{-3}$ for $\tau^- \to e^- \gamma$ events,
and
$\langle {M_{\mathrm{bc}}} \rangle = 1.78$ MeV/c$^{2}$ and $\langle {\Delta E/\sqrt{s}} \rangle = -0.6\times10^{-3}$ for $\tau^- \to \mu^- \gamma$ events.
For the \babar\ search~\cite{BaBar:2009hkt},
the resolutions for $\mathrm{m}_{\mathrm EC}$ and $\Delta {\rm{E}}$ are 8.6 MeV/c$^{2}$ and \mbox{42.1 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace} for $\tau^- \to e^- \gamma$,
and 8.3 MeV/c$^{2}$ and 42.2 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace for $\tau^- \to \mu^- \gamma$, respectively,
centered on 1777.3 MeV/c$^{2}$ and $-$21.4 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace for $\tau^- \to e^- \gamma$,
and 1777.4 MeV/c$^{2}$ and $-$18.3 \ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace for $\tau^- \to \mu^- \gamma$.
Shifts from zero for $\langle \Delta {\rm{E}} \rangle$
are mostly due to initial and final state~radiations.
The mass and energy kinematic variables typically have a small correlation
arising from initial and final state radiation,
as well as energy/momentum scale calibration effects.
For the \babar\ search~\cite{BaBar:2009hkt}, the~correlation was estimated to be
$-$8.5\% and $-$8.4\% for the $\tau^- \to e^- \gamma$ and $\tau^- \to \mu^- \gamma$ decays, respectively,
around the core region.
Without the beam-energy constraint, the~correlation between the invariant mass
and energy variables are typically much higher.
LFV process in $\tau$ decays containing a resonance in the final state
are identified by the presence of a peak in the invariant mass of the daughter particles
in the simulation of the signal process.
For example, distributions of invariant mass of the
$\pi^+\pi^-$, $K^+K^-$, $\pi^+\pi^-\pi^0$, $K^+\pi^-$, $\pi^+K^-$ and $\pi^+\pi^-$ systems
in the signal-side are studied {{and confirmed to contain the respective resonances}} in searches for
$\tau^- \to \mu^- \rho^0$,
$\tau^- \to \mu^- \phi$,
$\tau^- \to \mu^- \omega$,
$\tau^- \to \mu^- K^{\ast 0}$,
$\tau^- \to \mu^- \bar{K}^{\ast0}$ and
$\tau^- \to \mu^- f_0(980)$ decays, respectively,
performed by the Belle experiment~\cite{Belle:2011ogy, Belle:2008pdf}.
The selected mass regions ensure that the signal is unambiguously selected in the corresponding~searches.
\subsection{Background Suppression}
Background events containing leptons from decays of heavy quarks
are easily suppressed by appropriate cuts on Fox-Wolfram moments~\cite{Fox:1978vu},
and on the invariant mass of all decay products on the tag-side.
The characteristic difference between $\tau$-pairs events with LFV decays
and backgrounds consisting of generic $\tau$-pair, di-lepton,
two-photon production and $q\bar{q}$ processes
{{(where $q = u,~ d \mathrm{~or~} s$),}}
in the number of neutrinos in the signal-side and tag-side,
as defined by the event topology in Section~\ref{EvtTopo},
are shown in Table~\ref{NumNeutrino}.
\begin{table}[H]
\setlength{\tabcolsep}{6.2mm}
\caption{Number of neutrinos in the event for signal and background processes.}
\label{NumNeutrino}
\begin{tabular} {cccc}
\toprule
\textbf{ \# }\textbf{of }\boldmath{$\nu$}\textbf{'s} & \textbf{LFV Decays }& \textbf{Generic} \boldmath{$\tau$}\textbf{-Pair} & \textbf{Other Backgrounds} \\ \midrule
Signal-side & 0 & 1{--}2 & 0\\
Tag-side & 1{--}2 & 1{--}2 & 0\\\bottomrule
\end{tabular}
\end{table}
Since decay products of the $\tau$ decay via LFV in the signal-side
do not contain any neutrino, the~direction of the $\tau$ lepton
in the tag-side can be precisely obtained in the center-of-mass frame
by reversing the total momentum of the signal-side.
This allows for good kinematic reconstruction of the missing mass in the tag-side,
assuming that in the CM frame, the~tag-side $\tau$ momentum is opposite that of
the signal-side $\tau$ momentum and that its energy is constrained to be
half the center-of-mass energy.
Thus, {{selection of events with small values of}} the square of the missing mass ($m_\nu^2$) in the tag-side
play an important role in the suppression of the background events~\cite{Belle:2021ysv}.
Additional selection criteria are also used to suppress
the backgrounds in the different LFV decay modes,
which are mostly accidental in nature, except~in $\tau^- \to e^- \gamma$ and $\tau^- \to \mu^- \gamma$ searches.
The dominant background in the searches arise from $\tau^{+}\tau^{-}$ events decaying
via the $\tau^{\pm}\rightarrow e^{\pm}\nu_{e}\nu_{\tau}$
~($\tau^{\pm}\rightarrow\mu^{\pm}\nu_{\mu}\nu_{\tau}$)
channel with a photon coming from initial-state radiation or beam background.
The $e^{+}e^{-}\gamma$ and $\mu^{+}\mu^{-}\gamma$ events are subdominant,
and are estimated to contribute to <5\% of the total backgrounds
in the Belle search~\cite{Belle:2021ysv}.
Contributions from other sources of backgrounds,
such as two-photon and $q\bar{q}$ processes,
are estimated to be quite small in the signal~region.
Furthermore, each component of the background processes has distinctive features
as visible in their respective two-dimensional distributions
in the $(\Delta M, \Delta {\rm{E}})$ plane,
where $\Delta M$ denotes the difference between the characteristic mass
of the system of $\tau$-daughters and the well-known mass of the $\tau$-lepton
= $(1776.86 \pm 0.12)\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$~\cite{PDG2020},
and $\Delta {\rm{E}}$, as defined above. The~shapes of the leading backgrounds
in search of $\tau^- \to \mmm$ decays
as performed at the \babar\ experiment~\cite{BaBar:2010axs}
are shown in Figure~\ref{fig:taulll_bkg_dmde},
where the red box indicates the rectangular boundaries of a generic region
mostly populated by the signal processes. The~SM $\tau\tau$ background events
are generally restricted to {{small}} negative values of both $\Delta M$ and $\Delta {\rm{E}}$ variables,
because the {{reconstruction of signal}} event topology does not account for the {{neutrinos present in SM $\tau$ decays}}.
QED background events are mostly dominated by di-lepton production
as the main underlying hard process and typically lie within a narrow horizontal band
across the $\Delta M$ variable centered around slightly positive values of $\Delta {\rm{E}}$,
due to the presence of a pair of extra charged particles in such events.
The QCD background events from various $q\bar{q}$ processes
tend to populate the plane uniformly
across the $\Delta M$ variable and drop towards large values of the $\Delta {\rm{E}}$ variable.
The expected background rates inside the signal region can be obtained
by fitting the observed data in the $(\Delta M, \Delta {\rm{E}})$ plane
to a sum of probability density functions. Such data-driven estimates,
based on the shapes predicted by respective simulation samples
and validated by data-driven control regions, scale well with larger data statistics.
Thus, the~background uncertainties can be controlled in a statistical manner,
which is very useful in rare searches with high luminosity data~sets.
\begin{figure}[H]
\includegraphics*[width=\textwidth,height=.2\textheight]{figure5.jpg}
\caption{Two-dimensional distributions for $\tau\tau$, QED, and~$q\bar{q}$ background processes in the $(\Delta M, \Delta {\rm{E}})$ plane
obtained from simulated samples during analysis strategy development studies
for $\tau^- \to \mmm$ search at the \babar\ experiment~\cite{BaBar:2010axs}.
Rectangular box-shaped regions mostly populated by the signal processes are shown in red.}
\label{fig:taulll_bkg_dmde}
\end{figure}
\subsection{Upper Limit Estimation}
No excess of events has ever been observed in searches for LFV in $\tau$ decays.
The upper limit at 90\% confidence level (CL) on signal branching fraction (${\cal{B}}^{90}_{UL}$) is calculated as:
$${\cal{B}}^{90}_{UL} = \frac{S^{90}_{UL}}{2N_{\tau\tau} \epsilon},$$
where
$N_{\tau\tau}$ is the number of $\tau$-pairs produced,
$\epsilon$ is the reconstruction efficiency of the signal decay mode,
and $S^{90}_{UL}$ is the 90\% CL upper limit on number of signal events.
The factor of two enters into the denominator because either one of the two $\tau$ leptons produced in the event
can decay into the rare signal channel coming from~LFV.
\subsubsection{Number of $ \tau$-Pairs~Produced}
$N_{\tau\tau}$ is obtained by summing over
the product of luminosity $({\cal{L}})$ and the $\tau$-pairs production cross-section $(\sigma_{\tau\tau})$
at each of the center-of-mass energies ($\sqrt{s}$) where the search is conducted.
The cross-section of $\tau$-pair production in $e^-e^+$ annihilation follows a characteristic ${1/s}$ dependence,
but receives additional contribution from decays of the $\Upsilon(nS)$ resonances for $n = 1, ~2 {~\rm{and}~} 3$,
according to the known branching fractions~\cite{PDG2020}:
${\cal{B}}(\Upsilon (1S) \to \tau^-\tau^+) = (2.6 \pm 0.1) \%$,
${\cal{B}}(\Upsilon ({{2}}S) \to \tau^-\tau^+) = (2.0 \pm 0.2) \%$ and
${\cal{B}}(\Upsilon ({{3}}S) \to \tau^-\tau^+) = (2.3 \pm 0.3) \%$.
Beyond the open-beauty threshold,
contributions from resonances with $n\geq 4$ are~negligible.
The Belle experiment collected data with center-of-mass energies around the peak of
$\Upsilon (nS)$ resonances corresponding to luminosities of $5.7\ensuremath{\mbox{\,fb}^{-1}}\xspace$, $24.9\ensuremath{\mbox{\,fb}^{-1}}\xspace$ and $2.9\ensuremath{\mbox{\,fb}^{-1}}\xspace$ at $n = 1, ~2 {~\rm{and}~} 3$, respectively,
while the \babar\ experiment collected luminosities of $13.6\ensuremath{\mbox{\,fb}^{-1}}\xspace$ and $28.0\ensuremath{\mbox{\,fb}^{-1}}\xspace$
at center-of-mass energies corresponding to the $\Upsilon (nS)$ peak
at $n = 2{~\rm{and}~} 3$, respectively,
as reported in Table~3.2.1 in the reference~\cite{BaBar:2014omp}.
The statistical errors on these measured luminosities are much smaller than
the systematic errors, which are estimated to be 1.4\% at the Belle experiment,
and 0.7\% (0.6\%) at the \babar\ experiment for $n = 2{~\rm{and}~} 3$, respectively~\cite{BaBar:2014omp}.
The numbers of $\Upsilon (nS)$ produced in the Belle experiment are
$(102 \pm 2) \times 10^6$,
\mbox{$(158 \pm 4)$} $ \times 10^6$ and
$(11 \pm 0.3) \times 10^6$
for $n = 1, ~2 {~\rm{and}~} 3$, respectively,
while in the \babar\ experiment, the numbers are
$(98.3 \pm 0.9) \times 10^6$ and
$(121.3 \pm 1.2) \times 10^6$
for $n = 2{~\rm{and}~} 3$, respectively,
as obtained from Table~3.2.2 in the reference~\cite{BaBar:2014omp}.
By dividing the numbers of $\Upsilon (nS)$ resonances produced
with its corresponding luminosity,
the resonant production cross-sections are estimated
for $\Upsilon(1S)$ to be $(17.89 \pm 0.43)~nb$ in the Belle experiment,
for $\Upsilon(2S)$ to be $(6.35 \pm 0.18)~nb$ and $(7.23 \pm 0.08)~nb$
in the Belle and \babar\ experiments, and~ for $\Upsilon(3S)$ to be $(3.79 \pm 0.12)~nb$ and $(4.33 \pm 0.05)~nb$
in the Belle and \babar\ experiments, respectively.
For the purposes of averaging the measured values, the~observed differences at $n = 2{~\rm{and}~} 3$ resonances are accounted for
by calculating the PDG-style scale factors~\cite{PDG2020} $S=\sqrt{\chi^2/(N-1)}$ equal to 4.38 and 4.24, respectively,
with $N=2$. Thus, the~average values are estimated to be
$(7.08 \pm 0.33)~nb$ for $\Upsilon(2S)$ and $(4.24 \pm 0.20)~nb$ for $\Upsilon(3S)$.
The total $\sigma_{\tau\tau}$ is obtained by adding the contributions
to $\tau$-pairs production from the continuum~\cite{Banerjee:2007is}
and the contributions from the decays of the $\Upsilon(nS)$ resonances
for $n = 1, ~2 {~\rm{and}~} 3$.
The estimated values of total $\sigma_{\tau\tau}$
at the peak of $\Upsilon(nS)$ resonances
and at $60\ensuremath{\mathrm{\,Me\kern -0.1em V}}\xspace$ below the corresponding resonances (labelled with a “—off”)
are listed in Table~\ref{tab:cross}.
\begin{table}[H]
\setlength{\tabcolsep}{9mm}
\caption{$\sigma_{\tau\tau}$ at different center-of-mass energies corresponding to data-taking at the~B-Factories.}
\label{tab:cross}
\begin{tabular} {lclc}
\noalign{\hrule height 1pt}
\boldmath{$\sqrt{s}$} & \boldmath{$\sigma_{\tau\tau}$ $(nb)$} & \boldmath{$\sqrt{s}$} & \boldmath{$\sigma_{\tau\tau}$ $(nb)$} \\
\midrule
$\Upsilon(1S)$—off & $1.142 \pm 0.004$ & $\Upsilon(1S)$ & $1.593 \pm 0.021$ \\
$\Upsilon(2S)$—off & $1.026 \pm 0.004$ & $\Upsilon(2S)$ & $1.157 \pm 0.017$ \\
$\Upsilon(3S)$—off & $0.966 \pm 0.003$ & $\Upsilon(3S)$ & $1.052 \pm 0.014$ \\
$\Upsilon(4S)$—off & $0.928 \pm 0.003$ & $\Upsilon(4S)$ & $0.919 \pm 0.003$ \\
$\Upsilon(5S)$—off & $0.881 \pm 0.003$ & $\Upsilon(5S)$ & $0.873 \pm 0.003$ \\
\noalign{\hrule height 1pt}
\end{tabular}
\end{table}
\unskip
\subsubsection{Efficiency of Signal~Reconstruction}
{{\label{sec:signal_eff}}}
The signal reconstruction efficiency receives multiplicative reduction factors
corresponding to the application of trigger, acceptance, and~event topology requirements,
particle identification criteria, background suppression and
choice of the signal region in the two-dimensional plane
given by mass versus the normalized difference of energy of the $\tau$ decay products
and $\sqrt{s}/2$.
At Belle and \babar, the~signal efficiencies were estimated to lie approximately
between $2\%$ to $12\%$, depending on the different decay channels.
For example, the~overall signal efficiency estimated
for the search for $\tau^-\to e^- \eta^\prime$ reconstructed
via the $\eta^\prime\to\rho(\to\pi^-\pi^+)\gamma$ and $\eta^\prime\to\pi^-\pi^+\eta(\to\gamma\gamma))$ decay modes
is $(0.294 \times 1 \times 4.76 + 0.445 \times 0.3943 \times 4.27)\% = 2.1\%$~\cite{Belle:2007cio},
while the search for $\tau^-\to\mu^+e^-e^-$ decays have an efficiency of 11.5\%~\cite{Hayasaka:2010np}.
In the Belle II experiment, an~increase in the signal efficiency can be expected due to higher trigger efficiencies,
improvements in the vertex reconstruction, charged track and neutral meson reconstructions, and~particle identification.
Refinements in the analysis techniques will produce a more accurate understanding
of the physics backgrounds and would thus contribute to an increase in the signal detection efficiency, which directly translates into higher sensitivities
in searches for~LFV.
\subsubsection{Upper Limit on the Number of Signal~Events}
In the case of searches with very low counts,
the search becomes a single-bin counting experiment
following a Poisson probability distribution,
with the mean count given by the expected
number of background events ($b$)
and possibly some signal events ($s$).
The likelihood function $({\cal {L}})$ is thus described by:
$$ {\cal {L}}(s,b) = \frac{e^{-(s+b)}(s+b)^N}{N!},$$
where $N$ is the number of observed events.
If the experimental resolution of the discriminating variables allows multiple bins,
the difference in shapes of the discriminating variables between the signal and background distributions
can be exploited in an extended unbinned maximum likelihood fit using:
$$
{\cal {L}}(s,b)=\frac{e^{-(s+b)}} {N!}
\prod_{i=1}^{N} (s\cdot PDF^{sig}_i + b \cdot PDF^{bkg}_i) ,
$$
\noindent where $i$ indicates the $i$-th event, $PDF^{sig}$ and $PDF^{bkg}$
are the probability density functions (PDF) for signal and sum of all background process, respectively.
$S^{90}_{UL}$ is obtained by considering
$ {\cal {L}}(s) = \int_0^\infty {\cal {L}}(s,b) db$,
and integrating the likelihood ${\cal {L}}(s)$ up to the value that
includes 90\% of the total integral of the likelihood function,
following a flat prior Bayesian prescription~\cite{Helene:1982pb}.
Alternatively, following a Frequentist prescription~\cite{Narsky:1999kt},
a toy Monte Carlo approach is used to generate numerous samples
with sizes that follow a Poisson distribution about the mean value
being given by the number of observed events.
Each sample is then fitted to obtain the number of signal and background events using
the same extended unbinned maximum likelihood fit procedure
as that applied to the data.
$S^{90}_{UL}$ is obtained by varying the true branching fraction of the signal
such that 90\% of the samples
yield a fitted number of signal events greater than the number of signal events in the observed data sample.
In the unified approach for finding confidence levels~\cite{Feldman:1997qc},
the order of samples in the acceptance interval
for a specific value of the number of signal events
follows an ordering principle based on likelihood ratios,
where the denominator is determined by the best fit value in each~sample.
In order to have an unbiased estimate of the expected sensitivity, a~blinding procedure should be followed to predict the expected background rate
inside the signal region (SR),
which does not depend on the observed data inside a blinding region (BR),
defined as a part of a broad fit range (FR), but~hiding data events inside the SR.
For well-controlled modeling of the total background PDF,
the number of expected background events
($N^{bkg}_{SR}$) inside the SR can then be estimated directly
from the data $N^{data}_{FR-BR}$ outside the blinded region using the formula:
$$ N^{bkg}_{SR} = \frac{\int_{SR} PDF_{bkg}}
{\int_{FR-BR} PDF_{bkg}}
\times N^{data}_{FR-BR}$$
where $\int_{SR}PDF_{bkg}$ and $\int_{FR-BR}PDF_{bkg}$
are the integrals of the background probability density functions
over the signal region and the non-blinded parts of the fit~region.
\subsection{Systematic~Uncertainties}
In terms of the number of $\tau$ decays being studied,
$t = 2 N_{\tau\tau} \epsilon$,
the number of signal events are written as $(s = \mu t)$,
where $\mu$ is the branching fraction of the signal process
and the normalization factor ${t}$ includes uncertainities on
luminosity, cross-section and the signal efficiency.
The upper limit ${\cal{B}}^{90}_{UL}$ including all systematic effects
using the technique of Cousins and Highland~\cite{Cousins:1991qz},
is calculated by propagating all the measured uncertainties
onto the number of signal events $(s)$ and background events $(b)$.
Implementation of systematic effects
in the POLE (POisson Limit Estimator) program~\cite{Conrad:2002kn}
is based on the following likelihood function,
which is a convolution of a Poisson distribution
with two Gaussian resolution functions corresponding
to the signal normalization factor and background,
as described by the following formula:
$$
{{\cal L}}((\mu,t,b) = {(\mu t + b)^N e^{-(\mu t + b)} \over N!}
\frac{1}{2\pi\sigma_t\sigma_b}
e^{-{1\over 2}\left({\hat{t} - t\over\sigma_t}\right)^2
-{1\over 2}\left({\hat{b} - b\over\sigma_b}\right)^2},
$$
where $\hat t$ and $\hat b$ are the average estimates corresponding
to measured uncertainties of $\sigma_t$ and $\sigma_b$, respectively.
In searches for rare processes such as LFV in $\tau$ decays,
often a very small number of events are expected in the signal region.
Sometimes the sensitivity of the search cannot easily distinguish
a very small number of signal events
from the background-only hypothesis, and~inappropriately tends
to exclude an unusually small signal value.
To overcome such difficulties, the~upper limits can be calculated
using the $CL_s$ method~\cite{Junk:1999kv, Read:2002hq},
where the $CL_s$ is defined as the ratio of confidence levels
for the signal-plus-background hypothesis
normalized by the confidence level for the background-only hypothesis.
Asymptotic calculations of the likelihood ratios used as the test statistic
in such methods allow for a computationally efficient estimate
of the $CL_s$ intervals~\cite{Cowan:2010js}.
A Neyman construction~\cite{Neyman:1937uhy} of $CL_s$ upper limits
including systematic uncertainties is provided by the \texttt{HistFactory} implementation~\cite{Moneta:2010pm,Cranmer:2012sba},
based on the likelihood function $\mathcal{L}(\mu,\theta_j)$ defined as:
$$
\mathcal{L}(\mu, \theta_j) =
\prod_{\textrm{channel~}} \prod_{\textrm{category~}}
\left[ \prod_{i}
\text{Poisson}(N_i|\mu \cdot t_i + b_i)
\prod_{j}
\text{Lognormal}(\theta_j|0,1)
\right],
$$
where $N_i$ is the number of events observed in the $i^{th}$ bin with
signal normalization factor and background predictions given by $t_i$ and $b_i$,
respectively, of~a multicategorical search describing,
for example, different tag-side decay modes each with different sensitivity
over possibly multiple decay channels of the signal mode.
The systematic uncertainties are constrained by nuisance parameters
$\theta_j$ corresponding to various scale factors
as determined from dedicated calibration constants of efficiency measurements
and are obtained from simulation studies or analysis of control regions in the~data.
The \texttt{HistFactory} allows for the calculation of upper limits in both
Bayesian and Frequentist interpretations~\cite{Cranmer:2021urp},
with slightly different treatments of the nuisance parameters.
While in the former interpretation, the~nuisance parameters
are eliminated by marginalizing the posterior density,
using, for~example, Markov Chain Monte Carlo integration,
in the latter interpretation, the~nuisance parameters
are determined by profiling the likelihood function
based on auxiliary measurements, such as control regions,
side-bands, or~dedicated calibration measurements.
Some uncertainties arising from theoretical calculations
or ad hoc estimates are not statistical in nature
and thus are not associated with auxiliary measurements.
However, log-normal probability density functions of nuisance parameters
are used to constrain all the uncertainties, by~convention.
Bayesian limits can also be calculated using the Bayesian Analysis Toolkit~\cite{Caldwell:2008fw}.
\section{Current Status and Future~Prospects}
Summary of observed limits obtained by CLEO, \babar, Belle, ATLAS, CMS, and~LHCb experiments~\cite{HFLAV:2019otj}
are shown in Table~\ref{tab:TauDecays} and Figure~\ref{fig:TauLFV}, along with
projections for two illustrative scenarios of luminosity ${\cal{L}}$ = 5~\ensuremath{\mbox{\,ab}^{-1}}\xspace and 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace at the Belle II experiment~\cite{Belle-II:2022cgf}.
Projections are extrapolated from expected limits obtained at the Belle experiment.
The expected limits for $\tau^- \to \ell^-\gamma$ decays are obtained from Ref.~\cite{Belle:2021ysv}.
We assume the presence of irreducible backgrounds for $\tau^- \to \ell^-\gamma$ decays,
thus approximating the sensitivity to upper bounds
as proportional to $1/\sqrt{\cal{L}}$.
{{Given the expected number of background events in each channel from the previous searches at the Belle experiment and the improvements listed in Section~\ref{sec:signal_eff},
the background expectations corresponding to the integrated luminosity at Belle II for all other modes are still of the order of unity or less.
For such accidental backgrounds, the~sensitivity for upper bounds is proportional to $1/{\cal{L}}$, as~discussed in Section~8.2.1.3 of Ref.~\cite{Raidal:2008jk}.
The projections for the corresponding upper limits at Belle II are estimated using the Feldman and Cousins approach~\cite{Feldman:1997qc}.}
\begin{table}[H]
\caption{Current status of observed (obs) and expected (exp) upper limits (UL)~\cite{PDG2020,Belle-II:2022cgf}.
} \label{tab:TauDecays}
\begin{adjustwidth}{-\extralength}{0cm}
\setlength{\cellWidtha}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthb}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthc}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthd}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthe}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthf}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthg}{\fulllength/7-2\tabcolsep-0in}
\scalebox{1}[1]{\begin{tabularx}{\fulllength}{>{\PreserveBackslash\centering}m{\cellWidtha}>{\PreserveBackslash\centering}m{\cellWidthb}>{\PreserveBackslash\centering}m{\cellWidthc}>{\PreserveBackslash\centering}m{\cellWidthd}>{\PreserveBackslash\centering}m{\cellWidthe}>{\PreserveBackslash\centering}m{\cellWidthf}>{\PreserveBackslash\centering}m{\cellWidthg}}
\noalign{\hrule height 1pt}
\multicolumn{1}{c}{\textbf{}} & \multicolumn{3}{c}{\textbf{Observed Limits}} & \multicolumn{3}{c}{\textbf{Expected Limits}} \\ \midrule
\boldmath{$\tau^-\to$} & \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (obs)}& \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (exp)} \\ \midrule
$e^-\gamma$ & Belle~\cite{Belle:2021ysv} & 988~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 5.6$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50 \ensuremath{\mbox{\,ab}^{-1}}\xspace & 9.0$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2009hkt} & 516~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.3$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1996sqd} & 4.68~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 2.7$~\times~10^{-6}$ & & & \\\midrule
$\mu^-\gamma$ & Belle~\cite{Belle:2021ysv} & 988~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.2$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50 \ensuremath{\mbox{\,ab}^{-1}}\xspace & 6.9$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2009hkt} & 516~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.4$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1999lvl} & 13.8~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.1$~\times~10^{-6}$ & & & \\\midrule
$e^-\pi^0$ & Belle~\cite{Belle:2007cio} & 401~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 8.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 7.3$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2006jhm} & 339~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.3$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997gxa} & 4.68~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 3.7$~\times~10^{-6}$ & & & \\\midrule
$\mu^-\pi^0$ & Belle~\cite{Belle:2007cio} & 401~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.2$~\times~10^{-7}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 7.1$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2006jhm} & 339~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.1$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997gxa} & 4.68~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 4.0$~\times~10^{-6}$ & & & \\\midrule
$e^- K_S^0$ & Belle~\cite{Belle:2010rxj} & 671~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.6$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.0$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009qra} & 469~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.3$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:2002lxe} & 13.9~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 9.1$~\times~10^{-7}$ & & & \\\midrule
$\mu^- K_S^0$ & Belle~\cite{Belle:2010rxj} & 671~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.3$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.0$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009qra} & 469~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.0$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:2002lxe} & 13.9~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 9.5$~\times~10^{-7}$ & & & \\\midrule
$e^- \eta$ & Belle~\cite{Belle:2007cio} & 401~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 9.2$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.2$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2006jhm} & 339~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.6$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997gxa} & 4.68~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 8.2$~\times~10^{-6}$ & & & \\\midrule
$\mu^-\eta$ & Belle~\cite{Belle:2007cio} & 401~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 6.5$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 8.0$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2006jhm} & 339~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.5$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997gxa} & 4.68~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 9.6$~\times~10^{-6}$ & & & \\
\midrule
$e^-\eta^\prime$ & Belle~\cite{Belle:2007cio} & 401~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.6$~\times~10^{-7}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.2$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2006jhm} & 339~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.4$~\times~10^{-7}$ & & & \\\midrule
$\mu^-\eta^\prime$ & Belle~\cite{Belle:2007cio} & 401~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.3$~\times~10^{-7}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.2$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2006jhm} & 339~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.4$~\times~10^{-7}$ & & & \\\midrule
$e^- f_0(980)$ & Belle~\cite{Belle:2008pdf} & 671~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 6.8$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 9.5$~\times~10^{-10}$ \\\hline
$\mu^-f_0(980)$ & Belle~\cite{Belle:2008pdf} & 671~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 6.4$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 9.1$~\times~10^{-10}$ \\\hline
$e^- \rho^0$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.8$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 3.8$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.6$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 2.0$~\times~10^{-6}$ & & & \\
\noalign{\hrule height 1pt}
\end{tabularx}}
\end{adjustwidth}
\end{table}
\begin{table}[H]\ContinuedFloat
\small
\caption{{\em Cont.}}
\begin{adjustwidth}{-\extralength}{0cm}
\setlength{\cellWidtha}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthb}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthc}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthd}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthe}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthf}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthg}{\fulllength/7-2\tabcolsep-0in}
\scalebox{1}[1]{\begin{tabularx}{\fulllength}{>{\PreserveBackslash\centering}m{\cellWidtha}>{\PreserveBackslash\centering}m{\cellWidthb}>{\PreserveBackslash\centering}m{\cellWidthc}>{\PreserveBackslash\centering}m{\cellWidthd}>{\PreserveBackslash\centering}m{\cellWidthe}>{\PreserveBackslash\centering}m{\cellWidthf}>{\PreserveBackslash\centering}m{\cellWidthg}}
\noalign{\hrule height 1pt}
\multicolumn{1}{c}{\textbf{}} & \multicolumn{3}{c}{\textbf{Observed Limits}} & \multicolumn{3}{c}{\textbf{Expected Limits}} \\ \midrule
\boldmath{$\tau^-\to$} & \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (obs)}& \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (exp)} \\
\midrule
$\mu^-\rho^0$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.2$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 5.5$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.6$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 6.3$~\times~10^{-6}$ & & & \\\midrule
$e^-\omega$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.8$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.0$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2007amy} & 384~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.1$~\times~10^{-7}$ & & & \\ \midrule
$\mu^-\omega$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.7$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.4$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2007amy} & 384~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.0$~\times~10^{-7}$ & & & \\\midrule
$e^-K^{\ast 0}$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.2$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 6.7$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 5.9$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 5.1$~\times~10^{-6}$ & & & \\\midrule
$\mu^-K^{\ast 0}$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 7.2$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 9.3$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.7$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 7.5$~\times~10^{-6}$ & & & \\\midrule
$e^-\bar{K}^{\ast0}$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.4$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 6.2$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.6$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 7.4$~\times~10^{-6}$ & & & \\\midrule
$\mu^-\bar{K}^{\ast0}$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 7.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 8.5$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 7.3$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 7.5$~\times~10^{-6}$ & & & \\\midrule
$e^-\phi$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.1$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 7.4$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.1$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 6.9$~\times~10^{-6}$ & & & \\\hline
$\mu^-\phi$ & Belle~\cite{Belle:2011ogy} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 8.4$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 8.4$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2009wtb} & 451~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.9$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 7.0$~\times~10^{-6}$ & & & \\\midrule
$e^-e^+e^-$ & Belle~\cite{Hayasaka:2010np} & 782~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.7$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.7$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2010axs} & 468~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.9$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 2.9$~\times~10^{-6}$ & & & \\\midrule
$\mu^-e^+e^-$ & Belle~\cite{Hayasaka:2010np} & 782~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.8$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 2.9$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2010axs} & 468~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.2$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.7$~\times~10^{-6}$ & & & \\\hline
$e^-\mu^+\mu^-$ & Belle~\cite{Hayasaka:2010np} & 782~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.7$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.5$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2010axs} & 468~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.2$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.8$~\times~10^{-6}$ & & & \\
\midrule
$\mu^-\mu^+\mu^-$ & Belle~\cite{Hayasaka:2010np} & 782~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.1$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 3.6$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2010axs} & 468~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.3$~\times~10^{-8}$ & & & \\
& LHCb~\cite{LHCb:2014kws} & 3~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.6$~\times~10^{-8}$ & & & \\
& CMS~\cite{CMS:2020kwy} & 33~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 8.0$~\times~10^{-8}$ & & & \\
& ATLAS~\cite{ATLAS:2016jts} & 20~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.8$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.9$~\times~10^{-6}$ & & & \\\midrule
$e^+\mu^-\mu^-$ & Belle~\cite{Hayasaka:2010np} & 782~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.7$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 2.6$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2010axs} & 468~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.6$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.5$~\times~10^{-6}$ & & & \\\midrule
$\mu^+e^-e^-$ & Belle~\cite{Hayasaka:2010np} & 782~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.5$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 2.3$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2010axs} & 468~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.8$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.5$~\times~10^{-6}$ & & & \\
\noalign{\hrule height 1pt}
\end{tabularx}}
\end{adjustwidth}
\end{table}
\begin{table}[H]\ContinuedFloat
\small
\caption{{\em Cont.}}
\begin{adjustwidth}{-\extralength}{0cm}
\setlength{\cellWidtha}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthb}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthc}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthd}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthe}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthf}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthg}{\fulllength/7-2\tabcolsep-0in}
\scalebox{1}[1]{\begin{tabularx}{\fulllength}{>{\PreserveBackslash\centering}m{\cellWidtha}>{\PreserveBackslash\centering}m{\cellWidthb}>{\PreserveBackslash\centering}m{\cellWidthc}>{\PreserveBackslash\centering}m{\cellWidthd}>{\PreserveBackslash\centering}m{\cellWidthe}>{\PreserveBackslash\centering}m{\cellWidthf}>{\PreserveBackslash\centering}m{\cellWidthg}}
\noalign{\hrule height 1pt}
\multicolumn{1}{c}{\textbf{}} & \multicolumn{3}{c}{\textbf{Observed Limits}} & \multicolumn{3}{c}{\textbf{Expected Limits}} \\ \midrule
\boldmath{$\tau^-\to$} & \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (obs)}& \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (exp)} \\
\midrule
$e^-\pi^+\pi^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.3$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 5.8$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.2$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 2.2$~\times~10^{-6}$ & & & \\\midrule
$\mu^-\pi^+\pi^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.1$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 5.6$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.9$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 8.2$~\times~10^{-6}$ & & & \\\midrule
$e^-\pi^+K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.7$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 7.1$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.2$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 6.4$~\times~10^{-6}$ & & & \\\midrule
$\mu^-\pi^+K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 8.6$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.2$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.6$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 7.5$~\times~10^{-6}$ & & & \\
\midrule
$e^-K^+\pi^{-}$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.1$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 7.8$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.7$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 3.8$~\times~10^{-6}$ & & & \\\midrule
$\mu^-K^+\pi^{-}$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.5$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.2$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.2$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 7.4$~\times~10^{-6}$ & & & \\\midrule
$e^-K^+K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.4$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 6.5$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.4$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 6.0$~\times~10^{-6}$ & & & \\\midrule
$\mu^-K^+K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.4$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.1$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.5$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.5$~\times~10^{-5}$ & & & \\\midrule
$e^-K_S^0K_S^0$ & Belle~\cite{Belle:2010rxj} & 671~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 7.1$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 9.7$~\times~10^{-10}$ \\
& CLEO~\cite{CLEO:2002lxe} & 13.9~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 2.2$~\times~10^{-6}$ & & & \\\midrule
$\mu^-K_S^0K_S^0$ & Belle~\cite{Belle:2010rxj} & 671~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 8.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.1$~\times~10^{-9}$ \\
& CLEO~\cite{CLEO:2002lxe} & 13.9~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 3.4$~\times~10^{-6}$ & & & \\\midrule
$e^+\pi^-\pi^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.6$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.7$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 1.9$~\times~10^{-6}$ & & & \\\midrule
$\mu^+\pi^-\pi^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.9$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.5$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 7.0$~\times~10^{-8}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 3.4$~\times~10^{-6}$ & & & \\
\midrule
$e^+\pi^-K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.2$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 7.7$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.8$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 2.1$~\times~10^{-6}$ & & & \\\midrule
$\mu^+\pi^-K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.8$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 1.2$~\times~10^{-9}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.2$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 7.0$~\times~10^{-6}$ & & & \\\midrule
$e^+K^-K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.3$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 5.8$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.5$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 3.8$~\times~10^{-6}$ & & & \\\midrule
$\mu^+K^-K^-$ & Belle~\cite{Belle:2012unr} & 854~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.7$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 9.7$~\times~10^{-10}$ \\
& \babar~\cite{BaBar:2005yvr} & 221~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.8$~\times~10^{-7}$ & & & \\
& CLEO~\cite{CLEO:1997aqy} & 4.79~\ensuremath{\mbox{\,fb}^{-1}}\xspace& 6.0$~\times~10^{-6}$ & & & \\
\noalign{\hrule height 1pt}
\end{tabularx}}
\end{adjustwidth}
\end{table}
\begin{table}[H]\ContinuedFloat
\small
\caption{{\em Cont.}}
\begin{adjustwidth}{-\extralength}{0cm}
\setlength{\cellWidtha}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthb}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthc}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthd}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthe}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthf}{\fulllength/7-2\tabcolsep-0in}
\setlength{\cellWidthg}{\fulllength/7-2\tabcolsep-0in}
\scalebox{1}[1]{\begin{tabularx}{\fulllength}{>{\PreserveBackslash\centering}m{\cellWidtha}>{\PreserveBackslash\centering}m{\cellWidthb}>{\PreserveBackslash\centering}m{\cellWidthc}>{\PreserveBackslash\centering}m{\cellWidthd}>{\PreserveBackslash\centering}m{\cellWidthe}>{\PreserveBackslash\centering}m{\cellWidthf}>{\PreserveBackslash\centering}m{\cellWidthg}}
\noalign{\hrule height 1pt}
\multicolumn{1}{c}{\textbf{}} & \multicolumn{3}{c}{\textbf{Observed Limits}} & \multicolumn{3}{c}{\textbf{Expected Limits}} \\ \midrule
\boldmath{$\tau^-\to$} & \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (obs)}& \textbf{Experiment} & \textbf{Luminosity} & \textbf{UL (exp)} \\ \midrule
$\pi^-\bar{\Lambda}$ & Belle~\cite{Belle:2005exq} & 154~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.4$~\times~10^{-7}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 5.5$~\times~10^{-10}$ \\\hline
$\pi^-\Lambda$ & Belle~\cite{Belle:2005exq} & 154~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 7.2$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 5.4$~\times~10^{-10}$ \\\hline
$\bar{p}^-e^+e^-$ & Belle~\cite{Belle:2020lfn} & 921~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.0$~\times~10^{-10}$ \\\hline
$\bar{p}^-e^+\mu^-$ & Belle~\cite{Belle:2020lfn} & 921~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 2.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.4$~\times~10^{-10}$ \\\hline
$\bar{p}^-\mu^+e^-$ & Belle~\cite{Belle:2020lfn} & 921~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.8$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 4.4$~\times~10^{-10}$ \\\hline
$\bar{p}^-\mu^+\mu^-$ & Belle~\cite{Belle:2020lfn} & 921~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 1.8$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 7.4$~\times~10^{-10}$ \\
& LHCb~\cite{LHCb:2013fsr} & 1~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.3$~\times~10^{-7}$ & & & \\\midrule
$p^+e^-e^-$ & Belle~\cite{Belle:2020lfn} & 921~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 3.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 3.6$~\times~10^{-10}$ \\\midrule
$p^+\mu^-\mu^-$ & Belle~\cite{Belle:2020lfn} & 921~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.0$~\times~10^{-8}$ & Belle II~\cite{Belle-II:2022cgf} & 50~\ensuremath{\mbox{\,ab}^{-1}}\xspace & 8.3$~\times~10^{-10}$ \\
& LHCb~\cite{LHCb:2013fsr} & 1~\ensuremath{\mbox{\,fb}^{-1}}\xspace & 4.4$~\times~10^{-7}$ & & & \\
\noalign{\hrule height 1pt}
\end{tabularx}}
\end{adjustwidth}
\end{table}
A beam polarization upgrade of the SuperKEKB $e^-e^+$ collider
can enhance the sensitivity to LFV in $\tau$ decays
at the Belle II experiment to levels
beyond the ones listed in Table~\ref{tab:TauDecays}
and Figure~\ref{fig:TauLFV}.
The proposed upgrade~\cite{Banerjee:2022kfy}
will result in ${\sim}70\%$ longitudinal polarization
of the high energy electron beam,
which will influence the angular distribution of the $\tau$ decay products
in the SM $\tau$-pair backgrounds.
The characteristic $\tau$ polarization dependence of
the helicity angles of the $\tau$ decay products
with beam polarization can then be used to further suppress
the background in, for~example, $\tau^- \to \mu^- \gamma$ searches,
where one $\tau$ decays to a muon and a photon,
while the other $\tau$ decays to a pion and a neutrino,
the decay channel most sensitive to the polarization of the $\tau$ lepton.
Similar background suppression can also be obtained with the other decay modes,
which vary in their sensitivity to the $\tau$ polarization.
In general, the~maximal discriminating power is obtained by studying the polar angles
in the center-of-mass frame times the charge of the $\tau$ decay.
The “irreducible background” from $\tau^- \to \mu^- \nu \bar{\nu} \gamma$
decays are studied in Figure~\ref{fig:tau-lfv-exp}~\cite{Hitlin:2008gf}.
While the distributions of the backgrounds show marked differences
in the case of beam polarization
with respect to the case of no beam polarization,
the signal distribution modeled by {{uniform}} phase--phase does not change with beam polarization.
{{By removing events where the distribution of the irreducible background shows a rising trend near unity,}}
the background can be reduced significantly,
corresponding to a small loss in signal efficiency.
An optimization study has demonstrated that this would result
in approximately a 10\% improvement in the sensitivity to LFV.
Similar analyses are expected to yield comparable gain in sensitivities
for other decay~modes.
\begin{figure}[H]
\includegraphics[scale=1]{figure6.pdf}
\caption{
Current status of observed upper limits at CLEO, \babar, Belle, ATLAS, CMS,
and LHCb experiments~\cite{HFLAV:2019otj}
and projections of expected upper limits
at the Belle II experiment~\cite{Belle-II:2022cgf}.
}
\label{fig:TauLFV}
\end{figure}
\vspace{-6pt}
\begin{figure}[H]
\includegraphics[width=.75\textwidth]{figure7.png}
\caption{\label{fig:tau-lfv-exp}
Distribution of the cosine of the signal-side muon multiplied by the
muon charge for signal and background events with and without electron
beam polarization in the $\tau^- \to \mu^- \gamma$ search analysis~\cite{Hitlin:2008gf}.
}
\end{figure}
It is worth noting that the uniform phase space model of the signal distribution
is chosen because the underlying theory behind LFV is not known.
Different spin-dependent operators are predicted to give significantly different features
in the Dalitz plane of final state momenta distributions of, for~example,
$\tau^-\to\mu^-\mu^+\mu^-$ decays~\cite{Matsuzaki:2007hh, Dassinger:2007ru}.
One of the most interesting aspects of having the beam polarization
is the possibility to distinguish between these different new physics models
to understand the helicity structure of the couplings producing LFV in $\tau$ decays,
once such decays are~observed.
Belle II limits will probe predictions from several theoretical models,
as listed in Table~\ref{taulfv_predictions}.
Some theoretical expectations from different new physics models and
improvements on experimental limits over the last few decades for
$\tau^- \to \mu^- \gamma$~\cite{Hayes:1981bn,ARGUS:1992ggs,CLEO:1992orj,DELPHI:1995mws,CLEO:1996sqd,CLEO:1999lvl,
Belle:2003ynp,BaBar:2005wms,Belle:2007qih,BaBar:2009hkt,Belle:2021ysv}
and
$\tau^- \to \mmm$~\cite{Hayes:1981bn,CLEO:1989ikc,ARGUS:1992ggs,CLEO:1994bbg,CLEO:1997aqy,
Belle:2004vjj,BaBar:2003pme,BaBar:2007yte,Belle:2007diw,BaBar:2010axs,Hayasaka:2010np,
LHCb:2013fsr,LHCb:2014kws,ATLAS:2016jts,CMS:2020kwy} decays,
along with future prospects at Belle II~\cite{Belle-II:2022cgf}, are shown in Figure~\ref{fig:exp_vs_theory}.
\begin{figure}[H]
\centering
\includegraphics[width=.49\textwidth]{figure8a.jpg}
\includegraphics[width=.49\textwidth]{figure8b.jpg}
\caption{Evolution of experimental
bounds on $\tau^- \to \mu^- \gamma$~\cite{Hayes:1981bn,ARGUS:1992ggs,CLEO:1992orj,DELPHI:1995mws,CLEO:1996sqd,CLEO:1999lvl, Belle:2003ynp,BaBar:2005wms,Belle:2007qih,BaBar:2009hkt,Belle:2021ysv} (left)
and $\tau^- \to \mmm$~\cite{Hayes:1981bn,CLEO:1989ikc,ARGUS:1992ggs,CLEO:1994bbg,CLEO:1997aqy,
Belle:2004vjj,BaBar:2003pme,BaBar:2007yte,Belle:2007diw,BaBar:2010axs,Hayasaka:2010np,
LHCb:2013fsr,LHCb:2014kws,ATLAS:2016jts,CMS:2020kwy}
decays, future prospects at Belle II~\cite{Belle-II:2022cgf}
and some predictions from Table~\ref{taulfv_predictions}.
}
\label{fig:exp_vs_theory}
\end{figure}
\unskip
\section{Conclusions}
LFV in $\tau$ decays are unambiguous signatures of new physics,
and are thus of great experimental and theoretical interest.
Many models from supersymmetric scenarios to leptoquarks
predict LFV in $\tau$ decays at experimentally observable rates,
which will be probed at Belle II.
Searches for LFV in $\tau$ decays can discover new physics at the multi-TeV scale
by identifying the underlying mechanism beyond the SM,
or strongly constrain the flavor structure of TeV-scale extensions
beyond the SM~\cite{Husek:2020fru,Cirigliano:2021img},
as discussed in the context of different experimental efforts in Ref.~\cite{Banerjee:2022xuw}.
The first generation B-Factory experiments, Belle and \babar,
saw an order of magnitude improvement on the upper limit on LFV in $\tau$ decays
from $10^{-6}$ level down to $10^{-8}$ level.
The Belle II experiment will continue to improve the sensitivity
in searches of LFV in $\tau$ decays over the next decade.
The projected sensitivity at Belle II for LFV in $\tau$ decays with 50\ensuremath{\mbox{\,ab}^{-1}}\xspace of data
is at the $10^{-10}$--$10^{-9}$ level,
which constitutes one or two orders of magnitude of
improvement over the previous~experiments.
\vspace{6pt}
\funding{This project was supported by the U.S. Department of Energy under research Grant No.~DE-SC0022350.}
\institutionalreview{Not applicable.}
\informedconsent{Not applicable.}
\dataavailability{Not applicable.}
\acknowledgments{
The author acknowledges fruitful discussions with
Kiyoshi Hayasaka,
Kenji Inami,
John Micheal Roney,
Armine Rostomyan,
Michel Hern\'andez Villanueva,
and many others.
Sources of all figures and tables included for discussions have been appropriately cited,
and reproduced with permissions from the respective collaborations.
}
\conflictsofinterest{The author declares no conflict of~interest.}
\begin{adjustwidth}{-\extralength}{0cm}
\printendnotes[custom]
\reftitle{References}
|
\section{\label{sec:intro}Introduction}
Progress in high-power laser technology in recent decades has made it possible, through the generation of extraordinarily strong electromagnetic fields, to investigate radiation and particle-production processes in the nonlinear quantum regime~\cite{Dittrich:2000zu,piazza.rmp.2012,gonoskov.rmp.2022,fedotov.arxiv.2022,bula.prl.1996,burke.prl.1997,cole.prx.2018,poder.prx.2018,yoon.optica.2021, gonoskov.arxiv.2021}.
In addition, this has opened up new opportunities for the creation of exotic particle and radiation sources~\cite{gonoskov.prl.2014, stark.prl.2016, tamburini.scirep.2017, gonoskov.prx.2017, wallin.pop.2017, lei.prl.2018, vranic.scirep.2018, benedetti.np.2018, jansen.ppcf.2018, magnusson.prl.2019, magnusson.pra.2019}, as well as for studies of electron-positron plasmas \cite{efimenko.scirep.2018, efimenko.pre.2019}, which may help to understand various astrophysical processes~\cite{Turolla:2015mwa,Kaspi:2017fwg,Kim:2021kif}.
The nature of laser-matter (or laser-light) interactions is determined by {several} parameters, including the ratio between the electric field strength $E$ and the Schwinger, or critical, field strength $E_\text{cr} = m^2c^3/\left(e \hbar \right)$ (for $c$ the speed of light, $\hbar$ the reduced Planck constant, $m$ and $e>0$ the electron mass and charge), see also Appendix~\ref{sec:invariants}. When $E / E_\text{cr} \gtrsim 1$ nonlinear quantum effects are expected to be prominent, but the way this is achieved matters. Probing a subcritical field with ultrarelativistic particles, {for example, can `advance' the onset of those quantum effects that depend on the rest-frame (`r.f.') electric field strength via the quantum nonlinearity parameter, $\chi = \gamma E / E_\text{cr} = E_\text{r.f.}/E_\text{cr}$ with $\gamma \gg 1$ being the Lorentz factor.}
Experimental investigation of such effects is well underway~\cite{poder.prx.2018,cole.prx.2018, abramowicz.epjst.2021, yakimenko.prab.2019}. A different class of physical effects is manifested if we can achieve $E / E_\text{cr} \gtrsim 1$ in the absence of massive particles, i.e.~directly in the lab frame.
Such a critical field would be characterised by the invariants $F^2 = (E^2 - c^2 B^2)/E_\text{cr}^2$ or $G^2 = cB.E/E_\text{cr}^2$ satisfying $F,G \gtrsim 1$.
Critical and supercritical ($F,G \gg 1$) fields would modify not only the quantum dynamics of electrons and photons, but also those of heavier particles such as nuclei, {and indeed the QED vacuum itself.}
However, whether it is even possible to attain the needed high field strengths in the lab frame is an open question \cite{bulanov.jetp.2006,fedotov.prl.2010,bulanov.prl.2010s, gonoskov.prl.2013, baumann.scirep.2019, vincenti.prl.2019}. This is because such fields would be expected to trigger an electron-positron pair cascade, forming a dense pair plasma that would screen or absorb {the laser radiation being focused, preventing the} further increase of the field strength~\cite{fedotov.prl.2010,bulanov.prl.2010s, nerush.prl.2011}. Avoiding the triggering of such a cascade will be essential for maximizing the reachable field strength~\cite{gonoskov.prl.2013}.
In this paper we investigate the possibility to generate supercritical fields by a combination of three essential ideas: advanced focusing, plasma-based conversion of optical or near-IR light to XUV frequencies, and coherent combination of multiple laser pulses (see fig.~\ref{fig:concept}).
The conversion to higher frequency has been discussed as a means of reducing the focal volume which increases field strength at fixed power \cite{landecker.pr.1952, bulanov.prl.2003, naumova.prl.2004, gordienko.prl.2005, gonoskov.pre.2011, baumann.scirep.2019, vincenti.prl.2019} (a more detailed discussion can be found in~\cite{bulanov.ufn.2013}). Moreover, electromagnetic processes in high strength fields demonstrate a strong dependence on the field wavelength~\cite{zhang.pop.2020}.
In combination with $4\pi$ focusing, which itself reduces the focal volume, this maximises the electric or magnetic field while suppressing pair cascades, see Appendix~\ref{sec:paircascades}.
{Our goal here is to provide a far-future outlook on the field strengths that could be attained in `best case' scenarios which combine currently known concepts and approaches.} We demonstrate numerically that, given advanced focusing, the physics of laser-plasma interactions itself provides the possibility to reach $10E_\text{cr}$ already out at a laser power of \SI{20}{\peta\watt}. This should certainly be seen as an idealistic (theoretical) reference point, as we omit discussion of a variety of feasibility questions, but it does indicate that further consideration and technological efforts are warranted, with the hope that $E_\text{cr}$ could be attained at upcoming 10-PW-class laser facilities. This would open up new and exciting opportunities for scientific discoveries, in a regime previously considered to be unattainable. Giving a complete overview of physical applications of strong fields is of course not possible, and we will restrict ourselves {here} to examples from electron and nuclear physics. This paper is organised as follows. Section~\ref{sec:setups} concerns the combination of the three concepts mentioned above: advanced focusing~\ref{sec:advanced_focusing}, frequency upshifting through plasma-based conversion~\ref{sec:plasma_conversion}, and coherent combination of multiple laser pulses~\ref{sec:xuv_focusing}. Section~\ref{sec:processes} discusses the impact of such a supercritical field on nuclear and electron dynamics in ~\ref{sec:nuclear} and~\ref{sec:electron} respectively.
\section{\label{sec:setups}Setups}
\begin{figure}[t!]
\includegraphics[width=\textwidth]{Concept_6lasers.png}
\caption{The main principle behind maximizing field strength starting from laser sources with optical frequencies.}
\label{fig:concept}
\end{figure}
In this section we provide an order of magnitude estimate for the field strength hypothetically attainable with future laser systems, with the help of the known theoretical ideas
outlined above. We start by considering the concepts of $4\pi$ focusing and frequency up-shifting separately, before discussing the combination of the two.
\subsection{\label{sec:advanced_focusing}Advanced focusing}
The maximal attainable field strength for a given power of focused radiation is limited by the so-called dipole wave \cite{bassett.oa.1986} that can also be extended to time-limited solution known as the dipole pulse \cite{gonoskov.pra.2012}. The dipole wave can be seen as time-reversed emission of a dipole antenna and thus can be approximated by several focused beams or by focusing intensity-shaped radially-polarized beam with a parabolic mirror \cite{gonoskov.pra.2012, jeong.oe.2020}. Let us start from considering the benefits of using tight focusing of the laser radiation, characterized by small values of $f$-number $\fN$ or, equivalently, by large values of the divergence angle $\theta = \arctan\!\left(\fN^{-1}/2\right)$, where we for simplicity use the expression that implies $\theta < \pi/2$. To numerically ascertain the potential gain of using tight focusing, we set the initial electromagnetic field to be a two-cycle optical pulse propagating from a spherical surface of radius $r_0 = 16\lambda$, which can be considered large compared to the radiation wavelength $\lambda$ (the far-field zone):
\begin{align}
& \vec{E} = \frac{\vec{r} \times \left[\vec{\hat z} \times \vec{r}\right]}{\left|\vec{r} \times \left[\vec{\hat z} \times \vec{r}\right] \right|} S_r\left(r\right) S_{\alpha}\left(\alpha\right),\\
& \vec{B} = \frac{1}{c}\frac{\vec{\hat z} \times \vec{r}}{\left| \vec{\hat z} \times \vec{r} \right|} S_r\left(r + ct\right) S_{\alpha}\left(\alpha\right),
\end{align}
where the radial $S_r\left(r\right)$ and angular $S_{\alpha}\left(\alpha\right)$ shape functions are defined by:
\begin{align}
& {S_r\left(r\right)} = \sin\left( 2 \pi (r - r_0) /\lambda \right) \left\{
\begin{array}{ll}
\cos^2\left(\frac{\pi}{2}(r - r_0)/\lambda\right), & \left|r - r_0\right| \leq \lambda, \\
0, & \left|r - r_0\right| > \lambda,
\end{array}
\right.\\
& {S_\alpha\left(\alpha\right)} = \left\{
\begin{array}{ll}
1, & \alpha \leq \theta - \theta_s/2, \\
\sin^2\left(\frac{\pi}{2}\left(\alpha - \theta\right)/\theta_s\right), & \theta - \theta_s/2 < \alpha \leq \theta + \theta_s/2, \\
0, & \alpha > \theta + \theta_s/2,
\end{array}
\right.\\
& \alpha = \arctan\left(\sqrt{z^2 + y^2}/\left|x\right|\right).
\end{align}
In our setup the smoothing angle $\theta_s = 0.3$ eliminates sharp edges of the concave pulse within our model. We advance this field to the vicinity of the focal point using a spectral solver of Maxwell’s equations within the open-source package hi-$\chi$ \cite{hichi}. To reduce the amount of needed computational resources we also employ the module of contracting a spherical window that maps the concave region of the pulse to a thin layer of space with periodic boundary conditions~\cite{panova.as.2021}.
{The radiation intensity at focus is proportional to the power $\Power$ and inversely proportional to the focal spot area, which in turn scales as $\lambda^2$ with $\lambda$ being the radiation wavelength. It is thus possible to express the peak field strength at focus for arbitrary power $\Power$ and $\lambda$:}
\begin{equation}
\frac{E}{E_\text{cr}} = \frac{\delta}{4.1 \times 10^5}\left(\frac{\lambda}{1\:\mu\text{m}}\right)\left(\frac{\Power}{\SI{1}{\peta\watt}}\right)^{1/2} \;,
\label{delta_e_cr}
\end{equation}
{where a wavelength-agnostic, dimensionless parameter $\delta$ solely characterize the efficiency of focusing. Note that we define it so that $\delta \sqrt{\Power/\left(1\:\:\text{PW}\right)}$ gives field amplitude $E$ in relativistic units, i.e. in units of $mc\omega/e$, where $\omega$ is radiation frequency.} According to our simulations the focusing with \f{2} and \f{1} provides $\delta \approx 170$ and $\delta \approx 230$, respectively.
A significant improvement can be achieved by splitting the power into 6 pulses and focusing them with $\fN = 1$ ($2 \theta \approx 0.93 < 2 \pi/6$) to the same point symmetrically from different directions in the $x$-$y$ plane, so that the polarization vector for each pulse is orientated along the $z$-axis. For each pulse the power is then reduced by a factor of 6, but the strength of the field from each pulse is increased by a factor of 6 due to coherent summation of the field. As a result we have an increase by factor $\sqrt{6}$: $\delta\left(6 \times \f{1.0}\right) \approx 560$, which is relatively close to the theoretical maximum $\delta_\text{max} \approx 780$ provided by the dipole wave \cite{bassett.oa.1986, gonoskov.pra.2012}. We will use this 6 beam configuration as the main reference for future setups, whereas the configurations with larger number of beams can better sample the dipole wave and bring the value of $\delta$ even closer to $\delta_\text{max}$.
The maximal field strength is achieved either for electric or magnetic field component (pointing along the $z$-axis), whereas the other field component is close to zero in the center. The maximization of electric field with so-called electric dipole wave provides a strong, oscillating electric field that is especially interesting for enhancing production of electrons and positrons, as well as for trapping them by anomalous radiative trapping \cite{gonoskov.prl.2014} that in combination provides unique condition for the creation of radiation sources \cite{gonoskov.prx.2017} and extreme plasma states \cite{efimenko.scirep.2018, efimenko.pre.2019}. The maximization of magnetic field by the magnetic dipole wave can also be of interest for initiating extreme plasma dynamics \cite{bashinov.pre.2022} as well as for reaching strong fields with suppressed electromagnetic cascades in the center. {The interaction of an \emph{optical} dipole wave with a high energy electron beam leads to the generation of multi-GeV photon sources and can be used as a platform for the study of electromagnetic cascades, of both shower and avalanche type~\cite{magnusson.prl.2019,magnusson.pra.2019}.} Finally, a symmetric mixture of electric and magnetic dipole waves provides the optimal setup for attaining highest possible $\chi$ value for a given external beam of high-energy electrons \cite{olofsson.arxiv.2022}. Here we proceed our analysis for electric dipole wave.
\subsection{\label{sec:plasma_conversion}Plasma converter}
The idea and particular concepts for field intensification through plasma-based high-order harmonic generation and focusing have been being discussed by several research groups since the beginning of the 2000s. One possibility is to use the Doppler frequency up-shifting during the reflection of laser radiation from so-called relativistic flying mirrors formed either by the cusp preceding plasma wave breaking \cite{bulanov.prl.2003, martins.ps.2004, matlis.natphys.2006} or by the ejection of electrons from thin plasma layers \cite{kulagin.prl.2007, meyer-ter-vehn.epjd.2009, habs.apb.2008, bulanov.pla.2010, esirkepov.prl.2009}. In both cases a counter-propagating laser pulse is used to produce the flying mirror that can be shaped to focus the reflected radiation. Another possibility is to use highly-nonlinear reflection of laser radiation from dense plasma naturally formed by ionization of solid targets \cite{bulanov.pop.1994, lichters.pop.1996, vonderlinde.apb.1996}. The early discussions and models also appealed to the Doppler frequency up-shifting, but now during the reflection from oscillating effective boundary \cite{gordienko.prl.2004, baeva.pre.2006} that can also be shaped for harmonic focusing by tailoring the pulse intensity shape \cite{naumova.prl.2004, dromey.nphys.2009}. It was later recognized that the conversion can be more generally seen as coherent synchrotron emission (CSE) of electrons from a self-generated peripheral layer of electrons \cite{anderbrugge.pop.2010}, while the layer's spring-like dynamics and sought-after emission can be described by a set of differential equations forming so-called relativistic electronic spring (RES) model \cite{gonoskov.pre.2011, gonoskov.pop.2018, blanco.pop.2018}. Further studies \cite{blackburn.pra.2018} showed that optimal conversion achievable with an incidence angle of $50^\circ - 62^\circ$ and the density ramps achievable via tailored pulse contrast \cite{rodel.prl.2012}. Latest numerical studies exploiting plasma denting in combination with oblique incidence indicate the possibility of significant field intensification \cite{baumann.scirep.2019, vincenti.prl.2019}. Some of the reported numerical results are summarized in Table~\ref{tab:reported_simulations}. As a way to estimate future prospects we consider the conversion described in \cite{anderbrugge.pop.2010, gonoskov.pre.2011}.
We performed a number of simulations using 1D version of ELMIS PIC code \cite{gonoskov.phd.2013} (the oblique incidence is transformed to normal in a moving reference frame \cite{bourdier.pf.1983}). We assumed a single-cycle laser pulse ($\lambda = \SI{0.81}{\micro\m}$) interacting with a steep-front plasma surface with immobile ions.
{Many factors, including e.g.~the motion of ions, plasma spreading due to limited contrast, and pulse shape, can significantly affect both the increase of the amplitude and the optimal conditions for achieving it. However, the} physics of this process has been shown to be sufficiently robust to justify the considerations here as a good starting point for further studies~\cite{gonoskov.pop.2018, blackburn.pra.2018, bhadoria.pop.2022}.
The amplitude increase becomes larger with the increase of incident wave amplitude $a_{in}$, which we express in relativistic units \cite{bashinov.epjst.2014}. We consider two cases $a_{in} \approx 70$ ($I = \SI{e22}{\Wcm}$) and $a_{in} \approx 220$ ($I = \SI{e23}{\Wcm}$). For each case we fine tune the incidence angle $\alpha$ and the plasma density $n$ expressed in units of plasma critical density.
For $I = \SI{e22}{\Wcm}$ we find that the maximal amplitude increase of $8.4$ is achieved for $\alpha = 61.43^\circ$ and $n = 0.4125 a_{in}$, whereas for $I = \SI{e23}{\Wcm}$ the maximal amplitude increase of $16.1$ is achieved for the same incidence angle but for $n = 0.397 a_{in}$. For the latter, the resulted field distribution is show in fig.~\ref{peak}(b). The length of the generated pulses in these cases is less than \SI{1}{\nano\meter}, which corresponds to the XUV range.
\begin{table}[]
\centering
\begin{tabular}{c||c||c|c|c||c|c|c}
{\begin{tabular}[c]{@{}c@{}}Publication,\\geometry\end{tabular}} & laser& \multicolumn{3}{c||}{\begin{tabular}[c]{@{}c@{}}conversion parameters\\\end{tabular}} & \multicolumn{3}{c}{yield after focusing} \\
\cline{2-8}
& \begin{tabular}[c]{@{}c@{}}peak \\power, PW\end{tabular} & \begin{tabular}[c]{@{}c@{}}incident \\intensity, W/cm$^2$\end{tabular} & \begin{tabular}[c]{@{}c@{}}working plasma \\density, cm$^{-3}$\end{tabular} & \begin{tabular}[c]{@{}c@{}}incidence\\angle\end{tabular} & \begin{tabular}[c]{@{}c@{}}duration,\\as\end{tabular} & \begin{tabular}[c]{@{}c@{}}intensification\\factor\end{tabular} & \begin{tabular}[c]{@{}c@{}}peak intensity,\\W/cm$^2$\end{tabular} \\
\hhline{=::=::===::===}
\begin{tabular}[c]{@{}c@{}}Naumova (2004) \cite{naumova.prl.2004},\\plasma denting\end{tabular} & -- & $2 \times 10^{19}$ & $3 \times 10^{21}$ & $0^\circ$ & 200 & 2.5 & $5 \times 10^{19}$ \\
\hline
\begin{tabular}[c]{@{}c@{}}Gordienko (2005) \cite{gordienko.prl.2005},\\spherical converter\end{tabular} & $\sim 5 \times10^{-3}$ & $1.2\times 10^{19}$ & $5.5 \times 10^{21}$ & $0^\circ$ & $ \lesssim 40$ & $\sim 400$ & $\sim 6 \times 10^{21}$ \\
\hline
\begin{tabular}[c]{@{}c@{}}Gonoskov (2011) \cite{gonoskov.pre.2011},\\groove-shaped converter\end{tabular} & 10 & $5 \times 10^{22}$ & $0.85 \times 10^{23}$ & $62^\circ$ & $\sim 10$ & 4000 & $2 \times 10^{26}$ \\
\hline
\begin{tabular}[c]{@{}c@{}}Baumann (2019) \cite{baumann.scirep.2019},\\plasma denting\end{tabular} & 35 & $1.7 \times 10^{23}$ & $1.7 \times 10^{23}$ & $30^\circ$ & 150 & 16 & $2.7 \times 10^{24}$ \\
\hline
\begin{tabular}[c]{@{}c@{}}Vincenti (2019) \cite{vincenti.prl.2019},\\plasma denting\end{tabular} & 3 & $10^{22}$ & -- & $45^\circ$ & 100 & 1000 & $10^{25}$
\end{tabular}
\caption{Some of the reported numerical results on focusing plasma-generated XUV pulses.}
\label{tab:reported_simulations}
\end{table}
\subsection{\label{sec:xuv_focusing}Focusing of XUV pulses}
We now continue our analysis by considering the possibility of focusing the XUV pulses generated at the curved plasma surfaces of the 6 focusing mirrors with $\fN = 1$. We assume that the laser radiation is split into 6 beams, pre-focused and delivered so that the optimal conditions for the RES-converters are achieved at the plasma surfaces and the generated XUV pulses become focused at the central point. We assume that the conversion happens at the distance of \SI{6}{\micro\m} from the centre. We consider two cases: the total power $\Power$ is \SI{20}{\peta\watt} and \SI{200}{\peta\watt}, which results in the intensity of \SI{e22}{\Wcm} and \SI{e23}{\Wcm} at the plasma surfaces, respectively. A rough estimate for the peak field strength achievable in this configuration suggests that for $\Power = \SI{20}{\peta\watt}$ (\SI{200}{\peta\watt}) we can reach $a_{out} \sim 2500$ ($8000$) given in relativistic units for the wavelength $\lambda \sim \SI{1}{\nano\m}$, which is well above the Schwinger field strength in both cases. However, this estimate is not sufficient because different spectral fractions are focused to different diffraction-limited volumes. That is why we need to perform numerical calculation to perform estimations for these cases.
In order to resolve the singular XUV peak we use a sequence of adaptive sub-grids that are arranged in the following way. Firstly, we surround the XUV peak with a frame and deduce there the field multiplying by a mask function that smoothly goes from 1 to 0 and the ends of the frame. In such a way we cut out the XUV pulse and the remaining field with narrower spectral content can be sampled with the first grid. We then take the deduced field within the frame and repeat the procedure, introducing another subframe in a closer vicinity of the XUV peak and sampling the remaining field with another more thinner subgrid. We perform this procedure 7 times to reach a sufficient resolution, which in our case corresponds to the space step of \SI{0.064}{\nano\meter}. Every deduced field is advanced first analytically (as a spherical wave) to the distance of 4 frame lengths and then numerically using the spectral solver on the grid 128 $\times$ 512$^2$.
The result of our numerical calculation for $\Power = \SI{200}{\peta\watt}$ is shown in fig.~\ref{peak}. The peak field of 130~$E_\text{cr}$ is achieved in the centre within a volume of about few nanometers in size. The following fit can be used for estimates and calculations:
\begin{equation}\label{eq:XUV_shape}
\frac{E(r, t)}{E_\text{cr}} \approx \mathcal{A} \left(\left| \frac{r + ct}{R}\right|^{3/2} + 1\right)^{-1} \left(\left| \frac{ct}{D}\right| + 1 \right)^{-1},
\end{equation}
where $\mathcal{A} = 130$, $R = \SI{0.2}{\nano\meter}$ and $D = \SI{0.3}{\nano\meter}$ in this case (see the solid black curves in fig.~\ref{peak}~(c)). A similar result is obtained for the case of $\Power = \SI{20}{\peta\watt}$, for which we got best fit for $\mathcal{A} = 10$, $R = \SI{0.5}{\nano\meter}$ and $D = \SI{0.5}{\nano\meter}$.
The threshold for the cascade can be estimated as the equality of the volume size (distance to the centre) to the mean scale length of pair production. This estimate is shown in fig.~\ref{peak}~(c) with dashed black line and indicates that the region where the field reaches $E_s$ is too small for the occurrence of the cascade based on the Breit-Wheeler process.
We conclude this section by showing schematically the potential of reaching strong fields with different strategies based on a given value of total laser power of a laser facility (see fig.~\ref{map}). One can see that using tight focusing or, better, multiple colliding laser pulses (MCLP)\cite{bulanov.prl.2010} provides a substantial increase of the peak field, which is, however, well below the Schwinger field even in case of \SI{1}{\exa\watt} total power. The plasma converter can give a significant increase once the intensity of \SI{e22}{\Wcm} is reached, which can be provided by $\f{1.0}$ focusing already with the total laser power of about \SI{100}{TW}. The conversion at \SI{e23}{\Wcm} provides even larger boost. In both cases tight focusing of the generated XUV pulses can provide a significant increase of field strength beyond $E_\text{cr}$.
Certainly, this analysis is performed under the assumption of best-case scenario and the implementation of such a concept requires many technological advances. Among them, driving plasma conversion and reaching spatial-temporal synchronization of the generated XUV pulses appear to the the central difficulties. However, from our results we can draw a conclusion that achieving the needed spatio-temporal control in the domain of nanometer-attosecond could provide a pathway towards reaching the Schwinger field strength using the outlined concept based on high-intensity laser facilities.
\begin{figure*}
\includegraphics[width=\linewidth]{6xXUV.pdf}
\caption{%
The numerical result for the dipole focusing of XUV pulses (a). The total laser power of \SI{200}{\peta\watt} is split into 6 beams and each is focused to \SI{e23}{\Wcm} at \SI{7}{\micro\meter} from the focus, where the RES-converters provide amplitude boost by factor 15 and frequency upshift by factor $\sim 10^4$ (b). The conversion is followed by the MCLP (e-dipole) focusing using 6 beams at $\f{1.0}$. The dependency of the field strength on the $x$-coordinate (green curve), $z$-coordinate (blue curve) and time (red curve) is shown in panel (c) together with the fit (black solid curves) and the threshold for cascaded pair-generation (dashed black line).
}
\label{peak}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.5\linewidth]{E_of_P.pdf}
\caption{The prospects of reaching high field strength using tight focusing, multiple laser colliding pulses, the plasma conversion and their combination on the map of the attainable field strength and total power of laser facility. Two outlined options correspond to the use of the plasma conversion at $10^{22}$ and \SI{e23}{\Wcm}, respectively. The labels show the results of simulations by Gonoskov et al. \cite{gonoskov.pre.2011} (1), by Baumann et al. \cite{baumann.scirep.2019} (2) and by Vincenti \cite{vincenti.prl.2019} (3).}
\label{map}
\end{figure*}
{Therefore we estimate that delivering \SI{10}{\,[\si{\giga\eV}]} electrons to the strong-field region of the outlined setup would result in a $\chi$ of order $10^6$, for the case of $P = 20$~PW (see estimates in Appendix \ref{sec:paircascades}). We will investigate such possibilities in future work.}
\section{\label{sec:processes}Physical processes in supercritical fields}
Critical and supercritical fields open up the possibility to perform experiments in regimes that traditionally have not been available to light sources.
For the purpose of illustration, we briefly discuss a number of possible studies that could be performed using the extreme-field source we have outlined.
\subsection{\label{sec:nuclear}Nuclear dynamics}
Electric fields of the strength discussed in \cref{sec:setups} are sufficient to strip atoms;
the field strength necessary for barrier-suppression ionization of the deepest lying electron, $E_\text{BSI} \simeq (Z \alpha)^3 E_\text{cr} / 16$.
The bare nucleus can then be accelerated to relativistic velocity, in a single wave period, if the electric-field amplitude $E > E_\text{rel}$, where
\begin{equation}
\frac{E_\text{rel}}{E_\text{cr}} =
\frac{A m_p \omega}{e Z} =
3.8 \times 10^{-3} \, \frac{A}{Z \lambda [\mu\mathrm{m}]},
\end{equation}
and $Z$ and $A$ are the nucleus' atomic and mass numbers.
Thus a source of near-critical field could accelerate heavy nuclei from rest to normalised momentum $p / M = 225 (Z/A) (E/E_\text{cr}) \lambda [\mu\mathrm{m}] \gg 1$.
Stronger electric fields affect even the internal dynamics of the nucleus, by modifying the Coulomb barrier through which daughter particles tunnel.
For example, the characteristic electric field required to modify the $\alpha$-decay rate of an unstable nucleus, $E_\alpha$, can be estimated as~\cite{palffy.prl.2020}
\begin{equation}
\frac{E_\alpha}{E_\text{cr}} =
\frac{2 \sqrt{2} Q_\alpha^{5/2}}{3\pi \alpha^2 Z^2 Z_\text{eff} m_e^2 m_r^{1/2}}
\simeq 300\,\frac{Q^{5/2} [\text{MeV}]}{Z^2 Z_\text{eff}}
\end{equation}
where $Q$ is the energy of the $\alpha$ particle, $Z_\text{eff} = (2 A - 4 Z) / (A + 4)$, $Z$ and $A$ are the proton and mass numbers of the daughter nucleus, and $m_r$ is the reduced mass of the $\alpha$--daughter-nucleus system.
For polonium-$212$, which has a half life of $0.3$~ms, $Q \simeq 9.0$~MeV and $E_\alpha / E_\text{cr} \simeq 30$.
The correction to the decay rate $C = \exp[2 E(t) \cos\theta/ E_\alpha]$, where $\theta$ is the angle between the electric-field vector and the $\alpha$-emission direction.
Averaging over all $\theta$, we obtain $\avg{C}_\theta = \sinh[2 E(t) / E_\alpha] / [2 E(t) / E_\alpha]$.
Further averaged over a single cycle, with $E(t) = E_0 \sin \omega t$, we find that the modification to the decay rate is $\avg{C}_{\theta,t} \simeq 1.4$ for $E_0 = 30 E_\text{cr}$ and as much as $\avg{C}_{\theta,t} \simeq 21$ for $E_0 = 100 E_\text{cr}$.
We note that by exceeding $E_\alpha$ we enter a regime where the effect of the external field is no longer a small correction.
The same logic can be applied to $\beta$ decay, where
the characteristic electric field required to modify the decay rate is~\cite{akhmedov.arxiv.2011}:
\begin{equation}
\frac{E_\beta}{E_\text{cr}} = \left( \frac{2 Q_\beta} { m_e} \right)^{3/2}
\end{equation}
where $Q_\beta$ is the energy release associated with the decay.
In the case of tritium $Q_\beta = 18.6$~keV and $E_\beta = 0.02 E_\text{cr}$.
As this is a non-relativistic beta decay, $Q_\beta / m \ll 1$, the modification to the decay rate is $C \simeq (E_0 / E_\beta)^{7/3}$, for an applied electric field $E_0$ which satisfies $E_0 / E_\beta \gg 1$~\cite{akhmedov.arxiv.2011}:
at $E_0 = 0.1 E_\text{cr}$, $C \simeq 50$.
\subsection{\label{sec:electron}Electron dynamics}
A supercritical field structure of this type is a platform for investigating nonlinear quantum electrodynamics in a completely unexplored regime, either by probing it with externally injected electrons or by exploiting the nonlinear dynamics of virtual particles from the quantum vacuum.
Based on an analysis of quantum loop corrections to physical processes in constant, crossed fields, it has been conjectured that the relevant expansion parameter is not the fine structure constant $\alpha$, small, but rather $\alpha\chi^{2/3}$, which can become large in extremely strong fields~\cite{Ritus:1970radiative,Narozhnyi:1980dc}.
Hence the usual small expansion parameter of perturbative QED becomes, in principle, a large parameter for $\chi > 1600$.
In a collision between an electron beam of energy $\mathcal{E}$ and a supercritical electric field of magnitude $E$, we have that
\begin{equation}
\alpha \chi^{2/3} = 5.3 \, \mathcal{E}^{2/3} [10~\text{GeV}] \left(\frac{E}{E_\text{cr}}\right)^{2/3} \;.
\end{equation}
Higher-order corrections, normally thought of as \emph{suppressed} by powers of $\alpha$, are implied by the conjecture to become larger and larger as the order increases.
The technical implication is that the perturbative expansion of QED breaks down and needs (somehow) to be resummed~\cite{Heinzl:2021mji};
the physical implication is that QED enters a new `fully nonperturbative' regime in which it behaves as a strongly coupled theory~\cite{Fedotov:2016afw,fedotov.arxiv.2022}.
It is essential that large $\chi$ is reached not by simply increasing the particle energy $\mathcal{E}$ at low field strength, as the the Ritus-Narozhny conjecture only applies in the high-intensity (LCFA) regime where $a^3 / \chi \gg 1$~\cite{Podszus:2018hnz,Ilderton:2019kqp}.
Furthermore, the mitigation of radiative energy losses requires the field duration to be kept as short as possible: see alternative scenarios in~\cite{blackburn.njp.2019,baumann.scirep.2019, yakimenko.prl.2019}.
Nonlinear quantum dynamics are evident for pure EM fields as well, driven by virtual electron loops that modify the classical linearity of Maxwell's equations.
The nonlinear behaviour of a pure magnetic field of strength $B$, is controlled by the Heisenberg-Euler interaction Lagrangian (see. e.g., \cite[\SS 7]{fedotov.arxiv.2022}).
At one-loop order, $\mathcal{L} = m^4 (B / B_\text{cr})^4 / (360 \pi^2)$ for $B \ll B_\text{cr}$ and $\mathcal{L} = m^4 (B / B_\text{cr})^2 \ln(B / B_\text{cr}) / (24 \pi^2)$ for $B \gg B_\text{cr}$.
For supercritical magnetic fields, higher-order corrections grow logarithmically, with~\cite{karbstein.prl.2019}
\begin{equation}
\frac{{\cal L}^{n\text{-loop}}}{{\cal L}^{1\text{-loop}}}\sim \left[\frac{\alpha}{\pi}\ln\left( \frac{B}{B_\text{cr}}\right)\right]^{l-1}\,.
\end{equation}
Though this growth is slower than the power-law behaviour of higher-order corrections at ultralarge quantum parameter $\chi$, as predicted (above) in the Ritus-Narozhny conjecture, resummation is still required.
Investigating this non-perturbative, non-linear regime of electrodynamics motivates the creation of ultrastrong EM fields that are not probed by ultrarelativistic external particles.
\section{\label{sec:summary}Summary}
We have outlined how optimal configurations of laser systems and/or secondary sources could give us the opportunity to approach, or even exceed, the critical field of quantum electrodynamics. The configurations presented certainly constitutes immense engineering challenges, such as for timing and pointing stability, material engineering, vacuum properties etc, but could also be extremely rewarding as a scientific tool, if realised. {These feasibility questions should be addressed in future work. Our results nevertheless indicate that the presented concepts are promising and warrant further analysis.} Reaching such critical fields could give an opportunity to probe some of the most extreme environments in the universe, and investigate the behaviour of electrons, nuclei and the quantum vacuum under such conditions. We have given several examples of the use of such new photon sources for probing physical laws, ranging from electron and nuclear physics to probing the quantum vacuum.
\begin{acknowledgments}
This research was supported by
the Swedish Research Council Grants nos. 2016-03329 and 2020-06768 (T.G.B. and M.M.), and 2017-05148 (A.G.), as well as the U.S. Department of Energy Office of Science Offices of High Energy Physics and Fusion Energy Sciences (through LaserNetUS), under Contract No. DE-AC02-05CH11231 (S.S.B.).
Simulations were performed on resources provided by the Swedish National Infrastructure
for Computing (SNIC).
\end{acknowledgments}
|
\section{Introduction}
This paper discusses how reputational concerns impact the type of advice experts provide.
I frame advice as a ``rule'' or ``model'' that maps conditions to actions.
It makes a prediction of the \emph{ex post} ``correct'' action based on \emph{ex ante} unobserved conditions.
For example, it tells a patient whether to pursue medical treatment if they develop specific symptoms.
I assume that
\begin{enumerate}
\item[(i)]
the \emph{ex post} correct action cannot be predicted perfectly, and
\item[(ii)]
the advisee does not know what conditions are relevant if the expert does not include them in their advice.
\end{enumerate}
These assumptions reflect the situation faced by public health officials and policy-makers early in the COVID-19 pandemic.
Assumption~(i) reflects how the mapping between medical conditions (e.g., being immunocompromised) and optimal policies (e.g., getting vaccinated) was uncertain because the relevant science was unsettled.
It implies that advice can be ``wrong'' (i.e., recommend the \emph{ex post} incorrect action) even if it includes all relevant conditions.
Assumption~(ii) reflects how the novelty of the relevant science prevented advisees from comparing experts' advice with public knowledge.
It implies that \emph{if} advice is wrong then experts cannot be punished for excluding relevant conditions.
Intuitively, advisees cannot observe bad outcomes and attribute them to causes experts ``should have known about'' because advisees do not know the set of possible causes.%
\footnote{
This became less true as the pandemic progressed and people started sharing (scientific and anecdotal) information about vaccination outcomes.
}
Under assumptions~(i) and~(ii), experts make strategic choices about the type of advice they provide.
Advice may be unconditional (e.g., ``do $x$'') or have conditions (e.g., ``do $x$ if $y$ happens and $z$ does not'').
Including conditions tailors the advice and makes experts appear more confident.
However, this confidence may hurt experts' reputations if their advice turns out to be wrong.
Consequently, experts may exclude relevant conditions---that is, ``simplify'' their advice---to hide their confidence and preserve their reputation.
I analyze this tension between providing better advice and managing reputational risk.
Section~\ref{sec:model-setup} embeds this tension in a simple model.
An expert (he) tells an advisee (she) whether to take an action that may be good or bad.
Both parties want the advisee to take the action if and only if it is good.
The expert may provide a binary condition under which to take the action.
This condition predicts whether the action is good if the expert is competent but not if he is incompetent.
Neither the advisee nor the expert know if he is competent.
He chooses between a ``simple rule,'' which says to take the action, and a ``complex rule,'' which says to take the action under the provided condition.
If he chooses the complex rule then the advisee learns about the expert's competence by whether the rule predicts the correct action.
If he chooses the simple rule then she learns nothing about the expert's competence.
The expert trades off the complex rule's superior predictive power with its ability to reveal information about his competence.
Section~\ref{sec:model-analysis} analyzes the expert's trade-off.
Choosing the complex rule over the simple rule induces a mean-preserving spread in the posterior belief about the expert's competence.
This spread makes him better off if his payoff is convex in the posterior and worse off if it is concave in the posterior.
In the convex case, the expert always chooses the complex rule because it is more accurate and induces an expected reputational benefit.
In the concave case, he chooses the complex rule if and only if its accuracy benefit exceeds its reputational cost.
Increasing the condition's predictive power makes the complex rule more accurate and more diagnostic of the expert's competence.
If his payoff is concave in the posterior then this increase in predictive power makes the expert more willing to choose the complex rule if and only if his reputational concerns are sufficiently weak.
Section~\ref{sec:wage-example} considers a setting in which the expert earns a wage if and only if the posterior belief about his competence exceeds a threshold.
If this threshold equals the prior belief, then the expert chooses the complex rule if and only if the wage is sufficiently low.
If the advisee wants to solicit the complex rule but cannot change the wage, then she can
(i)~lower the threshold and guarantee earning the wage, or
(ii)~raise the threshold and guarantee \emph{not} earning the wage.
Both~(i) and~(ii) remove the reputational risk that prevents the expert from choosing the complex rule.
Section~\ref{sec:extensions} considers some extensions of my analysis and Section~\ref{sec:literature} compares it to related literature.
Section~\ref{sec:conclusion} concludes.
Appendix~\ref{sec:proofs} contains proofs of my mathematical claims.
\section{Model}
\label{sec:model-setup}
\paragraph{Environment}
There are two players: an expert (he) and an advisee (she).
The expert tells the advisee whether to take an action.
The action is either good ($A=1$) or bad ($A=0$).
The expert and advisee know $\Pr(A=1)=0.5$.
The expert may provide a condition under which to take the action.
This condition predicts $A$ if the expert is competent ($\theta=1$) but not if he is incompetent ($\theta=0$).
The condition $X_\theta\in\{0,1\}$ has prevalence $\Pr(X_\theta=1)=x\in(0,1)$ and covariance $\Cov(A,X_\theta)=\theta\sigma$ with $\sigma\in(0,0.25]$.
The expert wants the advisee to take the action if and only if it is good.
He offers one of two rules $r\in\{S,C\}$:
\begin{enumerate}
\item
``take the action'' (the ``simple'' rule $r=S$);
\item
``take the action if and only if $X_\theta=1$'' (the ``complex'' rule $r=C$).
\end{enumerate}
Thus, he provides the condition if and only if he chooses the complex rule.
The advisee can verify the condition's prevalence and covariance if and only if the expert provides it.
If he provides it then she observes its value between receiving and implementing his advice.
Let $Y\in\{0,1\}$ indicate whether the advisee takes the ``correct'' action.
Then
\begin{align}
\Pr(Y=1\,\vert\,r,\theta)
&= \begin{cases}
\Pr(A=1) & \text{if}\ r=S \\
\Pr(A=X_\theta) & \text{if}\ r=C
\end{cases} \notag \\
&= \begin{cases}
0.5 & \text{if}\ r=S \\
0.5+2\theta\sigma & \text{if}\ r=C
\end{cases} \label{eq:prob-Y-r-theta}
\end{align}
for each $\theta\in\{0,1\}$.
\paragraph{Beliefs}
Neither the advisee nor the expert know whether he is competent.
They have common prior $\Pr(\theta=1)=\pi_0\in(0,1)$.
Hence
\begin{align}
\Pr(Y=1\,\vert\,r)
&= \Pr(Y=1\,\vert\,r,\theta=1)\Pr(\theta=1)+\Pr(Y=1\,\vert\,r,\theta=0)\Pr(\theta=0) \notag \\
&= \begin{cases}
0.5 & \text{if}\ r=S \\
0.5+2\sigma\pi_0 & \text{if}\ r=C.
\end{cases} \label{eq:prob-Y-r}
\end{align}
The expert's ignorance of $\theta$ prevents him from trying to signal it.
Therefore, his chosen rule $r$ is independent of $\theta$.%
\footnote{
This remains true (in equilibrium) if the expert knows~$\theta$---see Section~\ref{sec:known-type}.
}
The players use~\eqref{eq:prob-Y-r-theta}, \eqref{eq:prob-Y-r}, and Bayes' rule to form a posterior belief
\begin{align}
\Pr(\theta=1\,\vert\,r,Y)
&= \frac{\Pr(Y\,\vert\,r,\theta=1)\Pr(\theta=1)}{\Pr(Y\,\vert\,r)} \notag \\
&= \begin{cases}
\pi_0 & \text{if}\ r=S \\
\frac{(1+4\sigma)\pi_0}{1+4\sigma\pi_0} & \text{if}\ r=C\ \text{and}\ Y=1 \\
\frac{(1-4\sigma)\pi_0}{1-4\sigma\pi_0} & \text{if}\ r=C\ \text{and}\ Y=0
\end{cases} \label{eq:posterior}
\end{align}
about $\theta$ after observing the chosen rule $r$ and the outcome $Y$.
For convenience, I let
\[ \pi_1(Y)\equiv\Pr(\theta=1\,\vert\,r=C,Y) \]
denote the random posterior belief about $\theta$ when the expert chooses the complex rule.
This belief has expected value
\begin{align}
\E[\pi_1(Y)]
&= \pi_1(1)\Pr(Y=1\,\vert\,r=C)+\pi_1(0)\Pr(Y=0\,\vert\,r=C) \notag \\
&= \Pr(Y=1\,\vert\,r=C,\theta=1)\pi_0+\Pr(Y=0\,\vert\,r=C,\theta=1)\pi_0 \notag \\
&= \pi_0. \label{eq:posterior-mean}
\end{align}
Thus, choosing the complex rule leads to unchanged beliefs \emph{on average}, while choosing the simple rule leads to unchanged beliefs independently of whether the advisee takes the correct action.
\paragraph{Payoffs}
The expert's payoff has two components:
\begin{enumerate}
\item
A unit ``accuracy payoff'' from the advisee taking the correct action;
\item
A ``reputation payoff'' $\psi(\Pr(\theta=1\,\vert\,r,Y))$ that depends on the posterior belief~\eqref{eq:posterior} about $\theta$.
\end{enumerate}
I assume $\psi:[0,1]\to\mathbb{R}$ is non-decreasing but has $\psi(0)<\psi(1)$.
This function captures the expert's extrinsic incentive to be viewed as competent.
This incentive may reflect the continuation value of employment in a dynamic setting with replaceable experts (see Section~\ref{sec:wage-example}).
Alternatively, it may reflect public officials (e.g., Anthony Fauci) wanting to foster trust in government institutions (e.g., the Centres for Disease Control).
The expert's total payoff equals the sum of his accuracy and reputation payoffs.
He chooses the rule $r\in\{S,C\}$ that maximizes his expected payoff
\[ \Phi(r,\sigma,\pi_0)\equiv \begin{cases}
0.5+\psi(\pi_0) & \text{if}\ r=S \\
0.5+2\sigma\pi_0+\Psi(\sigma,\pi_0) & \text{if}\ r=C,
\end{cases} \]
where
\begin{align*}
\Psi(\sigma,\pi_0)
&\equiv \E[\psi(\pi_1(Y))] \\
&= (0.5+2\sigma\pi_0)\psi\left(\frac{(1+4\sigma)\pi_0}{1+4\sigma\pi_0}\right)+(0.5-2\sigma\pi_0)\psi\left(\frac{(1-4\sigma)\pi_0}{1-4\sigma\pi_0}\right)
\end{align*}
is his expected reputation payoff from choosing the complex rule.
He breaks ties by choosing the complex rule.%
\footnote{
This is an arbitrary assumption.
It is not consequential for any of my results.
}
The advisee's payoff equals $Y$.
Therefore, she always follows the expert's advice.
\paragraph{Timing}
The model timing is as follows:
\begin{enumerate}
\item
The expert chooses the simple or complex rule.
\item
If the expert chooses the simple rule then the advisee takes the action.
If he chooses the complex rule then she observes $X_\theta$ and takes the action if and only if $X_\theta=1$.
\item
The expert and advisee observe $Y$, and form a posterior belief~\eqref{eq:posterior} about $\theta$.
\item
Payoffs are realized.
\end{enumerate}
\subsection{Discussion of modeling assumptions}
\paragraph{Behavioral limitations}
I assume the advisee is boundedly rational.
Her behavioral limitation is not knowing the set of conditions that could predict whether the action is good.
This knowledge is what makes the expert an ``expert.''
However, if the expert provides a condition then I assume the advisee can verify its potential predictive power.
This prevents the expert from mis-reporting the predictive power of the condition he provides.
Thus, assuming verifiability allows me to focus on the expert's decision to provide the condition and abstract from his decision about how much uncertainty to communicate.
Intuitively, the expert in my model generates hypotheses about the relationship between conditions and actions.
Choosing the complex rule corresponds to stating a hypothesis and referencing the scientific literature on which it is based.
The advisee reads that literature and verifies that the hypothesis is ``reasonable.''
However, she cannot tell if the hypothesis is true.
That skill is unique to competent experts.
The form of bounded rationality I assume is intentionally unrealistic.
For example, one reason we outsource vaccine guidance to public health agencies, rather than have the public engage with the relevant virological science themselves, is because that science is difficult to understand.
This difficulty may lead the public to endure cognitive costs or make mistakes that agencies want to prevent.
Thus, one explanation for why experts provide ``simple'' advice is that they internalize the advisee's cognitive burden and avoid sharing ``too much detail.''
However, this explanation is unsatisfying because one can rationalize any amount of detail suppression with a suitable assumption about the advisee's cognitive costs.
In contrast, my reputation-based explanation does not appeal to cognitive costs.
It is meant to be idealized: to show that experts may simplify their advice even if the advisee has \emph{zero} cognitive costs.
Section~\ref{sec:measurement-errors} considers an extension to my model in which the advisee makes mistakes when implementing the complex rule.
\paragraph{Expert does not know $\theta$}
I assume the expert does not know if he is competent.
This allows me to abstract from competence signaling motives because the expert has no information about his competence to signal.
It makes the expert's choice of model uninformative about his competence, and isolates the tension between predictive power and reputational risk.
Section~\ref{sec:known-type} considers the implications of letting the expert know $\theta$.
\paragraph{Expert has aligned preferences}
I assume the expert is ``unbiased'' in that he wants the advisee to take the action if and only if she thinks it is good.
This assumption is appropriate in the science communication setting that motivates my analysis: experts are intrinsically ``benevolent'' but face extrinsic incentives to preserve their reputation.
It allows me to abstract from persuasion motives \cite[as studied, e.g., by][]{Kamenica-Gentzkow-2011-AER} and lie detection \cite[see, e.g.,][]{Balbuzanov-2019-IJGT}.
\section{Analysis}
\label{sec:model-analysis}
\subsection{Expert's choice}
Proposition~\ref{prop:rule-choice} characterizes the expert's choice between the simple and complex rules.
For convenience, I let
\begin{align*}
\Delta\Phi(\sigma,\pi_0)
&\equiv \Phi(C,\sigma,\pi_0)-\Phi(S,\sigma,\pi_0) \notag \\
&= 2\sigma\pi_0+\Psi(\sigma,\pi_0)-\psi(\pi_0)
\end{align*}
denote the gain in the expert's expected payoff from choosing the complex rule over the simple rule.
He chooses the complex rule if and only if this gain is non-negative:
\begin{proposition}
\label{prop:rule-choice}
The expert chooses the complex rule if and only if
%
\begin{equation}
2\sigma\pi_0 \ge \psi(\pi_0)-\Psi(\sigma,\pi_0). \label{ineq:rule-choice}
\end{equation}
\end{proposition}
The left-hand side of~\eqref{ineq:rule-choice} equals the gain in the expert's expected accuracy payoff from choosing the complex rule over the simple rule.
The right-hand side equals the loss in his expected reputation payoff from choosing the complex rule over the simple rule.
If $\psi$ is convex then the expert's reputation gain from advising the correct action exceeds his reputation loss from advising the incorrect action.
As a result, he chooses the complex rule because it is more accurate and offers an expected reputation \emph{benefit}:
\begin{corollary}
\label{crly:rule-choice-convex}
If $\psi$ is linear or convex, then the expert chooses the complex rule.
\end{corollary}
In contrast, if $\psi$ is concave then the reputation gain from advising the correct action is smaller than the reputation loss from advising the incorrect action.
As a result, choosing the complex rule incurs an expected reputation \emph{cost}.
Choosing the simple rule avoids this cost because it prevents the advisee from learning about the expert's competence.
He chooses the simple rule if this incentive to hide his competence dominates his incentive to advise the correct action.
Suppose the expert's reputation payoff comprises a wage $w\ge0$ earned with some probability $R(\pi)$ that depends on the posterior belief $\pi$ about his competence.
For example, the advisee may decide whether to re-hire the expert based on his perceived competence relative to other experts (see Section~\ref{sec:wage-example}).
Corollary~\ref{crly:rule-choice-wR} states that if $R$ is concave in $\pi$ (e.g., since competition among more competent experts is more aggressive) then the expert chooses the complex rule if and only if $w$ is sufficiently low.
\begin{corollary}
\label{crly:rule-choice-wR}
Let $w\ge0$ and suppose $\psi(\pi)=w\,R(\pi)$ for all beliefs $\pi\in[0,1]$, where $R:[0,1]\to\mathbb{R}$ is non-decreasing and concave.
Then the expert chooses the complex rule if and only if
\[ w\le\frac{2\sigma\pi_0}{R(\pi_0)-\E[R(\pi_1(Y))]}. \]
\end{corollary}
\begin{figure}
\centering
\includegraphics{figures/sqrt-example-delta-Phi.pdf}
\caption{Values of $\Delta\Phi(\sigma,\pi_0)$ when $\psi(\pi)=w\sqrt\pi$, $\sigma\in\{0.1,0.2\}$, $\pi_0\in\{0.25,0.5,0.75\}$, and $w\in[0,10]$}
\label{fig:sqrt-example-delta-Phi}
\end{figure}
For example, suppose $R(\pi)\equiv\sqrt{\pi}$.
Figure~\ref{fig:sqrt-example-delta-Phi} plots the corresponding value of~$\Delta\Phi(\sigma,\pi_0)$ for combinations of $\sigma\in\{0.1,0.2\}$, $\pi_0\in\{0.25,0.5,0.75\}$, and $w\in[0,10]$.
This value falls when $w$ rises because the expected reputational cost of choosing the complex rule rises.
If $\sigma=0.2$ and $\pi_0=0.25$ (i.e., competence is valuable but rare) then the complex rule is good at diagnosing the expert's probable incompetence.
Consequently, the expert chooses the complex rule if and only if $w<3.07$; that is, his reputational concerns are sufficiently \emph{weak}.
This threshold value of $w$ rises when $\sigma$ falls to 0.1, which makes competence less valuable but also makes the complex rule less diagnostic.
The threshold also rises when $\pi_0$ rises, in part because $\psi$ becomes less concave near $\pi_0$.
This fall in local concavity lowers the reputational loss from advising the incorrect action relative to the gain from advising the correct action.
In this example, the expert chooses the complex rule only if the maximum reputation payoff~$w$ is many times larger than his accuracy payoff.
This difference in payoff magnitudes is plausible if $w$ represents the continuation value of employment.
This value includes the present value of the accuracy payoffs accrued from future advisory events.
If there are many such events and the expert is patient then $w$ can be large.
For example, if he offers advice at each date $t\in\{0,1,2,\ldots,10\}$ and has inter-temporal discount factor $\delta=0.95$ then the date~0 value of his date $t\ge1$ accuracy payoffs is $\delta(1-\delta^{10})/(1-\delta)\approx7.62$ times larger than the value of his date~0 accuracy payoff.
\subsection{Changes in parameter values}
I now analyze how the posterior belief $\pi_1(Y)$ and gain in expected payoff $\Delta\Phi(\sigma,\pi_0)$ from choosing the complex rule depend on the parameters $\sigma$ and $\pi_0$.
\begin{lemma}
\label{lem:posterior-comparative-statics}
$\pi_1(1)$ is increasing in $\sigma$ and $\pi_0$, whereas $\pi_1(0)$ is decreasing in $\sigma$ and non-decreasing in $\pi_0$.
\end{lemma}
\begin{figure}
\centering
\includegraphics{figures/posteriors.pdf}
\caption{Posterior belief $\pi_1(Y)$ for $Y\in\{0,1\}$ and $\sigma\in\{0.1,0.2\}$}
\label{fig:posteriors}
\end{figure}
Figure~\ref{fig:posteriors} shows how $\pi_1(1)$ and $\pi_1(0)$ grow with $\pi_0$ when $\sigma\in\{0.1,0.2\}$.
As $\sigma$ rises, the complex rule becomes less likely to lead to the incorrect action if the expert is competent.
Then observing $Y=0$ becomes more diagnostic of his \emph{in}competence if he chooses the complex rule.
Whether this dissuades him from choosing that rule depends on the shape of $\psi$:
\begin{proposition}
\label{prop:Delta-Phi-comparative-statics-sigma}
If $\psi$ is linear or convex, then $\Delta\Phi(\sigma,\pi_0)$ is increasing in $\sigma$.
\end{proposition}
Increasing the covariance $\sigma$ makes the complex rule (i)~more likely to lead to the correct action and (ii)~more diagnostic of the expert's competence.
These two forces on $\Delta\Phi(\sigma,\pi_0)$ act in the same direction when $\psi$ is convex.
In that case, the expert gains more from being seen as more competent than he loses from being seen as less competent.
Consequently, making the complex rule better at diagnosing his competence makes him more willing to choose that rule.
In contrast, forces (i) and (ii) act in \emph{opposite} directions when $\psi$ is concave.
In that case, the expert gains \emph{less} from being seen as more competent than he loses from being seen as less competent.
Consequently, making the complex rule better at diagnosing his competence makes him more willing to choose that rule if and only if his reputation concerns are sufficiently weak.
Figure~\ref{fig:sqrt-example-delta-Phi} demonstrates this condition: if $\psi(\pi)=w\sqrt{\pi}$ then $\Delta\Phi(0.2,0.5)>\Delta\Phi(0.1,0.5)$ when $w$ is small but $\Delta\Phi(0.2,0.5)<\Delta\Phi(0.1,0.5)$ when $w$ is large.
Increasing $\pi_0$ raises the probability that the complex rule has superior predictive power.
It also raises the expert's reputation payoff at each posterior belief about his competence.
However, this rise in reputation payoffs applies to the complex rule \emph{and} the simple model.
Consequently, the overall effect on the expert's willingness to choose the complex rule depends on how changing $\pi_0$ changes his reputation payoffs under each rule.
Proposition~\ref{prop:Delta-Phi-comparative-statics-prior-differentiable} characterizes this overall effect.
\begin{proposition}
\label{prop:Delta-Phi-comparative-statics-prior-differentiable}
Suppose $\psi$ is differentiable at $\pi_0$, $\pi_1(1)$, and $\pi_1(0)$.
Then $\Delta\Phi(\sigma,\pi_0)$ is increasing in $\pi_0$ if and only if
%
\begin{equation}
2\sigma+2\sigma\left[\psi(\pi_1(1))-\psi(\pi_1(0))\right]+\E\left[\psi'(\pi_1(Y))\parfrac{\pi_1(Y)}{\pi_0}\right]>\psi'(\pi_0), \label{ineq:Delta-Phi-comparative-statics-prior-differentiable}
\end{equation}
%
where $\psi'$ denotes the derivative of $\psi$.
\end{proposition}
The left-hand side of~\eqref{ineq:Delta-Phi-comparative-statics-prior-differentiable} decomposes the effect of increasing $\pi_0$ on the expert's expected payoff from choosing the complex rule.
The first term equals the increase in his expected accuracy payoff.
The second term equals the increase in his expected reputational payoff from being more likely to be competent.
The third term equals the increase in his expected reputation payoff induced by the increase in the posteriors $\pi_1(1)$ and $\pi_1(0)$.
The right-hand side of~\eqref{ineq:Delta-Phi-comparative-statics-prior-differentiable} equals the increase in the reputation payoff from choosing the simple rule.
Increasing $\pi_0$ increases $\Delta\Phi(\sigma,\pi_0)$ if and only if the sum of the three terms on the left-hand side of~\eqref{ineq:Delta-Phi-comparative-statics-prior-differentiable} exceeds the term on the right-hand side.
Proposition~\ref{prop:Delta-Phi-comparative-statics-prior-wR} analyzes the class of reputation payoff functions $\psi(\pi)=w\,R(\pi)$ considered in Corollary~\ref{crly:rule-choice-wR}, replacing the concavity restriction with a differentiability restriction.
\begin{proposition}
\label{prop:Delta-Phi-comparative-statics-prior-wR}
Let $w\ge0$ and suppose $\psi(\pi)=w\,R(\pi)$ for all beliefs $\pi\in[0,1]$, where $R:[0,1]\to\mathbb{R}$ is non-decreasing and differentiable.
Then either
%
\begin{enumerate}
\item
$\Delta\Phi(\sigma,\pi_0)$ is increasing in $\pi_0$, or
\item
There exists $w^*>0$ such that $\Delta\Phi(\sigma,\pi_0)$ is increasing in $\pi_0$ if and only if $w<w^*$.
\end{enumerate}
\end{proposition}
Intuitively, increasing $\pi_0$ raises the payoff from choosing the simple rule by a \emph{certain} amount but the payoff from choosing the complex rule by an \emph{uncertain} amount.
If $R$ is ``not too concave'' or the stakes are sufficiently low (i.e., $w$ is small) then the uncertain amount is more attractive.
\section{Wages and replacement thresholds}
\label{sec:wage-example}
This section considers a setting in which the expert earns wage $w>0$ if and only if the posterior belief about his competence exceeds a threshold $\pi^*\in(0,1)$.
This setting arises, for example, when the advisee faces an action in each of two periods, and she hires the expert to provide advice in the second period if and only if she believes he is more likely to be competent than another expert drawn from the market.
Then the expert's reputation payoff is given by
\begin{equation}
\psi(\pi)=\begin{cases}
w & \text{if}\ \pi\ge\pi^* \\
0 & \text{otherwise}
\end{cases} \label{eq:wage-example-psi}
\end{equation}
for all beliefs $\pi\in[0,1]$.
Lemma~\ref{lem:wage-example-expected-payoff} defines the expert's expected payoffs from choosing the simple and complex rules in this setting.
\begin{lemma}
\label{lem:wage-example-expected-payoff}
Suppose the reputation payoff function $\psi$ is defined as in~\eqref{eq:wage-example-psi}.
Then the expert's expected payoff from choosing the complex rule equals
%
\begin{equation}
\Phi(C,\sigma,\pi_0)=\begin{cases}
0.5+2\sigma\pi_0 & \text{if}\ \pi_0<\underline\pi \\
0.5+2\sigma\pi_0+w(0.5+2\sigma\pi_0) & \text{if}\ \underline\pi\le\pi_0<\overline\pi \\
0.5+2\sigma\pi_0+w & \text{if}\ \overline\pi\le\pi_0,
\end{cases} \label{eq:wage-example-expected-payoff-complex}
\end{equation}
%
where
%
\begin{equation}
\underline\pi\equiv\frac{\pi^*}{1+4\sigma(1-\pi^*)} \label{eq:wage-example-lower-threshold}
\end{equation}
%
and
%
\begin{equation}
\overline\pi\equiv\frac{\pi^*}{1-4\sigma(1-\pi^*)}. \label{eq:wage-example-upper-threshold}
\end{equation}
%
His expected payoff from choosing the simple rule equals
%
\begin{equation}
\Phi(S,\sigma,\pi_0)=\begin{cases}
0.5+w & \text{if}\ \pi_0\ge\pi^* \\
0.5 & \text{otherwise}.
\end{cases} \label{eq:wage-example-expected-payoff-simple}
\end{equation}
\end{lemma}
Suppose the expert chooses the simple rule.
Then the posterior belief about his competence equals the prior belief because no information about his competence is revealed.
Consequently, he receives the wage $w$ if and only if his prior exceeds the threshold $\pi^*$.
Now suppose the expert chooses the complex rule.
If $\pi_0<\underline\pi$ then the expert is guaranteed a reputation payoff of zero because the posterior belief $\pi_1(1)$ induced by the advisee taking the correct action is less than $\pi^*$.
Conversely, if $\pi_0\ge\underline\pi$ then the expert is guaranteed a reputation payoff of $w$ because the posterior belief $\pi_1(0)$ induced by the advisee taking the \emph{incorrect} action is greater than $\pi^*$.
If $\pi_0\in[\underline\pi,\overline\pi)$ then the expert receives a positive reputation payoff if and only if the advisee takes the correct action.
\begin{figure}
\centering
\includegraphics{figures/wage-example.pdf}
\caption{Graphs of $\Phi(C,\sigma,\pi_0)$ and $\Phi(S,\sigma,\pi_0)$ with $\psi$ defined as in~\eqref{eq:wage-example-psi} and with $\sigma\in\{0,1\}$, $w\in\{0.5,1\}$, and $\pi^*=0.5$}
\label{fig:wage-example}
\end{figure}
Figure~\ref{fig:wage-example} plots the expected payoffs~\eqref{eq:wage-example-expected-payoff-complex} and~\eqref{eq:wage-example-expected-payoff-simple} against $\pi_0$ when $\sigma\in\{0.1,0.2\}$, $w\in\{0.5,1\}$, and $\pi^*=0.5$.
It demonstrates three cases:
\begin{enumerate}
\item
$\Phi(C,\sigma,\pi_0)$ exceeds $\Phi(S,\sigma,\pi_0)$ for all $\pi_0$ (as when, e.g., $(\sigma,w)=(0.2,0.5)$).
\item
$\Phi(C,\sigma,\pi_0)$ and $\Phi(S,\sigma,\pi_0)$ ``cross'' at the replacement threshold $\pi_0=\pi^*$ and the discontinuity $\pi_0=\overline\pi$ (as when, e.g., $(\sigma,w)=(0.1,0.5)$);
\item
$\Phi(C,\sigma,\pi_0)$ and $\Phi(S,\sigma,\pi_0)$ intersect at some $\pi_0\in[\underline\pi,\overline\pi)$ (as when, e.g., $(\sigma,w)=(0.2,1)$).
\end{enumerate}
Proposition~\ref{prop:wage-example-choice} states that these are the \emph{only} three possible cases.%
\footnote{
Corollary~\ref{crly:rule-choice-wR} does not apply in this setting because $\psi$ is not concave.
}
\begin{proposition}
\label{prop:wage-example-choice}
Suppose the reputation payoff function $\psi$ is defined as in~\eqref{eq:wage-example-psi}.
Define
%
\begin{equation}
\pi^\dagger\equiv\frac{w}{4\sigma(1+w)}. \label{eq:wage-example-prior-dagger}
\end{equation}
%
If $\pi^\dagger\le\pi^*$ then the expert chooses the complex rule.
If $\pi^\dagger>\pi^*$ then he chooses the \emph{simple} rule if and only if $\pi^*\le\pi_0<\min\{\pi^\dagger,\overline\pi\}$, where $\overline\pi$ is defined as in Lemma~\ref{lem:wage-example-expected-payoff}.
\end{proposition}
If the wage $w$ is sufficiently low or the threshold $\pi^*$ is sufficiently high, then the expert receives a greater expected payoff from choosing the complex rule regardless of the prior $\pi_0$.
If $w$ is high and the prior lies between $\pi^*$ and $\min\{\pi^\dagger,\overline\pi\}$, then the reputational risk induced by choosing the complex rule dissuades the expert from making that choice.
If $\pi_0$ is the proportion of employable experts who are competent and the advisee replaces the expert with a random competitor, then she maximizes the probability of employing a competent expert by setting $\pi^*=\pi_0$.
In that case, if she wants the expert to choose the complex model then she must limit his reputational risk by offering a low wage:
\begin{corollary}
\label{crly:wage-example-wage-threshold}
Suppose the reputation payoff function $\psi$ is defined as in~\eqref{eq:wage-example-psi}.
If $\pi^*=\pi_0$ then the expert chooses the complex rule if and only if $w\le4\sigma\pi_0/(1-4\sigma\pi_0)$.
\end{corollary}
Alternatively, the wage may be fixed (due to, e.g., legal or institutional constraints).
Then the advisee's only way to solicit the complex rule is to vary the replacement threshold $\pi^*$.
She has two choices: set it low enough to guarantee that the expert receives the wage, or set it \emph{high} enough to guarantee that expert \emph{does not} receive the wage.
Both choices remove the expert's reputational risk.
The first choice removes the risk by fully insuring it.
The second choice removes the risk by making explicit that the advisee seeks the expert's advice as a ``one-off,'' effectively imposing a single-period term limit on the expert's employment.
\section{Extensions}
\label{sec:extensions}
This section considers several standalone extensions to my analysis in Sections~\ref{sec:model-setup} and~\ref{sec:model-analysis}.
\subsection{$\Pr(A=1)\not=0.5$}
\label{sec:prob-A}
Sections~\ref{sec:model-setup} and~\ref{sec:model-analysis} assume $\Pr(A=1)=0.5$.
This equates the simple and complex rules' expected accuracy payoffs when the expert is incompetent.
Consequently, it isolates the tension between the complex rule's superior predictive power and its ability to reveal the expert's competence.
If $\Pr(A=1)=a\in(0,1)$ then
\[ \Pr(A=X_\theta)=ax+(1-a)(1-x)+2\theta\sigma \]
and the condition~\eqref{ineq:rule-choice} under which the expert chooses the complex rule becomes
\begin{equation}
2\sigma\pi_0-(2a-1)(1-x)\ge\psi(\pi_0)-\Psi(\sigma,\pi_0). \label{eq:rule-choice-a}
\end{equation}
The left-hand side of~\eqref{ineq:rule-choice} exceeds the left-hand side of~\eqref{eq:rule-choice-a} if and only if $a>0.5$.
In that case, conditioning on $X_0$ lowers the expert's accuracy payoff because it lowers the probability that the advisee takes the action when it is good.
Conditioning on $X_1$ raises the expert's accuracy payoff if and only if the covariance $\sigma$ of $A$ and $X_1$ is sufficiently high.
This covariance depends on $a$ since it depends on the variance $\Var(A)=a(1-a)$.
Therefore, increasing $a$ changes both terms on the left-hand side of~\eqref{eq:rule-choice-a}.%
\footnote{
If $A$ and $X_1$ have correlation $\rho$ then $\sigma=\rho\sqrt{a(1-a)x(1-x)}$.
Then the left-hand side of~\eqref{eq:rule-choice-a} is increasing in $a$ if and only if
\[ \rho\pi_0(1-2a)>2(1-x)\sqrt{\frac{a(1-a)}{x(1-x)}}. \]
If $a\ge0.5$ then this condition never holds.
If $a<0.5$ then it holds if and only if $\rho\pi_0$ is sufficiently large.
}
Increasing $a$ also changes the right-hand side by changing the distribution of posterior beliefs $\pi_1(Y)$ and, in turn, the expected reputation payoff $\Psi(\sigma,\pi_0)$.
The overall effect of increasing $a$ on $\Delta\Phi(\sigma,\pi_0)$ is ambiguous; it depends on the parameters $x$, $\sigma$ and $\pi_0$, and the shape of $\psi$.
I do not elaborate on this dependency because it is not central to my analysis.
\subsection{Measurement errors}
\label{sec:measurement-errors}
One reason real-world experts may prefer the simple rule is that the advisee may make mistakes when implementing the complex rule.
For example, suppose the expert provides condition $X_\theta$ but the advisee observes $\hat{X}_\theta\in\{0,1\}$ with $\Pr(\hat{X}_\theta\not=X_\theta)=\epsilon\in(0,0.5]$.
Proposition~\ref{prop:measurement-errors-complex-accuracy} states that such measurement errors decrease the complex rule's accuracy.%
\footnote{
If $\Pr(A=1)\not=0.5$ then whether measurement errors decrease the complex rule's accuracy depends on the joint distribution of $A$ and $X_\theta$, and on the relative rates of false positives and negatives.
}
\begin{proposition}
\label{prop:measurement-errors-complex-accuracy}
Let $\epsilon\in[0,0.5]$.
Suppose $\hat{X}_\theta\in\{0,1\}$ is distributed such that $\Pr(\hat{X}_\theta\not=X_\theta)=\epsilon$ for each $\theta\in\{0,1\}$.
Then
%
\begin{equation}
\Pr(A=\hat{X}_\theta)=0.5+2\theta(1-2\epsilon)\sigma. \label{eq:measurement-errors-complex-accuracy}
\end{equation}
\end{proposition}
Measurement errors decrease the covariance between the observed condition and the correct action.
By Proposition~\ref{prop:Delta-Phi-comparative-statics-sigma}, this decrease makes the expert less willing to choose the complex rule if the reputation payoff function $\psi$ is linear or convex.
However, if $\psi$ is concave then measurement errors increase the expected reputation payoff from choosing the complex rule because they make that rule less diagnostic of the expert's competence.
Thus, introducing measurement errors may lead the expert to switch to the complex rule if the reputation benefit exceeds the accuracy cost.
In practice, measurement errors may arise exogenously from, for example, imprecise medical tests.
They may also arise endogenously from the advisee optimally allocating their scarce attention and mental resources.%
\footnote{
This possibility relates to \citeapos{Oprea-2020-AER} claim that complex rules are cognitively costly to implement.
}
For example, suppose the advisee receives benefit $b>0$ from taking the correct action and zero from taking the incorrect action.
Suppose further that she endures cost $c(0.5-\epsilon)^2$ from observing $\hat{X}_\theta$ with error $\epsilon\in[0,0.5]$, where $c\in(0,2b\sigma\pi_0)$.%
\footnote{
This upper bound on $c$ ensures $\epsilon^*<0.5$.
}
If the expert chooses the complex rule then the advisee chooses the error $\epsilon^*$ that maximizes
\[ b\left(0.5+2\sigma(1-2\epsilon)\pi_0\right)-c(0.5-\epsilon)^2, \]
which has unique maximizer
\[ \epsilon^*\equiv0.5-\frac{2b\sigma\pi_0}{c}. \]
Substituting $\epsilon=\epsilon^*$ into~\eqref{eq:measurement-errors-complex-accuracy} then gives
\[ \Pr(A=\hat{X}_\theta)=0.5+\frac{4b\theta\sigma^2\pi_0}{c}. \]
Thus, in this example, the accuracy of the complex rule is increasing in the advisee's benefit-cost ratio $b/c$.
Now suppose the expert knows $b$ and $c$, and anticipates the advisee's choice of error $\epsilon^*$ when choosing between rules.
Then it is possible the advisee is willing to ``pay'' for low errors because her benefit-cost ratio is high, but this payment leads the expert to avoid the complex rule because it becomes too diagnostic of his competence.
If the advisee wants to learn the expert's competence (e.g., because she will face a higher stakes choice in the future) then she may prefer to under-invest in measuring the condition correctly.
This would encourage the expert to choose the complex rule and reveal information about his competence.
\subsection{Expert knows $\theta$}
\label{sec:known-type}
Sections~\ref{sec:model-setup} and~\ref{sec:model-analysis} assume that the expert does not know whether he is competent.
This assumption allows me to isolate his trade-off between providing accurate advice and managing reputational risk.
If the expert knows his competence type then he also has a signaling motive.
I analyze that motive below.
Suppose the expert knows his competence type $\theta$ but the advisee does not.
Then his choice of rule may reveal information about $\theta$.
Lemma~\ref{lem:known-type-posterior} describes how the advisee incorporates that information into her posterior belief about $\theta$.
\begin{lemma}
\label{lem:known-type-posterior}
Suppose the expert knows $\theta$ but the advisee does not.
Let $\hat{q}_\theta\equiv\Pr(r=C\,\vert\,\theta)$ be the probability with which the advisee believes an expert with type $\theta$ chooses the complex rule.
Then her posterior belief about $\theta$ after observing the chosen rule $r\in\{S,C\}$ and its outcome $Y\in\{0,1\}$ is given by
%
\begin{equation}
\Pr(\theta=1\,\vert\,r,Y)
= \begin{cases}
\pi_S & \text{if}\ r=S \\
\frac{(1+4\sigma)\pi_C}{1+4\sigma\pi_C} & \text{if}\ r=C\ \text{and}\ Y=1 \\
\frac{(1-4\sigma)\pi_C}{1-4\sigma\pi_C} & \text{if}\ r=C\ \text{and}\ Y=0,
\end{cases} \label{eq:known-type-posterior}
\end{equation}
%
where
%
\begin{equation}
\pi_r=\begin{cases}
\frac{(1-\hat{q}_1)\pi_0}{(1-\hat{q}_1)\pi_0+(1-\hat{q}_0)(1-\pi_0)} & \text{if}\ r=S \\
\frac{\hat{q}_1\pi_0}{\hat{q}_1\pi_0+\hat{q}_0(1-\pi_0)} & \text{if}\ r=C
\end{cases} \label{eq:known-type-intermediate-belief}
\end{equation}
%
is her ``intermediate'' belief about $\theta$ after observing $r$ but before observing $Y$.
\end{lemma}
Equations~\eqref{eq:posterior} and~\eqref{eq:known-type-posterior} coincide when $p_0=p_1$; that is, when the advisee believes that competent and incompetent experts choose between rules with the same probabilities.
Then the choice of rule is uninformative and so $\pi_r=\pi_0$ for each $r\in\{S,C\}$.
I now study the perfect Bayesian equilibrium (henceforth ``equilibrium'') of the game in which the expert chooses a rule and the advisee forms a posterior belief about $\theta$.
In this game, the expert's strategy comprises (i)~a probability $p_\theta$ with which he chooses the complex rule given his type $\theta$ and (ii)~a conjecture $\hat{p}_{1-\theta}$ about the corresponding probability for the other type.
The advisee's strategy comprises a conjecture $\hat{q}_\theta$ about the probability for each type.
She uses this conjecture to form her posterior belief~\eqref{eq:known-type-posterior}.
In equilibrium, all conjectures are correct: $p_\theta=\hat{p}_\theta=\hat{q}_\theta$.
As a shorthand, I describe an equilibrium by the pair $(p_0,p_1)$ of probabilities that define it, and refer to $p_\theta$ as the $\theta$-type expert's ``strategy.''
This strategy is ``pure'' if $p_\theta\in\{0,1\}$ and ``mixed'' if $p_\theta\in(0,1)$.
It is his best-response to the strategy played by the other type, given the advisee's subsequent inference about $\theta$ and its reputational consequences.
Lemma~\ref{lem:known-type-expected-payoffs} defines the expected payoff that experts of each type maximize to determine their best-response.
\begin{lemma}
\label{lem:known-type-expected-payoffs}
Suppose the expert knows $\theta$ but the advisee does not.
Let $p_\theta$ be the equilibrium probability with which an expert with type $\theta$ chooses the complex rule.
Then such an expert has expected payoff
%
\begin{equation}
\Omega_\theta(\sigma,\pi_0,p_0,p_1)
\equiv p_\theta\Phi_\theta(C,\sigma,\pi_0,p_0,p_1)+(1-p_\theta)\Phi_\theta(S,\sigma,\pi_0,p_0,p_1), \label{eq:known-type-expected-payoff}
\end{equation}
%
where
%
\begin{equation}
\Phi_\theta(C,\sigma,\pi_0,p_0,p_1)
\equiv 0.5+2\sigma\theta+(0.5+2\sigma\theta)\psi\left(\frac{(1+4\sigma)\pi_C}{1+4\sigma\pi_C}\right)+(0.5-2\sigma\theta)\psi\left(\frac{(1-4\sigma)\pi_C}{1-4\sigma\pi_C}\right) \label{eq:known-type-expected-payoff-complex}
\end{equation}
%
is his expected payoff from choosing the complex rule with certainty,
%
\begin{equation}
\Phi_\theta(S,\sigma,\pi_0,p_0,p_1)\equiv 0.5+\psi\left(\pi_S\right) \label{eq:known-type-expected-payoff-simple}
\end{equation}
%
is his expected payoff from choosing the simple rule with certainty, and $\pi_C$ and $\pi_S$ are defined as in Lemma~\ref{lem:known-type-posterior}.
\end{lemma}
I first show that there are no pure separating equilibria.
This is because incompetent experts never play strategies that allow the advisee to diagnose their incompetence with certainty.
\begin{proposition}
\label{prop:known-type-pure-equilibrium}
Suppose the expert knows $\theta$ but the advisee does not.
If a pure strategy equilibrium exists, then it is pooling.
\end{proposition}
The game may have two, one, or no pure pooling equilibria.
It may also have mixed strategy equilibria.
The number and nature of equilibria depend on the parameters $\sigma$ and $\pi_0$, and the shape of the reputation payoff function $\psi$.
To gain traction, I focus on the ``replaceable expert'' setting discussed in Section~\ref{sec:wage-example}.
I assume the replacement threshold equals $\pi^*=\pi_0$.
Lemma~\ref{lem:known-type-wage-example-expected-payoffs} defines the expert's expected payoffs in this setting.
\begin{lemma}
\label{lem:known-type-wage-example-expected-payoffs}
Suppose the expert knows $\theta$ but the advisee does not.
Let the probabilities $p_\theta$ and functions $\Phi_\theta(r,\sigma,\pi_0,p_0,p_1)$ be defined as in Lemma~\ref{lem:known-type-expected-payoffs}, and let the reputation payoff function $\psi$ be defined as in~\eqref{eq:wage-example-psi} with $\pi^*=\pi_0$.
Then
%
\begin{equation}
\Phi_\theta(C,\sigma,\pi_0,p_0,p_1)
=\begin{cases}
0.5+2\sigma\theta & \text{if}\ p_1(1+4\sigma)<p_0 \\
0.5+2\sigma\theta+(0.5+2\sigma\theta)w & \text{if}\ p_1(1-4\sigma)<p_0\le p_1(1+4\sigma) \\
0.5+2\sigma\theta+w & \text{if}\ p_0\le p_1(1-4\sigma)
\end{cases} \label{eq:known-type-wage-example-expected-payoff-complex}
\end{equation}
%
and
%
\begin{equation}
\Phi_\theta(S,\sigma,\pi_0,p_0,p_1)
=\begin{cases}
0.5+w & \text{if}\ p_0\ge p_1 \\
0.5 & \text{otherwise}.
\end{cases} \label{eq:known-type-wage-example-expected-payoff-simple}
\end{equation}
\end{lemma}
Proposition~\ref{prop:known-type-wage-example-best-responses} determines competent and incompetent experts' best responses by maximizing their expected payoff $\Omega_\theta(\sigma,\pi_0,p_0,p_1)$ under the conditions specified in Lemma~\ref{lem:known-type-wage-example-expected-payoffs}.
\begin{proposition}
\label{prop:known-type-wage-example-best-responses}
Under the conditions specified in Lemma~\ref{lem:known-type-wage-example-expected-payoffs}, we have
%
\begin{equation}
\argmax_p\Omega_0(\sigma,\pi_0,p,p_1)=\{p_1\} \label{eq:known-type-wage-example-best-response-incompetent}
\end{equation}
%
and
%
\begin{align}
&\argmax_p\Omega_1(\sigma,\pi_0,p_0,p) \notag \\
&= \begin{cases}
\{0\} & \text{if}\ 1-p_0<4\sigma<1\ \text{and}\ \max\left\{2\sigma,4\sigma/(1-4\sigma)\right\}<w \\
\{0\}\cup\left[\frac{p_0}{1+4\sigma},p_0\right]\cup\{1\} & \text{if}\ 1-p_0<4\sigma<1\ \text{and}\ 2\sigma<w=4\sigma/(1-4\sigma) \\
\{1\} & \text{otherwise}.
\end{cases} \label{eq:known-type-wage-example-best-response-competent}
\end{align}
\end{proposition}
Incompetent experts best-respond by mimicking competent experts.
Thus, in any equilibrium, the advisee learns nothing from the choice of rule because competent and incompetent experts play the same strategies.
Yet she may learn from the \emph{outcome} of the complex rule---i.e., whether it leads to the correct action---if the expert chooses it.
Competent experts have a dominant strategy to choose the complex rule if (i)~$w\le 2\sigma$ or (ii)~$4\sigma<1$ and $w<4\sigma/(1-4\sigma)$---i.e., their reputation concerns are sufficiently weak.
Otherwise, they want to differentiate themselves from incompetent experts.
\begin{figure}
\centering
\includegraphics{figures/known-type-best-responses.pdf}
\caption{Experts' best-response curves when $\sigma\in\{0.1,0.2\}$ and $w\in\{0.5,1,5\}$}
\label{fig:known-type-best-responses}
\end{figure}
Figure~\ref{fig:known-type-best-responses} shows the best response curves for experts of each type $\theta\in\{0,1\}$ when $\sigma\in\{0.1,0.2\}$ and $w\in\{0.5,1,5\}$.
These curves do not depend on the prior $\pi_0$.
Their intersections correspond to equilibria.
If the wage $w$ is small and the covariance $\sigma$ is large, then the unique equilibrium involves pure pooling on the complex rule.
Increasing $w$ or decreasing $\sigma$ destroys this equilibrium.
Indeed, Corollary~\ref{crly:known-type-wage-example-equilibria} states that if the wage $w$ is large relative to the covariance $\sigma$ then the game has no equilibria:
\begin{corollary}
\label{crly:known-type-wage-example-equilibria}
Under the conditions specified in Lemma~\ref{lem:known-type-wage-example-expected-payoffs},
%
\begin{enumerate}
\item[(i)]
If $4\sigma<1$ and $\max\left\{2\sigma,4\sigma/(1-4\sigma)\right\}<w$, then there are no equilibria;
\item[(ii)]
If $4\sigma<1$ and $2\sigma<w=4\sigma/(1-4\sigma)$, then the pair $(p_0,p_1)$ defines an equilibrium if and only if $p_0=p_1\in(1-4\sigma,1]$.
\end{enumerate}
%
Otherwise, the pair $(p_0,p_1)=(1,1)$ defines the unique equilibrium.
\end{corollary}
Consider the knife-edge case in which the competent expert's condition is imperfectly predictive (i.e., $4\sigma<1$) and $2\sigma<w=4\sigma/(1-4\sigma)$.
If he can mimic an incompetent expert who chooses the complex rule with sufficiently high probability, then he is indifferent between doing so and choosing the complex rule for sure.
Otherwise, and outside this knife-edge case, the game either has no equilibrium or a unique equilibrium with pure pooling on the complex rule.
This pooling equilibrium exists only when the expert's reputational concerns (indexed by $w$) are small relative to his accuracy concerns (indexed by $\sigma$).
Thus, letting the expert know his type preserves the main insight from Sections~\ref{sec:model-setup}--\ref{sec:wage-example}: experts give complex advice only if they face low downside reputational risk.%
\footnote{
This insight echoes an insight from the innovation literature that failure tolerance tends to improve performance \citep[see, e.g.,][]{Azoulay-etal-2011-RAND,Manso-2011-JF,Tian-Wang-2014-RFStud}.
}
\section{Related literature}
\label{sec:literature}
My analysis is closely related to the literature on reputational cheap talk.
That literature explores how the incentive to appear informed distorts the equilibria of cheap talk games in which a sender shares his signal of an \emph{ex post} observable state.
\cite{Ottaviani-Sorensen-2006-JET,Ottaviani-Sorensen-2006-RAND} show that, in equilibrium, the sender does not report his signal truthfully if it depends on whether he is informed.
Instead, he manipulates his report to induce the most favorable belief about his informedness.
For example, he may bias his report toward the receiver's prior belief about the state to hedge against the report disagreeing with the state.%
\footnote{
\cite{Klein-Mylovanov-2017-JMathEcon} argue that this behavior is limited to models with short time horizons.
If the expert has a long horizon, then he may forgo ``impressing'' the receiver in the short run if he becomes privately pessimistic about his ability to impress in the long run.
}
This hedging behavior corresponds to choosing the simple rule in my model.
Indeed, one could interpret my model as an example of reputational cheap talk in which the expert reports information about the predictive power of the condition he identifies.
However, the expert in my model has reputational concerns associated with the correctness of the action taken as a consequence of his report, rather than the correctness of the report itself.%
\footnote{
This distinction between actions and information is the focus of \cite{Andrews-Shapiro-2021-ECTA}, who study statistical decision theoretic models of science communication.
}
The innovation of my paper is to show how reputational concerns influence the type of advice experts provide: do they make ``blanket recommendations,'' or do they provide conditions under which to take different actions?%
\footnote{
\cite{Levit-Tsoy-2020-AEJMicro} argue that experts make blanket recommendations to conceal their bias.
}
\cite{Backus-Little-2020-APSR} offer another model similar to mine: one of an unbiased expert who provides advice about an action but has reputational concerns.
In their model, the correct action may be impossible to identify and only competent experts know if identification is possible.
This makes competent experts averse to admitting ignorance because doing so makes them look incompetent.
Whereas in my model, the correct action \emph{is} possible to identify but only competent experts can identify it.
This makes competent experts \emph{pretend to be ignorant} (by choosing the simple rule) if the reputational cost of appearing incompetent is sufficiently large.
\citeauthor{Backus-Little-2020-APSR} and my conclusions differ because our definitions of ``competence'' differ.
My definition of ``expertise'' is similar to that used by \cite{Callander-etal-2021-JPE}.
They note that experts tend to provide ``referential'' advice that combines suggestions for what to do with examples of inferior alternatives.
The authors posit that expertise consists in knowing the relative merits of different suggestions.
Similarly, I posit that expertise consists in knowing conditions under which different suggestions are preferable.
\citeauthor{Callander-etal-2021-JPE} and my analyses differ in what motivates the expert.
They focus on the expert's desire to retain an informational advantage over their advisee.
In contrast, I allow the expert to forfeit this advantage because it makes his advice more likely to induce the correct the action and may yield reputational benefits.
\cite{Callander-etal-2021-JPE} build on the literature on disclosing verifiable information.
The strand of that literature closest to my analysis is due to \cite{Dye-1985-JAccRes}.
It considers the possibility that the sender is uninformed.
Then the receiver cannot distinguish between the sender withholding information and having none \citep[see, e.g.,][]{Dziuda-2011-JET}.
\citeauthor{Dye-1985-JAccRes} studies a model in which the receiver knows the sender has information but does not know its value.
This is similar to my model, in which the advisee knows the expert is aware of predictive conditions but does not know their predictive power.
\citeauthor{Dye-1985-JAccRes} and my models differ in that the expert I study has no notion of whether his information is ``good'' or ``bad'' for the advisee.
So it is not that competent experts ``can partially conceal bad news by pooling with those who are unable to disclose'' \cite[p.\! 226]{Grubb-2011-JEMS}.
Instead, competent and incompetent experts avoid disclosure because they want to avoid its reputational consequences.
My assumption that the expert does not know his competence type is typical in the literature on career concerns.
That literature explores how agents' desire to appear competent drives a wedge between their risk preferences and principals' \citep{Holmstrom-RicartiCosta-1986-QJE,Holmstrom-1999-REStud}.%
\footnote{
\cite{Dewatripont-etal-1999-REStud} generalize \citeapos{Holmstrom-1999-REStud} model through the lens of information design.
}
This wedge can lead to inefficient information seeking \citep{Levy-2004-EER,Milbourn-etal-2001-RAND,Scharfstein-Stein-1990-AER} and risk-taking \citep{Aghion-Jackson-2016-AEJMicro,Chevalier-Ellison-1997-JPE,Prendergast-Stole-1996-JPE}.
For example, \cite{Aghion-Jackson-2016-AEJMicro} analyze replacement schemes in a dynamic setting similar to that described in Section~\ref{sec:wage-example}.%
\footnote{
\cite{Tsuyuhaha-2012-EB} also analyzes this setting in a two-period model.
He focuses on replacing biased experts, whereas I focus on replacing incompetent experts.
}
They study a ``leader'' who takes a materially safe or risky action, where the cost of the risky action being bad exceeds the benefit of it being good.
This concavity in material payoffs drives the leader toward the safe action, which is less beneficial on average but reveals no information about his type.
In contrast, the expert in my model chooses between \emph{reputationally} safe and risky advice, and I show how his choice depends on the concavity in his reputational payoffs.%
\footnote{
\cite{Rappoport-2022-} also distinguishes between material and reputational payoffs.
}
Many insights from the career concerns literature apply to my model.
For example, \cite{Gibbons-Murphy-1992-JPE} explain that compensation contracts should be strongest for workers close to retirement because they have the weakest career concerns.
The advisee in my model may prefer an older expert with lower expected competence $\pi_0$ to a younger expert with higher expected competence.
This is because the older expert has a lower continuation value of employment, making him more willing to risk losing that value by choosing the complex rule.
I analyze the relationship between expected competence, continuation values, and rule choices in Propositions~\ref{prop:Delta-Phi-comparative-statics-prior-wR} and \ref{prop:wage-example-choice}.
More generally, the reputation payoff function $\psi$ acts as a reduced-form expression for the expert's compensation scheme.
Its shape determines the expert's choice.
For example, if $\psi$ is close to constant then the expert prefers to choose the complex rule and risk being ``wrong'' (see Corollary~\ref{crly:rule-choice-convex}).
This echoes another insight from the career concerns literature: weakening incentives, such as by offering tenure \citep{Aghion-Jackson-2016-AEJMicro} or severance pay \citep{Bonatti-Horner-2017-TE}, can lead to more efficient behavior.
\section{Conclusion}
\label{sec:conclusion}
This paper discusses how reputational concerns impact the type of advice that experts provide.
I show that experts may ``simply'' their advice by excluding relevant conditions if their payoff is sufficiently concave in posterior beliefs about their competence.
This simplification makes advice worse by pooling states in which actions should and should not be taken.
Advisees can prevent experts from giving simple advice by limiting their reputational risk: by reducing the payoff from being perceived as competent, or by making payoffs independence of perceived competence.
{
\small
\raggedright
\bibliographystyle{apalike}
|
\section{Introduction} \label{sec:intro}
The plasma emission (PE) refers to electromagnetic radiation at frequencies close to the plasma oscillation frequency ($\wpe$) and its harmonics, corresponding to the fundamental (F) and harmonic (H) plasma emission, respectively. The framework of the standard theory of PE was suggested more than 6 decades ago \citep{1958SvA.....2..653G}, which involves a multi-step nonlinear process of wave-particle and wave-wave interactions due to the presence of energetic beam electrons in plasmas. The theory starts from the beam-driven kinetic bump-on-tail instability that leads to enhanced Langmuir waves, followed by the decay of these waves or their scattering over ion-related fluctuations generating the fundamental (F) emission and/or the backward-propagating Langmuir wave, and the nonlinear interaction of forward and backward-propagating Langmuir waves generating the harmonic (H) emission \citep[e.g.,][]{melrose_emission_1980,Melrose1987,1987JPlPh..38..169C,1994ApJ...422..870R, 2012JGRA..117.4106S,2014JGRA..119...69S, 2013JGRA..118.4748L, 2014SoPh..289..951L,2015A&A...584A..83T,2017PNAS..114.1502C,2019JGRA..124.1475H}.
The standard theory has been widely used to explain radio bursts in space and astrophysical plasmas, such as solar radio bursts in terms of type I-V \citep[see, e.g.,][for latest observational studies]{2014ApJ...787...59C,2016ApJ...830L...2V,2019ApJ...870...30V,2017SoPh..292...82L,2017SoPh..292..194L}, emissions from planetary electron foreshocks \citep[e.g.,][]{1984JGR....89.6631E,1984GeoRL..11..869M,1987JPlPh..38..169C,2005JGRA..110.7107K,2017EPSC...11..585P} and the outer heliosphere \citep{1984Natur.312...27K,1994JGR....9914729Z,1995AdSpR..16i.279G,1998GeoRL..25.4433G,1995GeoRL..22.3433C,1998ApJ...506..456C,2002GeoRL..29.1143C,2022ApJ...928..125T}. The theory involves evolution of turbulence and complicated multi-step nonlinear processes. A proper numerical verification of the theory requires kinetic simulations by the nature of the problem. Both ion-scale and electron-scale kinetic physics should be included. This excludes any fluid/hybrid models. Only fully-kinetic methods can be used, including the Vlasov method and the PIC method.
The Vlasov method directly solves the Vlasov-Maxwell equations to resolve the velocity distribution function of particles. The weak-turbulence (WT) theory/simulation represents a variety of this method tailored for the study of plasma emission \citep[e.g.,][see \citet{2019ApJ...871...74L} for the complete list of references]{2006PhPl...13b2302Y,2012PhPl...19j2303Y,2015ApJ...806..237Z}.
The PIC simulations numerically resolve the motion of each particle and the evolution of the electro-magnetic fields by solving self-consistently the equation of particle motion and the full-set of the Maxwell equations. In comparison to the WT method, the PIC method has the advantage of making no approximations to the basic laws of mechanics and electro-magnetism. We therefore choose the PIC method for the present study. See \citet{2019ApJ...871...74L} for a detailed comparison of the two methods when being applied to the problem of plasma emission. Earlier studies along this line of research include \citet{2001JGR...10618693K}, \citet{2007P&SS...55.2336K}, \citet{2009ApJ...694..618R,2009JKPS...54..313R}, \citet{2010JGRA..115.1204U}, \citet{2012ApJ...751..145G,2012SoPh..280..551G}, and latest studies include \citet{2015A&A...584A..83T}, \citet{2019JGRA..124.1475H}, \citet{2020ApJ...891L..25N}, and \citet{2022ApJ...924L..34C}. The PIC simulations demand massive computing and turns out to be challenging as
elaborated below.
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item The number of macroparticles per cell per species (NPPCPS) should be large enough to lower the levels of numerical noise. Most earlier simulations have adopted values of NPPCPS less than a few hundred \citep[e.g.,][]{2001JGR...10618693K,2010JGRA..115.1204U} for economic reason. This results in relatively strong noise and low signal-noise ratio of the obtained F/H emissions since that are intrinsically weak. The situation becomes worse for weaker beam.
\item The simulation domain should be large enough to have a good resolution of wave number ($k$). This is necessary for the F emission whose frequency upon excitation is close to the cutoff ($\sim \wpe$). Since according to the standard theory of PE the F emission is given by the scattering of the Langmuir (L) wave over the low-frequency ion-acoustic (IA) mode, thus its frequency is the sum or difference of the frequencies of the L and IA modes. The IA frequency is negligible compared to the Langmuir frequency ($\sim \wpe$), the frequency of the F emission is therefore very close to $\wpe$, which is the cutoff frequency of the fast electro-magnetic mode propagating in unmagnetized plasmas. It is therefore characterized by a long wavelength ($\lambda$) and small $k$. According to the simplified dispersion relation of electromagnetic wave ($\omega^2 = \wpe^2 + k^2 c^2$), with $kc \le 0.2~\wpe$ one gets $\lambda \ge 1500~\lde$ (the electron Debye length) for coronal plasmas at 1--2 MK (with a thermal speed of $\sim$ 0.02 c). The minimum dimension of the domain should be larger than a few wavelengths for proper simulation \citep[see, e.g.,][]{2019JGRA..124.1475H}.
\item The temporal length of data used for the Fourier spectral analysis should be long enough to resolve modes with close frequencies. This is important since the F emission has almost the same frequency as the Langmuir mode. For instance, if the data length is taken to be $\sim$ 100 $\wpe^{-1}$, the two modes cannot be separated when their frequency difference is less than $0.1~\wpe$.
\end{enumerate}
The above elaborations point out the need of massive computation, while most earlier simulations did use insufficient number of macro-particles and limited domain. See the critical review given by \citet[][referred to as TT15 hereinafter]{2015A&A...584A..83T}. Note the domain is regarded as limited here if it is comparable to or only a few (say, $\sim $ 2--3) times larger than the the wavelength of the F emission. These limitations cause excessive numerical noise along dispersion curves and limited resolution of wave modes. The conclusions drawn are thus controversial regarding the generation of the F emission. For example, TT15 claimed that they presented ``the first self-consistent demonstration of F and H emission from a single-beam (unmagnetized) plasma system via fully kinetic PIC simulations'' --- in accordance with the standard theory of plasma emission. \citet{2019JGRA..124.1475H}, however, found that the F emission is hardly discernible according to their simulations with domain and NPPCPS being much larger than that of TT15. Thus, it remains open regarding whether the F emission does occur according to the standard theory (from the perspective of PIC verification), despite the positive statement of TT15 which is nevertheless inconclusive according to the following arguments.
\begin{enumerate}
\renewcommand{\labelenumi}{(\theenumi)}
\item The domain dimension of TT15 is 600 $\lde$, much less than the expected wavelength of the F emission. According to TT15, the expected $k \lde$ of the F emission is about 0.002, corresponding to a wavelength of a few thousands of $\lde$.
\item The duration of data for the Fourier analysis is 100 $\wpe^{-1}$, the corresponding spectral resolution is less than $0.1~\wpe$, too poor to separate the F emission from nearby Langmuir waves with the obtained $\omega$--$k$ diagram.
\item According to TT15, one evidence of the F emission is the growth of transverse component of magnetic field (see Figure 5 in TT15). Yet, such signal may have two contributions, according to our study presented here. One is given by the expected F emission, the other is by the Langmuir turbulence. The latter is found with $k \lde$ varying in a range of [0, 0.02], much broader than the expected $k \lde$ range of the F emission. Such signal is dominated by the Langmuir turbulence rather than by the F emission. This point cannot be inferred from earlier studies due to poor resolutions of both $\omega$ and $k$.
\end{enumerate}
Further, according to the dispersion relation of the fast electro-magnetic wave propagating in unmagnetized plasmas and the Faraday$'$s law, the relative strength of the wave E and B fields is determined by the ratio of $\omega^2 / k^2 c^2$. Thus, for the F emission with frequency ($\omega$) very close to $\wpe$, the E-field is much stronger than the B-field, so the F emission is dominated by the E-field. It is therefore critical to examine the E-field data to reveal its spectral characteristics and energy budget, to understand its generation mechanism.
In this study, we conduct the fully-kinetic PIC simulations in a domain much larger than that of TT15 to increase the spectral resolution, and with large value of NPPCPS to weaken the numerical noise. The main purpose is to verify the standard theory of plasma emission in unmagnetized plasmas energized by a single beam of energetic electrons. Signatures of the F emission and its generation process are highlighted to clarify existing contradiction in literature.
We first present the simulation method and parameter setup in the following section, together with the result for the thermal case, then present the cases with different number density ratio of beam and background electrons ($n_\mathrm{b}/n_\mathrm{0}$) in Sections 3 and 4. We adopt the same plasma and beam parameters for Case A as those in TT15 for a direct comparison. Cases with two values of $n_\mathrm{b}/n_\mathrm{0}$ (0.0057 and 0.05) have been modeled in TT15, while here an additional case with $n_\mathrm{b}/n_\mathrm{0} = 0.01$ is included to extend the analysis. The last section presents the conclusions and discussion.
\section{Parameter Setup and the Thermal Case}
Following our earlier studies of coherent emission in magnetized plasmas \citep[e.g.,][]{2020ApJ...891L..25N,2021ApJ...909L...5L,2021ApJ...909....3Y,2021ApJ...920L..40N,2022ApJ...924L..34C}, we continue to use the Vector-PIC \citep[VPIC:][]{5222734,2008PhPl...15e5703B,Bowers_2009} code released by the Los Alamos National Laboratory (LANL), which are 2D3V, i.e., two-dimensional in space and three-dimensional for particle velocity and electromagnetic fields. We present three cases (A, B, and C) with different $n_\mathrm{b}/n_\mathrm{0}$ (= 0.0057, 0.01, and 0.05) within simulation domain as large as $\sim$ 6000 $\lde$ $\times$ 6000 $\lde$. The cell size is set to be 2.929 $\lde$ and the number of cells to be 2048 along both $\hat{e}_x$ and $\hat{e}_z$ direction, where $\hat{e}_x$ ($\hat{e}_z$) represents the unit vector along the $x$ ($z$) direction. The beam is along $\hat{e}_z$, and $\vec k$ lies in the $xOz$ plane. The background electron-proton plasma is assumed to be unmagnetized and thermal, to compare directly with earlier studies. The time step is set to be 0.03 $\wpe^{-1}$. The NPPCPS are 4000 for background electrons, and 1000 for both background protons and beam electrons. The total number of macroparticles is 2.4 $\times~10^{10}$. In Table A1 of the Appendix, we have presented some setup parameters of several PIC studies on plasma emission for comparison. According to the table, the present study has almost the same size of simulation domain as that used by its sister paper \citep{2022ApJ...924L..34C} and the largest NPPCPS and total number of macroparticles among the listed studies.
The details of the simulation setup with the domain dimension, cell size, NPPCPS, resolvable ranges of wavenumber ($k$) and frequency ($\omega$) have been listed in Table 1. Parameters used by TT15 are also included. To illustrate the effect of the domain size, we include another case (Case A$^\prime$) within a smaller domain (1200 $\times$ 1200 $\lde^2$). Other parameters of Case A$^\prime$ are the same as Case A.
The beam electrons are represented with the following drifting Maxwellian distribution function
\begin{equation}
f_\mathrm{b} = A_\mathrm{b} \exp(-\frac{u^2_\perp}{2u^2_{0}} - \frac{(u_\parallel - u_\mathrm{d})^2}{2u^2_{0}})
\end{equation}
where $u_\parallel$ and $u_\perp$ are the parallel and perpendicular components of the momentum per mass, $u_d$ is the average drift momentum per mass of the beam electrons, $u_0$ is the thermal velocity of energetic electrons, and $A_\mathrm{b}$ is the normalization factor. The parameters for the background plasmas and the beam are taken from TT15, which are 2 MK ($\sim 0.02c$) for $T_\mathrm{e}$, $T_\mathrm{p} = 0.7~T_\mathrm{e}$, and 0.3c for the average drift speed of the beam and the beam temperature is set to be $T_\mathrm{e}$.
To tell the significance of plasma emission, it is essential to compare their spectral intensity with that given by the thermal case (Case T), in which wave-like noises exist along dispersion curves of intrinsic wave modes even without any energetic electrons. Other setup parameters of Case T are the same as those of Cases A--C. The modes attributed to energetic electrons must be considerably stronger than the corresponding numerical noise.
We start from analyzing Case T. See Figure A1 in the Appendix for the obtained $\omega$--$k$ diagram for Case T. To show the effect of NPPCPS we aslo present in Figure A1 another thermal case (T$^\prime$) with NPPCPS = 500 (for electrons) and 125 (for protons). The diagrams are along two propagating directions with $\theta = 30^\circ$ and $\theta = 80^\circ$. The color bar of the lower panels is 20 dB less than that of the upper two panels to show out weaker signals. Range of the normalized wavenumber ($k \lde$) for the uppermost panel is [$-$0.3, 0.3], while it is [$-$0.03, 0.03] for the lower three panels. Two conclusions can be drawn: (1) the numerical noise level is inverse-proportional to the NPPCPS, i.e., it is $\sim 9$ dB stronger in Case T$^\prime$ than in Case T; (2) there exist two high-frequency modes in unmagnetized plasmas (as expected), the electrostatic Langumuir mode and the fast electromagnetic mode. The Langmuir mode (the upper two panels) extends to frequencies above $\wpe$ due to the thermal effect, which is well-described by the following relation,
\begin{equation}
\omega^2 = \wpe^2 (1 + 3 k^2 \lde^2)
\end{equation}
The Langmuir wave exists continuously from superluminal to subluminal regimes, being perfectly electrostatic. The transverse mode is also dominated in energy by the $E_y$ and $E_z$ components that are stronger by about 10--20 times than the magnetic counterparts ($B_x$ and $B_y$). For Case T, the maximum intensity of $E_y$ is about -140 dB and $B_y$ about -150 dB, as read from the figure. These values represent the levels of modes relevant to numerical noises.
\section{Simulation Results for Case A and Comparison With TT15}
Figure 1 and the accompanying movie present the temporal evolution of the velocity distribution function (VDF) of electrons for Case A with $n_\mathrm{b}/n_\mathrm{0} = 0.0057$. Due to the rapid growth of the bump-on-tail instability, the beam electrons decelerate rapidly and diffuse towards the regime with lower $v_\parallel$. This occurs during the first few hundred $\wpe^{-1}$. The beam electrons also diffuse towards the regime with larger $v_\perp$, indicating efficient perpendicular heating. Such heating is induced by two processes, the perpendicular diffusion due to the bump-on-tail instability \citep[see][]{2020PhPl...27b0702H,2021SoPh..296...42M,2022ApJ...924L..34C} and the electromagnetic Weibel instability that was discovered by \citet{1959PhRvL...2...83W} for anisotropic plasmas with bi-Maxwellian distribution and by \citet{1959PhFl....2..337F} for plasmas with counter-streaming electrons. In our case, the instability is driven by the single electron beam, leading to the growth of the electromagnetic beam mode with significantly-enhanced $B_y$ component and perpendicular heating of the beam electrons \citep[see also][TT15]{2009ApJ...690..189K}. The VDF reaches the asymptotic state after $\sim$ 1000 $\wpe^{-1}$.
In Figure 2a, we show energy curves of the six field components and the negative change of the total kinetic energy of all electrons ($-\Delta E_k$) for Cases T and A. The beam-plasma interaction causes rapid rise of $E_z, E_x$, and $B_y$ within the first 80 $\wpe^{-1}$, while the other field components remain at the background levels. The maximum intensity of $E_z$ reaches about 13$\%$ of $-\Delta E_k$, while those of $E_x$ and $B_y$ reach about 2$\%$ and 0.01$\%$ of $-\Delta E_k$, respectively. As seen from the following analysis, the $E_x$ and $E_z$ components mainly correspond to the Langmuir mode while the $B_y$ component is energized by the low-frequency electromagnetic Weibel instability. The $E_x$ and $E_z$ intensities maintain an almost constant level before declining gradually after $\sim$ 1000 $\wpe^{-1}$, representative of the nonlinear evolution of turbulent plasmas in quasi-equilibrium.
In Figure 2b we have plotted the energy of field components associated with various wave modes. The temporal variation of mode energy is given by the inverse Fourier transform of the total spectral energy within the selected spectral regimes (see Figure A2). Among various wave modes, the strongest one is the BL mode given by the coupling of the beam mode and the intrinsic Langmuir mode. Its $k$ range is [0.08, 0.2] $\lambda_\mathrm{De}^{-1}$, and frequency range is [0.9, 1.3] $\wpe$. It propagates from $\theta = 0^\circ$ (i.e., along the beam) to $\theta = 60^\circ$. The energy curve of $E_z$ (Figure 2a) agrees with the Langmuir mode energy (Figure 2b).
We performed the exponential fittings to the energy profiles of the $E_z$ field and the beam-Langmuir (BL) mode (see the dashed lines overplotted in Figure 2). The obtained linear growth rate is 0.076 $\wpe^{-1}$ for $E_z$, and 0.053 $\wpe$ for the BL mode energy. The theoretical growth rate $\gamma$ and the corresponding frequency $\omega_\mathrm{r}$ versus $k \lde$, according to Equation (1) of the Appendix, has been plotted in the upper panels of Figure A3. The parameters used for the plots are identical to those of Case A. In Figure A3a, the maximum growth rate of the bump-on-tail instability ($\gamma_\mathrm{M}$) is 0.094 $\wpe$ at ($\omega_\mathrm{r}$, $k\lde$) = (0.96 $\wpe$, 0.069), and the range of significant growth (within which the growth rate is larger than, say, $\omega_\mathrm{rM}/2$) extends from $(\omega_\mathrm{r}$, $k\lde$) = (0.75 $\wpe$, 0.049) to (1.06 $\wpe$, 0.084). On the other hand, according to the PIC simulation (see Figure A3b for the dispersion diagram of Case A within the period of [0, 500 $\omega_\mathrm{pe}^{-1}$]), the BL mode presents the maximum growth around $(\omega_\mathrm{r}$, $k\lde$) = (0.98 $\omega_\mathrm{pe}$, 0.08), and the range of significant growth extends from $(\omega_\mathrm{r}$, $k\lde$) = (0.98 $\wpe$, 0.075) to (1.01 $\wpe$, 0.12). We concluded that the PIC simulation agrees reasonably well with the theoretical prediction.
Figure 3 presents the wave map in the $\vec k$ space. Figure 4 presents the corresponding $\omega$--$k$ diagrams along $\theta = 30^\circ$ and/or $80^\circ$ for the modes with frequencies close to $\wpe$, and Figure 5 presents the similar diagrams for the modes at the low frequency ($\ll \wpe$) and the harmonic frequency ($\sim 2 \wpe$). These figures (and accompanying movies) shall be combined to reveal the characteristics of each mode. According to Figures 3 and 4, there exist two other Langmuir components on the left (or backward) side of BL, referred to as the generalized Langmuir (GL) mode in total. One is forward-propagating with a narrow range of k ([-0.04, 0.04] $\lambda_\mathrm{De}^{-1}$), the other is backward-propagating with a larger range of k ([-0.2, -0.04] $\lambda_\mathrm{De}^{-1}$). The forward one is mostly superluminal. According to the movie accompanying Figure 3 and the energy curves of Figure 2, the backward one appears shortly after the onset of BL, while the forward one appears after 100 $\wpe^{-1}$. The two components have very similar energy profiles, both reaching the maximum around 1000 $\wpe^{-1}$, indicating the same physical origin.
From panels (b)--(d) of Figure 4, we observe significant enhancements along the corresponding dispersion curve of the F emission, i.e., the fast transverse electromagnetic mode around $\wpe$, from the $\omega$--$k$ diagrams of $E_z$ and $B_y$. According to Figure 2b, its energy (summed over all propagating directions) reaches about $10^{-5}~E_{k0}$, a fraction of the total energy of the H emission. As expected, the F emission is dominated in energy by its $E_z$ component, which is stronger than $B_y$ by about 2--3 orders in magnitude. It reaches the maximum intensity around 1000 $\wpe^{-1}$, with a variation trend similar to the H emission.
The harmonic (H) emission emerges as the circular feature after $\sim$ 500 $\wpe^{-1}$ in the maps of $E_z$, $E_x$, and $B_y$ (see Figures 3c and 5c, and accompanying movies), and saturates at the level of $10^{-5}~E_{k0}$ after 1000 $\wpe^{-1}$. Within the H circle, there appears enhanced quadrupolar $B_y$ feature which is mainly generated by the beam-driven electromagnetic Weibel instability at low frequency (see Figure 5b, see also TT15). The $B_y$ signal rises rapidly during the initial stage of the plasma-beam interaction, reaching the maximum intensity around 100 $\wpe^{-1}$.
Figure 5a presents the ion-acoustic (IA) mode that is very weak due to the strong Landau damping in plasmas with $T_\mathrm{p}/T_\mathrm{e}= 0.7$. See the lower two panels of Figure A3 in the Appendix for the frequency and growth rate of the IA mode within the background Maxwellian plasmas, evaluated according to Equation (2) in the Appendix.
Another interesting feature is the weak yet significant $B_y$ enhancement along the Langmuir dispersion curve (see Figure 4c). It occupies a much larger range of $k$ ([-0.04, 0.04] $\lambda_\mathrm{De}^{-1}$), with the spectral intensity weaker than the F emission. Yet, the total energy of the Langmuir $B_y$ component is much larger than that of the F emission. It reaches the maximum level ($\sim 10^{-6}~E_{k0}$) also around $1000~\wpe^{-1}$ (see Figure 2c), $\sim 10$ times stronger than the simultaneous F emission. Similar $B_y$ enhancement is also observed from Figure 5 of TT15. Yet, due to the poor spectral resolution there it is not possible to separate the $B_y$ component of the F emission from the prevailing Langmuir enhancement. Such $B_y$ enhancement is not associated with the Langmuir noise of thermal plasmas according to Figure A1.
To support the above argument, we conducted another simulation (Case A$^\prime$) within a much-smaller domain (1200 $\times$ 1200 $\lde^2$). See Table 1 for the simulation setup. The plasma-beam parameters are not changed. The obtained wave map in the $\vec k$ space is shown in Figure A4 of the Appendix for large (panel a) and small (panels b and c) range of $k$. The spectra of the three magnetic field components are presented in panel c. The obtained $\omega$--$k$ diagrams are presented in Figures A5 of the Appendix, along two propagating angles with $\theta = 30^\circ$ and $\theta = 80^\circ$. The accompanying movie presents the complete dispersion diagrams from $\theta = 0^\circ$ to $\theta = 90^\circ$. For direct comparison, the data duration used in the Fourier analysis is taken to be 100 $\wpe^{-1}$, the same as that used by TT15. We see that the circular H emission cannot be identified clearly, and there exist no regular wave enhancements along the dispersion curve of the F emission, no way to tell the significance of its $B_y$ field due to the presence of the overwhelming Langmuir $B_y$ component. Note that TT15 has mistakenly regarded the total $B_y$ enhancement to be the evidence of the F emission.
\section{Simulation Results for Cases B and C and Comparison with Case A}
Figures 6 and 7 present the wave distribution map in the $\vec k$ space and the $\omega$--$k$ diagram for Cases B ($n_\mathrm{b}/n_\mathrm{0} = 0.01$) and C ($n_\mathrm{b}/n_\mathrm{0} = 0.05$). The energy curves for Case B have been plotted in Figures 2c--2d. They should be combined with the results for Case A to understand the effect of varying $n_\mathrm{b}/n_\mathrm{0}$. Note that the illustrated spectral energy is `relative' since it has been normalized by the total initial beam energy ($E_{k0}$) of each individual case. $E_{k0}$ increases with increasing $n_\mathrm{b}/n_\mathrm{0}$, being $4.12~m_\mathrm{e} c^2$ for A, $7.25~m_\mathrm{e} c^2$ for B, and $37.7~m_\mathrm{e} c^2$ for C.
The most obvious result is that all modes discussed here, except the low-frequency electromagnetic Weibel mode, decrease in relative spectral energy with increasing $n_\mathrm{b}/n_\mathrm{0}$. All modes look very similar when $n_\mathrm{b}/n_\mathrm{0}$ increases from 0.0057 to 0.01, yet, when $n_\mathrm{b}/n_\mathrm{0}$ further increases to 0.05, only the forward BL mode together with a small part of the forward propagating GL wave remain while the backward one and the F/H emissions vanish. The Weibel instability responds to increasing $n_\mathrm{b}/n_\mathrm{0}$ differently, yielding stronger $B_y$ in terms of relative spectral energy. Further discussion on the Weibel instability is beyond the scope of the study.
As observed from Figure 7, the central frequency of BL in Case C is about 0.86 $\wpe$, being considerably lower than in Cases A and B. The difference from $\wpe$ is too large to allow the standard plasma emission process to occur. This agrees with TT15 that is based on simulations within a much smaller domain.
\section{Conclusions and Discussion}
Our main purpose is to verify the standard model of plasma emission which describes a multi-stage nonlinear process in the aftermath of the kinetic bump-on-tail instability of a single-beam plasma system. The primary outcome of the instability is the enhanced turbulence of electrostatic beam-Langmuir (BL) mode, which further interacts with other secondary fluctuations to yield the F and H radiations. Previous verification studies using PIC simulations, limited by available computational resources, have drawn contradictory statements regarding the significance of the F emission. In this study we employed a simulation domain and number of macroparticles that are among the largest ones of similar studies, to lower noise levels and achieve higher spectral resolution.
We found that for the ratio of number density ($n_\mathrm{b}/n_\mathrm{0}$) being less than 0.01, significant F emission can be generated. The F emission reaches an energy level of 10$^{-5}$ of the total initial beam energy $E_{k0}$ that is a fraction of the total energy of the H emission. The spectral energies of all wave modes relative to $E_{k0}$, except the beam-driven electromagnetic Weibel instability, decrease with increasing $n_\mathrm{b}/n_\mathrm{0}$. In agreement with TT15, for $n_\mathrm{b}/n_\mathrm{0} \approx 0.05$ the frequency of the BL mode is too low to meet the matching conditions of three-wave interaction, thus the backward-propagating Langmuir wave and the F and H emissions cannot be generated.
According to the simulation result, the IA mode is indeed extremely weak in plasmas with $T_\mathrm{p}/T_\mathrm{e}= 0.7$ due to the strong Landau damping effect (see Figure A3). Thus, the most promising process is the `d' or the decay process (BL $\rightarrow$ F + IA), rather than the `u' process (BL + IA $\rightarrow$ F). Due to the thermal effect, the maximum frequency of the BL mode is about 1.04 $\wpe$, exceeding the cutoff of the F emission by 0.04 $\wpe$. This also supports the occurrence of the `d' process, indicating that the thermal correction to the dispersion relation of Langmuir wave is critical to the emission process.
In the case with $n_\mathrm{b}/n_\mathrm{0} \le 0.01$, we found significant enhancement of the transverse component $B_y$ at frequency around $\wpe$ and in a $k$ range of [-0.02, 0.02] $\lambda_\mathrm{De}^{-1}$, with contributions from both the F emission and the beam-Langmuir turbulence. It remains unresolved regarding the generation mechanism of such Langmuir $B_y$ component (that is about three orders in magnitude weaker than the corresponding $E_z$ component). Note that in the simulation with only thermal plasmas the Langmuir noise manifests no signatures of $B_y$ (see Figure A1). Future study should explore how such $B_y$ component develops in such plasmas of turbulent equilibrium.
Using the same PIC code, \citet{2022ApJ...924L..34C} examined the plasma emission process in weakly magnetized plasmas ($\wpe/\Omega_\mathrm{ce} = 10$) interacting with a single beam of energetic electrons. They presented cases with arbitrarily different mass ratio ($m_\mathrm{p}/m_\mathrm{e}$) and found that with increasing $m_\mathrm{p}/m_\mathrm{e}$ the intensity of BL increases correspondingly, together with enhancements of Z mode and F/H emissions. This indicates that the latter modes (Z and F/H) are secondary products of the primary BL mode. This agrees with the study presented here.
Comparing the case for weakly-magnetized background plasmas \citep{2022ApJ...924L..34C} and Cases A and B presented here for unmagnetized plasmas, we find that the two sets of solution are quite similar. In particular, they are similar in the overall temporal evolution of the EVDF, the growth and characteristics of the beam-Langmuir wave, and backward and forward-propagating secondary generalized-Langmuir waves. The main difference between the two cases lies in the generation of the superluminal Z mode (with cutoff frequency below $\wpe$) and the whistler mode in \citet{2022ApJ...924L..34C} while in this study there present the generalized Langmuir wave and the electromagnetic mode induced by the beam-Weibel instability. These major similarities support the standard theory of plasma emission works for the two cases, that is, the electromagnetic decay to generate the F emission, and nonlinear coalescence of the BL wave with the backward-propagating Langmuir wave to generate the H emission.
The above argument does not support the plasma emission mechanism suggested by \citet{2020ApJ...891L..25N,2021PhPl...28d0701N} that the F emission is generated by almost-counterpropagating Z and W modes and the H emission by almost-counterpropagating upper-hybrid (UH) mode. Their conclusion is based on PIC simulations of plasmas interacting with trapped energetic electrons with the loss-cone type VDFs. There the primary mode is the UH mode which is the Langmuir wave with a large propagating angle relative to the background magnetic field. The UH mode and the secondary Z and W modes are excited directly through the electron cyclotron maser instability (ECMI). This suggests that the mechanism of plasma emission may be different for different VDFs of energetic electrons: for beam electrons the standard plasma emission mechanism is at work, while for trapped electrons the plasma emission induced by ECMI may be important.
\begin{acknowledgments}
This study is supported by NNSFC grants (11790303 (11790300), 11973031, and 11873036). The authors acknowledge the Beijing Super Cloud Computing Center (BSCC, URL: \url{http://www.blsc.cn/}) for computational resources, and LANL for the open-source VPIC code.
\end{acknowledgments}
|
\section*{Acknowledgments}
This material is based upon work supported by the U.S. Department of Energy (DOE), Office of Science, under contract number DE-AC02-06CH11357. This work is also supported by the U.S. Department of Energy, Office of Advanced Scientific Computing Research, Scientific Discovery through Advanced Computing (SciDAC) program. This work is also in part supported by NSF IIS-1910733 and DOE DE-SC0021015.
\input{ml-robustness-refs.bbl}
\end{document}
\section{Technical Background}
\label{sec:background}
We review the classic notion of robustness and the critical point tracking method by Guo {et al.}~\cite{GuoLenzXu2021}, referred to as the FTK algorithm in this paper.
\subsection{Robustness}
\label{sec:robustness}
\para{Degrees of critical points.}
Consider a continuous vector field $f: \Xspace \subseteq \Rspace^2 \to \Rspace^2$ defined on a 2D domain $\Xspace$. A \emph{critical point} $x \in \Xspace$ is an isolated zero in the vector field, that is, $|f(x)| = 0$.
A critical point $x$ in 2D can be classified with respect to its \emph{degree}, denoted as $\mm{\mathrm{deg}}(x)$, as the number of field rotations while traveling along a closed curve counterclockwise surrounding $x$ enclosing no other critical point.
In 2D, a saddle point has degree $-1$, whereas a source/sink/center has degree $+1$.
A connected component $C \subseteq \Xspace$ that contains $n$ critical points $\{x_1, x_2, \cdots, x_n\}$ has a degree that is the sum of the degrees of $x_i$, $\mm{\mathrm{deg}}(C)=\sum_{i=1}^n\mm{\mathrm{deg}}(x_i)$; see~\cite[page 134]{Hatcher2002} for a formal investigation of the degree of a continuous mapping. As illustrated in a 2D vector field in~\autoref{fig:example-mt}(A), $x_1$ and $x_3$ are centers with $+1$ degree, and $x_3$ and $x_4$ are saddles with $-1$ degree.
\begin{figure}[!t]
\centering
\includegraphics[width=0.98\columnwidth]{exampleMT.pdf}
\vspace{-2mm}
\caption{An augmented merge tree generated from a 2D vector field: (A) a continuous 2D vector field $f$; (B) relations among connected components of sublevel sets $\Xspace_r$; and (C) the augmented merge tree. Sources/sinks/centers are in red, saddles in blue.}
\label{fig:example-mt}
\vspace{-2mm}
\end{figure}
\para{Merge tree.}
The computation of robustness relies on the notion of an augmented merge tree.
Given a continuous 2D vector field $f:\Xspace \to \Rspace^2$, we can define a scalar field $f_0:\Xspace \to \Rspace$ that assigns the vector magnitude to each point $x \in \Xspace$, that is, $f_0(x) = ||f(x)||_2$.
Let $\Xspace_r=f_0^{-1} (-\infty, r]$ denote the \emph{sublevel set} of $f_0$ for some $r \geq 0$. $\Xspace_0$ is precisely the set of critical points of $f$.
In~\autoref{fig:example-mt}(B), $f_0$ is visualized using an orange to purple colormap, and certain sublevel sets $\Xspace_r$ are shown as colored curves.
We can construct a merge tree of $f_0$ that tracks the evolution of connected components in $\Xspace_r$ as $r$ increases.
Specifically, leaves in a merge tree represent the creation of a component at a local minimum of $f_0$, internal nodes represent the merging of components, and the root represents the entire space as a single component; see~\autoref{fig:example-mt}(C) for an example. Once the merge tree is constructed, it can be further augmented with the degrees of critical points (on leaves), and the degrees of components (on internal nodes).
As shown in~\autoref{fig:example-mt}(C), we use $C_i$ to represent the connected components of the sublevel sets of $\Xspace_r$ for some $r$, and augment the corresponding merge tree with $\deg(x_i)$ and $\deg(C_i)$ as attributes of the tree nodes.
For example, $C_1$ is one of the three components of $\Xspace_{r_1}$, which contains critical points $x_1$ and $x_2$.
Therefore, we have $\mm{\mathrm{deg}}(C_1)=\mm{\mathrm{deg}}(x_1)+\mm{\mathrm{deg}}(x_2)=0$.
\para{Robustness.}
The \emph{robustness} of a critical point is the function value of its lowest zero degree ancestor in the merge tree~\cite{WangRosenPrimoz2013}.
For the example in~\autoref{fig:example-mt}, the robustness of $x_1$ and $x_2$ is $r_1$ and the robustness of $x_3$ and $x_4$ is $r_3$, respectively.
We review some properties of robustness here for completeness; see~\cite{WangRosenPrimoz2013} for details.
Let us first define the concept of vector field \emph{perturbation}. A continuous mapping $h: \Xspace \to \Rspace^2$ is an \emph{$r$-perturbation} of $f$, if $d(f, h) \leq r$, where $d(f, h)=\sup_{x\in \Xspace}||f(x)-h(x)||_2$, where $\sup$ means supremum. Suppose a critical point $x$ of $f$ has robustness $r$, then we have:
\begin{lemma}[Critical Point Cancellation\cite{WangRosenPrimoz2013}]
\label{lemma:cancel}
Let $C$ be the connected component of $\Xspace_{r+\eta}$ containing $x$, for an arbitrarily small $\eta > 0$. Then, there exists an $(r+\eta)$-perturbation $h$ of $f$, such that $h^{-1}(0)\bigcap C=\emptyset$ and $h=f$ except possibly within the interior of $C$.
\end{lemma}
\begin{lemma}[Degree Preservation~\cite{WangRosenPrimoz2013}]
\label{lemma:degree}
Let $C$ be the connected component of $\Xspace_{r-\eta}$ containing $x$, for some $0<\eta < r$. For any $\varepsilon$-perturbation $h$ of $f$, where $\varepsilon<r-\eta$, $\mm{\mathrm{deg}}(h^{-1}(0)\bigcap C)= \mm{\mathrm{deg}}(C)$. If $C$ contains only one critical point $x$, $\mm{\mathrm{deg}}(h^{-1}(0)\bigcap C)=\mm{\mathrm{deg}}(x)$.
\end{lemma}
These two lemmas imply that the topological notion of robustness quantifies the stability of a critical point with respect to perturbations of the vector fields.
Intuitively, Lemma~\autoref{lemma:cancel} implies that a critical point $x$ with a robustness of $r$ may be canceled with a $(r+\eta)$-perturbation, for arbitrarily small $\eta>0$.
Lemma~\autoref{lemma:degree} states that $x$ may not be canceled with a $(r-\eta)$-perturbation.
\para{Limitations in computing the classic robustness.}
In practice, the robustness of a critical point depends on its cancellation partner(s) defined by the merge tree, whose locations may be influenced by the boundary condition of the known data domain.
To compute the classic notion of robustness, we use the known data domain to construct a single merge tree, as shown in~\autoref{fig:example-mt}(C).
If the domain is without boundary, we expect all critical points to have cancellation partners and all the robustness values to be finite (however, there is a technical condition on the domain for the algorithm to work, {i.e.}, trivial tangent bundle, which excludes the sphere).
If the domain has a boundary, a critical point may be canceled with a potentially far away critical point, based on the merge tree construction.
For example, $x_1$ has a partner $x_2$ in~\autoref{fig:example-mt}(B); however, in the cropped region (a), $x_1$ has a new partner $x_4$ since $x_4$ is the only candidate in (a) that may be canceled with $x_1$.
Furthermore, a critical point may have an infinite robustness value if it does not have a cancellation partner in the known data domain.
For example, $x_3$ has a partner in the original domain of ~\autoref{fig:example-mt}(B); however, it loses its partner in the cropped region (b).
These cases happen when the sublevel sets intersect the boundary of the domain where we have an incomplete picture of the flow behavior closer to the boundary.
We aim to mitigate some of these boundary effects by introducing the notion of multilevel robustness (see~\autoref{sec:ml-robustness}).
\subsection{Critical Point Trajectories}
Critical point tracking algorithms take a time-varying vector field as the input, and produce as the output 1D geometries that represent the trajectories of critical points in spacetime.
In general, our multilevel robustness framework may be used to enhance any critical point tracking result; we choose to use the recent FTK algorithm by Guo {et al.}~\cite{GuoLenzXu2021} for its simplicity and performance.
\begin{figure*}
\centering
\includegraphics[width=2.1\columnwidth]{MR.pdf}
\vspace{-4mm}
\caption{Classic robustness analysis of an instance from the $\ME$ dataset.
(A) A 2D vector field $f$ where critical point partners share the same color.
(B) Multiple sublevel sets of $f_0$, where critical points are colored by their degree ($+1$ in blue and $-1$ in red), and the size of each critical points is shown to be proportional to their robustness values.
(C) Part of an augmented merge tree of $f_0$ used in the robustness computation.
(D) Sublevel set $\Xspace_{0.13}$ in orange.
(E) Sublevel set $\Xspace_{0.43}$ in pink.}
\label{fig:classic-robustness}
\vspace{-2mm}
\end{figure*}
\para{Trajectories.}
Let $\hat{f}: \hat{\Xspace} = \Xspace \times \Rspace \to \Rspace^2$ denote a time-varying vector field over a 2D domain $\Xspace$, where $\hat{f}_t(x) = \hat{f}(x,t): \Xspace \to \Rspace^2$ represents a 2D vector field at time $t \in \Rspace$.
We define \emph{critical point trajectories} (or simply \emph{trajectories}) as the $0$-levelset of $\hat{f}$, $\hat{\Xspace}_0 := \hat{f}^{-1}(0,0)$, that is, the vicinity where both $x$- and $y$- components of $\hat{f}$ are $0$ and thus is the intersection of two isosurfaces of both vector components.
\para{Piecewise linear assumption.} The basic assumption of the tracking method is that $\hat{f}$ is \emph{piecewise linear in spacetime}. That is, $\hat{\Xspace}$ is a 3D simplicial complex consisting of a set of spacetime tetrahedra $\{T_i\}$ such that
$\hat{f}(x) = a_i x + b_i, x\in T_i\subset \hat{\Xspace}$,
where $a_i$ and $b_i$ are constants for each tetrahedron $T_i$, and $\hat{f}$ is $0$-continuous on cell boundaries.
If the linear system $\hat{f}=0$ in $T_i$ is nondegenerate, the $0$-levelset of $\hat{f}$ in $T_i$ may be analytically solved as a linear curve; otherwise, degenerate cases may be handled with the simulation of simplicity~\cite{EdelsbrunnerMucke1990}.
Therefore, trajectories can be extracted as 1D piecewise linear curves in 3D spacetime; see~\cite{GuoLenzXu2021} for details on the construction of spacetime simplicial complexes, handling of degenerates, and extraction of trajectories.
\para{Interpretation of trajectories.} Note that the $0$-levelset of $\hat{f}$ is not a bijection of time $t$ onto the trajectory; one may observe non-monotonous time along the same trajectory, such as a loop. The change of monotonicity typically indicates a bifurcation (split) or annihilation (merge).
Such events may reflect topological changes of the vector field or are simply caused by numerical instabilities in trajectory extraction. One may need to simplify, segment, and filter the trajectories to understand the vector field dynamics. To these ends, we demonstrate novel understanding of trajectories based on multilevel robustness, as demonstrated in the rest of this paper.
\section{Conclusion and Discussion}
\label{sec:discussion}
In this paper, we introduce a new multilevel robustness framework.
Our framework helps to mitigate the drawbacks of the classic robustness computation due to the boundary effect, and better differentiate the behaviors of critical points in terms of their multiscale stability.
We show that the statistical information of multilevel robustness, in particular, minimum multilevel robustness, can be integrated seamlessly with feature tracking algorithms such as FTK as a postprocessing step.
Our framework thus supports feature tracking, selection, and comparison, and improve the visual interpretability of vector fields from scientific simulations.
Modern heuristic tracking schemes detect tropical cyclones through a two-step procedure: first, isolated minima in the sea level pressure field are identified; second, an upper-level warm core criteria is used to filter out storms that are not tropical in nature.
Compared with such heuristics, our robustness-based framework has the advantage of identifying strong cyclonic features using only the wind vector fields.
There are a number of points for discussions, such as feature tracking in 3D, scalability, uncertainty visualization, and alternative strategies.
First, it is possible to extend our current approach to 3D vector fields, since robustness has been studied for critical points in 3D~\cite{SkrabaRosenWang2016}.
However, topological features such as vortices and vortex cores are arguably more interesting to study in 3D than critical points, where a notion of robustness has yet to be developed. This presents a current limitation of our framework.
Second, our current implementation approximates multilevel robustness with a discrete set of radii. Increasing the number of levels will require more computational resources, where advanced parallel/distributed computation may be needed (c.f.,~\autoref{fig:runningTime}, which uses an embarrassingly parallel approach).
Third, our work is motivated by the computation of classic robustness, which may produce artifacts when the sublevel sets containing critical points intersect the domain boundaries.
It would be interesting to consider alternative strategies. For instance, moving critical points out of the domain by a perturbation may change the structure of the underlying sublevel sets, and the corresponding merge tree may become inconsistent with the original (observable) data.
Fourth, the multilevel robustness is a natural candidate for uncertainty visualization, which is left for future work.
Finally, the classic robustness has been applied to data beyond climate science, such as vector fields from combustion simulation and tensor fields from materials science and diffusion tensor imaging. We believe a generalization of multilevel robustness to these datasets would be interesting but beyond the scope of the current paper. A main challenge is to study its correlation with physical quantities in these respective application domains.
\section{Introduction}
\label{sec:introduction}
The analysis and visualization of vector fields has seen widespread applications in science and engineering, including combustion, climate study, and ocean modeling.
With the increasing size and complexity of vector field data that arise from scientific simulations, vector field topology has been one of the most promising tools to describe and interpret vector field behavior by providing meaningful abstraction and summarization~\cite{PobitzerPeikertFuchs2011, BujackYanHotz2020}.
Critical points ({i.e.}, where a vector field vanishes) are core features of vector field topology.
To improve the visual interpretability of time-varying vector fields, a key challenge is feature tracking~\cite{PostVrolijkHauser2003} -- in particular, critical point tracking -- that is, to resolve the correspondences between critical points in successive time steps in the form of trajectories, and to understand the dynamic behavior of these trajectories via selections and comparisons.
The topological notion of robustness has been introduced recently to quantify the stability of critical points.
The robustness of a critical point is defined to be the minimum amount of perturbation to the vector field necessary to cancel it.
Robustness has been shown to be useful in feature extraction~\cite{WangBujackRosen2017} and simplification~\cite{SkrabaWangChen2014,SkrabaWangChen2015,SkrabaRosenWang2016} of vector field data.
In particular, Skraba and Wang inferred correspondences between critical points based on their closeness in stability, measured by robustness, instead of just distance proximities within the domain~\cite{SkrabaWang2014b}.
They obtained theoretical results by relating critical point tracking with the notion of robustness: roughly speaking, critical points with high robustness values could be tracked more easily and more accurately~\cite{SkrabaWang2014b}.
However, the results in~\cite{SkrabaWang2014b} were theoretical in nature, and bringing this theory to practice is nontrivial.
Vector field data generated from large-scale ocean, atmospheric, and fluid dynamics simulations contain features at different scales.
It is a common practice for researchers to study the data within a chosen domain of interest.
For critical points close to the boundary of the domain, we have an incomplete picture of flow behavior within their local neighborhoods.
Consequently, the computation of classic robustness may suffer from poor boundary conditions; for instance, a critical point may not find a cancellation partner or may be forced to cancel with another critical point that is far away in the known data domain (see~\autoref{sec:ml-robustness} for details).
Such phenomena decrease the effectiveness in robustness-based critical point tracking.
In this paper, we introduce \emph{multilevel robustness} for critical points, a ``scale-aware'' notion of robustness that accommodates the inherent multiscale nature of vector field data.
Multilevel robustness helps to mitigate the boundary effect suffered by the classic robustness computation.
More importantly, it can be integrated with existing feature tracking algorithms to improve feature tracking, selection, and comparison.
\para{Contributions.}
Building upon the theoretical basis established previously~\cite{SkrabaWang2014b}, the focus of this paper is to realize robustness-based critical point tracking in practice for large-scale scientific simulations. To that end,
\begin{itemize}
\item We introduce a multilevel robustness framework for the study of 2D time-varying vector fields. We compute the robustness of critical points across varying neighborhoods to capture the multiscale nature of the data and to mitigate the boundary effect suffered by the classic robustness computation.
\item We demonstrate that our proposed framework -- in particular, the minimum multilevel robustness -- can be combined with feature tracking algorithms such as FTK~\cite{GuoLenzXu2021} to improve the visual interpretability of vector fields in terms of feature tracking, selection, and comparison.
\item We observe, \emph{for the first time}, that the minimum multilevel robustness is highly correlated with physical quantities (such as maximum wind speed and mean sea-level pressure) used by domain scientists in studying a real-world tropical cyclone dataset.
\end{itemize}
The observation above is quite exciting as it implies that robustness -- a notion of feature stability derived based on vector field perturbation -- is highly correlated with scalar-valued physical quantities commonly used by domain scientists to study tropical cyclones, which helps to increase the physical interpretability of robustness.
\section{Our New Definition: Multilevel Robustness}
\label{sec:ml-robustness}
To mitigate the drawbacks of the classic robustness computation, we introduce a multilevel robustness framework.
In \autoref{fig:classic-robustness}, we give an example of a classic robustness analysis using a 2D vector field instance from the $\ME$ dataset (see~\autoref{sec:ME} for details).
We study the robustness of critical points that represent the centers of large-scale eddies.
In (A), we visualize cancellation partners in computing the classic robustness.
In (B), we visualize the critical points with radii proportional to their classic robustness values.
Specifically, a number of centers ({e.g.}, $x_0, x_1, \dots, x_{9}$) are shown to share the same lowest zero degree ancestor in the merge tree (C), thus, they are grouped together and have the same robustness value of $0.43$.
In other words, for any value $r < 0.43$, these critical points may not be canceled based on Lemma~\autoref{lemma:degree}.
Specifically, at $r = 0.13$, the sublevel set $\Xspace_{0.13}$ contains these centers in isolation, see (C) and (D).
Such a phenomenon happens for two reasons.
First, some of these critical points represent centers of large-scale eddies and are surrounded by flows of a large magnitude.
Imagining that these centers are sitting at the bottoms of deep wells (of the vector magnitude field), a large amount of perturbation is then needed to cancel these centers, and therefore they have high robustness values.
Second, the sublevel set $\Xspace_{0.43}$ is shown to intersect significantly with the domain boundary in (E), and some of these critical points become cancellation partners due to the boundary effect.
To mitigate these issues, we introduce the notion of multilevel robustness.
Roughly speaking, for a critical point $x \in \Xspace$, we define its multilevel robustness as a sequence of robustness values computed from its neighborhoods of increasing radii.
Formally, let $B_x(a)$ denote a ball of radius $a$ surrounding a critical point $x \in \Xspace$, that is,
$B_x(a) := \{x' \in \Xspace \mid ||x-x'|| \leq a\},$
where $||x-x'||$ represents the Euclidean distance between two points.
The multilevel robustness of $x$ is a function
\[R_x: [0,\infty) \to \Rspace,\]
where $R_x(a)$ is the (classic) robustness of $x$ computed {{w.r.t.}} the domain $B_x(a)$ for $a \in [0,\infty)$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{MR2.pdf}
\vspace{-6mm}
\caption{Multilevel robustness calculation of critical points in a 2D vector field from the $\ME$ dataset. (A) Multilevel neighborhoods for critical point $x_3$. Columns (B) and (C): $N$-level robustness and ground-truth robustness of critical points $x_3$ and $x_4$.}
\label{fig:classic-robustness2}
\vspace{-2mm}
\end{figure}
We compute $R_x$ at a discrete number of radii.
Assuming the domain $\Xspace$ contains $n$ critical points, then for a fixed critical point $x \in \Xspace$, as $a$ increases, its multilevel robustness will change at most $n-1$ times, since $x$ gets one more candidate of the cancellation partner as $B_x(a)$ passes through each critical point.
Computing the multilevel robustness $R_x$ exactly (considered as the ground truth) takes $O(n^2)$ time, which is impractical for complex data with a large number of critical points.
Therefore, in practice, we approximate $R_x$ by sampling a number ($N$) of radii.
\autoref{fig:classic-robustness2}(A) illustrates our method in calculating the multilevel robustness.
For a critical point $x$, we consider $N=10$ number of its neighborhoods at radius $\{a_0, \dots, a_{N-1}\}$, where each $a_i := L \times (i+1)/N$ for $L$ being the diameter of the domain $\Xspace$ (i.e.,~$L$ is the least upper bound of the set of all distances between pairs of points in the domain).
\autoref{fig:classic-robustness2}(A) shows the $N$ neighborhoods of a critical point $x_3$ at radii at $10\%, 20\%, \dots, 100\%$ of $L$, respectively.
At each fixed level $a_i$, we compute the classic robustness of $x_3$, giving rise to its multilevel robustness $R_{x_3}$.
We investigate multilevel robustness for critical points $x_3$ and $x_4$ as $N$ increases, see~\autoref{fig:classic-robustness2} (B) and (C) for $N = 10, 30$ and $50$, respectively.
Not surprisingly, $R_x(a)$ becomes a better approximation of the ground truth as $N$ increases.
On the other hand, $R_x(a)$ appears to converge to the ground truth when $N \leq 50$ for our datasets of interests.
Therefore, we use $N=50$ to compute multilevel robustness in the remainder of this paper.
There are a few benefits of using multilevel robustness $R_x$ for a critical point $x \in \Xspace$.
First, $R_x$ is better at differentiating different behaviors of critical points in terms of their multiscale stability.
As shown in \autoref{fig:classic-robustness2}, critical points $x_3$ and $x_4$ now exhibit different behaviors using $R_x$.
Second, statistical information, such as minimum, median, and maximum of $R_x$, could be used in analysis and visualization tasks.
Specifically, for the remainder of this paper, we work with the minimum of $R_x$ for critical point tracking, selection, and comparison, which is defined as
\[
\minR_{x} := \min_{a \in [0,~L)} R_x(a).
\]
$\minR_x$ captures the smallest possible robustness of $x$ with varying neighborhood sizes, and thus alleviates the artifacts induced by the boundary effects in classic robustness calculation.
In addition, $\minR_x$ is shown to be highly correlated with physical quantities employed by domain scientists who study tropical cyclones; compare~\autoref{fig:HK} for a concrete example.
\section{Related Work}
\label{sec:related-work}
We review related work on vector field topology, critical point tracking, and robustness of critical points.
\para{Vector field topology}
has been researched over the past decades since it was firstly introduced by Helman and Hesselink~\cite{HelmanHesselink1989}.
However, as pointed out by Pobitzer~{et al.}~\cite{PobitzerPeikertFuchs2011} and Bujack~{et al.}~\cite{BujackYanHotz2020}, vector field topology for time-varying flows remains a challenge. In particular, it is difficult to
interpret flow topology {{w.r.t.}} physical meaning in the time-varying setting~\cite{BujackYanHotz2020}.
In this paper, we focus on the tracking and visualization of critical points of time-varying vector fields, and investigate the potential relationship between the topological properties of critical points and physical quantities of relevance to real-world flow dataset.
\para{Critical point tracking}, which reconstructs the trajectories of critical points over time, may be achieved by proximity-, integral-, and interpolation-based methods.
Proximity-based critical point tracking includes the work of Helman and Hesselink~\cite{HelmanHesselink1989, HelmanHesselink1990}, which connects the critical points (singularities) from separate time steps based on proximity and region connectedness.
For integral-based critical point tracking approaches, Theisel and Seidel~\cite{TheiselSeidel2003} recast the tracking of critical points in a 2D vector field as an integration problem in a 3D field, called feature flow field (FFF), and computed feature trajectories based on tangent curves in FFF.
Weinkauf~{et al.}~\cite{WeinkaufTheiselVan2010} improved upon the FFF and presented a more stable formulation for tracking critical points by addressing instabilities in the numerical integration during the computation of tangent curves. This is followed by the work in~\cite{ReininghausKastenWeinkauf2011} that introduced a combinatorial version of FFF.
An example of interpolation-based method is from Tricoche~{et al.}~\cite{TricocheScheuermannHagen2001a, TricocheWischgollScheuermann2002}, who implemented the linear interpolation between time steps, which guarantees the existence of one critical point in each cell, and analyzed the cell faces to detect changes in the topology over time.
Analogously to~\cite{TricocheScheuermannHagen2001a}, Garth~{et al.}~\cite{GarthTricocheScheuermann2004} extended this approach and provided a critical point tracking algorithm for 3D time-varying vector fields.
Guo~{et al.}~\cite{GuoLenzXu2021} proposed a \emph{simplicial} spacetime meshing scheme for tracking critical points, referred to as the Feature Tracking Kit (FTK) framework, which is further reviewed in \autoref{sec:background}.
\para{Robustness of critical points} has been introduced recently to quantify the structural stability of critical points with respect to perturbations to the vector fields~\cite{SkrabaWangChen2014,SkrabaWangChen2015,SkrabaRosenWang2016}.
Robustness has been shown to be useful for the analysis and visualization of vector fields. For example, Wang~{et al.}~\cite{WangRosenPrimoz2013} studied how the robustness of a critical point evolves in the time-varying setting. Skraba and Wang~\cite{SkrabaWang2014b} showed potential usage of robustness in feature tracking, that is, critical points with high robustness values could be tracked more easily and more accurately.
Robustness is also used
for 2D~\cite{SkrabaWangChen2014,SkrabaWangChen2015} and 3D~\cite{SkrabaRosenWang2016} vector field simplification.
Lately, Wang~{et al.}~\cite{WangBujackRosen2017} further extended the classic definition of robustness to a Galilean invariant robustness framework that quantifies the stability of critical points across different frames of reference.
The notion of robustness was further extended to study the stability of degenerate points in tensor fields~\cite{WangHotz2017,JankowaiWangHotz2019}.
The concept of robustness, first introduced by Edelsbrunner~{et al.}~\cite{EdelsbrunnerMorozovPatel2011a, EdelsbrunnerMorozovPatel2011b}, is closely related to the notion of \emph{persistence}~\cite{EdelsbrunnerLetscherZomorodian2002} -- a common tool used to quantify feature importance.
In addition to robustness, other measures have been explored to characterize the importance of vector field critical points based on their lifetime~\cite{KastenHotzNoack2011} and scales~\cite{KleinErtl2007}.
Different to previous efforts, this paper introduces a new notation of \emph{multilevel robustness} for critical points.
Multilevel robustness studies the robustness of a critical point {{w.r.t.}} their local neighborhoods of varying sizes, and thus helps to mitigate the boundary effects suffered by classic robustness computation, and better differentiates the behaviors of critical points across multiple scales.
\section{Results with Large-Scale Simulations}
\label{sec:results}
\begin{figure*}[!t]
\centering
\includegraphics[width=2\columnwidth]{runningTime2.pdf}
\caption{Boxplots of running time with different neighborhood sizes for the (A) $\EW$, (B) $\EWL$, (C) $\HK$, and (D) $\ME$ datasets. }
\label{fig:runningTime}
\vspace{-2mm}
\end{figure*}
\begin{figure*}[!t]
\centering
\includegraphics[width=1.0\linewidth]{E3SMWindL-Selection.pdf}
\vspace{-4mm}
\caption{Feature tracking and selection for the $\EWL$ dataset. (A) A selected vector field with zoomed-in views, where
saddles are in blue, and critical points with degree $+1$ are in red.
(B) Tracking results with degree filter thresholds at $-1$ and $-0.2$ (from left to right) and stability filter thresholds at $0$, $0.2$, and $0.4$ (from top to bottom).}
\label{fig:EWL}
\vspace{-2mm}
\end{figure*}
\begin{figure*}[!t]
\centering
\includegraphics[width=1.0\linewidth]{HurricaneKatrina5.pdf}
\vspace{-4mm}
\caption{Feature tracking and selection for the $\HK$ dataset. (A) Selected vector fields with zoomed-in views.
(B) Tracking results with stability filter thresholds at $0$ and $0.1$ (from left to right) and degree filter thresholds at $-1$ and $-0.02$ (from top to bottom).}
\label{fig:HK}
\vspace{-2mm}
\end{figure*}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{HK-mainTrajectory2.pdf}
\vspace{-6mm}
\caption{Feature tracking result for the $\HK$ dataset. (A) Initial FTK trajectories. (B) Initial FTK trajectory that contains \emph{Katrina}. (C) Segmenting trajectory in (B) with multilevel robustness; new segments are assigned different colors.}
\label{fig:HK-mainTrajectory}
\vspace{-2mm}
\end{figure}
We demonstrate the use of multilevel robustness in feature tracking, selection, and comparison for large-scale scientific simulations.
\autoref{tab:data} lists some basic information for datasets used in this paper, including the number of time steps, grid nodes, grid cells (triangles), and critical points.
We also provide a brief running time analysis for all datasets based on each task discussed in \autoref{sec:vis} (i.e.,~a single task involves computing classic robustness of a given critical point at a fixed radius).
All tasks are arranged on a cluster with 664 nodes (128GB DDR4 and 36 cores). We utilize 16 nodes in all experiments, which means that at most $16 \times 36=575$ tasks can run at the same time.
Since the running time of each task is highly correlated with the size of neighborhood in the robustness calculation, \autoref{fig:runningTime} provides the box plots of running time at each level of robustness for all critical points in a given dataset; also see \autoref{tab:data} (last column) for the range of running time of tasks for each dataset.
In the following, all timestamps in the descriptions are represented in the universal coordinated time (UTC).
\begin{table}[t]
\caption{Datasets used in this paper. \#steps, \#nodes, \#cells, and \#CPs, respectively, mean the number of timesteps, number of vertices in the mesh, number of cells in the mesh, and number of critical points. Running times are in seconds.}
\label{tab:data}
\scriptsize%
\centering%
\begin{tabular}{cccccc
}
\hline
\textbf{Dataset }& \textbf{\#steps} & \textbf{\#nodes} & \textbf{\#cells} & \textbf{\#CPs} & \pbox{15cm}{\textbf{time per task}}\\
\hline
$\EW$&36&13,639&26,356&8-53&0.11-1.18\\
$\EWL$&36&99,007&195,602&238-425&0.11-27.72\\
$\HK$&216&50,861&100,800&208-501&0.11-15.97\\
$\ME$&$20\times 4$&37,929&75,063&18-65&0.11-1.44\\
\hline
\end{tabular}%
\end{table}
\subsection{Feature Tracking and Selection}
\para{E3SM Wind L dataset.}
We revisit the HighResMIP-v1.0 (1950-Control) dataset~\cite{CaldwellMametjanovTang2019} from
E3SM simulations.
Instead of truncating a small region for illustrative purposes in~\autoref{sec:vis}, we enlarge the $\EW$ dataset to the $\EWL$ dataset by choosing the region from $10$ S$^\circ$ to $50$ N$^\circ$ for latitude and from $80$ E$^\circ$ to $175$ E$^\circ$ for longitude.
As shown in~\autoref{fig:EWL}(B), with the appropriate trajectory segmentation, stability filtering, and degree filtering, our framework detects the trajectories of three main cyclones in the domain, denoted as $a$, $b$, and $c$.
Trajectory $a$ appears at the east of Japan from time step $0$, moves to the east, and disappears on the right boundary at time step $9$.
Trajectory $b$ exists from time step $0$ to $31$, and coincides with the selected cyclone trajectory of $\EW$ in~\autoref{fig:VisExample-Tracking} and~\autoref{fig:VisExample-Selection}.
Trajectory $c$ stays on the right bottom corner of the domain from time steps 0 to 21.
For further investigation of these trajectories, we visualize the vector fields associated with time step 0 in~\autoref{fig:EWL}(A), with color map based on the magnitude of the vector fields. We also give the zoomed-in views of the detected main features.
\para{Hurricane Katrina dataset.}
\emph{Hurricane Katrina} was a large and destructive Category 5 Atlantic hurricane that formed on August 23, 2005, and dissipated on August 31, 2005.
Our $\HK$ dataset is truncated from ECMWF Reanalysis v5 (ERA5), which is produced by the Copernicus Climate Change Service (C3S)~\cite{C3S}.
ERA5 provides hourly estimates of the global climate information covering the period from January 1950 to the present with the spatial grid resolution of 30 km. The rectangular region is centered at the southeast of the contiguous U.S. ($5$ N$^\circ$ to $50$ N$^\circ$ and $120$ W$^\circ$ to $50$ W$^\circ$); the time steps range from 12:00, August 23, 2005, to 23:59, August 31, 2005.
Since ERA5 uses a one-hour time gap, our $\HK$ dataset contains $9 \times 24 = 216$ instances.
We choose 10m zonal and meridional wind speed as the 2D vector field, since in the near-surface the hurricane core represents a region of strong convergence and associated vertical motion.
We start with the initial trajectories provided by FTK, shown in \autoref{fig:HK-mainTrajectory}(A).
Due to visual clutter among thousands of trajectories, it is hard to identify the principal features.
However, our multilevel robustness framework is able to detect four dominant features in the domain, after trajectory segmentation and filtering, shown as trajectories $a$ to $d$ in~\autoref{fig:HK}(B).
We visualize vector fields associated with time steps $15$, $65$, and $120$, and highlight these four features in the zoomed-in views; see~\autoref{fig:HK}(A).
In particular, trajectory $c$ contains critical points representing the center of \emph{Katrina}.
We perform a detailed analysis of the robustness-based segmentation of trajectory $c$.
As shown in~\autoref{fig:HK-mainTrajectory}(B), the initial FTK trajectory containing $c$ also contains a number of critical points that are not associated with \emph{Katrina}.
These critical points are part of the same initial trajectory with $c$.
Multilevel robustness is used to segment this initial trajectory by differentiating spurious features from the features of interest, using $k=0.5$ for the logistic transformation and $\sigma=0.2$ for the KDE.
As shown in~\autoref{fig:HK-mainTrajectory}(C), our approach extracts the trajectory $c$ that represents \emph{Katrina} with two segmentation points highlighted by orange arrows.
Trajectory $c$, which represents \emph{Katrina}, exists between time steps 33 and 201, which correspond to 9:00, August 24, 2005 and 9:00 August 31, 2005.
This means our framework does not capture \emph{Katrina} on August 23, 2005, when it was a tropical depression.
Taking a closer look at the critical points within the dashed white box from~\autoref{fig:HK}(A) at time step 15, it is hard for us to extract the low robust critical point that can represent \emph{Katrina} as there are many unstable features nearby.
However, this could be an artifact of the reanalysis product being used: since data assimilation in ERA5 occurs at 09:00, it is likely that the reanalysis product has been artificially adjusted to include \emph{Katrina}'s precursor.
Overall, our framework works well in detecting \emph{Katrina} when it strengthened into a tropical storm on the morning of August 24.
Based on input from domain scientist, our hypothesis of data preparation being an issue for Katrina is related to the disparate character of the storm at the hour of its first detection and the previous hour. The storm first appears when the data assimilation system is employed to generate new initial conditions for the forecast, suggesting that its development was not easy to predict from a forecast run initialized 12 hours earlier. Thus data preparation is likely one factor in the inability to extract a clear center when the storm is a tropical depression.
Tropical depressions are not well-organized systems, and whether or not the storm eventually develops a clear eye is highly dependent on the 3D evolution of the storm. So at this early stage, it is not surprising that a system that uses only a 2D slice of the wind field cannot detect the storm.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{PhysicalQuantity.pdf}
\vspace{-6mm}
\caption{ (A) A schematic in calculating a five degree region. (B) Minimum multilevel robustness $\minR_x$, (C) 5DMWS, and (D) MSLP along the Katrina trajectory.}
\label{fig:windSpeed}
\vspace{-2mm}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{PhysicalQuantity-abd.pdf}
\vspace{-6mm}
\caption{From top to bottom, $\minR_x$, 5DMWS, and MSLP along trajectories $a$, $b$, and $d$ from \autoref{fig:HK}. }
\label{fig:quantity-abd}
\vspace{-2mm}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\linewidth]{MPASOEddy.pdf}
\vspace{-6mm}
\caption{Feature comparison with multilevel robustness for the $\ME$ dataset. Left three columns: selected vector fields at time steps $0, 9$, and $19$, with $\omega =1.0, 1.5, 2.0$, and $2.5$. Critical points with degrees $+1$ and $-1$ are in red and blue, respectively.
Columns (A), (B) and (C): feature tracking and selection results and their zoomed-in views; trajectories are colored by their indices with radius proportional to $\minR_x$. (B) corresponds to the regions in (A) surrounded by red boxes, and (C) refers to the regions enclosed by blue boxes.}
\label{fig:ME}
\vspace{-2mm}
\end{figure*}
\para{Correlation with known physical quantities.}
We investigate the relationship between robustness and quantities that are used by the tropical cyclone research community, including five-degree maximum wind speed (5DMWS) and mean sea-level pressure (MSLP).
These two quantities are commonly used by domain scientists to detect, track, and evaluate tropical cyclones.
Compared with traditional cyclone tracking schemes, our multilevel robustness framework has the advantage of identifying cyclonic features using only the wind vector fields.
We observe a strong correlation between robustness and 5DMWS, which has been widely used in hurricane intensity metrics such as the Saffir-Simpson scale. As illustrated in~\autoref{fig:windSpeed}(A),
5DMWS is defined as the maximum wind speed within the five degree region of the hurricane center $x$.
As shown in~\autoref{fig:windSpeed}(B) and (C), the Pearson correlation coefficients between the $\minR_x$ curve
and 5DMWS is 0.95, suggesting a strong relationship between these two quantities.
Mean sea-level pressure (MSLP) is another scalar field that is frequently used by domain scientists in hurricane analysis. MSLP is connected to maximum wind speed via gradient wind balance, modified to account for surface friction; an empirical relationship connecting these two quantities is described in~\cite{Holland2008}.
Hurricanes higher on the Saffir-Simpson scale (with higher maximum wind speed) have lower MSLP at their center.
The MSLP along the detected Katrina trajectory is shown in~\autoref{fig:windSpeed}(D).
The Pearson correlation coefficient between the $\minR_x$ curve and MSLP is -0.83.
We further investigate the correlations between robustness with 5DMWS and MSLP for trajectories $a$, $b$, and $d$ from \autoref{fig:HK}.
As shown in \autoref{fig:quantity-abd} (1st row), $\minR_x$ has a strong correlation with 5DMWS and MSLP along trajectory $a$, where the correlation of the former is 0.94 and of the latter is -0.95.
Further investigation reveals that trajectory $a$ corresponds to Hurricane Hilary.
This experiment shows that robustness strongly correlates with physical quantities for trajectories $a$ and $c$, which correspond to Hurricane Hilary (category 2 hurricane) and Hurricane Katrina (category 5 hurricane) respectively.
However, such high correlations do not generalize to weaker storm systems.
The correlation between $\minR_x$ with 5DMWS is $0.764$ for trajectory $d$, and $-0.032$ for trajectory $b$, see \autoref{fig:quantity-abd} (2nd and 3rd row).
It turns out that trajectory $d$ represents the Tropical Storm Irwin, whereas trajectory $b$ can not be found in the National Hurricane Center's Tropical Cyclone Reports.
\subsection{Feature Comparison for Ensemble Dataset}
\label{sec:ME}
We demonstrate our framework in feature comparison in an ensemble of global ocean dataset (referred to as $\ME$), simulated with different wind stress parameters. The simulation code, MPAS-Ocean~\cite{GolazCaldwellVan2019, PetersenAsayBerres2019}, is a multiscale and unstructred mesh simulation for studying the ocean component of climate changes.
In this experiment, we utilized the standard low-resolution EC60to30 mesh, whose size of the cells along the coast varied from 60 km to 30 km.
Specifically, in the $\ME$ dataset, each of the four simulation runs captures 20-day ocean eddies with the \emph{bulk wind stress amplification parameter} $\omega$ varying from $1.0$ to $2.5$; the time-resolution of the data is 1 day.
We truncate the region near the equator in the Pacific Ocean ($15$ S$^\circ$ to $18$ N$^\circ$ for latitude and $170$ E$^\circ$ to $110$ W$^\circ$ for longitude), since this region contains many large eddies for feature comparison.
~\autoref{fig:ME} (left three columns) visualizes selected vector fields associated with time steps $0$, $9$, and $19$, for $\omega = 1.0, 1.5, 2.0$ and $2.5$, respectively.
In the rest of this section, we investigate the variability of features -- in particular, the centers of eddies -- induced by varying wind stress.
From the visualization of vector fields in~\autoref{fig:ME}, we obtain some preliminary observations: (1) critical point locations share a similar distribution at the beginning of each simulation (1st column); (2) as $\omega$ increases, vector field features show more variations as $t$ increases (3rd column); (3) a large value of $\omega$ leads to a higher flow magnitude. However, simply showing the locations of critical points in the ensemble has a limited effect on guiding parameter selection and post hoc analysis.
Instead, our framework can capture more variability across the four parameter settings for feature comparison.
To preserve the merging and splitting behavior of critical points, we set the threshold for degree filter at $-1$, thereby preserving saddles in the domain.
We also set $\sigma=0.2$ and the threshold for stability filter at $0.02$ to postprocess initial FTK trajectories and eliminate visual clutter.
The first observation from our framework is that trajectories have a shorter lifespan as $\omega$ increases; see the trajectories in~\autoref{fig:ME}(B) (from top to bottom) for examples.
The second observation is that as $\omega$ increases, a number of critical points have decreased robustness and more consistent cancellation partners, see~\autoref{fig:ME}(C).
The domain scientists pointed out that an increased wind stress will reduce the scales of existing eddies and suppress the development of larger scale eddies.
This also leads to a decrease in stability, measured by robustness, for some eddy centers as they interact more easily with nearby features, thus locating more consistent cancellation partners (see~\autoref{fig:ME}(C) at $\omega=2.5$).
As a consequence, some trajectories are filtered out in~\autoref{fig:ME}(A) as $\omega$ increases, since these trajectories become less stable.
To summarize, our feature comparison captures variability and stability among critical point trajectories under various parameter settings, which may help guide parameter selection ({e.g.}, maintaining a certain number of stable features) in scientific simulations.
\section{Method: Multilevel Robustness for Visualization Tasks}
\label{sec:vis}
With the newly introduced multilevel robustness framework, we develop its usage in visualization tasks.
Such a new notion of robustness can be combined seamlessly with any feature tracking algorithm.
We choose to integrate it with FTK~\cite{GuoLenzXu2021}, a state-of-the-art feature tracking technique.
In particular, we demonstrate that the minimum multilevel robustness $\minR_{x}$ can be integrated with the FTK algorithm to improve \myupdate{the original FTK} feature tracking and selection results for scientific simulations.
\para{Illustrative dataset.} In this section, we use an $\EW$ dataset to illustrate our method.
$\EW$ is a time-varying 2D vector field processed using a HiResMIP-v1.0 (1950-Control) dataset~\cite{CaldwellMametjanovTang2019} from the Energy Exascale Earth System Model (E3SM)~\cite{GolazCaldwellVan2019} project.
\myupdate{It has an approximate horizontal resolution of $0.25^\circ$ (degree) in the atmosphere ($28$ km grid spacing), with an ocean and sea ice grid of $18$ km in the mid-latitudes and $6$ km at the equator and poles.}
We truncate a rectangular region around south Asia ($10$ N$^\circ$ to $30$ N$^\circ$ and $105$ E$^\circ$ to $140$ E$^\circ$), and select $36$ time steps from September 18, the 26th run of the 1950-control dataset with 6 hours as the time gap.
We use \emph{UBOT} and \emph{VBOT} as 2D vector fields, which correspond to lowest model level zonal and meridional wind, respectively.
These instances describe the movement of a main cyclone, which forms in the Pacific ocean, passes through the Philippines (around time steps 15-18), makes landfall (time step 27), and dissipates (around time step 31) at the mainland of south Asia; \myupdate{ see~\autoref{fig:VisExample} (A), (C), and (D), which visualizes the vector fields associated with time steps $0$, $15$, and $27$. }
The cyclone of interest is indicated by the white arrows.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\columnwidth]{VisExample2.pdf}
\vspace{-2mm}
\caption{\myupdate{Critical points and FTK tracking results for the $\EW$ dataset. (A), (C), and (D): selected vector fields with their critical points. (B) Feature (i.e., critical point) tracking results using FTK; each trajectory is colored by the trajectory id.}}
\label{fig:VisExample}
\vspace{-2mm}
\end{figure}
\para{Initial computation of trajectories and multilevel robustness.}
The initial (critical point) trajectories of time-varying vector field data are computed by FTK~\cite{GuoLenzXu2021} \myupdate{(see~\autoref{fig:VisExample} (B))},
and we then use the method of Tricoche~{et al.}~\cite{TricocheScheuermannHagen2001b} to calculate degrees of critical points in individual time steps.
The computation of multilevel robustness is parallelized with Eden~\cite{SimmermanOsborneHuang2012}, which schedules and manages a number of small tasks on a high-performance computing cluster. In our implementation, each task is associated with one critical point and a neighborhood size. As a result, the robustness computation of $n$ critical points with $N$ levels leads to $n\times N$ independent tasks.
\subsection{Enhancing Feature Tracking with Multilevel Robustness}
In this section, we show that multilevel robustness -- in particular, the minimum multilevel robustness $\minR_x$ -- can significantly improve the feature tracking results.
Given the initial trajectory of a critical point $x$ together with its multilevel robustness $R_x$ over time, we may visualize the trajectory by encoding the statistical information of $R_x$ along the trajectory, such as its minimum, median, and maximum values.
\autoref{fig:VisExample-Clustering}(D) shows a visualization of these trajectories where the radius of each point along a trajectory is shown to be proportional to the minimum of its multilevel robustness $\minR_x$.
The main idea of feature tracking with multilevel robustness is to segment the initial trajectories (obtained by FTK or any other feature tracking algorithms) into multiple pieces with similar robustness values.
As discussed in~\autoref{sec:ml-robustness}, we prefer to use the minimum of multilevel robustness $\minR_x$ to quantify the stability of critical points, which alleviates the artifacts introduced by the boundary effect in classic robustness computation.
We demonstrate that our tracking strategy improves the initial FTK trajectories and captures stable features in the domain, for example, in tracking the main cyclone for the $\EW$ dataset.
We focus our analysis on an FTK trajectory that contains the main cyclone. The blue trajectory in~\autoref{fig:VisExample-Clustering}(A and D) shows the $\minR_x$ values along the trajectory.
Note that in \autoref{fig:VisExample-Clustering}(A), each trajectory is a parameterized curve, where an integer index (horizontal axis) corresponds to the parameter used in the parameterization; thus, each trajectory is not necessarily monotonic in time.
\myupdate{In the remaining of this section, we use indices to refer to nodes along a trajectory.}
As shown in~\autoref{fig:VisExample-Clustering}(A), $\minR_x$ decreases significantly at \myupdate{indices} $15$ and $16$, and its value remains low after \myupdate{index} $32$.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{KDEClustering.pdf}
\vspace{-6mm}
\caption{(A) Minimum multilevel robustness values $\minR_x$ for a critical point $x$ along the selected FTK trajectory. (B) Logistic transformation of $\minR_x$ along the FTK trajectory. (C) Segmentation of points in the trajectory based on kernel density estimation. (D) The selected FTK trajectory is highlighted in solid blue, whereas other trajectories are transparent.}
\label{fig:VisExample-Clustering}
\vspace{-2mm}
\end{figure}
Our first step is to segment a given trajectory into groups of critical points with similar robustness values.
This step is supported by the theoretical work in~\cite{SkrabaWang2014b}, where correspondences between critical points may be inferred based on their closeness in robustness.
To induce a segmentation more easily, we can amplify the signal $\minR_x$ with a logistic transformation.
Starting from a standard logistic function $s(z) = {1}/\left({1+e^{-k (z - z_0)}}\right)$, set $z = \minR_x$ at a fixed time step and $z_0 = 0$ (the minimal possible robustness value). Since $\minR_x \in [0, \infty)$, we have $s(z) \in [1/2, 1]$.
Introducing a normalization term, we have
\[
l(\minR_x)=\frac{2}{1+e^{-k\cdot \minR_x}}-1,
\]
so that $l(\minR_x) \in [0,1]$.
Here, $k$ is the logistic growth rate or the steepness of the curve of the function.
We set $k=0.5$ for most cases, and discuss the parameter choices later.
There are two justifications for using a logistic transformation.
First, $\minR_x$ may be infinity when a critical point $x$ cannot find a cancellation partner in the known data domain; $l(\minR_x)$ is thus constrained within the range $[0,1]$, making parameter selection easier.
Second, $l(\minR_x)$ is less sensitive {{w.r.t.}} to the changes in $\minR_x$, and therefore it focuses only on significant changes of $\minR_x$.
For example, as shown in \autoref{fig:VisExample-Clustering}(A) and (B), the logistic transformation $l$ maps the $\minR_x \in [0, \infty]$ to $l(\minR_x) \in [0,1]$.
Furthermore, it helps to differentiate unstable critical points along the trajectory from relatively stable ones.
As shown in~\autoref{fig:VisExample-Clustering}(B), there appears to be clear separations between indices $[15, 16]$ and $[31, 40]$ from the rest of the trajectory.
Our second step is to cluster critical points along a trajectory into different groups using $l(\minR_x)$ and kernel density estimation (KDE) with a Gaussian kernel.
As illustrated in~\autoref{fig:VisExample-Clustering}(C), by choosing an appropriate bandwidth parameter $\sigma$ for the Gaussian kernel, we can further cluster the critical points with indices $[15,16]$ and $[32, 40]$ from those with indices $[0,14]$ and $[17, 31]$.
$\sigma$ controls the smoothness of a KDE, where a small $\sigma$ leads to more segments.
For our experiments, we set $\sigma=0.2$ as the default value; see a later section for parameter tuning.
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{E3SMWind-Tracking.pdf}
\vspace{-6mm}
\caption{Comparing feature tracking results for a time-varying $\EW$ dataset. (A) Initial (classic) FTK trajectories; (B) trajectories obtained based on multilevel robustness, with zoomed-in view for indices 13-15. The vector field at index 14 is visualized in (C) with a zoomed-in view around the Philippians in (D).}
\label{fig:VisExample-Tracking}
\vspace{-2mm}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=2.01\columnwidth]{E3SMWind-parameter3.pdf}
\vspace{-1mm}
\caption{Parameter selection of $k$ and $\sigma$ for the $\EW$ dataset. (A) From top to bottom, feature tracking results for $k=0.1$ and $0.5$. From left to right, $\sigma=0.1$, $0.2$, and $0.4$. \myupdate{(B) Number of trajectories {{w.r.t.}}~parameter selection of $k$ and $\sigma$. (C) A zoomed-in view of (B) for the region surrounded by the dashed white box.}}
\label{fig:VisExample-parameter}
\vspace{-2mm}
\end{figure*}
Our third step is to refine critical point trajectories based on the clustering results.
Critical points belonging to the same cluster are reconstructed as a new trajectory by examining spatial faces and spacetime edges~\cite{GuoLenzXu2021}.
As illustrated in~\autoref{fig:VisExample-Tracking}(B), the selected blue FTK trajectory is segmented into three pieces: a orange trajectory connecting critical points of indices $[15,16]$, an red trajectory connecting critical points with indices $[31, 40]$, both with small robustness values; and the remaining blue trajectory with large robustness values.
In particular, the new blue trajectory in (B) is reconstructed by connecting critical points of indices $14$ and $16$ following the approach in~\cite{GuoLenzXu2021}.
Based on domain knowledge, a critical point representing the center of a cyclone should have a high stability measure across time before it hits the land and dissipates.
Take a close look at the trajectory at index 14. As shown in \autoref{fig:VisExample-Tracking} (D), there are two critical points with low stability near the landmass of the Philippines.
Since the classic FTK algorithm only considers the correspondences of critical points based on 0-levelset extraction, these two critical points are included in the initial trajectory in ~\autoref{fig:VisExample-Tracking}(A).
Furthermore, the critical points with indices $[31, 40]$ are likely unstable features when the cyclone makes landfall and dissipates.
Our feature tracking method is used as a postprocessing step to segment initial FTK trajectories into more meaningful segments, based on multilevel robustness.
In particular, we compare the initial trajectory with our new trajectory based on multilevel robustness in~\autoref{fig:VisExample-Tracking}(A)-(B).
The initial trajectory includes a pair of critical points on the island of the Philippines with low robustness values; whereas our method successfully tracks the main cyclone and removes these two critical points from the main trajectory, as indicated by an orange arrow in \autoref{fig:VisExample-Tracking}(B).
Also, our method splits the initial trajectory into a red trajectory when the cyclone hits the south of Asia, as indicated by the green arrow in \autoref{fig:VisExample-Tracking}(B), since the low robust tail of the initial trajectory is formed by unstable features on land; see critical points within the white box of~\autoref{fig:VisExample-Tracking}(C).
\para{Parameter selection for $k$ and $\sigma$.}
We now discuss how the choice of $k$ from the logistic transformation and bandwidth $\sigma$ from the KDE affect the feature tracking results.
$k$ is used to control the growth rate of the logistic transformation.
\autoref{fig:VisExample-parameter}\myupdate{(A)} shows the feature tracking results for $k=0.1$ and $0.5$.
When $k$ is relatively small ({e.g.}, $k=0.1$), the logistic transformation cannot differentiate stable features from unstable ones, regardless of the values of $\sigma$.
For example, for $k=0.1$ and $\sigma=0.1$, our method over-segments the initial trajectory.
As $\sigma$ increases (\autoref{fig:VisExample-parameter}\myupdate{(A)} 1st row), our method does not exclude unstable features on the island and those on the mainland.
On the other hand, we obtain reasonable (similar) feature tracking results for $k \in [0.3, 1.0]$.
This indicates that a slightly higher value of $k$ is effective in differentiating stable and unstable features. Therefore, we set $k=0.5$ for our experiments in~\autoref{sec:results}.
For a fixed $k$ value, \autoref{fig:VisExample-parameter}\myupdate{(A)} also shows the feature tracking results for $\sigma=0.1$, $0.2$, and $0.4$, respectively.
As discussed previously, a small $\sigma$ will likely introduce the over-segmentation of a given trajectory.
For example, when $\sigma=0.1$, the trajectory representing a merging behavior of a pair of critical points on the left bottom corner of $\EW$ is divided into two parts, as indicated by the orange arrows in \autoref{fig:VisExample-parameter}\myupdate{(A)} (1st column and 2nd row).
When $\sigma$ is large, the KDE curve becomes too smooth to differentiate stable and unstable features. As shown in \autoref{fig:VisExample-parameter}\myupdate{(A)} (3rd column and 2nd row), if we set $\sigma=0.4$, the blue trajectory is similar to the trajectory using classic FTK algorithm.
This means that our feature tracking under $\sigma=0.4$ fails to extract the main cyclone from other unstable features around the regions indicated by orange arrows.
Finally, since both KDE and $l(\minR_x)$ have a range of $[0,1]$, we set $\sigma=0.2$ as default since it works well in most cases considered in~\autoref{sec:results}.
\myupdate{Additionally, we use a heatmap that records the number of trajectories under various parameter settings in \autoref{fig:VisExample-parameter}(B), which provides a supplementary view for parameter selection.
The values of $k$ and $\sigma$ from \autoref{fig:VisExample-parameter}(A) are highlighted by crosses in \autoref{fig:VisExample-parameter}(B).
For parameter selection, we look for the regions where the number of trajectories remains relatively stable with a range of values for $k$ and $\sigma$.
For example, when $\sigma \geq 0.45$, our framework produces the same number of trajectories as the original FTK tracking result. It means that with a relatively high $\sigma$, our framework cannot differentiate between stable and unstable features, e.g., see the region surrounded by the green boundary in \autoref{fig:VisExample-parameter}(B).
On the other hand, a small $\sigma$ tends to over-segment the initial FTK tracking results, see the first row of \autoref{fig:VisExample-parameter}(B).
Our default values of $k$ and $\sigma$ come from the region surrounded by orange boundary in \autoref{fig:VisExample-parameter} (C). Any combination of $k$ and $\sigma$ from this region leads to the same post-processed feature tracking result.}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{E3SMWind-Selection.pdf}
\vspace{-4mm}
\caption{Feature selection for the $\EW$ dataset.
From top to bottom: degree filtering with thresholds at $-1$ ({i.e.}, keeping all saddles) and $-0.2$ respectively.
From left to right: stability filtering with thresholds at $0$, $0.2$, and $0.4$, respectively.
Trajectories are colored by the degrees of critical points, where red means $+1$ and blue means $-1$. The radius of a trajectory is proportional to $\minR_x$.}
\label{fig:VisExample-Selection}
\vspace{-2mm}
\end{figure}
\subsection{Feature Selection with Multilevel Robustness}
This section demonstrates feature selection aided by multilevel robustness. We introduce two filters, one based on $\minR_x$, and the other based on degree information.
Both filters help to reduce visual clutter and highlight dominant features in the domain.
Our first feature selection strategy is referred to as the \emph{stability filtering}.
For any trajectory, this strategy considers the topological notion of stability in terms of $l(\minR_x)$, as well as its temporal stability in terms of the lifespan.
Let $\gamma$ denote a parameterized trajectory, $|\gamma|$ is its total length.
Formally, we define the \emph{stability measure} of a trajectory $\gamma$ as follows:
\begin{equation}
\label{eqn:stability}
b(\gamma) := \frac{\sum_{x \in \gamma}l(\minR_x)}{|\gamma|} \cdot \frac{t_\gamma}{T},
\end{equation}
where $T$ is the temporal span of all trajectories ({e.g.}, $T=36$ for the $\EW$ dataset), and $t_\gamma$ is the temporal span of $\gamma$ in terms of the maximum difference between node indices.
The first term in Eqn.~\eqref{eqn:stability} captures the average pointwise stability (in a logistic scale), whereas the second term encodes the lifespan of the trajectory.
By definition, $b(\gamma)$ has a range in $[0, 1]$.
Our second feature selection strategy is referred to as the \emph{degree filtering}.
That is, we select trajectories based on their pointwise average degree.
Formally, for a trajectory $\gamma$, its \emph{average degree} is
\begin{equation}
d(\gamma) := \frac{\sum_{x \in \gamma} \mm{\mathrm{deg}}(x)}{|\gamma|},
\end{equation}
where $\mm{\mathrm{deg}}(x)$ is the degree of a critical point $x \in \gamma$.
Since a critical point may be of degree $+1$ or $-1$, $d(\gamma)$ has a range in $[-1, 1]$.
For our experiments involving cyclones and ocean eddies, we work primarily with critical points with a degree of $+1$, which correspond to centers of cyclones and eddies.
Domain scientists mainly care about centers in our applications.
For example, the trajectory that represents the main cyclone in the $\EW$ dataset contains critical points (centers) whose degrees are all $+1$.
Therefore, trajectory $\gamma_1$ within \autoref{fig:VisExample-Selection} has $d(\gamma_1) = +1$. For trajectory $\gamma_2$, $d(\gamma_2) = 0$, since critical points on its left branch have degrees $+1$ and critical points on its right branch have degrees $-1$.
Once a trajectory is enriched with a stability measure and an average degree, one may select features based on these criteria jointly or independently.
As illustrated in \autoref{fig:VisExample-Selection}, we successfully selected the trajectory that represents the main cyclone with degree filtering and stability filtering.
|
\section{Introduction and Related Work}
Nowadays High Energy Physics (HEP) analyses take advantage of Multivariate Analysis (MVA) techniques to optimize the discrimination between signal and background, preserving as much signal as possible. Indeed, running a classical cut-based selection on several physics observables would imply a severe reduction of both signal and background candidates, which would turn out to be a quite inefficient choice especially when signal events are rare as usually happens when performing Higgs-related studies.
On the other hand, HEP MVA focuses on using a pre-determined set of independent variables optimally combined to build discriminants which could effectively separate signal from background.
In this context, an Artificial Neural Network naturally implements a MVA since it is a complex Deep Learning computing system that mimics the biological neural networks receiving these variables as input, training and testing on them and, eventually, producing a single output for binary classification problems \cite{Peterson:1991wf,Glorot:2010}.
Generally, a Neural Network consists of nodes organized in layers. Each node receives either a feature of the problem or a weighted sum of the previous layer node output. It uses several hyper-parameters (e.g. learning rate, number of hidden layers, neurons per each layer, dropout fraction, the minimizer algorithm of the loss function, etc.) that are usually set manually in order to define the network architecture.
In our work, we used Monte Carlo generated events from non-resonant Higgs boson pair production analysis at the energies of the LHC, where one of the Higgs bosons decays into the four-lepton final state and the other one decays into a pair of b quarks as described in \autoref{signaltopology}. The signal and background data sets were generated to run the analysis on data corresponding to the integrated luminosity reached by the Large Hadron Collider (LHC) experiment at CERN during the 2018 Run II period using proton-proton collisions at a center of mass energy of 13 TeV. Furthermore, a DNN classifier, usually considered a very versatile and efficient method for these kind of problems, has been implemented by using the open-source library Keras and the open source platform for machine learning TensorFlow \cite{tensorflow2015-whitepaper,Keras}.
In particular, we show in \autoref{comments} the implementation and optimization of a DNN classifier for discrimination of our signal from SM background events. We then conclude with a brief outlook on our plans for future development.
\section{The Signal Topology}\label{signaltopology}
The discovery of the Higgs boson at the Large Hadron Collider in 2012 \cite{CMS:2012qbp,ATLAS:2012yve} opened a new frontier in HEP both for Standard Model (SM) and Beyond Standard Model (BSM) scenarios.
\begin{figure}
\begin{subfigure}{0.2\textwidth}
\includegraphics[width=\linewidth]{5.png}
\caption{} \label{fig:2a}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.2\textwidth}
\includegraphics[width=\linewidth]{18.png}
\caption{} \label{fig:2b}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.2\textwidth}
\includegraphics[width=\linewidth]{19.png}
\caption{} \label{fig:2c}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.2\textwidth}
\includegraphics[width=\linewidth]{20.png}
\caption{} \label{fig:2d}
\end{subfigure}\hspace*{\fill}
\caption{\small{(\ref{fig:2a}) Feynman diagram of the leading order HH vector boson fusion signal physics process for this data analysis; (\ref{fig:2b}), (\ref{fig:2c}), (\ref{fig:2d}) show the different double Higgs production mechanisms which are included in the Standard Model and Beyond Standard Model theories with different values for the Higgs self-couplings C$_{2V}$, C$_{V}$, $\kappa_{\lambda}$. }}\label{fig:2}
\end{figure}
Thus, a new era of ambitious studies has been opened up. After the precise measurement of the main parameters of the SM Higgs, one of the most important objectives to be accomplished is the measurement of the Higgs self-couplings, which are strictly related to the shape of the Higgs potential (see \autoref{fig:2}).
The Higgs self-coupling studies clearly involve the investigation of the pair production of Higgs bosons \cite{potentialatcolliders}. In contrast to what happens for the dominating production mechanism of gluon-gluon fusion, the production of two Higgs bosons via a vector-boson fusion (VBF) mode turns out to be a particularly important process for the determination of the quartic-Higgs coupling C$_{2V}$. In fact, in the VBF mechanism, the Higgs bosons are produced at leading order from heavy gauge bosons that are radiated off two quarks, which can be used as tags for jets to simplify the experimental identification and measurement.
Thus, the innovation for this study is related both to the technological point of view and to the originality of the physics analysis signature considered. In fact, differently from the single Higgs boson production modes, widely explored and studied during the Run I and II at the LHC, the double Higgs boson production via VBF in the four lepton plus 2 b-jets final state has not yet been investigated. This was mainly due to the small value of its cross section weighted with the branching ratios. Indeed, for the HH production via VBF, with the Higgs mass set to its best fit value of 125.09 GeV, the cross section at 13 TeV is around 1.723 fb \cite{crossectionsfirst} while the corresponding branching ratios are 2.79 $\times$ 10$^{-4}$ for H $\to$ ZZ$^{*}$ $\to$ 4l, with l = e, $\mu$, $\tau$, and 5.75 $\times$ 10$^{-1}$ for H $\to$ $b\overline{b}$ \cite{crossections}. This requires an exclusive event selection in order to efficiently perform a background rejection mostly from SM single Higgs processes and HH gluon-gluon fusion events. Nevertheless, the analysed double Higgs boson decay mode is highly interesting because of the four charged lepton Higgs boson decay mode and the two b-jets final state inclusion. The former is one of the highest signal-to-background (S/B) ratio and has an excellent invariant mass resolution, while the latter has the highest branching ratio among the Higgs boson decay modes.
In particular, the analysis has been applied to signal and background events for the channels 4e, 4$\mu$ and 2e2$\mu$ separately and the merged samples are used for training the Deep Learning algorithm. To properly prepare the data sets, we defined a signal region corresponding to a portion of the phase space which is expected to be populated with signal events from the model of interest, the VBF double Higgs production, while having low background rates. The event passes the signal selection if it includes at least one primary vertex, a Z candidate having an invariant mass 12 $< m_{ll(\gamma)} < $ 120 GeV/c$^{2}$, a ZZ candidate from a pair of Z bosons, which do not have common leptons (non-overlapping), a number of jets higher than 4 (see \autoref{fig:3}), $\eta <$4.7 and an angular separation between each jet and lepton of $\Delta R_{jl} >$ 0.3.
\begin{center}
\begin{figure}
\begin{subfigure}{0.35\textwidth}
\includegraphics[width=\linewidth]{6.png}
\caption{} \label{fig:3a}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.38\textwidth}
\includegraphics[width=\linewidth]{7.png}
\caption{} \label{fig:3b}
\end{subfigure}\hspace*{\fill}\\
\begin{subfigure}{0.37\textwidth}
\includegraphics[width=\linewidth]{8.png}
\caption{} \label{fig:3c}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.37\textwidth}
\includegraphics[width=\linewidth]{discriminant.png}
\caption{} \label{fig:3d}
\end{subfigure}\hspace*{\fill}
\caption{\small{Examples of normalized signal and main backgrounds data set distributions for some physical observables: (\ref{fig:3a}) shows the number of jets passing the requirements in \autoref{signaltopology}; the DNN supervised learning tool is trained by having as inputs the kinematic observables p$_{T}$, $\eta$, $\phi$ for the four charged leptons and for the six highest transverse momentum $p_{T}$ (\ref{fig:3b}, \ref{fig:3c}) jets plus their b-tagging score (\ref{fig:3d}). }} \label{fig:3}
\end{figure}
\end{center}
\begin{figure}
\begin{center}
\hspace*{\fill}\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\linewidth]{10.png}
\caption{} \label{fig:5a}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\linewidth]{11.png}
\caption{} \label{fig:5b}
\end{subfigure}\hspace*{\fill}\\
\hspace*{\fill}\begin{subfigure}{0.49\textwidth}
\includegraphics[width=\linewidth]{12.png}
\caption{} \label{fig:5c}
\end{subfigure}\hspace*{\fill}
\caption{\small{Examples of Deep Neural Network (DNN) hyper-parameters scan, respectively (\ref{fig:5a}) for the activation function (relu, selu and tanh were tested), (\ref{fig:5b}) minimization loss function algorithm (sgd and adadelta) and (\ref{fig:5c}) the network structure (the maximum number of hidden layers and neurons per each layer are respectively 4 and 1000, while the minimum values are 1 and 10). On the y-axis, the maximum of the purity ($\pi$), which is defined as the ability of the classifier not to label as signal an event that is background, times the signal efficiency metrics ($\epsilon_{S}$), which is the ability of the classifier to find all the signal samples, is plotted for the testing data set. This metrics is shown against the number associated to the i-th trained DNN configuration model (out of 722 DNNs). The best model corresponds to the highest purity times efficiency metrics, i.e. the model 620. }} \label{fig:5}
\end{center}
\end{figure}
\section{Multivariate Analysis and Results}\label{comments}
Training MVA methods is a procedure that always needs to be done in multiple parallel steps. Indeed, it is hard to guarantee that, for a set of inputs, a set of MVA hyper-parameters are the optimal choice, since there is a huge number of combinations that one can build. Those configurations (including the training data set size) can strongly affect the evolution of the MVA training.
In order to optimize the study, the framework developed in Python has the feature to build up, through Keras, different DNN architectures and perform multiple parallel scans. These can run over different sets of inputs and several parameters that need to be configured for the training (see \autoref{fig:5}).
The result of each NN training can be retrieved to produce plots that are used to validate and classify the quality of each training in terms of several evaluation metrics as it is shown in \autoref{fig:4}. In this work, the maximum of the efficiency times purity metrics, $\epsilon_{S} \times \pi$, is used to choose the best model from a parallel training launched on the ReCaS computing center. In this way, we also made over-training checks by overlapping the test and training curves.
Despite the signal rarity, good discrimination results were achieved in terms of an area under the ROC curve (AUC) of $ \sim$ 98\% and an AUC discrepancy of $ \sim$ 0.1\% between test and training data sets (it means no over-fitting issues are present). The reader should note that the associated production of the Higgs boson with a t$\bar{t}$ pair (Htt) is excluded from the training because of the high similarity with our signal. A separate discriminator should be trained for that background.
\begin{figure}
\hspace*{\fill}\begin{subfigure}{0.32\textwidth}
\includegraphics[width=\linewidth]{13.png}
\caption{} \label{fig:4a}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.33\textwidth}
\includegraphics[width=\linewidth]{15.png}
\caption{} \label{fig:4b}
\end{subfigure}\hspace*{\fill} \\
\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\linewidth]{14.png}
\caption{} \label{fig:4c}
\end{subfigure}\hspace*{\fill}
\begin{subfigure}{0.34\textwidth}
\includegraphics[width=\linewidth]{16.png}
\caption{} \label{fig:4d}
\end{subfigure}\hspace*{\fill}
\hspace*{\fill}\begin{subfigure}{0.3\textwidth}
\includegraphics[width=\linewidth]{17.png}
\caption{} \label{fig:4e}
\end{subfigure}\hspace*{\fill}
\caption{\small{Discriminating performance plots of the best Deep Neural Network model in terms of the area under the ROC curve (AUC), which is the probability that a classifier will rank a randomly chosen signal instance higher than a randomly chosen background one, for the merged backgrounds (\ref{fig:4a}) and the separated ones (\ref{fig:4b}), the loss function (\ref{fig:4c}), the signal efficiency times purity (\ref{fig:4d}) and the DNN score distributions for the training and the test samples.}} \label{fig:4}
\end{figure}
\section{Conclusions and Plans}
We presented the optimization and training of a DNN algorithm for discriminating very rare SM Monte Carlo signal events from SM background ones getting optimal results in terms of the most common evaluation metrics used in ML.
A similar binary classification task will be performed for discriminating the VBF HH production in Beyond Standard Model theories from SM bkgs as we started to do in \cite{thesis}. Since these samples have enhanced cross-sections, and it will be simpler to be selected. A Htt killer should be implemented in the future development of the analysis.
In addition to the scientific originality of the work itself, from these studies several exercises can be derived for both Analysis and Computing Schools, ensuring a deeper understanding of both the particle physics itself and the use of the most suitable ML tools. These techniques have been already applied on a different analysis by the authors and accepted to Hackathon INFN 2021 Competition, targeting master and Ph.D. students \cite{Hackathon}. Indeed, the organization of educational ML hands-on tutorials using the very latest particle physics analyses for young learners will be extended to the discussed double Higgs boson decay channel.
\ack
Authors would like to thank IT resources made available by ReCaS, a project funded by the MIUR (Italian Ministry for Education, University and Re-search) in the “PON Ricerca e Competitività 2007–2013-Azione I-Interventi di rafforzamento strutturale” PONa3\_00052, Avviso 254/Ric, University of Bari.
\section*{References}
|
\subsection{History of the Universe and Cosmology}
\begin{itemize}
\item Is the Hubble constant measured with low redshift probes different from the value inferred with $\Lambda$CDM normalized to the cosmic microwave background data?
\item Is the Hubble tension a footprint of physics beyond the Standard Model?
\item What is the absolute sum of neutrino mass? (given the lower limit of 0.06~eV from oscillations) Is the hierarchy normal or inverted?
\item What are the imprints of early Universe phase transitions and inflation in the stochastic gravitational-wave backgrounds?
\item What role do ultra-high-energy cosmic rays and advances in constraint-based modelling of Grand Unified Theories play in early Universe model building?
\end{itemize}
\subsection{Cosmic Probes of Dark Matter}
\begin{itemize}
\item Is there a portal connecting the dark and visible sectors?
\item What fraction of dark matter is held in primordial black holes? Are there currently evaporating primordial black holes?
\item Does the dark sector consist of a vast ensemble of particle species whose decay widths are balanced against their cosmological abundances?
\item What is the gravitational-wave signature of dark matter?
\item What are the gravitational-wave signatures of dilute dark matter distributions?
\end{itemize}
\subsection{Astroparticle Physics}
\begin{itemize}
\item What are the properties of Standard Model particles and their interactions beyond the reach of terrestrial accelerators? \label{item:particleProperty}
\item How do neutrino flavors mix at high energies? Are neutrinos stable? Are there hidden neutrino interactions with cosmic backgrounds?
\item Could an enhancement of strangeness production in hadronic collisions be the carrier of the observed muon deficit in air-shower simulations when compared to ultra-high-energy cosmic-ray data? Alternatively, do new particles and interactions exist at the highest energies?
\item How does matter behave in the center of neutron stars? What are the physical properties of matter at ultra-high density, large proton/neutron number asymmetry, and low temperature?
\item Do the Lorentz and CPT symmetries that underpin the Standard Model break down in extreme cosmic environments?
\item Does the QED domain (extreme magnetic fields) produce exotic particles or dark matter?
\end{itemize}
\subsection{Multimessenger Synergies in Particle Astrophysics}
\begin{itemize}
\item How are particles accelerated in the cosmos to ultra-high energies? Is the cosmic ray maximum energy a fingerprint of physics beyond the Standard Model?
\item What role do hadrons play in the extreme-energy Universe?
\item How does diffuse emission from different messengers and energies contribute to cosmic evolution?
\item How are Galactic TeVatrons and PeVatrons produced? Are gamma-ray halos a signal of physics beyond the Standard Model?
\item How are heavy elements formed?
\end{itemize}
\subsection{Architecture of Spacetime}
\begin{itemize}
\item What are the true degrees of freedom in gravitational-wave polarizations, how are gravitational waves produced and how do they propagate?
\item Is there a modification of General Relativity that successfully takes into account the effects ascribed to dark matter and dark energy?
\item Does the graviton have a mass, what is the speed of gravity, and is local Lorentz invariance a fundamental symmetry of nature?
\item Does General Relativity apply to electromagnetic and gravitational wave signals from dynamic black hole environments without modification?
\item What are the ``ab initio'' models of nonsingular, horizonless alternatives to black holes, and self-consistent predictions of the ringdown spectra and echo signal they might produce?
\item What is the space of low energy Effective Field Theories that admit an UV completion? What are the phenomenological implications of the Swampland conjectures for the topics discussed in this report?
\end{itemize}
\subsection{Production of the Heavy Elements}
\label{s:r-process}
The synthesis of the elements~\cite{burbidge1957synthesis, Frebel:2018slj} in the periodic table are part of the overall Hot Big Bang theory. After baryogenesis, the neutrinos decouple in the first two seconds, nearly freezing out the neutron/proton ratio. The light elements (hydrogen, helium, deuterium) are produced within the first several minutes and, after $\sim$300,000 years, the electrons and protons in the plasma recombine into neutral atoms, allowing the CMB to stream freely, enabling a host of high-precision cosmological measurements.
Elements in the periodic table up to iron are made in the hot cores of massive stars. Heavier elements are made in the slow neutron capture process (the \emph{s-process}), but this only accounts for around half of the heavy isotopes. The rest must be created by a high-density, rapid neutron capture process (the \emph{r-process}) either in the explosions of suernovae or through the merger of neutron star binaries.
After the spectacular binary neutron star merger GW170817~\cite{Monitor:2017mdv}, there has been a renewed interest in the so-called \emph{r-process}~\cite{Cowan:2019pkx}, by which $\approx$50\% of the heavy elements in the Universe are produced. In particular, primordial black holes could leave direct observational imprints of \emph{r-process} nucleosynthesis~\cite{Fuller:2017uyd}. In order to more precisely determine the contribution of various processes to the isotopic abundances, inputs from the EM/GW observations of neutron star mergers will have to be combined with precision measurements at accelerator facilities (e.g., the Facility for Rare Isotope Beams~\cite{Balantekin:2014opa}).
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{Figures/Origin_of_the_elements.pdf}
\caption[Origin of the elements]{Nucleosynthetic sources of elements in the Solar System. Each element in this periodic table is color-coded by the relative contribution of nucleosynthesis sources, scaled to the time of the Solar System formation. Only elements that occur naturally in the Solar System are shown; artificially made elements and elements produced only through radioactive decay of long-lived nuclei are shown in grey. Taken from Ref.~\cite{Johnson:2019}.}
\label{fig:periodic_table}
\end{figure}
The periodic table shown in Fig.~\ref{fig:periodic_table} summarizes the origin of the elements in the Solar System that we see today. Cosmic nucleosynthesis is one of the the challenges ahead for the multimessenger program~\cite{Diehl:2022jnq}.
\section{Architecture of Spacetime}
\label{s:spacetime}
General Relativity (GR) is an incredibly successful theory describing the relationship between mass-energy and spacetime curvature. With the recent explosion of gravitational-wave (GW) detections, the prospects for testing the fundamental structure of spacetime are now looming closer~\cite{Berti:2022wzk}. In the sections below, we describe some of the most prominent examples.
\subsection{The Birefringence of Spacetime}
High-precision GW measurements coupled with multimessenger astronomy allow one to search for violations of GR in the propagation of waves~\cite{Berti:2022wzk}. These effects can largely be parameterized into the basis of graviton mass, dispersion in the GW propagation (to be discussed in the following sections), and birefringence of spacetime. These effects are all absent in Einsteinian gravity. However, non-zero gravition mass and dispersion or birefringence of gravitational waves can be linked to violations of Lorentz and CPT symmetry~\cite{ONeal-Ault:2021uwu}. Thus, testing whether spacetime is birefringent may amount indirectly to testing local Lorentz invariance and CPT symmetry.
While parity symmetry is conserved in GR, GW birefringence arises in effective-field-theory extensions of GR when parity symmetry is broken. This causes the left- and right-handed polarizations to propagate differently from the source to the detector. Chern-Simons gravity~\cite{Alexander:2007kv, Yoshida:2017cjl}, Ho\v{r}ava-Lifshitz gravity~\cite{Horava:2009uw}, certain scalar-tensor theories of gravity~\cite{Crisostomi:2017ugk}, and the symmetric teleparallel equivalent of GR~\cite{Conroy:2019ibo} that have been proposed to account for dark matter and dark energy typically lead to parity violation. Moreover, such violations can also arise at large enough energy scales in quantum gravity theories such as the Loop Quantum Gravity and String Theory~\cite{Alexander:2007kv}.
Future GW detectors such as the pulsar timing arrays (PTAs), space-borne interferometers, and terrestrial laser interferometers will be able to fully constrain birefringence of the spacetime structure. To do so, it is necessary to observe GWs either for a long-enough duration or with enough number of non-collocated detectors so as to resolve their polarization states. LISA will track GWs from compact binary systems for years. As of 2022, the terrestrial detector network consists of the two LIGO detectors in the U.S. and the Virgo detector in Italy. Since the two LIGO detectors are nearly co-aligned, tests for non-GR polarizations have been limited. By the end of the decade, the KAGRA detector in Japan and the third LIGO detector in India~\cite{Saleem:2021iwi} should be coming on-line, allowing for the full measurement of all polarization modes. With more sensitive detectors such as LIGO Voyager~\cite{LIGO:2020xsf}, Einstein Telescope~\cite{Punturo:2010zz}, and Cosmic Explorer~\cite{Evans:2021gyd}, it would be possible to place exceedingly tight constraints on a host of alternative theories of gravity; even more so using multiband analyses~\cite{Cutler:2019krq, Muttoni:2021veo, Gupta:2020lxa} with space-borne interferometers.
\subsection{Modified Gravity as an Alternative to Dark Energy \& Dark Matter}
The $\Lambda$CDM model, based on the theory of GR, has been very successful in explaining the observable properties of big bang nucleosynthesis, cosmic microwave background (CMB) observations, and large-scale structure. This success is achieved at the price of assuming that the energy content of the Universe today is dominated by dark energy and dark matter. However, only the large-scale gravitational interaction of the dark components has been detected so far and their fundamental properties remain largely unknown. As of today, we do not even know if the dark components are associated with new elementary particles or represent a mirage produced by modifications of the laws of gravity. Over the past decade, various discrepancies have emerged between $\Lambda$CDM predictions and cosmological observations, e.g., the tensions in the Hubble expansion rate and the clustering of matter discussed in Sec.~\ref{s:H0tension}. Several modified gravity models have been constructed to resolve the $H_0$ and $S_8$ tensions, but there seems to be no consensus on a satisfactory solution to this problem yet~\cite{Abdalla:2022yfr}.
In the next decade, GW standard sirens are expected to provide strong constraints on dark energy, modified gravity, and dark matter and shed light on several other important aspects in cosmology (see Sec.\,IXA7 of Abdalla et al.~\cite{Abdalla:2022yfr} and references therein). Imprinted in the observed GWs is the nature of gravity. Thus, any modification of gravity beyond GR will leave a fingerprint in the GW signal.
Firstly, modified gravity theories are proposed mainly to explain the late-time acceleration of the Universe (dark-energy-dominated era), but they can also induce amplitude and phase corrections on the GW signal over cosmological volumes. The time variation of the gravitational constant could be inferred using a multimessenger~\cite{Engel:2022yig} approach, exploiting the unique relation between the GW luminosity distance, BAO angular scale, and the sound horizon at decoupling~\cite{Mukherjee:2020mha}.
Secondly, by changing the gravitational interaction in a binary system, one induces a change in the generation mechanism of the gravitational radiation. Such changes can be quantified through the parameterized post-Newtonian~\cite{Arun:2006hn, Mishra:2010tp} or post-Einsteinian framework~\cite{Yunes:2009ke}. Future terrestrial and LISA observations can lead to improvements of 2--4 orders of magnitude with respect to present constraints, while multiband observations can yield improvements of 1--6 orders of magnitude~\cite{Perkins:2020tra}.
Finally, an interesting possibility is the detection of stochastic gravitational waves. The existence of a stochastic background is a robust prediction of several well-motivated cosmological and astrophysical scenarios operating at both the early and late Universe~\cite{Maggiore:1999vm, Caprini:2018mtu, Giovannini:2019oii}. As previously described, the existence of such backgrounds can be probed with GW observatories on ground and in space as well as PTAs.
\subsection{The Graviton Mass}
Regardless of the specifics of the theory one considers, there are general properties of the graviton (understood as a gauge boson that mediates the gravitational interaction) that one may wish to measure or test to ensure our description is as prescribed by Einstein's theory. One such property is the graviton's mass which, according to GR, is exactly zero. Theories such as massive gravity~\cite{deRham:2014zqa} and bi-gravity~\cite{Crisostomi:2015xia} predict a non-zero value. In fact, many modified theories created to explain the present-day cosmic acceleration also predict deviations in the propagation of GWs~\cite{Cardoso:2002pa, Saltas:2014dha, Lombriser:2015sxa, Lombriser:2016yzn, Belgacem:2017ihm, Nishizawa:2017nef, Belgacem:2018lbp} and in the gravitational lensing of GWs~\cite{Congedo:2018wfn, Mukherjee:2019wcg, Mukherjee:2019wfw, Ezquiaga:2020spg}. Gravitational waves thus have the potential to place stringent bounds on the graviton mass because a non-zero value leads to a modified dispersion relation~\cite{Will:1997bb, Kostelecky:2016kfm}. On very general grounds that rely only on special relativity, a non-zero graviton mass implies that the GW frequency does not just depend on its wave-vector, but rather also on the mass, leading to a compression of the GW train that accumulates with distance travelled~\cite{Will:1997bb}.
Current GW observations are already placing constraints on the mass of the graviton, but much more can be achieved in the next decade. Current LIGO/Virgo observations have constrained the graviton mass to be less than $4.7 \times 10^{-23} \; {\rm{eV}}/c^2$~\cite{LIGOScientific:2019fpa}. Constraints on the mass of the graviton, however, can be shown to scale as $[f_{\rm low}/(D_L \rho)]^{1/2}$, where $D_L$ is the luminosity distance, $\rho$ is the SNR, and $f_{\rm low}$ is the lowest frequency detected~\cite{Perkins:2020tra}--- this is because the larger the distance, the longer the GW train compression can accumulate for, leading to a stronger constraint. As a result, in the next few years and then in the next decade, future observations with LIGO/Virgo/KAGRA/LIGO-India and XG ground-based detectors can place constraints better than $10^{-25} \; {\rm{eV}}/c^2$ and $10^{-26}$, respectively, while space-borne detectors like LISA can improve these constraints down to $3 \times 10^{-27} \; {\rm{eV}}/c^2$~\cite{Chamberlain:2017fjl, Perkins:2020tra}. These numbers are interesting because if one associates the late-time acceleration of the Universe to a non-zero graviton mass, then the graviton would have to be of the scale of the Hubble constant, $10^{-33}$~eV. By stacking events from LISA and XG detectors we may begin to approach this scale and thus confirm or rule out a non-zero graviton mass as an explanation for the late-time acceleration of the Universe.
Another property of the graviton as a particle that one may wish to probe is its group velocity in the high-energy limit $E \gg m_g$. In Einstein's theory, this group velocity is equal to the speed of light, but in other theories of gravity, this need not be the case~\cite{Baker:2017hug, Ezquiaga:2017ekz, Creminelli2017, Sakstein:2017xjx, Boran:2017rdn, Akrami:2018yjz}. For example, in the Einstein-\AE{}ther theory, the graviton travels at a constant group speed that is faster than the speed of light, avoiding causality violations~\cite{Jacobson:2000xp, Jacobson:2007veq}. The measurement of the speed of the graviton, unfortunately, is rather difficult because it requires that we compare the time of arrival of a GW to some other baseline. This is where multimessenger events shine. If an event produces both GWs and electromagnetic (EM) waves simultaneously, then one can, in principle, compare the speed of the GWs to the speed of the EM waves (i.e., the speed of light) by comparing their times of arrival.
This is exactly what was done with the first LIGO/Virgo binary neutron star observation, GW170817, which was accompanied by a short gamma-ray burst emitted shortly after the merger~\cite{Monitor:2017mdv, Mukherjee:2019wcg, Mukherjee:2019wfw, Baker:2020apq, Ezquiaga:2020dao}. This single observation was sufficient to infer that the speed of the graviton is equal to that of the photon to better than one part in $10^{15}$. Such a measurement had the effect of severely constraining a variety of modified theories of gravity. Future terrestrial observations with LIGO/Virgo/KAGRA/LIGO-India or with XG detectors will allow for additional measurements of the speed of the graviton along other lines of sight, and thus allow us to test local position invariance~\cite{Will:2005va, Yunes:2013dva}.
With LISA, we may detect supermassive black hole (BH) binaries at mHz frequencies and measure time delays between the arrivals of photons and gravitons. This will present some advantages. First, the longer timescales of these massive mergers can facilitate triggered EM precursor observations. The inevitable periodic modulations of the EM signal due to Doppler and lensing effects during the inspiral stage arise from the same orbital motion as the GWs and can be phased in a robust way without the need to model the astrophysical source in detail~\cite{Haiman:2017szj, Tang:2018rfm}. The measurements will also provide tighter limits, due to the high SNRs and large horizon distances achievable with LISA. The frequency dependence of the time delay would further probe Lorentz-violating theories~\cite{Kocsis:2007yu, Haiman:2009te, Mirshekari:2011yq}.
Some modifications of GR, invoked to explain the present-day cosmic acceleration, predict deviations between the propagation properties of EM radiation and GWs~\cite{Saltas:2014dha, Lombriser:2015sxa, Lombriser:2016yzn, Belgacem:2017ihm, Nishizawa:2017nef, Belgacem:2018lbp, Mastrogiovanni:2020gua}. A multimessenger, data-driven measurement of the running of the effective Planck mass and its redshift dependence is possible by combining three length scales, namely the GW luminosity distance, baryon acoustic oscillations (BAO), and the sound horizon from the CMB~\cite{Mukherjee:2020mha}. Sources detectable at higher redshifts (such as supermassive BH binaries) are most useful to measure the redshift dependence and running of the effective Planck mass. Such measurements may be possible by cross-correlating binary BHs with galaxies~\cite{Mukherjee:2020mha}. General relativity propagation effects could also be probed using other techniques--- e.g., by using the mass distribution of binary neutron stars~\cite{Finke:2021eio} and BHs~\cite{Leyde:2022orh}.
\subsection{Tests of Black Hole Dynamics}
The dynamical content of the underlying theory of gravity can be probed in violent, dynamical situations giving rise to strong bursts of GW emission. After the violent merger of two compact objects leading to BH formation, GR predicts the formation of a Kerr BH, wherein the spacetime is described by only two parameters. The relaxation to this state is described by a set of exponentially damped sinusoids (``ringdown'') whose frequencies and damping times depend only on the mass and spin~\cite{Kokkotas:1999bd, Berti:2009kk}. Since GW observations provide a measurement of frequencies and damping times, the ``ground state'' quasi-normal mode (QNM) allows us to infer the mass and spin. Any measurement of additional QNM frequencies (``excited states'') can then be used as a null test of the Kerr nature of the remnant.
The idea of treating BHs as ``gravitational atoms,'' thus viewing their QNM spectrum as a unique fingerprint of spacetime dynamics (in analogy with atomic spectra), is usually referred to as ``BH spectroscopy''~\cite{Dreyer:2003bv, Berti:2005ys, Berti:2007zu, Gossan:2011ha, Meidam:2014jpa}. The seeds of this idea were planted in the 1970s~\cite[see, e.g.,][for a detailed chronology]{Berti:2009kk}. Chandrasekhar and Detweiler developed various methods to compute the QNM spectrum, identifying and overcoming some of the main numerical challenges~\cite[see, e.g.,][]{Chandrasekhar:1975zza}. In particular, Detweiler concluded the first systematic calculation of the Kerr QNM spectrum~\cite{Detweiler:1980gk} with a prescient statement: {\em ``After the advent of gravitational wave astronomy, the observation of [the BH’s] resonant frequencies might finally provide direct evidence of BHs with the same certainty as, say, the 21 cm line identifies interstellar hydrogen.”}
Early estimates~\cite{Berti:2005ys,Berti:2007zu} showed that the detection and extraction of information from ringdown signals requires events whose signal-to-noise ratio (SNR) in the ringdown {\em alone} is larger than those achievable now (for example, the first GW detection (GW150914) had a combined SNR of $24$, with an SNR $\sim7$ in the ringdown phase~\cite{TheLIGOScientific:2016wfe, TheLIGOScientific:2016src}). There are claims that overtones have been detected in GW150914~\cite{Isi:2019aib} and higher modes have been measured in GW190521~\cite{Capano:2021etf}, but the detection of modes other than the fundamental is debatable at current SNRs~\cite{Bustillo:2020buq, LIGOScientific:2020tif, Cotesta:2022pci}.
Any deviation from the QNM spectrum of classical GR would indicate substructure of BH ``atoms'' inconsistent with the standard picture. In particular, a non-singular horizonless object would lead to different boundary conditions than the classical theory and departures from the BH QNM spectrum. In any case, conclusive tests should be achievable once LIGO and Virgo reach design sensitivity, and certainly with the next-generation (XG) observatories (Cosmic Explorer or the Einstein Telescope) or with space-based detectors such as LISA~\cite{Berti:2016lat}. If the frequencies turn out to be compatible with the predictions of GR, parametrized formalisms can be used to constrain theories of gravity that would predict different spectra~\cite{Cardoso:2019mqo, McManus:2019ulj, Maselli:2019mjd, Carullo:2021dui}.
The existence and properties of horizons can be inferred and quantified with a variety of observations~\cite{Cardoso:2019rvt}. It is believed that accreting horizonless objects would reach thermal equilibrium with the environment rather quickly, whereas accreting supermassive BHs do not--- the luminosity contrast between the central accreting object and its accretion disk imposes stringent constraints on the location and property of a putative surface~\cite{Broderick:2009ph, Cardoso:2019rvt}. However, constraints based on accretion models are model-dependent and have also been questioned~\cite{Carballo-Rubio:2018jzw}. They still leave open the possibility of a surface close to the would-be event horizon, as predicted in thin-shell gravastar models~\cite{Mazur:2004fk, Mazur:2015kia, Beltracchi:2021zkt, Beltracchi:2021lez}. The planned EHT and future surveys of tidal disruption events will improve current constraints on the location of a hypothetical surface by two orders of magnitude.
The EM observations above are done, essentially, in a fixed-background context in which the BH spacetime is an arena where photons propagate. One can also consider situations probing both the background {\it and} the field equations. A stellar-mass BH or a neutron star orbiting a supermassive BH will slowly inspiral due to emission of GWs, ``sweeping'' the near-horizon geometry and being sensitive to tiny near-horizon changes, such as tidal deformability or tidal heating, or to non-perturbative phenomena like resonances of the central object~\cite{Cardoso:2017cfl, Maselli:2017cmm, Cardoso:2019nis, Maggio:2021uge, Fang:2021iyf}. Accurate tracking of the GW phase by the future space-based detector LISA may constrain the location of a putative surface to Planckian levels~\cite{Cardoso:2019rvt}.
\subsection{Alternative Black Hole Models}
The absence of an horizon can also lead to smoking-gun effects in the GW signal. An ultracompact, horizonless vacuum object sufficiently close to the Kerr geometry outside the horizon behaves as a cavity for impinging GWs, which end up being trapped between the object's interior and its light ring~\cite{Cardoso:2016rao, Cardoso:2016oxy, Cardoso:2019rvt}. Thus, perturbations of such objects, and possibly mergers as well, lead to a GW signal which is---by causality principle---similar to that emitted by BHs on sufficiently small timescales. However, at late times, the signal trapped in the ``cavity'' leaks away as a series of ``echoes'' of the original burst, which may carry a significant amount of energy. LIGO/Virgo observations have so far shown no evidence for such echoes~\cite{LIGOScientific:2020tif, LIGOScientific:2021sio}. The absence of such structure in future observations by LIGO and LISA will allow the exclusion---or detection---of any significant structure a Planckian distance away from the Schwarzschild radius, with important implications for fundamental physics~\cite{Cardoso:2019rvt}.
Setting stringent constraints on the nature of compact objects---in particular quantifying the existence of horizons in the Universe---requires advanced detectors. It is also a challenging task from the modelling and computational point of view, as one needs: \textit{(i)} a physically motivated, well-posed theory solving, at least partially, the conceptual problems of GR; \textit{(ii)} the existence in such theories of ultracompact objects which arise naturally as the end-state of gravitational collapse; and \textit{(iii)} the solution of the relevant partial differential equations describing the mergers of such objects. There is pressing need for progress on all of these fronts to confront the increasingly precise data expected from a wide variety of new experimental facilities.
\subsection{Quantum Gravity Constraints on Low-Energy Dynamics}
\label{s:swampland}
Low-energy Effective Field Theories (EFTs) are the central tool in the theoretical description of low-energy particle physics, cosmology, and gravitational theories. Modern perception has it that the SM and Einstein gravity should both be understood as leading terms in an EFT expansion. It is thus of paramount importance to understand the space of allowed low-energy (IR) EFTs. Recently, there has been significant progress in understanding what is the space of low-energy EFTs that admit an UV completion.
Despite decades of research, a full-fledged theory of quantum gravity (QG) remains elusive. Nonetheless, along the way we have learned some generic features that a QG theory should possess. The Swampland Program seeks to delineate the boundary between the landscape of EFTs that are compatible with these features of QG and the swampland of EFTss that are not~\cite{Vafa:2005ui}. The set of QG features are sometimes referred as swampland conjectures~\cite{Palti:2019pca}. As illustrated in Fig.~\ref{fig:cone}, the swampland conjectures become more constraining as the energy at which the EFT should be valid increases, picking out, in the end, a (possibly unique) QG theory.
\begin{figure}[ht]
\centering
\includegraphics[width=0.7\textwidth]{swampland_figure.pdf}
\caption{The swampland and landscape of EFTs. The space of consistent EFTs forms a cone because swampland constraints become stronger at high energies. From Ref.~\cite{vanBeest:2021lhn}.}
\label{fig:cone}
\end{figure}
The accelerating expansion of the Universe is a phenomenon that is apparently IR but intrinsically UV. This cosmological hierarchy opens up the opportunity to probe physics beyond $\Lambda$CDM and the SM by analyzing some phenomenological implications of the swampland conjectures. For example, it has been conjectured that scalar field potentials $V$ that can be derived from putative QG theories obey the bound $V' \geq c V/M_{\rm Pl}$, where $c$ is a positive and dimensionless order one constant~\cite{Obied:2018sgi}. The most obvious consequence of this constraint is that de Sitter vacua are forbidden, ruling out the cosmological constant $\Lambda$ as a source of dark energy~\cite{Agrawal:2018own}. The model building of quintessence fields playing the role of dark energy has been featured extensively through the swampland program~\cite{Abdalla:2022yfr,Vafa:2019evj,OColgain:2018czj,Colgain:2019joh}. Of particular interest here, quintessence models tend to exacerbate the $H_0$ tension~\cite{Banerjee:2020xcn}. More generally, the swampland conjectures make it difficult for fundamental theories based on compactification from extra dimensions to accommodate a period of accelerated cosmic expansion~\cite{Montefalcone:2020vlu}. Such a restriction can be avoided in models whose internal space is not conformally Ricci flat~\cite{Anchordoqui:2020sqo}, e.g., the Salam-Sezgin model~\cite{Salam:1984cj}. Within this supergravity model, dark matter could acquire a mass term which depends on the value of the quintessence field~\cite{Anchordoqui:2019amx}, thus realizing an effective dark matter-dark energy coupling which
could help to reduce (though not fully eliminate) the $H_0$
tension~\cite{Agrawal:2019dlm}. Examined separately, the axion weak-gravity conjecture~\cite{Arkani-Hamed:2006emk} leads to a bound on early dark energy models proposed to resolve the $H_0$ tension~\cite{Rudelius:2022gyu}.
On a separate track, the distance conjecture~\cite{Ooguri:2006in,Lust:2019zwm}, combined with the cosmological hierarchy and bounds on deviations from Newton's law~\cite{Lee:2020zjt}, give rise to an exponentially light tower of states with two mass scales: {\it (i)}~the mass scale of states in the tower, $m \sim \Lambda^{1/4}/\lambda$, and {\it (ii)}~the scale at which the local EFT description breaks down, dubbed the species scale, $\hat M \sim \lambda^{-1/3} \ \Lambda^{1/12} \ M_{\rm Pl}^{2/3}$~\cite{Montero:2022prj}. For $\lambda \sim 10^{-3}$, $m \sim 1~{\rm eV}$ is of the order of the neutrino scale and $\hat M \sim 10^{10}~{\rm GeV}$ coincides with the sharp cutoff observed in UHECR data. This implies that the highest-energy cosmic rays could be an incisive probe of UV physics~\cite{Anchordoqui:2022ejw}. Moreover, this framework has interesting implications for the abundance of primordial black hole dark matter~\cite{Anchordoqui:2022txe}, while the excitations of the graviton in the bulk provide an alternative dark matter candidate~\cite{Gonzalo:2022jac} and a particular realization of the DDM scenario~\cite{Dienes:2011ja} discussed in Sec.~\ref{sec:DDM}.
The study of UV constraints on IR physics is a burgeoning field, with many new conceptual and technical developments. Promising future directions are summarized in Ref.~\cite{deRham:2022hpx}.
\section*{Executive Summary}
Cosmic Probes of Fundamental Physics take two primary forms: Very high energy particles and gravitational waves (GWs). Already today, these probes give access to fundamental physics not available by any other means, helping elucidate the underlying theory that completes the Standard Model~\cite{ParticleDataGroup:2020ssz}. The last decade has witnessed a revolution of exciting discoveries such as the detection of high-energy neutrinos~\cite{IceCube:2013low, IceCube:2014stg} and gravitational waves~\cite{LIGOScientific:2016aoc}. The scope for major developments in the next decades is dramatic, as detailed in this report. For example, precise measurements of the cosmic microwave background (CMB)~\cite{Planck:2018vyg} and large scale structure hint at complications in the concordance model of cosmology, which will be subjected to independent clarification within a few years thanks to the new cosmic probe of gravitational waves~\cite{Abdalla:2022yfr}. Another cosmic probe still in the incubation stage is the discovery and exploration of the cosmic neutrino background.
The very high energy particles we exploit include cosmic rays, gamma rays, and neutrinos. Their energies enable study of particle collisions at energies far above those accessible with laboratory measurements, and their enormous propagation distances enable extremely sensitive constraints to be placed on fundamental physics, including Grand Unified Theories (GUT) and Planck-scale phenomena such as Lorentz Invariance Violation and GUT-scale dark matter~\cite{Coleman:2022abf,Ackermann:2022rqc, Engel:2022bgx, Arguelles:2022xxa, Abraham:2022jse}. Ultra-high-energy cosmic rays (UHECRs) are observed with energies above $10^{11}~{\rm GeV}$~\cite{ParticleDataGroup:2020ssz}. When a $10^{11}~{\rm GeV}$ nucleus collides with an air nucleus in the upper atmosphere, the total center-of-mass (CM) energy is $1,700~{\rm TeV}$ -- nearly twice that in Pb-Pb collisions at the Large Hadron Collider (LHC). Of order 10\% of these UHECRs may be protons and, as techniques to constrain the nature of individual UHECRs improve and proton-induced collisions can be separately identified, $p$N collisions at $1,700~{\rm TeV}$ can be isolated. This is 120~TeV in the $p$-nucleon CM even without collective effects, and a host of new phenomena can be studied in the ultra-high-energy air showers, including black hole production, quark gluon plasma, and production of new long-lived heavy particles~\cite{Coleman:2022abf}. The already-established anomalous muon production in UHECR air showers (which has eluded explanation in models tuned to LHC data)~\cite{PierreAuger:2016nfk} guarantees that discoveries will be made.
In addition to studying hadron physics at ultra-high energies in UHECR air showers, experiments such as AugerPrime~\cite{PierreAuger:2016qzd}, GCOS~\cite{Horandel:2021prj} and POEMMA~\cite{POEMMA:2020ykm} can discriminate between nucleus-, photon- and neutrino-induced showers. Observations of photons and neutrinos at such energies enable unique probes of new physics including instanton-induced decay of super-heavy relics from the Big Bang~\cite{PierreAuger:2022wzk,PierreAuger:2022ibr}, cosmic strings~\cite{Berezinsky:2011cp}, Lorentz Invariance Violation~\cite{PierreAuger:2021tog}, and axion-photon conversion in large-scale magnetic fields fields~\cite{Csaki:2003ef,Gorbunov:2001gc}.
Neutrinos and gamma rays produced when ultra-high-energy cosmic rays interact with ambient gas and thermal photon backgrounds in their source environment also provide invaluable tests of fundamental symmetries~\cite{Engel:2022bgx, Ackermann:2022rqc, Arguelles:2022xxa, Abraham:2022jse}. (The boost factor of a neutrino of energy $10^7~{\rm GeV}$ is five orders of magnitude higher than has ever been observed for a proton primary!) Furthermore, cosmic neutrinos and gamma rays provide a unique indirect probe of particle dark matter~\cite{Engel:2022bgx, Ackermann:2022rqc, Coleman:2022abf}. Next-generation gamma-ray telescopes such as the Southern Wide-field Gamma-ray Observatory (SWGO)~\cite{Albert:2019afb, SWGOPBH}, the Cherenkov Telescope Array (CTA)~\cite{CTAConsortium:2017dvg}, and the All-sky Medium-Energy Gamma-ray Observatory (AMEGO)~\cite{2020SPIE11444E..31K} foresee significant improvements in sensitivity, effective area, and field of view. The
IceCube-Upgrade \cite{Ishihara:2019aao} and other upcoming cosmic neutrino experiments at GeV energies will provide neutrino oscillation sensitivity complementary to long baseline experiments. Future PeV-EeV neutrino experiments (as summarized in Figure~\ref{fig:scales}) will advance neutrino physics at energies beyond the reach of colliders by measuring the properties of standard model particles and their interactions, such as neutrino-nucleon cross sections. Observations of neutrino flavors and neutrino-antineutrino ratios have the potential to probe beyond-standard-model (BSM) neutrino physics, such as interactions with sterile neutrinos and unknown electrically neutral mediators.
Cosmic particle physics has an important synergy with accelerator-based particle physics. For instance, forward particle production plays a crucial role in astroparticle physics~\cite{Adhikari:LoI}, which will thereby benefit enormously from the far-forward experiments at the high-luminosity LHC to be studied at the Forward Physics Facility~\cite{Anchordoqui:2021ghd,Feng:2022inv}. Measurements of forward neutrinos will provide critical information to understand the anomalous muon production observed in UHECR air showers, while multi-faceted measurements of the properties of UHECR air showers -- as are now possible with state-of-the-art UHECR observatories -- will reciprocally inform theories of forward hadron production~\cite{Anchordoqui:2022fpn}. Another example is constraints on forward charm production using LHC neutrinos, being a key input for current and upcoming generations of large-scale neutrino telescopes.
Complementing very high energy particles as probes of new physics, we have Gravitational Waves. These can be measured in several frequency bands with few-to-tens of kilometer scale interferometers (LIGO/Virgo and future terrestrian GW observatories), pulsar timing arrays (PTAs), and future space-based interferometers such as the European Space Agency's Laser Interferometer Space Antenna (LISA, NASA) to be launched around 2037~\cite{Ballmer:2022uxx}. These different techniques allow observation of mergers between neutron stars and black holes in the mass range of less than 1 to $\sim$ 3,000 solar masses (terrestrial GW observatories), the merger of $10^4$--$10^7$ solar mass supermassive black holes (LISA) and the stochastic gravitational-wave background produced by a population of supermassive black holes (PTAs).
Future gravitational-wave observatories \cite{Evans:2021gyd, Punturo:2010zz} will measure the dense matter equation of state with exquisite precision \cite{Evans:2021gyd, Kashyap:2022wzr, Bogdanov:2022faf}, probe the QCD phase transition in neutron star cores \cite{Prakash:2021wpz}, explore dark matter in astrophysical environments \cite{Berti:2022wzk, Brito:2022lmd} and potentially discover primordial black holes in the dark ages \cite{Brito:2022lmd, Ng:2021sqn}, begin a new era in precision cosmology \cite{Abdalla:2022yfr}, open a new window for probing extreme gravitational phenomena in the early Universe~\cite{Caldwell:2022qsj, Berti:2022wzk, Foucart:2022iwu, Asadi:2022njl, Achucarro:2022qrl}, and have arguably unprecedented discovery potential \cite{Evans:2021gyd}. Already, LIGO and Virgo limits on the difference in the speed of light and gravitational waves from a single, well-measured, binary neutron star merger have dramatically reduced the model-space for theories of modified gravity. The measured tidal deformability in this same merger implies that neutron stars of $\sim 1.5$ solar masses have surprisingly similar radii to neutron stars of $\gtrsim 2 M_\odot$ measured by NICER, with important implications for quark matter~\cite{Raaijmakers:2019dks}. The large number of binary neutron star and neutron star-black hole mergers to be measured by LIGO/Virgo and future GW observatories will enable the exploration of transition between baryon and quark degrees of freedom and the quark matter equation of state to be mapped out in detail, confronting theoretical physics with data on non-perturbative QCD phenomena in an entirely new regime.
GW observatories will also strongly constrain cosmology, most immediately by offering a clean, largely systematic-free measurement of Hubble constant $H_0$ that should definitively settle the question of whether the expansion rate of the universe today is the same as or different from the value obtained from CMB measurements using $\Lambda$CDM. Furthermore, Cosmic Explorer and Einstein Telescope will enable structure and the black hole mass spectrum to be measured out to $z\approx 50$ and beyond, opening a completely new regime of testing $\Lambda$CDM and determining if there are primordial black holes.
The dawn of a new Astrophysical Multimessenger Era has been heralded by the recent co-detection of gamma rays and gravitational waves in a binary neutron star merger~\cite{LIGOScientific:2017ync}, the co-detection of gamma rays and neutrinos in a blazar flare~\cite{IceCube:2018dnn} and recent examples of neutrinos consistent with production in tidal disruption events~\cite{2021NatAs...5..510S,2022PhRvL.128v1101R}. Over the next decade, simultaneous observations with different techniques promise to reveal where these extreme-energy cosmic messengers come from, and how they came to be~\cite{Engel:2022yig}. Maximal exploitation of our cosmic probes will require a level of programmatic planning for complementarity between facilities with distinct goals, not previously attempted. The United States is well poised to lead this endeavor, through investment in facilities as well as the communities of scientists and specialists that build, maintain, and utilize them, but coordination between agencies will be indispensable to realizing this potential.
\clearpage
\tableofcontents
\clearpage
\pagenumbering{arabic}
\section{The Big Questions and Goals for the Next Decade}
\label{s:questions}
\noindent
\begin{minipage}{0.42\linewidth}
$\qquad$The seventh Cosmic Frontier (CF7), named Cosmic Probes of Fundamental Physics, was asked to summarize current knowledge and identify future opportunities (both experimental and theoretical) in the use of astrophysical and cosmological probes of fundamental physics. As a result of the breadth of this area of research, CF7 has been subdivided into five main topical areas: {\it (i)}~History of the Universe and Cosmology; {\it (ii)}~Cosmic Probes of Dark Matter; {\it (iii)}~Astroparticle Physics; {\it (iv)}~Multimessenger Synergies in Particle Astrophysics; and {\it (v)}~Architecture of Spacetime. All of these areas are aligned with the primary goals of High Energy Physics in general, and the APS Division of Particles and Fields (DPF) in particular.~~
\end{minipage}
\hfill
\begin{minipage}[r]{0.56\linewidth}
\centering
\captionsetup{type=figure}
\includegraphics[width=0.97\columnwidth]{Figures/CF7-Report_cord_v10.pdf}
\vskip-5pt
\captionof{figure}{Connections between messengers and fundamental physics topics. Current and future multimessenger landscapes are indicated by solid and dashed curves, respectively.}
\vskip+5pt
\label{fig:summary}
\end{minipage}
We received 12 White Papers addressing the current challenges and future opportunities in each of these fields as they relate to the other frontiers of High Energy Physics. We have identified 27 big picture questions that will shape the course of discovery of CF7 in the coming decade. The report is organized as follows. In
this section, we lay out the big questions and goals for the next decade. In Secs.~\ref{s:cosmo}, \ref{s:DM}, \ref{s:astropart}, \ref{s:multimessenger}, and \ref{s:spacetime}, we introduce the theoretical inputs needed for addressing these questions, with each subsection corresponding to a question in Sec.~\ref{s:questions}. In Sec.~\ref{s:experiments}, we go through a description of existing and future experiments where a U.S. contribution is sought. In Sec.~\ref{s:opportunities}, we identify important opportunities for complementarity with other frontiers. Finally, in Sec.~\ref{s:DEIA} we explore some challenges and limitations that professionals experience in their daily working practice to identify strategies for expanding diversity, equity, inclusion, and accessibility. A summary of the topics and their relevance to various messengers is shown in Figure~\ref{fig:summary}.
\input{BigQuestions}
\section{History of the Universe and Cosmology}
\label{s:cosmo}
The Universe is composed primarily of matter and energy we do not understand. Dark energy makes up about 70\% of the universe and we see its effects in the acceleration of the expansion of the universe over time, especially through measurements of high-redshift supernovae, anisotropies in the Cosmic Microwave Background, and the sub-critical density of large scale structure--- the distribution of galaxies and galaxy clusters. The past decade has seen renewed recognition for discoveries in dark energy's effects on the formation of the large-scale structure. Over the next decade, measurements of the Hubble constant, the cosmic microwave background, supernovae, and large-scale structure, especially in the ultraviolet band, will challenge the new gravity theories, the concordance model of cosmology, and even the Standard Model. These cosmic probes could provide complementary information about the unification of fundamental forces.
\subsection{The Hubble Tension}
\label{s:H0tension}
The $\Lambda$CDM model, in which the expansion of the Universe today is dominated by the cosmological constant $\Lambda$ and cold dark matter (CDM), is the simplest model that provides a reasonably good account of all astrophysical and cosmological observations~\cite{ParticleDataGroup:2020ssz}. However, over the last decade, various discrepancies have emerged. In particular, local measurements of the Hubble constant $H_0 = 100 h {\rm \,km\,s^{-1}\,Mpc^{-1}}$ are increasingly in tension with the value inferred from a $\Lambda$CDM fit to the cosmic microwave background (CMB) and baryon acoustic oscillation (BAO) data~\cite{Abdalla:2022yfr}. Throughout, we refer to the {\it Hubble tension} as the $5.0\sigma$ disagreement between the value inferred by the {\it Planck} Collaboration, $H_0=\left(67.27\pm0.60\right){\rm \,km\,s^{-1}\,Mpc^{-1}}$ at 68\% CL~\cite{Planck:2018vyg}, and the latest 2021 SH0ES Collaboration distance ladder constraint based on Type Ia supernovae (SNIa) calibrated by Cepheids, $H_0=(73.04 \pm 1.04){\rm \,km\,s^{-1}\,Mpc^{-1}}$ at 68\% CL~\cite{Riess:2021jrx}. However, these are not the only $H_0$ measurements--- there are actually two sets of measurements. Remarkably, all of the indirect model-dependent estimates at early times agree between themselves (such as those inferred from CMB and BAO experiments) and a similar agreement is reached by all of the direct late-time $\Lambda$CDM-independent measurements (such as distance ladders and strong lensing). Besides, an independent determination of $H_0$, based on the calibration of SNIa using the Tip of the Red Giant Branch, leads to $H_0=(72.4\pm2.0){\rm \,km\,s^{-1}\,Mpc^{-1}}$~\cite{Yuan:2019npk} and $H_0=(69.6 \pm 0.8\,({\rm stat}) \pm 1.7\,({\rm sys})){\rm \,km\,s^{-1}\,Mpc^{-1}}$~\cite{Freedman:2020dne}, both at 68\%~CL. A measurement that is independent of SNIa, based geometric distance measurements to megamaser-hosting galaxies gives $H_0 = (73.9 \pm 3.0)~{\rm \,km\,s^{-1}\,Mpc^{-1}}$~\cite{Pesce:2020xfe}. A collection of $H_0$ measurements is shown in Fig.~\ref{fig:whisker_H0}.
\begin{figure}[!ht]
\centering
\includegraphics[width=1.00\columnwidth]{H0-tension-v3.pdf}
\caption{68\% CL constraints on $H_0$ from different cosmological probes. (Adapted from \cite{Abdalla:2022yfr}).}
\label{fig:whisker_H0}
\end{figure}
Another seemingly different, but perhaps closely related, subject is the evidence of a growing tension between the Planck-preferred value and the local determination of $\sigma_8$, which gauges the amplitude of mass-density fluctuations when smoothed with a top-hat filter of radius $8h^{-1}~{\rm Mpc}$. More concretely, it is the combination $S_8 = \sigma_8 (\Omega_m/0.3)^{1/2}$ that is constrained by large-scale structure data, where $\Omega_m$ is the present-day value of the non-relativistic matter density parameter. On the assumption of $\Lambda$CDM, the Planck Collaboration reported $S_8 = 0.830 \pm 0.013$~\cite{Planck:2018vyg}, which is in $3\sigma$ tension with the result reported by KiDS-1000: $S_8 = 0.766^{+0.020}_{-0.014}$~\cite{KiDS:2020suj}. A collection of $S_8$ measurements is shown in Fig.~\ref{fig:whisker_S8}.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{Figures/fig_s8.pdf}
\caption{Constraints on $S_8$ and its corresponding 68\% error. We show the nominal reported values by each study, which may differ in their definition of the constraints.
The definition $S_8= \sigma_8 (\Omega_{\rm m}/0.3)^\alpha$ with $\alpha=1/2$ has been uniformly used for all points. In those cases where $\alpha \neq 1/2$ has been used in some references, the value of $S_8$ with $\alpha =1/2$ was recalculated (along with the uncertainties) using the constraints on $\sigma_8$ and $\Omega_{\rm m}$ shown in those references, assuming their errors are Gaussian. This concerns only 5 CC points where the published value of $\alpha$ was different from $1/2$ and the difference from the published $S_8$ (with different $\alpha$) is very small. The rest of the points are taken directly from the published values. Taken from Ref.~\cite{Abdalla:2022yfr}.}
\label{fig:whisker_S8}
\end{figure*}
The discrepancy in the value of $H_0$ inferred from model-independent and -dependent experiments (each sensitive to different physics and systematic errors) might be a hint that the standard $\Lambda$CDM model needs to be modified. However, more data are needed before we have a final verdict. An important role in reducing systematics of $H_0$ measurements will be played by gravitational-wave (GW) standard sirens (GWSS), the GW analog of astronomical standard candles~\cite{Abdalla:2022yfr, Berti:2022wzk, Engel:2022yig}. The amplitude of GWs is inversely proportional to the luminosity distance from the source, hence they can be used in conjunction with redshift information of the source location to probe the distance-redshift relation~\cite{Schutz:1986gp, Holz:2005df, Dalal:2006qt, Sathyaprakash:2009xt}. Observations of the binary neutron star merger GW170817, along with the redshift from its host galaxy (identified from the observation of an electromagnetic counterpart~\cite{LIGOScientific:2017ync}), yield $H_0=70_{-\,8}^{+12}{\rm km\,s^{-1}\,Mpc^{-1}}$ at 68\% CL~\cite{LIGOScientific:2017adf}. Despite the fact that the measurement has large uncertainties, it does not require any cosmic ``distance ladder'' and it is model-independent--- the absolute luminosity distance is directly calibrated by the theory of general relativity~\cite{Schutz:1986gp}. In other words, these GWSS are an ideal independent probe to weigh in on the Hubble tension. Around 50 additional observations of GWSS with electromagnetic counterparts would be needed to measure $H_0$ with a precision of $1$--$2\%$~\cite{Nissanke:2013fka, Chen:2017rfc, Mortlock:2018azx}. Complementary dark GWSS (GW sources without EM counterparts)~\cite{LIGOScientific:2018gmd, DES:2019ccw, Palmese:2019ehe, DES:2020nay, LIGOScientific:2021aug} are expected to provide percent-level uncertainty on $H_0$ after combining a few hundreds to thousands of events using the statistical host identification technique~\cite{Gray:2019ksv} or by identifying the host galaxy of the nearby sources~\cite{Borhanian:2020vyr}. Improved measurements of the Hubble constant will also come from future CMB experiments~\cite{Chang:2022tzj} (including the Simon Observatory~\cite{SimonsObservatory:2018koc} and CMB-S4~\cite{Abazajian:2019eic, Abazajian:2022nyh}) which, combined with gigantic cosmic surveys (such as Euclid~\cite{EUCLID:2020syl} and Rubin~\cite{Blum:2022dxi}), are expected to measure $H_0$ with an uncertainty of about 0.15\%~\cite{DiValentino:2020zio}. A thorough discussion of probes that will help reducing uncertainties in $H_0$, $\sigma_8$, and $S_8$, as well as other anomalies of lower statistical significance (see e.g.~\cite{Webb:2022mrw}) is given in Ref.~\cite{Abdalla:2022yfr}.
\subsection{Model Building a Breakout from the Hubble Tension}
Models addressing the $H_0$ tension are extremely difficult to concoct~\cite{Abdalla:2022yfr}. One promising class of models involves a boost in the expansion rate close to the epoch of matter-radiation equality to reduce the size of the baryon-photon sound horizon $r_d$ at recombination and increase the Hubble rate inferred from the CMB. Extra relativistic degrees of freedom at recombination (scalars, Weyl fermions, and/or vector particles~\cite{Anchordoqui:2011nh, Weinberg:2013kea, Baumann:2016wac, Brust:2013ova}) parametrized by the number of equivalent light neutrino species $N_{\rm eff}$~\cite{Steigman:1977kc} is one such possibility. For three families of massless (SM) neutrinos, $N_{\rm eff}^{\rm SM} \simeq 3.044$~\cite{Mangano:2005cc, Bennett:2020zkv}, and so the contribution of extra light relics to the cosmological energy density is usually expressed as $\Delta N_{\rm eff} = N_{\rm eff} - N_{\rm eff}^{\rm MS}$. Current data are only sensitive enough to detect additional relics that froze out after the quark-hadron transition, hence CMB-S4's ability to probe times well before that transition is a major advance~\cite{Dvorkin:2022bsc}. More concretely, CMB-S4 will constrain $\Delta N_{\rm eff} <0.06$ at 95\% C.L., achieving sensitivity to Weyl fermion and vector particles that froze out at temperatures a few hundred times higher than that of the QCD phase transition~\cite{Abazajian:2022nyh}. Another promising way to decrease $r_d$ is to include what is traditionally called ``early dark energy''~\cite{Poulin:2018cxd}. This type of models posit an additional energy density that briefly bumps up the expansion rate between the epoch of matter-radiation equality and recombination.
Deviations from $\Lambda$CDM that only affect pre-recombination physics have become more tightly constrained as the CMB data improve~\cite{Jedamzik:2020zmd, Lin:2021sfs}. Modifications of the late-time Universe have also been proposed~\cite{Salvatelli:2014zta, Berezhiani:2015yta, DiValentino:2017iww, DiValentino:2017rcr, Vattis:2019efj, Agrawal:2019dlm}. The basic idea behind this class of models is also simple: the matter-dark energy equality is shifted to earlier times than it otherwise would in $\Lambda$CDM to obtain a larger value of $H_0$. The challenge for this class of models is to increase $H(z)$ as $z \to 0$ while keeping a redshift-distance relation that is compatible with that inferred from the distance ladder~\cite{Benevento:2020fev, Alestas:2021luu}. This is because calibrated type Ia supernovae fundamentally tell us about their luminosity distances from us, which depends on the integrated expansion history and not just on $H_0$.
All in all, a plethora of new ideas have been put forward to ameliorate the $H_0$ tension~\cite{Abdalla:2022yfr, DiValentino:2021izs}, but as yet, none of the extant new physics models on this front have done so to a satisfactory degree~\cite{Schoneberg:2021qvd}. The resolution of this conundrum will likely require a coordinated effort from the side of theory and phenomenology (to construct model-independent consistency tests~\cite{Bernal:2021yli}), as well as data analysis and observation (to improve computational methods that could disentangle systematics).
\subsection{Inferring the Neutrino Mass from Cosmological Probes}
The total neutrino mass can be measured with the cosmological data thanks to the cosmic neutrino background created at early times and the growth of structures at late times (see~\cite{Lesgourgues:2012uu,Lattanzi:2017ubx,Green:2021gdc}).
The main cosmological probes that we can use for this purpose are the CMB and the large scale structure (LSS) data.
The cosmic neutrino background (CNB) is formed
when neutrinos decouple, that is when the rate of the weak interaction reactions, which keep neutrinos in equilibrium with the primordial plasma, becomes less than the expansion rate of the Universe, at a temperature of about 1~MeV. After neutrino decoupling, photons are heated by electron-positron annihilation. After the end of this process, the ratio between the temperatures of photons and neutrinos will be frozen, although they cool as the Universe expands. Therefore, we expect today a CNB at a temperature of $T_\nu=(4/11)^{1/3}T_\gamma \approx 1.95$~K.
The CNB has not been directly detected and it will be the goal of the PTOLEMY project~\cite{PTOLEMY:2019hkd}. However, in the meantime, we have an indirect detection with the measurement of the effective neutrino number $N_{\rm eff} = 2.92^{+0.36}_{-0.37}$ at 95\% CL from Planck~\cite{Planck:2018vyg}, at many $\sigma$ different from zero.
Neutrinos are the only particles in the standard model to have the transition between the relativistic and the non-relativistic regime.
When neutrinos are relativistic, they will contribute to the radiation content of the universe.
When they become non-relativistic, they will behave like matter contributing to the expansion of the Universe like baryons and cold dark matter. Neutrinos will only cluster on scales larger than their free streaming scale, thereby suppressing the structure formation at small scales, and affecting the large scale structures.
Since the CMB is formed at recombination, the effect of the neutrino mass can only manifest itself by changing the evolution of the background and introducing some secondary anisotropy corrections~\cite{Lesgourgues:2012uu,Lattanzi:2017ubx}.
Indeed, by varying their total mass we are changing the redshift of the matter-to-radiation equality $z_{eq}$, and the amount of matter density today $\omega_m = \omega_b + \omega_{cdm} + (\Sigma m_\nu)/93.14$~eV.
Therefore, the impact on the CMB will be the shift in the position of the peaks, the slope and the amplitude of the low-$\ell$ multipoles of the spectrum and the first peak, due to the ISW effect, and the damping of the high-$\ell$ tail, due to the lensing effect.
From Planck temperature and polarization spectra we have a very important upper limit on the total neutrino mass $\Sigma m_\nu<0.26$~eV~\cite{Planck:2018vyg}. This strong limit is completely due to the CMB gravitational lensing, indicating that we have a clear detection of this signal in the CMB spectra~\cite{Kaplinghat:2003bh}. In fact, the more massive the neutrino, the fewer structures we have, the less the CMB gravitational lensing should be. So a larger signal of the CMB lensing means a smaller neutrino mass. Given the neutrino effect on the structure formation, important observables are the LSS data, in particular the power spectrum of the non-relativistic matter fluctuations in Fourier space $P(k,z) = \langle | \delta_m (k,z)|^2 \rangle $, where $\delta_m = \delta \rho_m / \bar \rho_m$, and the two-point correlation function in the configuration space.
The shape of the matter power spectrum is a key observable for constraining neutrino masses with cosmological methods, and can be obtained with measurements of the CMB gravitational lensing, the clustering and the weak lensing of galaxies, and the galaxy cluster abundance~\cite{Hu:1997mj,Cooray:1999rv,Abazajian:2011dt,Chabanier:2019eai,TopicalConvenersKNAbazajianJECarlstromATLee:2013bxd,Green:2021gdc}.
Unfortunately, this is really difficult to derive in non-linear and mildly non-linear scales, and needs the help of perturbation theory or N-body simulations. On the other hand, the BAO peak of galaxy correlation function, corresponding to the acoustic scale at decoupling, is one of the prominent observables in today's cosmology and easier to obtain, and is very sensitive to massive neutrinos.
Combining Planck + BAO we get $\Sigma m_\nu<0.13$~eV~\cite{Planck:2018vyg}, because BAO data is directly sensitive to the free-streaming nature of neutrinos and helps in breaking the degeneracies among cosmological parameters.
Another important cosmological probe is the Redshift Space Distortions (RSD), which is obtained by analysing the clustering in redshift space, and it is the result of an anisotropic clustering along the radial direction because of the peculiar velocities~\cite{Hamilton:1997zq}. This RSD measures f$\sigma_8$, that is the product of the growth rate of structure (f) and the clustering amplitude of the matter power spectrum ($\sigma_8$). Massive neutrinos prefer a lower value for the f$\sigma_8$ data, so the inclusion of the latest RSD from eBOSS DR16~\cite{eBOSS:2020yzd} gives $\Sigma m_\nu < 0.087$~eV at 95\% CL~\cite{DiValentino:2021hoh}, disfavouring the minimal value allowed for Inverted Ordering (IO, $\Sigma m_\nu \gtrsim 0.1$~eV) at more than 2$\sigma$, but also the Normal Ordering (NO, $\Sigma m_\nu \gtrsim 0.06$~eV) at more than 68\% CL ($\Sigma m_\nu < 0.037$~eV). Current cosmological data do not allow to distinguish the ordering of the neutrino masses, but may give a preference for the NO when combined with oscillation and not oscillation data~\cite{Gerbino:2016ehw,Gariazzo:2018pei,Capozzi:2021fjo,Jimenez:2022dkn,Gariazzo:2022ahe}. It is worth underlining that in fact the total neutrino mass preferred by the cosmological data is zero or negative~\cite{eBOSS:2020yzd}, and
although this is not yet statistically significant, it shows a first hint of tension between cosmology and neutrino oscillation experiments.
These constraints could be drastically improved in the future.
Terrestrial CMB telescopes are currently the proposals with the highest probability of being realised. However, they need large angular scale measurements (such as Planck or future experiments) to measure the optical depth, that is strongly correlated with the neutrino masses, and a perfect a priori knowledge of the foregrounds.
The Simons Observatory~\cite{SimonsObservatory:2019qwx} aims to measure the total neutrino mass with an uncertainty $\sigma (\Sigma m_\nu)$ = 0.04~eV when combined with DESI BAO~\cite{DESI:2016fyo} and Rubin LSST~\cite{LSSTDarkEnergyScience:2018jkl} weak lensing data. The replacement of Planck with LiteBIRD’s future cosmic variance-limited measurements of the optical depth to reionisation SO can instead reach $\sigma (\Sigma m_\nu)$ = 0.02~eV.
CMB-S4 measurements~\cite{Chang:2022tzj} of the lensing power spectrum (or cluster abundances), when combined with BAO from DESI and the current measurement of the optical depth from Planck, will provide a constraint on the sum of neutrino masses with a $\sigma (\Sigma m_\nu)$ = 0.024~eV, improving to $\sigma (\Sigma m_\nu)$ = 0.014~eV with better measurements of the optical depth.
PICO~\cite{NASAPICO:2019thw}, a proposal for a future CMB satellite experiment, plus BAO from DESI (or Euclid) should reach an uncertainty $\sigma (\Sigma m_\nu)$ = 0.014~eV, i.e. a 4$\sigma$ detection of the minimum sum for the NO. A satellite experiment is the only instrument that can measure very precisely all these neutrino properties together with the optical depth with the same single dataset without calibration problems.
Finally, CMB-HD~\cite{CMB-HD:2022bsz}, a futuristic millimetre-wave survey, could get an uncertainty on $\sigma (\Sigma m_\nu)$ = 0.013~eV (at least 5$\sigma$ detection for the sum of the neutrino masses), measuring the gravitational lensing of the CMB and the thermal and kinetic SZ effect on small scales.
All of these constraints have been obtained by assuming the $\Lambda$CDM cosmological model, which is the mathematically simplest model among those introduced in the literature. However, it cannot yet explain key pillars in our understanding of the structure and evolution of the Universe, namely, dark energy, dark matter and inflation. For this reason, the anomalies and tensions we see between some parameters coming from different cosmological probes may indicate the need for a paradigm shift~\cite{Abdalla:2022yfr}. We have, in fact, the $H_0$ tension at 5$\sigma$ and the $S_8$ tension at the level of $2-3\sigma$, and both these parameters are very important for the determination of the total neutrino mass because they are strongly correlated with it.
Furthermore, we have the $A_{lens}$ consistency check which fails in the Planck data and, due to their correlation, the upper limits on the total neutrino mass are strongly weakened, up to a factor of 2 when $A_{lens}$ is free to vary.
Finally, the global tensions between CMB datasets, at the level of 2.6$\sigma$ assuming the $\Lambda$CDM model and the Suspiciousness statistics~\cite{Handley:2020hdp}, is translated in $1\sigma$ preference from the terrestrial CMB telescopes ACT-DR4 and SPT-3G data for a neutrino mass different from zero, which is very similar to the value obtained from a combination of Planck + CMB Lensing, when $A_{lens}$ is free to vary~\cite{DiValentino:2021imh}. Furthermore, if the neutrino limits are obtained in a 10 parameter model, ACT-DR4 and SPT-3G can host even larger neutrino masses~\cite{DiValentino:2021imh}, and when CMB and BAO constraints are considered in these extended cosmologies, they provide constraints on the $\Sigma m_\nu$ vs $H_0$ plane that clearly show a correlation between these two parameters, that is exactly the opposite of what is obtained with the standard $\Lambda$CDM model, allowing this combination to solve the $H_0$ tension with massive neutrinos.
To conclude, the indication for anomalies and tensions present in the cosmological data could significantly influence the current cosmological constraints on the neutrino properties, presenting a serious limitation to precision cosmology.
Until the nature of these anomalies (whether new physics or systematic errors) is clear, we should be very cautious when considering cosmological constraints.
\subsection{Imprints from the Early Universe on the Gravitational Wave Background}
\paragraph{Cosmic (super)strings}
Cosmic strings are topological defects that can form during phase transitions in the early Universe~\citep{Kibble:1976sj, Vilenkin:2000jqa}, and cosmic superstrings are the fundamental strings of string theory stretched to cosmological scales due to the expansion of the Universe~\citep{Jones:2002cv, Sarangi:2002yt, Dvali:2003zj, Jones:2003da, Copeland:2003bj, Jackson:2004zg}. In a cosmological setting, and for the simplest superstring models, cosmic string and superstring networks evolve in the same way. For a detailed review of cosmic (super)string network evolution and observational signatures, see, e.g.,~\citep{Copeland:2009ga}. Cosmic (super)strings can exchange partners when they meet and produce loops when they self-intersect. These loops then oscillate and lose energy to GWs generating bursts and a stochastic background~\citep{Berezinsky:2001cp,Damour:2000wa,Damour:2001bk,Damour:2004kw,Siemens:2006vk,Siemens:2006yp}--- signals that can be potentially detected by space-based and terrestrial GW detectors and pulsar timing arrays~\citep{Pastorello:2019akb, Blanco-Pillado:2017rnf}. Strings are characterized by their mass per unit length $\mu$, which is normally given in terms of the dimensionless parameter $G\mu/c^2$, the ratio of the string energy scale to the Planck scale squared.
The cosmic string GW spectrum is broad-band, spanning many orders of magnitude in frequency, and hence accessible to a number of GW experiments including the Laser Interferometer Gravitational-wave Observatory (LIGO)~\cite{Harry:2010zz} and Virgo~\cite{VIRGO:2014yos}, Laser Interferometer Space Antenna (LISA)~\cite{LISA:2017pwj}, and the Pulsar Timing Arrays (PTAs)~\cite{Hobbs:2009yy}. PTAs are currently the most sensitive experiment for the detection of cosmic (super)strings and have set the most stringent bounds on the energy scale and other model parameters. The best limit on the string tension, $G\mu/c^2 < 5.3(2) \times 10^{-11}$, is several orders of magnitude better than constraints from CMB data, and comes from the NANOGrav Collaboration~\citep{NANOGRAV:2018hou}. In fact, PTAs might have already seen the first hints of a stochastic background~\cite{NANOGrav:2020bcs}. A definitive detection of GWs from cosmic (super)strings would be transformative for fundamental physics which could be enabled by PTAs over the next decade.
\paragraph{Primordial gravitational waves from inflation}
The inflationary paradigm that the very early Universe saw a period of exponential expansion accounts for the observed homogeneity, isotropy, and flatness~\citep{Brout:1977ix, Starobinsky:1980te, Kazanas:1980tx, Sato:1980yn, Guth:1980zm, Linde:1981mu, Albrecht:1982wi}. Additionally, by expanding quantum fluctuations present in the pre-inflationary epoch, inflation seeds the density fluctuations that evolve into the large-scale structures we see in the Universe today~\citep{Mukhanov:1981xt, Hawking:1982cz, Guth:1982ec, Starobinsky:1982ee, Bardeen:1983qw} and produces a stochastic background of GWs~\citep{Starobinsky:1979ty, Rubakov:1982df, Abbott:1984fp}, which have so far eluded detection. This GW background is broad-band, like the one produced by cosmic strings, and potentially detectable by multiple experiments.
Detecting primordial GWs from inflation has been a critical objective of CMB experiments for some time~\citep{Kamionkowski:2015yta}. The CMB is sensitive to the lowest frequency portion of the GW spectrum from inflation, and CMB data can be used to constrain the tensor-to-scalar ratio, which is the ratio of the size of GWs produced to that of scalar perturbations that seed density fluctuations, as described above. For standard inflation models, the GW background in the PTA band is likely to be fainter than that of supermassive black hole binaries depending on the nature of the latter at lowest frequencies~\citep{Sampson:2015ada}. In addition, some inflationary models have a spectrum that rises with frequency. Thus, GW detectors operating at higher frequencies than CMB experiments, like PTAs and space-based and terrestrial interferometers, can be used to constrain the shape of the inflationary GW spectrum. Indeed, PTA, CMB, and GW interferometer data across 29 decades in frequency have already begun to place stringent limits on such models~\citep{Lasky:2015lej}, though future observations are necessary to detect the background or tighten the constraints on model parameters.
\paragraph{Gravitational waves from phase transitions}
The early Universe may have experienced multiple phase transitions as it expanded and cooled. Depending on the detailed physical processes that occur during a phase transition, GWs can be generated with wavelengths of order the Hubble length at the time of the phase transition. That length scale, suitably redshifted, translates into a GW frequency today. Thus, GW experiments at different frequencies today probe horizon-sized physical processes that occurred at different times in the early Universe, with higher frequency experiments probing earlier and earlier times.
For example, the nanohertz frequency band accessible to PTAs maps onto the era in the early Universe when the quantum chromodynamics (QCD) phase transition took place, about $10^{-5}$~s after the Big Bang. The horizon at that time was on the order of 10~km, and any GWs generated at that length scale at that time would today be stretched to about 1~pc (or 3~light-years), which corresponds to GW frequencies of about 10~nHz, and lie within the PTA sensitivity band. The possibility that interesting QCD physics can result in a GW signal detectable by PTAs was first pointed out by Witten in the 1980s~\citep{PhysRevD.30.272}. More recently, Caprini et al.~\citep{Caprini:2010xv} considered the possibility of a first-order phase transition at the QCD scale. In standard cosmology, the QCD phase transition is only a cross-over and we do not expect it to generate GWs. However, if the neutrino chemical potential is sufficiently large, it can become first order (it is worth pointing out that if sterile neutrinos form dark matter, we expect a large neutrino chemical potential). There is also the possibility that the fluctuations of gluon fields could generate scalar GWs from the conformal anomaly in the quark-gluon plasma phase~\cite{Mottola:2016mpl}. Thus, PTAs provide a window onto physical processes occurring in the Universe at the time of the QCD phase transition or before and could detect GWs from a first order phase transition at that time.
\subsection{Ultra-High-Energy Cosmic Rays as Probes of the Early Universe}
\label{s:SHDM}
The motivations for super-heavy dark matter (SHDM) particles were recently revived by the possibility that new physics could only manifest at the Planck scale or at the scale of Grand Unified Theories (GUTs)~\cite{Coleman:2022abf, Anchordoqui:2021crl}. This possibility is motivated not only by the absence of any sign of new physics at the TeV scale, but also by the precise measurements of the mass of the Higgs boson and the Yukawa coupling of the top quark that make it possible to extrapolate the SM all the way to the Planck mass without encountering any inconsistency that would make the electroweak vacuum of the SM unstable. This vacuum lies, in fact, close to the boundary between stability and metastability~\cite{Degrassi:2012ry}.
Super-heavy dark matter particles that are only gravitationally coupled could have been produced at the end of inflation via the freeze-in mechanism, which relies on annihilations of SM fields to populate the dark sector. An interesting consequence is that, so as to produce enough such very weakly coupled heavy particles, the reheating temperature must be relatively high, which implies a tensor/scalar ratio of the primordial modes possibly detectable in the power spectrum of the CMB. The limits inferred from the Planck satellite on this ratio thus constrain the possible phase space for the mass of the particles and the value of the Hubble rate at the end of inflation~\cite{Garny:2015sjg}.
Another possibility to constrain these models is to look for the secondary products produced via particle decay. In the minimalist benchmark described above, dark matter (DM) particles are protected in the perturbative domain by the conservation of quantum numbers, and so would only decay through non-perturbative effects. One of these effects is due to the non-trivial vacuum structure of non-commutative gauge theories and the possibility of the generation of one quantum number for the benefit of another through the change of configuration of gauge fields by tunnel effect (instantons)~\cite{Kuzmin:1997jua}. This mechanism offers the possibility of providing metastable particles, which can produce detectable secondaries.
If SHDM particles decay into SM fields, then a flux of ultra-high-energy photons could be observed preferentially from regions of denser DM density, such as the center of our Galaxy~\cite{Aharonian:1992qf, Berezinsky:1997hy, Birkel:1998nx, Evans:2001rv, Aloisio:2006yi}. AugerPrime is in a prime position to collect a large exposure in the direction of the Galactic center~\cite{PierreAuger:2016qzd}. Indeed the non-observation of photons in Auger data has allowed limits to be set on the gauge coupling in the dark sector~\cite{PierreAuger:2022wzk,PierreAuger:2022ibr}, which are complementary to those obtained via the tensor/scalar ratio of the primordial modes. With increased sensitivity to the tensor-to-scalar ratio on the one hand and to ultra-high-energy photons thanks to the planned extreme-energy cosmic-ray observatories in the next decade~\cite{POEMMA:2020ykm, Horandel:2021prj} on the other, the GUT parameter space will continue to shrink towards the low-mass particle range and/or small gauge coupling values.
While the observation of ultra-high-energy photons could open a window to explore high-energy gauge interactions and possibly GUTs in the early Universe, the observation of extreme-energy neutrinos could provide a method of searching for strongly coupled string moduli, which complements searches based on gravitational effects of cosmic strings (including structure formation, CMB data, gravitational radiation, and
gravitational lensing)~\cite{Anchordoqui:2018qom}. In particular, the future Probe of Extreme Multi-Messenger Astrophysics (POEMMA)~\cite{POEMMA:2020ykm} will be able to detect extreme energy neutrinos from cosmic strings with $G \mu/c^2 \sim 10^{-20}$~\cite{Berezinsky:2011cp}. Thus, POEMMA will be sensitive to dimensionless string tensions down to 9 orders of magnitude below the current upper limit from the NANOGrav Collaboration~\citep{NANOGRAV:2018hou}.
\section{Cosmic Probes of Dark Matter}
\label{s:DM}
A second component of the dark, or hidden, sector is dark matter. Dark matter is distinct from dark energy in that it is evidenced by its gravitational pull on celestial objects, but its nature has otherwise eluded searches during decades. However, over the past decade, new tools and techniques to search for dark matter have come to the fore to probe new parameter spaces. Searches for primordial black holes, which should radiate both baryonic and dark matter as they decay, are prominent in high-energy gamma rays, and gravitational-wave facilities now look for evidence of dilute distributions and coalescence of dark matter. Connections between Standard Model particles and the hidden sector may now be revealed through a variety of messengers, including key measurements of cosmic rays, neutrinos, gravitational waves, gamma rays, and other photon wavelengths.
\subsection{Connection between Visible and Hidden Sectors}
\label{s:portals}
The nature of dark matter (DM) is one of the great fundamental puzzles of particle physics and cosmology~\cite{Boddy:2022knd,Ando:2022kzd,Carney:2022gse,Aramaki:2022zpw,Leane:2022bfm,Baryakhtar:2022hbu}. The DM distribution in galaxies and other virialized systems is a powerful indicator of its nature and a portal towards understanding the DM phenomenon~\cite{Salucci:2018hqu}. The annihilations and decays of DM could produce visible particles over a wide range of energy scales, which subsequently decay producing a range of visible secondary particles. Long-standing efforts have been dedicated to searches for such signals in photons, cosmic rays, and neutrinos, and future experiments offer the prospect of significantly improved sensitivity.
Indirect searches for DM based on gamma-ray, cosmic-ray, and neutrino signals are highly complementary~\cite{PerezdelosHeros:2020qyt}. For example, the production of high-mass quarks and gluons leads to copious production of gamma rays, neutrinos, antiprotons, and antinuclei, while the production of electrons or muons leads to strong signals in searches for cosmic-ray positrons. Dark matter decaying or annihilating into neutrinos can be well constrained by high-energy neutrino telescopes. In scotogenic models where the neutrinos mass is achieved via interaction with DM, neutrinos might be the principal portal to the dark sector. Combining constraints from all these channels allows us to avoid blind spots in sensitivity, and probe the lifetime or annihilation rate of DM in broadest possible range of scenarios. UHECR experiments could be also sensitive to interactions induced by macroscopic dark quark nuggets~\cite{Bai:2018dxf} in the atmosphere, offering further windows to identify the nature of DM~\cite{Coleman:2022abf, Anchordoqui:2021xhu}.
Searches for DM often rely critically on an understanding of astrophysical backgrounds or systems, including diffuse astrophysical backgrounds, as discussed in Sec.~\ref{subsec:diffuseBG}, and emission by individual astrophysical sources, as discussed in Sec.~\ref{subsec:TeVatronPeVatron}. Poorly understood systematic errors associated with multimessenger astrophysics can be the major limiting factor for sensitivity to dark matter signals. In the event of a possible detection of DM in an astrophysical data set, searches for multimessenger counterpart signals will be crucial in determining whether the apparent detection is truly associated with DM, and if so, determining the properties of that DM.
\subsection{Primordial Black Holes as Dark Matter}\label{s:PBHs}
Primordial Black Holes (PBHs) have long been considered as plausible cold DM (CDM) candidates~\cite{Carr:1975qj, Brito:2022lmd}, potentially forming a significant fraction of the DM~\cite{Bird:2016dcv, Ali-Haimoud:2017rtz, Raidal:2017mfl, Raidal:2018bbj, Vaskonen:2019jpv, Atal:2020igj, DeLuca:2020qqa, Wong:2020yig, Franciolini:2021tla}. The detection of binary black holes of masses in excess of 30 $M_\odot$ has brought renewed attention to this possibility by positing that progenitor black holes in these systems could be primordial in origin~\cite{Bird:2016dcv, Clesse:2016vqa, Sasaki:2016jop}. It is not possible to obtain conclusive evidence for PBHs in these detections as astrophysical models are able to explain their existence~\cite{Clesse:2020ghq}. The primordial origin of binary black holes would be compelling if either of their masses are below one solar mass or if they arise in the dark ages when stars could not have produced black hole binaries~\cite{Ng:2021sqn, Ng:2022agi, Raidal:2017mfl, Raidal:2018bbj, Vaskonen:2019jpv, Atal:2020igj, Mukherjee:2021ags, Mukherjee:2021itf, Ng:2021sqn, Franciolini:2021xbq}. PBHs can also be distinguished on the basis of their source properties such as mass, and eccentricity spin, the redshift evolution of BBH merger rates, and their spatial distribution, though a firm detection is required.
PBHs have been predicted as a generic outcome of density perturbations in the early Universe~\cite{1967SvA....10..602Z, 1971MNRAS.152...75H, 1975ApJ...201....1C, 1980PhLB...97..383K, 1985MNRAS.215..575K, Carr:2005zd, Clesse:2016vqa, Sasaki:2018dmp, Sasaki:2016jop, Raidal:2017mfl, Raidal:2018bbj, Vaskonen:2019jpv, Gow:2019pok, Jedamzik:2020ypm, Jedamzik:2020omx, DeLuca:2020jug, Atal:2020igj, DeLuca:2020qqa, Clesse:2020ghq}. The formation of black holes in the early Universe appears to be quite generic~\cite{Garcia-Bellido:1996mdl} and does not require special conditions such as large density fluctuations of matter. Large, non-Gaussian exponential tails in the density fluctuations arising from quantum processes during inflation could produce them~\cite{Ezquiaga:2019ftu}. Moreover, even the known thermal history of the Universe may play an important role by providing the required lack of pressure to allow gravitational collapse at certain well-defined epochs in the evolution of the Universe~\cite{Carr:2019kxo}; these are the Electroweak and quantum chromodynamics (QCD) epochs at the time of $e^+ e^-$ annihilation, which generate a multimodal PBH mass function with peaks at $10^{-5},\ 1,\ 10^{2},\ \mathrm{and}\ 10^{6}\ M_\odot$. These black holes come with different fractional abundances depending on the underlying inflationary potential and this may be used as a window into the early Universe and fundamental physics. For example, if an excess of $10^{-10}\ M_\odot$ is found in microlensing events, or in the induced Stochastic Gravitational Wave Background (SGWB) at LISA frequencies, we may infer the existence of new fundamental particles at scales above those reached by present particle physics accelerators, which become non-relativistic and momentarily decrease the radiation pressure at a time when the mass within the horizon is precisely that mass scale.
The possibility that PBHs may be lurking in the dark Universe as building blocks of the CDM fluid is extremely attractive~\cite{Clesse:2017bsw}. In fact, the non-Gaussian exponential tails mentioned above may give rise to enhanced clustering, which leads, after recombination, to a population of PBH clusters with intermediate masses of order $10^{6}\ M_\odot$ that could be searched for with the microlensing of quasars around clusters or the perturbations they induce on stellar tidal streams around our galaxy and Andromeda~\cite{Montanari:2020gcr}. These clusters could explain where most of the mass in the halo of galaxies is, thus evading the microlensing limits coming from stars in the Large Magellanic Cloud and the Galactic bulge~\cite{Calcino:2018mwh}, which had been used in the past to rule out PBHs as the main component of CDM~\cite{Wyrzykowski:2019jyg}. Moreover, such dense objects may help explain many unexpected correlations in the radio and X-ray backgrounds at high redshift~\cite{Kashlinsky:2016sdv, Kashlinsky:2019kac}, as well as the unusually high number of massive galaxies and quasars at high redshift, unaccounted for by the standard $\Lambda$CDM scenario.
There is, nowadays, a great opportunity for testing all these ideas with new astrophysical and cosmological observations. For example, if PBHs existed before recombination, they should have left their imprint in an excess of injected energy in the plasma in the form of spectral distortions at high frequencies that a CMB experiment dedicated to it could detect~\cite{Chluba:2019kpb}. One could further use the James Webb Space Telescope (JWST) to look for the first stars and galaxies at redshifts bigger than 10 or 20, confirming the role of black holes in early star formation \cite{Hasinger:2020ptw,Cappelluti:2021usg}.
Moreover, GWs may also allow us to detect PBHs over a wide range of masses, being complementary to other proposed probes and able to distinguish between BHs of astrophysical origin and PBHs using either resolved events~\cite{Hall:2020daa, Wong:2020yig, DeLuca:2021wjr, Hutsi:2020sol, Franciolini:2021tla} or the stochastic GW background~\cite{Mandic:2016lcn, Mukherjee:2021ags, Mukherjee:2021itf}. With the advent of the next generation of GW antennas, like the Cosmic Explorer~\cite{Evans:2021gyd} and Einstein Telescope~\cite{Punturo:2010zz, Maggiore:2019uih} on the ground and LISA in space~\cite{Barausse:2020rsu}, we should be able to reach black hole fusions at redshifts $z\sim 100,$ where no plausible stellar evolution could have generated such a population, and thereby convincingly proving their primordial nature. A further hint at their primordial origin, which would link their formation with the cosmic history, would be the discovery of the induced SGWB from second-order perturbations of large-amplitude fluctuations entering at the same time as the formation of PBHs~\cite{Garcia-Bellido:2017aan}, when the size of the horizon redshifted today gives LISA frequencies (mHz), or perhaps in the PTA range (nHz). Such a discovery would open a new window into the early Universe, where we could explore independent constraints on the existence of PBHs~\cite{Kohri:2018awv, Espinosa:2018eve, Wang:2019kaf} such as the non-Gaussian character of the fluctuations giving rise to the PBH mass spectrum, as well as the number of relativistic species present at that time~\cite{Carr:2019kxo}, well beyond the present reach with particle accelerators. In addition, terrestrial GW detectors, LISA and PTAs, could give us conclusive evidence of whether or not PBHs form a significant fraction of DM in a wide range of masses~\cite{Singh:2020wiq,LIGOScientific:2021job,LIGOScientific:2019kan}.
The detection of PBHs of any size would acutely constrain our understanding of the physics of the early Universe. Such a monumental reward motivates the search for signs of PBHs across all facets of the multimessenger landscape, including the hunt for gamma-ray and neutrino signatures of PBH evaporation. The prediction that a black hole will thermally radiate (evaporate) with a blackbody temperature inversely proportional to its mass was first calculated by Hawking~\cite{Hawking1974}--- the emitted radiation consisting of all fundamental particles with masses less than $\sim$T$_\mathrm{BH}$~\cite{MacGibbon1990}. While Hawking radiation for black holes in the stellar mass range and above is nearly negligible, this process dominates the evolution of lower-mass PBHs over time. PBHs with initial masses of $\sim$10$^{14}$--10$^{15}$~g should be expiring today, producing short bursts (lasting a few seconds) of high-energy radiation in the GeV--TeV energy range~\cite{MacGibbon2008, Ukwatta:2015tza}. Their final moments would thus be an ideal phenomenon to observe with current space-based and terrestrial gamma-ray telescopes, as well as neutrino observatories~\cite{FermiPBH1, FermiPBH2, HAWCPBH, Tavernier2021HESSPBH, Archambault:2017asc, Dave:2019epr}. Improvements in sensitivity, effective area, and field of view seen with the proposed next generation of gamma-ray telescopes, such as the Southern Wide-field Gamma-ray Observatory (SWGO)~\cite{Albert:2019afb, SWGOPBH}, the Cherenkov Telescope Array (CTA)~\cite{CTAConsortium:2017dvg}, and the All-sky Medium-Energy Gamma-ray Observatory (AMEGO)~\cite{2020SPIE11444E..31K}, present a boundless new frontier for discovery beyond the Standard Model and characterization of early Universe conditions with PBHs. While this mass regime is not currently a candidate for PBHs as dark matter, confirmation of an evaporation signal from a PBH of any size would lend significant credence to that dark matter model.
For reference, it is worth pointing out the Snowmass paper on Primordial Black Holes, Ref.~\cite{Bird:2022wvk}, which specifically focuses on these natural candidates for dark matter, describes the science cases (the origin of PBH dark matter in the early Universe) and the existing observational constraints, and expands on the theoretical work and data analysis required to improve the constraints and/or enable a possible detection in the future.
\subsection{Properties of Non-Minimal Dark Sectors}
\label{sec:DDM}
For several decades, it has been suspected that the dark sector consists of one stable weakly interacting massive particle. However, some critical thinking was recently adopted to build up a more generic view of the hidden sector in which a given dark matter particle need not be stable if its abundance at the time of its decay is sufficiently small. Dynamical Dark Matter (DDM) is a framework for non-minimal dark sectors which posits that the dark matter in the Universe comprises a vast ensemble of interacting fields $\chi_\ell$ with a variety of different masses, lifetimes, and cosmological abundances~\cite{Dienes:2011ja}. In general, the mass spectra and corresponding lifetimes and abundances of the individual states within the DDM ensemble turn out to be tied together through scaling relations involving only a few scaling exponents. As a result, the DDM ensemble is described by only a few free parameters, rendering the DDM framework every bit as constrained and predictive as more traditional dark matter scenarios.
The DDM framework might be experimentally tested and constrained through dark matter direct- and indirect-detection~\cite{Dienes:2012cf, Dienes:2013lxa, Boddy:2016fds, Boddy:2016hbp} experiments, and at colliders~\cite{Curtin:2018ees, Dienes:2019krh, Dienes:2021cxr}. Since there may be a large number of transitions between the ensemble of DDM states, there may be a variety of lifetimes and long-lived particles which, on decay, can produce spectacular signals at the Forward Physics Facility (FPF)~\cite{Feng:2022inv}. DDM scenarios can also leave observable imprints across the cosmological timeline, stretching from structure formation~\cite{Dienes:2020bmn, Dienes:2021itb} all the way to late-time supernova recession data~\cite{Desai:2019pvs} and unexpected implications for evaluating Ly-$\alpha$ constraints~\cite{Dienes:2021cxp}. Such dark sectors also give rise to new theoretical possibilities for stable mixed-component cosmological eras~\cite{Dienes:2021woi}. DDM scenarios also bring about enhanced complementarity relations~\cite{Dienes:2014via, Dienes:2017ylr} between different types of experimental probes.
DDM scenarios in which the constituents decay entirely within the dark sector---i.e., to final states comprising other, lighter ensemble constituents and/or dark radiation---are particularly challenging to test. Nevertheless, there exist observational handles that can be used to probe and constrain DDM ensembles that decay primarily via ``dark-to-dark'' decay processes of this sort, and thus potentially permit us to distinguish them from traditional DM candidates.
Dark-to-dark decays of this sort modify the way in which the expansion rate of the Universe, as described by the Hubble parameter $H(z)$, evolves with redshift. These modifications, in turn, affect the functional relationship between the redshifts $z$ and luminosity distances $D_L(z)$ of Type-Ia supernovae~\cite{Desai:2019pvs}. Since the dark-to-dark decays of a DDM ensemble alter the dependence of $H(z)$ on $z$, the DDM framework can potentially also provide a way of addressing the $H_0$ tension~\cite{Abdalla:2022yfr}. The advantage of a DDM ensemble relative to a single decaying dark matter species is that the timescale across which the decays have a significant impact on the expansion rate can be far broader. Nevertheless, DDM models in which the $\chi_\ell$ decays into dark radiation via a two-body process of the form $\chi_\ell \to \psi \bar \psi$, where $\psi$ is a massless dark-radiation field, cannot ameliorate the $H_0$ tension~\cite{Anchordoqui:2022gmw}. Models in which the $\chi_\ell$ decays primarily via intra-ensemble processes---e.g., of the form $\chi_\ell \rightarrow \chi_m \bar{\psi} \psi$, where $\psi$ once again denotes a dark-radiation field---could be more promising~\cite{Anchordoqui:2020djl}. Such decays endow the final-state $\chi_m$ with non-negligible velocities, thereby modifying the equation of state $w_m(z)$ for each ensemble constituent and modifying the DM velocity distribution of the ensemble as a whole. Moreover, complementary scattering processes of the form $\chi_\ell \psi \rightarrow \chi_m \psi$, through which the different ensemble constituents interact with the dark radiation, could potentially also help to ameliorate the $\sigma_8$ tension in the same way that they do in partially acoustic DM scenarios~\cite{Chacko:2016kgg}.
A concise description (put together in 13 ``take-away lessons'' for Snowmass 2021~\cite{Dienes:2022zbh}) of collective phenomena that can arise in dark sectors, which contain a large number of states, underscores the need to maintain a broad perspective when contemplating the possible signals and theoretical possibilities associated with non-minimal hidden sectors.
\subsection{Gravitational-Wave Probes of Coalescing Dark Matter}
Some of the DM may have clustered gravitationally in the early Universe, forming compact dark objects. These structures may cause a transient magnification of light from distant stars via microlensing, which remains one of the most powerful techniques to constrain compact dark objects in a wide range of masses~\cite{1986ApJ...304....1P}.
DM clumps near (or within) the Earth can alter the planet's tidal field---which is well monitored for decades and therefore well constrained---or cause sudden accelerations, leading to interesting constraints on asteroid-like clumps~\cite{Seto:2007kj, Namigata:2022vry, Kashiyama:2018gsh}. Albeit small, the interaction cross section of DM with Standard Model fields can lead to the deposition of small DM cores at the center of stars~\cite{Press:1985ug}, with capture rates that can be enhanced by the large density of white dwarfs and neutron stars. For fermionic fields, the accumulation of DM could eventually lead to cores more massive than the Chandrasekhar limit, collapse of the DM core to a black hole (BH), and eventually to the disruption of the star by accretion onto the newly formed BH~\cite{Goldman:1989nd}. For bosonic DM, this fate may be eluded via gravitational cooling~\cite{Brito:2015yga}.
Another possibility is that standard CDM models could produce small-scale clumps. A CDM clump moving near the Earth or a pulsar produces an acceleration that could be measurable in PTA data, providing an opportunity to test the CDM paradigm~\citep{Siegel:2007fz, Kashiyama:2018gsh}.
The possibility of compact objects harboring DM cores is intriguing. If these cores are sufficiently massive, the star is effectively described by a different equation of state and its properties change. The coalescence of DM stars will differ from the prediction of standard GR, leading to peculiar signatures in the GW signal close to merger~\cite{Ellis:2017jgp, Bezares:2019jcb}. In fact, DM clumps can also form in isolation and bind to compact stars in their vicinity. Compact DM cores orbiting neutron stars (either in their exterior or in their interior) may give rise to detectable signals in our Galaxy~\cite{Horowitz:2019aim}.
The general GW signatures of the coalescence of DM clumps or ``blobs'' have been explored by various authors~\cite[see, e.g.,][]{Giudice:2016zpa, Diamond:2021dth}, but precise calculations of the signal from the coalescence of two DM clumps require an underlying theory with a well-posed initial value problem. One example are compact configurations made of self-gravitating scalar fields, also known as boson stars~\cite{Palenzuela:2017kcg, Cardoso:2017cfl, Sennett:2017etc, Bustillo:2020syj}.
\subsection{The Gravitational-Wave Signatures of Dilute Dark Matter}
One of the most solid experimental pillars of modern physics is the equivalence principle, which ensures that all forms of matter couple universally to gravity. Even if DM does not form compact objects, dilute DM configurations must still interact gravitationally; dense DM spikes can then develop in the vicinity of isolated compact bodies such as BHs~\cite{Gondolo:1999ef, Sadeghian:2013laa}. Massive BHs are expected to be present at the center of many galaxies. In these environments the DM density should therefore be substantially higher than in the Solar System. Compact objects (BHs or neutron stars) moving in these dense DM environments will be subject to accretion and dynamical friction, leading to small changes in their dynamics that require a detailed understanding of the physics involved in these processes. Preliminary studies indicate that DM-induced changes in the GW phase of compact objects could be detectable by next-generation GW interferometers~\cite{Barausse:2014tra, Cardoso:2019rou, Kavanagh:2020cfn, Annulli:2020lyc, Annulli:2020ilw, Traykova:2021dua, Vicente:2022ivh}.
If DM has a very large Compton wavelength, as in the case of ``fuzzy'' ultralight DM fields of mass $10^{-23}$--$10^{-22}$~eV, it may give rise to small pressure oscillations at low frequency (e.g., of the order of nHz), that could affect the motion of stars and binary systems~\cite{Khmelnitsky:2013lxt, Porayko:2018sfa}. These minute changes can be tracked with PTA experiments. These oscillations can also affect the GW detectors themselves: the direct couplings to the beam splitter of GW detectors can be used to set stringent constraints on the abundance and coupling strength of DM~\cite{Vermeulen:2021epa, Pierce:2018xmy}.
\paragraph{Nonperturbative effects: ultralight bosonic fields}
The simplest possibility for new matter sectors are bosonic or fermionic degrees of freedom minimally coupled to gravity. These fields could form all or part of the DM. Their scale is set by their mass $\mu$, which could range from cosmological scales to very heavy particles~\cite{Dine:1982ah, Preskill:1982cy, Arvanitaki:2009fg}. Bosonic fields with Compton wavelengths comparable to the Schwarzschild radius of astrophysical BHs of mass $M$, i.e., $GM\mu/(c\hbar)\sim 1$, can trigger a new fascinating phenomenon caused by the existence of {\it ergoregions} around spinning BHs~\cite{Penrose:1969pc, Brito:2015oca}. Spinning BHs can spontaneously transfer their rotational energy to a boson ``condensate'' or ``cloud'' co-rotating with the BH and carrying a significant fraction of its angular momentum. The bosonic cloud is a classical object of size much larger than the BH itself, and it can contain up to $10\%$ of its mass~\cite{Brito:2015oca}. The BH/cloud system is similar to a huge gravitational ``lighthouse'' which extracts energy from the BH by emitting a nearly monochromatic GW signal.
Proposed ways to rule out or constrain light bosons as DM candidates include~\cite{Brito:2015oca}:
\begin{itemize}
\item[(i)] Monitoring the spin and mass distribution of astrophysical BHs. Measurements of highly spinning BHs will immediately rule out fields with Compton wavelengths comparable to the horizon radius, as these BHs should have been spun down on relatively short timescales~\cite{Brito:2015oca,Baryakhtar:2017ngi,Davoudiasl:2019nlo}.
\item[(ii)] Direct searches for the resolvable or stochastic monochromatic GW signals produced by the boson cloud~\cite{Arvanitaki:2010sy, Arvanitaki:2014wva, Brito:2014wla, Arvanitaki:2016qwi, Brito:2017wnc, Brito:2017zvb, Ghosh:2018gaw}, which are now routinely carried out by the LIGO/Virgo collaboration~\cite{LIGOScientific:2021jlr}.
\item[(iii)] Searches for electromagnetic (EM) emission from BH/boson cloud systems. Axion-like particles have been proposed in many theoretical scenarios, including variations of the original solution to the strong CP problem of QCD. Self-interactions and couplings with Standard Model fields can lead to periodic bursts of light, ``bosenovas,'' and other interesting phenomenology~\cite{Yoshino:2012kn, Ikeda:2018nhb}. In addition, axion-like particles should couple to photons and produce preferentially polarized light~\cite{Chen:2019fsq}.
\item[(iv)] Observations of peculiar stellar distributions around massive BHs. The nonaxisymmetric boson cloud can cause a periodic forcing of other orbiting bodies, possibly leading to Lindblad or corotation resonances where stars can cluster~\cite{Ferreira:2017pth, Boskovic:2018rub}.
\end{itemize}
These are only some of the possible strategies. Superradiance does not require any ``seed" boson abundance: any vacuum fluctuation will lead to energy extraction and exponential growth in time. In this sense, BHs are natural particle detectors, complementary to terrestrial colliders~\cite{Brito:2015oca, Barack:2018yly}. It is important to remark that the complementary role of the different GW and EM instruments necessary to probe the large range of mass/energy scales--- astrophysical BHs span about ten orders of magnitude in mass, thus allowing us to constrain ultralight bosonic fields across ten orders of magnitude in mass (or energy).
Most of the discussion above was focused on a neutral DM environment and gravitational dynamics. Another possibility is that beyond-the-Standard-Model fermions may carry a fractional electric charge or be charged under a hidden $U(1)$ symmetry~\cite{DeRujula:1989fe, Perl:1997nd}. Modified theories of gravity can also lead to compact stars or BHs carrying nonzero scalar charges~\cite{Damour:1993hw, Doneva:2017bvd, Silva:2017uqg}. In all of these theoretical scenarios, BHs and compact stars can carry non-negligible charges that would lead to different inspiral and merger signals~\cite{Zilhao:2012gp, Cardoso:2016olt, Alexander:2018qzg, Kopp:2018jom, Dror:2019uea, Bozzola:2020mjx, Maselli:2021men}; GW observations can be used to reveal or constrain these charges and the underlying theories.
\section{Astroparticle Physics}
\label{s:astropart}
In order to understand the fundamental physical forces at play in sources that human hands did not create, there is a need to collaborate with astronomers and astrophysicists to discover the nature of matter and emission from cosmic sources. This is akin to understanding the machinery along a particle accelerator to put the detection in context. While the specialties of building detectors and interpreting the results go hand in hand for particle and astroparticle physics, specialities in particle instrumentation and astrophysical progenitors require more intentional collaboration for a common goal. Nevertheless, the inclusion of each area of expertise is vital to broadening our understanding of physics as a species since each piece of the Standard Model puzzle is necessary for its completion.
\subsection{Properties of Standard Model Particles and their Interactions}
\label{s:astropart_1}
The history of cosmic-ray and neutrino studies has witnessed many discoveries central to the progress of High Energy Physics, from the watershed identification of new elementary particles in the early days, to the confirmation of long-suspected neutrino oscillations, to measuring cross sections and accessing particle interactions far above accelerator energies. There have recently been two major achievements towards this progress: {\it (i)}~the measurement of the proton-proton cross section at $\sqrt{s} \sim 75~{\rm TeV}$~\cite{PierreAuger:2012egl, TelescopeArray:2015oxb, Abbasi:2020chd}, which provides evidence that the proton behaves as a black disk at asymptotically high energies~\cite{Block:2012nj, Block:2015mjw}, and {\it (ii)}~the measurements of both the charged-current neutrino-nucleon cross section~\cite{IceCube:2017roe, Bustamante:2017xuy} and the neutral-to-charged-current cross-section ratio~\cite{Anchordoqui:2019ufu} at $\sqrt{s} \sim 1~{\rm TeV}$, which provide restrictive constraints on fundamental physics at sub-fermi distances. Moreover, at ultra-high energies, neutrino interactions probe the structure of the proton in kinematic regions that cannot be explored by accelerator experiments. In the coming decade, neutrino-nucleon cross sections will be probed well above the energy scale of colliders, testing many allowed novel-physics scenarios; see Fig.~\ref{fig:cross_section_uhe} and Refs.~\cite{Ackermann:2022rqc,Denton:2020jft,Esteban:2022uuw,Valera:2022ylt} for details. Additionally, the inelasticity distribution of events detected at neutrino telescopes has also been envisioned as an important tool for revealing new physics processes~\cite{Anchordoqui:2006wc}. Current IceCube data in the TeV--PeV range are in good agreement with the SM predictions~\cite{IceCube:2018pgc}. Finally, the cosmic neutrino observatories such as the upcoming IceCube Upgrade will provide complimentary sensitivity to neutrino oscillation measurements by long baseline oscillation facilities, as shown in Fig.~\ref{fig:oscillation} and discussed in the white papers \citep{Ackermann:2022rqc, Abraham:2022jse}.
Above $\sim 1$~PeV, $W$-boson production becomes relevant from two processes: electron anti-neutrino scattering on atomic electrons and neutrino-nucleus interactions in which the hadronic coupling is via a virtual photon. The former produces the distinct Glashow resonance at 6.3~PeV~\cite{Glashow:1960zz}. The latter can reach up to 5--10\% of the deep-inelastic-scattering cross section in the PeV range~\cite{Seckel:1997kk, Alikhanov:2015kla, Gauld:2019pgt, Zhou:2019vxt, Zhou:2019frk}. IceCube recently reported the detection of a particle shower compatible with a Glashow resonance event~\cite{IceCube:2021rpz}. $W$-boson production~\cite{Seckel:1997kk, Alikhanov:2015kla, Zhou:2019vxt, Zhou:2019frk} can play a significant role in the detection of tau neutrinos from cosmic origin~\cite{Soto:2021vdc}. Future analyses with more neutrino data above PeV energies will better probe these effects.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{Figures/cross_sections_whitepaper.pdf}
\caption{Neutrino-nucleon cross section measurements, compared to deep-inelastic-scattering (DIS) cross section predictions from Ref.~\cite{Bertone:2018dse} (BGR18). In the TeV range, FASER and FASER$\nu$ have started measurements~\cite{Arakawa:2022rmp}. Measurements in the TeV--PeV range are based on IceCube showers~\cite{Bustamante:2017xuy, IceCube:2020rnc} and tracks~\cite{IceCube:2017roe}. Projected measurements at energies above 100~PeV~\cite{Valera:2022ylt} are based on 10~years of operation of the radio component of IceCube-Gen2, assuming a resolution in energy of 10\% and a resolution in zenith angle of 2°. Since the flux at these energies remains undiscovered, projections for the measurement of the cross section are for different flux predictions. From Ref.~\cite{Ackermann:2022rqc}, adapted from Ref.~\cite{Valera:2022ylt}.}
\label{fig:cross_section_uhe}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.7\textwidth]{Figures/Upgrade_NuMu_Disappearance_Sensitivity_May2022.png}
\caption{
Measurements of $\sin^2 (\theta_{23})$ and $\Delta m^2_{32}$ with the IceCube Upgrade (inner fiducial volume) with 3 years of data, comparing with long-baseline neutrino oscillation facilities and other Cherenkov detectors \cite{Super-Kamiokande:2017edb, OPERA:2018nar, Haegel:2017Ck, Whitehead:2016xud}. From Ref.~\cite{Ishihara:2019aao, Stuttard22}.
}
\label{fig:oscillation}
\end{figure}
LHC experiments provide a laboratory for measurements relevant to understand the subtleties of astroparticle physics. Atmospheric neutrinos, produced in the interaction of cosmic rays with nuclei in the Earth’s atmosphere and the subsequent decay of mesons, are an irreducible background to searches for cosmic neutrinos (see e.g.~\cite{IceCube:2021uhz}). An accurate understanding of the physics of cosmic sources therefore requires an in-depth understanding of the atmospheric neutrino flux. The CERN's Forward Physics Facility (FPF) will provide key information to reduce the uncertainties for cosmic neutrino searches in the context of multimessenger astrophysics~\cite{Feng:2022inv, Anchordoqui:2021ghd,Bai:2022jcs,Jeong:2021vqp}. More concretely, LHC neutrinos to be measured at the FPF experiments could give critical information on charm production at Feynman-$x$ close to 1. This process could potentially become a source of background for cosmic neutrinos above $10^{7.3}~{\rm GeV}$~\cite{Anchordoqui:2022ivb} and, at the moment, we have no supportive data and no theory for the process.
\subsection{Beyond-Standard-Model Neutrino Physics}
\label{s:nuosc}
Neutrinos could have beyond-Standard-Model (BSM) interactions and, if the coupling strengths are weak or if heavy particles mediate the interactions, these interactions may only manifest themselves in the high-energy (HE) and ultra-high-energy (UHE) neutrino sector. Possible scenarios include BSM neutrino interactions with dark matter---including heavy dark matter---and with sterile neutrinos.
While the SM allows for interactions among neutrinos, these interactions are all highly suppressed by the electroweak scale. It is still unknown whether there are additional, BSM secret interactions solely among neutrinos that are stronger. In a UV-complete BSM model, this implies the existence of some new electrically neutral mediator, significantly lighter than the Z boson, that couples to neutrinos.
New interactions of this type can significantly modify the character of the HE and UHE neutrino flux arriving at the Earth by the scattering of the HE/UHE neutrinos off of nearly at-rest cosmic neutrino background (C$\nu$B) neutrinos. Such spectral distortion features may appear in the PeV--EeV regime, depending on the absolute mass of neutrinos~\cite{Hooper:2007jr, Lykken:2007kp, Ioka:2014kca, Ng:2014pca, Ibe:2014pja, Blum:2014ewa, DiFranzo:2015qea, Cherry:2016jol, Kelly:2018tyg, Barenboim:2019tux, Murase:2019xqi, Bustamante:2020mep, Creque-Sarbinowski:2020qhz}. The secret neutrino interactions are independently motivated by the neutrino mass generation mechanism~\cite{Blum:2014ewa}, muon $g-2$ anomaly~\cite{Araki:2015mya}, small-scale problems in dark matter substructures~\cite{vandenAarssen:2012vpm, Tulin:2017ara}, and apparent Hubble tension~\cite{Cyr-Racine:2013jua, Kreisch:2019yzn, Blinov:2019gcj,Carpio:2021jhu}.
In addition, flavor and $\nu/\bar{\nu}$ ratios provide complementary probes of new neutrino physics and neutrino production mechanisms~\cite{Ackermann:2022rqc, Engel:2022bgx, Arguelles:2022xxa, Abraham:2022jse}. Due to the fact that neutrinos are predominately expected from the decays of muons and charged pions, the nominal expectation is that only electron and muon neutrinos are generated at the sources and that $\nu$ and $\bar{\nu}$ are produced in comparable numbers. After leaving the sources, oscillations over cosmological distances are expected to distribute the flux nearly evenly among all flavors by the time the neutrinos reach Earth. In reality, however, different neutrino production channels become accessible at different energies and, as a result, the flavor and $\nu/\bar{\nu}$ ratios should vary with energy~\cite[see, e.g.,][]{Anchordoqui:2003vc, Anchordoqui:2004eb, Lipari:2007su, Bustamante:2015waa}. Following this, the expected flavor ratios at Earth might deviate from a democratic flavor composition, and may do so as a function of energy. Hence, the flavor ratios measured at Earth~\cite{IceCube:2015rro} combined with information about the values of the neutrino mixing parameters~\cite{Esteban:2020cvm} can be used to infer the flavor ratios at the sources~\cite{Palomares-Ruiz:2015mka, Bustamante:2019sdb, Song:2020nfh}. However, as discussed in Sec.~\ref{s:portals}, large deviations are possible in some BSM scenarios (e.g., neutrino decay, pseudo-Dirac states, new neutrino interactions with dark matter or sterile neutrinos, violation of Lorentz and CPT symmetries) which can alter the oscillation parameters~\cite{Beacom:2002vi, Beacom:2003eu, Barenboim:2003jm, Beacom:2003nh, Anchordoqui:2005ey, Bustamante:2010nq, Arguelles:2015dca, Shoemaker:2015qul, Rasmussen:2017ert, Ahlers:2018yom, Denton:2018aml, Ahlers:2020miq,Abdullahi:2020rge}. Large event statistics and complementary flavor-specific detection techniques are needed to identify flavor-specific signals and to measure the flavor composition statistically in a sample of collected events.
Constraints on BSM neutrino interactions using current IceCube data have been derived in Refs.~\cite{Bustamante:2020mep, Esteban:2021tub}. Future detectors, with improvements in particular in detector energy resolution and capability to identify neutrino flavor, are crucial to probing the BSM neutrino physics.
The ANtarctic Impulsive Transient Antenna (ANITA) has observed two anomalous events, which qualitatively look like air showers initiated by energetic ($\sim 500~{\rm PeV}$) particles that emerge from the ice along trajectories with large elevation angles ($\sim 30^\circ$ above the horizon)~\cite{ANITA:2016vrp,ANITA:2018sgj}. As was immediately noted by the ANITA Collaboration, these events may originate in the atmospheric decay of an upgoing tau-lepton produced through a charged current interaction of a tau-neutrino inside the Earth. However, for the angles inferred from ANITA observations, the ice would be well screened from up-going high-energy neutrinos by the underlying layers of Earth, challenging SM explanations~\cite{Romero-Wolf:2018zxt}. As of today, the origin of these anomalous events remains unclear; follow-up observations of these unusual events by EUSO-SPB2~\cite{Eser:2021mbp} and PUEO~\cite{PUEO:2020bnn} are well-motivated~\cite{Ackermann:2022rqc}.
\subsection{The Muon Puzzle of Ultra-High-Energy Cosmic Rays}
\label{s:muonpuzzle}
The muonic component of cosmic-ray air showers is generally used as a probe of the hadronic interactions during the cascade development~\cite{Coleman:2022abf}. Various measurements of atmospheric muons with energies $1 \lesssim E_\mu/{\rm GeV} \lesssim 10$ have revealed a discrepancy between simulated and observed muon production in air showers. The highest energy cosmic rays currently observed by the Pierre Auger Observatory (Auger) show a significant discrepancy in the shower muon content when compared to predictions of LHC-tuned hadronic event generators~\cite{PierreAuger:2014ucz, PierreAuger:2016nfk}. More concretely, the analysis of Auger data suggests that the hadronic component of showers (with primary energy $10^{9.8} < E/{\rm GeV} < 10^{10.2}$) contains about $30\%$ to $60\%$ more muons than expected with a significance somewhat above $2.1\sigma$. The discrepancy between experiment and simulations has also been observed in the Telescope Array data analysis at the same energy range~\cite{TelescopeArray:2018eph}. Auger findings have also been recently confirmed by studying air-shower measurements over a wide range of energies. The muon deficit between simulation and data, dubbed the {\it muon puzzle}, seems to start at $E \sim 10^8~{\rm GeV}$, increasing noticeably as primary energy grows, with a slope that was found to be significant at $\sim8\sigma$~\cite{Albrecht:2021cxw}. However, the muon deficit of simulated events has not been observed in IceTop data with $10^{6.4} < E/{\rm GeV} < 10^{8.1}$~\cite{IceCube:2022yap}. It is noteworthy that, in this energy range, the cosmic-ray spectrum has a significant contribution of nulcei~\cite{Coleman:2022abf}, and so the center-of-mass energy per nucleon pair of the showers in the IceTop sample is well below those of LHC collisions. In contrast, the center-of-mass energy per nucleon of a $10^{10.3}~{\rm GeV}$ helium nucleus incident upon a nucleon in the atmosphere is 100~TeV, and so many secondary interactions would be above the LHC center-of-mass energy~\cite{Allen:2013hfa}. Within this decade, ongoing detector upgrades of existing facilities---such as AugerPrime~\cite{PierreAuger:2016qzd} and IceCube-Gen2~\cite{IceCube-Gen2:2020qha}---will enhance the precision of air-shower measurements and reduce uncertainties in the interpretation of muon data. In particular, as a part of the upcoming AugerPrime upgrade, each surface station will have additional detectors that will provide complementary measurements of the incoming shower particles, consequently leading to improved reconstruction of muons and electromagnetic particles~\cite{PierreAuger:2016qzd}. This will allow for the measurement of the properties of extensive air showers initiated by the highest energy cosmic rays with unprecedented precision, providing a unique probe of hadronic collisions at center-of-mass energies that surpass the LHC energy.
In solving the muon puzzle, one has to simultaneously get good agreement with the measurements of the distribution of the depth of shower maximum $X_{\rm max}$~\cite{PierreAuger:2014sui} and the fluctuations in the number of muons~\cite{PierreAuger:2021qsd}. A thorough phenomenological study has shown that an unrivaled solution to the muon deficit, compatible with the observed $X_{\rm max}$ distributions, is to reduce the transfer of energy from the hadronic shower into the electromagnetic shower by reducing the production or decay of neutral pions~\cite{Allen:2013hfa}. Hence, the amount of forward strangeness production could be of particular relevance in addressing the muon puzzle~\cite{Allen:2013hfa, Anchordoqui:2016oxy, Baur:2019cpv, Anchordoqui:2022fpn}. Strangeness production is traced by the ratio of charged kaons to pions, for which the ratio of electron and muon neutrino fluxes is a proxy that will be measured by the FPF experiments~\cite{Feng:2022inv, Anchordoqui:2021ghd}. Electron neutrino fluxes are a measurement of kaons, whereas both muon and electron neutrinos are produced via pion decay. However, $\nu_\mu$ and $\nu_e$ populate different energy regions, which can help to disentangle them. In addition, neutrinos from pion decay are more concentrated around the line-of-sight than those of kaonic origin, given that $m_\pi < m_K$, and thus neutrinos from pions obtain less additional transverse momentum than those from kaon decays. Thereby, the closeness of the neutrinos to the line-of-sight or, equivalently, their rapidity distribution, can be used to disentangle different neutrino origins to get an estimate of the pion-to-kaon ratio. This implies that measurements at the FPF will improve the modeling of high-energy hadronic interactions in the atmosphere, reduce the associated uncertainties of air-shower measurements, and thereby help to understand the properties of cosmic rays, such as their energy and baryonic structure, which is crucial to discover their origin.
There is also the possibility that the muon puzzle does not originate from an incomplete understanding of the forward particle physics. If this were the case, future ultra-high-energy cosmic-ray (UHECR) measurements would provide a unique probe of BSM physics with a high potential for discovery~\cite{Coleman:2022abf}.
\subsection{The Nature of Matter in Neutron Star Interiors}
\label{s:NS_EOS}
The nature of matter at ultra-high densities ($\rho_s > 2.8 \times 10^{14}$\,g\,cm$^{-3}$), large proton/neutron number asymmetry, and low temperatures ($\lesssim 10^{10}~{\rm K}$) is, at present, one of the major outstanding problems in modern physics, owing to a number of challenges both in the experimental and theoretical realms~\cite[see, e.g.,][for a review]{Bogdanov:2022faf, Watts:2016uzu}. A plethora of well-motivated theoretical predictions for the state of matter in this temperature-density regime have been proposed, ranging from normal nucleonic matter, to particle exotica such as hyperons, deconfined quarks, color superconducting phases, and Bose-Einstein condensates (for a review see~\cite{Watts:2016uzu}. Matter in this extreme regime is known to only exist stably in the cores of neutron stars (NSs).
\vskip5pt
\noindent
\begin{minipage}{0.48\textwidth}
Neutron stars are host to the densest matter in the Universe. The density increases toward the center of the star, reaching densities of 5--10\,$n_s$, that is, several times the nuclear saturation density of $n_s = 0.16/$fm$^3$. At these densities, we currently do not know how matter behaves, what the phase structure is, or what the dynamic degrees of freedoms are. Neutron stars offer a unique laboratory to study strongly interacting matter and the underlying theory of QCD in the most extreme conditions.
\end{minipage}
\hfill
\begin{minipage}{0.48\textwidth}
\centering
\includegraphics[width=0.98\textwidth]{NS_profile}
\vskip-0.5cm
\captionsetup{type=figure}
\captionof{figure}{The structure of a neutron star as predicted by theory.}
\label{fig:structure}
\end{minipage}
\vskip5pt
They have the potential to facilitate the discovery of novel exotic phases of matter in their cores, including the appearance of strangeness in the form of hyperonic matter and ultimately the melting of nucleonic structure, giving rise to novel forms of cold quark matter.
The structure of NSs is determined by the competition between self gravity and pressure of strong nuclear interactions keeping the star in a hydrostatic equilibrium. This interaction is described in the simplest case of non-rotating NSs by the Tolman-Oppenheimer-Volkoff (TOV) equations~\cite{Oppenheimer:1939ne, Tolman:1939jz}, which map the equation of state (EOS) of dense nuclear matter to the macroscopic properties of NSs, making the EOS the primary object of interest for the nuclear physics of NSs. A large community effort has been put to investigating the EOS using different \emph{ab-initio} calculations as well as various models.
The evidence from observations, in particular with the advent of multimessenger astronomy, is more complicated than the simple static systems described by the TOV equations and a significant push has been made in the past years to numerically solve the combined Einstein and relativistic-fluid-dynamic equations~\cite{Romatschke:2017ejr}. Fig.~\ref{fig:structure} displays a schematic figure of the NS structure. The crust and the outer core down to a depth of roughly $0.5$~km, where densities are of the order of nuclear density, is under good theoretical control~\cite{Baym:1971pw, Negele:1971vb}. Beyond that, our understanding of the structure relies on theoretical extrapolations. In particular, the phase of the inner core is currently unknown. Fig.~\ref{fig:eosphysics} is a schematic view of the hypothesized phase diagram of QCD. It is a firm prediction of QCD wherein, at sufficiently high temperatures and/or densities, ordinary hadronic matter melts to a partonic form of matter--- Quark Matter.
In the regime of high temperatures and low baryon densities, the deconfiement transition to Quark Matter is well studied using lattice field theory~\cite{Fodor:2004nz,Aoki:2006we, Karsch:2001cy} and its existence is confirmed in two decades of experimentation with ultra-relativistic heavy-ion collisions~\cite{Margetis:2000sv} at the RHIC and the LHC. Further experimental program runs at the LHC aim to quantify the transport properties and the conditions of the onset of Quark Matter.
The deconfinement transition is a cross-over at low baryon densities but it is been long hypothesized that the transition becomes stronger with increasing baryon density. New theoretical arguments based on topological features of QCD have been recently put forward supporting the first-order-nature of the transition~\cite{Cherman:2018jir, Fukushima:2010bq}. The beam energy scan (BES) program at the RHIC~\cite{STAR:2017sal} and the future FAIR facility~\cite{CBM:2016kpk} are geared to discover the critical point separating the crossover transition from the first-order transition. The discovery of the critical point would have a profound impact on physics of NSs.
At very-high densities, owing to an attractive interaction between quarks in QCD, it is expected that Quark Matter is in the form of a color superconductor, fundamentally affecting the transport properties of Quark Matter~\cite{Alford:2007xm, Alford:1997zt}. Based on large-$N_c$ arguments, it has also been speculated~\cite{McLerran:2018hbz} that, at low temperatures, there may be a further intermediate phase that is still confined but where the chiral symmetry is restored. Owing to the similarities with both the hadronic and Quark Matter phases, this hypothetical phase is dubbed the quarkyonic phase.
Gravitational-wave observations of binary neutron star mergers can constrain the cold state of ultra-high density matter in neutron stars from tidal effects during the inspiral phase of the binary made possible with GW170817 \cite{LIGOScientific:2018cki, De:2018uhw, Bauswein:2017vtn}, as well as observe the dynamics of the hot, dense matter after merger, which will become possible with the next generation of gravitational-wave detectors.
\begin{figure*}[t!]
\centering
\includegraphics[width=0.8\textwidth]{ns710403.f6.jpeg}
\caption{Schematic view of the QCD phase diagram. The figure highlights regions probed by the RHIC, LHC, FAIR, and FRIB experiments, regions of validity for lattice QCD and chiral EFT, and environments reached in neutron stars, supernovae, and neutron star mergers. Abbreviations: EFT=effective field theory, QCD=quantum chromodynamics.[Borrowed from \url{https://www.annualreviews.org/doi/full/10.1146/annurev-nucl-102419-041903}]}
\vspace{-0.3cm}
\label{fig:eosphysics}
\end{figure*}
Constraints on the EOS from mass and radius observations of NSs can be complemented with constraints on the nuclear-symmetry energy obtained from nuclear experiments and \textit{ab-initio} neutron matter theory. It has been known for some time that there is a high degree of correlation between neutron star radii and the pressure of neutron star matter slightly above the nuclear saturation density ($n_s=0.16$ fm$^{-3}$ or $\rho_s=2.7\times10^{14}$ g cm$^{-3}$). The NS matter pressure at $n_s$ is nearly completely determined by the slope $L$ of the symmetry energy at the same density~\cite{2001ApJ...550..426L}, which is experimentally probed by nuclear binding energies, dipole polarizabilities and neutron skin thicknesses of neutron-rich nuclei, and, theoretically, from neutron matter studies.
An important set of measurements in nuclear astrophysics comes from studying the timing properties of radio pulsars--- rotating NSs that emit beamed radiation along their magnetic poles. An early example is the discovery and analysis of the ``Hulse-Taylor" radio pulsar-binary system which (indirectly) confirmed of the existence of gravitational radiation~\cite{Weisberg:1981bh, Taylor:1982zz} and produced mass estimates with high precision. The recent discoveries of additional relativistic pulsar orbits and high-mass NSs, as well as the eventual detection of nanohertz-frequency gravitational radiation through pulsar timing, show that radio pulsars continue to serve as ideal laboratories for fundamental physics.
Space-borne observations at X-ray energies offer various means for obtaining strong constraints on the allowed dense-matter EOS, providing unique insight into the high-density, low-temperature region of the QCD phase diagram. While current telescopes have made important headway, they lack the required capabilities to fully exploit the information about the dense-matter EOS encoded in the observed X-ray emission from NSs. This important undertaking requires a new generation of X-ray facilities with at least an order-of-magnitude improvement in sensitivity relative to current observatories, while also offering the high time resolution required for effective studies of rapidly spinning NSs.
Collectively, electromagnetic (radio and X-ray) and gravitational-wave astrophysical measurements, combined with terrestrial laboratory constraints, hold the promise to provide definitive empirical constraints on the true nature of the densest matter in the Universe~\cite{Miller:2019cac, Raaijmakers:2019qny, 2020ApJ...892...55J, 2020PhRvD.101l3007L}.
\subsection{Tests of Lorentz and CPT invariance}
\label{s:LIV}
Both Lorentz and CPT symmetries are fundamental to our understanding of the SM and General Relativity~\cite{Coleman:1998ti, Stecker:2017gdy, Altschul2011}. Lorentz invariance (LI)---one of the main symmetries that govern the SM of elementary particles---requires the structure of spacetime to be the same for all observers. However, proposed Grand Unified Theories suggest that our understanding of spacetime symmetries may be incomplete and that fundamental modifications to the Lorentz symmetry could be made to account for quantum effects, thereby potentially violating this symmetry when approaching the Planck scale~\cite{Addazi:2021xuf}. Lorentz invariance violation (LIV) at a high-enough energy scale could actually arise in loop quantum gravity or string theory~\cite{Alfaro:2004aa, AmelinoCamelia2001, Bluhm:2013mu, Calcagni:2016zqv, Colladay:1998fq, EllisMavromatosNanopoulos1999, GambiniPullin1999, Kostelecky:1988zi, Nambu1968, Potting2013}. Closely intertwined with Lorentz symmetry is the CPT symmetry--- the established discrete spacetime symmetry of charge, parity, and time reversal. In a local quantum theory, it is impossible to violate CPT invariance without also breaking LI~\cite{Greenberg:2002uu}. Thus, many tests of LI can also be interpreted as tests of CPT.
Even a small violation of LI could easily affect the propagation of particles on a cosmological scale~\cite{Coleman:1998ti, Aloisio:2000cm}. Moreover, at extreme energies, like those available in the collision of ultra-high-energy cosmic rays in the Earth's atmosphere, one could also expect a change in the interactions driving the air-shower development due to LIV~\cite{Coleman:2022abf, Tomar:2017mgc}.
As there are several signatures of the violation of these fundamental symmetries, there are a variety of tests that may be performed to search for them. The Pierre Auger Collaboration has derived limits on LIV by comparing the energy spectrum and cosmic-ray composition with upper limits on the photon flux~\cite{PierreAuger:2021tog} and by comparing Monte Carlo expectations to muon fluctuation measurements~\cite{PierreAuger:2021mve}. In the years ahead, the most restrictive bounds on LIV could be coming from UHECR experiments~\cite{PierreAuger:2021tog, Coleman:2022abf}. Additionally, as discussed in Sec.~\ref{s:nuosc}, the flavor ratio of cosmic neutrinos provides a powerful test of Lorentz and CPT symmetries.
Precise measurements of very-high-energy photons can also be used to test LIV~\cite[see, e.g.,][]{Vasileiou:2013vra, HAWC:2019gui, LHAASO:2021opi}. One consequence of LIV is that photons of sufficient energy are unstable and decay over short timescales~\cite{Martinez-Huerta:2017ulw}. This means that high-energy photons from astrophysical objects may decay well before they can arrive at Earth. Constraints to the LIV energy scale have been established by looking at the highest-energy photons from the Crab nebula, eHWC J1825-134, and LHAASO J2032+4102~\cite{LHAASO:2021opi, HAWC:2019gui}. However, higher limits are expected from continued observations of even more high-energy sources, such as RXJ1713.7-3946, with upcoming observatories including the Southern Wide-field Gamma-ray Observatory (SWGO)~\cite{Albert:2019afb, Hinton:2021rvp, Schoorlemmer:2019gee} and the Cherenkov Telescope Array (CTA)~\cite{CTAConsortium:2017dvg}. The higher the energy of a detected gamma ray and the narrower its energy uncertainty, the more stringent the constraints would be. Thus, instruments optimized at the highest energies, such as SWGO, LHAASO~\cite{LHAASO:2019qtb}, and CTA, would be optimal instruments to search for LIV signatures.
\subsection{Production of Exotic Particles in the QED Domain}
Exotic quantum electrodynamics (QED) processes may operate in extremely strong magnetic fields with $B> B_{\rm cr} = 4.4\times 10^{13}$~G, when $h\nu_B \sim m_e c^2$ is achieved.
Magnetars, a topical subclass of neutron stars with surface fields exceeding $10^{14}$~G \cite{Harding:2006qn, Mereghetti:2008je, Turolla:2015mwa, Kaspi:2017fwg}, provide a cosmic lab to test QED in this domain. The potential action of exotic QED mechanisms of photon splitting and magnetic pair creation yields distinctive imprints on magnetar polarization and Comptonization, which may be observed in the sub-MeV waveband \cite{Wadiasingh:2017rcq, Hu:2019nyw} by future MeV telescopes like the All-sky Medium-Energy Gamma-ray Observatory (AMEGO)~\cite{AMEGO:2019gny}.
\section{Multimessenger Synergies in Particle Astrophysics}
\label{s:multimessenger}
Multimessenger astrophysics encompasses the measurement of any cosmic event with more than one type of signal--- photons, gravitational waves, neutrinos, or cosmic rays. The dawn of the modern Multimessenger Era was heralded by the co-detection of gamma rays and gravitational waves in a binary neutron star merger~\cite{LIGOScientific:2017ync} and by the co-detection of gamma rays and neutrinos in a blazar flare~\cite{IceCube:2018dnn}. We learned more from each of those singular co-detection events than a decade of astrophysical observations could have told us with photons alone. Over the next decade, multimessenger detections will become more important to accelerating the rate of discoveries in cosmic particle physics by constraining coincident event types from different messengers simultaneously. The United States currently leads efforts in multimessenger astrophysics through the investments DOE, NSF-Physics, and NASA have made over the past several decades. Maintaining U.S. primacy in this field will require the support of a well-balanced program of facilities across all messengers in complementarity with our collaborators around the world and leadership in the rigorous task of coordinating between them. The following section highlights several compelling astroparticle physics areas that are best addressed with multimessenger methods.
\subsection{Pinpointing the Sources of the Highest-Energy Cosmic Rays}
\label{UHECRsources}
It is well known that the cosmic microwave background (CMB) makes the Universe opaque to the propagation of ultra-high-energy cosmic rays (UHECRs). The so-called GZK interactions of cosmic rays above the photopion production threshold (or nucleus photodisintegration) and the relic photons lead to a sharp cutoff in the UHECR spectrum above about $10^{10.6}~{\rm GeV}$~\cite{Greisen:1966jv, Zatsepin:1966jv}. It was recently noted that the characteristic cosmic-ray energy of the GZK cutoff could coincide with the species scale ($\hat M \sim 10^{10}~{\rm GeV}$), where
physics becomes strongly coupled to gravity~\cite{Montero:2022prj}. This suggests that the cosmic-ray maximum energy may be driven by the species scale. Hence, aside from its astrophysical motivations, understanding the origin of the abrupt cutoff observed in the UHECR flux around $10^{10.6}~{\rm GeV}$~\cite{PierreAuger:2008rol, HiRes:2007lra} could have direct applications to probe BSM physics and posses a huge challenge for UHECR experiments within the next decade. In particular, a precise characterization of the source spectra of the highest-energy cosmic rays has the potential for breakthrough results in fundamental physics~\cite{Anchordoqui:2022ejw}.
Rapid progress in computational high-energy astrophysics is dramatically advancing the study of acceleration mechanisms. Some of the current contenders for acceleration mechanisms and source types are shock acceleration in systems ranging from the large-scale shocks surrounding galaxy clusters~\cite{Kang:1996rp,Ryu:2003cd} to internal or external shocks of starburst-superwinds~\cite{Anchordoqui:1999cu, Anchordoqui:2018vji, Anchordoqui:2020otc}, active galactic nuclei (AGN)~\cite{Biermann:1987ep, Takahara:1990he, Rachen:1992pg, Blandford:2018iot, Matthews:2018laz, Matthews:2018rpe, Eichmann:2022ias} or gamma-ray burst~\cite{Waxman:1995vg, Vietri:1995hs, Wang:2007xj, Murase:2008mr, Baerwald:2013pu, Globus:2014fka, Zhang:2017moz} jets, and the jets of tidal disruption events (the transient cousins of AGN jets)~\cite{Farrar:2008ex,Farrar:2014yla,Pfeffer:2015idq}. Other contenders are shear acceleration~\cite{Rieger:2004jz,Kimura:2017ubz} and one-shot mechanisms such as ``espresso"~\cite{Caprioli:2015zka}, in which an AGN or other jet boosts a galactic cosmic ray of the host galaxy; electromotive force acceleration as in fast-spinning pulsars~\cite{Blasi:2000xm, Fang:2012rx, Fang:2013cba} and magnetars~\cite{Arons:2002yj}, black holes~\cite{Blandford:1977ds, Neronov:2009zz}, and potentially reconnection, explosive reconnection, gap and/or wakefield acceleration~\cite{Chen:2002nd, Murase:2009pg, Ebisuzaki:2013lya}. The abundance of possibilities suggests there may well be multiple sources of UHECRs---some of which may be transient---making the identification of sources even more challenging and essential~\cite{Coleman:2022abf, Engel:2022yig}. The Study of particle acceleration in astrophysical plasmas is a near-term application of the accelerator physics as pointed out by the Snowmass 2021 Accelerator Frontier White Paper: ``Near Term Applications driven by Advanced Accelerator Concepts" \citep{Emma:2022zdv}.
On the experimental side, the Pierre Auger Collaboration (Auger) has discovered a large-scale dipole anisotropy above $10^{9.9}~{\rm GeV}$ with a significance $> 6\sigma$~\cite{PierreAuger:2017pzq}. Given that the dipole direction is $\sim 115^\circ$ away from the Galactic center, this is evidence of the extragalactic origin of cosmic rays above this energy threshold. Intriguingly, the dipole direction is not aligned with the CMB dipole, the local matter over-density, or any obvious individual source. A further analysis finds a $4\sigma$ significance for correlation of cosmic rays above $10^{10.6}~{\rm GeV}$ with a model based on a catalog of bright starburst galaxies and a $3.1\sigma$ correlation with a model based on a \textit{Fermi}-LAT catalog of jetted AGNs~\cite{PierreAuger:2018qvk,PierreAuger:2022axr}. The best-fit Gaussian angular scales correspond to a top-hat radii of $25^\circ$ and the signal fractions range from 5--10\%. Most of the anisotropy signal comes from the so-called Centaurus region (which contains the jetted AGN Centaurus A as well as the starburst galaxies NGC4945 and M83). The starburst model also benefits from one prominent source candidate, NGC253, being close to the southern Galactic pole where a warm spot of Auger events is found. When data from the Telescope Array are included in the analysis, the correlation with starburst galaxies is mildly stronger than the Auger-only result with a post-trial significance of $4.2\sigma$~\cite{TelescopeArray:2021gxg}. Continuing operation of Auger should yield a significance level of $5\sigma$ for the Centaurus region excess by the end of 2025 ($\pm 2$ calendar years), possibly preceded by a similar significance milestone in the correlation with the starburst catalog if the warm spot continues to grow~\cite{Coleman:2022abf}.
Interaction between UHECRs and the CMB leads to the production of ultra-high-energy (UHE) neutrinos~\cite{Berezinsky:1969erk}. The so-called GZK process is effective when the energies of UHECR nucleons are higher than $\sim 5\times 10^{10}$~GeV and the corresponding cosmogenic neutrinos see their main flux around and below $\sim 10^{9}$~GeV~\cite{Stecker:1978ah,Yoshida:1993pt, Takami:2007pp, Anchordoqui:2007fi, Kotera:2010yn, Ahlers:2010fw, AlvesBatista:2018zui}. Upper limits on the cosmogenic neutrino fluxes have been obtained by IceCube~\cite{IceCube:2018fhm}, the Pierre Auger Observatory~\cite{PierreAuger:2019ens}, and ANITA~\cite{Gorham:2019guw}. Cosmogenic neutrinos when combined with UHECR observations could provide a unique multimessenger signature of GZK interactions, but as of today no neutrino has been observed with energy above $10^{7}~{\rm GeV}$.
Sources of Galactic UHECR neutrons, when combined with the antineutrino flux resulting from neutrons decaying on flight at lower energies provide a unique beam to test neutrino oscillations, as the expected Earthly neutrino flavor ratio differs from the nearly even distribution among electron, muon, and tau flavors (1:1:1) of astrophysical neutrinos originating via charged pion decay~\cite{Anchordoqui:2003vc}.
\subsection{Probing Extreme-Energy Hadron Acceleration and Interaction with Neutrinos}
While gamma rays may be produced by both leptonic process, such as inverse Compton scattering of background photons, and hadronic process, such as pion decay, high-energy neutrinos may only be produced when hadronic cosmic rays interact with surrounding matter ($pp$) and light ($p\gamma$). Thus, high-energy neutrinos provide a unique probe of hadron acceleration and interaction in astrophysical environments.
The origin of the bulk of the high-energy neutrinos remains unknown \cite{IceCube:2019cia}, though hints to the first sources have been found. The coincident observations of a high-energy neutrino event, IceCube-170922A, with X-rays and gamma rays from the blazar TXS 0506 +056~\cite{IceCube:2018cha, IceCube:2018dnn} make this blazar the first candidate high-energy neutrino source. In addition, the ten-year point-source searches with IceCube indicated that NGC~1068 is the most significant steady source of neutrinos at a significance of $\sim3\sigma$~\cite{IceCube:2019cia}.
Neutrinos are an important probe of dense environments that are not visible with photons. Interestingly, gamma-ray limits and observations of these early sources indicate that the energy carried by hadrons must be significantly higher than that carried by leptons. Models that may explain the observed neutrinos require a large baryonic loading, i.e., a large fraction of the available energy imparted to cosmic rays, which may be theoretically challenging.
The flux and spectral index of the TeV--PeV diffuse neutrino background are comparable to that of the GeV--TeV diffuse gamma-ray background, and the latter tightly constraining the flux of the electromagnetic cascades of the gamma-ray counterparts of high-energy neutrinos. The current IceCube measurements already indicate that unless new physics processes are at play~\cite{Anchordoqui:2021dls}, the bulk of the neutrino sources are likely opaque to gamma rays \cite{Murase:2015xka, Fang:2022trf}. Future observation of TeV and sub-TeV neutrinos may confirm the present indications that neutrinos originate in cosmic environments that are optically thick to GeV--TeV gamma rays. Such gamma-ray-obscured sources may be bright in 1--100~MeV energies and be observed by future MeV gamma-ray facilities like the All-sky Medium-Energy Gamma-ray Observatory (AMEGO)~\cite{AMEGO:2019gny}. This is a clear example of the predictive power of multimessenger science, which will be testable within this decade.
Future firm detections of high-energy neutrino sources and characterization of their spectra are crucial to the understanding of hadron acceleration and interaction in the cosmos. The Snowmass 2021 whitepaper ``Snowmass2021 Cosmic Frontier: Advancing the Landscape of Multimessenger Science in the Next Decade"~\cite{Engel:2022yig} discusses the current and future multimessenger network and the collaboration and infrastructure needed for successful multimessenger observations of neutrino sources.
With the improved statistics, sensitivity, and sky coverage offered by upcoming neutrino experiments, we can expect to expand our view of the neutrino sky, including firmly establishing neutrino sources. Next-generation telescopes currently in th planning stage or under construction, as discussed in the Snowmass 2021 whitepaper ``High-Energy and Ultra-High-Energy Neutrinos"~\cite{Ackermann:2022rqc}, will allow detailed studies of high-energy neutrinos, including their energy spectrum, flavor composition, and the identity of their sources.
\subsection{Diffuse Backgrounds}
\label{subsec:diffuseBG}
Diffuse astrophysical backgrounds arise in all of the astrophysical messengers, not just due to the limitations of the resolutions of current detectors, but as an indication of large-scale and diffuse structure in the Universe. These diffuse backgrounds are studied extensively for individual messengers, but future insights to the origin of the cosmos may arise from considering their similarities and collaboration across diffuse working groups for each messenger~\cite{Engel:2022yig}.
{\bf Diffuse gamma-ray background} (DGRB). The DGRB is defined as a smooth residual component of the measured gamma-ray emission emerging after the subtraction of known sources of gamma rays, including both point-like and extended sources. The unresolved gamma-ray background (UGRB) can be explained by the cumulative emission of randomly distributed gamma-ray sources whose flux is below the sensitivity of the observing instrument. The UGRB between 100~MeV and 800~GeV is measured by the {\it Fermi} Large-Area Telescope (LAT)~\cite{Fermi-LAT:2010pat, Fermi-LAT:2014ryh}. The UGRB is expected to be contributed to largely by the faint subgroups of the bright gamma-ray source populations, including blazars~\cite{Ando:2006mt, Fermi-LAT:2015otn, DiMauro:2017ing} and star-forming galaxies~\cite{2010ApJ...722L.199F, Linden:2016fdd, Roth:2021lvk}. More exotic scenarios may contribute, as well-annihilating or decaying particles of dark matter in extragalactic halos may explain the diffuse backgrounds~\cite{Camera:2012cj, Camera:2014rja, Shirasaki:2014noa, Lisanti:2017qlb}.
{\bf Diffuse Supernova Neutrino Background} (DSNB). The detection of 25~MeV neutrinos from SN1987A in the Large Magellanic Cloud marked the first time neutrinos were detected from a massive star undergoing core collapse~\cite{Kamiokande-II:1987idp, Bionta:1987qt}. While the low galactic supernova rate requires much larger neutrino detectors to detect more supernovae neutrinos from nearby galaxies (1--10~Mpc), another avenue to study these explosions is available through the detection of the DSNB, which consists of MeV neutrinos from all past core-collapse supernovae. The discovery prospects of the DSNB in the next decade are promising with the gadolinium-enhanced Super-Kamiokande detector~\cite{Beacom:2003nk, Super-Kamiokande:2021the}, Jiangmen Underground Neutrino Observatory (JUNO~\cite{JUNO:2015zny}), and Hyper-Kamiokande~\cite{Abe:2011ts}.
{\bf Astrophysical Diffuse Neutrino Background}. The flux of diffuse neutrinos at TeV--PeV energies of astrophysical origin has been measured by IceCube with a significance well above $5\sigma$~\cite{IceCube:2013low, IceCube:2014stg}. Specifically, the flux has been measured using a sample of high-energy neutrinos, which includes both tracks and cascades with interaction vertices within the instrumented volume~\cite{IceCube:2020wum}, a sample of up-going tracks (mostly muon neutrinos)~\cite{IceCube:2021uhz}, a sample of cascade-like events (mostly electron and tau neutrinos)~\cite{IceCube:2020acn}, and a sample of tracks that start within the instrumented volume~\cite{IceCube:2018pgc}. An apparent slight tension between the different measurements could be due to differences in flavor composition, energy range, the accounting of atmospheric backgrounds, and the spectral model used. Future statistics and analyses with improved calibration and simulations will lead to improvements of the accuracy of the measurement and a reduction of systematic uncertainties.
{\bf Cosmogentic Neutrino Background}. Whether the diffuse astrophysical neutrino flux extends to higher energies is unknown. Determining if this flux has or does not have a cutoff in the 10--100~PeV range is crucial for understanding the physics underlying UHECR accelerators and identifying source classes. Studying the diffuse neutrino flux in this energy regime also opens an avenue to probe fundamental neutrino physics and BSM physics at an energy scale that would be otherwise unreachable.
{\bf Galactic Diffuse Emission.} In addition to the extragalactic diffuse emission, Galactic diffuse emission of gamma rays and neutrinos is produced by energetic cosmic rays interacting with the interstellar medium and radiation fields in our Galaxy. The Galactic diffuse gamma-ray emission has been measured by {\it Fermi}-LAT between 0.1~GeV and 1~TeV and by H.E.S.S.~\cite{HESS:2014ree} and HAWC~\cite{HAWC:2021bvb} above 1~TeV. The Galactic diffuse neutrino emission has been constrained by IceCube and ANTARES~\cite{IceCube:2017trr, ANTARES:2018nyb}, but is expected to be detected in the near future~\cite{Fang:2021ylv}.
The perspective of an {\bf indirect detection of dark matter} with the diffuse backgrounds depends on the level of the understanding of the astrophysical sources of astroparticles including, but not limited to, the fraction of astrophysical contribution, the faint end of the luminosity function of the astrophysical contributors, and the cosmological evolution of the source classes.
\subsection{Galactic TeVatrons and PeVatrons}
\label{subsec:TeVatronPeVatron}
The recent launch and operation of wide-field air-shower observatories, including the Tibet AS$\gamma$, HAWC, and LHAASO experiments, has opened up the view of the Universe in the very-high-energy (0.1--100~TeV) and ultra-high-energy ($> 100$~TeV) (note that these definitions of the energy ranges are adopted in gamma-ray astrophysics) regimes with unprecedented sensitivities. Ultra-high-energy gamma rays are produced by cosmic-ray protons and electrons at PeV energies. Detecting PeV proton accelerators, a.k.a. PeVatrons, are crucial to solving the long-standing puzzle of the ``knee" feature in the Galactic cosmic-ray spectrum. Several candidates have been identified so far~\cite{HAWC:2018gwz, Abeysekara:2021yum, TibetASg:2021kgt}, though more discoveries of sources and differentiation between leptonic and hadronic scenarios are needed to identify the highest-energy Galactic accelerators. Future VHE and UHE detectors with improved sensitivities, like the Southern Wide-field Gamma-ray Observatory (SWGO)~\cite{Albert:2019afb} and the Cherenkov Telescope Array (CTA)~\cite{CTAConsortium:2017dvg}, and neutrino experiments at TeV--PeV like IceCube-Gen2~\cite{IceCube-Gen2:2020qha}, KM3NeT~\cite{KM3Net:2016zxf}, P-ONE~\cite{P-ONE:2020ljt}, and Baikal-GVD~\cite{Baikal-GVD:2019kwy}, have the potential to unveil the nature of PeVatrons.
Dozens of VHE and UHE sources have been discovered by HAWC~\cite{HAWC:2020hrt} and LHAASO \cite{2021Natur.594...33C}, including many new ones that were not seen in other wavelengths. In particular, detection of few-degrees-extended gamma-ray emission, called halos, were first reported by HAWC~\cite{HAWC:2017kbo} around Geminga and Monogem, the two closest middle-aged pulsars, that could contribute to the positron excess measured by PAMELA~\cite{PAMELA:2013vxg} and AMS-02~\cite{PhysRevLett.122.041102}. More TeV halos have then been found by HAWC~\cite{HAWC:2020hrt} and LHAASO~\cite{LHAASO:2021crt}. The small angular size of the gamma-ray halos challenges traditional views of particle diffusion in the interstellar medium~\cite{Hooper:2017gtd, Tang:2018wyr, Fang:2018qco, DiMauro:2019yvh, DiMauro:2019hwn} and, so far, no convincing theoretical explanation of this effect has been proposed~\cite{Giacinti:2019nbu, Lopez-Coto:2017pbk, Evoli:2018aza, Liu:2019zyj, Recchia:2021kty, DeLaTorreLuque:2022chz}. The unexpectedly efficient confinement of electrons and positrons by pulsars could limit the astrophysical interpretation to the positron flux and hint at the necessity of exotic physics~\cite{HAWC:2017kbo}. Better understanding of the TeV halo population and their forming mechanism with future wide-field gamma-ray experiments is thus needed for the indirect detection of dark matter~\cite{Engel:2022yig}.
\input{HeavyElements}
\input{Spacetime}
\section{Current and Future Experiments}
\label{s:experiments}
A new age of precision cosmology and elucidation beyond the Standard Model with astroparticle physics has just begun. In the coming decades, there will be a large set of new probes to determine the cosmological parameters with unprecedented rigor, as well as an array of multimessenger experiments to discover new and exciting physics at energies not achievable by terrestrial accelerators. In this context, programmatic balance is imperative (see Fig.~\ref{fig:gantt_chart}).
The individual instruments involved in the multimessenger program are some of the most finely tuned that human hands have developed. Current gamma-ray, neutrino, cosmic-ray, and gravitational-wave facilities plan generally to increase their spectral coverage over the next two decades, while proposed, but currently unfunded, future facilities go well beyond just picking up at the sunset of their predecessors--- they are poised to unravel the mysteries of Universe and serve as fertile grounds for the discovery of new and exciting physics.
\begin{figure}[ht!]
\includegraphics[width=1.0\textwidth]{Figures/gantt_chart_panelled2D-cropped.pdf}
\caption{Timeline of current and proposed photon, gravitational-wave (GW), neutrino, and cosmic-ray (CR) facilities. Hatched regions indicate energies which proposed experiments would observe that would not be simultaneously observed by any current facilities. Over time, most messengers plan to increase their spectral coverage. The photon frame in blue illustrates continuous multi-wavelength coverage for the next two decades, with the glaring exception of MeV, GeV, and ultra-high-energy gamma rays. This impending gamma-ray gap is concerning to the broader multimessenger community. From Ref.~\cite{Engel:2022yig}.}
\label{fig:gantt_chart}
\end{figure}
The importance and great benefit of the involvement of DOE National Laboratories with those future experiments, both cosmological and astrophysical, cannot be overstated. The wealth of knowledge employed at these laboratories can be put to use by these experimental collaborations to achieve greater theoretical and technical progress than previously envisaged. This relationship is also incredibly symbiotic. As seen with the relationships between, e.g., Los Alamos National Laboratory and the HAWC Observatory, SLAC and \textit{Fermi}, Argonne National Laboratory and VERITAS, Fermilab and the Pierre Auger Observatory and Dark Energy missions, etc., this partnership enables the labs to work on smaller-scale experiments, in addition to their larger projects, and is a lucrative pathway for the recruitment and retention of highly skilled scientific minds to National Laboratories.
As elaborated in greater detail below and elsewhere in this Report, each of the Cosmic Probes brings unique access to one or more aspects of physics of the Standard Model and BSM physics and merits support as part of the HEP mission. Reflecting the maturity of the respective fields, the 4 types of Cosmic Probes (photons, neutrinos, UHECRs and GW) have different needs for development and increased US support in the next decades:
\begin{itemize}[noitemsep,topsep=0pt]
\item Next-stage gamma and neutrino investments should continue to be supported by US commitments including NASA and DOE, including infrastructure and financing. DOE investments in technology development for MeV gamma ray detection in colliders
and in next-generation air shower gamma-ray detectors will benefit this field as well.
\item UHECRs give unique access to UHE phenomena, but facilities are currently mainly funded by Europe; it would benefit the US HEP community to maintain and grow US involvement in the next generation.
\item Cosmic Explorer is US-lead, with international participation. Cosmic Explorer is probably the most dramatic new opportunity in the entire Cosmic Frontier portfolio. It is at a critical moment when additional involvement of HEP physicists and support for infrastructure and R\&D will have disproportionate returns.
\end{itemize}
\paragraph{Gamma-ray facilities} Gamma rays are vital messengers that carry information about an abundance of key scientific goals, both within our Galaxy and from the far reaches of extragalactic space.
They bring messages about naturally occurring particle acceleration throughout the Universe in environments so extreme they cannot be reproduced on Earth for a closer look and provide a window into beyond-the-Standard-Model Physics. Gamma-ray astrophysics is so complementary with collider work that particle physicists and astroparticle physicists are often one in the same, thus their facilities are vital tools in elucidating the mysteries of beyond-the-Standard-Model physics and astroparticle physics and for the discovery of new physics~\cite{Engel:2022bgx}.
While photons at different energies provide different pieces to each scientific puzzle, with no image being able to be completed with only a single input, the GeV-to-TeV-and-beyond energy regime hosts a highly successful set of current experiments, such as HAWC~\cite{HAWC:2019xhp}, VERITAS~\cite{2015ICRC...34..771P}, MAGIC~\cite{2016APh....72...76A}, H.E.S.S.~\cite{2006A&A...457..899A}, and LHAASO. Through their strict limits on PBHs, axion-like particles, CPT violation, and LIV, as well as observations of TeVatrons and PeVatrons, the discovery of gamma-ray halos, and countless other exciting scientific feats, these facilities have proven that gamma-ray facilities, especially those observing at the highest energies, are a force to be reckoned with. Complementary with these experiments and carrying on their heavy scientific loads in the next decade are two proposed future experiments: the Southern Wide-field Gamma-ray Observatory (SWGO)~\cite{Albert:2019afb, Hinton:2021rvp, Schoorlemmer:2019gee} and the Cherenkov Telescope Array (CTA)~\cite{CTAConsortium:2017dvg}. Building upon lessons learned from the current observatories and their predecessors, these facilities will have unprecedented sensitivity to the highest energies and are critical to carrying on the legacy of science at the forefront of particle and astroparticle physics.
Notably, current MeV and GeV gamma-ray facilities are expected to end before 2030 with no long-term plan to fill that gap in coverage that will impact, intrinsically, MeV and GeV science as well as make it impossible to collaborate with other wavelengths and messengers, effectively ending multimessenger science as we currently conceive of it. The timeline of current and proposed photon, gravitational-wave, neutrino, and cosmic-ray facilities is shown in Fig.~\ref{fig:gantt_chart}.
\begin{figure}[ht!]
\includegraphics[width=0.5\textwidth, trim = 14cm 0cm 14cm 0cm, clip]{multi_ch2-plot_lost_capability}
\includegraphics[width=0.5\textwidth, trim = 14cm 0cm 14cm 0cm, clip]{multi_ch2-plot_simplified}
\includegraphics[width=0.5\textwidth, trim = 14cm 0cm 14cm 0cm, clip]{multi_ch3-plot_lost_capability}
\includegraphics[width=0.5\textwidth, trim = 14cm 0cm 14cm 0cm, clip]{multi_ch3-plot_simplified}
\caption{{\it Top panels:} Connections between messengers and fundamental physics topics. {\it Bottom panels:} Connections between messengers and particle astrophysics topics. {\it Left panels:} Future multimessenger landscape with current facilities that are planned to continue operating and future facilities that are already funded. {\it Right panels:} Future multimessenger landscape with enhanced capabilities provided by proposed facilities. From Ref.~\cite{Engel:2022yig}.}
\label{fig:chord_plots}
\end{figure}
The loss of instrumental coverage in the MeV--GeV gap has broad implications for the goals of fundamental physics through the study of astronomical objects. Gamma rays are pivotal in the study of every major physics question in the coming decade. The lack of planned funding for this photon band should be truly alarming to those who have borne witness to the magnitude of recent multimessenger discoveries. The possible connections between fundamental physics questions, the astronomical objects through which they are studied, and observations that probe them by messenger and energy are shown in Fig.~\ref{fig:chord_plots}, where we note the potential loss of scientific excellence if key instrument classes are not prioritized over the next decade. For details, see Refs.~\cite{Engel:2022yig, Engel:2022bgx}.
There are many facility concepts in progress to improve cost and sensitivity for the gamma-ray band in this decade. A key area of investment for the future of multimessenger astrophysics is gamma-ray detector technology. Many aspects of instrumentation and software pipelines for cosmic gamma-ray detectors are nearly identical to those used in colliders, making this technology development extremely relevant to the broader particle physics community. For details, see Ref.~\cite{Engel:2022bgx}.
\paragraph{Neutrino facilities} The rich experimental program of neutrino-detection facilities is encapsulated in Fig.~\ref{fig:scales} and summarized in Ref.~\cite{Ackermann:2022rqc}. The next decade will result in the construction of multiple high-energy neutrino detectors spanning complementary regions of the sky, with differing sensitivity to different energy ranges between TeV and EeV, and complementary flavor-identification capabilities.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{scales.png}
\caption{Distribution of neutrino sources in energy and distance traveled to the detector, and experiments aimed at detecting them
that are presently in different stages of planning, design, and construction.
From Ref.~\cite{Ackermann:2022rqc}.}
\label{fig:scales}
\end{figure}
The neutrino oscillation program is entering a precision era, where the known parameters are being measured with an ever increasing accuracy. The IceCube Upgrade will provide the first precision measurement of the number of tau neutrinos appearing as a result of these oscillations~\cite{Ishihara:2019aao}. A measurement inconsistent with the poorly constrained current theory would be a smoking gun pointing to undiscovered types of neutrinos or to new physics.
\noindent
\begin{minipage}{0.48\textwidth}
The wide range of neutrino energies and traveled distances allow us to explore neutrino properties, their interactions, and fundamental symmetries across a wide breadth of parameter space.
\vskip20pt
Since neutrinos are neutral and weakly interacting, they carry information about the physical conditions at their points of origin; at the highest energies, even from powerful cosmic accelerators at the edge of the observable Universe. Due to the fact that they travel unscathed for the longest distances---up to a few Gpc, the size of the observable Universe---even tiny effects can accumulate and become observable. The potential for searches of beyond-SM physics in a wide energy range is illustrated in Fig.~\ref{fig:models} and summarized in Refs.~\cite{Ackermann:2022rqc, Arguelles:2022xxa, Abraham:2022jse}.
\end{minipage}
\noindent
\hfill
\begin{minipage}{0.48\textwidth}
\centering
\captionsetup{type=figure}
\includegraphics[width=0.9\textwidth]{model_classification.pdf}
\captionof{figure}{\label{fig:models}Models of new neutrino physics and other new physics classified according to the stage at which they act---at production, during propagation, and at detection---and what feature they affect---energy spectrum, arrival directions, flavor composition, and arrival times. From~Ref.~\cite{Arguelles:2019rbn}.}
\end{minipage}
\paragraph{Ultra-high-energy cosmic-ray experiments (UHECR)} In the coming decade, UHECR experiments will employ the three major detection techniques: extensive surface detector arrays, high-resolution air-fluorescence detectors, and radio detectors. The UHECR particle physics roadmap is specified in Fig.~\ref{fig:UHECRroadmap} and summarized in Ref.~\cite{Coleman:2022abf}.
\begin{figure}[!htb]
\centering
\includegraphics[width=\textwidth]{UHECRroadmap.pdf}
\vspace{-2mm}
\caption{Upgraded and next-generation UHECR experiments with their defining features, scientific goals relevant to the APS DPF, and timeline. From Ref.~\cite{Coleman:2022abf}.}
\label{fig:UHECRroadmap}
\vspace{-2mm}
\end{figure}
To address the paradigm shift arising from the results of the current generation of experiments, three upgrades are either planned or already underway. TA$\times$4, a 4-fold expansion of the Telescope Array, will allow for Auger-like exposure in the Northern Hemisphere with the aim of identifying (classes of) UHECR sources and further investigating potential differences between the Northern and Southern skies~\cite{TelescopeArray:2021dri}. AugerPrime, the upgrade of Auger, focuses on achieving sensitivity to the cosmic-ray baryonic composition for each shower, measured by its upgraded surface detector through multi-hybrid observations~\cite{PierreAuger:2016qzd}. IceCube-Gen2, IceCube’s planned upgrade, will include an expansion of the surface array to measure cosmic rays with energies up to a few EeV, providing a unique laboratory to study cosmic-ray physics such as the insufficiently understood prompt particle decays in extensive air showers~\cite{IceCube-Gen2:2020qha}. It will also be used to study the transition from galactic to extragalactic sources by combining the mass-sensitive observables of the surface and deep in-ice detectors. The upgrades benefit from recent technological advances, including the resurgence of the radio technique as a competitive method, and the development of machine learning as a powerful new analysis technique.
Looking into the future ahead, the POEMMA mission~\cite{POEMMA:2020ykm} and the multi-site Giant Radio Array for Neutrino Detection (GRAND)~\cite{GRAND:2018iaj} are two instruments that will measure both ultra-high-energy neutrinos and cosmic rays. Thanks to their large exposure, both POEMMA and GRAND will be able to search for UHECR sources and ZeV particles beyond the flux suppression. The Global Cosmic Ray Observatory (GCOS), a $40, 000~{\rm km}^2$ ground array likely split into at least two locations, one or more of them possibly co-located with a GRAND site, will be a purposely built precision multi-instrument ground array~\cite{Horandel:2021prj}. Its design will need to meet the goal of $< 10\%$ muon-number resolution to leverage our improved understanding of hadronic interactions. With these capabilities, GCOS will be able to study particle and BSM physics at the Energy Frontier while determining the cosmic-ray baryonic composition on an event-by-event basis to enable rigidity-based studies of UHECR sources at the Cosmic Frontier.
As we discussed in Sec.~\ref{s:portals}, all of these UHECR experiments will have sensitivity to signals of ssuper-heavy dark matter (SHDM) and macroscopic dark quark nuggets. Indeed, UHECR observatories will offer a unique probe of the dark matter mass spectrum near the GUT scale. The origin of SHDM particles can be connected to inflationary cosmologies and their decay to instanton-induced processes, which would produce a cosmic flux of ultra-high-energy neutrinos and photons. While their non-observation sets restrictive constraints on the gauge couplings of the dark matter models, the unambiguous detection of a single ultra-high-energy photon or neutrino would be a game changer in the quest to identify the dark matter properties. In particular, as we discussed in Sec.~\ref{s:SHDM}, AugerPrime will achieve a world-leading sensitivity to indirect detection of SHDM particles by searching for SHDM decay products coming from the direction of the Galactic center~\cite{PierreAuger:2022wzk}.
In addition, AugerPrime will provide a unique probe of hadronic interaction models at center-of-mass energies and kinematic regimes
not accessible at terrestrial colliders, as well as high-resolution measurements of the proton-air inelastic cross section $\sigma_{p-{\rm Air}}$. Hadronic interaction models, continuously informed by new accelerator data, play a key role in our understanding of the physics driving the production of extensive air showers induced by UHECRs in the atmosphere. Thanks to ever-more-precise measurements from UHECR experiments, there are now strong indications that our understanding is incomplete. In particular, all of the hadronic models underestimate the number of muons produced in air showers, hinting at new particle physics processes at the highest energies. Reducing the systematic uncertainties between models and incorporating the missing ingredients are major goals at the interface of the field of UHECRs and particle physics. The on-going AugerPrime upgrade will give each surface detector muon separation capabilities, allowing for high precision air-shower measurements connected to the muon puzzle~\cite{Albrecht:2021cxw} and probes of BSM physics; see Sec.~\ref{s:muonpuzzle}. The general strategy to solve the muon puzzle relies on the accurate determination of the energy scale combined with a precise set of measurements over a large parameter space, that can together disentangle the electromagnetic and muon components of extensive air showers. A muon-number resolution of $< 15\%$ is within reach with upgraded detectors in the next decade using hybrid measurements. Achieving the prime goal of $< 10\%$ will likely require a purposely-built next-generation observatory.
Additionally, as we discussed in Sec.~\ref{s:LIV}, UHECR experiments will provide the most restrictive bounds on violations of CPT and Lorentz invariance. Finally, the identification of the UHECR population could provide a direct probe of the species scale that could rule the cutoff energy of cosmic-ray accelerators~\cite{Montero:2022prj}; see Secs.~\ref{UHECRsources} and \ref{s:swampland}. Altogether, UHECR observatories offer an unparalleled opportunity to address basic problems of fundamental physics.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.99\textwidth]{Figures/FPFMap.pdf}
\caption{The Forward Physics Facility, a proposed new underground cavern located near the LHC tunnel at CERN. The FPF will house a diverse set of experiments in the far-forward region and will detect TeV-energy neutrinos, constrain forward hadron production, and probe proton and nuclear structure, with synergies with many astroparticle experiments. Adapted from Ref.~\cite{Feng:2022inv}.
}
\label{fig:FPFMap}
\end{figure}
\paragraph{Forward Physics Facility} The Forward Physics Facility (FPF) is a proposed underground cavern at the Large Hadron Collider (LHC) at CERN that will house a suite of far-forward experiments during the High-Luminosity LHC era from $\sim$ 2030-2042~\cite{Anchordoqui:2021ghd,Feng:2022inv}. The preferred site for the FPF is along the beam collision axis, 617-682 m west of the ATLAS experiment; see Fig.~\ref{fig:FPFMap}. FPF experiments, such as FASER$\nu$, Advanced SND, and FLArE, will detect $\sim 10^6$ neutrino interactions at TeV energies, filling the gap between current fixed-target accelerator experiments and astroparticle measurements; see Fig.~\ref{fig:scales}. In addition, the FPF will expand our understanding of proton and nuclear structure and the strong interactions to new regimes, and carry out world-leading searches for a wide range of new phenomena.
The FPF provides opportunities for interdisciplinary studies at the intersection of high-energy particle physics and modern astroparticle physics. Cosmic rays enter the atmosphere with energies up to $10^{11}$ GeV and beyond, where they produce large cascades of high-energy particles. The development of these extensive air showers is driven by hadron-ion collisions under low momentum transfer in the non-perturbative regime of QCD. Measurements at the FPF will improve the modeling of high-energy hadronic interactions in the atmosphere, reduce the associated uncertainties of air shower measurements, and thereby help to understand the properties of cosmic rays, such as their energy and mass, which is crucial to discovering their origin. Moreover, atmospheric muons and neutrinos produced in these extensive air showers in the far-forward region are the main background for searches of high-energy astrophysical neutrinos with large-scale neutrino telescopes, including IceCube and KM3NET. The FPF will help to understand the atmospheric neutrino flux and reduce the uncertainties for astrophysical neutrino searches in the context of multi-messenger astrophysics.
\begin{figure}[thb!]
\begin{center}
\includegraphics[width=0.50\textwidth]{sources-space-snowmass-v2.pdf}
\includegraphics[width=0.49\textwidth]{sources-ground-snowmass.pdf}
\end{center}
\caption{Strain sensitivity of various detectors and the expected signal strengths for different classes of sources plotted for the space-based LISA (left panel) and terrestrial detectors (right panel). See text for explanation of various sources plotted on the two diagrams.}
\label{fig:sense}
\end{figure}
\paragraph{Gravitational-wave facilities}
Gravitational wave detectors are sensitive to the signal amplitude and not energy or intensity. Consequently, an increase in the sensitivity of a gravitational-wave detector by a factor of 10 magnifies the accessible volume of the Universe by a factor of 1000 for low redshifts, where the geometry of the Universe is approximately Euclidean. All in all, Cosmic Explorer, conceived to be ten times bigger than Advanced LIGO, can observe essentially the entire universe for mergers of black holes and neutron stars.
The next generation of gravitational-wave observatories can explore a wide range of fundamental physics phenomena throughout the history of the Universe. These phenomena include access to the Universe's binary black hole population throughout cosmic time, to the Universe's expansion history independent of the cosmic distance ladders, to stochastic gravitational waves from early Universe phase transitions, to warped spacetime in the strong-field and high-velocity limit, to the equation of state of nuclear matter at neutron-star and post-merger densities, and to dark matter candidates through their interaction in extreme astrophysical environments or their interaction with the detector itself. A comparison of the strain sensitivity of these proposed detectors is shown in Fig.~\ref{fig:sense} and summarized in Ref.~\cite{Ballmer:2022uxx}. The right plot in Fig.~\ref{fig:sense} shows the sensitivity curves of advanced LIGO (aLIGO) and the next-generation laser interferometers: Cosmic Explorer, LIGO Voyager, and Einstein Telescope. Also shown on the same plot are the spectral densities of typical sources: GW150914 and GW170817 detected by the LIGO-Virgo Scientific Collaboration, binary neutron star (BNS) mergers at 450~Mpc and redshift of 2, GW150914 if it were at $z=10$, the Crab pulsar assuming an ellipticity of $\epsilon = 10^{-6},$ strengths of rotating neutron stars at 10~kpc for ellipticities $10^{-6}$ and $10^{-8},$ the neutron star in the low-mass X-ray binary Sco-X1 and other similar systems in the Galaxy (LMXBs), stochastic backgrounds of flat power spectrum $\Omega_{\rm GW}=10^{-9}$ and $\Omega_{\rm GW}=10^{-11}$, and radiation from quakes in neutron stars that deposit an energy $E \sim 10^{-12}\,M_\odot$ in gravitational waves. The left plot shows the sensitivity curve for the Laser Interferometer Space Antenna (LISA) together with coalescences of supermassive black hole binaries of various masses, inspiral of a 10\,$M_\odot$ black hole into a $10^6\, M_\odot$ black hole at $z=1$ (EMRI), the Galactic white dwarf binary (WDB) background as well as resolvable white dwarf binaries, AM Cn systems, and ultra-compact X-ray binaries. It is assumed that continuous waves from isolated neutron stars, white dwarf binaries, and stochastic backgrounds are integrated for a year, except for Sco-X1, for which an integration time of one week is assumed. Also see Ref.~\cite{Ballmer:2022uxx} for the sensitivity curves of the proposed Neutron Star Extreme Matter Observatory (NEMO) and MAGIS atom interferometers.
In the U.S., the proposed Cosmic Explorer observatory is designed to have ten times the sensitivity of Advanced LIGO and will push the reach of GW astronomy towards the edge of the observable Universe (redshift $z \sim 100$)~\cite{Reitze:2019iox, Evans:2021gyd}. Binary neutron star mergers at cosmological distances will be observable with Cosmic Explorer and LIGO Voyager. A network consisting of Cosmic Explorer in the U.S. and Einstein Telescope in Europe would detect $\sim 10^5$ binary neutron star mergers per year, with a median redshift of $\sim 1.5$ (close to the peak of star formation) and a horizon of $z \sim 9$~\cite{Borhanian:2022czq}. Approximately 200 of these binary neutron stars would be localized every year to better than one square degree, enabling followup with telescopes with small fields of view. The improved low-frequency sensitivity of next-generation detectors allows them to detect and localize sources prior to merger. A rough timeline of the various gravitational-wave detectors is given in Fig.\,\ref{fig:GW-obs-timeline}. The current plan in the US is to maximize the observation in the LIGO Facilities until Cosmic Explorer is observing. While there will be some breaks to further improve the sensitivity, actual observing time will be prioritized in coordination with the other terrestrial detectors of that epoch: LIGO-India, Virgo, and KAGRA.
In order to realize the full potential of current and future observatories improved waveform models would be needed to meet the greater sensitivity of next generation observatories. A new generation of numerical-relativity codes capable of achieving greater accuracy, smaller systematic bias and larger computational speeds, should be developed \cite{Foucart:2022iwu}. At the same time, it is important to harness analytical tools from high-energy physics, e.g. scattering amplitudes and effective field theory, and develop a framework for computing gravitational-wave signals from binary black holes and neutron stars \cite{Buonanno:2022pgc}. The synergy between the gravitational-wave and high-energy physics communities will help build waveform models that will be more accurate and mitigate systematic bias.
\begin{figure}[hbt!]
\centering
\includegraphics[width=\columnwidth]{GWobs-timeline}
\caption[GW Observatories' Timeline]{Timeline of current and proposed GW observatories. The LIGO, Virgo, and KAGRA timelines are taken from the frequently updated joint run planning page~\url{https://observing.docs.ligo.org/plan}.
The LISA launch date and mission lifetime are taken from the ESA-LISA Factsheet~\url{https://www.esa.int/Science\_Exploration/Space\_Science/LISA\_factsheet}.
}
\label{fig:GW-obs-timeline}
\end{figure}
\paragraph{Cosmological probes} The present tensions and discrepancies among different cosmological measurements, in particular the $H_0$ tension as the most significant one, offer crucial insights in our understanding of the Universe. In the near future, we expect precise measurements of the expansion and growth history over a large range of experiments. In Table~\ref{tab:cosmo1}, we provide a list of all these multi-frequency/multimessenger experiments together with the most influential probes and space missions from the last two decades. In Table~\ref{tab:cosmo2}, the experiments are grouped by their ``driving science'': detection of the redshifted 21 cm line in neutral hydrogen, BAO, redshift space distortion (RSD), cosmic chronometers (CC), CMB, distance ladder, fast radio bursts (FRB), GW, quasars, redshift drift, spectral distortions (SDs), supernovae (SNe), time-delay cosmography, time-lag cosmography, varying fundmental constants, and weak lensing (WL). A detailed description of these experiments is provided in Ref.~\cite{Abdalla:2022yfr} and in the CF6 report.
\begin{table}[ht]
\caption{{Cosmological probes. From Ref.~\cite{Abdalla:2022yfr}. }}
\label{tab:cosmo1}
\begin{center}
\scalebox{0.6}{
\begin{tabular}{|c|l|l|l|}
\hline
Acronym & Experiment & Website & Status \\
\hline
4MOST & 4-metre Multi-Object Spectroscopic Telescope & \href{https://www.eso.org/sci/facilities/develop/instruments/4MOST.html}{https://4MOST} & expected 2023\\
ACT & Atacama Cosmology Telescope &
\href{https://act.princeton.edu}{https://act.princeton.edu} & ongoing \\
ANDES & ArmazoNes high Dispersion Echelle Spectrograph & \href{https://elt.eso.org/instrument/ANDES/}{ https://ANDES} & planned\\
ATLAS Probe & Astrophysics Telescope for Large Area Spectroscopy Probe & \href{https://atlas-probe.ipac.caltech.edu/}{https://atlas-probe} & proposed \\
BAHAMAS & BAryons and HAloes of MAssive Systems & \href{https://www.astro.ljmu.ac.uk/~igm/BAHAMAS}{https://BAHAMAS} & 2017-2018\\
BICEP & Background Imaging of Cosmic Extragalactic Polarization & \href{http://bicepkeck.org/}{http://bicepkeck.org} & ongoing
\\
BINGO & Baryon Acoustic Oscillations & \href{https://bingotelescope.org/}{https://bingotelescope.org} & planned
\\[-4pt]
& from Integrated Neutral Gas Observations & &
\\
BOSS & Baryon Oscillations Spectroscopy Survey & \href{https://cosmology.lbl.gov/BOSS/}{https://BOSS} & ongoing \\
CANDELS & Cosmic Assembly Near-infrared Deep &\href{https://www.ipac.caltech.edu/project/candels}{https://candels}&\\[-4pt]
& Extragalactic Legacy Survey & & \\
CCHP & Carnegie-Chicago Hubble Project & \href{https://carnegiescience.edu/projects/carnegie-hubble-program}{https://carnegiescience.edu}&\\
CE & Cosmic Explorer & \href{https://cosmicexplorer.org}{https://cosmicexplorer.org}& planned\\
CFHT & Canada-France-Hawaii Telescope & \href{https://www.cfht.hawaii.edu}{https://cfht.hawaii.edu}& ongoing\\
CHIME & Canadian Hydrogen Intensity Mapping Experiment & \href{https://chime-experiment.ca/en}{https://chime-experiment.ca} & ongoing \\
CLASS & Cosmology Large Angular Scale Surveyor & \href{https://sites.krieger.jhu.edu/class/}{https://class} & ongoing \\
CMB-HD & Cosmic Microwave Background-High Definition & \href{https://cmb-hd.org}{https://cmb-hd.org} & proposed\\
CMB-S4 & Cosmic Microwave Background-Stage IV & \href{https://cmb-s4.org}{https://cmb-s4.org} & planned 2029-2036\\
COMAP & CO Mapping Array Pathfinder & \href{https://comap.caltech.edu}{https://comap.caltech.edu} & ongoing\\
DECIGO & DECi-hertz Interferometer Gravitational wave Observatory & \href{https://decigo.jp/index_E.html}{https://decigo.jp} & planned \\
DES & Dark Energy Survey & \href{https://www.darkenergysurvey.org}{https://darkenergysurvey.org} & ongoing \\
DESI & Dark Energy Spectroscopic Instrument & \href{https://www.desi.lbl.gov}{https://desi.lbl.gov}& ongoing\\
dFGS & 6-degree Field Galaxy Survey & \href{http://www.6dfgs.net}{http://6dfgs.net} & 2001-2007\\
eBOSS & Extended Baryon Oscillations Spectroscopy Survey & \href{https://www.sdss.org/surveys/eboss/}{https://eboss} & 2014-2019\\
ELT & Extremely Large Telescope & \href{https://elt.eso.org}{https://elt.eso.org} & planned 2027 \\
ESPRESSO & Echelle SPectrograph for Rocky Exoplanets & \href{https://www.eso.org/sci/facilities/paranal/instruments/espresso.html}{https://espresso.html} &ongoing \\[-4pt]
& and Stable Spectroscopic Observations & &\\
ET & Einstein Telescope & \href{http://www.et-gw.eu}{http://www.et-gw.eu} & planned \\
{\it Euclid} & {\it Euclid} Consortium & \href{https://www.euclid-ec.org}{https://www.euclid-ec.org} & planned 2023 \\
Gaia& Gaia &
\href{https://sci.esa.int/web/gaia/}{https://gaia} & ongoing\\
GBT& Green Bank Telescope &
\href{https://greenbankobservatory.org/science/telescopes/gbt/}{https://greenbankobservatory.org} & ongoing\\
GRAVITY& General Relativity Analysis via VLT InTerferometrY&
\href{https://www.mpe.mpg.de/ir/gravity}{https://gravity} & ongoing\\
GRAVITY+& upgrade version of GRAVITY& \href{https://www.mpe.mpg.de/ir/gravityplus }{https://gravityplus} & planned\\
HARPS & High Accuracy Radial-velocity Planet Searcher & \href{https://www.eso.org/sci/facilities/lasilla/instruments/harps.html}{https://harps.html} & ongoing\\
HIRAX & Hydrogen Intensity and Real-time Analysis eXperiment & \href{https://hirax.ukzn.ac.za}{https://hirax.ukzn.ac.za} & planned \\
HIRES & HIgh Resolution Echelle Spectrometer &\href{https://www2.keck.hawaii.edu/inst/hires/}{https://hires}&ongoing\\
H0LiCOW & $H_0$ Lenses in Cosmograil's Wellspring & \href{https://shsuyu.github.io/H0LiCOW/site/}{https://H0LiCOW} &\\
HSC & Hyper Suprime-Cam & \href{https://hsc.mtk.nao.ac.jp/ssp/survey}{https://hsc.mtk.nao.ac.jp} & finished\\
HST & Hubble Space Telescope & \href{https://www.nasa.gov/mission_pages/hubble}{https://hubble} & ongoing\\
KAGRA&Kamioka Gravitational wave detector&\href{https://gwcenter.icrr.u-tokyo.ac.jp/en/organization}{https://kagra}&expected 2023\\
KiDS & Kilo-Degree Survey & \href{http://kids.strw.leidenuniv.nl}{http://kids} & ongoing\\
JWST & James Webb Space Telescope & \href{https://jwst.nasa.gov/content/webbLaunch/index.html}{https://jwst.nasa.gov} & ongoing\\
LIGO & Laser Interferometer Gravitational Wave Observatory & \href{https://www.ligo.caltech.edu}{https://ligo.caltech.edu} & ongoing \\
LIGO-India &Laser Interferometer Gravitational Wave Observatory India&\href{https://www.ligo-india.in}{https://ligo-india.in}& planned\\
LiteBIRD & Lite (Light) satellite for the studies of B-mode polarization & \href{https://www.isas.jaxa.jp/en/missions/spacecraft/future/litebird.html}{https://litebird.html} & planned \\[-4pt]
& and Inflation from cosmic background Radiation Detection & & \\
LISA & Laser Interferometer Space Antenna & \href{https://lisa.nasa.gov}{https://lisa.nasa.gov} & planned\\
LGWA & Lunar Gravitational-Wave Antenna &\href{http://socrate.cs.unicam.it/index.php}{http://LGWA} & proposed\\
MCT & CLASH Multi-Cycle Treasury & \href{https://www.stsci.edu/~postman/CLASH/}{https://CLASH} & \\
MeerKAT & Karoo Array Telescope & \href{https://www.sarao.ac.za/science/meerkat/}{https://meerkat} & ongoing\\
NANOGrav & North American Nanohertz Observatory for Gravitational Waves & \href{http://nanograv.org/}{http://nanograv.org/} & ongoing\\
OWFA & Ooty Wide Field Array & \href{http://rac.ncra.tifr.res.in/ort.html}{http://ort.html} & planned\\
OWLS & OverWhelmingly Large Simulations & \href{https://virgo.dur.ac.uk/2010/02/12/OWLS}{https://OWLS} &\\
Pan-STARRS & Panoramic Survey Telescope and Rapid Response System & \href{https://panstarrs.stsci.edu}{https://panstarrs.stsci.edu} & ongoing \\
PFS & Subaru Prime Focus Spectrograph & \href{https://pfs.ipmu.jp}{https://pfs.ipmu.jp} & expected 2023
\\
{\it Planck} & {\it Planck} collaboration & \href{https://www.esa.int/Science_Exploration/Space_Science/Planck}{https://www.esa.int/Planck} & 2009-2013
\\
POLARBEAR & POLARBEAR & \href{http://bolo.berkeley.edu/polarbear/}{http://polarbear} & finished \\
PUMA &Packed Ultra-wideband Mapping Array & \href{http://puma.bnl.gov}{http://puma.bnl.gov} & planned
\\
{\it Roman}/WFIRST & Nancy Grace {\it Roman} Space Telescope & \href{http://roman.gsfc.nasa.gov}{http://roman.gsfc.nasa.gov} & planned\\
{\it Rubin}/LSST & {\it Rubin} Observatory Legacy Survey of Space and Time & \href{https://www.lsst.org}{https://lsst.org} & expected 2024-2034\\
SDSS & Sloan Digital Sky Survey & \href{https://www.sdss.org}{https://sdss.org} & ongoing\\
SH0ES & Supernovae $H_0$ for the Equation of State & \href{https://archive.stsci.edu/proposal_search.php?id=10802\&mission=hst}{https://SH0ES-Supernovae} & \\
SKAO & Square Kilometer Array Observatory & \href{https://www.skatelescope.org}{https://skatelescope.org} & planned\\
Simons Array & Simons Array & \href{http://bolo.berkeley.edu/polarbear/}{http://simonarray} & in preparation\\
SLACS & Sloan Lens ACS & \href{https://web.physics.utah.edu/~bolton/slacs/What\_is\_SLACS.html}{https://SLACS.html} & \\
SO & Simons Observatory & \href{https://simonsobservatory.org}{https://simonsobservatory.org} & expected 2024-2029\\
\hline
\end{tabular}}
\end{center}
\end{table}
\begin{table}[ht]
\begin{center}
\scalebox{0.6}{
\begin{tabular}{|c|l|l|l|}
\hline
Acronym & Experiment & Website & Status \\
\hline
SPHEREx & Spectro-Photometer for the History of the Universe, Epoch of Reionization, &\href{https://www.jpl.nasa.gov/missions/spherex}{https://spherex} & expected 2025\\
& and Ices Explorer & &\\
SPIDER & SPIDER & \href{https://spider.princeton.edu/}{https://spider} & planned \\
SPT & South Pole Telescope & \href{https://pole.uchicago.edu}{https://pole.uchicago.edu} & ongoing\\
STRIDES & STRong-lensing Insights into Dark Energy Survey & \href{https://strides.astro.ucla.edu}{https://strides.astro.ucla.edu} & ongoing \\
TDCOSMO & Time Delay Cosmography & \href{http://www.tdcosmo.org}{http://tdcosmo.org} & ongoing\\
uGMRT & Upgraded Giant Metre-wave Radio Telescope & \href{https://www.gmrt.ncra.tifr.res.in/}{https://gmrt.ncra.tifr.res.in} & ongoing \\
UNIONS & The Ultraviolet Near- Infrared Optical Northern Survey & \href{https://www.skysurvey.cc}{https://skysurvey.cc} & \\
UVES & Ultra Violet Echelle Spectrograph & \href{https://www.eso.org/public/teles-instr/paranal-observatory/vlt/vlt-instr/uves/}{https://uves} & ongoing\\
VIKING & VISTA Kilo-degree Infrared Galaxy Survey & \href{http://horus.roe.ac.uk/vsa/}{http://horus.roe.ac.uk/vsa/} & ongoing\\
Virgo & Virgo& \href{https://www.virgo-gw.eu}{https://virgo-gw.eu}& ongoing\\
VLA & Very Large Array & \href{https://public.nrao.edu/telescopes/vla/}{https://vla} & ongoing \\
VLBA & Very Long Baseline Array & \href{https://public.nrao.edu/telescopes/vlba/}{https://vlba} & ongoing \\
VLT &Very Large Telescope & \href{https://www.eso.org/public/teles-instr/paranal-observatory/vlt/}{https://vlt} & ongoing \\
WFC3 & Wide Field Camera 3 & \href{https://www.stsci.edu/hst/instrumentation/wfc3}{https://wfc3} & ongoing \\
WMAP & Wikilson Microwave Anisotropy Probe & \href{https://map.gsfc.nasa.gov}{https://map.gsfc.nasa.gov} & 2001-2010\\
YSE & Young Supernova Experiment & \href{https://yse.ucsc.edu}{https://yse.ucsc.edu} & ongoing \\
ZTF & Zwicky Transient Facility & \href{https://www.ztf.caltech.edu}{https://ztf.caltech.edu} & ongoing \\
\hline
\end{tabular}}
\end{center}
\end{table}
\begin{table*}[ht]
\caption{Cosmological probes grouped by their driven science. From Ref.~\cite{Abdalla:2022yfr}.}
\label{tab:cosmo2}
\begin{center}
\scalebox{0.65}{
\begin{tabular}{|l|l|l|}
\hline
Science & Facilities \\
\hline
21 cm & BINGO, CHIME, GBT, HIRAX, MeerKAT, OWFA, PUMA , SKAO, uGMRT\\
BAO and RSD & 4MOST, BINGO, CHIME, COMAP, DESI, Euclid, HIRAX, PFS, {\it Roman}, {\it Rubin}, SKAO, SPHEREx \\
CC & ATLAS, Euclid, SPHEREx\\
CMB & ACT, BICEP/Keck, CMB-HD, CMB-S4, LiteBIRD, SO, SPT \\
Distance ladder & ELTs, Gaia , GBT, JWST, LIGO, {\it Roman}, {\it Rubin}, VLA, VLBA \\
FRB & CHIME \\
GW & Cosmic Explorer, DECIGO , ET, LGWA, LIGO/Virgo/KAGRA/LIGO-India, LISA, Taiji, TianQin\\
Quasars & GRAVITY+ \\
Redshift drift & ANDES, ELTs, SKAO\\
SDs & SuperPIXIE \\
SNe & {\it Rubin}, {\it Roman}, YSE, ZTF \\
Time Delay cosmography & Euclid, Pan-STARRS, {\it Roman}, {\it Rubin}, SKAO, ZTF\\
Time Lag cosmography & {\it Rubin} \\
Varying fundamental constant & ANDES, ELTs, ESPRESSO\\
WL & 4MOST, CFHT,DES, Euclid, HSC, KiDS, Pan-STARRS , {\it Roman}, {\it Rubin}, SKAO, UNIONS\\
\hline
\end{tabular}}
\end{center}
\end{table*}
\section{Connections to other Snowmass Frontiers}
\label{s:opportunities}
Seeking the fundamental nature of matter and associated mysteries bridges the Theory, Accelerator, Energy, Instrumentation, Neutrino, Computational, and Cosmic Frontiers, thus connecting astroparticle physics and accelerator-based particle physics. Ergo, the study of astroparticle physics can have significant implications in the search for physics beyond the SM at the LHC and future colliders. Correspondingly, LHC experiments provide the laboratory for measurements relevant to understand the subtleties of astroparticle physics. We have provided specific examples of this synergy in Secs.~\ref{s:astropart_1} and \ref{s:muonpuzzle}, where we discussed the relation between the Forward Physics Facility (FPF) with neutrino telescopes and cosmic-ray observatories. This has been discussed in more detail in Refs.~\cite{Feng:2022inv, Coleman:2022abf}. All of these specific examples are, of course, also related to the Computational Frontier as explained in Ref.~\cite{Coleman:2022abf}.
There is also a strong synergy between cosmological and laboratory searches in new physics~\cite{Abazajian:2022ofy}. The relation between the Cosmic and Neutrino Frontiers has been emphasized in Sec.~\ref{s:nuosc}, with typical examples ranging from measurements of neutrino oscillation parameters to understanding the properties of neutrino masses~\cite{Abraham:2022jse, Ackermann:2022rqc, Arguelles:2022xxa} and bounds on the neutrino mass sum inferred from cosmological observations~\cite{Abdalla:2022yfr}.
The phenomenological implications of the Swampland program provide a strong connection between the Theory and the Cosmic Frontiers. The swampland conjectures seem to pose an interesting challenge for inflation, particle phenomenology, and the cosmological hierarchy problem. In Sec.~\ref{s:swampland}, we briefly related some of these topics, which are discussed at length in Refs.~\cite{Abdalla:2022yfr, deRham:2022hpx, Achucarro:2022qrl}. At the same time, the interface between early Universe cosmology and fundamental theories of particle physics ties up the Cosmic and Energy Frontiers~\cite{Flauger:2022hie, Agrawal:2022rqd}. Finally, searches for signals of particle dark matter and light relics provide the connector between the Accelerator, Energy, and Cosmic Frontiers~\cite{Feng:2022inv, Dienes:2022zbh, Dvorkin:2022jyg, Ando:2022kzd}.
Beyond the fundamental scientific complementarity between studying terrestrial and astroparticle physics, there is a deep connection between the instrumentation built for each subfield. The future of multimessenger astrophysics hangs on the development of new gamma-ray detector technology in the MeV and GeV range because all of the modern multimessenger co-detections involve gamma-rays in this range, and no new, long-term facilities are currently planned to replace those we may lose over the next decade. Gamma-ray detectors, discussed in Sec.~\ref{s:experiments} and further in Refs.~\cite{Engel:2022yig,Engel:2022bgx}, are developed and built using techniques, materials, and understanding from particle physics, and vice versa. The instrumentalists who build each type of detector move fluidly from one field to the other, giving particle physicists the opportunity to work on smaller experiments in astroparticle physics, and passing new technology development back into larger particle-physics experiments. Due to the interconnected nature of gamma-ray and collider detector technology development, this is a key area for collaborative investment across agencies in the coming decade, and could define multimessenger astrophysics for many decades to come.
\section{Diversity, Equity, Inclusion, and Accessibility}
\label{s:DEIA}
Peak scientific excellence for any country begins and develops in key government investment targeted at actions available today to produce tangible advancement tomorrow. While we often think of these investments in terms of technology development, flagship facilities, and the returns they enable, as a nation, we cannot afford to overlook investment in our scientific workforce. While this conversation must include broader generational aspects of fair compensation, scientific literacy in public education, and reasonable access to higher education without a lifetime of debt, there are also specific barriers to some people whose abilities would benefit national excellence in science, who are too often excluded for untenable reasons, collectively referred to as Diversity, Equity, Inclusion, and Accessibility (DEIA).
Diversity, equity, inclusion, and accessibility are fundamental elements of a modern and innovative workplace, school, or team. Cosmic Probes of Fundamental Physics and the Cosmic Frontier, more generally, are well poised to bring new ideas from the extensive literature on DEIA topics to the fore of the broader physics community. It is necessary to be mindful of cultural bias and the impact of personal experiences on student recruiting, training and opportunities, as well as on the retention of more senior trainees and experts. Prioritizing the recruitment and retention of bright minds over the shape or color of the body they come in or their socioeconomic background is of vital and imminent importance to innovation and excellence. This is an argument for providing the educational, mentoring, and community support to individuals in achieving their highest potential because that raises, rather than lowers the bar for academic, scientific, and competitive achievement for the nation as a whole.
For funding agencies, one of the most impactful changes that should be made over the next decade is to keep track of demographic information for collaborations they fund and for PIs (and other key leadership roles) specifically. That demographic information should include gender and ethnicity at a minimum, but may also include career stage, sexuality, institution type, and other items they might find relevant. The idea here being that there is demonstrated gender bias in awards and leadership roles as well as suspected racial bias, especially in small and mid-sized awards, which is where an earlier career person might start to build up their grant portfolio. Keeping those statistics and making them public in aggregate will allow the community to push for other changes and measure if they are effective, e.g., dual anonymous reviews or inclusion of DEIA service in science grants. At present, DEIA considerations are not a component of funding decisions for the grant host, but they sneak into the results through bias prone review processes. Advisors who tell their students for their (the students') own benefit not to spend time on activities that will not pan out as part of their career are not categorically in the wrong. The system they are advising for needs to support the work of tracking, studying, and supporting DEIA goals, so we ask for that support. The first step is the most important: track the demographics to map where the money goes. In the future, we hope to see hiring and performance reviews for scientists and researchers to include a component evaluating their service to DEIA in the same way that we often consider other services to the community, like mentoring and serving as a reviewer.
Recommendations for building a culture of equitable access and success for marginalized members in today’s Particle Physics Community have been presented in various Snowmass whitepapers and seminars~\cite{Engel:2022yig, Assamagan:2022ztm, Georgi:2022jfv}.
\clearpage
\bibliographystyle{JHEP}
|
\section{Introduction}
Quantum memories and transducers are enabling technologies for the global quantum internet. Optical quantum memories that coherently store and recall unknown states of a traveling optical field on demand \cite{Lvovsky2009} are a component of quantum repeaters \cite{briegel1998quantum,childress2005fault, sangouard2011quantum, asadi2018quantum} to enhance the entanglement distribution range, and therefore the secure communication range, of quantum networks \cite{kimble2008quantum} as well as a means of synchronizing processes with optical quantum computers. Microwave quantum memories, on the other hand, can interface with microwave frequency material qubits, including superconducting circuit qubits---arguably the most advanced quantum computing platform at present---to extend circuit depth and reduce no-op error rates \cite{PhysRevLett.127.140503}.
Long-coherence spin ensembles have been proposed as quantum memories for both applications, with considerable research into optical quantum memories in particular. To date, several quantum memory protocols including electromagnetically induced transparency (EIT) \cite{fleischhauer2005electromagnetically}, atomic frequency comb (AFC) \cite{afzelius2009multimode}, controlled reversible inhomogeneous broadening (CRIB) \cite{kraus2006quantum}, and Autler-Townes Splitting (ATS) \cite{Saglamyurek2018a} have been proposed based on platforms such as rare-earth ions \cite{ma2021one, hedges2010efficient}, NV centres in diamond \cite{heshami2014raman}, atomic vapours \cite{guo2019high, katz2018light}, and Bose-Einstein condensates (BEC) \cite{zhang2009creation, saglamyurek2021storing}. However, few of these systems operate in a telecommunications (telecom) band. Extending the quantum internet to global scale requires optical quantum memories compatible with satellite communications as well as terrestrial optical fibre networks.
Similarly, efficient and coherent conversion between gigahertz microwave and terahertz optical signals is essential for networking microwave material qubits \cite{Lauk2020}. Transduction from microwave to telecom photons could enable, for example, networking superconducting quantum processors and spin quantum processors \cite{Vandersypen2019,Xue2022} via the global quantum internet. There are several remarkable experimental demonstrations \cite{HaiTao2022, higginbotham2018harnessing, fan2018superconducting, vogt2019efficient} as well as theoretical approaches \cite{obrien2014, li2017quantum, asadi2022proposal} on reversibly converting between microwave and optical fields. In general, quantum transduction can be carried out through intermediate systems such as atomic ensembles \cite{obrien2014,Fernandez-Gonzalvo2015, HaiTao2022,vogt2019efficient}, electro-optical systems \cite{soltani2017efficient,fan2018superconducting}, electro-optomechanical \cite{stannigel2010optomechanical, higginbotham2018harnessing}, and others \cite{hisatomi2016bidirectional, das2017interfacing}.
Ensembles of atoms or atom-like systems combining long-coherence microwave spin qubits and optical transitions are therefore a versatile `swiss-army knife' platform suitable for microwave and optical memory as well as microwave to optical transduction. Identifying and developing long-coherence quantum systems combining narrow and absorptive microwave and telecom optical transitions is an important milestone for designing a quantum network. However, to date few suitable systems have been identified. Memory and transduction proposals of this kind have focused almost exclusively on Er$^{3+}$, which combines narrow MW and telecom optical transitions in many crystal hosts \cite{Williamson2014a,obrien2014,Fernandez-Gonzalvo2015}. Although the optical transition is comparatively weak, significant optical depths have been observed with concentrated Er$^{3+}$ ensembles in large crystals \cite{xie2021}.
The newly rediscovered silicon $T$ centre offers native O-band telecom access and long-lived spins operating at microwave frequencies \cite{Bergeron2020, higginbottom2022}. Notably, silicon is an attractive host material for such a system, as it is already the substrate of choice for a wide variety of quantum platforms including ion traps~\cite{Cho2015}, quantum dots~\cite{Veldhorst2017}, superconducting qubits~\cite{Arute2019}, and, naturally, impurity and defect spins in silicon~\cite{Pla2012} and silicon photonics~\cite{Wang2020}. The $T$ centre integrates with silicon photonic devices on CMOS-compatible wafers as a waveguide and fibre-networked quantum information platform \cite{higginbottom2022} that could further mediate between other silicon quantum technologies packaged on-chip, such as superconducting, trapped ion and gate-defined quantum dot qubits, for a complete, all-silicon hybrid quantum information platform.
In this paper we introduce the prospects of $T$ centres for microwave quantum memory, optical quantum memory, and microwave-optical quantum transducers. We will first briefly review the known properties of the $T$ centre as revealed by recent studies. Then, we further characterize $T$ centre ensembles experimentally towards the goal of quantum memories and transducers. We measure $T$ centre optical depth (OD) in \cref{sec:OD}, microwave coherent population trapping (CPT) in \cref{sec:CPT}, and microwave Autler-Townes splitting in \cref{sec:ATS}. Informed by these measurements, in \cref{sec:memory}, we estimate the efficiency of applicable optical memory schemes including EIT and ATS. In \cref{sec:mw} we derive the resonator requirements for efficient $T$ centre microwave quantum memories. In \cref{sec:transduction}, we consider the potential of combining these ATS schemes to design a transduction protocol. Finally, we conclude with future directions in \cref{sec:discussion}.
\section{The $T$ centre}
\label{sec:t_review}
The $T$ centre is a luminescent silicon defect with a sharp emission line at $935$~meV, in the telecommunications O band. Early work \cite{Irion1985a,Safonov1996b} established that $T$ centres comprise two nonequivalent carbon atoms and a hydrogen atom. Some of the $T$ centre parameters required to calculate quantum memory and transduction efficiencies, including the level structure, optical linewidths and lifetimes, and spin coherence times have recently been established \cite{Bergeron2020,MacQuarrie2021,higginbottom2022}. The remaining key parameters---optical depth, microwave coupling strengths and two-photon microwave linewidth---are obtained in this work and put to use in proof-of-principle quantum memory precursor experiments.
Spectroscopy of $T$ centre ensembles \cite{Bergeron2020} established that the ground state has a tightly bound s\,=\,1/2 electron coupled to the hydrogen nuclear s\,=\,1/2 spin through an anisotropic hyperfine interaction. In its optically excited state a bound exciton is formed, the electron spins form a singlet state, and there remains an unpaired s\,=\,3/2 hole. The four-fold degeneracy of free holes in silicon is lifted in the reduced symmetry of the $T$ centre to form two distinct excited state doublets labeled TX$_0$ and TX$_1$ separated by $1.76$~meV. Thermal excitation between these two states is negligible below $\sim2$~K. Studies of TX$_0$ with ensembles and single centres revealed remarkable optical linewidths sufficient to resolve the ground state electron spin splitting at low fields \cite{Bergeron2020,higginbottom2022}.
The TX$_0$ lifetime is 0.94~$\upmu$s in bulk silicon \cite{Bergeron2020}. The radiative efficiency, $\eta_\mathrm{R}$, of this transition is not precisely known. To date, measurements with bulk ensembles have not found evidence of non-radiative relaxation \cite{Bergeron2020}. First-principles theoretical calculations indicate $0.19 < \eta_\mathrm{R} < 0.72$ \cite{Dhaliah2022} and single-centre photon fluorescence rates bound $\eta_\mathrm{R} \geq 0.03$ \cite{higginbottom2022}. The zero-phonon fraction, or Debye-Waller (DW) factor, is known to be $\eta_\mathrm{DW}=0.23$ \cite{Bergeron2020}.
\begin{figure}[]
\includegraphics[width=\linewidth]{Fig1_system_schematic.pdf}
\caption{Schematic energy level diagram of the $T$ centre 935.1\,meV TX$_0$ optical transition. Under magnetic field, the unpaired electron spin of the paramagnetic ground state $T$ and unpaired hole spin of the excitonic excited state TX$_0$ split into doublets. The anisotropic hole spin determines 12 orientational subsets for a given magnetic field orientation, with optical transitions A$_i$--D$_i$ for subset $i$ that may be optically resolved. The nuclear spin transitions are resolvable with microwave addressing.}
\label{fig:1}
\end{figure}
The schematic level structure in \cref{fig:1} shows TX$_0$ under an applied magnetic field, with four resolved optical transitions A--D between the ground state electron and excited state hole states. The $T$ centre has 24 possible orientations forming 12 inequivalent inversion symmetric subsets $i$, each with its own set of optical transitions A$_i$--D$_i$ determined by the effective anisotropic hole Land\'{e} factor $0.85 < g_{\mathrm{H}i} < 3.5$ for a particular magnetic field orientation. Optically detected magnetic resonances (ODMR) reveal ground state hyperfine splitting under microwave or RF excitation. Both the `allowed' nuclear spin preserving microwave transitions, MW$_\mathrm{\Uparrow}$ and MW$_\mathrm{\Downarrow}$, and the `forbidden' microwave transitions, X$_\mathrm{m}$ and X$_\mathrm{p}$, are observed to be strong. The $T$ centre therefore boasts both optical and microwave $\Lambda$ and V transitions for interfacing between travelling fields and spins. The nuclear spin states of the hydrogen nucleus, which have not yet been optically resolved, would become optically accessible in lifetime-limited homogeneous ensembles.
Isotopically purified \ensuremath{^{28}\text{Si}}{} improves upon both the spin and optical properties of the $T$ centre. Removing $^{29}$Si atoms (spin 1/2 nuclei) reduces the magnetic noise bath and improves native spin coherence times. Removing mass variations due to both $^{29}$Si and $^{30}$Si dramatically reduces the optical inhomogeneous broadening. Ensemble optical inhomogeneous linewidths as low as $33$~MHz have been observed in \ensuremath{^{28}\text{Si}}{} $T$ centre ensembles \cite{Bergeron2020}. Measurements of spectral diffusion have indicated that the \ensuremath{^{28}\text{Si}}{} ensemble optical linewidths may be close to diffusion limited, and that spectral diffusion remains low in natural silicon samples even as the inhomogeneous linewidth increases to $9$~GHz \cite{MacQuarrie2021}. Hahn echo measurements in \ensuremath{^{28}\text{Si}}{} crystals revealed $T$ centre electron and nuclear spin coherence times ($T_2$) of $2.1$~ms and $1.1$~s respectively \cite{Bergeron2020}. Both $T$ centre spins offer a memory advantage compared to superconducting qubits, and the nuclear spin coherence time is furthermore sufficient for repeaters in terrestrial quantum networks.
In the following sections we extend these studies by characterizing the optical absorption properties of \ensuremath{^{28}\text{Si}}{} $T$ centre ensembles to determine the requirements for $T$ centre optical quantum memories. We further implement the first coherent $\Lambda$ and V schemes for $T$ centres as a precursor to quantum memory and transduction protocols.
\section{Optical depth}
\label{sec:OD}
An efficient quantum memory fundamentally requires signal absorption approaching 100\% for the reversible transfer of information. Here we report the first optical absorption measurements of $T$ centres and extrapolate what optical depths are feasible within the near term for $T$ centre quantum memories. An $\approx5$~mm thick \ensuremath{^{28}\text{Si}}{} sample was electron-irradiated and annealed according to the procedure in \cref{sec:Methods} to produce a significant concentration of $T$ centres. We measure the complete photoluminescence (PL) spectrum under above-band excitation with Fourier transform infrared (FTIR) spectroscopy, \cref{fig:5}(a). Detail about the prominent TX$_0$ ZPL is shown in \Cref{fig:5}(b). In addition to the PL spectrum (blue), we resolve the ZPL by photoluminescence excitation (PLE) spectroscopy, resonantly exciting the TX$_0$ with a scanning laser and detecting fluorescence from the $T$ centre phonon sideband to determine an ensemble ZPL linewidth of $56$~MHz (orange). This linewidth is an inhomogeneous factor of 331 over the lifetime-limited linewidth, and slightly larger than the best measured \cite{Bergeron2020}. The corresponding FTIR optical transmission spectrum (black), taken with a continuum light source and $1.4$~K sample temperature, is presented on a separate axis of \Cref{fig:5}(b). Absorption at the $T$ centre ZPL is evident with peak absorption of 0.27(1)\%, limited by the instrument resolution. Assuming the true absorption linewidth is identical to the PL linewidth, we infer a corrected peak absorption of 0.93(3)\% corresponding to peak resonant OD $d_\mathrm{corr.} = 0.009$ and resonant absorption coefficient $\alpha_\mathrm{corr.} = 0.017$/cm.
\begin{figure}[]
\includegraphics[width=\linewidth]{Fig2_ensemble_absorption.pdf}
\caption{(a) PL spectrum of a $T$ centre ensemble in \ensuremath{^{28}\text{Si}}. (b) High-resolution PL spectrum about the TX$_0$ zero phonon peak (blue) and PLE spectrum of the same sample (orange). The corresponding absorption spectrum (black) is instrument resolution limited. We calculate an area-preserving corrected absorption spectrum (dashed black) with linewidth and shape matched to the PLE spectrum.}
\label{fig:5}
\end{figure}
This first absorption measurement allows us to assess the feasibility of $T$ centre ensembles in the complete absorption (OD$>5$) regime required for optical quantum memories. First, we can compare the concentration of this sample to theoretical limits and concentrations observed in other work. The $T$ centre concentration $[c]$ can be determined from the integrated absorption line area $\int \! \alpha \,\mathrm{dV}$ shown in Fig.~\ref{fig:5} \cite{Hilborn1982}
\begin{equation}
[c] = \frac{g_1}{g_2} \frac{8 \pi n^2 \tau_\mathrm{ZP} \int \alpha \mathrm{dV} }{h \lambda^2},
\label{eq:lifetime}
\end{equation}
\noindent where $g_{1,2}$ are the degeneracy of the ground and excited states, respectively, $n = 3.45$ is the silicon index of refraction at cryogenic temperatures, $\tau_\mathrm{ZP} = \tau/\left(\eta_\mathrm{R} \eta_\mathrm{DW} \right)$ is the ZP radiative lifetime, $h$ is Planck's constant, and $\lambda = 1326$~nm is the wavelength.
Assuming the demonstrated upper (lower) bound radiative efficiency $\eta_\mathrm{R} = 1 (0.03)$ we determine the $T$ centre bulk concentration $[c] = 8.2 (2)\times10^{10}$~cm$^{-3}$ ($2.7 (0.7) \times 10^{12}$~cm$^{-3}$). This is, to our knowledge, the highest concentration of $T$ centres measured in \ensuremath{^{28}\text{Si}}{}. In this case concentration is limited by the very low residual carbon contamination of our sample (given in \cref{sec:Sample}).
Concentrations up to $100$ times larger than this lower concentration bound, at least $1.7\times10^{13}$~cm$^{-3}$, have been produced in silicon photonics using a carbon implantation recipe \cite{higginbottom2022}. Orders of magnitude improvement in concentration have also been achieved using ion implantation for other radiation damage centres \cite{Murata2011,Rotem2007}. This bodes well for the feasibility of \ensuremath{^{28}\text{Si}}{} ensembles with considerable optical depth. We expect an order of magnitude improvement over the measured $T$ concentrations in Ref.~\cite{higginbottom2022} is possible. In the unit radiative efficiency case, a $15$~mm $^{28}$Si crystal with such a $T$ concentration will have optical depth $d=27$ suitable for efficient quantum memory. We explore the optimum memory schemes for such an ensemble in \cref{sec:memory}. It remains to be seen what additional inhomogeneous broadening is introduced in [c]$\sim 10^{14}$~cm$^{-3}$ $T$ ensembles, which could reduce the peak resonant OD. However it is also possible that further material or implant optimization will reduce the inhomogeneous linewidth, with corresponding OD improvements.
In addition to concentration or sample length improvements, optical resonators can be used to increase the resonant OD of spin-photon ensembles. Many of the most efficient transduction demonstrations leverage optical resonators \cite{higginbotham2018harnessing,fan2018superconducting}, although microwave to optical upconversion has been demonstrated with $82$\% efficiency with a cavity-free atomic ensemble \cite{HaiTao2022}. Silicon is a superb platform for monolithic resonators due to the extremely low intrinsic loss coefficient of silicon in the telecommunications bands ($1.6\times 10^{-5}$ at $1326$~nm \cite{Green2008}). Silicon Fabry-Perot and whispering gallery mode \cite{Markosyan2018} resonators have been demonstrated with finesse many orders of magnitude beyond the $F\sim1000$ required for efficient optical quantum memory with even the limited concentration of our \ensuremath{^{28}\text{Si}}{} $T$ sample. These resonators were manufactured from intrinsic FZ silicon, but no performance reduction is expected using \ensuremath{^{28}\text{Si}}{}. Even with the broader inhomogeneous linewidths that have been measured so far, high density $T$ centres implanted in natural silicon samples ($9$~GHz linewidth \cite{MacQuarrie2021}) can be enhanced to optical depth sufficient for quantum memory with only the same $F\sim1000$ low finesse cavities.
Finally, ensembles can be incorporated into integrated silicon on insulator (SOI) silicon photonic crystal \cite{Asano2017} or waveguide ring cavities. Typical intrinsic waveguide losses in such devices are $\sim2.5$~dB/cm, imposing an upper limit on the effective ensemble length. Integrated ensembles therefore require higher concentrations than the factor of ten improvement forecast above, or narrower integrated inhomogeneous linewidths than those demonstrated in devices to date ($18$--$50$ GHz) \cite{MacQuarrie2021,higginbottom2022}. In any case, it is necessary to consider how the inequivalent $T$ centre orientational subsets will impact memory performance. Memory schemes utilizing only a single orientation will suffer an optical depth penalty as high as a factor of $12$, but possibly less depending on orientation degeneracy for a chosen magnetic field direction. This orientation penalty would be reduced by the (untested) ability to form a given orientation preferentially.
\section{Microwave coherent population trapping}
\label{sec:CPT}
As a precursor to quantum memory protocols, we next demonstrate microwave (MW) coherent population trapping (CPT) \cite{Arimondo1976} with another \ensuremath{^{28}\text{Si}}{} $T$ centre ensemble. CPT is a quantum phenomenon in which an equilibrium atomic superposition is prepared by two coherent electromagnetic fields \cite{Gray1978}. CPT underpins three-level coherent atom-light phenomena, including EIT \cite{Boller1991a} and stimulated Raman adiabatic passage (StiRAP). Observing an EIT window requires sufficient ensemble absorption, but the accompanying CPT can be observed even in low optical depth ensembles by ODMR spectroscopy.
The use of optical fields to generate coherent dark states has been extensively studied \cite{Alzetta1976}. In comparison, relatively little effort has been focused on MW CPT \cite{Rakhmatullin1998,Wei1999,Childress2010,Novikov2015,Jamonneau2016}.
The optical ground state of the $T$ centre under magnetic field $B_0$ boasts two possible microwave $\Lambda$ schemes for coherently coupling the nuclear spin states $\ket{1}$ and $\ket{2}$ via the higher energy electron spin states $\ket{3}$ and $\ket{4}$ as shown in \cref{fig:1}. The $T$ centre is therefore a potential platform for $\Lambda$ microwave quantum memory schemes.
To demonstrate this capability a \ensuremath{^{28}\text{Si}}{} $T$ sample, prepared by the same method, is chosen for its narrow optical inhomogeneous linewidth of $33$~MHz. A static magnetic field $B_0 = 80$~mT is applied along the [110] axis of the sample, splitting the TX$_0$ transition. Under this magnetic field $10$ of the $12$ inequivalent orientational subsets $i$ produce their own set of four optically resolved transitions A$_i$, B$_i$--D$_i$ determined by the Land\'{e} $g$ tensor of the anisotropic hole. Following Ref.~\cite{Bergeron2020}, which used the same magnetic field configuration, we label the two remaining optically unresolved orientations $i=1$,$1'$. We may select distinct subsets for ODMR measurement by resonant optical excitation. For the following ODMR measurements we choose the $i=1$,$1'$ subsets. Although these subsets are not optically resolved \cite{Bergeron2020}, we will resolve the distinct orientations by CPT.
\begin{figure}[t!]
\includegraphics[width=\linewidth]{Fig3_mw_cpt_lambda.pdf}
\caption{ODMR spectra show the sideband photoluminescence as a function of MW probe detuning. Resonant optical excitation of TX$_0$ C provides electron-spin selective readout from $\ket{\uparrow_\mathrm{e}}$. (a) Two of the four resonances MW$_{\Uparrow /\Downarrow}$, X$_\mathrm{p,m}$ are visible depending on the configuration of a MW pump (pump frequency arrow matches corresponding spectra colour). The outer resonances are MW $\Lambda$ configurations. (b, c) High resolution ODMR spectra of the $\Lambda$ resonances show narrow CPT dips at two-photon resonance for each of the 1 and 1' orientation sub-ensembles. Inset Level schemes show the corresponding optical and MW configuration.}
\label{fig:2}
\end{figure}
A `readout' laser resonant with C$_1$ selects these two optically degenerate subsets for ODMR spectroscopy. In the absence of any additional field the readout simply pumps these two optically degenerate orientational subsets into the electron spin state $\ket{\downarrow_e}$ and this hyperpolarized ensemble does not fluoresce. Adding a single pump MW field that selects only one of the two hyperfine split states is still insufficient to prevent hyperpolarization, instead pumping the system to whichever of hyperfine states $\ket{1}$,$\ket{2}$ remains unaddressed. In these ODMR measurements a MW `pump' is set resonantly with one of the four available MW transitions and a second MW `probe' scans over all four transitions. Continuous photoluminescence is detected when hyperfine states $\ket{1}$ and $\ket{2}$ are separately addressed by the MW fields. The ODMR spectra in \cref{fig:2}(a) show two resonances per pump frequency. As expected resonances involving the `forbidden' X$_\mathrm{p,m}$ are possible, but weaker and narrower than resonances with only MW$_{\Downarrow, \Uparrow}$ (linewidths of 462(13), 506(9), 698(8) and 583(7)~kHz respectively).
CPT is evident in each of the two possible MW $\Lambda$ configurations, shown inset to \cref{fig:2}(a). When the MW fields are in two-photon resonance CPT produces a dark state superposition of $\ket{1}$ and $\ket{2}$ that does not couple to the readout laser. Higher resolution ODMR spectra in the region of the two-photon resonances, \cref{fig:2}(b) and (c), show CPT luminescence dips with $16$($2$)~kHz linewidth---significantly narrower than the ODMR lines. For each configuration in \cref{fig:2} there are two dark resonances, one for each of the distinct orientational subsets 1 and 1' and separated by $242(1)$~kHz, below the inhomogeneous linewidth and unresolvable by above-band PL. The low-power CPT linewidths are nuclear spin coherence limited, which is already known to be longer than $1$~s at this temperature \cite{Bergeron2020}. However, at the microwave powers required to produce sufficient luminescence signal these CPT linewidths are broadened by the microwave Rabi frequencies $\Omega_{\Uparrow/\Downarrow}$, $\Omega_\mathrm{p/m}$.
The observed ODMR and CPT profiles in \cref{fig:2} are the sum of fluorescence from the two near-degenerate orientational sub ensembles 1 and 1', in proportion determined by relative concentration and dipole orientation. The precise ODMR peak associated with each ensemble is unknown. When one ensemble is in two-photon dark resonance, fluorescence from the remaining out-of-resonance ensemble is still present, with the effect that neither CPT dip apparently achieves complete fluorescence extinction as expected in the ideal low-power case.
\begin{figure}[]
\includegraphics[width=\linewidth]{Fig4_mw_cpt_vee.pdf}
\caption{ODMR spectra with optical readout via TX$_0$ B$_1$. In this configuration CPT is possible in each of the two V schemes, inset (a) and (b) alongside the corresponding data. Coherences are prepared in the opposite electron spin state from \cref{fig:2}}
\label{fig:3}
\end{figure}
Optically pumping transition C empties states $\ket{3}$ and $\ket{4}$, rapidly destroying any coherence $\hat{\rho}_{34} = \braket{3|\hat{\rho}|4}$, where $\hat{\rho}$ is the density matrix of the centre, and power broadening the MW transitions. However the choice to read using optical transition C is arbitrary, and we can invert the system to achieve two-photon coherence via the less common V scheme. Choosing transition B instead empties $\ket{1}$ and $\ket{2}$ and removes coherence $\hat{\rho}_{12}$. This configuration sustains nuclear spin coherences of the $\ket{\uparrow_\mathrm{e}}$ hyperfine manifold, and we therefore expect CPT in V two-photon resonance, as shown in \cref{fig:3}. The V CPT linewidths are $17(2)$~kHz, identical to the $\Lambda$ case. The pump-probe MW frequency difference at each $\Lambda$ or V CPT condition determines effective hyperfine constants $-2.93(1)$~MHz and $-2.57(1)$~MHz for the 1 and 1' sub-ensembles, respectively.
Dark CPT steady states decouple from the two fields in either $\Lambda$ or V configurations and cause a corresponding EIT window in the ensemble absorption. Absorption and dispersion manipulation by such electromagnetically induced transparency (EIT) is the operational principle of EIT quantum memories. In \cref{sec:memory} we consider the prospects for $T$ centre EIT memories (both optical and microwave), but first we perform one final preparatory measurement. In the same $\Lambda$ or V configuration, but in the limit of increasing pump power, the CPT window transitions to an Autler-Townes split transition, which is itself the basis of an Autler-Townes quantum memory.
\section{Autler-Townes splitting}
\label{sec:ATS}
The operational bandwidth of EIT quantum memories is limited by the two-photon transition linewidth. Autler-Townes (AT) quantum memories utilize selective absorption by the spectrally distinct AT-split dressed states of a single transition to circumvent this limit \cite{Saglamyurek2018a}. With a modified ODMR scheme we can observe AT splitting of MW transitions in the same $T$ ensemble. For this measurement a slightly smaller magnetic field is chosen, $B_0 = 60$~mT along the [110] axis of the sample.
AT dressed states are split by the pump Rabi frequency $\Omega_\mathrm{pump}$. Rabi frequencies fitted to the CPT notches above, and observed by EPR with the same apparatus \cite{Bergeron2020}, indicate that attainable AT splittings $\Omega_\mathrm{pump} < 0.2$~MHz will not be clearly resolved with the ODMR schemes above. Instead, we apply an alternative ODMR scheme in which a radio frequency (RF) probe field scans over one of the NMR transitions. By this method we isolate a single sub-ensemble and achieve ODMR linewidths narrow enough to measure the transition to AT splitting as $\Omega_\mathrm{pump}$ increases. \Cref{fig:4}(a) shows ODMR spectra of the transition $\ket{3}$--$\ket{4}$ as the MW power resonant with MW$_\Uparrow$ increases. The RF probe field is swept across the $\ket{\uparrow_\textrm{E}}$ nuclear transition between states $\ket{3}$ and $\ket{4}$. As the MW pump power increases, the RF resonances split by $\Omega_\mathrm{MW}$, the microwave Rabi frequency.
We measure AT splitting up to $55(1)$~kHz, demonstrating the basic mechanism required for microwave AT quantum memories. Splitting is limited in this case by the available MW power and inefficient resonator design. Larger splittings for improved optical and MW memory bandwidths are possible. In the following section we compare EIT and AT memory models to assess the prospects of $T$ centre memories for microwave and optical fields, and determine the optimal bandwidth for operating $T$ centre memories at feasible optical depths.
\begin{figure}[]
\includegraphics[width=\linewidth]{Fig5_mw_ats.pdf}
\caption{Autler-Townes splitting of the MW$_\Downarrow$ transition as a function of MW$_\Downarrow$ power. (a) Spectra taken by scanning a non-phase coherent RF source over the nuclear flip transition. ODMR PL signal is generated by exciting to the TX$_0$ state by a laser resonant with the B$_1$ optical transition (scheme shown inset in (b)). As the MW$_\Uparrow$ power (text labels) increases, AT splitting becomes clear. Dashed lines are a split-peak fit to each dataset. (b) Splitting from fits (equal to the MW$_\Uparrow$ Rabi frequency $\Omega_\mathrm{pump}$) as a function of resonant power.}
\label{fig:4}
\end{figure}
\section{Free-space optical memory }
\label{sec:memory}
The coherence times of the $T$ centre electron and nuclear spins, $<2.1$~ms and $<1.1$~s respectively, are appealing for quantum memory in either optical quantum networks or quantum processors. It remains to determine which of the many quantum memory schemes are best suited to the optical and microwave properties of $T$ centre ensembles determined in the literature and in the measurements above. Storage of a weak signal (i.e., probe) pulse using a relatively strong control field in a $\Lambda$-type system has been studied using different protocols such as EIT, off-resonant Raman, and ATS \cite{lukin2003colloquium,Lvovsky2009,reim2010towards,Saglamyurek2018a}.
Keeping in mind that the nuclear splitting is not optically resolvable, there are four optical transitions per $T$ centre orientation that can be addressed individually and formed into two possible optical $\Lambda$ configurations addressing the electron spin coherence. In the ideal lifetime-limited case a $T$ centre optical quantum memory could directly address the longer-lived nuclear spin coherence, as in the MW memory proposal below, for storage times beyond $1$~s. However even the electron spin coherence time of $\approx2$~ms is sufficient for a basic repeater demonstrations and swapping between the electron and nuclear spins is possible.
In this section we consider theoretically the $\Lambda$-type system comprising two long-lived ground states $\ket{g_1} = \ket{\uparrow_e}$ and $\ket{g_2} = \ket{\downarrow_e}$ that are optically connected to a common excited state $\ket{\downarrow_h}$. We assume both electron and nuclear spins are initially polarized. We then use a strong control field with the Rabi frequency of $\Omega$ in resonance with the optical transition $B$, to store a weak signal pulse coupled to the $A$ transition (see \cref{fig:1}).
Based on the optical depth measurement performed in \cref{sec:OD}, $T$ centre ensembles could be made with optical depth sufficient for resonant optical quantum memory schemes given feasible concentration improvements. However, absent cavity-enhanced absorption, even these optical depths are insufficient for off-resonant schemes including gradient echo memory (GEM). Resonant memory schemes that efficiently utilize the available optical depth are best suited to near-term $T$ centre ensembles. In the following subsections we consider in detail the prospects of EIT and ATS free-space optical quantum memory schemes. Although \ensuremath{^{28}\text{Si}}{} $T$ centre ensembles exhibit optical inhomogeneous broadening of at least 33 MHz \cite{Bergeron2020}, throughout this
analysis we consider spectral tailoring of the ensemble bounded from below by the homogeneous linewidth $\Gamma/2\pi$. Therefore, we neglect the effect of inhomogeneous broadening in our estimations.
\subsection{EIT Memory}
In the EIT protocol, ramping down the control field power gradually to zero allows the adiabatic transfer of signal coherence to the spin-wave mode. EIT can be implemented effectively in both narrowband, i.e. $B_{\text{sig}}<\Gamma/2\pi$, and broadband, i.e. $B_{\text{sig}}>\Gamma/2\pi$, signal regimes where $B_{\text{sig}}$ is the bandwidth of the signal at full-width half-maximum, and $\Gamma/2\pi$ is the linewidth of the signal transition. However, this protocol is best suited for the narrowband regime where control field optimization is achievable at any optical depth. Eliminating signal absorption and maintaining adiabatic evolution with a broad transparency window requires both a very large optical depth and a strong control field \cite{wei2020broadband, rastogi2019discerning}.
In general, EIT efficiency optimization is achievable through control field and/or signal pulse optimization \cite{gorshkov2007freespace,novikova2007optimal,novikova2008optimal}. The consistency of these optimization methods has been demonstrated in Ref. \cite{phillips2008optimal}. In the narrowband regime and for $d>20$, control field optimization is achievable by keeping the process adiabatic while satisfying the condition $\tau_d/\tau_{\text{sig}}\approx2$, where $\tau_d=d\,\Gamma/\Omega^2$ is the group delay and $\tau_{\text{sig}}$ is the signal time at FWHM \cite{rastogi2019discerning}. Considering the probe field with a Gaussian temporal profile, the control optical Rabi frequency $\Omega$ should therefore satisfy the condition $\Omega^2 \approx d \,\Gamma B_{\text{sig}} /0.88$.
For the $T$ centre with $d=27$, corresponding to feasible near-term optical depths as discussed in \cref{sec:OD}, the maximum achievable efficiency of the EIT protocol for the forward retrieval can be estimated as $\sim71\%$ \cite{phillips2008optimal}. The memory efficiency for the backward propagation depends on the ground states splitting $ \omega_{g_1 g_2}$ (as the non-zero splitting breaks the conservation of momentum in backward
retrieval \cite{gorshkov2007freespace}). Therefore, for a non-zero splitting, the backward efficiency could be lower than the efficiency of the forward retrieval unless we make $\sqrt{d}\gg L \omega_{g_1 g_2} /c$ where $L$ is the length of the medium and $c$ is the speed of light.
This EIT memory efficiency estimate applies for short storage times. Reduction of the efficiency due to the spin decoherence can be taken into account by adding a term $\text{exp}(-\gamma_s t)$ to the efficiency during the storage time \cite{phillips2008optimal,gorshkov2007freespace}.
Other imperfections such as four-wave mixing \cite{phillips2008optimal, lauk2013fidelity} may further reduce the experimentally achievable efficiency.
\subsection{ATS Memory}
\label{ssec:ATS-memory}
Unlike the EIT protocol, which relies on the adiabatic elimination of the atomic polarization mode, in ATS memory polarization mediates the non-adiabatic coherence exchange between the signal pulse and spin modes. In the broadband regime, ATS is less demanding in terms of technical requirements such as optical depth and control field power compared to the EIT memory. However, in the narrowband regime, the efficiency of the ATS memory is close to zero as the average coherence time of the transitions in resonance with the signal and control fields becomes shorter than the interaction time (i.e., signal pulse duration).
Assuming the system is initially prepared in the state $\ket{g_1}$, we can use dynamically controlled ATS lines produced by a strong
control field that drives the $\ket{g_2}$--$\ket{e}$ transition to absorb
a weak signal pulse in resonance with the $\ket{g_1}$--$\ket{e}$ transition \cite{Saglamyurek2018a}.
Absorption of the signal pulse by ATS peaks (with a peak separation equal to the control Rabi frequency) will then map its coherence to a collective state between $\ket{g_1}$ and $\ket{e}$ (i.e., polarization mode), which subsequently evolves into a collective spin-excitation between the ground states. After the write (i.e., storage) process, we abruptly switch the control field off to trap the coherence, wait for storage time, and turn it on again for the read-out (i.e., retrieval) process. Using the control-field optimization, one can increase the spectral overlap between the ATS peaks separation and the signal bandwidth to maximize the signal pulse absorption. Here, the control-field optimization requires the control pulse area for both write and read stages to be $2\pi$ \cite{rastogi2019discerning}. Using the ATS memory protocol for the $T$ centre, a relatively high efficiency is achievable for broadband light storage and retrieval.
\begin{figure}[]
\includegraphics[width=\linewidth]{Fig6_freespace_memory.pdf}
\caption{Free-space forward $\eta_f$ and backward $\eta_m$ memory efficiency with respect to the ATS factor for $d=27$ in the broadband regime.}
\label{fig:ATS}
\end{figure}
In Ref \cite{saglamyurek2018coherent}, ATS memory efficiency in the free space has been discussed in terms of the ATS factor $F=\Omega/\Gamma$ (or equivalently $F=2\pi B_{\text{sig}}/\Gamma$).
As shown in Fig.\ref{fig:ATS}, for the $T$ centre with $d=27$,
the maximum efficiency of $\eta_{\text{b},\text{max}}=72.6\%$ ($\eta_{\text{f},\text{max}}=46.9\%$) for $F=4$ ($F=7.25$) is achievable in the backward (forward) propagation. For the optical memory, assuming $\Gamma/2\pi=27$ MHz \cite{MacQuarrie2021}, these correspond to the Rabi frequency of $\Omega=2\pi\times108$ and $2\pi\times196$ MHz for the backward and forward modes, respectively.
\section{Microwave memory}\label{sec:mw}
We now turn our attention to $T$ centre prospects for MW quantum memories. In \cref{sec:CPT,sec:ATS} we have measured coherent MW interactions in $T$ centre ensembles from the CPT to AT regimes. MW AT splitting such as we observed is a precursor to the demonstration of a $T$-ATS MW memory. We will treat such an ATS memory theoretically to determine the prospects of MW memories with near term $T$ centre ensembles. For this study of the MW ATS memory we choose one of the two possible MW $\Lambda$ schemes shown in \cref{sec:CPT}, assigning $\ket{1}$ and $\ket{2}$ as ground states $\ket{g_1}$ and $\ket{g_2}$, respectively, and $\ket{3}$ as excited state $\ket{e}$. We begin with an initial state $\ket{g_1}$, and use dynamically controlled ATS lines produced by a control field, as observed in \cref{sec:ATS}, to store a signal field resonant with the MW$_\Uparrow$ transition (see Fig.\ref{fig:1}). So that the coherence of the signal pulse is mapped into the collective nuclear spin excitations of the atoms. The theoretical treatment from Ref. \cite{saglamyurek2018coherent} applies to the ATS MW memory as well as the optical memory considered above. For this case lower MW transition linewidths, as measured in \cref{sec:CPT}, yield lower $\Omega$ for the MW ATS memory.
However in general the weakness of magnetic dipole transitions requires a MW resonator to enhance the MW-ensemble coupling strength. To describe this resonantly enhanced interaction we can use the cavity input-output formalism i.e, $\hat{E}_\text{out}=-\hat{E}_\text{in}+\sqrt{2\kappa}\hat{E}$, and Heisenberg equations of motion \cite{Saglamyurek2018a,gorshkov2007photon}
\begin{equation}
\begin{aligned}
\dot{\hat{P}}&=-\gamma_e \hat{P}+i g(t)\sqrt{N}\hat{E}+\frac{i}{2} \Omega \hat{S},\\
\dot{\hat{S}}&= \frac{i}{2} \Omega^* \hat{P},\\
\dot{\hat{E}}&= i g(t)\sqrt{N}\hat{P}-\kappa\hat{E}+\sqrt{2\kappa}\hat{E}_\text{in},
\end{aligned}
\label{Hem}
\end{equation}
where $\hat{P}$ and $\hat{S}$ are the polarization and spin-wave operators, $\hat{E}$ is the cavity field which we assume can be adiabatically eliminated,
$\kappa$ is the cavity decay rate, $\gamma_e=\Gamma/2$ is the decoherence rate of the polarization mode
and $g(t)$ is the time-dependent light-matter coupling per emitter and $N$ is the number of atoms in the ensemble. Note that here we have ignored the Langevin noise operators (i.e., the incoming noise is vacuum) and once again assumed the spin-wave decoherence rate is negligible as per \cref{sec:memory}.
To estimate the overall efficiency of the memory in the presence of a resonator, we follow the same optimal scheme as in Ref \cite{gorshkov2007photon} for the non-adiabatic or fast limit $\Omega> \Gamma C$, where $C$ is the cooperativity parameter $C=Ng^2/\kappa\gamma_{e}$. The overall efficiency is $\eta=\eta_s\eta_r $, the product of the optimal storage efficiency $\eta_s$ (average ratio of stored excitations to incoming photons) and the optimal retrieval efficiency $\eta_r$ (average ratio of re-emitted photons to stored excitations).
In the optimal regime, we compute the retrieval efficiency as
\begin{equation}
\eta_r=\int_{0}^{t_r}dt\frac{2N}{\kappa}g^2(t)e^{-\int_{0}^{t}dt^\prime 2(Ng^2(t^\prime)/\kappa\,+\,\gamma_{e})},\label{eta_r}
\end{equation}
where $t_r$ is the effective time elapsed during the retrieval process. For simplicity, it is assumed that the retrieval process starts at $t=0$ rather than at some time $t_s>\tau$ where $\tau$ is the duration of the signal, and that $S(0)=1$ and $P(0)=0$.
It has been shown that using the optimal strategy the storage and retrieval efficiencies are analogous \cite{gorshkov2007universal, heshami2012controllable, gorshkov2007photon}. Hence, in the optimal regime, the overall efficiency can be estimated
\begin{equation}
\eta \simeq\frac{C^2}{(1+C)^2} \left(1-e^{-2\gamma_{e}(1+C)t_s}\right)\left(1-e^{-2\gamma_{e}(1+C)t_r}\right).
\label{eq:eff}
\end{equation}
In deriving Eq.\ref{eq:eff}, it is assumed that during the retrieval process all spin mode excitations have been evolved back to the polarization mode (i.e., no excitations are left in the spin mode). For fixed and finite values of the effective times, as the cavity cooperativity increases, the exponentially decaying terms of the Eq.\ref{eq:eff} go to zero, and hence, the first term dominates. Equivalently, in Eq.\ref{eta_r}, if we initially assume $t_r\rightarrow \infty$ such that no excitations are left in the T ensemble i.e., $P(\infty)=0$, the optimal retrieval efficiency reduces to $C/(1+C)$ and the overall efficiency becomes $C^2/(1+C)^2$ (see Fig.\ref{fig:transductionefficency}) \cite{gorshkov2007photon}. With this maximum possible efficiency, one can store photons with $\tau\approx 1/(C\gamma_{e})$ where the polarization decay rate is negligible.
\begin{figure}[]
\includegraphics[width=\linewidth]{Fig7_cavity_memory.pdf}
\caption{Memory efficiency as a function of the resonator cooperativity. The solid line shows the efficiency given by equation (\ref{eq:eff}) assuming $t_s=t_r=1/(2\gamma_{e})$. The dashed line corresponds to the $t_r\rightarrow \infty$ (i.e., $\eta=C^2/(1+C)^2$).}
\label{fig:transductionefficency}
\end{figure}
From this analysis we can see that resonator cooperativities of order $10$ are required to exceed $80$\% memory efficiency with a $T$ centre ensemble. Although we have introduced them in the context of the MW memory scheme, Eqs.\ref{Hem} apply to resonantly enhanced optical memories as well. In the following section we will propose to combine resonantly enhanced MW and optical memories to realize a $T$ centre MW to optical transducer with $T$ centre ensembles.
\section{Microwave-to-optical transduction}
\label{sec:transduction}
Motivated by the optical and microwave memory performance estimates above, we will now look at combining the capabilities to form a $T$ centre ensemble microwave-to-optical transducer that stores MW photons as spin-wave excitations and recalls them optically. For this operation we will consider both resonantly enhanced optical and microwave ATS memories.
Transduction may be achieved by way of two $\Lambda$ schemes, optical and microwave, addressing the same spin coherence. If this can not be achieved, it is necessary to transfer the coherence by an intermediate step. In the following, we elaborate on three possible approaches.
First, we may consider the MW and optical $\Lambda$ schemes set out in \cref{sec:memory} and \cref{sec:mw}. The MW configuration utilizes the long-lived nuclear spin coherence $\rho_{12}$, however, as discussed in \cref{sec:memory}, the hyperfine splitting between $\ket{1}$ and $\ket{2}$ is not optically resolved in $T$ centre ensembles measured to date. Transduction from this microwave interface to an optical field using the same optical $\Lambda$ scheme we utilized in \cref{sec:memory} for an optical AT memory requires one to (i) store MW photons in resonance with the $\ket{1}-\ket{3}$ transition as nuclear spin-wave excitations between the states $\ket{1}$ and $\ket{2}$, (ii) transfer the nuclear spin wave $\rho_{12}$ to an electron spin wave $\rho_{13}$ and (iii) recall the stored pulse using a control field in resonance with the optical transition $B_1$. Transferring coherence between the nuclear and electron spin waves (step (ii)) requires only coherent control of the $T$ centre ground states, which has already been demonstrated over ensembles in Ref.~\cite{Bergeron2020}.
Alternatively, one may consider schemes that avoid step (ii) above. One option is to use the electron spin for MW storage via a different $\Lambda$ configuration i.e. $\ket{1}$ and $\ket{3}$ as the ground states, and $\ket{4}$ as the excited state. This configuration allows us to store MW photons in resonance with the $\ket{1}$--$\ket{4}$ transition while the control field is resonant with the $\ket{3}$--$\ket{4}$ transition. Note that the decoherence rate of the $\ket{1}$--$\ket{3}$ and $\ket{1}$--$\ket{4}$ transitions are comparable. In general this results in ATS line broadening and reduced efficiency \cite{Saglamyurek2018a}. However, in the limit that electron spin coherence time is very long, as it is in the $T$ centre \cite{Bergeron2020} such ATS line broadening is negligible.
Finally, the transfer step is unnecessary when the hyperfine splitting of the nuclear spin wave is optically resolved. For example, in a scheme that utilizes larger hyperfine splittings due to nuclear spins at the carbon sites in $^{13}$C isotopic variants of $T$ \cite{Bergeron2020}, at higher magnetic fields, or by improving the optical linewidth such that the hydrogen hyperfine splitting is resolved. This way one can consider $\ket{1}$ and $\ket{2}$ as the joint ground states, and $\ket{3}$ and $\ket{\downarrow_\mathrm{h}}$ as the excited state of the MW and optical $\Lambda$ systems, respectively.
\subsection{Transduction efficiency and fidelity}
Optimal efficiency of the transducer, where MW and optical transitions are coupled to the respective cavities, can be estimated using the same optimal strategy as for the ATS memory except that there is now an additional mode mismatch factor $\eta_m$ that should be taken into account. As a result, the overall efficiency would be $\eta=\eta_s\eta_r\eta_m$. The mode mismatch factor depends on the design (geometry) of the cavities. In the ideal case, where all atoms are placed in the maximum of both optical and MW fields, the mode mismatch factor can reach unity. In this case and for $t_r\rightarrow \infty$, the optimal efficiency of the transducer is estimated as $C_sC_r/(1+C_s)(1+C_r)$, where $C_s$ and $C_r$ are the cooperativity of the MW and optical cavities, repetitively.
The other metric to quantify the quality of transduction is the overlap between the transduced signal and the transduced noises. We refer to this measure as the transduction fidelity. In order to compute the fidelity, one could solve the system dynamics by taking several important imperfections such as spin decoherence, microwave thermal photons, and inhomogeneous broadening into account to determine the fidelity. As discussed before, inhomogeneous broadening can be dealt with by spectral tailoring. Therefore, here we only consider the thermal photons in the microwave cavity. Spin decoherence is also negligible given that the process can be much faster than the spin coherence times \cite{Bergeron2020}. For this calculation, we consider a transduction system consisting of $\ket{1}$ and $\ket{2}$ as the ground states, and $\ket{3}$ and $\ket{\downarrow_h}$ as the excited state of the microwave and optical $\Lambda$ subsystems, respectively.
The initial microwave thermal occupation $\bar{n}_{\text{th}}=1/(e^{(\hbar\omega_{1, 3}/k_B T)}-1)$ is estimated to be $0.0045$ with $T=20$ mK and $\omega_{1, 3}=2\pi\times 2.25$ GHz \cite{Bergeron2020} which is the splitting between $\ket{1}$ and $\ket{3}$. Since the initial microwave photon occupation is quite small, and it does not affect the quantum system dynamics significantly, we can treat it separately from the signal. The fidelity can be computed using the signal-to-noise ratio (SNR) \cite{asadi2022proposal}:
\begin{equation}
F=\frac{1}{1+\text{SNR}^{-1}}.
\label{eq:fid}
\end{equation}
Given that the thermal noise and signal can be treated independently, it is valid to compute the SNR even before the transduction process starts. In this way, the SNR is given by:
\begin{equation}
\text{SNR}=\frac{n_{\text{sig}}}{n_{\text{th}}},
\end{equation}
where $n_{\text{sig}}$ is the mean occupation of signal in microwave cavity. For a single-photon signal, $n_{\text{sig}}=B_{\text{sig}}/\kappa_s$, where $\kappa_s$ is the microwave cavity decay rate. This is valid in the regime that $B_{\text{sig}}\ll\kappa_s$. Given that $n_{\text{th}}\ll n_{\text{sig}}$, the fidelity shown in Eq.(\ref{eq:fid}) can be simplified as:
\begin{equation}
F\approx 1-\frac{n_{\text{th}}}{n_{\text{sig}}}=1-\frac{n_{\text{th}}\kappa_s}{B_{\text{sig}}}.
\end{equation}
Since we are interested in the case where the transduction efficiency is maximized, $B_{\text{sig}}$ can be expressed as $B_{\text{sig}}=C_s\gamma_{e_s}$. Hence, in the regime of maximum efficiency, the fidelity is explicitly given by:
\begin{equation}
F\approx 1-\frac{n_{\text{th}}\kappa_s}{C_s\gamma_{e_s}}.
\label{eq:ff}
\end{equation}
Clearly, as the cavity cooperativity increases, the fidelity increases, and the efficiency also increases as shown in Fig. \ref{fig:transductionefficency}.
The other way to compute the fidelity is to estimate the SNR after the transduction, which should be equivalent to the approach discussed above. In this way, we can model the microwave thermal photons as dark counts \cite{Ji2022proposalroom}. The dark count rate is given by $D=R\eta$ with the transduction efficiency $\eta$ and the average rate of thermal photons $R$. It describes the effective number of photons emitted as background noise in the optical cavity per unit of time. The probability distribution associated with this dark count rate is given by a Poisson distribution $P(T_{\text{tr}},D)=D^nT^n_{\text{tr}}e^{-DT_{\text{tr}}}/n!$ where $n$ is the number of dark counts and $T_{\text{tr}}$ is the total transduction time. Thus, this time can also be written as $T_{\text{tr}}=t_s+t_r$.
Then, for a single-photon transduction, the SNR takes the following form:
\begin{equation}
\text{SNR}=\frac{\eta n_{\text{sig}}}{\overline{N}_D},
\end{equation}
where $\overline{N}_D$ is the mean number of dark counts. Specifically, at a given time $T_{\text{tr}}$ the SNR becomes:
\begin{equation}
\text{SNR}=\frac{\eta n_{\text{sig}}}{DT_{\text{tr}}},
\end{equation}
where $DT_{\text{tr}}$ is the mean number of dark counts that follow the Poisson distribution. It can be further simplified as $\text{SNR}=n_{\text{sig}}/RT_{\text{tr}}$. In the regime of the maximum efficiency, $RT_{\text{tr}}=n_{\text{th}}$, we recover $\text{SNR}=n_{\text{sig}}/n_{\text{th}}$, which leads to the fidelity shown in Eq.(\ref{eq:ff}).
\section{Discussion} \label{sec:discussion}
The $T$ centre appears to combine several properties appealing for both microwave and quantum memories, and transduction between them, including long lived spins and telecom optical emission. We have considered possible quantum memory and transduction schemes for ensembles of $T$ centres, and performed preliminary measurements with ensembles in \ensuremath{^{28}\text{Si}}{} to establish properties informing memory and transduction device design.
Absorption measurements indicate a resonant optical depth of $0.009$ in a narrow-linewidth ensemble. Cautiously forecasting a small improvement over $T$ centre concentrations that have already been demonstrated in devices, we project that an optical depth of $27$ with a similar sample is possible. This optical depth is sufficient for efficient optical quantum memory by EIT or ATS with estimated a memory efficiency as high as $73\%$.
In addition to the potential for optical memory, the $T$ centre's optical ground state spin structure features $\Lambda$ and V configurations suitable for storing MW fields. We demonstrated these configurations by ODMR spectroscopy, and measured CPT windows in $\Lambda$ and V configurations as a precursor to an EIT memory for MW fields. Fluorescence SNR requirements prohibit us from reaching the low-power limit, but even so we measured CPT linewidths of $17(2)$~kHz, significantly below the ODMR linewidth, and resolved orientation sub-ensembles that were degenerate to MW ODMR precision. We demonstrated microwave ATS up to $55(1)$~kHz splitting. We then discussed a cavity-enhanced MW memory protocol and arrived at an efficiency estimate as a function of cavity cooperativity. In particular, MW memory efficiencies of greater than $80\%$ are possible for $C=10$.
In general, the operation of MW and optical $T$ centre quantum memories can be combined to transduce between the microwave and optical fields. Transduction requires either (i) an optical quantum memory addressing the nuclear spin coherence $\rho_{12}$, (ii) a microwave memory addressing an electron spin coherence such as $\rho_{13}$ via an NMR control field, or (iii) coherent transfer between $\rho_{12}$ and $\rho_{13}$ at an intermediate stage. (iii) requires only the coherent control of $T$ ensembles that has already been experimentally demonstrated \cite{Bergeron2020}. Furthermore, (i) is within reach of schemes utilizing the known $^{13}$C $T$ centre variants.
We proposed a transduction protocol where both microwave and optical transitions are coupled to resonators, and discussed the efficiency and fidelity of the system. Mutually compatible MW and optical cavities with $C = 10$ would yield microwave to optical transduction efficiencies of $83\%$. Necessary future work includes further improving ensemble concentrations, and engineering optical and microwave resonators meeting these metrics, but an in-principle $T$ centre transduction demonstration could be achieved with designs available in the literature.
We have demonstrated the potential of this new spin-photon platform for quantum memory, and identified pathways towards efficient microwave-to-optical transduction with $T$ centres. Further developing and transplanting this approach into on-chip devices, for example with integrated silicon photonic resonators combining moderate mode confinement and high quality factors, would yield an on-chip spin-photon interface for quantum memories and microwave to optical transducers. These devices could network directly between quantum information platforms that share the same technologically and commercially advanced silicon chip platform (including superconducting qubits, quantum dots, trapped ions, and silicon impurity spins) and the telecom optical fibres of a global quantum internet.
\section*{Acknowledgements}
This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC), the New Frontiers in Research Fund (NFRF), the National Research Council (NRC) of Canada through its High-Throughput Secure Networks (HTSN) challenge program, the Canada Research Chairs program (CRC), the Canada Foundation for Innovation (CFI), the B.C. Knowledge Development Fund (BCKDF), and the Canadian Institute for Advanced Research (CIFAR) Quantum Information Science program. D.B.H is supported by the Banting Fellowship program.
The $^{28}$Si samples used in this study were prepared from the Avo28 crystal produced by the International Avogadro Coordination (IAC) Project (2004--2011) in cooperation among the BIPM, the INRIM (Italy), the IRMM (EU), the NMIA (Australia), the NMIJ (Japan), the NPL (UK), and the PTB (Germany). We thank Alex English of Iotron Industries for assistance with electron irradiation.
\section*{APPENDIX A: SAMPLE PREPARATION} \label{sec:Sample}
All measurements are performed on one of two \ensuremath{^{28}\text{Si}}{} crystals, previously described in Ref. \cite{Bergeron2020}. These have been purified to 99.995\% \ensuremath{^{28}\text{Si}}{} isotope proportion by the Avogadro project and then electron irradiated and thermally annealed to produce $T$. In addition to isotopic purity these samples have low chemical impurity, with less than $10^{14}$~oxygen/cm$^3$. $T$ centres are formed from naturally present low-level carbon impurities, which differ between the two samples. The sample used for absorption measurements has $1.5\times10^{15}$~carbon/cm$^3$, the sample used for CPT and ATS measurements has $5\times10^{14}$~carbon/cm$^3$.
\section*{APPENDIX B: METHODS}\label{sec:Methods}
\textit{Cryogenics}
\vspace{2mm}\newline
The samples are loosely mounted in a liquid Helium immersion dewar such that the crystal is not strained. The temperature of the Helium-4 bath is adjusted between $1.4$--$4.2$~K by pressure control.
\vspace{5mm}
\textit{Photoluminescence spectra}
\vspace{2mm}\newline
The sample is illuminated with an above-bandgap $1047$~nm laser and the resulting broadband photoluminescence is directed into a Bruker IFS 125 HR Fourier Transform infrared (FTIR) spectrometer with a CaF$_2$ beam splitter and liquid nitrogen cooled Ge photodetector.
\vspace{5mm}
\textit{Absorption measurement}
\vspace{2mm}\newline
The sample is illuminated along its longest side by a broadband light source with beam size smaller than the sample cross section. Light passes through the sample and two cryostat windows before being directed into the Bruker FTIR spectrometer. The light source spectrum is free of structure in the vicinity of the $T$ centre absorption lines and is removed by subtracting a linear fit about the absorption dip location.
\vspace{5mm}
\textit{Photoluminescence excitation}
\vspace{2mm}\newline
Photoluminescence excitation spectra are recorded by scanning a narrow, tunable laser over the TX$_0$ ZPL and detecting lower energy photons emitted into the TX phonon sideband. Resonant excitation is performed with a Toptica DL100 tunable diode laser. The optical power is amplified up to $100$~mW with a Thorlabs BOA1017P and the beam is expanded to diameter $2$--$4$~mm. Non-resonant excitation light due to spontaneous emission in the laser and amplifier is removed by a combination of an Edmund Optics 87-830 $1350 \pm 12.5$~nm and Iridian Spectral Technologies DWDM $1329\pm0.5$~nm bandpass filters. The two filters are tilted so as to tune them onto the TX$_0$ resonance.
Stray excitation light is removed from the detection path by two Semrock BLP02-1319R-25 $1319$~nm longpass rejection filters, which in fact block the excitation light at TX$_0$ with OD$=4.5$ at normal incidence, but pass light with wavelength longer than $1350$ nm. A $1375$($50$)~nm bandpass filter further selects a detection window in the TX phonon sideband that excludes a silicon Raman replica of the pump laser at $1426$~nm. The fluorescence rate in this detection window is measured by an IDQuantique ID230 avalanche photodiode with 25\% quantum efficiency.
\vspace{5mm}
\textit{Optically Detected Magnetic Resonance}
\vspace{2mm}\newline
Magnetic resonance experiments are performed with the sample in an adjustable magnetic field applied by an iron core electromagnet within the PLE apparatus as described above. The sample is held at the centre of a split-ring resonator with a resonant frequency of $2.25$~GHz and $10$~MHz bandwidth. MW signals from by two SRS SG384 and SG386 signal generators are combined with a ZB2PD-63-S+ power splitter and then amplified up to 1W of power by ZHL-16W-43-S+ and ZHL-1-2W+ amplifiers as required.
\FloatBarrier
|
\subsection*{prediction of distribution of delta states from loss}
TODO: SEAN
\begin{figure*}[tbp]
\centering
\includegraphics[scale=1]{delta_dist_v2.pdf}
\caption{\textbf{$\delta$ state hopping from single photon loss.} a Plot of the $C_{\delta}$ coefficients from eq. TODO for different $delta$ values assuming a pair coherent state amplitude of $\gamma = 1.25$ and loss rates of $(200\mu s)^{-1}$ and $(216\mu s)^{-1}$ for cavity a and b respectively. b) Diagrammatic representation of single photon loss causing hopping to different $\delta$ states where the right (left) pointing red arrows represent a photon loss event from cavity b (a) with the respective rates as indicated on the figure.
}
\end{figure*}
\begin{figure*}[tbp]
\centering
\includegraphics[scale=.45]{dm_reconstruction.pdf}
\caption{\textbf{Density matrix reconstruction} a) Comparison of the $Re(\alpha)$ vs. $Re(\beta)$ 2-D Wigner cut from experiment (left) against the reconstructed density matrix (right). b) Color plot of the reconstructed density matrix with only the elements corresponding to the $delta = -1,0,1$ states.
}
\end{figure*}
\begin{figure*}[tbp]
\centering
\includegraphics[scale=.38]{dm_reconstruction_v3.pdf}
\caption{\textbf{Density matrix reconstruction} a-d) Comparison of the 2-D Wigner cuts from experiment (left) against the reconstructed density matrix (right) for $Re(\alpha) vs. Re(\beta)$, $Re(\alpha) vs. Im(\alpha)$, $Re(\alpha) vs. Im(\beta)$, $Im(\alpha) vs. Im(\beta)$ respectively. b) Color plot of the reconstructed density matrix with only the elements corresponding to the $delta = -1,0,1$ states.
}
\end{figure*}
Given a pair coherent state that is allowed to freely evolve with single photon loss from each cavity while being stabilized by two photon loss, we would like to be able to predict the distribution of the populations of each $\delta$ state by finding the $C_{\delta}$ coefficients in the following equation:
\begin{equation}
\rho(t=\infty) = \sum_{\delta=-\infty}^\infty C_\delta \ket{\gamma, \delta}\bra{\gamma, \delta}
\end{equation}
We will be working in the regime where the two photon loss stabilization is much stronger than the single photon loss ($\kappa_{ab}>>\kappa_{a},\kappa_{b}$) so we can assume that after a single photon loss event happens, creating a change in $\delta$ by 1, the state can be treated as though it is effectively instantaneously stabilized to a pair coherent state of the new $\delta$ while maintaining the state amplitude $\gamma$.
With this assumption in place, we can treat the problem as a simple equilibrium solution to the different delta states and their respective probabilities to jump to neighboring states based off the photon populations in each cavity, as illustrated in Fig.~E1(b). We can represent the photon populations at a given $\gamma$ and $\delta$ ($\Bar{n}_{a}(\gamma ,\delta)$) in cavity $a$ with the following expression ~\cite{agarwal_generation_1986}:
\begin{align}
\Bar{n}_{a}(\gamma ,\delta)=\frac{N_{\gamma,\delta}^{2}}{N_{\gamma,\delta+1}^{2}}|\gamma|^{2}+\delta ,\\\nonumber N_{\gamma,\delta} = ((i|\gamma|)^{-\delta}J_{\delta}(2i|\gamma|))^{-1/2}
\end{align}
Where $J_{\delta}$ is the bessel function of the first kind of order $\delta$. One can then get the photon number in $b$ ($\Bar{n}_{b}(\gamma ,\delta)$) by subtracting the $\delta$ from Eq.~(16). We can now construct the rates of photon loss from cavity $a/b$ with a $\gamma$ and $\delta$ being $\kappa_{i}\Bar{n}_{a/b}(\gamma,\delta)$. Treating each state then with an outgoing rate due to photon loss and an incoming rate from photon loss of neighboring states allows us to create the following equilibrium equation for each $\delta$ value:
\begin{align}
C_{\delta}(\kappa_{a}\Bar{n}_{a}(\gamma ,\delta)+\kappa_{b}\Bar{n}_{b}(\gamma ,\delta))-C_{\delta+1}(\kappa_{a}\Bar{n}_{a}(\gamma ,\delta)) \\\nonumber - C_{\delta-1}(\kappa_{b}\Bar{n}_{b}(\gamma ,\delta)) = 0
\end{align}
Numerically solving this set of coupled equations for all the desired $\delta$ states yields a solution for the distribution of $C_{\delta}$ values with an example distribution plotted out in Fig.~E1(b).
\begin{figure*}[tbp]
\centering
\caption{\textbf{Time domain measurement of the pair coherent state at longer times. a} Showing the population of each state $\delta=0$ state with time and decaying over long times, the best pair coherent state is observed at $\Delta t=????$, \textbf{b} We also measure the population of $\delta=+1,-1$ which is the sum of population of all the states with one more photon in cavity A and cavity B respectively. \textbf{c}Similarly we also measure the $\delta=0, 2$ states over long timescale to find that the population becomes equal after $150\mu s$ suggesting that we do not have a pair coherent state that can live for longer timescales like these.
}
\end{figure*}
\subsection*{math dump}
\begin{equation*}
\xi\approx\sqrt{\Delta_\text{Stark}/2 \alpha_q}
\end{equation*}
\begin{equation}
g_{ab} = \sqrt{\frac{\chi_{qa} \chi_{qb} \chi_{rq}\Delta_\text{Stark}}{\chi_{qq}^2}}
\end{equation}
\begin{equation}
k_{ab} = 4\frac{\chi_{qa} \chi_{qb} \chi_{rq}}{\chi_{qq}^2 \kappa_r}\Delta_\text{Stark}
\end{equation}
\begin{equation}
k_{ab} = 4\frac{\chi_{qa} \chi_{qb} \chi_{rq}}{\chi_{qq}^2 \kappa_r}\Delta_\text{Stark}
\end{equation}
From fit: $\epsilon_{ab} \approx (15.9 kHz) * 2 \pi$, $\kappa_{ab} \approx (12.7 kHz) * 2\pi$ which gives
$|\gamma| = 2 \frac{\epsilon_{ab}}{\kappa_{ab}} = 2.5$
$\epsilon_{ab} = -2ig_{ab} \epsilon_d/\kappa_r$ and pair photon dissipation $\kappa_{ab} = 4 |g_{ab}|^2/\kappa_r$
\begin{equation}
g_{ab} = \sqrt{\kappa_r \kappa_{ab}}/2
\end{equation}
\begin{equation}
|\epsilon_d| = \frac{\epsilon_{ab}\kappa_r}{2\g_ab}= \epsilon_{ab}\sqrt{\kappa_r/\kappa_{ab}}
\end{equation}
\begin{equation}
\Dot{\rho} = -i(\hat{H}_{eff}\rho - \rho \hat{H}_{eff}^\dagger) + \sum_j \kappa_j \hat{L}_j \rho \hat{L}_j^\dagger
\end{equation}
\begin{equation}
\hat{H}_{eff} = \hat{H} - i \sum_J \frac{\kappa_j}{2}\hat{L}_j^\dagger \hat{L}_j
\end{equation}
\begin{equation}
\hat{H}_{eff} = \epsilon_{ab} \hat{a}\hat{b} + \epsilon_{ab}^* \hat{a}^\dagger\hat{b}^\dagger + K_{eff}\hat{a}^\dagger\hat{b}^\dagger \hat{a}\hat{b} - i \frac{\kappa_{ab}}{2}\hat{a}^\dagger\hat{b}^\dagger \hat{a}\hat{b}
\end{equation}
\begin{align}
\Dot{\rho} &= (\epsilon_{ab} \gamma - \epsilon_{ab}^* \gamma^* + i \kappa_{ab} |\gamma|^2)\rho \\\nonumber &+ (\epsilon^* + K_{eff}\gamma - i\frac{\kappa_{ab}}{2} \gamma)\hat{a}^\dagger\hat{b}^\dagger \rho \\\nonumber &- (\epsilon + K_{eff}\gamma^* + i\frac{\kappa_{ab}}{2} \gamma^*) \rho \hat{a}\hat{b}
\end{align}
\begin{equation}
\gamma = \frac{\epsilon^*}{\frac{i\kappa}{2}-K_{eff}}
\end{equation}
\begin{equation}
\rho(t=\infty) = \sum_{\delta=-\infty}^\infty C_\delta \ket{\gamma, \delta}\bra{\gamma, \delta}
\end{equation}
\begin{equation}
\hat{P}_J = \hat{P}_a\hat{P}_b = e^{i\pi \hat{a}^\dagger \hat{a}} e^{i\pi \hat{b}^\dagger \hat{b}}
\end{equation}
\subsection{Density Matrix Reconstruction}
TODO: SEAN
To reconstruct the density matrix from a set of wigner tomography slices with varying displacements on both cavities, we implemented a general N-mode numerical density matrix reconstruction method outlined in~\cite{reinhold_controlling_2019}. We can treat the wigner tomography data as general measurements $M$ where we can express the measurement outcomes $x$ in terms of the density matrix $\rho$ as:
\begin{equation}
x = Tr{(M\rho)}=\sum_{ij} M_{ij}\rho_{ji}\equiv \langle\braket{M}{\rho}\rangle
\end{equation}
Then taking $N_{exp}$ measurements gives us the following relation:
\begin{equation}
\vec{x} = (\sum_{i}^{N_{exp}}\ket{x_{i}}\langle\bra{M_{i}})\ket{\rho}\rangle \equiv \mathscr{M}\ket{\rho}\rangle
\end{equation}
Since $\mathscr{M}$ will not generally be a square matrix, we can take a pseudo inverse to get the closest approximate solution using the Moore-Penrose pseudo-inverse:
\begin{equation}
minimize|\mathscr{M}\ket{\rho}\rangle - \vec{x}| \rightarrow \ket{\rho_{least-sq}}\rangle = (\mathscr{M}\mathscr{M}^{T})^{-1}M^{T}\vec{x}
\end{equation}
To try to constrain the physicality of the reconstruction, we added a lagrange multiplier constraint to the reconstruction that fixes the density matrix on diagonal population values to what was observed in the transmon spectroscopy data plotted in Fig.~2(b):
\begin{equation}
\begin{pmatrix}
\mathscr{M}^{T}\mathscr{M} & C^{T}\\
C & 0
\end{pmatrix}
\begin{pmatrix}
\vec{\rho}\\
\vec{\lambda}
\end{pmatrix}
=
\begin{pmatrix}
\mathscr{M}^{T}\vec{x}\\
\vec{d}
\end{pmatrix}
\end{equation}
Where $C\vec{\rho} = \vec{d}$ represents the constraints. Since the input diagonals are all real and with the trace equaling one, further physicality constraints on the matrix, such as positive definite eigenvalues and $Tr(\rho)=1$, are unnecessary.
The density matrix yielded from this procedure is plotted in a reduced form showing only the $\delta = -1,0,1$ components in Fig.~E2(b), with a side by side wigner cut comparison against the experimental values displayed in Fig.~E2(a).\\
\section{Additional data}
\subsection{Superposition of different delta}
\begin{figure}[tbp]
\centering
\includegraphics[scale=.95]{fig9_09122022_2.pdf}
\caption{Comparing the Wigner function of a superposition of two PCS in the experiment and the simulation.
(a)
Measured Re[$\alpha$]-Re[$\beta$] cut of the joint Wigner function $W(\alpha,\beta)$ for an experimental two-cavity state of the expected form $(\ket{\gamma e^{i\phi}, 0}+e^{i\varphi}\ket{\gamma,1})/\sqrt2$. The cavity rotating frame used for the tomography (phase definition of $\alpha, \beta$) is defined such that $\gamma$ is real.
The Wigner function is measured at $t_{w}=4.5$ $\mu$s after the $t=15$ $\mu$s of stabilizing pump on a $(\ket{00}+\ket{01})/\sqrt{2}$. The cavity free evolution during $t_w$ causes non-zero $\phi$ and $\varphi$.
(b) Simulated Wigner function of the same two-PCS superposition state using the same $t_{w}=4.5$ $\mu$s and the same stabilizing pumping parameters. The cavity frame for tomography is similarly defined for make $\gamma$ real in order to simulate the same condition as in the experiment. The relative phases of the superposition are consistent with the experiment as demonstrated by the similar features in the Wigner function.
}\label{fig:superposition}
\end{figure}
\end{document}
\section{Introduction}
The use of continuous-variable states of bosonic modes as a platform for quantum information processing, originating in quantum optics~\cite{dodonov_nonclassical_2002, braunstein_quantum_2005}, is rapidly advancing in superconducting circuit quantum electrodynamics (QED)~\cite{joshi_quantum_2021, ma_quantum_2021}. While many of the exotic states envisioned decades ago remain challenging to implement in the optical domain, they have become practical and valuable resources in the microwave domain thanks to the ability to engineer a wide range of mode couplings and nonlinearities in Josephson circuits~\cite{blais_circuit_2020}. For example, the Schr\"odinger cat states~\cite{vlastakis_deterministically_2013, wang_schrodinger_2016,grimm_stabilization_2020} and the Gottesman-Kitaev-Preskill grid states~\cite{campagne-ibarcq_quantum_2020} have not only been realized but also actively pursued for encoding logical qubits with error suppression or correction capabilities.
One interesting class of bosonic states yet to be studied experimentally is the pair coherent state (PCS), an example of a Barut-Girardello generalized coherent state~\cite{barut_new_1971}. This state gained theoretical interest as an example of a highly entangled two-mode state~\cite{agarwal_generation_1986, agarwal_Quantitative_2005} and was initially proposed to explain a suppression of amplified spontaneous emission in an atomic system~\cite{malcuit_suppression_1985}. Analogous to the Glauber coherent state, $\ket{\alpha} = \mathcal{N} \sum_{n=0}^{\infty} \frac{\alpha^{n}}{\sqrt{n!}}\ket{n}$, a PCS can be written in the Fock state basis of two harmonic oscillators ($a$ and $b$) as:
\begin{equation}
\ket{\gamma,\delta} = \mathcal{N} \sum_{n=0}^{\infty} \frac{\gamma^{n+\delta/2}}{\sqrt{n!(n+\delta)!}}\ket{n+\delta}_a\ket{n}_b
\label{eq:PC_state}
\end{equation}
where $\delta$ is an integer describing the photon number difference (PND) between the two modes, $\gamma$ a complex number describing the amplitude and phase of the state, and $\mathcal{N}$ a normalization factor. This state is both an eigenstate of the pair photon annihilation operator $\hat{a}\hat{b}$ and the PND operator $\hat{\delta}=\hat{a}^\dagger \hat{a} - \hat{b}^\dagger \hat{b}$:
\begin{equation}
\hat{a}\hat{b} \ket{\gamma,\delta} = \gamma \ket{\gamma,\delta},
\hat{\delta}\ket{\gamma,\delta} = \delta\ket{\gamma,\delta}.
\end{equation}
A PCS is inseparable and already in the form of a Schmidt decomposition~\cite{agarwal_Quantitative_2005}. The $\ket{\gamma, 0}$ state resembles a two-mode squeezed state in terms of photon number correlation~\cite{zhang_entanglement_2018} but has a Poisson-like photon number distribution approximately centered around $\gamma$.
Pair coherent states form the basis of a recently-proposed quantum error correction (QEC) code called the pair cat code~\cite{albert_pair-cat_2019}. This code promises autonomous QEC of all types of first-order physical errors associated with loss \& dephasing by encoding a logical qubit in the superposition of pair coherent states $|\pm\gamma,\delta\rangle$ of two oscillators.
The scheme involves stabilizing PND to correct the quantum jumps of photon loss events while simultaneously stabilizing the pair coherent state manifold with a two-mode four-photon dissipation process.
Despite progress in universal control~\cite{ heeres_implementing_2017, gao_entanglement_2019, chakram_multimode_2022} and measurement feedback~\cite{ofek_extending_2016, campagne-ibarcq_quantum_2020, ma_error-transparent_2020, essig_multiplexed_2021} in bosonic cavities systems, there is a clear need for new tools for multi-mode quantum operations. For example, unitary preparation of a PCS using a dispersively coupled ancilla qubit with the current standard technique, numerical optimal control pulses~\cite{heeres_implementing_2017}, becomes prohibitively difficult for modest photon numbers. On the other hand, reservoir engineering~\cite{poyatos_quantum_1996} has been of particular interest for its ability to stabilize non-classical oscillator states in a resource-efficient manner~\cite{kienzler_quantum_2015}. Moreover, engineered dissipation can provide us with not only encoded qubits with high noise bias~\cite{mirrahimi_dynamically_2014,albert_pair-cat_2019} but also a set of bias-preserving gates for hardware efficient quantum computing~\cite{guillaud_repetition_2019, puri_bias-preserving_2020, chamberland_building_2022, yuan_construction_2022}. In circuit QED, realization of nonlinear dissipation operators have led to stabilization of the cat-state manifolds~\cite{leghtas_confining_2015, touzard_coherent_2018, lescanne_exponential_2020} and autonomous QEC of photon losses~\cite{gertler_protecting_2021} in a single cavity. However, engineered nonlinear dissipation across two cavity modes remains to be explored.
\begin{table*}[tbp]
\caption{Comparison of multi-photon steady states under driven dissipative processes constructed with different operators, $\hat{o}=\hat{a},\hat{a}^2$ and $\hat{a}\hat{b}$. For each of the three cases, we consider cavity
dynamics following the master equation $\Dot{\rho}=-\frac{i}{\hbar}[\hat{H_0}+\hat{H_d},\rho]+\mathcal{D}(\rho)$ written in the rotating frame of the drives, where $\mathcal{D}$ is the Lindblad superoperator with $\hat{o}$ as its jump operator.
In the textbook example of the driven damped quantum oscillator, the mode detuning $\Delta$ and single-photon loss $\kappa$ jointly counter the driving force and determine the complex amplitude of the unique steady state of the system.
In two-photon dynamics, the complex amplitude of the steady-states is analogously determined by the one-mode or two-mode squeezing drives countered by the corresponding two-photon loss rates and confining Kerr Hamiltonian~\cite{gautier_combined_2022}. Our experiment produces PCS under both dissipative and cross-Kerr Hamiltonian confinement, with the effective cross-Kerr $K_\mathrm{eff}$ a few times stronger than the pair dissipation $\kappa_{ab}$. Notably, two steady states exist (with even or odd photon number parity $\Pi$) for the case of cat-state stabilization while there are infinitely many steady states (with different photon number difference $\delta$) for the case of stabilizing pair coherent states.
}
\centering
\begin{tabular}{c c c c c}
\hline\hline\\[-2ex]
Process category & Drive Hamiltonian $\hat{H}_d/\hbar$ & \,\,\, Dissipator $\mathcal{D}$ \,\,\, & Hamiltonian $\hat{H_0}/\hbar$ & Steady State(s) \\[0.5ex]
\hline\\[-1.5ex]
Single-photon & $\epsilon^*\hat{a}+\epsilon\hat{a}^\dagger$ & $\kappa\mathcal{D}[\hat{a}]$ & $\Delta \hat{a}^\dagger\hat{a}$ & Coherent state $|\alpha\rangle$, $\alpha = \frac{\epsilon}{i\kappa/2-\Delta}$ \\[0.8ex]
Single-mode two-photon & $\epsilon_2^*\hat{a}^2+\epsilon_2\hat{a}^{\dagger 2}$ ~\cite{slusher_observation_1985} & $\kappa_2\mathcal{D}[\hat{a}^2]$ \cite{leghtas_confining_2015} & $K_{a} \hat{a}^{\dagger2}\hat{a}^2$ ~\cite{grimm_stabilization_2020} & Cat states $|\alpha,\Pi=\pm1\rangle$, $\alpha = \sqrt{\frac{\epsilon_2}{i\kappa_2/2-K_{a}}}$\\[0.8ex]
Two-mode pair-photon & $\epsilon_{ab}^*\hat{a}\hat{b}+\epsilon_{ab}\hat{a}^\dagger\hat{b}^\dagger$ ~\cite{heidmann_observation_1987} & $\kappa_{ab}\mathcal{D}[\hat{a}\hat{b}]$ & $K_\text{eff} \hat{a}^\dagger\hat{a}\hat{b}^\dagger\hat{b}$ & Pair coherent states $|\gamma, \delta\in\mathcal{Z}\rangle$, $\gamma = \frac{\epsilon_{ab}}{i\kappa_{ab}/2-K_\text{eff}}$\\[1ex]
\hline
\end{tabular}
\label{table:PC_principle}
\end{table*}
Similarly, characterization of multi-mode bosonic states poses substantial challenges beyond their single-mode counterparts. Although multi-mode Wigner tomography using joint parity or joint photon number measurements has been previously demonstrated~\cite{wang_schrodinger_2016, chakram_multimode_2022, wang_deterministic_2011, wollack_quantum_2022},
such process shares similar scalability challenges as in multi-qubit state tomography and additionally lacks the convenience to arrange orthogonal measurement basis for efficient information extraction.
In order to explore the space of multi-mode bosonic QEC codes such as the pair-cat code or the two-mode binomial~\cite{chuang_bosonic_1997,michael_new_2016} and GKP codes~\cite{royer_encoding_2022}, it is crucial to develop efficient tools to characterize the relevant metrics of the states.
In this work, we present an experimental realization of the pair coherent state and efficient characterization of its coherence.
We expand the toolbox of quantum reservoir engineering by realizing an effective dissipation operator
that removes photon pairs from two superconducting cavities and stabilizes the complex amplitude of the pair coherent state. This pair-photon dissipation is realized using a pumped superconducting transmon ancilla for nonlinearity and a short-lived oscillator mode as the reservoir.
To characterize the two-cavity state, we use three levels of the transmon ancilla to isolate selected subspaces of the large Hilbert space before Ramsey-style tomographic measurements. We use this subspace tomography technique to independently measure the quantum coherence between individual pairs of Fock components
of the two-cavity state. Our characterization reveals the phase distortion and limited coherence of the stabilized PC state, which we attribute to the spurious cavity-reservoir cross-Kerr interactions. We further demonstrate an effective method of measuring PND without fine matching of system parameters, which may be used for discrete or continuous tracking of error syndromes in future implementation of the pair cat code.
This Article is organized as the following: In Section II we
discuss a model of the pair photon driven dissipation process in our circuit QED system while introducing our experimental setup. In Section III we present experimental characterization of our PCS stabilization process. This includes measurements of the pair photon population dynamics, direct PND measurements, a demonstration of manifold stabilization, and join Wigner tomography. In Section IV we introduce and implement the subspace tomography that leads to quantitative understanding of the non-ideality of the stabilized PCS. We conclude in Section V with a brief summary and outlook.
\section{PCS stabilization scheme in circuit QED}
Our approach to generate and stabilize a pair coherent state
is based on the application of a pair photon drive counter-balanced
by engineered pair photon dissipation
and a cross Kerr interaction. To understand how the PCS naturally emerges as the steady state under their combined effect, the two-mode system dynamics can be compared or mapped to: 1) the textbook example of the stabilized Glauber coherent state of a driven damped oscillator, and 2) previous demonstrations of stabilized single-mode cat-state manifolds~\cite{leghtas_confining_2015, grimm_stabilization_2020}. The correspondence of relevant Hamiltonian and dissipator terms are listed in Table~\ref{table:PC_principle}.
While the pair-photon drive $\hat{H}_d=\epsilon_{ab}\hat{a}^\dagger\hat{b}^\dagger+c.c.$, also known as the two-mode squeezing drive~\cite{heidmann_observation_1987}, has long been a workhorse in quantum optics, stabilization of the PCS requires either a strong (non-unitary) pair photon loss mechanism or a strong (unitary) cross-Kerr interaction (strong relative to the single-photon decay rates). These two possible strategies are analogous to the stabilized ``dissipative cat"~\cite{leghtas_confining_2015,lescanne_exponential_2020} and ``Kerr cat"~\cite{puri_engineering_2017,grimm_stabilization_2020}, respectively, in the single-mode two-photon processes. In fact, the dissipative and unitary effects play the roles of real and imaginary components of the restoring force and are mutually compatible (see Table~\ref{table:PC_principle}). Inspired by both types of cat-state stabilization, our system combines both effects while generalizing this coherent stabilization to the two mode scenario.
The 3d circuit QED architecture is ideal for realizing this driven dissipation process due to the availability of strong coupling between modes, the four-wave-mixing capability of the Josephson junction, and the wide range of mode lifetimes achievable in the same system.
\begin{figure}[tbp]
\centering
\includegraphics[scale=1]{fig1_08202022.pdf
\caption{Pair coherent state generation: system and protocol. (a) Cartoon of the 3D cQED system containing two high-Q cylindrical post cavities $a$ and $b$ (blue and orange), the transmon ancilla $q$ (magenta), and the stripline low-Q reservoir $r$ (green). (b) Mixing drive $p$, with frequency $\omega_p \approx \omega_a + \omega_b - \omega_r$, coherently converts reservoir excitations with pairs of $a$ and $b$ excitations mediated by the ancilla junction.
The same ancilla junction also converts the reservoir decay (dotted green arrow) into an effective pair photon dissipative process (dotted blue/orange arrows). After adiabatic elimination of the reservoir, the storage cavity dynamics are shown in the dotted box. (c) Cartoon of the mode frequencies and linewidths (not to scale). Strong off-resonance CW mixing drive $p$ and weak on-resonance ($\omega_d\approx\omega_r$) reservoir drive $d$ are shown as vertical arrows. (d) Schematic diagram of $\delta=0$ states under pair photon drive (double-headed arrow) and pair dissipation (dotted arrow). Fock states are written as $\ket{n_a n_b}$ and vertical bars represent a PCS distribution.
}
\label{fig:PC_1}
\end{figure}
As shown in Fig.~\ref{fig:PC_1}(a), our system contains two cylindrical post cavity modes $a$ and $b$ (with single photon loss rates $\kappa_a/2\pi=0.30$ kHz and $\kappa_b/2\pi=0.74$ kHz), a stripline resonator $r$ (with decay rate $\kappa_{r}/2\pi = 0.78$ MHz) used for readout and as a Markovian reservoir, and an ancilla transmon $q$.
The device architecture is similar to that in Ref.~\cite{wang_schrodinger_2016} except that the two cavity posts share the same elliptical cavity body to allow strong transmon-cavity couplings with a relatively small transmon antenna (see Appendix \ref{Setup}). The cavity/resonator modes have annihilation operators $\hat{a}$, $\hat{b}$, $\hat{r}$.
The leading order terms of the static system Hamiltonian in the rotating frame are:
\begin{align}
\hat{H_0} = \hat{H}_{\textrm{disp}}+\hat{H}_{\textrm{sk}}+\hat{H}_\textrm{rk}
\label{eq:H0}
\end{align}
where
\begin{align}
\hat{H}_\textrm{disp}/\hbar =
- \sum_{m=a,b,r}
\big(\chi_{m}\ket{e}\bra{e}+\chi^{f}_{m}\ket{f}\bra{f}\big) \hat{m}^\dagger\hat{m}
\end{align}
are the dispersive interaction terms between the lowest three transmon levels ($\ket{g}, \ket{e}, \ket{f}$) and the cavity/resonator modes, which we use for characterization of the cavity bosonic states,
\begin{equation}
\hat{H}_\textrm{sk}/\hbar = - K_{ab} \hat{a}^\dagger\hat{a} \hat{b}^\dagger\hat{b} - \frac{K_{aa}}{2} (\hat{a}^\dagger\hat{a})^2 - \frac{K_{bb}}{2} (\hat{b}^\dagger\hat{b})^2
\label{eq:H_SK}
\end{equation}
are the cross-Kerr and self-Kerr nonlinearities of the storage cavities which contributes to PCS stabilization, and
\begin{equation}
\hat{H}_\textrm{rk}/\hbar = - (K_{ar} \hat{a}^\dagger\hat{a} + K_{br} \hat{b}^\dagger\hat{b}) \hat{r}^\dagger\hat{r} - \frac{K_{rr}}{2} (\hat{r}^\dagger\hat{r})^2
\label{eq:H_RK}
\end{equation}
are the Kerr terms involving the reservoir mode, which are spurious nonlinearities in this experiment. The device is measured in a dilution refrigerator at a nominal base temperature of 20 mK. All device parameters are listed in the Appendix Table II, with the general rates hierarchy of $\chi_{m} \gg \kappa_r \gg K_{mn} \gg \kappa_{a,b}$ (where $m,n=a,b,r$).
\begin{figure*}[tbp]
\centering
\includegraphics[scale=0.8]{fig2_09172022.pdf}
\caption{Pair coherent state characterization: population and time dynamics. (a) Pulse sequence using selective ancilla rotation to analyze photon population dynamics. We use an ancilla $\pi$-pulse with a frequency as shown to measure the population of a two-cavity state $\ket{mn}$ after waiting for 1 $\mu$s for the reservoir to relax. (b) Populations of two-cavity states $\ket{00}$ through $\ket{44}$ after $t=15$ $\mu$s of pumping at a different mixing pump frequency (while keeping the reservoir drive constant). The rest of the experiments are carried out under the condition shown as $\Delta\omega_\mathrm{fwm}=0$ here.
(c) Transmon spectroscopy after 15 $\mu$s of pumping with cw mixing and reservoir tones. Dark vertical bars correspond to $\delta=0$ states while light bars correspond to error states caused by single photon loss. (d) Time-domain photon population measurements performed starting in vacuum and pumping with both cw tones for a variable time $t$.
The inset shows a time-domain measurement of the pair photon decay (without pair photon drives) by keeping only the mixing pump on for a variable amount of time after the two-tone pumping for $t=15$ $\mu$s (vertical dashed line).
Dashed curves are numerical fits to the pair photon dynamics model Eq.~(\ref{eq:rho_full}) plus single-photon losses, including only two free parameters: $\epsilon_{ab}/2\pi = 99$ kHz and $\kappa_{ab} = (12.7$ $\mu s)^{-1}$.
(e) Measurement of the photon number difference $\delta$, where the probability of $\delta=\pm1$ is measured over variable time $t$. The dashed lines are simulation results with the same parameters as in (d).
}
\label{fig:PC_2}
\end{figure*}
To implement the pair photon excitation and dissipation, we apply two stabilization drives to our system: a strong off-resonance pump
that coherently converts reservoir photons with pairs of photons in $a$ and $b$, and a weaker drive approximately on resonance with the reservoir (Fig.~\ref{fig:PC_1}b,c).
Under these two drives, the Hamiltonian gains an interaction term under the rotating wave approximation:
\begin{equation}
\hat{H}_{int}/ \hbar = g_{ab} \hat{a}^\dagger \hat{b}^\dagger \hat{r} + \epsilon_d \hat{r}^\dagger + h.c.
\label{eq:H_int}
\end{equation}
where $\epsilon_d$ is the rate of the reservoir drive, and $g_{ab}$ is the four-wave mixing rate activated by the off-resonance pump. If we assume the transmon remains in the ground state and then adiabatically eliminate the reservoir due to its fast relative dynamics, we obtain our desired form of the Lindblad master equation for the reduced density matrix $\rho$ of the two storage cavities:
\begin{align}
\frac{\partial \rho}{\partial t} =
-&\frac{i}{\hbar} \big[ (\epsilon_{ab} \hat{a}^\dagger\hat{b}^\dagger
+ K_\text{eff} \hat{a}^\dagger\hat{a} \hat{b}^\dagger\hat{b} + h.c.), \rho\big]\nonumber\\
&+\mathcal{D} \big[\sqrt{\kappa_{ab}} \hat{a}\hat{b}+\zeta_a\hat{a}^\dagger\hat{a}+\zeta_b\hat{b}^\dagger\hat{b}\big](\rho)
\label{eq:rho_full}
\end{align}
where we have combined the storage Kerr terms ($\hat{H}_\textrm{sk}$) as a single $K_\text{eff}$ term (valid in the limit of $\gamma \gg \delta$). $\mathcal{D}[\hat{o}](\rho)$ is the Lindblad superoperator with a composite jump operator $\hat{o}=\sqrt{\kappa_{ab}} \hat{a}\hat{b}+\zeta_a\hat{a}^\dagger\hat{a}+\zeta_b\hat{b}^\dagger\hat{b}$, where $\zeta_a$ and $\zeta_b$ are complex coefficients related to $\hat{H}_\mathrm{rk}$. In the limit of $g_{ab}\gg K_{ar},K_{br}$, we have $\zeta_a, \zeta_b=0$, and this jump operator is reduced to the desirable form of two-photon loss, $\hat{o}\propto\hat{a}\hat{b}$, and the system is stabilized to the PCS manifold as described in Table~\ref{table:PC_principle}.
Unlike previous analyses of two-photon driven dissipation~\cite{leghtas_confining_2015, lescanne_exponential_2020}, we emphasize that Eq.~(\ref{eq:rho_full}) is valid even when the reservoir mode is driven far away from vacuum (\textit{i.e.}~displaced to a large coherent state when $\epsilon_d > \kappa_r$) as long as the adiabatic condition, $\kappa_r\gg g_{ab}, K_{mn}$, is satisfied.
The effective pair photon drive $\epsilon_{ab}$ and pair photon loss rate $\kappa_{ab}$ both increase with the four-wave mixing rate, $\epsilon_{ab} = -2ig_{ab} \epsilon_d/\kappa_r$, $\kappa_{ab} = 4 |g_{ab}|^2/\kappa_r$, while only the pair photon drive increases with the reservoir drive. Initially unexpectedly, the reservoir nonlinearity, $\hat{H}_\textrm{rk}$, enters as dephasing-like modifications added to the pair-photon loss operator $\sqrt{\kappa_{ab}}\hat{a}\hat{b}$ in Eq.~(\ref{eq:rho_full}), which will contribute significantly to the experimental outcome. See Appendix \ref{Derivation} for analytical derivation of the system dynamics.
\section{General characterization of pair photon stabilization}
\subsection{Pair Photon Population Dynamics}
To characterize the pair photon driven dissipative dynamics, we first measure the two-cavity photon number distributions for any prepared state. To do this we perform spectroscopy of the ancilla transmon whose frequency is shifted by $-\chi_{a,b}$ for every photon in cavity $a$, $b$. A frequency-selective rotation of the ancilla at a detuning $\Delta\omega_q = -n_a\chi_a-n_b\chi_b$ maps the probability of being in Fock state $\ket{n_a n_b}$ to the ancilla excitation, which can then be read out. Here, we have written the two-cavity Fock state $\ket{n}_a\otimes\ket{m}_b$ as $\ket{nm}$ in short, a convention that will be used for the rest of the Article. To find an optimal condition to create a PCS in the presence of ac-Stark shift, we sweep the frequency of the four-wave mixing pump while measuring populations of various $\delta=0$ states (Fig.~\ref{fig:PC_2}b).
After 15 $\mu$s of pumping with calibrated drive rates of $g_{ab}/2\pi=60$ kHz and $\epsilon_{d}/2\pi=780$ kHz, we observe the full spectroscopy of the ancilla (Fig.~\ref{fig:PC_2}c) which illustrates the Poisson-like distribution of a PCS. We can also track these photon populations over time to understand how the system converges to a quasi-steady state with a pair photon drive and pair photon dissipation (Fig.~\ref{fig:PC_2}d) or how it decays under the pair photon dissipation alone (inset of Fig.~\ref{fig:PC_2}d).
Fitting these time-domain data of pair photon population dynamics to numerical simulations of Eq.~(\ref{eq:rho_full}) (Appendix \ref{Simulations}), we extract the pair photon dissipation rate $\kappa_{ab}/2\pi=12.5$ kHz (from Fig.~\ref{fig:PC_2}d inset) and subsequently the pair photon drive rate $\epsilon_{ab}/2\pi=99$ kHz (from Fig.~\ref{fig:PC_2}d), in good agreement with values expected from adiabatic elimination. Together with the effective cross-Kerr of the system ($K_\textrm{eff}/2\pi=-86$ kHz), these rates are significantly faster than the undesirable single photon loss rates $\kappa_a$ and $\kappa_b$. Therefore, we experimentally create a quasi-steady state resembling a PCS with $|\gamma|=2.3$ and $\delta=0$ under these driven dissipation conditions before single photon loss eventually decoheres the state by altering $\delta$.
We note that different combinations of $n_a$ and $n_b$ can in principle result in similar dispersive shift to the ancilla, causing ambiguity in our measurement of cavity photon population. However, for our system with $3\chi_{qa} \approx \chi_{qb}$, this ambiguity arises only when the underlying Fock state $\ket{n_a n_b}$ deviates from its expected $\delta$ by at least 4, which is highly unlikely to occur over time scales shorter than the single photon losses.
A prominent feature of the population dynamics in Fig.~\ref{fig:PC_2}(d) is oscillations between states which are then damped to a steady state. The oscillations arise from the under-damped Kerr dynamics of the storage cavities, $\kappa_{ab}<|K_\textrm{eff}|$, and are due to the sudden turn-on of the stabilization drives not allowing the storage cavities to adiabatically evolve in the ground state of the Kerr Hamiltonian as in Ref.~\cite{grimm_stabilization_2020}. Nevertheless, the pair photon dissipation plays a crucial role in relaxing the system to the steady state, allowing us to by-pass the otherwise slow ramping of the pump tones required for adiabatic state preparation. This hybrid implementation of dissipation and Hamiltonian stabilization is analogous to recent proposals of combining Kerr cat qubits with engineered dissipation to further improve robustness against unwanted excitations~\cite{gautier_combined_2022, guillaud_repetition_2019, putterman_stabilizing_2022}. The spurious dephasing-like contributions in the jump operator in Eq.~(\ref{eq:rho_full}) also contributes to the convergence to a steady state, which, however, is accompanied with loss of coherence as will be discussed later.
\subsection{Measurement of Photon Number Difference}
Single photon loss is a fundamental decoherence channel in superconducting cavities which is not stabilized by the pair-photon driven dissipation. Quantum non-demolition measurement of photon number difference (PND) followed by autonomous or digital feedback is a crucial component of the pair-cat QEC code~\cite{albert_pair-cat_2019}.
An existing proposal of PND measurement requires an exact negative $\chi$-matching condition, $\chi_{a} = -\chi_{b}$,
which allows direct mapping of the probability of the two-cavity state in a targeted $\delta$ onto the ancilla excitations via a simple selective rotation~\cite{albert_pair-cat_2019}.
For a transmon ancilla, this negative $\chi$ matching condition does not occur naturally but can be achieved with additional strong off-resonant pumps~\cite{rosenblum_cnot_2018}. It should be noted that managing multiple strong off-resonant pumps in cQED devices without incurring instabilities remains an active challenge~\cite{lescanne_escape_2019}.
Here we demonstrate an alternative method of PND measurement without $\chi$ matching: We apply a comb of number-selective $\pi$-pulses to excite the ancilla at frequencies corresponding to the dispersive shifts for all cavity states (below a reasonable truncation) with a targeted $\delta$.
Subsequent readout of the ancilla informs whether the PND of the two-cavity state is equal to $\delta$ or not.
For instance, we could superimpose pulses at $\omega_q-n\chi_{a}-(n+1)\chi_{b}$, where $0\leq n \leq 5$, to inquire whether the system is in $\delta=1$. In Fig.~\ref{fig:PC_2}(c) we use this measurement to probe the populations in the $\delta=\pm 1$ states, which shows the effect of single photon loss. While our experimental setup does not have a quantum-limited amplifier to perform the necessary demonstration, this technique of PND measurement is in principle quantum non-demolition as long as the phases of the $\pi$-pulses are tuned to be effectively equal after compensating for any ac Stark shift and Kerr effects. It can then be used to repetitively monitor single photon loss from a state with known $\delta$
so long as $\ket{00}$, $\ket{10}$, and $\ket{01}$ are number resolved and a recovery operation can be applied before the next photon loss occurs. Therefore, it provides a practical avenue for QEC of single photon loss in the pair cat code, hence relaxing one of its demanding requirements for implementation. The comb-based PND measurement may also be converted to an autonomous QEC protocol to correct photon loss errors~\cite{gertler_protecting_2021}.
\begin{figure}[t]
\centering
\includegraphics[scale=1.01]{fig3_08202022.pdf}
\caption{The manifold of stabilized pair coherent states. (a) Conceptual representation of the different $\delta$ subspaces of the two-cavity system under pair photon driven dissipation. Each subspace has a unique steady state $\ket{\gamma_\infty, \delta}$, which functions as an attractor in a generalized phase plane. In general, the system is stabilized to the manifold of states represented by the vertical string $\gamma=\gamma_\infty$.
(b, c) Ancilla spectroscopy that demonstrates the photon number distribution of the stabilized (b) $\delta=1$ and (c) $\delta=-1$ PCS. These measurements are performed by preparing initial states of $\ket{10}$ and $\ket{01}$ respectively and then applying the exact same stabilization drives as used in Fig.~\ref{fig:PC_2}(c) for $t=15$ $\mu$s.
For each plot, dark vertical lines indicate the desired states while light vertical lines mark error states due to single photon loss or imperfect initial state preparation. }
\label{fig:PC_3}
\end{figure}
\subsection{Manifold Stabilization}
An important property of the pair photon driven dissipation process is conservation of the photon number difference $\delta$:
it stabilizes $\gamma$ while allowing $\delta$ to be the one and only degree of freedom inherited from arbitrary initial values and may allow quantum operations in the presence of stabilizing pumps. For example, the photon number distribution in Fig.~\ref{fig:PC_2} corresponds to the unique steady state of the two-cavity system within the $\delta=0$ subspace, and its $\delta$ is inherited from the initial vacuum state before the pair-photon pumping is applied. To create a PCS with $\delta=1$ or $-1$, we can use a Selective Number-dependent Arbitrary Phase (SNAP) gate~\cite{heeres_cavity_2015} to prepare a $\ket{10}$ or $\ket{01}$ initial state before applying the same pumping conditions. The resultant photon number distributions are shown in Fig.~\ref{fig:PC_3}, corresponding to PCS of $\ket{\delta =1, |\gamma|\approx2.4}$ and $\ket{\delta =-1, |\gamma|\approx2.2}$ respectively.
More generally, the pair photon dynamics of Eq.~(\ref{eq:rho_full}), in the ideal limit of $\zeta_a,\zeta_b=0$, confines the two-cavity state to a quantum manifold (Hilbert subspace) spanned by a series of PCS with a fixed $\gamma$ and an arbitrary integer $\delta$. Any coherent superpositions of these PCS are steady states allowed by the driven dissipation although the coherence is protected against single photon loss, which shift the value of $\delta$ between neighboring integers.
\begin{figure*}[tbp]
\centering
\includegraphics[scale=.78]{fig4_09152022.pdf}
\caption{Joint Wigner tomography of pair coherent states. (a) The pulse sequence for Wigner function measurements. The two cw tones are applied for $t=15$ $\mu$s to prepare and stabilize a state that is then evolved under no pump tones for $t_w=4.5$ $\mu$s to reach the first phase revival of an approximate PCS under cavity Kerr dynamics (and for the reservoir to relax to vacuum). The joint parity mapping uses a Ramsey sequence with evolution time of $\pi/\chi_{ef}$, where $\chi_{ef}\approx\chi_a^f-\chi_a\approx\chi_b^f-\chi_b$. (b) Wigner function, $\frac{\pi^2}{4}W(\alpha, \beta)$, of the stabilized $\delta=0$ state with constant displacement amplitudes $|\alpha| = |\beta| = 0.3$ and sweeping over the angle of both displacements.
(c) Measured Re$[\alpha]$-Re$[\beta]$ cut of the joint Wigner function for the approximate $\ket{\gamma=2.3, \delta=0}$ PCS (with the real axes redefined such that $\arg[\gamma]=0$). This can be compared to the same Wigner function cuts of (d) the ideal PCS $\ket{2.3, 0}$ and (e) the experimentally measured density matrix block [as shown in (j)] of this stabilized state.
(f), (h) Measured Re$[\alpha]$-Re$[\beta]$ cut of the joint Wigner function for the approximate $\ket{\gamma=2.4, \delta=1}$ and $\ket{\gamma=2.2, \delta=-1}$ PCS stabilized under the same condition, to be compared to (g), (i) the corresponding ideal PCS respectively. (j) The $\delta=0$ sector of the density matrix of the stabilized $\delta=0$ state at the same effective free evolution time, measured with quantum subspace tomography (\textit{Note:} independent from the Wigner measurements) as discussed in Section IV. The blue bars are the diagonal elements indicating population distribution whereas the off-diagonal coherence elements are shown by orange (green) for nearest (next-nearest) neighboring Fock-state pairs. Other coherence elements are small and hence not measured in the experiment. Phases of the coherence elements are shown numerically over the bars.
}\label{fig:PC_4}
\end{figure*}
\subsection{Joint Wigner Measurements}
To demonstrate the non-classicality and coherence of these stabilized states, we measure the joint Wigner function of the two-cavity state $\rho$~\cite{cahill_density_1969}:
\begin{equation}
W(\alpha,\beta) = \frac{4}{\pi^2}\textrm{Tr}[\rho \hat{D}_a(\alpha)\hat{D}_b(\beta)\hat{P}_J\hat{D}^\dagger_b(\beta)\hat{D}^\dagger_a(\alpha)]
\end{equation}
where $\hat{P}_{J} = e^{\pi i (\hat{a}^\dagger\hat{a}+\hat{b}^\dagger\hat{b})}$ is the joint photon number parity operator, and $\hat{D}_a(\alpha),\hat{D}_b(\beta)$ are phase-space displacement operators of the two cavities respectively. We take advantage of our approximately matched dispersive shift of the $\ket{e}-\ket{f}$ transmon transition: $\chi^f_a-\chi_a=\chi^f_b-\chi_b$ to measure the joint photon number parity, whose expectation value following cavity displacements in the 4D phase space ($\alpha,\beta$) can be directly scaled to the joint Wigner function~\cite{wang_schrodinger_2016, milman_proposal_2005}.
In Fig.~\ref{fig:PC_4}(b), with a 2d angular cut of the joint Wigner function at fixed displacement amplitudes, we first demonstrate an interesting property of the PCS (or any eigenstates of PND in general): invariance with respect to the differential cavity phase: $W(\alpha,\beta)=W(\alpha e^{i\phi},\beta e^{-i\phi})$. This is because the product of cavity rotation operators $\hat{R}_a(\phi)\hat{R}_b(-\phi)=e^{i\phi\hat{a}^\dagger\hat{a}}e^{-i\phi\hat{b}^\dagger\hat{b}}=e^{i\phi\hat{\delta}}\equiv\mathbb{I}$ for a PCS up to a global phase. On the other hand, oscillation with respect to the total phase is a signature of phase coherence of the two-cavity state. Choosing the common phase $\frac{1}{2}(\arg[\alpha]+\arg[\beta])$ where the Wigner function takes the maximum value, we show measured 2d phase-space cuts of the joint Wigner function of the experimental $\delta=0$, +1, and -1 PCS to be compared to the closest-matched ideal PCS (Fig.~\ref{fig:PC_4}(d-i)). These Wigner data are acquired after $t=15$ $\mu$s of stabilizing drives and an additional $t_w=4.5$ $\mu$s of wait time. We note that $t_w\gtrsim 5/\kappa_r=1$ $\mu$s is necessary to allow the reservoir photons to fully decay, and the wait time here is chosen to reach the first phase revival of an approximate PCS after cavity Kerr dynamics~\cite{kirchmair_observation_2013}. The characteristic interference fringes of the PCS are clearly visible, indicating appreciable coherence and consistent multi-photon phases. The even-parity ($\delta=0$) and odd-parity ($\delta=\pm1$) states show opposite Wigner contrasts as expected, and the differences in the slope of the fringes are striking features indicating the different PND of the states.
The technique of direct joint Wigner tomography in principle allows full reconstruction of the two-cavity density matrix but faces a multitude of practical challenges.
Firstly, tomographic reconstruction requires a very large amount of data. Even if we impose a low cutoff of 4 photons per cavity, the tomography still requires accurate measurements at a minimum of 625 (typically many more) different phase-space sampling points.
Secondly, the joint parity measurement contains some photon-number-dependent systematic errors caused by the storage-readout cross-Kerr, the 6th-order dispersive shift of the transmon, and non-perfect $\chi$-matching for joint parity extraction. Last but not least, even with sufficient brute force, it is challenging to prevent small stochastic and systematic errors from propagating badly in the matrix pseudo-inversion problem of reconstruction due to the presence of singularities. In our experiment, primarily to limit spurious readout signals due to a large $K_{br}$, we have limited our cavity displacements to $|\alpha|, |\beta|<1$ in Wigner function measurements as in Fig.~4, which contains nearly all the salient features of the PCS. However, a full Wigner reconstruction would require cumulative measurements of highly diluted features over a much larger extent of the 4D phase space.
\begin{figure*}[tbp]
\centering
\includegraphics[scale=.35]{Fig_6_v2.pdf}
\caption{Subspace tomography protocol. (a) Example schematic for tomography in subspace $S=\textrm{span}\{\ket{1,1},\ket{2,2}\}$. Number-selective transmon $\pi$-pulses (2 $\mu$s-long Gaussian) are applied to isolate the two Fock components in the $\ket{e}$ manifold. Cavity displacement pulses are then applied to induce interference, followed by mapping an observable $\hat{M}$ (e.g. the joint parity) to $\ket{f}$.
(b) Spectroscopy of the $\ket{e}-\ket{f}$ transition of the transmon after the cavity displacements $\hat{D}_a(|\alpha|e^{i\phi}) \hat{D}_b(|\beta| e^{i\phi})$
in (a) with two different displacement phases $\phi = \phi_0$ and $\phi_0+\pi/2$ where $\phi_0$ is chosen to minimize the $\hat{M}=\hat{P}_{11}$ measurement, thus $\phi_0+\pi/2$ will maximize it. The difference in photon number distribution population in the two cases demonstrates the presence of interference. The $\ket{21}$ and $\ket{12}$ peaks overlap significantly due to the approximately matched dispersive shift of the $\ket{e}-\ket{f}$ transmon transition.
(c) The population of $\ket{11}$, the sum of the populations of $\ket{21}$ and $\ket{12}$, and the photon number parity of the displaced state in the $\ket{e}$ manifold, measured as a function of the common displacement phase $\phi$. All fit curves correspond to the scaled coherence element $\rho_{11,22}/\sqrt{\rho_{11}\rho_{22}} = 0.63 e^{0.84i}$.
}
\label{fig:subspace_tomo}
\end{figure*}
\section{Quantum Subspace tomography}
Quantitative insight into our experimental two-cavity state is enabled by a new characterization tool focusing on specific subspaces of interest. The goal is to completely characterize the projected density matrix $\rho_{SS}=\hat{P}\rho \hat{P}$, with projection operator $\hat{P}$ for subspace $S$ of the system. For $d$-dimensional subspace $S$, we only need to perform $d^2$ measurements for subspace tomography. Using a three-level ancilla, one can effectively isolate $S$ from the rest of the Hilbert space using the $\ket{g}-\ket{e}$ transition of the ancilla, and using the $\ket{e}-\ket{f}$ transition to perform tomography. A general protocol for this method is discussed in Appendix \ref{Protocol}.
\subsection{Implementation of 2d Subspace Tomography}
We demonstrate this subspace tomography technique with a direct and self-calibrated measurement of the pair-wise coherence between the constituent states of the stabilized $\delta=0$ state. This is realized by entangling only two Fock components, e.g.~$\ket{11}$ and $\ket{22}$, with $\ket{e}$ of the transmon using number-selective $\pi$-pulses. Within this 2d subspace, since the diagonal elements of the density matrix, which we denote as $\rho_{11}$ and $\rho_{22}$, have been measured via transmon spectroscopy (Fig.~2), the complex-valued off-diagonal element $\rho_{11,22}$ is the only unknown.
We then displace both cavities $\hat{D}_a(|\alpha|e^{i\phi}) \hat{D}_b(|\beta| e^{i\phi})$ with fixed amplitudes (typically $|\alpha|$, $|\beta|<0.5$) and vary the common phase $\phi$,
which redistributes the photon populations in the $\ket{e}$ manifold and shows constructive or destructive interference in each Fock component depending on $\rho_{11,22}$ and $\phi$ [Fig.~\ref{fig:subspace_tomo}(b)].
Using number-selective $\ket{e}$-$\ket{f}$ rotation of the transmon, we can measure the population of any individual target state after displacement, such as $\ket{11e}$, which oscillates as a function of $\phi$. Alternatively, we can also measure the oscillation in other observables, such as the joint parity [Fig.~\ref{fig:subspace_tomo}(c)]. The amplitude and phase of such oscillations can be directly converted to $\rho_{11,22}$ following a comparison with easily calculable properties of the state $D(\alpha,\beta)(\sqrt{\rho_{11}}\ket{11}+\sqrt{\rho_{22}}\ket{22})$.
Repeating the same procedure for different pairs of states, we obtain the $\delta=0$ block of the system density matrix by direct measurements of its individual off-diagonal elements [Fig.~4(f)]. In practice, we run numerical calculation to find the optimal $|\alpha|$, $|\beta|$ and measurement observable for each pair of states to maximize the visibility of the oscillations.
\begin{figure*}[tbp]
\centering
\includegraphics[scale=.75]{fig7_09082022.pdf}
\caption
(a) Phases of selected off-diagonal density matrix elements of the two-cavity state, measured with the $d=2$ subspace tomography, as a function of wait time $t_w$ after the state has been prepared by the stabilizing drives for $t=15$ $\mu$s. The phase evolution agrees with the cavity Kerr terms $\hat{H}_{sk}$ and the cavity detuning in the drive frame. Dashed lines are linear fits, which allow us to interpolate or extrapolate the phases of these states to different $t_w$. The vertical dashed line corresponds to the time at which the density matrix sector is plotted in Fig.~\ref{fig:PC_4}(j). In order to compare with Wigner data in Fig.~\ref{fig:PC_4}, note that the effective free-evolution time of the two-cavity state here is $t_w+2$ $\mu$s, accounting for the duration of the selective transmon $\ket{g}-\ket{e}$ pulses in the subspace tomography protocol (Fig.~\ref{fig:subspace_tomo}(a)).
(b, c) Phases and relative coherence of selected off-diagonal density matrix elements of the two-cavity state for different stabilization pumping time $t$, measured with subspace tomography after $t_w=1$ $\mu$s.
Dashed lines: numerical simulation results of the storage pair photon dynamics according to Eq.~(\ref{eq:rho_full}), which does not contain the reservoir mode and captures all prominent features of the data.
\label{fig:coh_over_time}
}
\end{figure*}
Experimental implementation of the subspace tomography technique requires careful tracking of rotating frames and phase accumulation of different branches of states. To ensure self-consistent phase measurements, we work in the two-cavity rotating frame set by two cw stabilization drives (the drive frame), and require the four frequencies of the tomographic pulses precisely sum up to 0 (Appendix Fig.~\ref{fig:phase_loop}). Since the pulse frequencies for selective transmon rotations
have no freedom for adjustment, we apply our fast cavity displacement drives at deliberately detuned frequencies in order to exactly compensate for the irregular detunings of all other tones due to various Stark shifts and higher-order nonlinearity of the system (Appendix \ref{Freq-matching}). This strategy allows us to measure the phases of off-diagonal density matrix elements that are unambiguously defined relative to the two stabilization drives. As shown in Fig.~\ref{fig:coh_over_time}(a), following $t=15$ $\mu$s pumping, our measured phases of $\rho_{00,11}, \rho_{11,22}, \rho_{22,33}, \rho_{33,44}$ all evolve over wait time $t_w$ as expected from the frequencies of the corresponding transitions in the drive frame. Note that their frequencies differ due to the Kerr Hamiltonian of the storage cavities [Eq.~(\ref{eq:H_SK})]. Fig.~\ref{fig:coh_over_time}(b) further shows that the measured phases are constant over extended pumping times $t$.
A key advantage of this subspace tomography technique of $d=2$ is that it maps individual density matrix elements of interest to individual experimental signatures in an intuitive manner, making the experimental uncertainties of the tomographic measurements highly transparent. For example, from the sinusoidal fits in Fig.~\ref{fig:subspace_tomo}(b), we can unambiguously report the uncertainly of the amplitude (3\%) and phase (0.05 radian) for $\rho_{12,21}$,
which would be extremely challenging to obtain in traditional tomographic reconstruction.
Furthermore, it is relatively straightforward to extend our scheme to $d>2$ by applying more than 2 selective ancilla $\pi$-pulses in isolating the subspace, although extra care make be taken in accounting for phase accumulations if the selective pulses are not exactly equally spaced in frequency.
\subsection{Understanding Deviations from Ideal PCS}
With the capability of direct pair-wise coherence measurements, we can track the amplitude and phase of selected coherence elements in the two-cavity state as a function of stabilization pumping time $t$. We find that the cavities indeed converge to a quasi-steady state (except for the slow process of single photon loss) with persistent coherence (Fig.~\ref{fig:coh_over_time}(c)). However, the magnitude of the steady-state coherence is far lower than the ideal values $|\rho_{nn,mm}|=\sqrt{\rho_{nn}\rho_{mm}}$. Moreover, the stabilized cavity phases, which we
can extrapolate backwards in time from data such as Fig.~\ref{fig:coh_over_time}(a), also differ considerably from the ideal PCS, which should have equal superposition phases between neighboring Fock components, i.e.~$\arg[\rho_{nn,(n+1)(n+1)}]=\arg[\gamma]$). Due to the Kerr evolution of the storage cavities during idle time ($K_\textrm{eff}=-86$ kHz), we expect a stabilized PCS to collapse and revive every 5.8 $\mu$s~\cite{kirchmair_observation_2013}, making $t_w=5.8$ $\mu$s the first optimal timing to observe the PCS in Wigner measurements. However, the measured pair-wise phases of the stabilized state suggests that the two-cavity state most resembles a PCS at $t_w\approx4.5$ $\mu$s, which has been confirmed by the joint Wigner measurements. This skewness in phases and loss of coherence is caused by the relatively large cross-Kerr term of the reservoir mode, $K_{br}/2\pi = 58$ kHz, which is comparable to the competing mixing drive rate $g_{ab}/2\pi$ (while $K_{ar}/2\pi = 7$ kHz is much smaller).
Remarkably, this spurious effect from the reservoir cross-Kerr can be well captured by the simple reduced model of the pair-photon dynamics Eq.~(\ref{eq:rho_full}), which features a composite jump operator $\hat{o}=\sqrt{\kappa_{ab}} \hat{a}\hat{b}+\zeta_a\hat{a}^\dagger\hat{a}+\zeta_b\hat{b}^\dagger\hat{b}$. Through adiabatic elimination of the $r$ mode (Appendix B), we find $\zeta_b=\sqrt{4K^2_{br}r_0^2/\kappa_r}$ (and similarly for $\zeta_a$), where $r_0$ is the complex displacement amplitude of the $r$ mode during the stabilization process. This coefficient is identical to that of photon shot-noise dephasing of qubits~\cite{gambetta_qubit-photon_2006}, with $K_{br}$ taking the place of the qubit-cavity dispersive shift. Indeed, if the mixing drive is turned off (hence $\kappa_{ab}=0$), any initial storage cavity state $\rho$ should be dephased by the fluctuating photon numbers in a driven $r$ mode, which can be described by the usual dephasing jump operators $\hat{a}^\dagger\hat{a}$ and $\hat{b}^\dagger\hat{b}$. In our experiment, $|r_0|=1.8$,
which gives rise to a dephasing rate $|\zeta_b|^2/2\approx 2\pi\cdot28$ kHz while the effect of $\zeta_a$ is much less significant. Intuitively, the finite steady-state coherence of the system in Fig.~\ref{fig:coh_over_time} can be understood from the competition between the PCS stabilization mechanism and the spurious dephasing effects.
It is important to note that the pair-photon dissipation and the two dephasing-like terms constitute a single jump operator $\hat{o}$, not separate ones. This is because the three events are all directly originating from the same stochastic event of reservoir photon loss. As a result, the relative phases between the terms in $\hat{o}$ play a crucial role in skewing the stabilized state from the ideal PCS, including a larger state size, a broader distribution of photon numbers, and the varying phases between neighboring Fock components.
A numerical simulation of Eq.~(\ref{eq:rho_full}) using this composite jump operator reproduces the prominent coherence elements of the pumped two-cavity state, which is in good agreement with the experimental data in Fig.~\ref{fig:coh_over_time} with no free parameters except for a two-cavity global frame rotation of 1.5 radian (which can be attributed to the difference in cable electrical length between the stabilization drive lines and the cavity displacement lines together with a small additional cavity rotation during the reservoir ring-down before the tomography).
To improve the quality of the stabilized PCS, it is necessary to suppress the spurious dephasing terms in the composite jump operator $\hat{o}$. Ideally, this can be accomplished by engineering a Josephson circuit without the reservoir cross-Kerr, analogous to the asymmetrically threaded SQUID circuit for dissipative cat-state stabilization~\cite{lescanne_exponential_2020}. In our current transmon-based device, this may be accomplished by either an extra Kerr-cancellation pump tone~\cite{zhang_drive-induced_2022} or by detuning the two stabilization drives a few (reservoir) linewidths away to reduce photon number fluctuations in the $r$ mode. The former will also suppress $|K_\textrm{eff}|$ and hence boosts the state size and the dissipative nature of the stabilization scheme. The latter will suppress $\kappa_{ab}$ and move our PCS stabilization deeper into the regime relying on Hamiltonian confinement. Given the full understanding of the pair photon dynamics at the conclusion of this experiment, we believe both strategies can yield improved fidelity for the stabilized PCS in our device, although new forms of nonlinear ancillas are needed to realize spurious-free pair photon processes at faster rates.
\section{Summary and outlook}
There is a growing history of circuit QED experiments inspired by the field of quantum optics. The convenience of microwave systems and the powerful nonlinearity of superconducting circuitry has paved new ways
to the study of exotic bosonic states that may be otherwise prohibitively challenging to implement.
Owing to the identical nature of bosons, bosonic quantum states offer the opportunities of hardware-efficient dissipation engineering schemes, but also poses unique challenges requiring new characterization techniques. Both aspects are reflected in our experiments. Unlike earlier demonstrations of two-mode non-Gaussian states such as the entangled cat state~\cite{wang_schrodinger_2016} or the N00N state~\cite{wang_deterministic_2011}, the PCS should contain more than one e-bit (EPR pair) of entanglement. It would be interesting to further develop efficient tools to calibrate and distill the entanglement in such states.
The pair cat code, a recent addition to the zoo of bosonic QEC codes, offers the tantalizing prospect to correct both photon loss and dephasing errors (to the first order) autonomously and fault-tolerantly~\cite{albert_pair-cat_2019}.
Our experimental realization of the PCS, demonstration of the first cross cavity dissipator, and introduction of a convenient (although not fault-tolerant) PND measurement are all valuable steps towards a pair cat code.
However, fully implementing and utilizing the advantages of the pair cat code will require schemes to suppress forward propagation of ancilla errors and upgrading the pair photon dissipation to a four photon cross cavity dissipator, which involves a leap in experimental complexity.
As a first attempt towards a two-mode bosonic code, our work suggests several challenges moving from one mode to two. Unlike the concatenation of a smaller multi-qubit QEC code to a larger one, two-mode bosonic code states generally cannot be created from concatenating single-cavity bosonic qubits using cavity-cavity logic gates. Generation of two-mode states in circuit QED requires specific cross-cavity nonlinear interactions, which has been implemented in our work with a relatively simple hardware setup but not without spurious effects and difficult parameter trade-offs. Future experiments will have to tailor the cross-cavity nonlinear interactions likely using Josephson circuits with a strong 3rd-order nonlinearity~\cite{bergeal_phase-preserving_2010} while suppressing unwanted 4th-order terms.
In addition, since two-mode states reside in a much larger Hilbert space, their implementation requires both effective confinement to preferred subspaces and efficient methods to diagnose imperfections. To the latter, our work demonstrates a valuable tool in subspace tomography, which provides a general framework for future experiments to isolate key components of a high-dimensional quantum state for characterization.
\subsection*{Acknowledgements}
We thank Michel Devoret, Ioannis Tsioutsios, Akshay Koottandavida and Yu-Xin Wang for helpful discussions. We thank Juliang Li and Xiaowei Deng for experimental assistance. This research was supported by the US Department of Energy (DE-SC0021099). Fabrication and initial characterization of the device was supported by the Air Force Office of Scientific Research (FA9550-18-1-0092). LJ would like to acknowledge the support from the ARO (W911NF-18-1-0212),
AFOSR MURI (FA9550-19-1-0399) and the Packard Foundation (2020-71479).\\
\noindent
$\dagger$ There authors contributed equally.
\noindent
$*$ Correspondence and requests for materials should be addressed to Chen Wang, <EMAIL>.
\begin{appendix}
\renewcommand\thefigure{\arabic{figure}}
\renewcommand{\figurename}{\textbf{Appendix Fig.}}
\renewcommand{\tablename}{\textbf{Appendix Table}}
\section{Experimental Setup and Methods}\label{Setup}
\subsection{Device architecture and transmon fabrication}
The system consists of two post cavities dispersively coupled to a fixed frequency transmon ancilla which is then dispersively coupled to a stripline $\lambda/2$ resonator that is used for readout and as reservoir in the experiment. The posts are made with high-quality Aluminum inside a cavity as shown in Fig. ~\ref{fig:fridge_alCav}b. The cavity has a tunnel opening on one of the sides which is where the sapphire chip containing the qubit and resonator is held with a chip clamp. This high-quality superconducting cavity is made with 5N ($99.999\%$) purity Aluminum and it shields the qubit from magnetic field lines that can degrade the coherence. The ancilla qubit and the cavity modes each have their own drive ports. The readout signal is collected from another port that is coupled strongly with the stripline resonator. The coupling strength between all the modes and the coherence numbers for the modes are listed as in the Table. ~\ref{table:parameters}. The setup with the cavities and chip was initially simulated in the electromagnetic solver software called Ansys HFSS to design the frequencies of the modes and interactions between them that are needed for the experiment.
\par The simulated chip design is fabricated in a clean room facility for nanofabrication using the standard procedure. We use a 30keV JEOL JSM-7001F SEM to perform electron beam lithography to define the transmon and the stripline resonator in one step. The aluminum thin film is evaporated on the sapphire chip using Plassys MEB550S. The transmon has a single Al-AlO$_x$-Al Josephson junction produced by the Dolan bridge method where the aluminum is deposited in two different angles while allowing some time in between for it to oxidize.
\subsection{Measurement setup}
The measurement setup is designed to generate microwave signals that drive all the four modes in the system and to receive signals from the readout port that are digitized to read the qubit state. We use the basic microwave engineering technique of mixing signals from the signal generator and the arbitrary waveform generator (AWG) using an IQ-mixer to modulate the sidebands. The filtered RF output from each mixer then goes into the fridge to the input ports as shown in Fig. ~\ref{fig:fridge_alCav}a. The signal collected from the readout port goes to a 3-port mixer for demodulation and then to a digitizer. In addition to drives for the modes of the system the setup also produces continuous four-wave mixing tone and a reservoir tone that satisfy the frequency matching condition as discussed in the main text to create a PCS. Inside the fridge all the input signals are attenuated by 20 dBm at 4K plate, 10 dBm at Still plate and 30 dBm at the MXC, which then go through the eccosorbs and last level of filtering at MXC before going into the input ports.
\par We want to study the coherence of the prepared entangled state in the storage cavities by probing them to measure 2D slices of joint wigners as shown in Fig.~\ref{fig:PC_4}b-d. Using different generators to pump and probe the cavities would cause the phase of the state in the storage cavities to not be locked to the cavity displacement drives used to measure Wigner functions, thus smearing out any phase coherence. We therefore eliminate a phase degree of freedom by generating cavity B drive from a reservoir, cavity A and the four-wave mixing generators. As shown in Fig.~\ref{fig:fridge_alCav}a the signal from the generators for the four-wave mixing pump and cavity A drive are mixed with a 3-port mixer which, with a low pass filter, produces an output signal at $\omega_{a}-\omega_{p}$. This signal is again mixed with the signal from the reservoir drive generator with a 3-port mixer and a low pass filter on the output to produce $\omega_{r}-\omega_{a}+\omega_{p}$, thus generating a local oscillator for the IQ-mixer modulating the sideband for driving cavity B.
\begin{table}[bp]
\caption{Measured System parameters.\\
$^\dagger$ Values measured in the presence of the relatively strong four-wave mixing tone. \\
$^*$ $K_{aa}$ is estimated from the measured $K_{ab}$ and $K_{bb}$.
}
\centering
\begin{tabular}{c c c c c}
\hline\hline\\[-2ex]
& Symbol & Value \\
\hline\\[-2ex]
Transmon frequency & $\omega_q/2\pi$ & 5378 MHz\\
Transmon anharmonicity & $\alpha_q/2\pi$ & 204 MHz \\
Transmon $T_1$ & $T_{1q}$ & $40$ $\mu$s \\
Transmon $T_2^*$ Ramsey & $T^*_{2q}$ & $18$ $\mu$s \\
Transmon $T_2$ Echo & $T_{2q}$ & $60$ $\mu$s \\
Transmon $\ket{e}_q$ population & & 2\% $(3.5\%^{\dagger})$ \\
\hline\\[-2ex]
Reservoir frequency & $\omega_r/2\pi$ & 7409 MHz \\
Reservoir-transmon coupling & $\chi_{qr}/2\pi$ & 3.09 MHz \\
Reservoir anharmonicity & $K_{rr}/2\pi$ & 12 kHz \\
Reservoir decay rate & $\kappa/2\pi$ & 0.78 MHz \\%$204$ ns \\
\hline\\[-2ex]
Cavity $a$ frequency & $\omega_a/2\pi$ & 4072 MHz \\
Cavity $a$-transmon coupling & $\chi_{qa}/2\pi$ & 1.89 MHz \\
Cavity $a$ anharmonicity & $K_{aa}/2\pi$ & 8 kHz$^{*}$ \\
Cavity $a$ $T_1$ & $T_{1a}$ & $530$ $\mu$s \\
Cavity $a$ $T_2$ & $T_{2a}$ & $400$ $\mu$s \\
Cavity $a$ $\ket{1}_a$ population & & $\sim$2\%\\%($4\%^{\ddagger}$) \\
\hline\\[-2ex]
Cavity $b$ frequency & $\omega_b/2\pi$ & 6094 MHz \\
Cavity $b$-transmon coupling & $\chi_q/2\pi$ & 6.26 MHz \\
Cavity $b$ anharmonicity & $K_{bb}/2\pi$ & 71 kHz (81 kHz$^{\dagger}$)\\
Cavity $b$ $T_1$ & $T_{1b}$ & $216$ $\mu$s \\
Cavity $b$ $T_2$ & $T_{2b}$ & $200$ $\mu$s \\
Cavity $b$ $\ket{1}_b$ population & & $\sim$1\%\\% ($2\%^{\ddagger}$)\\
\hline\\[-2ex]
Cavity $a$-$b$ cross-Kerr & $K_{ab}/2\pi$ & 48 kHz (53 kHz$^{\dagger}$)\\
Cavity $a$-$r$ cross-Kerr & $K_{ar}/2\pi$ & 7 kHz\\
Cavity $b$-$r$ cross-Kerr & $K_{br}/2\pi$ & 58 kHz\\
\hline
\end{tabular}
\label{table:parameters}
\end{table}
\begin{figure*}[tbp]
\centering
\includegraphics[scale=0.6]{fridge_diagram_08052022.pdf}
\caption{(a) Fridge and Measurement Wiring Diagram: We have four signal generators which, with eight AWG channels, make input signals for all the modes and FWM tone. As discussed in the supplementary, the measurement setup shows that the drive signal for Cavity $b$ is generated by mixing the signals from cavity $a$, reservoir and the FWM generators with 3-port mixers and using the correct sideband filters. (b) Picture of high purity aluminum cavity used in the experiment. The posts that give the storage cavity modes are seen inside the aluminum body with the sapphire chip entering the cavity from a side tunnel. The chip is held by an aluminum clamp and the transmon qubit antenna is at equal distance from both the posts.
The cavity is mounted on a bracket that is placed inside a high-permeability magnetic shield thermalized to the mixing chamber.
}\label{fig:fridge_alCav}
\end{figure*}
\subsection{Pump Tuneup Procedure}
Given the presence of the storage-reservoir cross-Kerr interaction, there is no perfect experimental pumping condition that can exactly stabilize the PCS. In our experimental procedure, with the intention of studying driven dissipative dynamics, we aimed to maximize the pair-photon dissipation rate $\kappa_{ab}$ while empirically obtaining a steady state with relatively large photon numbers and minimal undesirable heating effects.
In order to maximize the ratio of the two-photon dissipation rate to the single-photon loss rate in the cavities $\kappa_{ab}/\kappa_{a,b}$, a relatively strong four-wave mixing (FWM) pump amplitude $g_{ab}$ is needed. We set the power of the FWM pump at a value where we start to observe a small measurable rise of the excited state population of the ancilla transmon. The power of the reservoir drive is set to displace the reservoir to a few photons in its steady state, as can be measured via the Stark shift and dephasing of the transmon. We start by setting the frequency of the reservoir and FWM drive to be on resonance with the reservoir $\omega_{r}$ and the FWM condition $\omega_{p}=\omega_{a}+\omega_{b}-\omega_{r}$, respectively. We also use spectroscopy measurements to calibrate the Stark-shifted mode frequencies, which give us improved estimates of the pumping frequency condition.
These two crudely set tones allow preparation of some non-trivial
pair photon states (e.g.~with substantial $\ket{11}$ and $\ket{22}$ states) after 10-20 $\mu$s of pumping time.
To calibrate the optimal frequency for the FWM pump that maximizes $\kappa_{ab}$, we start with the above crudely-prepared state, and then attempt to apply only the mixing pump at varying frequencies to evacuate photons in pairs. We finely sweep through the FWM frequencies, and measure the population of $\ket{00}$ after a fixed amount of pumping time. The FWM detuning $\Delta_{p}$ at which we get the largest $\ket{00}$ state gives the maximum $\kappa_{ab}$. Experimentally we found $\Delta_{p}=-125$ kHz relative to expected value from undriven mode frequencies $\omega_{p}=\omega_{a}+\omega_{b}-\omega_{r}$, primarily due to the Stark shift.
We can extract the two-photon dissipation rate $\kappa_{ab}$ from these decay measurements by fitting them to simulations as discussed in Appendix C.
The next step is to tune the amplitude and frequency of the reservoir drive to make a relatively large $\delta=0$ pair photon state with Poisson-like distribution of Fock state amplitudes in line with the PCS distribution.
We found that the optimal reservoir drive frequency is about -185 kHz detuned from its bare frequency, which can be explained by the cross-Kerr effect $\hat{H}_{rk}$ in the presence of storage photons and the ac Stark from the pump tone.
Finally, in a verification experiment depicted in Fig.~\ref{fig:PC_2}(b), we measure the cavity photon population distribution after 15 $\mu$s of stabilization drives with a fixed reservoir drive condition while varying FWM frequency. This figure shows that indeed the pair-photon drive is also most efficient when we pump at the FWM frequency tuned up to the previously determined $\Delta_{p}$ based on maximizing pair-photon dissipation.
\section{Driven Hamiltonian and Adiabatic elimination}\label{Derivation}
\subsection{Rotating frame Hamiltonian transformation}
In this section, we derive the four-wave mixing effect of the off-resonant pump tone on the Josephson circuit under the rotating wave approximation, arriving at the drive Hamiltonian in the rotating frame Eq.~(\ref{eq:H_int}). This procedure is similar to various previous experiments~\cite{leghtas_confining_2015, gao_programmable_2018, gertler_protecting_2021}.
We start by writing our Hamiltonian as a sum of our four modes (reservoir: $r$, qubit: $q$, storage a: $a$, storage b: $b$) coupled to a josephson junction with two drives (FWM pump: $p$, reservoir drive: $d$) applied to the reservoir mode~\cite{blais_circuit_2020}:
\begin{align}
\hat{H}_{full}/\hbar = \sum_{m=q,r,a,b}\bar{\omega_{m}}\hat{m}^\dagger\hat{m} -
E_{J}(cos(\hat{\phi})+\hat{\phi}^{2}/2) \\\nonumber +2Re(\epsilon_{p}e^{-i\omega_{p}t}+\epsilon_{d}e^{-i\omega_{d}t})(\hat{r}+\hat{r}^\dagger)
\end{align}
\begin{equation}
\hat{\phi} = \sum_{m=q,r,a,b}\phi_{m}(\hat{m}^\dagger+\hat{m})
\end{equation}
The first term in Eq.~(B1) represents the linear character of each mode corresponding to $a_{i}$. The Josephson junction is represented by the cosine term with $E_{J}$ as the Josephson energy and $\hat{\phi}$ as the phase across the junction decomposed into the contributions from each mode with contribution of $\phi_{m}$ to the zero point fluctuations of $\hat{\phi}$. The two drive terms are acting on the reservoir mode with complex amplitudes $\epsilon_{d}$, $\epsilon_{p}$ and frequencies $\omega_{d}$, $\omega_{p}$, respectively, where the pump is a strong off resonant tone and the drive is a weak near resonant tone with the reservoir mode.
We are in a regime where the following inequality holds:
\begin{equation}
\omega_{p},\omega_{d},\bar{\omega_{m}} \gg \epsilon_{p} \gg \frac{E_{J}}{\hbar} ||\hat{\phi}||^{4}/4!
\end{equation}
We now want to make a change of frame using the following unitary to eliminate the fastest time scales:
\begin{equation}
U = e^{i\bar{\omega_{q}}t\hat{q}^\dagger\hat{q}}e^{i\omega_{d}t\hat{r}^\dagger\hat{r}}e^{i\omega_{da}t\hat{a}^\dagger\hat{a}}e^{i\omega_{db}t\hat{b}^\dagger\hat{b}}e^{-\Tilde{\xi_p}\hat{r}^\dagger+\Tilde{\xi_p}^{*}\hat{r}}
\end{equation}
We are rotating out the qubit frequency $\bar{\omega_{q}}$ simply for convenience as it will not change the rest of the analysis in any meaningful way. We then rotate out the drive frequency $\omega_{d}$ on the reservoir mode so the drive term is stationary in the chosen frame. We then want to put our storage modes in the drive frame of $\omega_{d}+\omega_{p}$. We can do this by rotating out frequency $\omega_{da}(\omega_{db})$ from mode $a(b)$. Any $\omega_{da},\omega_{da}$ that satisfy $\omega_{
da}+\omega_{db} = \omega_{d}+\omega_{p}$ will put us in the correct drive frame, thus there is some freedom as to which frequencies we choose. Here we choose a constant offset($\frac{\Delta_{sd}}{2}$) of $\omega_{da},\omega_{db}$ from $\bar{\omega_{a}},\bar{\omega_{b}}$, respectively, such that $\bar{\omega_{a}} -\omega_{da} = \bar{\omega_{b}} - \omega_{db} = \frac{\Delta_{sd}}{2}$. Finally we want to apply a displacement unitary to eliminate the time dependant amplitude ($\Tilde{\xi_{p}}$) in the reservoir caused by the pump tone to bring the effect of this amplitude into the $\hat{\phi}$ operator allowing the effects on the other modes to become directly apparent. Looking at times on the order or greater than $1/\kappa_{r}$ we can ignore the reservoir transient dynamics and look at the steady state response of our displacement amplitude as done in \cite{leghtas_confining_2015}:
\begin{align}
\Tilde{\xi_{p}}&=\xi_{p}e^{-i\omega_{p}t},\\\nonumber
\xi_{p} &= \frac{-i\epsilon_{p}}{\kappa_{r}/2 + i(\bar{\omega_{r}}-\omega_{p})}\approx \frac{-i\epsilon_{p}}{\kappa_{r}/2 + i(\omega_{r}-\omega_{p})}
\end{align}
The Hamiltonian in this new frame ($\hat{H}_{full}'$) now becomes:
\begin{align}
\hat{H}_{full}'/\hbar =& (\bar{\omega_{r}}-\omega_{d})\hat{r}^\dagger\hat{r}+\Delta_{sd}\hat{a}^\dagger\hat{a}\\\nonumber&+\Delta_{sd}\hat{b}^\dagger\hat{b} - \frac{E_{J}}{\hbar}(cos(\Tilde{\hat{\phi}})+\Tilde{\hat{\phi}}^{2}/2)
\end{align}
\begin{align}
\Tilde{\hat{\phi}} = \sum_{m=q,r,a,b}\phi_{m}(\Tilde{\hat{m}}+\Tilde{\hat{m}}^\dagger) + (\Tilde{\xi_{p}}+\Tilde{\xi_{p}}^{*})\phi_r
\end{align}
\begin{align}
&\Tilde{\hat{q}}=\hat{q}e^{-i\bar{\omega_{q}}t},\Tilde{\hat{r}}=\hat{r}e^{-i\omega_{d}t}, \\\nonumber &\Tilde{\hat{a}}=\hat{a}e^{-i\omega_{da}t},\Tilde{\hat{b}}=\hat{b}e^{-i\omega_{db}t}
\end{align}
We now expand the cosine term to fourth order and keep only non-rotating terms in alignment with the rotating wave approximation and separate the Hamiltonian into 3 parts ($\hat{H}_{full}' = \hat{H}_{freq} + \hat{H}_{Kerr} + \hat{H}_{drive}$) as defined below:
\begin{align}
\hat{H}_{freq} = & (- \chi_{qr}|\xi_{p}|^{2})\hat{q}^\dagger\hat{q} \\\nonumber
& + (\omega_{r}-\omega_{d}-2K_{rr}|\xi_{p}|^{2})\hat{r}^\dagger\hat{r}\\\nonumber
& + (\Delta_{sd}-K_{ar}|\xi_{p}|^{2})\hat{a}^\dagger\hat{a}\\\nonumber
& + (\Delta_{sd}-K_{br}|\xi_{p}|^{2})\hat{b}^\dagger\hat{b}
\end{align}
\begin{align}
\hat{H}_{Kerr} =& \sum_{m=r,a,b}-\frac{K_{mm}}{2}{(\hat{m}^\dagger\hat{m})^{2}}-\alpha_{q}(\hat{q}^\dagger\hat{q})^{2}-\chi_{qr}\hat{q}^\dagger\hat{q}\hat{r}^\dagger\hat{r} \nonumber\\
&-K_{ar}\hat{a}^\dagger\hat{a}\hat{r}^\dagger\hat{r}-K_{br}\hat{b}^\dagger\hat{b}\hat{r}^\dagger\hat{r}-K_{ab}\hat{b}^\dagger\hat{b}\hat{a}^\dagger\hat{a}
\end{align}
\begin{align}
\hat{H}_{int} = g_{ab}\hat{a}^\dagger\hat{b}^\dagger\hat{r} + \epsilon_{d}\hat{r}^\dagger + h.c.
\end{align}
$H_{freq}$ gives the frequency shifts to all the elements and the shifts from terms containing $|\xi_{p}|^2$ are the AC Stark shift effects from the pump tone. $H_{Kerr}$ corresponds to both self and cross-Kerr coupling terms where $K_{mm}=\frac{E_{J}}{\hbar}\phi_{m}^{4}/2, K_{mm'}=\frac{E_{J}}{\hbar}\phi_{m}^{2}\phi_{m'}^{2}$ where $m\neq m'$ and $g_{ab} = \phi_a \phi_b \phi_r^{2} \xi_p$.Note $H_{0}$ in Eq.~(\ref{eq:H0}) is analogous to $H_{Kerr}$ just with a truncation of the allowed qubit states to the $\ket{f}$ state.
\subsection{Semi-classical analysis of reservoir state}
In this section we
perform a semi-classical analysis of the Langevin equations of motion for the reservoir and storage modes and show that the impact of the storage states on the reservoir dynamics is quite small and warrants treating the effect as a small perturbation. This condition forms the basis for
the adiabatic elimination of the reservoir mode in the following section.
We first will assume the qubit is in the ground state during the stabilization. This leads us to work with the Hamiltonian:
\begin{align}
\hat{H}=& \Delta_d\hat{r}^\dagger\hat{r} + \Delta_a\hat{a}^\dagger\hat{a}
+ \Delta_b\hat{b}^\dagger\hat{b} \\\nonumber &- \sum_{m=r,a,b}\frac{K_{mm}}{2}{(\hat{m}^\dagger\hat{m})^{2}} - K_{ab}\hat{a}^\dagger\hat{a}\hat{b}^\dagger\hat{b} \\\nonumber &- K_{ar}\hat{a}^\dagger\hat{a}\hat{r}^\dagger\hat{r} - K_{br}\hat{a}^\dagger\hat{a}\hat{r}^\dagger\hat{r} + \hat{H}_{int}
\end{align}
where $\Delta_d = (\omega_{r}-\omega_{d}-2K_{rr}|\xi_{p}|^{2}),\Delta_a = (\Delta_{sd}-K_{ar}|\xi_{p}|^{2}),\Delta_b = (\Delta_{sd}-K_{br}|\xi_{p}|^{2})$. For the following analysis, for simplicity we will focus on analyzing the $\delta = 0$ storage state, but the conclusion, as will be explained later, is general. In this state our $\hat{a}^\dagger\hat{a}$ and $\hat{b}^\dagger\hat{b}$ terms are equivalent, thus we can group the storage cavity self and cross Kerr terms into one Kerr term, $K_\textrm{eff}\hat{a}^\dagger\hat{a}\hat{b}^\dagger\hat{b}, K_\textrm{eff}=-K_{ab}-K_{aa}/2-K_{bb}/2$. This yields the following simplified Hamiltonian ($\hat{H}_{\delta 0}$) for a $\delta = 0$ state:
\begin{align}
\label{eq:Hd0}
\hat{H}_{\delta 0}=& \Delta_d\hat{r}^\dagger\hat{r} + \Delta_a\hat{a}^\dagger\hat{a}
+ \Delta_b\hat{b}^\dagger\hat{b} \\\nonumber &- \frac{K_{rr}}{2}{(\hat{r}^\dagger\hat{r})^{2}} + K_\mathrm{eff}\hat{a}^\dagger\hat{a}\hat{b}^\dagger\hat{b} \\\nonumber &- K_{ar}\hat{a}^\dagger\hat{a}\hat{r}^\dagger\hat{r} - K_{br}\hat{a}^\dagger\hat{a}\hat{r}^\dagger\hat{r} + \hat{H}_{int}
\end{align}
We now want to motivate the statement that the effect of the storage modes on the reservoir is quite small from both the effects of the pumping $(g_{ab})$ and the reservoir storage cross Kerr $(K_{ar},K_{br})$.
The pumping effects can be seen by looking at the Langevin equations for the classical amplitude analogs of $\langle\hat{r}\rangle \rightarrow r, \sqrt{\langle \hat{a}\hat{b}\rangle}\rightarrow s$ in this Hamiltonian. We can then get the steady state solutions to these Langevin equations:
\begin{align}
\label{eq:Langevin}
\frac{d}{dt}r & = -i\epsilon_d-(\kappa_r/2+i\Delta_{d})r-ig^*_{ab}s^2=0 \\\nonumber
\frac{d}{dt}s & = -2ig_{ab}s^*r+2iK_\mathrm{eff}|s|^2s=0
\end{align}
From this we can algebraically solve for the steady-state mode amplitude:
\begin{align}
r = \frac{\epsilon_d}{-\frac{|g_{ab}|^2}{K_\mathrm{eff}}+\frac{i(\kappa_r+2i\Delta_d)}{2}}
\end{align}
We are in the limit of $|\kappa_r+2i\Delta_d| \gg \frac{2|g_{ab}|^2}{K_\mathrm{eff}}$ so we can look at the mode amplitude as an uncoupled driven mode with a perturbation from the coupling to the storage cavity as:
\begin{align}
\label{eq:r0_def}
& r \approx r_0 + \delta_{disp}\\\nonumber
& r_0 = \frac{2\epsilon_d}{i(\kappa_r+2i\Delta_d)} \\\nonumber
& \delta_{disp} = -\frac{4\epsilon_d|g_{ab}|^2}{K_\textrm{eff}(\kappa_r+2i\Delta_d)^2}, |\delta_{disp}| \ll 1
\end{align}
For our experimental parameters, $|r_0|=1.8$. We note that the storage-reservoir Langevin equation Eq.~(\ref{eq:Langevin}) takes a similar form as previous single-mode two-photon driven dissipation experiments~\cite{leghtas_confining_2015, lescanne_exponential_2020}, but our system is in a different parameter regime due to the presence of a large $K_\mathrm{eff}$. In the absence of $K_\mathrm{eff}$, the steady-state $r$ would be close to 0 even if $|r_0|>1$.
While this was derived under the steady state and with a $\delta=0$ state to more concretely illustrate the point, one can more generally consider the `pull' the storage modes have on the reservoir as $|g_{ab}|^2/K_\mathrm{eff}$ and compare this to the loss rate $\kappa_{r}$ and see that we are in a regime where $\kappa_{r}\gg|g_{ab}|^2/K_\mathrm{eff}$, thus the effect of the storage modes on the reservoir state via the pumping effects is quite small. Now the reservoir storage cross Kerr effects must be taken into account. The cross Kerr values will simply add a further detuning on top of $\Delta_{d}$ to the reservoir amplitude for each number state of the storage. Since our pair coherent state amplitude $\gamma$ is on the order of the photon number in each mode in the system for the $\delta = 0$ state we are analyzing, we will use this to represent the average photon number in the inequality. Thus, we must satisfy the inequality $\kappa_{r} \gg \gamma(K_{ar}+K_{br})$ to ensure the effects of the storage reservoir cross kerr can be treated as a perturbation on top of the bare reservoir amplitude. Since our parameter regime satisfies $\kappa_{r} \gg \gamma(K_{ar}+K_{br}),|g_{ab}|^2/K_\mathrm{eff}$, we are well motivated to treat the reservoir amplitude as an uncoupled amplitude $r_{0}$ with a perturbation added to account for the effect from the storage modes.
\subsection{Adiabatic elimination of the reservoir mode}
In this section we use the above analysis of the reservoir dynamics as an uncoupled amplitude with a small perturbation accounting for the storage mode effects to adiabatically eliminate the reservoir mode. We will do this by going into a displaced frame of the reservoir mode using the uncoupled mode ampliude $r_{0}$ in which the resulting reservoir state is solely from the perturbation effects of the storage mode and hence remains close to vacuum. This small resulting reservoir amplitude allows us to write the density matrix as a perturbative expansion in the reservoir states. Then using the fast time scale of $\kappa_{r}$ we can adiabatically eliminate the reservoir mode to arrive at Eq.~(\ref{eq:rho_full}).
Using the above analysis of the reservoir mode, we can apply a unitary transformation $U_{disp}=e^{-r_{0}\hat{r}^\dagger+r_{0}^{*}\hat{r}}$ to $\hat{H}$ resulting in $\hat{H}' = U_{disp}\hat{H}U_{disp}^\dagger$ to go into a frame where the coherent state displacement on the reservoir mode is represented by just $\delta_{disp}$ in addition to the effect on the amplitude from the reservoir storage cross kerr effects which are also small:
\begin{align}
\hat{H}' =&
\Delta_d(\hat{r}^\dagger\hat{r} + \hat{r}^\dagger r_{0} +\hat{r}r_{0}^* +|r_0|^2) + \Delta_a\hat{a}^\dagger\hat{a}
\\\nonumber&+ \Delta_b\hat{b}^\dagger\hat{b} - \sum_{m=r,a,b}\frac{K_{mm}}{2}{(\hat{m}^\dagger\hat{m})^{2}} - K_{ab}\hat{a}^\dagger\hat{a}\hat{b}^\dagger\hat{b}\\\nonumber & -(K_{ar}\hat{a}^\dagger\hat{a} + K_{br}\hat{b}^\dagger\hat{b})(\hat{r}^\dagger\hat{r} + \hat{r}^\dagger r_{0} +\hat{r}r_{0}^* +|r_0|^2) \\\nonumber &+ g_{ab}\hat{a}^\dagger\hat{b}^\dagger(\hat{r}+r_{0}) + \epsilon_{d}(\hat{r}^\dagger+r_{0}^*) \\\nonumber &+ g_{ab}^*\hat{a}\hat{b}(\hat{r}^\dagger+r_{0}^*) + \epsilon_{d}^*(\hat{r}+
r_{0})
\end{align}
One can first omit all constant energy offset terms and then combine the $\Delta_{d}r_{0}\hat{r}^\dagger + h.c.$ terms with the resulting $\kappa_r$ Lindbladian loss operator terms of the same order from the same unitary transformation to cancel out the $\epsilon_{d}$ terms, yielding the following simplified $\hat{H}'$:
\begin{align}
\hat{H}' =&
\Delta_d\hat{r}^\dagger\hat{r} + \Delta_a\hat{a}^\dagger\hat{a}
+ \Delta_b\hat{b}^\dagger\hat{b} \\\nonumber& - \sum_{m=r,a,b}\frac{K_{mm}}{2}{(\hat{m}^\dagger\hat{m})^{2}} - K_{ab}\hat{a}^\dagger\hat{a}\hat{b}^\dagger\hat{b}\\\nonumber& - (K_{ar}\hat{a}^\dagger\hat{a} + K_{br}\hat{b}^\dagger\hat{b})(\hat{r}^\dagger\hat{r} + \hat{r}^\dagger r_{0} +\hat{r}r_{0}^* +|r_0|^2) \\\nonumber& + g_{ab}\hat{a}^\dagger\hat{b}^\dagger(\hat{r}+r_{0}) + g_{ab}^*\hat{a}\hat{b}(\hat{r}^\dagger+r_{0}^*)
\end{align}
Now, we can group terms that are small and say that the following dimensionless quantities,$\frac{\gamma(K_{ar}+K_{br})}{\kappa_{r}},\delta_{disp},\kappa_{a(b)}/\kappa_{r},g_{ab}/\kappa_{r},K_{aa(bb)}/\kappa_{r}$, are on the order of $\epsilon$ where $\epsilon\ll1$ and $r_{0}\approx\mathcal{O}(1)$. Since, as explained earlier, our reservoir state size is small, $~\mathcal{O}(\epsilon)$, in this displaced frame so to good approximation we can decompose our full density matrix with both storage modes and the reservoir ($\rho_{abr}$) as follows, where $\rho_{ij}$ signifies the reduced density matrix element of the storage cavities alone entangled to the reservoir state $\ket{i}\bra{j}$:
\begin{align}
\rho_{abr} = &\rho_{00}\ket{0}\bra{0} + \epsilon(\rho_{01}\ket{0}\bra{1}+\rho_{10}\ket{1}\bra{0})\\\nonumber &+\epsilon^2(\rho_{11}\ket{1}\bra{1}+\rho_{02}\ket{0}\bra{2}+\rho_{20}\ket{2}\bra{0})
\end{align}
We will now take $\Delta_d=(\Delta_a-K_{ar}|r_0|^2)=(\Delta_b-K_{br}|r_0|^2)=0$ because experimentally we drive the reservoir on resonance after the Stark shift is accounted for and we have calibrated our drives such that $2\Delta_{sd} = K_{ar}|r_0|^2+K_{br}|r_0|^2+K_{ar}|\xi_{p}|^{2}+K_{br}|\xi_{p}|^{2}$. For a $\delta = 0$ state, this calibrated equality is analogous to setting $(\Delta_a-K_{ar}|r_0|^2)=(\Delta_b-K_{br}|r_0|^2)=0$. This brings us to the following master equation for our density matrix $\rho_{abr}$:
\begin{align}
\frac{d}{dt}\rho_{abr} = -i[\hat{H}',\rho_{abr}] + \kappa_r D[\hat{r}]\rho_{abr} \\\nonumber+ \kappa_a D[\hat{a}]\rho_{abr} +
\kappa_b D[\hat{b}]\rho_{abr}
\end{align}
We can now use this to derive the dynamics of just the storage cavities ($\rho$) from getting $\mathrm{Tr}_{r}[\rho_{abr}] = \rho_{00}+\epsilon^{2}\rho_{11}$ up to second order in $\epsilon$. First, looking at the evolution of $\rho_{00}$ by multiplying our master equation by $\bra{0}$ and $\ket{0}$, we get:
\begin{align}
\frac{1}{\kappa_{r}}\frac{d}{dt}\rho_{00} = & -\frac{i}{\kappa_{r}}\bra{0}[\hat{H}',\rho_{abr}]\ket{0} + \epsilon^{2}\rho_{11} \\\nonumber
& + \frac{\kappa_a}{\kappa_r}D[\hat{a}]\rho_{abr} +
\frac{\kappa_b}{\kappa_r}D[\hat{b}]\rho_{abr} \\\nonumber
= & -i\epsilon^2(\hat{A}^{\dagger}\rho_{10}-\rho_{01}\hat{A}) \\\nonumber
& -i[-\frac{K_{aa}}{2\kappa_r}(\hat{a}^{\dagger}\hat{a})^{2}-\frac{K_{bb}}{2\kappa_r}(\hat{b}^{\dagger}\hat{b})^{2} - \frac{K_{ab}}{\kappa_r}\hat{a}^{\dagger}\hat{a}\hat{b}^{\dagger}\hat{b} \\\nonumber
& +\frac{g_{ab}}{\kappa_r}\hat{a}^{\dagger}\hat{b}^{\dagger}(r_{0}) + \frac{g_{ab}}{\kappa_r}^*\hat{a}\hat{b}(r_{0}^{*}),\rho_{00}] \\\nonumber
& + \epsilon^{2}\rho_{11}+
\frac{\kappa_a}{\kappa_r}D[\hat{a}]\rho_{abr} +
\frac{\kappa_b}{\kappa_r}D[\hat{b}]\rho_{abr}
\end{align}
where $A\equiv \frac{1}{\epsilon\kappa_r}(g_{ab}^{*}\hat{a}\hat{b}-K_{ar}r_{0}\hat{a}^{\dagger}\hat{a}-K_{br}r_{0}\hat{b}^{\dagger}\hat{b})$ thus making $||A|| \approx \mathcal{O}(1)$. Now we can find similar evolution equations for $\rho_{10}$ and $\rho_{11}$, working only up to zeroth order in $\epsilon$:
\begin{align}
\frac{1}{\kappa_{r}}\frac{d}{dt}\rho_{10} & = -i\hat{A}\rho_{00} - \frac{1}{2}\rho_{10} +\mathcal{O}(\epsilon)
\\\nonumber
\frac{1}{\kappa_{r}}\frac{d}{dt}\rho_{11} & = -i(\hat{A}\rho_{01}-\rho_{01}\hat{A}^{\dagger})-\rho_{11} +\mathcal{O}(\epsilon)
\end{align}
Looking at the the equation for $\rho_{10}$, we can see that the first term involving $\rho_{00}$ is time dependant, making an exact solution difficult, but the time rate of change can be seen to be slow, on the order of $\epsilon$, whereas the second term, which acts as a damping term, is of order unity. Since the effective driving term is slowly changing and the damping term is relatively much stronger, we can impose an adiabatic elimination and take $\rho_{10}$ to be in it's steady state. The same argument would apply to $\rho_{11}$, yielding:
\begin{align}
\rho_{10} & = -2i\hat{A}\rho_{00} + \mathcal{O}(\epsilon) \\\nonumber
\rho_{11} & = -i(\hat{A}\rho_{01}-\rho_{10}\hat{A}^{\dagger}) + \mathcal{O}(\epsilon)
= 4\hat{A}\rho_{00}\hat{A}^{\dagger} + \mathcal{O}(\epsilon)
\end{align}
substituting the above expressions for $\rho_{00}$,$\rho_{10}$ and$\rho_{01}$ into eq.~B25 yields the following master equation:
\begin{align}
\label{eq:rho_full_supp}
\frac{d}{dt}\rho = & -i[\hat{H}_{red},\rho] + \kappa_a D[\hat{a}]\rho +
\kappa_b D[\hat{b}]\rho \\\nonumber
& + D[\frac{2 g_{ab}^{*}}{\sqrt{\kappa_{r}}}\hat{a}\hat{b} - \frac{2K_{ar}r_0}{\sqrt{\kappa_r}}\hat{a}^{\dagger}\hat{a} - \frac{2K_{br}r_0}{\sqrt{\kappa_r}}\hat{b}^{\dagger}\hat{b}]\rho
\end{align}
where
\begin{align}
\label{eq:Hred}
\hat{H}_{red} = & -\frac{K_{aa}}{2}(\hat{a}^{\dagger}\hat{a})^{2}-\frac{K_{bb}}{2}(\hat{b}^{\dagger}\hat{b})^{2} \\\nonumber
& - K_{ab}\hat{a}^{\dagger}\hat{a}\hat{b}^{\dagger}\hat{b} + \epsilon_{ab}\hat{a}^{\dagger}\hat{b}^{\dagger} + \epsilon_{ab}^*\hat{a}\hat{b}
\end{align}
\begin{align}
\epsilon_{ab} = g_{ab}r_{0}, \;\; \kappa_{ab} = \frac{4|g_{ab}|^{2}}{\kappa_r}
\end{align}
\subsection{Ideal steady state PCS without reservoir cross Kerr}
In this section we show that the ideal pair-photon driven dissipation process, in the absence of spurious reservoir cross Kerr and single photon loss, stabilizes a pair-coherent state as discussed in Table I.
Starting from Eq.~(\ref{eq:rho_full_supp}) and Eq.~(\ref{eq:Hred}), this ideal scenario results in equation of motion:
\begin{equation}
\Dot{\rho} = -i(\hat{H}_{ideal}\rho - \rho \hat{H}_{ideal}^\dagger) + \kappa_{ab} \hat{a}\hat{b} \rho \hat{a}^\dagger\hat{b}^\dagger
\end{equation}
\begin{equation}
\hat{H}_{ideal} = \epsilon_{ab}^* \hat{a}\hat{b} + \epsilon_{ab} \hat{a}^\dagger\hat{b}^\dagger + K_\mathrm{eff}\hat{a}^\dagger\hat{b}^\dagger \hat{a}\hat{b} - i \frac{\kappa_{ab}}{2}\hat{a}^\dagger\hat{b}^\dagger \hat{a}\hat{b}
\end{equation}
If the system is in a pair coherent state $\ket{\gamma, \delta}$, we can write the instantaneous density matrix time evolution as:
\begin{align}
\Dot{\rho} =&(\epsilon_{ab}^* \gamma - \epsilon_{ab} \gamma^* + i \kappa_{ab} |\gamma|^2)\rho \\\nonumber &+ (\epsilon_{ab} + K_\mathrm{eff}\gamma - i\frac{\kappa_{ab}}{2} \gamma)\hat{a}^\dagger\hat{b}^\dagger \rho \\\nonumber &- (\epsilon_{ab}^* + K_\mathrm{eff}\gamma^* + i\frac{\kappa_{ab}}{2} \gamma^*) \rho \hat{a}\hat{b}
\end{align}
Setting the terms that distort $\rho$ to 0 to give a steady-state solution and the following expression for the PCS state size $\gamma$:
\begin{equation}
\gamma = \frac{\epsilon_{ab}}{\frac{i\kappa_{ab}}{2}-K_\mathrm{eff}}
\end{equation}
\section{Numerical simulation methods}\label{Simulations}
We numerically simulate the PCS stabilization with two-photon pumping and dissipation that compares well with the experiment. The data shown in Fig.~\ref{fig:PC_2}(d) is a measurement of storage cavity population while pumping the PCS for variable time and the inset is measuring the population when only the FWM drive is on causing two-photon dissipation starting in a PCS. The Hamiltonian used in the simulation to fit the two-photon pumping and dissipation is unitary transformed so that the PCS is stationary in the drive frame. The reservoir mode is adiabatically eliminated resulting in a reduced Hamiltonian used in master equation Eq.~\ref{eq:rho_full} in the main text. We rewrite the master equation with the single photon loss operators for both the storage cavities,
\begin{align}
\partial_t \rho = -&i [\hat{H}_{sk} + (\epsilon_{ab} \hat{a}^{\dagger}\hat{b}^{\dagger} + h.c.), \rho] \nonumber\\ &+ \mathcal{D} [\sqrt{\kappa_{ab}}\hat{a}\hat{b}+\zeta_{a}\hat{a}^{\dagger}\hat{a} +\zeta_{b}\hat{b}^{\dagger}\hat{b}]\rho \nonumber\\ &+
\mathcal{D} [\sqrt{\kappa_{a}}\hat{a}]\rho +
\mathcal{D}[\sqrt{\kappa_{b}}\hat{b}]\rho
\label{eq:SimME}
\end{align}
where $\hat{H}_{sk}$ is defined in Eq.~(\ref{eq:H_SK}) and, $$\zeta_{a}=-\frac{2K_{ar}r_{0}}{\sqrt{\kappa_{r}}}, \zeta_{b}=-\frac{2K_{br}r_{0}}{\sqrt{\kappa_{r}}}.$$ The storage cavity single photon loss rates, $\kappa_{a}=1/T_{1a}$, $\kappa_{b}=1/T_{1b}$, Kerr and cross-Kerr values measured during pumping are listed in Table. \ref{table:parameters}. The complex parameter $r_{0}=\frac{2\epsilon_{d}}{i\kappa_{r}-2\Delta_{d}}$, first defined in Eq.~(\ref{eq:r0_def}), is the amplitude of the reservoir displacement. We displace the reservoir with $\epsilon_{d}/2\pi=780$ kHz at a detuning $\Delta_{d}/2\pi=-185$ kHz which with a $\kappa_{r}/2\pi=780$ kHz results in a $|r_{0}|=1.8$.
\begin{figure*}[tbp]
\centering
\includegraphics[scale=0.75]{fig8_09202022.pdf}
\caption{Density Matrix from experiment vs simulation. (a) The experimental density matrix measured using the quantum subspace tomography method discuss in the main text for $d=2$ after pumping for 15 $\mu$s and a wait time of $t_{w}=4.5$ $\mu$s. (b) The simulated density matrix with the system parameters and rates extracted from the time domain measurement fittings as in Fig.~\ref{fig:PC_2}. This is also after a 15 $\mu$s of pumping and a wait time of 4.5 $\mu$s.
}\label{fig:density_mat}
\end{figure*}
Starting in a PCS $\ket{\gamma, \delta=0}$ as shown in Fig.~\ref{fig:PC_2}(c) we measure the population of states at different times $t_{w}$ for which only the FWM tone is on (Fig.~\ref{fig:PC_2}(d inset)). In the absence of the reservoir drive the FWM drive should only cause two-photon dissipation $\mathcal{D}[\sqrt{\kappa_{ab}}\hat{a}\hat{b}]$. Using Eq.~(\ref{eq:SimME}) with $\kappa_{ab}$ as the only fit parameter and no drive term $\epsilon_{ab}=0$ the time scale for two-photon dissipation can be extracted. The dephasing rates $\zeta_{a}=\zeta_{b}=0$ since there are no photons in the reservoir and thus no dephasing caused by the storage-reservoir cross Kerr $K_{ar/br}$. In the second part we fix the two photon-dissipation $\kappa_{ab}/2\pi = 12.5$ kHz and vary $\epsilon_{ab}$. We estimate the dephasing rates from cavity $a$ and $b$, $|\zeta_{a}|^{2}/2=2\pi\times0.4$ kHz, $|\zeta_{b}|^{2}/2=2\pi\times28$ kHz respectively. We find the two-photon pumping rate to be $\epsilon_{ab}/2\pi=99$ kHz. The oscillations in the pumping time domain data are the result of the weak dissipation rate compared to the $K_\mathrm{eff}$. Simulation shows that with $K_\mathrm{eff}\gg\kappa_{ab}$, the damping of the oscillation from two-photon dissipation alone should be slower than we see in the experiment and the state is predicted by the equation for $\gamma$ as in Table.~\ref{table:PC_principle}. The state is smaller with more PCS like distribution. However, the modified dissipation operator to include dephasing effects of the cross-Kerr as derived in Appendix section~\ref{Derivation} produces faster damping, mimicking the experimental PCS distribution and size.
Using the equation, $\kappa_{ab}/2\pi = 4|g_{ab}|^2/\kappa_{r}=12.5$ kHz, we find $g_{ab}/2\pi=50$ kHz. The Stark shift we measure on the qubit due to the strong cw tone is $\sim4.3$ MHz, that makes the fourth order FWM term in the cosine expansion $g_{ab}/2\pi=60$ kHz. This value of $g_{ab}$ estimates the $\epsilon_{ab}/2\pi=120$ kHz and $\kappa_{ab}/2\pi=18.4$ kHz which is in agreement with the simulation to a good approximation. With the extracted two-photon dissipation and pumping rates we simulate the coherence and phases over time in a subspace of the state Fig.~\ref{fig:coh_over_time}(b,c).
\section{General Protocol of subspace tomography}\label{Protocol}
Using a three-level ancilla, we may use the following protocol for quantum subspace tomography:
\begin{figure}[tbp]
\centering
\includegraphics[scale=.37]{fig9_subspace_tomo.pdf}
\caption{General protocol of subspace tomography. We are illustrating the density matrices above in a very loose manner where $\rho_{ss}$ and $\rho_{\bar{s}\bar{s}}$ are blocks of the total storage density matrix $\rho$. There are coherence elements between these two blocks that are not depicted because they do not impact the experiment at hand. The goal here is to isolate $RD\rho_{ss}D^{\dagger}R$ in the $\ket{f}$ subspace using the operations $V_{1},V_{2},V_{3}$ as shown in the image.
}
\end{figure}
\begin{enumerate}\addtocounter{enumi}{-1}
\item Start with the system-ancilla state $\rho \otimes \ket{g}\bra{g}$.
\item Perform the unitary evolution $\hat{V}_1=\exp\big[ -i\frac{\pi}{2}\hat{P}\otimes (\ket{e}\bra{g}+\ket{g}\bra{e}) \big]$, which transforms the full density matrix into $\rho_1=\hat{V}_1(\rho\otimes\ket{g}\bra{g})\hat{V}_1^\dagger=\rho_{SS}\otimes\ket{e}\bra{e}+(...)$, where the terms (...) do not contribute to $\ket{e}$ state population or later measurement.
\item Perform the unitary evolution that preserves the ancilla basis states, $\hat{V}_2=\sum_{j=g,e,f}\hat{U}_j\otimes\ket{j}\bra{j}$. For a bosonic system, we may just apply displacement operation $\hat{U}_e=\hat{D}$, which transforms the density matrix into $\rho_2=\hat{V}_2\rho_1\hat{V}_2^\dagger=\hat{D}\rho_{SS}\hat{D}^\dagger\otimes\ket{e}\bra{e}+(...)$.
\item Choose the readout projector operator $\hat{R}$ (e.g.~photon number state projection $\ket{nn}\bra{nn}$ and perform the unitary evolution $\hat{V}_3 = \exp\big[ -i\frac{\pi}{2}\hat{R}\otimes (\ket{f}\bra{e}+\ket{e}\bra{f}) \big]$. The density matrix becomes $\rho_3=\hat{V}_3\rho_2\hat{V}_3^\dagger=\hat{R}\hat{D}\rho_{SS}\hat{D}^\dagger\hat{R}\otimes\ket{f}\bra{f}+(...)$.
\item Measure the probability in $\ket{f}$ state, Tr$[\rho_3\ket{f}\bra{f}]=\textrm{Tr}[\hat{R}\hat{D}\rho_{SS}\hat{D}^\dagger\hat{R}]$, which is sufficient for subspace tomography.
\end{enumerate}
In principle, a measurement of the ancilla after Step 2 (post select on $\ket{e}$) can be used to physically project the cavity state to $\rho_\mathcal{SS}$ for subsequent subspace tomography. However, this requires highly ideal measurement properties including high single-shot fidelity, quantum non-demolition on the ancilla, and no spurious back-action on the cavity system. In comparison, in the protocol above, the only requirement on the measurement is some degree of distinguishing ability between $\ket{e}$ and $\ket{f}$. The measurement outcome
can be scaled from calibrated readout contrast between $\ket{e}$ and $\ket{f}$, and any spurious readout signal from $\ket{g}$ only contribute to a background independent of $\hat{D}$ and $\hat{R}$.
\section{Frequency matching in subspace tomography}\label{Freq-matching}
In a typical Ramsey-type experiment, one considers the rotating frame of a qubit or a cavity set by its first excitation pulse, and the phases of its subsequent control pulses (usally at the same frequency) at any later time can be defined relative to the first pulse and programmed in the same rotating frame. In our subspace tomography protocol, because individual pair-wise coherence measurements involve ancilla rotation pulses of different frequencies, it is non-trivial to ensure that: 1) The subspace tomography pulse sequence carried out at any time ($t$ or $t_w$) informs the superposition phase of cavity states consistently in a pre-defined frame, and 2) The superposition phase of different Fock pairs are extracted consistently in the same rotating frame and hence can be combined in the same density matrix.
In our experiment, we enforce a ``closed-loop" frequency-matching condition when choosing the frequency of the cavity displacement pulses ($\omega_3$ and $\omega_4$) in each of the 2d subspace tomography measurements,
\begin{equation}
\omega_3+\omega_4 = \omega_p+\omega_d - (\omega_1-\omega_2)
\end{equation}
where $\omega_p$ and $\omega_d$ are the stabilization drive frequencies. $\omega_1, \omega_2$ are the frequencies of the photon-number selective ancilla pulses and they must precisely match the dispersively-shifted ancilla frequencies, with the dispersive shift to be referred as $-\chi_{jk}$ for the Fock state $\ket{jk}$ of the cavities. The closed loop in measuring the $\ket{11}$-$\ket{22}$ interference as an example is illustrated in Appendix Fig.~\ref{fig:phase_loop}. While one could in principle account for accumulated phases from the relative timing and relative frequency difference of individual pulses without this condition, this strategy allows us conveniently measure the two-cavity state in the ``pump frame" defined by the two stabilization drives self-consistently. It is important to note that the main reason this equality is important to experimentally satisfy is that the dispersive shifts on the qubit due to different number states ($\chi_{jk}$) are not equal to the corresponding multiples of $\chi_{qa}$ and $\chi_{qb}$, i.e.~$\chi_{jk}\neq j\chi_{qa} + k\chi_{qb}$, due to higher 6th order terms in the cosine expansion of our hamiltonian that have been neglected. Due to these neglected higher order terms, $\omega_{3}$ and $\omega_{4}$ must be determined from the experimentally measured $\chi_{jk}$ values for each pair of number states to satisfy the equality.
\begin{figure}[tbp]
\centering
\includegraphics[scale=0.45]{fig_9_v2.pdf}
\caption{Visual depiction of the pulse sequence used for the subspace tomography to determine the $\rho_{11,22}$ element while tracking the $\ket{12}$ state as a specific example.
The frequencies of the four pulses applied in this tomographic analysis and the frequencies of the original cavity stabilization drives satisfy the ``closed-loop" condition for phase locking: $\omega_1-\omega_2=\omega_p+\omega_d-\omega_{3}-\omega_{4}$ which is the same effective equality as in Eq.~(\ref{eq:simple phase match}). Note that $\omega_1-\omega_2 = \chi_{22}-\chi_{11}\approx\chi_{qa}+\chi_{qb}$ but is not exactly equal because of the 6th-order dispersive shifts.
\label{fig:phase_loop}
}
\end{figure}
In the following, we show how our subspace tomography protocol faithfully extracts the superposition phase $\Phi_{jj}-\Phi_{kk}$ between two Fock components $\ket{jj}$ and $\ket{kk}$ of the cavity state as defined in the pump frame. (The analysis is easily extensible to $\delta\neq0$ states.) In this analysis, we work in a rotating frame where the transmon rotates with frequency $\omega_q$ and the two cavities rotate by a combined frequency $\omega_p+\omega_d$. (More specifically, Cavity $a$ rotates by $\omega_a-\Delta_{sd}/2$ and Cavity $b$ rotates by $\omega_b-\Delta_{sd}/2$ where $\Delta_{sd}=(\omega_a+\omega_b)-(\omega_p+\omega_d)$). In this rotating frame, the frequencies of the four tomography tones, $\omega_1, \omega_2, \omega_3, \omega_4$ are on the order of the dispersive shifts, and the eigen-frequencies of the joint Fock states $\ket{jj}$ are:
\begin{align}
\omega_{jjg} = j\Delta_{sd} - K_{ab}j^{2} - (K_{aa}+K_{bb})j(j-1)/2
\end{align}
\begin{align}
\omega_{jje} = j\Delta_{sd} - K_{ab}j^{2} - (K_{aa}+K_{bb})j(j-1)/2 - \chi_{jj}
\end{align}
for the transmon in $\ket{g}$ and $\ket{e}$, respectively.
We now track the phase accumulation on the $\ket{jj}$ component of the cavity state over the 3 time steps of a subspace tomography, first, the wait time $(t_{w}$ a few $\mu$s after the PCS is made, next, the duration of our selective qubit pulse $(\delta t_{q}=2$ $\mu$s) that excites it to be entangled with $\ket{e}$ and finally, the duration of our displacement cavity pulses ($\delta t_{d}=24$ ns) to create interference of $\ket{jje}$ versus a different cavity state. The short duration and therefore the broad selectivity of our displacement cavity pulses gives us the freedom to vary their frequencies to satisfy this closed loop condition without needing to worry about their frequencies being too far from the desired cavity states to be displaced.
During the first wait time period ($t_{w}$) the qubit remains in $\ket{g}$ thus accumulating a phase $\omega_{jjg}t_{w}$. During the qubit pulse step ($\delta t_{q}$) the generator for the qubit pulse is rotating in a different frame with a frequency offset of $-\chi_{jj}$ thus leading to a phase offset of $-\chi_{jj} (t_{w}+\delta t_{q})$ from the phase imparted by the generator since the generator would have been accumulating this phase through both time periods $t_{w}$ and $\delta t_{q}$. We can use the same time steps here for different generators in different frames because they all share a common reference of how t=0 is defined, where any reference point of t=0 is valid provided it is consistent among all generators. For our final time period we have a state phase accumulation from the state with the qubit in $\ket{e}$ giving $\omega_{jje}\delta t_{d}$ in addition to the phase imparted by our displacement pulse. The displacement pulses are in different rotating frames with a combined frequency of $\omega_{3}+\omega_{4}$. Taking the state that we are interfering with as $\ket{kk}$, this will impart a phase of $(\omega_{3}+\omega_{4})|k-j|(t_{w} + \delta t_{q} + \delta t_{d})$, where the $|k-j|$ factor comes from the fact that the phase gets imparted for each excitation traded. With this, we can subtract the accumulated phases between $\ket{jj}$ and $\ket{kk}$ over the total time $t_{tot} = t_{w} + \delta t_{q} + \delta t_{d}$ where we are tracking state $\ket{jj}$ so that is the only one that will recieve the $\omega_{3}+\omega_{4}$ term. The chosen state to track here will not matter so this is still a general treatment, with the phase subtraction being:
\begin{align} \label{eq:phase match}
&\Phi_{jj} - \Phi_{kk} + (\omega_{jje} - \omega_{kke} + (\omega_{3}+\omega_{4})|k-j|)t_{tot}\nonumber \\
= &\Phi_{jj} - \Phi_{kk} + (\omega_{jjg} - \omega_{kkg} + \nonumber\\&(\chi_{kk}- \chi_{jj} + |k-j|(\omega_{3}+\omega_{4}))t_{tot}
\end{align}
It is clear that we can simply calibrate our $\omega_{3},\omega_{4}$ for each set of interference measurements on states $\ket{jj},\ket{kk}$ to satisfy the equality:
\begin{align} \label{eq:simple phase match}
\chi_{jj}- \chi_{kk}=|k-j|(\omega_{3}+\omega_{4})
\end{align}
so we can neglect a large part of Eq.~(\ref{eq:phase match}), leaving only $\Phi_{jj} - \Phi_{kk} + (\omega_{jjg} - \omega_{kkg})t_{tot}$. It is important to note that this cancellation we have imparted gets rid of any terms that could have contributions from higher order neglected terms mentioned earlier and also allows an arbitrary choice of a consistent $t=0$ definition across generators.
This leaves us with the simple linear phase accumulation terms $\omega_{jjg},\omega_{kkg}$ that can be simply calculated and subtracted away, leaving the desired phase difference of $\Phi_{jj}-\Phi_{kk}$. This linear phase accumulation from $\omega_{jjg}-\omega_{kkg}$ can be seen for different states in Fig.~\ref{fig:coh_over_time}(a) at different $t_{w}$ values, allowing one to linearly extrapolate the original phase from the $t_{tot}$ times.
One can think of this method in a more direct intuitive way by looking at the case of isolating $\ket{11}$ and $\ket{22}$ to determine the coherence element $\rho_{11,22}$ and seeing that on top of the desired $\Phi_{22}-\Phi_{11}$ and $(\omega_{22g}-\omega_{11g})t_{tot}$ phase contributions there will be a constant phase accumulation over time with frequency of $-\chi_{22}$ and $-\chi_{11}$ on the $\ket{11}$ and $\ket{22}$ states, respectively. This is because while the cavity state $\ket{11}(\ket{22})$ is entangled with the qubit in $\ket{g}$ the generator that will send the pulse out to flip the state is rotating with a frequency of $-\chi_{11}(-\chi_{22})$, thus imparting the accumulated phase from this frequency difference and after the rotation, there will be the same frequency difference from the fact that the qubit is now in $\ket{e}$ so it will also rotate at a frequency of $-\chi_{11}(-\chi_{22})$ in this frame so in effect this frequency accumulation will always be present. As mentioned earlier, there will additionally be the phase imparted from the displacement pulse, so one simply needs to tune the $\omega_{3}, \omega_{4}$ values to satisfy the above inequality in Eq.~(\ref{eq:simple phase match}) to cancel out the effects from the $-\chi_{11},-\chi_{22}$ terms, thus leaving only the desired phase difference of $\Phi_{22}-\Phi_{11}$ and linear phase accumulation from $\omega_{22g}-\omega_{11g}$ that can be easily subtracted away.
\end{appendix}
|
\section{Introduction}
In 2002, Fomin and Zelevinsky initiated the study of cluster algebras as a framework for understanding total positivity in semisimple groups \cite{FZ1}.
In the following year, they classified cluster algebras of finite type by finite Coxeter groups via Cartan matrices \cite{FZ2}, and they showed that the number of clusters is given by the Coxeter-Catalan number \cite{FZ_y}.
Moreover, they constructed a simplicial complex, called the \emph{cluster complex}, consisting of ``compatible'' almost positive roots of the corresponding Cartan type.
The original enumeration was type-by-type, and an open problem was to provide uniform bijections to other Coxeter-Catalan objects.
One well-studied candidate was the $W$-noncrossing partition lattice, which we define below.
Let $W$ be a finite Coxeter group, and denote the set of reflections in $W$ by $T$.
We note that, in particular, $T$ contains the set of simple reflections of $W$, and thus it generates $W$.
As a consequence, we can represent each element $w\in W$ as word $t_1t_2 \ldots t_k$ in $T$.
We say that $t_1t_2 \ldots t_k$ is \emph{reduced} if there does not exist another expression $t'_1t'_2\ldots t'_j$ for $w$ in terms of $T$ with $j<k$.
We define the \emph{absolute order} of $W$ by $x \le y$ if and only if any reduced $T$-word for $x$ occurs as a subword for some reduced $T$-word for $y$.
The $W$-noncrossing partitions lattice, denoted $\mathrm{NC(W,c)}$, is the maximal interval $[e, c]$ from the identity element $e$ to a Coxeter element $c$.
See \cite{Armstrong} for more detailed background.
The classical noncrossing partition lattice (of type A) was studied as poset in 1972 by Kreweras \cite{kreweras}, who also introduced a certain anti-isomorphism of noncrossing partitions called the \emph{Kreweras Complement}.
However, again, there was no uniform (type-free) proof that the $W$-noncrossing partition lattices were in fact lattice-posets.
In 2007, Brady and Watt provided a uniform proof of this fact, and also constructed the cluster complex in terms of noncrossing partitions \cite{BradyWatt}.
Building on this construction, Athanasiadis, Brady, and Watt gave a uniform EL-labeling of the noncrossing partition lattice in terms of the reflections of $W$ \cite{ABW}.
Independently, Reading introduced a new Coxeter-Catalan family---$c$-Cambrian lattices \cite{cambrian}.
Each $c$-Cambrian lattice of type $W$ is a lattice quotient of the (right) weak order on $W$, and can be realized as a subposet of $c$-sortable elements.
Moreover, in \cite{cambrian2}, Reading constructed two uniformly defined bijections from the set of $c$-sortable elements (of type $W$): one to the facets of the cluster complex; and the other to the noncrossing partition lattice $\mathrm{NC(W,c)}$.
The latter bijection is essentially encoded by a certain minimal join-representation called the \emph{canonical join representation} (see Definition~\ref{def:canonical_join_rep}.
The elements of a canonical join representation are called \emph{join-irreducible} (see Definition~\ref{def: irreducible}).
The main goal of this paper is to extend both the EL-labeling of Athanasiadis, Brady, and Watt and to study the canonical join representations that are central to Reading's bijections in the context of representation theory.
Throughout our paper we take $\Lambda$ to be a finite dimensional algebra over an arbitrary field $K$.
When $\Lambda$ is hereditary of Dynkin type $W$, Ingalls and Thomas showed that the noncrossing partition lattice $\mathrm{NC(W,c)}$ is isomorphic to the lattice of wide subcategories of $\mathsf{mod} \Lambda$ \cite{IT}.
This was later generalized to all hereditary artin algebras in \cite{IS}.
In this context, chains in the lattice of wide subcategories are in bijection with \emph{exceptional sequences}. Using the \emph{$\tau$-exceptional sequences} of \cite{BM_exceptional}, this bijection was generalized to non-hereditary algebras \cite{BM_wide,BuH}.
Our first main result uses these bijections to extend the EL-labeling given by Athanasiadis, Brady, and Watt.
\begin{thmIntro}[Theorem~\ref{thm:mainA}]\label{thm:intro:mainA}
Let $\Lambda$ be a representation-directed algebra. Then there exists a partial order on the bricks in $\mathsf{mod}\Lambda$ which makes the labeling of $\mathsf{wide}(\mathsf{mod}\Lambda)$ from $\tau$-exceptional sequences into an EL-labeling.
\end{thmIntro}
We note that Athanasiadis, Brady, and Watt's EL-labeling comes from a linear ordering of the reflections $T$ called a \emph{reflection order}.
When $\Lambda$ is finite type hereditary, the set of reflections of $W$ are in bijection with the set of bricks in $\mathsf{mod}\Lambda$.
Thus we obtain the original EL-labeling of $\mathrm{NC(W,c)}$ as a special case of our theorem.
Reading's constructions were also translated into the language of quiver representations by Ingalls and Thomas in \cite{IT}, who showed that if $\Lambda$ is representation-finite hereditary, then the lattice of torsion classes for $\mathsf{mod} \Lambda$ is isomorphic to a $c$-Cambrian lattice.
More recently, canonical join-representations in type A were modeled by certain noncrossing arc diagrams \cite{reading_arcs}, from which it was observed that a set of join-irreducible elements is a canonical join representation if and only if each pair of join-irreducible elements if a canonical join representation.
We call this the \emph{flag property}.
In \cite{emily_canonical}, it was shown that a large class of finite lattices called \emph{semidistributive lattices} have the flag property (and indeed such lattices are characterized by this property).
The class of finite semidistributive lattices includes Reading's $c$-Cambrian lattices as well as the lattices of torsion classes of $\tau$-tilting finite algebras (see Section~\ref{sec:lattice}). The flag property was also shown to hold for (possibly infinite) lattices of torsion classes of arbitrary finite-dimensional algebras in \cite{BCZ}. This was done by establishing a correspondence between canonical join representations and collections of hom-orthogonal bricks, or \emph{semibricks}, in $\mathsf{mod}\Lambda$.
We now extend this result to a larger class of infinite semidistributive lattices.
Recall that an element $j\in L$ is \emph{completely join-irreducible} provided that $j$ covers precisely one element in $L$.
\begin{thmIntro}[Corollary~\ref{cor:flag}]\label{thm:intro:mainB}
Let $L$ be a well-separated completely semidistributive lattice.
Then a collection of completely join-irreducible elements is a canonical join representation if and only if each pair of elements is a canonical join representation.
\end{thmIntro}
Completely semidistributive lattices come equipped with a pair of inverse bijections, which we call $\kappa$ and $\kappa^d$, from the set of completely join-irreducible elements of $L$ to the set of completely meet irreducible elements of $L$:
\begin{eqnarray*}
\kappa(j) &=& \mathrm{m\text{-}label}[j_*,j] = \max\{y \in L \mid j \wedge y = j_*\}\\
\kappa^d(m) &=& \mathrm{j\text{-}label}[m,m^*] = \min\{y \in L \mid m \vee y = m^*\}.
\end{eqnarray*}
(See Section~\ref{sec:lattice} for a more detailed explanation of these formulas.) We show that $\kappa$ gives a criterion for testing when join-irreducible elements are ``compatible'' in the sense that they are a canonical join representation. In particular, this criterion implies the flag property in Theorem~\ref{thm:intro:mainB}.
\begin{thmIntro}[Theorem~\ref{thm:covers_canonical}, simplified]\label{thm:intro:mainC}
Let $L$ be a well-separated completely semidistributive lattice.
For any set $A$ of completely join-irreducible elements of $L$, the following are equivalent.
\begin{enumerate}
\item $\Join A$ is a canonical join representation.
\item For all $i \neq j \in A$ one has $i \leq \kappa(j)$.
\end{enumerate}
\end{thmIntro}
The maps $\kappa$ and $\kappa^d$ have recently played an important role in dynamical algebraic combinatorics and in Coxeter-Catalan combinatorics.
See e.g. \cite[Section~1.2]{DW} and the refences therein.
In \cite{BTZ}, explicit formulas for $\kappa$ are given in the case where $L$ is a lattice of torsion classes.
More recently, Enomoto used $\kappa$ to define a new partial order on the elements of a completely semidistributive lattice $L$, which he calls the $\kappa$-order \cite{enomoto}. (This order is also defined for a generalization of finite semidistributive lattices in \cite{DW}.) Enomoto further showed that if $L$ is the lattice of torsion classes of an abelian length category, then the $\kappa$-order is isomorphic to the lattice of wide subcategories.
For our last main result, we use the $\kappa$-map to give a combinatorialization of the $\tau$-exceptional sequences of $\tau$-tilting finite. We first explain how the ``extended $\kappa$-map'' $\overline{\kappa}^d$ is related to the Auslander-Reiten translation in Section~\ref{sec:combinatorial_tau_tilting}. In particular, we show that the $\tau$-tilting reduction of Jasso \cite{jasso} (see Section~\ref{sec:tau_exceptional}), and moreover the ``wide intervals'' of \cite{AP}, are related to the operator $\overline{\kappa^d}$. This motivates the definition of a \emph{$\kappa^d$-exceptional sequence}, given in Definition~\ref{def:kappa_exceptional} of Section~\ref{sec:kappa_exceptional}. Roughly speaking, while a $\tau$-exceptional sequence makes use of $\tau$-rigid modules to interatively move to smaller wide subcategories of a module category, a $\kappa^d$-exceptional sequence uses completely
join-irreducible elements to iteratively move to smaller ``nuclear intervals'' (Definition~\ref{def:nuclear}) of a finite semidistributive lattice. Our last main result shows that, up to the brick-$\tau$-rigid correspondence of \cite{DIJ}, the sets of $\kappa^d$-exceptional sequences and $\tau$-exceptional sequences coincide for $\tau$-tilting finite algebras.
\begin{thmIntro}[Theorem~\ref{thm:mainD}]\label{thm:intro:mainD}
Let $\Lambda$ be a $\tau$-tilting finite algebra. Then there is a bijection between the set of $\kappa^d$-exceptional sequences in the lattice of torsion classes $\mathsf{tors}\Lambda$ and the set of $\tau$-exceptional sequences in $\mathsf{mod}\Lambda$. In particular, up to the correspondence between indecomposable $\tau$-rigid modules and completely join-irreducible torsion classes, the $\tau$-exceptional sequences of $\mathsf{mod}\Lambda$ are completely determined by the lattice of torsion classes.
\end{thmIntro}
We conclude the paper with detailed examples of $\kappa^d$-exceptional sequences and a brief section on our planned future work.
\subsection*{Acknowledgements}
EJH is grateful to Erlend D. B{\o}rve, Aslak Bakke Buan, H{\aa}vard Utne Terland, and Hugh Thomas for several insightful conversations related to this work. The authors are also thankful to Haruhisa Enomoto for helpful discussions.
\section{Semidistributive lattices and the kappa map}\label{sec:lattice}
In this section we review the fundamental definitions for posets and lattice-posets.
A key concept we use throughout is the notion of an \emph{edge-labeling} for a poset $P$.
Recall that an element $y$ \emph{covers} $x$ in $P$ if and only if $y>x$ and there does not exist $z\in P$ with $y>z>x$.
We write $y{\,\,\,\cdot\!\!\!\! >\,\,} x$, and we also say that $x$ is \emph{covered by} $y$, and the pair form a \emph{cover relation}. We use the \emph{Hasse quiver} of $P$ as a visual representation. This is the directed graph $\mathrm{Hasse}(P)$ whose vertices correspond to elements of $P$ such that there is an arrow $x \rightarrow y$ whenever $y {\,\,<\!\!\!\!\cdot\,\,\,} x$. We will use the poset whose Hasse quiver is shown in Figure~\ref{fig:run_ex} as a running example. The elements of this poset are drawn as circled nodes. The labeling of each arrow in the Hasse quiver with an elements $j_i$ is an example of the following, where the partial order on $\{j_1,j_2,j_3,j_4\}$ is inherited from $P$.
\begin{figure}
\begin{tikzpicture}
\begin{scope}[decoration={
markings,
mark=at position 0.65 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-3,4) node [midway,above left] {$j_1$};
\draw[postaction=decorate,dashed](0,6) -- (0,4) node [midway,left] {$j_2$};
\draw[postaction=decorate](0,6) -- (3,4) node [midway,above right] {$j_3$};
\draw[postaction=decorate](-3,4) -- (0,2) node [near start,right] {$j_3$};
\draw[postaction=decorate,dashed](0,4) -- (-3,2) node [near start,left] {$j_1$};
\draw[postaction=decorate](3,4) -- (0,2) node [near start, left] {$j_1$};
\draw[postaction=decorate,dashed](0,4) -- (3,2) node [near start,right] {$j_3$};
\draw[postaction=decorate](-3,4) -- (-3,2) node [midway, left] {$j_2$};
\draw[postaction=decorate](3,4) -- (3,2) node [midway, right] {$j_2$};
\draw[postaction=decorate](-3,2) -- (-1.5,1) node [midway,left] {$j_4$};
\draw[postaction=decorate](-1.5,1) -- (0,0) node [midway,left] {$j_3$};
\draw[postaction=decorate](0,2) -- (0,0) node [midway,left] {$j_2$};
\draw[postaction=decorate](3,2) -- (0,0) node [midway,below right] {$j_1$};
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node[draw,circle,fill=white] at (-3,2) {$j_4$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (3,2) {$j_1$};
\node[draw,circle,fill=white] at (-1.5,1) {$j_3$};
\node[draw,circle,fill=white] at (3,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-3,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\end{tikzpicture}
\caption{Our main running example of a lattice poset.}\label{fig:run_ex}
\end{figure}
\begin{figure}
\begin{tikzpicture}
\begin{scope}[decoration={
markings,
mark=at position 0.5 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,16) -- (-4,8) node [midway,above left] {$(1,-1)$};
\draw[postaction=decorate] (-4,8) -- (0,0) node [midway,below left] {$s$};
\draw[postaction=decorate] (4,1) -- (0,0) node [near end,below right] {$(1,-1)$};
\draw[postaction=decorate] (4,3) -- (4,1) node [midway,right] {$(2,-1)$};
\draw[postaction=decorate] (4,4) -- (4,3) node [near start,right] {$(3,-1)$};
\draw[postaction=decorate] (2,7) -- (4,6) node [midway,below left] {$\{\lambda\}$};
\draw[postaction=decorate] (6,7) -- (4,6) node [midway,below right] {$\{\mu\}$};
\draw[postaction=decorate] (4.7,6.7) -- (4,6);
\draw[postaction=decorate] (3.3,6.7) -- (4,6);
\draw[postaction=decorate] (4,10)--(2,9) node [midway,above left] {$\{\mu\}$};
\draw[postaction=decorate] (4,10) -- (6,9) node [midway,above right] {$\{\lambda\}$};
\draw[postaction=decorate] (4.4,9.6)--(4.7,9.2);
\draw[postaction=decorate] (3.6,9.6)--(3.4,9.2);
\draw[postaction=decorate] (4,12.8) -- (4,12) node [near end, right] {$(2,1)$};
\draw[postaction=decorate] (4,15) -- (4,13) node [midway,right] {$(1,1)$};
\draw[postaction=decorate] (0,16) -- (4,15) node [midway,above right] {$s$};
\end{scope}
\node[draw,fill=white] at (0,0) {$\hat{0}$};
\node[draw,fill=white] at (0,16) {$\hat{1}$};
\node[draw,fill=white] at (4,1) {$(1,-1)$};
\node[draw,fill=white] at (4,3) {$(2,-1)$};
\node[draw,fill=white] at (4,15) {$(1,1)$};
\node[draw,fill=white] at (4,13) {$(2,1)$};
\node[draw,fill=white] at (4,6) {$\emptyset$};
\node[draw,fill=white] at (2,7) {$\{\lambda\}$};
\node[draw,fill=white] at (6,7) {$\{\mu\}$};
\node[draw,fill=white] at (2,9) {$\mathbb{P}^1(K) \setminus\{\mu\}$};
\node[draw,fill=white] at (6,9) {$\mathbb{P}^1(K)
\setminus\{\lambda\}$};
\node[draw,fill=white] at (4,10) {$\mathbb{P}^1(K)
$};
\node[draw,fill=white] at (-4,8) {$s$};
\node at (4,5) {$\vdots$};
\node at (4,11.25) {$\vdots$};
\node at (4,8) {$\vdots$};
\node at (3,8) {$\vdots$};
\node at (2,8) {$\vdots$};
\node at (1,8) {$\vdots$};
\node at (5,8) {$\vdots$};
\node at (6,8) {$\vdots$};
\node at (7,8) {$\vdots$};
\node at (3.5,7) {$\cdots$};
\node at (4.5,7) {$\cdots$};
\node at (1,7) {$\cdots$};
\node at (7,7) {$\cdots$};
\node at (4,9) {$\cdots$};
\node at (0.5,9) {$\cdots$};
\node at (7.5,9) {$\cdots$};
\node at (4,6.5) {$\cdots$};
\node at (4,9.5) {$\cdots$};
\end{tikzpicture}
\caption{The Hasse quiver of the lattice described in Example~\ref{ex:lattices}(2).}\label{fig:kronecker}
\end{figure}
\begin{definition}\label{def:edge-label}
Let $P$ be a poset and let $\mathcal{E}(P) = \{[x,y]:\,x{\,\,<\!\!\!\!\cdot\,\,\,} y \text{ in $P$}\}$.
An \emph{edge-labeling} of $P$ is a function $\mathrm{label}: \mathcal{E}(P) \to Q$, where $Q$ is a poset.
\end{definition}
Recall that a \emph{lattice-poset} $L = (L,\leq)$, or simply a lattice, is a poset satisfying the following two conditions:
First, each pair of elements $x$ and $y$ in $L$ has a unique smallest common upper bound called the \emph{join}, and denoted $x\vee y$ or $\Join \{x,y\}$;
second, each pair $x$ and $y$ in $L$ has a unique greatest common lower bound called the \emph{meet}, and denoted $x\wedge y$ or $\bigwedge \{x,y\}$.
Observe that for any finite subset $X\subseteq L$ the join $\Join X$ and the meet $\bigwedge X$ both exist. If an addition the join $\Join X$ and meet $\bigwedge X$ exist for \emph{any} subset $X \subseteq L$, then $L$ is called a \emph{complete} lattice. Unless otherwise stated, all of the lattices in this paper will be complete. In particular, each of our lattices has a unique smallest element $\hat{0}$ and a unique largest element $\hat{1}$.
We adopt the convention that $\Join \emptyset = \hat{0}$, and $\bigwedge \emptyset = \hat{1}$.
\begin{example}\label{ex:lattices}\
\begin{enumerate}
\item Our main running example of a lattice is the poset shown in Figure~\ref{fig:run_ex}. For example, in this poset we have $j_3 \vee j_2 = m_1$ and $j_3 \wedge j_2 = \hat{0}$. The labels on the Hasse arrows of Figure~\ref{fig:run_ex} are explained in Example~\ref{ex:label}.
\item Let $K$ be an algebraically closed field. We define a lattice $L_{kr}(K)$ as follows. As a set, we have $L_{kr}(K) = \mathbb{N} \times \{-1,1\} \cup 2^{\mathbb{P}^1(K)} \cup \{\hat{0},\hat{1},s\}$, where $\mathbb{P}^1(K) = K \cup \{\infty\}$ is the projective line over $K$ and $\hat{0}, \hat{1}$, and $s$ are formal symbols. The relation $\leq$ is defined as follows.
\begin{enumerate}
\item For all $x \in L_{kr}(K)$: $\hat{0} \leq x \leq \hat{1}$.
\item For all $n, m \in \mathbb{N}$: $(n, -1) \leq (m, -1)$ if and only if $n \leq m$ in the usual sense.
\item For all $n, m \in \mathbb{N}$: $(n, -1) \leq (m, -1)$ if and only if $n \geq m$ in the usual sense.
\item For all $S, S' \in 2^{\mathbb{P}^1(K)}$: $S \leq S'$ if and only if $S \subseteq S'$.
\item For all $n \in \mathbb{N}$ and $S \in 2^{\mathbb{P}^1(K)}$: $(n,-1) \leq S \leq (n,1)$.
\end{enumerate}
The Hasse diagram of $L_{kr}(K)$ is shown in Figure~\ref{fig:kronecker} (see also \cite[Example~1.3]{thomas_intro}). The symbols $\lambda$ and $\mu$ denote generic elements of $\mathbb{P}^1(K)$. The labels on the Hasse arrows are explained in Example~\ref{ex:label}. Note that the interval $[\emptyset,\mathbb{P}^1(K)]$ is isomorphic to the boolean lattice on $\mathbb{P}^1(K)$, and that there are no cover relations of the form $u {\,\,<\!\!\!\!\cdot\,\,\,} \emptyset$ or $\mathbb{P}^1(K) {\,\,<\!\!\!\!\cdot\,\,\,} u$ in $L_{kr}(K)$.
\item Two other examples of lattices which will feature in this paper are the lattices of torsion classes and of wide subcategories of a finite-dimensional algebra. In particular, $L_{kr}(K)$ is isomorphic to the lattice of torsion classes of the path algebra of the Kronecker quiver, see Example~\ref{ex:kronecker}. These lattices will be formally introduced in Section~\ref{sec:background}.
\end{enumerate}
\end{example}
As we discuss in Section~\ref{sec:torsion}, the lattice of torsion classes also has a property called \emph{complete semidistributivity}, in which certain elements called \emph{completely join-irreducible} and \emph{completely meet-irreducible} play a role analogous to prime numbers in Number theory.
\begin{definition}\label{def: irreducible}
An element $j$ in a lattice $L$ is \emph{join-irreducible} provided that whenever $j=\Join X$ for a finite subset $X\subseteq L$ we have $j\in X$.
We say that $j$ is \emph{completely join-irreducible} provided that for any subset $X\subseteq L$ one has that $j = \Join X$ if and only if $j \in X$.
Equivalently, $j$ is completely join-irreducible if and only if there is a unique element $j_*$ such that $j{\,\,\,\cdot\!\!\!\! >\,\,} j_*$.
The notions of meet-irreducible and completely meet-irreducible are defined dually, by substituting ``$\bigwedge$'' for ``$\Join$''.
An element $m$ is completely meet-irreducible if and only if there is a unique element $m^*$ such that $m{\,\,<\!\!\!\!\cdot\,\,\,} m^*$.
\end{definition}
\begin{notation}
We denote by $\cji(L)$ and $\cmi(L)$ the sets of completely join-irreducible and completely meet-irreducible elements, respectively.
\end{notation}
\begin{example}\label{ex:jirr}\
\begin{enumerate}
\item Note that the convention that $\Join \emptyset = \hat{0}$ means that $\hat{0}$ is not completely join-irreducible, even though it cannot be written as a join of strictly smaller elements. Similarly $\hat{1}$ is never completely meet irreducible.
\item Let $L$ be the lattice in Figure~\ref{fig:run_ex}. Then $$\cji(L) = \{j_1,j_2,j_3,j_4\},\qquad\qquad\cmi(L) = \{m_1,m_2,m_3,j_3\}.$$
\item Let $L_{kr}(K)$ be the lattice in Figure~\ref{fig:kronecker}. Then
\begin{eqnarray*}
\cji(L_{kr}(K)) &=& \{s\} \cup \mathbb{N} \times \{-1,1\} \cup \{\{\lambda\} \mid \lambda \in \mathbb{P}^1(K)\},\\
\cji(L_{kr}(K)) &=& \{s\} \cup \mathbb{N} \times \{-1,1\} \cup \{\mathbb{P}^1(K) \setminus \{\lambda\} \mid \lambda \in \mathbb{P}^1(K)\}.
\end{eqnarray*}
Moreover, the element $\emptyset$ (resp. $\mathbb{P}^1(K)$) is join-irreducible (resp. meet-irreducible), but not completely join-irreducible (resp. completely meet-irreducible). Indeed, we have $\emptyset = \bigvee\{(n,-1)\mid n \in \mathbb{N}\}$, but there is no finite subset $X \subseteq L_{kr}(K) \setminus\{\emptyset\}$ such that $\emptyset = \bigvee X$.
\end{enumerate}
\end{example}
A lattice $L$ is \emph{completely semidistributive} if for any subset $Y\subseteq L$, and elements $x$ and $w$ in $L$, both of the following implications are true. (Note that by $x \vee Y = w$ we mean $x \vee y = w$ for all $y \in Y$, and likewise for $x \wedge Y = w$.)
\begin{equation}\label{jsd}
\text{If $x\vee Y = w$, then $x\vee\left(\bigwedge Y\right) = w$}\tag{$SD_\vee$}
\end{equation}
\begin{equation}\label{msd}
\text{If $x\wedge Y = w$, then $x\wedge\left(\Join Y\right) = w$}\tag{$SD_\wedge$}
\end{equation}
Some important examples of completely semidistributive lattices are lattice of torsion classes of finite-dimensional algebras (see \cite[Theorem~4.5]{GM} and \cite[Theorem~1.3]{DIRRT}) and posets of regions of simplicial hyperplane arrangements (see e.g. \cite[Corollary~9-3.9]{reading_book}).
\begin{example}
The lattices from Figures~\ref{fig:run_ex} and~\ref{fig:kronecker} are completely semidistributive. For example, in the lattice from Figure~\ref{fig:run_ex}, we have $m_1 \vee \{m_2,m_3\} = \hat{1}$ and $m_1 \vee (m_2 \wedge m_3) = m_1 \vee j_1 = \hat{1}$.
\end{example}
As we discuss in Section~\ref{sec:bricks}, cover relations in the lattice of torsion classes have a natural edge-labeling by certain indecomposable modules called bricks.
This is actually a special case of a more general phenomenon which we recall now.
\begin{definition-theorem}\label{def:j-label}
Let $L$ be a completely semidistributive lattice and let $u {\,\,<\!\!\!\!\cdot\,\,\,} v$ be a cover relation in $L$. By \cite[Lemma~3.11]{RST}, the set $\{y \in L \mid y \vee u = v\}$ contains a minimum element $j$ which is completely join-irreducible and satisfies $j_* \leq u$. We call $j$ the \emph{join-irreducible label} of the cover relation $u {\,\,<\!\!\!\!\cdot\,\,\,} v$ and denote $j = \mathrm{j\text{-}label}[u,v]$.
\end{definition-theorem}
\begin{example}\label{ex:label}
The cover relations in Figures~\ref{fig:run_ex} and~\ref{fig:kronecker} are decorated by their join-irreducible labels.
\end{example}
\begin{remark}\label{rem:m-label}\
\begin{enumerate}
\item Note that $\cji(L)$ inherits a partial order from $L$. Thus the association $[u,v] \mapsto \mathrm{j\text{-}label}[u,v]$ can be seen as a special case of Definition~\ref{def:edge-label}.
\item One can also define a labeling $\mathrm{m\text{-}label}$ of cover relations by completely meet-irreducible elements using the dual of Definition-Theorem~\ref{def:j-label}. We note that $\mathrm{m\text{-}label}[u,v] = \mathrm{j\text{-}label}^d[v,u]$; that is the meet-irreducible label of the cover relation $[u,v]$ is the same as the join-irreducible label of the cover relation $[v,u]$ in the dual lattice $L^d$.
\end{enumerate}
\end{remark}
\begin{notation}
We extend Definition-Theorem~\ref{def:j-label} as follows. Given a relation $x \leq y \in L$, we denote
$$\mathrm{j\text{-}label}[x,y] = \left\{\mathrm{j\text{-}label}[u,v] \mid x \leq u {\,\,<\!\!\!\!\cdot\,\,\,} v \leq y\right\}.$$
That is, $\mathrm{j\text{-}label}[x,y]$ is the set of completely join-irreducible elements which label some cover relation in the interval $[x,y]$. Note that we have made a slight abuse of notation in that if the cover relation $u {\,\,<\!\!\!\!\cdot\,\,\,} v$ is labeled by $j$, then we use $\mathrm{j\text{-}label}[u,v]$ to denote both $j$ and $\{j\}$.
\end{notation}
Restricting to cover relations involving completely join- and meet-irreducible elements, Definition-Theorem~\ref{def:j-label} and its dual allow one to construct inverse bijections $\kappa: \cji(L) \rightarrow \cmi(L)$ and $\kappa^d: \cmi(L) \rightarrow \cji(L)$ such that
\begin{eqnarray}
\kappa(j) &=& \mathrm{m\text{-}label}[j_*,j] = \max\{y \in L \mid j \wedge y = j_*\}\label{eqn:kappa}\\
\kappa^d(m) &=& \mathrm{j\text{-}label}[m,m^*] = \min\{y \in L \mid m \vee y = m^*\}.\label{eqn:kappad}
\end{eqnarray}
The bijections $\kappa$ and $\kappa^d$ are sometimes referred to as either \emph{rowmotion} or the \emph{$\kappa$-maps} of $L$. See e.g. \cite[Theorem~9.3]{thomas_intro} for a proof that these maps are indeed inverse bijections.
\begin{example}\label{ex:kappa}
\begin{enumerate}
\item Let $L$ be the lattice in Figure~\ref{fig:run_ex}. Then $\kappa(j_4) = j_3$ and $\kappa(j_i) = m_i$ for $i \in \{1,2,3\}$.
\item Let $L_{kr}(K)$ be as in Figure~\ref{fig:kronecker}. Then $\kappa(s) = (1,1)$, $\kappa(1, -1) = s$, $\kappa(n, -1) = (n-1, -1)$ for $n > 1$, $\kappa(n, 1) = (n+1, 1)$ for $n \in \mathbb{N}$, and $\kappa(\{\lambda\}) = \mathbb{P}^1(K) \setminus \{\lambda\}$ for $\lambda \in \mathbb{P}^1(K)$.
\end{enumerate}
\end{example}
\begin{remark}
When $L$ is finite, the existence of the bijections $\kappa$ and $\kappa^d$ is equivalent to (complete) semidistributivity, see \cite[Theorem~2.28]{RST}. On the other hand, there exist infinite lattices which are not completely semidistributive for which these bijections are still well-defined. See \cite[Example~3.18]{RST} for an example.
\end{remark}
The following, most of which is contained in \cite{enomoto}, will be useful in computing the join-irreducible labeling.
\begin{proposition}\label{prop:cover}
Let $L$ be a completely semidistributive lattice and let $x \leq y \in L$. Then the following hold.
\begin{enumerate}
\item $\mathrm{j\text{-}label}[x,y] = \{j \in \cji(L) \mid j \leq y \text{ and }\kappa(j) \geq x\}$.
\item Suppose $x \leq \kappa(j)$. Then there is a cover relation $(x \vee j) \wedge \kappa(j) {\,\,<\!\!\!\!\cdot\,\,\,} x \vee j$ with $\mathrm{j\text{-}label}[(x \vee j) \wedge \kappa(j), x \vee j] = j$.
\item Suppose $x {\,\,<\!\!\!\!\cdot\,\,\,} y$. Then the following are equivalent.
\begin{enumerate}
\item $\mathrm{j\text{-}label}[x,y] = j$.
\item $\mathrm{m\text{-}label}[x,y] = \kappa(j)$.
\item $x \vee j = y$ and $x \wedge j = j_*$.
\item $y \wedge \kappa(j) = x$ and $y \vee \kappa(j) = \kappa(j)^*$.
\item $x \vee j = y$ and $y \wedge \kappa(j) = x$.
\end{enumerate}
\end{enumerate}
\end{proposition}
\begin{proof}
Items (1) and (2) are Theorem~3.14 and Lemma~3.15 in \cite{enomoto}, respectively. The equivalence $(3a \iff 3c)$ is also \cite[Lemma~2.7]{enomoto}. This also implies the equivalence $(3b \iff 3d)$ by duality.
$(3a \implies 3e)$: Suppose $\mathrm{j\text{-}label}[x,y] = j$. By (1), this means $j \leq y$ and $\kappa(j) \geq x$. Now by the definition of $\kappa$, we have $j \not\leq \kappa(j)$, and so $j \not \leq x$ and $\kappa(j) \not\geq y$. Since $x {\,\,<\!\!\!\!\cdot\,\,\,} y$ is a cover relation, it follows that $x \vee j = y$ and $y \wedge \kappa(j) = x$.
$(3e \implies 3c)$: Suppose $x \vee j = y$ and $x \wedge \kappa(j) = x$. In particular, this means $j \leq y$ and $\kappa(j) \geq x$. Since $x {\,\,<\!\!\!\!\cdot\,\,\,} y$ is a cover relation, it follows from (1) that $\mathrm{j\text{-}label}[x,y] = j$.
The proofs of $(3b \implies 3e)$ and $(3e \implies 3d)$ are dual to those above.
\end{proof}
To conclude this section, we recall the definitions of canonical join representations and the extended kappa-map of \cite{BTZ}.
\begin{definition}\label{def:canonical_join_rep}
Let $L$ be a complete lattice and $x \in L$.
\begin{enumerate}
\item A \emph{join representation} of $x$ is an equation $x = \Join A$ where $A \subseteq L$. We say this join representation is \emph{irredundant} if for any $j \in A$ one has $x \neq \Join(A \setminus \{j\})$.
\item Given two join representations $x = \Join A = \Join B$, we say that $\Join A$ \emph{refines} $\Join B$ if for every $a \in A$ there exists $b \in B$ such that $a \leq b$.
\item A join representation $x = \Join A$ is called a \emph{canonical join representation} if $A$ is an antichain and $\Join A$ refines every join representation of $x$. In this case, we say that $x$ is \emph{canonically join-representable} and write $A = \CJR(x)$. We call the elements of $A$ the \emph{canonical joinands} of $x$ and say that $A$ \emph{joins canonically}.
\item \emph{(Canonical) meet representations}, $\CMR(x)$, etc. are all defined dually.
\end{enumerate}
\end{definition}
\begin{example}\label{ex:can_join_rep}
Let $L$ be the lattice in Figure~\ref{fig:run_ex}. Then every element of $L$ has both a canonical join representation and a canonical meet representation. As some explicit examples:
\begin{enumerate}
\item Then there are exactly three irredundant join representations of $m_1$, namely $m_1 = \Join\{m_1\}$, $m_1 = \Join\{j_2,j_4\}$, and $m_1 = \Join\{j_2,j_3\}$. The canonical join representation of $m_1$ is $m_1 = j_2 \vee j_3$. An example of a non-irredundant join representation of $m_1$ is $m_1 = \Join\{j_2,j_3,j_4\}$.
\item The canonical meet representation of $\hat{0}$ is $\hat{0} = \bigwedge\{m_1,m_2,m_3\}$.
\item Each $j_i \in \cji(L)$ has canonical join representation $j_i = \Join\{j_i\}$. Likewise each completely meet irreducible element is its own canonical meet representation.
\item For all $x \in L$ it follows from \cite[Theorem~5.10]{RST} that $$\CJR(x) = \{j \in \mathsf{cj\text{-}rep}(L) \mid \exists u {\,\,<\!\!\!\!\cdot\,\,\,} x \text{ s.t. }j = \mathrm{j\text{-}label}[u,v] \}.$$
This relies on the fact that $|L| < \infty$. We will discuss how this generalizes to the infinite case in Theorem~\ref{thm:covers_canonical}.
\end{enumerate}
\end{example}
\begin{example}\label{ex:no_can_join}
Let $L_{kr}(K)$ be as in Figure~\ref{fig:kronecker}. Then
\begin{enumerate}
\item $x := \emptyset$ does not have a canonical join representation. Indeed, let $x = \Join A$ be a join representation of $x$. If $x \in A$, let $B = \{(n,-1) \mid n \in \mathbb{B}\}$. Then $x = \Join B$ and $\Join A$ does not refine $\Join B$. Indeed, there does not exist $b \in B$ such that $x \leq b$. If, on the other hand, $x \notin A$, then there must be an infinite subset $S \subseteq \mathbb{N}$ such that $\{(n,-1) \mid n \in S\} \subseteq A$. In particular, this means $A$ is not an antichain. Note also that $\mathbb{P}^1(K)$ does not have a canonical meet representation for a similar reason.
\item $\CJR(S) = \Join\{\{\lambda\} \mid \lambda \in S\}$ for all $\emptyset \neq S \in \mathbb{P}^1(K)$.
\end{enumerate}
\end{example}
We emphasize that the join representation $\hat{0} = \Join\emptyset$ vacuously satisfies Definition~\ref{def:canonical_join_rep}(3). Likewise, we have a canonical meet representation $\hat{1} = \bigwedge\emptyset$.
\begin{notation}
Let $L$ be a complete lattice. We denote by $\mathsf{cj\text{-}rep}(L)$ (respectively $\mathsf{cm\text{-}rep}(L)$) the subposet of canonically join-representable (resp. canonically meet-representable) elements.
\end{notation}
Before commenting on these definitions, we recall the following.
\begin{lemma}\cite[Lemma~4.5]{enomoto}\label{lem:canonical_join_rep}
Suppose that $L$ is a complete lattice and that $x \in \mathsf{cj\text{-}rep}(L)$. Then $\CJR(x) \subseteq \cji(L)$; that is, every canonical joinand of $x$ is completely join-irreducible.
\end{lemma}
\begin{remark}\label{rem:canonical_join_rep}
The definition of a canonical join representation we have used in this paper agrees with those in e.g. \cite{emily_canonical,enomoto,gorbunov,reading_book,RST}. This differs slightly from the definition used in \cite{BCZ,BTZ}, where one assumes only that $\Join A$ refines every irredundant join representation of $x$. On the other hand, we will show in Theorem~\ref{thm:covers_canonical} that the definitions coincide under the additional assumption that every element of $A$ is completely join-irreducible. For example, in the setting of Example~\ref{ex:no_can_join}, the only irredundant join representation of $x$ is $\bigvee\{x\}$. This means $x$ has a canonical join representation as defined in \cite{BCZ,BTZ}. On the other hand, we showed in Example~\ref{ex:no_can_join} that $x$ does not have a canonical join representation as defined in the present paper. This is consistent with the fact that $x$ is not compeltely join-irreducible, and therefore $\Join\{x\}$ does not satisfy the hypotheses of our Theorem~\ref{thm:covers_canonical}. See \cite[Remark~4.3]{enomoto}, Theorem~\ref{thm:brick_labeling}, and Theorem~\ref{thm:covers_canonical} for additional discussion.
\end{remark}
In Section~\ref{sec:flag} we will explore the relationship between the existence of canonical join representations and that of cover relations. For now, we recall the following result.
\begin{proposition}\cite[Theorem~1]{gorbunov}\label{prop:covers_canonical_existence}
Let $L$ be a completely semidistributive lattice and let $x \in L$. Then $x \in \mathsf{cj\text{-}rep}(L)$ if and only if for all $y < x$ there exists a cover relation $y < z {\,\,<\!\!\!\!\cdot\,\,\,} x$. In particular, if $L$ is finite then $\mathsf{cj\text{-}rep}(L) = L = \mathsf{cm\text{-}rep}(L)$.
\end{proposition}
We are now prepared to state the final definition of this section, see \cite[Definition~1.0.3]{BTZ}.
\begin{definition}\label{def:kappa_bar}
Let $L$ be a completely semidistributive lattice and let $x \in \mathsf{cj\text{-}rep}(L)$. We define
$$\overline{\kappa}(x) := \bigwedge \{\kappa(j) \mid j \in \CJR(x)\}.$$
For $x \in \mathsf{cm\text{-}rep}(L)$ we define $\overline{\kappa}^d(x)$ dually.
\end{definition}
The maps $\overline{\kappa}: \mathsf{cj\text{-}rep}(L) \rightarrow L$ and $\overline{\kappa}^d: \mathsf{cm\text{-}rep}(L) \rightarrow L$ are sometimes referred to as the \emph{extended kappa-maps}.
\begin{example}\label{ex:kappa_bar}
Let $L$ be the lattice in Figure~\ref{fig:run_ex}. Written as a permutation in cycle notation, we then have
$\overline{\kappa} = (\hat{0},\hat{1})(j_1,m_1)(j_2,m_2)(j_3,m_3,j_4).$
\end{example}
\begin{remark}
In \cite[Question~4.13]{enomoto}, Enomoto asks whether the image of $\overline{\kappa}$ lies in $\mathsf{cm\text{-}rep}(L)$ (as is the case for lattices of torsion classes, see \cite[Theorem~4.19]{enomoto} and \cite[Corollary~4.4.3]{BTZ}). When the answer to Enomoto's question is ``yes'', one can adapt the arguments of \cite[Section~4.1]{enomoto} to conclude that $\overline{\kappa}: \mathsf{cj\text{-}rep}(L) \rightarrow \mathsf{cm\text{-}rep}(L)$ is a bijection and moreover that for every $x \in \mathsf{cj\text{-}rep}(L)$ there is an induced bijection $\kappa|_{\CJR(x)}: \CJR(x) \rightarrow \CMR(\overline{\kappa}(x))$. In Section~\ref{sec:kappa_order}, we show that this is indeed the case under the additional assumption that $L$ is \emph{well-separated}.
\end{remark}
\section{Wide subcategories, torsion classes, and $\tau$-rigid modules}\label{sec:background}
Let $\Lambda$ be a finite-dimensional basic algebra over a field $K$. We denote by $\mathsf{mod}\Lambda$ the category of finitely generated (left) $\Lambda$-modules. When we speak of a subcategory of $\mathsf{mod}\Lambda$, we will always mean a full subcategory which is closed under isomorphisms.
We recall that a subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ is called \emph{wide} if it is closed under kernels, cokernels, and extensions. Equivalently, $\mathcal{W}$ is an exact-embedded abelian subcategory. Note that it is an immediate consequence of the definitions that a subcategory $\mathcal{U} \subseteq \mathcal{W}$ is a wide subcategory of $\mathcal{W}$ if and only if it is a wide subcategory of $\mathsf{mod} \Lambda$.
Given a fixed wide subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$, we denote by $\mathsf{wide}(\mathcal{W})$ the poset of wide subcategories of $\mathcal{W}$ under the inclusion order. It is well known that this poset is a (complete) lattice, with the meet operation being the intersection. By the previous paragraph, we can then identify $\mathsf{wide}(\mathcal{W})$ with the interval $[0,\mathcal{W}] \subseteq \mathsf{wide}(\mathsf{mod}\Lambda)$.
In the remainder of this section, we recall background information about torsion classes and $\tau$-tilting theory. As we will see in the sequel, this paper will be concerned with the $\tau$-tilting theory \emph{within} certain wide subcategories. Thus our convention will be to work within a wide subcategory $\mathcal{W}$ of $\mathsf{mod}\Lambda$ throughout this section. To recover the original constructions, one can take $\mathcal{W} = \mathsf{mod}\Lambda$.
\begin{remark}
Many of the wide subcategories in this paper satisfy an additional property known as \emph{functorial finiteness}. While the technical definition of functorial finiteness will not be needed in this paper, it is well-known that such subcategories are precisely those which are themselves equivalent to categories of finitely-generated modules over finite-dimensional algebras, see e.g. \cite[Proposition~4.12]{enomoto_ff}. This allows any statement which is known for such module categories to be extended to any functorially finite wide subcategory. More generally, every wide subcategory is an \emph{abelian length category}, and often this is enough to recover results which have been shown for module categories, see e.g. \cite[Section~2.2]{enomoto}. One notable exception is $\tau$-tilting theory (discussed in Section~\ref{sec:tau_tilting}), which is intimately related with the notion of projective objects, and therefore will require the assumption of functorial finiteness.
\end{remark}
\subsection{Torsion pairs}\label{sec:torsion}
We begin by recalling the definition and basic properties of torsion pairs. The material in this section is standard, and can be found for example in \cite[Section~VI.1]{ASS}.
Given two subcategories $\mathcal{C}, \mathcal{D} \subseteq \mathsf{mod}\Lambda$, we denote
\begin{eqnarray*}
\mathcal{C}^{\perp_\mathcal{D}} &:=& \{X \in \mathcal{D} \mid \mathrm{Hom}_\Lambda(-,X)|_\mathcal{C} = 0\},\\
\lperpD{\mathcal{C}} &:=& \{X \in \mathcal{D} \mid \mathrm{Hom}_\Lambda(X,-)|_\mathcal{C} = 0\}.
\end{eqnarray*}
That is, $\mathcal{C}^{\perp_\mathcal{D}}$ (resp. $\lperpD{\mathcal{C}}$) denotes the subcategory consisting of those objects in $\mathcal{D}$ which admit no nonzero morphisms from (resp. to) the objects in $\mathcal{C}$.
Now fix a wide subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$. A \emph{torsion pair in $\mathcal{W}$} is a pair $(\mathcal{T}, \mathcal{F})$ of subcategories of $\mathcal{W}$ such that $\mathcal{T}^{\perp_\mathcal{W}} = \mathcal{F}$ and $\lperpW{\mathcal{F}} = \mathcal{T}$. The subcategory $\mathcal{T}$ is called a \emph{torsion class of $\mathcal{W}$} and the subcategory $\mathcal{F}$ is called a \emph{torsion-free class of $\mathcal{W}$}.
Given an arbitrary subcategory $\mathcal{C}$, we write $\Gen\mathcal{C}$ (resp. $\Cogen \mathcal{C}$) for the subcategory of quotients (resp. submodules) of direct sums of modules in $\mathcal{C}$. More generally, if $\mathcal{C}$ is contained in some wide subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$, we denote $\Gen_\mathcal{W}\mathcal{C} := \mathcal{W} \cap \Gen\mathcal{C}$ and $\Cogen_\mathcal{W}\mathcal{C} := \mathcal{W} \cap \Cogen\mathcal{C}$.
It is well known that $\mathcal{C} \subseteq \mathcal{W}$, is a torsion class (resp. a torsion-free class) in $\mathcal{W}$ if and only if $\mathcal{C}=\Gen_\mathcal{W}\mathcal{C}$ (resp. $\mathcal{C} = \Cogen_\mathcal{W}(\mathcal{C})$) and $\mathcal{C}$ is closed under extensions.
Given a torsion pair $(\mathcal{T}, \mathcal{F})$ in $\mathcal{W}$ and a module $M \in \mathcal{W}$, there exist unique modules $t_\mathcal{T}(M) \in \mathcal{T}$ and $f_\mathcal{F}(M) \in \mathcal{F}$ which fit into an exact sequence of the form
\begin{equation}\label{eqn:canonical} 0 \rightarrow t_\mathcal{T}(M) \xrightarrow{\iota} M \xrightarrow{q} f_\mathcal{F}(M) \rightarrow 0. \end{equation}
This is called the \emph{canonical exact sequence} of $M$ associated to the torsion pair $(\mathcal{T},\mathcal{F})$.
We denote by $\mathsf{tors}(\mathcal{W})$ the poset of torsion classes of $\mathcal{W}$ with the inclusion order. As mentioned in Section~\ref{sec:lattice}, it is well known that $\mathsf{tors}(\mathcal{W})$ is a completely semidistributive lattice. We will discuss the semidistributivity property further in Section~\ref{sec:bricks}.
\begin{example}\label{ex:kronecker}
Consider the Kronecker quiver $Q = \left(\begin{tikzcd}1 \arrow[r,yshift = 0.1cm]\arrow[r,yshift = -0.1cm] & 2\end{tikzcd}\right)$ and the corresponding path algebra $\Lambda = KQ$. (For simplicity, we will assume that $K$ is algebraically closed when discussing this example.) Then $\mathsf{tors}(\mathsf{mod}\Lambda)$ is isomorphic to the lattice $L_{kr}(K)$ shown in Figure~\ref{fig:kronecker}. See \cite[Example~1.3]{thomas_intro} for a detailed description of this isomorphism.
\end{example}
The following will be useful throughout this paper. Note that given an arbitrary subcategory $\mathcal{C} \subseteq \mathsf{mod}\Lambda$, the notation $\mathsf{Filt}(\mathcal{C})$ refers to the subcategory consisting of those modules $M \in \mathsf{mod}\Lambda$ which admit a finite filtration
$$0 = M_0 \subsetneq M_1 \subsetneq \cdots \subsetneq M_k = M$$
such that $M_j/M_{j-1} \in \mathcal{C}$ for all $j$.
\begin{proposition}\cite[Proposition~2.1]{thomas_intro}
Let $\mathcal{C} \subseteq \mathcal{W}$ be any subcategory. Then $\mathsf{Filt}(\Gen_\mathcal{W}\mathcal{C})$ is the smallest torsion class in $\mathcal{W}$ which contains $\mathcal{C}$. Dually, $\mathsf{Filt}(\Cogen_\mathcal{W}\mathcal{C})$ is the smallest torsion-free class in $\mathcal{W}$ which contains $\mathcal{C}$.
\end{proposition}
\subsection{$\tau$-rigid modules}\label{sec:tau_tilting}
In this section, we recall the basic notions of $\tau$-tilting theory as constructed in \cite{AIR}. In this section, $\mathcal{W}$ will denote a functorially finite wide subcategory of $\mathsf{mod}\Lambda$.
To make the notion of $\tau$-tilting theory in a wide subcategory precise, we fix the following notation.
\begin{notation}
Fix a finite dimensional algebra $\Lambda_\mathcal{W}$ and inverse equivalences of categories $F: \mathcal{W} \rightarrow \mathsf{mod}\Lambda_\mathcal{W}$ and $G:\mathsf{mod}\Lambda_W \rightarrow \mathcal{W}$. Let $\tau_{\Lambda_\mathcal{W}}$ denote the Auslander-Reiten translate in $\mathsf{mod}\Lambda_\mathcal{W}$. Then the Auslander-Reiten translate in $\mathcal{W}$ is the assignment $\tau_\mathcal{W}: \mathcal{W} \rightarrow \mathcal{W}$ given by $\tau_\mathcal{W}(M) = G(\tau_{\Lambda_\mathcal{W}}(FM))$. We note that $\tau_\mathcal{W}$ is defined only on objects and is well-defined only up to isomorphism.
\end{notation}
We now recall the definition of a $\tau$-rigid module.
\begin{definition}
Let $M \in \mathcal{W}$ be a basic module. Then $M$ is \emph{$\tau$-rigid} (in $\mathcal{W}$), in symbols $M \in \tau\text{-}\mathrm{rigid}(\mathcal{W})$, if $\mathrm{Hom}_\Lambda(M, \tau_\mathcal{W} M) = 0$.
Let $\mathrm{rk}(\mathcal{W})$ denote the number of non-isomorphic modules which are simple in $\mathcal{W}$.
If $M\in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ and $\mathrm{rk}(M) = \mathrm{rk}(\mathcal{W})$, then $M$ is \emph{$\tau$-tilting} (in $\mathcal{W}$).
\end{definition}
\begin{remark}
Note that we have adopted the convention of \cite{BM_exceptional} in using the term ``$\tau$-rigid (in $\mathcal{W}$)'' rather than ``$\tau_\mathcal{W}$-rigid''. Nevertheless we caution that there may be modules $M \in \mathcal{W}$ which are $\tau$-rigid in $\mathcal{W}$, but not $\tau$-rigid in $\mathsf{mod}\Lambda$.
\end{remark}
We also fix the following definition for use in Section~\ref{sec:kappa_tau}.
\begin{notation}\label{not:tau_bar}
Let $M \in \mathcal{W}$. Let $\tau_\mathcal{W}$ and $\nu_\mathcal{W}$ denote the Auslander-Reiten translation and Nakayama functor in $\mathcal{W}$. (Recall that $\nu_\mathcal{W}$ sends the indecomposable projective $P(i)$ to the corresponding indecomposable injective $I(i)$.) If $M$ is indecomposable, we denote
$$\overline{\tau_\mathcal{W}}M = \begin{cases} \tau_\mathcal{W} M & M \text{ is not projective in }\mathcal{W}\\
\nu_\mathcal{W} M & M \text{ is projective in }\mathcal{W}.\end{cases}$$
We then extend this definition additively to define $\overline{\tau_\mathcal{W}} M$ when $M$ is not indecomposable.
\end{notation}
Before continuing, we recall the following characterization of Auslander and Smal\o, which will be useful in several of our proofs.
\begin{proposition}\cite[Proposition~5.8]{AS} \label{prop:AStau}
Let $M, N \in \mathcal{W}$. Then $\mathrm{Hom}(N,\tau_\mathcal{W} M) = 0$ if and only if $\mathrm{Ext}^1_\Lambda(M,N') = 0$ for every $N' \in \Gen_\mathcal{W}(N)$.
\end{proposition}
For torsion classes, we take the following as our definition of functorial finiteness. See \cite[Sections~2.2-2.3]{AIR} for justification for doing so.
\begin{definition-theorem}\cite[Sections~2.2-2.3]{AIR}\label{thm:functorially_finite}
Let $\mathcal{T} \subseteq \mathcal{W}$ be a torsion class of $\mathcal{W}$. Then $\mathcal{T}$ is \emph{functorially finite in $\mathcal{W}$} if and only if there exists $M \in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ such that $\mathcal{T} = \Gen_\mathcal{W} M$.
\end{definition-theorem}
As we have stated Theorem~\ref{thm:functorially_finite}, the module $M$ satisfying $\mathcal{T} = \Gen_\mathcal{W} M$ may not be unique. We use the following to formulate the uniqueness result we will use in this paper.
\begin{definition}\
\begin{enumerate}
\item Let $M\in\tau\text{-}\mathrm{rigid}(\mathcal{W})$. Write $M = \bigoplus_{i = 1}^k M_i$ as a direct sum of indecomposable modules. We say that $M$ is \emph{gen-minimal} if for all $j$ we have that $M_j \notin \Gen_\mathcal{W}\left(\bigoplus_{i \neq j} M_i\right)$. Equivalently, for all $j$ we have that $\Gen_\mathcal{W} M \neq \Gen_\mathcal{W}\left(\bigoplus_{i \neq j} M_i\right)$.
\item Let $\mathcal{C} \subseteq \mathcal{W}$ be a subcategory, and let $M \in \mathcal{C}$ be indecomposable. We say that $M$ is (indecomposable) \emph{split projective} in $\mathcal{C}$ if every epimorphism in $\mathcal{C}$ with target $M$ is split.
\end{enumerate}
\end{definition}
By combining \cite[Lemma~2.8]{IT} with \cite[Theorem~2.7]{AIR}, we obtain the following.
\begin{proposition}\label{prop:functorially_finite}
The association $M \mapsto \Gen_\mathcal{W} M$ is a bijection between the set of gen-minimal $\tau$-rigid modules in $\mathcal{W}$ and the set of functorially finite torsion classes in $\mathcal{W}$. The inverse sends a (functorially finite) torsion class $\mathcal{T} \subseteq \mathcal{W}$ to the direct sum of the indecomposable modules which are split projective in $\mathcal{T}$.
\end{proposition}
We conclude this section with the following.
\begin{definition-theorem}\label{def:tau_tilting_finite}
We say the algebra $\Lambda$ is \emph{$\tau$-tilting finite} if the following equivalent conditions hold.
\begin{enumerate}
\item The lattice $\mathsf{tors}(\mathsf{mod}\Lambda)$ is finite.
\item There are finitely many functorially finite torsion classes in $\mathsf{mod}\Lambda$.
\item Every torsion class of $\mathsf{mod}\Lambda$ is functorially finite.
\item The lattice $\mathsf{wide}(\mathsf{mod}\Lambda)$ is finite.
\item There are finitely many functorially finite wide subcategories in $\mathsf{mod}\Lambda$.
\item There are finitely many $\tau$-rigid modules in $\mathsf{mod}\Lambda$.
\item There are finitely many $\tau$-tilting modules in $\mathsf{mod}\Lambda$.
\end{enumerate}
\end{definition-theorem}
\begin{proof}
The equivalence of (1), (2), (3), (6), and (7) can be found in \cite{DIJ} and the implication $(4 \implies 5)$ is trivial. The implications $(1 \implies 4)$ and $(5 \implies 2)$ follow from the fact that there is an injective map $\mathsf{wide}(\mathsf{mod}\Lambda) \rightarrow \mathsf{tors}(\mathsf{mod}\Lambda)$, that every functorially finite torsion class is in the image of this map, and that the preimage of any functorially finite torsion class must be functorially finite, see \cite{IT,MS}.
\end{proof}
\subsection{Bricks and semibricks}\label{sec:bricks}
Another important class of modules in $\tau$-tilting theory are the \emph{bricks}, defined as follows.
\begin{definition}
We say a module $B \in \mathsf{mod}\Lambda$ is a \emph{brick} if $\mathrm{End}_\Lambda(B)$ is a division algebra. We say a set $\mathcal{S}$ of bricks is a \emph{semibrick} if for all $B \neq C \in \mathcal{S}$ we have $\mathrm{Hom}_\Lambda(B,C) = 0 = \mathrm{Hom}_\Lambda(C,B)$.
\end{definition}
\begin{notation}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a wide subcategory. We denote by $\mathsf{brick}(\mathcal{W})$ and $\mathsf{sbrick}(\mathcal{W})$ the set of bricks and semibricks in $\mathcal{W}$, respectively.
\end{notation}
\begin{remark}\label{rem:brick}
We note that the property of being a brick does not change when one replaces $\mathsf{mod}\Lambda$ with a subcategory. In other words, given a wide subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$, one has $\mathsf{brick}(\mathcal{W}) = \mathcal{W} \cap \mathsf{brick}(\mathsf{mod}\Lambda)$ and $\mathsf{sbrick}(\mathcal{W}) = \mathcal{W} \cap \mathsf{sbrick}(\mathsf{mod}\Lambda)$.
\end{remark}
In light of the previous remark, it is sometimes more intuitive to work with (semi)bricks than with $\tau$-rigid modules. The ``brick-$\tau$-rigid correspondence'' of Demonet-Iyama-Reiten \cite{DIJ}, and its generalization to semibricks due to Asai \cite{asai}, offers further motivation for the study of bricks and semibricks:
\begin{theorem}\label{thm:DIJ}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a functorially finite wide subcategory.
\begin{enumerate}
\item \cite[Lemma~4.3]{DIJ} Let $M\in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ be indecomposable, and denote $$r(M) = \{g \in \mathrm{Hom}_\Lambda(M,M) \mid 0 \neq g(M) \neq M\}.$$ Then $$\beta(M):=M/\left(\sum_{g \in r(M)} \mathrm{im} g\right)$$
is a brick which satisfies $\mathsf{Filt}\Gen_\mathcal{W}(\beta(M)) = \Gen_\mathcal{W}(M)$.
\item \cite[Theorem~4.1]{DIJ} The association $\beta$ is an injection from the set of indecomposable modules which are $\tau$-rigid in $\mathcal{W}$ to the set of bricks in $\mathcal{W}$. The image of this injection is the set of bricks $B \in \mathcal{W}$ for which the torsion class $\mathsf{Filt}\Gen_\mathcal{W}(B)$ is functorially finite in $\mathcal{W}$.
\item \cite[Theorem~1.3]{asai} Let $M\in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ be gen-minimal and decompose $M = \bigoplus_{i = 1}^k M_i$ as a direct sum of indecomposable modules. For $j \in \{1,\ldots,k\}$, denote $\mathcal{F}_j:= \left(\bigoplus_{i \neq j} M_i\right)^{\perp_\mathcal{W}}$. Then
$\mathcal{X}(M) := \left\{\beta\left(f_{\mathcal{F}_j}M_j\right)\right\}$
is a semibrick which satisfies $\mathsf{Filt}\Gen_\mathcal{W}(\mathcal{X}(M)) = \Gen_\mathcal{W}(M)$.
\item \cite[Theorem~1.3]{asai} The association $\mathcal{X}$ is a injection from the set of gen-minimal $\tau$-rigid modules in $\mathcal{W}$ to the set of semibricks in $\mathcal{W}$. The image of this injection is the set of semibricks $\mathcal{X} \subseteq \mathcal{W}$ for which the torsion class $\mathsf{Filt}(\Gen_\mathcal{W} \mathcal{X})$ is functorially finite in $\mathcal{W}$.
\end{enumerate}
\end{theorem}
\begin{remark}\label{rem:sbrick_unique}
Note in particular that Theorem~\ref{thm:DIJ}(4) implies that if $\mathcal{S}, \mathcal{S}' \subseteq \mathcal{W}$ are semibricks which satisfy $\mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S}) = \mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S}')$ (and this torsion class is functorially finite), then $\mathcal{S} = \mathcal{S}'$. It is also shown in \cite{asai} that this remains true without the assumption that $\mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S})$ be functorially finite, but this generalization no longer follows from Theorem~\ref{thm:DIJ} in its stated form.
\end{remark}
We refer to $\beta(M)$ (resp. $\mathcal{X}(M)$) as the \emph{brick corresponding to $M$} (resp. \emph{semibrick corresponding to $M$}). We note that if $M$ is (indecomposable) $\tau$-rigid in both $\mathcal{W}$ and $\mathcal{V}$, then $\beta(M)$ and $\mathcal{X}(M)$ remain the same whether computed in $\mathcal{W}$ or $\mathcal{V}$. That is, the formulas for $\beta(M)$ and $\mathcal{X}(M)$ do not depend on the wide subcategory $\mathcal{W}$. Both the domains and images of $\beta$ and $\mathcal{X}$, on the other hand, do. We explain this further in Remark~\ref{rem:beta_domain} and Proposition~\ref{prop:sf-brick}.
\begin{notation}
Let $\mathcal{W}$ be a functorially finite wide subcategory, let $B \in \mathcal{W}$ be a brick, and suppose $\mathsf{Filt}\Gen_\mathcal{W}(B)$ is functorially finite in $\mathcal{W}$. We denote by $\beta^{-1}_\mathcal{W}(B)$ the unique indecomposable module $M \in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ and satisfies $\beta(M) = B$. We define $\mathcal{X}^{-1}_\mathcal{W}(\mathcal{S})$ analogously.
\end{notation}
\begin{remark}\label{rem:beta_domain}
It is important to specify $\mathcal{W}$ in the notation $\beta^{-1}_\mathcal{W}$. Indeed, there are examples where $\beta_\mathcal{W}^{-1}(B) \neq \beta_{\mathcal{U}}^{-1}(B)$. For example, let $\Lambda$ be the preprojective algebra of type $A_3$, take $\mathcal{W} = \mathsf{mod}\Lambda$, and take $\mathcal{U} = \mathsf{Filt}(P_2/S_2)$. Then $\beta_\mathcal{W}^{-1}(P_2/S_2) = P_2$ and $\beta_\mathcal{U}^{-1}(P_2/S_2) = P_2/S_2$.
\end{remark}
Another important use of bricks is in interpreting the join-irreducible labeling of the lattice of torsion classes. In particular, we consider the following definition.
\begin{definition}\label{def:min_extending}\cite[Definition~1.1 and~2.10]{BCZ}
Let $\mathcal{W}$ be a wide subcategory and let $\mathcal{T} \in \mathsf{tors}\mathcal{W}$ be a torsion class.
\begin{enumerate}
\item A module $M \in \mathcal{W}$ is called a \emph{minimal extending module} for $\mathcal{T}$ if the following hold:
\begin{enumerate}
\item If $X$ is a proper factor of $M$, then either $X \in \mathcal{T}$ or $X \notin \mathcal{W}$.
\item For every $X \in \mathcal{T}$ and every nonsplit exact sequence $0 \rightarrow M \rightarrow E \rightarrow X \rightarrow 0$, one has $E \in \mathcal{T}$.
\item $M \in \mathcal{T}^{\perp_\mathcal{W}}$.
\end{enumerate}
\item Denote $\mathcal{F} = \mathcal{T}^{\perp_\mathcal{W}}$. A module $M \in \mathcal{W}$ is called a \emph{minimal co-extending module} for $\mathcal{F}$ if the following hold:
\begin{enumerate}
\item If $X$ is a proper submodule of $M$ then either $X \in \mathcal{F}$ or $X \notin \mathcal{W}$.
\item For every $X \in \mathcal{F}$ and every nonsplit exact sequence $0 \rightarrow X\rightarrow E \rightarrow M \rightarrow 0$ one has $E \in \mathcal{F}$.
\item $M \in \mathcal{T} = \lperpW{\mathcal{F}}$.
\end{enumerate}
\end{enumerate}
\end{definition}
We also need the following notation.
\begin{notation}\label{not:kappaW}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a wide subcategory. We denote by $\kappa_\mathcal{W}$ and $\overline{\kappa_\mathcal{W}}$ the kappa-map of the lattice $\mathsf{tors}(\mathcal{W})$.
\end{notation}
We conclude by recalling some of the main results of the papers \cite{BCZ,BTZ,DIRRT}. See also \cite[Theorem~2.15]{enomoto} for the explicit extension of Theorem~\ref{thm:min_extending} to arbitrary abelian length categories (and thus arbitrary wide subcategories).
\begin{theorem}\label{thm:min_extending}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a wide subcategory.
\begin{enumerate}
\item \cite[Proposition~3.1]{BCZ}\cite[Theorem~3.3(c)]{DIRRT} There is a bijection $\mathsf{brick}(\mathcal{W}) \rightarrow \cji(\mathsf{tors}\mathcal{W})$ given by $B \mapsto \mathsf{Filt}(\Gen_\mathcal{W} B)$.
\item \cite[Theorem~A]{BTZ} Let $B \in \mathsf{brick}(\mathcal{W})$. Then $\kappa_\mathcal{W}(\mathsf{Filt}(\Gen_\mathcal{W} B)) = \lperpW{B}$. In particular, there is a bijection $\mathsf{brick}\mathcal{W} \rightarrow \cmi(\mathsf{tors}(\mathcal{W}))$ given by $B \mapsto \lperpW{B}$.
\end{enumerate}
\end{theorem}
\begin{notation}\label{not:brick_lab}
Let $\mathcal{U} \subseteq \mathcal{T}$ be torsion classes in $\mathsf{tors}(\mathcal{W})$. We denote $$\mathrm{br\text{-}label}[\mathcal{U},\mathcal{T}] := \{B \mid \mathsf{Filt}(\Gen_\mathcal{W} B) \in \mathrm{j\text{-}label}[\mathcal{U},\mathcal{T}]\}.$$
\end{notation}
\begin{remark}\label{rem:brick_lab}\
\begin{enumerate}
\item The association of a cover relation $\mathcal{U} {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{T}$ to the brick $\mathrm{br\text{-}label}[\mathcal{U},\mathcal{T}]$ is sometimes called the \emph{brick labeling} of $\mathsf{tors}(\mathcal{W})$. Asai introduced this labeling for the subposet of functorially finite torsion classes in \cite{asai} (see also \cite{BST} for a geometric interpretation). For non-functorially finite torsion class, the brick labeling was introduced independently in \cite{BCZ,DIRRT}.
\item While it is often unimportant in practice, to satisfy Definition~\ref{def:edge-label}, one should place a partial order on the set of bricks, see Definition~\ref{def:edge-label}. One possibility is to set $C \preceq B$ whenever $\mathsf{Filt}(\Gen_\mathcal{W} B) \subseteq \mathsf{Filt}(\Gen_\mathcal{W} C)$. Indeed, we will use a refinement of this partial order to construct an EL-labeling in Section~\ref{sec:el}, see Remark~\ref{rem: rho}.
\end{enumerate}
\end{remark}
Our next result is essentially contained in \cite[Section~4.3]{BTZ}, albeit with a different definition of canonical join representation. We thus provide an outline of a proof for the convenience of the reader. See also \cite[Section~4.2]{enomoto}.
\begin{theorem}\label{thm:brick_labeling}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a wide subcategory and let $\mathcal{T} \in \mathsf{tors}(\mathcal{W})$. Then
\begin{enumerate}
\item $\mathcal{T} \in \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$ if and only if there exists a semibrick $\mathcal{S} \in \mathsf{sbrick}(\mathcal{W})$ such that $\mathcal{T} = \mathsf{Filt}(\Gen_\mathcal{W} \mathcal{S})$. Moreover, if such a semibrick exists then $\mathcal{S}$ is the set of minimal co-extending modules for $\mathcal{T}^{\perp_\mathcal{W}}$ and $\CJR(\mathcal{T}) = \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \in \mathcal{S}\}$.
\item The map $\overline{\kappa_\mathcal{W}}$ induces a bijection $\mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W})) \rightarrow \mathsf{cm\text{-}rep}(\mathsf{tors}\mathcal{W})$. Moreover, for $\mathcal{T} \in \mathsf{cj\text{-}rep}(\mathcal{W})$, one has $\CMR(\overline{\kappa}_\mathcal{W}(\mathcal{T})) = \{\lperpW{B} \mid \mathsf{Filt}(\Gen_\mathcal{W} B) \in \CJR(\mathcal{T})\}$.
\end{enumerate}
\end{theorem}
\begin{proof}
Suppose that $\mathcal{T}\in \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$.
Then by Proposition~\ref{prop:covers_canonical_existence}, for any any torsion class $\mathcal{U}\lneq \mathcal{T}$ there exists $\mathcal{T}'$ such that $\mathcal{U}\le \mathcal{T}' {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{T}$.
By \cite[Corollary~3.9]{BCZ}, for $\mathcal{S}$ the set of minimal co-extending modules for $\mathcal{T}^{\perp_\mathcal{W}}$, the join $\Join \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \in \mathcal{S}\}$ is an irredundant join-representation of $\mathcal{T}$ and it refines all other irredundant join-representations of $\mathcal{T}$.
Moreover, by Theorem~\ref{thm:min_extending}, $\mathsf{Filt}(\Gen_\mathcal{W} B)$ is completely join-irreducible, for each $B\in \mathcal{S}$, and $\mathcal{S}$ is a semibrick by \cite[Proposition~3.5]{BCZ}.
We will show in Theorem~\ref{thm:covers_canonical} that this implies $\CJR(\mathcal{T}) = \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \in \mathcal{S}\}$.
For the reverse implication, assume there exists such a semibrick $\mathcal{S} \in \mathsf{sbrick}(\mathcal{W})$ such that $\mathcal{T} = \mathsf{Filt}(\Gen_\mathcal{W} \mathcal{S})$.
By \cite[Proposition~3.7]{BCZ}, $\mathcal{T}= \Join \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \in \mathcal{S}\}$ is irredundant, and refines all other irredundant join-representations of $\mathcal{T}$.
As before, Theorem~\ref{thm:min_extending} and Theorem~\ref{thm:covers_canonical} implies that $\CJR(\mathcal{T}) = \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \in \mathcal{S}\}$.
Thus $\mathcal{T}\in \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$.
(2) The fact that $\overline{\kappa_\mathcal{W}} = \Join \{\lperpW B \mid \mathsf{Filt}(\Gen_\mathcal{W} B) \in \CJR(\mathcal{T})\}$ is \cite[Corollary~4.4.3]{BTZ}. Moreover, by \cite[Proposition~4.4.5]{BTZ}, we have that $\{B \mid \mathsf{Filt}(\Gen_\mathcal{W} B)\}$ is the set of minimal extending modules for the torsion class $\overline{\kappa_\mathcal{W}}(\mathcal{T})$. The fact that $\overline{\kappa_\mathcal{W}}$ is a bijection and that $\overline{\kappa_\mathcal{W}}(\mathcal{T})$ has the desired canonical meet representation thus follows from the dual of (1).
\end{proof}
To conclude this section, we briefly recall the connection between torsion classes and wide subcategories in terms of the so-called ``Ingalls-Thomas correspondences'' of \cite{IT,MS}. This result is explicit as \cite[Theorem~4.17]{enomoto}, but also appears implicitly in \cite[Section~3.2 and Corollary~5.1.8]{BTZ}.
\begin{theorem}\label{thm:IT}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a wide subcategory. Then the association $\mathcal{V} \mapsto \mathsf{Filt}(\Gen_\mathcal{W} \mathcal{V})$ is a bijection $\mathsf{wide}(\mathcal{W}) \rightarrow \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$. The inverse sends $\mathcal{T} \in \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$ to
$$\alpha(\mathcal{T}) := \{X \in \mathcal{T} \mid \ker(f) \in \mathcal{T} \text{ for all } f: X \rightarrow Y \text{ with } Y \in \mathcal{T}\}.$$
\end{theorem}
\section{$\tau$-perpendicular subcategories and $\tau$-exceptional sequences}\label{sec:tex}
In this section, we briefly review the \emph{$\tau$-perpendicular subcategories} of Jasso \cite{jasso} and the \emph{$\tau$-exceptional sequences} of Buan and Marsh \cite{BM_exceptional}. We then describe how the brick-$\tau$-rigid correspondence can be used to interpret $\tau$-exceptional sequences as a ``brick labeling'' of the lattice of wide subcategories.
\subsection{$\tau$-perpendicular subcategories}\label{sec:tau_perp}
\begin{definition}\label{def:tau_perp}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a functorially finite wide subcategory.
\begin{enumerate}
\item Let $M \in \mathcal{W}$ be $\tau$-rigid in $\mathcal{W}$. The \emph{$\tau$-perpendicular subcategory} of $M$ (in $\mathcal{W}$) is
$$\mathcal J_\mathcal{W}(M) := (M^{\perp_\mathcal{W}}) \cap (\lperpW{(\tau_\mathcal{W} M)}).$$
\item Let $\mathcal{U} \subseteq \mathcal{W}$ be a (functorially finite) wide subcategory. We say that $\mathcal{U}$ is a \emph{$\tau$-perpendicular subcategory} of $\mathcal{W}$ if there exists a $\tau$-rigid module $M \in \mathcal{W}$ such that $\mathcal{U} = \mathcal J_\mathcal{W}(M)$.
\end{enumerate}
\end{definition}
\begin{example}\label{ex:tau_perp}\
\begin{enumerate}
\item By taking $M = \Lambda$ and $M = 0$, respectively, we can see both the 0-subcategory and $\mathsf{mod}\Lambda$ as $\tau$-perpendicular subcategories of $\mathsf{mod}\Lambda$. Moreover, all ``right finite'' and ``left finite'' wide subcategories are $\tau$-perpendicular. See \cite[Lemma~4.3]{BuH}.
\item In general, there are functorially finite wide subcategories which are not $\tau$-perpendicular. See \cite[Example~3.13]{asai}.
\item If $\Lambda$ is $\tau$-tilting finite, then every wide subcategory is a $\tau$-perpendicular subcategory. This is shown explicitly in \cite[Theorem~4.18]{DIRRT}, and also follows from the fact that if $\Lambda$ is $\tau$-tilting finite, then every wide subcategory of $\mathsf{mod}\Lambda$ is both left and right finite.
\end{enumerate}
\end{example}
\begin{remark}
We note that the definition of a $\tau$-perpendicular category given in \cite{BuH} is given in terms of $\tau$-rigid \emph{pairs} rather than $\tau$-tilting modules. It turns out, however, that the two definitions are equivalent. This fact is implicit in many places, including the seminal work \cite{jasso}. For convenience, we give a brief explanation here. Our argument is based off of \cite[Theorem~6.4]{BuH}, which is an extension of \cite[Theorem~4.3]{BM_wide}.
Let $\mathcal{U} \subseteq \mathcal{W}$. We suppose that $\mathcal{U}$ satisfies the definition of a $\tau$-perpendicular subcategory given in \cite{BuH}; that is, that there exists a module $M$ which is $\tau$-rigid in $\mathcal{W}$ and a module $P$ which is projective in $\mathcal{W}$ such that $\mathrm{Hom}_\Lambda(P,M) = 0$ and $\mathcal{U} = \mathcal J_\mathcal{W}(M) \cap P^{\perp_\mathcal{W}}$. Since projective modules are $\tau$-rigid, it follows that $\mathcal{V} := P^{\perp_\mathcal{W}} = \mathcal J_\mathcal{W}(P)$ is a $\tau$-perpendicular category. Moreover, by \cite[Theorem~6.4]{BuH} (see also \cite[Proposition~2.3]{AIR} and \cite[Lemma~3.8]{BM_wide}) we have that $M$ is $\tau$-rigid in $\mathcal{V}$ and satisfies $\mathcal J_\mathcal{V}(M) = \mathcal J_\mathcal{W}(M)$. Applying Theorem~\cite[Theorem~6.4]{BuH} once again, we conclude that there exists a module $M'$ such that $P\oplus M'$ is $\tau$-rigid in $\mathcal{W}$ and $\mathcal J_\mathcal{W}(P\oplus M') = \mathcal{U}$. This shows that any subcategory satisfying the definition given in \cite{BuH} also satisfies Definition~\ref{def:tau_perp}. Conversely, it is clear that any subcategory of the form $\mathcal J_\mathcal{W}(M)$ can also be written in the form $\mathcal J_\mathcal{W}(M) \cap P^{\perp_\mathcal{W}}$ by taking $P = 0$.
\end{remark}
Implicit in Definition~\ref{def:tau_perp}(2) is the fact that every $\tau$-perpendicular subcategory is a functorially finite wide subcategory. This follows from the work of Jasso \cite{jasso}, where it is also shown that if $M \in \mathcal{W}$ is $\tau$-rigid, then \begin{equation}\label{eqn:jasso}\mathrm{rk}(\mathcal J_\mathcal{W}(M)) + \mathrm{rk}(M) = \mathrm{rk}(\mathcal{W}).\end{equation}
The following is critical, and ultimately allows us to refer to the poset of $\tau$-perpendicular subcategories of $\mathsf{mod}\Lambda$.
\begin{theorem}\cite[Corollary~6.7]{BuH}\label{thm:iterated_tau_perp} Let $\mathcal{U} \subseteq \mathcal{V} \subseteq \mathcal{W}$ be a chain of subcategories. If $\mathcal{U}$ is $\tau$-perpendicular in $\mathcal{V}$ and $\mathcal{V}$ is $\tau$-perpendicular in $\mathcal{W}$, then $\mathcal{U}$ is $\tau$-perpendicular in $\mathcal{W}$.
\end{theorem}
We are now ready to formally introduce the poset of $\tau$-perpendicular subcategories.
\begin{proposition}\label{prop:poset}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a functorially finite wide subcategory, and let $\tau\text{-}\mathsf{perp}(\mathcal{W})$ be the set of $\tau$-perpendicular subcategories of $\mathcal{W}$. Define a relation $\leq_\tau$ on $\tau\text{-}\mathsf{perp}(\mathcal{W})$ so that $\mathcal{U} \leq_\tau \mathcal{V}$ if and only if $\mathcal{U}$ is a $\tau$-perpendicular subcategory of $\mathcal{V}$. Then $\leq_\tau$ is a partial order on $\tau\text{-}\mathsf{perp}(\mathcal{W})$.
\end{proposition}
\begin{proof}
Reflexivity follows from the fact that $\mathcal{V} = \mathcal J_\mathcal{V}(0)$ for all $\mathcal{V} \in \tau\text{-}\mathsf{perp}(\mathcal{W})$. Transitivity is shown in Theorem~\ref{thm:iterated_tau_perp}. It thus remains only to show that $\leq_\tau$ is antisymmetric. To see this, we note that if $\mathcal{U} \lneq_\tau \mathcal{V}$, then $\mathrm{rk}(\mathcal{U}) < \mathrm{rk}(\mathcal{V})$ by Equation~\ref{eqn:jasso}.
\end{proof}
\begin{remark}
If $\Lambda$ is $\tau$-tilting finite, then every wide subcategory is a $\tau$-perpendicular subcategory and $\leq_\tau$ coincides with the inclusion order. (This follows from Example~\ref{ex:tau_perp}(3).) We conjecture that $\leq_\tau$ concides with the inclusion order more generally, even though $\tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda) \neq \mathsf{wide}(\mathsf{mod}\Lambda)$ outside of the functorially finite case. See \cite[Conjecture~6.8]{BuH}.
\end{remark}
From now on, we use $\tau\text{-}\mathsf{perp}(\mathcal{W})$ to refer to the poset $(\tau\text{-}\mathsf{perp}(\mathcal{W}),\leq_\tau)$. An example for the Kronecker path algebra will be discussed in Example~\ref{ex:kronecker_kappa} after we introduce the ``$\kappa$-order'' in Section~\ref{sec:kappa_order}. The following is direct consequence of Theorem~\ref{thm:iterated_tau_perp}.
\begin{corollary}\label{cor:iterated_tau_perp}
Let $\mathcal{V} \subseteq \mathcal{W}$ be a chain of $\tau$-perpendicular subcategories. Then
$$\tau\text{-}\mathsf{perp}(\mathcal{V}) = \{\mathcal{U} \in \tau\text{-}\mathsf{perp}(\mathcal{W}) \mid \mathcal{U} \leq_\tau \mathcal{V}\}$$
as partially ordered sets.
\end{corollary}
In particular, Corollary~\ref{cor:iterated_tau_perp} justifies writing $\mathcal{V} \leq_\tau \mathcal{W}$ without specifying the ambient category.
\subsection{$\tau$-exceptional sequences}\label{sec:tau_exceptional}
We are now prepared to recall the definition of a $\tau$-exceptional sequence from \cite{BM_exceptional}
\begin{definition}\label{def:tau_seq}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a functorially finite wide subcategory. An ordered sequence $(M_k,\ldots,M_1)$ of indecomposable modules is a \emph{$\tau$-exceptional sequence} in $\mathcal{W}$ if $M_1$ is $\tau$-rigid in $\mathcal{W}$ and $(M_k,\ldots,M_2)$ is a $\tau$-exceptional sequence in $\mathcal J_\mathcal{W}(M_1)$.
\end{definition}
\begin{remark}\
\begin{enumerate}
\item In order to clarify the recursive definition, we note that $(M_k,\ldots,M_1)$ is a $\tau$-exceptional sequence if and only if there exists a sequence $\mathcal{W}_{k-1} \subseteq \cdots \subseteq\mathcal{W}_0$ of (functorially finite) wide subcategories such that (i) $\mathcal{W}_0 = \mathcal{W}$, and (ii) For all $i$ we have that $M_i$ is $\tau$-rigid in $\mathcal{W}_{i-1}$ and that $\mathcal{W}_i = \mathcal J_{\mathcal{W}_{i-1}}(M_i)$.
In particular, if $(M_k,\ldots,M_1)$ is a $\tau$-exceptional sequence, then for $i < j$ we have $\mathrm{Hom}_\Lambda(M_i,M_j) = 0$ because $M_j\in M_i^\perp$ and $\mathrm{Ext}^1_\Lambda(M_i,M_j) = 0$ since $M_j \in \prescript{\prescript{}{(\mathcal{W}_{i-1})}{\perp}}{}{\tau_{\mathcal{W}_{i-1}} M_i}$ (see also Proposition~\ref{prop:AStau}). Moreover, for all $i$ we have $\mathrm{Ext}^1_\Lambda(M_i,M_i) = 0$ since $M_i$ is $\tau$-rigid in $\mathcal{W}_{i-1}$.
\item When $\Lambda$ is hereditary, a sequence $(M_k,\ldots,M_1)$ of indecomposables is a $\tau$-exceptional sequence if and only if it is an \emph{exceptional sequence}; i.e., if and only if each $M_i$ is a brick, $\mathrm{Hom}(M_i,M_j) = 0$ for $i < j$, and $\mathrm{Ext}^n(M_i,M_j) = 0$ for $i \leq j$ and $n > 0$. Exceptional sequences over hereditary algebras are a classical object of study, but the condition on Ext vanishing is often quite restrictive for non-hereditary algebras. Since this condition is void for $n > 1$ in the hereditary case, one could also restrict the vanishing condition on Ext to only $n = 1$ to obtain {\c S}en's ``weak exceptional sequences'' \cite{sen}. (These are themselves a subclass of the well-studied ``stratifying systems'', see e.g. the introduction of \cite{MT} and the references therein.) In case the bricks and $\tau$-rigid modules of an algebra coincide, every $\tau$-exceptional sequence will be a weak exceptional sequence, but one consequence of {\c S}en's work is that the converse does not hold in general.
\end{enumerate}
\end{remark}
We denote by $\tau\text{-}\mathsf{ex}(\mathcal{W})$ the set of $\tau$-exceptional sequences in $\mathcal{W}$.
\begin{definition}\cite[Definition~3.1]{MT}\label{def:TF_adm}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a functorially finite wide subcategory. An ordered sequence $(M_k,\ldots,M_1)$ of indecomposable modules is a \emph{TF-admissible ordered $\tau$-rigid module} in $\mathcal{W}$ if $\bigoplus_{j = 1}^k M_j$ is $\tau$-rigid in $\mathcal{W}$ and for all $j \in \{2,\ldots,k\}$, one has
$$M_j \notin \Gen_\mathcal{W}\left(\bigoplus_{i = 1}^{j-1} M_i\right).$$
\end{definition}
We denote by $\mathsf{tfo}\mathcal{W}$ the set of TF-admissible ordered $\tau$-rigid modules in $\mathcal{W}$.
\begin{remark}
If $\bigoplus_{j = 1}^k M_j$ is gen-minimal, then $(M_k,\ldots,M_1)$ is a TF-admissible ordered $\tau$-rigid module. The converse, however, is not true. For example, let $\Lambda = K(1 \rightarrow 2 \rightarrow 3)$ and take $\mathcal{W} = \mathsf{mod}\Lambda$. Then $P_1 \oplus (P_1/P_3)$ is a $\tau$-rigid module which is not gen-minimal. We then have that $(P_1,P_1/P_3)$ is a TF-admissible ordered $\tau$-rigid module in $\mathsf{mod}\Lambda$, but that $(P_1/P_3,P_1)$ is not. In particular, we emphasize that the property of being gen-minimal does not depend on how the indecomposable direct summands are ordered, while the property of being TF admissible does.
\end{remark}
\begin{definition}
Let $\mathcal{P}$ be a poset. We say that a chain $(x_0 {\,\,<\!\!\!\!\cdot\,\,\,} x_1 {\,\,<\!\!\!\!\cdot\,\,\,} \cdots {\,\,<\!\!\!\!\cdot\,\,\,} x_m)$ is a \emph{saturated top chain} of $\mathcal{P}$ if $x_m$ is maximal in $\mathcal{P}$ and each relation $x_i {\,\,<\!\!\!\!\cdot\,\,\,} x_{i + 1}$ is a cover relation.
\end{definition}
We denote by $\mathsf{stop\text{-}chain}(\mathcal{P})$ the set of saturated top chains of $\mathcal{P}$. We conclude this section by describing bijections between the sets $\mathsf{tfo}(\mathcal{W})$, $\tau\text{-}\mathsf{ex}(\mathcal{W})$, and $\mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}(\mathcal{W}))$ which have appeared in the literature. We note that all three of the bijections in the following theorem are built upon Jasso's seminal work on reduction of $\tau$-rigid modules \cite{jasso}.
\begin{theorem}\label{thm:tf_tau}
Let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a wide subcategory.
\begin{enumerate}
\item \cite[Theorem~5.1]{MT} There is a bijection $\chi: \mathsf{tfo}(\mathcal{W}) \rightarrow \tau\text{-}\mathsf{ex}(\mathcal{W})$ given as follows. Let $(M_k,\ldots,M_1) \in \mathsf{tfo}(\mathcal{W})$. For $j \in \{1,\ldots,k\}$, let $\mathcal{F}_j := \left(\bigoplus_{i < j} M_i\right)^{\perp_\mathcal{W}}$ and let $N_j := f_{\mathcal{F}_j}(M_j)$. (Note that $N_1 = M_1$.) Then $\chi(M_k,\ldots,M_1) := (N_k,\ldots,N_1)$.
\item \cite[Theorem~10.1]{BM_wide} There is a bijection $\psi: \tau\text{-}\mathsf{ex}(\mathcal{W}) \rightarrow \mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}(\mathcal{W}))$ given as follows. Let $(M_k,\ldots,M_1) \in \tau\text{-}\mathsf{ex}(\mathcal{W})$, and denote $\mathcal{W}_0 := \mathcal{W}$. For $j \in \{1,\ldots,k\}$, iteratively define $\mathcal{W}_j = \mathcal J_{\mathcal{W}_{j-1}}(M_j)$. Then $\psi(M_k,\ldots,M_1) := (\mathcal{W}_k {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_{k-1} {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \cdots {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_0)$.
\item \cite[Theorem~4.3]{BM_wide} Let $(M_k,\ldots,M_1) \in \mathsf{tfo}(\mathcal{W})$. For $j \in \{1,\ldots,k\}$, let $\mathcal{W}_j = \mathcal J_\mathcal{W}\left(\bigoplus_{i \leq j} M_i\right)$. Then $\psi \circ \chi(M_k,\ldots,M_1) = (\mathcal{W}_k {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_{k-1} {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \cdots {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_1 {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W})$.
\end{enumerate}
\end{theorem}
As a special case, note that $\tau$-exceptional sequences of length one are precisely the indecomposable $\tau$-rigid modules. In particular, suppose
$\mathcal{U} {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}$ is a cover relation in $\tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$, or equivalently that $(\mathcal{U} {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}) \in \mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}(\mathcal{W}))$. Then by Theorem~\ref{thm:tf_tau}(2), there exists a unique indecomposable module $M \in \mathcal{W}$ which is $\tau$-rigid and satisfies $\mathcal{U} = \mathcal J_\mathcal{W}(M)$. This leads to the following definition.
\begin{definition}\label{def:brick_label}
Let $\mathcal{U} {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}$ be a cover relation in $\tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$, and let $M \in \mathcal{W}$ be the unique indecomposable $\tau$-rigid module which satisfies $\mathcal{U} = \mathcal J_\mathcal{W}(M)$. We refer to $\beta(M)$ as the \emph{brick label} of $U {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}$.
\end{definition}
\begin{remark}\label{rem:brick_label}
We have chosen to label $\mathcal{U} {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}$ with $\beta(M)$ rather than with $M$ for two reasons. First, as discussed in Remark~\ref{rem:brick}, the property of being a brick is stable under passing to a wide subcategory, while the property of being $\tau$-rigid may not be. Second, we have by construction that $\Gen_\mathcal{W} M = \mathsf{Filt}(\Gen_\mathcal{W} M)$ is completely join-irreducible. Thus one can read $\beta(M)$ from the brick labeling of $\mathsf{tors}(\mathcal{W})$, while in general finding $M$ requires more information. This is the approach we will use to construct ``$\kappa$-exceptional sequences'' in Section~\ref{sec:kappa_exceptional}.
\end{remark}
The brick-labeling of $\tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$ will play a major role in the remainder of this paper. Indeed, we will show in Section~\ref{sec:el} that for many algebras, this labeling can be made into an ``EL-labeling''. In order to do so, it is useful to understand which bricks can appear as brick labels in $\tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$. To that end, we have the following definition.
\begin{definition}\label{def:sfbrick}
Let $B \in \mathcal{W}$ be a brick. We say that $B$ is an \emph{sf-brick} if there exists a semibrick $\mathcal{X}$ with $B \in \mathcal{X}$ such that $\mathsf{Filt}(\Gen_\mathcal{W} \mathcal{X})$ is functorially finite.
\end{definition}
We conclude this section with the following.
\begin{proposition}\label{prop:sf-brick}
Let $B \in \mathcal{W}$ be a brick. Then there exists a cover relation $U {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{V}$ in $\tau\text{-}\mathsf{perp}(\mathcal{W})$ with label $B$ if and only if $B$ is an sf-brick.
\end{proposition}
\begin{proof}
Suppose first that $B$ labels a cover relation $\mathcal{U} {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{V}$ in $\tau\text{-}\mathsf{perp}(\mathcal{W})$. If $\mathcal{V} = \mathcal{W}$, this means $\mathsf{Filt}(\Gen_\mathcal{W} B) = \Gen_\mathcal{W}(\beta^{-1}_\mathcal{W}(B))$ is a functorially finite torsion class of $\mathcal{W}$ and we are done. Otherwise, there exists a $\tau$-exceptional sequence $\sigma \in \tau\text{-}\mathsf{ex}(\mathcal{W})$ of the form $\sigma = (\beta^{-1}_\mathcal{V}(B), M_{k-1},\ldots,M_1)$. Let $\chi^{-1} (\sigma) = (N_k,\ldots,N_1) \in \mathsf{tfo}(\mathcal{W})$ be the corresponding TF-admissible ordered $\tau$-rigid module in $\mathcal{W}$, and denote $N := \bigoplus_{i = 1}^k N_i$. By Theorem~\ref{thm:tf_tau}, it follows that $\mathcal{U} = \mathcal J_\mathcal{W}(V)$. Now let $N' \in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ be the unique gen-minimal $\tau$-rigid module which satisfies $\Gen_\mathcal{W}(N') = \Gen_\mathcal{W}(N)$. Then $N_k$ must be a direct summand of $N'$ by the assumption that $(N_k,\ldots,N_1) \in \mathsf{tfo}(\mathcal{W})$. It then follows from Theorem~\ref{thm:DIJ}(3) that $B \in \mathcal{X}(N')$ and that $\mathsf{Filt}(\Gen_\mathcal{W}(\mathcal{X}(N'))) = \Gen_\mathcal{W}(N')$ is functorially finite in $\mathcal{W}$.
Now suppose that $B$ is an sf-brick and let $\mathcal{S} \in \mathsf{sbrick}(\mathcal{W})$ be a semibrick with $B \in \mathcal{S}$ such that $\mathsf{Filt}(\Gen_\mathcal{W} \mathcal{S})$ is functorially finite. By Theorem~\ref{thm:DIJ}(3), there exists $M \in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ which is gen-minimal and satisfies $\Gen_\mathcal{W}(M) = \mathsf{Filt}(\Gen_\mathcal{W} \mathcal{S})$. Since $M$ is gen-minimal, any ordering of its direct summands gives a TF-admissible ordered $\tau$-rigid module. Thus we can choose some $\rho = (M_k,\ldots,M_1) \in \mathsf{tfo}(\mathcal{W})$ such that $M = \bigoplus_{i = 1}^k M_i$ and $B = \beta\left(f_{\mathcal{F}_k} M_k\right)$ in the notation of Theorem~\ref{thm:DIJ}(3). By Theorem~\ref{thm:tf_tau}, this means that $B$ is the brick label of the cover relation $\mathcal J_\mathcal{W}(M) {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal J_\mathcal{W}\left(\bigoplus_{i < k} M_i\right)$ in $\tau\text{-}\mathsf{perp}(\mathcal{W})$.
\end{proof}
\begin{remark}
We note that there may exist sf-bricks $B$ for which the torsion class $\mathsf{Filt}(\Gen_\mathcal{W} B)$ is not functorially finite. See \cite[Example~3.13]{asai} for an example.
\end{remark}
\section{An EL-labeling for the lattice of wide subcategories}\label{sec:el}
In this section, we consider a special type of $\tau$-exceptional sequence, which we call \emph{hom-orthogonal chains}. We show that (up to permutation) at most one hom-orthogonal chain can exist between two $\tau$-perpendicular subcategories $\mathcal{U} \leq_\tau \mathcal{W}$. We then give examples of algebras where hom-orthogonal chains always exist. As a consequence, we conclude the the lattices of wide subcategories of these algebras admit EL-labelings.
We begin by recalling the definition of an edge-lexicographic (EL) labeling from \cite[Definition~3.2.1]{wachs}.
Let $P$ be a poset with smallest element $\hat{0}$ and largest element $\hat{1}$.
We let $\mathrm{label}: P\to Q$ be an edge-labeling, where $Q$ is a poset.
Frequently, $Q$ is taken to be the integers with their usual total order.
We will take $Q$ to be the set of bricks of $\Lambda$, which we will totally order (see Definition~\ref{def: rho}).
To each maximal chain $\sigma = (\hat{0}=x_{k+1} {\,\,<\!\!\!\!\cdot\,\,\,} x_k {\,\,<\!\!\!\!\cdot\,\,\,} x_{k-1}{\,\,<\!\!\!\!\cdot\,\,\,} \cdots {\,\,<\!\!\!\!\cdot\,\,\,} x_1 {\,\,<\!\!\!\!\cdot\,\,\,} x_0= \hat{1})$, we associate a tuple
$(\mathrm{label}(x_{k}, x_{k-1}]), \mathrm{label}([x_{k-1} x_{k-2}]), \ldots, \mathrm{label}([x_1, x_0]))$.
We totally order the set of all such tuples in reflected lexicographic order (reading left-to-right).
Equivalently, this is the usual lexicographic order, reading tuples from right-to-left.
We say a $\sigma$ is \emph{increasing} if $\mathrm{label}([x_{i+1}, x_{i}]) > \mathrm{label}([x_{i}, x_{i-1}])$ for each $i\in [1,k]$.
\begin{definition}\label{def:EL-label}
We say that $\mathrm{label}:P\to Q$ is an edge-lexicographic (EL) labeling for $P$ provided that for each interval $[x,y]$ in $P$ there is a unique increasing maximal chain, and furthermore that this increasing chain is smallest in the reflected lexicographic order.
\end{definition}
We now recall the definitions of $\beta, \mathcal{X}, \chi$, and $\psi$ from Theorems~\ref{thm:DIJ} and~\ref{thm:tf_tau}.
\begin{definition}
Let $\mathcal{W} \in \tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$, and let $\sigma \in \mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}\mathcal{W})$. Consider the corresponding $\tau$-exceptional sequence $\psi^{-1}(\sigma) =:(U_k,\ldots,U_1)$. We say that $\sigma$ is a \emph{hom-orthogonal chain} if $\mathrm{Hom}_\Lambda(\beta(U_j),\beta(U_i)) = 0$ for all $1 \leq i < j \leq k$.
\end{definition}
The following technical lemma is the basis of showing that, up to permutation, there can exist at most one hom-orthogonal chain with a designated target category.
\begin{lemma}\label{lem:unique_hom_ortho_chain}
Let $\mathcal{W} \in \tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$, let $\sigma = (\mathcal{W}_k {\,\,<\!\!\!\!\cdot\,\,\,} \cdots {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{W}_0 = \mathcal{W}) \in \mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}\mathcal{W})$, and suppose that $\sigma$ is a hom-orthogonal chain. Let $\psi^{-1}(\sigma)=:(U_k,\ldots,U_1)$ and $(\psi\circ\chi)^{-1}(\sigma)=:(M_k,\ldots,M_1)$ be the $\tau$-exceptional sequence and TF-admissible ordered $\tau$-rigid module corresponding to $\sigma$. Denote $M := \bigoplus_{i = 1}^k M_i$. Then $\mathcal{S} := \{\beta(U_k),\ldots,\beta(U_1)\}$ is the unique semibrick which satisfies $\mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S}) = \Gen_\mathcal{W} M$.
\end{lemma}
\begin{proof}
To show that $\mathcal{S}$ is a semibrick, it suffices to show that $\mathrm{Hom}(U_i, U_j) = 0$ for $i \neq j$. If $i < j$, this follows from the fact that $U_j \in \mathcal J_{\mathcal{W}_{i-1}}(U_i) \subseteq U_i^\perp$. If $i > j$, this follows from the assumption that $\sigma$ is hom-orthogonal.
It remains to show that $\mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S}) = \Gen_\mathcal{W}\left(\bigoplus_{i = 1}^k M_i\right)$. (The uniqueness of $\mathcal{S}$ then follows from Remark~\ref{rem:sbrick_unique}.) We first note that, by the definition of $\chi$, we have $U_i \in \Gen_\mathcal{W}(M_i)$ for each $i$. Since torsion classes are closed under quotients in $\mathcal{W}$, this implies that $\mathsf{Filt}\Gen_\mathcal{W}(\mathcal{S}) \subseteq \Gen_\mathcal{W}\left(\bigoplus_{i = 1}^k M_i\right)$. It thus suffices to show that $M_i \in \mathsf{Filt}\Gen_\mathcal{W}(\mathcal{S})$ for all $i$. We proceed by induction on $i$. For $i = 1$, we have $M_i = U_i$, so there is nothing to show. For the inductive step, denote $M_{<i}:= \bigoplus_{j = 1}^{i-1} M_i$. we consider the canonical exact sequence
$$0 \rightarrow \ker(q) \rightarrow M_i \xrightarrow{q} U_i \rightarrow 0$$
associated to the torsion pair $(\Gen_\mathcal{W} M_{<i} ,(M_{<i})^{\perp_\mathcal{W}})$. Then $\ker(q) \in \Gen_\mathcal{W} M_{<i} \subseteq \mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S})$ by the induction hypothesis. Since torsion classes are closed under extensions, this implies that $M_i \in \mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S})$, as desired.
\end{proof}
\begin{proposition}\label{prop:uniqueness}
Let $\mathcal{U} \leq_\tau \mathcal{W}$ be $\tau$-perpendicular subcategories, and suppose there exist hom-orthogonal chains
$\sigma = (\mathcal{U} = \mathcal{W}_k {\,\,<\!\!\!\!\cdot\,\,\,} \cdots {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{W}_0 = \mathcal{W})$ and $\rho = (\mathcal{U} = \mathcal{V}_k {\,\,<\!\!\!\!\cdot\,\,\,} \cdots {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{V}_0 = \mathcal{W})$. Denote $\psi^{-1}(\sigma) =: (U_k,\ldots,U_1)$ and $\psi^{-1}(\rho) =: (V_k,\ldots,V_1)$ the $\tau$-exceptional sequences corresponding to $\sigma$ and $\rho$, respectively. Then there exists a bijection $f: \{1,\ldots,k\} \rightarrow \{1,\ldots,k\}$ such that $\beta(U_i) = \beta(V_{f(i)})$ for all $i$.
\end{proposition}
\begin{proof}
Let $M$ be the direct sum of the modules which are indecomposable split projective in $\lperpW{U}$. Then by Lemma~\ref{lem:unique_hom_ortho_chain} we have that $\{\beta(U_k),\ldots,\beta(U_1)\} = \mathcal{X}(M) = \{\beta(V_k),\ldots,\beta(V_1)\}$.
\end{proof}
We now establish sufficient criteria for the existence of hom-orthogonal chains.
\begin{definition}\label{def: rho}
We say a total order $\preceq$ on the set of sf-bricks in $\mathsf{mod}\Lambda$ is \emph{reverse-hom-orthogonal} or \emph{rho} if for all sf-bricks $B \neq C \in \mathsf{mod}\Lambda$ we have that $B \preceq C$ implies $\mathrm{Hom}_\Lambda(C,B) = 0$.
\end{definition}
\begin{remark}\label{rem: rho}
Let $B$ and $C$ be sf-bricks, and suppose that $\mathsf{Filt}(\Gen B) \subseteq \mathsf{Filt}(\Gen C)$. Then in particular $B \in \mathsf{Filt}(\Gen C)$, and so $\mathrm{Hom}_\Lambda(C,B) \neq 0$. Thus one must have $C \preceq B$ in any rho order. In particular, by identifying bricks with completely join-irreducible objects, any rho order must refine the reverse of the partial order inherited from $\mathsf{tors}(\mathsf{mod}\Lambda)$. See Remark~\ref{rem:brick_lab}(2).
\end{remark}
The main example of algebras admitting an rho order are the so-called \emph{representation-directed} algebras. These algebras are characterized by having no cycles in their Auslander-Reiten quivers. As a consequence of this definition, it turns out that every representation-directed algebra is representation finite (and thus $\tau$-tilting finite), and that every indecomposable module over a representation-directed algebra is a $\tau$-rigid sf-brick. See \cite[Corollary~IX.3.4]{ASS} and
\cite[Proposition~7.1]{TW2}. Some examples of representation-directed algebras are hereditary and tilted algebras of Dynkin type. See \cite[Proposition~IX.6.5]{ASS}. More generally, we have that an algebra $\Lambda$ admits an rho order if and only if the lattice $\mathsf{tors}\Lambda$ is \emph{extremal}. See \cite[Appendix~A]{keller}.
We now show that the existence of an rho order guarantees the existence of hom-orthogonal chains. To be precise, we consider the following definitions. Note that the following notations are well-defined by Proposition~\ref{prop:sf-brick}.
\begin{definition} Let $\preceq$ be an rho order on the set of sf-bricks in $\mathsf{mod}\Lambda$, and let $\mathcal{V} \leq_\tau \mathcal{W}$ be $\tau$-perpendicular subcategories.
\begin{enumerate}
\item Let $\sigma = (\mathcal{V} = \mathcal{W}_k {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \cdots {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_0 = \mathcal{W})$ and $\rho = (\mathcal{V} = \mathcal{V}_k {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \cdots {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{V}_0 = \mathcal{W})$ be chains in $\mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}\mathcal{W})$ which end in $\mathcal{V}$. Denote $\psi^{-1}(\sigma) =: (M_k,\ldots,M_1)$ and $\psi^{-1}(\rho) =: (N_k,\ldots,N_1)$ the $\tau$-exceptional sequences corresponding to $\sigma$ and $\rho$. We write $\sigma \preceq_{lex} \rho$ if $(\beta(M_k),\ldots,\beta(M_1)) \preceq_{lex} (\beta(N_k),\ldots,\beta(N_1))$, where $\preceq_{lex}$ is the (reflected) lexicographic order on $k$-tuples of bricks.
\item Let $\sigma = (\mathcal{V} = \mathcal{W}_k {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \cdots {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_0 = \mathcal{W})$ be a chain in $\mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}\mathcal{W})$ which ends in $\mathcal{V}$. We say $\sigma$ is \emph{increasing} if $\beta(M_j) \preceq \beta(M_i)$ for all $1 \leq i < j \leq k$.
\end{enumerate}
\end{definition}
\begin{lemma}\label{lem:existence}
Let $\preceq$ be an rho order on the set of sf-bricks in $\mathsf{mod}\Lambda$, and let $\mathcal{V} \leq_\tau \mathcal{W}$ be wide subcategories. Let $\sigma = (\mathcal{V} = \mathcal{W}_k {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \cdots {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_0 = \mathcal{W})$ be smallest with respect to $\preceq_{lex}$ amongst the chains in $\mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}\mathcal{W})$ which end in $\mathcal{V}$. Then
\begin{enumerate}
\item $\sigma$ is increasing.
\item $\sigma$ is a hom-orthogonal chain.
\item $\sigma$ is the unique increasing chain in $\mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}\mathcal{W})$ which ends in $\mathcal{V}$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) Denote by $(M_k,\ldots,M_1):= \psi^{-1}(\sigma)$ the $\tau$-exceptional sequence corresponding to $\sigma$. We prove the result by
induction on $k:= \mathrm{rk}(\mathcal{W}) - \mathrm{rk}(\mathcal{V})$.
There is nothing to prove if $k = 1$, so suppose $k > 1$. By assumption, $\beta(M_1)$ is minimal amongst those bricks $B$ for which $\mathcal{V} \subseteq \mathcal J_\mathcal{W}(\beta_\mathcal{W}^{-1}B)$. It follows that $\psi(M_k,\ldots,M_2)$ is minimal amongst the chains in $\mathsf{stop\text{-}chain}(\mathcal J(M_1))$ which end at $\mathcal{V}$. This means $\psi(M_k,\ldots,M_2)$ is increasing by the induction hypothesis. It remains only to show that $\beta(M_1) \preceq \beta(M_2)$.
By Theorem~\ref{thm:tf_tau}, there exists $N \in \mathcal{W}$ such that $N\oplus M_1$ is $\tau$-rigid in $\mathcal{W}$ and satisfies $\mathcal J_\mathcal{W}(N \oplus M_1) = \mathcal J_{\mathcal J_\mathcal{W}(M_1)}(M_2)$ and $M_2 \in \Gen_{\mathcal J_\mathcal{W}(M_1)}N$. In particular, $\beta(M_2) \in \mathsf{Filt}(\Gen_\mathcal{W}\beta(N))$, and so $\mathrm{Hom}(\beta(N),\beta(M_2)) \neq 0$. It follows that $\beta(N) \preceq \beta(M_2)$. Finally, we have that $\mathcal{V} \leq_\tau \mathcal J_\mathcal{W}(N)$, and so $\beta(M_1) \preceq \beta(N)$ by the minimality of $\beta(M_1)$. This concludes the proof.
(2) It is immediate from the definitions that any increasing chain must be hom-orthogonal.
(3) Let $\rho$ be an increasing chain in $\mathsf{stop\text{-}chain}(\tau\text{-}\mathsf{perp}\mathcal{W})$ which ends in $\mathcal{V}$. We will show that $\rho = \sigma$. Let $(N_k,\ldots,N_1) = \psi^{-1}(\rho)$ the the $\tau$-exceptional sequence corresponding to $\rho$. Note that by (1), we have that both $\sigma$ and $\rho$ are hom-orthogonal chains. Thus by Proposition~\ref{prop:uniqueness} there exists a bijection $f: \{1,\ldots,k\} \rightarrow \{1,\ldots,k\}$ such that $\beta(M_i) = \beta(N_{f(i)})$ for all $i$. Since both $\sigma$ and $\rho$ are increasing, this is only possible if $f$ is the identity. We conclude that $\psi^{-1}(\sigma) = \psi^{-1}(\rho)$, which implies the result.
\end{proof}
We now conclude our first main theorem.
\begin{theorem}[Theorem~\ref{thm:intro:mainA}]\label{thm:mainA}
Suppose the set of sf-bricks in $\mathsf{mod}\Lambda$ admits an rho order $\preceq$. Then the brick-labeling of $\tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$ is an EL-labeling with respect to the total order $\preceq$.
\end{theorem}
\begin{proof}
The existence and uniqueness of a increasing chain are precisely Lemma~\ref{lem:existence} parts (1) and (3), respectively.
\end{proof}
In particular, Theorem~\ref{thm:mainA} implies that hereditary and tilted algebras of Dynkin type have lattices of wide subcategories which admit EL-labelings.
\section{The flag property for canonical join representations}\label{sec:flag}
Let $L$ be a completely semidistributive lattice. Generalizing the finite case from \cite{emily_canonical}, we recall that the \emph{canonical join complex} of $L$ is the simplicial complex with underlying set $\cji(L)$ such that a subset $A \subseteq \cji(L)$ spans a simplex if and only if $\Join A$ is a canonical join representation. The main result of \cite{emily_canonical} is that if $L$ is finite, then the canonical join complex is flag.
Now let us also recall from \cite{BCZ} that when $L = \mathsf{tors}\Lambda$ for some finite-dimensional algebra $\Lambda$, we can identify $\cji(L)$ with the set of bricks in $\mathsf{mod}\Lambda$. It is then shown in \cite[Theorem~1.8]{BCZ} that a set of bricks $A$ joins canonically if and only if $A$ is a semibrick (see Theorem~\ref{thm:brick_labeling}). This is also a flag condition, in the sense that $A$ is a semibrick if and only if every subset $B \subseteq A$ with $|B| = 2$ is a semibrick. In particular, the canonical join complex of $\mathsf{tors}\Lambda$ is also flag.
The purpose of this section is to generalize these results by proving the flag condition for a large family of (infinite) completely semidistributive lattices. In doing so, we also give a positive answer to \cite[Question~4.13]{enomoto} for this family (see Corollary~\ref{cor:extended_kappa_bij}). The lattices we consider are defined as follows.
\begin{definition}\label{def:kappa_lattice}
Let $L$ be a completely semidistributive lattice. We say that $L$ is a \emph{well-separated completely semidistributive lattice}, abbreviated \emph{ws-csd-lattice} if for any $x \not\leq y \in L$ there exists $j \in \cji(L)$ such that $j \leq x$ and $y \leq \kappa(j)$.
\end{definition}
\begin{remark}\label{rem:kappa_lattice}\
We refer readers to \cite[Sections~1 and~3.1]{RST} for a detailed discussion of the well-separated property and other conditions on infinite semidistributive lattices. In particular, we point out the following.
\begin{enumerate}
\item Every finite (completely) semidistributive lattice is a ws-csd-lattice.
\item Both lattices of torsion classes are always ws-csd lattices, see \cite[Section~8.2]{RST}.
\item Suppose $y \lneq x \in L$. Then the element $j$ coming from Definition~\ref{def:kappa_lattice} must lie in $\mathrm{j\text{-}label}[y,x]$ by Proposition~\ref{prop:cover}(1). In particular, this means the interval $[y,x]$ contains a cover relation. Lattices with this property are sometimes called ``weakly atomic'', see \cite[Section~3.1]{RST}.
\end{enumerate}
\end{remark}
We now prove the main result of this section, which generalizes \cite[Theorem~5.13]{RST} and \cite[Proposition~4.9]{enomoto}.
\begin{theorem}[Theorem~\ref{thm:intro:mainC}]\label{thm:covers_canonical}
Let $L$ be a ws-csd-lattice. Let $A \subseteq \cji(L)$ and $x = \Join A$. Then the following are equivalent.
\begin{enumerate}
\item $A$ joins canonically; i.e., $A = \CJR(x)$.
\item $\Join A$ is an irredundant join representation of $x$ which refines every other irredundant join representation of $x$.
\item For all $i\neq j \in A$, one has $i \leq \kappa(j)$.
\item $A$ is an antichain and every $j \in A$ labels a cover relation of the form $u {\,\,<\!\!\!\!\cdot\,\,\,} x$.
\end{enumerate}
Moreover, if (1-4) hold, then $A$ is precisely the set of elements which label cover relations of the form $u {\,\,<\!\!\!\!\cdot\,\,\,} x$.
\end{theorem}
\begin{proof}
Note that the moreover part can be found as \cite[Lemma~4.10]{enomoto}.
$(1 \implies 2)$: Suppose $A$ join canonically and suppose for a contradiction that $\Join A$ is not irredundant. Thus there exists $j \in A$ such that $x = \Join (A\setminus\{j\})$. By definition, $\Join A$ must refine $\Join (A\setminus\{j\})$, so there exists $i \in A \setminus \{j\}$ such that $j \leq i$. This means that $A$ is not an antichain, which contradicts Definition~\ref{def:canonical_join_rep}(2).
$(2 \implies 3)$: This follows using an argument nearly identical to that of \cite[Lemma~4.8]{enomoto}. Indeed, suppose (2) holds and that there exist $i, j \in A$ such that $i \not\leq \kappa(j)$. As in the proof of \cite[Lemma~4.8]{enomoto}, this means that $i \vee j = i \vee j_*$ by \cite[Lemma~2.57]{FJN}. Denoting $B:= (A \setminus\{j\}) \cup \{j_*\})$, we thus have a join representation $x = \Join B$.
We will show that $\Join B$ is irredundant.
Observe that $\Join (B\setminus\{j_*\}) = \Join A\setminus \{j\} \lneq \Join A$ because $\Join A$ is irredundant.
Now, consider $\Join (B\setminus \{i\})$ for some $i\neq j_*$.
Then $\Join (B\setminus \{i\}) \leq \Join (A\setminus \{i\}) \lneq \Join A$ because $\Join A$ is irredundant.
Therefore $\Join B$ is irredundant.
By assumption we have that $\Join A$ refines $\Join B$. Since $j \not\leq j_*$, this means there exists $j \neq i \in A$ such that $j \leq i$, which contradicts that $\Join A$ is irredundant.
$(3 \implies 4)$: Suppose (2) holds. The fact that $A$ is an antichain is immediate from the fact that $i \not \leq \kappa(j)$. It then follows from Proposition~\ref{prop:cover}(2) that every $j \in A$ labels a cover relation of the form $u {\,\,<\!\!\!\!\cdot\,\,\,} x$.
$(4 \implies 1)$: The proof is similar to that of \cite[Proposition~3.7]{BCZ}. Let $x = \Join B$ be a join representation. We will show that $\Join A$ refines $\Join B$.
Let $j \in A$, and let $u {\,\,<\!\!\!\!\cdot\,\,\,} x$ be labeled by $j$. By Proposition~\ref{prop:cover}(1), this means $\kappa(j) \geq u$. Now there exists $b \in B$ such that $b \not \leq u$ (otherwise we would have $\Join B = u$).
By Proposition~\ref{prop:cover}(3), we have $u= \kappa(j)\wedge x$ and $b\le x$.
Thus also we have $b \not \leq \kappa(j)$. Definition~\ref{def:kappa_lattice} then says that there exists $i \in \cji(L)$ such that $i \leq b$ and $\kappa(j) \leq \kappa(i)$. We will show that $i = j$, which will imply that $\Join A$ refines $\Join B$.
First note that $x \not \leq \kappa(i)$ since $i \leq x$. Thus $x \wedge \kappa(j) \leq x \wedge \kappa(i) \lneq x$. By Proposition~\ref{prop:cover}(3), this means $u = x \wedge \kappa(j) = x \wedge \kappa(i)$. It follows that $i \not \leq u$, and so $x = u \vee i$. Using Proposition~\ref{prop:cover}(3) once again, we conclude that $i = \mathrm{j\text{-}label}[u, x] = j$.
\end{proof}
As a consequence, we obtain the flag condition for the canonical join complex.
\begin{corollary}[Theorem~\ref{thm:intro:mainB}]\label{cor:flag}
Let $L$ be a ws-csd-lattice. Then the canonical join complex of $L$ is a flag simplicial complex.
\end{corollary}
\begin{proof}
Let $A \subseteq \cji(L)$. It is clear from Condition~2 in Theorem~\ref{thm:covers_canonical} that $A$ joins canonically if and only if every subset $B \subseteq A$ with $|B| = 2$ joins canonically. This implies the result.
\end{proof}
\begin{remark}
Suppose $L = \mathsf{tors}\Lambda$ is the lattice of torsion classes of a finite-dimensional algebra. Consider two distinct bricks $B \neq C \in \mathsf{mod}\Lambda$. Then by \cite[Theorem~8.6]{RST}, we have that $\mathsf{Filt}(\Gen B) \leq \kappa(\mathsf{Filt}(\Gen C))$ if and only if $\mathrm{Hom}_\Lambda(B,C) = 0$. That is, Condition~2 in Theorem~\ref{thm:covers_canonical} is precisely the semibrick condition on bricks used in \cite{BCZ}.
\end{remark}
We conclude this section with the following, which shows that the answer to \cite[Question~4.13]{RST} is ``yes'' when $L$ is a ws-csd lattice.
\begin{corollary}\label{cor:extended_kappa_bij}
Let $L$ be a ws-csd-lattice and let $x \in \mathsf{cj\text{-}rep}(L)$. Then $\overline{\kappa}(x) \in \mathsf{cm\text{-}rep}(L)$ and $\CMR(\overline{\kappa}(x)) = \{\kappa(j) \mid j \in \CJR(x)\}$. In particular, the extended kappa-maps $\overline{\kappa}$ and $\overline{\kappa}^d$ give inverse bijections between $\mathsf{cj\text{-}rep}(L)$ and $\mathsf{cm\text{-}rep}(L)$.
\end{corollary}
\begin{proof}
Let $x \in \mathsf{cj\text{-}rep}(L)$, and denote $A = \{\kappa(j) \mid j \in \CJR(x)\}$. Then for all $\kappa(i) \neq \kappa(j) \in A$, Condition~2 in Theorem~\ref{thm:covers_canonical} tells us that $\kappa(i) \geq \kappa^d(\kappa(j))$. This precisely says that the dual of Condition~2 holds for $\overline{\kappa}(x) = \bigwedge A$, and so $\bigwedge A$ is a canonical meet representation. The fact that $\overline{\kappa}$ and $\overline{\kappa}^d$ are inverse bijections then follows from duality.
\end{proof}
\section{The $\kappa$-order and core label orders}\label{sec:kappa_order}
In this section, we recall three additional partial orders which can be placed on the canonically join-representable elements of a ws-csd lattice $L$: the $\kappa$-order, the upper core label order, and the lower core label order (or shard intersection order). We show that the $\kappa$-order is the common refinement of the two core label orders and give necessary and sufficient conditions for these orders to coincide.
\subsection{Labels of intervals}\label{sec:intervals}
Fix a ws-csd lattice $L$. We note that not all of the definitions in this section will require the assumption that $L$ be well-separated, but we keep this assumption on $L$ for readability.
As a starting point, we need the following result, which essentially follows from \cite[Theorem~4.3]{RST}.
\begin{lemma}\label{prop:intervals}
Let $x \leq y \in L$. Then the interval $[x,y]$ is also a ws-csd-lattice with the join and meet operations inherited from $L$. Moreover:
\begin{enumerate}
\item There is a bijection $\mathrm{j\text{-}label}[x,y] \rightarrow \cji[x,y]$ given by $j \mapsto x \vee j$.
\item Identifying each $j$ with its image under this bijection, the label of any cover relation in the interval $[x,y]$ is the same whether computed in $L$ or in the sublattice $[x,y]$.
\end{enumerate}
\end{lemma}
As a consequence of Lemma~\ref{prop:intervals}, every atom of the sublattice $[x,y]$ is of the form $x \vee j$ for some $j \in \cji(L)$. Thus the following is well-defined.
\begin{definition}\label{def:atom}
Let $x \leq y \in L$. We say a completely join-irreducible element $j \in \mathrm{j\text{-}label}[x,y]$ is an \emph{atom} of the interval $[x,y]$ if $x \vee j$ is an atom in the sublattice $[x,y]$. We denote by $\mathrm{atom}[x,y]$ the set of atoms in $[x,y]$.
Dually, we say a completely meet-irreducible element $\kappa(j) \in \mathrm{m\text{-}label}[x,y]$ is a \emph{coatom} of $[x,y]$ if $y\wedge \kappa(j)$ is a coatom in the sublattice $[x,y]$, and we denote by $\mathrm{coatom}[x,y]$ the set of coatoms in $[x,y]$.
\end{definition}
\begin{example}
Let $L$ be the lattice in Figure~\ref{fig:run_ex}, and consider the interval $[j_4,\hat{1}]$. The elements which are complely join-irreducible in $[j_4,\hat{1}]$ are then $m_1 = j_4 \vee j_2$ and $m_2 = j_4 \vee j_1$. By identifying $m_1$ with $j_2$ and $m_2$ with $j_1$, it is then possible to verify (1) and (2) in Lemma~\ref{prop:intervals}. Moreover, we have $\mathrm{atom}[j_4,\hat{1}] = \{j_1,j_2\}$.
\end{example}
We now consider special types of intervals in which the atoms and coatoms are in bijection with one another.
\begin{definition}\label{def:nuclear}
Let $x \leq y \in L$. We say that $[x,y]$ is a \emph{nuclear interval} if
$$x = \bigwedge \{y \wedge \kappa(j) \mid \kappa(j) \in \mathrm{coatom}[x,y]\}$$
and for all $x \leq z < y$ there exists $\kappa(j) \in \mathrm{coatom}[\hat{0},\hat{1}]$ such that $z \leq \kappa(j)$.
Dually we say that $[x,y]$ is a \emph{conuclear interval} if $$x = \Join \{y \vee j \mid j \in \mathrm{atom}[x,y]\}$$
and for all $x < z \leq y$ there exists $j \in \mathrm{atom}[\hat{0},\hat{1}]$ such that $j \leq z$.
\end{definition}
\begin{lemma}\label{lem:nuclear_canonical_join_rep}\
\begin{enumerate}
\item Suppose $[\hat{0},\hat{1}]$ is nuclear. Then $\hat{0} \in \mathsf{cm\text{-}rep}(L)$ and $\CMR(\hat{0}) = \mathrm{coatom}[\hat{0},\hat{1}]$.
\item Suppose $[\hat{0},\hat{1}]$ is conuclear. Then $\hat{1} \in \mathsf{cj\text{-}rep}(L)$ and $\CJR(\hat{1}) = \mathrm{atom}[\hat{0},\hat{1}]$.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove only (2) since (1) is dual. The assumption that $[\hat{0},\hat{1}]$ is conuclear implies that $\hat{1} = \Join \{j \mid j \in \mathrm{atom}[\hat{0},\hat{1}]\}$. Now for $i \neq j \in \mathrm{atom}[\hat{0},\hat{1}]$, it is clear that $i \vee j$ is a canonical join representation (because it is an irredundant join of atoms). The result thus follows from Corollary~\ref{cor:flag}.
\end{proof}
\begin{example}\label{ex:nuclear}
Let $L$ be the lattice in Figure~\ref{fig:run_ex}. Since $|L| < \infty$, every interval in $[x,y] \subseteq L$ satiisfies that (a) if $x < z \leq y$ then there exists $j \in \mathrm{atom}[x,y]$ such that $x \vee j \leq z$, and (b) if $x \leq z < y$ then there exists $\kappa(j) \in \mathrm{coatom}[x,y]$ such that $z \leq y \wedge \kappa(j)$. On the other hand, we have that $[\hat{0},m_1]$ is both nuclear and conuclear, while the interval $[j_3, \hat{1}]$ is neither.
\end{example}
As a special case of these constructions, we recall the following from \cite{AP}. See also \cite[Theorem~4.12, Proposition~4.13]{DIRRT} and \cite[Theorem~3.12]{jasso}, which explicitly address the case of functorially finite torsion classes (which will ultimately be all we need in Section~\ref{sec:kappa_tau}). For readability, given a wide subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$, we denote by $\mathrm{br\text{-}label}_\mathcal{W}$, $\mathrm{atom}_\mathcal{W}$, etc. the relevant constructions in the lattice $\mathsf{tors}(\mathcal{W})$.
\begin{theorem}\label{thm:asai_pfeifer_1}
Let $\Lambda$ be a finite dimensional algebra, let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ a wide subcategory, and let $\mathcal{U} \subseteq \mathcal{T}$ be torsion classes in $\mathsf{tors}(\mathcal{W})$. Then the following hold.
\begin{enumerate}
\item \cite[Theorem~5.2]{AP} The following are equivalent.
\begin{enumerate}
\item The interval $[\mathcal{U},\mathcal{T}]$ is nuclear.
\item The interval $[\mathcal{U},\mathcal{T}]$ is conuclear.
\item $\mathcal{U}^{\perp_\mathcal{W}} \cap \mathcal{T} \in \mathsf{wide}(\mathcal{W})$.
\end{enumerate}
\item \cite[Theorem~4.2]{AP} Suppose the equivalent conditions from (1) hold. Then there is a label-preserving isomorphism of lattices $[\mathcal{U},\mathcal{T}] \rightarrow \mathsf{wide}(\mathcal{U}^{\perp_\mathcal{W}} \cap \mathcal{T})$ given by $\mathcal{T}' \mapsto \mathcal{T}' \cap \mathcal{U}^{\perp_\mathcal{W}}.$ That is, for any cover relation $\mathcal{U} \subseteq \mathcal{U}' {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{T}' \subseteq \mathcal{T}$ in $\mathsf{tors}(\mathcal{W})$, one has that $\mathrm{br\text{-}label}_\mathcal{W}[\mathcal{U}',\mathcal{T}'] = \mathrm{br\text{-}label}_{\mathcal{U}^{\perp_\mathcal{W}} \cap \mathcal{T}}[\mathcal{U}' \cap \mathcal{U}^{\perp_\mathcal{W}}, \mathcal{T}' \cap \mathcal{U}^{\perp_\mathcal{W}}]$. In particular: \begin{enumerate}
\item $\mathrm{br\text{-}label}_\mathcal{W}[\mathcal{U},\mathcal{T}] = \mathsf{brick}(\mathcal{U}^{\perp_\mathcal{W}} \cap \mathcal{T})$.
\item $\mathrm{atom}_\mathcal{W}[\mathcal{U},\mathcal{T}] = \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \text{ is simple in }\mathcal{U}^{\perp_\mathcal{W}} \cap \mathcal{T}\}.$
\end{enumerate}
\end{enumerate}
\end{theorem}
The following two results extend Theorem~\ref{thm:asai_pfeifer_1}(1) to the generality of ws-csd lattices.
\begin{proposition}\label{prop:nuclear}
The following are equivalent.
\begin{enumerate}
\item $[\hat{0},\hat{1}]$ is a nuclear interval.
\item $[\hat{0},\hat{1}]$ is a conuclear interval.
\item $\hat{1} \in \mathsf{cj\text{-}rep}(L)$ and $\overline{\kappa}(\hat{1}) = \hat{0}$.
\item $\hat{0} \in \mathsf{cm\text{-}rep}(L)$ and $\overline{\kappa}^{d}(\hat{0}) = \hat{1}$.
\end{enumerate}
Moreover, if (1)-(4) all hold, then $\mathrm{coatom}[\hat{0},\hat{1}] = \{\kappa(j) \mid j \in \mathrm{atom}[\hat{0},\hat{1}]\}$.
\end{proposition}
\begin{proof}
The equivalence $(3 \iff 4)$ is an immediate consequence of Corollary~\ref{cor:extended_kappa_bij}, so we prove only $(2 \implies 3 \implies 1)$ and the moreover part.
The direction $(1\implies 4 \implies 2)$ is similar by duality.
$(2\implies 3)$: Suppose that $[\hat{0},\hat{1}]$ is conuclear. Then $\hat{1} \in \mathsf{cj\text{-}rep}(L)$ and $\CJR(\hat{1}) = \mathrm{atom}[\hat{0},\hat{1}]$ by Lemma~\ref{lem:nuclear_canonical_join_rep}.
We claim that $\overline{\kappa}(\hat{1}) = \bigwedge \mathrm{coatom}[\hat{0},\hat{1}]$. Indeed, the moreover part of Theorem~\ref{thm:covers_canonical} says that there is a bijection $\sigma$ between the atoms and coatoms of $[\hat{0},\hat{1}]$ such that for each $j \in \mathrm{atom}[\hat{0},\hat{1}]$ the cover relation $[\sigma(j),\hat{1}]$ is labeled by $j$. Proposition~\ref{prop:cover}(1) then implies that $\sigma(j) = \kappa(j)$, which proves the claim.
Now suppose for a contradiction that $\hat{0} \neq \overline{\kappa}(\hat{1})$. Then since $[\hat{0},\hat{1}]$ is conuclear, there exists $j \in \mathrm{atom}[\hat{0},\hat{1}]$ such that $j \leq \overline{\kappa}(\hat{1})$. On the other hand, we have $\overline{\kappa}(\hat{1}) \leq \kappa(j)$ by the claim. Since we know $j \not \leq \kappa(j)$, we have reached a contradiction.
$(3 \implies 1)$:
Suppose (3) holds. Then Proposition~\ref{prop:covers_canonical_existence} tells us that for every $\hat{0} \leq z < \hat{1}$ there exists $\kappa(j) \in \mathrm{coatom}[\hat{0},\hat{1}]$ such that $z \leq \kappa(j)$. Moreover, as in the proof of $(2 \implies 3)$, Theorem~\ref{thm:covers_canonical} implies that there is a bijection $\sigma: \CJR(\hat{1}) \rightarrow \mathrm{coatom}[\hat{0},\hat{1}]$ such that for all $j \in \CJR(\hat{1})$ the cover relation $\sigma(j) {\,\,<\!\!\!\!\cdot\,\,\,} \hat{1}$ is labeled by $j$. Proposition~\ref{prop:cover} then implies that $\sigma(j) = \kappa(j)$. It follows that $\hat{0} = \overline{\kappa}(\hat{1})$ is the meet of the coatoms of $L$.
Finally, we prove the moreover part. Suppose (1)-(4) all hold and let $j \in \mathrm{atom}[\hat{0},\hat{1}]$. We showed in the proof of $(2\implies 3)$ that $j$ labels a cover relation $\kappa(j) {\,\,<\!\!\!\!\cdot\,\,\,} \hat{1}$, which by definition says that $\kappa(j) \in \mathrm{coatom}[\hat{0},\hat{1}]$. The fact that every coatom is of this form then follows from the fact that conditions (1)-(4) taken together are self-dual.
\end{proof}
\begin{corollary}\label{cor:nuclear}
Let $x \leq y \in L$. Then $[x,y]$ is a nuclear interval if and only if it is a conuclear interval. Moreover, in this case we have $\mathrm{coatom}[x,y] = \{\kappa(j) \mid j \in \mathrm{atom}[x,y]\}$.
\end{corollary}
\begin{proof}
By Lemma~\ref{prop:intervals}, we have that the sublattice $[x,y]$ is a ws-csd lattice. The result thus follows by applying Proposition~\ref{prop:nuclear} to this sublattice.
\end{proof}
We are now prepared for the main definitions of this section.
\begin{definition}\label{def:pop}
Let $x \in \mathsf{cj\text{-}rep}(L)$.
\begin{enumerate}
\item We denote by $\mathrm{pop}_\downarrow(x)$ the meet of the elements which are covered by $x$; that is, $\mathrm{pop}_\downarrow(x) = x \wedge \left(\bigwedge \mathrm{coatom}[\hat{0},x]\right)$.
\item We denote by $\mathrm{pop}_\uparrow(x)$ the join of the elements which cover $x$; that is, $\mathrm{pop}_\uparrow(x) = x \vee \left(\Join \mathrm{atom}[x,\hat{1}]\right)$.
\item The \emph{lower core} of $x$ is the interval $\mathrm{core}_\downarrow(x) := [\mathrm{pop}_\downarrow(x),x]$. The \emph{lower core label set} of $x$ is $\mathrm{core}\text{-}\mathrm{lab}_\downarrow(x) := \mathrm{j\text{-}label}(\mathrm{core}_\downarrow(x))$.
\item The \emph{upper core} of $x$ is the interval $\mathrm{core}_\uparrow(x) := [\overline{\kappa}(x),\mathrm{pop}_\uparrow(\overline{\kappa}(x)]$. The \emph{upper core label set} of $x$ is $\mathrm{core}\text{-}\mathrm{lab}_\uparrow(x) := \mathrm{j\text{-}label}(\mathrm{core}_\uparrow(x))$.
\item We denote $W(x) := \mathrm{core}\text{-}\mathrm{lab}_\uparrow(x) \cap \mathrm{core}\text{-}\mathrm{lab}_\downarrow(x)$.
\end{enumerate}
\end{definition}
The operators ``$\mathrm{pop}_\downarrow$'' and ``$\mathrm{pop}_\uparrow$'' are sometimes called the \emph{pop-stack sorting operators}, and have been of recent interest in the field of dynamical combinatorics. See e.g. \cite[Section~1.2]{DW} and there references therein. The name ``core'' comes from the relation between these operators and the \emph{core-label order}, which we discuss in Section~\ref{sec:kappa_order}
\begin{remark}\label{rem:atoms} Recall that the labels of the upper covers of $\overline{\kappa}(x)$ coincide with those of the lower covers of $x$. Thus one can think of $\mathrm{core}_\downarrow(x)$ as the interval ``spanned'' by the lower covers of $x$. With this perspective, $\mathrm{core}_\uparrow(x)$ is ``spanned'' by the same join-irreducible elements. Said differently, one has $\mathrm{atom}(\mathrm{core}_\downarrow(x)) = \mathrm{atom}(\mathrm{core}_\uparrow(x))$.
\end{remark}
Note in particular that $\mathrm{pop}_\downarrow(\hat{0}) = \hat{0}$, and so $\mathrm{core}\text{-}\mathrm{lab}_\downarrow(\hat{0}) = \emptyset$. Moreover, we have that $\mathrm{core}_\downarrow(\hat{1}) = L$ (and thus that $\mathrm{core}\text{-}\mathrm{lab}_\uparrow(\hat{1}) = \cji(L)$) if and only if $L$ is (co)nuclear
\begin{example}\label{ex:pop}
Let $L$ be the lattice in Figure~\ref{fig:run_ex} and $x = m_1$. Recall that $\overline{\kappa}(x) = j_1$. Then
$\mathrm{pop}_\downarrow(x) = \hat{0}$ and $\mathrm{pop}_\uparrow(\overline{\kappa}(x)) = \hat{1}$. (Note in particular that $\mathrm{pop}_\uparrow(\mathrm{pop}_\downarrow(x)) = \hat{1} \neq x$.) We then have
$$\mathrm{core}\text{-}\mathrm{lab}_\downarrow(x) = \{j_2,j_3,j_4\} \qquad\qquad \mathrm{core}\text{-}\mathrm{lab}_\uparrow(x) = W(x) = \{j_2,j_3\}.$$
\end{example}
We now give an explicit example in the lattice of torsion classes. Recall the notation $\alpha(-)$ from Theorem~\ref{thm:IT}.
\begin{example}\label{ex:asai_pfeifer}
Let $\Lambda$ be a finite dimensional algebra and $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ a wide subcategory. Let $\mathcal{T} \in \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$. Then:
\begin{enumerate}
\item By \cite[Theorem~6.7]{AP}, ${\mathrm{core}_\downarrow}_\mathcal{W}(\mathcal{T}) = [{\mathrm{pop}_\downarrow}_\mathcal{W}(\mathcal{T}), \mathcal{T}]$ is a nuclear interval which satisfies $\mathcal{T} \cap ({\mathrm{pop}_\downarrow}_\mathcal{W}(\mathcal{T}))^{\perp_\mathcal{W}} = \alpha(\mathcal{T})$. In particular, ${\mathrm{br\text{-}label}}_\mathcal{W}({\mathrm{core}_\downarrow}_\mathcal{W}(\mathcal{T})) = \mathsf{brick}(\alpha(\mathcal{T}))$ by Theorem~\ref{thm:asai_pfeifer_1}(2a).
\item By construction and Theorem~\ref{thm:asai_pfeifer_1}(2b), we have $$\mathrm{atom}(\mathrm{core}_\downarrow(\mathcal{T})) = \mathrm{atom}(\mathrm{core}_\uparrow(\mathcal{T})) = \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \text{ is simple in }\alpha(\mathcal{T})\}.$$ Thus Theorem~\ref{thm:asai_pfeifer_1}(2) tells us there are (label-preserving) isomorphisms
$${\mathrm{core}_\downarrow}_\mathcal{W}(\mathcal{T}) \cong \mathsf{tors}(\alpha(\mathcal{T})) \cong {\mathrm{core}_\uparrow}_\mathcal{W}(\mathcal{T}).$$
In particular, this means $W_\mathcal{W}(\mathcal{T}) = \{\mathsf{Filt}(\Gen_\mathcal{W} B) \mid B \in \mathsf{brick}(\alpha(\mathcal{T}))\}$. This is our justification for using the symbol $W(x)$ in Definition~\ref{def:pop}(5).
\end{enumerate}
\end{example}
If $L$ is finite, the following is a straightforward consequence of the definitions. More generally, the difficulty lies in showing that the intervals in question are (co)atom-based.
\begin{proposition}\label{prop:core_atomic}
Let $x \in \mathsf{cj\text{-}rep}(L)$. Then both $\mathrm{core}_\downarrow(x)$ and $\mathrm{core}_\uparrow(x)$ are (co)nuclear.
\end{proposition}
\begin{proof}
We prove that $\mathrm{core}_\downarrow(x)$ is nuclear, as the result for $\mathrm{core}_\uparrow(x)$ then follows from Corollary~\ref{cor:extended_kappa_bij} and duality. Let $A = \CJR(x)$ be the canonical join representation of $x$ in the lattice $L$, and let $A' = \{\mathrm{pop}_\downarrow(x) \vee j \mid j \in A\}$. We claim that $A'$ is the canonical join representation of $x$ in the sublattice $\mathrm{core}_\downarrow(x)$. Indeed, we already know that every $\mathrm{pop}_\downarrow(x) \vee j \in A'$ is completely join-irreducible in $\mathrm{core}\text{-}\mathrm{lab}_\downarrow(x)$ by Lemma~\ref{prop:intervals}, and it is clear that $x = \Join A'$. Thus let $x = \Join B$ be a join representation of $x$ in the sublattice $\mathrm{core}_\downarrow(x)$. Then $\Join B$ is also a join representation of $x$ in $L$. Thus for every $j \in A$, there exists $b \in B$ such that $j \leq b$. Since also $\mathrm{pop}_\downarrow(x) \leq b$ by assumption, this means $x \vee j \leq b$ and so $A'$ refines $B$. This proves the claim.
It follows that for all $z$ with $\mathrm{pop}_\downarrow(x) \leq z < x$ there exists $\kappa(j) \in \mathrm{coatom}[\mathrm{pop}_\downarrow(x),x]$ such that $z \leq \kappa(j)$ by Proposition~\ref{prop:covers_canonical_existence} (since $x\in \mathsf{cj\text{-}rep}[\mathrm{core}\text{-}\mathrm{lab}_\downarrow(x)]$).
By definition $\mathrm{pop}_\downarrow(w)$ is equal to the meet of the coatoms of $\mathrm{core}_\downarrow(x)$.
Thus, $\mathrm{core}_\downarrow(x)$ is nuclear, hence also conuclear by Proposition~\ref{prop:nuclear}.
\end{proof}
The following can be deduced immediately from Proposition~\ref{prop:cover}.
\begin{proposition}\label{prop:core_formulas}
Let $x \in \mathsf{cj\text{-}rep}(L)$. Then
\begin{enumerate}
\item $\mathrm{core}\text{-}\mathrm{lab}_\downarrow(x) = \{j \in \cji(L) \mid j \leq x \text{ and } \kappa(j) \geq \mathrm{pop}_\downarrow(x)\}.$
\item $\mathrm{core}\text{-}\mathrm{lab}_\uparrow(x) = \{j \in \cji(L) \mid j \leq \mathrm{pop}_\uparrow(\overline{\kappa}(x)) \text{ and } \kappa(j) \geq \overline {\kappa}(x)\}$.
\item $W(x) = \{j \in \cji(L) \mid j \leq x \text{ and } \kappa(j) \geq \overline{\kappa}(x)\}$.
\end{enumerate}
\end{proposition}
As a consequence, we obtain the following relationship between $\mathrm{pop}_\downarrow$ and $\overline{\kappa}$. We note that this result has previously been shown for finite semidistributive lattices in \cite[Theorem~9.1]{DW}.
\begin{proposition}\label{prop:pop_formula}
Let $x \in L$. Then the following hold.
\begin{enumerate}
\item If $x \in \mathsf{cj\text{-}rep}(L)$, then $\mathrm{pop}_\downarrow(x) = x \wedge \overline{\kappa}(x)$ and $\CMR(\overline{\kappa}(x)) = \mathrm{coatom}(\mathrm{core}_\downarrow(x))$. Moreover, if in addition $\mathrm{pop}_\downarrow(x) \in \mathsf{cm\text{-}rep}(L)$, then $\CMR(\mathrm{pop}_\downarrow(x)) \supseteq \CMR(\overline{\kappa}(x))$, with equality if and only if $\mathrm{pop}_\downarrow(x) = \overline{\kappa}(x)$.
\item If $x \in \mathsf{cm\text{-}rep}(L)$, then $\mathrm{pop}_\uparrow(x) = x \vee \overline{\kappa}^{d}(x)$ and $\CJR(\overline{\kappa}^d(x)) = \mathrm{atom}(\mathrm{core}_\uparrow(x))$. Moreover, if in addition $\mathrm{pop}_\uparrow(x) \in \mathsf{cj\text{-}rep}(L)$, then $\CJR(\mathrm{pop}_\uparrow(x)) \supseteq \CJR(\overline{\kappa}^d(x))$, with equality if and only if $\mathrm{pop}_\uparrow(x) = \overline{\kappa}^d(x)$.
\end{enumerate}
\end{proposition}
\begin{proof}
We prove only (1) as the proof of (2) is dual. Consider the interval $\mathrm{core}_\downarrow(x) = [\mathrm{pop}_\downarrow(x),x]$, which is (co)nuclear by Proposition~\ref{prop:core_atomic}.
Hence, by Lemma~\ref{lem:nuclear_canonical_join_rep}, in the sublattice $\mathrm{core}_\downarrow(x)$, the canonical meet representation of $\mathrm{pop}_\downarrow(x)$ is given by the coatoms of $\mathrm{core}_\downarrow(x)$.
Then by the moreover part of Theorem~\ref{thm:covers_canonical}, we have \begin{equation}\label{eqn:coatom}\mathrm{coatom}(\mathrm{core}_\downarrow(x)) = \{\kappa(j) \mid j \in \CJR(x)\}.\end{equation}
Thus $\overline{\kappa}(x)\wedge x = \bigwedge \{x \wedge \kappa(j) \mid \kappa(j) \in \mathrm{coatom}(\mathrm{core}_\downarrow(x))\}$, which is equal to $\mathrm{pop}_\downarrow(x)$ by definition.
Moreover, Corollary~\ref{cor:extended_kappa_bij} and Equation~\ref{eqn:coatom} imply that $\CMR(\overline{\kappa}(x)) = \mathrm{coatom}(\mathrm{core}_\downarrow(x))$.
It remains to prove the moreover part.
Suppose $\mathrm{pop}_\downarrow(x) \in \mathsf{cm\text{-}rep}(L)$ and let $\kappa(j) \in \CMR(\overline{\kappa}(x))$. Then $j \in \mathrm{atom}(\mathrm{core}_\downarrow(x))$ by Equation~\ref{eqn:coatom} and the moreover part of Corollary~\ref{cor:nuclear}. By Definition, this means $j$ labels a cover relation of the form $\mathrm{pop}_\downarrow(x) {\,\,<\!\!\!\!\cdot\,\,\,} u$, and so $\kappa(j) \in \CMR(\mathrm{pop}_\downarrow(x))$ by the dual of Theorem~\ref{thm:covers_canonical}.
\end{proof}
We conclude this section with two results which will be useful in relating the $\kappa$-order with the core label orders.
\begin{proposition}\label{prop:K_sets_join}
Let $x \in L$. Then the following hold.\begin{enumerate}
\item If $x \in \mathsf{cj\text{-}rep}(L)$, then $\CJR(x) \subseteq W(x)$ and $\Join W(x) = x$.
\item If $x \in \mathsf{cm\text{-}rep}(L)$, then
$\{j \mid \kappa(j) \in \CMR(x)\} \subseteq \mathcal{W}(\overline{\kappa}^d(x))$ and $\bigwedge \{\kappa(j) \mid j \in \mathcal{W}(\overline{\kappa}^d(x))\} = x$.
\item If $x \in \mathsf{cj\text{-}rep}(L)$, then $\bigwedge \{\kappa(j) \mid j \in W(x)\} = \overline{\kappa}(x)$.
\end{enumerate}
\end{proposition}
\begin{proof}
(1) The fact that $\CJR(x) \subseteq W(x)$ is an immediate consequence of condition (3) in Theorem~\ref{thm:covers_canonical}. This in particular implies that $\bigvee W(x) \geq x$. The fact that $\bigvee W(x) \leq x$ then follows from the fact that every $j \in W(x)$ satisfies $j \leq x$ by Proposition~\ref{prop:core_formulas}.
(2) Denote $A = \{\kappa(j) \mid j \in \mathcal{W}(\overline{\kappa}^d(x))\}$. Let $\kappa(j) \in \CMR(x)$. Then $j \in \CJR(\overline{\kappa}^d(x)) \subseteq \mathcal{W}(\overline{\kappa}^d(x))$ by Corollary~\ref{cor:extended_kappa_bij} and (1). This in particular implies that $\bigwedge A \leq x$. The fact that $\bigwedge A \geq x$ then follows from the fact that every $\kappa(j) \in A$ satisfies $\kappa(j) \geq \overline{\kappa}(\overline{\kappa}^d(x)) = x$ by Proposition~\ref{prop:core_formulas} and Corollary~\ref{cor:extended_kappa_bij}.
(3) By Corollary~\ref{cor:extended_kappa_bij}, this follows by applying (2) to the element $\overline{\kappa}(x)$.
\end{proof}
\begin{remark}
Proposition~\ref{prop:K_sets_join}(1) can be seen as an extension of Example~\ref{ex:asai_pfeifer} in the following sense. Let $\mathcal{T} \in \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$. Then the canonical joinands of $\mathcal{T}$ are precisely the atoms of ${\mathrm{core}_\downarrow}_\mathcal{W}(\mathcal{T})$, which correspond to the simple objects in $\alpha(\mathcal{T})$ by Example~\ref{ex:asai_pfeifer}(2). Moreover, we have that $\mathsf{Filt}(\Gen_\mathcal{W} (\alpha(\mathcal{T}))) = \Join W_\mathcal{W}(\mathcal{T}) = \mathcal{T}$ by Theorem~\ref{thm:IT}. We also note that the explicit description of the $\kappa$-map for $\mathsf{tors}(\mathcal{W})$ given in Theorem~\ref{thm:BTZ_A} could also be used to interpret Proposition~\ref{prop:K_sets_join}(2,3) as extending Example~\ref{ex:asai_pfeifer} in similar ways.
\end{remark}
\begin{proposition}\label{prop:join_irreps_cores}
Let $j \in L$ be completely join-irreducible. Then
\begin{enumerate}
\item $\mathrm{core}_\downarrow(j) = [j_*,j]$.
\item $\mathrm{core}_\uparrow(j) = [\kappa(j),\kappa(j)^*]$.
\item $\mathcal{W}(j) = \mathrm{core}\text{-}\mathrm{lab}_\uparrow(j) = \mathrm{core}\text{-}\mathrm{lab}_\downarrow(j) = \{j\}.$
\end{enumerate}
\end{proposition}
\begin{proof}
All three items follow immediately from the definitions.
\end{proof}
\subsection{The $\kappa$-order and the core label orders}
We now use the constructions in Section~\ref{sec:intervals} to study three different partial orders on the set $\mathsf{cj\text{-}rep}(L)$.
\begin{definition}
Let $L$ be a ws-csd lattice.
\begin{enumerate}
\item The \emph{lower core label order} is $(\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\downarrow})$, where $x \leq_{\mathrm{clo}_\downarrow} y$ if and only if $\mathrm{core}\text{-}\mathrm{lab}_\downarrow(x) \subseteq \mathrm{core}\text{-}\mathrm{lab}_\downarrow(y)$.
\item The \emph{upper core label order} is $(\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\uparrow})$, where $x \leq_{\mathrm{clo}_\uparrow} y$ if and only if $\mathrm{core}\text{-}\mathrm{lab}_\uparrow(x) \subseteq \mathrm{core}\text{-}\mathrm{lab}_\uparrow(y)$.
\item The \emph{$\kappa$-order} is $(\mathsf{cj\text{-}rep}(L),\leq_\kappa)$, where $x \leq_\kappa y$ if and only if $x \leq y$ and $\overline{\kappa}(y) \leq \overline{\kappa}(x)$.
\end{enumerate}
\end{definition}
\begin{example}\label{ex:clo}
Let $L$ be the lattice from Figure~\ref{fig:run_ex}. The upper and lower core label orders and the $\kappa$-order of $L$ are shown in Figure~\ref{fig:ex_clo}. We note that $(L,\leq_{\mathrm{clo}_\downarrow}) \neq (L,\leq_\kappa)$ as shown in \cite[Example~4.28]{enomoto}. On the other hand, we have that $(L,\leq_{\mathrm{clo}_\uparrow}) = (L,\leq_\kappa)$. Note also that $(L,\leq_{\mathrm{clo}_\downarrow})$ is not a lattice (as is shown in \cite[Figure~7]{muhle}), but that $(L,\leq_{\mathrm{clo}_\downarrow}) = (L,\leq_\kappa)$ is a lattice.
\end{example}
\begin{figure}
\begin{tikzpicture}
\begin{scope}[decoration={
markings,
mark=at position 0.7 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-1.5,4);
\draw[postaction=decorate,dashed](0,6) -- (0,4);
\draw[postaction=decorate](0,6) -- (1.5,4);
\draw[postaction=decorate](-1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,2) -- (0,0);
\draw[postaction=decorate](0,2) -- (0,0);
\draw[postaction=decorate](1.5,2) -- (0,0);
\draw[postaction=decorate,bend right,smooth](0,6) to (-3,3);
\draw[postaction=decorate,bend right,smooth](-3,3) to (0,0);
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node [draw,circle,fill=white] at (-3,3) {$j_4$};
\node[draw,circle,fill=white] at (-1.5,2) {$j_3$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (1.5,2) {$j_1$};
\node[draw,circle,fill=white] at (1.5,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-1.5,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\begin{scope}[shift = {(8,0)}]
\begin{scope}[decoration={
markings,
mark=at position 0.7 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-1.5,4);
\draw[postaction=decorate,dashed](0,6) -- (0,4);
\draw[postaction=decorate](0,6) -- (1.5,4);
\draw[postaction=decorate](-1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,2) -- (0,0);
\draw[postaction=decorate](0,2) -- (0,0);
\draw[postaction=decorate](1.5,2) -- (0,0);
\draw[postaction=decorate](-1.5,4) to (-3,2);
\draw[postaction=decorate,dashed](0,4) to (-3,2);
\draw[postaction=decorate](-3,2) to (0,0);
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node [draw,circle,fill=white] at (-3,2) {$j_4$};
\node[draw,circle,fill=white] at (-1.5,2) {$j_3$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (1.5,2) {$j_1$};
\node[draw,circle,fill=white] at (1.5,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-1.5,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\end{scope}
\end{tikzpicture}
\caption{The core label orders and $\kappa$-order for the lattice in Figure~\ref{fig:run_ex}. The left diagram is $(L,\leq_{\mathrm{clo}_\uparrow}) = (L,\leq_{\kappa})$ and the right diagram is $(L,\leq_{\mathrm{clo}_\downarrow})$}\label{fig:ex_clo}
\end{figure}
The (reverse of) the lower core label order was first introduced under the same \emph{shard intersection order} by Reading in \cite{reading} in the case where $L$ is the weak order on a finite Coxeter group. Reading then extended to the poset of regions of any simplicial hyperplane arrangement in \cite[Section~9-7.4]{reading_book}, where he also noted that one could make the same construction for any congruence uniform lattice. M\"uhle then went on to characterize precisely for which finite congruence uniform lattices the lower core label order is a lattice in \cite{muhle}. The lower core label order was further generalized to the class of (finite) ``semidistrim" lattices under the name ``row-core label order" by Defant and Williams in \cite[Section~11.4]{DW}. In the same paper, they also introduce the ``pop-core label order", which coincides with the $\kappa$-order of a finite semidistributive lattice. In the infinite case, Enomoto introduced (and named) the $\kappa$-order in \cite{enomoto}. One of his main results is that when $L$ is the lattice of torsion classes, the lower core label order and the ordering of the corresponding wide subcategories all coincide. More precisely, he proved the following.
\begin{theorem}\cite[Corollary~4.26]{enomoto}\label{thm:enomoto}
Let $\Lambda$ be a finite-dimensional algebra and $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ a wide subcategory. Then $\leq_{\mathrm{clo}_\downarrow}$ and $\leq_\kappa$ coincide on $\mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$. Moreover, the association $\mathcal{T} \mapsto \alpha(\mathcal{T})$ induces a lattice isomorphism $(\mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W})),\leq_\kappa)\rightarrow \mathsf{wide}(\mathcal{W})$.
\end{theorem}
\begin{example}\label{ex:kronecker_kappa}
Let $L_{kr}(K)$ be as in Figure~\ref{fig:kronecker} and let $KQ$ be the path algebra over the Kronecker quiver as in Example~\ref{ex:kronecker}. Recall from Examples~\ref{ex:no_can_join}(2) and~\ref{ex:jirr} that $\mathsf{cj\text{-}rep}(L_{kr}(K)) = L_{kr}(K) \setminus \{\emptyset\}$. Moreover, by Example~\ref{ex:kronecker} we have that $L_{kr}(K) \cong \mathsf{tors}(\mathsf{mod} KQ)$, and so the orders $\leq_\kappa$, $\leq_{\mathrm{clo}_\downarrow}$, and $\leq_{\mathrm{clo}_\uparrow}$ coincide by Theorem~\ref{thm:enomoto}. (This fact is also straightforward to verify directly.) These three orders are then given as follows:
\begin{enumerate}
\item Restricted to $(2^{\mathbb{P}^1(K)} \setminus \{\emptyset\}) \{\hat{0},\hat{1},s\}$, one has $x \leq_\kappa y$ if and only if $x \leq y$, where $\leq$ is the order in $L_{kr}(K)$.
\item For $(n,t) \in \mathbb{N} \times \{-1,1\}$, one has $\hat{0} \leq_\kappa (n,t) \leq_\kappa \hat{1}$.
\end{enumerate}
In addition, it can be shown that $\tau\text{-}\mathsf{perp}(\mathsf{mod} KQ)$ is isomorphic to the restriction of the $\kappa$ order of $L_{kr}(K)$ to $\mathbb{N}\times \{-1,1\} \cup \{\hat{0},\hat{1},s\}$. See \cite[Section~7]{BuH}.
\end{example}
As a lattice-theoretic analog of Theorem~\ref{thm:enomoto}, we have the following.
\begin{proposition}\label{prop:K_sets}
Let $x, y \in \mathsf{cj\text{-}rep}(L)$. Then $x \leq_\kappa y$ if and only if $W(x) \subseteq W(y)$.
\end{proposition}
\begin{proof}
Suppose first that $x \leq_\kappa y$. Since $x \leq y$, we have
$$\{j \mid j \leq x\} \subseteq \{j \mid j \leq y\}.$$
Likewise since $\overline{\kappa}(x) \geq \overline{\kappa}(y)$, we have
$$\{j \mid \kappa(j) \geq \overline{\kappa}(x)\} \subseteq \{j \mid \kappa(j) \geq \overline{\kappa}(y)\}.$$
Taking intersections, Proposition~\ref{prop:core_formulas} thus implies that $W(x) \subseteq W(y)$.
Now assume that $W(x) \subseteq W(y)$. It then follows from Proposition~\ref{prop:K_sets_join} that $x = \Join W(x) \leq \Join W(y) = y$ and that $\overline{\kappa}(x) = \bigwedge \kappa(W(x)) \geq \bigwedge \kappa(W(y)) = \overline{\kappa}(y)$. We conclude that $x \leq_\kappa y$.
\end{proof}
The $\kappa$-order also satisfies the following duality.
\begin{proposition}\label{prop:self_dual}
The map $\overline{\kappa}$ induces a poset isomorphism $(\mathsf{cj\text{-}rep}(L),\leq_\kappa) \rightarrow (\mathsf{cj\text{-}rep}(L^d), \leq_{\kappa^d})$ between the $\kappa$-order of $L$ and that of $L^d$.
\end{proposition}
\begin{proof}
This is an immediate consequence of the fact that $\kappa^d$ is precisely the $\kappa$-map for the lattice $L^d$. More explicitly, suppose $x \leq_\kappa y$ in the lattice $L$. That is, $x, y \in \mathsf{cj\text{-}rep}(L)$, $x \leq_L y$ and $\overline{\kappa}(x) \geq_L \overline{\kappa}(y)$. By Corollary~\ref{cor:extended_kappa_bij}, this means $\overline{\kappa}(x), \overline{\kappa}(y) \in \mathsf{cm\text{-}rep}(L) = \mathsf{cj\text{-}rep}(L^d)$, $\overline{\kappa}^d(\overline{\kappa}(x)) \geq_{L^d} \overline{\kappa}^d(\overline{\kappa}(y))$ and $\overline{\kappa}(x) \leq_{L^d} \overline{\kappa}(y)$; i.e., that $\kappa(x) \leq_{\kappa^d} \kappa(y)$ in the lattice $L^d$.
\end{proof}
\begin{remark}
Let $\Lambda$ be a finite-dimensional algebra and $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ a wide subcategory. It is well known that the torsion-free classes of $\mathcal{W}$ form a lattice $\mathsf{tors}(\mathcal{W})$ which is dual to $\mathsf{tors}(\mathcal{W})$. Proposition~\ref{prop:self_dual} then implies that the $\kappa$-orders of $\mathsf{tors}(\mathcal{W})$ and $\mathsf{torf}(\mathcal{W})$ are isomorphic. In particular, both will be isomorphic to the lattice $\mathsf{wide}(\mathcal{W})$ by Theorem~\ref{thm:enomoto}. This is also implicit in \cite[Remark~4.20]{enomoto}.
\end{remark}
We now obtain the following, where the implication $(2a \implies 2b)$ can be found as \cite[Proposition~4.30]{enomoto}.
\begin{proposition}\label{prop:orders_coincide}
Let $L$ be a ws-csd lattice. Then:
\begin{enumerate}
\item The following are equivalent:
\begin{enumerate}
\item $(\mathsf{cj\text{-}rep}(L),\leq_\kappa) = (\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\downarrow})$.
\item $W(x) = \mathrm{core}\text{-}\mathrm{lab}_\downarrow(x)$ for all $x \in \mathsf{cj\text{-}rep}(L)$.
\end{enumerate}
\item The following are equivalent:
\begin{enumerate}
\item $(\mathsf{cj\text{-}rep}(L),\leq_\kappa) = (\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\uparrow})$.
\item $W(x) = \mathrm{core}\text{-}\mathrm{lab}_\uparrow(x)$ for all $x \in \mathsf{cj\text{-}rep}(L)$.
\end{enumerate}
\item If $(\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\uparrow}) = (\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\downarrow})$, then the properties (1a), (1b), (2a), and (2b) all hold.
\end{enumerate}
\end{proposition}
\begin{proof}
We prove only (1) and (3), since the proof of (2) is analogous to that of (1).
(1) The fact that $(1b \implies 1a)$ follows from Proposition~\ref{prop:K_sets} and the definition of $W(x)$. To see that $(1a \implies 1b)$, assume (1a) and let $x \in \mathsf{cj\text{-}rep}(L)$. By the definitions, we need only show that $\mathrm{core}\text{-}\mathrm{lab}_\downarrow(x) \subseteq W(x)$. To see this, let $j \in \mathrm{core}\text{-}\mathrm{lab}_\downarrow(x)$. Then $j \leq_{\mathrm{clo}_\downarrow} x$ by Proposition~\ref{prop:join_irreps_cores}. Thus by assumption, we have that $j \leq_\kappa x$, and so $j \in W(x)$ by another application of Proposition~\ref{prop:join_irreps_cores}.
(3) Suppose that $(\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\uparrow}) = (\mathsf{cj\text{-}rep}(L),\leq_{\mathrm{clo}_\downarrow})$ and let $x \in \mathsf{cj\text{-}rep}(L)$. We will show that $\mathrm{core}\text{-}\mathrm{lab}_\uparrow(x) = \mathrm{core}\text{-}\mathrm{lab}_\downarrow(x)$. By the Definition of $W(x)$, this will imply that properties (1b) and (2b) must hold, thus proving the result.
Let $j \in \cji(L)$. By Propositions~\ref{prop:join_irreps_cores} and the assumption that the two core label orders coincide, we thus have that
\begin{eqnarray*}
j \in \mathrm{core}\text{-}\mathrm{lab}_\downarrow(x) & \iff & j \leq_{\mathrm{clo}_\downarrow} x\\
& \iff & j \leq_{\mathrm{clo}_\uparrow} x\\
&\iff& j \in \mathrm{core}\text{-}\mathrm{lab}_\uparrow(x).
\end{eqnarray*}
This concludes the proof.
\end{proof}
One may hope that Proposition~\ref{prop:orders_coincide} can be modified into a statement about when the three partial orders on $\mathsf{cj\text{-}rep}(L)$ are isomorphic as abstract posets. The following example shows that this it is not immediately obvious how to do so.
\begin{example}\label{ex:weird_clo}
Let $L$ be the lattice shown in Figure~\ref{fig:run_ex} with an extra element $j_5$ whose cover relations are $m_3 {\,\,<\!\!\!\!\cdot\,\,\,} j_5 {\,\,<\!\!\!\!\cdot\,\,\,} \hat{1}$. The lattice $L$, both of the core label orders, and the $\kappa$-order are all shown in Figure~\ref{fig:weird_clo}. We observe that the two core label orders are isomorphic as abstract posets, where the isomorphism exchanges $j_4$ and $j_5$. On the other hand, neither core label order is isomorphic to the $\kappa$-order.
\end{example}
\begin{figure}
\begin{tikzpicture}
\begin{scope}[decoration={
markings,
mark=at position 0.65 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-3,4) node [midway,above left] {$j_1$};
\draw[postaction=decorate,dashed](0,6) -- (0,4) node [midway,left] {$j_2$};
\draw[postaction=decorate](0,6) -- (1.5,5) node [midway,above right] {$j_3$};
\draw[postaction=decorate](1.5,5) -- (3,4) node [midway,above right] {$j_5$};
\draw[postaction=decorate](-3,4) -- (0,2) node [near start,right] {$j_3$};
\draw[postaction=decorate,dashed](0,4) -- (-3,2) node [near start,left] {$j_1$};
\draw[postaction=decorate](3,4) -- (0,2) node [near start, left] {$j_1$};
\draw[postaction=decorate,dashed](0,4) -- (3,2) node [near start,right] {$j_3$};
\draw[postaction=decorate](-3,4) -- (-3,2) node [midway, left] {$j_2$};
\draw[postaction=decorate](3,4) -- (3,2) node [midway, right] {$j_2$};
\draw[postaction=decorate](-3,2) -- (-1.5,1) node [midway,left] {$j_4$};
\draw[postaction=decorate](-1.5,1) -- (0,0) node [midway,left] {$j_3$};
\draw[postaction=decorate](0,2) -- (0,0) node [midway,left] {$j_2$};
\draw[postaction=decorate](3,2) -- (0,0) node [midway,below right] {$j_1$};
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node[draw,circle,fill=white] at (-3,2) {$j_4$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (3,2) {$j_1$};
\node[draw,circle,fill=white] at (-1.5,1) {$j_3$};
\node[draw,circle,fill=white] at (1.5,5) {$j_5$};
\node[draw,circle,fill=white] at (3,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-3,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\begin{scope}[shift = {(8,-8)}]
\begin{scope}[decoration={
markings,
mark=at position 0.7 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-1.5,4);
\draw[postaction=decorate,dashed](0,6) -- (0,4);
\draw[postaction=decorate](0,6) -- (1.5,4);
\draw[postaction=decorate](-1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,2) -- (0,0);
\draw[postaction=decorate](0,2) -- (0,0);
\draw[postaction=decorate](1.5,2) -- (0,0);
\draw[postaction=decorate](-1.5,4) to (-3,2);
\draw[postaction=decorate,dashed](0,4) to (-3,2);
\draw[postaction=decorate](-3,2) to (0,0);
\draw[postaction=decorate, bend left](0,6) to (3,3);
\draw[postaction=decorate, bend left](3,3) to (0,0);
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node [draw,circle,fill=white] at (-3,2) {$j_4$};
\node[draw,circle,fill=white] at (-1.5,2) {$j_3$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (1.5,2) {$j_1$};
\node[draw,circle,fill=white] at (1.5,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-1.5,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\node[draw,circle,fill=white] at (3,3) {$j_5$};
\end{scope}
\begin{scope}[shift = {(0,-8)}]
\begin{scope}[decoration={
markings,
mark=at position 0.7 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-1.5,4);
\draw[postaction=decorate,dashed](0,6) -- (0,4);
\draw[postaction=decorate](0,6) -- (1.5,4);
\draw[postaction=decorate](-1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,2) -- (0,0);
\draw[postaction=decorate](0,2) -- (0,0);
\draw[postaction=decorate](1.5,2) -- (0,0);
\draw[postaction=decorate](-1.5,4) to (-3,2);
\draw[postaction=decorate,dashed](0,4) to (-3,2);
\draw[postaction=decorate](-3,2) to (0,0);
\draw[postaction=decorate, bend left](0,6) to (3,3);
\draw[postaction=decorate, bend left](3,3) to (0,0);
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node [draw,circle,fill=white] at (-3,2) {$j_5$};
\node[draw,circle,fill=white] at (-1.5,2) {$j_3$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (1.5,2) {$j_1$};
\node[draw,circle,fill=white] at (1.5,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-1.5,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\node[draw,circle,fill=white] at (3,3) {$j_4$};
\end{scope}
\begin{scope}[shift = {(8,0)}]
\begin{scope}[decoration={
markings,
mark=at position 0.7 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-1.5,4);
\draw[postaction=decorate,dashed](0,6) -- (0,4);
\draw[postaction=decorate](0,6) -- (1.5,4);
\draw[postaction=decorate](-1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (0,2);
\draw[postaction=decorate,dashed](0,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,4) -- (-1.5,2);
\draw[postaction=decorate](1.5,4) -- (1.5,2);
\draw[postaction=decorate](-1.5,2) -- (0,0);
\draw[postaction=decorate](0,2) -- (0,0);
\draw[postaction=decorate](1.5,2) -- (0,0);
\draw[postaction=decorate,bend left](0,6) to (3,3);
\draw[postaction=decorate,bend left](3,3) to (0,0);
\draw[postaction=decorate, bend right](0,6) to (-3,3);
\draw[postaction=decorate, bend right](-3,3) to (0,0);
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node [draw,circle,fill=white] at (-3,3) {$j_4$};
\node[draw,circle,fill=white] at (-1.5,2) {$j_3$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (1.5,2) {$j_1$};
\node[draw,circle,fill=white] at (1.5,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-1.5,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\node[draw,circle,fill=white] at (3,3) {$j_5$};
\end{scope}
\end{tikzpicture}
\caption{A semidistributive lattice $L$ with its join-irreducible labeling (top left), the corresponding $\kappa$-order (top right), upper core label order (bottom left), and lower core label order (bottom right). }\label{fig:weird_clo}
\end{figure}
\section{$\kappa^d$-exceptional sequences}\label{sec:kappa_tau}
The purpose of this section is to relate $\tau$-exceptional sequences to the operators $\overline{\kappa}^d$ and $\mathrm{pop}_\uparrow$ on the lattice $\mathsf{tors}\Lambda$. More precisely, let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a functorially finite wide subcategory and suppose $M \in \mathcal{W}$ is $\tau$-rigid (in $\mathcal{W}$). In Section~\ref{sec:combinatorial_tau_tilting}, we describe the torsion classes $\lperpW{(\tau_\mathcal{W} M)}$ and $\lperpW{(\overline{\tau_\mathcal{W}}M)}$ as coming from applying ${\mathrm{pop}_\uparrow}_\mathcal{W}$ and $\overline{\kappa_\mathcal{W}}^d$ to the torsion class $\Gen_\mathcal{W} M$. (Recall the definition of $\overline{\tau_\mathcal{W}}$ from Notation~\ref{not:tau_bar} and that the subscript $\mathcal{W}$ means the lattice being considered is $\mathsf{tors}(\mathcal{W})$.) In Section~\ref{sec:kappa_exceptional}, we then use this description to explain how all of the $\tau$-exceptional sequences in $\mathcal{W}$ can be read directly from the brick labeling of $\mathsf{tors}(\mathcal{W})$. The result is a ``combinatorialization'' of the notion of a $\tau$-exceptional sequence given in Definition~\ref{def:kappa_exceptional}.
\subsection{Rowmotion and the AR translate}\label{sec:combinatorial_tau_tilting}
In this section, we describe the combinatorial relationship between the torsion classes $\Gen_\mathcal{W} M$, $\lperpW{(\tau_\mathcal{W} M)}$, and $\lperpW{(\overline{\tau_\mathcal{W}} M)}$. We fix for the duration of this section a functorially finite wide subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$. The starting point is the following, which describes how the map $\overline{\kappa}^{d}$ is computed on the lattice of torsion classes.
\begin{theorem}\cite[Theorem~B]{BTZ}\label{thm:BTZ_A} Let $\mathcal{T} \in \mathsf{cm\text{-}rep}(\mathsf{tors}(\mathcal{W}))$, so that $\mathcal{T} = \lperpW{\mathcal{S}}$ for some semibrick $\mathcal{S} \in \mathsf{sbrick}(\mathcal{W})$. Then $\overline{\kappa_\mathcal{W}}^{d}( \lperpW{\mathcal{S}}) = \mathsf{Filt}(\Gen_\mathcal{W}\mathcal{S})$.
\end{theorem}
We also need the following. This result is well-known, but we include an explanation of how it can be deduced from explicit results in the literature.
\begin{theorem}\label{thm:regular}
Let $\mathcal{T} \in \mathsf{tors}(\mathcal{W})$ be a functorially finite torsion class. Then $\mathcal{T} \in \mathsf{cj\text{-}rep}(\mathsf{tors}\mathcal{W})\cap \mathsf{cm\text{-}rep}(\mathsf{tors}\mathcal{W})$ and $|\mathrm{CJR}(\mathcal{T})| + |\mathrm{CMR}(\mathcal{T})| = \mathrm{rk}(\mathcal{W})$. Moreover, if $\mathcal{T} = \Gen_\mathcal{W} M$ with $M \in \mathcal{W}$ a gen-minimal $\tau$-rigid module, then $|\mathrm{CJR}(\mathcal{T})| = \mathrm{rk}(M)$.
\end{theorem}
\begin{proof}
Since $\mathcal{T}$ is functorially finite, we can write $\mathcal{T} = \Gen_\mathcal{W} M$ with $M \in \mathcal{W}$ a gen-minimal $\tau$-rigid module. By \cite[Theorem~3.1]{DIJ}, given an inclusion $\mathcal{U} \subsetneq \mathcal{T}$ there exists a cover relation $\mathcal{V} {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{T}$ such that $\mathcal{U} \subseteq \mathcal{V}$, and the dual property holds as well. By Proposition~\ref{prop:covers_canonical_existence} and its dual, this means that $\mathcal{T} \in \mathsf{cj\text{-}rep}(\mathsf{tors}\mathcal{W}) \cap \mathsf{cm\text{-}rep}(\mathsf{tors}\mathcal{W})$.
Now recall from Theorem~\ref{thm:covers_canonical} that the canonical joinands of $\mathcal{T}$ are in bijection with cover relations of the form $\mathcal{U} {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{T}$ and the canonical meetands of $\mathcal{T}$ are in bijection with cover relations of the form $\mathcal{T} {\,\,<\!\!\!\!\cdot\,\,\,} \mathcal{U}$. The fact that $|\mathrm{CJR}(\mathcal{T})| + |\mathrm{CMR}(\mathcal{T})| = \mathrm{rk}(\mathcal{W})$ is thus a consequence of the fact that the mutation graph of support $\tau$-tilting pairs is $\mathrm{rk}(\mathcal{W})$-regular, see \cite[Section~2]{AIR}.
Finally, we recall from Theorem~\ref{thm:DIJ}(3) that there is a semibrick $\mathcal{X}(M)$ with $|\mathcal{X}(M)| = \mathrm{rk}(M)$ and $\mathsf{Filt}(\Gen_\mathcal{W}\mathcal{X}(M)) = \Gen_\mathcal{W} M$.
By Theorem~\ref{thm:brick_labeling}(1), this means the canonical join representation of $\mathcal{T}$ is
$$\mathcal{T} = \Join \{\mathsf{Filt} \Gen_\mathcal{W}(B) \mid B \in \mathcal{X}(M)\},$$
and so $|\CJR(\mathcal{T})| = \mathrm{rk}(M)$.
\end{proof}
In particular, Theorem~\ref{thm:regular} implies that both $\overline{\kappa_\mathcal{W}}(\mathcal{T})$ and $\overline{\kappa_\mathcal{W}}^d(\mathcal{T})$ are defined when $\mathcal{T}$ is functorially finite.
Together, Theorems~\ref{thm:BTZ_A} and~\ref{thm:regular} allow us to prove the following.
\begin{corollary}\label{cor:kappa_proj}
Let $P$ be projective in $\mathcal{W}$. Then
\begin{enumerate}
\item $\overline{\kappa_\mathcal{W}}^{d}(\Gen_\mathcal{W} P) = P^{\perp_\mathcal{W}} =\lperpW{(\overline{\tau_\mathcal{W}} P)}$, and
\item ${\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} P) = \mathcal{W}$
\end{enumerate}
\end{corollary}
\begin{proof}
First note that $P$ is $\tau$-rigid in $\mathcal{W}$, and so $\Gen_\mathcal{W} P \in \mathsf{cj\text{-}rep}(\mathsf{tors}\mathcal{W})$. Now let $\mathcal{S}$ be the set of modules which are simple in $\mathcal{W}$. We can partition $\mathcal{S} = \mathcal{S}_1 \sqcup \mathcal{S}_2$, where $\mathcal{S}_1$ contains those modules whose ($\mathcal{W}$-)projective covers are direct summands of $P$ and $\mathcal{S}_2 = \mathcal{S} \setminus \mathcal{S}_1$. Then $\mathcal{S}_1 \subseteq \Gen_\mathcal{W} P = \lperpW{\mathcal{S}_2}$ and $P^{\perp_\mathcal{W}} = \mathsf{Filt}\mathcal{S}_2 = \mathsf{Filt}(\Gen_\mathcal{W} \mathcal{S}_2)$.
Theorem~\ref{thm:BTZ_A} then implies that $\overline{\kappa_\mathcal{W}}^{d}(\Gen_\mathcal{W} P) = P^{\perp_\mathcal{W}}$. Finally, by Proposition~\ref{prop:pop_formula} we have that $\mathcal{S} \subseteq (\Gen_\mathcal{W} P) \vee (P^{\perp_\mathcal{W}}) = {\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} P)$. Since torsion classes are closed under extensions, this implies that ${\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} P) = \mathcal{W}$.
\end{proof}
The main result of this section is the following.
\begin{proposition}\label{prop:kappa_tau}
Let $M \in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ be gen-minimal. Then ${\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} M) = \lperpW{(\tau_\mathcal{W} M)}$.
\end{proposition}
We emphasize that we are using $\tau$, not $\overline{\tau}$, in Proposition~\ref{prop:kappa_tau}. In particular, since any basic projective module is also gen-minimal, Proposition~\ref{prop:kappa_tau} gives an alternative proof of the fact that ${\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} P) = \mathcal{W}$ whenever $P$ is projective in $\mathcal{W}$.
\begin{proof}
Since $\mathcal J_\mathcal{W}(M) = (M^{\perp_\mathcal{W}}) \cap (\lperpW{(\tau_\mathcal{W} M)})$ is a wide subcategory, it follows from Theorem~\ref{thm:asai_pfeifer_1}(1) that $[\Gen_\mathcal{W} M\lperpW{(\tau_\mathcal{W} M)}]$ is a nuclear interval. By definition, this means $$\lperpW{(\tau_\mathcal{W} M)} = \Gen_\mathcal{W} M \vee \left\{j \ \middle| \ \kappa_\mathcal{W}(j) \in \CMR_\mathcal{W}(\Gen_\mathcal{W} M) \text{ and } (\Gen_\mathcal{W} M) \vee j \leq \lperpW{(\tau_\mathcal{W} M)}\right\}.$$
To prove the result, it therefore suffices to show that $(\Gen_\mathcal{W} M) \vee j \leq \lperpW{(\tau_\mathcal{W} M)}$ for every $\kappa_\mathcal{W}(j) \in \CMR_\mathcal{W}(\Gen_\mathcal{W} M)$.
Now recall from Theorems~\ref{thm:min_extending} and~\ref{thm:BTZ_A} that the elements of $\CMR_\mathcal{W}(\Gen_\mathcal{W} M)$ are precisely the torsion classes $\mathcal{W} \cap \lperp{X}$ for $X$ a minimal extending module of $\Gen_\mathcal{W} M$. Let $X$ be such a minimal extending module.
Now, every proper factor of $X$ which lies in $\mathcal{W}$ must lie in $\Gen_\mathcal{W} M \subseteq \lperpW{(\tau_\mathcal{W} M)}$.
By Proposition~\ref{prop:AStau}, this means that $\mathrm{Hom}_\Lambda(X,\tau_\mathcal{W} M) \neq 0$ if and only if $\mathrm{Ext}^1_\Lambda(M,X) = 0$.
Assume for the sake of contradiction that there is a nonsplit exact sequence $X\hookrightarrow E \twoheadrightarrow M$. Then by the definition of a minimal extending module, we have that $E \in \mathcal{T}$. In particular, we have that $M$ is not split projective in $\mathcal{T}$, which contradicts Proposition~\ref{prop:functorially_finite}.
\end{proof}
Combining the results of this section, we have the following.
\begin{corollary}\label{cor:kappa_indec}
Let $M \in \tau\text{-}\mathrm{rigid}(\mathcal{W})$ be indecomposable. Then exactly one of the following holds.
\begin{enumerate}
\item $M$ is projective in $\mathcal{W}$, $\overline{\kappa_\mathcal{W}}^{d}(\Gen_\mathcal{W} M) = M^{\perp_\mathcal{W}} = \lperpW{(\overline{\tau_\mathcal{W}} M)}$, and ${\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} M) = \mathcal{W} = \lperpW{(\tau_\mathcal{W} M)}$.
\item $M$ is not projective in $\mathcal{W}$ and ${\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} M) = \overline{\kappa_\mathcal{W}}^{d}(\Gen_\mathcal{W} M) = \lperpW{(\tau_\mathcal{W} M)}$.
\end{enumerate}
\end{corollary}
\begin{proof}
If $M$ is projective in $\mathcal{W}$, the result is the special case of Corollary~\ref{cor:kappa_proj} where $P$ is indecomposable. If $M$ is not projective, then by Proposition~\ref{prop:kappa_tau} we have that ${\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} M) = \lperpW{(\tau_\mathcal{W} M)}$. Thus it remains only to show that $\lperpW{(\tau_\mathcal{W} M)} = \overline{\kappa_\mathcal{W}}^{d}(\Gen_\mathcal{W} M)$.
First note that $\Gen_\mathcal{W} M \in \mathsf{cm\text{-}rep}(\mathsf{tors}(\mathcal{W}))$ and that $|\CJR_\mathcal{W}(\overline{\kappa_\mathcal{W}}^d(\Gen_\mathcal{W} M))| = \mathrm{rk}(\mathcal{W})-1$ by Theorem~\ref{thm:regular} and Corollary~\ref{cor:extended_kappa_bij}. We also know that $\lperpW{(\tau_\mathcal{W} M)} \in \mathsf{cj\text{-}rep}(\mathsf{tors}(\mathcal{W}))$ because it is functorially finite, see \cite[Theorem~2.10]{AIR}. Since $M$ is not projective, we have $\lperpW{(\tau_\mathcal{W} M)} \neq \mathcal{W}$, and so $|\CJR_\mathcal{W}( \lperpW{(\tau_\mathcal{W} M}))| \leq \mathrm{rk}(\mathcal{W}) - 1$ by Theorem~\ref{thm:regular}. It then follows from Proposition~\ref{prop:pop_formula}(2) that the sets $\CJR_\mathcal{W}(\lperpW{(\tau_\mathcal{W} M)}) = \CJR_\mathcal{W}({\mathrm{pop}_\uparrow}_\mathcal{W}(\Gen_\mathcal{W} M))$ and $\CJR_\mathcal{W}(\overline{\kappa_\mathcal{W}}^d(\Gen_\mathcal{W} M))$ must coincide, and so $\lperpW{(\tau_\mathcal{W} M)} = \overline{\kappa_\mathcal{W}}^{d}(\Gen_\mathcal{W} M)$.
\end{proof}
\begin{remark}
Corollary~\ref{cor:kappa_indec}(2) is only true under the assumption that $M$ is indecomposable. For example, let $\Lambda = K(1 \rightarrow 2 \rightarrow 3)$ and let $M = I_2 \oplus S_2$. Then $\Gen M = \lperp{P_1}$. Thus Theorem~\ref{thm:BTZ_A} tells us that $\overline{\kappa}^{d}(\Gen M) = \Gen P_1$, while Proposition~\ref{prop:kappa_tau} tells us that $\mathrm{pop}_\uparrow(\Gen M) = \lperp{\tau M} = \lperp{(P_2 \oplus P_3)} = \Gen(P_1 \oplus S_2)$.
\end{remark}
\subsection{$\kappa^d$-exceptional sequences}\label{sec:kappa_exceptional}
We recall that for $\Lambda$ an arbitrary finite-dimensional algebra, understanding the $\tau$-exceptional sequences for $\mathsf{mod}\Lambda$ requires one to understand which wide subcategories are $\tau$-perpendicular subcategories and which torsion classes are functorially finite. Currently, there is no purely combinatorial criteria which allows one to recover these special subcategories directly from the lattice of torsion classes. As such, we restrict to the case where the lattice of torsion classes $\mathsf{tors}(\mathsf{mod}\Lambda)$ is finite; i.e., we assume that $\Lambda$ is a $\tau$-tilting finite algebra (see Definition-Theorem~\ref{def:tau_tilting_finite}). For such algebras, every torsion class is functorially finite and every wide subcategory is a $\tau$-perpendicular subcategory. In particular, Proposition~\ref{prop:kappa_tau} allows one to compute ${\mathrm{pop}_\uparrow}_\mathcal{W}(\mathcal{T})$ for any wide subcategory $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ and any torsion class $\mathcal{T}\in \mathsf{tors}(\mathcal{W})$. This is the motivation for the following (recursive) definition.
\begin{definition}\label{def:kappa_exceptional}
Let $L$ be a finite semidistributive lattice and let $(j_k,\ldots,j_1)$ be a sequence of (completely) join-irreducible elements of $L$. We recursively call $(j_k,\ldots,j_1)$ a \emph{$\kappa^d$-exceptional sequence} if both of the following hold.
\begin{enumerate}
\item $j_i \in \mathrm{core}\text{-}\mathrm{lab}_\uparrow(j_1) = \mathrm{j\text{-}label}[j_1,\mathrm{pop}_\uparrow(j_1)]$ for all $i > 1$.
\item $(j_k \vee j_1,\ldots, j_2 \vee j_1)$ is a $\kappa^d$-exceptional sequence in the lattice $\mathrm{core}_\uparrow(\overline{\kappa}^{-1}(j_1)) = [j_1,\mathrm{pop}_\uparrow(j_1)]$.
\end{enumerate}
\end{definition}
\begin{remark}
Recall from Lemma~\ref{prop:intervals} that $\mathrm{core}_\uparrow(j_1) = [j_1,\mathrm{pop}_\uparrow(j_1)]$ is itself a finite semidistributive lattice, and that supposing $j_i \in \mathrm{core}\text{-}\mathrm{lab}_\uparrow(j_1)$ is equivalent to supposing that $j_i \vee j_1 \in \cji(\mathrm{core}_\uparrow(j_1))$. The recursive part of the definition can then be seen as ``cutting out'' successively smaller (interval) sublattices from $L$.
\end{remark}
We defer giving examples of $\kappa^d$ to exceptional sequences to Section~\ref{sec:kappa_exceptional_ex}. To conclude the present section, we build towards our final main them. In particular, we need the following, which combines Theorem~\ref{thm:asai_pfeifer_1} with Corollary~\ref{cor:kappa_indec}.
\begin{corollary}\label{cor:reduction}
Suppose $B \in \mathcal{W}$ is a brick and that $\mathsf{Filt}(\Gen_\mathcal{W} B)$ is functorially finite. Then there is a label-preserving bijection $\mathrm{core}_\uparrow(\mathsf{Filt}(\Gen_\mathcal{W} B)) \cong \mathcal J_\mathcal{W}(\beta_\mathcal{W}^{-1}(B)).$
\end{corollary}
\begin{proof}
Since $\mathsf{Filt}(\Gen_\mathcal{W} B)$ is functorially finite, it follows from Theorem~\ref{thm:DIJ}(2) that $B$ is in the image of the bijection $\beta_\mathcal{W}$, and thus that $\beta_\mathcal{W}^{-1}(B)$ is an indecomposable $\tau$-rigid module which satisfies $\Gen_\mathcal{W}(\beta_\mathcal{W}^{-1}(B)) = \mathsf{Filt}(\Gen_\mathcal{W} B)$. Moreover, we have $\mathrm{pop}_\uparrow(\mathsf{Filt}(\Gen_\mathcal{W} B)) = \lperpW{(\tau_\mathcal{W} (\beta_\mathcal{W}^{-1}(B)))}$ by Proposition~\ref{prop:kappa_tau}. The result then follows from Theorem~\ref{thm:asai_pfeifer_1}(2)
\end{proof}
We are now prepared to prove our final main theorem.
\begin{theorem}[Theorem~\ref{thm:intro:mainD}]\label{thm:mainD}
Let $\Lambda$ be a $\tau$-tilting finite algebra, and let $\mathcal{W} \subseteq \mathsf{mod}\Lambda$ be a wide subcategory. Then there is a bijection $\rho$ from the set of $\tau$-exceptional sequences for $\mathcal{W}$ to the set of $\kappa^d$-exceptional sequences for $\mathsf{tors}(\mathcal{W})$ given as follows. Let $(M_k,\ldots,M_1)\in \tau\text{-}\mathsf{ex}(\mathcal{W})$ and let $\psi(M_k,\ldots,M_1) = (\mathcal{W}_k{\,\,<\!\!\!\!\cdot\,_\tau\,\,} \cdots {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_1 {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}_0 = \mathcal{W})$ be the corresponding saturated top chain in $\tau\text{-}\mathsf{perp}(\mathcal{W})$. (See Theorem~\ref{thm:tf_tau}.) For $1 \leq i \leq k$, denote $\mathcal{T}_i = \Gen_{\mathcal{W}_{i-1}} M_i \in \mathsf{tors}(\mathcal{W}_{i-1})$. Then
$$\rho(M_k,\ldots,M_1) = (\mathcal{T}_k,\ldots,\mathcal{T}_1).$$
Moreover, for each $1 \leq i \leq k$, one has $\mathrm{br\text{-}label}_{\mathcal{W}_{i-1}}[(\mathcal{T}_i)_*,\mathcal{T}_i] = \beta(M_i)$.
\end{theorem}
\begin{proof}
Note that since $\Lambda$ is $\tau$-tilting finite, we have that every torsion class is functorially finite (Definition-Theorem~\ref{def:tau_tilting_finite}), every wide subcategory is a $\tau$-perpendicular subcategory (Example~\ref{ex:tau_perp}), and every brick is an sf-brick. In particular, for every $\mathcal{V} \in \mathsf{wide}(\mathcal{W})$ there are bijections from the set of indecomposable modules in $\tau\text{-}\mathrm{rigid}(\mathcal{V})$ to $\cji(\mathsf{tors}(\mathcal{V}))$ and $\mathsf{brick}(\mathcal{V})$ given by $M \mapsto \Gen_\mathcal{V} M$ and $M \mapsto \beta(M)$. (See Theorem~\ref{thm:regular}.)
We now prove the result by induction on $\mathrm{rk}(\mathcal{W})$. The result is trivial for $\mathrm{rk}(\mathcal{W}) = 0$, so suppose that $\mathrm{rk}(\mathcal{W}) > 0$ and that the result holds for all $\mathcal{V}$ with $\mathrm{rk}(\mathcal{V}) < \mathrm{rk}(\mathcal{W})$. Now by Definition~\ref{def:brick_label}, $\beta(M_1)$ is the (unique) brick label of the cover relation $\mathcal{W}_1 {\,\,<\!\!\!\!\cdot\,_\tau\,\,} \mathcal{W}$ in $\tau\text{-}\mathsf{perp}(\mathcal{W})$. In particular, $M_1$ bijectively determines both $\mathcal{T}_1$ and $\mathcal{W}_1$. We also have $\mathsf{Filt}(\Gen_\mathcal{W} \beta(M_1)) = \mathcal{T}_1$ and ${\mathrm{pop}_\uparrow}_\mathcal{W}(\mathcal{T}_1) = \lperpW(\tau_\mathcal{W} M_1)$ by Corollary~\ref{cor:kappa_indec}. This in particular means that ${\mathrm{br\text{-}label}}_\mathcal{W}[(\mathcal{T}_1)_*,\mathcal{T}_1] = \beta(M_1)$.
Now by Corollary~\ref{cor:reduction}, we have a label-preserving isomorphism $\mathrm{core}_\uparrow(\mathcal{T}_1) \cong \mathsf{tors}(\mathcal{W}_1)$. Also, by assumption we have that $(M_k,\ldots,M_2) \in \tau\text{-}\mathsf{ex}(\mathcal{W}_1)$ and that $\rho(M_k,\ldots,M_2) = (\mathcal{T}_k,\ldots,\mathcal{T}_2)$. As $\mathrm{rk}(\mathcal{W}_1) = \mathrm{rk}(\mathcal{W}) - 1$ by Equation~\ref{eqn:jasso}, the induction hypothesis implies that $\rho$ is a bijection for the wide subcategory $\mathcal{W}_1$ and that $\mathrm{br\text{-}label}_{\mathcal{W}_{i-1}}[(\mathcal{T}_i)_*,\mathcal{T}_i] = \beta(M_i)$ for $1 < i \leq k$. Since $\mathcal{W}_1$ was bijectively determined by $M_1$, this implies the result.
\end{proof}
\begin{remark}
In particular, Theorem~\ref{thm:mainD} implies that if one identifies every indecomposable $\tau$-rigid module with the corresponding completely join-irreducible torsion class, then the $\tau$-exceptional sequences of $\mathsf{mod}\Lambda$ can be determined from only from the underlying lattice structure of the lattice of torsion classes.
\end{remark}
\subsection{Examples of $\kappa^d$-exceptional sequences}\label{sec:kappa_exceptional_ex}
In this section, we describe the $\kappa^d$-exceptional of the lattices in Figures~\ref{fig:run_ex} and~\ref{fig:weird_clo}(top left). We then observe to what extent these sequences can be used to label the corresponding core label orders and discuss our work in progress towards formalizing these results.
\begin{example}\label{ex:detailed_ex_1}
Let $L$ be as in Figure~\ref{fig:run_ex}. Then $L$ has seven maximal $\kappa^d$-exceptional sequences, in the sense that they cannot be extended by adding more terms on the left. They are:
$$(j_1,j_2,j_4)\qquad\qquad (j_1,j_3,j_2)\qquad\qquad (j_2,j_1,j_4) \qquad\qquad (j_2,j_1,j_4)$$
$$(j_3,j_1,j_2) \qquad\qquad (j_3,j_2,j_1) \qquad\qquad (j_4,j_3).$$
We observe the following:
\begin{enumerate}
\item The number of maximal $\kappa^d$ exceptional sequences coincides with the number of maximal chains in the upper core label order $(L,\leq_{\mathrm{clo}_\uparrow})$, which is redrawn in Figure~\ref{fig:labeled_clo}
\item There is a bijection between $\cji(L)$ and the set of cover relations in $(L,\leq_{\mathrm{clo}_\uparrow})$ of the form $u {\,\,<\!\!\!\!\cdot\,\,\,} \hat{1}$ which sends $j_i$ to $\overline{\kappa}^d(j_i)$. This allows us to label each such cover relation in the upper core label order. See Figure~\ref{fig:labeled_clo}.
\item The pattern in (2) is inherited by each sublattice of the form $\mathrm{core}_\uparrow(j_i)$. Moreover, passing to such a sublattice respects the upper core label order in the following way. Denote $L' = [j_i,\mathrm{pop}_\uparrow(j_i)]$. Then by identifying $\cji(L')$ with $\mathrm{core}\text{-}\mathrm{lab}_\uparrow(j_i)$ as in Lemma~\ref{prop:intervals}, we can then identify the upper core label order $(L',\leq_{\mathrm{clo}_\uparrow})$ with the interval $[\hat{0},\overline{\kappa}^d(j_i)] \subseteq (L,\leq_{\mathrm{clo}_\uparrow})$.
\item Observation (2) allows us to label every cover relation in $(L,\leq_{\mathrm{clo}_\uparrow})$, see Figure~\ref{fig:labeled_clo}. Moreover, the lattice $L$ is extremal (there is a maximal chain which has all of the completely join-irreducible elements as labels). From the maximal chain on the left side of $L$, we define a total order $\preceq$ on $\cji(L)$ by $j_1 \preceq j_2 \preceq j_4 \preceq j_3$. It is straightforward to verify that this total order makes the labeling of $(L,\leq_{\mathrm{clo}_\uparrow})$ into an EL-labeling.
\item There does not seem to be a natural way to turn the $\kappa^d$-exceptional sequences into an EL-labeling of the lower core label order $(L,\leq_{\mathrm{clo}_\downarrow})$, which coincides with the $\kappa$-order as shown in Figure~\ref{fig:ex_clo}. Indeed, all seven maximal chains of $(L,\leq_{\mathrm{clo}_\downarrow})$ contain three cover relations, while only six of the maximal $\kappa^d$-exceptional sequences contain three elements.
\end{enumerate}
\end{example}
\begin{figure}
\begin{tikzpicture}
\begin{scope}[decoration={
markings,
mark=at position 0.65 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-3,4) node[midway, above]{$j_1$};
\draw[postaction=decorate,dashed](0,6) -- (0,4) node[midway,left]{$j_2$};
\draw[postaction=decorate](0,6) -- (3,4) node[midway,above right]{$j_4$};
\draw[postaction=decorate](-3,4) -- (0,2) node[near start, right]{$j_3$};
\draw[postaction=decorate,dashed](0,4) -- (-3,2) node [near start, left]{$j_1$};
\draw[postaction=decorate](3,4) -- (0,2) node [near start,left]{$j_1$};
\draw[postaction=decorate,dashed](0,4) -- (3,2) node [near start,right]{$j_3$};
\draw[postaction=decorate](-3,4) -- (-3,2) node [midway,left]{$j_2$};
\draw[postaction=decorate](3,4) -- (3,2) node [midway, right]{$j_2$};
\draw[postaction=decorate](-3,2) -- (0,0) node[midway,above]{$j_3$};
\draw[postaction=decorate](0,2) -- (0,0) node [midway,left]{$j_2$};
\draw[postaction=decorate](3,2) -- (0,0) node[midway,below right]{$j_1$};
\draw[postaction=decorate,bend right,smooth](0,6) to (-5,3);
\draw[postaction=decorate,bend right,smooth](-5,3) to (0,0);
\node at (-3,0.5) {$j_4$};
\node at (-3,5.5) {$j_3$};
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node [draw,circle,fill=white] at (-5,3) {$j_4$};
\node[draw,circle,fill=white] at (-3,2) {$j_3$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (3,2) {$j_1$};
\node[draw,circle,fill=white] at (3,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-3,4) {$m_1$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\end{tikzpicture}
\caption{A labeling of the upper core label order of the lattice in Figure~\ref{fig:run_ex} by $\kappa^d$-exceptional sequences}\label{fig:labeled_clo}
\end{figure}
\begin{example}\label{ex:detailed_ex_2}
Let $L$ be as in the top left of Figure~\ref{fig:weird_clo}. Then $L$ has ten maximal $\kappa^d$-exceptional sequences, in the sense that they cannot be extended by adding more terms on the left. They are:
$$(j_5,j_2,j_1) \qquad\qquad (j_3,j_5,j_1) \qquad\qquad (j_2,j_3,j_1) \qquad\qquad (j_5,j_1,j_2) \qquad\qquad (j_3,j_5,j_2)$$
$$(j_1,j_3,j_2) \qquad\qquad (j_4,j_3) \qquad\qquad (j_2,j_1,j_4) \qquad\qquad (j_1,j_2,j_4) \qquad\qquad (j_3,j_5).$$
We observe the following:
\begin{enumerate}
\item There are nine maximal chains in $(L,\leq_{\mathrm{clo}_\uparrow})$ (redrawn in Figure~\ref{fig:labeled_clo_2}), but there are ten maximal $\kappa^d$-exceptional sequences. On the other hand, one could argue that $(j_3,j_5)$ should not be counted as maximal since it can be extended to the right to obtain $(j_3,j_5,j_2)$.
\item The element $\hat{1}$ covers four elements in $(L,\leq_{\mathrm{clo}_\uparrow})$. These cover relations are of the form $\overline{\kappa}(j_i) {\,\,<\!\!\!\!\cdot\,\,\,} \hat{1}$ for $i \neq 5$, see Figure~\ref{fig:labeled_clo_2}. In particular, if we ignore $(j_3,j_5)$ as suggested by (1), then the cover relations are in bijection with those join-irreducible elements which can appear on the left of a maximal $\kappa^d$-exceptional sequence.
\item As in Example~\ref{ex:detailed_ex_1}(3), for any $i \in \{1,\ldots,5\}$ we can identify the upper core label order of $[j_i,\mathrm{pop}_\uparrow(j_i)]$ with the interval $[\hat{0},\overline{\kappa^d}(j_i)] \subseteq (L,\leq_{\mathrm{clo}_\uparrow})$. We can thus label $(L,\leq_{\mathrm{clo}_\uparrow})$ with the maximal $\kappa^d$-exceptional sequences other than $(j_3,j_5)$, as shown in Figure~\ref{fig:labeled_clo_2}. \item Both the cover relations $[j_3,m_1]$ and $[j_3,m_2]$ in Figure~\ref{fig:labeled_clo_2} are labeled with $j_5$. This seems to indicate that the maximal chains of $(L,\leq_{\mathrm{clo}_\uparrow})$ cannot be build from bottom-up using $\kappa$-exceptional sequences.
\item The lattice $L$ is not extremal. However, the total order $\preceq$ on $\cji(L)$ given by $j_1 \preceq j_5 \preceq j_2 \preceq j_4 \preceq j_3$ still makes the labeling in Figure~\ref{fig:labeled_clo_2} into an EL-labeling.
\end{enumerate}
\end{example}
\begin{figure}
\begin{tikzpicture}
\begin{scope}[decoration={
markings,
mark=at position 0.65 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-3,4) node[midway, above]{$j_1$};
\draw[postaction=decorate,dashed](0,6) -- (0,4) node[midway,left]{$j_2$};
\draw[postaction=decorate](0,6) -- (3,4) node[midway,above right]{$j_4$};
\draw[postaction=decorate](-3,4) -- (0,2) node[near end, above]{$j_3$};
\draw[postaction=decorate,dashed](0,4) -- (-3,2) node [near end, right]{$j_5$};
\draw[postaction=decorate](3,4) -- (0,2) node [near start,left]{$j_1$};
\draw[postaction=decorate,dashed](0,4) -- (3,2) node [near start,right]{$j_3$};
\draw[postaction=decorate](-3,4) -- (-3,2) node [near end,above right]{$j_5$};
\draw[postaction=decorate](3,4) -- (3,2) node [midway, right]{$j_2$};
\draw[postaction=decorate](-3,2) -- (0,0) node[midway,above]{$j_3$};
\draw[postaction=decorate](0,2) -- (0,0) node [midway,left]{$j_2$};
\draw[postaction=decorate](3,2) -- (0,0) node[midway,below right]{$j_1$};
\draw[postaction=decorate,bend left,smooth](0,6) to (5,3);
\draw[postaction=decorate,bend left,smooth](5,3) to (0,0);
\draw[postaction=decorate](-3,4)--(-6,2) node[midway,above left]{$j_2$};
\draw[postaction=decorate,dashed](0,4)--(-6,2) node[near start,above]{$j_1$};
\draw[postaction=decorate](-6,2)--(0,0) node[midway,below left]{$j_5$};
\node at (3,0.5) {$j_4$};
\node at (3,5.5) {$j_3$};
\end{scope}
\node[draw,circle,fill=white] at (0,0) {$\hat{0}$};
\node [draw,circle,fill=white] at (-6,2) {$j_5$};
\node[draw,circle,fill=white] at (-3,2) {$j_3$};
\node[draw,circle,fill=white] at (0,2) {$j_2$};
\node[draw,circle,fill=white] at (3,2) {$j_1$};
\node[draw,circle,fill=white] at (3,4) {$m_3$};
\node[draw,circle,fill=white] at (0,4) {$m_2$};
\node[draw,circle,fill=white] at (-3,4) {$m_1$};
\node[draw,circle,fill=white] at (5,3) {$j_4$};
\node[draw,circle,fill=white] at (0,6) {$\hat{1}$};
\end{tikzpicture}
\caption{A labeling of the upper core label order of the lattice in Figure~\ref{fig:run_ex} by $\kappa^d$-exceptional sequences}\label{fig:labeled_clo_2}
\end{figure}
\section{Discussion and future work}
Theorem~\ref{thm:mainD} shows that, at least for $\tau$-tilting finite algebras, $\tau$-exceptional sequences can be seen as a special case of a more general combinatorial construction. A natural question is whether other constructions from $\tau$-tilting yield similar combinatorial generalizations. Towards this end, in a future paper we will discuss how to extending Definition~\ref{def:kappa_exceptional} to obtain \emph{signed} $\kappa^d$-exceptional sequences. Alternatively, this amounts to describing when a completely join-irreducible element in a ws-csd lattice is (relatively) ``projective''.
We will also look into extending Theorem~\ref{thm:mainA} in two different ways. On the representation theory side, there exist algebras whose sf-bricks do not admit an rho order, but whose $\tau$-excpetional sequences can still be used to describe an EL-labeling of the poset $\tau\text{-}\mathsf{perp}(\mathsf{mod}\Lambda)$. A simple example is the preprojective algebra of type $A_2$, see Figure~\ref{fig:preproj}. On the combinatorics side, let $L$ be a finite semidistributive lattice. In both Examples~\ref{ex:detailed_ex_1} and~\ref{ex:detailed_ex_2} we showed that the maximal chains of $(L,\leq_{\mathrm{clo}_\uparrow})$ can be labeled by $\kappa^d$-exceptional sequences. Moreover, in both cases, there is a total order $\preceq$ on $\cji(L)$ which makes this labeling into an EL-labeling. In our future work, we will examine when these phenomenon hold more generally. In particular, we will discuss how $\kappa^d$-exceptional sequences are related to pulling triangulations of \emph{Newton polytopes} (also known as \emph{Harder-Narasimhan polytopes} or \emph{submodule polytopes}, see e.g. \cite{BKT,fei1}), as well as other convex polytopes coming from hyperplane arrangements.
\begin{figure}
\begin{tikzpicture}
\begin{scope}[decoration={
markings,
mark=at position 0.65 with {\arrow[scale=1.5]{>}}}
]
\draw[postaction=decorate] (0,6) -- (-4,4) node [midway,above left] {$P_1$};
\draw[postaction=decorate](0,6) -- (-1.5,4) node [midway,left] {$S_1$};
\draw[postaction=decorate](0,6) -- (1.5,4) node [midway,right] {$S_2$};
\draw[postaction=decorate](0,6) -- (4,4) node [midway,above right] {$P_2$};
\draw[postaction=decorate](-4,4) -- (0,2) node [midway,below left] {$S_1$};
\draw[postaction=decorate](-1.5,4) -- (0,2) node [midway,left] {$P_2$};
\draw[postaction=decorate](1.5,4) -- (0,2) node [midway, right] {$P_1$};
\draw[postaction=decorate](4,4) -- (0,2) node [midway,below right] {$S_2$};
\end{scope}
\node[draw,fill=white] at (0,2) {$0$};
\node[draw,fill=white] at (-4,4) {$\mathrm{add}(S_2)$};
\node[draw,fill=white] at (-1.5,4) {$\mathrm{add}(P_1)$};
\node[draw,fill=white] at (1.5,4) {$\mathrm{add}(P_2)$};
\node[draw,fill=white] at (4,4) {$\mathrm{add}(S_1)$};
\node[draw,fill=white] at (0,6) {$\mathsf{mod}\Lambda$};
\end{tikzpicture}
\caption{The lattice of wide subcategories of the preprojective algebra of type $A_2$ and its labeling by $\tau$-exceptional sequences. As a quotient of a path algebra, we have $\Lambda = (1 \leftrightarrows 2)/R^2$, where $R$ is the ideal generated by the arrows. The total order $P_1 \preceq S_2 \preceq P_2 \preceq S_1$ makes this into an EL-labeling. }\label{fig:preproj}
\end{figure}
\bibliographystyle{amsplain}
|
\section{Introduction}
Although the major part of the post-reionization neutral hydrogen (HI) has been destroyed during the Epoch of Reionization by radiation \citep{haermmerle2020}, a small fraction of HI ($x_{\textrm{HI}}\sim2\%$) was shielded inside extremely dense regions ($N_{\textrm{HI}}>2.10^{20}\textrm{cm}^{-2}$)\footnote{$N_{\textrm{HI}}=\int ds\ n_{\textrm{HI}}$, with $n_{\textrm{HI}}$ being HI number density.} protecting them \citep{wolfe2005,bird2014,mingfengho2021}. These regions are known as Damped Ly$\alpha$ (DLA). As they contain a large fraction of the HI post-reionization, they are a direct proof of the distribution of neutral gas and matter as a consequence. The HI remaining in such regions has a higher spin temperature than CMB temperature \citep{kanekar2003}, making it possible to detect their 21 cm line emissions.
Its abundance in the Universe offsets the low probability of the 21 cm emission. Even though 21 cm signals are weak, one can use temporal integration of their photons through \emph{Intensity Mapping} (IM) method in radio frequency to cover a larger and deeper field of the sky. This is an efficient and fast way to compensate its lower angular resolution. It is unnecessary to identify each source because all sources in the same pixel will contribute together. The major difficulty in IM is that the redshifted 21 cm signals are observed together with other different astrophysical sources of emission in the same %
frequency band.
These unwanted emissions, generally called \emph{foregrounds}, cannot be distinguished by the instrument and are usually several orders of magnitude higher than the target signal.
Accurate identification of the 21 cm signal over the sky is a statistical
process and the data need to be pre-processed in order to measure the foreground contribution to observations. The inhomogeneous spatial response of the feed horns and their spectral dependence in a single-dish radio telescope spread an instantaneous observation of the survey and leads to a more complex estimation process \citep{Matshawule}. The instrument behavior also introduces correlated and uncorrelated spectral noise, that is an additional pre-processing difficulty.
To identify 21 cm signals between redshifts 0.13 and 0.45 (corresponding to 980-1260 MHz) in the southern celestial hemisphere of the sky, BINGO Telescope \citep{BINGO_I} will operate with 28 vertically moving feed horns in a single-dish drift-scan IM observation mode \citep{BINGO_III} aiming at a homogeneous covering between -22.5$^o$ and -7.5$^o$ in declination. The project will explore HI emissions to identify primordial information of the Universe that remains over distributions of the matter and can be used both as cosmological constraint parameters and as a cosmological rule. This informational quantity is called \emph{Baryonic Acoustic Oscillations} (BAO). However, a possible identification of the BAO is strongly dependent on the quality of the recovery of the 21 cm signal. Then, it is essential for precise pre- and post-processing of the observed data, as well as a strong control and understanding of the systematics.
A set of BINGO papers have been published describing its current status, particularly the status of its pipeline. The \emph{Generalized Needlet Internal Linear Combination} (GNILC) is a component separation developed by \citep{Remazeilles:2011} and successfully used in the Planck data analysis effort (see, e.g., \citep{GNILCplanck2016}. It is currently the only algorithm used to estimate the foreground contribution to the measurements made by BINGO. \citep{olivari2016,BINGO_IV,BINGO_V}.
In this framework, this work is a natural result of the BINGO pipeline construction, introducing two additional methods. It results from the construction of a pipeline module for \emph{Foreground Removal} and \emph{21 cm signal estimates} that can be explored with different algorithms and in different mathematical domains. Testing how it works and identifying possible problems is the first step for this module. At this point, in the BINGO pipeline construction, it is also crucial to identify which algorithm may be more suitable for testing, both in terms of accuracy and computational cost. Also, implementing and testing different component separation methods is a safe procedure for consistency and redundancy during the analysis of real data.
All algorithms used in this work are blind, i.e. there is no prior information from the source of the emissions from the sky. Each one exploits different spectral and statistical information to estimate the foreground influence, such as \emph{Generalized Morphological Component Analysis} (GMCA) \citep{bobin2007} using morphological diversity and sparsity, \emph{Fast Independent Component Analysis} (FastICA) \citep{maino2002} using statistical independence between non-Gaussian physical sources, and GNILC, which estimates the mixing matrix dimension in both harmonic and spatial domains. At this stage, to maintain more control over the known effects, we have only used uncorrelated thermal noise, leaving frequency dependent contributions such as 1/f noise for later work. The main beam for each feed horn is assumed to be Gaussian and the same for all frequencies.
An important aspect of including several independent component separation methods in the BINGO pipeline is that it will allow us to perform cross-consistency checks on the measured signal and, thereby, get more confidence on the reliability of a future detection of the BAO signal by BINGO
This paper is organized as follows. We begin by describing the sources of the emissions and instrumental contamination that build the data set used (section \ref{Section: 2}). Then, we considered the mathematical model of the sky observation by survey and its posterior instrumental contamination (section \ref{Section: 3}). We show the idea behind each algorithm in section \ref{Section: 4}. We calculate the 21 cm signal after estimating and removing foreground signal in section \ref{Section: 5}. Subsequently we show the tools used to quantify the behavior of each algorithm in removing unwanted signals (section \ref{Section: 6}). The main analysis follows in (section (\ref{Section: 7}). Finally, some discussion and conclusions are done in section (\ref{Section: 8}).
\section{21 cm and Foreground Components}
\label{Section: 2}
In a radio frequency context, noise contributes to the observed signal besides the 21 cm radiation. Separating into 21 cm, foregrounds and instrumentation signals, we first give a short description of what composes our data set. We used the set of simulations described in \citep{BINGO_V}.
In Fig. \ref{fig: Foregrounds}, we can see all components used to perform the foreground removal analysis and estimate 21 cm signal. All maps were masked with a binary mask that allows only regions between -24$^0$ and -8$^0$ in declination.
\subsection{21 cm signals}
Atomic neutral hydrogen (HI) emits a characteristic signal from the hyperfine splitting of the 1S ground state. Such signal has a wavelength of 21 cm, corresponding to the frequency $\nu_{10}\ =$ 1420 MHz in the rest frame. Even with a very low probability of transition ($\sim 10^{-15}$ s), in an astrophysical context, HI clouds contain a large amount of HI that can be excited - either by radiative transition or by collision - and emit radiation \citep{field1958}.
Let $A_{10}$, $m_{\textrm{HI}}$, $n_{\textrm{HI}}$, and $\Omega_{\textrm{HI}}$ be the spontaneous emission coefficient of the 21 cm transition, the HI atom mass, the density number of HI atoms, and the density parameter of HI, respectively; and then, following the description in \citep{Furlanetto2006,pritchard2012}, we can write the observed (average) 21 cm brightness temperature at low-redshifts, $z<2$, as
\begin{eqnarray}
T_{\textrm{HI}}(z) &=& \left(\frac{9hc^3A_{10}}{256\pi^2G k_{\textrm{B}}\nu_{10}^2 m_{\textrm{HI}}} \right) \frac{\Omega_{\textrm{HI}}(z)}{(1+z)^2}\frac{H_{0}^2}{\| dv_{\|}/d\chi \|}
\label{Eq_5}
\end{eqnarray}
where $G$, $c$, and $h$ are the gravitational, speed of light, and Planck (fundamental) constants, respectively. $H_0$ is the Hubble constant. $\| dv_{\|}/d\chi \|$ is the gradient of the specific velocity field along the line of sight. In comoving coordinates $\chi = a^{-1}s$, with the scale factor $a =a(z)$.
Different mechanisms along the
line of sight can cause fluctuations in the HI distribution, which, to first order, can be described as $n_{\textrm{HI}}(z,\hat{n})=\overline{n}_{\textrm{HI}}(z) \left(1+\delta_{\textrm{HI}}(z,\hat{n})\right)$ in the $\hat{n}$ direction. As a consequence, we have a first order temperature perturbation, $T_{\textrm{HI}}(z,\hat{n}) = \overline {T}_{\textrm{HI}}(z) \left (1+\delta T_{\textrm{HI}} (z,\hat{n}) \right)$. Such a perturbation term is one of the main BINGO goals, as shown in \citep{hall2013, BINGO_VII}, we can describe its first-order perturbation in the conformal Newtonian gauge as a composition of the different physical effects
\begin{eqnarray}
\delta T_{\textrm{HI}} (z,\hat{n}) &=& \delta_{\textrm{n}} - \frac{1}{\mathcal{H}}\hat{\textbf{n}}\cdot\left(\hat{\textbf{n}}\cdot \nabla\hat{\textbf{v}}\right)\nonumber\\
& &+ \left(\frac{\textrm{d}}{\textrm{d}\eta}\ln\left(a^3\bar{\textrm{n}}_{\textrm{HI}}\right)-\frac{\dot{\mathcal{H}}}{\mathcal{H}}-2\mathcal{H}\right)\delta \eta + \frac{1}{\mathcal{H}}\dot{\Phi} + \Psi
\label{eqn: deltaT}
\end{eqnarray}
where the first term expresses the contribution from perturbation of the HI density, the second term is related to the \emph{redshift space distortion} from peculiar velocities of the sources. The third term between parenthesis is the zero-order brightness temperature calculated at the perturbed conformal time $\eta$ of the observed redshift. The last two terms come from the Newtonian gauge, where $\Psi$ is the gravitational potential, and $\Phi$ is a local perturbation of the average scale factor. The first of the final term on the Eq. \ref{eqn: deltaT} is associated with the integrated Sachs-Wolf effect, and the other to conversion between increments in redshift to radial distances in the gas frame. $\mathcal{H}$ is the Hubble function in the conformal time.
It is possible to simulate the $\delta T_{\textrm{HI}}$ related to HI temperature and its $\textrm{HI}$-$\textrm{HI}$ angular power spectrum. In our case, we used the \texttt{UCLCl} code \citep{mcleod2017} to generate the $\textrm{HI}$-$\textrm{HI}$
angular power spectra for different redshift bins. As shown in \citep{BINGO_VII}, this code generates output closer to that calculated from Eq. \ref{eqn: deltaT} with a result with a deviation of less than 1\%. \texttt{UCLCl} $\textrm{HI}$-$\textrm{HI}$ angular power spectrum is used by FLASK code\footnote{\href{http://www.astro.iag.usp.br/~flask/}{FLASK webpage}} to create log-normal maps for different tomographic redshift bins. Using a log-normal distribution to describe the temperature fields over the desired space is a better approximation than Gaussian one, as described in \citep{FLASK}. This is mainly because log-normal distribution prevents the field from taking unrealistic values.
\subsection{Foreground Signals}
The 21 cm signals are extremely weak compared to other astrophysical and cosmological signals detected by the receivers. BINGO data will reside in a low-frequency range when compared to PLANCK \citep{planck}. In that range, the main sources originating from the sky, in addition to 21 cm, are called foregrounds. They are divided into Galactic, extragalactic, and cosmological emissions.
In general, Galactic emissions come from the Galaxy's interstellar medium (ISM). Cold molecular and atomic clouds constitute them. The medium between those clouds is partially ionized, most likely formed by supernovae. Such environments are heavily concentrated on the Galactic plane. Then we can see in Fig. \ref{fig: Foregrounds} that significant emissions in radio frequency come from this region. Emissions that contain stronger intensity in this frequency range are mainly such that when energetic charged particles move through the Galactic magnetic field (synchrotron); when free electrons interact with ions in the Galactic ionized media (free-free); and emission because of spinning nano dust grains (anomalous microwave emission). On the other hand, extragalactic emissions are mainly emitted from active galactic nuclei (AGN) and star-forming galaxies (SFG) in radio frequency \citep{chapman2019foregrounds}. AGN emits synchrotron radiation from matter accretion by its central supermassive black holes ejecting it in a beam perpendicular to the accretion plan. SFG produces synchrotron emissions as our galaxy and free-free from regions of ionized hydrogen. In this work, we are assuming extragalactic emissions as radio point sources. This emission component comprises two kinds of unresolved sources: a set of faint objects that cannot be resolved individually and the brightest extragalactic objects that can be identified from catalogs at different frequencies. Lastly, we have assumed that cosmological signals are emitted from very high-z, and in this work composed of cosmic microwave background (CMB).
In summary, we assumed five foreground components: synchrotron, free-free, anomalous microwave emission (AME), radio point sources (FRPS), and CMB. Each foreground component map was generated by Planck Sky Model (PSM) code \citep{PSM}, as described in \citep{BINGO_V}.
\begin{figure*}[h]
\centering
\includegraphics[width=18cm]{images/MapsALLmasked_cartview.jpg}
\caption{Intensity maps in the BINGO sky region are described in antenna temperature: observed (top left), synchrotron (top middle), free-free (top right), radio point sources (middle left), AME (center), CMB (middle right), HI (bottom left), white noise (bottom center) and, HI + white noise (bottom right). The region in gray color represents part of the Galaxy masked.}
\label{fig: Foregrounds}
\end{figure*}
\subsection{Thermal noise}
Aiming to analyze the consistency and understanding of the BINGO pipeline, we assumed the instrumental noise to be thermal only. This noise was considered for a year of observing the BINGO optical design in its Phase 1 as described in \citep{BINGO_II}.
It is crucial to keep in mind that one year of observation gives us a low level of signal stacked from the sky and, consequently, a low signal-noise ratio. BINGO expects to operate for five years, which will result in a better sensitivity than assumed in this work. But, for now, it is essential to understand how the first year and this stage of the pipeline work.
\section{Modeling signals from the sky}
\label{Section: 3}
The \emph{drift-scan} IM survey cannot distinguish different signal sources from the sky and cannot specifically identify HI emissions. However, in radio astronomy, in general, there is no correlation between emissions from different parts of the sky and between different frequencies (see Fig. \ref{fig: components_with_smooth_spectra}), since this occurs when the emission originates from the sum of independent systems that do not have time/phase correlations with each other \citep{liu2020}. Besides that, foregrounds have smooth spectral information, that can be modelled and used to estimate the foreground contributions. Such a process is called \emph{Foreground Removal}. This process results in a signal composed of 21 cm + (instrumental) noise since their (almost) Gaussian behavior is not modelled by blind algorithms.
In summary, the estimation of 21 cm maps using a blind algorithm can be separated into two steps: (1) estimating foreground contribution to the data by exploring its spectral information; and (2) statistically estimating 21 cm maps upon adjoining noise.
The sky signals can modelled considering the contribution of the components described in sec. \ref{Section: 2} to a sky position $p$ (here represented by a pixel with coordinates $(\phi, \theta)$, in a frequency channel $\nu_{i}$. The observed signal can be described as a linear combination of contributions from independent sources $x\left(\nu_{i},p\right)$.
Assuming there are $N_{\textrm{s}}$ physical sources modelled by the blind algorithms and further signals that will not be modelled, described by $x_{21\textrm{cm}}$ and $n_{\textrm{Inst}}$, namely the 21 cm and the instrumental noise, we can thus write the observation as:
\begin{eqnarray}
x\left(\nu_{i},p\right) &=& \sum_{j=1}^{N_{\textrm{s}}}y_j\left(\nu_i,p\right) + x_{21\textrm{cm}}\left(\nu_i, p\right) + n_{\textrm{Inst}}\left(\nu_i, p\right)\\
&=& \sum_{j=1}^{N_{\textrm{s}}}y_j\left(\nu_i,p\right) + n\left(\nu_i, p\right),
\end{eqnarray}
\noindent where $y_j$ is $j$-th foreground component and $\nu_i$ describes the $i$-th frequency or channel. We condense 21 cm + instrumental noise into the same $n$ notation since neither is modelled by blind algorithms. We now explore the fact that the foregrounds have smooth spectral information for each sky direction $p$.
\begin{eqnarray}
x\left(\nu_i,p\right) &=& \sum_{j=1}^{N_{\textrm{s}}}a_{j}\left(\nu_i, p\right)s_j\left(p\right) + n\left(\nu_i,p\right) \nonumber\\
&=& \sum_{j=1}^{N_{\textrm{s}}}a_{ij}\left(p\right)s_j\left(p\right) + n\left(\nu_i,p\right),
\end{eqnarray}
\noindent where $a_{j}\left(p\right)$ describes the spectral law of $j$-foreground component at $p$-direction, and $s_j(p)$ its spatial response at $p$-direction. For different frequencies or channels, we can explore the multichannel information together and write the latter equations in a matrix formulation. If we express the different observations with the same frequency $\nu_i$ as $\textbf{x}\left(\nu_i\right) = \left[x\left(\nu_i,p_1\right), \dots, x\left(\nu_i,p_k\right), \dots\right]^{\textrm{T}}$, this vector covers all positions in the sky. Each such vector corresponds to a HEALPix map \citep{HEALpix} such that we can build an $\mathbb{X}$ matrix with each row corresponding to a map produced for a different frequency, that is, $\mathbb{X} = \left[\textbf{x}\left(\nu_0\right), \textbf{x}\left(\nu_1\right),\cdots, \textbf{x}\left(\nu_{n_{\textrm{ch}}-1}\right)\right]^{\textrm{T}}$. Taking the same idea for the foreground and residual contributions, we can write\footnote{Note that $\mathbb{N} = \mathbb{X}_{21\textrm{cm}} + \mathbb{N}_{\textrm{Inst}}$, with multichannel data matrices for 21 cm and instrumental noise.}
\begin{eqnarray}
\mathbb{X} = \mathbb{A}\mathbb{S} + \mathbb{N},
\end{eqnarray}
where $\mathbb{S}$ is called the \emph{spatial response} and $\mathbb{A}$ the \emph{mixing matrix}. The former has in each row the spatial response of each foreground component and the latter has each column corresponding to a spectral law. Thus, $\mathbb{A}$, $\mathbb{S}$, and $\mathbb{X}$ ($\mathbb{N}$) have dimensions of $N_{\textrm{ch}}\times n_{\textrm{s}}$, $n_{\textrm{s}}\times N_{\textrm{pix}}$, and $N_{\textrm{ch}}\times N_{\textrm{pix}}$, respectively. We assume that $N_{\textrm{pix}}$ is the number of pixels. $N_{\textrm{s}}$ is the actual number of foreground components, which is unknown, while $n_{\textrm{s}}$ is the number of templates used to describe the total foreground emission. On the other hand with the blind algorithm we cannot know the number of sources and we take $n_{\textrm{s}} \in \left[1,N_{\textrm{ch}}\right]$. Therefore, if the algorithm for a given $n_{\textrm{s}}$ estimates $\hat{\mathbb{A}}$ and $\hat{\mathbb{S}}$, it does not exactly estimate each astrophysical component of the foregrounds, but their effectiveness in a nonphysical description of the foregrounds,
\begin{equation}
\hat{\mathbb{Y}} = \hat{\mathbb{A}}\hat{\mathbb{S}}\nonumber
\end{equation}
Estimating $\mathbb{A}$ (and $\mathbb{S}$) is a typical linear inversion problem and different methods/algorithms explore different ways to perform this task.
If the number of channels and sources is equal ($N_{\textrm{ch}}=n_{\textrm{s}}$), the mixing matrix determinant exists, and does not have any noise contribution. In that case, there is a unique solution given by exactly $\mathbb{A}^{-1}$ that provides the spatial response. However, there is a signal not modelled by the algorithm and the mixing matrix dimension is not necessarily (almost never is) equal to the number of frequency/channels. In that case, the mixing matrix has no inverse. \emph{A priori}, a blind algorithm does not define neither estimates the mixing matrix dimensions\footnote{GNILC uses the \emph{Akaike Information Criterium} \citep{AIC1974} to estimate the dimension of the mixing matrix in both harmonic and spatial domains.}. It is precisely the way of estimating the mixing matrix that characterizes each algorithm/method. One general way to solve that problem is looking for a $\mathbb{W}$ filter that, upon operating on the observational matrix, $\mathbb{X}$, gives us the estimated foreground matrix. We can understand that it downweights $\mathbb{N}$, preserving only the foreground signals
\begin{figure}[h]
\centering
\includegraphics[width=8cm]{images/Map__decMinus12dot33__ra20dot91.jpg}
\includegraphics[width=8cm]{images/spectral__decMinus12dot33__ra20dot91__00.jpg}\\
\includegraphics[width=8cm]{images/spectral__decMinus12dot33__ra20dot91__1.jpg}
\includegraphics[width=8cm]{images/spectral__decMinus12dot33__ra20dot91__2.jpg}
\caption{First image shows a red circle corresponding to (RA, DEC)=($-20.91^{\textrm{o}}, 12.33^{\textrm{o}}$) in celestial coordinate. In this position on the sky it is observed different emission within 960-1260 MHz. (Second row) CMB, Point Sources, AME, Free-Free and Synchrotron have smooth spectral response in that range, here divided inside 30 channels. (Third row) on the other hand, 21 cm integrated and (bottom) white noise.}
\label{fig: components_with_smooth_spectra}
\end{figure}
\begin{eqnarray*}
\hat{\mathbb{S}} &=& \mathbb{W}\mathbb{X}\\
&=& \left(\mathbb{W}\mathbb{A}\right)\mathbb{S} + \mathbb{W}\mathbb{N}
\end{eqnarray*}
Following \citep{hyvarinen2001}, a natural way to solve that problem is trying to minimize the $\mathbb{N}$ contribution by a least-squares criterion. This is a deterministic approach to the estimation problem. No assumptions on the probability distribution are necessary,
\begin{eqnarray*}
\nabla_{\mathbb{S}}\bigg|_{\hat{\mathbb{A}},\hat{\mathbb{S}}}\left( \frac{1}{2} \|\mathbb{N} \|^2 \right)= 0,
\end{eqnarray*}
that gives us
\begin{eqnarray}
\mathbb{W} = \mathbb{A}^+ = \left(\hat{\mathbb{A}}^{\textrm{T}}\hat{\mathbb{A}}\right)^{-1}\hat{\mathbb{A}}^{\textrm{T}}
\label{eqn:W_LS}
\end{eqnarray}
known as \emph{Moore-Penrose pseudoinverse}, $\mathbb{A}^{+}$.
We used a canonical metric with $\ell_2$ in the minimization, $\|\cdot\|=\|\cdot\|_{\ell_2}$\footnote{Let $\mathbb{B}$ be an arbitrary matrix. We can define the canonical metric as $\|\mathbb{B}\|^2=\mathbb{B}^{\textrm{T}}\mathbb{B}$}, so far. But, we can also use another metric like e.g., $\|\cdot\|_{\textrm{Q}}$\footnote{Let $\mathbb{B}$ and $\mathbb{Q}$ be arbitrary matrices. $\|\mathbb{B}\|_{\mathbb{Q}}=\mathbb{B}^{\textrm{T}}\mathbb{Q}\mathbb{B}$.}, with some $\mathbb{Q}$ matrix. In this case, the minimization is a generalization of the previous norm and is called \emph{Generalized Least Squares} (GLS).
So
\begin{eqnarray*}
\nabla_{\mathbb{S}}\bigg|_{\hat{\mathbb{A}},\hat{\mathbb{S}}}\left( \frac{1}{2} \|\mathbb{N} \|_{\mathbb{Q}}^2 \right)= 0.
\end{eqnarray*}
Therefore, the filter is
\begin{eqnarray}
\mathbb{W} = \left(\hat{\mathbb{A}}^{\textrm{T}}\mathbb{Q}\hat{\mathbb{A}}\right)^{-1}\left(\hat{\mathbb{A}}^{\textrm{T}}\mathbb{Q}\right),
\label{eqn:W_GLS}
\end{eqnarray}
which also known as \emph{Gauss-Markov estimator} in Information Theory.
Here, we use the definition of the difference between observational matrix and estimated matrix of foregrounds as a residual matrix given by
\begin{equation*}
\mathbb{R} = \mathbb{X} - \hat{\mathbb{Y}} = \left(1-\mathbb{W}_{\textrm{FG}}\right)\mathbb{X},
\end{equation*}
where $\mathbb{W}_{\textrm{FG}} \doteq \mathbb{A}\mathbb{W}$.
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/Eigenvalues.jpg}
\caption{Normalized eigenvalues of the observational covariance matrix. Given that mixing matrix dimension is limited by number of channel, there are only three eigenvalues contributing significantly which corroborates with better results of FastICA and GMCA.}
\label{fig:eigenvalues}
\end{figure}
\section{Foreground Removal methods}
\label{Section: 4}
To begin with, in order to treat the sky images we must account for the undesirable foreground contribution to the total observed signal. We use here three blind methods: the algorithms do not assume any previous knowledge about the emission sources. Their general description and assumptions will be given below.
\subsection{Fast Independent Component Analysis}
\emph{Independent Component Analysis} (ICA) is an algorithm applied to astrophysical (and cosmological) observations to model or remove foregrounds using the hypothesis that astrophysical sources are statistically (and mutually) independent. Signals from different sources being statistically independent mean that different signals do not contain information from one another. In mathematical jargon, when two signals are (statistically) independent, their joint distribution function is the product of their distributions.
ICA-based algorithms look for a linearly transformed matrix $\mathbb{\hat{S}}=\mathbb{W}\mathbb{X}$, where each row is composed of a transformed vector and all components are mutually independent \citep{hyvarinen1999}. It is essential to point out here that this work uses a noise-free ICA-based algorithm. ICA-based algorithms just seek to model foregrounds, and non-smooth spectral components are leading as rest of the modelling. Let $\mathbf{w}_{i}$ be the $i$th row of $\mathbb{W}$, i.e., $\mathbb{W} = \left[\mathbf{w}_1, \mathbf{w}_2, \dots, \mathbf{w}_{N_s} \right]^{\textrm{T}}$. We can look for
\begin{equation}
\mathbf{y}_i=\sum_{j=1}^{N_{\textrm{ch}}}W_{ij}\mathbf{x}_j=\mathbf{w}_{i}^{\textrm{T}}\mathbb{X}
\end{equation}
as an independent component, that is equivalent to look for it as a maximally non-Gaussian component \citep{hyvarinen2001}. For each iteration, the ICA-based algorithm look for a new transformed variable that is more independent than the variable iterated previously, in the sense of minimizing the negentropy defined ahead. Thus, if we look for something like $\mathbf{y}_i = \mathbf{w}_{i}^{\textrm{T}}\mathbb{X} = (\mathbf{w}_{i}^{\textrm{T}}\mathbb{A})\mathbf{s}$, but for residuals, the operator $\mathbf{e} = (\mathbf{w}^{\textrm{T}}\mathbb{A})^{\textrm{T}}$ assumes the value 1 in $i$-th position of the vector, and 0 otherwise. That occurs when $\mathbf{w}_{i}$ is one of the rows of the inverse of the mixing matrix. But neither mixing matrix necessarily has an inverse nor we are dealing with a noise-free inverse problem, so we cannot know exactly what $\mathbf{w}$'s are and we need to find an estimator for them. By the central limit theorem, the linear combination $\mathbf{e}^{\textrm{T}}\mathbf{s}$ is more Gaussian than any of the components of $\mathbf{s}$ becoming least Gaussian when it is (approximately) equal to one of the components of $\mathbf{s}$, as, e.g., when $\mathbf{e}^{\textrm{T}}\mathbf{s}$ is (close to) $s_i$. When this happens, we get the $\mathbf{e}_{i}$ corresponding to $\mathbf{w}_i$.
Therefore, we need to maximize the non-Gaussianity of the $\mathbf{e}^{\textrm{T}}\mathbf{s}$ to get each component corresponding to a respective $\mathbf{e}_i$, or equivalently $\mathbf{w}_i$. That is nothing more than a search for a direction for which the components are maximally non-Gaussian.
\citep{hyvarinen2001} suggested the \emph{negentropy} function as such a measure, in contrast to the higher-order (statistical) method as \emph{kurtosis} (the forth-order cumulant of a random variable) which by that \emph{"is not a robust measure of nongaussianity"}. The negentropy is defined as
\begin{equation*}
J\left(\mathbf{\xi}\right) = H \left(\mathbf{\xi}_{\textrm{G}}\right) - H\left(\mathbf{\xi}\right),
\end{equation*}
where $H$ is the entropy function, $\mathbf{\xi}$ a vector, and the subscript $\textrm{G}$ identifies a Gaussian random variable of the same covariance matrix as $\mathbf{\xi}$. The disadvantage of using negentropy is that it is computationally hard to calculate. So it is necessary to estimate it through some \emph{contrast function} $g$,
\begin{equation*}
J\left(\mathbf{\xi}\right) \sim \Big[\textrm{E}\big\{g\!\left(\mathbf{\xi}\right)\big\} - \textrm{E}\{g\!\left(\mathbf{\nu_{\textrm{G}}}\right)\}\Big]^2
\end{equation*}
for non-quadratic function $g$, and a Gaussian variable $\nu_{\textrm{G}}$ of zero mean and unit variance. The ICA algorithm used in this paper was the FastICA, a fast fixed-point algorithm that uses negentropy as a measure of non-Gaussianity and $\xi=\mathbf{w}^{\textrm{T}}\mathbf{z}$, where $\mathbf{z}$ is the whitened data of $\mathbf{x}$. In this work, we have used $g(\cdot) = \log\cosh(\cdot)$, with 20 maximum number of iterations and 0.01 of tolerance with FastICA by scikit-learn package\footnote{\url{https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.FastICA.html}} to get the mix matrix and so to estimate the $\mathbb{W}$. For a complete description of the algorithm, please see \citep{hyvarinen2001}.
\subsubsection{FastICA as an optimization problem}
It is straightforward to note that it is possible to describe the FastICA algorithm as an optimization problem. It aims to maximize the non-Gaussianity of the signals from negentropy we also constraint $\mathbb{W}$ matrix to be orthogonal. We define
\begin{eqnarray}
\{\mathbb{W}\} &=&\underset{\mathbb{W}}{\textrm{argmax}}\Bigg \{ J\left(\mathbb{W}^{\textrm{T}}\mathbb{X}\right) + \lambda\|\mathbb{W}\mathbb{W}^{\textrm{T} }- \mathbb{I} \| \Bigg\}
\end{eqnarray}
where $\textrm{argmax}$ finds the variable $\mathbb{W}$ that maximizes the relation inside the brackets.
FastICA was introduced for the separation of astrophysical components in \citep{maino2002}. Some works used FastICA in 21 cm cosmology in the reionization \citep{chapman2012} and post-reionization \citep{carucci2020} contexts.
It is important to note that FastICA estimates neither 21 cm nor thermal noise signals: these are residuals of the process of the foreground estimation. FastICA gives a unique solution to the problem. Each column of $\mathbb{A}$ is unique up to the signal.
\subsection{Generalized Morphological Component Analysis}
Generalized Morphological Component Analysis (GMCA) is based on two pieces of information: \emph{morphological diversity} and \emph{sparsity}. We will briefly describe these properties below.
\subsubsection{Sparsity and morphological diversity}
A signal $\mathbf{y}$ is said to be \emph{sparse} in a dictionary\footnote{A Dictionary $\mathcal{D}$ is a collection of parameterized waveforms $\{\phi_{\gamma}; \gamma \in \Gamma\}$. The decomposition of a signal $s$ on a dictionary $\mathcal{D}$ is described in an approximated way as $$s = \sum_{\gamma \in \Gamma}\alpha_{\gamma}\phi_{\gamma} + \textrm{Res},$$ where here we represent $\textrm{Res}$ as the residual of the representation. Examples of dictionaries are: Fourier, Wavelet, Gabor, others. For a more detailed description, we recommend the reader to see \citep{chen1995,chen2001}. } $\boldsymbol{\Phi}$ if it can be represented a by few elements of $\boldsymbol{\Phi}$. For instance, the signal $\mathbf{y}$ can be represented by
\begin{equation*}
\mathbf{y}= \boldsymbol{\alpha\Phi} = \sum_{\gamma \in \Gamma}\alpha^{\gamma}\phi_{\gamma}
\end{equation*}
where $\alpha^{\gamma}$ is $\gamma$th coefficient of the representation. If $\mathbf{y}$ is sparse in $\boldsymbol{\Phi}$ the number of nonzero coefficients is small. More precisely, $\mathbf{y}$ is $k$-sparse in $\boldsymbol{\Phi}$ if the number of non vanishing coefficients is $k$. However, signals of practical interests, in general, are not strictly sparse. Therefore, it is important to find representations in which the signals are sparse. In this sense using wavelets for information compressing is useful since they are very well located in both spatial (or time) and frequency domains \citep{starckWT1997,starckIUWT2007}. Because of this feature, wavelets are useful in the sparse representation of a signal. Wavelets on the sphere are also well located in both harmonic and pixel domains, been thus of widespread use in order to analyse astrophysical maps.
The first paper to use sparsity as a source BSS problems was \citep{Zibulevsky2000}. The \citep{starckMCA2004redundant} shows the method of \emph{Morphological Component Analysis} (MCA). It uses the sparsity idea for signals as linear instantaneous mixture from different sources through overcomplete dictionaries for representation. The method takes advantage of the sparse representation to separate features based on their morphologies. That is, sparsity and \emph{morphological diversity} are extensively used. Its multichannel extension gives rise to \emph{Multichannel}-MCA (MMCA), which assumes each signal source $s_j$ can be sparsely well represented in a given dictionary $\boldsymbol{\Phi}_j$. Finally, \citep{bobin2007} shows the GMCA, that expands MMCA to assume $s_j$ can be represented not only a possible dictionary, but by a linear combination of different dictionaries (here assumed as being orthogonal bases), that is, a superdictionary $\mathcal{D} = [\boldsymbol{\Phi}^{\textrm{T}}_1, \dots, \boldsymbol{\Phi}^{\textrm{T}}_D]^{\textrm{T}}$.
Each source $s_j$ has a representation given by
\begin{eqnarray}
s_{j} &=& \sum_{k=1}^{D}\phi_{jk} = \sum_{k=1}^{D}\alpha_{j}^{\ k}\boldsymbol{\Phi}_{k},
\end{eqnarray}
or, equivalently, in a matrix representation,
\begin{eqnarray}
\mathbb{S}&=& \boldsymbol{\alpha}\mathcal{D}.
\end{eqnarray}
We can say that $s_j$ is modeled as a linear combination of D-\emph{morphological components}, $\{\phi_{jk} \}_{k=1}^{D}$, each being morphological component sparse in a given orthogonal base.
\subsubsection{GMCA as an optimization problem}
When we assume that each row (signal) of $\mathbb{S}$, is sparse in $\mathcal{D}$, we inevitably run into the problem of having more than one solution because there is more than one representation of each signal, since $\mathcal{D}$ is overcomplete. Thus we need to add some constraint to get a single solution. Indeed, since we want the most sparse representations, we seek for the representations with the least number of the coefficients. The original problem was proposed using a constraint as $\|\alpha\|_{\ell_0}$ that is a $\ell_0$-regularization problem, but this one is not computationally efficient, since it is a non-convex and combinatorial problem. We can ou problem from a an ${\ell_0}$ to a $\ell_1$-regularization problem using $\|\alpha\|_{\ell_1}$,
\begin{eqnarray}
\{\hat{\mathbb{A}},\boldsymbol{\hat{\alpha}}\} &=&\underset{\mathbb{A},\boldsymbol{\alpha}}{\textrm{argmin}}\Bigg \{ \|\mathbb{X}-\mathbb{A}\boldsymbol{\alpha}\mathcal{D}\|^2_{\textrm{F},\mathbb{C}^{-1}_{\mathbb{N}}} + 2\lambda\sum_{j=1}^{N_s}\|\alpha_{\!j}\|_{\ell_1}\Bigg \}
\end{eqnarray}
with the norm of the first term being the Frobenius norm with respect to inverse of the noise covariance matrix\footnote{Let $\mathbb{Z}$ be a matrix. $\mathbb{Q}$-weighted Forbenius norm will be defined as $\|\mathbb{Z}\|_{\textrm{F},\mathbb{Q}}^2=\textrm{Tr}\left(\mathbb{Z}\mathbb{Q}^{-1}\mathbb{Z}^{\textrm{T}}\right)$. In case $\mathbb{Q}$ is the identity matrix, the norm will be just the trace and we denote it by $\|\cdot\|_{\textrm{F}}$.}, $\mathbb{C}_{\mathbb{N}}$, and $\lambda$ is a regularization parameter important to the thresholds. We do not detail the algorithm, which can be seen in \citep{bobin2007,GMCA2008SZ}; we can resume the algorithm idea, in each $i$th iteration, by starting with $\mathbb{A}^{(i-1)}$ to estimate $\mathbb{S}^{(i)}$, and then update the mix matrix to new estimation, $\mathbb{A}^{(i)}$. Which can be done first by applying hard-thresholding operator in $\mathbb{S}^{(i-1)}$ description estimate in $\mathcal{D}$, and latter backward transform to real domain: $\mathbb{S}^{(i)} = \textrm{HT}_{\lambda^{(i-1)}}\!\left[\mathbb{A}^{(i-1)+}\mathbb{X}\mathcal{D}^{\textrm{T}}\right]\mathcal{D},$ where $\lambda^{(i-1)}$ is the $i$th threshold. Thereafter, we need to update the mix matrix $\mathbb{A}^{(i)} = \mathbb{X}\mathbb{S}^{(i)+}$.
In this work, we used starlet\footnote{Also known as \emph{Isotropic Undecimated Wavelet Transform} on the Sphere.} as a dictionary \citep{starckIUWT2007}, that is a well known spherical wavelet transform in astronomical image analysis, since it explores the features from astrophysical objects that are more or less isotropic (in most cases). We used a starlet decomposition with a scale of decomposition as can be seen in the Fig. \ref{fig:star1}, where there is a map with 21cm + foregrounds (first row), its scale coefficient map (second row) and wavelet coefficient map (third row).
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/STARLETS_bin10_mapTotal.jpg}
\includegraphics[width=7cm, height=3cm]{images/starletfilters_bin10_mapTotal.jpg}
\caption{The first row is the input map composed of all foregrounds and 21 cm emission. The input map transformed by starlet transformation can be seen in the second and third rows, where we have a starlet coarse coefficient and the starlet detail coefficient maps, respectively. The last row represents (blue) spherical scaling function and (orange) spherical wavelet function at the null scales.}
\label{fig:star1}
\end{figure}
\subsection{Generalized Needlet Internal Linear Combination}
\label{sec:GNILC0}
As the effective ratio between 21 cm and observational signals depends on the pixels (i.e., direction), the number of the (nonphysical) templates to describe the contribution of the foregrounds changes across the sky. This fact can be checked if we look at the ratio in the directions of the Galaxy and outside of Galaxy. Also, although thermal noise and 21 cm signals follow a Gaussian distribution among channels, foregrounds do not, and that ratio depends on the spectrum too. Then, it is essential to have a detailed description of the foregrounds contribution to both spatial and spectral information. Moving the spectral analysis to the harmonic domain, we can use the localized filtering needlets feature \citep{NeedletsCMB2008}. In the Fig. \ref{fig:need1}, we can see the specific needlets described in \citep{NILC2013} through a set of cosine filters with \emph{bandcenters}\footnote{The band centres determine the peak, minimum and maximum multipoles in the needlet filter described in \citep{NILC2013}} 0, 128 and 383, each corresponding to a finite number of the multipoles (\emph{needlet bands}). Each needlet band decomposes the observational map in different scales, as shown in Fig. \ref{fig:need1}.
\cite{Delabrouille2009} applied the \emph{Internal Linear Combination} (ILC) in the needlet space for the WMAP data, called it \emph{Needlet}-ILC (NILC) to extract the CMB maps. \cite{Remazeilles:2011} generalized it as a multidimensional filter in a new algorithm called \emph{Generalized}-NILC (GNILC). He modelled the observation as multidimensional components (templates) instead of one single template.
\subsubsection{GNILC as an optimization problem}
\label{sec:GNILC}
Roughly speaking, GNILC algorithm can be resumed as a search for a filter that has a unitary response for the target signal and blocks the others. Following \citep{Remazeilles:2011} we can described it as an optimization problem,
\begin{eqnarray}
\{\hat{\mathbb{A}},\hat{\mathbb{S}}\} = \underset{\mathbb{A},\mathbb{S}}{\textrm{argmin}}\Bigg \{ \|\mathbb{W}\|^2_{\textrm{F},\mathbb{R}^{-1}_{\mathbb{X}}} + \|\Lambda^{\textrm{T}}\left(\mathbb{A}-\mathbb{W}\mathbb{A}\right)\|_{\textrm{F}}\Bigg \},
\end{eqnarray}
where $\Lambda$ is a matrix containing Lagrange multipliers. The solution of this problem is exactly Eq. \ref{eqn:W_GLS} with $\mathbb{Q}=\mathbb{R}^{-1}_{\mathbb{X}}$. However, GNILC does that in needlet space; that is, using the covariance matrix of the needlet coefficients in each range of multipoles, obtaining a filter per needlet bands, and then estimating foregrounds contribution in those. Finally, applying inverse needlet transform in these last maps to estimate the foreground maps in real space.
GNILC uses a prior angular power spectrum of the target signal to measure the signal-noise ratio and separate the degree of freedom of the target and unwanted signals. That is, it uses the prior template of the 21-cm signal to measure the degrees of freedom of the foreground contribution, $N_s$, in the observation per needlet bands and pixel.
For 21 cm maps, GNILC was first used by \cite{olivari2016}. Unlike the previous GNILC works, \cite{olivari2016} uses $N_s$ to represent degrees of freedom of foregrounds + instrumental noise, and estimates its value using the \emph{Akaike Information Criterium} (AIC) method \citep{AIC1974,AICbozdogan1987model}. Here, we separate the foregrounds from the 21 cm and (instrumental) noise. They are projected onto a $(n_{\textrm{ch}} - N_s)$-dimensional subspace. In other words, GNILC operates a \emph{singular value decomposition} (SVD) of the whitened data covariance matrix,
\begin{equation}
\mathbb{\hat{R}}^{-1/2}_{\textrm{prior}}\mathbb{R_{\mathbb{X}}}\mathbb{\hat{R}}^{-1/2}_{\textrm{prior}} = \mathbb{\hat{U}}\mathbb{\hat{D}}\mathbb{\hat{U}}^{\dagger},
\label{eq:SVDdec}
\end{equation}
where $\mathbb{\hat{R}}_{\textrm{prior}}$ is the covariance matrix of the prior template, and $\mathbb{\hat{U}}$ and $\mathbb{\hat{D}}$ are matrices composed of eigenvectors and eigenvalues of the term on the left hand in Eq. \ref{eq:SVDdec}, respectively. The matrix of the eigenvectors is split into a part of foregrounds (FG) contribution and another part of 21cm + noise (21cm+N), $\mathbb{\hat{U}}=\left[\mathbb{\hat{U}}_{\textrm{FG}},\mathbb{\hat{U}}_{21\textrm{cm}+\textrm{N}}\right]$. The AIC estimates the dimension of the 21cm + noise space, that is, the dimension of $\mathbb{\hat{U}}_{21\textrm{cm}+\textrm{N}}$. Thus, GNILC obtains the mixing matrix by
\begin{eqnarray}
\mathbb{\hat{A}}=\mathbb{\hat{R}}_{\textrm{prior}}^{1/2}\mathbb{\hat{U}}_{21\textrm{cm}+\textrm{N}},
\end{eqnarray}
With the mixing matrix estimated and the needlet coefficient covariance matrix, GNILC estimates the multidimensional ILC filter in each pixel and bandwidth, identified here by $j$,
\begin{eqnarray}
\hat{\mathbf{x}}^{(j)}_{21\textrm{cm}+N}(p) = \mathbb{W}^{(j)}(p)\mathbf{y}^{(j)}(p).
\end{eqnarray}
with $\hat{\mathbf{x}}^{(j)}_{21\textrm{cm}+N}(p)$ the multichannel 21cm + noise estimation per pixel and in $j$th bandwidth. After doing all the needlet bands and estimating all the target signal per pixels and range of multipoles, GNILC performs the inverse needlet transform (INT) and gives the 21cm+noise maps in the pixel space,
\begin{equation*}
\{ \hat{\mathbf{x}}^{(j)}_{21\textrm{cm}+N}(p) \} \stackrel{\textrm{INT}}{\longrightarrow} \hat{\mathbf{x}}_{21\textrm{cm}+N}(p)
\end{equation*}
For a complete description of the algorithm, please see \citep{BINGO_V}. As a prior, we used a template for each channel of an arbitrary realization of 21cm + white noise.
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/GNILCneedlets_bin10_mapTotal_BINGOskycover.jpg}
\includegraphics[width=7cm, height=3cm]{images/GNILCneedletfilters_bin10_mapTotal.jpg}
\caption{The first row is the input map composed of all foregrounds and 21 cm emission. The input map transformed by needlet transformation can be seen in the three next images. The second row is the needlet coefficient map from null bandcenter, the third for bandcenter 128, and the forth for 383. The last plot represents (blue) spherical scaling function and (orange) spherical wavelet function.}
\label{fig:need1}
\end{figure}
\subsection{General Observations}
It is worth stressing that we first used all maps in all channels using HEALPix NSIDE=512 resolution, convoluted with a main Gaussian beam with FWHM=40 arcmin
Then the convoluted maps are degraded to NSIDE=256. This is an idealistic assumption since not all feed horn beams have the exact Gaussian shape. Indeed, although the central feed horn in the focal plane has a very Gaussian main beam, the farther the feed horn is from the center of the focal plane, the less Gaussian its main beam will be. Furthermore, the way reflectors are constructed also significantly alter the main beam of the feed horns. The BINGO reflectors will be built modularly, connecting $\sim$ 1x2 m$^2$ panels. This construction affects the beam shape, mainly the more distant lobes. \citep{BINGO_III} describes the BINGO optical design. The description of the effects of building the reflectors will be described in a future paper.
The main beams can also depend on the frequency. These different effects lead to a more complex foreground removal process, and the algorithms need to be adapted to that situation. We shall deal with these issues in future works.
\vspace{0.3cm}
\section{Noise debias process}
\label{Section: 5}
After subtracting the foreground contributions to the "observed" maps, we need to estimate the 21 cm signal, extracting it from a residual map composed of 21-cm signal and instrumental noise.
\begin{equation*}
\hat{\mathbb{X}}_{\textrm{FG}} \doteq \hat{\mathbb{Y}} = \mathbb{W}_{\textrm{FG}}\mathbb{X}.
\end{equation*}
Therefore,
\begin{equation*}
\hat{\mathbb{X}}_{21\textrm{cm}+\textrm{N}} = \mathbb{X} - \mathbb{W}_{\textrm{FG}}\mathbb{X}.
\end{equation*}
We used between 25 and 400 different simulated 21 cm and noise maps. Estimating 21 cm power spectrum is done here using ANAFAST, an angular power spectrum estimator from HEALPix. Let $a^{\textrm{(F)}}_{\ell,m}$ be the harmonic coefficients for a \emph{F} field. The angular power spectrum for \emph{F}
is given by
$C_{\ell}^{\textrm{(F)}}
=(2\ell + 1)^{-1}
\sum_{m=-\ell}^{\ell}
a_{\ell m}^{\textrm{(F)}\dagger}
a^{\textrm{(F)}}_{\ell m}$.
From now on we rely on the notation ${}_{\left(\textrm{b}\right)}^{\left(a\right)}\hat{C}^{\left(c\right)}_{\ell}$
where $a$, $b$ and $c$ denote realization, algorithm used to estimate the foreground contribution (GNILC, GMCA or FastICA) and type of map (for instance, 21 cm and/or noise N), respectively. The "hat" $\hat{}$ defines the estimated value of the angular power spectrum derived from the debiasing process. Therefore, for a specific $M$ algorithm and a specific $L_i$ realization, the estimated 21 cm angular power spectrum is:
\begin{eqnarray*}
{}_{\left(\textrm{M}\right)}^{\left(L_i\right)}\hat{C}^{\left(21cm\right)}_{\ell} &=& \frac{{}_{\left(\textrm{M}\right)}^{\left(L_i\right)}C^{\left(\textrm{R}\right)}_{\ell}}{S_{\ell}} - \bigg \langle {}^{\left(L_j\right)}C^{\left(\textrm{N}\right)}_{\ell} \bigg \rangle_{L_j}^{j \neq i},
\end{eqnarray*}
where R is the residual, that is, the components not modeled by the algorithm (21 cm and N) plus foregrounds modeling error; and
\begin{eqnarray}
S_{\ell} &=& \Bigg\langle \frac{{}_{\left(\textrm{M}\right)}^{\left(L_j\right)}C^{\left(\textrm{R}\right)}_{\ell}}{{}^{\left(L_j\right)}C^{(\textrm{21cm + N})}_{\ell}} \Bigg \rangle_{L_j}^{j \neq i},
\end{eqnarray}
and $\langle \rangle_{L_j}^{j \neq i}$ describes the average over all realizations except the $L_i$ realization; that is, for $L_j\in\left(L_0, \cdots,L_{i-1},L_{i+1}, \cdots,L_{n_r-1}\right)$.
\section{Statistical tools}
\label{Section: 6}
We used a resampling method to estimate the variance of each angular power spectrum recovered after foreground removal and noise debiasing. We tested both \emph{Jackknife} and \emph{Bootstrap} methods and did not identify any relevant distinction between them. So we decided to use Jackknife because it is computationally faster.
We applied $\chi^2$ per multipole tests to assess the statistical quality of the results.
\subsection{Jackknife variance estimator}
\label{sec:jack}
The Jackknife method is a resampling strategy that works by sequentially deleting one observation at a time in a given data set, then recomputing the desired statistics. In our case, we want to estimate the variance of the angular power spectrum for each multipole and channel. We thus define a vector containing the angular power spectrum with fixed multipole $\ell$ and channel $\nu$ and all, except the specific $i$-th, $N$ different realizations.
\begin{eqnarray*}
{}^{\left(-i\right)}\textbf{C}_{\ell} \doteq \left({}^{\left(L_0\right)}C_{\ell}, \cdots,{}^{\left(L_{i-1}\right)}C_{\ell},{}^{\left(L_{i+1}\right)}C_{\ell}, \cdots, {}^{\left(L_{n_r-1} \right)}C_{\ell}\right)^{\textrm{T}}.
\end{eqnarray*}
Thus, we can build $n_r$ new samples from the original one. Let $G$ be a function that gives some estimator associated with the sample. We can define the estimators by
\begin{eqnarray*}
{}^{(-i)}\hat{\theta}_{(\ell)} &=& G\left({}^{\left(-i\right)}\textbf{C}_{\ell}\right).
\end{eqnarray*}
It is convenient to define a matrix composed of all estimator values
\begin{eqnarray*}
^{(-i)}\hat{\Theta} = \left({}^{(-i)}\hat{\theta}_{(\ell_{\textrm{min}})},\ \cdots, {}^{(-i)}\hat{\theta}_{(\ell_{\textrm{max}})}\right)^{\textrm{T}}
\end{eqnarray*}
Assuming $G$ to be the function that calculates the mean value,
\begin{equation}
{}^{(-i)}\hat{\theta}_{(\ell)} = \frac{1}{n_r}\sum_{\underset{j\neq i}{j=0}}^{n_r -1}{}^{\left(L_{j}\right)}C_{\ell},
\end{equation}
we can get the Jackknife variance by
\begin{eqnarray}
\hat{\sigma} = \frac{n_r-1}{n_r} \sum_{j=0}^{n_r-1}\left({}^{(-j)}\hat{\Theta} - \langle \Theta \rangle\right)\left({}^{(-j)}\hat{\Theta} - \langle \Theta \rangle\right)^{\textrm{T}},
\label{Eq:Jackknife}
\end{eqnarray}
where
\begin{eqnarray*}
\langle \Theta \rangle := \frac{1}{n_r} \sum_{j=0}^{n_r-1} {}^{(-j)}\hat{\Theta}.
\end{eqnarray*}
\subsection{The $\chi^2$ test
To quantify our results, we have used the $\chi^2$ test, which is a measure of the goodness-of-fit of the model to the estimated data. We also used the Jackknife variance described in sec. \ref{sec:jack} (Eq. \ref{Eq:Jackknife}) as an estimator of the variance for each angular power spectrum per multipole and per channel.
Considering the vector of the angular power spectrum of a $X$ map ($X$ is arbitrary here) on a specific channel $\nu$ as being $\textbf{C}^{X}\left(\nu\right) = \left(C^{X}_{\ell_{\textrm{min}}}\left(\nu\right), C^{X}_{\ell_{\textrm{min}}+1}\left(\nu\right), \cdots, C^{X}_{\ell_{\textrm{max}}}\left(\nu\right) \right)$, we can calculate the $\chi^2$ for a specific $i$-realization, $L_{i}$, and $M$-algorithm, i.e., $M \in \{\textrm{GNILC}, \ \textrm{GMCA},\ \textrm{FastICA}\}$, as\\
\begin{eqnarray}
{}_{\left(M\right)}^{\left(L_i\right)}\chi^{2}\left(\nu, \ell\right) = \frac{\left({}^{(L_i)}_{(M)}\hat{C}_{\ell}^{21cm}\left(\nu\right) - {}^{(L_i)}C_{\ell}^{\textrm{input}}\left(\nu\right)\right)^2}{\hat{\sigma}_{\ell}^2\left(\nu\right)}
\label{eq: xi2}
\end{eqnarray}
with $\hat{\sigma}^2\left(\nu\right) = \left(\hat{\sigma}^2_{\ell_{\textrm{min}}}\left(\nu\right), \cdots, \hat{\sigma}^2_{\ell_{\textrm{max}}}\left(\nu\right)\right)$ vector calculated by Jackknife resampling method.
Other interesting measures are those which can quantify a similar $\chi^2$ quantity per channel and a general $\chi^2$, considering all contributions in multipoles and channels. The former can be defined as the average of all $\chi^2$ in the same channel summing over all multipoles between a minimum and maximum values, $\ell_{\textrm{min}}$ and $\ell_{\textrm{max}}$, respectively, and dividing by its dimension,
\begin{eqnarray}
{}_{\left(M\right)}^{\left(L_i\right)}\chi^{2}_{\textrm{eff}}\left(\nu\right) = \frac{1}{n_{\ell}} \sum_{\ell = \ell_{\textrm{min}}}^{\ell_{\textrm{max}}} {}_{\left(M\right)}^{\left(L_i\right)}\chi^{2}\left(\nu, \ell\right),
\label{eq: xi2eff}
\end{eqnarray}
where $n_{\ell}$ is the length of $\ell = \left(\ell_{\textrm{min}},...,\ell_{\textrm{max}}\right)$. The general $\chi^2$ is the average of the values of both channels and multipoles,
\begin{eqnarray}
{}_{\left(M\right)}^{\left(L_i\right)}\chi^{2}_{\textrm{overall}} = \frac{1}{n_{\nu}}\sum_{\nu=0}^{n_{\nu}-1}{}_{\left(M\right)}^{\left(L_i\right)}\chi^{2}_{\textrm{eff}}\left(\nu\right).
\end{eqnarray}
Another measurement which we used, similar to the \ref{eq: xi2eff}, is the mean over channels
\begin{eqnarray}
{}_{\left(M\right)}^{\left(L_i\right)}\chi^{2}_{\ell} = \frac{1}{n_{\nu}} \sum_{\nu=0}^{n_{\nu}-1} {}_{\left(M\right)}^{\left(L_i\right)}\chi^{2}\left(\nu, \ell\right),
\label{eq: xi2l}
\end{eqnarray}
\section{Results}
\label{Section: 7}
\subsection{Influence of mixing matrix dimension on the GMCA and FastICA estimations}
The choice of sky masking, on top of the fixed BINGO survey area, is an important decision which impacts the recovery of the signal and reduces the contamination by different foregrounds. Fig. \ref{fig: FGcls} shows the angular power spectrum of the foreground components chosen for this analysis. Dashed lines correspond to a power spectrum of the full sky and full lines correspond to a partial sky (BINGO survey area with the region around the Galactic plane also removed). The "unmasked" power spectrum show three components (synchrotron, free-free and FRPS) with significantly more power than the other two (AME and CMB).
They overlap each other and the contributions of FRPS and free-free switch positions after masking the sky to recompute the power spectrum. Synchrotron emission clearly dominates the sky in both cases, and masking the sky increases the contribution of FRPS, reducing its difference from the free-free emission, with both contributing almost equally to the total power at $l \le 10$.
\begin{figure*}[h]
\centering
\includegraphics[width=18cm,height=7cm]{images/FGCls.jpg}
\caption{Angular power spectrum masked (solid line) and unmasked (dashed line) within BINGO coverage region for all foreground components assumed in this work: CMB (gray), AME (orange), Free-Free (red), FRPS (green), and Synchrotron (blue).}
\label{fig: FGcls}
\end{figure*}
In contrast to the GNILC algorithm, FastICA and GMCA do not have a native way to measure the foreground contribution on the observation maps, i.e., they do not estimate $N_{s}$. Please note that $N_s$ is the number of foreground components of the simulated sky, as described in sec. \ref{sec:GNILC}, and $n_s$ is the number of non-physical templates assumed to represent the total foreground signals. As in the context of blind foreground removal, it is not known \emph{a priori} which foreground components make up the observations. Therefore, it is necessary to have a way to choose $n_s$. GNILC uses AIC method per needlet band and per pixel. For the GMCA and FastICA, we computed the normalized eigenvalues from the observational covariance matrix, as shown in Fig. \ref{fig:eigenvalues}. The 21 cm + instrumental noise are subdominant in the observation data, since foregrounds are stronger than their brightness temperature by up to six orders of magnitude. As a consequence, the eigenvalues are almost all dominated by the foregrounds. Then, we can see three or four eigenvalues corresponding to almost all information contained in the observations. However, this choice has a subtle issue because the $n_s$ value may be larger or smaller than needed. If the $n_s$ is larger, the foreground removal process can be so aggressive as to extract also the 21 cm signal. If the value is smaller, the foreground removal process would still keep some foreground signal in the data, contaminating the 21 cm + instrumental noise information. It is expected that the impact of such a choice affects the final results even after noise debiasing and an analysis using different $n_s$ is performed.
We used three configurations for FastICA and GMCA, $n_s \in \{2,3,4\}$. All FastICA analysis was done with 400 realizations and in pixel space. The quality of the results was expressed as $\chi^2_{\textrm{eff}}$, corresponding to $\chi^2$ per channel, in Fig. \ref{fig: chi2_eff_ns_ICA}. The variance was estimated by Jackknife, taking different realizations as the reference in the noise debiasing process and generating $\chi^2_{\textrm{eff}}$ for each configuration. But we used the same realization for each $n_s$ plots in Fig. \ref{fig: chi2_eff_ns_ICA}. The black dashed line represents $\chi^2_{\textrm{eff}}=1$. For $n_s=2$, although $\chi^2_{\textrm{eff}}$ of the FastICA configuration is compatible with 1 for all channels, its variance is large enough to be compatible with values larger than 1. Only two non-physical templates to describe the foreground influence lead to an amount of foreground information leakage to 21 cm + instrumental noise, and then the 21 cm angular power spectrum estimated will be larger.
On the other hand, if we take $n_s=4$, the variances of the $\chi^2_{\textrm{eff}}$ are almost all smaller than $n_s=2$ case and all values are compatible at one sigma level with $\chi^2_{\textrm{eff}}=1$. Even for $n_s=4$, the variance values are large, and the third channel is representative of the amplitude of the variance because there are more fluctuations between the realizations assumed and, most likely that those amplitude is due to the more aggressive estimation of the foregrounds, leading to a loss of 21 cm signal. As this addition has a minor eigenvalue, just a subtle change occurs. The red line describes the best results with small variances and almost all values compatible with one, as shown in the below representation in Fig. \ref{fig: chi2_eff_ns_ICA}.
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/ICA_chi2eff_ns.jpg}
\caption{Effective chi-square - that is, chi-square over multipoles per channel - for three different numbers of mixing matrix dimensions through FastICA algorithm: 2 (black), 3 (red), and 4 (blue). The plot below is an enlargement of the plot above in order to focus on red solid line behavior. The error bars were obtained from the Jackknife method over 400 realizations.}
\label{fig: chi2_eff_ns_ICA}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/GMCA_chi2eff_ns.jpg}
\caption{Effective chi-square - that is, chi-square over multipoles per channel - for three different numbers of mixing matrix dimensions through GMCA algorithm: 2 (black), 3 (red), and 4 (blue). The plot below is an enlargement of the plot above in order to focus on red solid line behavior. The error bars were obtained from the Jackknife method over 400 realizations.}
\label{fig: chi2_eff_ns_GMCA}
\end{figure}
The explanation for $n_s=3$ being the best choice is because synchrotron, free-free, and FRPS emissions are the dominant components of emission in the declination range between -24$^o$ and -8$^o$ in celestial coordinates. There is a strong influence not only by synchrotron and free-free but by FRPS too. FRPS importance is shown in Fig. \ref{fig: allvar}, where we have a measurement of the total variance of the temperature fields given by
\begin{eqnarray}
\langle\Delta^2\rangle(\ell_{max}) = \int_{2}^{\ell_{max}}d\ln\ell\ \frac{\ell(\ell+1)}{2\pi}C_{\ell},
\label{total_variance_temperature}
\end{eqnarray}
where $\ell_{max}\in\{2, 3,...,300\}$. Basically, the integrated term above is the variance per $\ln\ell$, and the area below that term is the total variance. That figure is for our tenth channel. It is possible to see three principle foreground components in our sky region, which explains why we need three non-physical templates to describe the foreground contribution to the observational data.
Fig. \ref{fig: allvar} shows that the total variance of the 21cm signal (+ noise) exceeds the total AME variance when integrating multipoles up to an $\ell_{max}> 20$ and the total CMB variance when integrating multipoles up to an $\ell_{max}> 300$. That is crucial information because that range is where happen better estimation of the 21 cm angular power spectrum. Below that range, there is a higher variance of the angular power spectrum estimated, and above one, the estimation is worst with a smaller variance, as shown in Fig. \ref{fig: chi2l_ns}. $\chi^2_{\ell}$ is similar to Eq. \ref{eq: xi2eff} but over channels and not multipoles, as described in Eq. \ref{eq: xi2l}.
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/allVars.jpg}
\caption{The total variance of the temperature field (Eq. \ref{total_variance_temperature}) of each foreground component and of the components that are not modeled by the algorithms. Correspond to the foreground synchrotron (blue), Free-Free (orange), FRPS (green), AME (red), and CMB (violet). The total variance of the sum of 21cm signal and white noise is shown in dashed brown. Solid lines represent the component with higher contributions to the signal and the dashed one the lower contributions.}
\label{fig: allvar}
\end{figure}
Similar results were obtained with GMCA but in sparse decomposition in the starlet transform. The results for GMCA are shown in Fig. \ref{fig: chi2_eff_ns_GMCA}, and the $\chi^2_{\textrm{overall}}$ values are showed in Table \ref{tab: GMCAxICAchi2overall_ns}. In the table, it is possible to see that FastICA and GMCA are statistically equal when no other parameters are changed.
\begin{center}
\begin{table}
\centering \begin{tabular}{ |c|c|c| }
\hline
\multicolumn{1}{|c|}{$n_s$}&\multicolumn{1}{|c|}{FastICA}&\multicolumn{1}{|c|}{GMCA} \\ \hline
2 & 1.8(4) & 1.8(4)\\
3 & 1.02(4)& 1.02(4)\\
4 & 1.1(2) & 1.1(2)\\
\hline
\end{tabular}
\caption{$\chi^{2}_{\textrm{overall}}$ for different number of mixing matrix dimension assuming 400 realizations.}
\label{tab: GMCAxICAchi2overall_ns}
\end{table}
\end{center}
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/chi2l_ns.jpg}
\caption{$\chi^2_{\ell}$ is the $\chi^2(\nu,\ell)$ mean over the channels. There are three different plots - displayed in three windows- each one with a certain number of non-physical templates (ns) to reconstruct the foreground contribution with FastICA and then debias the residuals for 400 realizations. Each window (row) represents a range of the multipole: 0-20, 21-300, and 301-400. The region within 20-300 corresponds to the better results for reconstruction; at the same time, it is the region where the 21cm + white noise (prior) is higher than AME and smaller than CMB emissions.}
\label{fig: chi2l_ns}
\end{figure}
\subsection{Influence of the number of realizations on the estimations}
Employing FastICA and GMCA with $n_s=3$, but changing the number of simulations, we would inquired about the dependence of which algorithm of $N_r$. It is essential because in this period when BINGO pipelines are been built it is necessary to run 21 cm estimation several times. But to know how many simulations are sufficient to get an accurate estimate makes analysis processes faster.
In Fig. \ref{fig: chi2_eff_nr_CS}, we see $\chi^2_{\textrm{eff}}$ for five different numbers of simulations for the three methods. Regarding FastICA and GMCA values, it is possible to verify that only for 400 realizations both algorithms have a statistically similar shape. FastICA seems to converge faster to the shape of the $\chi^2_{\textrm{eff}}$ with 400 realizations than GMCA. The results become clearer in Table \ref{tab: chi2overall_Ns}, where it is shown that only 25 realizations for all results of the algorithms are not compatible with one. All other numbers are statistically compatible. GMCA and FastICA only have the same shape with 400 realizations and $n_s=3$. It seems for a higher number of realizations on the estimation, in our configuration, both algorithms converge to the same shape and as the same $\chi_ {\textrm{overall}}^2$.
\begin{figure*}[t]
\centering
\includegraphics[width=9cm]{images/ICA_chi2eff_ns3_nr.jpg}
\includegraphics[width=9cm]{images/GMCA_chi2eff_ns3_nr.jpg}\\
\includegraphics[width=9cm]{images/GNILC_chi2eff_nr.jpg}
\caption{Effective $\chi^2$ per channel for FastICA (top left), GMCA (top right), and GNILC (bottom) for five number of simulations used: 25 (cyan), 50 (orange), 100 (red), 200 (blue), and 400 (black).}
\label{fig: chi2_eff_nr_CS}
\end{figure*}
\begin{center}
\begin{table}
\centering \begin{tabular}{ |c|c|c|c| }
\hline
\multicolumn{1}{|c|}{$n_r$ }&\multicolumn{1}{|c|}{GMCA}&\multicolumn{1}{|c|}{FastICA}&\multicolumn{1}{|c|}{GNILC} \\ \hline
400 & 1.02(4) & 1.02(4) & 1.04(4) \\
200 & 1.04(4) & 1.01(4) & 1.04(4) \\
100 & 1.03(3) & 1.01(4) & 1.03(3) \\
50 & 1.01(3) & 0.99(3) & 1.01(3) \\
25 & 0.97(4) & 0.95(3) & 0.97(4) \\
\hline
\end{tabular}
\caption{$\chi^{2}_{\textrm{overall}}$ for different numbers of simulations assuming GMCA and FastICA with dimension}
\label{tab: chi2overall_Ns}
\end{table}
\end{center}
\subsection{Comparison between GNILC, GMCA and FastICA algorithms}
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/CS_chi2_other.jpg}
\caption{The plots are the $\chi^2$ (Eq. \ref{eq: xi2}) heat map of the three algorithms. The vertical axis corresponds to the channel, and the horizontal one to the reconstructed HI angular power spectrum corresponding to a specific multipole. The redder values concerning HI angular power spectrum are harder to estimate.}
\label{fig: heatmapXi2}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{images/Cls_problem2.jpg}\\
\includegraphics[width=9cm]{images/Cls_problem1.jpg}
\caption{Plots using GMCA with $n_r$ = 3 and 400 realizations. The top plot is for 5th channel ($\nu$ = 4), and the bottom one for 13th channel ($\nu$ = 12). On the left plots are multipoles between 2-300, and on the right, between 30-100. The reconstructed HI angular power spectrum and the original one are represented by blue and red colors, respectively. The green color represents the white noise angular power spectrum. The vertical black dashed line corresponds to the multipole 47 (top plots) and 181 (bottom plots).}
\label{fig: problems}
\end{figure}
The estimation of the 21 cm angular power spectrum for the first, tenth, twentieth, and thirtieth channels can be seen in Appendix \ref{apendice: 1} for all algorithms (blue color), with respective angular power spectrum from fiducial maps (red color), and white noise angular power spectrum (green color). Here, we will concentrate on quantifying the results. Analyzing the $\chi^2$ for 400 simulations to estimate the values, in Fig. \ref{fig: heatmapXi2} there is a heatmap representation where the redder color represents more difficulty to estimate the angular power spectrum value. There are some regions where there are highlighted redder values, as around $(\nu,\ell) = (4, 47)$ and $(12, 181)$, represented on the top and bottom in Fig. \ref{fig: problems}, respectively. These plots explicitly show the effect of the lower signal-noise ratio in view of the short observation in time (one year). The top plot has the right representation where the estimated signal follows up the noise bump. The bottom plot shows the same effect, but for the signal-noise relation even lower. Another crucial point to highlight is that only cosmic variance contributes to noise here, leading to further difficulty in obtaining a better $\chi^2$ value.
In the Fig. \ref{fig: chi2_eff_CS_comparison} the very similar result for GMCA and FastICA is clear. At the bottom of the figure, we show the percentage ratio between the latest algorithms, where there is no statistical difference between them. But if we see Fig. \ref{fig: GMCAxFastICA_FG}, it shows the estimation of the ninth channel and shows that the algorithms do not work equally to estimate the foreground contribution. In that figure, GMCA and FastICA were run with three non-physical templates for estimating the foreground. We compared the input foreground maps with the estimated one for each algorithm (upper plots) and their respective residuals (central plots). The bottom right plot in that figure represents the difference between GMCA and FastICA estimation. It is possible to see that there are big differences where the foreground emission is higher. FastICA foreground estimation is slightly higher in the Galactic region. Fig. \ref{fig: GMCAxFastICA_Residuals} shows the residuals for each algorithm, and it is possible to see a slight difference between them.
\begin{figure}[h]
\centering
\includegraphics[width=9cm,height=6cm]{images/CS_chi2eff_diff_model3.jpg}
\caption{At the top, there is a comparison of effective $\chi^2$ of the GMCA (solid blue line), FastICA (dashed orange line), and GNILC (solid gray line). At the bottom, we represent the residual difference between the effective $\chi^2$ of FastICA and GMCA.}
\label{fig: chi2_eff_CS_comparison}
\end{figure}
\begin{center}
\begin{table}
\centering \begin{tabular}{ |c|c|c|c| }
\hline
\multicolumn{1}{|c|}{ }&\multicolumn{1}{|c|}{GMCA}&\multicolumn{1}{|c|}{FastICA}&\multicolumn{1}{|c|}{GNILC} \\ \hline
GMCA & 0 &0.00(5) & -0.02(6)\\
FastICA & 0.00(5)& 0 & -0.02(6)\\
GNILC & 0.02(6)& 0.02(6)& 0\\
\hline
\end{tabular}
\caption{Measuring the difference of Akaike Information Criterium (AIC) from two methods ($\Delta \textrm{AIC}$) is here equivalent to measuring the difference of their overall chi-square. $\Delta \textrm{AIC}>0$ is evidence in favor of the higher chi-square. Comparing the three methods two-by-two and obtaining their errors using Jackknife technique results in all compatibility with 0.}
\label{tab: chi2overall_Ns}
\end{table}
\end{center}
Again, there is no difference between the studied algorithms in our configuration. They are statistically compatible. What is better expressed by the Table \ref{tab: chi2overall_Ns}, that is, the difference between AIC of the algorithms for the last assumed configuration. The values in the table are
\begin{eqnarray}
\Delta\textrm{AIC}_{jk} = {}_{\left(M_j\right)}^{\left(L_i\right)}\chi^{2}_{\textrm{overall}} - {}_{\left(M_k\right)}^{\left(L_i\right)}\chi^{2}_{\textrm{overall}},
\label{Eq: Delta_AIC}
\end{eqnarray}
where $j$ is the result from $M_j$ algorithm and represents the algorithm in the row, $k$ is the same but for the column. $L_i$ is a specific realization chosen. $\Delta\textrm{AIC}_{jk}>0$, which can be represented as "evidence in favor" of the algorithm $M_j$, compared to the $M_k$ algorithm. The table presents no evidence in favor of any of the algorithms
We used a \emph{Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz} and in relation to computational costs, FastICA is the most efficient among them taking less than one minute per run. While, the GMCA and GNILC take between
6-10 minutes.
\subsection{Influence of binned multipoles}
Another important information is how binning of the multipoles affects the estimation. Binning helps compressing data and turns all processes faster. However we have an accurate description of how our choice behavior is essential. Binning here is when we take the weighted average of the angular power spectrum per range of the multipoles,
\begin{eqnarray}
\hat{C}_{\Delta\ell_i} = \frac{\sum_{\ell \in {\Delta\ell}_i}\left(2\ell + 1\right)\hat{C}_{\ell}}{\sum_{\ell \in {\Delta\ell_i}}\left(2\ell + 1\right)},
\end{eqnarray}
where $\Delta\ell_i$ is $i$th multipole bin that can be of different widths.
We calculated binning effect on the estimation using FastICA with 400 simulations and $n_s=3$. Each bin size, minimum and maximum multipoles, number of the multipoles used, and respective $\chi^2_{\textrm{overall}}$ are in Table \ref{tab: chi2overall_binned}.
Fig. \ref{fig: del_l_ICA} represents (top left) the first one hundred angular power spectrum, and (top right) between multipoles 100 and 300, to estimate values with all multipoles and for different $\Delta\ell$. At the bottom, are evaluation of the $\chi^2_{\textrm{eff}}$ and their relative difference concerning $\Delta\ell=1$. It is possible to see that there are high shifts of the values for $\Delta\ell=12$ mainly for lower and higher channels. Examining $\chi^2_{\textrm{overall}}$ it is possible to identify inconsistency of this binning with in one to two sigmas. By Table. \ref{tab: Delta_AIC_binned}, we can see at one sigma level the evidence to $\Delta\ell = 1$ and $2$ comparing to $12$. Among 1, 2, and 4, there is no statistical evidence in favor of any choice. Therefore, between $\Delta\ell=1$ and $4$, it has 765 and 191 multipoles - that is, the last is four times smaller - there is no statistical evidence in favor of using each other and $\Delta\ell=4$ is compatible with $\chi^2_{\textrm{overall}}=1$ up two sigmas.
\begin{center}
\begin{table}
\centering \begin{tabular}{ |c|c|c|c|c| }
\hline
\multicolumn{1}{|c|}{$\Delta_{\ell}$ }&\multicolumn{1}{|c|}{$\ell_{\textrm{min}}$}&\multicolumn{1}{|c|}{$\ell_{\textrm{max}}$}&\multicolumn{1}{|c|}{$n_{\ell}$}&\multicolumn{1}{|c|}{$\chi^{2}_{\textrm{overall}}$} \\ \hline
1& 2 & 766 & 765 & 1.02(4) \\
2& 2 & 764 & 382 & 1.05(5) \\
4& 2 & 762 & 191 & 1.07(6) \\
12& 2 & 746 & 63 & 1.19(8) \\
\hline
\end{tabular}
\caption{Overall chi-square for different lengths of multipole binning through FastICA with 400 realizations and dimension of mixing matrix of 3. }
\label{tab: chi2overall_binned}
\end{table}
\end{center}
\begin{center}
\begin{table}
\centering \begin{tabular}{ |c|c|c|c|c| }
\hline
\multicolumn{1}{|c|}{$\Delta \ell$}&\multicolumn{1}{|c|}{1}&\multicolumn{1}{|c|}{2}&\multicolumn{1}{|c|}{4}&\multicolumn{1}{|c|}{12} \\ \hline
1 & 0 & -0.03(6)& -0.05(7) & -0.17(9)\\
2 & 0.03(6)& 0 & -0.03(7)& -0.15(9)\\
4 & 0.05(7)& 0.03(7)& 0 & -0.1(1)\\
12 & 0.17(9)& 0.15(9)& 0.1(1) & 0 \\
\hline
\end{tabular}
\caption{Difference of AIC ($\Delta \textrm{AIC}$) of different multipole binning. Each table value resulted from difference of both row and column overall chi-square multipole binning, with error estimated from Jackknife method between different realizations. }
\label{tab: Delta_AIC_binned}
\end{table}
\end{center}
\section{Discussions and Conclusions}
\label{Section: 8}
Our original aim was to build a module of foreground removal and 21 cm estimation for BINGO pipeline that could be user-friendly and flexible enough to explore different methods and mathematical domains, as well as connectable to other modules of the BINGO data analysis pipelines. The next step was to test different module configurations to compare the previous situations and frameworks. This paper results from this second step and the first tests, comparing blind algorithms already implemented. So far, the works on 21 cm post-reionization in the BINGO configuration were published using only GNILC. This work is the first to add new foreground removal methods and compare their performance in the current status of the pipeline. It is a natural result of the BINGO pipeline construction. We investigated three blind foreground removal algorithms applied to BINGO Telescope configuration. It is essential to keep in mind that our analysis did not assume effects from instrumental noise different from white noise. Effects due to the construction of the mirrors lead to changes on the main beam and its side lobes and dependence of the beam with frequency and feed horn location. All of this additional information will be done elsewhere.
In the literature, there are reports of similar behavior between GMCA and ICA-based algorithms \citep{bobin2007,cunnington2021}, and the necessity of a more accurate measure to identify a distinction between them. Here, we could see that both algorithms converge to the equivalent results when which used optimal configuration, but there is a visible difference between them. A possibility for future work is to analyze if the property explored by each algorithm (statistical independence and sparsity) are relevant to the point that can help us identifying foreground contribution. The covered region and the bandwidth assumed can also be influenced in that case. Otherwise, introducing other realistic instrumental effects as correlated band noise, side lobes, realistic non-symmetric main beam, different beams per feed horn, and frequency leads to a more subtle problem beyond our scope. Introducing these effects will change how each algorithm and in which assumed space one is working. Different shapes of the beams per channel and focal plane position are better treated in harmonic space (where GNILC works) than pixel space (FastICA). Assuming only thermal noise as instrumental noise and one year of BINGO coverage, we can get the first study to perform and check possible initial difficulties, and check exactly how each algorithm works in a first and preliminary situation. In one year of the coverage, the 21 cm signal collected cannot be enough to have a good statistics, mainly because the signal-to-noise ratio is still low. This could be seen in Fig. \ref{fig: heatmapXi2} and in Fig. \ref{fig: problems}, where a higher $\chi^2$ value is associated with a higher value of the thermal noise signal even for multipole not really high.
Upon checking $\chi^{2}_{\textrm{eff}}$ for GMCA and FastICA, we learn that the mean value per channel is better described assuming three non-physical templates (Table \ref{tab: GMCAxICAchi2overall_ns}). What is explained by there are three main foreground components stronger in the BINGO region. Now, for different simulations, it is possible to see that GMCA and FastICA converge to the equivalence values for 400 simulations with $n_s = 3$. That is, possibly both converge to the equivalent values in their optimal case. Finally, comparing three algorithms no significant difference in this pipeline step shows up. Comparisons between all algorithms with $\Delta \textrm{AIC}$ are compatible with zero. That is, there is no any preference for any algorithm.
Again, in this step of the pipeline, it is essential to do many simulations and tests, and for that use, a faster algorithm saves time. Therefore, FastICA can be justified by its easier and faster work without losing information. Each run with FastICA takes less than one minute in a \textsc{Intel(R) Xeon(R) CPU E5-2640 v4 \@ 2.40GHz} in 64 bits. While the GMCA and GNILC take between 6-10 minutes. Also, binning the $C_{\ell}$ with $\Delta\ell = 2$ reduces the number of multipoles to half and maintains statistical quality.
Lastly, the results got in this step of the pipeline help us better understanding the influence of thermal noise for one year of coverage and compare the three blind algorithms to estimate foreground contribution to the coverage sky. All algorithms are statistically equivalent in this step, what we can explore to use the fastest one. However, we should keep in mind that for more complex foreground models and instrument models, one of the algorithms may perform better than another one. Especially, if there are significant variations of the foreground properties inside the BINGO region, a constant global $n_s=3$ value over all pixels like what is assumed by FastICA may lead to poorer results. Also, using a set of angular power spectrum binned with $\Delta \ell = 2$ reduces to half the number of multipoles values and maintains the statistical information.
\begin{figure*}[t]
\centering
\includegraphics[width=18.4cm, height=5.6cm]{images/cls_binned_ICAbin11.jpg}\\
\centering \includegraphics[width=10cm, height=8cm]{images/ICA_del_l_chi2eff_error_2.jpg}
\caption{The upper left and right plots correspond to the reconstructed HI angular power spectrum with and without using the function for compressing data. In solid black color is the reconstructed HI without using the process of binning. In the dotted orange line is the HI angular power spectrum for the binning length equal to one (which should be the same as the previous one). The other lines correspond to the binning lengths 2 (blue), 4 (purple), and 12 (red). The bottom center plots are the $\chi^2_{\textrm{eff}}$ for HI angular power spectrum through the function for compressing data and the residual difference with the length equal to one.}
\label{fig: del_l_ICA}
\end{figure*}
\begin{acknowledgements}
The BINGO project is supported by São Paulo Research Foundation (FAPESP) grant 2014/07885-0 and by Fapesq-PB. A.M. acknowledges and thank the support from Brazilian agency CNPq for the financial support, Isabella Carucci for making your GMCA codes available to us, and Lucas Olivari for his firsts ideas and suggestions. K.S.F.F. thanks São Paulo Research Foundation (FAPESP) for financial support through grant 2017/21570-0. L.S. is supported by the National Key R\&D Program of China (2020YFC2201600) and NSFC grant 12150610459.
C.A.W. acknowledges CNPq grants 313597/2014-6 and 407446/2021-4. A.R.Q. acknowledges the support of Fapesq-PB for the financial support. M.R. would like to thank the Spanish Agencia Estatal de Investigación (AEI, MICIU) for the financial support provided under the project with reference PID2019-110610RB- C21, L.B. acknowledges Ministerio de Ciência e Tecnologia - Brasil for financial support,
C.P.N. thanks São Paulo Research Foundation (FAPESP) for financial support through grant 2019/06040-0.
\end{acknowledgements}
\begin{appendix}
\section{Estimated HI angular power spectrum}
The accuracy of the HI reconstruction strongly depends on different things: number and type of foregrounds, type of instrumental noises, beam shape, map resolutions, sidelobe effects, polarization leakage, channel, signal-to-noise ratio, the algorithm used to estimate foregrounds, number of simulations in noise debias process, among others.
In this work, we have assumed all maps at equal resolution (40 arcmins) with a Gaussian beam, five foreground sources (section), only white noise as instrumental noise, no polarization leakage and sidelobe effects, three different algorithms, and up to 400 simulations.
We represent the reconstruction of each algorithm with 400 realizations for four different channels (first, 10th, 20th, and 30th channel) in Fig. \ref{fig: recGMCA} (GMCA), \ref{fig: recGNILC} (GNILC), and \ref{fig: recFastICA} (FastICA). GMCA and FastICA using $n_r=3$
\label{apendice: 1}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=9cm, height=5cm]{images/GMCA_bin0.jpg}
\includegraphics[width=9cm, height=5cm]{images/GMCA_bin9.jpg}\\
\includegraphics[width=9cm, height=5cm]{images/GMCA_bin19.jpg}
\includegraphics[width=9cm, height=5cm]{images/GMCA_bin29.jpg}
\end{center}
\caption{
The graphs are the reconstructed HI angular power spectrum for the first (top left), 10th (top right), 20th (bottom left), and 30th (bottom right) channels using GMCA with 400 realizations and $n_s$=3. The blue color line represents the HI information reconstructed, the original information in red color, and the white noise information in green color. Below each graph is the residual difference between reconstructed and original HI information.}
\label{fig: recGMCA}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=9cm, height=5cm]{images/GNILC_bin0.jpg}
\includegraphics[width=9cm, height=5cm]{images/GNILC_bin9.jpg}\\
\includegraphics[width=9cm, height=5cm]{images/GNILC_bin19.jpg}
\includegraphics[width=9cm, height=5cm]{images/GNILC_bin29.jpg}
\end{center}
\caption{
The graphs are the reconstructed HI angular power spectrum for the first (top left), 10th (top right), 20th (bottom left), and 30th (bottom right) channels using GNILC with 400 realizations. The blue color line represents the HI information reconstructed, the original information in red color, and the white noise information in green color. Below each graph is the residual difference between reconstructed and original HI information.}
\label{fig: recGNILC}
\end{figure*}
\begin{figure*}[t]
\begin{center}
\includegraphics[width=9cm, height=5cm]{images/ICA_bin0.jpg}
\includegraphics[width=9cm, height=5cm]{images/ICA_bin9.jpg}\\
\includegraphics[width=9cm, height=5cm]{images/ICA_bin19.jpg}
\includegraphics[width=9cm, height=5cm]{images/ICA_bin29.jpg}
\end{center}
\caption{
The graphs are the reconstructed HI angular power spectrum for the first (top left), 10th (top right), 20th (bottom left), and 30th (bottom right) channels using FastICA with 400 realizations and $n_s$=3. The blue color line represents the HI information reconstructed, the original information in red color, and the white noise information in green color. Below each graph is the residual difference between reconstructed and original HI information.}
\label{fig: recFastICA}
\end{figure*}
\section{Comparison of GMCA and FastICA maps}
Similar results obtained by GMCA and ICA algorithm have been published before. The exact reason behind the similar results is not clear yet. This similarity explains the indifference in starlet or real space results in this work. Although we have shown the statistical equivalence between the algorithms, their results for just one realization are not the same.
In Fig \ref{fig: GMCAxFastICA_FG}, we represent the reconstructed foreground emission by algorithms e their comparisons. The reconstructions are worse at the edge of the mask and in the Galaxy region. The unmasked Galaxy part represents a poor estimate by both algorithms. When we look at the reconstructions one by one, it looks the same, but when we look at the difference between them (bottom-right image), we can see a subtle difference in the Galaxy region.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{images/FGrec_ch9_cartview_withoutnames.jpg}
\caption{In the first row, there are the foreground emissions reconstructed by (top left) FastICA and (top center) GMCA and the (top right) original foreground emissions. In the plots below, we compare the quality of reconstruction, and the difference between algorithm reconstruction: (bottom left) original subtracted by GMCA reconstructed, (bottom center) original subtracted by FastICA reconstructed, and GMCA minus the FastICA result.}
\label{fig: GMCAxFastICA_FG}
\end{figure*}
Fig. \ref{fig: GMCAxFastICA_Residuals} shows the residual of the algorithms. When we compare the residual from both algorithms (on the right), we can see that in the Galaxy region, there are negative values. These are the same regions with positive difference in Fig \ref{fig: GMCAxFastICA_FG}. Therefore, the GMCA is subtly underestimating emissions from the Galaxy region.
\begin{figure*}[t]
\centering
\includegraphics[width=18cm]{images/Residual_ch9_cartview.jpg}
\caption{The plots represent the algorithm residuals, (left) FastICA and (right) GMCA, and the (bottom center) difference between both residuals. All maps are of channel 9.}
\label{fig: GMCAxFastICA_Residuals}
\end{figure*}
\end{appendix}
\bibliographystyle{aa}
|
\section{Introduction}
Resonance energy transfer (RET) is a fundamental process in photophysics and has attracted considerable attention in a variety of fields owing to its extensive applications in biological and chemical sensing \cite{Wu2020,Chen2013,Chang1995,Zadran2012,Kaminska2021,Shi2015,Xia2009,Hildebrandt2017,Bhuckory2019}, molecular imagining~\cite{Li2020,Sasmal2016,Ha2001,Zhou2019,Choi2009}, and photovoltaics~\cite{Scully2007,Liu2005,Shankar2009,Li2015,Heidel2007}.
To understand the mechanism of RET, numerous theoretical studies have been conducted based on the electric-dipole approximation (EDA) since 1940s~\cite{Dexter1953,Skourtis2016,Murphy2004,Olaya-Castro2011,Duque2015,Weeraddana2016,Salam2018,Salam2022,Andrews2022}.
For example, F\"orster developed a concept of spectral overlap for the description of RET~\cite{Forster1948,Forster1949}, Andrew established a unified theory of radiative and nonradiative RET~\cite{Andrews1989,Andrews2004,Daniels2003} (which corresponds to the resonant dipole-dipole interaction in free space) from molecular quantum electrodynamics~\cite{Craig1998}, and Welsch \textit{et al.} incorporated the effect of dielectric environment into the resonant dipole-dipole interaction (RDDI)~\cite{dung2002,dung2002intermolecular}.
These theories not only advance the understanding of RET but also successfully capture the main features of RET.
Recently, experimental studies have shown that the RET rates between two entities (molecules~\cite{Komarala2008,L.-Viger2011,Georgiou2021,Andrew2004,Hichri2021}, semiconductors~\cite{Li2015}, biomolecules~\cite{Zhang2007,Stobiecka2015}, etc.) can be significantly influenced by the presence of polaritons, which provides a new perspective on exploring light-matter interaction in spatially
dependent electric fields. It is well-known that the EDA cannot be used to describe RET in the following scenarios, (i) the distance between two entities is insufficient~\cite{Munoz-Losa2009,Andrews2011,Beljonne2009,Zheng2014,Andrews2009}, and (ii) the surrounding electromagnetic field of entities varies drastically~\cite{Zhang2020,NasiriAvanaki2018}.
In the former scenario, several approaches have been successively proposed and applied to some representative systems, e.g., transition monopole theory for chlorophylls~\cite{Chang1999}, line-dipole approximation for conjugated polymers~\cite{Beenken2004}, and transition density cube (TDC) method for pigments of light-harvesting complex~\cite{Krueger1998}, but these approaches cannot describe the retardation effect, i.e., the mechanism of radiative RET.
In the later scenario, a straightforward improvement is to consider the quadrupolar interaction~\cite{Scholes1997,NasiriAvanaki2018,Salam2005}, but the convergence of multipolar expansion depends on material structures.
Moreover, the methods used to address the scenarios (i) and (ii) cannot describe the influence of polaritons (photons dressed by dielectric environments) on RET. Therefore, to address the above issues, our strategy is to start from the framework of macroscopic quantum electrodynamics \cite{Vogel2006,Buhmann2012} (which enables us to incorporate the effect of dielectric environments) and derive an explicit RET-rate expression by employing a transition current density approach (which enables us to capture the retardation effects).
In this article, the main goal is to establish a generalized theory of RET beyond the EDA and allow us to study RET between two entities with material structures in spatially dependent vacuum electric fields.
The structure of this article is organized as follows.
In Sec.~\ref{sect:theory}, we begin from the total Hamiltonian of polaritons and point charges with interactions introduced via the minimal coupling procedure.
Next, we derive the RET rate of a pair of molecules expressed by the molecular transition current density by expanding Born series to the second order.
Moreover, we adopt the Condon-like approximation in order to separate the electronic and nuclear degrees of freedom, and then derive a formula which allows us to acquire the transition current density via \textit{ab initio} calculations.
In Sec.~\ref{sect:app}, we demonstrate the generality of our theory by comparing them to previous works, including the TDC method~\cite{Krueger1998} and plasmon-coupled RET \cite{Hsu2017}.
In the last section, we present a brief summary of this research.
\section{Theory}
\label{sect:theory}
\subsection{Total Hamiltonian}
In the non-relativistic regime, we consider a collection of point charges with the quantized electromagnetic fields in the presence of linear, dispersive, and absorbing media (polaritons).
Hence, the total Hamiltonian comprises a polariton Hamiltonian and a Hamiltonian of point charges (pc),
$\hat{H}_\mathrm{tot}=\hat{H}_\mathrm{pol}+\hat{H}_\mathrm{pc}$.
To properly describe the quantum behavior of polaritons, we adopt the quantization framework of macroscopic quantum electrodynamics~\cite{dung2002intermolecular,Vogel2006,Buhmann2012} and express the polariton Hamiltonian as
\begin{align}
\hat{H}_\mathrm{pol} = \int \dd[3]{\vb{r}} \int_0^\infty \dd{\omega}\hbar\omega \hat{\mathbf{f}}^\dagger(\mathbf{r},\omega)\cdot\hat{\mathbf{f}}(\mathbf{r},\omega),
\end{align}
where $\hat{\mathbf{f}}(\mathbf{r},\omega)$ is a vector bosonic operator that obeys the following commutation relations,
\begin{subequations}
\begin{align}
\left[\hat{f}_k^{\mathstrut{}}(\mathbf{r},\omega),\hat{f}_{k'}^\dagger(\mathbf{r'},\omega')\right] &= \delta_{kk'}\delta(\mathbf{r-r'})\delta(\omega-\omega'),\\
\left[\hat{f}_k(\mathbf{r},\omega),\hat{f}_k(\mathbf{r'},\omega')\right] &= 0.
\end{align}
\label{Eq:commutation}%
\end{subequations}
For the Hamiltonian of point charges, the interaction between point charges and quantized electromagnetic fields is introduced through the minimal coupling procedure in the Coulomb gauge,
\begin{align}
\nonumber
\hat{H}_\mathrm{pc} &=
\sum_{n} \frac{1}{2m_n}\left[\vphantom{\frac{1}{2m_n}}\hat{\mathbf{p}}_n - q_n \hat{\mathbf{A}}(\hat{\mathbf{r}}_n) \right]^2
+\sum_{n<m} \hat{V}(\hat{\vb{r}}_n,\hat{\vb{r}}_m)\\
&~+\sum_{n} q_n\hat{\varphi}(\hat{\vb{r}}_n).
\end{align}
The first term in $\hat{H}_\mathrm{pc}$ describes the kinetic energy of point charges, where $m_n$, $q_n$, $\hat{\vb{r}}_n$, and $\hat{\vb{p}}_n$ are the mass, charge number, position operator, and canonical momentum operator of the $n$-th point charge, respectively.
The second term, $\hat{V}(\hat{\vb{r}}_n,\hat{\vb{r}}_m)$, describes the Coulomb interaction between the $n$-th and $m$-th point charges.
The final term is the interaction between point charges and scalar potential from media.
In the Coulomb gauge, the scalar potential operator $\hat{\varphi}(\vb{r})$ is associated with the transverse auxiliary electric-field operator $\hat{\mathbf{E}}^\perp (\mathbf{r},\omega)$ [Eq.~(\ref{Eq:defgradphi})].
Similarly, the vector potential operator $\hat{\vb{A}}(\vb{r})$ is associated with the longitudinal auxiliary electric-field operator $\hat{\mathbf{E}}^\parallel (\mathbf{r},\omega)$ [Eq.~(\ref{Eq:defA})].
\begin{subequations}
\begin{align}
\label{Eq:defgradphi}
-\nabla \hat{\varphi}(\vb{r}) &= \int_0^\infty \dd{\omega}
\left[\hat{\mathbf{E}}^\parallel (\mathbf{r},\omega) + \mathrm{H.c.}\right]
,\\
\label{Eq:defA}
\hat{\mathbf{A}}(\mathbf{r}) &= \int_0^\infty \dd{\omega}
\left[(i\omega)^{-1}\hat{\mathbf{E}}^\perp(\mathbf{r},\omega) + \mathrm{H.c.}\right].
\end{align}
\label{Eq:FieldsinFeq}%
\end{subequations}
Here, the transverse (longitudinal) component is defined by the auxiliary electric-field operator $\hat{\mathbf{E}} (\mathbf{r},\omega)$,
\begin{align}
\hat{\mathbf{E}}^{\perp(\parallel)}(\mathbf{r},\omega)= \int \dd[3]{\vb{r}'} \tensor{\boldsymbol{\delta}}\vphantom{\delta}^{\perp(\parallel)}(\vb{r-r}')\cdot\hat{\mathbf{E}}(\mathbf{r}',\omega),
\end{align}
where $\tensor{\boldsymbol{\delta}}\vphantom{\delta}^{\parallel}(\vb{r-r}')$ and $\tensor{\boldsymbol{\delta}}\vphantom{\delta}^{\perp}(\vb{r-r}')$ denote the longitudinal and transverse dyadic delta functions, respectively.
Moreover, the auxiliary electric-field operator in the frequency domain is defined by \cite{Dung1998,dung2002intermolecular}
\begin{align}
\nonumber
&\hat{\mathbf{E}}(\mathbf{r},\omega)\\
&=i\sqrt{\frac{\hbar}{\pi\epsilon_0}}\frac{\omega^2}{c^2} \int \dd[3]{\mathbf{r'}} \sqrt{\epsilon_\mathrm{I}(\mathbf{r'},\omega)}~\tensor{\mathbf{G}}(\mathbf{r},\mathbf{r'},\omega)\cdot\hat{\mathbf{f}}(\mathbf{r'},\omega).
\label{Eq:Edef}
\end{align}
Here, $\epsilon_\mathrm{I}(\vb{r}',\omega)=\mathrm{Im}\left[\epsilon(\vb{r}',\omega)\right]$ denotes the imaginary part of the permittivity function, and $\tensor{\mathbf{G}}(\mathbf{r},\mathbf{r'},\omega)$ is the dyadic Green's function of macroscopic Maxwell's equations, i.e.,
\begin{align}
\left[
\frac{\omega^2\epsilon(\vb{r},\omega)}{c^2}-\nabla\times\nabla\times
\right]
\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)=-\tensor{\vb{I}}\delta(\vb{r}-\vb{r}'),
\end{align}
where $\tensor{\vb{I}}$ is a 3-by-3 identity matrix.
On the basis of the total Hamiltonian, we further categorize the point charges according to their belonging entities because the main purpose of this study is to focus on the RET between the entities, where the entities can be atoms, molecules, 2D materials, etc.
Accordingly, the total Hamiltonian is rewritten as
\begin{align}
\nonumber
\hat{H}_\mathrm{pc}&\approx
\sum_M\hat{H}_M + \sum_{M<M'}\hat{V}_{MM'}\\
&~+
\sum_M\sum_{\xi\in M} q_\xi\hat{\varphi}(\hat{\vb{r}}_\xi) -
\sum_M\sum_{\xi\in M} \frac{q_\xi}{m_\xi}\hat{\vb{A}}(\hat{\vb{r}}_\xi)\cdot\hat{\vb{p}}_\xi,
\label{Eq:Hpcapprox}
\end{align}
where $M$ denotes the index of entities and $\xi\in M$ denote the index of point charges (including nuclei and electrons) in the entity $M$.
It is worth reminding that we assume the number density of polaritons during the RET processes is sufficiently low so that the weak-field approximation \cite{Schatz2002} is valid, thus omitting the quadratic term $q_\xi^2\hat{\vb{A}}^2(\hat{\vb{r}}_\xi)/2m_\xi$.
Here, the Hamiltonian $\hat{H}_M$ of the entity $M$ and the Coulomb interaction between two entities $\hat{V}_{MM'}$ are defined by
\begin{align}
\hat{H}_M &=
\sum_{\xi\in M}\frac{\hat{\mathbf{p}}_\xi^2}{2m_\xi}
+\sum_{\xi<\zeta} \hat{V}(\hat{\vb{r}}_\xi,\hat{\vb{r}}_{\zeta}),
\label{Eq:HMdef}\\
\hat{V}_{MM'} &=
\sum_{\xi\in M}\sum_{\zeta\in M'}
\hat{V}(\hat{\vb{r}}_\xi,\hat{\vb{r}}_{\zeta}),
\end{align}
respectively.
In the last term of Eq.~(\ref{Eq:Hpcapprox}), we use the relation, $\hat{\vb{p}}_\xi\cdot\hat{\vb{A}}(\hat{\vb{r}}_\xi)=\hat{\vb{A}}(\hat{\vb{r}}_\xi)\cdot\hat{\vb{p}}_\xi$, which holds as we choose the Coulomb gauge.
In addition, it is convenient to define the charge density operator ${\hat{\rho}}_M(\vb{r})$ and the current density operator ${\hat{\vb{j}}}_M(\vb{r})$ of the entity $M$,
\begin{align}
\label{Eq:Mrhodef}
{\hat{\rho}}_M(\vb{r}) &\equiv \sum_{\xi\in M} q_\xi
\delta(\mathbf{r}-\hat{\mathbf{r}}_\xi),\\
{\hat{\vb{j}}}_M(\vb{r}) &\equiv
\sum_{\xi\in M}\frac{q_\xi}{2m_\xi}\Big[
\delta(\vb{r}-\hat{\vb{r}}_\xi)\hat{\vb{p}}_\xi+
\hat{\vb{p}}_\xi\delta(\vb{r}-\hat{\vb{r}}_\xi)
\Big],
\label{Eq:Mcurrentdef}
\end{align}
so that the interaction Hamiltonian between polaritons $\hat{V}_{\mathrm{pol},M}$ and the entity $M$ can be expressed in a continuous form,
\begin{align}
\nonumber
\hat{V}_{\mathrm{pol},M}&=
\sum_{\xi\in M} \left[q_\xi\hat{\varphi}(\hat{\vb{r}}_\xi) - \frac{q_\xi}{m_\xi}\hat{\vb{A}}(\hat{\vb{r}}_\xi)\cdot\hat{\vb{p}}_\xi\right]\\
&=
\int \dd[3]{\vb{r}}
\left[\hat{\varphi}(\mathbf{r}){\hat{\rho}}_M(\mathbf{r})-
\hat{\mathbf{A}}(\mathbf{r})\cdot{\hat{\vb{j}}}_M(\vb{r})\right].
\label{Eq:Hintdef}
\end{align}
Note that the order of delta functions and canonical momentum operators in Eq.~(\ref{Eq:Mcurrentdef}) should be done with care because $\hat{\vb{r}}_\xi$ and $\hat{\vb{p}}_\xi$ are not commutative. The details can be found in Appendix \ref{sect:deltap}.
Here, we would like to emphasize that the interaction form in Eq.~(\ref{Eq:Hintdef}) is the complete couplings between polaritons and molecules, i.e., the couplings beyond the commonly used EDA.
In other words, the information of molecular structures is preserved in Eq.~(\ref{Eq:Hintdef}) because of the spatial dependence of the charge and current density operators.
Finally, the total Hamiltonian is reorganized as the following compact form,
\begin{align}
\hat{H}_\mathrm{tot}
&\approx\hat{H}_\mathrm{pol} + \sum_M\left[\hat{H}_M +
\hat{V}_{\mathrm{pol},M}\right] + \sum_{M<M'}V_{MM'}.
\label{Eq:GeneralFormH}
\end{align}
Remind that Eq.~(\ref{Eq:GeneralFormH}) is approximate because we omit $\hat{\vb{A}}^2(\vb{r})$.
It is worth pointing out that the above Hamiltonian is a general form for entities coupled to the quantized electromagnetic fields in the presence of dispersive and absorbing media, which can also be applied to investigate other photophysical (photochemical) topics such as spontaneous emission and electron transfer under the influence of polaritons.
\subsection{Transfer Rate of a Two-Entity System}
Now, we focus on the RET rate between a pair of entities with the assistance of polaritons in the incoherent limit.
According to the definition in Eq.~(\ref{Eq:GeneralFormH}), the total Hamiltonian in the two-entity case reads
\begin{align}
\hat{\mathcal{H}}&\equiv\hat{H}_\mathrm{tot}(M=\{A,B\})
= \hat{\mathcal{H}}_0 + \hat{\mathcal{H}}_1,
\end{align}
with the unperturbed Hamiltonian $\hat{\mathcal{H}}_0$ and interaction Hamiltonian $\hat{\mathcal{H}}_1$ described by
\begin{align}
&\hat{\mathcal{H}}_0 = \hat{H}_\mathrm{pol} + \hat{H}_A + \hat{H}_B,\\
&\hat{\mathcal{H}}_1 = \hat{V}_{\mathrm{pol},A} + \hat{V}_{\mathrm{pol},B}
+ \hat{V}_{AB}.
\end{align}
In this system, we consider the initial (final) state, which is the direct product of the polaritonic vacuum state and the energy eigenstates $A$ and $B$,
\begin{subequations}
\begin{align}
\ket{{i}} &=
\ket{a'}\otimes\ket{b}\otimes\ket{\left\{0\right\}},
~~ E_i=E_{a'}+E_b,\\
\ket{{f}} &=
\ket{a}\otimes\ket{b'}\otimes\ket{\left\{0\right\}}.
~~ E_f=E_a+E_{b'},
\end{align}
\end{subequations}
Here, $\ket{a(a')}$ and $\ket{b(b')}$ are the eigenkets of the Hamiltonians $\hat{H}_{A}$ and $\hat{H}_{B}$, and their corresponding energy are $E_{a(a')}$ and $E_{b(b')}$.
Also, we denote the energy of the initial (final) state to $E_i$ ($E_f$), and we require that $E_{a'}>E_a$ and $E_{b'}>E_b$.
By expanding Born series up to the second order in the time-dependent perturbation theory~\cite{CohenTannoudji1998}, the total RET rate $\Gamma$ is expressed as follows,
\begin{align}
\Gamma = \sum_{f,i} P_i \Gamma_{fi},\quad
\Gamma_{fi} = \frac{2\pi}{\hbar}\abs{\bra{{f}}\hat{\mathcal{T}}\ket{{i}}}^2\delta(E_{f}-E_{i}),
\label{Eq:ratedef}
\end{align}
where $P_i$ is the probability of the inital state and the transition operator $\hat{\mathcal{T}}$ is given by
\begin{align}
\hat{\mathcal{T}} &= \hat{\mathcal{T}}_1 + \hat{\mathcal{T}}_2
=
\hat{\mathcal{H}}_1 + \hat{\mathcal{H}}_1\hat{\mathcal{G}_0}\hat{\mathcal{H}}_1.
\end{align}
Here, $\hat{\mathcal{G}}_0$ is the retarded Green's operator, which is defined by
\begin{align}
\hat{\mathcal{G}}_0=\frac{1}{E_{i}-\hat{\mathcal{H}}_0+i\eta},
\quad \eta\rightarrow{0}^+.
\end{align}
Now, we evaluate the total transition amplitude $\bra{{f}}\hat{\mathcal{T}}\ket{{i}}$, and divide $\bra{{f}}\hat{\mathcal{T}}\ket{{i}}$ into two parts, $\matrixel{f}{\hat{\mathcal{T}}_1}{i}$ and $\matrixel{f}{\hat{\mathcal{T}}_2}{i}$.
First, in the transition amplitude of $\mathcal{T}_1$, it is not difficult to obtain that only $\hat{V}_{AB}$ contributes to transition amplitude,
\begin{align}
\nonumber
\bra{{f}}\hat{\mathcal{T}}_1\ket{{i}}=&\bra{{f}}\hat{V}_{AB}\ket{{i}}\\
=&
\bra{a;b'}
\sum_{\xi\in A}
\sum_{\zeta\in B}
\frac{q_\xi q_{\zeta}}
{4\pi\epsilon_0\abs{\hat{\mathbf{r}}_\xi-\hat{\mathbf{r}}_{\zeta}}}
\ket{a';b},
\label{Eq:T1}
\end{align}
where $\epsilon_0$ denotes the vacuum permittivity.
Equation (\ref{Eq:T1}) clearly shows that the first-order perturbation excludes the interplay of molecules and polaritons
due to $\hat{V}_{\mathrm{pol},A}$ $\hat{V}_{\mathrm{pol},B}$ as indirect couplings between two entities.
This is the reason that the second-order perturbation is required in our theory.
Second, the transition amplitude of $\mathcal{T}_2$ contains the following non-zero terms,
\begin{align}
\nonumber
\bra{{f}}\hat{\mathcal{T}}_2\ket{i}
&=\bra{f}\hat{V}_{AB}\hat{\mathcal{G}}_0\hat{V}_{AB}\ket{i}\\
\nonumber
&~+ \bra{f}\hat{V}_{\mathrm{pol},A}\hat{\mathcal{G}}_0 \hat{V}_{\mathrm{pol},B}\ket{{i}}\\
&~+ \bra{f}\hat{V}_{\mathrm{pol},B}\hat{\mathcal{G}}_0
\hat{V}_{\mathrm{pol},A}\ket{{i}}.
\label{Eq:defT2}
\end{align}
It is worth pointing that $\bra{f}\hat{V}_{AB}\hat{\mathcal{G}}_0\hat{V}_{AB}\ket{i}$ is neglectable if one of the following statements is established: (i) the ground- (excited-) state permanent polarization density of entities is weak, (ii) the entities are far apart, resulting in negligible Coulomb interactions.
At the present stage, we neglect the contribution of $\bra{f}\hat{V}_{AB}\hat{\mathcal{G}}_0\hat{V}_{AB}\ket{i}$ and evaluate the second-order transition amplitude by using the spectral representation of the retarded Green's operator,
\begin{widetext}
\begin{align}
\nonumber
\bra{{f}}\hat{\mathcal{T}}_2\ket{i}&\approx\bra{f}\hat{V}_{\mathrm{pol},A}\hat{\mathcal{G}}_0 \hat{V}_{\mathrm{pol},B}\ket{{i}} + \bra{{f}}\hat{V}_{\mathrm{pol},B}
\hat{\mathcal{G}}_0\hat{V}_{\mathrm{pol},A}\ket{{i}}\\
\nonumber
&= \sum_{l=1}^3\int \dd[3]{\vb{s}}\int_0^\infty \dd{\omega} \frac{\bra{a;\left\{0\right\}}\hat{V}_{\mathrm{pol},A}\ket{a';\left\{1_l(\vb{s},\omega)\right\}}
\bra{b';\left\{1_l(\vb{s},\omega)\right\}}\hat{V}_{\mathrm{pol},B}\ket{b;\left\{0\right\}}}{E_{b}-E_{b'}-\hbar\omega+i\eta}\\
&~+ \sum_{l=1}^3\int \dd[3]{\vb{s}}\int_0^\infty \dd{\omega} \frac{\bra{b';\left\{0\right\}}\hat{V}_{\mathrm{pol},B}\ket{b;\left\{1_l(\vb{s},\omega)\right\}}
\bra{a;\left\{1_l(\vb{s},\omega)\right\}}\hat{V}_{\mathrm{pol},A}\ket{a';\left\{0\right\}}}{E_{a'}-E_{a}-\hbar\omega+i\eta}.
\label{Eq:HGH1}
\end{align}
\end{widetext}
\noindent
In $\bra{f}\hat{V}_{\mathrm{pol},A}\hat{\mathcal{G}}_0 \hat{V}_{\mathrm{pol},B}\ket{{i}} + \bra{{f}}\hat{V}_{\mathrm{pol},B}
\hat{\mathcal{G}}_0\hat{V}_{\mathrm{pol},A}\ket{{i}}$, we consider the contribution from the two intermediate states, $\ket{a',b';\left\{1_l(\vb{s},\omega)\right\}}$ and $\ket{a,b;\left\{1_l(\vb{s},\omega)\right\}}$. The former corresponds to a counter-rotating-wave process and the latter corresponds to a rotating-wave process.
Recall that the single-polariton Fock state in Eq.~(\ref{Eq:HGH1}) is defined by $\ket{\left\{1_l(\vb{s},\omega)\right\}}=f_l^\dagger(\vb{s},\omega)\ket{\left\{0\right\}}$,
which is interpreted as single-polariton density with the polarization component $l$ at the frequency $\omega$ and at the position $\vb{s}$.
Furthermore, to adequately cope with $\hat{V}_{\mathrm{pol},M}$, one can define the auxiliary current density of the entity $M$,
\begin{subequations}
\begin{align}
&{\mathcal{J}}_{mm'}(\vb{r};\omega)\equiv
\frac{\omega}{\omega_{m'm}}
{\vb{j}}_{mm'}^\parallel(\vb{r})+
{\vb{j}}_{mm'}^\perp(\vb{r}),\\
&{\mathcal{J}}_{m'm}(\vb{r};\omega)\equiv
\frac{\omega}{\omega_{m'm}}
{\vb{j}}_{m'm}^\parallel(\vb{r})+
{\vb{j}}_{m'm}^\perp(\vb{r}),
\end{align}
\label{Eq:AuxJdef}%
\end{subequations}
where $\omega_{m'm}=\omega_{m'}-\omega_{m}$ and ${\vb{j}}_{mm'(m'm)}^{\parallel(\perp)}(\vb{r})$ as the longitudinal (transverse) part of the transition current density ${\vb{j}}_{mm'(m'm)}(\vb{r})=\matrixel{m(m')}{\hat{\vb{j}}_M(\vb{r})}{m'(m)}$.
For convenience, we restrict the denominator $\omega_{m'm}$ in Eq.~(\ref{Eq:AuxJdef}) to be positive and $\omega_{mm'}$ to be negative.
Therefore, according to the auxiliary transition current density, each element of $\hat{V}_{\mathrm{pol},M}$ in Eq.~(\ref{Eq:HGH1}) can be expressed as
\\
\begin{widetext}
\begin{subequations}
\begin{align}
&\bra{a;\left\{0\right\}}
\hat{V}_{\mathrm{pol},A}
\ket{a';\left\{1_l(\vb{s},\omega)\right\}}=
-\int\dd[3]{\vb{r}}\int_0^\infty \dd{\omega'}
(i\omega')^{-1}
\bra{\left\{0\right\}}\hat{\mathbf{E}}(\mathbf{r},\omega')\ket{\left\{1_l(\vb{s},\omega)\right\}}\cdot{\mathcal{J}}_{aa'}(\vb{r};-\omega'),\\
&\bra{b';\left\{1_l(\vb{s},\omega)\right\}}
\hat{V}_{\mathrm{pol},B}
\ket{b;\left\{0\right\}}=
\int\dd[3]{\vb{r}}\int_0^\infty \dd{\omega'}
(i\omega')^{-1}
\bra{\left\{1_l(\vb{s},\omega)\right\}}\hat{\mathbf{E}}^\dagger(\mathbf{r},\omega')\ket{\left\{0\right\}}\cdot{\mathcal{J}}_{b'b}(\vb{r};-\omega'),\\
&\bra{b';\left\{0\right\}}
\hat{V}_{\mathrm{pol},B}
\ket{b;\left\{1_l(\vb{s},\omega)\right\}}=
-\int\dd[3]{\vb{r}}\int_0^\infty \dd{\omega'}
(i\omega')^{-1}
\bra{\left\{0\right\}}\hat{\mathbf{E}}(\mathbf{r},\omega')\ket{\left\{1_l(\vb{s},\omega)\right\}}\cdot{\mathcal{J}}_{b'b}(\vb{r};\omega'),\\
&\bra{a;\left\{1_l(\vb{s},\omega)\right\}}
\hat{V}_{\mathrm{pol},A}
\ket{a';\left\{0\right\}}=
\int\dd[3]{\vb{r}}\int_0^\infty \dd{\omega'}
(i\omega')^{-1}
\bra{\left\{1_l(\vb{s},\omega)\right\}}\hat{\mathbf{E}}^\dagger(\mathbf{r},\omega')\ket{\left\{0\right\}}\cdot{\mathcal{J}}_{aa'}(\vb{r};\omega'),
\end{align}
\label{Eq:HABint}%
\end{subequations}
where the details can be found in Appendix \ref{sect:coupling}.
Note that the integral variable $\omega'$ differs from the frequency $\omega$ in the single-polariton Fock state.
Incidentally, Eqs.~(\ref{Eq:HABint}a) and (\ref{Eq:HABint}b) correspond to the counter-rotating-wave term while Eqs.~(\ref{Eq:HABint}c) and (\ref{Eq:HABint}d) correspond to the rotating-wave term.
Substituting Eqs.~(\ref{Eq:Edef}) and (\ref{Eq:HABint}) into Eq.~(\ref{Eq:HGH1}) with mathematical operations and using the identity to contract two dyadic Green's functions,
\begin{align}
\mathrm{Im}~\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)
&=\int\dd[3]{\vb{s}}\frac{\omega^2\epsilon_\mathrm{I}(\vb{s},\omega)}{c^2}~
\tensor{\vb{G}}(\vb{r},\vb{s},\omega)\cdot
\tensor{\vb{G}}\vphantom{G}^\dagger(\vb{r}',\vb{s},\omega),
\label{Eq:2Gidentity}
\end{align}
we finally obtain that the transition amplitude of $\mathcal{T}_2$ becomes
\begin{align}
\nonumber
\bra{{f}}\hat{\mathcal{T}}_2\ket{i}
&=
\frac{\hbar}{\pi\epsilon_0c^2}
\int\dd[3]{\vb{r}}\int \dd[3]{\vb{r}'}
\int_0^\infty \dd{\omega}\\
&~\times\left\{
\frac{
{\mathcal{J}}_{aa'}(\vb{r};-\omega)
\cdot\mathrm{Im}~\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)\cdot
{\mathcal{J}}_{b'b}(\vb{r}';-\omega)}
{\hbar(\omega_{b}-\omega_{b'}-\omega)+i\eta}+
\frac{
{\mathcal{J}}_{b'b}(\vb{r};\omega)
\cdot\mathrm{Im}~\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)\cdot
{\mathcal{J}}_{aa'}(\vb{r}';\omega)}
{\hbar(\omega_{a'}-\omega_{a}-\omega)+i\eta}\right\}.
\label{Eq:T2ImG1}
\end{align}
\end{widetext}
To further simplify the transition amplitude in Eq.~(\ref{Eq:T2ImG1}), we next evaluate the $\omega$-integral by contour integration in the complex domain.
We evaluate the $\omega$-integral by extending the interval $[0,\infty)$ to the whole real axis through the transformation $\omega\rightarrow-\omega$ to the first term in Eq.~(\ref{Eq:T2ImG1}) and using the identity $\mathrm{Im}~\tensor{\vb{G}}(\vb{r},\vb{r}',-\omega)=-\mathrm{Im}~\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)$~\cite{dung2002intermolecular,Buhmann2012}.
Next, by exchanging the inner product order of the dyadic Green's function (by Onsager reciprocity~\cite{Buhmann2012}), we extend the $\omega$-integral to the whole real axis,
\begin{align}
\nonumber
\bra{{f}}\hat{\mathcal{T}}_2\ket{i}
&=
\frac{\hbar}{\pi\epsilon_0c^2}
\int\dd[3]{\vb{r}}\int \dd[3]{\vb{r}'}
\int_{-\infty}^\infty \dd{\omega}\\
\label{Eq:T2temp}
&~\times
\frac{
{\mathcal{J}}_{b'b}(\vb{r};\omega)
\cdot\mathrm{Im}~\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)\cdot
{\mathcal{J}}_{aa'}(\vb{r}';\omega)}
{\hbar(\omega_{\mathrm{T}}-\omega)+i\eta\sgn{\omega}},
\end{align}
where $\sgn{z}\equiv z/\abs{z}$ is a sign function.
Here, we assume that the transition frequency of entities $A$ and $B$ is the same, $\omega_{\mathrm{T}}=\omega_{a'}-\omega_{a}=\omega_{b'}-\omega_{b}$, because we focus on the process of ``resonance" energy transfer.
Moreover, because $\mathrm{Im}~\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)$ is not holomorphic, we rewrite Eq.~(\ref{Eq:T2temp}) as
\begin{align}
\bra{{f}}\hat{\mathcal{T}}_2\ket{i}
&=
\frac{\hbar}{\pi\epsilon_0}
\int\dd[3]{\vb{r}}\int \dd[3]{\vb{r}'}
\int_{-\infty}^\infty \dd{\omega}I(\vb{r},\vb{r}',\omega),
\end{align}
where
\begin{align}
\nonumber
&I(\vb{r},\vb{r}',\omega)\\
\nonumber
&=\frac{1}{2ic^2}
\Bigg\{
\frac{
{\mathcal{J}}_{b'b}(\vb{r};\omega)
\cdot
\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)
\cdot
{\mathcal{J}}_{aa'}(\vb{r}';\omega)}
{\hbar(\omega_{\mathrm{T}}-\omega)+i\eta\sgn{\omega}}\\
&\hspace{1 cm}-\frac{
{\mathcal{J}}_{b'b}(\vb{r};\omega)
\cdot
\tensor{\vb{G}}\vphantom{G}^*(\vb{r},\vb{r}',\omega)
\cdot
{\mathcal{J}}_{aa'}(\vb{r}';\omega)}
{\hbar(\omega_{\mathrm{T}}-\omega)+i\eta\sgn{\omega}}
\bigg\}.
\label{Eq:Idef}
\end{align}
By using the fact that the dyadic Green's function $\tensor{\vb{G}}\vphantom{G}(\vb{r},\vb{r}',\omega)$ is holomorphic in the upper complex half plane~\cite{dung2002intermolecular}, we choose the path, as shown in Fig.~\ref{Fig:contour}, and evaluate the $\omega$-integral.
Hence, the $\omega$-integral becomes
\begin{align}
\int_{-\infty}^{\infty} I(\vb{r},\vb{r}',\omega) \dd{\omega}=
\oint_{C} I(\vb{r},\vb{r}',\Omega) \dd{\Omega}-
\int_{C_2} I(\vb{r},\vb{r}',\Omega) \dd{\Omega}.
\label{Eq:T2contour}%
\end{align}
According to the result derived in Appendix \ref{sect:Gamma2}, the contour integral gives the result
\begin{align}
\label{Eq:T2Gamma2}
\left.\bra{{f}}\hat{\mathcal{T}}_2\ket{i}\right|_{C_2}
&=
\frac{1}{\epsilon_0\omega_\mathrm{T}^2}
\int\dd[3]{\vb{r}}
{\vb{j}}_{b'b}^\parallel(\vb{r})\cdot
{\vb{j}}_{aa'}^\parallel(\vb{r}),\\
&=
\bra{a;b'}
\hat{V}_{AB}
\ket{a';b},
\label{Eq:TA_Gamma2}
\end{align}
\begin{figure}[!ht]
\hspace{10 mm
\centering
\includegraphics[width=.33\textwidth]{BS2contour.pdf}
\caption{Illustration of the contour adopted in the $\omega$-integral. The total closed contour $C$ is equal to $C_1+C_2$. A singularity at $\Omega=\omega_\mathrm{T}+i\eta$ is located in the upper complex half plane.}
\label{Fig:contour}%
\end{figure}%
In Eq.~(\ref{Eq:T2Gamma2}), the coupling of longitudinal transition current densities between entities $A$ and $B$ can be further reduced to the Coulomb interaction form (the derivation details can be found in Appendix~\ref{sect:longcurrent}), as a consequence of the Coulomb gauge. For the contour integral of the closed path $C=C_1+C_2$, according to the residue theorem, we can obtain
\begin{align}
\nonumber
\oint_{C} I(\vb{r},\vb{r}',\Omega) \dd{\Omega}
&= 2\pi i~\mathrm{Res}[I(\vb{r},\vb{r}',\Omega),\omega_\mathrm{T}+i\eta/\hbar]\\
\nonumber
&=\frac{\pi}{\hbar c^2}
{\vb{j}}_{b'b}(\vb{r})
\cdot \tensor{\vb{G}}(\vb{r},\vb{r}',\omega_{\mathrm{T}})\cdot
{\vb{j}}_{aa'}(\vb{r}'),
\end{align}
which indicates that
\begin{align}
\nonumber
\left.\bra{{f}}\hat{\mathcal{T}}_2\ket{i}\right|_{\Gamma}&=
\frac{1}{\epsilon_0c^2}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}\\
&~\times
{\vb{j}}_{b'b}(\vb{r})
\cdot \tensor{\vb{G}}(\vb{r},\vb{r}',\omega_{\mathrm{T}})\cdot
{\vb{j}}_{aa'}(\vb{r}').
\label{Eq:T2Gamma}
\end{align}
Recall that the auxiliary current density defined in Eq.~(\ref{Eq:AuxJdef}) gives
${\mathcal{J}}_{b'b}(\vb{r};\omega_{\mathrm{T}})={\vb{j}}_{b'b}(\vb{r})$ and ${\mathcal{J}}_{aa'}(\vb{r}';\omega_{\mathrm{T}})={\vb{j}}_{aa'}(\vb{r}')$.
Finally, according to Eqs.~(\ref{Eq:T1}), (\ref{Eq:T2contour}), (\ref{Eq:T2Gamma2}), and (\ref{Eq:T2Gamma}), we obtain the total transition amplitude,
\begin{align}
\nonumber
\bra{{f}}\hat{\mathcal{T}}\ket{{i}}
&=\frac{1}{\epsilon_0c^2}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}\\
&~\times
{\vb{j}}_{b'b}(\vb{r})
\cdot \tensor{\vb{G}}(\vb{r},\vb{r}',\omega_{\mathrm{T}})\cdot
{\vb{j}}_{aa'}(\vb{r}').
\label{Eq:TAFinal}
\end{align}
Note that the transition amplitudes in Eqs.~(\ref{Eq:T1}) and (\ref{Eq:T2Gamma2}) mutually cancel out.
Finally, we obtain the RET rate in terms of transition current density in the following expression,
\begin{widetext}
\begin{align}
\Gamma &= \frac{2\pi}{\hbar^2}\int_0^\infty \dd{\omega}
\sum_{(a,b')}\sum_{(a',b)} P_{a'}P_{b}
\abs{
\frac{1}{\epsilon_0c^2}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}
{\vb{j}}_{b'b}(\vb{r})
\cdot \tensor{\vb{G}}(\vb{r},\vb{r}',\omega)\cdot
{\vb{j}}_{aa'}(\vb{r}')
}^2
\delta(\omega_{a'}-\omega_a-\omega)\delta(\omega_{b'}-\omega_b-\omega),
\label{Eq:RETfin}
\end{align}
\end{widetext}
where $(a',b)$ [$(a,b')$] denotes the grouped indices of initial (final) state, and $P_{a'}$($P_b$) denotes the initial-state probability distribution of the entity $A$($B$).
Recall that the transition frequency of $A$ and $B$ are the same, $\omega_{a'}-\omega_{a}=\omega_{b'}-\omega_{b}$.
In the current stage, we have derived an explicit form of the RET rates in terms of transition current density, as shown in Eq.~(\ref{Eq:RETfin}), but this equation cannot be used to directly evaluate RET rates in material systems via \textit{ab initio} methods. To solve this issue, we adopt a further approximation in the next section.
\subsection{Transition Current Density and Molecule}
To evaluate the transition current density via \textit{ab initio} methods, it is necessary to separate the electronic and nuclear degrees of freedom first (approach to recapturing the vibronic effect can be found in Ref.~\cite{Nafie1997,Freedman1998,Freedman1997}).
In the same spirit of the Condon approximation~\cite{Condon1928,Nitzan2006}, we approximate the transition current density to
\begin{align}
\nonumber
\vb{j}_{mm'}(\vb{r})&=\matrixel{m}{\hat{\vb{j}}_M(\vb{r})}{m'}\\
&\approx
\matrixel*{\phi_{M,\gamma}}{\hat{\vb{j}}_M(\vb{r})}{\phi_{M,\gamma'}}_{\{\vb{R}\}
\braket{\chi_{M,\nu}}{\chi_{M,\nu'}}.
\label{Eq:CondonlikeApprox}
\end{align}
where $\ket*{\phi_{M,\gamma}}$ denotes the $\gamma$-th electronic state, and $\ket{\chi_{M,\nu}}$ is the $\nu$-th nuclear state associated with the $\gamma$-th electronic state.
Here, the subscript $\{\vb{R}\}$ represents the electronic element is evaluated under a specific nuclear coordinates $\{\vb{R}\}$.
The Condon-like approximation allows us to separate the electronic and nuclear degrees of freedoms. As a result, we can focus only on the electronic transition current density, and the contribution of the nuclear part can be attributed to the nuclear wavefunction overlap $\braket{\chi_{M,\nu}}{\chi_{M,\nu'}}$.
As a consequence, the quantum number $m$ is now assigned by the two indices, $m\rightarrow(\gamma,\nu)$.
In addition, because most \textit{ab initio} calculations are performed in the coordinate space, the projection of the states to the position-spin coordinates is required.
After taking the projection and considering the antisymmetric property of electrons, we finally get the electronic transition current density (detail derivations can be found in Appendix \ref{sect:TCD})
\begin{align}
\nonumber
\vb{j}_{\gamma\gamma'}^M(\vb{r})
&\equiv\matrixel*{\phi_{M,\gamma}}{\hat{\vb{j}}_M(\vb{r})}{\phi_{M,\gamma'}}_{\{\vb{R}\}}\\
\nonumber
&=
\frac{-i\hbar e N_\mathrm{el}}{2m_\mathrm{el}
\bigg[
\tilde{\phi}_{M,\gamma}^*(\vb{r};\{\vb{R}\})\nabla\tilde{\phi}_{M,\gamma'}(\vb{r};\{\vb{R}\})\\
&\hspace{1.3 cm}-
\tilde{\phi}_{M,\gamma'}(\vb{r};\{\vb{R}\})\nabla\tilde{\phi}_{M,\gamma}^*(\vb{r};\{\vb{R}\})
\bigg],
\label{Eq:j_ewfn}
\end{align}
where $N_\mathrm{el}$ is the total number of electrons in molecule $M$, $e$ is the elementary charge, and $m_\mathrm{el}$ is the electron mass.
Remind that the gradient operator only operate on $\vb{r}$, not $\{\vb{R}\}$.
Moreover, $\tilde{\phi}_{M,\gamma}(\vb{r},\{\vb{R}\})$ is the single-electron wavefunction of the electronic state $\gamma$, which is defined by
\begin{align}
\nonumber
&\tilde{\phi}_{M,\gamma}^*(\vb{x}_1;\{\vb{R}\})\tilde{\phi}_{M,\gamma'}(\vb{x}_1;\{\vb{R}\})\\
&\equiv
\int \mathcal{D}\{\vb{x}_{\mu\neq1}\}~\phi_{M,\gamma}^*(\{\vb{x}_\mu\};\{\vb{R}\})\phi_{M,\gamma'}(\{\vb{x}_\mu\};\{\vb{R}\}).
\label{Eq:eff1ewfn}
\end{align}
In Eq.~(\ref{Eq:eff1ewfn}), $\{\vb{x}_\mu\}$ denotes the set of position-spin coordinates, and $\phi_{M,\gamma}(\{\vb{x}_\mu\};\{\vb{R}\})$ denotes the multi-electron wavefunction of the electronic state $\gamma$, which is parameterized by a specific nuclear coordinates $\{\vb{R}\}$. The integration symbol represents a series of integrals except for $\vb{x}_1$,
\begin{align}
\nonumber
\int \mathcal{D}\{\vb{x}_{\mu\neq1}\}\equiv
\int \dd{\vb{x}_2} \dots \int \dd{\vb{x}_{N_\mathrm{el}}}.
\end{align}
Specially, in molecular systems, it is common to approximate the single-electron wavefunctions as molecular orbitals (MO), which can be obtained from \textit{ab initio} calculations.
If the transition between the electronic ground state and the first excited state plays the most important transition,
one can assume that the single-electron wavefunction can be properly described by
\begin{subequations}
\begin{align}
\tilde{\phi}_{M,\mathrm{e}}(\vb{r};\{\vb{R}\})&\approx
\phi_{M,\mathrm{LUMO}}(\vb{r};\{\vb{R}\}),
&\gamma' \equiv \mathrm{e},\\
\tilde{\phi}_{M,\mathrm{g}}(\vb{r};\{\vb{R}\})&\approx
\phi_{M,\mathrm{HOMO}}(\vb{r};\{\vb{R}\}),
&\gamma \equiv \mathrm{g},
\end{align}
\label{Eq:MOapprox}%
\end{subequations}
where LUMO and HOMO are the abbreviation of the lowest unoccupied MO and the highest occupied MO, respectively.
In other words, the transition current density for the molecule $M$ is determined by its HOMO, LUMO, and the nuclear wavefunction overlap.
Under the approximation in Eq.~(\ref{Eq:MOapprox}), the transition current density can be expressed as
\begin{align}
\vb{j}_{mm'}(\vb{r})&=
\braket{\chi_{M,\nu}}{\chi_{M,\nu'}}~
\vb{j}_\mathrm{ge}^M(\vb{r}).
\end{align}
According to Eqs.~(\ref{Eq:j_ewfn}) and~(\ref{Eq:MOapprox}), one can obtain $\vb{j}_\mathrm{ge}^M(\vb{r})$ as follows,
\begin{align}
\nonumber
\vb{j}_\mathrm{ge}^M(\vb{r}) &=
\frac{-i\hbar e N_\mathrm{el}}{2m_\mathrm{el}}\\
\nonumber
&~\times
\bigg[
\phi_{M,\mathrm{HOMO}}^*(\vb{r};\{\vb{R}\})\nabla\phi_{M,\mathrm{LUMO}}(\vb{r};\{\vb{R}\})\\
&\hspace{.31 cm}-
\phi_{M,\mathrm{LUMO}}(\vb{r};\{\vb{R}\})\nabla\phi_{M,\mathrm{HOMO}}^*(\vb{r};\{\vb{R}\})
\bigg].
\end{align}
Eventually, the RET rates in Eq.~(\ref{Eq:RETfin}) can be expressed and interpreted as a generalized spectral overlap between two molecules and electromagnetic coupling factor $F(\omega)$,
\begin{align}
\Gamma &= \frac{2\pi}{\hbar^2}\int_0^\infty \dd{\omega}
\mathcal{L}_B^{\mathrm{abs}}(\omega)
F(\omega)
\mathcal{L}_A^{\mathrm{em}}(\omega),
\label{Eq:RETCondonfin}
\end{align}
with
\begin{align}
\nonumber
\mathcal{L}_A^{\mathrm{em}}(\omega)&\equiv
\sum_{\alpha,\alpha'}
P_{\alpha'}
\abs{\braket{\chi_{A,\alpha}}{\chi_{A,\alpha'}}}^2
\delta(\omega_{\alpha'}-\omega_\alpha-\omega),\\
\nonumber
\mathcal{L}_B^{\mathrm{abs}}(\omega)&\equiv
\sum_{\beta,\beta'}
P_{\beta}
\abs{\braket{\chi_{B,\beta'}}{\chi_{B,\beta}}}^2
\delta(\omega_{\beta'}-\omega_\beta-\omega),
\end{align}
and
\begin{align}
\nonumber
F(\omega
&=
\abs{\frac{1}{\epsilon_0c^2}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}
{\vb{j}}_\mathrm{eg}^B(\vb{r})
\cdot \tensor{\vb{G}}(\vb{r},\vb{r}',\omega)\cdot
{\vb{j}}_\mathrm{ge}^A(\vb{r}')}^2.
\end{align}
The schematic illustration of $F(\omega)$ is shown in Fig.~\ref{Fig:Illust}a.
Note that $\mathcal{L}_B^{\mathrm{abs}}(\omega)$ and $\mathcal{L}_A^{\mathrm{em}}(\omega)$ are related to the absorption spectrum of $B$ and the emission spectrum of $A$, respectively.
\begin{figure*}[!t]
\centering
\includegraphics[width=1\textwidth]{SchematicIllustration_mu.pdf}
\caption{Schematic diagram of the electromagnetic coupling factors in three different theoretical approaches: (a) transition current density, (b) transition density, and (c) transition dipole moment.}
\label{Fig:Illust}
\end{figure*}
\section{Applicability to Former Theories}
\label{sect:app}
In Sec.~\ref{sect:theory}, we have shown how to derive the RET rate based on the transition current density approach in the framework of macroscopic quantum electrodynamics.
Furthermore, to demonstrate the generality of our theory, in this section, we will prove that Eq.~(\ref{Eq:RETCondonfin}) can recover to the main results in previous studies: the TDC method~\cite{Krueger1998} and the plasmon-coupled resonance energy transfer~\cite{Hsu2017}.
\subsection{Transition Density Cube Method}
When two entities are close, Krueger \textit{et al.} have developed the TDC method \cite{Krueger1998} in the electrostatic limit and demonstrated how to calculate the Coulomb coupling between the pigments of the bacterial light-harvesting complex.
To recover the main result given in the TDC method, we consider the free-space dyadic Green's function $\tensor{\vb{G}}_0(\vb{r},\vb{r}',\omega)$ and adopt the electrostatic approximation,
\begin{align}
\tensor{\vb{G}}(\vb{r},\vb{r}',\omega)\rightarrow
\lim_{k\rightarrow 0} \tensor{\vb{G}}_0(\vb{r},\vb{r}',\omega)=
\nabla\nabla\frac{1}{4\pi k^2R},
\end{align}
where $k=\omega/c$, $R=\abs{\vb{r}-\vb{r}'}$, and $\nabla\nabla$ is a dyadic operator.
Thus, under the electrostatic limit, the electromagnetic coupling factor $F(\omega)$ becomes $F_\mathrm{TD}$ (TD is the acronym for the transition density) as follows,
\begin{align}
\nonumber
F_\mathrm{TD}&\equiv\lim_{k\rightarrow0}F(\omega)\\
\nonumber
&=\abs{\frac{1}{\epsilon_0 c^2}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}
\nabla\cdot{\vb{j}}_\mathrm{eg}^B(\vb{r})
\frac{1}{4\pi k^2R}
\nabla'\cdot{\vb{j}}_\mathrm{ge}^A(\vb{r}')}^2\\
&= \abs{\frac{1}{4\pi\epsilon_0}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}
\rho_\mathrm{eg}^B(\vb{r})g(\vb{r},\vb{r}')\rho_\mathrm{ge}^A(\vb{r}')}^2,
\label{Eq:TDCcoupling}
\end{align}
where $\rho_\mathrm{ge}^A(\vb{r}')$ and $\rho_\mathrm{eg}^B(\vb{r})$ are transition densities of the entity A and the entity B, respectively, and $g(\vb{r},\vb{r}')=1/\abs{\vb{r}-\vb{r}'}$ is the scalar Green's function of the Poisson equation.
Note that the derivation of Eq.~(\ref{Eq:TDCcoupling}) requires the continuity equation.
Obviously, $F_\mathrm{TD}$ is exactly the continuous form of the electronic coupling corresponding to Eq.~(5) in Ref.~\cite{Krueger1998}.
Incidentally, to numerically perform the integral in Eq.~(\ref{Eq:TDCcoupling}), Krueger \textit{et al.} discretized the space into sufficiently small cubes shown in Fig.~\ref{Fig:Illust}b, hence the name transition density cube (TDC).
Furthermore, because $F_\mathrm{TD}$ is independent of frequency, it can be taken out from the generalized spectral overlap, and the RET rate becomes
\begin{align}
\Gamma = \frac{2\pi}{\hbar^2}F_\mathrm{TD}J,
\qquad J = \int_0^\infty\dd{\omega}
\mathcal{L}_B^{\mathrm{abs}}(\omega)
\mathcal{L}_A^{\mathrm{em}}(\omega),
\label{Eq:TDCrate}
\end{align}
where $F_\mathrm{TD}$ corresponds to $\abs{V}^2$ in Eq.~(8) in Ref.~\cite{Krueger1998}.
Note that the prefactor of Eq.~(8) in Ref.~\cite{Krueger1998} is slightly different from that in Eq.~(\ref{Eq:TDCrate}) because of their spectral overlap integrals in different units ($J$ in Ref.~\cite{Krueger1998} expressed in $\bar{\nu} = \omega/2\pi c$).
In addition, we would like to point out that $J$ defined in Eq.~(\ref{Eq:TDCrate}) is associated with the well-known spectral overlap in F\"orster's theory.
In brief, the TDC method is the electrostatic limit of our theory in free space (i.e., neglect
of the retardation effect in homogeneous, non-dispersive, and non-absorbing media).
\subsection{Plasmon-Coupled Resonance Energy Transfer}
Equation~(\ref{Eq:RETCondonfin}) is general for us to describe RET in linear, dispersive, and absorbing media. In other words, the effect of plasmon polaritons can be included in our theory, indicating that the main result of plasmon-coupled RET should be able to be recovered.
The influence of plasmon polaritons on the RET rates under the EDA have been discussed in various electromagnetic environments~\cite{Ding2017,Ding2018,Wu2018,Lee2020,Wei2021}.
To obtain the RET rate [Eq.~(2)] in Ref.~\cite{Hsu2017}, we start from the definition of the current density and apply the relation $\hat{\vb{p}}_\xi=\comm*{\hat{\vb{r}}_\xi}{\hat{H}_M}$,
\begin{align}
\nonumber
{\hat{\vb{j}}}_M(\vb{r})
&=
\sum_{\xi\in M}\frac{q_\xi}{2i\hbar}\bigg\{
\delta(\vb{r}-\hat{\vb{r}}_\xi)\comm{\hat{\vb{r}}_\xi}{\hat{H}_M}+
\mathrm{H.c.}
\bigg\}\\
\nonumber
&\approx
\sum_{\xi\in M}\frac{q_\xi}{2i\hbar}\bigg\{
\delta(\vb{r}-\vb{r}_M)\comm{\hat{\vb{r}}_\xi-\vb{r}_M}{\hat{H}_M}+
\mathrm{H.c.}
\bigg\}\\
&=
\frac{1}{i\hbar}\comm{\hat{\boldsymbol{\mu}}_M}{\hat{H}_M}
\delta(\vb{r}-\vb{r}_M)
\label{Eq:jpdapprox}
\end{align}
Here, we introduce the center of mass $\vb{r}_M$ for the entity $M$ and make the point-dipole approximation to get the dipole operator $\hat{\boldsymbol{\mu}}_M$.
Through calculating $\bra{m(m')}{\hat{\vb{j}}}_M(\vb{r})\ket{m'(m)}=\vb{j}_{mm'(m'm)}(\vb{r})$ in Eq.~(\ref{Eq:jpdapprox}), where $\ket{m(m')}$ is the $m$-th ($m'$-th) energy eigenstate of an entity $M$, we can obtain
\begin{align}
\vb{j}_{mm'(m'm)}(\vb{r})=
\mp i\omega_\mathrm{T}\boldsymbol{\mu}_{mm'(m'm)}
\delta(\vb{r}-\vb{r}_M),
\label{Eq:dipoleelement}
\end{align}
Next, we take the Condon approximation to the transition dipole $\boldsymbol{\mu}_{mm'(m'm)}$, it is straightforward to obtain
\begin{align}
\boldsymbol{\mu}_{mm'(m'm)} \approx
\boldsymbol{\mu}_{\mathrm{ge(eg)}}^M
\braket{\chi_{M,\nu(\nu')}}{\chi_{M,\nu'(\nu)}},
\label{Eq:CondonApprox}
\end{align}
where $\boldsymbol{\mu}_{\mathrm{ge(eg)}}^M$ is the electronic transition dipole.
Substituting Eqs.~(\ref{Eq:dipoleelement}) and (\ref{Eq:CondonApprox}) into Eq.~(\ref{Eq:RETfin}), we obtain that
\begin{align}
\label{Eq:RETRDDI}
\Gamma
&= \frac{2\pi}{\hbar^2}\int_0^\infty \dd{\omega}
\mathcal{L}_B^{\mathrm{abs}}(\omega)
F_\mathrm{RDDI}(\omega)
\mathcal{L}_A^{\mathrm{em}}(\omega),
\end{align}%
where
\begin{align}
F_\mathrm{RDDI}(\omega)=
\abs{
\frac{\omega^2}{\epsilon_0c^2}
\boldsymbol{\mu}_\mathrm{eg}^B
\cdot \tensor{\vb{G}}(\vb{r}_B,\vb{r}_A,\omega)\cdot
\boldsymbol{\mu}_\mathrm{ge}^A
}^2
\end{align}%
is exactly the form of resonant dipole-dipole interaction (RDDI) in Ref.~\cite{Hsu2017}.
Here, we would like to mention that the definition of $F_\mathrm{RDDI}(\omega)$ is slightly different from that of coupling factor in Ref.~\cite{Hsu2017} because the magnitudes of the transition dipoles of molecules $A$ and $B$ in Ref.~\cite{Hsu2017} have been incorporated into the emission and absorption lineshape functions, i.e.,
\begin{subequations}
\begin{align}
W_A^\mathrm{em}(\omega
&= \frac{2\pi}{\hbar^2}
\abs{\boldsymbol{\mu}_\mathrm{ge}^A}^2
\mathcal{L}_A^{\mathrm{em}}(\omega),\\
W_B^\mathrm{abs}(\omega
&=\frac{2\pi}{\hbar^2}
\abs{\boldsymbol{\mu}_\mathrm{eg}^B}^2
\mathcal{L}_B^{\mathrm{abs}}(\omega).
\end{align}
\label{Eq:lineshape}%
\end{subequations}
According to Eq.~(\ref{Eq:lineshape}), it is obvious that Eq.~(\ref{Eq:RETCondonfin}) covers the main result in Ref.~\cite{Hsu2017}.
Moreover, it is worth pointing out that if we further impose the electrostatic limit on Eq.~(\ref{Eq:RETRDDI}), we can obtain the famous F\"orster theory, as discussed in Ref.~\cite{Hsu2017}.
To emphasize the difference among these RET theories, we provide a schematic diagram, shown in Fig.~\ref{Fig:Illust}, that depicts the key concepts of the three electromagnetic coupling factors in RET theories.
\begin{table*}[!t]
\caption{Summary of RET theories with and without approximations.}
\label{table:1}
\begin{ruledtabular}
\begin{tabular}{lcll}
Method & Approximation & Rate & Coupling Form\\
\hline
\footnote{Transition current density}TCD & - &
$\displaystyle
\Gamma = \frac{2\pi}{\hbar^2}\int_0^\infty \dd{\omega}
\mathcal{L}_B^{\mathrm{abs}}(\omega)
F(\omega)
\mathcal{L}_A^{\mathrm{em}}(\omega)$
&
$F(\omega)=\displaystyle
\abs{\frac{1}{\epsilon_0c^2}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}
{\vb{j}}_\mathrm{eg}^B(\vb{r})
\cdot \tensor{\vb{G}}(\vb{r},\vb{r}',\omega)\cdot
{\vb{j}}_\mathrm{ge}^A(\vb{r}')}^2$\\
\footnote{Transition density}TD & Electrostatic &
$\displaystyle\Gamma = \frac{2\pi}{\hbar^2}
F_\mathrm{TD}
\int_0^\infty\dd{\omega}
\mathcal{L}_B^{\mathrm{abs}}(\omega)
\mathcal{L}_A^{\mathrm{em}}(\omega)$
&
$\displaystyle
F_\mathrm{TD}=\abs{\frac{1}{4\pi\epsilon_0}
\int\dd[3]{\vb{r}} \int\dd[3]{\vb{r}'}
\rho_\mathrm{eg}^B(\vb{r})
g(\vb{r},\vb{r}')
\rho_\mathrm{ge}^A(\vb{r}')}^2$ \\
\footnote{Transition dipole moment}TDM & Point dipole &
$\displaystyle\Gamma
= \frac{2\pi}{\hbar^2}\int_0^\infty \dd{\omega}
\mathcal{L}_B^{\mathrm{abs}}(\omega)
F_\mathrm{RDDI}(\omega)
\mathcal{L}_A^{\mathrm{em}}(\omega)$
&
$\displaystyle
F_\mathrm{RDDI}(\omega)=
\abs{
\frac{\omega^2}{\epsilon_0c^2}
\boldsymbol{\mu}_\mathrm{eg}^B
\cdot \tensor{\vb{G}}(\vb{r}_B,\vb{r}_A,\omega)\cdot
\boldsymbol{\mu}_\mathrm{ge}^A
}^2$\\
\footnote{F\"orster resonance energy transfer}FRET & \makecell{Electrostatic\\\& point dipole} &
$\displaystyle\Gamma = \frac{2\pi}{\hbar^2}
F_\mathrm{FRET}
\int_0^\infty\dd{\omega}
\mathcal{L}_B^{\mathrm{abs}}(\omega)
\mathcal{L}_A^{\mathrm{em}}(\omega)$
&
%
$\displaystyle F_\mathrm{FRET}=
\frac{\footnote{$\kappa$ is the orientation factor.}\kappa^2}{4\pi\epsilon_0R^6}
\abs{\boldsymbol{\mu}_\mathrm{eg}^B}^2
\abs{\boldsymbol{\mu}_\mathrm{ge}^A}^2
$
\end{tabular}
\end{ruledtabular}
\end{table*}
\section{Conclusion}
Resonant dipole-dopole interaction has been widely used to describe light-matter interaction in physics, chemistry, and material sciences; however, RDDI is only a first-order approximation describing the light-matter interaction and cannot contain the structural information of entities, e.g., atoms, molecules, quantum dots, 2D materials.
To include the spatial-dependent interaction, we developed the generalized RET theory based on the concept of transition current density in the presence of linear, dispersive, and absorbing media within the framework of macroscopic quantum electrodynamics.
By expanding the Born series up to the second order in the time-dependent perturbation theory, we successfully derived the RET rate in the generalized-coupling expression of the transition dipole moment.
Furthermore, by applying the Condon-like approximation to the transition current density, we separated the electronic and nuclear degrees of freedom and showed that the transition current density can be described by the HOMO and LUMO, which can be obtained from \textit{ab initio} calculations.
Moreover, we expressed the RET rates in terms of the generalized spectral overlap, as shown in Eq.~(\ref{Eq:RETCondonfin}).
Finally, to demonstrate the validity and generality of Eq.~(\ref{Eq:RETCondonfin}), we proved that the present theory can be reduced to the main results in the previous studies, including TDC and plasmon-coupled RET. The comparison of several representative RET theories are summarized in Table \ref{table:1}.
In short, in the framework of macroscopic quantum electrodynamics, the current approach provides one key step beyond the traditional RET theory based on RDDI because Eq.~(\ref{Eq:RETCondonfin}) not only serves as a generalized version (i.e., containing retardation effect) of the TDC method, but also includes the influence of photonic environments, e.g., polartions.
The generalized RET theory beyond RDDI has been presented in this work. However, this study is just the beginning, and several issues are worth further exploration.
First, in the present theory, we do not consider the mechanism of Dexter energy transfer~\cite{Dexter1953,Skourtis2016,Murphy2004,Olaya-Castro2011} (i.e., electron exchange between two entities).
This mechanism becomes important when the wavefunction overlap of two entities cannot be negligible.
Second, the quantum dynamics of RET cannot be described in the present theory due to the limitation of Fermi's golden rule.
How to generalize the theory to include quantum dynamics is an intriguing but challenging issue.
In the end, we leave the numerical demonstration to the future study and hope that the present theory will inspire further investigation into the basic theory of energy transfer and its applications.
\begin{acknowledgments}
Hsu thanks Academia Sinica (AS-CDA-111-M02) and the Ministry of Science and Technology of Taiwan (110-2113-M-001-053 and 111-2113-M-001-027-MY4) for the financial support.
\end{acknowledgments}
|
\subsection{3D Forward Dynamics}\label{sec:approach_dyna}
At the core of our model lies a transformation prediction module. Given an RGB-D frame $\mathbf{x}_t = (\mathbf{I}_t, \mathbf{D}_t)$, a latent variable $\mathbf{z}_t$, and optionally an $n$-dimensional continuous action $\mathbf{a}_t$ as inputs, T3VIP learns a motion embedding from which it decomposes a scene into $K$ object masks $\mathbf{M}_t$ (including background) and predicts a 3D rigid body transformation $[\mathbf{R}_t,\mathbf{T}_t] \in \mathbf{SE}(3)$ per object. Using camera intrinsics, our model first converts the depth map $\mathbf{D}_t$ of the current frame into an ordered 3D point cloud $\mathbf{P}_t=(X_t, Y_t, Z_t)$, where each point contains the 3D coordinates of the scene, and then moves this point cloud according to the predicted object masks and $\mathbf{SE}(3)$ transformations to generate a transformed point cloud:\looseness=-1
\begin{equation}\label{eq:transformed_pointcloud}
\hat{\mathbf{P}}_{t+1} = \sum_{k=1}^{K} \mathbf{M}_{t}^{k} (\mathbf{R}_t^{k} \mathbf{P}_{t} + \mathbf{T}_t^{k}).
\end{equation}
Since our transformation module explicitly learns the 3D motion of point cloud segments and utilizes information from the previous point cloud to construct the transformed point cloud, it remains invariant to the visual appearance of objects. Please note that $\hat{\mathbf{P}}_{t+1}$ is no longer an ordered point cloud since the transformation layer moves each point in 3D and the previous x/y-axis ordering of points are not valid anymore. Hence, we do not directly obtain the next depth map $\hat{\mathbf{D}}_{t+1}$, but a transformed depth map $\hat{\mathbf{D}}^{\prime}_{t+1}$ that is unordered.
Nevertheless, we can compute a binary occlusion mask $\hat{\mathbf{O}}_t \in \mathbb{R}^{1 \times H \times W}$ analytically by performing a perspective projection test of the transformed point cloud. This mask addresses points that are occluded in the initial frame but will be visible in the next frame. Furthermore, we can compute the scene flow $\hat{\mathbf{V}}_t \in \mathbb{R}^{3 \times H \times W}$ which is the 3D motion of scene points with respect to the camera and optical flow $\hat{\mathbf{U}}_t \in \mathbb{R}^{2 \times H \times W}$ which is the projection of 3D motion onto the image plane as follows:\looseness=-1
\begin{equation}
\renewcommand*{\arraystretch}{1.3}
\hat{\mathbf{V}}_{t} = \hat{\mathbf{P}}_{t+1} - \mathbf{P}_t
\quad\text{and}\quad
\hat{\mathbf{U}}_{t} = \mathbf{w}(\hat{\mathbf{V}}_{t}),
\end{equation}
where $\mathbf{w}$ is the projection layer, which utilizes the camera intrinsics to project the scene flow over the camera plane. \looseness=-1
Our forward dynamics model closely resembles the forward model of Hind4sight-Net~\cite{nematollahi2020hindsight}, with the difference that T3VIP employs convolutional LSTMs to leverage the spatial invariance of frame representations for multi-step future prediction. Compared to CDNA~\cite{finn2016unsupervised} proposed by Finn \textit{et al. } for 2D video prediction, our model makes the same architectural choices for encoding the frames and actions and decoding the masks of the objects, but it differs in how it decodes the motion. T3VIP predicts 3D rigid body transformations instead of 2D normalized convolution kernels.\looseness=-1
\subsection{Generator}\label{sec:approach_gen}
To generate the next RGB-D frame, our model first uses the predicted optical flow $\hat{\mathbf{U}}_t$ to forward-warp the transformed depth map $\hat{\mathbf{D}}^{\prime}_{t+1}$ and the RGB image $\mathbf{I}_t$ to the coordinates aligned with the predicted 3D motion:\looseness=-1
\begin{equation}
\renewcommand*{\arraystretch}{1.3}
\hat{\mathbf{I}}^{fw}_{t+1} = \overrightarrow{\sigma} (\mathbf{I}_t, \hat{\mathbf{U}}_t)
\quad\text{and}\quad
\hat{\mathbf{D}}^{fw}_{t+1} = \overrightarrow{\sigma} (\hat{\mathbf{D}}^{\prime}_{t+1}, \hat{\mathbf{U}}_t),
\end{equation}
where $\overrightarrow{\sigma}$ is the splatting operator proposed by~\cite{niklaus2020softmax}, and $\hat{\mathbf{I}}^{fw}_{t+1}$ and $\hat{\mathbf{D}}^{fw}_{t+1}$ are the forward-warped RGB image and depth map. While we now have the next frame generated via the predicted 3D motion, we need to address occluded regions that are missing in the new frame. To this end, our model employs an RGB-D inpainter, which takes the motion embedding as input and predicts an RGB image $\hat{\mathbf{I}}^{in}_{t+1}$, and a depth map $\hat{\mathbf{D}}^{in}_{t+1}$. Finally, T3VIP generates the next RGB-D frame $\hat{\mathbf{x}}_{t+1}$ as follows:\looseness=-1
\begin{align} \begin{split}
\hat{\mathbf{I}}_{t+1} &= (1 - \hat{\mathbf{O}}_t) \cdot \hat{\mathbf{I}}^{fw}_{t+1} + \hat{\mathbf{O}}_t \cdot \hat{\mathbf{I}}^{in}_{t+1},
\end{split} \\ \begin{split}
\hat{\mathbf{D}}_{t+1} &= (1 - \hat{\mathbf{O}}_t) \cdot \hat{\mathbf{D}}^{fw}_{t+1} + \hat{\mathbf{O}}_t \cdot \hat{\mathbf{D}}^{in}_{t+1}.
\end{split} \end{align}
Our binary occlusion mask makes sure that the model mainly uses the predicted 3D transformations to generate the next frame and only occluded points of the scene get inpainted.\looseness=-1
\subsection{Modeling Stochasticity}\label{sec:approach_sto}
To enable our model to grasp the stochastic nature of the real world during training, we introduce a latent variable $\mathbf{z}_t$ into our recurrent generative model and sample from $p(\rgbdx{c}{T}\mid \rgbdx{0}{c-1}, \latentx{0}{T-1})$. This hinders the generative model from directly maximizing the likelihood of the data due to the latent's dependency on $p(\rgbdx{0}{T})$. To overcome this problem, we follow the approach proposed by Babaeizadeh \textit{et al. }~\cite{babaeizadeh2018stochastic} to approximate the posterior with an inference network $q_{\phi}(\mathbf{z}_t \mid \rgbdx{0}{T})$ and optimize the variational lower bound of the log-likelihood. In order to encourage the latent variable to discover the stochastic information between frames during training, the inference network takes the entire video sequence as input and outputs the parameters of a conditionally Gaussian distribution $\mathcal{N}(\mu_\phi(\rgbdx{0}{T}), \sigma_\phi(\rgbdx{0}{T}))$. However, at test time, we sample the latents from an assumed prior which in our case is a fixed unit Gaussian ${\mathcal{N}(\mathbf{0}, \mathbf{I})}$. The optimization of our model involves minimizing the reconstruction loss between the predicted and ground-truth frames and the KL divergence between the approximated posterior and the assumed prior.\looseness=-1
\begin{equation}
\mathcal{L}_{\mathit{rec}}(\rgbdx{0}{T}) = \mathbb{E}_{\substack{\rgbdx{0}{T} \\
\mathbf{z}_{t} \sim {q_{\phi}}}} \left[\sum_{t=0}^{T-1} ||\mathbf{x}_t - G_{\theta}(\mathbf{x}_0, \mathbf{z}_{0:t-1})||_{\alpha} \right],
\end{equation}
\begin{equation}
\mathcal{L}_{\mathit{kl}}(\mathbf{x}_{0:T}) = \mathbb{E}_{\rgbdx{0}{T}}\left[\sum_{t=0}^{T-1} D_{\mathit{KL}}\left(q_{\phi}(\mathbf{z}_t \mid \rgbdx{0}{T}) \mid\mid p(\mathbf{z})\right)\right],
\end{equation}
where $\alpha$ is a hyperparameter specifying if the pixel-wise absolute ($\alpha=1$) or squared ($\alpha=2$) error is used to reconstruct frames. Our inference network $q_{\phi}$ closely mimics SV2P~\cite{babaeizadeh2018stochastic}, with the difference that ours incorporates RGB-D sequences of the video instead of RGB.\looseness=-1
\subsection{3D Point Cloud Alignment Loss}\label{sec:approach_3d}
Our approach learns the 3D scene dynamics of the real-world without requiring any external trackers and by solely relying on unlabeled RGB-D videos. In this regard, our optimization formulation leverages observational cues to enforce an explicit geometric constraint on predicting the forward dynamics. Concretely, we impose consistency between the transformed $\hat{\mathbf{P}}_{t+1}$ and observed $\mathbf{P}_{t+1}$ point clouds at each time step. Since the transformed point cloud is unordered, we employ the k-Nearest-Neighbors (kNN) algorithm to find the point-wise data association between the transformed and observed point clouds. Hence, kNN takes as input two point clouds $\hat{\mathbf{P}}_{t+1}$ and $\mathbf{P}_{t+1}$, and for each point in each point cloud, it first finds the nearest neighbor in the other point set based on their distance to the camera and then forms a distance transform by summing up the euclidian distance of corresponding points. Since kNN is not necessarily a symmetric function, we calculate the distance transforms in both directions. We define the distance transforms between the point sets as follows:\looseness=-1
\begin{align} \begin{split}
\mathbf{J}_{t+1}^{\hat{\mathbf{P}} \veryshortarrow \mathbf{P}} &= \min_{x^{\prime},y^{\prime}} \sum_{x,y} \|\hat{\mathbf{P}}_{t+1}^{xy} - \mathbf{P}_{t+1}^{x^{\prime}y^{\prime}}\|_{2},
\end{split} \\ \begin{split}
\mathbf{J}_{t+1}^{\mathbf{P} \veryshortarrow \hat{\mathbf{P}}} &= \min_{x^{\prime},y^{\prime}} \sum_{x,y} \|\mathbf{P}_{t+1}^{xy} - \hat{\mathbf{P}}_{t+1}^{x^{\prime}y^{\prime}}\|_{2},
\end{split} \end{align}
where $x$ and $y$ are the grid coordinates of points, $x^{\prime} \in (x-k, x+k)$ and $y^{\prime} \in (y-k, y+k)$ are the coordinates of the data associated points, and $k$ is the kNN parameter. Finally, we sum the distance transforms to define the point cloud alignment loss:
\begin{equation}
\label{eq:knn_dist}
\mathcal{L}_{\mathit{knn}}(\mathbf{x}_{0:T}) = \mathbb{E}_{\rgbdx{0}{T}}\left[\sum_{t=0}^{T-1} \mathbf{J}_{t+1}^{\hat{\mathbf{P}} \veryshortarrow \mathbf{P}} + \mathbf{J}_{t+1}^{\mathbf{P} \veryshortarrow \hat{\mathbf{P}}}\right].
\end{equation}
Please note that compared with Hind4sight-Net~\cite{nematollahi2020hindsight}, which utilizes a chamfer distance for the point cloud alignment loss, our kNN formulation is computationally more efficient since it finds the associated point in a smaller $k \times k$ window instead of an $H \times W$ window required by the chamfer distance.\looseness=-1
\subsection{Edge-aware Smoothness Loss}\label{sec:approach_edge}
T3VIP explicitly reasons about the 3D transformations of the scene segments in a self-supervised fashion and outputs the scene flow and optical flow as emergent properties. As an additional proxy loss for our self-supervised framework, we adopt an edge-aware second-order smoothness regularization~\cite{tomasi1998bilateral}, to encourage piecewise smoothness of geometry and motion. Our smoothness loss term measures the difference between spatially neighboring points in the scene and optical flow, adaptively weighted by the image gradients:
\begin{equation*}
\mathcal{L}_{\mathit{fs}}(\mathbf{x}_{0:T}) =\mathbb{E}_{\rgbdx{0}{T}} \left[\sum_{t=0}^{T-1} \sum_{x,y} {\big\lvert}{\nabla^{2} {\mathbf{F}}_{t}^{xy}}{\big\lvert} \times \exp{\left(-|\nabla \mathbf{I}_{t}^{xy}|\right)} \right],
\end{equation*}
where $|\cdot|$ denotes element-wise absolute value, $\nabla$ is the vector differential operator, and ${\mathbf{F}}_{t}$ could be either the scene or optical flow. Intuitively, by utilizing an edge-aware smoothness penalty, we leverage the observation that motion boundaries and edges present in the image usually coincide.\looseness=-1
\subsection{Full Model}\label{sec:approach_full}
Our full model combines all the above-mentioned objectives to learn a stochastic 3D video prediction model from unlabeled RGB-D data. Namely, our model aims to reconstruct RGB images ($\mathcal{L}_{rec}^{I}$) and depth maps ($\mathcal{L}_{rec}^{D}$), enforce the consistency of predicted and observed point clouds ($\mathcal{L}_{knn}$), encourage scene flow ($\mathcal{L}_{fs}^{s}$) and optical flow smoothness ($\mathcal{L}_{fs}^{o}$) and fit the prior distribution ($\mathcal{L}_{kl}$). Hence, the full objective of our T3VIP model is:\looseness=-1
\begin{equation*}
\label{eq:full_loss}
\mathcal{L} = \lambda_{1}\mathcal{L}_{\mathit{rec}}^{I} + \lambda_{2}\mathcal{L}_{\mathit{rec}}^{D} + \lambda_{3}\mathcal{L}_{\mathit{knn}} + \lambda_{4}\mathcal{L}_{\mathit{fs}}^{s} + \lambda_{5}\mathcal{L}_{\mathit{fs}}^{o} + \lambda_{6}\mathcal{L}_{\mathit{kl}},
\end{equation*}
where $\lambda_{1,\ldots ,6}$ are hyperparameters representing the relevance of each loss term.\looseness=-1
\subsection{Hyperparameter Optimization}\label{sec:approach_hpo}
Since observational statistics of a robot depend greatly on the specific environment that it collects data from and its onboard sensors, the best hyperparameter configuration of a model usually varies across different datasets. Manual hyperparameter optimization (HPO) is thus cumbersome and requires expert knowledge. To overcome this bottleneck, we equip our model with automated HPO (AutoML) techniques~\cite{hutter2019automated}, which have been shown to determine well-performing hyperparameters automatically. They improve performance substantially over manual tuning not just in supervised learning \cite{Feurer2019} but also in Model-based Reinforcement Learning and Planning \cite{zhang2021importance} which is closer to our setting. More concretely, we tune our model for automatically setting the learning rate, $\alpha$ (L1 or L2 reconstruction loss), and the $\lambda_{1, \ldots , 6}$ hyperparameters (importance weights of losses). We tested the Ray Tune \cite{moritz2018ray} implementations of ASHA \cite{li2020system}, BOHB \cite{falkner2018bohb} and HEBO~\cite{cowen2020hebo} and selected ASHA as it fully utilized the available parallel resources. ASHA is an asynchronous version of HyperBand \cite{hyperband_jmlr_17}, a multi-fidelity HPO approach. As such, it terminates runs with poorly performing configurations after having run for smaller budgets while promoting well performing configurations to run for larger budgets.\looseness=-1
\subsection{Evaluation Metrics}
Following prior work on stochastic video prediction evaluation~\cite{babaeizadeh2018stochastic}, we sample 100 latent values from the prior distribution per video and report the best sample results. More concretely, for the predicted RGB images, we report PSNR and SSIM and VGG cosine similarity~\cite{zhang2018unreasonable} scores. VGG similarity has been shown to coincide better with human perceptual judgments~\cite{zhang2018unreasonable}. We further measure the standard quality metrics of predicted depth maps~\cite{eigen2015predicting} and report two widely-accepted error metrics: root mean square error (RMSE) and absolute relative error (AbsRel). Finally, we quantitatively evaluate the success rate of a model-predictive control algorithm that employs our 3D-aware model to plan for reaching 3D goal points.\looseness=-1
\subsection{Datasets}
We evaluate our model extensively with one real-world and two synthetic datasets. Please note that we use all datasets with the spatial resolution of $64 \times 64$. Datasets are as following:\looseness=-1
\noindent\textbf{DexHand}: Inspired by OpenAI dexterous in-hand manipulation~\cite{andrychowicz2020learning}, we collected a synthetic RGB-D dataset of a Shadow Hand robot manipulating a cube towards arbitrary goal configurations. This dataset consists of about $10000$ videos, each video including $25$ RGB-D frames. DexHand is challenging as the robot has $24$ degrees of freedom, and there can be a significant amount of motion and occlusion between consecutive frames.\looseness=-1
\noindent\textbf{CALVIN}~\cite{mees2021calvin}: This synthetic dataset includes $24$ hours of unstructured play data collected via teleoperating a Franka Emika Panda robot arm to manipulate objects in four visually distinct 3D environments. CALVIN is specifically interesting for us as it allows us to train our predictive model in one environment and test its invariance to object appearance in another unseen environment. Moreover, we use the CALVIN environment to perform model-predictive control via the learned 3D world model and plan for action trajectories. Concretely, our train and validation sets are from the CALVIN environment \textit{Env C}, and our test set is from \textit{Env D}.\looseness=-1
\noindent\textbf{Omnipush}~\cite{bauza2019omnipush}: Although there has been a considerable effort in the robotics community to collect real-world robot interactions~\cite{finn2016unsupervised, dasari2019robonet}, a significant limitation is the lack of depth modality in these datasets. To the best of our knowledge, Omnipush is the only real-world dataset that provides RGB-D videos recorded via a static camera looking towards the workspace of a robot pushing differently shaped objects. Omnipush is a challenging dataset as it consists of noisy actions and observations and reflects the stochastic dynamics of the real world very well. We use the first dataset split, consisting of 70 objects without extra weight. We use 50 objects for the training set, 10 for validation, and 10 for test sets.\looseness=-1
\begin{figure}[t]
\centering
\includegraphics[width=0.75\columnwidth]{Images/t3vip_omni_corr_v3.pdf}
\caption{Spearman rank correlation across pairs of different budgets in epochs: The rank correlation for a pair of budgets can be read in the respective cell for a given row and column. While $r$ represents the correlation value and $p$ the p-value, $n$ is the number of configurations that were common to the given pair of budgets. For instance, the rank correlation between the pair of budgets $2$ and $200$ can be read in the cell corresponding to the first row and fourth column as $0.700$ with a p-value of $0.188$ and $5$ common configurations that were promoted from the budget of $2$ epochs to the budget of $200$ epochs. As can be seen, the correlations across different pairs of budgets are fairly high and positive, implying that multi-fidelity HPO is efficient at finding well-performing configurations in our experiments.\looseness=-1}
\label{fig:hpo_corr}
\vspace*{-8mm}
\end{figure}
\begin{figure*}[t!]
\centering
\includegraphics[scale=0.64]{Images/comparisons_v5.pdf}
\caption{Qualitative results of T3VIP compared with SV2P~\cite{babaeizadeh2018stochastic} on DexHand, CALVIN, and Omnipush datasets, respectively. T3VIP predicts long-range RGB-D frames and outputs sharp scene and optical flows and sparse binary occlusion masks as emergent properties. T3VIP deals better with occlusions and generates sharper RGB images than the baseline. Please use the color wheel \raisebox{-0.10cm}{\includegraphics[height=0.35cm]{Images/wheel.pdf}} to validate how our computed occlusion mask is compliant with our predicted flow field, e.\,g. when the robot goes to the right (reddish color), the occlusion mask indicates that the robot's left side has been occluded before and now is visible and should get inpainted. A flow field with several distinct colors shows that our model decomposes the scene into objects and reasons about their 3D motion separately.\looseness=-1}
\label{fig:qual}
\vspace*{-8mm}
\end{figure*}
\vspace*{-6mm}
\subsection{Hyperparameter Optimization}\label{sec:expt_hpo}
As mentioned in \secref{sec:approach_hpo}, we employ ASHA~\cite{li2020system} separately on each dataset to automatically find hyperparameter configurations that lead to high-quality RGB-D predictions on that dataset. Thus, we consider the sum of PSNR scores for both the predicted RGB images and depth maps as the optimization metric of ASHA. We observed that the hyperparameter configurations found on smaller budgets still perform well after training for the full budget, and this allowed us to perform HPO much cheaper than using the full budget for every training run. The analysis of the HPO for the experiment on the Omnipush dataset in \figref{fig:hpo_corr} shows that the Spearman rank correlation across the different pairs of budgets is quite high (often 0.7 and above).
This intuitively explains the performance transfer from the max budget given to the HPO to the full budget of the final evaluation. Across all three datasets, we observed that ASHA finds not only one but several configurations that result in learning the 3D scene dynamics and consequently accurate RGB-D video predictions.\looseness=-1
\subsection{Comparisons and Ablations}
In our experiments, we compare our stochastic RGB-D video prediction model to the following well-established RGB video prediction baselines:\looseness=-1
\noindent\textbf{CDNA}: Deterministic video prediction model proposed in~\cite{finn2016unsupervised} that predicts the motion of pixels and transforms them from previous images to construct future images. CDNA has a similar architecture as our forward dynamics model, with the difference that it predicts motion via 2D convolution kernels and our model explicitly predicts 3D rigid body transformations.\looseness=-1
\noindent\textbf{SNA}: Inspired by~\cite{ebert2017self}, we empower the CDNA baseline to address occlusions by adding a residual at each time step via a skip connection from the first image in the video sequence and an additional predicted mask.\looseness=-1
\noindent\textbf{SV2P}: Stochastic video prediction model proposed in~\cite{babaeizadeh2018stochastic} which employs the deterministic CDNA model to perform next frame generation, but this time conditioned on stochastic latent variables sampled from a prior distribution. In this comparison, we adapted SV2P to use SNA instead of CDNA to enable it to reason about the occlusions.\looseness=-1
We follow the training recipe of~\cite{babaeizadeh2018stochastic} and train all models for $200K$ steps. We condition all the models on the first two frames and train them to predict the next ten frames. Quantitative results of each model in predicting future frames conditioned on the robot's actions are reported in ~\tabref{tab:quan-eval}. Our T3VIP successfully makes long-range RGB-D predictions across all three datasets via
reasoning about the 3D motion, while the baselines cannot reason about the scene's geometry. Furthermore, our model significantly outperforms baselines on the DexHand dataset (where much self-occlusion is present) in the RGB metrics while achieving on-par performance on CALVIN and Omnipush datasets. Although SNA shows improved performance compared to CDNA in handling occlusions, our model analytically computes a binary occlusion mask based on the predicted 3D motion and has superior performance in handling the missing regions of the scene. We validated our superior occlusion handling also on Omnipush and CALVIN by always skipping two frames in the datasets between the frames used for training the models. This skipping causes more motion and consequently more occlusion between consecutive frames, and~\tabref{tab:quan-eval} shows that our model outperforms the baselines in such scenarios.\looseness=-1
Our qualitative evaluation (displayed in~\figref{fig:qual}) shows that T3VIP produces sharp flow fields and sparse occlusion masks. This indicates that our model learns to reason about the 3D dynamics of objects in the scene and generates the next frame primarily using the predicted 3D motion and only inpainting those sparse occluded regions. Besides predicting long-range future depth maps, our model also produces sharper RGB images than baselines. The main reason for this is that beyond the reconstruction loss, our model also encourages sharp scene flow and optical flow fields.\looseness=-1
\bigbreak
\begin{table}[b!]
\vspace*{-10mm}
\footnotesize
\centering
\setlength\tabcolsep{1.3pt}
\renewcommand{\arraystretch}{0.95}
\begin{tabular}{c|c|c|ccc|cc}
\toprule
\multirow{2}{*}{\centering \textbf{Dataset}} & \textbf{Test} & \multirow{2}{*}{\centering \textbf{Model}} & \multicolumn{3}{c|}{\cellcolor{red!25}\textbf{RGB}} & \multicolumn{2}{c}{\cellcolor{blue!25}\textbf{Depth}}\\
& \textbf{Steps} & & \cellcolor{red!25}\textbf{PSNR}$\uparrow$ & \cellcolor{red!25}\textbf{SSIM}$\uparrow$ & \cellcolor{red!25}\textbf{VGG}$\uparrow$ & \cellcolor{blue!25}\textbf{\space RMSE}$\downarrow$ & \cellcolor{blue!25}\textbf{\space AbsRel}$\downarrow$\\
\midrule
\parbox[t]{2mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{\parbox{1.0cm}{\centering DexHand}}}} & \multirow{4}{*}{\centering 25} &
CDNA~\cite{finn2016unsupervised} & $20.254$ & $0.803$ & $0.825$ & $\times$ & $\times$\\
& & SNA~\cite{ebert2017self} & $20.643$ & $0.814$ & $0.836$ & $\times$ & $\times$\\
& & SV2P~\cite{babaeizadeh2018stochastic} & $20.892$ & $0.810$ & $0.836$ & $\times$ & $\times$\\
& & T3VIP (Ours) & $\textbf{23.153}$ & $\textbf{0.889}$ & $\textbf{0.909}$ & $\textbf{0.024}$ & $\textbf{0.005}$\\
\midrule
\midrule
\parbox[t]{2mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{\parbox{1.2cm}{\centering CALVIN $skip=0$}}}}& \multirow{4}{*}{\centering 60} &
CDNA~\cite{finn2016unsupervised} & $22.374$ & $0.847$ & $0.875$ & $\times$ & $\times$\\
& & SNA~\cite{ebert2017self} & $22.279$ & $0.842$ & $0.881$ & $\times$ & $\times$\\
& & SV2P~\cite{babaeizadeh2018stochastic} & $22.363$ & $0.844$ & $\textbf{0.882}$ & $\times$ & $\times$\\
& & T3VIP (Ours) & $\textbf{22.692}$ & $\textbf{0.849}$ & $0.857$ & $\textbf{0.133}$ & $\textbf{0.006}$\\
\midrule
\midrule
\parbox[t]{2mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{\parbox{1.2cm}{\centering CALVIN $skip=2$}}}}& \multirow{4}{*}{\centering 60} &
CDNA~\cite{finn2016unsupervised} & $19.397$ & $0.740$ & $0.779$ & $\times$ & $\times$\\
& & SNA~\cite{ebert2017self} & $19.403$ & $0.739$ & $0.787$ & $\times$ & $\times$\\
& & SV2P~\cite{babaeizadeh2018stochastic} & $19.408$ & $0.741$ & $0.791$ & $\times$ & $\times$\\
& & T3VIP (Ours) & $\textbf{21.208}$ & $\textbf{0.779}$ & $\textbf{0.801}$ & $\textbf{0.125}$ & $\textbf{0.007}$\\
\midrule
\midrule
\parbox[t]{2mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{\parbox{1.2cm}{\centering Omnipush $skip=0$}}}} & \multirow{4}{*}{\centering 60} &
CDNA~\cite{finn2016unsupervised} & $\textbf{25.239}$ & $0.877$ & $0.848$ & $\times$ & $\times$\\
& & SNA~\cite{ebert2017self} & $23.855$ & $0.880$ & $0.871$ & $\times$ & $\times$\\
& & SV2P~\cite{babaeizadeh2018stochastic} & $23.853$ & $\textbf{0.884}$ & $0.874$ & $\times$ & $\times$\\
& & T3VIP (Ours) & $24.261$ & $0.879$ & $\textbf{0.893}$ & $\textbf{0.053}$ & $\textbf{0.023}$\\
\midrule
\midrule
\parbox[t]{2mm}{\multirow{4}{*}{\rotatebox[origin=c]{90}{\parbox{1.2cm}{\centering Omnipush $skip=2$}}}} & \multirow{4}{*}{\centering 20} &
CDNA~\cite{finn2016unsupervised} & $25.304$ & $0.890$ & $0.865$ & $\times$ & $\times$\\
& & SNA~\cite{ebert2017self} & $25.452$ & $0.903$ & $0.891$ & $\times$ & $\times$\\
& & SV2P~\cite{babaeizadeh2018stochastic} & $25.730$ & $\textbf{0.907}$ & $0.891$ & $\times$ & $\times$\\
& & T3VIP (Ours) & $\textbf{26.053}$ & $0.899$ & $\textbf{0.911}$ & $\textbf{0.050}$ & $\textbf{0.020}$\\
\bottomrule
\end{tabular}
\caption{Comparison between T3VIP and 2D baselines on average scores over the sequence of predicted videos. Please note that although all models are trained to predict the next ten time steps, we test them for longer horizons.}
\label{tab:quan-eval}
\end{table}
\begin{figure*}[t]
\centering
\subfloat[Evaluation of a 3D servoing experiment conducted via a 2D (SV2P) and a 3D world model (T3VIP)]{
\includegraphics[scale=0.67]{Images/planning_v2.pdf}\label{fig:plan_scenario}}
\subfloat[Overall success rate on 2D and 3D domains]{
\includegraphics[scale=0.67]{Images/plan_total.pdf}\label{fig:overall_success}}
\caption{(a) The first row visualizes the robot's viewpoint, whereas the second row visualizes an alternative view (the robot does not have access to this view). We observe that while SV2P reaches the pixel coordinates of the target, it misses the target 3D point considerably. However, T3VIP which is 3D-aware successfully reaches the target and solves the task in both 2D and 3D domains. (b) T3VIP significantly outperforms SV2P in 3D servoing experiments.}
\label{fig:planning}
\vspace{-0.8cm}
\end{figure*}
To analyze the influence of our different loss functions on the learned 3D dynamics and quality of RGB-D video prediction, we conducted an ablation experiment on the Omnipush dataset (see~\tabref{tab:ablation}). Our results indicate that a model (M1) that only aims to reduce the reconstruction loss is not effective in learning the 3D dynamics of the scene. Furthermore, a model that additionally regularizes the scene flow to be smooth (M2) was unsuccessful in reasoning about the scene dynamics. Although the third variant of our model (M3) that leverages point cloud alignment loss to enforce geometric consistency learns the dynamics of the scene, it leads to non-smooth flow fields. This is expected because the nearest-neighbor data association is prone to spurious matches, especially with noisy depth measurements. We found that variants that utilize 3D point cloud alignment and regularize for scene flow (M4) and both optical and scene flow fields (M5) lead to better reasoning about the dynamics and sharper predicted frames. Finally, our final model (M6) that also reasons about the stochasticity and minimizes the KL divergence between the approximated posterior and the assumed prior leads to the best RGB-D prediction performance.\looseness=-1
We also evaluated our model performance on Omnipush dataset without conditioning on actions, see~\figref{fig:act_free}. We observed that although the performance degrades compared to the action-conditioned setting, our model can still reasonably predict the future RGB-D frames. More specifically, we demonstrate that our full model, which captures stochasticity, outperforms a deterministic version of our model (T3VIP-D).\looseness=-1
\begin{figure}[h]
\vspace*{-3mm}
\centering
\includegraphics[width=0.65\columnwidth]{Images/act_free_v2.eps}
\caption{Comparison between T3VIP and baselines on action-free prediction of frames on Omnipush dataset. T3VIP-D is a deterministic variant of our model.\looseness=-1}
\label{fig:act_free}
\vspace*{-5mm}
\end{figure}
\begin{table}[t]
\vspace*{2mm}
\centering
\footnotesize
\setlength\tabcolsep{0.05pt}
\begin{tabular}{ccccccc|ccc|cc}
\toprule
\centering \textbf{Model} & \textbf{$\mathcal{L}_{rec}^{I}$} & \textbf{$\mathcal{L}_{rec}^{D}$} & \textbf{$\mathcal{L}_{knn}$} & \textbf{$\mathcal{L}_{fs}^{s}$} & \textbf{$\mathcal{L}_{fs}^{o}$} & \textbf{$\mathcal{L}_{kl}$} & \cellcolor{red!25}\textbf{PSNR}$\uparrow$ & \cellcolor{red!25}\textbf{SSIM}$\uparrow$ & \cellcolor{red!25}\textbf{VGG}$\uparrow$ & \cellcolor{blue!25}\textbf{RMSE}$\downarrow$ &
\cellcolor{blue!25}\textbf{AbsRel}$\downarrow$\\
\midrule
M1 & \checkmark & \checkmark & - & - & - & - & $\times$ & $\times$ & $\times$ & $\times$ & $\times$\\
M2 & \checkmark & \checkmark & - & \checkmark & - & - & $\times$ & $\times$ & $\times$ & $\times$ & $\times$\\
M3 & \checkmark & \checkmark & \checkmark & - & - & - & $24.92$ & $0.88$ & $0.86$ & $0.06$ & $0.02$\\
M4 & \checkmark & \checkmark & \checkmark & \checkmark & - & - & $25.29$ & $0.88$ & $0.89$ & $0.05$ & $0.02$\\
M5 & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & - & $25.53$ & $0.89$ & $0.90$ & $0.05$ & $0.02$\\
M6 & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & \checkmark & $\textbf{26.05}$ & $\textbf{0.90}$ & $\textbf{0.91}$ & $\textbf{0.05}$ & $\textbf{0.02}$\\
\bottomrule
\end{tabular}
\caption{Ablation study: Our full model leverages all observational cues to reason about the 3D dynamics of the scene and predict future RGB-D frames.\looseness=-1}
\label{tab:ablation}
\vspace*{-2mm}
\end{table}
\subsection{Model-Predictive Control}
To evaluate the effectiveness of our 3D-aware world model and compare it with a 2D world model (SV2P), we conduct a 3D servoing experiment in the CALVIN environment. The goal of this experiment is for the robot arm to reach the user-defined goal point as closely as possible. To this end, we employ iCEM~\cite{pinneri2020sample}, which utilizes the learned world model to search for the best action trajectory that leads to the desired goal point. Please note that although the goal point is in 3D, a 2D world model is unaware of the target's depth measurement and can only comprehend the corresponding target pixel. Our model uses its predicted transformed point cloud to track the progress towards the target point, whereas SV2P uses its predicted 2D convolutional kernels to track the progress towards the target pixel. We use a discounted sum of trajectory distances to the target (point distance for T3VIP and pixel distance for SV2P) as the cost function. To conduct an extensive evaluation of the performances of T3VIP versus SV2P, we created a series of 10 experiment setups, where each setup consists of a random configuration of the scene and robot arm, a random goal point in 3D, and its corresponding pixel coordinates. We executed each experiment setup five times for each model with a fixed budget of 100 steps. As the distance of the robot end-effector to the goal point varies for each experiment setup, we consider an experiment successful if the end-effector reaches within 10 percent of the distance from the goal and meets this condition at least five times within an experiment. We observe that although SV2P can successfully reach the pixel coordinates of the goal point, it often misses reaching the user-defined 3D target (see~\figref{fig:plan_scenario}), as it suffers from the inherent ambiguity of 2D vision in capturing depth. In contrast, our model comprehends the goal in 3D, effectively predicts the 3D dynamics of the robot's end-effector, and successfully reaches the target point. \figref{fig:overall_success} shows the overall success of both models in reaching the target in 2D and 3D domains across all the experiments. Videos of these experiments are available at \url{http://t3vip.cs.uni-freiburg.de}.\looseness=-1
\section{Video Prediction}
\subsection{Network Details}
\subsection{Additional Experiments}
\section{Hyperparameter optimization}\label{sec:append_hpo}
\subsection{Training Details}
report found hyperparameters
\section{Planning}
\subsection{Model-Predictive Control}
write down the planning algorithm
\begin{table}[htbp]
\centering
\catcode`,=\active
\def,{\char`,\allowbreak}
\renewcommand\arraystretch{1.0}
\begin{tabular}{p{3.5cm}<{\raggedright} p{1.5cm}<{\raggedleft} }
\toprule
\textbf{Hyperparameters} & \textbf{Values}\\
\midrule
population & 150 \\
minimum std & 0.001 \\
maximum std & 1.0 \\
elite fraction & 0.1 \\
horizon & 5 \\
maximum iterations & 5 \\
internal momentum (alpha) & 0.1 \\
external momentum (beta) & 0.5 \\
colored noise exponent & 2 \\
population decay factor & 0.9 \\
cost function decay factor & 0.8 \\
\bottomrule
\end{tabular}
\caption{iCEM hyperparameters}
\label{tab:planner_hyper}
\end{table}
\section{0pt}{12pt plus 4pt minus 2pt}{0pt plus 2pt minus 2pt}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{amsfonts}
\usepackage{bm}
\usepackage{siunitx}
\sisetup{detect-weight=true, detect-family=true}
\usepackage{cancel}
\usepackage[inline]{enumitem}
\DeclareMathOperator*{\argmax}{argmax}
\usepackage{microtype}
\input{commands}
\usepackage[hidelinks]{hyperref}
\usepackage{xcolor}
\hypersetup{
colorlinks,
linkcolor={red!50!black},
citecolor={blue!50!black},
urlcolor={blue!80!black}
}
\usepackage{cleveref}
\usepackage[table]{xcolor}
\title{\LARGE \bf T3VIP: Transformation-based 3D Video Prediction}
\author{Iman Nematollahi\textsuperscript{1}, Erick Rosete-Beas\textsuperscript{1}, Seyed Mahdi B. Azad\textsuperscript{1}, Raghu Rajan\textsuperscript{1}, \\ Frank Hutter\textsuperscript{1,2}, Wolfram Burgard\textsuperscript{3
\thanks{\textsuperscript{1}University of Freiburg. \textsuperscript{2}Bosch Center for AI. \textsuperscript{3}University of Technology Nuremberg. This work has been supported by German Federal Ministry of Education and Research under contract number 01IS18040B-OML and also by the Deutsche Forschungsgemeinschaft under grant number 417962828.\looseness=-1}
}
\begin{document}
\maketitle
\thispagestyle{empty}
\pagestyle{empty}
\begin{abstract}
For autonomous skill acquisition, robots have to
learn about the physical rules governing the 3D world dynamics
from their own past experience to predict and reason
about plausible future outcomes. To this end, we propose a transformation-based 3D video prediction (T3VIP) approach that explicitly models the 3D motion by decomposing a scene into its object parts and predicting their corresponding rigid transformations. Our model is fully unsupervised, captures the stochastic nature of the real world, and the observational cues in image and point cloud domains constitute its learning signals. To fully leverage all the 2D and 3D observational signals, we equip our model with automatic hyperparameter optimization (HPO) to interpret the best way of learning from them. To the best of our knowledge, our model is the first generative model that provides an RGB-D video prediction of the future for a static camera. Our extensive evaluation with simulated and real-world datasets demonstrates that our formulation leads to interpretable 3D models that predict future depth videos while achieving on-par performance with 2D models on RGB video prediction. Moreover, we demonstrate that our model outperforms 2D baselines on visuomotor control. Videos, code, dataset, and pre-trained models are available at \url{http://t3vip.cs.uni-freiburg.de}.\looseness=-1
\end{abstract}
\section{Introduction}\label{sec:intro}
\input{Sections/1_Introduction}
\section{Related Work}\label{sec:related}
\input{Sections/2_RelatedWork}
\section{Problem Formulation}\label{sec:problem}
\input{Sections/3_ProblemFormulation}
\section{T3VIP }\label{sec:approach}
\input{Sections/4_Approach}\label{sec:approach}
\section{Experimental Evaluation}\label{sec:experiments}
\input{Sections/5_Experiments}
\newpage
\section{Conclusions}\label{sec:conclusion}
\input{Sections/6_Conclusions}
\footnotesize
\bibliographystyle{IEEEtran}
|
\section{Problem Statement}
The use of machine learning (ML) in safety-critical applications has moved beyond autonomous driving and entered new domains such as avionics. Although the state-of-the-art paradigm considers the safety engineering of ML to be a continuous engineering process (cf. ISO 21448~\cite{SOTIF} appendix D), one of the core challenges is to accelerate the continuous engineering paradigm. This can be reflected in our use case of integrating ML-based vision for an automated guided vehicle (AGV) within the domain of factory automation. Each factory can have different environmental conditions such as the width of the corridor, and the integrated mobility platforms may have different resource consumption constraints and operational speed. Altogether, when the ML component is safe to use on one shop floor with a particular robot configuration, the transition from one shop floor to another may imply that the established safety argument is no longer valid.
\section{Automating Continuous ML-Safety Assurance}
We present a workflow on how to extend from standard ML-Ops (i.e., set of practices that aims to deploy and maintain ML models in production) with an explicit connection to the assurance case.
\paragraph{Baseline} The following list of information can be considered as a basis required for the ML to be used in a safety-critical application, i.e., the basic snapshot to be demonstrated to the certification authority.
\begin{enumerate}
\item The current ML system (including the pre- and post-processor) being deployed.
\item The safety case (in the form of Goal-Structuring-Notation - GSN~\cite{kelly2004goal}) associated with the ML on how one argues that the ML engineering process is rigorous and the ML output insufficiency is not a source of creditable harm.
\item For each evidence listed in the safety case, the associated document or results produced by 3rd party tools demonstrating that the evidence is fulfilled.
\end{enumerate}
\paragraph{Discovering potential change requests} The first step of the workflow is to know when a change will occur. Newly acquired information such as publicly available incident reports, context evolution (change of factory setup or regulations), or onboard monitoring (e.g., near-collision or out-of-distribution detection) can be used as sources to start the change analysis.
\vspace{1mm}
\paragraph{Automation via formal specification, semantic tags, and evidence evaluation}
When a potential change is uncovered, its impact on the safety case must be assessed. Recall that a GSN maintains a tree-like structure specifying why the deployed ML is safe. Edges are introduced to connect a particular goal to the underlying supporting evidences or decomposed sub-goals. Therefore, understanding the change impact can be reduced to examining the validity of each evidence or assumption located at the leaf of the GSN tree.
To automate the process, we try to formalize each evidence and assumption using mathematical logic or measurable metrics, such that one can precisely evaluate if a particular piece of evidence, under the change, has turned invalid / false. As an example, the original GSN may contain a claim stating that \emph{the false-positive is only a performance issue rather than a safety issue}. We sharpen the supporting argument using mathematical logic (here with substantial simplification) and reach the following statement:
\vspace{-2mm}
\begin{equation}\label{eq:FPisfine}
\begin{split}
\forall t: d_{AGV, agent_{rear}}(t) \geq d_{B, agent_{rear}} \\ \wedge \; FP_{ML}(t, img) \wedge detected_{fusion}(t)\\ \rightarrow (\forall t' \in [t, t+ t_{B, AGV}]: d_{AGV,agent_{rear}}(t') > 0)
\end{split}
\end{equation}
\vspace{-2mm}
\begin{table*}[t]
\centering
\caption{Predicates and the associated semantic tags for the translated formula}\label{table:predicate}
\vspace{-2mm}
\begin{tabular}{|l|l|l|}\hline
\textbf{Predicate / Constant} & \textbf{Explaination} & \textbf{Semantic Tags} \\\hline
$d_{AGV, agent_{rear}}(t)$ & Distance between AGV and the rear agent at time $t$ & distance, AGV, rear-agent \\\hline
$d_{B, agent_{rear}}$ & Maximum braking distance for the rear agent & rear-agent, braking distance \\\hline
$FP_{ML}(t, img)$ & With input $img$ at time $t$, the ML has produced a false positive & false positive, detection \\\hline
$detected_{fusion}(t)$ & Sensor fusion indicates that an object is in the front and close-by & fusion, detection \\\hline
$t_{B, AGV}$ & Maximum required braking time for the AGV & AGV, braking time \\\hline
\end{tabular}
\vspace{-5mm}
\end{table*}
The explanation of each predicate or constant is shown in Table~\ref{table:predicate}. Intuitively, the formula states an assumption where the rear agent should keep a safe distance, such that when the ML component detects a non-existent object and the AGV triggers a stop, it is impossible for the AGV to be hit from behind.
While the formal specification can assist in understanding the impact over each evidence, the overall activity can also be challenging when one needs to examine each evidence in the complex argumentation tree. We therefore associate each evidence or assumption with a set of \emph{\textbf{semantic tags}} (cf. Table~\ref{table:predicate} for all semantic tags used in Formula~\ref{eq:FPisfine}), store all relevant information using a database, and utilize queries to automatically highlight evidences to be considered.
Finally, by concretizing evidences into mathematical formula, it is also possible to \emph{\textbf{directly evaluate the validity of the formula under change}}. As an example, the braking distance of the AGV is associated with the initial speed of the AGV, which can be formulated using Newton's laws. As safety can be characterized using the travelled distance of the AGV and the agent until the AGV reaches a full stop (similar concept available at responsibility-sensitive safety~\cite{shalev2017formal}), one can derive\footnote{Simple automatic dynamic derivation can be implemented via \texttt{eval()} in python, while complex derivation requires automatic theorem provers.} that increasing the detection frame rate (which is a constant stored in the argumentation) only reduces the travel distance of the AGV and safety is thus not compensated.
\paragraph{Incorporating changes}
To incorporate changes and to again make the ML-system arguably safe, we consider a three stage process with increased difficulty.
\begin{enumerate}
\item The simplest case is related to the situation where, by analyzing the safety case, each listed evidence remains valid. One can directly change the parameters stored in the safety case and close the task\footnote{As there can be potential gaps in the argumentation, the impact analysis may be further complemented by re-doing (part of) the tests conducted in the evidence construction process.}.
\item Subsequently, consider if it is possible to perform improvement (e.g., collect more data) such that each evidence listed in the safety case turns valid again. Tools with intelligent test case generation capabilities (e.g.,~\cite{cheng2019nn}) can increase the degree of automation.
\item Finally, situations exist where it is impossible to incorporate the change by keeping the same assurance case structure\footnote{As an example, when one changes the ML function from object detection to semantic segmentation, as the set of ML output insufficiencies are different the safety argumentation needs to be reworked.}. When the assurance case
structure needs substantial change, human-in-the-loop
modification may be unavoidable.
\end{enumerate}
\vspace{-1mm}
\section{Concept Validation}
\vspace{-1mm}
We have built a research prototype
with the help of multiple open-source tools. We used D3.js
for building a front-end visualizing the complete GSN tree (Figure~\ref{figure:gsn} shows an example from~\cite{cheng2019nn} where one piece of evidence has turned invalid), where the tree and all evidences are dynamically linked to a NoSQL database.
Apart from formally specified formulae, we also connected the evidence with the ML testing tool \emph{nn-dependability-kit}~\cite{cheng2019nn} which can provide various types of metrics such as data completeness.
\begin{figure}[t]
\centering
\includegraphics[width=0.95\columnwidth]{fig/GSN_3_modified.png}
\vspace{-4mm}
\caption{Validating the concept with an assist of a front-end, where evidence (circle at the bottom left) being impacted is highlighted by the tool in red. }
\label{figure:gsn}
\vspace{-5mm}
\end{figure}
The database allows easy query for semantic tags to identify nodes in the argumentation tree that a change may impact. Additionally, it allows to flexibly update the status (true or false) of an evidence, either automatically (via the automatic evaluation of an embedded formula), with an external connection to the database (3rd party tool used to demonstrate an evidence is fulfilled), or manually when required.
\section{Concluding Remarks}
Our proposed methodology aims at the efficient iteration of the ML-Ops where safety argumentation is an integral part of the engineering. The methodology is particularly suitable in the initial prototyping phase, where safety should be part of the engineering process, but the system architecture can continuously evolve. Realizing the concept into tools, the result also facilitates communication between ML engineers and safety analysts working on seemly unrelated domains. It also leads to various improvement potentials, such as using natural language processing to reduce human efforts further.
\bibliographystyle{abbrv}
|
\subsection*{Exciton generation in TMDs}
Transition metal dichalcogenides (TMDs) have attracted attention due to their interesting electronic and optical properties. In monolayer form, they exhibit a direct bandgap \cite{Mak2010}, responsible for radiative electron-hole recombination \cite{Splendiani2010}. TMDs support excitons with high binding energies due to quantum confinement \cite{Chernikov2014}, making their optical properties to be dominantly defined by their excitonic states \cite{Jung2018, Hsu2019, Ermolaev2020}. Exciton generation in TMDs was first studied through optical excitation \cite{Mak2010,Kozawa2014,Splendiani2010,Lien2019,Amani2015,Kim2021}. Later, combination with conductors (i.e. graphene) and insulators (i.e. hexagonal boron nitride) \cite{Novoselov2005,Geim2013} allowed the design and fabrication of light-emitting van der Waals (vdW) heterostructures with TMDs as the optically active material \cite{Withers2015,Paur2019,PenaRoman2020,Binder2017,Palacios-Berraquero2016,Jauregui2019,Sundaram2013,Ross2014,Uddin2022,Withers2015a,Fu2021,Binder2019,Clark2016, Lopez2022,Roman2022,Feng2022}. In these studies, excitons are electrically generated through direct charge injection \cite{Withers2015,Paur2019,PenaRoman2020,Binder2017, Palacios-Berraquero2016,Jauregui2019,Sundaram2013,Ross2014,Uddin2022,Withers2015a,Fu2021,Binder2019,Clark2016,Lopez2022,Roman2022} or through charge impact at high alternating voltages \cite{Feng2022}. Here, we introduce a different way to generate excitons in TMD based tunneling devices. We use the rich platform of 2D materials to demonstrate excitonic light emission from tunnel junctions where the TMD is located outside of the current pathway.
The studied device structure is illustrated in Fig.~\ref{fig1}A. The TMD and the tunnel junction are electrically decoupled, preventing exciton generation through direct electron-hole injection. The emission spectrum of such a device features a distinctive peak at the exciton energy, as shown in Fig.~\ref{fig1}B. We attribute the generation of excitons to energy transfer (ET) from tunneling electrons. ET from a donor to an acceptor is mediated by dipole-dipole coupling \cite{Forster1964,Novotny2012}. It is an electromagnetic effect that is dominant at the near-field region of an emitter and has been extensively studied in biological systems \cite{Weiss1999}, molecular assemblies \cite{Dale1976}, solid-state quantum dots \cite{Kagan1996} and photosynthetic membranes \cite{van2000photosynthetic}. It has been also used for color conversion and broadband source design as an energy exchange channel between different fluorophores \cite{Achermann2006,Pimputkar2009}. Recent STM studies discuss the possibility of ET in tunneling systems \cite{Lopez2022,Roman2022} without however providing any conclusive observation. Here we make use of atomically thin 2D materials and the flat interfaces between them to preserve optical coupling in near-field distances between a TMD and a tunnel junction. Our results reveal the role of ET between electrons and excitons in tunneling-based light emitting devices.
\begin{figure*}
\centering
\includegraphics[width=0.7\textwidth]{figure1.pdf}
\caption{\textbf{Exciton generation in TMD-coupled tunnel junctions.} (a) Device composed of a WSe$_2$-coupled graphene-hexagonal boron nitride (hBN)-gold tunnel junction. The WSe$_2$ layer is stacked on top of the junction. When a bias voltage $V\mathrm{_{b}}$~ is applied between graphene and gold, excitons are generated in WSe$_2$. (b) Measured spectrum at room temperature of the electrically generated emission for $V\mathrm{_{b}}$~=~\SI{2}{V}. The peak position matches the WSe$_2$ exciton energy and the spectral width matches the thermal energy $k\mathrm{_BT}$.}
\label{fig1}
\end{figure*}
\subsection*{TMD-coupled tunnel junctions}
Our devices are fabricated by stacking mechanically exfoliated flakes of 2D materials with a dry pickup and transfer technique (see supplementary materials for details). The resulting tunnel junctions consist of a graphene and a gold electrode separated by a hexagonal boron nitride layer (hBN) of \SI{2.3}{nm} thickness acting as a tunneling barrier. The TMD is stacked on top of the junction and optionally separated from the graphene electrode by a second hBN layer.
Figure~\ref{fig2}A shows a false-color microscope image of the device with an outline of the different flakes. The device area is indicated by a white dashed box. The tunnel junction area is defined by the overlap of the gold electrode and the graphene flake. The WSe$_2$ flake is only partially covering the tunneling area. This creates two different tunneling regions, with and without WSe$_2$ on top. We electrically and optically characterize the device by applying a bias voltage ($V\mathrm{_{b}}$) between the gold and graphene electrodes, as indicated in Fig.~\ref{fig2}A, and we collect the emitted light with a high-numerical aperture objective (see supplementary materials). The current and conductance measurements suggest that tunneling occurs between graphene and gold (see Fig. S1A). The spectra emitted from the whole tunneling area for various $V\mathrm{_{b}}$~values are presented in Fig.~\ref{fig2}B. There are two distinct contributions in the spectra. A broad emission at lower energies and a narrow emission peak at \SI{1.63}{eV} (\SI{762}{nm}). The broad emission shifts to higher photon energies as $V\mathrm{_{b}}$~is increased. This behavior suggests that the emission originates from inelastic electron tunneling that couples to radiative modes of the free-space continuum (\textit{photon-coupling}) with e$V\mathrm{_{b}}$~being the cut-off energy of emission \cite{Parzefall2015,Parzefall2019}. The narrow peak at \SI{1.63}{eV} corresponds to the neutral 1s exciton of WSe$_2$ and its narrow width of around \SI{25}{meV} can be attributed to graphene filtering \cite{Lorchat2020}. It has been shown that charge transfer between graphene and TMDs, filters the TMD photoluminescence spectra from any charged exciton contributions \cite{Lorchat2020}. To further our understanding, we study the emission distribution for different spectral regions. Real space images of the photoemission are presented in Fig.~\ref{fig2}C, where we observe that the area covered by WSe$_2$ dominates the emission and in Fig.~\ref{fig2}D, where the emission is filtered by a band-pass filter at \SI{900}{nm}. This is a spectral region where the excitonic contribution is much reduced and the emission extends over both device areas. This indicates that the broad low energy contribution indeed originates from \textit{photon-coupling} and is not associated with the presence of WSe$_2$. To understand the angular emission of the device we record Fourier-space images. Fig.~\ref{fig2}E shows the Fourier space image of the full emission spectrum whereas Fig.~\ref{fig2}F depicts the Fourier space image of the bandpass filtered light around \SI{900}{nm}. For the spectrally filtered measurement at 900 nm, the emission vanishes at the center of the Fourier space ($k_x=k_y=0$) (Fig.~\ref{fig2}F) indicative for an out-of-plane dipole orientation. This is characteristic of inelastic electron tunneling where the transition dipole is oriented along the electron path. On the other hand, when measuring the full spectrum, the Fourier space image (Fig.~\ref{fig2}E) is not zero at $k_x=k_y=0$, which indicates an in-plane contribution. This extra contribution is associated with the excitonic emission of WSe$_2$, which is known to originate from in-plane transitions \cite{Schuller2013}. We emphasize that in our devices any exciton generation due to electron-hole injection is avoided due to the WSe$_2$ being outside of the tunneling pathway. This is further supported by the fact that very similar emission spectra are observed for negative $V\mathrm{_{b}}$~(see Fig. S2). For these reasons we conclude that excitons are generated in the TMD by ET from tunneling electrons. \\
\indent
We continue to investigate this exciton generation process by considering the band-structure diagram shown in Fig.~\ref{fig3}A. Upon application of $V\mathrm{_{b}}$, electrons tunnel from one electrode to the other, elastically or inelastically. In the latter case, the transition dipole associated with the energy loss $\Delta E$ couples to available optical modes of the environment. Some of these modes are radiative (photons) as in the case of the broad emission observed in Fig.~\ref{fig2}A, or non-radiative, as in the case of ET. This ET mechanism generates excitons in WSe$_2$ and their spontaneous decay contributes to the narrow emission peak in the spectra of Fig. \ref{fig2}B. Thus, the excitonic emission observed in our device is a two step process, in which the transition energy is first transferred to the WSe$_2$ exciton and then spontaneously emitted (\textit{exciton-coupling}). \\
\indent
Next, we compare the ET efficiency to the \textit{photon-coupling} efficiency. To do that we first break down the processes to the involved interactions (see pictorial representation in Fig.~\ref{fig3}B) and we assign conversion efficiencies. Both \textit{photon-} and \textit{ exciton-coupling} processes describe an electron-to-photon ($e\mhyphen\gamma$) conversion. \textit{Photon-coupling} is a first order process for which we define an $e\mhyphen\gamma$ conversion efficiency as $\eta\mathrm{_{e\mhyphen\gamma}}$. \textit{Exciton-coupling} is a second order process in which we can assign a combined efficiency $\eta^{\prime}\mathrm{_{e\mhyphen\gamma}} = \eta\mathrm{_{e\mhyphen x}}\cdot\eta\mathrm{_{x\mhyphen\gamma}}$, where $\eta\mathrm{_{e\mhyphen x}}$ is the ET efficiency and $\eta\mathrm{_{x\mhyphen\gamma}}$ is the exciton to photon conversion efficiency. From the spectra shown in Fig.~\ref{fig2}B, we can infer a value of 4.3 for the ratio between the photon-coupled emission and the exciton-coupled emission. By taking into account the values for WSe$_2$ PL efficiency and the graphene quenching (see supplementary materials) we arrive at an estimation of the ratio $\frac{\eta_{e\mhyphen x}}{\eta_{e\mhyphen\gamma}} \cong 10^4$. This result indicates that ET is orders of magnitude more efficient than direct photon emission. In order to shed light on this surprising finding we analyze in the next section the density of optical states near a monolayer TMD.
\begin{figure*}
\centering
\includegraphics[width=0.7\textwidth]{figure2.pdf}
\caption{\textbf{ WSe$_2$-coupled tunnel junction.} (a) Optical microscope image in false-color of a fabricated WSe$_2$-Graphene-hBN-gold device. The different flakes are outlined. The encapsulating~hBN~flake is not shown for clarity. The dashed frame indicates the device area depicted in (c) and (d). (b) Recorded emission spectra from the whole device area. Contributions from radiative mode coupling (\textit{photon-coupling}) and non-radiative mode coupling (\textit{exciton-coupling}) are observed. Real space (c,d) and Fourier space (e,f) images of the emission from the dashed box in (a) for $V\mathrm{_{b}}$~=~\SI{2}{V}. The full emission spectrum is used in (c) and (e) whereas a bandpass filter with center wavelength 900 nm and 40 nm FWHM is used in (d) and (f). Only half of every Fourier space is imaged for ease of comparison. The white solid lines are the rotational averages of the two Fourier spaces in (e) and (f).}
\label{fig2}
\end{figure*}
\subsection*{Optical density of states near monolayer TMDs}
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{figure3.pdf}
\caption{ \textbf{Photon and exciton coupling of inelastic tunneling electrons.} (a) Band diagram of the TMD-coupled tunnel junction. A voltage $V\mathrm{_{b}}$~ is applied across the hBN~barrier between graphene and gold electrodes. Part of the electrons tunnel inelastically. They either couple to radiative modes (1) or couple non-radiatively to the TMD via ET (2). The generated excitons decay radiatively making the ET process detectable in the farfield. (b) Pictorial representation of the physical processes involved in the light emission. In (1) electron-hole recombination is followed by photon emission. In (2) ET is depicted as a virtual photon $\gamma_v$ that couples to the TMD and creates an electron-hole pair (x). (c) Angular spectral density of~$\mathrm{\rho_{opt}}$~for the case of an out-of-plane dipole placed \SI{2}{\nano \meter} away from a monolayer WSe$_2$. Optical properties are taken from \cite{Jung2018}. The contribution of non-radiative modes ($\mathrm{\rho_{nr}}$) is substantially higher than the contribution of radiative modes ($\mathrm{\rho_{r}}$). The calculation assumes a photon energy of \SI{1.68}{eV}, matching the exciton energy of \cite{Jung2018}. The x-axis is normalized by the vacuum wavenumber $k_0$ and the y-axis is normalized by $\rho_0/k_0$, where $\rho_0$ is the vacuum LDOS. Inset in (c) illustrates the calculated system.}
\label{fig3}
\end{figure*}
When a dipole interacts with an absorbing material at a distance much smaller than the wavelength, the largest fraction of the dissipated power is associated with non-radiative energy transfer (ET) from the dipole to the material~\cite{Ford1984,Andrews2004}. The rate of this process is related to the local density of optical states (LDOS)~$\mathrm{\rho_{opt}}$. In Fig. \ref{fig3}C we show the calculated angular spectral density of the LDOS, $\frac{\mathrm{d \rho\mathrm{_{opt}}} }{\mathrm{d} k\mathrm{_\parallel}}$, as a function of $k_\parallel$ for the simple case of a dipole at \SI{2}{nm} distance from a TMD monolayer. Here, $k_\parallel$ is the wavevector component along the plane of the TMD monolayer (see inset in Fig. \ref{fig3}C) and $k\mathrm{_0}$ is the wavenumber for a photon energy equal to the exciton energy of WSe$_2$. The calculated~$\mathrm{\rho_{opt}}$~for $k_\parallel < k_0$ includes radiative modes associated with photon emission ($\mathrm{\rho_{r}}$) whereas the region $k\mathrm{_\parallel} > k\mathrm{_0}$ accounts for non-radiative modes ($\mathrm{\rho_{nr}}$) associated with ET. $\mathrm{\rho_{nr}}$~is orders of magnitude higher than $\mathrm{\rho_{r}}$. Moreover, the influence of the TMD on $\mathrm{\rho_{r}}$~ enhancement is negligible. Hence, the TMD introduces mainly additional non-radiative decay channels that are associated with near-field ET ($\rho\mathrm{_{opt}} \cong \rho\mathrm{_{nr}}$), and is the reason for the high $\frac{\eta\mathrm{_{e\mhyphen x}}}{\eta\mathrm{_{e\mhyphen\gamma}}}$ ratio. The ET process in our measurements is highly efficient and shows that near-field interactions at nanoscale distances are dominating the LDOS. Here, we probe this non-radiative interaction by using a direct-gap semiconductor whose luminescence is a direct measure for the ET efficiency. In essence, the TMD acts as a receiving optical antenna that enhances the LDOS and converts non-radiative modes to a measurable signal. We refer to this process as electron-photo-luminescence (ePL) since it is triggered by an electron that transfers energy through optical modes to WSe$_2$.
\subsection*{Dependence on TMD-Graphene separation}
To further test our interpretation, we study the dependence of the excitonic emission on the distance between the TMD and the tunnel junction. A device schematic is shown in Fig.~\ref{fig4}A where we vary the coupling to the TMD by changing the thickness $d_\mathrm{s}$ of a hBN~spacer layer (s-hBN) in different device regions. Figure~\ref{fig4}B shows the emitted light from a device where the s-hBN~ presents steps in thickness in the tunnel junction area. Four regions of emission are created. Three with different s-hBN~ thicknesses (6, 5 and 4 layers) and one without WSe$_2$ on top. We observe that the emission gets stronger the thinner the s-hBN is. As evidenced by the spectra shown in Fig.~\ref{fig4}C the enhanced emission can be entirely attributed to the strength of the excitonic peak. We thus find that the excitonic contribution increases for thinner s-hBN~ in contrast to the broadband background that doesn't show any major change. A comparison of the excitonic emission intensity $I\mathrm{_x}$ for different $d\mathrm{_s}$ at $V\mathrm{_{b}}$~=~\SI{2}{V} is given in Fig.~\ref{fig4}D. To compare different devices we normalize $I\mathrm{_x}$ by the area of emission, the PL efficiency and the background level (see supplementary materials). The data points in Fig.~\ref{fig4}D clearly show that the rate of exciton generation, and hence the ET rate, decay superlinearly with the separation $d_\mathrm{s}$.
$I\mathrm{_x}$ can be modeled by calculating the exciton generation rate $\Gamma\mathrm{_x}$. By using inelastic tunneling theory \cite{Parzefall2017} $\Gamma\mathrm{_x}$ assumes the following expression:
\begin{equation}
\begin{aligned}
\Gamma\mathrm{_x} (V\mathrm{_b},d\mathrm{_s}) \propto & \\ \int_{0}^{\infty}\eta\mathrm{_{abs}}(\omega)&\frac{\gamma\mathrm{^{0}_{inel}}(\omega,V\mathrm{_b})}{\rho\mathrm{_{0}}(\omega)} [\rho\mathrm{_{opt}}(\omega,d\mathrm{_s}) -\rho\mathrm{_{opt}}(\omega,\infty)]d\omega
\end{aligned}
\label{eq1}
\end{equation}
\noindent
where $\omega$ is the angular frequency, $\eta\mathrm{_{abs}}$ is the absorption spectrum of monolayer WSe$_2$ \cite{Kozawa2014}, $\gamma\mathrm{^{0}_{inel}}$ is the inelastic tunneling spectral rate in vacuum and $\rho\mathrm{_{0}}$ is the vacuum LDOS. We approximate the LDOS responsible for exciton generation by $\rho\mathrm{_{opt}}(\omega,d\mathrm{_s})-\rho\mathrm{_{opt}}(\omega,\infty)$ which assumes that the states provided by the TMD, leading to exciton generation, vanish for $d\mathrm{_s} \shortrightarrow \infty$. See supplementary materials for more information on the calculation. The calculated $\Gamma\mathrm{_x}$ as a function of distance $d\mathrm{_s}$ and for $V\mathrm{_{b}}$~=~\SI{2}{V} is plotted in Fig.~\ref{fig4}D. It agrees well with the distance dependent measurement. Interestingly, both $\Gamma\mathrm{_x}$ and the measured data can be described by a simple inverse square law, as shown in Fig.~\ref{fig4}D. Similar observations were reported for ET between optically excited TMDs \cite{Karmakar2022}. The good agreement between our theoretical model and our experimental measurements supports our interpretation.
\begin{figure*}
\centering
\includegraphics[width=0.7\textwidth]{figure4.pdf}
\caption{\textbf{Distance dependent energy transfer. }(a) Schematic of a TMD separated from the tunnel junction by a spacer-hBN~(s-hBN). The coupling strength between TMD and the tunnel junction is varied using different s-hBN~thicknesses $d_\mathrm{s}$. (b) Real-space image of emission from the WSe$_2$/s-hBN/Gr/b-hBN/gold device. Strongest emission is observed from the thinnest s-hBN~region. The emission is filtered with a bandpass filter ($\lambda_{center}$ = \SI{750}{\nano\meter} and fwhm = \SI{40}{\nano\meter}). The scale bar is \SI{4}{\micro\meter}. (c) Spectra measured on areas with different s-hBN~thickness and for $V\mathrm{_{b}}$~=~\SI{2}{V}. The broad spectral background corresponds to direct photon coupling and is also observed in areas with no WSe$_2$. (d) Excitonic emission I$_x$ as a function of $d\mathrm{_s}$. The values are normalized with the PL efficiency, the emission area and the intensity of background emission. The measurement point at $d\mathrm{_s}$ = \SI{0}{nm} refers to the device presented in Fig.~\ref{fig2}. The three points between $d\mathrm{_s}$ = \SI{1}{nm} and \SI{2.5}{nm} refer to the device presented in Fig.\ref{fig4}B and the point at $d\mathrm{_s}$ = \SI{5}{nm} refers to a third device, for which spectra are presented in Fig. S2. The dashed curve is the result of Eq.~\ref{eq1} and the solid curve is a fit with an inverse square distance function. $\mathrm{D+d_s}$ is the distance between the TMD and the center of the b-hBN layer as illustrated in (a).}
\label{fig4}
\end{figure*}
\subsection*{Monolayer vs bilayer TMD}
We continue our study by using a different TMD (MoSe$_2$) and by comparing ET for monolayers (1L) and bilayers (2L). Figure~\ref{fig5}A shows a schematic of the device and the measured ePL spectra. The emission for 2L MoSe$_2$ ($I\mathrm{_{2L}^{ePL}}$) is less intense than for 1L MoSe$_2$ ($I\mathrm{_{1L}^{ePL}}$) owing to the indirect bandgap of 2L MoSe$_2$. To compare this ET-mediated photoluminescence (ePL) with optically excited photoluminescence (PL) we optically excite the two regions with a laser and measure the corresponding emission spectra $I\mathrm{_{1L}^{PL}}$ and $I\mathrm{_{2L}^{PL}}$, respectively. The measurements are presented in Fig.~\ref{fig5}B. The center wavelengths and the shapes of PL and ePL spectra match very well, which supports the interpretation that ePL corresponds to spontaneous exciton emission. We note however, that the ratios $I\mathrm{_{1L}^{PL}}$/$I\mathrm{_{2L}^{PL}}$ and $I\mathrm{_{1L}^{ePL}}$/$I\mathrm{_{2L}^{ePL}}$ are different. Interestingly, the $I\mathrm{_{1L}^{ePL}}$/$\mathrm{I_{2L}^{ePL}}$ ratio depends on $V\mathrm{_{b}}$. In fact, $I\mathrm{_{2L}^{ePL}}$ is stronger than $I\mathrm{_{1L}^{ePL}}$ for low voltages and a crossing occurs near $V\mathrm{_{b}}$~=~\SI{1.75}{V}, as shown in Fig.~\ref{fig5}C, which depicts the integrated spectra for different voltages. By employing the model in Eq.~\ref{eq1} we are able to reproduce this effect. The result of this calculation is given in the inset of Fig.~\ref{fig5}C. The reason for this voltage dependence are the different optical properties of 1L and 2L MoSe$_2$, i.e. the absorption spectrum $\eta\mathrm{_{abs}}$ of 2L MoSe2, which features a cut-off energy at lower energies and stronger absorption than the 1L MoSe$_2$. This explains the earlier onset-voltage for $I\mathrm{_{2L}^{ePL}}$ in Fig.~\ref{fig5}C. Moreover, the two curves are also affected by the dependence of the LDOS on flake thickness. In fact, our model predicts that the rate of ET to a 2L TMD is lower compared to a 1L despite the fact that the 2L TMD is thicker. This surprising behavior has been studied in the past in optically excited systems \cite{Prins2014}.
\begin{figure*}[t!]
\centering
\includegraphics[width=1\textwidth]{figure5.pdf}
\caption{\textbf{Tunnel junctions coupled to monolayer and bilayer MoSe$_2$.} (a) ePL emission spectra from a device with monolayer (1L) and bilayer (2L) MoSe$_2$. Spectra taken at $V\mathrm{_{b}}$~=~\SI{1.9}{V}. The inset shows the device schematic. (b) PL emission from 1L and 2L regions of the same device. (c) Integrated intensities of the ePL spectra in (a) as a function of $V\mathrm{_{b}}$. The inset shows corresponding theoretical curves based on Eq.\ref{eq1}.}
\label{fig5}
\end{figure*}
\subsection*{Discussion}
The exciton generation mechanism, discussed in this work, is an electromagnetic phenomenon that is usually neglected in the analysis of tunneling driven systems. The tunneling probability of an electron relies on the available electronic states but also optical states. In semiconducting materials, exciton resonances provide non-radiative optical states that couple strongly with inelastic electrons increasing the tunneling probability. This leads to near-field generation of excitons even for applied voltages below the electronic bandgap of the material, thus contributing to sub-bandgap emission. Such low voltage emission has been previously studied with interpretations varying from direct exciton formation \cite{Binder2017} to Auger scattering \cite{Binder2019}. Here we show that ET from tunneling electrons can contribute to this sub-bandgap emission.
\subsection*{Conclusions}
We investigated ET from a tunnel junction to TMD excitons. Our studies are based on vdW tunnel devices in which the TMD is placed outside the electronic pathway, ensuring that there is no direct charge injection into the TMD. We studied photoemission spectra as well as real space and Fourier space emission patterns and we concluded that excitons are generated when the tunnel junction is under bias. This surprising observation is understood by ET from tunneling electrons to excitons. Emission from excitons is observed even when the TMD is separated from the junction by a hBN~spacer layer. Our calculations reveal that non-radiative modes of the LDOS are strongly increased at the exciton energy. These non-radiative modes are responsible for the efficient energy transfer from the tunnel junction to the TMD. Our theoretical model based on inelastic tunneling theory and near-field coupling is in agreement with our experimental measurements. \\ \indent The electrical generation of excitons via ET provides new perspectives for the development of optical sources, detectors and sensors. For example, it can be employed for truly chip-scale light sources with applications in electrical pumping scheme for vdW heterostructure lasers \cite{Ye2015,Paik2019} or optical sensing schemes \cite{Oh2021} that do not rely on external laser sources.
\section*{Acknowledgments}
The authors would like to thank Mathieu Luisier, Achint Jain, Martin Frimmer, Ronja Khelifa, Anna Kuzmina, Shengyu Shan and Massimiliano Rossi for fruitful discussions. This study was supported by funding from ETH Zurich under ETH Grant No. ETH-15 19-1 SYNEMA, the ETH Zurich Foundation project number 2013-08 (11) with a donation from the Stavros Niarchos Foundation, and the Swiss National Science Fund under grant number $200020\_192362$. K.W. and T.T. acknowledge support from the JSPS KAKENHI (Grant Numbers 19H05790, 20H00354 and 21H05233).
\section*{Author contributions}
S.P., L.W. and L.N. conceived the experiment. S.P. fabricated the devices, performed the measurements, analysed the data, developed the theoretical model and performed the calculations. L.W. supported in the device fabrication. L.W. and L.N. helped with the data interpretation. K.W. and T.T. provided the high-quality hBN~crystals, L.N. initiated and supervised the project. S.P. wrote the paper and all authors discussed the results and worked on the manuscript.
\subsection*{Materials and methods}
\paragraph*{\textbf{Sample fabrication}}
MoSe$_2$, WSe$_2$, graphene and hBN~flakes are mechanically exfoliated from bulk crystals on Si/SiO$_2$ substrates with the scotch tape method. The hBN~flakes are exfoliated in air the other flakes in argon (Ar) atmosphere. The Si/SiO$_2$ substrates are first cleaned in an oxygen plasma asher for 5 minutes. The heterostructures are fabricated by a dry pick-up and transfer technique \cite{Zomer2014} as follows: A poly-dimethylsiloxane (PDMS) stamp, covered by a thin polycarbonate (PC) film, is used to sequentially pick-up all flakes from their substrate at 80 °C in Ar atmosphere in a glovebox. The resulting stack is then transferred on top of pre-patterned gold electrodes on a glass substrate by heating up the substrate to 170 °C in order to melt the PC film and allow it to detach from the PDMS stamp. In that way, the stack is transferred on top of the gold electrodes. Later, the PC film is dissolved in a chloroform bath for 30 minutes. The gold electrodes are pre-patterned by photolithography followed by e-beam evaporation of 5 nm Ti (adhesion layer) and 50 nm gold. Metal lift-off is done in an acetone bath at 60 °C. \\ \\
\paragraph*{\textbf{Electrical and optical characterization}}
The gold electrodes on the sample's substrate are wire-bonded to adhesive copper pads that allow conductive contact to co-axial connections to a Keithley Source Meter 2602B, that is used as a voltage source and current meter. The sample is then put, facing down, on a Nikon TE300 inverted microscope in ambient conditions. The light emitted by the device under bias voltage is collected by a 100x objective with 0.9 NA. All spectra are measured by imaging the real space of emission at the entrance of a Princeton Instruments Acton SpectraPro 300i spectrometer. The real and Fourier space (back-focal plane) of emission is imaged at the entrance of an Andor iXon Ultra camera. Finally, all PL measurements were collected by optically exciting the sample with a 532 nm solid state laser. \\ \\
\paragraph*{\textbf{Density of optical states}}
The dissipated power $P$ of a point electric dipole is calculated by \cite{Novotny2012}
\begin{equation}
P = \frac{1}{2}\omega~ \mathbf{p}~ Im \{\mathbf{E(r_0)}\}
\end{equation}
\noindent
where $\omega$ is the angular frequency of the dipole's emission, $\textbf{p}$ is the dipole moment and $\textbf{E}$ is the electric field at the dipoles origin \textbf{$r\mathrm{_0}$}. The dipole is assumed to be situated at the center of the tunneling hBN~layer. The electric field $\textbf{E(r$_0$)}$ is calculated by solving the electromagnetic wave equation for a multi-layer structure. The optical density of states $\mathrm{\rho_{opt}}$~is then calculated through its relation to P \cite{Novotny2012}
\begin{equation}
\frac{\rho\mathrm{_{opt}}}{\rho\mathrm{_{0}}} = \frac{P}{P\mathrm{_{0}}}
\end{equation}
\noindent
with $P\mathrm{_{0}}$ being the dissipated power for a point dipole in vacuum. \\ \\
\paragraph*{\textbf{Exciton to photon coupling ratio}}
We calculate the ratio $\eta\mathrm{_{e\mhyphen x}}/\eta\mathrm{_{e\mhyphen\gamma}}$ using the following relationship,
\begin{equation}
\frac{I\mathrm{_x}}{I\mathrm{_\gamma}} = \frac {\eta\mathrm{_{e\mhyphen x}}\cdot\eta\mathrm{_{x\mhyphen\gamma}}\cdot\eta\mathrm{_{col}^{\shortrightarrow}}}{ \eta\mathrm{_{e\mhyphen\gamma}}\cdot\eta\mathrm{_{col}^{\shortuparrow}}}
\end{equation}
\noindent
where $I\mathrm{_x}$ is the exciton coupled emission, $I\mathrm{_\gamma}$ is the photon coupled emission, $\eta\mathrm{_{col}^{\shortrightarrow}}$ and $\eta\mathrm{_{col}^{\shortuparrow}}$ is the collection efficiency for an in-plane and an out-of-plane dipole, respectively. The $\eta\mathrm{_{x\mhyphen\gamma}}$ expresses the PL efficiency of the WSe$_2$ flake which is measured to be around $1.5\cdot 10^{-5}$. The low value results from graphene quenching \cite{Lorchat2020}. \\ \\
\paragraph*{\textbf{Exciton generation rate}}
The exciton generation rate calculated in Eq. 1 in the main text involves the inelastic tunneling rate spectrum $\gamma\mathrm{_{inel}}$ which is a function of the electronic density of states of the right and left electrodes ($\rho\mathrm{_R}$ and $\rho\mathrm{_L}$, respectively),the optical density of states $\rho\mathrm{_{opt}}$ and the applied $V\mathrm{_{b}}$. It is calculated by \cite{Parzefall2019a,Parzefall2017},
\begin{equation}
\gamma\mathrm{_{inel}}(\hslash\omega) = \frac{\rho\mathrm{_{opt}}(\hslash\omega)\pi e^2}{3\hslash\omega m^2\epsilon\mathrm{_0}}\int_{\hslash\omega}^{eV\mathrm{_b}}|\altmathcal{P}(E,\hslash\omega)|^2\rho\mathrm{_R}(E-\hslash\omega)\rho\mathrm{_L(E)}dE
\label{eq2}
\end{equation}
\noindent
where $\altmathcal{P}$ is the momentum matrix element, \textit{m} is the electrons mass, $\epsilon_0$ is the vacuum permittivity and \textit{e} is the elementary charge. \\ \\
\paragraph*{\textbf{Emission comparison between different devices}}
To be able to compare the rate of excitonic emission of different devices we normalize the measured emission intensity $I\mathrm{_x}$ by the area of emission and by the relative PL efficiency of the flake. Finally we normalize by the photon-coupled background emission $I\mathrm{_\gamma}$ in every device. This normalization is necessary because in every device the tunneling barrier thickness may vary by $\pm \SI{0.33}{nm}$. This has a minor effect on the optical coupling but it can drastically affect $\gamma\mathrm{_{inel}}$ through the momentum matrix element $\altmathcal{P}$ (Eq.~\ref{eq2}). This impacts the exciton generation rate $\Gamma \mathrm{_x}$ making the direct comparison between different devices impossible. Since $I\mathrm{_\gamma}$ is directly proportional to $\altmathcal{P}$ we use it to normalize the $I\mathrm{_x}$ emission and be able to compare devices with different TMD to tunnel junction distance (Fig.~4 in the main text). \\ \\
\subsection*{Conductance measurements}
We electrically characterize the device by applying a bias voltage ($V\mathrm{_{b}}$) between the gold and graphene electrodes as indicated in Fig.~\ref{figs0}A. The current and conductance measurements are shown in Fig.~\ref{figs0}B. The triple minima in the conductance curve around 0 V are associated with the intrinsic graphene doping and the phonon induced tunneling that governs the transport of such devices, as it has been investigated in previous studies \cite{Zhang2008, Vdovin2016, Parzefall2019}. This indicates that tunneling indeed occurs between graphene and gold and is phonon-mediated. Moreover, for $V\mathrm{_{b}}$~$<$1.5 V the conductance shows a linear behavior. In this low voltage regime compared to the hBN barrier (\SI{5}{eV}) the conductance curve is associated to the linear density of states of graphene. Interestingly, around $V\mathrm{_{b}}$ = \SI{1.5}{V} and \SI{-1.7}{V}, the slope of the conductance increases. This behavior is not observed in devices without a WSe$_2$ and can be associated with an increased rate due to interaction of the tunneling electrons with excitonic states.
\begin{center}
\begin{figure}[h!]\center
\includegraphics[width=0.7\textwidth]{figureS1.pdf}
\caption{ \textbf{Electrical characterization.} Device schematic of a TMD coupled tunneling junction. Current and conductance (dI/dV) measurements as a function of $V\mathrm{_{b}}$~.}
\label{figs0}
\end{figure}
\end{center}
\pagebreak
\subsection*{Exciton generation in both polarities}
As discussed in the main text the exciton generation process occurs for both bias polarities. Figure~\ref{figs1} shows spectra for both polarities where the narrow emission at exciton wavelengths is This indicates excitons are not generated by electron-hole injection since such a process, if possible, would be strictly polarity dependent.
\begin{center}
\begin{figure}[h!]\center
\includegraphics[width=0.8\textwidth]{figureS2.pdf}
\caption{\textbf{Spectral symmetry with bias voltage.} (a,b) Photoemission spectra from the TMDC coupled tunneling junction presented in Fig.~2A in the main text for (a) positive bias voltage and (b) negative bias voltage. Similar spectral characteristics are observed. This observation further supports the argument that there excitons are not created by charge injection as such a process would be polarity dependent.}
\label{figs1}
\end{figure}
\end{center}
\newpage
\subsection*{TMD coupled device with 5 nm spacer hBN}
\begin{center}
\begin{figure}[h!]\center
\includegraphics[width=0.9\textwidth]{figureS3.pdf}
\caption{\textbf{TMD coupled device with 5 nm spacer hBN.} (a) Device schematic. The WSe$_2$ is separated from the tunneling junction by a spacer hBN with 5 nm thickness. The WSe$_2$ flake is stacked on top in a way that two areas are formed: with and without WSe$_2$. (b,c) Photoemission spectra from the two areas (b) without WSe$_2$ and (c) with WSe$_2$. A small contribution in excitonic wavelengths is still observed even though the WSe$_2$ flake is 5 nm away from the tunneling junction.}
\label{figs2}
\end{figure}
\end{center}
|
\section{Introduction}
\label{sec:intro}
Simulation of fluid flow for practical applications is characterised by high Reynolds number turbulent flows over complex geometries\cite{homiff,karniadakis1999simulating}. The temporal and spatial scales of motion in such flows span several orders of magnitude\cite{moin_DNS_research}. The cost of Direct Numerical Simulations (DNS), which attempt to capture all energy containing lengthscales in the flow, grows superlinearly with the Reynolds number of the flow\cite{kolmogorov, dns_primer, orszag_1970}, limiting the scope of fully resolving simulations to canonical geometries at medium Reynolds numbers\cite{moin_DNS_research}.
Turbulence closure models have been developed for cases where resolving and evolving the entirety of velocity and pressure spectra is not possible. Techniques such as Large Eddy Simulations (LES) and Reynolds Averaged Navier-Stokes (RANS) allow for cheaper calculations of high Reynolds number flows by resolving only a portion of the spectrum and `modeling' the rest \cite{pope}. Put simply, the goal of turbulence modeling is to find a closed system that can describe the time-evolution of observables, such as the resolved velocity spectrum, via a `closure' map that allows one to untangle the dependence of said observables on unresolved variables\cite{sagaut2006large,adams2007mathematics,ml_stat_closure}.
The problem of turbulence closure modeling has been extensively studied for over a century, dating back to Boussinesq's eddy viscosity hypothesis in 1877\cite{Layton2014THE1B}. More recently, machine learning (ML) based methods have utilized for modeling fluid flow problems\cite{Brunton2020,vinuesa2021potential,turb_modeling_data,kochkov2021machine,watt2021correcting,Vlachas2018,Thuerey_2020,portwood2021interpreting,Wang2020}. The success of machine learning is built upon the ability of deep neural networks to approximate functions in high-dimensional spaces with a number of parameters that does not scale exponentially with the dimension of the problem space\cite{weinan}. As the underlying problem in closure modeling is of finding the functional mapping between high-dimensional vector spaces,
deep neural networks can be utilized in a supervised learning framework.
However, the performance of machine learning models to turbulence modeling has largely been unsatisfactory. A probable reason is that while neural networks are touted to break the curse of dimensionality, the cost of attaining an accurate model may be larger. The cost of generating high-fidelity data for the model does scale with dimensionality\cite{pope}, and even if data is readily available, machine learning models need to be trained on large number of samples to accurately approximate even simple functions. In other words, an optimal closure model may be present in the space of functions that a neural network architecture can express, but there is no clear path towards realizing it.
Some success has been seen, however, by augmenting neural network models with physical priors. For example, Ling et al.\cite{ling_kurzawski_templeton_2016,milani2021} improved upon conventional machine learning models by embedding Galilean invariance in the prediction of Reynolds stress tensor for Reynolds Averaged Navier-Stokes simulations. Shankar et al.~\cite{shankar2020learning} enhanced the output of a convolutional neural network model on homogeneous isotropic turbulence by imposing the divergence-free constraint on the model output. Both approaches can be understood as attempts to narrow down the space of functions that a neural network architecture can represent.
Rackauckas et al.\cite{universal_diffeq} further develops inductive biases in data-driven learning by directly embedding trainable parameters in a model's governing differential equations. Training such parameters against a loss function that depends on the solution to the governing system requires passing gradient information through a differential equation solve. Therefore a solver that is compatible with automatic-differentiation toolchains in needed. Directly backpropagating gradients is computationally expensive for large solves, so the adjoint optimization method\cite{Chen,farrell2013automated,errico1997adjoint,sigmund2013topology} is utilized. CFD codes such as SU2 have utilized the adjoint method for sensitivity analysis and shape optimization\cite{su2}.
This idea is formalized into the notion of Differentiable Physics\cite{diffphys}, defined as the set of scientific avenues emerging from the marriage of state-of-the-art partial differential equation (PDE) solvers with differentiable programming. The promise of differentiable physics is that machine learning models can be developed to learn specific unknown or residual physics, leading to more interpretable models than classical ML approaches. We envision the landscape of data-driven models on an axis of increasing inductive biases in \autoref{fig:dp}.
\begin{figure*}
\centering
\includegraphics[width=0.8\linewidth]{plots/diffphys.pdf}
\caption{Data-driven models can incorporate different inductive biases, from locality in convolutional networks, symmetries in equivariant networks, to differential equation forms in universal differential equations. Models on the right are constrained by known physics, such as a dynamical system or advective and dissipative terms, leading to interpretable models where the learning problem has been restricted to unknown or residual physics.}
\label{fig:dp}
\end{figure*}
In this work, we comprehensively test the ability of ML techniques to develop turbulence closure models that are resolution independent, interpretable in the language of flow physics, and generalizeable across a range of Reynolds numbers. We consider the 1D Burgers system as the ideal test problem for modeling the energy spectra observed in advection-dominated turbulence problems, and narrow our focus to the task of subgrid-stress modeling. We test a range of differentiable physics models, that vary from a black-box approach to training coefficients in state-of-the-art turbulence models. Experiments are conducted over a range of Reynolds numbers and grid resolutions to assess the generalizability of the model, and over multiple resolution-independent architectures. We find that physics-based inductive biases are critical to the development of data-efficient and accurate ML closures, and that these ML closures can outperform state-of-the-art closure baselines in a wide range of flows.
\section{Background}
\label{sec:background}
\subsection{Burgers Turbulence}
Turbulent flows are characterised by a competition between viscous forces, which damp out velocity fluctuations by converting kinetic energy into heat, and inertial forces, such as gravitation or pressure gradients. These tend to generate and preserve velocity and pressure fluctuations\cite{wall-bounded-turb}. In low Reynolds number flows, viscous forces dominate and the flow is near perfectly damped, meaning that any arbitrary fluctuation in the velocity field would be smoothed out in no time\cite{damped}. The velocity fields adjusts almost instantaneously to any changes in the pressure gradient that drives the flow. At high Reynolds numbers, however, viscous forces may not be strong enough to dampen out velocity fluctuations, and even tiny disturbances may cause the entire flow to destabilise into turbulence\cite{instabilitiy}. Energy dissipation primarily happens at the smaller scales as the rate of dissipation scales with the wavenumber squared.
We restrict the scope of this work to the task of subgrid stress modeling in advection-dominated turbulence problems.
The unforced, viscous Burgers problem, \autoref{eqn:burg}, an advection-diffusion type problem\cite{burg} that exhibits an energy cascade similar to the Navier-Stokes system, is the perfect test bench. The Burgers system is governed by
\eqn{\label{eqn:burg}
\partial_t u + u\partial_x u = \nu \partial_{xx}u,
}
where $u(x,t)$, the solution to the Burgers system, represents a time-varying velocity field and $\nu$ is the kinematic viscosity. The 1D viscous Burgers system is a canonical PDE that is widely used as a simplified case study for subgrid model development and analysis \cite{falkovich2006lessons,burg,love1980subgrid,labryer2015framework}. The nonlinear advective term coupled with viscous dissipation leads to similar multiscale phenomena observed in more complex flow dynamics. \autoref{eqn:burg} is deterministic which allows us to directly compare trajectories rather than statistical averages, speeding up model training. We insulate the problem from arbitrary effects of boundary conditions by considering \autoref{eqn:burg} on the entire real line, $\R$, which is approximated by periodic boundary conditions on an interval. The Burgers equation can also be written in its conservative form
\begin{equation}
\label{eqn:burg-conservative}
\partial_{t} u + \frac{1}{2}\partial_{x} u^2 = \nu\partial_{xx}u ,
\end{equation}
which is typically more amenable to numerical calculation.
\subsection{Large Eddy Simulations}
As high Reynolds number flows have a wide energy spectrum, DNS of such flows is often computationally intractable, and a turbulence model is needed. Large Eddy Simulations (LES) resolves the flow variables starting from the intertial range up to a cut-off length, $\Delta$, that depends upon the computation grid\cite{lesieur1996new,boris1992new}. Critically, an LES calculation ignores the smallest lengthscales in the flow, which are most computationally expensive to resolve. Filtering is accomplished by
\eqn{
\label{eqn:filter}
u(\vect{x}, t) &= \bar{u}(\vect{x}, t) + u'(\vect{x}, t)\\
\bar{u}(\vect{x}, t) &= G_\Delta \star u(\vect{x},t),
}
where $G_\Delta$ is a low-pass filtering operator. Quantities of interest are interpolated onto a computational grid that can resolve flow features only up to lengthscale $\Delta$. The grid has much lower resolution than what is needed to fully resolve the flow, and evolving flow variables on this grid lead to saving computational resources.
Filtering the Burgers equation in this manner leads to
\eqn{\label{eqn:burg-LES}
\partial_t \bar{u} + \frac{1}{2}\partial_x \bar{u}^2 = \nu \partial_{xx} \bar{u} - \frac{1}{2}\partial_x\bar{u'^2},
}
the LES equations for the Burgers system. Evolving $\bar{u}$ per \autoref{eqn:burg-LES} is not possible as one does not have access to the subgrid stress term, ${\eta}\defeq\frac{1}{2}\bar{{u}'{u}'}$, and ignoring it leads to an incorrect flow field and numerical instabilities. For the Navier-Stokes equations, ignoring subgrid stresses leads to numerical instabilities spurious high-frequency modes\cite{fischer-mullen-filter}. The closure problem of LES is of modeling the effect of ${\eta}$ on the dynamics of $\bar{u}$. This is called subgrid stress modeling. As one does not have access to the unresolved quantities, subgrid stress models are formulated in terms of the resolved quantities.
Kolmogorov hypothesized that the finer scales are largely similar\cite{kolmogorov}, and are primarily responsible for energy dissipation. One approach to closure modeling is to take out small amounts of energy from the high-frequency modes in $\bar{{u}}$ by applying a mild low-pass filter at every timestep. Fisher et al.\cite{fischer-mullen-filter} has applied this methodology to the Navier-Stokes equations in the Computational Fluid Dynamics (CFD) code \textsc{NEK5000}\cite{nek5000}, and Layton et al.\cite{layton2010temporally} has prove existence of unique strong solutions to the resulting continuum model for the same.
Another approach is to directly model the commutative error term, ${\eta}$ in terms of $\bar{{u}}$ with a map $\eta(\bar{u})$, and evolve the flow variables as per \autoref{eqn:burg-LES}. Hypothesizing a functional form for additive closures of this sort is highly problem dependent. A common hypothesis is the eddy viscosity family of models\cite{frisch1995turbulence,SCHMITT2007617} that introduce an ``effective'' or ``eddy'' viscosity, $\nu_t$, to the momentum equation for enhancing dissipation. This leads to a functional form of ${\eta} = \partial_x\nu_t\partial_x\bar{u}$, where $\nu_t$ could be a fixed scalar based on known flow physics and computational grid, or be defined by a transport equation. In \autoref{sec:experiment} we consider several such models and apply them to the Burgers problem.
\section{Method}
\label{sec:method}
\subsection{Numerical methods}
We compute solutions to the Burgers equation via a pseudospectral method \cite{pseudospectral}. The periodic 1D domain is discretized with a regular grid of $N$ grid points. The periodic nature of the solutions allow for expansion in a Fourier basis, as per
\begin{equation}
u(x,t)=\sum_{k=-N/2}^{N/2}\hat{u}_k(t)e^{ikx} .
\end{equation}
The linear term of the equation can be computed pointwise in Fourier space
\begin{equation}
\partial_{xx}\hat{u}(k,t) = -k^2\hat{u}(k,t) ,
\end{equation}
however the nonlinear term introduces interactions between all the wavemodes and thus requires more care for efficient computation.
Therefore, it is more suitable to compute the quadratic term $u^2$ pointwise in real space at the $N$ collocation points, then compute the spatial derivative in Fourier space.
The spatial discretization allows for a method of lines approach to time evolution. The discrete Fourier modes from the spectral basis expansion produces a system of $N$ ordinary differential equations (ODEs) that can be solved with standard ODE integration schemes. We choose the Tsitouras adaptive 5th order scheme \cite{tsit5}.
\subsection{Problem statement}
Accurate simulation of the governing Burgers equation imposes certain restrictions on the temporal and spatial discretization. As we use an adaptive time-stepping scheme with fixed error tolerance, we focus on errors arising from the spatial discretization. The spatial grid should be sufficiently dense to resolve the smallest dissaptive length scales in the flow, typically much smaller than the characteristic length scales of the flow.
We wish to approximate solutions to the Burgers equation on a considerably coarser grid of $M$ points, by exploring a family of data-driven models enabled by algorithmic differentiation of the temporal and spatial discretization schemes. Each of the models is described by a PDE or set of PDEs, where certain terms in the equations are computed by deep neural networks. Thus, we enable significant interpretability of the models as network outputs can be readily identified as components of traditional closure modeling approaches.
\subsection{Models}
In this section, we provide a description of the various models that we use in our experiments. First, we make clear our notational conventions.
$u(x,t)$ is the spatiotemporal velocity field we are modeling. $\bar u(x,t)$ is the filtered velocity field on the LES grid. We drop the parentheses for brevity. $\eta$ represents an unknown closure variable we are modeling. The exact description of this variable differs across models. In some models, an additional transport equation is introduced to govern the dynamics of the closure variable. The $ \overset{\bullet}{} $ and the $\nabla$ symbols correspond to the time derivative and spatial derivative of the field respectively. A subscript 0 represents the initial condition of the field.
Neural networks are indicated with subscript $\theta$. These functions are parameterized by a large number of weights with functional form described in the subsequent section. Network inputs and outputs are written as:
\begin{equation}
c,d=f_\theta(a,b;e) ,
\end{equation}
where $a$ and $b$ are explicit inputs to the network, and $c$ and $d$ are outputs of the network with implicit dependence on variable $e$. For example, $\eta=f_\theta(u;x)$ represents a network that takes the spatially varying velocity field as input and produces a spatially varying $\eta$ field as output.
The ground truth equation, or \textit{none} model, is given by:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u
\end{equation}
Since there is no closure term or trainable networks, we use this as a baseline comparison for our learned models.
We also compare our models to a couple baseline closure models \cite{Maulik2018}, the constant Smagorinsky model (further denoted as \textit{smag-const}), and the dynamic Smagorinsky model (further denoted as \textit{smag-dyn}).
The constant Smagorinsky model\cite{smagorinsky1963general} is given by:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\nu_t\nabla \bar u)
\end{equation}
\begin{equation}
\nu_t = (C_s\delta x)^2|\nabla \bar u| ,
\end{equation}
where $\nu_t$ is the eddy viscosity, $C_s$ is the Smagorinsky constant, and $\delta x$ is the grid spacing. While there are no trainable networks in this model, $C_s$ can be optimized via hand tuning or gradient descent.
The dynamic Smagorinsky model \cite{germano1991dynamic,lilly1992proposed} is similar to the constant model, however the term $(C_s\delta x)^2$ is determined from the state of the system. Studies have shown that $C_s$ is not constant and may be variable with different flow characteristics \cite{smagorinsky1993some}. The model uses a second test filter, denoted by $\sim$ with filter scale $\tilde{\delta} x$. We fix the filter ratio $\kappa=\tilde{\delta} x/\delta x =2$. Specifically,
\begin{equation}
(C_s\delta x)^2 = \frac{\langle HM \rangle}{\langle M^2 \rangle} ,
\end{equation}
\begin{equation}
H = \nabla(\widetilde{\bar{u}}^2/2) - \nabla(\widetilde{\bar{u}^2}/2) ,
\end{equation}
\begin{equation}
M = \kappa^2 \nabla(|\nabla \widetilde{\bar{u}}|\nabla \widetilde{\bar{u}}) - \nabla(\widetilde{|\nabla \bar{u}|\nabla \bar{u}}) ,
\end{equation}
where $\langle \rangle$ denotes averaging over the spatial domain.
The rest of the models in this section contain trainable networks in functional forms that rely on varying degrees of assumptions typically used in closure analysis. At one extreme, we have model \textit{resnet}, which uses the classical ResNet architecture \cite{resnet} to predict the coarse-grained flow field at the next time step:
\begin{equation}
(\bar{u}_{t+\Delta t}, \eta_{t+\Delta t}) = (\bar{u}_{t}, \eta_{t}) + f_\theta(\bar u_t,\eta_t,\nu;x)\Delta t,
\end{equation}
\begin{equation}
\eta_0 = g_{\theta}(\bar u_0; x),
\end{equation}
where $\Delta t$ is a coarse time step. This model makes no assumptions about the physical description of the system, including the fact that it may be described as a PDE. The closure variable $\eta$ is used only to lift the model into a higher-dimensional space governed by the neural network. The network may choose to learn $\eta_t=0$, or it may leverage $\eta$ in the computation of $\bar{u}_t$. Projecting or transforming an input into a higher-dimensional subspace is a common practice in ML, where a large latent space of hidden channels eases learning. This is a core component of augmented neural ODEs \cite{anode}, which have shown to generalize better than standard neural ODEs.
\textit{anode}, an augmented neural ODE, is a small modification to \textit{resnet} that assumes an ODE inductive bias. The model is given by:
\begin{equation}
\dot{\bar{u}}, \dot \eta = f_\theta(\bar u,\eta,\nu;x)
\end{equation}
\begin{equation}
\eta_0 = g_{\theta}(\bar u_0; x),
\end{equation}
where $f_\theta$ explicitly describes the RHS of an ODE governing the discretized flow field. The dynamics of $\bar u$ and $\eta$ are determined purely by the network $f_\theta$ with no additional terms. This involves only a small change in the code of \textit{resnet}, changing the solver from explicit Euler to an adaptive scheme.
Filtering the governing equations as in \autoref{eqn:burg-LES} introduces an additional closure term to the equations. Model \textit{direct} models the entirety of this term directly as an additive correction:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \eta
\end{equation}
\begin{equation}
\eta = f_\theta(\nabla \bar u, \nu; x).
\end{equation}
Model \textit{transport-I} leverages the fact that the closure term can be written as the divergence of an unknown subgrid stress. Here, $\eta$ represents this stress analogue in 1D, which is additionally transported with another dynamical equation:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla\eta
\end{equation}
\begin{equation}
\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta
\end{equation}
\begin{equation}
\alpha, \beta = f_\theta(\nabla \bar u, \nabla \eta, \nu; x)
\end{equation}
\begin{equation}
\eta_0 = g_{\theta}(\bar u_0, \nu; x).
\end{equation}
The coefficients of the dissipative and convective terms in the $\eta$ transport equation are given by neural networks.
Model \textit{transport-II} uses the Boussinesq hypothesis that the subgrid stress has a linear relationship with the strain in the flow. $\eta$ represents an eddy viscosity analogue:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\eta\nabla \bar u)
\end{equation}
\begin{equation}
\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta
\end{equation}
\begin{equation}
\alpha, \beta = f_\theta(\nabla \bar u, \nabla \eta, \nu, \delta x; x)
\end{equation}
\begin{equation}
\eta_0 = g_{\theta}(\bar u_0, \nu; x).
\end{equation}
Model \textit{transport-I-p} is nearly identical to \textit{transport-I}, however the $\eta$ transport equation has an additional production term $\gamma$:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla\eta
\end{equation}
\begin{equation}
\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta + \gamma
\end{equation}
\begin{equation}
\alpha, \beta, \gamma = f_\theta(\nabla \bar u, \nabla \eta, \nu; x)
\end{equation}
\begin{equation}
\eta_0 = g_{\theta}(\bar u_0, \nu; x).
\end{equation}
Model \textit{transport-II-p} is equivalent to \textit{transport-II} except for the additional production term:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\eta\nabla \bar u)
\end{equation}
\begin{equation}
\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta + \gamma
\end{equation}
\begin{equation}
\alpha, \beta, \gamma = f_\theta(\nabla \bar u, \nabla \eta, \nu; x)
\end{equation}
\begin{equation}
\eta_0 = g_{\theta}(\bar u_0, \nu; x).
\end{equation}
The last model, \textit{smag-ml}, attempts to improve on the standard constant Smagorinsky model by learning a spatially varying Smagorinsky constant, computed via neural network:
\begin{equation}
\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\nu_t\nabla \bar u)
\end{equation}
\begin{equation}
\nu_t = (\eta\delta x)^2|\nabla \bar u|
\end{equation}
\begin{equation}
\eta = f_\theta(\nabla \bar u, \nu, \delta x; x).
\end{equation}
All of the models used in our analysis are tabulated in \autoref{tab:models} of the appendix.
\subsection{Neural network architecture}
To leverage our choice of pseudospectral discretization, we use Fourier Neural Operators (FNO) as our network architecture \cite{FNO}. All functions with subscript $\theta$ in the previous section correspond to an FNO architecture. FNOs combine pointwise in real space and pointwise in Fourier space linear operations with pointwise nonlinearities in real space to enable learning of a large suite of resolution-invariant function classes, including linear and nonlinear operators.
Given an input feature $h^n \in \mathbb{R}^{N\times c_{in}}$, where $c_{in}$ is the number of channels or scalar fields, an FNO layer is defined as:
\begin{equation}
h^{n+1} = \sigma(Bh^n + \mathcal{F}^{-1}(W_k\mathcal{F}(h^n)_k)),
\label{eqn:fno}
\end{equation}
where $B \in \mathbb{R}^{c_{in} \times c_{out}}$ is a 2-dimensional tensor that is contracted along $c_{in}$ identically pointwise in real space and $W \in \mathbb{R}^{k_{max} \times c_{in} \times c_{out}}$ is a 3-dimensional tensor that is contracted along $c_{in}$ pointwise in Fourier space for each wavemode $k$ up to $k_{max}$. Therefore, the Fourier transform of the input is truncated at $k_{max}$. The output of these operations are summed and a pointwise nonlinearity $\sigma$ is applied. The parameters of the layer are given by the tensors $W$ and $B$. \autoref{fig:fno} illustrates the algorithm for $\eta = f_\theta(\nabla \bar u, \nu, \delta x; x)$ in the \textit{smag-ml} model.
\begin{figure*}
\centering
\includegraphics[width=0.8\linewidth]{plots/FNO.pdf}
\caption{An example schematic of the FNO neural network architecture for the function $f_\theta$ in \textit{smag-ml}. At each time step, the network takes as input $\nabla \bar u$, $\nu$, and $\delta x$, computing the $\eta$ field that corresponds to a spatially varying subgrid term. Within the network, there are two FNO layers that include a local linear bias from matrix $B$ and a transformation of the Fourier modes $\hat h_k$ with matrix $W_k$. These are summed and passed through a non-linearity to the next layer.}
\label{fig:fno}
\end{figure*}
\subsection{Optimization}
Network parameters are optimized with respect to the loss function:
\begin{equation}
L = \sum_{i=1}^{T}\sum_{j=1}^{M} \frac{(\bar u_{model}(x_j,t_i)-\bar{u}_{DNS}(x_j,t_i))^2}{M*T}
\end{equation}
\begin{equation}
\hat{\bar{u}}_{DNS}=G\star \hat{u}_{DNS},
\end{equation}
where $u_{DNS}$ is the ground truth velocity field resolved on a grid of $N$ points, $\bar{u}_{DNS}$ is filtered using sharp low-pass filter $G$ with cutoff $k=M/2$, $x_j$ correspond to the $M$ collocation points on the coarse grid, and $t_i$ correspond to the $T$ timesteps in the discretized solution trajectory.
Optimization is achieved with the gradient-based optimizer ADAM, shown to be effective for deep learning applications \cite{ADAM}.
Gradients are computed with the automatic differentiation (AD) platform available in \textit{Pytorch} via reverse-mode autodifferentiation, or backpropagation \cite{pytorch}. We show a schematic of the training procedure in \autoref{fig:schematic}.
\begin{figure*}
\centering
\includegraphics[width=0.8\linewidth]{plots/schematic.pdf}
\caption{A schematic of the training procedure for learning the parameters in the data-driven models. An initial velocity field is evolved (offline) according to the governing equations on a DNS-resolving grid and filtered to generate ground truth data. During the optimization loop, the same field is filtered and evolved with the data-driven model to compute an estimate for the filtered field. The gradients are backpropagated through the ODE solve and spatial discretization to update the parameters.}
\label{fig:schematic}
\end{figure*}
Most of the operations in our solution algorithm consist of basic linear operations and pointwise nonlinearities, operations that have formed the foundation of classical multi-layer perceptron machine learning architectures. Deep learning libraries have readily supported backpropagation through these operations for many years and thus we do not discuss implementation details further. However, we highlight two slightly more complex components of the algorithm corresponding to aspects of the spatial and temporal discretization and discuss their gradient computation.
\subsubsection{Gradients of the Fourier transform}
In reverse-mode gradient accumulation \cite{speelpenning}, we wish to compute vector-Jacobian products (VJPs) \cite{pytorchAD,griewank} of each function $f$ in the algorithm:
\begin{equation}
\text{VJP} = \mathbf{v}^T f'(\mathbf{x}),
\end{equation}
given the cotangent vector $\mathbf{v}$, input $\mathbf{x}$, and Jacobian $f'(\mathbf{x})$. The VJP can then be passed to the previous function in the computational graph to compute the next VJP.
We leverage spectral basis expansion to compute spatial derivatives in the flow and inside the FNO layers. This is implemented with a Discrete Fourier Transform (DFT) operation. Since the DFT is ultimately a linear operation
\begin{equation}
\mathcal{F}_{DFT}(\mathbf{x}) = \mathbf{F}\mathbf{x},
\end{equation}
with DFT matrix $\mathbf{F}$, the Jacobian of the transform is simply the matrix $\mathbf{F}$. To compute the VJP, we can equivalently perform matrix-vector multiplication with the adjoint operator $\mathbf{F^*}$. Since $\mathbf{F^*}=N\mathbf{F}^{-1}$, we can compute the VJP as an inverse Fourier transform multiplied by the size of the signal. The benefit of this approach comes from the ability to implement Fast Fourier Transform \cite{fftw} (FFT) algorithms during the backwards pass as well, instead of constructing the entire DFT matrix explicitly.
\subsubsection{Propagating gradients through time integration}
Time evolution of the velocity field is accomplished by solving the ODE resulting from spatial discretization of the field using an arbitrary integration method. A naive implementation of AD would propagate gradients through each of basic operations in the solver. Depending on the solver tolerance, order, and trajectory length, this may result in a very large number of operations, and all intermediate steps must be saved in memory for AD to work. In our experiments, backpropagating through even moderately sized trajectory lengths, networks, and batch sizes could exceed the 16GBs of memory on our NVIDIA V100 GPU using naive autograd.
Therefore, a more memory efficient approach is desired. We use the adjoint method in time, which has been proposed for use in machine learning applications by Chen et. al \cite{Chen}.
The forward ODE system is given by the evolution of the discrete Fourier modes:
\begin{equation}
\frac{d\hat{\mathbf{u}}_k(t)}{dt} = f(\hat{\mathbf{u}}_k(t),\theta),
\end{equation}
where $\theta$ represents any learnable parameters in the model. The loss can be written as an integral:
\begin{equation}
L = \int_{t_0}^{t_T} l(\hat{\mathbf{u}}_k(t)) dt .
\end{equation}
The adjoint method produces a second ODE system that evolves the adjoint variable $\lambda$ backwards in time, given by the equation:
\begin{equation}
\frac{d\mathbf{\lambda}(t)}{dt} = -\lambda(t)^T \frac{\partial f(\hat{\mathbf{u}}_k(t),\theta)}{\partial \hat{\mathbf{u}}_k} + \frac{\partial l(\hat{\mathbf{u}}_k(t))}{\partial \hat{\mathbf{u}}_k},
\end{equation}
and an integral:
\begin{equation}
\frac{dL}{d\theta} = - \int_{t_T}^{t_0} \lambda(t)^T \frac{\partial f(\hat{\mathbf{u}}_k(t),\theta)}{\partial \theta} dt,
\end{equation}
to compute the parameter sensitivities. The RHS of the equations involve VJPs that are efficiently evaluated with AD.
The adjoint system can again be solved with an arbitrary solver, potentially even different than the forward solver.
Specifically, we use the interpolated adjoint method, which reduces the memory footprint by constructing an interpolation of the forward solution, which appears in the adjoint equations. During the forward pass, the solution is saved at a selected few timesteps and reused in the adjoint computation via interpolation during the backwards pass.
\section{Experiments and results}
\label{sec:experiment}
To assess the quality and practicality of our proposed models, we seek to study model performance across a range of Burgers solutions.
\subsection{Viscosity generalization and temporal stability}
In this experiment, we examine model generalizability with respect to initial conditions of the velocity field and the viscosity parameter $\nu$. In addition, we investigate stability of the models by extrapolating in time relative to the training data.
The models are trained using the optimization procedure described in the previous section on a ground truth dataset of 100 Burgers solutions. We vary the initial conditions of the velocity field by randomizing the Fourier modes of the initial field. The $k$th Fourier mode is denoted by $\hat u_k$. The dataset is generated with the parameters tabulated in \autoref{tab:model_params}.
\renewcommand{\arraystretch}{1.5}
\begin{table}[h]
\centering
\begin{tabularx}{0.7\linewidth}{cm{1pt}|m{2pt}X}
\textbf{Parameter} &&& \textbf{Range} \\ \hline
$\nu$ &&& $[10^{-4},10^{-3})$\\
$\hat u_k$ &&& $\left([-0.5,0.5)+[-0.5,0.5)i\right)e^{-0.5k}$, $k=0..32$\\
$t$ &&& $[0,3)$\\
$N$ &&& 4096\\
$M$ &&& 64\\
\end{tabularx}
\caption{Parameters and their distributions used to generate the training set for the first experiment, including the viscosity $\nu$, the Fourier modes of the initial velocity field $\hat u_k$, the timespan for integration $t$, the DNS grid resolution $N$, and the coarse grid resolution $M$.}
\label{tab:model_params}
\end{table}
The neural networks in the trainable models contain 2 FNO layers with $k_{max}=16$, 128 hidden channels, and appropriate input and output sizes. The first layer uses a ReLU nonlinearity while the second uses the identity. The models are trained for 500 epochs using the ADAM optimizer with learning rate varying from $10^{-3}$ to $10^{-1}$ depending on the model. Generally, models such as \textit{smag-ml} that are more constrained by inductive biases required larger learning rates.
Model performance is gauged using the mean-squared error of the predictions relative to the ground truth solution on an unseen test set. We leverage two test sets to assess generalizability across viscosities and stability in time. The first test set consists of $25 \times 25$ solutions -- 25 discrete viscosities in the range of $10^{-4.5}$ -- $10^{-2}$ and 25 trajectories, or different initial conditions, per viscosity. We compute the model error for each prediction and plot the average ensemble error as a function of viscosity.
The second test set contains 25 trajectories at $\nu=10^{-3}$. Here, we evolve the trajectories between $t=[0,10)$, over 3 times longer than the training set, to see how the models behave outside of the training regime. The average error across all trajectories is plotted as a function of time.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{plots/models.pdf}
\caption{Blue shaded regions indicate the section of the plot corresponding to the training regime, while red shading indicates a test regime. (a) Plot of the ensemble mean-squared error relative to the ground truth DNS for each model as a function of viscosity. All models generally improve over the no closure baseline except $anode$, but only $smagorinsky$ improves of the constant Smagorinsky model. (b) Ensemble MSE for each model as a function of time. The inclusion of a production term gives rise to anomalous behavior in the extrapolation region.}
\label{fig:all_models}
\end{figure*}
Two non-trainable models \textit{none} and \textit{smag-const} provide some bounding benchmarks for model comparison. In \autoref{fig:all_models}(a), at the upper end, the \textit{none} model with no closure results in high error at low viscosities, where the dissipative length scales are small, with progressively lower error at higher viscosities where the dissipative length scales become larger. An effective closure model should at the very least perform better than this, otherwise no additional physics have been learned to improve over the baseline governing equations solved on a coarse grid. At the lower end, the \textit{smag-const} model is a well-known closure model that has been shown to perform well in a wide range of viscosities. An ideal model would produce lower loss than this baseline, indicating that the neural closure functional form can show improvements over classical closure models from theory.
Model performance is delineated well by model accuracy relative to the baselines in \autoref{fig:all_models}(a). Model \textit{anode} has the highest loss, greater than the \textit{none} baseline, thus showing no improvement over the governing equations. Given that the \textit{anode} model contains no inherent physics or structural form to the PDE, this result is unsurprising as the model must learn both the governing dynamics and any closure approximations. Despite \textit{resnet} having a similar architecture to \textit{anode}, \textit{resnet} can achieve marginally better accuracy, although it is still no better than the \textit{none} baseline. Theoretically, \textit{anode} should be able to learn the same optima as \textit{resnet}, since \textit{anode} has the same functional form as \textit{resnet}, except solved using an adaptive time-stepper. However, actually reaching those optima can be challenging during optimization due to the ODE inductive bias.
Models \textit{transport-I} and \textit{transport-I-p} show marginal improvement over \textit{none}, with \textit{transport-I-p} the better of the two. These models were particularly difficult to train, producing very stiff ODEs that made integration difficult. This is apparent in the fact that some of the viscosities are not shown as the stiffness of those systems led to intractable solutions. We hypothesize that this could be due to overfitting the training set, since all of the training samples could be solved in a reasonable amount of time.
Model \textit{direct} lies firmly between the two benchmarks. Accuracy is relatively uniform with little dependency on viscosity. Thus, while \textit{direct} consistently improves over \textit{none} at low viscosities, the advantage decreases with increasing viscosity, becoming a disadvantage at the highest viscosities.
Models \textit{transport-II} and \textit{transport-II-p} have accuracies comparable to \textit{smag-const} and perform almost identically to each other within the viscosity test. Generally, we see a consistent positive bias over the baseline closure benchmark that diminishes only at the highest viscosities.
Model \textit{smag-ml} is the only model with lower error than \textit{smag-const} everywhere in \autoref{fig:all_models}(a). Given that the model leverages the functional form of the classical Smagorinsky model and only adds a spatial dependency, again this result is perhaps unsurprising. However, it is promising that the improvements persist beyond the training region.
Overall, besides models \textit{none}, \textit{transport-I}, and \textit{transport-I-p}, accuracy is mostly independent of the viscosity within this range. While models have been trained on viscosities between $10^{-4}$ and $10^{-3}$, they can be used on viscosities well outside this range, with little to no erroneous behavior outside the training regime.
\autoref{fig:all_models}(b) shows model accuracy on the second test set. The blue region indicates the time span the model was trained on, while the red region shows the extrapolation range. The model hierarchy by accuracy is generally preserved here, i.e. \textit{anode} performs the worst, \textit{smag-ml} the best, and \textit{transport-II} and \textit{transport-II-p} are comparable to \textit{smag-const}.
However, the stability test highlights the effect of the production term in the models. The models largely have a monotonically decreasing slope in the error over time, except for \textit{anode}, \textit{transport-I-p}, and \textit{transport-II-p}, which display aberrant behavior in the extrapolation range. The inclusion of the production term results in improvements within the training time regime, and in the case of \textit{transport-I} and \textit{transport-I-p}, the difference is significant. Outside of this range, the production term can hinder the solution and lead to growing errors. Deviations between \textit{transport-II} and \textit{transport-II-p} can be seen at $t>5$, and error from \textit{transport-I-p} grows to match \textit{transport-I} at $t=10$.
\begin{table}[h]
\centering
\begin{tabularx}{0.6\linewidth}{rm{2pt}|m{2pt}X}
\textbf{Model} &&& \textbf{Relative evaluation time} \\ \hline
\textit{none} &&& 1.0 \\
\textit{anode} &&& 2.54 \\
\textit{resnet} &&& 0.37 \\
\textit{direct} &&& 2.97 \\
\textit{transport-I} &&& 3.36 \\
\textit{transport-I-p} &&& 3.40 \\
\textit{transport-II} &&& 3.48 \\
\textit{transport-II-p} &&& 3.52 \\
\textit{smag-ml} &&& 3.24 \\
\textit{smag-const} &&& 1.36 \\
\end{tabularx}
\caption{Table of evaluation times for the models on the second test set. Times are normalized by the \textit{none} model, which is the cheapest differential equation-based model to compute.}
\label{tab:model_evals}
\end{table}
\autoref{tab:model_evals} shows the model evaluation times relative to \textit{none}. All of the models except for \textit{resnet} are more expensive than this baseline, which is expected considering that the models require more operations and Fourier transforms to compute than the governing dynamics. We found that the model evaluation times have limited correlation with the number of operations needed at each time step, and are more heavily influenced by the stiffness of the resulting ODE system, given the adaptive integration scheme. Thus, \textit{resnet} is much more efficient than other models because it requires the fewest function evaluations, coinciding with the fixed Euler time steps.
We also show training curves for this experiment in \autoref{fig:model_training}, including the train loss and training time as a function of epoch.
We point out the train loss of \textit{resnet} in \autoref{fig:model_training}(a), which is vastly different than the test loss in \autoref{fig:all_models}, different behavior from the other models tested. We believe the use of differential equation inductive biases in the other models close the learning gap between training and testing data distributions, promoting data-efficient and generalizable models.
\autoref{fig:model_training}(b) clearly demonstrates the effect of ODE stiffness. Since the number of operations per function evaluation in each model is fixed, one might expect the training time to be linear with epoch. However, as models like \textit{transport-I} and \textit{direct} evolve in training, the curves become steeper, indicating more function evaluations required for time integration.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{plots/model_training.pdf}
\caption{Training curves for the models in the first experiment. (a) MSE loss of the models, normalized by \textit{none}, as a function of epoch. There is significant variability between the models in time to convergence and magnitude of the overall change in loss. (b) Training time as a function of epoch. For some models, the computational cost of evaluation remains constant, leading to linear behavior. For others, the computational cost increases with training iteration, demonstrating non-linear behavior with a positive second derivative.}
\label{fig:model_training}
\end{figure*}
\subsection{Resolution invariance}
In the previous experiment, we have shown that many of the models generalize well over a wide range of viscosities and time. In this experiment, we wish to examine our approach with regards to the coarse-grain scale. In the previous experiment, we fixed $M=64$, here we will vary $M$ and thus the cutoff of the low-pass filter, $k=M/2$.
Again, we generate a training dataset of 100 Burgers solutions. Since we desire for all the cutoffs $M/2$ to lie in the inertial region of the flow spectrum, we slightly modify the distribution of the initial condition, the time span for integration, and the ground truth grid density $N$, realizing a wider inertial range. We divide the set into 4 batches with $M=64,128,256,512$. The dataset parameters are tabulated in \autoref{tab:res_params}.
\begin{table}[h]
\centering
\begin{tabularx}{0.7\linewidth}{cm{1pt}|m{2pt}X}
\textbf{Parameter} &&& \textbf{Range} \\ \hline
$\nu$ &&& $[10^{-4},10^{-3})$\\
$\hat u_k$ &&& $[-0.05,0.05)+[-0.05,0.05)i$, $k=0..32$\\
$t$ &&& $[0,6)$\\
$N$ &&& 8192\\
$M$ &&& 64, 128, 256, 512\\
\end{tabularx}
\caption{Parameters and their distributions used to generate the training set for the second experiment, including the viscosity $\nu$, the Fourier modes of the initial velocity field $\hat u_k$, the timespan for integration $t$, the DNS grid resolution $N$, and the coarse grid resolutions $M$.}
\label{tab:res_params}
\end{table}
We train the two best performing models from the previous experiment, \textit{transport-II} and \textit{smag-ml}, as well as two modifications to the \textit{smag-ml} model to understand the impact of certain neural network architectural choices. These models are trained for 200 epochs, which was sufficient for convergence, again using the ADAM optimizer with learning rate varying from $10^{-3}$ to $10^{-1}$ depending on the model.
Specifically, we add two models, \textit{smag-ml-small} and \textit{smag-ml-local}. \textit{smag-ml-small} reduces the number of hidden channels in the network to 2 from 128. Since the FNO architecture requires each of the channels to be Fourier transformed and inverse transformed at each timestep, the size of the hidden layer has a large impact on the computational efficiency of the model. Therefore, we wish to see how small the model can be made without sacrificing accuracy. \textit{smag-ml-local} switches the FNO layers for standard linear layers resulting in a purely local closure correction. The FNO is a non-local operator given the pointwise multiplication in Fourier space. Replacement with a linear layer gives only pointwise multiplication by a linear operator in real space. Thus, we can compare the efficacy of a non-local functional form with a local functional form.
To test the models, we generate a new test set that includes variation in $M$. Here, we look at a set of $13 \times 25$ trajectories -- 13 discrete $M$'s in the range of $2^6$-$2^9$, including interpolation between the training discretizations, and 25 samples per $M$. The $\nu$ and $\hat u_k$ values are drawn from the same distributions as the training set, and $t=[0,20)$.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{plots/resolution.pdf}
\caption{The ensemble mean-squared error of various models normalized by the MSE of the constant Smagorinsky model, one of the baseline we hope to improve over. The adaptive ML Smagorinsky models can reduce the error up to 40\%, with \textit{smag-ml-small} demonstrating the best performance. \textit{smag-dyn} can only achieve lower error than \textit{smag-const} at high resolutions, where the inertial region is more fully resolved.}
\label{fig:resolution}
\end{figure*}
\autoref{fig:resolution} shows the MSE of the models normalized by the constant Smagorinsky baseline. Since the training dataset encompasses variations in both $\nu$ and $M$, we train the Smagorinsky constant via gradient descent on the same training dataset as the models to obtain the optimal constant instead of hand-tuning. The training procedure converged clearly to a value of $C_s=0.4$ as shown in \autoref{fig:Cs}. We also include a comparison to the non-trainable \textit{smag-dyn} baseline, given that most of the models in this experiment are Smagorinsky model derivatives.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{plots/Cs.pdf}
\caption{Convergence of the Smagorinsky constant $C_s$ to 0.4 as optimized by gradient descent on the training dataset.}
\label{fig:Cs}
\end{figure}
Similar to the previous experiment, \autoref{fig:resolution} shows that model \textit{transport-II} results in greater error than \textit{smag-const}, while the adaptive Smagorinsky approaches can improve over the constant model. Even when the ideal $C_s$ is fit to the data, the spatiotemporally varying models outperform the baseline. Within the adaptive trainable Smagorinsky models, the local correction can improve over the constant baseline at all resolutions, though not as much as the other two models. The two non-local models generally perform better than all others, which fits with our understanding that turbulence is a non-local phenomenon. However, surprisingly, the \textit{smag-ml-small} model ultimately shows a universal advantage over both the dynamic and constant Smagorinsky baselines. Comparing the training and validation losses of the models provides some insight. The overparameterized \textit{smag-ml} model overfits the training set resulting in worse performance on the test set. While this may be ameliorated by a larger training dataset, we do not pursue this route and leave this study for future work. Instead, we show that we can learn an effective and computationally efficient correction to the baseline Smagorinsky model with limited training data, due to our differentiable physics approach to end-to-end training of ML models.
\begin{table}[h]
\centering
\begin{tabularx}{0.6\linewidth}{rm{2pt}|m{2pt}X}
\textbf{Model} &&& \textbf{Relative evaluation time} \\ \hline
\textit{none} &&& 1.0 \\
\textit{smag-const} &&& 1.33 \\
\textit{smag-dyn} &&& 2.35 \\
\textit{smag-ml} &&& 3.65 \\
\textit{smag-ml-small} &&& 3.47 \\
\textit{smag-ml-local} &&& 1.77 \\
\textit{transport-II} &&& 3.97 \\
\end{tabularx}
\caption{Table of evaluation times for the models used in the resolution experiment. Times are normalized by the \textit{none} model, which is the cheapest to compute.}
\label{tab:res_evals}
\end{table}
\autoref{tab:res_evals} lists the normalized evaluation times for the models in this experiment. While the FNO-based models are unequivocally more expensive than any of the baselines, \textit{smag-ml-local} provides an interesting compromise between computational cost and accuracy, achieving lower error than the baselines everywhere except for the highest resolutions, where \textit{smag-dyn} improves over \textit{smag-const}.
The training curves for the models are shown in \autoref{fig:resolution_training}. From \autoref{fig:resolution_training}(a), convergence was achieved in 200 epochs and from \autoref{fig:resolution_training}(b), these models did not display the same kind of non-linear behavior in the training time. \textit{smag-ml-small} was much faster to train than \textit{smag-ml}, despite the comparable evaluation times.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{plots/res_training.pdf}
\caption{Training curves for the models in the second experiment. (a) MSE loss of the models, normalized by \textit{none}, as a function of epoch. The inductive biases of these models lead to fast convergence. (b) Training time as a function of epoch. Although \textit{smag-ml} and \textit{smag-ml-small} have comparable final evaluation costs, the differences become significant during training, in terms of overall optimization time.}
\label{fig:resolution_training}
\end{figure*}
Lastly, we show some examples of predictions from model \textit{smag-ml-small} compared to \textit{smag-const}. \autoref{fig:64} displays results from $M=64$, where the cutoff wavenumber is close to the energy containing region of the spectra. \autoref{fig:512} presents results from $M=512$, where the cutoff wavenumber is on the other side of the inertial region close to the dissipative region. The snapshots of the velocity field in (a) and the Smagorinsky constant and eddy viscosity in (b) are taken at $t=10$. (c) compares the DNS and model energy spectra at $t=10$, where the cutoff wavenumber is indicated by the model spectrum. (d) illustrates the MSE loss of the model \textit{smag-ml-small} and \textit{smag-const} over time, including the relative loss of the model to the constant Smagorinsky baseline.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{plots/64.pdf}
\caption{Statistics and snapshots comparing results from closure model \textit{smag-ml-small} to a baseline and DNS with $M=64$. (a) Sample snapshot of the velocity field from DNS, the \textit{smag-const} baseline, and \textit{smag-ml-small}. (b) Sample snapshot of the closure model eddy viscosity field and spatially varying Smagorinsky constant from \textit{smag-ml-small}. (c) Snapshot of the energy spectra of the sample from DNS and \textit{smag-ml-small}. The cutoff $M/2$ lies in the very upper region of the inertial range. (d) MSE of the model and baseline over time.}
\label{fig:64}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{plots/512.pdf}
\caption{Statistics and snapshots comparing results from closure model \textit{smag-ml-small} to a baseline and DNS with $M=512$. (a) Sample snapshot of the velocity field from DNS, the \textit{smag-const} baseline, and \textit{smag-ml-small}. (b) Sample snapshot of the closure model eddy viscosity field and spatially varying Smagorinsky constant from \textit{smag-ml-small}. (c) Snapshot of the energy spectra of the sample from DNS and \textit{smag-ml-small}. The cutoff $M/2$ lies very close to the dissipative range. (d) MSE of the model and baseline over time.}
\label{fig:512}
\end{figure*}
\section{Conclusions}
\label{sec:concl}
In this work, we systematically investigate the efficacy of various machine learning based closure models for Burgers turbulence using an end-to-end learning framework. We aim to develop a better understanding of how the functional form of the data-driven closure impacts its ability to accurately model subgrid scales, and seek to design a data-driven closure that is broadly applicable and ideally outperforms baseline closure models. Our models are discretization independent and thus defined by their PDE form and neural network hyperparameters. Models are trained in an end-to-end fashion, meaning the entire closure PDE is solved with a pseudo-spectral method at each iteration of the optimization procedure. Differentiable programming closes the gap between typical \textit{a priori} learning and separate \textit{a posteriori} testing that is common in ML closure modeling today. Given a differentiable solution algorithm, models can be trained via gradient descent directly on an \textit{a posteriori} loss function. This approach allows for great flexibility in model design. Just as neural network hyperparameters can be easily tuned and adjusted, the PDE form of the closure can be modified to incorporate various degrees of known physics or physical assumptions. Once the high level PDE has been defined, the system is solved and evaluated on the loss function, the MSE of the velocity field relative to a ground truth DNS simulation.
We asses the models on a wide range of Burgers systems to determine their generalizability. Given the spread of Reynolds numbers encountered in flow problems and the need for a closure model that can accommodate this range, we test our models' performance on unseen Burgers systems with viscosities spanning 2.5 orders of magnitude, including interpolation of the training viscosities and extrapolation outside of this distribution. We also check the stability of the models by extrapolation in time.
We find that models that incorporate more physical assumptions result in lower errors on the test set than those without such inductive biases. For example, use of the Boussinesq hypothesis in \textit{transport-II} produces significantly better approximations than \textit{transport-I}, which only models a stress analogue directly. The \textit{resnet} and \textit{anode} models have no assumptions of the PDE form, meaning they must learn all the complex dynamics of the Burgers system with limited training data, resulting in unsuccessful models. Increasing the size of the training set may have allowed for improved learning, however we are specifically interested in data-efficient models given the cost of generating ground truth DNS simulations. At the other extreme, we leverage all the assumptions of the baseline Smagorinsky model and only apply a small modification to learn a spatially varying Smagorinsky constant. Here, we are able to show universal improvement over the baseline Smagorinsky models at all viscosities.
We take our analysis further by demonstrating the resolution invariance of our closure approach. Subgrid closures model flow scales smaller than their cutoff wavenumber, $k=M/2$ in our case. This cutoff should lie in the inertial region of the energy spectra, but where exactly within the region it lies may be variable depending on the coarse grid resolution and system parameters. A universal closure model should be able to adapt to different cutoff wavenumbers, which impacts the range of energy scales it must model. In our second experiment, we examine the adaptive trainable Smagorinsky models in more detail with regards to different coarse grid resolutions. We train the model on four resolutions ranging from cutoffs very close to the dissipative region to cutoffs close to the energy containing region. We show that the model can continue to outperform the constant model at all resolutions, including interpolation between the training resolutions. To the best of our knowledge, this is the first use of a differentiable physics based closure on varying uniform grids.
Finally, we show the impact of two neural network architectural choices -- the size of the hidden layer in the network and the non-locality of the FNO functional form. It is typical in machine learning approaches to create very large networks with 10s of thousands of parameters to create very expressive networks that can act as universal approximators. However, we have included a great deal of inductive bias in our model simply by leveraging the Smagorinsky form, meaning our network only has to learn a small correction to see improvement. We demonstrate that a small network of 520 parameters is enough to learn an effective correction, even showing lower loss than the larger model which can overfit the small training set. Lastly, we see the impact of the non-local FNO by testing a model with linear layers. While indeed the local adaptive model improves over the baseline, the advantage is not as significant as the non-local model, although its computational efficiency is greater.
Closure modeling remains a significant challenge in the fluid dynamics community. Machine learning is enabling a new class of data-driven models, but their performance on arbitrary flows can be difficult to characterize. We use the Burgers system as a test bed for developing subgrid models in a differentiable framework, training on an \textit{a posteriori} loss and testing on a large variety of Burgers systems. We show that a simple correction to existing closure models using data-efficient neural networks can result in better performance.
\newpage
\section{Acknowledgements}
This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE 1745016 awarded to VS. The authors from CMU acknowledge the support from the Technologies for Safe and Efficient Transportation University Transportation Center, and Mobility21, A United States Department of Transportation National University Transportation Center. This work was supported in part by Oracle Cloud credits and related resources provided by the Oracle for Research program. RM is supported by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research, under Contract~DE-AC02-06CH11357.
\setlength\rotFPtop{-6pt}
\begin{sidewaystable}
\centering
\caption{Model descriptions}
\label{tab:models}
\begin{tabularx}{\textwidth}{|r|X|m{72pt}|m{72pt}|m{72pt}|X|}
\hline
\textbf{Name} & \centering\textbf{Equations} & \centering\textbf{Time-stepper} & \centering\textbf{Layer type} & \centering\textbf{Hidden channels} & \textbf{Notes} \\
\hline \hline
\textit{none} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u$ &
Tsit5 & N/A & N/A & \\ \hline
\textit{smag-const} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\nu_t\nabla \bar u)$,\newline
$\nu_t = (C_s\delta x)^2|\nabla \bar u|$ &
Tsit5 & N/A & N/A & This model contains no neural networks, but $C_s$ is a potentially trainable parameter.\\ \hline
\textit{smag-dyn} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\nu_t\nabla \bar u)$,\newline
$\nu_t = (C_s\delta x)^2|\nabla \bar u|$, \newline
$(C_s\delta x)^2 = \frac{\langle HM \rangle}{\langle M^2 \rangle}$, \newline
$H = \nabla(\widetilde{\bar{u}}^2/2) - \nabla(\widetilde{\bar{u}^2}/2)$, \newline
$M = \kappa^2 \nabla(|\nabla \widetilde{\bar{u}}|\nabla \widetilde{\bar{u}}) - \nabla(\widetilde{|\nabla \bar{u}|\nabla \bar{u}})$ &
Tsit5 & N/A & N/A & \\ \hline \hline
\textit{resnet} &
$(\bar{u}_{t+\Delta t}, \eta_{t+\Delta t}) = (\bar{u}_{t}, \eta_{t}) + f_\theta(\bar u_t,\eta_t,\nu;x)\Delta t$, \newline
$\eta_0 = g_{\theta}(\bar u_0; x)$ &
Euler & FNO & 128 & \\ \hline
\textit{anode} &
$\dot{\bar{u}}, \dot \eta = f_\theta(\bar u,\eta,\nu;x)$, \newline
$\eta_0 = g_{\theta}(\bar u_0; x)$ &
Tsit5 & FNO & 128 & \\ \hline
\textit{direct} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \eta$, \newline
$\eta = f_\theta(\nabla \bar u, \nu; x)$ &
Tsit5 & FNO & 128 & \\ \hline
\textit{transport-I} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla\eta$, \newline
$\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta$,\newline
$\alpha, \beta = f_\theta(\nabla \bar u, \nabla \eta, \nu; x)$,\newline
$\eta_0 = g_{\theta}(\bar u_0, \nu; x)$ &
Tsit5 & FNO & 128 & \\ \hline
\textit{transport-II} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\eta\nabla \bar u)$, \newline
$\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta$,\newline
$\alpha, \beta = f_\theta(\nabla \bar u, \nabla \eta, \nu, \delta x; x)$,\newline
$\eta_0 = g_{\theta}(\bar u_0, \nu; x)$ &
Tsit5 & FNO & 128 & The input $\delta x$ to $f_\theta$ is only used in the resolution study.\\ \hline
\textit{transport-I-p} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla\eta$, \newline
$\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta + \gamma$,\newline
$\alpha, \beta, \gamma = f_\theta(\nabla \bar u, \nabla \eta, \nu; x)$,\newline
$\eta_0 = g_{\theta}(\bar u_0, \nu; x)$ &
Tsit5 & FNO & 128 & \\ \hline
\textit{transport-II-p} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\eta\nabla \bar u)$, \newline
$\dot \eta = \alpha \cdot \nu\nabla^2\eta - \beta \cdot \bar u\nabla\eta + \gamma$,\newline
$\alpha, \beta, \gamma = f_\theta(\nabla \bar u, \nabla \eta, \nu; x)$,\newline
$\eta_0 = g_{\theta}(\bar u_0, \nu; x)$ &
Tsit5 & FNO & 128 & \\ \hline
\textit{smag-ml} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\nu_t\nabla \bar u)$, \newline
$\nu_t = (\eta\delta x)^2|\nabla \bar u|$,\newline
$\eta = f_\theta(\nabla \bar u, \nu, \delta x; x)$ &
Tsit5 & FNO & 128 & The input $\delta x$ to $f_\theta$ is only used in the resolution study.\\ \hline
\textit{smag-ml-local} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\nu_t\nabla \bar u)$, \newline
$\nu_t = (\eta\delta x)^2|\nabla \bar u|$,\newline
$\eta = f_\theta(\nabla \bar u, \nu, \delta x; x)$ &
Tsit5 & Linear & 128 & \\ \hline
\textit{smag-ml-small} &
$\dot{\bar{u}} = \nu\nabla^2\bar u - \bar u\nabla \bar u + \nabla(\nu_t\nabla \bar u)$, \newline
$\nu_t = (\eta\delta x)^2|\nabla \bar u|$,\newline
$\eta = f_\theta(\nabla \bar u, \nu, \delta x; x)$ &
Tsit5 & FNO & 2 & \\ \hline
\end{tabularx}
\end{sidewaystable}
\clearpage
\section{}
\subsection{}
\subsubsection{}
|
\section{Preliminaries}\label{preliminaries}
In this section we give some definitions in our terminology and collect facts which are used in the paper.
\subsection{Notation and definitions}
Let $A$ be an Artin $R$-algebra, that is, $R$ is a commutative Artin ring and $A$ is an $R$-algebra which is finitely generated as an $R$-module. As usual, $A^{op}$ stands for the opposite algebra of $A$. Denote by $A$-mod the category of finitely generated left $A$-modules. The kernel, image and cokernel of a homomorphism $f$ in $A$-mod are denoted by ${\rm Ker}(f)$, ${\rm Im}(f)$ and ${\rm Coker}(f)$, respectively.
Let $T$ be an $A$-module. We define
$$^{\perp}({_A}T):=\{M\in A{\text-}{\rm mod}\ |\ \mbox{\rm Ext}_{A}^i(M,T)=0\ for \ all\ i\geqslant1 \}.$$
Similarly, $({_A}T)^{\perp}$ can be defined. Denote by add$(_{A}T)$ the full subcategory of $A$-mod consisting of direct summands of finite direct sums of copies of $T$. The $n$-th syzygy of ${_A}T$ is denoted by $\Omega^{n}_{A}(T)$ for each $n\in\mathbb{N}$.
\subsection{Derived categories of exact categories}\label{DCEC}
An exact category $\mathcal{E}$ (in the sense of Quillen) is by definition an additive category endowed with a class of conflations closed under isomorphism and satisfying certain axioms (see \cite{Quillen,keller} for details). When the additive category is abelian, the class of conflations coincides with the class of short exact sequences. An additive functor $F:\mathcal{E}\to \mathcal{E'}$ between exact categories $\mathcal{E}$ and $\mathcal{E'}$ is said to be \emph{exact} if it
sends the conflations in $\mathcal{E}$ to the ones in $\mathcal{E'}$.
Let $\mathcal{E}$ be an exact category and $\mathcal{F}$ a full subcategory of $\mathcal{E}$. If $\mathcal{F}$ is closed under extensions in $\mathcal{E}$, then $\mathcal{F}$, endowed with the conflations in $\mathcal{E}$ having their terms in $\mathcal{F}$ , is an exact category, and the inclusion $\mathcal{F}\subseteq\mathcal{E}$ is
a fully faithful exact functor. In this case, $\mathcal{F}$ is called a \emph{fully exact subcategory} of $\mathcal{E}$ (see \cite[Section 4]{keller}). In an abelian category $\mathcal{A}$ with enough projective objects, resolving subcategories are fully exact subcategories.
Throughout the paper, we always regard resolving subcategories of $\mathcal{A}$ as exact categories.
Let $\mathcal{A}$ be an abelian category and $\mathcal{E}$ a fully exact subcategory of $\mathcal{A}$. Denote by $\mathscr{C}(\mathcal{E})$ the category of complexes over $\mathcal{E}$.
A complex $X\in{\mathscr{C}(\mathcal{E})}$
is said to be \emph{strictly exact} if it is an exact complex over $\mathcal{A}$ and all of its boundaries belong to $\mathcal{E}$. Let $\mathscr{K}_{ac}(\mathcal{E})$ be the full
subcategory of $\mathscr{K}(\mathcal{E})$ consisting of those complexes which are isomorphic to strictly exact complexes. Then
$\mathscr{K}_{ac}(\mathcal{E})$ is a full triangulated subcategory of $\mathscr{K}_{ac}(\mathcal{E})$ closed under direct summands. The \emph{unbounded derived
category} of $\mathcal{E}$, denoted by $\mathscr{D}(\mathcal{E})$, is defined to be the Verdier quotient of $\mathscr{K}(\mathcal{E})$ by $\mathscr{K}_{ac}(\mathcal{E})$. Similarly, the
bounded-below, bounded-above and bounded derived categories $\mathscr{D}^{+}(\mathcal{E})$, $\mathscr{D}^{-}(\mathcal{E})$ and $\mathscr{D}^{b}(\mathcal{E})$ can be defined.
In Section \ref{K-theory}, we need the following result (see \cite[Section 11]{keller} and \cite[Proposition A.5.6]{Positselski}).
\begin{lem}\label{lemma:2.3'} Let $\mathcal{F}$ and $\mathcal{E}$ be fully exact subcategories of $\mathcal{A}$ with $\mathcal{F} \subseteq \mathcal{E}$. Assume that $\mathcal{E}$ is closed under direct summands in $\mathcal{A}$ and the following two conditions hold:
\begin{enumerate}
\item[(a)] For an exact sequence $0 \to X \to Y \to Z \to 0$ in $\mathcal{A}$ with $X\in{\mathcal{E}}$, if $Y,Z\in{\mathcal{F}}$, then $X\in{\mathcal{F}}$.
\item[(b)] There is a natural number $n$ such that, for each object $E\in{\mathcal{E}}$, there is a long exact sequence in $\mathcal{A}$
$$0\to F_{n}\stackrel{f_{n}}\to\cdots\to F_{1}\stackrel{f_1}\to F_{0}\stackrel{f_0}\to E\to 0$$
\noindent with $F_{i}\in{\mathcal{F}}$ and ${\rm Im}(f_i)\in{\mathcal{E}}$ for all $0 \leqslant i \leqslant n$.
\end{enumerate}
\noindent Then the inclusion $\mathcal{F} \subseteq \mathcal{E}$
induces a triangle equivalence $\mathscr{D}(\mathcal{F})\to \mathscr{D}(\mathcal{E})$ which can be restricted to an equivalence $\mathscr{D}^{*}(\mathcal{F})\to \mathscr{D}^{*}(\mathcal{E})$ for any $*\in\{+,-,b\}$.
\end{lem}
\subsection{(semi-)Gorenstein-projective modules and $\mathcal{X}$-(co)resolutions}
A complete projective resolution over $A$ is by definition a (double infinite) exact complex
$P^{\bullet}: \cdots \to P^{-1}\stackrel{d_P^{-1}}\to P^{0}\stackrel{d_P^{0}}\to P^{1}\to\cdots$
of projective $A$-modules such that the complex $\mbox{\rm Hom}_{A}(P^{\bullet},A)$, obtained by applying $\mbox{\rm Hom}_A(-,A)$ to $P^{\bullet}$, is again exact.
\begin{definition} \emph{\cite{EJ1995,EJ}} An $A$-module $M$ is said to be \emph{Gorenstein-projective} if there is a complete projective resolution $P^{\bullet}$ over $A$ such that $M$ is isomorphic to the image of $d_P^{-1}: P^{-1}\to P^0$.
\end{definition}
Gorenstein-projective modules were called modules of G-dimension zero in \cite{AB1969} or totally reflexive modules in \cite[Section 2]{AM2002}.
A generalization of Gorenstein-projective modules is the following.
\begin{definition}\cite{RZ1} An $A$-module $M$ is said to be \emph{semi-Gorenstein-projective} provided that $\mbox{\rm Ext}_{A}^{i}(M,A)=0$ for all $i\geqslant1$.
\end{definition}
Denote by $\mathcal{GP}(A)$ and $\mathcal{SGP}(A)$ the categories of Gorenstein-projective and semi-Gorenstein-projective $A$-modules, respectively.
Then $\mathcal{GP}(A)$ and $\mathcal{SGP}(A)$ are resolving subcategories of $A\ensuremath{\mbox{{\rm -mod}}}$. Moreover,
$\mathcal{GP}(A)\subseteq \mathcal{SGP}(A)$, but the converse of the inclusion is not true in general (see \cite{RZ1} for examples).
Let $\mathcal{X}$ be a full subcategory of $A$-mod. An \emph{$\mathcal{X}$-resolution} of an $A$-module $M$ is by definition
a long exact sequence of $A$-modules $\cdots\to X_2\to X_{1}\to X_{0}\to M\to 0$
with $X_{i}\in{\mathcal{X}}$ for all $i$. The length of the resolution is defined to be the supremum of $i$ such that $X_i\neq0$.
Now, the \emph{$\mathcal{X}$-resolution dimension} of $M$, denote by res.dim$_{\mathcal{X}}(M)$,
is defined to be the minimal natural number $n$ such that $M$ has an $\mathcal{X}$-resolution
of length $n$, or $\infty$ if no such $n$ exists.
The concept of $\mathcal{X}$-coresolutions and $\mathcal{X}$-coresolution dimension of $M$ can be defined dually.
The following result is known in the literature.
\begin{lem}\label{lemma:2.1} Let $\mathcal{X}$ be a resolving subcategory of $A$-{\rm mod} and $M$ an $A$-module. The following are equivalent for a natural number $n$:
\begin{enumerate}
\item {\rm res.dim}$_{\mathcal{X}}(_{A}M)\leqslant n$.
\item $\Omega^n_{A}(M)\in{\mathcal{X}}$.
\item For any $\mathcal{X}$-resolution $\cdots\to X_2\to X_{1}\to X_{0}\to M\to 0$ of $M$, ${\rm Ker}(X_{n-1}\to X_{n-2})\in \mathcal{X}$.
\end{enumerate}
\end{lem}
Associated with $\mathcal{X}$, there are
$$\mathcal{X}^{\leqslant n}(A):=\{M\in{A\textrm{-mod}} \ | \ \textrm{res.dim}_{\mathcal{X}}(_{A}M)\leqslant n\},$$
$$\mathcal{X}^{<\infty}(A):=\{M\in{A\textrm{-mod}} \ | \ \textrm{res.dim}_{\mathcal{X}}(_{A}M)<\infty\}.$$
Clearly, if $\mathcal{X}$ is a resolving subcategory of $A$-{\rm mod}, then so are $\mathcal{X}^{\leqslant n}(A)$ and $\mathcal{X}^{<\infty}(A)$.
Denote by $\mathcal{P}(A)$ the category of projective $A$-modules. For simplicity, we set
$$\mathcal{P}^{\leqslant n}(A):=\mathcal{P}(A)^{\leqslant n}(A),\quad \mathcal{P}^{<\infty}(A):=\mathcal{P}(A)^{<\infty}(A),$$
$$\mathcal{GP}^{\leqslant n}(A):=\mathcal{GP}(A)^{\leqslant n}(A),\quad \mathcal{GP}^{<\infty}(A):=\mathcal{GP}(A)^{<\infty}(A),$$
$$\mathcal{SGP}^{\leqslant n}(A):=\mathcal{SGP}(A)^{\leqslant n}(A),\quad \mathcal{SGP}^{<\infty}(A):=\mathcal{SGP}(A)^{<\infty}(A).$$
\subsection{Resolving dualities induced by tilting modules}
An $A$-module $T$ is called an \emph{$n$-tilting} module (see \cite{Brenner,Happel-Ringel,Happel1988,Miyashita}) if the following conditions are satisfied:
\begin{enumerate}
\item[(T1)]
$T\in{\mathcal{P}^{\leqslant n}(A)}$, that is,
the projective dimension of $_{A}T$, denoted by proj.dim$(_{A}T)$, is at most $n$.
\item[(T2)] $\mbox{\rm Ext}_{A}^{j}(T,T)=0$ for all $j\geqslant1$.
\item[(T3)] There exists an exact sequence of $A$-modules
$0\to {_{A}A}\to T_{0}\to\cdots\to T_{n}\to 0$
with $T_{i}\in{\textrm{add}(_{A}T)}$ for all $0\leqslant i\leqslant n$.
\end{enumerate}
A module $_{A}T$ is said to be \emph{tilting} if it is $n$-tilting for some natural number $n$. If, in addition,
$\mathcal{P}^{<\infty}(A)\subseteq{^{\perp}(_{A}T)}$, then $_{A}T$ is said to be \emph{strong tilting} (see \cite{AR1991}).
Given a tilting module $_{A}T$ with $B:=\textrm{End}_{A}(T)^{op}$, we see that $T_{B}$ is
also a tilting module and $A$ is isomorphic to $\textrm{End}_{B^{op}}(T)$ as algebras.
The following theorem is readily deduced from \cite[Theorem 3.5]{Miyashita}.
\begin{thm}{\rm (Miyashita's duality)}\label{thm:2.3} For a tilting bimodule $_{A}T_{B}$, let
\vspace{2mm}
\begin{center}{$\mathcal{C}:={^{\bot}(_{A}T)}\cap{\mathcal{P}^{<\infty}(A)}$ and $\mathcal{D}:={^{\bot}(T_{B})}\cap{\mathcal{P}^{<\infty}(B^{op})}$.}
\end{center}
\vspace{2mm}
Then the restricted Hom-functors $\mbox{\rm Hom}_{A}(-,T)|_{\mathcal{C}}:\mathcal{C}\to\mathcal{D}$ and $\mbox{\rm Hom}_{B^{op}}(-,T)|_{\mathcal{D}}:\mathcal{D}\to\mathcal{C}$
are inverse resolving dualities. Further, if $_{A}T_{B}$ is {strong tilting}, then there are inverse resolving dualities
$\mathcal{P}^{<\infty}(A)\simeq{\mathcal{P}^{<\infty}(B^{op})}$.
\end{thm}
Recently, Huisgen-Zimmermann has supplemented Miyashita's duality by showing that resolving dualities between subcategories of $\mathcal{P}^{<\infty}(A)$ and $\mathcal{P}^{<\infty}(B^{op})$ are always afforded by tilting bimodules (see \cite[Theorem 1]{Huisgen}).
\begin{thm}{\rm (Huisgen-Zimmermann's correspondence)} \label{thm:2.4}Let $\mathcal{C} \subseteq \mathcal{P}^{<\infty}(A)$ and $\mathcal{D} \subseteq \mathcal{P}^{<\infty}(B^{op})$ be resolving
subcategories of $A\textrm{-}{\rm mod}$ and $B^{op}\ensuremath{\mbox{{\rm -mod}}}$, respectively.
If $F:\mathcal{C}\to\mathcal{D}$ and $G:\mathcal{D}\to\mathcal{C}$ are inverse resolving dualities, then there exists a tilting bimodule $_{A}T_{B}$ such that
\begin{enumerate}
\item[(a)] $F\cong\mbox{\rm Hom}_{A}(-,T)|_{\mathcal{C}}$ and $G\cong\mbox{\rm Hom}_{B^{op}}(-,T)|_{\mathcal{D}}$,
\item[(b)] $\mathcal{C}={^{\bot}(_{A}T)\cap{\mathcal{P}^{<\infty}(A)}}$ and $\mathcal{D}={^{\bot}(T_{B})}\cap{\mathcal{P}^{<\infty}(B^{op})}$, and
\item[(c)] $\mathcal{C}$ consists of those modules $M$ which have finite ${\rm add}(_{A}T)$-coresolutions and $\mathcal{D}$ consists of those modules $N$ which have finite ${\rm add}(T_{B})$-coresolutions.
\end{enumerate}
\end{thm}
By Theorem \ref{thm:2.4}(c), we have ${^{\bot}(_{A}T)\cap{\mathcal{P}^{<\infty}(A)}}={^{\bot}(_{A}T)}\cap\mathcal{P}^{\leqslant \ell}(A)$ and ${^{\perp}(T_{B})}\cap{\mathcal{P}^{<\infty}(B^{op})}={^{\perp}(T_{B})}\cap{\mathcal{P}^{\leqslant \ell}(B^{op})}$, where
$\ell$ denotes the projective dimension of ${_A}T$.
\section{Correspondence between Wakamatsu tilting modules and resolving dualities}\label{proof}
In the section, we discuss resolving subcategories of module categories related to Wakamatsu tilting modules
and establish relationships between resolving dualities and Wakamatsu tilting modules.
In case that these dualities can be restricted to resolving subcategories of modules with finite projective, Gorenstein-projective or semi-Gorenstein-projective dimension, the associated Wakamatsu tilting modules are shown to be tilting. In particular, we show Theorem \ref{THM} and Corollary \ref{GV}.
\subsection{Basic facts on Wakamatsu tilting modules}
Let $T$ be an $A$-module. We denote by cogen$^{*}(_{A}T)$ the full subcategory of $A$-mod consisting of modules $M$
which admits an exact sequence of $A$-modules
$0\to M\to T_{0}\to T_{1}\to T_2\to \cdots$ with $T_{i}\in{\textrm{add}(_{A}T)}$ for all $i\geqslant 0$ such that applying the functor $\mbox{\rm Hom}_{A}(-,T)$ to the sequence still yields an exact sequence. Equivalent characterizations of cogen$^{*}(_{A}T)$ are given in the following result (for example, see
\cite[Lemmas 2.2 and 2.4]{Mabiao}).
\begin{lem}\label{lemma:3.1} Let $B={\rm End}_{A}(T)^{op}$. For an $A$-module $M$, the following statements are equivalent.
\begin{enumerate}
\item $M\in{{\rm cogen}^{*}(_{A}T)}$.
\item The map $\sigma_{M}:M\to \mbox{\rm Hom}_{B^{op}}(\mbox{\rm Hom}_{A}(M,T),T)$, $m\mapsto[f\mapsto f(m)]$ for $m\in M$ and $f\in\mbox{\rm Hom}_{A}(M,T)$ is an isomorphism and $\mbox{\rm Hom}_{A}(M,T)\in{^{\perp}(T_{B})}$.
\item The canonical maps $$\mbox{\rm Ext}^{i}_{A}(N,M)\to \mbox{\rm Ext}^{i}_{B^{op}}(\mbox{\rm Hom}_{A}(M,T),\mbox{\rm Hom}_{A}(N,T))$$
induced by the functor $\mbox{\rm Hom}_A(-,T)$ are isomorphisms for all $N\in{^{\perp}(_{A}T)}$ and $i\geqslant 0$.
\end{enumerate}
\end{lem}
In the introduction, we have defined $${\mathcal{W}{(_{A}T)}}:={^{\perp}(_{A}T)}\cap {\rm cogen}^{*}(_{A}T).$$
Clearly, $T\in {\mathcal{W}{(_{A}T)}}$ if and only if $\mbox{\rm Ext}_A^i(T,T)=0$ for all $i\geqslant 1$. Thanks to Lemma \ref{lemma:3.1}(3),
the functor $\mbox{\rm Hom}_A(-,T):A\ensuremath{\mbox{{\rm -mod}}}\to B^{op}\ensuremath{\mbox{{\rm -mod}}}$ can be restricted to a fully faithful functor ${\mathcal{W}{(_{A}T)}}\to B^{op}\ensuremath{\mbox{{\rm -mod}}}$ which preserves extension groups of modules. However, in general, ${\mathcal{W}{(_{A}T)}}$ is not a resolving subcategory of $A$-mod since it
may not contain projective $A$-modules. Following \cite{Wakamatsu}, an $A$-module $T$ is said to be \emph{Wakamatsu tilting} if
$T\in {\mathcal{W}{(_{A}T)}}$ and $_{A}A\in{\mathcal{W}{(_{A}T)}}$.
This is also equivalent to the following two conditions:
$(1)$ $\textrm{End}_{B^{op}}(T)\cong A$, where $B=\textrm{End}_{A}(T)^{op}$;
$(2)$ $\mbox{\rm Ext}_{A}^{i}(T,T)=0=\mbox{\rm Ext}_{B^{op}}^{i}(T,T)$ for all $i\geqslant1$.
\noindent By these conditions, if $_{A}T$ is Wakamatsu tilting, then $T_{B}$ is also Wakamatsu tilting. So, we can say $T$ is a Wakamatsu tilting $A$-$B$-bimodule.
Now, we collect some basic properties of Wakamatsu tilting modules.
\begin{lem}{\rm(}\cite[Proposition 2.11]{Mantese}{\rm)}\label{proposition:3.2}
Let $T$ be a Wakamatsu tilting $A$-module. Then:
(1) $\mathcal{W}{(_{A}T)}$ is a resolving subcategory of $A$-{\rm mod}.
(2) $\mathcal{W}{(_{A}T)}\cap{\mathcal{W}{(_{A}T)}}^{\perp}={\rm add}(_{A}T)$ and
$^{\perp}({\mathcal{W}{(_{A}T)}}^{\perp})=\mathcal{W}{(_{A}T)}$.
(3) $T$ is an injective cogenerator for $\mathcal{W}{(_{A}T)}$, that is, for any $X\in \mathcal{W}{(_{A}T)}$,
there is an exact sequence $0\to X\to I_0\to X_1\to 0$ in $A\ensuremath{\mbox{{\rm -mod}}}$
such that $I_0\in\mbox{\rm add}(_AT)$ and $X_1\in \mathcal{W}{(_{A}T)}$.
\end{lem}
For simplicity, sometimes we write $(X,Y)$ for $\mbox{\rm Hom}_{A}(X,Y)$ in our proofs for $A$-modules $X$ and $Y$.
\begin{lem}\label{prop:3.3} Let $_{A}T$ be a Wakamatsu tilting module. Then:
(1) $^{\perp}(_{A}T)\cap{\mathcal{P}^{\leqslant n}(A)}=
\mathcal{W}{(_{A}T)}\cap{\mathcal{P}^{\leqslant n}(A)}$ for any $n\geqslant0$;
(2) $_{A}T\in{\mathcal{P}^{<\infty}(A)}$ if and only if $_{A}T\in{\mathcal{GP}^{<\infty}(A)}$ and $\mathcal{GP}(A)\subseteq{\mathcal{W}{(_{A}T)}}$;
(3) If $_{A}T\in{\mathcal{P}^{<\infty}(A)}$, then
$\mathcal{W}{(_{A}T)}\cap{\mathcal{GP}^{\leqslant n }(A)}={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant n }(A)}$ for any $n\geqslant0$.
(4) If $_{A}T$ is a tilting module of projective dimension $\ell$, then $\mathcal{W}{(_{A}T)}\subseteq{\mathcal{GP}^{\leqslant \ell}(A)}$.
\end{lem}
\begin{proof} (1) It suffices to prove $^{\perp}(_{A}T)\cap{\mathcal{P}^{\leqslant n}(A)}\subseteq
{{\rm cogen}^{*}(_{A}T)}$. Let $M\in{^{\perp}(_{A}T)}\cap{\mathcal{P}^{\leqslant n}(A)}$. Then there exists an exact sequence in $A$-mod $$0\to P_{n}\to\cdots\to P_1\to P_0\to M\to 0,$$
where $P_{i}$ is projective for all $0\leqslant i\leqslant n$. This yields the following exact sequence in $B^{op}\ensuremath{\mbox{{\rm -mod}}}$:
$$0\to \mbox{\rm Hom}_{A}(M,T)\to \mbox{\rm Hom}_{A}(P_{0},T)\to \mbox{\rm Hom}_{A}(P_{1},T)\to \cdots \to \mbox{\rm Hom}_{A}(P_{n},T)\to 0.$$
Since $\mbox{\rm Hom}_{A}(P_{i},T)\in{\textrm{add}(T_{B})}\subseteq{^{\perp}(T_{B})}$ for all $0\leqslant i\leqslant n$ and since ${^{\perp}(T_{B})}$ is a resolving subcategory of $B^{op}\ensuremath{\mbox{{\rm -mod}}}$, we have $\mbox{\rm Hom}_{A}(M,T)\in{^{\perp}(T_{B})}$. It follows that
there is a commutative diagram with exact rows in $A$-mod:
$$\xymatrix{
0\ar[r] &P_{n}\ar[d]^{\sigma_{P_{n}}}\ar[r]&\cdots \ar[r]&P_{0}\ar[r]\ar[d]^{\sigma_{P_{0}}}&M\ar[r]\ar[d]^{\sigma_{M}}&0\\
0\ar[r] &((P_{n},T),T)\ar[r]&\cdots \ar[r]&((P_{0},T),T)\ar[r]&((M,T),T)\ar[r]&0.}$$
Since $\sigma_{P_{i}}$ is an isomorphism for all $0\leqslant i\leqslant n$, we see that $\sigma_{M}$ is an isomorphism.
Thus $M\in{{\rm cogen}^{*}(_{A}T)}$ by Lemma \ref{lemma:3.1}.
(2) Assume that $_{A}T\in{\mathcal{GP}^{<\infty}(A)}$ and $\mathcal{GP}(A)\subseteq{\mathcal{W}{(_{A}T)}}$. Since $_{A}T\in{\mathcal{GP}^{<\infty}(A)}$, we see from \cite[Lemma 2.17]{cfh} that there is an exact sequence
$0\to T\to H\to G\to 0 $ in $A$-mod, where $H\in{\mathcal{P}^{<\infty}(A)}$ and $G\in{\mathcal{GP}(A)}$. By $\mathcal{GP}(A)\subseteq{\mathcal{W}{(_{A}T)}}$,
the exact sequence $0\to T\to H\to G\to 0$ splits. Thus $_{A}T$ is isomorphic to a direct summand of $H$, which forces $T\in{\mathcal{P}^{<\infty}(A)}$.
Conversely, assume $_{A}T\in{\mathcal{P}^{<\infty}(A)}$. Let $M$ be an $A$-module in $\mathcal{GP}(A)$. There exists a complete projective resolution over $A$
$$P^{\bullet}:\cdots \to P^{-1}\stackrel{d^{-1}}\to P^{0}\stackrel{d^{0}}\to P^{1}\stackrel{d^{1}}\to\cdots$$
such that $M\cong \textrm{Im}(d^{-1})$. Set $K^{i}:=\textrm{Im}(d^{i})$ for all $i\in{\mathbb{Z}}$. Then $K^{i}\in{\mathcal{GP}(A)}$. Since $\mathcal{GP}(A)\subseteq{^{\perp}(_{A}T)}$, the complex $\mbox{\rm Hom}_{A}(P^{\bullet},T)$ is again exact. Hence we have the following commutative diagrams
$$\xymatrix{
0\ar[r] &M\ar[d]^{\sigma_{M}}\ar[r]&P^{0} \ar[d]^{\cong}\ar[r]&K^{1}\ar[r]\ar[d]^{\sigma_{K^{1}}}&0\\
0\ar[r]&((M,T),T)\ar[r]&((P^{0},T),T)\ar[r]&(((K^{1},T),T)}$$
and
$$\xymatrix{
0\ar[r] &K^{1}\ar[d]^{\sigma_{K^{1}}}\ar[r]&P^{1} \ar[d]^{\cong}\ar[r]&K^{2}\ar[r]\ar[d]^{\sigma_{K^{2}}}&0\\
0\ar[r]&((K^{1},T),T)\ar[r]&((P^{1},T),T)\ar[r]&((K^{2},T),T).}$$
This implies that both $\sigma_{K^{1}}$ and $\sigma_{M}$ are injective, and therefore $\sigma_{M}$ is an isomorphism by the snake lemma. Similarly, we can show that $\sigma_{K^{i}}:K^{i}\to((K^{i},T),T)$ is an isomorphism for any $i\in{\mathbb{Z}}$. Thus
$\mbox{\rm Ext}^{i}_{B}(\mbox{\rm Hom}_{A}(M,T),T)=0$ for all $i\geqslant1$. By Lemma \ref{lemma:3.1}, $M\in{{\rm cogen}^{*}(_{A}T)}$.
Since $M\in{{^{\perp}(_{A}T)}}$, we have $M\in{\mathcal{W}{(_{A}T)}}$.
(3) It suffices to show ${^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant n }(A)}\subseteq\mathcal{W}{(_{A}T)}$.
Let $M\in{^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant n }(A)}$. By \cite[Lemma 2.17]{cfh}, there exists an exact sequence $0\to M\to H\to G\to 0$ in $A$-mod with {\rm proj.dim$(_{A}H)\leqslant n$} and $G\in{\mathcal{GP}(A)}$. Since $G, M\in{^{\perp}(_{A}T)}$, we have $H\in {^{\perp}(_{A}T)}$. This forces $H\in{\mathcal{W}{(_{A}T)}}$ by $(1)$. Since $G\in{\mathcal{W}{(_{A}T)}}$ by (2), $M\in{\mathcal{W}{(_{A}T)}}$ by Lemma \ref{proposition:3.2}.
(4) Let $M\in\mathcal{W}{(_{A}T)}$. There exists a long exact sequence
$0\to M\to X_{0}\to X_{1}\to \cdots$
in $A$-mod with $X_{i}\in{\textrm{add}(_{A}T)}$ for all $i\geqslant0$ such that it stays exact after applying $\mbox{\rm Hom}_{A}(-,T)$. Since proj.dim$(_{A}T)=\ell$, we see from the horseshoe lemma that there exists an exact sequence in $A\ensuremath{\mbox{{\rm -mod}}}$
$$0\to \Omega^{\ell}_{A}(M)\to Q_{0}\to Q_{1}\to \cdots,$$
where $Q_{i}$ is projective for each $i\geqslant0$. Recall that there exists an exact sequence of $A$-modules
$0\to {_{A}A}\to T_{0}\to\cdots\to T_{\ell}\to 0$
with $T_{i}\in{\textrm{add}(_{A}T)}$ for all $0\leqslant i\leqslant \ell$. Since $M\in{^{\perp}(_{A}T)}$, there are isomorphisms $\mbox{\rm Ext}_{A}^{i}(\Omega^{\ell}_{A}(M),A)\cong\mbox{\rm Ext}_{A}^{\ell+i}(M,A)\cong\mbox{\rm Ext}_{A}^{i}(M,T_{\ell})=0$
for all $i\geqslant1$. This implies $M\in{\mathcal{GP}^{\leqslant \ell}(A)}$.
\end{proof}
The following corollary is an immediate consequence of Lemma \ref{prop:3.3}.
\begin{cor}\label{cor:3.5} Let $_{A}T$ be a Wakamatsu tilting module. Then:
(1) $^{\perp}(_{A}T)\cap{\mathcal{P}^{<\infty}(A)}=
\mathcal{W}{(_{A}T)}\cap{\mathcal{P}^{<\infty}(A)}$.
(2) If $_{A}T\in{\mathcal{P}^{<\infty}(A)}$, then
$\mathcal{W}{(_{A}T)}\cap{\mathcal{GP}^{<\infty }(A)}={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{<\infty}(A)}$.
(3) If $_{A}T$ is a tilting module of projective dimension $\ell$, then
$$\mathcal{W}{(_{A}T)}={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant \ell }(A)}={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{<\infty}(A)},$$
$$\mathcal{W}{(T_B)}={^{\perp}(T_{B})}\cap{\mathcal{GP}^{\leqslant \ell}(B^{op})}={^{\perp}(T_{B})}\cap{\mathcal{GP}^{<\infty}(B^{op})}.$$
\end{cor}
\subsection{Proofs of Theorem \ref{THM} and Corollary \ref{GV}}
We begin this subsection with a proof of Theorem \ref{THM}(1).
{\bf Proof of Theorem \ref{THM}(1)}. Recall from \cite[Theorem 23.5]{Anderson} that, given inverse (not necessarily resolving) dualities $F:\mathcal{C}\to\mathcal{D}$ and $G:\mathcal{D}\to\mathcal{C}$ with ${_A}A\in\mathcal{C}$ and $B_B\in\mathcal{D}$, there exists a faithfully balanced bimodule $_{A}T_{B}$ with $T_{B}\cong{F(_{A}A)}\in{\mathcal{D}}$ and $_{A}T\cong{G(B_{B})}\in{\mathcal{C}}$ such that $F\cong\mbox{\rm Hom}_{A}(-,T)|_{\mathcal{C}}$ and $G\cong\mbox{\rm Hom}_{B^{op}}(-,T)|_{\mathcal{D}}$. Now, assume that $F$ and $G$ are resolving dualities. They are exact functors between exact categories $\mathcal{C}$ and $\mathcal{D}$. Let $\delta: 0\to T\to X\to M\to0$ be an exact sequence in $A$-mod with $M\in{\mathcal{C}}$. Since $\mathcal{C}$ is closed under extensions in $A$-mod, $X$ belongs to $\mathcal{C}$. As $F$ is exact, the sequence $F(\delta): 0\to F(M)\to F(X)\to F(T)\to 0$ is exact. It follows from $F(T)\cong {FG(B_{B})}\cong{B_{B}}$ that $F(\delta)$ splits in $B^{op}\ensuremath{\mbox{{\rm -mod}}}$. Note that $GF$ is naturally isomorphic to the identity functor of $\mathcal{C}$. Thus $\delta$ splits in $A$-mod. This implies that $\mbox{\rm Ext}_{A}^{1}(M,T)=0$ for all $M\in{\mathcal{C}}$. Since $\mathcal{C}\subseteq A\ensuremath{\mbox{{\rm -mod}}}$ is a resolving subcategory, $\Omega^{i}_{A}(M)\in{\mathcal{C}}$ for all $M\in{\mathcal{C}}$ and $i\geqslant1$. Consequently, $\mbox{\rm Ext}_{A}^{i}(M,T)\cong{\mbox{\rm Ext}_{A}^{1}(\Omega^{i-1}_{A}(M),T)=0}$. This yields $\mathcal{C}\subseteq{^{\perp}(_{A}T)}$. In particular, $_AT\in{^{\perp}(_{A}T)}$. Similarly, one can prove $\mathcal{D}\subseteq{^{\perp}(T_{B})}$, which gives rise to $T_B\in {^{\perp}(T_{B})}$. Since ${_A}T{_B}$ is faithfully balanced, it is Wakamatsu tilting.
To prove (c), it suffices to show the inclusion $\mathcal{C}\subseteq{\mathcal{W}{(_{A}T)}}$, while the inclusion $\mathcal{D}\subseteq{\mathcal{W}{(T_{B})}}$ can be shown dually.
In fact, by $\mathcal{C}\subseteq{^{\perp}(_{A}T)}$, we only need to show $\mathcal{C}\subseteq$ cogen$^{*}(_{A}T)$.
For each $X\in \mathcal{C}$, let
$$\cdots\to Q_{1}\stackrel{g_{1}}\to Q_{0}\stackrel{g_{0}}\to F(X)\to 0$$
be a projective resolution of $F(X)$ in $B^{op}\ensuremath{\mbox{{\rm -mod}}}$. Clearly, $F(X)\in{\mathcal{D}}$ and $Q_{i}\in{\mathcal{D}}$ for all $i\geqslant0$. Since $\mathcal{D}$ is a resolving subcategory of $B^{op}$-mod, $\textrm{Ker}(g_{i})\in{\mathcal{D}}$ for all $i\geqslant0$. By the exactness of $G$, we obtain an exact sequence of $A$-modules $0\to X\to G(Q_{0})\to G(Q_{1})\to \cdots$
in which $G(Q_{i})\in{\textrm{add}(G(B_{B}))=\textrm{add}(_{A}T)}$ for all $i\geqslant0$.
Thus the exactness of $F$ implies $X\in$ cogen$^{*}(_{A}T)$. This finishes the proof of Theorem \ref{THM}(1). \hfill$\Box$
\medskip
To show Theorem \ref{THM}(2), we first show the following result.
\begin{lem}\label{lem:3.6} Let $_{A}T_{B}$ be the Wakamatsu tilting bimodule associated with inverse resolving dualities $F:\mathcal{C}\to \mathcal{D}$ and $G:\mathcal{D}\to \mathcal{C}$ in Theorem \ref{THM}(1).
Suppose that $\mathcal{U}\subseteq A\ensuremath{\mbox{{\rm -mod}}}$ and $\mathcal{V}\subseteq B^{op}\ensuremath{\mbox{{\rm -mod}}}$ are resolving subcategories.
Let $m$ and $n$ be natural numbers. Then:
$(1)$ $F$ and $G$ restrict to inverse dualities
$\mathcal{C}\cap\mathcal{U}^{\leqslant n}(A)\simeq\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})$
if and only if the following condition holds:
$$(\ast)\quad F(\mathcal{C}\cap\mathcal{U})\subseteq{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}\quad\mbox{and}\quad G(\mathcal{D}\cap\mathcal{V})\subseteq{\mathcal{C}\cap\mathcal{U}^{\leqslant n}(A)}.$$
$(2)$ Assume that the condition $(\ast)$ holds. The following statements are true.
\quad\; $(a)$ $\mathcal{C}\cap{\mathcal{U}^{<\infty}(A)}=\mathcal{C}\cap{\mathcal{U}^{\leqslant n}(A)}$. If $\mathcal{U}\subseteq{\mathcal{C}}$, then $\mathcal{C}\cap{\mathcal{U}^{<\infty}(A)}={^{\perp}(_{A}T)\cap{\mathcal{U}^{<\infty}(A)}}.$
\quad\; $(b)$ $\mathcal{D}\cap{\mathcal{V}^{<\infty}(B^{op})}=\mathcal{D}\cap{\mathcal{V}^{\leqslant m}(B^{op})}$. If $\mathcal{V}\subseteq{\mathcal{D}}$, then $\mathcal{D}\cap{\mathcal{V}^{<\infty}(B^{op})}={^{\perp}(T_{B})\cap{\mathcal{V}^{<\infty}(B^{op})}}.$
\quad\; $(c)$ If $\mathcal{U}\subseteq{\mathcal{SGP}(A)}$ and $\mathcal{V}\subseteq{\mathcal{SGP}(B^{op})}$, then $_{A}T_{B}$ is a tilting bimodule.
\end{lem}
\begin{proof} (1) The necessity of $(1)$ is clear since $\mathcal{U}\subseteq \mathcal{U}^{\leqslant n}(A)$ and $\mathcal{V}\subseteq \mathcal{V}^{\leqslant m}(B^{op})$. It suffices to show the sufficiency of $(1)$. Assume that the condition $(\ast)$ holds.
Let $M\in\mathcal{C}\cap\, \mathcal{U}^{\leqslant n}(A)$.
We need to show $F(M)\in{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}$.
Let $\cdots\to P_1\to P_0\to M\to 0$ be a projective resolution of $_{A}M$ and let $s:=\textrm{max}\{m,n\}$. Since $\mathcal{C}$ and $\mathcal{U}$ are resolving subcategories of $A\ensuremath{\mbox{{\rm -mod}}}$, $\Omega^{s}_{A}(M)\in{\mathcal{C}\cap\mathcal{U}}$ by Lemma \ref{lemma:2.1}. It follows from $F(\mathcal{C}\cap\mathcal{U})\subseteq{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}$ that $F(\Omega^{s}_{A}(M))\in{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}$.
Since $M\in\mathcal{C}\subseteq{^{\perp}(_{A}T)}$ by Theorem \ref{THM}(1), there is an exact sequence in $B^{\rm op}\ensuremath{\mbox{{\rm -mod}}}$:
$$0\to F(M)\to F(P_0)\to F(P_1)\to\cdots \to F(P_{s-1})\to F(\Omega^{s}_{A}(M))\to 0.$$
Observe that both $\mathcal{D}$ and $\mathcal{V}^{\leqslant m}(B^{op})$ are resolving subcategories of $B^{op}\ensuremath{\mbox{{\rm -mod}}}$, and so is
their intersection ${\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}$. Since $F(P_i)\in F(\mbox{\rm add}(_{A}A))\subseteq F(\mathcal{C}\cap\mathcal{U})$ for all $0\leqslant i\leqslant s-1$, we have $F(M)\in{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}$.
(2) It is enough to prove (a) and (c) since (b) can be shown dually.
(a) To prove $\mathcal{C}\cap{\mathcal{U}^{<\infty}(A)}=\mathcal{C}\cap{\mathcal{U}^{\leqslant n}(A)}$, it suffices to prove $\mathcal{C}\cap{\mathcal{U}^{<\infty}(A)}\subseteq\mathcal{C}\,\cap{\mathcal{U}^{\leqslant n}(A)}$. Let $X\in{\mathcal{C}\cap{\mathcal{U}^{<\infty}(A)}}$. Then there is a natural number $t$ such that $\Omega^{t}_{A}(X)\in{\mathcal{C}\cap\mathcal{U}}$
by Lemma \ref{lemma:2.1}. It follows from the proof of the sufficiency of $(1)$ that $F(X)\in{{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}}$. Since $X\cong{GF(X)}$, we have $X\in{\mathcal{C}\cap{\mathcal{U}^{\leqslant n}(A)}}$.
Suppose $\mathcal{U}\subseteq{\mathcal{C}}$. By Theorem \ref{THM}(1), we only need to check ${^{\perp}(_{A}T)\cap{\mathcal{U}^{<\infty}(A)}}\subseteq{\mathcal{C}\cap{\mathcal{U}^{<\infty}(A)}}$. Let $Y\in{^{\perp}(_{A}T)\cap{\mathcal{U}^{<\infty}(A)}}$. By Lemma \ref{lemma:2.1}, we have an exact sequence in $A$-mod:
$$0\to U\to Q_{t-1}\to \cdots\to Q_1\to Q_0\to Y\to 0,$$
where $U\in{\mathcal{U}}$ and $Q_i\in\mbox{\rm add}(_{A}A)$ for all $0\leqslant i\leqslant{t-1}\in\mathbb{N}$. By $Y\in{^{\perp}(_{A}T)}$, there is an exact sequence in
$B^{op}\ensuremath{\mbox{{\rm -mod}}}$:
$$0\to F(Y)\to F(Q_0)\to F(Q_1)\to \cdots\to F(Q_{t-1})\to F(U)\to 0.$$
Clearly, $F(Q_{i})\cong\mbox{\rm Hom}_{A}(Q_{i},T)\in{\textrm{add}(T_{B})}$ for all $0\leqslant i\leqslant{t-1}$. Since $\mathcal{U}\subseteq{\mathcal{C}}$, the condition $(\ast)$ implies $F(U)\in F(\mathcal{U})\subseteq{{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}}$. Similarly, we can show $F(Y)\in{{\mathcal{D}\cap\mathcal{V}^{\leqslant m}(B^{op})}}$.
By the exactness of $G$, the following sequence
$$0\to GF(U)\to GF(Q_{t-1})\to \cdots\to GF(Q_1)\to GF(Q_0)\to GF(Y)\to 0$$
is exact in $A\ensuremath{\mbox{{\rm -mod}}}$. Recall that $GF$ is naturally isomorphic to the identity functor of $\mathcal{C}$.
Since $U\in\mathcal{U}\subseteq\mathcal{C}$ and $P_i\in\mathcal{C}$ for all $i$, we have $Y\cong GF(Y)\in G(\mathcal{D})\subseteq{{\mathcal{C}}}$.
(c) Since $_{A}T\cong G(B_{B})$ by Theorem \ref{THM}(1), we obtain $_{A}T\in{\mathcal{C}\cap\mathcal{U}^{\leqslant{s}}(A)}$. As $\mathcal{C}$ and $\mathcal{U}$ are resolving subcategories of $A\ensuremath{\mbox{{\rm -mod}}}$, it follows from Lemma \ref{lemma:2.1} that $\Omega^{i}_{A}(T)\in{\mathcal{C}\cap\mathcal{U}}$ for all $i\geqslant s$. In particular, $\Omega^{s+1}_{A}(T)\in {\mathcal{C}\cap\mathcal{U}}$. By $(\ast)$, $F(\Omega^{s+1}_{A}(T))\in{\mathcal{D}\cap\mathcal{V}^{\leqslant s}(B^{op})}$.
Moreover, from the minimal projective resolution $\cdots\to P_2\to P_1\to P_0\to {_A}T\to 0$ of the module ${_A}T$, we obtain an exact sequence
$0\to B\to F(P_{0})\to F(P_{1})\to F(P_2)\to \cdots$ in $B^{op}\ensuremath{\mbox{{\rm -mod}}}$, where $F(P_{i})\cong\mbox{\rm Hom}_{A}(P_{i},T)\in{\textrm{add}(T_{B})}$ for all $i\geqslant0$. Since $\mathcal{D}\subseteq {^{\perp}(T_{B})}$ by Theorem \ref{THM}(1)(c) and $F(\Omega^{s+1}_{A}(T))\in\mathcal{D}$, it is clear that
$\mbox{\rm Ext}_{B^{op}}^{j}(F(\Omega^{s+1}_{A}(T)),F(P_{i}))=0$ for all $j\geqslant1$ and $i\geqslant0$.
Consequently, there are isomorphisms
$$\mbox{\rm Ext}_{B^{op}}^{1}\big(F(\Omega^{s+1}_{A}(T)),F(\Omega^{s}_{A}(T))\big)\cong
\mbox{\rm Ext}_{B^{op}}^{s+1}\big(F(\Omega^{s+1}_{A}(T)),B\big)\cong\mbox{\rm Ext}_{B^{op}}^1\big(\Omega_{B^{op}}^s(F(\Omega^{s+1}_{A}(T))),B\big).$$
By Lemma \ref{lemma:2.1}, $\Omega_{B^{op}}^s(F(\Omega^{s+1}_{A}(T)))\in\mathcal{V}$. Since $\mathcal{V}\subseteq{\mathcal{SGP}(B^{op})}={^{\perp}(B_B)}$ by assumption, $\mbox{\rm Ext}_{B^{op}}^1\big(\Omega_{B^{op}}^s(F(\Omega^{s+1}_{A}(T))),B\big)=0$, and further $\mbox{\rm Ext}_{B^{op}}^{1}\big(F(\Omega^{s+1}_{A}(T)),F(\Omega^{s}_{A}(T))\big)=0$
which leads to $F(\Omega^{s}_{A}(T))\in{\textrm{add}(T_{B})}$. Since $\Omega^{s}_{A}(T)\in \mathcal{C}$, we have $\Omega^{s}_{A}(T)\cong GF(\Omega^{s}_{A}(T))\in\mbox{\rm add}(_{A}A)$. This means that ${\rm proj.dim}(_AT)\leqslant s$. Similarly, we can show ${\rm proj.dim}(T_{B})\leqslant s$. Thus ${_A}T{_B}$ is tilting.
\end{proof}
{\bf Proof of Theorem \ref{THM}(2).}
The sufficiency of $(2)$ is a direct consequence of Lemma \ref{lem:3.6}. Now, assume that $_{A}T_{B}$ is a tilting bimodule with
$\ell={\rm proj.dim}(_AT)$. By Theorem \ref{THM}(1) and Corollary \ref{cor:3.5}(3), $\mathcal{C}\subseteq\mathcal{W}({_A}T)\subseteq\mathcal{GP}^{\leqslant \ell}(A)$ and $\mathcal{D}\subseteq\mathcal{W}(T_B)\subseteq \mathcal{GP}^{\leqslant \ell}(B^{op})$.
It follows that $\mathcal{C}\cap\mathcal{GP}^{\leqslant \ell}(A)=\mathcal{C}$ and $\mathcal{D}\cap\mathcal{GP}^{\leqslant \ell}(B^{op}) =\mathcal{D}$.
Since $\mathcal{GP}^{\leqslant \ell}(A)\subseteq\mathcal{SGP}^{\leqslant \ell}(A)$, we also have $\mathcal{C}\cap\mathcal{SGP}^{\leqslant \ell}(A)=\mathcal{C}$ and $\mathcal{D}\cap\mathcal{SGP}^{\leqslant \ell}(B^{op})=\mathcal{D}$.
Thus the assertions on (semi-)Gorenstein-projective modules in the necessity of $(2)$ automatically hold.
Let $\mathcal{U}:=\mbox{\rm add}(_AA)$ and $\mathcal{V}:=\mbox{\rm add}(B_B)$.
Since $\mathcal{C}\subseteq A\ensuremath{\mbox{{\rm -mod}}}$ is a resolving subcategory, we see that
$F(M)\in{\textrm{add}(G(B_{B}))=\textrm{add}(_{A}T)}$ for any $M\in{\mathcal{U}\cap\mathcal{C}}$. This forces $F(M)\in{\mathcal{D}\cap\mathcal{V}^{\leqslant \ell}}(B^{op})$. Dually, $G(N)\in{\mathcal{C}\cap\mathcal{U}^{\leqslant \ell}}(A)$ for any $N\in{\mathcal{V}\cap\mathcal{D}}$. By Lemma \ref{lem:3.6}(1), $F$ and $G$ can be restricted to inverse resolving dualities
$\mathcal{C}\cap{\mathcal{P}^{\leqslant \ell}(A)}\simeq \mathcal{D}\cap{\mathcal{P}^{\leqslant \ell}(B^{op})}$.
Further, by Lemma \ref{lem:3.6}(2), $\mathcal{C}\cap{\mathcal{P}^{<\infty}(A)}=\mathcal{C}\cap{\mathcal{P}^{\leqslant \ell}(A)}$ and $\mathcal{D}\cap{\mathcal{P}^{\leqslant \ell}(B^{op})}= \mathcal{D}\cap{\mathcal{P}^{<\infty}(B^{op})}$. This finishes the proof of $(2)$.\hfill$\Box$
\smallskip
{\bf Proof of Corollary \ref{GV}.}
(1) follows from Corollary \ref{cor:3.5}(3) and the resolving dualities $\mathcal{W}{(_{A}T)}\simeq\mathcal{W}{(T_{B})}$ by \cite[Theorem 1.5(d)]{GRS}.
(2) Suppose {$\mathcal{C}\subseteq {\mathcal{GP}^{\leqslant n }(A)}$ and $\mathcal{D}\subseteq {\mathcal{GP}^{\leqslant m}(B^{op})}$} for some $n, m\in\mathbb{N}$. Then $\mathcal{C}\cap\mathcal{GP}^{\leqslant n}(A)=\mathcal{C}$ and $\mathcal{D}\cap\mathcal{GP}^{\leqslant m}(B^{op}) =\mathcal{D}$. By Theorem \ref{THM}(2), ${_A}T_{B}$ is a tilting bimodule.
In Lemma \ref{lem:3.6}, we can take $\mathcal{U}:=\mathcal{GP}(A)$ and $\mathcal{V}:=\mathcal{GP}(B^{op})$, and then
the condition $(\ast)$ is satisfied. Further, assume {$\mathcal{GP}(A)\subseteq\mathcal{C}$ and $\mathcal{GP}(B^{op})\subseteq\mathcal{D}$. By Lemma \ref{lem:3.6}(2), $\mathcal{C}={^{\perp}(_{A}T)\cap{\mathcal{GP}^{<\infty}(A)}}$ and $\mathcal{D}={^{\perp}(_{A}T)\cap{\mathcal{GP}^{<\infty}(B^{op})}}$.
Since $T$ is tilting, we see from Corollary \ref{cor:3.5}(3) that $\mathcal{C}={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant \ell}(A)}$ and $\mathcal{D}={^{\perp}(T_{B})}\cap{\mathcal{GP}^{\leqslant \ell}(B^{op})}$. \hfill$\Box$
\medskip
Finally, we provide several equivalent characterizations for Wakamatsu tilting bimodules to be tilting. This may be helpful for understanding the Wakamatsu tilting conjecture from the viewpoint of resolving dualities.
\begin{cor}\label{cor:3.8} A Wakamatsu tilting bimodule $_{A}T_{B}$ is tilting if and only if
the functors $\mbox{\rm Hom}_{A}(-,T)$ and $\mbox{\rm Hom}_{B^{op}}(-,T)$ can be restricted to any one of inverse resolving dualities of the following types:
\begin{enumerate}
\item $\mathcal{W}{(_{A}T)}\cap{\mathcal{P}^{< \infty }(A)}\simeq \mathcal{W}{(T_{B})}\cap{\mathcal{P}^{<\infty}(B^{op})};$
\item $^{\perp}(_{A}T)\cap{\mathcal{P}^{< \infty }(A)}\simeq {^{\perp}(T_{B})\cap{\mathcal{P}^{<\infty}(B^{op})}};$
\item $\mathcal{W}{(_{A}T)}\cap{\mathcal{GP}^{\leqslant n}(A)}\simeq \mathcal{W}{(T_{B})}\cap{\mathcal{GP}^{\leqslant m}(B^{op})}$
for some $n, m\in\mathbb{N}$;
\item $^{\perp}(_{A}T)\cap{\mathcal{GP}^{\leqslant n }(A)}\simeq {^{\perp}(T_{B})\cap{\mathcal{GP}^{\leqslant m}(B^{op})}}$
for some $n, m\in\mathbb{N}$;
\item $\mathcal{W}{(_{A}T)}\cap{\mathcal{SGP}^{\leqslant n}(A)}\simeq \mathcal{W}{(T_{B})}\cap{\mathcal{SGP}^{\leqslant m}(B^{op})}$
for some $n, m\in\mathbb{N}$.
\end{enumerate}
\end{cor}
\begin{proof}
Recall that the functors $F:=\mbox{\rm Hom}_{A}(-,T)$ and $G:=\mbox{\rm Hom}_{B^{op}}(-,T)$ between $A\ensuremath{\mbox{{\rm -mod}}}$ and $B^{op}\ensuremath{\mbox{{\rm -mod}}}$ can be restricted to inverse resolving dualities $\mathcal{W}{(_{A}T)}\simeq\mathcal{W}{(T_{B})}$ by \cite[Theorem 1.5(d)]{GRS}. So, the necessity of Corollary \ref{cor:3.8} follows from Theorem \ref{THM}(2) and Lemma \ref{prop:3.3}(1)(3). As to the sufficiency of Corollary \ref{cor:3.8}, all the cases except $(4)$ are direct consequences of Theorem \ref{THM}(2) and Lemma \ref{prop:3.3}(1). It remains to show the case $(4)$. Suppose that $F$ and $G$ are restricted to the inverse resolving dualities $^{\perp}(_{A}T)\cap{\mathcal{GP}^{\leqslant n }(A)}\simeq {^{\perp}(T_{B})\cap{\mathcal{GP}^{\leqslant m}(B^{op})}}$ for some $n, m\in\mathbb{N}$. By Theorem \ref{THM}(1), $^{\perp}(_{A}T)\cap{\mathcal{GP}^{\leqslant n }(A)}\subseteq{\mathcal{W}{(_{A}T)}}$ and
${^{\perp}(T_{B})\cap{\mathcal{GP}^{\leqslant m}(B^{op})}}\subseteq{\mathcal{W}{(T_{B})}}$.
This implies that $^{\perp}(_{A}T)\cap{\mathcal{GP}^{\leqslant n }(A)}=\mathcal{W}(_{A}T)\cap{\mathcal{GP}^{\leqslant n}(A)}$ and $^{\perp}(T_{B})\cap{\mathcal{GP}^{\leqslant n }(B^{op})}=\mathcal{W}(T_{B})\cap{\mathcal{GP}^{\leqslant n}(B^{op})}$, and then we return to the case $(3)$.\end{proof}
\begin{rem}\label{generalization}
In Theorem \ref{THM}, if $\mathcal{C}=A\ensuremath{\mbox{{\rm -mod}}}$ and $\mathcal{D}=B^{op}\ensuremath{\mbox{{\rm -mod}}}$, then ${_A}T$ and $T_B$ are injective cogenerators since $A\ensuremath{\mbox{{\rm -mod}}}={\mathcal{W}{(_{A}T)}}$ and $B^{op}\ensuremath{\mbox{{\rm -mod}}}={\mathcal{W}{(T_{B})}}$.
By Corollary \ref{cor:3.8}(2), we obtain Miyachita's duality (see Theorem \ref{thm:2.3}).
Further, if $F:\mathcal{C}\to \mathcal{D}$ is a resolving duality with $\mathcal{C} \subseteq \mathcal{P}^{<\infty}(A)$ and $\mathcal{D} \subseteq \mathcal{P}^{<\infty}(B^{op})$, then the conditions $(a)$ and $(b)$ in Huisgen-Zimmermann's correspondence (see Theorem \ref{thm:2.4}) also follow from Theorem \ref{THM} and Lemma \ref{lem:3.6}.
\end{rem}
\section{Applications to establish homological invariances under tilting}\label{K-theory}
In the section, we apply the resolving dualities established in the former section to homological invariants of algebras.
On the one hand, we employ Corollary \ref{GV} to construct triangle equivalences of derived categories of Gorenstein-projective modules and to show that higher algebraic $K$-groups of Gorenstein-projective modules are invariant under tilting. On the other hand, we show that semi-derived Ringel-Hall algebras of Gorenstein-projective modules over algebras are preserved under tilting (see Theorem \ref{thm:4.10}), which generalizes some results in \cite[A5]{Lu2022}.
\subsection{Derived equivalences and algebraic K-groups of
Gorenstein-projective modules}
Throughout the section, for a small exact category $\mathcal{E}$, we denote by
$K(\mathcal{E})$ the $K$-theory space of $\mathcal{E}$ in the sense of Quillen (see \cite{Quillen}), and by $K_{n}(\mathcal{E})$ the $n$-th homotopy group of $K(\mathcal{E})$ (called the $n$-th algebraic $K$-group of $\mathcal{E}$).
Let us recall two classical results on algebraic $K$-theory of exact categories. One is usually called the ``resolution theorem" (see, for example, \cite[Section 4]{Quillen}); the other conveys that algebraic $K$-groups of exact categories are invariant under dualities.
\begin{lem}\label{lemmma:5.1} Let $\mathcal{E}'$ be a full subcategory of a small exact category $\mathcal{E}$. Assume that the
following two conditions hold:
\begin{enumerate}
\item[(a)] If $X\rightarrowtail Y \twoheadrightarrow Z$ is a conflation in $\mathcal{E}$ with $Z\in{\mathcal{E}}$, then $Y\in{\mathcal{E}}$ if and only if $X\in{\mathcal{E}}$.
\item[(b)] For any object $M \in{\mathcal{E}}$, there is an exact sequence in $\mathcal{E}$:
$$0\to M_{n}\to M_{n-1}\to\cdots \to M_{1} \to M_{0}\to M\to 0$$
such that $M_{i}\in{\mathcal{E}}$ for all $0 \leqslant i \leqslant n$.
\end{enumerate}
\noindent Then the inclusion $\mathcal{E}'\subseteq \mathcal{E}$ of exact categories induces a homotopy equivalence of $K$-theory space
$K(\mathcal{E}')\stackrel{\simeq}\longrightarrow K(\mathcal{E}).$ In particular, $K_n(\mathcal{E}')\simeq K_n(\mathcal{E})$ for all $n\in\mathbb{N}$.
\end{lem}
\begin{lem}\label{lemma:5.2} If $F:\mathcal{A}_{1}\to\mathcal{A}_{2}$ is a duality of small exact categories, then
$K_{n}(\mathcal{A}_{1})\simeq{K_{n}(\mathcal{A}_{2})}$ for all $ n\in{\mathbb{N}}$.
\end{lem}
\begin{proof} Since $F$ induces an equivalence $\mathcal{A}_{1}\stackrel{\simeq}\longrightarrow {\mathcal{A}_{2}}^{op}$ of small exact categories, we obtain $K_{n}(\mathcal{A}_{1})\simeq{K_{n}({\mathcal{A}_{2}}^{op})}$ for each $n\in\mathbb{N}$. Now,
Lemma \ref{lemma:5.2} follows from $K_{n}(\mathcal{A}_{2})\simeq{K_{n}({\mathcal{A}_{2}}^{op})}$ (see \cite[Section 2]{Quillen}).
\end{proof}
{\bf Proof of Corollary \ref{Cor-1}(1)(2)}.
\begin{proof}
$(1)$ Let $\ell:={\rm proj.dim}(_AT)$. Set $\mathcal{A}_{1}:={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant\ell}(A)}$ and $\mathcal{A}_{2}:={^{\perp}(T_{B})}\cap{\mathcal{GP}^{\leqslant\ell}(B^{op})}$. By Corollary \ref{GV}(1), $F:=\mbox{\rm Hom}_{A}(-,T):\mathcal{A}_{1}\to\mathcal{A}_{2}$ is a duality of small exact categories. Let $G:=\mbox{\rm Hom}_{B^{op}}(-,B):\mathcal{GP}(B^{op})\to \mathcal{GP}(B)$. Then $G$ is also a duality of small exact categories. Since $\mathcal{GP}(A)\subseteq \mathcal{A}_{1}$ and $\mathcal{GP}(B^{op})\subseteq \mathcal{A}_{2}$, there is a diagram of exact categories:
\begin{equation}\label{diagram:4.0}
\begin{aligned}
\xymatrix@C=3em@R=3em{
&\mathcal{GP}(A) \ar[r]^{\subseteq}& \mathcal{A}_{1}\ar[d]^{F}\\
\mathcal{GP}(B)& \mathcal{GP}(B^{op})\ar[l]_-{G}^-{\simeq}\ar[r]^{\subseteq}&\mathcal{A}_{2}.}
\end{aligned}
\end{equation}
Clearly, $\mathcal{A}_{1}\subseteq A\ensuremath{\mbox{{\rm -mod}}}$ and $\mathcal{B}_{1}\subseteq B^{op}\ensuremath{\mbox{{\rm -mod}}}$ are resolving subcategories.
Now, we apply Lemma \ref{lemma:2.3'} to the inclusions of exact categories
in \eqref{diagram:4.0}, and then obtain the following diagram of triangle equivalences of
unbounded derived categories of exact categories:
$$
\begin{aligned}
\xymatrix@C=3em@R=3em{
&\mathscr{D}(\mathcal{GP}(A) )\ar[r]^{\;\simeq}&\mathscr{D}(\mathcal{A}_{1})\ar[d]^{\simeq}\\
\mathscr{D}(\mathcal{GP}(B))&\mathscr{D}(\mathcal{GP}(B^{op}))^{op}\ar[l]_{\simeq}\ar[r]^{\quad\simeq}&\mathscr{D}(\mathcal{A}_{2})^{op}.}
\end{aligned}
$$
This implies a triangle equivalence $\mathscr{D}(\mathcal{GP}(A))\simeq{\mathscr{D}(\mathcal{GP}(B))}$, which can be restricted to an equivalence $\mathscr{D}^{*}(\mathcal{GP}(A))\simeq\mathscr{D}^{*}(\mathcal{GP}(B))$ for any $*\in\{+,-,b\}$
by Lemma \ref{lemma:2.3'}.
$(2)$ By Lemma \ref{lemmma:5.1}, $K_{n}(\mathcal{GP}(A))\simeq{K_{n}(\mathcal{A}_{1})}$ and $K_{n}(\mathcal{GP}(B^{op}))\simeq{K_{n}(\mathcal{A}_{2})}$. By Lemma \ref{lemma:5.2}, $K_{n}(\mathcal{A}_{1})\simeq{K_{n}(\mathcal{A}_{2})}$ and $K_{n}(\mathcal{GP}(B))\simeq{K_{n}(\mathcal{GP}(B^{op}))}$. Thus $K_{n}(\mathcal{GP}(A))\simeq{K_{n}(\mathcal{GP}(B))}$.
\end{proof}
By use of resolving dualities, we can show the following result, of which all the assertions except the equivalence $\mathscr{K}(\mathcal{P}(A))\simeq{\mathscr{K}(\mathcal{P}(B))}$ are known (for example, see \cite{Happel1988,Rickard,Daniel}).
\begin{cor}\label{Projective} Let $A$ and $B$ be algebras and $_{A}T_{B}$ a tilting bimodule. Then:
\begin{enumerate}
\item There is a triangle equivalence $\mathscr{K}(\mathcal{P}(A))\simeq{\mathscr{K}(\mathcal{P}(B))}$ which
can be restricted to an equivalence $\mathscr{K}^{*}(\mathcal{P}(A))\simeq \mathscr{K}^{*}(\mathcal{P}(B))$ for any
$*\in\{+,-,b\}$.
\item $K_{n}(\mathcal{P}(A))\simeq{K_{n}(\mathcal{P}(B))}$ for any $n\in{\mathbb{N}}.$
\end{enumerate}
\end{cor}
\begin{proof}
Note that $\mathscr{K}^{*}(\mathcal{P}(A))= {\mathscr{D}^{*}(\mathcal{P}(A))}$ for any $*\in\{\emptyset,+,-,b\}$.
Set $\mathcal{B}_{1}:={^{\perp}(_{A}T)}\cap{\mathcal{P}^{\leqslant\ell}(A)}$ and $\mathcal{B}_{2}:={^{\perp}(T_{B})}\cap{\mathcal{P}^{\leqslant\ell}(B^{op})}$ with $\ell:={\rm proj.dim}(_AT)$. Then $\mathcal{B}_{1}={^{\perp}(_{A}T)}\cap{\mathcal{P}^{<\infty}(A)}$ and $\mathcal{B}_{2}={^{\perp}(T_{B})}\cap{\mathcal{P}^{<\infty}(B^{op})}$ by Theorem \ref{thm:2.4}(c), and $F:\mathcal{B}_{1}\to\mathcal{B}_{2}$ is a duality of small exact categories by Theorem \ref{thm:2.4} (see also Corollary \ref{cor:3.8}(2)). Now, in the proof of Corollary \ref{Cor-1}(1)(2), we replace $\mathcal{A}_{1}$, $\mathcal{A}_{2}$, $\mathcal{GP}(A)$ and $\mathcal{GP}(B^{op})$ with $\mathcal{B}_{1}$ and $\mathcal{B}_{2}$, $\mathcal{P}(A)$ and $\mathcal{P}(B^{op})$, respectively, and then show Corollary \ref{Projective} similarly.
\end{proof}
\subsection{Semi-derived Ringel-Hall algebras of weakly $1$-Gorenstein exact categories}\label{Ringel-Hall-algebras}
In the section, we first recall the definition of semi-derived Ringel-Hall algebras of weakly $1$-Gorenstein exact categories from \cite{Lu2022}, and then
introduce a new definition for these algebras (up to isomorphism) which behaves better under dualities (see Proposition \ref{prop:4.6}).
Let $k:=\mathbb{F}_{q}$ be a finite field and $\mathcal{A}$ a small exact category linear over $k$. For each $X\in{\mathcal{A}}$, we define
\begin{center}
{${\rm Ext}\textrm{-}{\rm proj.dim}X:=\textrm{min}\{i\in{\mathbb{N}} \ | \ \textrm{Hom}_{\mathscr{D}(\mathcal{A})}(X,Y[j])=0\ \textrm{for} \ \textrm{all} \ Y\in{\mathcal{A}} \ \textrm{and} \ \textrm{all} \ j>i\}$,
\vspace{2mm}
${\rm Ext}\textrm{-}{\rm inj.dim}X:=\textrm{min}\{i\in{\mathbb{N}} \ | \ \textrm{Hom}_{\mathscr{D}(\mathcal{A})}(Y,X[j])=0\ \textrm{for} \ \textrm{all} \ Y\in{\mathcal{A}}\ \textrm{and} \ \textrm{all} \ j>i\}$}.
\end{center}
Following the Appendix of \cite{Lu2022}, there are four subcategories of $\mathcal{A}$:
\vspace{2mm}
\begin{center}{$\mathcal{P}^{\leqslant{i}}(\mathcal{A})=\{X\in{\mathcal{A}\ |\ {\rm Ext}\textrm{-}{\rm proj.dim}X\leqslant i}\}$,
\vspace{2mm}
$\mathcal{I}^{\leqslant{i}}(\mathcal{A})=\{X\in{\mathcal{A}\ |\ {\rm Ext}\textrm{-}{\rm inj.dim}X\leqslant i}\}$,
\vspace{2mm}
$\mathcal{P}^{<\infty}(\mathcal{A})=\{X\in{\mathcal{A}\ |\ {\rm Ext}\textrm{-}{\rm proj.dim}X<\infty}\}$,
\vspace{2mm}
$\mathcal{I}^{<\infty}(\mathcal{A})=\{X\in{\mathcal{A}\ |\ {\rm Ext}\textrm{-}{\rm inj.dim}X<\infty}\}$.}
\end{center}
The category $\mathcal{A}$ is said to be \emph{weakly Gorenstein} if $\mathcal{P}^{<\infty}(\mathcal{A})=\mathcal{I}^{<\infty}(\mathcal{A})$; \emph{weakly $d$-Gorenstein} if it is weakly Gorenstein and $\mathcal{P}^{<\infty}(\mathcal{A})=\mathcal{P}^{\leqslant d}(\mathcal{A})=\mathcal{I}^{\leqslant d}(\mathcal{A})$.
We consider an exact category $\mathcal{A}$ satisfying the following conditions:
\begin{enumerate}
\item[(E-a)] $\mathcal{A}$ is a small exact category with finite morphism spaces and finite extension spaces, i.e.,
$$|\mbox{\rm Hom}_{\mathcal{A}}(M,N)|<\infty, \quad |\mbox{\rm Ext}_{\mathcal{A}}^{1}(M,N)|<\infty;$$
\item[(E-b)] $\mathcal{A}$ is linear over $k=\mathbb{F}_{q}$;
\item[(E-c)] $\mathcal{A}$ is weakly 1-Gorenstein;
\item[(E-d)] for any object $X\in{\mathcal{A}}$, there exists a deflation $P_{X}\to X$ with $P_{X}\in{\mathcal{P}^{<\infty}(\mathcal{A})}$.
\end{enumerate}
In this case, $\mathcal{P}^{<\infty}(\mathcal{A})=\mathcal{P}^{\leqslant 1}(\mathcal{A})=\mathcal{I}^{<\infty}(\mathcal{A})=\mathcal{I}^{\leqslant 1}(\mathcal{A})$.
Clearly, if $A$ is a finite-dimensional algebra over $k$, then the Frobenius category $\mathcal{GP}(A)$ satisfies {\rm(E-a)-(E-d)}.
If, in addition, $A$ is $1$-Gorenstein (that is, both $_AA$ and $A_A$ have injective dimension at most $1$), then the abelian category $A\ensuremath{\mbox{{\rm -mod}}}$ also satisfies {\rm(E-a)-(E-d)}. The following result supplies a class of weakly 1-Gorenstein exact categories which may be neither Frobenius nor abelian categories in general.
\begin{lem}\label{1-weak-G}
Let $A$ be a finite-dimensional algebra over the field $k$ and $_{A}T_{B}$ a tilting bimodule with ${\rm proj.dim}(_AT)\leqslant 1$. Define
$\mathcal{A}:={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant 1}(A)}$. Then $\mathcal{A}$ is weakly 1-Gorenstein satisfying {\rm(E-a)-(E-d)}
and $\mathcal{P}^{<\infty}(\mathcal{A})={^{\perp}(_{A}T)}\cap{\mathcal{P}^{\leqslant 1}(A)}$.
\end{lem}
\begin{proof}
By Corollary \ref{cor:3.5}(3), $\mathcal{A}=\mathcal{W}(_{A}T)$. Since $\mathcal{A}$ is a resolving subcategory of $A\ensuremath{\mbox{{\rm -mod}}}$ by Lemma \ref{proposition:3.2}(1), it is a small exact category of which the projective objects are exactly projective $A$-modules. By Lemma \ref{proposition:3.2}(3), {\rm add}$(_{A}T)$ equals the full subcategory of $\mathcal{A}$ consisting of injective objects. So, $\mathcal{A}$ as an exact category has enough projective objects and injective objects. Clearly,
$\mathcal{P}^{<\infty}(\mathcal{A})=\mathcal{A}\cap\mathcal{P}^{<\infty}(A)={^{\bot}(_{A}T)}\cap\mathcal{P}^{\leqslant 1}(A)$. Further, since ${\rm proj.dim}(_AT)\leqslant 1$, it follows from Theorem \ref{thm:2.4}(c) that ${^{\bot}(_{A}T)}\cap\mathcal{P}^{\leqslant 1}(A)={^{\bot}(_{A}T)}\cap\mathcal{P}^{<\infty}(A)$ which consists of those $A$-modules having finite $\mbox{\rm add}({_A}T)$-coresolutions. Thus $\mathcal{P}^{<\infty}(\mathcal{A})=\mathcal{I}^{<\infty}(\mathcal{A})$, that is, $\mathcal{A}$ is weakly Gorenstein.
Since $\mathcal{P}^{<\infty}(\mathcal{A})\subseteq \mathcal{P}^{\leqslant 1}(A)$, we have
$\mathcal{P}^{<\infty}(\mathcal{A})=\mathcal{P}^{\leqslant 1}(\mathcal{A})$. Moreover, by the resolving duality $\mbox{\rm Hom}_A(-,T): {^{\bot}(_{A}T)}\cap\mathcal{P}^{\leqslant 1}(A)\to {^{\bot}(T_B)}\cap\mathcal{P}^{\leqslant 1}(B^{op})$ in Theorem \ref{thm:2.3}, each object of ${^{\bot}(_{A}T)}\cap\mathcal{P}^{\leqslant 1}(A)$ has an $\mbox{\rm add}({_A}T)$-coresolution of length at most $1$. This implies $\mathcal{I}^{<\infty}(\mathcal{A})=\mathcal{I}^{\leqslant 1}(\mathcal{A})$, and therefore $\mathcal{A}$ is weakly 1-Gorenstein.
Since $\mathcal{A}\subseteq A\ensuremath{\mbox{{\rm -mod}}}$ is a resolving subcategory and $k$ is a finite field, it can be checked that $\mathcal{A}$ satisfies {\rm(E-a)-(E-d)} .
\end{proof}
Now, let $\mathcal{A}$ be an exact category which satisfies {\rm(E-a)-(E-d)}. Denote by ${\rm Iso}(\mathcal{A})$ the set of isomorphism classes of objects in $\mathcal{A}$ and by $K_{0}(\mathcal{A})$ the Grothendieck group of $\mathcal{A}$. Let $\mathcal{H}(\mathcal{A})$ be the \emph{Ringel-Hall algebra} of $\mathcal{A}$, that is, $\mathcal{H}(\mathcal{A})=\bigoplus_{[M]\in{\rm Iso}(\mathcal{A})}\mathbb{Q}[M]$ as $\mathbb{Q}$-vector spaces with the multiplication given by
$$[M]\diamond[N]:=\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}\displaystyle\frac{|\mbox{\rm Ext}_{\mathcal{A}}^{1}(M,N)_{L}|}{|\mbox{\rm Hom}_{\mathcal{A}}(M,N)|}[L]$$
where $\mbox{\rm Ext}_{\mathcal{A}}^{1}(M,N)_{L}$ stands for the subset of ${\mbox{\rm Ext}_{\mathcal{A}}^{1}(M,N)}$ parameterizing all extensions in which the middle term is isomorphic to $L$.
Then $\mathcal{H}(\mathcal{A})$ is a $K_{0}(\mathcal{A})$-graded algebra.
For $M\in{\mathcal{A}}$ and $K\in{\mathcal{P}^{\leqslant 1}(\mathcal{A})}$, define
$$\langle K,M\rangle=\displaystyle\textrm{dim}_{k}\mbox{\rm Hom}_{\mathcal{A}}(K,M)-\displaystyle\textrm{dim}_{k}\mbox{\rm Ext}_{\mathcal{A}}^1(K,M),$$
$$\langle M,K\rangle=\displaystyle\textrm{dim}_{k}\mbox{\rm Hom}_{\mathcal{A}}(M,K)-\displaystyle\textrm{dim}_{k}\mbox{\rm Ext}_{\mathcal{A}}^1(M,K).$$
These formulas descend bilinear forms (called \emph{Euler forms}), again denoted
by $\langle \cdot,\cdot\rangle$, on the Grothendieck groups $K_{0}({\mathcal{P}^{\leqslant 1}(\mathcal{A})})$ and $K_{0}(\mathcal{A})$.
To introduce semi-derived Ringel-Hall algebras of weakly 1-Gorenstein exact categories, we
first recall the definition of (left or right) denominator subsets of rings and their relations with Ore localizations.
Let $R$ be a ring with identity and $S$ a subset of $R$ closed under multiplications with $1\in{S}$. Following \cite[Chapter 4]{Lam}, $S$ is called a \emph{left denominator subset} of $R$ if the following conditions hold:
\begin{enumerate}
\item[(i)] For any $a\in{R}$ and $s\in{S}$, the intersection $Sa\cap{Rs}$ is not empty;
\item[(ii)] For any $r\in{R}$, if $rt=0$ for some $t\in{S}$, there exists some $t'\in{S}$ such that $t'r=0$.
\end{enumerate}
If $S$ satisfies only the condition (i), then $S$ is called a \emph{left Ore subset} of $R$. Similarly, we can define right denominator sets and right Ore sets. Now, Ore's localization theorem states that
\begin{enumerate}
\item the left Ore localization $[S^{-1}]R$ exists if and only if $S$ is a left denominator subset of $R$;
\item the right Ore localization $R[S^{-1}]$ exists if and only if $S$ is a right denominator subset of $R$.
\end{enumerate}
If $S$ is a left and right denominator subset of $R$,
then $[S^{-1}]R$ is called the \emph{Ore localization} of $R$ at $S$.
In this case, up to isomorphism of rings, $[S^{-1}]R$, $R[S^{-1}]$ and the universal localization $R_{S}$ of $R$ at $S$ are
the same (for example, see \cite[Section 2.2]{chen-xi}).
Let $I(\mathcal{A})$ be the two-sided ideal of $\mathcal{H}(\mathcal{A})$ generated by
$$\{[L]-[K\oplus M] \ | \ \exists \ \textrm{an exact sequence}\ 0\to K\to L\to M\to 0 \ \textrm{with}\ K\in{{\mathcal{P}^{\leqslant 1}(\mathcal{A})}}\}.$$
We consider the following multiplicatively closed subset of the quotient $\mathcal{H}(\mathcal{A})/I(\mathcal{A})$ of $\mathcal{H}(\mathcal{A})$ by $I(\mathcal{A})$:
$$\mathcal{S}_{\mathcal{A}}:=\{a[K]\in{\mathcal{H}(\mathcal{A})/I(\mathcal{A})} \ |\ a\in{\mathbb{Q}^{\times}},\ K\in{{\mathcal{P}^{\leqslant 1}(\mathcal{A})}}\}.$$
\begin{lem}\label{lem:4.2}{\rm(}\cite[Proposition A5]{Lu2022}{\rm)}
$\mathcal{S}_{\mathcal{A}}$ is a right denominator subset of $\mathcal{H}(\mathcal{A})/I(\mathcal{A})$. Equivalently, the right Ore localization $(\mathcal{H}(\mathcal{A})/I(\mathcal{A}))[\mathcal{S}^{-1}_{\mathcal{A}}]$ of $\mathcal{H}(\mathcal{A})/I(\mathcal{A})$ with respect to $\mathcal{S}_{\mathcal{A}}$ exists.
\end{lem}
Following \cite{Lu2022} (also cf. \cite{Bridgeland, Gorsky, LP21}), the algebra $(\mathcal{H}(\mathcal{A})/I(\mathcal{A}))[\mathcal{S}^{-1}_{\mathcal{A}}]$ is called the \emph{semi-derived Ringel-Hall algebra} of $\mathcal{A}$ and denoted by $\mathcal{SDH}(\mathcal{A})$.
Since the opposite category $\mathcal{A}^{op}$ of $\mathcal{A}$ is also a weakly $1$-Gorenstein exact category, the algebra
$\mathcal{SDH}(\mathcal{A}^{op})$ is well defined. However, at the present time, it is not clear whether $\mathcal{SDH}(\mathcal{A})\cong (\mathcal{SDH}(\mathcal{A}^{op}))^{op}$ as algebras because the definition of $\mathcal{SDH}(\mathcal{A})$ seems not to be left-right symmetric.
To solve this problem, we will introduce a new definition of $\mathcal{SDH}(\mathcal{A})$ up to isomorphism of algebras.
Let $J(\mathcal{A})$ be the two-sided ideal of $\mathcal{H}(\mathcal{A})$ generated by
$$\{[L]-[K\oplus M] \ | \ \exists \ \textrm{exact} \ \textrm{sequence}\ 0\to M\to L\to K\to 0 \ \textrm{with}\ K\in{{\mathcal{I}^{\leqslant 1}(\mathcal{A})}}\}.$$
Then $I({\mathcal{A}}^{op})\cong{J(\mathcal{A})}$ and there is an isomorphism of algebras:
$\mathcal{H}(\mathcal{A})/J(\mathcal{A})\cong{(\mathcal{H}(\mathcal{A}^{op})/I(\mathcal{A}^{op}))^{op}}.$
Similarly, we consider the multiplicatively closed subset $\mathcal{R}_{\mathcal{A}}$ of $\mathcal{H}(\mathcal{A})/J(\mathcal{A})$:
$$\mathcal{R}_{\mathcal{A}}:=\{a[K]\in{\mathcal{H}(\mathcal{A})/J(\mathcal{A})} \ |\ a\in{\mathbb{Q}^{\times}},\ K\in{{\mathcal{I}^{\leqslant 1}(\mathcal{A})}}\}.$$
The following result is the dual of Lemma \ref{lem:4.2}.
\begin{lem}\label{lem:4.4}
$\mathcal{R}_{\mathcal{A}}$ is a left denominator subset of $\mathcal{H}(\mathcal{A})/J(\mathcal{A})$. Equivalently, the left Ore localization $[\mathcal{R}^{-1}_{\mathcal{A}}](\mathcal{H}(\mathcal{A})/J(\mathcal{A}))$ of
$\mathcal{H}(\mathcal{A})/J(\mathcal{A})$ with respect to $\mathcal{R}_{\mathcal{A}}$ exists. Moreover, there is an isomorphism $[\mathcal{R}^{-1}_{\mathcal{A}}](\mathcal{H}(\mathcal{A})/J(\mathcal{A}))\cong{(\mathcal{SDH}(\mathcal{A}^{op}))^{op}}$ of algebras.
\end{lem}
Now, we consider the quotient $\mathcal{H}(\mathcal{A})/(I(\mathcal{A})+J(\mathcal{A}))$ of $\mathcal{H}(\mathcal{A})$ by the ideal $I(\mathcal{A})+J(\mathcal{A})$ and
its multiplicatively closed subset
$$\Phi_{\mathcal{A}}:=\{a[K]\in{\mathcal{H}(\mathcal{A})/(I(\mathcal{A})+J(\mathcal{A}))} \ |\ a\in{\mathbb{Q}^{\times}},\ K\in{{\mathcal{P}^{\leqslant 1}(\mathcal{A})}}\}.$$
\begin{lem}\label{lem:4.5}
$(1)$ $\Phi_{\mathcal{A}}$ is a left and right denominator subset of the algebra $\mathcal{H}(\mathcal{A})/(I(\mathcal{A})+J(\mathcal{A}))$.
$(2)$ There are isomorphisms of algebras:
\begin{equation*}
\begin{aligned}
\mathcal{SDH}(\mathcal{A})&\cong{(\mathcal{H}(\mathcal{A})/(I(\mathcal{A})+J(\mathcal{A})))[\Phi^{-1}_{\mathcal{A}}]}\\
&\cong{[\Phi^{-1}_{\mathcal{A}}](\mathcal{H}(\mathcal{A})/(I(\mathcal{A})+J(\mathcal{A})))}\\
&\cong{[\mathcal{R}^{-1}_{\mathcal{A}}](\mathcal{H}(\mathcal{A})/J(\mathcal{A}))}\\
&\cong{(\mathcal{SDH}(\mathcal{A}^{op}))^{op}}.
\end{aligned}
\end{equation*}
\end{lem}
\begin{proof} (1) Let $H:=\mathcal{H}(A)$, $I:=I(\mathcal{A})$ and $J:=J(\mathcal{A})$.
For any $K\in{{\mathcal{P}^{\leqslant 1}(\mathcal{A})}}$ and $M\in{\mathcal{A}}$, it follows from \cite[Lemma A4]{Lu2022} that $[M]\diamond[K]=q^{-\langle M,K\rangle}[M\oplus K]$ in $H/I$. Dually, $[K]\diamond[M]=q^{-\langle K,M\rangle}[M\oplus K]$ in $H/J$. Thus $q^{\langle M,K\rangle}[M]\diamond[K]=q^{\langle K,M\rangle}[K]\diamond[M]$
in $H/(I+J)$. This implies that $\Phi_{\mathcal{A}}$ is a left and right Ore subset of the algebra $H/(I+J)$. By a similar argument as in the proof of \cite[Proposition A5]{Lu2022}, one can further show that $\Phi_{\mathcal{A}}$ is a left and right denominator subset of $H/(I+J)$.
(2) By $(1)$, there is an isomorphism of algebras
${(H/(I+J))[\Phi^{-1}_{\mathcal{A}}]}\cong{[\Phi^{-1}_{\mathcal{A}}](H/(I+J))}.$
By Lemma \ref{lem:4.4}, it is enough to show the algebra isomorphism $$\mathcal{SDH}(\mathcal{A})\cong{(H/(I+J))[\Phi^{-1}_{\mathcal{A}}]}.$$ The algebra isomorphism ${[\Phi^{-1}_{\mathcal{A}}](H/(I+J))}\cong{[\mathcal{R}^{-1}_{\mathcal{A}}](H/J)}$ can be proved dually.
Let $\lambda_{1}:H/I\to\mathcal{SDH}(\mathcal{A})$ and $\lambda:H/(I+J)\to(H/(I+J))[\Phi^{-1}_{\mathcal{A}}]$
be the localizations and let $\pi_{1}:H/I\to H/(I+J)$ be the canonical surjection.
Since $\pi_{1}(\mathcal{S}_{\mathcal{A}})=\Phi_{\mathcal{A}}$, there is a unique homomorphism of algebras
$$\sigma:\mathcal{SDH}(\mathcal{A}) \to{(H/(I+J))[\Phi^{-1}_{\mathcal{A}}]}$$ such that $\sigma\lambda_{1}=\lambda\pi_{1}$. By the statements following
\cite[Lemma A8]{Lu2022}, we have $\lambda_{1}(J)=0$ in $\mathcal{SDH}(\mathcal{A})$. Then there exists a unique homomorphism of algebras $$\widetilde{\lambda_{1}}:{(H/(I+J))[\Phi^{-1}_{\mathcal{A}}]}\to \mathcal{SDH}(\mathcal{A})$$ such that $\lambda_{1}=\widetilde{\lambda_{1}}\pi_{1}$, and hence $\widetilde{\lambda_{1}}([K])$ is invertible in $\mathcal{SDH}(\mathcal{A})$ for any $[K]\in{\Phi_{\mathcal{A}}}$. This implies that $\lambda$ induces a unique homomorphism of algebras $$\rho:{(H/(I+J))[\Phi^{-1}_{\mathcal{A}}]}\to \mathcal{SDH}(\mathcal{A})$$ such that $\rho\lambda=\widetilde{\lambda_{1}}$. So we have following commutative diagram:
$$
\xymatrix@C=3em@R=3em{
H/I \ar[d]_{\pi_{1}} \ar[r]^{\lambda_{1}}
& \mathcal{SDH}(\mathcal{A})\ar@<.5ex>@{.>}[d]^{\sigma} \\
H/(I+J)\ar@{.>}[ur]^{\widetilde{\lambda_{1}}}\ar[r]_{\lambda}
& (H/(I+J))[\Phi^{-1}_{\mathcal{A}}]\ar@<.5ex>@{.>}[u]^{\rho}.}
$$
Since $\rho\sigma\lambda_{1}=\rho\lambda\pi_{1}=\widetilde{\lambda_{1}}\pi_{1}=\lambda_{1}$, it follows from the the universal property of $\lambda_{1}$ that $\rho\sigma=\textrm{Id}$. On the other hand, $\sigma\rho\lambda\pi_{1}=\sigma\widetilde{\lambda_{1}}\pi_{1}=\sigma\lambda_{1}=\lambda\pi_{1}$. Since $\pi_{1}$ is surjective, $\sigma\rho\lambda=\lambda$. By the universal property of $\lambda$, we have $\sigma\rho=\textrm{Id}$. Thus $\sigma$ and $\rho$ are isomorphisms of algebras. \end{proof}
Thanks to Lemma \ref{lem:4.5}, up to isomorphism of algebras, we can define the \emph{semi-derived Ringel-Hall algebra}
of $\mathcal{A}$ to be the algebra $(\mathcal{H}(\mathcal{A})/(I(\mathcal{A})+J(\mathcal{A})))[\Phi^{-1}_{\mathcal{A}}]$. This definition is left-right symmetric and applied to show the following result.
\begin{prop}\label{prop:4.6} Let $F:\mathcal{A}_{1}\to\mathcal{A}_{2}$ be a duality of weakly $1$-Gorenstein exact categories. Then there exists an isomorphism of algebras
\vspace{2mm}
\begin{center}
{$\Upsilon_{F}:\mathcal{SDH}(\mathcal{A}_{1})\stackrel{\simeq}\longrightarrow (\mathcal{SDH}(\mathcal{A}_{2}))^{op}$ \\
\vspace{2mm}
$[M]\mapsto [F(M)]$.}
\end{center}
\end{prop}
\begin{proof}
Note that $F$ induces an equivalence $\mathcal{A}_{1}\stackrel{\simeq}\longrightarrow {\mathcal{A}_{2}}^{op}$ of weakly $1$-Gorenstein exact categories. Then $F$ induces an isomorphism of algebras: $\mathcal{SDH}(\mathcal{A}_{1})\stackrel{\simeq}\longrightarrow{\mathcal{SDH}({\mathcal{A}_{2}}^{op})}.$ By Lemma \ref{lem:4.5}(2), ${\mathcal{SDH}({\mathcal{A}_{2}}^{op})}\simeq (\mathcal{SDH}(\mathcal{A}_2))^{op}$. Thus Proposition \ref{prop:4.6} holds.
\end{proof}
\subsection{Invariance of semi-derived Ringel-Hall algebras under tilting} \label{Ringel-Hall-algebras-Gorenstein}
In the section, our main result is the following theorem which contains Corollary \ref{Cor-1}(3) in the introduction.
\begin{thm}\label{thm:4.10} Let $A$ be a finite-dimensional algebra over $k$ and $_{A}T_{B}$ a tilting bimodule with ${\rm proj.dim}(_AT)\leqslant1$. Then there exists an isomorphism of algebras:
\vspace{2mm}
\begin{center}
{$\Xi:\;\mathcal{SDH}(\mathcal{GP}(A))\stackrel{\simeq}\longrightarrow \mathcal{SDH}(\mathcal{GP}(B))$ \\
\vspace{2mm}
$ [G]\mapsto q^{-\langle L,G\rangle}[\mbox{\rm Hom}_{A}(T,L)]^{-1}\diamond[\mbox{\rm Hom}_{A}(T,Z)]$,}
\end{center}
where $f:G\to Z$ is a minimal left $(_{A}T)^{\perp}$-approximation of $G$ and
$L={\rm Coker}(f)$.
\end{thm}
When both $A$ and $B$ are $1$-Gorenstein algebras, Theorem \ref{thm:4.10} is exactly \cite[Corollary A23]{Lu2022}. To show Theorem \ref{thm:4.10} for general algebras, we establish a crucial result as follows.
\begin{prop}\label{prop:4.7} Let $A$ be a finite-dimensional algebra over $k$ and $_{A}T_{B}$ a tilting bimodule with ${\rm proj.dim}(_AT)\leqslant1$. Set $\mathcal{A}:={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant 1}(A)}$ and $\mathcal{B}:=\mathcal{GP}(A)$. Then the embedding $\phi:\mathcal{H}(\mathcal{B})\to \mathcal{H}(\mathcal{A})$ induces an algebra isomorphism
$\widetilde{\phi}:\mathcal{SDH}(\mathcal{B})\to\mathcal{SDH}(\mathcal{A}).$
Furthermore, the inverse of $\widetilde{\phi}$ is given by $\widetilde{\psi}:[M]\mapsto{q^{-\langle M,H_{M}\rangle}[G_{M}]\diamond[H_{M}]^{-1}}$, where $M\in{\mathcal{A}}$, $G_{M}\in{\mathcal{B}}$ and $H_{M}\in{\mbox{\rm add}(_{A}A)}$ such that they fit into a short exact sequence $0\to H_{M}\to G_{M}\to M \to 0$ of $A$-modules.
\end{prop}
\begin{proof} Clearly, $\mathcal{B}$ is weakly $1$-Gorenstein satisfying {\rm(E-a)-(E-d)}. By Lemma \ref{1-weak-G} and its proof, $\mathcal{A}$ is also weakly $1$-Gorenstein satisfying {\rm(E-a)-(E-d)} and ${\mathcal{P}^{{\leqslant 1}}(\mathcal{A})}={^{\bot}(_{A}T)}\cap\mathcal{P}^{\leqslant 1}(A)$. This means that $\mathcal{SDH}(\mathcal{B})$ and $\mathcal{SDH}(\mathcal{A})$ are well defined.
Moreover, since each object $M\in\mathcal{A}$ has Gorenstein dimension at most $1$, the exact sequence in Proposition \ref{prop:4.7} always exists. By a similar proof of \cite[Theorem A15]{Lu2022}, one can check that $\widetilde{\phi}$ is a surjective homomorphism of algebras and the map $$\psi:\mathcal{H}(\mathcal{A})\to\mathcal{SDH}(\mathcal{B}),\;\;[M]\mapsto{q^{-\langle M,H_{M}\rangle}[G_{M}]\diamond[H_{M}]^{-1}}$$ is well defined.
Next we claim that $\psi$ is a homomorphism of algebras. It suffices to show that
\begin{equation}\label{4.0}\psi([M]\diamond[N])=\psi([M])\diamond\psi([N])
\end{equation}
for all $M,N\in{\mathcal{A}}$.
For this aim, we fix two exact sequences in $A$-mod:
\begin{equation}\label{4.1}0\to H_{1}\to G_{1}\to M\to0,
\end{equation}
\begin{equation}\label{4.2} 0\to H_{2}\to G_{2}\to N\to0,
\end{equation}
where $H_{1},H_{2}\in{\mbox{\rm add}(_{A}A)}$ and $G_{1},G_{2}\in{\mathcal{B}}$. Applying $\mbox{\rm Hom}_{A}(-,N)$ and $
\mbox{\rm Hom}_{A}(G_1,-)$ to the sequences \eqref{4.1} and \eqref{4.2}, respectively, we obtain the following diagram
\begin{equation}\label{4.3}
\begin{aligned}
\xymatrix{
\mbox{\rm Ext}_{A}^{1}(M,N) \ar[dr] \ar[r]^{\Theta}
& \mbox{\rm Ext}_{A}^{1}(G_{1},G_{2}) \ar[d]^{\cong}\ar[r] &0 \\
& \mbox{\rm Ext}_{A}^{1}(G_{1},N). }
\end{aligned}
\end{equation}
Now, let $0\to N\to L\to M\to 0$ be an exact sequence in $\mathcal{A}$. By \eqref{4.3}, there is commutative diagram with exact rows and columns:
\begin{equation}\label{4.4}
\begin{aligned}
\xymatrix{&0\ar[d]&0\ar[d]&0\ar[d]\\
0\ar[r]&H_2\ar[r]\ar[d]&H_{L}\ar[d]\ar[r]&H_1\ar[d]\ar[r]&0\\
0\ar[r]&G_2\ar[r]\ar[d]&G_{L}\ar[d]\ar[r]&G_1\ar[d]\ar[r]&0\\
0\ar[r]&N\ar[r]\ar[d]&L\ar[d]\ar[r]&M\ar[d]\ar[r]&0\\
&0&0&\;0.}
\end{aligned}
\end{equation}
Note that $\langle L,H_{L}\rangle={\langle M+N,H_{1}+H_{2}\rangle}={\langle M,H_{1}\rangle}+{\langle M,H_{2}\rangle}+{\langle N,H_{1}\rangle}+{\langle N,H_{2}\rangle}$. Since $\mathcal{A}$ and $\mathcal{B}$ are fully exact subcategories of $A\ensuremath{\mbox{{\rm -mod}}}$, there are equalities
\begin{center}
{$|\mbox{\rm Ext}_{\mathcal{A}}^{1}(X_1,X_2)|=|\mbox{\rm Ext}_{A}^{1}(X_1,X_2)|$, \quad $|\mbox{\rm Hom}_{\mathcal{A}}(X_1,X_2)|=|\mbox{\rm Hom}_{A}(X_1,X_2)|$,
\vspace{2mm}
$|\mbox{\rm Ext}_{\mathcal{B}}^{1}(Y_1,Y_2)|=|\mbox{\rm Ext}_{A}^{1}(Y_1,Y_2)|$, \quad $|\mbox{\rm Hom}_{\mathcal{B}}(Y_1,Y_2)|=|\mbox{\rm Hom}_{A}(Y_1,Y_2)|$ }
\end{center}
for all $X_1,X_2\in{\mathcal{A}}$ and $Y_1,Y_2\in{\mathcal{B}}$.
Thus
\begin{equation*}
\begin{split}
\psi([M]\diamond[N])&=\psi(\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}\displaystyle\frac{|\mbox{\rm Ext}_{A}^{1}(M,N)_{L}|}{|\mbox{\rm Hom}_{A}(M,N)|}[L])\\
&=\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}\displaystyle\frac{|\mbox{\rm Ext}_{A}^{1}(M,N)_{L}|}{|\mbox{\rm Hom}_{A}(M,N)|}\displaystyle q^{-\langle L,H_{L}\rangle}[G_{L}]\diamond[H_{L}]^{-1}\\
&=(\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}\displaystyle\frac{|\mbox{\rm Ext}_{A}^{1}(M,N)_{L}|}{|\mbox{\rm Hom}_{A}(M,N)|}q^{-\langle L,H_{L}\rangle}[G_{L}])\diamond[H_{1}\oplus H_{2}]^{-1}\\
&=(\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}\displaystyle\frac{|\mbox{\rm Ext}_{A}^{1}(M,N)_{L}|}{|\mbox{\rm Hom}_{A}(M,N)|}q^{-{\langle M,H_{1}\rangle}-{\langle M,H_{2}\rangle}-{\langle N,H_{1}\rangle}-{\langle N,H_{2}\rangle}}[G_{L}])\diamond[H_{1}\oplus H_{2}]^{-1}.
\end{split}
\end{equation*}
Since $q^{\langle H_1,G_{2}\rangle}[H_1]\diamond[G_2]=[H_1\oplus G_2]=q^{\langle G_2,H_{1}\rangle}[G_2]\diamond[H_1]$, it follows that
\begin{equation*}
\begin{split}
\psi([M]) \diamond \psi([N])&=q^{-{\langle M,H_{1}\rangle}-{\langle N,H_{2}\rangle}}[G_1]\diamond{[H_1]^{-1}}\diamond[G_2]\diamond{[H_2]^{-1}}\\
&=q^{-{\langle M,H_{1}\rangle}-{\langle N,H_{2}\rangle}+{\langle H_{1},G_{2}\rangle}-{\langle G_{2},H_{1}\rangle}}[G_1]\diamond[G_2]\diamond{[H_1]^{-1}}\diamond{[H_2]^{-1}}\\
&=q^{-{\langle M,H_{1}\rangle}-{\langle N,H_{2}\rangle}+{\langle H_{1},G_{2}\rangle}-{\langle G_{2},H_{1}\rangle}}[G_1]\diamond[G_2]\diamond([H_2]\diamond[H_1])^{-1}\\
&=q^{-{\langle M,H_{1}\rangle}-{\langle N,H_{2}\rangle}+{\langle H_{1},G_{2}\rangle}-{\langle G_{2},H_{1}\rangle}+{\langle H_{2},H_{1}\rangle}}[G_1]\diamond[G_2]\diamond[H_2\oplus H_1]^{-1}.
\end{split}
\end{equation*}
Consequently, to prove \eqref{4.0}, we only need to check
\begin{align}\label{4.6}
q^{-{\langle M,H_{2}\rangle}-{\langle N,H_{1}\rangle}}\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}\displaystyle\frac{|\mbox{\rm Ext}_{A}^{1}(M,N)_{L}|}{|\mbox{\rm Hom}_{A}(M,N)|}[G_{L}]&
=q^{{\langle H_{1},G_{2}\rangle}-{\langle G_{2},H_{1}\rangle}+{\langle H_{2},H_{1}\rangle}}[G_1]\diamond[G_2].
\end{align}
Now, we set $K:=\ker(\Theta)$ in the diagram \eqref{4.3}. Then there exists an exact sequence
$$0\to \mbox{\rm Hom}_{A}(M,N)\to \mbox{\rm Hom}_{A}(G_1,N) \to \mbox{\rm Hom}_{A}(H_1,N)\to K\to0.$$
Thus
\begin{equation}\label{4.7}
|K|=q^{{\rm dim}_{k}\scriptstyle{\rm Hom}_{A}(H_1,N)-{\rm dim}_{k}\scriptstyle{\rm Hom}_{A}(G_1,N)+{\rm dim}_{k}\scriptstyle{\rm Hom}_{A}(M,N)}.
\end{equation}
For any $\delta':0\to G_2\to W\to G_1\to 0$ in the set $\mbox{\rm Ext}_{A}^{1}(G_1,G_2)_{W}$, one can show
$$|\{\delta\in{\mbox{\rm Ext}_{A}^{1}(M,N)}\ | \ \Theta(\delta)=\delta'\}|=|K|.$$ This leads to
$$|\{\delta\in{\mbox{\rm Ext}_{A}^{1}(M,N)}\ | \ \Theta(\delta)\in{\mbox{\rm Ext}_{A}^{1}(G_1,G_2)_{W}}\}|=|K||\mbox{\rm Ext}_{A}^{1}(G_1,G_2)_{W}|$$
and therefore \begin{equation}\label{4.8}
\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}|\mbox{\rm Ext}_{A}^{1}(M,N)_{L}|[G_{L}]=
|K|\sum_{[W]\in{{\rm Iso}(\mathcal{A})}}|{\mbox{\rm Ext}_{A}^{1}(G_1,G_2)_{W}}|[W].
\end{equation}
Since $G_1\in\mathcal{B}$ and $H_{2}\in{\mbox{\rm add}(_{A}A)}$, the equality $\mbox{\rm Ext}_A^1(G_1, H_2)=0$ holds.
This implies that the sequence $0\to\mbox{\rm Hom}_{A}(G_1,H_{2})\to\mbox{\rm Hom}_{A}(G_1,G_{2}) \to \mbox{\rm Hom}_{A}(G_1,N)\to0$ is exact.
Consequently,
\begin{equation}\label{4.9}
\displaystyle{\rm dim}_{k}\mbox{\rm Hom}_{A}(H_1,N)-\displaystyle{\rm dim}_{k}\mbox{\rm Hom}_{A}(G_1,N)=
{\langle H_{1},N\rangle}+\displaystyle{\rm dim}_{k}\mbox{\rm Hom}_{A}(G_1,H_{2})-\displaystyle{\rm dim}_{k}\mbox{\rm Hom}_{A}(G_1,G_2).
\end{equation}
So we have
\begin{equation*}
\begin{aligned}
&\hspace{4mm}q^{-{\langle M,H_{2}\rangle}-{\langle N,H_{1}\rangle}}\sum_{[L]\in{{\rm Iso}(\mathcal{A})}}\displaystyle\frac{|\mbox{\rm Ext}_{A}^{1}(M,N)_{L}|}{|\mbox{\rm Hom}_{A}(M,N)|}[G_{L}]\\
&=q^{-{\langle M,H_{2}\rangle}-{\langle N,H_{1}\rangle}+{\rm dim}_{k}\scriptstyle{\rm Hom}_{A}(H_1,N)-{\rm dim}_{k}\scriptstyle{\rm Hom}_{A}(G_1,N)}\sum\limits_{[W]\in{{\rm Iso}(\mathcal{B})}}|\mbox{\rm Ext}_{A}^{1}(G_1,G_2)_{W}|[W]\quad(\mbox {by}\ \eqref{4.7},\;\eqref{4.8})\\
&=q^{-{\langle M,H_{2}\rangle}-{\langle N,H_{1}\rangle}+{\langle H_{1},N\rangle}+{\rm dim}_{k}\scriptstyle{\rm Hom}(G_1,H_{2})}\sum\limits_{[W]\in{{\rm Iso}(\mathcal{B})}}\displaystyle\frac{|\mbox{\rm Ext}_{A}^{1}(G_1,G_2)_{W}|}{\mbox{\rm Hom}_{A}(G_1,G_2)}[W] \quad(\mbox {by}\ \eqref{4.9})\\
&=q^{-{\langle M,H_{2}\rangle}-{\langle N,H_{1}\rangle}+{\langle H_{1},G_2\rangle}-{\langle H_{1},H_2\rangle}+{\rm dim}_{k}\scriptstyle{\rm Hom}_{A}(G_1,H_{2})}[G_1]\diamond[G_2] \quad(\mbox {since}\ {\langle H_{1},G_2\rangle}={\langle H_{1},H_{2}+N\rangle})\\
&=q^{{\langle H_{1},G_2\rangle}-{\langle N,H_{1}\rangle}}q^{-({\langle M,H_{2}\rangle}+{\langle H_{1},H_2\rangle}-{\rm dim}_{k}\scriptstyle{\rm Hom}_{A}(G_1,H_{2}))}[G_1]\diamond[G_2] \\
&=q^{{\langle H_{1},G_2\rangle}-{\langle N,H_{1}\rangle}}[G_1]\diamond[G_2] \quad(\mbox {since}\ {\langle G_{1},H_2\rangle}={\langle M+H_{1},H_{2}\rangle})\\
&=q^{{\langle H_{1},G_2\rangle}-{\langle G_{2},H_{1}\rangle}+{\langle H_{2},H_{1}\rangle}}[G_1]\diamond[G_2] \quad(\mbox {since}\ {\langle G_{2},H_1\rangle}={\langle H_{2}+N,H_{1}\rangle}).
\end{aligned}
\end{equation*}
This shows that \eqref{4.6} is true, and thus \eqref{4.0} is also true. So, $\psi$ is a homomorphism of algebras.
Finally, we show that $\psi$ factorizes through the canonical surjection $\mathcal{H}(\mathcal{A})\to \mathcal{H}(\mathcal{A})/I(\mathcal{A})$.
Suppose $N\in{\mathcal{P}^{{\leqslant 1}}(\mathcal{A})}$. Since $H_2\in\mbox{\rm add}(_{A}A)$, we have $G_{2}\in{\mathcal{P}^{{\leqslant 1}}(A)}$ (see the first column in the diagram \eqref{4.4}). As $G_{2}$ lies in $\mathcal{B}$, it is projective. This implies
$G_{L}\cong{G_{1}\oplus G_{2}}$, and therefore
$$\psi([L])={q^{-\langle L,H_{L}\rangle}[G_{L}]\diamond[H_{L}]^{-1}}=
{q^{-\langle M\oplus N,H_{1}\oplus H_{2}\rangle}[G_{1}\oplus G_{2}]\diamond[H_{1}\oplus H_{2}]^{-1}}=\psi([N\oplus M]).$$
Consequently, $\psi$ induces a homomorphism of algebras
$\psi':\mathcal{H}(\mathcal{A})/I(\mathcal{A})\to\mathcal{SDH}(\mathcal{B}).$
Since $\psi([K])$ is invertible in $\mathcal{SDH}(\mathcal{B})$ for any $K\in{\mathcal{P}^{{\leqslant 1}}(\mathcal{A})}$, $\psi'$ induces a unique homomorphism of algebras
$\widetilde{\psi}:\mathcal{SDH}(\mathcal{A})\to\mathcal{SDH}(\mathcal{B}).$
Clearly, $\widetilde{\psi}\widetilde{\phi}={\rm Id}$, which means that $\widetilde{\phi}$ is injective. Since $\widetilde{\phi}$ is surjective, it is an isomorphism of algebras.
\end{proof}
A consequence of Proposition \ref{prop:4.7} is the following.
\begin{cor}\label{corollary:4.9} Let $A$ be a finite-dimensional algebra over $k$ and $_{A}T$ a strong tilting module with ${\rm proj.dim}(_AT)\leqslant 1$. Then there exists an isomorphism of algebras:
$\mathcal{SDH}(\mathcal{GP}(A))\cong\mathcal{SDH}(\mathcal{GP}^{\leqslant 1}(A)).$
\end{cor}
\begin{proof} By Proposition \ref{prop:4.7}, it suffices to show ${^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant 1}(A)}={\mathcal{GP}^{\leqslant 1}(A)}$. Let $M\in{\mathcal{GP}^{\leqslant 1}(A)}$. By \cite[Lemma 2.17]{cfh}, there exists an exact sequence $0\to M\to H\to G\to 0$ in $A$-mod with {\rm proj.dim$(_{A}H)\leqslant 1$} and $G\in{\mathcal{GP}(A)}$. It follows from Lemma \ref{prop:3.3}(2) that $G\in{\mathcal{W}{(_{A}T)}}$. Since $_{A}T$ is a strong tiling module with ${\rm proj.dim}(_AT)\leqslant 1$, we have $H\in{^{\perp}(_{A}T)}$. Hence $M\in{^{\perp}(_{A}T)}$. This implies ${\mathcal{GP}^{\leqslant 1}(A)}\subseteq{{^{\perp}(_{A}T)}}$.
\end{proof}
{{\bf Proof of Theorem \ref{thm:4.10}}.}
\begin{proof}
Let $\mathcal{A}_{1}:={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant 1}(A)}$ and $\mathcal{A}_{2}:={^{\perp}(T_{B})}\cap{\mathcal{GP}^{\leqslant 1}(B^{op})}$. By Corollary \ref{GV}(1), $F:=\mbox{\rm Hom}_{A}(-,T):\mathcal{A}_{1}\to\mathcal{A}_{2}$ is a duality of weakly $1$-Gorenstein exact categories. Moreover, $G:=\mbox{\rm Hom}_{B^{op}}(-,B):\mathcal{GP}(B^{op})\to \mathcal{GP}(B)$ is also a duality of weakly $1$-Gorenstein exact categories. By Propositions \ref{prop:4.7} and \ref{prop:4.6}, we obtain the following diagram of isomorphisms of algebras:
$$
\xymatrix@C=3em@R=3em{
\mathcal{SDH}(\mathcal{GP}(A)) \ar[r]^-{\widetilde{\phi}_{A}}
& \mathcal{SDH}(\mathcal{A}_{1})\ar[d]^{\Upsilon_{F}} \\
& (\mathcal{SDH}(\mathcal{A}_{2}))^{op}\ar[d]^{(\widetilde{\psi}_{B^{op}})^{op}} \\
\mathcal{SDH}(\mathcal{GP}(B))
& (\mathcal{SDH}(\mathcal{GP}(B^{op}))^{op}\ar[l]_-{(\Upsilon_{G})^{op}}}
$$
Define $$\Xi:={(\Upsilon_{G})^{op}}{(\widetilde{\psi}_{B^{op}})^{op}}{\Upsilon_{F}}{\widetilde{\phi}_{A}}: \;\mathcal{SDH}(\mathcal{GP}(A)) \to \mathcal{SDH}(\mathcal{GP}(B)).$$
Then $\Xi$ is an isomorphism of algebras. It remains to show that, for any $G\in{\mathcal{GP}(A)}$,
$$\Xi(G)=q^{-\langle L,G\rangle}[F(L)]^{-1}\diamond[F(Z)],$$
where $L$ and $Z$ are given in Theorem \ref{thm:4.10}.
In fact, since $G\in\mathcal{GP}(A)$, there is an exact sequence $0\to K\to P\to G\to 0$ in $A$-mod such that $P\in\mbox{\rm add}(_AA)$ and $K\in\mathcal{GP}(A)$. As $_{A}T$ is $1$-tilting, there is an exact sequence $0\to P\to T_{0}\to T_{1}\to0$ in $A$-mod with $T_{0},T_{1}\in{\mbox{\rm add}(_{A}T)}$. By pushout, we construct the diagram $(\sharp)$:
$$\xymatrix{&&0\ar[d]&0\ar[d]&\\
0\ar[r]&K\ar[r]\ar@{=}[d]&P\ar[r]\ar[d]&G
\ar[r]\ar[d]& 0\\
0\ar[r]&K\ar[r]&T_{0}\ar[d]\ar[r]&X
\ar[r]\ar[d]& 0\\
&&T_{1}\ar@{=}[r]\ar[d]&T_{1}\ar[d]&\\
&&0&\ 0.&\\
}$$
Since $T$ is $1$-tilting, $(_{A}T)^{\perp}={\rm Gen}(_AT)$, the smallest full subcategory of $A\ensuremath{\mbox{{\rm -mod}}}$ containing $_AT$ and being closed under direct sums and quotients. This forces $X\in{(_{A}T)^{\perp}}$. Since $T_1\in\mbox{\rm add}(_AT)\subseteq (_{A}T)^{\perp}$, the third column in the diagram $(\sharp)$ implies that the map
$G\to X$ is a left $(_{A}T)^{\perp}$-approximation of $G$. Now, let $f:G\to Z$ be a minimal left $(_{A}T)^{\perp}$-approximation with $L:={\rm Coker}(f)$. Then $f$ is injective; $Z$ and $L$ are isomorphic to direct summands of $X$ and $T_1$, respectively. In particular, $Z\in{\rm Gen}(_AT)$ and $L\in\mbox{\rm add}(_AT)$. Since $\mathcal{GP}(A)\cup \mbox{\rm add}(_AT)\subseteq \mathcal{A}_1$ and $\mathcal{A}_1$ is closed under extensions in $A\ensuremath{\mbox{{\rm -mod}}}$, we have $Z\in\mathcal{A}_1$. By $Z\in{\rm Gen}(_AT)$, there exists an exact sequence $0\to N\to H\to Z\to 0$ of $A$-modules with $H\in{\mbox{\rm add}(_{A}T)}$. Since $\mathcal{A}_1\subseteq A\ensuremath{\mbox{{\rm -mod}}}$ is a resolving subcategory, $N\in\mathcal{A}_{1}$. So, the sequence is exact in $\mathcal{A}_1$. Applying $F$ to the sequence yields an exact sequence $0\to F(Z)\to F(H)\to F(N)\to 0$ in $\mathcal{A}_2$. From $F(H)\in\mbox{\rm add}(B_B)$ and $F(N)\in{\mathcal{GP}^{\leqslant 1}(B^{op})}$, it follows that $F(Z)\in{\mathcal{GP}(B^{op})}$. Hence we have an exact sequence in $\mathcal{A}_2$
\begin{equation}\label{4.8.1}
0\longrightarrow F(L)\longrightarrow F(Z)\longrightarrow F(G)\longrightarrow 0
\end{equation}
such that $F(Z)$ is Gorenstein projective and $F(L)$ is projective. Moreover, there are equalities
\begin{equation}
\begin{aligned}\label{4.8.2}
\langle F(G),F(L)\rangle&=\displaystyle\textrm{dim}_{k}\mbox{\rm Hom}_{B^{op}}(F(G),F(L))-
\displaystyle\textrm{dim}_{k}\mbox{\rm Ext}^{1}_{B}(F(G),F(L))\\
&=\displaystyle\textrm{dim}_{k}\mbox{\rm Hom}_{A}(L,G)-
\displaystyle\textrm{dim}_{k}\mbox{\rm Ext}^{1}_{B}(L,G)\quad(\mbox{by}\ \mbox{Lemma} \ \ref{lemma:3.1})\\
&=\langle L,G\rangle.
\end{aligned}
\end{equation}
Thus
\begin{equation*}
\begin{aligned}
\Xi([G])&=((\Upsilon_{G})^{op}}{(\widetilde{\psi}_{B^{op}})^{op}}{\Upsilon_{F}}{\widetilde{\phi}_{A})([G])\\
&={(\Upsilon_{G})^{op}}{(\widetilde{\psi}_{B^{op}})^{op}}([F(G)])\\
&={(\Upsilon_{G})^{op}}(q^{-\langle F(G),F(L)\rangle}[F(L)]^{-1}\diamond[F(Z)])\quad(\mbox{by}\ \eqref{4.8.1})\\
&=q^{-\langle F(G),F(L)\rangle}[GF(L)]^{-1}\diamond[GF(Z)]\\
&=q^{-\langle F(G),F(L)\rangle}[\mbox{\rm Hom}_{A}(T,L)]^{-1}\diamond[\mbox{\rm Hom}_{A}(T,Z)]\quad(\mbox{by}\ \mbox{Lemma} \ \ref{lemma:3.1})\\
&=q^{-\langle L,G\rangle}[\mbox{\rm Hom}_{A}(T,L)]^{-1}\diamond[\mbox{\rm Hom}_{A}(T,Z)]\quad(\mbox{by}\ \eqref{4.8.2}).
\end{aligned}
\end{equation*}
This finishes the proof of Theorem \ref{thm:4.10}.
\end{proof}
\begin{cor}\label{corollary:4.9'} Let $_{A}T_{B}$ be a tilting bimodule with ${\rm proj.dim}(_AT)\leqslant 1$.
\begin{enumerate}
\item {\rm(}\cite[Theorem A15]{Lu2022}{\rm)} If $A$ is a finite-dimension 1-Gorenstein $k$-algebra, then there exists an isomorphism of algebras:
$\mathcal{SDH}(A)\cong\mathcal{SDH}(\mathcal{GP}(A)).$
\item {\rm(}\cite[Theorem A22]{Lu2022}{\rm)} If $A$ and $B$ are finite-dimension 1-Gorenstein $k$-algebras, then there exists an isomorphism of algebras: $\mathcal{SDH}(A)\cong\mathcal{SDH}(B)$.
\end{enumerate}
\end{cor}
\begin{proof}
Let $A$ be a finite-dimension 1-Gorenstein $k$-algebra. Then $A$-mod is a weakly 1-Gorenstein satisfying {\rm(E-a)-(E-d)}. Recall that $\mathcal{SDH}(A):=\mathcal{SDH}(A\ensuremath{\mbox{{\rm -mod}}})$ is the semi-derived Ringel-Hall algebra of $A$.
Set $_{A}T:=\mbox{\rm Hom}_{k}(A,k)$, the ordinary injective cogenerator for $A\ensuremath{\mbox{{\rm -mod}}}$.
Then $_{A}T$ is a strong tilting module and ${\rm proj.dim}(_AT)\leqslant 1$. Moreover, $A\ensuremath{\mbox{{\rm -mod}}}={\mathcal{GP}^{\leqslant 1}(A)}={^{\perp}(_{A}T)}\cap{\mathcal{GP}^{\leqslant 1}(A)}$. Now, $(1)$ is a direct consequence of Corollary \ref{corollary:4.9} and (2) follows from $(1)$ and Theorem \ref{thm:4.10}.
\end{proof}
\begin{remark}\label{Different}
We point out that our proof of Corollary \ref{corollary:4.9'} is different from the proof given in \cite{Lu2022}.
$(a)$ In the proof of \cite[Theorem A15]{Lu2022}, under the assumption that $A$ is $1$-Gorenstein, the map $\psi:\mathcal{H}(\mathcal{A})\to\mathcal{SDH}(\mathcal{GP}(A))$ (see the proof of Proposition \ref{prop:4.7} for $T:=\mbox{\rm Hom}_{k}(A,k)$) was shown to induce a unique morphism of $\mathcal{T}(A)$-bimodules $\widetilde{\psi}:\mathcal{SDH}(\mathcal{A})\to\mathcal{SDH}(\mathcal{GP}(A))$ by using an explicit description of $\mathcal{SDH}(\mathcal{A})$ as a $\mathcal{T}(A)$-bimodule (see \cite[Proposition A13]{Lu2022}), where $\mathcal{A}=A\ensuremath{\mbox{{\rm -mod}}}$ and $\mathcal{T}(A):=\mathcal{SDH}(\mathcal{P}^{\leqslant 1}(A))$ is a subalgebra of $\mathcal{SDH}(A)$. In our proof, we have shown that $\psi:\mathcal{H}(\mathcal{A})\to\mathcal{SDH}(\mathcal{GP}(A))$ for a general $1$-tilting module ${_A}T$ (see the proof of Proposition \ref{prop:4.7}) is an algebra homomorphism and automatically induces an algebra homomorphism $\widetilde{\psi}:\mathcal{SDH}(\mathcal{A})\to\mathcal{SDH}(\mathcal{GP}(A))$
which is the inverse of $\widetilde{\phi}$. Moreover, the proof of Proposition \ref{prop:4.7} does not involve the structure of $\mathcal{SDH}(\mathcal{A})$ as a $\mathcal{T}(A)$-bimodule.
$(b)$ In the proof of \cite[Theorem A22]{Lu2022}, when both $A$ and $B$ are $1$-Gorenstein, the additive equivalence between the coresolving subcategory $({_A}T)^{\perp}:=\{X\in A\ensuremath{\mbox{{\rm -mod}}}\mid \mbox{\rm Ext}_A^1(T,X)=0\}$ of $A\ensuremath{\mbox{{\rm -mod}}}$ and the resolving subcategory $\mathcal{Y}:=\{Y\in B\ensuremath{\mbox{{\rm -mod}}}\mid{\rm Tor}_1^B(T,Y)=0\}$ of $B\ensuremath{\mbox{{\rm -mod}}}$ (by the Brenner-Butler tilting theorem) was applied to establish the second isomorphism in the following algebra isomorphisms:
$$\mathcal{SDH}(A)\cong\mathcal{SDH}((_{A}T)^{\perp})\cong
\mathcal{SDH}(\mathcal{Y})\cong\mathcal{SDH}(B),$$
where $(_{A}T)^{\perp}$ and $\mathcal{Y}$ are weakly $1$-Gorenstein exact categories, and the first isomorphism is induced from the inclusion $({_A}T)^{\perp}\subseteq A\ensuremath{\mbox{{\rm -mod}}}$ (see \cite[Proposition A21]{Lu2022}). However, for a general algebra $A$, the category $(_{A}T)^{\perp}$ may not be weakly $1$-Gorenstein, and therefore $\mathcal{SDH}((_{A}T)^{\perp})$ is not well defined in general.
In our proof of Corollary \ref{corollary:4.9'}(2), we use the resolving dualities in Corollary \ref{GV}(1) and establish a series of algebra isomorphisms:
$$\mathcal{SDH}(A)\cong\mathcal{SDH}(\mathcal{GP}(A))\cong\mathcal{SDH}(\mathcal{GP}(B))\cong\mathcal{SDH}(B).$$
\end{remark}
\smallskip
{\bf Acknowledgements.} The research was supported by the National Natural Science Foundation of China (Grant 12122112, 12031014 and 12171206). Also, the author J. S. Hu thanks the Natural Science Foundation of Jiangsu Province (Grant No. BK20211358) and Jiangsu 333 Project for partial support.
|
\section{Introduction}
\label{sec:intro}
\IEEEPARstart{H}{uman} action recognition plays an important role in video surveillance, human-machine interaction, and sport video analysis~\cite{herath2017going}.
Different modality information, such as appearance, depth, optical flows, and body skeletons~\cite{bhardwaj2019efficient} has been used for human action recognition.
Among them, the skeleton consists of compact positions of major body joints~\cite{zhang2020semantics} and can provide highly effective information on human motion underlying different actions~\cite{johansson1973visual,bian2021structural}.
Skeleton-based action recognition is robust to appearance inconsistencies, different environments, and varying illuminations and is getting more accessible with the rapid development of sensor technology for capturing the skeleton.
3D skeletons simultaneously captured for the same person from different views are usually different~\cite{zhang2019view}, as shown in Figure~\ref{skeleton}, even if we try to transform them to the same coordinates. There are many reasons accounting for this phenomenon, such as altered reference coordinates, different occluded joints in different views, and inaccurate human pose estimation.
In practice, the skeleton data used for action recognition may be captured from different and even time-varying views, and such view variance can easily lead to incorrect feature representation and
recognition~\cite{zhang2019view,nie2019view}. Nowadays, view-invariant action recognition is still a challenging problem.
\begin{figure}
\centering
\includegraphics[trim={0 0 0 0},clip,scale=0.85]{figs/skeleton1.png}
\caption{An illustration of view variance of skeleton data: 3d skeleton data simultaneously captured for a same person, but from different views, are usually different due to altered reference coordinates, different occluded joints, or inaccurate human pose estimation.}
\label{skeleton}
\vspace{0pt}
\end{figure}
A commonly used strategy to improve view invariance in skeleton-based action recognition is to perform frame-level or sequence-level pre-processing of skeleton transformation~\cite{liu2016spatio,song2017end,liu2017global,zhang2019view}. Nevertheless, the frame-level pre-processing transforms the skeleton to the body center with the upper-body orientation aligned and results in partial loss
of relative motion information. To further preserve the motion information, sequence-level pre-processing performs the same transformation on all frames with the same parameters derived in the first frame. However, since the human body is non-rigid, the definition of body plane by the specific joints is not always suitable for the purpose of orientation alignment. Furthermore, it is almost impossible to eliminate the structural differences in different viewpoints through these transformations. Manually designing a view-invariant representation of the action is another approach, such as the displacement of joints within one frame or between frames~\cite{yang2012eigenjoints}, the histogram of joint orientations~\cite{xia2012view}, and some higher-level features like Lie group~\cite{vemulapalli2014human} and the covariant matrix of joints~\cite{hussein2013human}, but they could only deal with small
view changes. While the deep learning-based methods vastly outperform the traditional hand-crafted feature-based methods, most of them just rely on training on a large number of labeled samples taken from various views and cannot achieve actual view invariance in the underlying representation learning.
Most of the state-of-the-art methods for skeleton-based action recognition use supervised deep learning, which requires large-scale annotated data samples for training~\cite{liu2020disentangling,Cheng2020Skeleton}.
To address this problem, several recent studies attempt to leverage unsupervised learning for skeleton-based action recognition~\cite{zheng2018unsupervised,lin2020ms2l,su2020predict}.
In these studies, deep representations are learned for skeleton data sequences in terms of tasks like human motion prediction or regeneration, without using any action labels for supervision.
For algorithm evaluation, a simple linear classifier is finally trained for action recognition based on both the learned representations and action labels of the training data.
At present, there is still a relatively obvious performance gap between the supervised and unsupervised methods for skeleton-based action recognition. One primary cause lies in that the existing unsupervised skeleton representation learning cannot effectively exclude all the irrelevant factors for action recognition, such as view variation, skeleton deformation, and noise.
For this reason, we propose a new approach to enhancing representation learning by tackling view variation in skeleton-based action recognition
without using any manual action labels. Since the training data are unlabeled and have been already formated and related in data collection and prepossessing,
we follow previous studies~\cite{li2018unsupervised,ji2021view}, where a surrogate task is designed to exploit the inherent structure of unlabeled multi-view data for representation learning, by calling our proposed method unsupervised in this paper.
In this paper, we adopt contrastive representation learning to enhance the view-invariant unsupervised skeleton-based action recognition. More specifically, we develop a multi-view spatial-temporal graph (ST-Graph) contrastive representation learning (CRL) approach, in which the training loss function is defined to maximize the mutual information between features learned from skeletons that are simultaneously captured for the same person from different views. Such multi-view representations of the same person are pulled closer to each other in the embedding space through network training.
Furthermore, the proposed loss takes the form of a global-local contrastive one, which can also model the multi-scale co-occurrence relationships between the spatial and temporal domains.
In the testing stage, just like in previous works we only take one skeleton data sequence captured from an unknown view as the input of the network for skeleton-based action recognition.
We conduct comprehensive evaluation and analysis in our experiments to demonstrate that the proposed method can better learn view-invariant representations for improving the performance of skeleton-based human action recognition. The proposed method achieves a new state-of-the-art performance of unsupervised skeleton-based action recognition on two widely used multi-view benchmarks under the linear evaluation protocol.
The main contributions of this paper are as follows:
\begin{itemize}
\item[-] A contrastive learning framework for explicit learning of view-invariant representations for skeleton-based action recognition is proposed.
\item[-] We introduce a local-global spatial-temporal graph contrastive loss, combined with task uncertainty, to model the multi-scale co-occurrence relationship between spatial and temporal domains.
\item[-] Compared with existing methods that do not use ground-truth action labels in training, the proposed algorithm significantly boosts the performance on two widely used benchmarks of PKUMMD and NTU RGB+D,
\end{itemize}
The remainder of the paper is organized as follows. Section 2 gives a brief review of the related work on skeleton-based action recognition and contrastive learning. In Section 3, we describe our proposed multi-view contrastive representation learning approach. Section 4 describes the benchmark datasets and experimental setting and reports the experiment results, followed by a brief conclusion in Section 5.
\section{Related Work}
\subsection{Skeleton-based Action Recognition}
Skeleton-based action recognition is a very active and burgeoning area of research, due to its effective representation of motion dynamics. Much of the traditional skeleton-based action recognition work focuses on designing effective handcrafted features, especially the joint or body part-based features~\cite{vemulapalli2014human,yang2012eigenjoints,xia2012view,hussein2013human}. New methods have recently emerged in the literature to address the skeleton-based action representation with deep learning, including Recurrent Neural Network (RNN), Convolutional Neural Networks (CNN), and Graph Convolutional Network (GCN). Most of them aim to find more effective ways to model temporal and spatial information of skeleton sequences. The structure of RNN is suitable for processing sequential data and prior works have shown that RNN is especially good for handling varying-length skeleton sequences~\cite{wang2017modeling}. To extract discriminative spatial and temporal features of different actions, Song et al.~\cite{song2017end} propose a spatial and temporal attention module to assign different importance to each joint and frame within a sequence on top of RNN. CNN has the intrinsic ability to learn structural information from 2D or 3D grids, and it has also been used to encode skeleton sequences as pseudo-images for spatial-temporal representation learning~\cite{ke2017new}. Liu et al.~\cite{liu2017enhanced} firstly transform skeleton sequence into a series of color images, and then enhance visual and motion local patterns through mathematical morphology, finally propose a multi-stream CNN-based model to extract and fuse deep features from the enhanced-color images. GCN is the generalization of CNN to graphs and it can well represent the joint-based skeleton data. Therefore the use of GCN can automatically capture the patterns embedded in the spatial configuration of the joints as well as their temporal dynamics~\cite{yan2018spatial,liu2020disentangling,Cheng2020Skeleton}.
Cheng et al.~\cite{Cheng2020Skeleton} take novel shift graph operations and lightweight point-wise convolutions to replace regular graph convolutions. This way it reduces computation cost and provides flexible receptive fields for both the spatial graph and the temporal graph.
To avoid the laborious labeling of large-scale skeleton data, unsupervised skeleton-based action recognition has been studied by many researchers~\cite{lin2020ms2l,su2020predict}.
It performs the feature learning by an encoder-decoder structure, the input of which is a masked or original skeleton sequence, and the goal of training is to reconstruct the skeleton sequences from the encoded features. For the same reason, we focus on enhancing the view-invariant representative learning for skeleton-based action recognition without any manual action labeling using the GCN-based network in this paper.
\subsection{Contrastive Learning}
Contrastive learning aims to pull together an anchor and a “positive” sample in embedding space while pushing apart the anchor from many “negative” samples~\cite{khosla2020supervised}.
Therefore, contrastive losses are adopted to learn effective representations for pretext tasks in an unsupervised fashion.
Closely related to contrastive learning is the family of losses based on metric distance learning or triplets that depend on class labels to supervise the choice of positive and negative pairs~\cite{schroff2015facenet}. The key distinction between triplet losses and contrastive losses is that the former use exactly one positive and one negative pair per anchor and the positive pair of them is chosen from the same class and the negative pair is chosen from different classes. Contrastive learning generally uses just one positive pair for each anchor sample, selected using either co-occurrence~\cite{hjelm2018learning,henaff2020data} or data augmentation~\cite{chen2020simple}. The introduction of contrastive learning leads to a surge of interest in unsupervised visual representation learning~\cite{chen2020simple}. Wu et al.~\cite{wu2018unsupervised} maximize distinction between instances via a novel nonparametric softmax formulation and use a memory bank to store the instance class representation vector. For effective similarity measurement between samples in low-dimensional embedding space, other work explores the use of in batch samples for negative sampling instead of a memory bank~\cite{ye2019unsupervised,ji2019invariant}. Recently, researchers have attempted to relate the success of their methods to the maximization of mutual information between latent representations~\cite{bachman2019learning,henaff2020data}.
In probability theory and information theory, the mutual information of two random variables is a measure of their mutual dependence~\cite{mutualinformation}. It has important applications to contrastive learning~\cite{chen2020simple}. By maximizing mutual information between node and graph representations, some works, focusing on general graphs, have achieved state-of-the-art results in unsupervised node and graph classification tasks~\cite{velivckovic2018deep,sun2019infograph}. Maximizing mutual information between features extracted from multiple views of a shared context is analogous to human learning to represent observations generated by a shared cause driven by a desire to predict other related observations~\cite{bachman2019learning}. Aiming at a specific graph structure and task, we introduce a multi-view spatial-temporal graph contrastive representation learning method for view-invariant skeleton-based action recognition in this paper.
\begin{figure*}
\centering
\includegraphics[trim={0 0 0 0},clip,scale=0.6]{figs/method1.png}
\caption{The overall pipeline of the proposed multi-view ST-Graph CRL for view-invariant skeleton-based action recognition. $\mathbf{x}_a^{\mathrm{view1}}$ and $\mathbf{x}_a^{\mathrm{view2}}$ are from any two views of the multi-view skeleton sequence $\mathbf{X}_a$. $\mathbf{x}_b^{\mathrm{view*}}$ is from any view of the multi-view skeleton sequence $\mathbf{X}_b$. This approach pulls together skeletons simultaneously captured for the same person from different views in embedding space, while pushing apart the others.}
\label{method1}
\vspace{0pt}
\end{figure*}
\section{Multi-View ST-Graph Contrastive Representation Learning}
Inspired by recent contrastive learning algorithms, we propose an approach to learning view-invariant representations without any manual action labeling by maximizing the mutual information between skeleton sequences that are simultaneously taken for the same person but from different views, via a global-local contrastive loss in the latent space. The overall pipeline of the proposed approach is illustrated in Figure~\ref{method1}.
Specifically, a stochastic data augmentation module $\varphi(\cdot)$ that transforms any given data example randomly to encourage learning a more robust representation for the downstream task. Then, an ST-GCN structural encoder $g(\cdot)$ extracts representation vectors from augmented data examples. We maximize the representation agreement between samples simultaneously taken for the same person but from different views on a global level and a local level. On the global level, a small neural network projection head $f_\psi(\cdot)$ maps the representations to a latent space by applying a global contrastive loss. On the local level, as shown in Figure~\ref{method2}, an ST-Graph partitioning function $\rho(\cdot)$ splits the graph structural representation of the whole skeleton sequence into multi-local subgraphs, and then a projection head $f_\phi(\cdot)$ maps the representations to a latent space by applying a local contrastive loss. Moreover, to effectively combine global and local contrastive losses, we adjust their relative weights based on task uncertainty.
\begin{figure*}
\centering
\includegraphics[trim={0 0 0 0},clip,scale=0.7]{figs/method2.png}
\caption{An more detail illustration of local contrastive in Figure~\ref{method1}. ST-Graph structural representations $\mathbf{h}_a^{\mathrm{view1}}$ and $\mathbf{h}_a^{\mathrm{view2}}$ are from any two views of the multi-view skeleton sequence $\mathbf{X}_a$ while $\mathbf{h}_b^{\mathrm{view*}}$ is from any view of the multi-view skeleton sequence $\mathbf{X}_b$. This loss aims to pull together skeleton sequence regions simultaneously captured for the same person from different views in embedding space, while pushing apart the others.}
\label{method2}
\vspace{0pt}
\end{figure*}
Before getting into the details of the approach, we state the main notations. Similar to previous studies~\cite{Cheng2020Skeleton, zhang2020semantics}, we organize skeleton sequence of an action sample as an undirected spatial-temporal graph $\mathbf{x} = (\mathbf{\mathcal{J},\mathcal{E}})$, where $\mathcal{J} = \{\mathbf{j}_{ti}\mid t=1,...,T; i=1,..., M\} $ denotes a set of vertices, corresponding to $T$ frames and $M$ body joints per frame, and $\mathcal{E}$ is the set of edges, indicating the connections between nodes. Then, we represent a multi-view skeleton sample as $\mathbf{X}=\{\mathbf{x}^v\}_{v=1}^{V}$, where $V$ represents the number of viewpoints, which could be as many as needed, and $v$ indicates the specific $v$-th viewpoint. For many multi-view skeleton samples, we also use $\mathbf{x}_i^v$ to denote the $v$-th view of the $i$-th multi-view skeleton sample $\mathbf{X}_i$.
\subsection{Multi-view skeletal data augmentation.} Data augmentation aims to create novel and realistically rational data by applying a certain transformation to the original training data without affecting their semantic meanings.
It has been demonstrated that contrastive learning usually needs stronger data augmentation than supervised learning~\cite{chen2020simple}. Meanwhile, for specific graphs, certain data augmentations might be more effective than the others~\cite{you2020graph}.
Let an augmented skeleton sequence be $\hat{\mathbf{x}}_i^v = \varphi(\mathbf{x}_i^v)$, where $\varphi(\cdot)$ is the augmentation function. In this paper, we apply temporal subgraph as the data augmentation for multi-view CRL, with definitions as follows: it samples a segment from $\mathbf{x}_i^v$ along the temporal dimension. As the length of a skeleton sequence is fixed to 100 frames, we randomly sample 95 consecutive frames and then extend it to 100 frames by linear interpolation. This data augmentation increases the robustness of action recognition when the starting and ending frames of the action cannot be accurately determined and the skeleton sequences captured from different views do not have perfect temporal alignment.
\subsection{ST-GCN structural encoder.} The ultimate goal of ST-Graph CRL is to train a powerful skeleton sequence encoder $g(\cdot)$ to get view-invariant representation for skeleton-based action recognition without any manual action labeling. Specifically, to effectively model the co-occurrence relationships among joints in both spatial and temporal domains, we apply an ST-GCN structural encoder, which extracts representation $\mathbf{h}_i^v$ from the augmented skeleton sequence $\hat{\mathbf{x}}_i^v$. Specifically, it contains two parts: spatial graph convolution and temporal graph convolution.
For spatial graph convolution, the neighbor set of joints is defined as an adjacent matrix $\mathbf{A} \in \{0,1\}^{M \times M}$ according to $\mathcal{E}$, which is typically partitioned into 3 partitions: the centripetal group containing neighboring nodes that are closer to the skeleton center, the node itself and otherwise the centrifugal group. For individual skeleton, let $\mathbf{F} \in \mathbb{R} ^{M \times C}$ and $\mathbf{F}^{\prime} \in \mathbb{R} ^{M \times C^{\prime}}$ denote the input and output feature during the processing respectively, where $C$ and $C^{\prime}$ are the input and output feature dimensions. The graph convolution is computed as:
\begin{equation}
\mathbf{F}^{\prime} = \sum_{p\in\mathcal{P}} \bar{\mathbf{A}}_{p}\mathbf{F}\mathbf{W}_p,
\end{equation}
where $\mathcal{P} = \{\mathrm{root, centripetal, centrifugal}\}$ denotes the spatial partitions, $\bar{\mathbf{A}}_{p} = \mathbf{\Lambda}_p^{-\frac{1}{2}}\mathbf{A}_p \mathbf{\Lambda}_p^{-\frac{1}{2}}\in \mathbb{R} ^{M \times M}$ is the normalized adjacent matrix and $\mathbf{\Lambda}_p^{ij} = \sum_{j}(\mathbf{A}_p^{ij})+\alpha$, $\alpha$ is set to 0.001 to avoid empty rows. $\mathbf{W}_p\in\mathbb{R} ^{1\times1\times C \times C^{\prime}} $ is the weight of the $1\times1$ convolution for each partition group. For temporal dimension, we construct temporal graph by connecting identical joints in consecutive frames and use regular 1D convolution on the temporal dimension as the temporal graph convolution.
The ST-GCN structural encoder comprises a series of dynamic spatial-temporal graph convolution blocks stacked one above the other.
In this form, there existed many specific models with subtle differences~\cite{yan2018spatial, shi2019two, Cheng2020Skeleton}. The proposed approach does not place any restriction on the ST-GCN structural encoder, as long as it maintains the feature of the spatial-temporal graph structure. In our implementation, we adopt the network recently proposed by~\cite{Cheng2020Skeleton} as the ST-GCN structural encoder.
\subsection{ST-Graph partitioning function.}
As stated in Li et al.~\cite{li2018deeper}, the graph convolution operation can be considered Laplacian smoothing for node features over graph topology. The Laplacian smoothing computes the new node features as the weighted average of itself and its neighbors. It helps make nodes in the same cluster tend to learn similar representations. Nevertheless, it may also lead to the over-smoothing
problem and make nodes indistinguishable as the number of network layers increases. Meanwhile, it may concentrate more on node features and make the learned embeddings lack structural information. In short, ST-GGN can handle most simple cases but may ignore local details on a complicated graph.
Given the above problems, we enhance the representation by giving more consideration to specific characteristics of local regions. Specifically, we include an ST-Graph partitioning function $\rho(\cdot)$ to split the feature of the whole skeleton sequence $\mathbf{h}_i^v$ into multi-local subgraphs $\mathbf{l}_{i,s}^v, s\in[1,...,S]$, where $S$ represents the number of generated subgraphs, $i$ and $s$ indicate sample index and subgraph index, respectively. The choice of partitioning strategies has a strong impact on not only the performance of recognition networks but also the design of the networks~\cite{fan2020application}. Several graph partitioning algorithms have already been developed and they are often either edge cut~\cite{andreev2006balanced}, which evenly partitions vertices and cuts edges, or vertex cut~\cite{bourse2014balanced}, which evenly partitions edges by replicating vertices. There have also been hybrid algorithms~\cite{li2019topox}, which cut both edges and vertices. In this paper, we adopt two simple rule-based edge cut style partitioning strategies to segment the skeleton spatial-temporal feature graph. Specifically, vertices of the ST-Graph are evenly partitioned into $S$ segments along the spatial dimension or the temporal dimension by cutting edges, as shown in figure~\ref{partition}.
\begin{figure}
\centering
\includegraphics[trim={0 0 0 0},clip,scale=0.65]{figs/graphcut.png}
\caption{ST-Graph spatial or temporal partitioning strategies. The spatial-temporal feature graph are evenly partitioned along different dimensions by cutting edges.}
\label{partition}
\vspace{0pt}
\end{figure}
\subsection{Projection head.}
Recent work by Chen et al.~\cite{chen2020simple} found that mapping features to another latent space before contrastive loss calculation can be more effective. In this way, the features before a nonlinear projection are the learned representations, where information loss of raw data induced by the contrastive loss can be relieved. Therefore, in this paper, the representations $\mathbf{h}_i^v$ and $\mathbf{l}_{i,m}^v$ are mapped to another latent space through an MLP with one hidden layer, respectively. We name this module as projection head and add it to global and local contrastive learning subnetworks. Meanwhile, a global pooling is performed on $\mathbf{h}_i^v$ and $\mathbf{l}_{i,m}^v$ to get a fixed dimension feature vector for each ST-Graph to aggregate the node features before the projection head. Formally, the process is defined as:
\begin{equation}
\resizebox{0.9\linewidth}{!}{$
\displaystyle
\begin{split}
\mathbf{G}_i^v = f_\psi(\mathrm{pool}(\mathbf{h}_i^v)) = \mathbf{W}^{(\psi, 2)}\sigma(\mathbf{W}^{(\psi, 1)}\mathrm{pool}(\mathbf{h}_i^v)), \\
\mathbf{L}_{i,m}^v = f_\phi(\mathrm{pool}(\mathbf{l}_{i,m}^v)) = \mathbf{W}^{(\phi, 2)}\sigma(\mathbf{W}^{(\phi, 1)}\mathrm{pool}(\mathbf{l}_{i,m}^v)),
\end{split}$}
\end{equation}
where $f_\psi(\cdot)$ and $f_\phi(\cdot)$ represent global and local projection heads. $\mathbf{G}_i^v$ and $\mathbf{L}_{i,m}^v$ are the global and local representations in another latent space. $\mathrm{pool}(\cdot)$ is a global pooling function. $\sigma$ is a ReLU nonlinearity and $\mathbf{W}$'s are learned weights of MLP. Note that the output of $\mathrm{pool}(\mathbf{h}_i^v)$ is named as $\mathbf{H}_i^v$, which is the representation we learned for the skeleton sequence based on ST-Graph CRL.
\subsection{Global-Local contrastive learning.}
A global representation can well capture the common knowledge of action patterns among all the regions in the skeleton sequence and hence possesses the nice merit in terms of model generalization while a local representation targets personalization of individual regions. As mentioned above, we propose several ST-Graph partitioning strategies to segment the graph into multiple local subgraphs. In this section, a global-local contrastive learning loss is proposed to effectively model the multi-scale co-occurrence relationship between spatial and temporal domains in the ST-Graph. For this, we define different positive pairs in global and local scenarios and maximize the consistency between the positive pairs compared with corresponding negative pairs using global and local contrastive loss functions. Meanwhile, the two contrastive loss functions are combined with task uncertainty in order to balance the trade-off between generalization and personalization of representation.
\textbf{Global contrastive loss.} Given two global representations $\mathbf{G}_a^{v1}$ and $\mathbf{G}_b^{v2}$, we specify that they form a positive pair if $a$ is equal to $b$, else they form a negative pair. It means multiple skeleton sequences, if simultaneously taken for the same person from different views, will be pulled together in embedding space, otherwise will be pulled apart, which is shown in Figure~\ref{method1}. Therefore, not only skeleton representations can be effectively learned without any action label information, but also their view-invariant property of them can be enhanced during multi-view contrastive learning. To achieve this, we adopt the normalized temperature-scaled cross-entropy loss~\cite{chen2020simple}. Specifically, we randomly sample a minibatch of $N$ examples and define the contrastive prediction task on pairs of skeleton sequences. Note that each example consists of $V$ skeleton sequences collected from $V$ different views, resulting in $VN$ data points. Given $V$ positive pairs in an example, we treat the other $V(N-1)$ data points within a minibatch as negative examples. Let $\mathbf{u}$ and $\mathbf{v}$ denote representations of two data points. To measure similarity, we define $\mathrm{sim}(\mathbf{u}, \mathbf{v}) = \mathbf{u}^\top \mathbf{v} / \left\|\mathbf{u}\right\| \left\|\mathbf{v}\right\|$ that denotes the dot product between $\ell_2$ normalized $\mathbf{u}$ and $\mathbf{v}$. Then, the global loss function for positive pairs of example $i$ is defined as
\begin{equation}
\begin{aligned}
\resizebox{.91\linewidth}{!}{$
\displaystyle
\begin{split}
& \ell_{i}^{\mathrm{global}}=-\log\\ & \frac{\sum_{v1,v2=1}^{V} \mathbbm{1}{[v1 \neq v2]}\exp \left(\operatorname{sim}\left(\mathbf{G}_{i}^{v1}, \mathbf{G}_{i}^{v2}\right) / \tau\right)}{\sum_{k=1}^{N}\sum_{v1,v2=1}^{V} \mathbbm{1}{\begin{tiny}
\begin{bmatrix}
k \neq i \\
v1 \neq v2
\end{bmatrix}
\end{tiny}} \exp \left(\operatorname{sim}\left(\mathbf{G}_{i}^{v1}, \mathbf{G}_{k}^{v2}\right) / \tau\right)},
\end{split}
$}
\end{aligned}
\end{equation}
where $\mathbbm{1}{[v1 \neq v2]} \in \{0, 1\}$ is an indicator function evaluating to 1 if $v1 \neq v2$, $\mathbbm{1}{\begin{tiny}
\begin{bmatrix}
k \neq i \\
v1 \neq v2
\end{bmatrix}
\end{tiny}}$ is also an indicator function evaluating to 1 if $k \neq i$ and $v1 \neq v2$ are simultaneously satisfied, otherwise evaluating to 0. $\tau$ denotes a temperature parameter. For a minibatch, the global contrastive loss $\mathcal{L}_{\mathrm{global}}$ is computed across all examples,
\begin{equation}
\mathcal{L}_{\mathrm{global}} = \frac{1}{VN}\sum_{m=1}^{N}\ell_{m}^{\mathrm{global}},
\end{equation}
where $N$ is the batchsize.
\textbf{Local contrastive loss.} Local contrastive loss is calculated among the local representations, as illustrated in Figure~\ref{method2}. Given two local representations $\mathbf{L}_{a,s1}^{v1}$ and $\mathbf{L}_{b,s2}^{v2}$, we specify that they form a positive pair if both $a=b$ and $s1=s2$ are satisfied, else they form a negative pair. From the composition of the positive and negative pairs, the contrastive loss achieves the same effect as the global one in the local scale when subgraph indices are consistent for all pairs. Besides, it can also handle the over-smoothing and the structural information lacking problems by contrasting among local regions in a sequence when sample indices are consistent for all pairs. The definition of local contrastive loss is the same as the global one. But because of the extra subgraph dimension, there are $VS$ positive pairs and $VS(N-1)$ negative pairs in a sample. Formally, the local contrastive loss function for positive pairs of example $i$ is defined as
\begin{equation}
\begin{aligned}
\resizebox{.91\linewidth}{!}{$
\displaystyle
\begin{split}
&\ell_{i}^{\mathrm{local}}=-\log\\& \frac{\sum_{s=1}^{S}\sum_{v1,v2=1}^{V} \mathbbm{1}{[v1 \neq v2]} \exp \left(\operatorname{sim}\left(\mathbf{L}_{i,s}^{v1}, \mathbf{L}_{i,s}^{v2}\right) / \tau\right)}{\sum_{k=1}^{N}\sum_{s1,s2=1}^{S}\sum_{v1,v2=1}^{V}
\mathbbm{1}{\begin{tiny}
\begin{bmatrix}
k \neq i \\
s1 \neq s2\\
v1 \neq v2
\end{bmatrix}
\end{tiny}} \exp \left(\operatorname{sim}\left(\mathbf{L}_{i,s1}^{v1}, \mathbf{L}_{k,s2}^{v2}\right) / \tau\right)},
\end{split}
$}
\end{aligned}
\end{equation}
where $\mathbbm{1}{\begin{tiny}
\begin{bmatrix}
k \neq i \\
s1 \neq s2\\
v1 \neq v2
\end{bmatrix}
\end{tiny}}$ is an indicator function that needs the three inequalities are simultaneously true. $S$ is the number of split subgraphs. For a minibatch, the local contrastive loss $\mathcal{L}_{\mathrm{local}}$ also needs to be computed across all examples,
\begin{equation}
\mathcal{L}_{\mathrm{local}} = \frac{1}{SVN}\sum_{m=1}^{N}\ell_{m}^{\mathrm{local}}.
\end{equation}
Corresponding to spatially or temporally partitioning the ST-Graph into multiply subgraphs, the notation of local contrastive loss is $\mathcal{L}_{\mathrm{spalocal}}$ or $\mathcal{L}_{\mathrm{temlocal}}$.
\textbf{Global-Local contrastive loss.} Global-Local contrastive loss is concerned about jointly optimizing the related global and local contrastive loss functions. In this paper, the popular approach of using a linear combination of them as a total loss function is abandoned. Because manually tuning their weight hyper-parameters is expensive and intractable. Instead, following the work of~\cite{wang2020m2grl}, we adjust each loss's relative weight in the total loss function by deriving a multi-task loss function based on maximizing the Gaussian likelihood with task-dependent uncertainty during model training. We define the global-local contrastive loss $\mathcal{L}$ as follows:
\begin{equation}
\mathcal{L} = \frac{1}{\sigma_1^2}\mathcal{L}_{\mathrm{global}} + \frac{1}{\sigma_2^2}\mathcal{L}_{\mathrm{local}}+
\log(\sigma_1^2) + \log(\sigma_2^2),
\end{equation}
where $\sigma_1$ and $\sigma_1$ associate with the task uncertainty and can be interpreted as the relative weights of respective loss terms. $\log(\sigma_1^2)$ and $\log(\sigma_2^2)$ serve as regularizers to avoid over-fitting. All network parameters and the uncertainty task weights are trainable and optimized by gradient back propagation.
The proposed multi-view ST-Graph CRL is summarized as Algorithm~\ref{alg}.
\begin{algorithm}
\caption{Multi-view spatial-temporal graph contrastive representation learning algorithm}
\label{alg}
\textbf{Input}: Augmentation $\varphi(\cdot)$, global pooling $\mathrm{pool}(\cdot)$, ST-Graph partitioning function $\rho(\cdot)$, ST-GCN structural encoder $g(\cdot)$, global and local projection heads $f_\psi(\cdot)$ and $f_\phi(\cdot)$, training multi-view skeleton sequences $\{\mathbf{X}_i = \{\mathbf{x}_i^v\}_{v=1}^{V} \}_{i=1}^{N}$, global contrastive loss $\mathcal{L}_{\mathrm{global}}$, local contrastive loss $\mathcal{L}_{\mathrm{local}}$, similarity measurement function $sim(\cdot)$.\\
\textbf{Parameters}: Learnable relative weight parameters for global and local contrastive loss: $\sigma_1$ and $\sigma_2$; number of views $V$; number of split subgraphs $S$; number of samples in one batch $K$; temperature parameter $\tau$.
\begin{algorithmic}[1]
\WHILE{sampled batch~$\{\{\mathbf{x}_i^v\}_{v=1}^{V}\}_{i=1}^{K}$}
\WHILE{$i=1$ to $K$}
\WHILE{$v=1$ to $V$}
\STATE $\mathbf{h}_i^v = g(\varphi(\mathbf{x}_i^v))$
\STATE $\mathbf{G}_i^v = f_\psi(\mathrm{pool}(\mathbf{h}_i^v))$
\STATE $\{\mathbf{l}_{i,s}^v\}_{s=1}^S = \rho(\mathbf{h}_i^v)$
\WHILE{$s=1$ to $S$}
\STATE $\mathbf{L}_{i,s}^v = f_\phi(\mathrm{pool}(\mathbf{l}_{i,s}^v))$
\ENDWHILE
\ENDWHILE
\ENDWHILE
\WHILE{$i=1$ to $K$}
\STATE
$\ell_{i}^{\mathrm{global}}=-\log$
$\frac{\sum_{v1,v2=1}^{V} \mathbbm{1}{[v1 \neq v2]}\exp \left(\operatorname{sim}\left(\mathbf{G}_{i}^{v1}, \mathbf{G}_{i}^{v2}\right) / \tau\right)}{\sum_{k=1}^{K}\sum_{v1,v2=1}^{V} \mathbbm{1}{\begin{tiny}
\begin{bmatrix}
k \neq i \\
v1 \neq v2
\end{bmatrix}
\end{tiny}} \exp \left(\operatorname{sim}\left(\mathbf{G}_{i}^{v1}, \mathbf{G}_{k}^{v2}\right) / \tau\right)}$
\STATE
$\ell_{i}^{\mathrm{local}}=-\log$
$\frac{\sum_{s=1}^{S}\sum_{v1,v2=1}^{V} \mathbbm{1}{[v1 \neq v2]} \exp \left(\operatorname{sim}\left(\mathbf{L}_{i,s}^{v1}, \mathbf{L}_{i,s}^{v2}\right) / \tau\right)}{\sum_{k=1}^{K}\sum_{s1,s2=1}^{S}\sum_{v1,v2=1}^{V}
\mathbbm{1}{\begin{tiny}
\begin{bmatrix}
k \neq i \\
s1 \neq s2\\
v1 \neq v2
\end{bmatrix}
\end{tiny}} \exp \left(\operatorname{sim}\left(\mathbf{L}_{i,s1}^{v1}, \mathbf{L}_{k,s2}^{v2}\right) / \tau\right)}$
\ENDWHILE
\STATE $\mathcal{L}_{\mathrm{global}} = \frac{1}{VK}\sum_{m=1}^{K}\ell_{m}^{\mathrm{global}}$
\STATE $\mathcal{L}_{\mathrm{local}} = \frac{1}{SVK}\sum_{m=1}^{K}\ell_{m}^{\mathrm{local}}$
\STATE $\mathcal{L} = \frac{1}{\sigma_1^2}\mathcal{L}_{\mathrm{global}} + \frac{1}{\sigma_2^2}\mathcal{L}_{\mathrm{local}}+
\log(\sigma_1^2) + \log(\sigma_2^2)$
\STATE update networks $g(\cdot)$, $f_\psi(\cdot)$, $f_\phi(\cdot)$, $\sigma_1$ and $\sigma_2$ to minimize $\mathcal{L}$
\ENDWHILE
\STATE \textbf{return} encoder model $g(\cdot)$, and throw away projection heads $f_\psi(\cdot)$ and $f_\phi(\cdot)$
\end{algorithmic}
\end{algorithm}
\section{Experiment}
\subsection{Dataset}
We evaluate the proposed method on two public available multi-view action recognition benchmarks: NTU RGB+D~\cite{shahroudy2016ntu} and PKUMMD~\cite{Liu2017PKU}. We briefly describe them below.
\textbf{NTU RGB+D (NTU).} NTU is a large-scale multi-modal action recognition dataset. It is composed of 56,880 samples over 60 classes captured from 40 distinct subjects and three Kinect cameras. Each action in the samples involves one or two people. The dataset is very challenging due to the large intra-class and view variations. The original paper of the NTU recommends two benchmarks: 1) Cross-subject (CS): all samples from a selected group of subjects are used for training and the rest samples for testing. 2) Cross-view (CV): the training set contains samples that are captured
by cameras 2 and 3, and the testing set contains videos
that are captured by camera 1. We follow this convention and report performance on both benchmarks.
\textbf{PKUMMD.} PKUMMD is a new large-scale benchmark for continuous multi-modality 3D human action understanding and covers a wide range of complex human activities with well-annotated information. It contains almost 20,000 action instances in 51 action categories, performed by 66 subjects in three different view Kinect sensors. PKUMMD consists of two subsets: PKUMMD-I is an easier subset for action recognition, while PKUMMD-II is more challenging with more skeleton noise caused by large view variation. We conduct experiments under the cross-subject protocol on the two subsets.
\subsection{Implementation Details}
\subsubsection{Pre-training without any action label information}
In ST-Graph CRL, an ST-GCN structural encoder $g(\cdot)$, a global projection head $f_\phi(\cdot)$ and a local projection head $f_\psi(\cdot)$ are pre-trained using multi-view skeleton sequences without any action label information. We use SGD with Nesterov momentum 0.9 to pre-train them for 40 epochs. The learning rate is set to 0.1 and divided by 10 at epoch 20, 30, and 35. The batch size is set to 16 for all experiments. The sequence length $T$ is set to 100. The temperature parameter for global-local contrastive loss is set to 0.07. The number of subgraph $S$ is set to 5. $V$ is set to 2, which means each sample includes two skeleton sequences, simultaneously taken from different views.
\subsubsection{Evaluation Protocol}
To validate the effectiveness of the proposed representation learning method, we follow the linear evaluation protocol~\cite{wang2020m2grl,chen2020simple}, which is commonly used to evaluate unsupervised learning methods. In this way, a linear classifier attached to the frozen encoder model $g(\cdot)$ is trained with the annotated dataset. We report Top-1 accuracy on the testing set as a quantitative evaluation indicator. The classifier is trained for 45 epochs, with the learning rate divided by 10 at epoch 25, 35, and 40. The other settings remain the same as the pre-training.
\subsection{Comparison Experiments}
To quantitatively evaluate the performance, Table~\ref{state-of-the-art} and Table~\ref{crossdataset} list the linear evaluation results of ST-Graph CRL and other state-of-the-art unsupervised methods on PKUMMD and NTU benchmarks. The model which only trains the linear classifier and freezes the randomly initialized encoder is denoted as ST-Graph Rand. We regard this model as one of our baselines. The models implementing ST-Graph contrastive learning in single-view and multi-view scenarios are denoted as ST-Graph CRL SV and ST-Graph CRL MV, respectively. In the single view version, we maximize the mutual information between skeleton ST-Graph representations of one augmented instance and another augmented instance of an identical skeleton sequence, to learn inherent action patterns of different skeleton transformations. For the evaluation of P\&C FW on the action recognition task, we reproduce the coder of P\&C FW with a linear evaluation protocol. The temporal subgraph is the default data augmentation method we adopt in these experiments.
\begin{table*}
\caption{Comparison of action recognition performance of the proposed ST-Graph CRL and other state-of-the-art methods.}
\centering
\begin{tabular}{clrrrrr}
\toprule
Supervised &Models & PKUMMD-I & PKUMMD-II & NTU (CS) & NTU (CV)\\
\midrule
Yes & ST-Graph & 94.45 & 56.75 & 87.82 & 95.13 \\
No &ST-Graph Rand & 30.14 & 10.58 & 19.55 & 23.23 \\
No&LongT GAN~\cite{zheng2018unsupervised} & 67.70 & 25.95 & 52.14 & - \\
No&P\&C FW~\cite{su2020predict} & 67.62 & 35.90 & 32.50 & 35.67 \\
No&$\mathrm{M}^2$SL~\cite{lin2020ms2l} & 64.86 & 27.63 & 52.55 & - \\
No&CAE+~\cite{rao2021augmented} & - & - & 58.50 & 64.80 \\
No&ST-Graph CRL SV (\textbf{Ours}) & 68.42 & 31.80 & 60.24 & 59.79 \\
No&ST-Graph CRL MV(\textbf{Ours}) & \textbf{83.62} & \textbf{39.89} & \textbf{74.71} & \textbf{82.62} \\
\bottomrule
\end{tabular}
\label{state-of-the-art}
\vspace{0pt}
\end{table*}
\subsubsection{Comparison with State-of-the-art} In existing studies~\cite{su2020predict,lin2020ms2l}, the pre-training and evaluation are usually conducted on the same dataset. An overall summary of the results is given in Table~\ref{state-of-the-art}, where the proposed method has returned significantly improved performance in the unsupervised methods that do not use action labels for training. As we can see, ST-Graph CRL MV is far beyond the performance of random baseline and other state-of-the-art unsupervised methods and greatly reduces the gap to the models trained with action annotation. NTU(CV) is a suitable benchmark to evaluate the model's robustness to the viewpoint difference. Here, we can see that model’s Top-1 accuracy of ST-Graph CRL MV in NTU(CV) is 82.62\%, while ST-Graph Rand and P\&C FW are only 23.23\% and 35.67\%, respectively. Therefore, the multi-view contrastive learning significantly improved the view-invariant property of skeleton representation. Even in a single view scenario, under the truly unsupervised setting, the performances of ST-Graph CRL SV are quite outstanding, which performs better than almost all the baselines. It achieves high recognition accuracies of 60.21\% and 59.79\% on NTU(CS) and NTU(CV), respectively, which proves that our global-local contrastive learning of augmented skeletons of the same sample also works well. From the comparison of ST-Graph CRL SV and MV, we can see that substantial improvements are made in each benchmark. It proves that CRL between the multi-view skeletons brings in a giant performance leap for unsupervised skeleton-based action recognition.
\begin{table*}
\caption{Performance of transfer learning setting and linear evaluation.}
\centering
\begin{tabular}{clrrr}
\toprule
Supervised & Models & PKUMMD-I & PKUMMD-II\\
\midrule
Yes & ST-Graph & 90.56 & 55.01 \\
No & P\&C FW~\cite{su2020predict} & 63.31 & 23.61 \\
No & ST-Graph CRL SV (\textbf{Ours}) & 76.29 & 39.83 \\
No & ST-Graph CRL MV(\textbf{Ours}) & \textbf{82.21} & \textbf{46.98} \\
\bottomrule
\end{tabular}
\label{crossdataset}
\vspace{0pt}
\end{table*}
\subsubsection{Transfer learning performance.} To further evaluate whether the proposed ST-Graph CRL can gain knowledge of related tasks, we investigate the transfer learning performance of our model~\cite{lin2020ms2l}. As the representations learned from large-scale data are more generalizable, we regard the NTU as the source dataset and PKUMMD-I and PKUMMD-II as the target datasets. We conduct the pre-training on the source dataset and the evaluation of target datasets. Under this setting, the samples used for pre-training and linear evaluation are completely different in terms of viewpoints, action patterns, and so on, which is more in accordance with the practical scenarios. The results are summarized in Table~\ref{crossdataset}, from which ST-Graph CRL MV gets better results of 82.21\% for PKUMMD-I and 46.98\% for PKUMMD-II when models are pre-trained without action annotations. Apart from that, together with Table~\ref{state-of-the-art}, we can see that the accuracies of P\&C FW are reduced from 67.62\% and 35.90\% to 63.31\% and 23.61\%, respectively, while most ST-Graph CRLs are boosted from 68.42\%, 31.80\%, 83.62\%, and 39.89\% to 76.29\%, 39.83\%, 82.21\%, and 46.98\%, respectively, when the training and testing datasets are from consistent to inconsistent. Meanwhile, the performances of models pre-trained with action annotations also decrease in this transfer learning setting from 94.45\% and 56.75\% to 90.56\% and 55.01\%, respectively. One possible reason is that ST-Graph CRL can take advantage of a larger training set more effectively with less influence from the data distribution difference between different datasets. It can be concluded that the representations learned in ST-Graph CRL have a good generalization ability.
\subsection{Ablation Experiments}
For a specific ST-Graph structural encoder, the performance of ST-Graph CRL is mainly determined by the following four components: multi-view skeleton contrastive mechanism, data augmentation, projection head, and global-local contrastive loss. From the results of ST-Graph CRL SV and ST-Graph CRL MV in Table~\ref{state-of-the-art} and Table~\ref{crossdataset}, the performance of the multi-view skeleton contrastive mechanism is shown to be impressive in all cases. To further assess the other factors, we conduct several ablation experiments on NTU with a linear evaluation protocol.
\begin{table}
\caption{Analysis of global-local contrastive learning loss function.}
\centering
\begin{tabular}{lrr}
\toprule
Loss Function & NTU (CS) & NTU (CV) \\
\midrule
$\mathcal{L}_{\mathrm{global}}$ & 69.69 & 73.70 \\
$\mathcal{L}_{\mathrm{temlocal}}$ & 67.36 & 74.88 \\
$\mathcal{L}_{\mathrm{spalocal}}$ & 61.04 & 65.89 \\
$\mathcal{L}_{\mathrm{global}} + \mathcal{L}_{\mathrm{temlocal}}$ & \textbf{74.71} & \textbf{82.62} \\
$\mathcal{L}_{\mathrm{global}} + \mathcal{L}_{\mathrm{spalocal}}$ & 73.62 & 79.04 \\
$\mathcal{L}_{\mathrm{global}}+ \mathcal{L}_{\mathrm{temlocal}} + \mathcal{L}_{\mathrm{spalocal}}$ & 74.21 & 81.54 \\
\bottomrule
\end{tabular}
\label{globallocalloss}
\vspace{0pt}
\end{table}
\begin{table}
\caption{Performance by using different methods to combine the global and local contrastive learning losses.}
\centering
\begin{tabular}{llrr}
\toprule
Viewpoint & Loss combination & NTU (CS) &NTU (CV)\\
\midrule
Single-view & Linear & 54.12 & 56.23 \\
& Task uncertainty & 60.24 & 59.79 \\
Multi-view & Linear & 71.06 & 78.68 \\
& Task uncertainty & \textbf{74.71} & \textbf{82.62} \\
\bottomrule
\end{tabular}
\label{architecture}
\vspace{0pt}
\end{table}
\subsubsection{The effect of global-local contrastive loss.} In this experiment, we evaluate different forms of the contrastive loss function. Experimental results are summarized in Table~\ref{globallocalloss}. Based on the results, we make the following observations. As the accuracy of $\mathcal{L}_{\mathrm{temlocal}}$ is higher than $\mathcal{L}_{\mathrm{spalocal}}$ by 6.32\%(CS) and 8.99\%(CV), the temporal splitting method is superior to the spatial splitting in this experiment for local contrastive loss. The impacts of the global and the local losses are different but complementary. Compared with using only one of them, the combined global-local loss function $\mathcal{L}_{\mathrm{global}} + \mathcal{L}_{\mathrm{temlocal}}$ leads to substantially better performance in two benchmarks, i.e., 74.71\%(CS) and 82.62\%(CV).
In Table~\ref{globallocalloss}, it can be found that $\mathcal{L}_{\mathrm{spalocal}}$ only produces poor performance. We think the reason might be related to the multi-view action
datasets. Specifically, as most multi-view action datasets do not provide the corresponding relation between persons in different views, the spatial partitioning strategy is likely to lead to a phenomenon that the positive pairs of local parts are from different persons when an action is performed by two people. In this case, the effect of $\mathcal{L}_{\mathrm{spalocal}}$ is inconsistent with our expectation and one of its impacts, learning a fine-grained view of irrelevant representation in the spatial dimension, will fail, while others still work. However, when combined with $\mathcal{L}_{\mathrm{global}}$ and $\mathcal{L}_{\mathrm{temlocal}}$, the main role of $\mathcal{L}_{\mathrm{spalocal}}$ is reflected in learning a fine-grained view irrelevant representation in the spatial dimension. This is why its accuracy goes down. Therefore, the default form of global-local contrastive loss consists of $\mathcal{L}_{\mathrm{global}}$ and $\mathcal{L}_{\mathrm{temlocal}}$, combined with task uncertainty in this paper. We also compare linear and task uncertainty based methods to combine the global and local contrastive losses. Note that all the weight parameters are uniformly set to 1 in the linear combination method. Results are shown in Table~\ref{architecture}, from which we can see that the task uncertainty based combination method outperforms the linear combination methods in both single-view and multi-view scenarios.
\subsubsection{Analysis of the projection head.} We study the importance of including a projection head, i.e. $f_\psi(\cdot)$ and $f_\phi(\cdot)$. Table~\ref{projectionhead} shows the linear evaluation results using two different architectures for the head: identity mapping and the nonlinear projection with one additional hidden layer. We can observe that a nonlinear projection head, regardless of its output representation dimension, performs better than identity mapping in terms of recognition accuracy. Therefore, it can be concluded that the hidden layer before the projection head is a better representation than the layer after.
\begin{table*}
\caption{Linear evaluation of representations with different projection heads and various dimensions of output. The representation, before projection, is 256-dimensional here.}
\centering
\begin{tabular}{lcccccc}
\toprule
Projection head & Identity Mapping & \multicolumn{5}{c}{Nonlinear projection}\\
\midrule
Output dimension & 256 & 32 & 64 & 128 & 256 & 512\\
\midrule
Accuracy & 66.21 & 74.29 & 74.31 & 74.39 & 74.38 & \textbf{74.71} \\
\bottomrule
\end{tabular}
\label{projectionhead}
\vspace{0pt}
\end{table*}
\subsubsection{The effects of data augmentation.}
Apart from the temporal subgraph, we also explore other four popular skeleton data augmentations in experiments including node dropping, node perturbation, view rotation, and skeleton shearing, with definitions as follows:
Node dropping. It randomly discards body joints in the input skeleton sequence $\mathbf{x}_i^v$. Specifically, with a $50\%$ chance, we randomly drop $10\%$ of nodes, where the corresponding joint coordinates are set to zero. It is a common phenomenon that a subset of joints, e.g., those occluded ones, cannot be detected. The augmentation of node dropping enables the crucial action patterns can still be learned from a subset of joints.
Node perturbation. The coordinates of joints are perturbed using a normal Gaussian distribution. The mean of the distribution is set to 0 while the standard deviation is set to 0.05. The detected joint locations, even for those without occlusion, always contain errors due to sensor and estimation accuracies in practice. The augmentation of node perturbation enables the action recognition to be robust to such errors.
View rotation. It randomly rotates the joint coordinates in a skeleton sequence along three axes in terms of a rotation matrix. Specifically, we randomly select three degrees $\alpha, \beta, \gamma$, all uniformly in the range of $[-17^\circ, 17^\circ]$ for each sequence. Three basic rotation matrices with rotation angles about X, Y, Z axis are given as follows:
\begin{equation}
\begin{aligned}
\resizebox{.5\linewidth}{!}{$
\displaystyle
\begin{split}
& \mathbf{R}_{\mathrm{X}}(\alpha) = \begin{bmatrix}
1 & 0 & 0 \\
0 & \mathrm{cos}\alpha & \mathrm{sin}\alpha \\
0 & -\mathrm{sin}\alpha & \mathrm{cos}\alpha
\end{bmatrix},\\&
\mathbf{R}_{\mathrm{Y}}(\beta) = \begin{bmatrix}
\mathrm{cos}\beta & 0 & -\mathrm{sin}\beta \\
0 &1 &0 \\
\mathrm{sin}\beta & 0 & \mathrm{cos}\beta
\end{bmatrix},\\&
\mathbf{R}_{\mathrm{Z}}(\gamma) = \begin{bmatrix}
\mathrm{cos}\gamma & \mathrm{sin}\gamma & 0 \\
-\mathrm{sin}\gamma & \mathrm{cos}\gamma & 0 \\
0 & 0 & 1
\end{bmatrix}.
\end{split}
$}
\end{aligned}
\end{equation}
Based on these three basic rotation matrices, the final rotation matrix is
\begin{equation}
\mathbf{R} = \mathbf{R}_{\mathrm{X}}(\alpha)\mathbf{R}_{\mathrm{Y}}(\beta)\mathbf{R}_{\mathrm{Z}}(\gamma).
\end{equation}
We apply the rotation matrix $\mathbf{R}$ to the original coordinates of the skeleton sequence and get the transformed coordinates. It simulates the view changes of the camera. This augmentation enables the action recognition to be robust to the camera view changes.
Skeleton shearing. It slants the shape of a skeleton at a random angle. The shearing factors are drawn from a uniform distribution in $[0.01, 0.1]$. The transformation matrix can be written as
\begin{equation}
\mathbf{S} = \begin{bmatrix}
1 & s_\mathrm{X}^\mathrm{Y} & s_\mathrm{X}^\mathrm{Z} \\
s_\mathrm{Y}^\mathrm{X} & 1 & s_\mathrm{Y}^\mathrm{Z} \\
s_\mathrm{Z}^\mathrm{X} & s_\mathrm{Z}^\mathrm{Y} & 1
\end{bmatrix},
\end{equation}
where $s_\mathrm{X}^\mathrm{Y}$, $s_\mathrm{X}^\mathrm{Z}$, $s_\mathrm{Y}^\mathrm{X}$, $s_\mathrm{Y}^\mathrm{Z}$, $s_\mathrm{Z}^\mathrm{X}$, $s_\mathrm{Z}^\mathrm{Y}$ are shearing factors.
All joint coordinates of the original skeleton sequence are transformed with the shearing matrix $\mathbf{S}$. The augmentation of skeleton shearing further increases the robustness of action recognition to more nonrigid transformations of the skeleton sequence.
\begin{table}
\caption{Performance of multi-view ST-Graph~CRL using different augmentation strategies.}
\centering
\begin{tabular}{lrr}
\toprule
Augmentation & NTU (CS) & NTU (CV) \\
\midrule
Original & 69.85 & 78.35 \\
Node Dropping & 69.05 & 77.37 \\
Node perturbation & 66.26 & 74.99 \\
View rotation & 70.37 & 77.99 \\
Shear & 69.56 & 77.55 \\
Temporal subgraph & \textbf{74.71} & \textbf{82.62} \\
\bottomrule
\end{tabular}
\label{augmentation}
\vspace{0pt}
\end{table}
We denote the model without any data augmentation as the original. The results are shown in Table~\ref{augmentation}. Much to our surprise, compared with directly using the original sequence, only the temporal subgraph strategy to ST-Graph CRL can significantly improve the accuracy by 5.26\%(CS) and 4.27\%(CV). Two possible reasons are: 1) defining precise frame-level starting and ending time for action is almost impossible, and 2) it is hard to achieve a strict temporal alignment of skeleton sequences captured by multiple cameras. Applying inconsistent temporal subgraphs for different views can improve the robustness of these unavoidable problems without breaking their original relationships. The counterproductive of other data augmentations maybe because the original correspondences in spatial structure among skeletons, which are simultaneously taken from different views, are destroyed after these random transformations. For example, compared with other augmentations, node perturbation, which changes the values of joints with a normal Gaussian distribution, is most damaging to spatial structure correspondences and leads to the sharpest performance drop of 3.59\%(CS) and 3.36\%(CV). Thus, the temporal subgraph is the default data augmentation we adopt in this work.
\subsection{Visualization of Skeleton Representation.}
The superior performance of ST-Graph CRL over the existing methods is largely due to the use of the multi-view skeleton contrastive mechanism. Hence apart from the quantitative evaluation, we also visualize the feature changes by using this mechanism. We randomly select ten classes in the NTU testing set and visualize the TSNE-embeddings of the features obtained from P\&C FW~\cite{su2020predict}, ST-Graph CRL SV, and ST-Graph CRL MV for the same skeleton sequences in Figure~\ref{tsne}. Here we observe that even in this 2D embedding it is evident that the features for different classes are better separated by using ST-Graph CRLs than using P\&C FW. Points of different colors are mixed up in (a) while they are more separated in (b) and (c). Meanwhile, points of the same color in (c) are more concentrated than those in (b). For example, there is a clear line among points with different colors at the bottom right of (c) while they are mixed up at the bottom left of (b). This supports the conclusion that CRL between multi-view skeletons makes the learned representation more discriminative.
\begin{figure*}
\centering
\subfigure[P\&C FW~\cite{su2020predict}]{\includegraphics[trim={80 80 30 80},clip,scale=0.3]{figs/PWCS.png}}
\subfigure[ST-Graph CRL SV]{\includegraphics[trim={80 80 30 80},clip,scale=0.3]{figs/singleview.png}}
\subfigure[ST-Graph CRL MV]{\includegraphics[trim={30 80 80 80},clip,scale=0.3]{figs/multiview.png}}
\vspace{-10pt}
\caption{TSNE-embedding visualizations of the learned representations from 10 classes randomly selected in NTU(CS) testing set.}
\label{tsne}
\vspace{0pt}
\end{figure*}
\section{Conclusion}
In this paper, we studied the problem of view-invariant skeleton-based action recognition by learning effective representations without any manual action labeling. Based on ST-GCN structural encoder, a multi-view spatial-temporal graph contrastive representation learning approach was developed to maximize the mutual information between the representations extracted from multiple skeleton data simultaneously taken from different views. Specifically, we explored five popular skeleton data augmentation methods and found only temporal subgraph can make a positive role in multi-view CRL. Then, to support our global-local CRL, partitioning functions were designed to segment ST-Graph into multiple subgraphs along spatial or temporal dimensions and projection heads were added to map the learned representations to another latent space. Besides, we proposed a local-global spatial-temporal graph contrastive loss, combined with task uncertainty, to model the multi-scale co-occurrence relationship between spatial and temporal domains. Experiments on two multi-view action datasets showed that our proposed approach, no matter in single-view or multi-view scenarios, got competitive performance compared with the random baseline and other state-of-the-art unsupervised skeleton-based action recognition methods.
In the future, we will explore new approaches to effectively handle multi-view multi-person scenarios.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\small
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The ability to purposefully verify your identity in the digital realm is the foundation for most, if not all, online activities. This online identity is something that we protect at varying levels depending on the negative impact of a compromise on our lives. With a larger portion of our lives being transacted online, people are getting more savvy in the way they manage their digital identities. It used to be common practice to use the same `simple' password for all your accounts. Now that is seen as a recipe for disaster, knowing that if one online account is leaked, then all your accounts are potentially exposed \cite{Beck2016}. Organizations are helping users by moving them toward safer online identity practices requiring multi-factor authentication and stronger passwords. These steps have helped improve online identity security immensely, but the architecture of the standard online identity management system is inherently insecure. The development of blockchain and decentralized identities has enabled the creation of the Self-Sovereign Identity (SSI) framework. The SSI framework has huge possibilities of implementations across many domains where there is a drive toward individual privacy \cite{Tykn2022}.
The development of these frameworks aims to improve the overall security of a user's Personal Identifiable Information (PII). Examples of PII include a user's name, address, passport number, driver’s license number, taxpayer identification number, patient identification number, financial records, etc. Personal Health Information (PHI) is one of the most sought after PII by hackers for the primary reason that identities used in healthcare are tied to Social Security Numbers (SSN) which never change for an individual \cite{Accoutnable2022}. Other PHI, such as home address, and phone number are used by organizations to determine identity can change.
Impersonation using stolen PHI puts the individual in a scenario which can lead to severe financial loss. SSI usage in the healthcare industry could help secure user data and limit the proliferation of patient PHI. Healthcare providers that utilize an SSI model would only have access to patient PHI with the approval/permission from the patient.
Similar to Public Key Infrastructure (PKI), SSI uses asymmetric keys as the encryption architecture as well as digital signatures to ensure secure communication between peers \cite{Sorokin2021}.
PKI utilizes a group of \textit{central} authorities to store the association database of identities, certificates, and public encryption keys. Conversely, SSI employs Decentralized Ledger Technology (DLT) to associate Decentralized Identifiers (DID) and public encryption keys. In an SSI implementation, DID and the public portion of their symmetric key pair are the only publicly accessible information associated with an entity. The entity retains their identity information details which they share as they see fit.
This paper describes the SSI architecture and its major components, along with a comparison between the mechanics used in PKI and SSI. We also examine how SSI can be utilized in different domains like healthcare, finance, retail, and government, where identity authentication and data security are expected and in some cases required by regulation.
The architecture presented in this paper is an identity framework utilizing blockchain, Decentralized Identity (DID), as well as other decentralized and cryptographic technologies. The framework is focused on the idea of \textit{self-sovereignty} of one's identity. Self-sovereignty can be defined as having control over the specifics of your own identity. This includes the user's ability to be able to share specific identity attributes and acquired credentials depending on the intended recipient of the information. The user will present a profile to the recipient, where the profile includes only those identity details required by the recipient for a particular operation.
The main goals of the SSI architecture is to fulfilling the CIA triad; Confidentiality, Integrity, and Availability. Other frameworks fulfill these goals through different architectural implementations relying on different structures and technologies to meet these goals. The presented framework prioritizes information security and user control of access.
The overall impact and proliferation of SSI through different domains has not been fully realized. Most domains have some level of need for associating identity with a collection of data or transactions. SSI has the ability to deliver personal information protection and control. Many systems built on the SSI framework are being developed to meet these different needs with the promise of higher level of security and greater control afforded to the user.
The rest of the paper is organized as follows - Section \ref{relwork} presents some background and related works. We describe the SSI architecture and its components in Section \ref{arch}. SSI is compared with PKI in Section \ref{comp}. Section \ref{use} covers a number of development efforts utilizing SSI that are taking place across multiple domains. We discuss the impact of SSI on the identity industry in Section \ref{discuss}. We conclude in Section \ref{conc}.
\section{Related Work}\label{relwork}
In this section, we present some related work and background on Identity, Distributed Ledger Technologies and Blockchains, Authentication, and Trust.
\subsection{Identity before Blockchains}
Public Key Infrastructure (PKI) has been developed over the last 35 years, built utilizing centuries of research in cryptography \cite{stapleton2012concise}. Many people consider PKI to be the defacto solution for managing entity identities on the internet. PKI is used to encrypt communication as well as establishing identity of users and systems.
The history of the concept of Self-Sovereign Identity (SSI) can be traced back to 1991 and the Pretty Good Privacy (PGP) project \cite{OpenPGP2016}. The idea of a Web-of-Trust (WoT) was created as a PGP project \cite{Stahl2018}. In WoT there are a few entities/other users trusted by a particular user.
As this user continues to add trusted entities to their WoT, other users/entities trust them as well. These other users have a collection of trusted entities/users that is different from the first user. In this model there is a relationship established (e.g. primary, secondary, tertiary, ...) between all users of the PGP project and results in the creation of a full WoT. The development of WoT was a monumental step in decentralized authentication.
In 1996, Carl Ellison presented the idea of identity without certification authorities \cite{Ellison1996}. Ellison presented the idea of binding the certificate of the issuer in a manner that is recognized by the verifier. The idea of not having a central authority to be the issuer and verifier was an innovative concept that was years ahead of its time. The decentralized aspect of the authorization is clearly seen in the current SSI architecture.
Microsoft proposed a federated ID model when they created MS Passport \cite{MSPowerU2016}. The concept was for a single login to allow access to multiple systems. A good idea to improve user experience and ease access to multiple systems/sites. However, the design relied upon Microsoft as the central authority to manage authentication and authorization. This model was no better, and in some ways worse than a traditional central authority.
\subsection{Decentralized Identity, Distributed Ledger Technologies, and Blockchains}
The next big development in decentralized identity was Identity Commons \cite{allen2016} and the
creation of the Internet Identity Workshop (IIW) working group in 2005. The working group put forward the idea of user-centric identity. There is a clear connection between these ideas and the decentralized architecture in the SSI framework. The IIW has long supported OpenID which an individual could utilize as a psuedo-SSI. The main issue is that to truly mimic SSI using OpenID, a user would need advanced technical knowledge to implement it properly.
Nakamoto wrote a paper in 2008 describing the bitcoin blockchain \cite{Nakamoto2008}. The decentralized nature of blockchain along with the addition of technological advancements has allowed for the development of SSI. Leveraging the creation of blockchain along with the WoT and IIW, the Decentralized Identity Foundation (DIF) was formed in 2017 \cite{DIF2021}. The DIF is a consortium of technology innovators in SSI. The member of DIF include, but not limited to Microsoft, Hyperledger (Linux Foundation), Accenture, and Sovrin. Hyperledger and Sovrin are both non-profit organizations that rely on contributing member to work collaboratively to define and development SSI frameworks and best practices. Sovrin utilizes the Hyperledger Indy codebase for it system \cite{Tobin2017}.
The Hyperledger Indy framework is being used in a number of development/research efforts. The open source aspect of these framework makes them a great platform for researchers to investigate, propose, and develop new innovations in the SSI framework. Being part of Hyperledger or Sovrin enables organizations to help drive development as well as provide technical guidance with regard to the direction that the projects should follow \cite{Tobin2017}.
There are businesses that are developing what could be considered SSI, but because of the centralized nature of the company which controls the authorization of the users, they cannot truly be considered SSI. One such implementation is 1Kosmos which utilizes a digital wallet on your smart phone to manage PII and credentials \cite{1kosmos2022},
\begin{figure*}[ht]
\centering
\includegraphics[scale=.95]{Arch.png}
\caption{\textbf{SSI Architecture Flow} Here, Agent B is verifying Agent A. Agent A, acting on behalf of an entity, and utilizing the data in Wallet A, (1) key B\_public is retrieved by Agent A and (2) used to encrypt and send a message, a verifiable credential (VC-A1) to Agent B. The message is (3) received by Agent B and decrypted by its private key B\_secret. Entity B then (4) verifies the credential VC-A1 against the Verifiable Data Registry.}
\label{fig:Arch.png}
\end{figure*}
\subsection{Authentication}
The concept of authentication is closely related to identity. An entity can claim an identity \cite{BlueCheck2021}; however, until that claim is authenticated, it is just a claim and therefore access to protect objects (systems, data) should not be granted. Passwords have traditionally been used in digital systems to authenticate a user's identity. Passwords are characterized as knowledge based identity authentication. PINs, passwords, and personal knowledge answers all fit into the knowledge based category. Two other type of authentication categories exist; `What you have?', and `Who you are?'. `What you have?' relates to a physical object of which you have possession. This can be a mobile phone, memory/smart card, or another certificate bearing object. `Who you are?' relates to biometric information. This information can be fingerprint, iris scan, or facial scan. Utilizing more than one of these categories of authentication is known as Multi-Factor Authentication (MFA). With the advent software that is able `crack' passwords, MFA has become a defacto standard by which modern systems are secured, especially healthcare and financial systems. The ability prove identity to access secure systems is critical to maintaining confidentiality as part of an overall security program.
\subsection{Trust}
Within an SSI architecture, there is a level of trust that is required. The trust within the peer-to-peer relationship that is used for communication and identity verification is developed by utilizing other peer relationships to confirm the veracity of an entity's claim \cite{Stahl2018}. The Web of Trust (WoT) project presents a model of trust based on reputation. Utilizing nodes within the WoT allows for trust to be extended to parties that may not have a direct relationship with an entity. The challenge in the WoT are nodes that try to manipulate the trust network to allow more trust to be offered to an entity that should not be trusted. In a peer-to-peer relationship, extending trust must be guarded.
\section{Self-Sovereign Identity (SSI) Architecture}\label{arch}
The foundation of SSI is the decentralized identifiers architecture presented by W3.org \cite{w3c2022}. Core components include Decentralized Identifiers (DID) v1.0, architecture, data model, and representations. The task is to present credentials to another entity for verification/authentication. Figure \ref{fig:Arch.png} shows a simple communication between two DIDs sending a verifiable credential. {Here, Agent B is the verifier and Agent A needs to be verified.} Agent A, acting on behalf of entity A, and utilizing the data in Wallet A, (1) public key B\_public is retrieved by A and (2) used to encrypt a message, a verifiable credential (VC-A1) to B. The message is (3) received by B and decrypted by its private key B\_secret. Entity B then (4) verifies the credential VC-A1 against the Verifiable Data Registry.
The SSI architecture is composed of seven key technologies defined by the W3C. The seven technologies are:
\begin{itemize}
\item Decentralized Identifiers
\item Verifiable Credentials
\item Decentralized Public Key Infrastructure
\item Blockchain and Distributed Ledger Technology
\item Verifiable Data Registry
\item Agents
\item Digital Wallets
\end{itemize}
Next week describe each of these in detail.
\subsection{Decentralized Identifiers (DID)}
Decentralized Identifiers (DID) are globally unique identifiers structured similar to a universally unique identifier (UUID) but with modifications to enable cryptographic identity/security \cite{w3c2022}. The first difference is there is no centralized authority managing the identities. DID's use a decentralized mechanism like a Decentralized Ledger Technology (DLT) \cite{Frankenfield2021}. The other difference is that the DID has cryptographic properties. Key pairs are generated as part of the DID address and can be used to prove ownership/identity in the same fashion as digital signatures. The DID Auth protocol is a challenge-response process and therefore has the ability to replace the username/password structure of a centralized authentication system \cite{Gataca2021}. The immutability of the blockchain and the use of cryptogrpahic proofs to affirm ownership makes the DID platform a rich environment for the development of authoritative identity verification.
\begin{figure}[hb]
\centering
\includegraphics[scale=0.75]{VC-flow.png}
\caption{\textbf{Issue/Holder/Registry} Within the process of issuing verifiable credentials, the Issuer (1) issues credentials to the Holder which is the entity being authenticated using the credential. The Holder (3) submits the credentials and its use schema to the Verifiable Data Registry, where the Issuer (2) verifies the identities and use schema of the credentials. The Holder (4) also send the credentials in a presentation to the Verifier, where the Verifier (5) send the identities to the Verifiable Data Registry where the identity and schema are verified.}
\label{fig:VC-flow.png}
\end{figure}
\subsection{Verifiable Credentials}
Verifiable credentials (VC) are exactly what one would assume them to be, credentials that are able to be verified by the issuers of the credential. There are three main parts of the VC; the issuer, the holder, and the registry (See Figure \ref{fig:VC-flow.png}). The issuer defines the credential scheme to the registry as well as issuing a specific credential to the holder. The holder will register the credential with the registry. To complete the transaction of verification, a third party verifier will receive a credential presentation from the holder, and then confirm authenticity of the credential with the registry \cite{Bolgouras2022}. The credential presentation is a `view' of the credential that is defined as specific attributes of the credential that the verifier requires to be seen to confirm authenticity of the credential. The benefit of this limited view is that the presentation does not need to disclose all the information in the credential, but only what is agreed upon by the holder and verifier as to what information is `enough' to confirm the veracity of the credential.
\subsection{Decentralized Public Key Infrastructure (DPKI)}
Decentralized Public Key Infrastructure (DPKI) is a part of the overall decentralized model of identity verification \cite{Singh2017}. While the DPKI is not directly related to the individual identity verification, it is a significant security system in which the SSI systems rely upon to provide system/URI identity verification. In the event that the PKI is compromised, then the SSI environment is able to be breached, and therefore bring into question the veracity of the whole system. DPKI resolves multiple vulnerabilities of current Public Key Infrastructure Certificate Authority (PKI/CA) systems. DPKI utilizes blockchain and smart contracts to provide system identity confirmation.
\subsection{Blockchain and Distributed Ledger Technology (DLT)}
What is commonly known as blockchain is a Distributed Ledger Technology (DLT). Blockchain gained notoriety with the popularization of Bitcoin. Bitcoin was an early implementation of cryptocurrency based on a 2008 whitepaper written by Satoshi Nakamoto \cite{Nakamoto2008}. Cryptocurrencies utilize the trust and integrity of the blockchain to manage/document the transactions between parties. Because of the distributed nature of blockchain, availability is also provided in the architecture. Confidentiality is not a concern for cryptocurrencies which rely on the ability to see all transactions. SSI utilizes the all three aspects of the security triad; Confidentiality, Integrity, and Availability.
Blockchain can be implemented in four configurations, where there are two settings for two parameters. The four configurations are as follows:
\begin{itemize}
\item Public - Permissionless
\item Public - Permissioned
\item Private - Permissionless
\item Private - Permissioned
\end{itemize}
While cryptocurrencies are usually Public - Permissionless where they can be seen and added to by anyone, SSI implementations are typically Public - Permissioned which can be seen by anyone, but one needs special permission to add to the blockchain. \cite{Tobin2017} The Private - Permissioned is able to support the integrity desired in an SSI. Each block in the blockchain contains hashed transactions creating an immutable block as part of the distributed ledger.
\subsection{Verifiable Data Registry}
A Verifiable Data Registry manages the life-cycle and verification of credentials. The database system that acts as a verifiable data registry can take multiple forms driver's licenses, passports, and other government ID databases are examples of verifiable data registries \cite{Dunphy2018}.
\subsection{Agents}
Agents are software applications that run either in the cloud or on an `edge' device such as a mobile device or browser \cite{Reed2016}. The agents acts on behalf of an entity to access digital wallets as well as performing other cryptographic operations. The benefit of the agents is the increased privacy provided by using the agent.
\subsection{Digital Wallets}
Digital Wallets can be software or hardware implementation of a data store. Digital wallets work in concert with agents to perform multiple functions related to the management of cryptographic information such as digital signatures, key pairs life-cycle management. The digital wallet can take many forms including being supported as a Software as a Service (SaaS) function \cite{Soltani2021}. SSI Wallets can contain multiple verified credentials as well a other data structure. Figure \ref{fig:Wallet.png} shows the primary data structures utilized in an SSI Wallet. Items stored include DID, crypto-keys, multiple verifiable credentials, and other data. The other data mentioned can be Non-Fungible Tokens (NFTs), specific cryptocurrencies, or other personal data.
\begin{figure}[hb]
\centering
\includegraphics[width=8cm]{Wallet.png}
\caption{\textbf{Digital Wallet} Information stored within an SSI Wallet. Items stored include DID, crypto-keys, multiple verifiable credentials, and other data.}
\label{fig:Wallet.png}
\end{figure}
\section{Comparison with Public Key Infrastructure (PKI)}\label{comp}
\begin{figure}[ht]
\centering
\includegraphics[width=8cm]{PKI.png}
\caption{\textbf{PKI Architecture} (1) Entity submits CSR (Certificate Signing Request) to RA (Registration Authority). (2) RA sends ID approval to CA (Certificate Authority). (3) CA sends certificates to VA (Verification Authority). (4) CA sends keys back to the requesting entity. (5) Keys/Signature are sent to or installed on a server or website. (6) Keys are sent to the VA for verification. (7) Verification/approval is returned confirming the authenticity of the keys/signature.}
\label{fig:PKI.png}
\end{figure}
Public Key Infrastructure (PKI) can be characterized as a central store of identity registration where cryptographic keys are associated with a registered identity. Identities can be users or devices. The inclusion of a PKI certificate is used to confirm the authorized identity associated with the certificate. There is a hierarchy of trust that is established within the PKI services that allow for the life-cycle management of the certificates as well as the authenticity verification of the certificates.
PKI utilizes a challenge-response process to verify the identity authenticity of the certificate holder (See Figure \ref{fig:PKI.png}). This is similar to the DID authentication process which is also a challenge-response loop.
A key \textit{difference} between SSI and PKI is the decentralization of SSI. The decentralization of SSI allows for exclusive peer-to-peer coordination of transactions \cite{Singh2017}. These transactions can include identity verification as well as exchange of information or digital asset. For PKI, there is an exclusive function performed that relates to cryptography. The certificate and key pair that are issued as a result of registration with a PKI Certificate Authority (CA) can be used for encryption as well as identity verification function as in digital signatures. SSI has the ability within its architecture to support any type of digital asset as well as tokens such as NFTs. The construct of a smart contract within the SSI architecture allows for enumerable implementations supporting peer relationship agreements within the blockchain. PKI certificates/cryptogrpahy are used almost exclusively for encryption and identity verification.
Another key difference between SSI and PKI is the hierarchical nature of PKI as opposed to the flat peer-to-peer nature of SSI. Within the PKI, there is a structure of hierarchy that has the root Certificate Authority (CA) at the top of the structure. {Root CAs maintain the database of certificates and share it with subordinate CAs, which are the systems that respond to queries}. The Registration Authority (RA) receives requests for new certificates as well as performing a number of other administrative tasks. They also offload the life-cycle management of certificates from the CAs.
The central database of a PKI makes the CA's a significant target for attackers who wish to leverage vulnerabilities in the security of the PKI for their advantage. Compromising a CA would result in an attackers ability to direct users to malicious websites masquerading as actual websites. The decentralized nature of SSI as well as the use of blockchain as the data store creates a immutable data structure that distributed across the internet. To compromise this architecture, an attacker would have to be able to compromise multiple systems simultaneously which high unlikely if not completely impossible.
\section{Usecases}\label{use}
While SSI is still considered in it's infancy, there are a number of development efforts that are taking place across multiple domains. The immutability and security of the SSI architecture make it a compelling framework for many functions where users wish to move to a fully digital environment. In this section, we describe some usecases possible in healthcare, finance, retail, and government.
\subsection{Healthcare}
The creation and adoption of the Health Insurance Portability and Accountability Act (HIPAA) \cite{HIPAA2003} and Health Information Technology for Economic and Clinical Health (HITECH) Act \cite{HITECH2017} instituted guidelines on the protection and portability of Personal Healthcare Information (PHI). These laws outline how healthcare providers must manage patient information. Business associates are companies and individuals who utilize, manage, or engage in business with healthcare organizations and access, store, process, or control patient information. Utilizing SSI to store PHI would enable patients to have their medical history immediately available when seeing a new doctor or visiting an emergency room. There is also the aspect of increased security afforded when using SSI as opposed to a central database. SSI also allows a patient to control which personal and health information is shared with a healthcare provider or payer.
The authentication of patient identity/credential would follow the steps in figure \ref{fig:Healthcare.png}. (1) Patient requests a credential from an authority able to issue the required credential. (2) Verifier send credential and its own signature to a DLT to be used as a point of verification for the issued credentials. (3) Verifier issues credential to patient where the patient stores them in their digial wallet. (4) Patient sends request to Provider, with Patient credential issued from Verifier, requesting access to Patient data. (5) Provider sends credentials to DLT for verification where DLT responds with approval. (6) Provider replies with access granted.
\begin{figure}[htp]
\centering
\includegraphics[scale=0.7]{Healthcare.png}
\caption{\textbf{Healthcare SSI Architecture} (1) Patient requests a credential (2) Verifier send credential and signature to DLT (3) Verifier issues credential to patient (4) Patient sends request to Provider with Patient credentials, requesting access (5) Provider sends credentials to DLT for verification where DLT responds with approval (6) Provider replies with access granted.}
\label{fig:Healthcare.png}
\end{figure}
\subsection{Finance}
The ability for banks and other financial institutions to leverage SSI could bring about significant improvements to the customer experience \cite{Keil2022}. Utilizing SSI, banks would be able to streamline the audit process as well as the ability to onboard new customers. Couple this improved user experience with the improved security and lower back-office costs, the motivation to incorporate SSI into the banking industry has tangible and quantifiable benefits.
\subsection{Retail}
There has been significant development in the area of digital currency that impact the retail industry. The idea of making expenditures as an individual without the insertion of a middle-man to manage the transfer of funds has been widely accepted. However, SSI goes beyond just the ability to make a peer-to-peer transaction that is safe, secure, and private \cite{Tobin2017}. SSI being incorporated into the retail/e-commerce ecosystem could have a significant impact on several aspects of these financial transaction. A few benefits are the increased security of the financial transaction, the reduction in fraudulent purchases, and the lack of need to store customer payment information. This last benefit would reduce the number of lost/stolen customer payment information, therefore improving the purchasing experience for the customer as well as the retailer.
\subsection{Government}
In our modern societies we rely on the governments to provide individuals with credentials that affirm our claim to our personal identity. In the USA, prospective employees must submit government approved documentation to demonstrate one's identity as well as nationality. A few of the documents that are acceptable to be used for identity proofing are a state issued driver's license, social security card, and a passport. These forms of credentials are difficult, but not impossible to forge. The use of SSI to maintain a digital wallet with verified credentials from a government agency would provide a significant improvement to the identity proofing process as well as securing specific PII that would not need to be held in the government database. India has a national digital identification system which issues a residents an Aadhaar number \cite{India2019}. To receive an Aadhaar number, a person must provide the following demographic and biometric information:
\begin{itemize}
\item Name
\item Date of Birth (verified) or Age (declared)
\item Gender
\item Address
\item Mobile Number (optional) and Email ID (optional)
\item Ten Fingerprints
\item Two Iris Scans
\item Facial Photograph
\end{itemize}
With this information being collected and technology being developed to help secure personal information, there have been a number of proposals to coalesce multiple government identities housed on one card. A proposal in 2019 suggested that this one card could hold voter registration, bank card, passport, and Aadhaar card information \cite{Misra2019}.
A significant challenge perceived in the use of SSI are the vulnerabilities that can be found in the system components. As some of these SSI components like digital wallets store all consolidated PII, an implementation software bug or an exploitable system vulnerability, would result in significant personal and financial concern to the user.
\section{Discussion}\label{discuss}
As is the case in the development of many technologies, the concepts for SSI have long been established before the foundational technologies were developed to efficiently and effectively implement the innovative ideas. SSI has been gaining market exposure over the last 5 years including consortiums being established 10 years ago but finally releasing alpha level solutions/products in the past 5 years. Many corporations are looking at SSI to deliver passwordless authentication solutions. Security professionals and attackers have known for years that users are the biggest security vulnerability. This is why phishing emails remain one of the top attack vectors every year. One of the main exploitation methods within phishing emails is an embedded link to a fake login page. A user entering in their username/password is handing their credentials over to attacker that they can later use to login into company resources and conduct a malicious attack. If the company had setup passwordless/SSI authentication, then there would be no credentials to steal and therefore the phishing attack would fail to exploit the user.
While the SSI model presents the idea that an individual controls their credentials and attributes, that is not always 100\% true. When credentials are verified, there is an issuer that has the ability to revoke/expire these credentials. Therefore, there is a decentralized ledger and entity controlled wallet that manages the presentation of credentials, there are central authorities that are able to revoke credentials.
\section{Conclusions}\label{conc}
After almost 30 years since the creation of the concepts that form the foundation of SSI, the identity industry is investing significantly in development of frameworks to support its use across multiple domains. The development of blockchain and decentralized identity create a synergy in technology that is able to support the notion that an individual should have the ability to control how their personal information is shared with other people, organizations, and companies. The ability to verify an individual's claim of identity and their personal credentials without the need to reference a central authority creates a network of trust that can return the power of identity back into the hands of the individual.
In this paper, we have presented a SSI architecture which will support multiple domain requirements as well as the similarities and differences to the current PKI architecture being used across the internet today. We also discussed specific benefits and usecases possible in healthcare, finance, retail, and government. We elaborate on how these industries could revolutionized the customer/user experience in authentication.
\bibliographystyle{unsrt}
|
\section{Introduction}
Motivated by \cite{bogdan}, in this paper we identfy the dynamics generated by sample paths of a Gaussian process $f = \setof{f(x)}_{x\in X}$ conditioned to pass through a set of data points ${\mathcal T} = \setof{(x_n,y_n)}_{n=0}^N$ where $X=[x_0,x_N]$ (we specify the mean and covariance of this process precisely in Section~\ref{Sec: brownian}).
More precisely, we provide a characterization of local and global dynamics on $X$ and an exact probability that this description is valid for a random sample path from $f$.
Although for the sake of simplicity we remain in a probabilistic setting, and avoid describing any sampling procedures, we are motivated by the problem of understanding a dynamical system that we are able to access only through finite samples of the same form as ${\mathcal T}$.
To understand the significance of our results (and those of \cite{bogdan}) let us first consider a typical approach to this problem.
Given the data set ${\mathcal T}$, a variety of well-known techniques are available to generate an surrogate function $\hat{f}$ \cite{gramacy}.
If $\hat{f}$ is accepted as the model, then the standard theoretical and computational techniques of nonlinear dynamics can be applied.
However, because dynamical systems are subject to bifurcations, it is not clear that the dynamics generated by $\hat{f}$ and the true dynamics that generated the data are the same, i.e., conjugate.
In fact, the behavior of nonlinear dynamics is sufficiently rich that distinguishing conjugacy classes requires uncountably many invariants \cite{conj_prob} and thus identifying the conjugacy class is impossible with finite data.
With this in mind we propose to characterize the dynamics via the concept of a Morse tiling \cite{lattice3}.
\begin{defn}
Let $g: Y \to Y$ be a continuous function on a compact metric space $Y$ and ${\mathsf Q}$ a finite partially ordered set with partial order $\preceq$.
A {\bf Morse tiling} of $Y$ for $g$ is a decomposition of $Y$ into a collection
\[
{\mathcal M} = \setof{M(q) \mid q \in {\mathsf Q}}
\]
of regular closed sets with disjoint interiors with the following property.
Let $y \in Y$.
If $g^n(y) \in M(p)$, $g^m(y) \in M(q)$, and $n<m$, then $q\preceq p$.
The sets $M(q)$ are {\bf Morse tiles} for $g$ and the partially ordered set ${\mathsf Q}$ is a {\bf Morse graph} for $g$.
\end{defn}
The partial order on ${\mathsf Q}$ implies that if the trajectory leaves a Morse tile, then it cannot return.
Therefore, recurrent dynamics of $g$ occurs within individual Morse tiles.
This provides a global decomposition of the dynamics where the gradient-like behavior is characterized by the Morse graph.
As is described in Section~\ref{Sec: conley} an algebraic topological invariant, the Conley index, can be assigned to each Morse tile.
The significance of this is that the Conley index can be used to deduce the existence of interesting invariant sets for $g$ such as fixed points, periodic orbits, heteroclinic orbits, bistability, and chaotic dynamics \cite{mischmroz}.
The main result of this paper (see Theorem~\ref{Thm:TileBnd} for the technical details) is that we can produce a Morse tiling (as well as the associated Conley indices) and provide an exact formula for the probability that this Morse tiling is valid for a random sample path from $f$.
There is an important caveat: our Gaussian process $f$ is a sequence of Brownian bridges with variance parameter $\sigma^2$ interpolating the set ${\mathcal T}$.
An outline of the paper is as follows.
In Section~\ref{Sec: brownian} we state all of our probabilistic assumptions along with a simple proposition that is necessary for our computations.
For the convenience of the reader we recall in Section~\ref{Sec: order} basic definitions and concepts from order theory.
In Section~\ref{Sec: conley} we recall basic ideas from dynamical systems and combinatorial Conley theory.
The main results of the paper are presented in Section~\ref{Sec: results}.
Section~\ref{Sec: examples} covers several examples designed to illustrate the techniques used in this paper.
Finally, we provide some concluding remarks and comment on future directions in Section~\ref{Sec: concl}.
\section{Brownian Paths and Excursion Bounds}\label{Sec: brownian}
In this section we state our basic probabilistic assumptions and a proposition that will be used extensively throughout the article. We begin by recalling the definitions of some simple Gaussian processes.
A Gaussian process is uniquely specified by its mean and covariance function \cite{adtay}. The Gaussian process $W=\setof{W(x)}_{x \geq 0}$ such that $\text{E}(W(x)) = 0$ and $\text{Cov}(W(x),W(x')) =\min(x,x')$ is called the {\bf standard Weiner process} or {\bf standard Brownian motion}. More generally, if $x_a,x_b,y_a,y_b \in {\mathbb{R}}$ and $x_a < x_b$ then the Gaussian process $Z = \setof{Z(x)}_{x \in [x_a,x_b]}$ with $\text{E}(Z(x)) = y_a$ and $\text{Cov}(Z(x),Z(x')) = \sigma^2 \min(x,x')$ for all $x,x' \in [x_a,x_b]$, where $\sigma >0$,
is the {\bf Weiner process (or Brownian motion) on the interval $[x_a,x_b]$, starting at $(x_a,y_a)$, with variance parameter $\sigma^2$}. A {\bf Brownian bridge $B = \setof{B(x)}_{x \in [x_a,x_b]}$ from $(x_a,y_a)$ to $(x_b,y_b)$ with variance parameter $\sigma^2$} has the law of $Z$ conditioned to take the value $y_b$ when $x = x_b$. Then $B$ is a Gaussian process satisfying $\text{E}(B(x)) = y_a + \frac{x - x_a}{x_b - x_a}(y_b - y_a)$ for all $x\in [x_a,x_b]$ and $\text{Cov}(B(x),B(x')) = \sigma^2 \frac{(x - x_a)(x_b - x')}{x_b - x_a}$ for $x_a \leq x \leq x' \leq x_b$.
With these definitions in mind we discuss the Gaussian process $f$ that we study in this paper.
We begin with a collection of $N+1$ points, ${\mathcal T} = \setof{(x_n,y_n)}_{n=0}^N$, where we assume that $x_n < x_m$ for $n < m$; recall that we have defined $X:= [x_0,x_N]$.
For simplicity of exposition we also assume that
$$ x_0 < y_n < x_N \text{ for } 0\leq n \leq N.$$
As indicated in the introduction, we will assume that $f$ is Brownian motion with variance parameter $\sigma^2$ conditioned on the events $\setof{f(x_n) = y_n}_{n=0}^N$. More specifically, $f = \setof{f(x)}_{x\in X}$ is the Gaussian process with mean $\mu(x):=\text{E}(f(x))$ and covariance $\kappa(x,x'):=\text{Cov}(f(x),f(x'))$ given by
\begin{itemize}
\item $\mu(x) = y_{n-1} + \frac{x - x_{n-1}}{x_n - x_{n-1}}(y_n - y_{n-1})$ for all $x \in [x_{n-1},x_n]$ and
\item $\kappa(x,x') =
\begin{cases}
\sigma^2 \frac{(x - x_{n-1})(x_n - x')}{x_n - x_{n-1}}, & x_{n-1} \leq x \leq x' \leq x_{n}\\
0, & \text{otherwise}
\end{cases}$.
\end{itemize}
Note that we can think of this process as a sequence of $N$ independent Brownian bridges $B_n$ from $(x_{n-1},y_{n-1})$ to $(x_n,y_n)$ with variance parameter $\sigma^2$.
Our ultimate goal is to characterize the dynamics of $f$ on $X$. In order to do so we will need to know the probability that $f$ stays between two particular threshold values on various combinations of intervals. That is, we are interested in computing the probabilities of events $$ S_n(\alpha,\beta) := \setof{f(x) \in (\alpha,\beta)\,|\, x \in [x_{n-1},x_n]}.$$
Because the process $f$ may be considered a sequence of independent Brownian bridges, this probability is equivalent to $\mathbb{P}( B_n(x) \in (\alpha, \beta) \,|\, x \in [x_{n-1},x_n])$. We can obtain this value exactly, which is a key motivation for our choice of Gaussian process.
In order to obtain $\mathbb{P}(S_n(\alpha,\beta))$ we will make use of the function $\pi:{\mathbb{R}}^4 \to {\mathbb{R}}$ defined by
\[
\pi(a,b,c,d) := \sum_{m=1}^\infty
\exp\left( -2[m^2ab + (m-1)^2cd + m(m-1)(ad + c b)]\right)
\]
\[
+ \exp\left( -2[(m-1)^2ab + m^2cd + m(m-1)(ad + c b)]\right)
\]
\[
- \exp\left( -2[m^2(ab+cd) + m(m-1)ad + m(m+1)c b]\right)
\]
\[
- \exp\left( -2[m^2(ab+cd) + m(m+1)ad + m(m-1)c b]\right).
\]
\begin{lem}[\cite{doob}, (4.3)]\label{Lem: Doob_exact}
Let $a,c \geq 0$ and $b,d > 0$. Then
\[
\mathbb{P}\left[\sup_{0\leq x<\infty}\{W(x) - (ax + b)\} \geq 0 \,\text{OR}\, \inf_{0\leq x \leq \infty}\{ W(x) + c x +d\} \leq 0 \right]= \pi(a,b,c,d).
\]
\end{lem}
Using Lemma \ref{Lem: Doob_exact} we are able to derive the following equation.
\begin{prop}\label{prop:BasicBound}
Assume that $\alpha < \min(y_{n-1},y_n) \leq \max(y_{n-1}, y_n) < \beta$. Then
$$\mathbb{P}\left( S_n(\alpha,\beta) \right)
= 1 - \pi\left(\frac{(\beta - y_n)}{\sigma \sqrt{x_n - x_{n-1}}}, \frac{(\beta - y_{n-1})}{\sigma \sqrt{x_n - x_{n-1}}} ,\frac{(y_n-\alpha)}{\sigma \sqrt{x_n - x_{n-1}}}, \frac{(y_{n-1} - \alpha)}{\sigma \sqrt{x_n - x_{n-1}}}\right).$$
\end{prop}
Notice that if either $\alpha$ or $\beta$ is in the interval $[\min(y_{n-1},y_n) , \max(y_{n-1}, y_n)]$ then the probability is trivially zero.
\begin{proof}
In order to simplify the notation we will prove the result for a Brownian bridge $B$ with variance $\sigma^2$ from $(0,p)$ to $(T,q)$; no generality is lost with this shift. It is straightforward to verify that $B$ may be represented as
$$B(x) = \frac{T-x}{\sqrt{T}}\sigma W\left(\frac{x}{T-x}\right) + p + x\left(\frac{q - p}{T}\right)$$
with $B(T) = q$. We will use the transformation $s = \frac{x}{T-x}$. Thus we see the following:
\begin{align*}
\sup_{x \in [0,T)} B(x) &= \sup_{x \in [0,T)} \frac{T-x}{\sqrt{T}}\sigma W(\frac{x}{T-x}) + p + x\left(\frac{q-p}{T}\right)\\
&= \sup_{s \in [0,\infty)} \left( \frac{T - \frac{sT}{1+s}}{\sqrt{T}}\right) \sigma W(s) + p + \frac{Ts}{1+s}\left(\frac{q-p}{T}\right)\\
&= \sup_{s \in [0,\infty)} \left( \frac{\sigma \sqrt{T}}{1+s}\right) W(s) + p + \frac{s(q-p)}{1+s}.
\end{align*}
From this characterization we see that the following events are equivalent.
\begin{align*}
\sup_{x \in [0,T)} B(x) \leq \beta &\iff \sup_{s \in [0,\infty)}\left[ \left( \frac{\sigma \sqrt{T}}{1+s}\right) W(s) + p + \frac{s(q-p)}{1+s} \right] \leq \beta \\
&\iff \sup_{s \in [0,\infty)}\left[ \left( \frac{\sigma \sqrt{T}}{1+s}\right) W(s) + (p - \beta) + \frac{s(q-p)}{1+s} \right] \leq 0\\
&\iff \sup_{s \in [0,\infty)}\left[ \left( \sigma \sqrt{T}\right) W(s) + (p - \beta)(1+s) + s(q-p) \right] \leq 0\\
&\iff \sup_{s \in [0,\infty)}\left[ \left( \sigma \sqrt{T}\right) W(s) -\left( s(\beta - q) + (\beta - p) \right)\right] \leq 0\\
&\iff \sup_{s \in [0,\infty)}\left[ W(s) -\left( s\frac{(\beta - q)}{\sigma \sqrt{T}} + \frac{(\beta - p)}{\sigma \sqrt{T}} \right)\right] \leq 0\\
\end{align*}
Similary, we get that
\[
\inf_{x \in [0,T)} B(x) \geq \alpha \iff \inf_{s \in [0,\infty)}\left[ W(s) + \left( s\frac{(q-\alpha)}{\sigma \sqrt{T}} + \frac{(p - \alpha)}{\sigma \sqrt{T}} \right) \right] \geq 0.
\]
The result then follows from Lemma \ref{Lem: Doob_exact}.
\end{proof}
\section{Order Theory}\label{Sec: order}
In this section we recall some basic definitions and notations from order theory that are used in the following section on computational Conley theory.
For a more complete introduction the reader is refered to \cite{davey:priestley}.
\begin{defn}
A {\bf lattice} ${\mathsf L}$ is a set with the commutative and associative binary operations $\land,\lor: {\mathsf L} \times {\mathsf L} \to {\mathsf L}$ which satisfy the following absorption axiom for all $a,b,c \in {\mathsf L}$:
\[
a \land (a \lor b) = a = a \lor (a \land b)
\]
A lattice is {\bf distributive} if for each $a,b,c \in {\mathsf L}$ it also satisfies the additional axiom
\[
a \lor (b \land c) = (a \lor b) \land (a \lor c).
\]
A lattice is {\bf bounded} if there exist {\bf neutral} elements ${\bf 0},{\bf 1} \in {\mathsf L}$ with the property that
\[
{\bf 0} \land a = {\bf 0}, \hspace{0.5cm} {\bf 0} \lor a = a, \hspace{0.5cm} {\bf 1} \land a = a, \hspace{0.5cm} {\bf 1} \lor a = {\bf 1}
\]
for all $a \in {\mathsf L}$.
\end{defn}
All lattices used in this paper are both bounded and distributive.
In the introduction we described our main results using the idea of a poset. We formally record the definition of this object here.
\begin{defn}
A {\bf partially ordered set} or {\bf poset} ${\mathsf Q} $ is a set with an order relation $\leq$ satisfying the three properties:
\begin{enumerate}
\item {\bf Reflexivity:} $a \leq a$
\item {\bf Transitivity:} $a\leq b$ and $b \leq c$ $\implies$ $ a\leq c$
\item {\bf Anti-symmetry:} $a \leq b$ and $b \leq a$ $\implies$ $a=b$
\end{enumerate}
\end{defn}
Given $q\in {\mathsf Q}$ the {\bf downset} of $q$ is defined as $\downarrow(q):= \setof{q' \in {\mathsf Q} \mid q' \leq q}$.
An element $c$ of a lattice ${\mathsf L}$ is {\bf join-irreducible} if $c \neq {\bf 0}$ and $c = a \lor b$ implies that $c = a$ or $c = b$ for all $a,b \in {\mathsf L}$.
We denote the set of join-irreducible elements of ${\mathsf L}$ by ${\mathsf J}({\mathsf L})$
Any lattice ${\mathsf L}$ has a naturally induced partial order relation $\leq$; for any $a,b \in {\mathsf L}$,
\[
a \leq b \iff a \land b = a.
\]
Since ${\mathsf J}({\mathsf L})\subset {\mathsf L}$ it inherits the partial order relation on ${\mathsf L}$.
Furthermore, an element $c\in{\mathsf L}$ is join-irreducible if and only if there exists a unique element $\overleftarrow{c} \in {\mathsf L}$ such that $\overleftarrow{c} < c$ and there is no $a \in {\mathsf L}$ with $a\neq \overleftarrow{c}$ such that $\overleftarrow{c} < a < c$; the element $\overleftarrow{c} \in {\mathsf L}$ is called the {\bf immediate predecessor} of $c \in {\mathsf J}({\mathsf L})$.
Before closing this section let us specify some notation. Given any lattice ${\mathsf L}$, the notation ${\mathsf L}' \hookrightarrow {\mathsf L}$ will indicate that ${\mathsf L}'$ is a sublattice of ${\mathsf L}$; that is, ${\mathsf L}'$ is a subset of ${\mathsf L}$, ${\mathsf L}'$ is itself a lattice and there is an inclusion morphism from ${\mathsf L}'$ to ${\mathsf L}$. We will further assume that any sublattice ${\mathsf L}' \hookrightarrow {\mathsf L}$ introduced in this paper contains the same neutral elements ${\bf 0}$ and ${\bf 1}$ that bounded the original lattice ${\mathsf L}$.
\section{Conley Theory}\label{Sec: conley}
Conley theory has two components: (i) a framework for global decompositions of dynamics, and (ii) algebraic topological tools for reconstructing dynamics.
For this paper the first is represented by the Morse tiling and the second by the Conley index.
For the sake of simplicity we present this theory in the setting of one-dimensional maps where the starting point is the set of data points ${\mathcal T}$ (see \cite{bogdan} for a more general dimension independent discussion).
As indicated in the introduction given the set of points ${\mathcal T} = \setof{(x_n,y_n)}_{n=0}^N$ we define the phase space of interest to be $X=[x_0,x_n]$.
We decompose $X$ as a simplicial complex ${\mathcal X} = {\mathcal X}({\mathcal T})$ with vertices ${\mathcal X}^{(0)} := \setof{x_n}_{n=0}^N$ and edges ${\mathcal X}^{(1)} := \setof{[x_{n-1},x_{n}]}_{n=1}^{N}$.
Viewing the face relation as a partial order $\downarrow([x_{n-1},x_{n}])= \setof{[x_{n-1},x_{n}], x_{n-1},x_n}$.
We use \emph{combinatorial multivalued maps} to model the dynamics.
In particular for the dynamics computations we use ${\mathcal F}^\text{top}\colon {\mathcal X}^{(1)} \rightrightarrows {\mathcal X}^{(1)}$ a set valued function ${\mathcal F}^\text{top}(\xi)\subset {\mathcal X}^{(1)}$.
To simplify the discussion concerning the Conley index we restrict our attention to \emph{interval valued} maps, i.e., for each interval $\xi\in {\mathcal X}^{(1)}$, ${\mathcal F}^\text{top}(\xi) = \cup_{i=1}^j [x_i,x_{i+1}]$, i.e., the image of ${\mathcal F}^\text{top}(\xi)$ is an interval.
In order to perform the algebraic topological computations that determine the Conley index we extend ${\mathcal F}^\text{top}$ to ${\mathcal F}\colon {\mathcal X}\rightrightarrows{\mathcal X}$ by setting
\begin{align*}
{\mathcal F}([x_{n-1},x_n]) &= {\mathcal F}^\text{top}([x_{n-1},x_n]) \\
{\mathcal F}(x_n) &= \downarrow \left({\mathcal F}^\text{top}([x_{n-1},x_n]) \cup {\mathcal F}^\text{top}([x_{n},x_{n+1}]) \right)
\end{align*}
A combinatorial multivalued map ${\mathcal F}^\text{top}\colon {\mathcal X}^{(1)} \rightrightarrows {\mathcal X}^{(1)}$ is an \emph{outer approximation} of $g\colon X\to X$ if $g(x)\in \Int({\mathcal F}^\text{top}(\xi))$ for $x\in \xi$.
To tie these combinatorial multivalued maps to the information provided by ${\mathcal T}$ we make use
of the surrogate map $\mu$ and define
\begin{equation}
\label{eq:minimalMV}
{\mathcal F}_\mu^\text{top}(\xi):=\setof{\xi' \in {\mathcal X}^{(1)} \mid \xi' \cap \mu(\xi) \neq \emptyset}.
\end{equation}
As is shown in \cite{KMV0} ${\mathcal F}_\mu^\text{top}$ is an outer approximation of $\mu$.
Note that any \emph{enclosure} of ${\mathcal F}_\mu^\text{top}$, i.e., ${\mathcal F}^\text{top}\colon {\mathcal X}^{(1)} \rightrightarrows {\mathcal X}^{(1)}$ such that ${\mathcal F}_\mu^\text{top}\subset {\mathcal F}^\text{top}(\xi)$ for every $\xi\in{\mathcal X}^{(1)}$, is an outer approximation of $\mu$.
\begin{rem}
For the remainder of this paper we restrict our attention to multivalued maps ${\mathcal F}^\text{top}$ that are enclosures of ${\mathcal F}_\mu^\text{top}$.
\end{rem}
To see how these combinatorial constructions relate to continuous dynamics, recall that a closed set $K\subset Y$ is an {\bf attractor block} for a continuous function $g\colon Y \to Y$ if
\[
g(K)\subset \Int(K)
\]
where $\Int(K)$ denotes the interior of $K$.
The set of all attractor blocks for $g$ forms a bounded and distributive lattice \cite{lattice1} and is denoted by ${\mathsf{ ABlock}}(g)$.
The bounding elements of ${\mathsf{ ABlock}}(g)$ are ${\bf 1} = Y$ and ${\bf 0} = \emptyset$.
Given ${\mathcal F}^\text{top}\colon {\mathcal X}^{(1)} \rightrightarrows {\mathcal X}^{(1)}$ define ${\mathsf{ Invset}}^+({\mathcal F}^\text{top}):=\setof{{\mathcal S} \subset {\mathcal X}^{(1)} \mid {\mathcal F}^\text{top}({\mathcal S}) \subset {\mathcal S}}$; software which computes ${\mathsf{ Invset}}({\mathcal F}^\text{top})$ is available \cite{CMGDB}.
We leave it to the reader to check that if ${\mathcal F}^\text{top}$ is an enclosure of ${\mathcal F}_\mu^\text{top}$, then ${\mathsf{ Invset}}^+({\mathcal F}^\text{top})\hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_\mu^\text{top})$.
As is shown in \cite{lattice2}, ${\mathsf{ Invset}}^+({\mathcal F}^\text{top})$ is bounded distributive lattice with $\wedge =\cap$, $\vee = \cup$, ${\bf 0} = \emptyset$ and ${\bf 1} = X$.
Of fundamental importance, as it leads to Proposition~\ref{prop:tiling}, is the following result of \cite{lattice2}:
if ${\mathcal F}^\text{top}$ is an outer approximation of $g\colon X\to x$, then ${\mathsf{ Invset}}^+({\mathcal F}^\text{top}) \hookrightarrow {\mathsf{ ABlock}}(g)$.
\begin{prop}
\label{prop:tiling}
Given data set ${\mathcal T} = \setof{(x_n,y_n)}_{n=0}^N$, simplicial complex ${\mathcal X}({\mathcal T})$, and combinatorial multivalued map ${\mathcal F}^\text{top}\colon {\mathcal X}^{(1)} \rightrightarrows {\mathcal X}^{(1)}$ that is an enclosure of ${\mathcal F}_\mu^\text{top}$, consider a lattice ${\mathsf K}$ such that ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}^\text{top}) \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_\mu^\text{top}) \hookrightarrow {\mathsf{ ABlock}}(\mu)$.
For each $K \in {\mathsf J}({\mathsf K})$, define
\begin{equation}\label{eq: tile}
M(K):= \text{cl}(K\setminus \overleftarrow{K})
\end{equation}
and further let
\begin{equation}\label{eq: tiling}
{\mathcal M}({\mathsf K}) := \setof{M(K) \mid K \in {\mathsf J}({\mathsf K})}.
\end{equation}
If ${\mathcal F}^\text{top}$ is an outer approximation of $g\colon X\to X$, then ${\mathcal M}({\mathsf K})$ is a Morse tiling for $g$ with Morse graph ${\mathsf J}({\mathsf K})$.
\end{prop}
For a proof of Proposition~\ref{prop:tiling} in a more general setting see \cite{lattice3}.
However, the intuition behind the result is simple.
Let $x\in X$. Then $x\in M(K)$ for some $K\in{\mathsf J}({\mathsf K})$. This implies that $x\in K\in {\mathsf{ ABlock}}(g)$. Let $n>0$ and assume $g^n(x) \in M(K')$. If $K\neq K'$, then $K'\subset K$ and hence $K' \leq K$.
We now turn to the Conley index and again begin our discussion on the purely combinatorial level.
We assume that we are given ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}^\text{top})$.
We define an {\bf index pair} for ${\mathcal F}^\text{top}$ to be a pair ${\mathcal K} = (K_1,K_0)$ where $K_1,K_0 \in {\mathsf{ Invset}}^+({\mathcal F}^\text{top})$ and $K_0 \subset K_1$.
Then ${\mathcal F}$ induces a map on homology, ${\mathcal F}*:H_*(\downarrow(K_1),\downarrow(K_0)) \to H_*(\downarrow(K_1),\downarrow(K_0))$.
We define the {\bf Conley index} of ${\mathcal K}$ to be the shift equivalence class of ${\mathcal F}_*$ and denote it by $\mathop{\rm Con}\nolimits_*({\mathcal K})$ \cite{mischaikow:weibel:22}.
In particular we can identify each $K \in {\mathsf J}({\mathsf K})$ with the index pair $(K,\overleftarrow{K})$, and so we declare the Conley index of $M(K)$ to be
\begin{equation}\label{eq: index}
\mathop{\rm Con}\nolimits_*(M(K)):= \mathop{\rm Con}\nolimits_*(\downarrow(K),\downarrow(\overleftarrow{K})).
\end{equation}
Turning to continuous dynamics consider a continuous map $g\colon X\to X$.
Given $N \subset X$, the {\bf maximal invariant set} contained in $N$ is given by
\[
\mathop{\mathrm{Inv}}\nolimits(N,g):=\setof{x \in N \mid \exists \upsilon:{\mathbb{Z}} \to N \text{ such that } \upsilon(0) = x \text{ and } \upsilon(k+1) = g(\upsilon(k)) \text{ for all } k \in {\mathbb{Z}} }.
\]
A compact set $N \subset X$ is an {\bf isolating neighborhood} if $\mathop{\mathrm{Inv}}\nolimits(N,g) \subset \Int(N)$.
If $N$ is an isolating neighborhood, then the homology Conley index $\mathop{\rm Con}\nolimits_*(\mathop{\mathrm{Inv}}\nolimits(N,g))$ is well defined \cite{mischmroz}.
To tie together the combinatorial and continuous theory we note that if ${\mathcal F}^\text{top}$ is an outer approximation of $g\colon X\to X$, then $K,\overleftarrow{K}\in {\mathsf{ ABlock}}(g)$.
Therefore, $g\colon (K,\overleftarrow{K})\to (K,\overleftarrow{K})$ and $(K,\overleftarrow{K})$ is an index pair for the classical Conley theory \cite{mischmroz}.
Finally, ${\mathcal F}*:H_*(\downarrow(K_1),\downarrow(K_0)) \to H_*(\downarrow(K_1),\downarrow(K_0))$ and $g_*\colon H_*(K,\overleftarrow{K})\to H_*(K,\overleftarrow{K})$. Thus we obtain the following result.
\begin{prop}\label{Prop: index}
$\mathop{\rm Con}\nolimits_*(M(K))$ is shift equivalent to $\mathop{\rm Con}\nolimits_*(\mathop{\mathrm{Inv}}\nolimits(M(K),g))$.
\end{prop}
\section{Results}\label{Sec: results}
Our construction of a Morse tiling of $X$ for a function $g:X \to X$ relied on a lattice of attractor blocks ${\mathsf K}$ for $g$. Since we are interested in obtaining a Morse tiling of $X$ for a sample path from $f$ we would therefore like to know the probability that a finite sublattice ${\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(\mu)$ is also a lattice of attractor blocks for $f$.
\begin{rem}
Since attractor blocks are defined for $f$ only if $f \in C(X,X)$, we are interested in the event
\[
(f \in C(X,X)) \cap (f(K) \subset \text{Int}(K)\forall K \in {\mathsf K})
\]
that for the sake of simplicity we denote by $({\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f))$.
Since ${\mathsf{ ABlock}}(\mu)$ is bounded from above by ${\bf 1} = X$, and we have assumed that all sublattices will be bounded by the same elements as the original lattice, we have that ${\bf 1} = X \in {\mathsf K}$.
Thus the event $(f(K) \subset \text{Int}(K)\forall K \in {\mathsf K})$ implies the event $(f(x) \in X \forall x \in X)$.
Therefore, since $f$ is continuous with probability $1$,
\[
\mathbb{P}({\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f)) = \mathbb{P}((f(K) \subset \text{Int}(K)\forall K \in {\mathsf K})).
\]
\end{rem}
Let ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_\mu^\text{top}) \hookrightarrow {\mathsf{ ABlock}}(\mu)$; our goal is to compute $\mathbb{P}({\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f))$. We need notation to describe the elements of ${\mathsf K}$ in terms of the edges in ${\mathcal X}$.
Thus, for fixed $K\in {\mathsf K}$ we write $K = \bigcup_{m=1}^{M_K} J_m$ where the
\[
J_m = [x_{i_m},x_{j_m}]
\]
are disjoint closed intervals.
Define the indexing sets ${\mathcal I}_m(K) = {\mathcal I}_m = \setof{n \mid i_m < n \leq j_m}$ for $1\leq m\leq M_K$.
By assumption $K$ is an attractor block for $\mu$, and therefore for each $m$ there is some unique $\tau(m) \in \setof{1,\cdots, M_K}$ such that $\mu(J_m) \subset \Int(J_{\tau(m)})$.
Let ${\mathcal I}(K) = \cup_{m=1}^{M_K} {\mathcal I}_m$ and define the maps $\alpha_{K}:{\mathcal I}({K}) \to {\mathbb{R}}$ and $\beta_{K}:{\mathcal I}({K}) \to {\mathbb{R}}$ as follows.
\begin{equation}
\label{eq:alphaDefn}
\text{
If $n\in {\mathcal I}_{m}$ then $\alpha_{K}(n) := x_{i_{\tau(m)}}$ and $\beta_{K}(n) := x_{j_{\tau(m)}}$.
}
\end{equation}
Notice that $\alpha_{K}(n)$ (resp. $\beta_{K}(n)$) is simply the minimum (resp. maximum) of the connected component of ${K}$ which contains $y_{n-1},y_n$.
The event that $K$ is an attracting block for $f$ is equivalent to $\bigcap_{n\in {\mathcal I}({K})} S_n(\alpha_{K}(n),\beta_{K}(n))$. Therefore we have the following result.
\begin{prop}\label{Prop:AttBlBound}
\begin{equation}
\mathbb{P}( f({K}) \subset \Int({K})) =\prod_{n \in {\mathcal I}({K})} \mathbb{P}(S_n(\alpha_{K}(n),\beta_{K}(n)))
\end{equation}
\end{prop}
We now aim to extend Proposition~\ref{Prop:AttBlBound} to all of ${\mathsf K}$ simultaneously.
Let ${\mathcal I} := \setof{1,\ldots, N}$ and define the map $\gamma:{\mathcal I} \to {\mathsf K}$ by
\[
\gamma(n):=\min\setof{K \in {\mathsf K} \,|\, n \in {\mathcal I}(K)}
\]
where the minimum is taken with respect to the partial order of ${\mathsf K}$. Note that this minimum is well-defined because if $n \in {\mathcal I}(K)$ and $n \in {\mathcal I}(K')$ then $K\cap K' \in {\mathsf K}$ and $n \in {\mathcal I}(K \cap K')$.
Observe that the event $S_n(\alpha_{\gamma(n)}(n),\beta_{\gamma(n)}(n))$ implies the event $S_n(\alpha_{K}(n),\beta_{K}(n))$ for all $K$ such that $n \in {\mathcal I}(K)$; this statement holds because $n\in {\mathcal I}(K)\cap {\mathcal I}(K')$ and $K < K'$ implies that $\alpha_{K'}(n) \leq \alpha_{K}(n) < \beta_{K}(n) \leq \beta_{K'}(n)$. Finally, we define the maps $\alpha:{\mathcal I} \to {\mathbb{R}}$ and $\beta:{\mathcal I} \to {\mathbb{R}}$ by
\[
\alpha(n):= \alpha_{\gamma(n)}(n), \hspace{.5cm} \beta(n):= \beta_{\gamma(n)}(n).
\]
With this notation introduced we may now state one of the key results of this paper.
\begin{thm}\label{Thm:LattBound}
Let ${\mathsf K}$ be a sublattice of ${\mathsf{ Invset}}^+({\mathcal F})$. The probability that ${\mathsf K}$ is a lattice of attractor blocks for $f$ is given by the following equation.
\begin{equation}
\mathbb{P}( {\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f)) =\prod_{n \in {\mathcal I}} \mathbb{P}(S_n(\alpha(n),\beta(n)))
\end{equation}
\end{thm}
\begin{proof}
Recall that the lattice ${\mathsf K}$ is naturally imbued with a partial order $\leq$. Choose a linear extension $\leq'$ of the partial order $\leq$. This allows us to write ${\mathsf K} = \setof{K_q \mid 1 \leq' q \leq' Q}$ where the labelling of the sets $K_q$ respects the partial order of ${\mathsf K}$; that is, if $K_p \subset K_q$ then $p \leq q$.
For each $1\leq q \leq Q$ and each $n\in {\mathcal I}$, define $S_n^q$ to be the event $f([x_{n-1},x_n]) \subset (\alpha_{K_q}(n),\beta_{K_q}(n))$. Let $V_q$ be the event that $K_q \in {\mathsf{ ABlock}}(f)$; then $V_q =\cap_{n \in {\mathcal I}(K_q)} S_n^q$. Our goal is to compute
$$\mathbb{P}( {\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f)) = \mathbb{P}(\cap_{q=1}^Q V_q) = \mathbb{P}(V_1)\mathbb{P}(V_2|V_1)\mathbb{P}(V_3|V_1\cap V_2)\cdots\mathbb{P}(V_Q|\cap_{q=1}^{Q-1}V_q).$$
In order to compute that value we will use the following intermediary step for any $1\leq P \leq Q$.
\begin{align*}
\mathbb{P}(V_P | \cap_{q=1}^{P-1}V_q) &= \mathbb{P}(\cap_{n\in {\mathcal I}(K_p)} S_n^P | (\cap_{n \in {\mathcal I}(K_1)} S_n^1)\cap(\cap_{n \in {\mathcal I}(K_2)} S_n^2)\cap\cdots\cap(\cap_{n \in {\mathcal I}_{P-1}} S_n^{k-1}))\\
&= \mathbb{P}(\cap_{n\in {\mathcal I}(K_P)} S_n^P | \cap_{n \in (\cup_{p = 1}^{P-1}{\mathcal I}(K_p))}S_n(\alpha(n),\beta(n)))\\
&= \mathbb{P}(\cap_{n\in {\mathcal I}(K_P)\setminus (\cup_{p = 1}^{P-1}{\mathcal I}(K_p))} S_n^P)\\
&= \mathbb{P}(\cap_{n\in {\mathcal I}(K_P)\setminus (\cup_{p = 1}^{P-1}{\mathcal I}(K_p))} S_n(\alpha(n),\beta(n)))\\
&= \prod_{n\in {\mathcal I}(K_P)\setminus (\cup_{p = 1}^{P-1}{\mathcal I}(K_p))}\mathbb{P}(S_n(\alpha(n),\beta(n)))\\
\end{align*}
The result then follows from this computation:
\begin{align*}
\mathbb{P}(\cap_{q=1}^Q V_q)
&= \mathbb{P}(V_1)\mathbb{P}(V_2|V_1)\mathbb{P}(V_3|V_2\cap V_1)\cdots \mathbb{P}(V_Q|\cap_{q=1}^{Q-1}V_q)\\
&= \prod_{q = 1}^Q \prod_{n \in {\mathcal I}(K_q)\setminus(\cup_{p = 1}^{q-1}{\mathcal I}(K_p))} \mathbb{P}(S_n(\alpha(n),\beta(n)))\\
&= \prod_{n \in {\mathcal I}} \mathbb{P}(S_n(\alpha(n),\beta(n)))
\end{align*}
\end{proof}
As indicated in Section~\ref{Sec: conley}, knowing a lattice of attractor blocks ${\mathsf K}$ for a function allows us to determine a Morse tiling of the domain for that function. In order to compute the Conley index of each of the associated Morse tiles, however, we must have an outer approximation of the function of interest; therefore we will now construct a combinatorial multivalued map ${\mathcal F}_{\mathsf K}^\text{top}:{\mathcal X}^{(1)}\rightrightarrows {\mathcal X}^{(1)}$ that is an outer approximation of $f$ whenever ${\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f)$.
In fact, ${\mathcal F}_{\mathsf K}^\text{top}$ is essentially defined by the construction used in Theorem~\ref{Thm:LattBound}. For each $\xi_n:=[x_{n-1},x_n] \in {\mathcal X}^{(1)}$ let
\begin{equation}
\label{eq:cFsK}
{\mathcal F}_{\mathsf K}^\text{top}(\xi_n):=[\alpha(n),\beta(n)].
\end{equation}
\begin{prop}\label{Prop: outappx}
Let ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_\mu)$ and define ${\mathcal F}_{\mathsf K}$ by Equation~\eqref{eq:cFsK}. The following properties hold.
\begin{enumerate}
\item ${\mathcal F}_{\mathsf K}^\text{top}$ is an enclosure of ${\mathcal F}_\mu^\text{top}$.\label{enclosure}
\item ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_{\mathsf K}^\text{top})$\label{invset}
\item ${\mathcal F}_{\mathsf K}^\text{top}$ is an outer approximation of $f$ if and only if $({\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f))$.\label{outappx}
\end{enumerate}
\end{prop}
\begin{proof}
Properties \ref{enclosure} and \ref{invset} follow directly from the construction of ${\mathcal F}_{\mathsf K}^\text{top}$ and Theorem~\ref{Thm:LattBound}. Further, if ${\mathcal F}_{\mathsf K}^\text{top}$ is an outer approximation of $f$ then ${\mathsf{ Invset}}^+({\mathcal F}_{\mathsf K}) \hookrightarrow {\mathsf{ ABlock}}(f)$ \cite{lattice2} and so ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_{\mathsf K})\hookrightarrow {\mathsf{ ABlock}}(f)$.
Therefore it remains only to show that ${\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f)$ implies that ${\mathcal F}_{\mathsf K}^\text{top}$ is an outer approximation for $f$; we prove this by the contrapositive and assume that ${\mathcal F}_{\mathsf K}^\text{top}$ is not an outer approximation for $f$. Then there exists $x$ such that $f(x) \not \in \Int({\mathcal F}_{\mathsf K}^\text{top}(\xi_n))$ where $x \in \xi_n = [x_{n-1},x_n]$. Consider $\gamma(n) \in {\mathsf K}$ and write $\gamma(n) = \bigcup_{m=1}^{M_{\gamma(n)}} J_m$. Let $m'$ be the integer such that $\xi_n \subset J_{m'}$. By construction ${\mathcal F}_{\mathsf K}^\text{top}$ is constant on each $J_m$ and hence $f(x) \not \in \Int({\mathcal F}_{\mathsf K}^\text{top}(J_{m'}))= \Int({\mathcal F}_{\mathsf K}^\text{top}(\xi_n))$. Again by construction ${\mathcal F}_{\mathsf K}^\text{top}(J_{m'}) = J_{\tau(m')}$ and so $f(x) \not \in \Int(J_{\tau(m')})$; thus $f(J_{m'}) \not \subset \Int(J_{\tau(m')})$ and so $f(K) \not \subset \Int(K)$ (here we use the fact that the $J_m$ are disjoint intervals, and so $f(K) \subset \Int(K)$ only if each of these intervals map into the interior of another one).
\end{proof}
We now obtain the main result of this paper as a direct consequence of Theorem~\ref{Thm:LattBound} and Propositions~\ref{prop:tiling}, \ref{Prop: index}, and \ref{Prop: outappx}.
\begin{thm}\label{Thm:TileBnd}
Let ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_\mu^\text{top})$ and define $M(K)$ and $\mathop{\rm Con}\nolimits_*(M(K))$ for each $K \in {\mathsf J}({\mathsf K})$ by Equations~\eqref{eq: tile} and \eqref{eq: index} respectively. Then with probability
\[\prod_{n \in {\mathcal I}} \mathbb{P}(S_n(\alpha(n),\beta(n)))\]
${\mathcal M}({\mathsf K}) = \setof{M(K) \mid K \in {\mathsf J}({\mathsf K})}$ is a Morse tiling of $X$ for $f$ and $\mathop{\rm Con}\nolimits_*(M(K))$ is the Conley index of each Morse tile.
\end{thm}
\section{Examples}\label{Sec: examples}
This section contains three examples that help to explain the results of Section~\ref{Sec: results}. In each example we plot the mean $\mu$ of the Gaussian process $f$ as well as a random sample path from $f$. The sample path plotted does not come into our analysis of the examples in any way and is merely included in order to illustrate the ideas at play.
In this first example we begin with an attractor block identified for the mean $\mu$ that is a single closed interval; a single closed interval is the simplest form of an attractor block possible. We will apply Proposition~\ref{Prop:AttBlBound} and give the probability that this interval is an attractor block for the Gaussian process of interest.
\begin{ex}\label{Ex: intvl}
Consider
\begin{align*}
{\mathcal T} &= \setof{(x_n,y_n) \mid x_n = n/10}_{n=0}^{10}\\
&= \{(0,.4509),(.1,.4999),(.2,.4613),(.3,.455),(.4,.5185),\\
&\, (.5,.4987),(.6,.5398),(.7,.5147),(.8,.5397),(.9,.5221),(1,.5331)\}.
\end{align*}
We will let $f$ be Brownian motion with variance parameter $\sigma^2 = 1$ conditioned to pass through ${\mathcal T}$. The mean function of $f$, $\mu$, is the piecewise linear function indicated in blue, shown in Figure~\ref{Fig:ex1_all}. The combinatorial multivalued map ${\mathcal F}_\mu^\text{top}$ is shown in Figure~\ref{Fig:ex1_F}; $K := X = [0, 1]$ is an element of ${\mathsf{ Invset}}^+({\mathcal F}_\mu^\text{top})$ and hence $K\in {\mathsf{ ABlock}}(\mu)$. More directly, we can see that $K$ is an attractor block for $\mu$ precisely because $0 < f(x_n) < 1$ for all $n \in \setof{0,\cdots,10}$, and the piecewise linearity of $\mu$ then implies that $ 0 < f(x) < 1$ for all $x \in [0,1]$.
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{ex1_F.png}
\caption{The combinatorial multivalued map ${\mathcal F}_\mu^\text{top}$ is shown above. Cells on the horizontal axis map to the collection of cells on the vertical axis that indicated by the grey coloring.}
\label{Fig:ex1_F}
\end{subfigure}
\hfill
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{ex1_main.png}
\caption{The interval $[0, 1]$ is an attracting block for the sample path because the sample path lies entirely between $0$ and $1$ on the interval $[0,1]$. The dashed lines define ${\mathcal F}_{\mathsf K}^\text{top}$.}
\label{Fig:ex1_main}
\end{subfigure}
\caption{}
\label{Fig:ex1_all}
\end{figure}
In this case, $\alpha_K(n)$ is identically $\alpha:= 0$, and $\beta_K(n)$ is identically $\beta:=1$. By Proposition~\ref{Prop:AttBlBound} the probability that $K \in{\mathsf{ ABlock}}(f)$ is
\[
\prod_{n=1}^{10}
1 - \pi\left(\frac{(\beta - y_n)}{\sigma \sqrt{x_n - x_{n-1}}}, \frac{(\beta - y_{n-1})}{\sigma \sqrt{x_n - x_{n-1}}} ,\frac{(y_n-\alpha)}{\sigma \sqrt{x_n - x_{n-1}}}, \frac{(y_{n-1} - \alpha)}{\sigma \sqrt{x_n - x_{n-1}}}\right)
\approx
0.8586.
\]
Alternatively, we can view the set ${\mathsf K} = \setof{\emptyset, K=X}$ as a lattice of attractor blocks for $\mu$ and apply Theorem~\ref{Thm:LattBound} to determine that $\mathbb{P}({\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f)) = 0.8586$ as well. In this example the Morse tiling of $X$ is trivial and ${\mathcal M}({\mathsf K}) = \setof{M(K)=X}$. Also,
\[
\mathop{\rm Con}\nolimits_k(M(K)) = \begin{cases}
\mathop{\mathrm{id}}\nolimits & \text{if $k=0$} \\
0 & \text{otherwise}
\end{cases}
\]
which implies the existence of a fixed point \cite{szymczak:96,day:frongillo:trevino}.
\end{ex}
In this next example we consider a larger lattice of sets and use Theorem~\ref{Thm:LattBound} in order to identify the probability that this lattice is made up of attractor blocks for $f$. While the probabilistic ideas are not really any different from the preceding example, the indexing required to keep track of this more sophisticated structure is more complicated. We remark that this example demonstrates how the techniques developed in this paper may be used to identify bistability in a system.
\begin{ex}
We consider the set
\begin{align*}
{\mathcal T} &= \setof{(x_n,y_n) \mid x_n = n/10}_{n=0}^{10}\\
&= \{(0,.2005),(.1,.225),(.2,.2057),(.3,.2025),(.4,.2343),\\
&\, (.5,.5243),(.6,.8449),(.7,.8324),(.8,.8448),(.9,.8361),(1,.8416)\}
\end{align*}
shown in Figure~\ref{Fig:ex2_all}, and let $f$ be Brownian motion with variance parameter $\sigma^2 = 1/16$ conditioned to pass through ${\mathcal T}$. We err on the side of verbosity in analyzing this data in order to be totally clear how the indexing works when dealing with a general lattice of attracting blocks.
Using ${\mathcal F}_\mu^\text{top}$, shown in Figure~\ref{Fig:ex2_F}, we identify five attracting blocks for the posterior mean $\mu$ which form a lattice ${\mathsf K}$:
$$ K_0 = \emptyset, K_1 = [x_{0},x_{4}], K_2 = [x_{6},x_{10}], K_3 = K_1 \cup K_2, K_4 = [x_{1}, x_{10}]$$
Note that $K_0 \subset K_1 \subset K_3 \subset K_4$ and $K_0 \subset K_1 \subset K_2 \subset K_4$, but $K_1$ and $K_2$ are incomparable.
The indexing sets are
\begin{align*}
{\mathcal I}(K_1) & = \setof{1,2,3,4}, \\
{\mathcal I}(K_2) & = \setof{7,8,9,10}, \\
{\mathcal I}(K_3) & = \setof{1,2,3,4,7,8,9,10}, \\
{\mathcal I}(K_4) &= \setof{1,2,3,\cdots,9,10}.
\end{align*}
For $ q\neq 3$, each component $K_q$ consists only of a single connected component and thus the maps $\alpha_{K_q}:{\mathcal I}(K_q) \to {\mathbb{R}}$, $\beta_{K_q}:{\mathcal I}(K_q) \to {\mathbb{R}}$ are constant maps:
$$ \alpha_{K_1} \equiv x_{0}, \beta_{K_1} \equiv x_{4}$$
$$ \alpha_{K_2} \equiv x_{6}, \beta_{K_2} \equiv x_{10}$$
$$ \alpha_{K_4} \equiv x_{0}, \beta_{K_4} \equiv x_{10}$$
The maps $\alpha_{K_3}$ and $\beta_{K_3}$ are slightly more complicated since $K_3$ has two connected components, but we do not need to describe them because $\gamma(n) \neq K_3 $ for any $n \in {\mathcal I}$. We do need to know $\gamma(n)$ for each $n \in {\mathcal I}$:
\[
\gamma(n) =
\begin{cases}
K_1, & n \in {\mathcal I}(K_1)\\
K_2, & n \in {\mathcal I}(K_2)\\
K_4, & \text{otherwise}
\end{cases}
\]
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{ex3_F.png}\caption{The combinatorial multivalued map ${\mathcal F}_\mu^\text{top}$.}
\label{Fig:ex2_F}
\end{subfigure}
\hfill
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{ex3_main.png}
\caption{${\mathsf K}$ is a lattice of attracting blocks for the sample path shown because the path remains between the indicated (dashed) bounds; these bounds define ${\mathcal F}_{\mathsf K}^\text{top}$.}
\label{Fig:ex2_main}
\end{subfigure}
\caption{}
\label{Fig:ex2_all}
\end{figure}
With this information we are able to calculate $\mathbb{P}({\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f))$:
\begin{small}
\[
\prod_{n \in {\mathcal I}}
1 - \pi\left(\frac{(\beta(n) - y_n)}{\sigma \sqrt{x_n - x_{n-1}}}, \frac{(\beta(n) - y_{n-1})}{\sigma \sqrt{x_n - x_{n-1}}} ,\frac{(y_n-\alpha(n))}{\sigma \sqrt{x_n - x_{n-1}}}, \frac{(y_{n-1} - \alpha(n))}{\sigma \sqrt{x_n - x_{n-1}}}\right)
\approx 0.9989
\]
\end{small}
Having determined the probability that ${\mathsf K}$ is a lattice of attractor blocks for $f$ we now provide the associated Morse tiling and Conley indices and discuss what these features indicate about the dynamics of $f$.
The lattice ${\mathsf K}$ gives the Morse tiling
\[
{\mathcal M}({\mathsf K}) = \setof{M(K_1) = K_1, M(K_2) = K_2, M(K_4) = \text{cl}(K_4\setminus K_3) \mid M(K_4) > M(K_1), M(K_4)>M(K_2)}.
\]
The Conley indices of these Morse tiles are
\[
\mathop{\rm Con}\nolimits_k(M(K_1)) = \mathop{\rm Con}\nolimits_k(M(K_2)) = \begin{cases}
\mathop{\mathrm{id}}\nolimits & \text{if $k=0$} \\
0 & \text{otherwise.}
\end{cases}
\quad\text{and}\quad
\mathop{\rm Con}\nolimits_k(M(K_4)) = \begin{cases}
\mathop{\mathrm{id}}\nolimits & \text{if $k=1$} \\
0 & \text{otherwise}
\end{cases}
\]
This Morse tiling indicates that the dynamical system is bistable, with attractors that contain at least one fixed point in $K_1$ and $K_2$ \cite{szymczak:96,day:frongillo:trevino}.
Moreover, there is a non-trivial invariant set in $(x_4,x_6)$ with the Conley index of a repelling fixed point. By Theorem~\ref{Thm:TileBnd}, all of this information is valid for $f$ with probability $0.9989$.
\end{ex}
Before beginning this final example we make a more general comment. Given a dynamical system $g:Y \to Y$ and a subset $Y'\subset Y$ such that $g(Y') \subset Y'$, the function $g|_{Y'}:Y' \to Y'$ defines a dynamical system. Such a restricted system is sometimes of interest when one is concerned only with the local dynamics in the region $Y'$. These local dynamics can also be understood using Conley theory and the probabilistic techniques that we develop in this paper. In particular, if one is interested only in understanding the dynamics of $f|_{X'}$, where $X' = \bigcup_{n \in {\mathcal I}'} [x_{n-1},x_n]$ for ${\mathcal I}' \subset \setof{1,\cdots,N}$, then we can analyze the dynamics of $X'$ in the same manner as we analyze the dynamics of $X$. That is, if we let ${\mathcal X}' $ be a simplicial complex with vertices ${\mathcal X}'^{(0)} := \setof{x_n}_{n \in {\mathcal I}'}$ and edges ${\mathcal X}'^{(1)} := \setof{[x_{n-1},x_{n}]}_{n\in {\mathcal I}'}$, and for ${\mathcal F}={\mathcal F}_\mu$ or ${\mathcal F}_{\mathsf K}$ we define the combinatorial multivalued map ${\mathcal F}':{\mathcal X}'\rightrightarrows{\mathcal X}'$ by ${\mathcal F}'(\xi):={\mathcal F}(\xi)\cap{\mathcal X}'$, then we are able to give a Morse tiling of $X'$ for $f|_{X'}$ using the same methodology as we did for $X$ and $f$.
In this final example we exploit this perspective. We begin by analyzing the dynamics of a Gaussian process defined on the interval $X = [0,1]$. We will see that the lattice of attractor blocks ${\mathsf K}$ identified for the mean $\mu$ that we define has a fairly low probability of being a lattice of attractor blocks for $f$. However, we will then note that if we restrict our view to a more local perspective, and instead analyze the same system on the interval $X' = [0.2,0.8]$, we have a reasonably high probability that $f|_{X'}$ contains a periodic orbit. Hopefully this example demonstrates how an individual who is interested only in certain local information--in this case, a periodic orbit--may increase the probability of seeing the dynamics of interest by restricting their view to a smaller region of phase space.
\begin{ex}\label{ex: per_orb}
Consider the set
\begin{align*}
{\mathcal T} &= \setof{(x_n,y_n) \mid x_n = n/10}_{n=0}^{10}\\
&= \{(0,.03),(.1,.07),(.2,.6853),(.3,.6999),(.4,.6884),\\
&\, (.5,.501),(.6,.255),(.7,.3185),(.8,.2987),(.9,.95),(1,.97)\},
\end{align*}
shown in Figure~\ref{Fig:ex3_all}. Let $f$ be Brownian motion on $X = [0,1]$ with variance parameter $\sigma^2 = 1/16$ conditioned to interpolate ${\mathcal T}$. The mean function $\mu$ of $f$ is the piecewise linear function indicated in blue. The lattice ${\mathsf K} = \setof{K_0 = \emptyset, K_1 = [0.2,0.4] \cup [0.6,0.8], K_2 = [0,2,0.8], K_3 = X}$ is a lattice of attractor blocks for $\mu$; we determine this information because ${\mathsf K} \hookrightarrow {\mathsf{ Invset}}^+({\mathcal F}_\mu^\text{top})$, where ${\mathcal F}_\mu^\text{top}$ is shown in Figure~\ref{Fig:ex3_F}. We observe that $K_1$ contains a periodic orbit for $\mu$ and $K_2 \setminus K_1$ contains a repelling fixed point. Our goal in this example is to show that there is a relatively high probability that $f$ has a periodic orbit and an isolated invariant with the index of an unstable fixed points in these same regions.
For each of the connected intervals $K_q$, the maps $\alpha_{K_q}$ and $\beta_{K_q}$ are constant maps whose images are, respectively, the left and right endpoints of $K_q$. However, for $K_1$ the situation is somewhat more complicated. Note that $\mu([0.2,0.4]) \subset (0.6,0.8)$ and $\mu([0.6,0.8]) \subset (0.2,0.4)$. Thus
\[
\alpha_{K_1}(n) = \begin{cases}
0.6, & n = 3,4\\
0.2, & n = 7,8
\end{cases}
\]
and
\[
\beta_{K_1}(n) = \begin{cases}
0.8, & n = 3,4\\
0.4, & n = 7,8
\end{cases}.
\]
Therefore we have that
\[
\alpha(n) = \begin{cases}
0, & n = 1,2,9,10\\
0.6, & n = 3,4\\
0.2, & n = 5,6,7,8
\end{cases}
\]
and
\[
\beta(n) = \begin{cases}
1, & n = 1,2,9,10\\
0.8, & n = 3,4,5,6\\
0.4, & n = 7,8
\end{cases}.
\]
With these maps defined we are able to use Theorem~\ref{Thm:LattBound} to compute
\[
\mathbb{P}({\mathsf K} \hookrightarrow {\mathsf{ ABlock}}(f)) \approx 0.1199.
\]
This probability is relatively low, but we observe that a sample path is most likely to leave the necessary bounds on $K_4\setminus K_2$; for instance, $\mathbb{P}(S_{10}(\alpha(10),\beta(10)) = 0.3812$. Because we are not concerned with the behavior of $f$ in this region we restrict our view to the domain $X':= K_2$. For this domain, the lattice ${\mathsf K}'= \setof{K_0 = \emptyset, K_1, K_2}$ is a full lattice, bounded from above by the domain itself. Now Theorem~\ref{Thm:LattBound} implies that
\[
\mathbb{P}({\mathsf K}' \hookrightarrow {\mathsf{ ABlock}}(f|_{X'})) \approx 0.6283.
\]
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{ex2_F.png}
\caption{The combinatorial multivalued map ${\mathcal F}_\mu^\text{top}$.}
\label{Fig:ex3_F}
\end{subfigure}
\hfill
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=\textwidth]{ex2_main.png}
\caption{The dashed lines define ${\mathcal F}_{\mathsf K}^\text{top}$. Note that ${\mathcal F}_{\mathsf K}^\text{top}$ is not an outer approximation for the sample path shown, but ${\mathcal F}_{{\mathsf K}'}^\text{top}$ is an outer approximation of the sample path restricted to $K_2= [0.2,0.8]$.}
\label{Fig:ex3_main}
\end{subfigure}
\caption{Note that ${\mathsf K}$ is not a lattice of attracting blocks for the sample path shown because the path leaves the required bounds. However, $K_1$ and $K_2$ are both attractor blocks for the sample path. We conclude that the sample path contains a periodic orbit in $K_1$ and an isolated invariant set with the Conley index of an unstable fixed point; the same is true of $f$ with probability $0.6283$. }
\label{Fig:ex3_all}
\end{figure}
The Morse tiling associated with ${\mathsf K}'$ is
\[{\mathcal M}({\mathsf K}') = \setof{M(K_1) = K_1, M(K_2) = \text{cl}(K_2 \setminus K_1) \mid M(K_2) > M(K_1)};
\]
by Theorem~\ref{Thm:TileBnd}, this ${\mathcal M}({\mathsf K}')$ is a Morse tiling for $f$ with probability $0.6283$. Using ${\mathcal F}_{{\mathsf K}'}^\text{top}$, we are able to compute the Conley indices
\[
\mathop{\rm Con}\nolimits_k(M(K_1)) = \begin{cases}
\begin{bmatrix}
0 & 1 \\
1 & 0
\end{bmatrix} & \text{if $k=0$} \\
0 & \text{otherwise.}
\end{cases}
\quad\text{and}\quad
\mathop{\rm Con}\nolimits_k(M(K_2)) = \begin{cases}
-\mathop{\mathrm{id}}\nolimits & \text{if $k=1$} \\
0 & \text{otherwise}
\end{cases}.
\]
These indices imply the existence of a periodic orbit for $f$ in the interior of $K_1$ whenever ${\mathcal F}_{{\mathsf K}'}^\text{top}$ is an outer approximation of $f|_{K_2}$ \cite{szymczak:96,day:frongillo:trevino}; thus we conclude that $f$ has a periodic orbit with at least probability $0.6283$. Therefore, while we cannot make particularly strong claims about the global dynamics of $f$ all of $X$, we are able to identify interesting features of the dynamics on $X' = [0.2,0.8]$ with a reasonable probability.
\end{ex}
\section{Future Directions and Comparisons to Other Approaches}\label{Sec: concl}
We begin this section by comparing our results to those found in \cite{bogdan}. The spirit of both papers is the same. In each of them, combinatorial Conley index theory and Gaussian processes are combined in order to identify dynamics with rigorous probabilities. However, there are two key differences between these, reflecting a trade-off between the methods used in each case.
The first difference is that the results of \cite{bogdan} apply to a larger set of Gaussian processes. The dynamics of any Gaussian process with a covariance kernel that is at least four times differentiable may be studied using the results of that paper. This requirement means that the Gaussian process $f$ that is analyzed in this paper cannot be studied using the methods of \cite{bogdan}, but the majority of Gaussian processes used in applications--including any process using the squared exponential covariance kernel--can be analyzed using those methods.
On the other hand, the results of \cite{bogdan} are asymptotic. In that paper, the set of points ${\mathcal T}$ that is used to condition the Gaussian process represents sampled data. The main theorem establishes a procedure that characterizes dynamics where, given a fixed confidence level $\delta\in (0,1)$, it is proven that for a large enough sample ${\mathcal T}$ the characterization is accurate with confidence level $\delta$. However, for any single fixed sample ${\mathcal T}$, the method cannot rigorously identify the confidence that the characterization is accurate. By contrast, this work can identify the probability that the provided characterization of dynamics is accurate for a fixed ${\mathcal T}$ and the Gaussian process $f$ interpolating ${\mathcal T}$.
Future work should focus on extending the techniques from this paper to higher dimensional data and more general covariance structures. In order to replicate the results of this paper for more general Gaussian process requires obtaining some understanding of the events $S_n$ in this more general setting. One natural approach to addressing this problem would be to use the maxima and minima of a Gaussian process in order to estimate the probability of $S_n$. That is, for a general Gaussian process $G$ with parameter space ${\mathbb{R}}$, if we let $U_n(\beta) = \setof{\sup_{x\in[x_i,x_j]} G(x) \leq \beta}$ and $L_n(\beta) = \setof{\inf_{x\in[x_i,x_j]} G(x) \geq \alpha}$, then $\mathbb{P}(S_n(\alpha,\beta)) \geq 1 - \mathbb{P}(U_n(\beta)) - \mathbb{P}(L_n(\alpha))$; therefore if we can know (or bound) these probabilities we should be able to obtain lower bounds on the probability that a lattice of attractor blocks identified for the mean of the Gaussian process is also a lattice of attractor blocks for the true map. One possible route to obtain such bounds would be through Rice's formula \cite{adtay}.
Extending the results to higher dimensional data sets again requires obtaining (or bounding) probabilities like $S_n$. As mentioned earlier, a closed set $K$ is an attractor block for a map $g$ if $g(K) \subset \Int(K)$; this characterization does not depend on the dimension. We should be able to give a bound on the probability of such events in any finite dimension $d$ if we can bound the probability that hypercubes map into other hypercubes. That is, we would like to know or estimate $$\mathbb{P}(f(\prod_{k=1}^d[a_k,b_k]) \subset \prod_{k=1}^d[\alpha_k,\beta_k]).$$ While this cubical approach will not allow us to perfectly represent any attractor block in higher dimensions (which have no geometric constraints in general) if we can obtain formulas like this then we will likely be able to extend the results of this paper to higher dimensions with a reasonable level of generality.
Finally, we note that one obvious extension of this work is to develop a statistical procedure which allows us to analyze data sets of the same form as ${\mathcal T} = \setof{(x_n,y_n)}_{n=0}^N$. While we worked in a purely probabilistic setting here, we remarked in the introduction that such analysis is our ultimate motivation, and thus future work should attempt to make this generalization.
\section{Acknowledgements}
The authors would like to thank Harry van Zanten, Ying Hung, and Kasper Larsen; our conversations on Gaussian processes were very valuable in crafting this paper.
C.T. was partially supported by HDR TRIPODS award CCF-1934924. K.M. was partially supported by the National Science Foundation under awards DMS-1839294 and HDR TRIPODS award CCF-1934924, DARPA contract HR0011-16-2-0033, and NIH 5R01GM126555-01. K.M. was also supported by a grant from the Simons Foundation.
|
\section{Introduction}
In today's modern world of 24/7 surveillance, vision sensory data \textcolor{black}{are} widely used to monitor activities automatically and report them to connected departments for counter actions. Automated surveillance is a major problem and concern of computer vision experts, where deep models have achieved tremendously precise results in many computer vision problems, ranging from object detection to complex multiple activities prediction and perception. The achievements of computer vision techniques are mainly due to the excessive demand of automated video analytic methods for enormous applications such as video indexing and retrieval, video summarization~\cite{wu2020dynamic}, action and activity recognition \cite{dang2020sensor}, and video classification~\cite{sasithradevi2020video}. In video classification domain, action and activity recognition has received significant attention in research community due to its applications to medical healthcare, security, sports and entertainment, among many others. Where violence detection (VD) is among other video classification major groups focused on detection of harmful and brutal events patterns from an input video. Broader-level categorization of video classification domains is given in Figure~\ref{fig:videoclassificationdomains}.
\begin{figure}
\centering
\smartdiagram[bubble diagram]
{ \textbf{Video} \\ \textbf{classification} \\, \\ \\ \\ Activity \\ Recognition \\ \cite{dang2020sensor,pareek2021survey} \\ \\ \\,
\\ \textit{Violence}\\\textit{Detection \cite{sumon2020violence}} \\, \\ \\ Anomaly \\ Recognition \\ \cite{ullah2021efficient,ullah2020one} \\ , Others \cite{spolaor2020systematic,shen2020advance} \\ (retrieval and indexing) \\}
\caption{Video classification categorization into sub-groups. The diameter of the circle indicates the maturity level of the field which is selected based on its reputation, number of datasets, and endorsement of related research community in terms of citations. The italic text refers to the currently observed topic. }
\label{fig:videoclassificationdomains}
\end{figure}
VD refers to abnormal patterns selection$/$extraction from a normal surveillance video that could occur for a very short or a prolonged time duration. Its early detection assists in either prevention or reducing the residual damage in terms of human lives and their properties by initiating countermeasures such as notifying the nearest concerned departments for responsive actions. Violence can be any abnormal brutal action by humans including hitting someone, damaging, destroying, and injuring.
\textcolor{black}{Video data (VD) based on video data is applicable to a large number of real-world scenarios. It is evident that the occurrence of violence is extremely rare when compared to the normal pattern of continuous video surveillance. As a result, deploying human resources to monitor video streams could be impractical, as it would require repetitive practice in order to detect abnormal patterns. Currently, automated video detection techniques are of great assistance in monitoring surveillance video streams in real-time. The main goal of a VD system is to detect and report any abnormal behavior, which is deviated from the normal pattern of behavior. Therefore, instead of human intervention, VD can be used effectively in real-world surveillance scenarios. \textcolor{black}{Based on the categorized video data collected from CCTV cameras, the VD techniques utilise different soft computing mechanisms to learn the patterns and analyze the types of motions associated with normal and abnormal behaviors ~\cite{chang2022hybrid} in order to detect violent and non-violent situations using the collected data ~\cite{mumtaz2022fast}}} VD using \textcolor{black}{soft computing techniques} techniques from video data is in practice since decades \cite{datta2002person}, initiated based on traditional image processing techniques. The early VD literature has consideration of various attributes for decision making such as acceleration of human motion, their appearance, and motion flow, among many others. The generic steps involved in baseline research include video data preprocessing, features extraction, and its final classification into violent or non-violent segments. The preprocessing step involves data cleansing and video segmentation. Features extraction refers to individual-level (spatial) frames processing using certain features extractors such as histogram oriented gradients (HOG), motion, speed, and optical flow. The final classification of these features is based on an existing model's weights and configuration acquired using the training set of the dataset. The models used in pioneering VD research are machine learning based techniques such as support vector machine (SVM). These steps are given in Figure~\ref{fig:VDsteps}.
\begin{figure}
\centering
\smartdiagram[descriptive diagram]
{
{Video data preprocessing, {Data cleansing, video segmentation, objects of interest detection (humans).}},
{Features extraction,{Low and high-level spatial or temporal features extraction. Some methods consider spatio-temporal features.}},
{Trained model's prediction,{Trained weights based classification. It can be a trained SVM, 2D or 3D CNN model.}},
}
\caption{Major steps involved in VD. }
\label{fig:VDsteps}
\end{figure}
With the advent and advancements of deep learning in computer vision domains, researchers also employed it for VD. The basic pipeline of deep learning based VD is closely similar to the traditional mechanism, but the absence of preprocessing in most of the deep learning based strategies is a highlighted distinction. In most of the cases, a deep VD model has an end-to-end architecture to input a video (sequence of frames) or a single frame and output their probability of being violent or normal.
While there has been extensive research in activity recognition domain \cite{dang2020sensor}, VD, being its sub-domain has not been explored to its value in real-world problems. Although abnormal activity recognition too has been studied widely, but that is different from violent activity recognition. There exist several challenges in VD domain, the foremost of which is the limited datasets publicly available. Real-time performance with significant accuracy to be implemented in a smart city's surveillance scenarios has yet to be achieved. Embedded devices with intelligent vision have been recently applied to many emerging computer vision domains, need to be applied in VD domain.
The mentioned challenges and their further extension along with future derivations based on the current literature are provided in this review article. The proposed survey flows smoothly from traditional machine learning based techniques towards the recent deep learning based models. Further, we explain basic steps and conclude the generic flow of VD methods, supported by baseline references for further guidance of readers. Our summarized contributions can be enumerated as follows:
\begin{enumerate}
\item Detailed analysis of existing VD literature, discussion on representative methods from classified VD categories, as given in Figure~\ref{fig:structureofsurvey}.
\item Explanation of advantages and drawbacks of various VD methods based on their utilized features.
\item Performance evaluation of representative VD methods \textcolor{black}{and discussion about how and why some methods achieved higher accuracy with fewer parameters.}
\item Comprehensive future insights for interested readers and domain experts. The future directions are derived based on the analyzed literature and supported by appropriate references.
\end{enumerate}
\textcolor{black}{The VD techniques that broadly falls under the scope of video classification are introduced in this section. The major contributions of our survey are highlighted in bulleted form.} The rest of the paper has 4 sections. Section 2 corresponds to the detailed explanation of the current achievements in VD domain. In Section 3, we best explain the available datasets and discuss the performance evaluation of different methods on these datasets. In Section 4, we highlight the current challenges in VD domain, each challenge followed by its corresponding future research guidelines. Finally, we conclude our article in Section 5 by explaining the derivation of this research.
\begin{figure*}[!h]
\centering
\includegraphics[width=\textwidth]{Structureofsurvey.png}
\caption{The overall structure of the proposed survey.}
\label{fig:structureofsurvey}
\end{figure*}
\begin{figure*}
\centering
\resizebox{\columnwidth}{!}{
\begin{forest}
for tree={
align=center,
parent anchor=south,
child anchor=north,
font=\sffamily,
edge={thick, -{Stealth[]}},
l sep+=10pt,
edge path={
\noexpand\path [draw, \forestoption{edge}] (!u.parent anchor) -- +(0,-10pt) -| (.child anchor)\forestoption{edge label};
},
if level=0{
inner xsep=0pt,
tikz={\draw [thick] (.south east) -- (.south west);}
}{}
}
[Violence Detection Literature
[Traditional\\Methods
[Machine\\Learning\\~\cite{hussain2020multi}
[SVM\\~\cite{gao2016violence,lohithashva2020violent,dhiman2017high,mabrouk2017spatio}]
[Others\\~\cite{mishra2018automated,mabrouk2017spatio}]]
[Handcrafted\\Features\\~\cite{chen2011violence,xu2014violent}
[Motion Patterns\\and Optical Flow\\~\cite{chen2011violence,gracia2015fast,datta2002person,mahadevan2010anomaly,chen2008recognition,zhang2016new,febin2020violence}]
[Mathematical\\Modeling\\~\cite{nguyen2005learning}]
]
]
[Deep Learning\\Models
[Temporal\\Learning\\~\cite{ding2014violence,ullah2019violence,li2019efficient}]
[Spatio-temporal\\Models\\~\cite{traore2020violence,ullah2020cnn,ullah2021intelligent}\\~\cite{ullah2021ai,chang2022hybrid,qiu2017learning,ullah2021intelligent}\\~\cite{vosta2022cnn,vosta2022cnn,peixoto2019toward,traore2020violence,vijeikis2022efficient}\\]
[Others\\(2D CNN etc.)\\~\cite{mu2016violent,meng2017trajectory,xia2018real}\\~\cite{freire2022inflated,naik2022automated,mumtaz2022fast,mumtaz2018violence}\\~\cite{hanson2018bidirectional,hussain2022real}]
]
]
]
\end{forest}
}
\caption{A comprehensive form of overall VD literature.}
\label{fig:VDLiterature}
\end{figure*}
\begin{figure*}[!h]
\centering
\includegraphics[width=\textwidth]{Prismamodel.png}
\caption{The overall process of research articles retrieval from different databases.}
\label{fig:prismamodel}
\end{figure*}
\textcolor{black}{
\section{Review Methodology}
The VD domain using video data for surveillance applications is in action since 2002, when Datta~\latinphrase{et~al.}\xspace~\cite{datta2002person} presented their first person-to-person VD approach. Afterwards, in the era between 2002 and 2013, mainstream techniques are based handcrafted features and completely rely on the domain knowledge to constitute low-level features from the input videos. With the advancements in deep learning models, the very first approach utilizing 3D ConvNets was introduced by Ding~\latinphrase{et~al.}\xspace~\cite{ding2014violence}, that works independently without any prior knowledge. Following VD using 3D ConvNets, many researchers introduced deep learning-based techniques, that show significantly better results when compared to the hand-engineered features methods. }
\textcolor{black}{
In this review, the baseline research contributions in VD domain are discussed, leading to the current state-of-the-art approaches and advanced deep models. The research articles are retrieved from several famous databases such as \emph{Web of Science} and \emph{Google Scholar}. The overall PRISMA flow diagram for the research articles retrieval process is visualized in Figure~\ref{fig:prismamodel}. Overall there are lots of articles retrieved during the initial search, that are first scrutinized using the title. For instance, the research contributions mentioning non-surveillance domains violence are not included in our review. Similarly, lots of low-level features-based articles are available in the early VD literature, but the ones with comparatively higher number of citations are discussed in the proposed survey.}
\section{Current Achievements in VD Domain} \label{sec:Relatedworks}
The growth and technological advancements in the field of computer vision has widely impacted the surveillance applications~\cite{dehingia2022help} positively, providing various services such as continuous video data monitoring using VD and reporting. There are a lot of existing techniques focusing on these services, that are broadly categorized into handcrafted (conventionally known as low-level features and machine learning) and deep learning-based methods, discussed separately below and given hierarchically in Figure~\ref{fig:VDLiterature}. \textcolor{black}{
The manuscript retrieves contributions based on popularity amongst computer vision and visual perception experts. Each piece of research is reviewed for relevance after retrieval from renowned research repositories such as \emph{Web of Science}, where we used different keywords such as \emph{violence detection in video data, video violence detection, violence detection,~\latinphrase{etc.}\xspace}. The literature review in this proposal focuses on papers that have been published between 2002 and the present date. It is notable that deep learning results are more prevalent during the retrieval procedure. Finally, the relevance of each research is checked~\latinphrase{i.e.}\xspace using the mentioned technique's strategy, input horizon, the intermediate procedure, and finally the output results. For instance, research articles focused on VD using sensors other than vision are not considered in the proposed survey. Similarly, hybrid techniques considering various sensory data are also excluded from discussion in this research.}
\begin{table*}[ht]
\centering
\caption{Traditional ML representative VD methods selected based on their popularity among VD experts.}
\resizebox{\textwidth}{!}{
\begin{tabular}{m{1.5cm}||m{4.5cm}m{3cm}m{3cm}m{3cm}}
\toprule
\textbf{Ref.} & \textbf{Strategy} & \textbf{Horizon} & \textbf{Dataset$/$Availability} & \textbf{Remarks}
\\ \midrule
Chen \latinphrase{et~al.}\xspace \cite{chen2011violence} & Violence classification using SVM & Movies data & Movies used in experimentation$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & In traditional machine learning mechanisms, mainstream researchers utilize the potentials of SVM to detect violence in surveillance videos. \\ \midrule
Hassner \latinphrase{et~al.}\xspace \cite{hassner2012violent} & VIolent Flows (ViF) descriptor are used to classify violent or nonviolent using linear SVM & Surveillance & Action Similarity Labeling Challenge (ASLAN) and Hockey dataset$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & This is a challenging dataset against other datasets due to data diversity and reduced amount of videos present in this dataset. \\ \midrule
Deniz \latinphrase{et~al.}\xspace \cite{deniz2014fast} & Spatio-temporal features and acceleration computation for VD & Surveillance and sports & UCF101, Hockey Fight, and Movies data$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Fast VD method \\ \midrule
Gracia \latinphrase{et~al.}\xspace \cite{gracia2015fast} & Motion blob and acceleration measure vector method is used for fight detection & Surveillance crowd & UCF-101, Hockey Fight, and Movies$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Fast computation over the cost of accuracy. \\ \midrule
Gao \latinphrase{et~al.}\xspace \cite{gao2016violence} & Motion information with AdaBoost algorithm for features selection and linear SVM for classification & Surveillance & Crowd violence and hockey fight$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Hybrid motion features with AdaBoost and Linear SVM performs well against SOTA. \\ \midrule
Bilinski \latinphrase{et~al.}\xspace \cite{bilinski2016human} & Proposed improved fisher vector to recognized violence in videos & Surveillance & Violent-Flows dataset, Hockey Fight dataset and Movies dataset$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Improved efficiency with violence detection and recognition potentials. \\ \midrule
Li \latinphrase{et~al.}\xspace \cite{li2016detecting} & Proposed Multi-modal features model by exploiting subclass visualization to manually labeling violence in videos & Surveillance & MediaEval 2015 violence dataset$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Diverse set of sub-classes video annotation and classification over MediaEval 2015 dataset \latinphrase{i.e.}\xspace extending this dataset. \\ \midrule
Dhiman \latinphrase{et~al.}\xspace \cite{dhiman2017high} & Proposed integrated features-based on average energy images framework to recognize abnormal activities & Surveillance & R fall detection and KARD dataset$/\times$ & Hybrid features from Kinect and visual sensor are reduced using PCA and classified via SVM. \\ \midrule
Malek \latinphrase{et~al.}\xspace \cite{al2017novel} & Temporal differencing algorithm and motion regions are located using gaussian function and SVM for normal and abnormal activities recognition & Surveillance & Private students recorded data$ /\times$& After VD, retrieval of the detected objects from database for detailed analysis. \\ \midrule
Mishra \latinphrase{et~al.}\xspace \cite{mishra2018automated} & Bag of words representation technique is used for detection and classification of art material videos & Surveillance & KTH$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ and proposed Judo and Taekwondo$/\times$ & Simple martial arts videos classification using statistical machine learning. \\ \midrule
Song \latinphrase{et~al.}\xspace \cite{song2018multi} & Single temporal and multi-frameworks based on learning or semantic modeling technique for high level activity analysis & Surveillance & BEHAVE dataset $/\times$, \newline YouTube data $/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$, and \newline NUS–HGA dataset$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Graph trees are used in the framework and testing on real-world YouTube videos show promising generalization abilities of this method. \\ \midrule
Mabrouk \latinphrase{et~al.}\xspace \cite{mabrouk2017spatio} & Crowd$/$non-crowd violence detection using covariance matrix and spatio-temporal features. Optical flow of HOG features & Crowded $/$ Non-crowded areas VD & Hockey fight and violent flows dataset $/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & In this method an SVM classifier is trained to classify diverse data, yielding poorly generalized model, showing 88.6 $/$ \% accuracy for hockey fight dataset \\ \midrule
Lohithashva \latinphrase{et~al.}\xspace \cite{lohithashva2020violent} & LBP and GLCM features with supervised classifiers for VD & Surveillance and movies & Hockey fight and violent-flow $/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Low-level features show reduced performance and limited generalization abilities. \\ \midrule
\end{tabular}
}
\label{tab:MLforVD}
\end{table*}
\subsection{Machine Learning Techniques for VD}
These are baseline and early VD methods~\cite{de2010violence}, considering purely image processing and pattern recognition methods to predict violence and abnormal patterns using either filtering techniques motion information. This group of study is effective for simple VD problems with only abnormal motion patterns, where limited to work in situation with instant normal motion and complex real-world situation. In low-level features-based approaches, authors generally use human trajectories and motion information from visual data \cite{chen2011violence}. In motion information, most of the methods use optical flow or its several invariants. In supervised learning-based traditional mechanisms, mainstream researchers utilize the potentials of Support Vector Machine (SVM). Some of the research in VD domain is based on multi-modalities such as aural and video data are employed to detect violence in surveillance videos. In the subsequent paragraphs, these methods are discussed briefly.
The human violence refers to fist fighting, hitting other objects, and many other similar actions that occur in surveillance videos. To detect such activities automatically, a person-on-person violence detection method is introduced in \cite{datta2002person}, that is functional for video data. The authors utilized motion trajectories information along with the limb’s orientations information in their method to detect violence effectively. There is an acceleration measure vector in their method which contains information about the magnitude of motion and jerk is defined here using temporal derivative of the acceleration motion vector. This method is completely based on low-level features \latinphrase{i.e.}\xspace background subtraction and acceleration measure vector computation.
Continuing research using low-level features, Nguyen \latinphrase{et~al.}\xspace \cite{nguyen2005learning} used hidden Markov model for activities recognition that may contain violence. Their main novelty and contribution lie in construction of a particle filter with an approximation inference scheme that computes the distributions of hidden Markov model at constant time to receive new observation and data, making the system functional in real-time.
Heading onward towards more sophisticated and reliable anomaly and violence detection methods \cite{mahadevan2010anomaly}, Mahadevan \latinphrase{et~al.}\xspace introduced a novel framework for anomaly detection in crowded scenes~\cite{naik2022automated}, that is a very challenging task to achieve. Their research concludes that a set of three properties acts as the main constituent of any anomaly detection system. These properties include integrated modeling of dynamics and appearance of the ongoing scene in an input video, temporal, and spatial abnormalities. Their model is completely based on phenomenon that the normal behavior has a specific pattern and contains a mixture of predicted, but at the same time dynamic texture. While detecting any outlier in this pattern or dynamics indicates anomaly.
In continuation with handcrafted features-based methods, motion vectors are widely used as well in the early related literature. As Chen \latinphrase{et~al.}\xspace utilized human behavior using binary location motion patters and decide about an aggressive behavior \cite{chen2008recognition}. Similarly, optical flow-based methods perform comparatively better. For instance, in a baseline research \cite{zhang2016new}, the authors used Gaussian Model of Optical Flow to effectively extract candidate violence regions, that are adaptively modeled as a deviation from any normal behavior being observed in the current scene. These candidate violence regions are then sampled to generate or detect violence.
A deep study of the aforementioned methods reveals their several drawbacks, which limits their adaptation and deployment in real-world surveillance scenarios. Most of these methods are designed to function for static image data, that is the foremost concern of today’s smart surveillance, where actions and events occur in a sequence of images. Therefore, these methods do not result in trusty-worthy decisions due to their spatial image processing. Another major concern related to these methods is their features representation approaches, which are very confined and are extremely low-level to be trusted for smart surveillance scenarios. In advanced machine learning based methods, the SVM~\latinphrase{etc.}\xspace perform well for classifications and activity recognition tasks~\cite{hussain2020multi}, but when the features are extracted using handcrafted techniques, the output results are poorly generated. The traditional methods are comprehensively tabulated in Table \ref{tab:MLforVD}, where we also indicate the used dataset of each method. To overcome the limitations mentioned above, deep learning performs very well, and deep learning-based approaches are discussed in the subsequent section.
\subsection{Deep Learning for VD}
The emergence of deep learning models in several computer vision domains \textcolor{black}{motivated researchers to apply it} for enormous surveillance applications including violence detection etc. Video analysis is a challenging problem due to the involvement of complex image data and its huge processing for useful information extraction. \textcolor{black}{Video analysis comprises several major domains, where primarily it is composed of video classification~\cite{wang2018appearance,tran2019video,li2020smallbignet}, segmentation~\cite{gao2022deep,wang2021survey}, among many others. Video classification~\cite{yue2015beyond} is an interesting domain with main focus on effective temporal contents representation that significantly contributes to precise video label prediction. Although the early approaches in video classification are based on simple CNNs~\cite{karpathy2014large}, but the recent methods employ various temporal~\cite{carreira2017quo} and spatio-temporal~\cite{hussain2022vision} strategies for video classification. Video classification is further divided into several major domains such as activity recognition, anomaly detection and recognition, and violence detection and recognition.}
Violence occurs in sequence of frames, falling under the broad domain of \textcolor{black}{video} classification. Sequential frames data, compared to single image data is much more complex due to involvement of time series analysis and its patterns and structure learning for final output generation. Motion features are also used in sequential form such as in bag of words, as presented in an existing research~\cite{fu2017automatic}. The main strategy used in this system is to analyze the motion to detection fight events. Although, motion features involve sequential patterns processing, but still motion features seem to perform poorly, as huge motion in sequential frames do not necessarily guarantee abnormal events i.e., it can be normal patterns. Otherwise, anomaly or violence may occur with slight motion, therefore limiting its usage.
Towards violence detection, deep learning-based methods~\cite{roman2020violence} have completely replaced the traditional working of hand-crafted features-based algorithms with convincingly precise output results. As mentioned earlier, violence occurs in sequence of frames, therefore, many researchers use spatial as well as temporal features to generate optimal results. Spatial features are concerned with the frames-level data, while temporal features works best for sequence analysis. In activity classification domain, the concept of hybrid features, i.e., integrating spatial and temporal features is very common. For instance, Amira Ben and Zagrouba~\cite{mabrouk2017spatio} employed spatio-temporal optical flow in their research to detect violence. Their method estimates the magnitude and orientation of optical flow, followed by its descriptor and both these (estimator and descriptor) are used to train an SVM classifier, showing 88.6\% accuracy for hockey fight dataset. This method is based on traditional machine learning, yielding poorly generalized model, as SVM is not a good option to classify diverse data.
In a follow up research~\cite{ullah2019violence}, C3D features are used to train a deep learning model for violence detection. Deep learning features are used as intermediate representation of sequential information, followed by SoftMax classifier to extract the probability of the class having maximum chances of occurrence. This research shows higher computational complexity and is not real-time, as reported by authors in their experiments as 25 FPS. In another recent research, authors used deep learning features with long short-term memory (LSTM) for anomaly detection~\cite{ullah2020cnn}. The spatio features representation in this method is achieved using the famous ResNet-50 model, where the authors extracted already learned features from ResNet-50 pretrained model in their method. LSTM is used in this method for sequential learning, where the learned features are stacked together to form a sequence of 15 frames, that are passed to bi-directional LSTM for anomalous activities patterns learning.
The aforementioned methods tend to tackle the challenges of hand crafted-features based violence detection systems, where they perform much better to classify violent activities. But the time complexity of these models restrict their practical deployment in real-life surveillance scenarios, such as implementation in smart cities surveillance etc. In addition, mainstream deep models provide only violence detection or classification services, demanding human experts for its analysis and reporting ultimately to the concerned departments.
\begin{table*}
\centering
\caption{Deep learning representative VD methods selected based on their popularity among VD experts.}
\resizebox{\textwidth}{!}{
\begin{tabular}{m{1.7cm}||m{2.8cm}m{3cm}m{3cm}m{4.5cm}}
\toprule
\textbf{Ref.} & \textbf{Strategy} & \textbf{Horizon} & \textbf{Dataset$/$Availability} & \textbf{Limitations}
\\ \midrule
Ding~\latinphrase{et~al.}\xspace~\cite{ding2014violence} & 3D convolution with back propagation strategy & Sports violence, applicable to surveillance & Hockey fight$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Deep 3D model to directly process raw inputs for features extraction without any prior knowledge. \\ \midrule
Meng~\latinphrase{et~al.}\xspace~\cite{meng2017trajectory} & Integrating frame of trajectory and Deep CNN to detecting human violent behavior in videos & Surveillance and movies & Hockey Fights and Crowd Violence dataset$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Utilizing hand-crafted and deep features in a unified framework, yielding supremacy over SOTA using Hockey Fight and Crowd Violence dataset but lower generalization potentials. \\ \midrule
Mu~\latinphrase{et~al.}\xspace~\cite{mu2016violent} & CNN are used to detect violent detection based on acoustic information & Movies & MediaEval 2015$/\times$& This framework has limited potentials for real-world surveillance due to the modalities used as input. \\ \midrule
Xia~\latinphrase{et~al.}\xspace~\cite{xia2018real} & Bi-channels CNN with SVM are used for violence detection & Surveillance & Hockey fight, violent flow$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Real-time performance (30 (fps) using appearance and motion features, fed to SVM for classification, where SVM can be replaced with more robust sequential deep model. \\ \midrule
Serrano~\latinphrase{et~al.}\xspace~\cite{serrano2018fight} & Proposed hybrid features based framework for violent behavior & Surveillance and movies & Hockey, Movies and Behave dataset$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Hybrid (handcrafted$/$learned features) with computational efficiency for VD task. The authors oppose the idea of only using motion features directly. \\ \midrule
Ullah~\latinphrase{et~al.}\xspace~\cite{ullah2019violence} & 3D CNNs with Softmax classifier for VD & Surveillance & Hockey fight, violence in movies, and violent crowd datasets.$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & Almost real-time method, has higher accuracy but not tested on surveillance-specific data. \\ \midrule
Li~\latinphrase{et~al.}\xspace~\cite{li2019efficient} & Proposed 3D convolutional neural networks exclusively for encoding temporal information & Surveillance and movies & Hockey Fights, Movies Dataset and and VIolent-Flows Dataset $/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ & The internal design comprises bottleneck units for motion patterns with DenseNet structure for channel's interaction, thereby capturing spatial and temporal information effectively but highly computational. \\ \midrule
Traore~\latinphrase{et~al.}\xspace~\cite{traore2020violence} & RNNs with 2D CNN for VD & Surveillance & Hockey fight, violent flow$/\tikz\fill[scale=0.3](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;$ \newline Real-life violence situations datasets$/\times$ & Optical flow along with RNN for temporal information and 2D CNN for spatial information produces effective results, where the authors achieved best results over three benchmark datasets yet the framework is not tested on real-world videos. \\ \midrule
\end{tabular}
}
\label{tab:DLforVD}
\end{table*}
\subsubsection{Spatial deep models for VD}
Spatial deep models are those utilizing frame-level features to classify activities into violent or non-violent. Spatial methods using deep learning are rare, as VD patterns occur in sequence of frames, therefore, most of the existing research is based on sequential features.
Working with spatial data is not always effective for VD task~\cite{khan2019cover}, as proved from experimental results of Sumon \latinphrase{et~al.}\xspace~\cite{sumon2020violence}. The authors experimented on various deep learning models including VGG-19 and ResNet using transfer learning strategy to extract features from the last layer followed by fully connected layer to from a VD classification model. Their results reported significant training accuracy but the testing phase indicated the poor generalization performance of such models, questioning their practical applicability.
\subsubsection{Deep sequence learning for VD}
Deep sequence learning approaches are comparatively robust towards diverse type of environments, holding representative features extraction potentials~\cite{accattoli2020violence}. One of the main reasons of their better generalization abilities is that violent activities occurs in sequence of frames, and these techniques process a single sequence while classifying violent activity. Thus, better representation is achieved for each sequence due to abrupt patterns easily recognizable by deep model when compared with the normal sequence of frames.
Working in sequence learning direction, the existing literature has either considered deep features of a sequence of frames with LSTM~\cite{abdali2019robust} or employed 3D CNNs~\cite{ullah2019violence} to process a sequence of predefined number of frames. For instance, a former research has considered~\cite{sudhakaran2017learning} frame-level features followed by an LSTM with convolutional gates for aggregation. The authors claimed to represent motion patterns effectively using this strategy, that being proved from experimental results, where they received convincing accuracy. LSTM is widely used in such cases of sequence classification~\cite{hussain2019cloud} in many computer vision tasks, where the preprocessing features varies from method to method. For instance, Fenil~\latinphrase{et~al.}\xspace extracted local HOG features and fed them to a bi-directional LSTM to retrieve information in both forward and reverse directions~\cite{fenil2019real}.
\color{black}
The current achievements in VD domain were discussed in this section, where it is evident from existing literature that mainstream state-of-the-art techniques are based on deep learning models. Currently, the spatio-temporal methods perform best in classifying violent activities, otherwise the 3D ConvNets-based methods have higher number of parameters yet lowered performance. The 3D ConvNets mainly focus on the temporal information and learning but limited spatial representation reduces their performance.
\color{black}
\section{Datasets and Performance Evaluation of VD Techniques}
The datasets particularly relating VD domain are very scarce in existing state-of-the-art literature. Mainstream datasets are either extracted from several sports or movies, hardly resembling real-world surveillance scenarios. Although the patterns relate fight, but still the camera motion and other similar parameters affect a model's performance whenever employed for real-world monitoring. The most challenging datasets so far in VD domain is Violent-Flows and Real-World Fighting (RWF), that contain real-world video footage of violence in crowd, recorded using surveillance low-resolution cameras. The overall datasets are given in Figure \ref{fig:VDDatasets}.
\begin{figure}
\centering
\smartdiagram[constellation diagram]{VD \\ Datasets, Hockey Fight \cite{nievas2011violence}, Violent Flows$/$Violent Crowd \cite{hassner2012violent}, Violence in \\ Movies \cite{nievas2011violence}, RWF-2000 \cite{cheng2021rwf}, \textbf{UCF Crime \cite{sultani2018real}}, \textbf{UT-Interaction \cite{ryoo09}}, \textbf{Two-person interaction \cite{yun2012two}}}
\caption{The available VD datasets. Datasets written in bold letters indicate that these are not purely VD datsets, but contain some classes with human violence and can be utilized in VD domain.}
\label{fig:VDDatasets}
\end{figure}
\subsection{Violence in Movies}
As indicated by the name, this dataset is retrieved from movies with fight scenes and actions \cite{nievas2011violence}. This is comparatively smaller in volume, containing only 200 video clips of size 360 by 250 pixels. The first-person motion in this dataset is negligible or even no motion at all in some scenarios, making it very easy to detect the patterns with abnormal behavior i.e., the ones with fight.
\subsection{Violent Crowd$/$Violent Flows}
This dataset is a challenging one as it comprises several categories of violence such as violence in sports and in crowds \cite{hassner2012violent}. The videos are downloaded from YouTube and the overall dataset consists of five sets of video clips, while each one has two classes i.e., violent and non-violent. This dataset is comparatively challenging against other datasets due to data diversity and the reduced amount of videos present in this dataset.
\subsection{Hockey Fight}
This is a very common and widely used dataset for violence detection by most of the methods in VD related literature. The dataset was proposed by Nievas \latinphrase{et~al.}\xspace \cite{nievas2011violence}, containing 1000 short video clips. The videos presented in this dataset were previously record in National Hockey League. The dataset is evenly balanced with 500 videos categorized as fight and the remaining 500 videos are placed in non-fight category i.e., normal videos. Each of the video clip contains round about 50 frames and the resolution of these videos is constant, that is 360 by 288. As mentioned, the dataset is recorded in hockey ground, indicating the nature of videos as indoor. This dataset is not much challenging as the videos have very simple activities patterns, therefore, most of the recent methods with deep learning strategies have achieved round about 98\% accuracies on this dataset.
The accuracy of representative VD methods over Hockey fight dataset are given in Figure \ref{fig:hockeyfightevaluation}. Since the nature of the dataset is very simple \latinphrase{i.e.}\xspace binary classification problem, therefore, many methods have scored significantly higher accuracy.
\begin{figure}
\centering
\begin{tikzpicture}
\begin{axis}
[ybar,
width=0.9\columnwidth,
ylabel={\ Accuracy (\%)},
xlabel={\ Methods},
enlarge x limits = 0.1,
xtick=data,
x tick label style={rotate=45,anchor=east},
nodes near coords,
nodes near coords align={vertical},
symbolic x coords=
{
Hassner et al.,
Ding et al.,
Bilinski et al.,
Mabrouk et al.,
Sudhakaran et al.,
Xia et al.,
Ullah et al. (a),
Li et al.,
Lohithashva et al.,
Traoré et al.,
Ullah et al. (b),
Ullah et al. (c),
Freire et al.,
},
nodes near coords,
every node near coord/.append style={rotate=90, anchor=west}
]
\addplot [draw=black, line width=.2mm, semithick, pattern = north east lines, pattern color = black] coordinates
{
(Hassner et al., 58.2)
(Ding et al., 91.0)
(Bilinski et al., 93.4)
(Mabrouk et al., 88.6)
(Sudhakaran et al., 97.1)
(Xia et al., 95.9)
(Ullah et al. (a), 96.0)
(Li et al., 98.3)
(Lohithashva et al., 91.5)
(Traoré et al., 96.5)
(Ullah et al. (b), 98.5)
(Ullah et al. (c), 98.2)
(Freire et al., 99.4)
};
\end{axis}
\end{tikzpicture}
\caption{Performance evaluation of representative VD methods over Hockey fight dataset. The methods are sorted year-wise, starting left-side from 2012 and so on. The references are given as: Hassner et al.~\cite{hassner2012violent},
Ding et al.~\cite{ding2014violence},
Bilinski et al.~\cite{bilinski2016human},
Mabrouk et al.~\cite{mabrouk2017spatio},
Sudhakaran et al.~\cite{sudhakaran2017learning},
Xia et al.~\cite{xia2018real},
Ullah et al. (a)~\cite{ullah2019violence},
Li et al.~\cite{li2019efficient},
Lohithashva et al.~\cite{lohithashva2020violent},
Traoré et al.~\cite{traore2020violence}, Ullah et al. (b)~\cite{ullah2021ai}, Ullah et al. (c)~\cite{ullah2021intelligent}, Freire~\latinphrase{et~al.}\xspace~\cite{freire2022inflated}.}
\label{fig:hockeyfightevaluation}
\end{figure}
\subsection{RWF-2000}
The contents of mainstream existing VD datasets are prepared in some controlled environments, where the quality and visibility of the actors are distinct. However, violence activities in real-world surveillance pose various resolution, small scale, and different viewpoints. Therefore, to make the violence more authentic and realistic RWF dataset is prepared from the YouTube repository, particularly, surveillance CCTV cameras. It consists of 2000 video clips with five seconds duration and variable resolutions. It has two classes \latinphrase{i.e.}\xspace violent and non-violent activity. This dataset is very challenging because some videos are in poor imaging quality due to dark environment, fast moving actors, and lighting blur. Performance evaluation of representative methods such as ConvLSTM~\cite{tran2015learning}, C3D model~\cite{sudhakaran2017learning}, and I3D architecture~\cite{carreira2017quo} is given in Figure~\ref{fig:rwfevaluation}.
\begin{figure}
\centering
\begin{tikzpicture}
\begin{axis}
[ybar,
width=0.9\columnwidth,
ymajorgrids=true,
xmajorgrids=true,
ylabel={\ Accuracy (\%)$/$Parameters(M)},
xlabel={\ Methods},
enlarge x limits = 0.1,
xtick=data,
legend style={at={(0.82,0.70)},anchor=north},
x tick label style={rotate=45,anchor=east},
symbolic x coords=
{
Tran et al.,
Sudhakaran et al.,
Carreira et al. (a),
Carreira et al. (b),
Carreira et al. (c),
Cheng et al.
},,
nodes near coords,
every node near coord/.append style={rotate=90, anchor=west}
]
\addplot
[draw=black, line width=.2mm, semithick, pattern = vertical lines, pattern color = black]
coordinates
{
(Tran et al.,82.75)
(Sudhakaran et al.,77.0)
(Carreira et al. (a),85.75)
(Carreira et al. (b),75.50)
(Carreira et al. (c),81.50)
(Cheng et al.,87.25)
};
\addplot
[draw=black, line width=.2mm, semithick, pattern = grid, pattern color = black]
coordinates
{
(Tran et al.,94.8)
(Sudhakaran et al.,94.8)
(Carreira et al. (a),12.3)
(Carreira et al. (b),12.3)
(Carreira et al. (c),24.6)
(Cheng et al.,0.27)
};
\legend {Accuracy,Parameters(M)}
\end{axis}
\end{tikzpicture}
\caption{Performance evaluation of representative VD methods over RWF-2000 dataset, where results are reported based on findings of a prior research \cite{cheng2021rwf}. Carreira~\latinphrase{et~al.}\xspace \cite{carreira2017quo} refers to I3D features-based VD, where (a) refers to only RGB data, (b) indicates only optical flow, and (c) represents two stream deep model \latinphrase{i.e.}\xspace combining RGB and optical flow. The research methods are presented by Tran~\latinphrase{et~al.}\xspace~\cite{tran2015learning},
Sudhakaran~\latinphrase{et~al.}\xspace~\cite{sudhakaran2017learning},
Carreira~\latinphrase{et~al.}\xspace~\cite{carreira2017quo}, and
Cheng~\latinphrase{et~al.}\xspace~\cite{cheng2021rwf}, respectively.}
\label{fig:rwfevaluation}
\end{figure}
\color{black}
\subsection{Results Discussion}
VD approaches in the early days depended on low-level hand-crafted features such as appearance, histograms of oriented gradients, and different types of motion measurement techniques, such as optical flow. In contrast, complex violent patterns that deviate from normal motion are quite hard to detect using these low-level features. That is why traditional low-level features-based techniques are comparatively less effective on simple datasets such as Hockey fight, which have conventional low-level features-based techniques. The motion patterns in Hockey fight dataset are not very complex, yet the best traditional features-based methods such MoSIFT+HIK~\cite{nievas2011violence}, ViF~\cite{hassner2012violent}, and MoSIFT+KDE+Sparse Coding~\cite{xu2014violent} techniques scored 90.9\%, 82.9\%, and 94.3\% accuracy, respectively. Similarly, on Violent-Flows dataset, the best hand-crafted features methods~\latinphrase{i.e.}\xspace ViF~\cite{hassner2012violent}, MoSIFT+KDE+Sparse Coding~\cite{xu2014violent}, and ViF+OViF~\cite{gao2016violence} achieved 81.3\%, 89.05\%, and 88\% accuracy, respectively. Although the mentioned accuracy and performance is higher, but still the deployment of these methods in real-world environments is questionable. As the real-world actions and events are complex as well as their patterns are significantly varying from those of the Hockey fight dataset.
In contrast, the performance of VD techniques emerged significantly with the usage of deep sequential learning algorithms. The emerging recurrent neural networks~\cite{zaremba2014recurrent} are widely adopted in many video classification tasks~\cite{ullah2021conflux,ramanujam2021human} due to their long-range accurate temporal learning abilities. The fact that temporal dependencies as well as spatial information play an important role in sequential patterns classification is evident from Figure~\ref{fig:hockeyfightevaluation} and~\ref{fig:rwfevaluation}. These figures show that the best performance is achieved by deep models containing spatial and temporal learning potentials. For instance, Ullah~\latinphrase{et~al.}\xspace~\cite{ullah2021ai} proposed convolutional LSTM followed by GRU and fully connected layers to classify violent and non-violent classes. The authors' motivation of choosing convolutional LSTM with GRU is to generate effective spatial representation of each frame in video using CNNs and encode the temporal changes and learn long-range dependencies using recurrent neural networks.
In terms of time complexity and the number of parameters, mainstream deep learning-based VD models share limited information~\cite{ullah2019violence} in their research, but the architecture mostly suggest very high number of parameters. For instance, multi-stream convolutional and optical flow networks operations along with multi-layers of LSTMs are highly computational but their details are skipped in the relevant research article~\cite{ullah2021intelligent}. VD experts in computer vision domain mostly focus on the accurate predictions of their method, rather pointing out the real-time decision making of a technique. The number of parameters for different deep models can be analyzed from Figure~\ref{fig:rwfevaluation}, where most of the deep models focus on the accuracy of their method. The number of parameters are mostly high, except some representative methods that are able to provide a balanced trade-off between the number of parameters and the accuracy of the model. For instance, Carreira~\latinphrase{et~al.}\xspace~\cite{carreira2017quo} I3D features-based mechanism to compute VD has 12.3 million parameters when computed using only RGB or optical flow input, while their two stream model has two times parameters (24.6 million) compared to the individual streams. The best single stream accuracy is 85.75\% for RGB input, while the two stream architecture's accuracy is 81.5 and the optical flow only mechanism produced the lowest 75.5\% accuracy. Although these methods have comparatively lower parameters against state-of-the-art, as given in Figure~\ref{fig:rwfevaluation}. The best achievement and trade-off between parameters and accuracy is provided by Cheng~\latinphrase{et~al.}\xspace~\cite{cheng2021rwf}, where the authors achieve 87.25\% accuracy with only 0.27 million parameters. They achieved reduced parameters by employing the concept of depth-wise separable convolutions from Pseudo-3D Residual Networks~\cite{qiu2017learning} and MobileNet~\cite{howard2017mobilenets} to adopt the 3D convolutional layers in their proposed model, which tremendously reduced their model parameters without compromising on the performance loss.
\color{black}
\section{Challenges and Future Directions in VD Domain}
Action and activity recognition is one of the most interesting and hot area of research, covering several major domains. One of these major domains is VD, that seems not to be fully covered to the level of its requirements in today's smart surveillance systems. For instance, there have been extremely challenging datasets available for action and activity recognition, while in contrast, there are only few for VD domain. Similarly, there are many interesting breakthroughs in the field of activity recognition such as future activity perceptions, activities localization, multi-view action and activity recognition. VD domain is being deprived of such breakthroughs, as it lacks multi-view challenging datasets recorded in real-world scenarios such as UCF anomaly detection~\cite{roka2022review} and recognition dataset \cite{sultani2018real}. Similarly, the accuracy of the current methods is so perfect on the available datasets but these data does not seem to be representative of real-world because these data are retrieved or recorded from sports videos. The major challenges in VD domain are covered in the subsequent paragraphs along with possible future directions.
\subsection{Scarcity of Challenging VD Data}
The video data available for use in VD methods is very limited \latinphrase{i.e.}\xspace only few datasets are there to be used for comparison with SOTA methods. Practical applicability of action and activity recognition methods highly depends on the used data while training. For instance, diverse set of videos while training a deep model or learning motion patterns of huge set of violent videos results in highly generic trained weights or a mathematical model for practical usage in real-world scenarios.
\subsection{Evaluation of VD Methods}
So far, confusion matrix, accuracy, and F1 metrics are commonly used to evaluate VD model or compare its performance against SOTA methods. Current VD trend is only focused on accuracy of any model, ignoring the worth of time complexity for resource-restricted environments. Furthermore, the evaluation is mostly performed using test set of the same dataset, that is a biased decision until the test set is marginally different from the training set, evaluating a model's generalization.
\subsection{Future VD Prediction}
Future activity prediction has gained certain importance recently with many applications to autonomous driving for human location and activity perception. The idea of future activities prediction is also required in VD domain, predicting future violent actions with significant accuracy will lead to prevention of the receipted action or at least quick responsive actions initiation.
\subsection{Generalization Potentials of VD Methods}
Generalization abilities of any VD method has direct relation with its applicability, the more is a model generic and performs best towards unseen data, the greater are the chances to detect violence in early stages. Such type of task can be achieved by specially designing a neural network to function completely or consider the importance of certain parameters including human locations, their distance among each other, and their motion trajectories. Such parameters have importance to detect violence and helps in distinguishing abnormal violence patterns form real-time normal video frames. Existing research on anomaly detection by Ullah \latinphrase{et~al.}\xspace can be a good source to integrate various importance features and gain enhanced and generalized model \cite{ullah2021cnn}. Similar research is applied to activity recognition domain \cite{ullah2018activity}.
\subsection{Relating Anomaly Detection with VD}
Anomaly detection domain has broader number of classes such as fire, arson, shoplifting, and others including human violence, comprising assault and fighting as well \cite{ullah2021efficient}. So far, VD literature is based on just binary classification problem \latinphrase{i.e.}\xspace detecting whether there is a human fight or not, which is a simple problem for today's computer vision algorithms. Complex problems such as further classifying the type of fight and its intensity level are future directions yet to be covered, demanding experts' attention. In this regard, end-to-end deep learning models can be employed to classify any abnormal fight action and relate its intensity by using the same models' parameters.
\subsection{Edge-based VD Methods}
To ensure the safety of people, eradicate or reduce the risk of the violence, it is necessary to equip the vision sensors in smart cities with highly computational embedded devices. Besides the embedded devices, the vision sensors themselves can be made smart enough to process video data effectively for any kind of prior violence acts detection or violence detection while it's happening.
It is very challenging task to achieve effective VD over the edge networks due to higher computational cost of mainstream deep learning CNN \cite{suleiman2017towards} models \cite{lygouras2019unsupervised}. Furthermore, online learning over the edge devices is also a challenging task as the traditional training is performed over high-computation devices. Future directions in terms of VD is to acquire optimized online learning techniques that are able to learn over the edge devices \cite{cui2019stochastic,jin2021learning}.
\subsection{Reinforcement Learning for VD}
Reinforcement learning is widely been used in many computer vision problems \cite{dogru2021actor,hafiz2021reinforcement}, where it has not been explored for the problem of VD in real-world environments. Reinforcement learning techniques have surely accomplished complex tasks with higher effective and efficient results. The current limited adaptability with the environment challenge for VD methods in real-world can be easily solved with the assistance of reinforcement learning methods \cite{agarwal2021contrastive,sonar2021invariant}.
\subsection{Vision Transformers in VD Domain}
Vision transformers in sequence-based problems have shown tremendous performance, particularly, for image recognition and detection tasks \cite{dosovitskiy2020image,zhang2021vit}. Similarly, TimeSformer are introduced for precise video classification tasks such as action and activity recognition and video understanding \cite{bertasius2021space}. There are many other recent technologies such as human brain-inspired spiking neural networks \cite{pavlidis2005spiking}, and explainable artificial intelligence \cite{arrieta2020explainable,rojat2021explainable} models, that are barely explored for the VD task. These possible future directions need the research attention in near future to enrich this domain and assist significantly in automated surveillance, thereby protecting and ensuring the safety of humans.
\subsection{\textcolor{black}{Computation Friendly VD Methods}}
\textcolor{black}{The computation resources acquired by the current VD methods are not environment friendly and most of the techniques utilize computationally complex models to detect violent activities. Such methods practical applicability is questionable as in most of the surveillance environments real-time decision making is required for quick responsive actions and counter-measures. Therefore, future research from computation perspective with end-to-end resources friendly deep models can be applied in VD domain.}
\section{Conclusive Remarks}
The presented review paper highlights the current achievements of computer vision experts in VD domain and highlights the major challenges with future research directions. Concluding the overall review, it suggests the in the early VD research, traditional features such as motion vectors, acceleration information, instant change detection filters are utilized to identify violence. With the success of deep learning in many domains, the traditional processing is replaced by the powerful features representation potentials of 2D CNNs, that too comes with some limitations such as non-effective complex large motion patterns representation, sometimes huge time complexity, under representation of events, and the foremost problem of thirst of data. VD literature advanced to a new level with the usage of 3D CNNs for video data representation, which consider time series frames for video classification tasks and performs well for VD, but with the huge cost of computation, requiring complex GPUs for output generation.
\color{black}
There is a higher level of accuracy and precision in the testing results for most of the VD datasets surveyed in this study. However, the existing methods do not focus enough on the training and evaluation of their models in terms of the environment that they will live in in the real world. It can for instance happen that in real world surveillance scenarios, a scene may contain multiple abrupt changes, such as the sudden appearance of irrelevant objects that instantly change the motion patterns, and hence destabilize the model's output prediction. Thus, there is a need to adapt deep models in terms of non-stationary surveillance environments in order to cope with the current situation. Furthermore, it is also necessary to use VD techniques focusing on meaningful information extraction. This is to analyze the person's behavior after its successful localization has been accomplished, finally resulting in successful VD and localization, respectively.
\color{black}
\section*{Conflict of Interest}
The authors declare that they have no known conflict of interest.
\bibliographystyle{spmpsci}
|
\section{Introduction}
\label{section:Introduction}
Multimedia transmission now dominates the internet. While video transmission and storage indisputably command more resources, audio is also a resource hog, and like video, is important to consumers. Modern media consumers are increasingly savvy about audio technology and expect a high quality of experience (QoE) when listening using increasingly high-resolution and high-fidelity systems, whether they be on mobile devices or in their living rooms. Therefore, sound-quality evaluation tests are critical since they provide the necessary user feedback that drives improvements in these technologies. However, subjective listening tests demand a large amount of time and effort. Thus, there is a significant impetus to develop and deploy efficient and accurate audio quality assessment models that can be used to monitor and control end-user QoE. Furthermore, since bandwidth-hungry spatial audio is becoming more pervasive~\cite{NetflixTechBlog}, perceptual audio rate control has become more important and should be a significant factor in QoE optimization, especially when the available bandwidth becomes limited. Many lossy speech and audio codecs are not able to exactly preserve the positions of the various sources within the auditory scene or they may even discard the spatial information altogether if the available data rate is too small~\cite{audio_codec_artifacts},\cite{AESartifacts}. Furthermore, the assessment of binaural audio quality is receiving increasing interest since spatial features in the latest codecs are of increasing importance. Any decrease in spatial quality results in a significant decrease in presence and immersion. These issues greatly motivate us to focus on and develop a consistent and reliable stereo audio quality metric. We are interested in assessing spatial quality in a way that ensures how accurately the sources are positioned as compared to the unencoded reference as well as the effects of audio fidelity degradation due to audio coding.
Quality metrics for audio coding have been primarily designed to evaluate mono audio codecs, the earliest being PEAQ~\cite{PEAQstandard},\cite{peaq}. The fundamental principle of PEAQ is the calculation of so-called model output variables (MOVs) of a monaural hearing model, comparing these MOVs of the reference signal and the degraded signal and feeding these differences into a shallow neural network (NN) that is trained based on the known results of numerous listening tests. PEAQ does offer the possibility to evaluate stereo signals. However, for stereo, two monaural hearing models are used in parallel~\cite{PEAQstandard}, i.e., no inter-channel cues or spatial aspects are considered. The exception is the basic version of PEAQ, where, with the 2 (out of 11) MOVs, a sort of “worse ear”-approach~\cite{PEAQstandard} is used to look for perceivable distortions. Following PEAQ, POLQA~\cite{polqa} and ViSQOL~\cite{Hines20156:ViSQOLAudio01},\cite{7940042},\cite{v3} were developed. POLQA was primarily developed for speech and does not include any spatial aspects in its design. In ViSQOL, left, right, mid, and side channels were evaluated and studied, but ultimately, the final design considered the mid-channel only~\cite{7940042},\cite{ViSQOLgithub}. Thus, none of the popular and standardized audio quality metrics consider spatial aspects in their design. Since the fidelity of the spatial information is not explicitly considered, this may lead to faulty quality estimates~\cite{peaq-mc}.
Several research papers~\cite{peaq-mc},\cite{peaq-binaural},\cite{seo2013perceptual} have focused on learning a spatial audio quality metric from known binaural auditory cues~\cite{Blauert} like Interaural Level Differences (ILD), Interaural Time Differences (ITD) and Cross- Correlation (IACC) between signals entering the left and the right ear. Although early work to extend PEAQ to spatial audio was published~\cite{peaq-mc} it did not yield a recommendation. Schäfer et al.~\cite{peaq-binaural} presented an extension to PEAQ, that utilized MOVs from a binaural hearing model which is fed to a NN together with the result of PEAQ. In a similar spirit,~\cite{seo2013perceptual} extended PEAQ by a binaural auditory model, allowing predicting the audio quality of multichannel audio codecs. Delgado and Herre~\cite{Fhg_DirecLoudMap} proposed utilizing distortion between directional loudness maps of reference and coded audio as an indicator for spatial distortion, and in addition, utilized the objective scores predicted by PEAQ and POLQA to predict the overall objective quality. In their evaluation they specifically excluded speech signals. All the mentioned approaches can be considered as an add-on to the standard PEAQ (and POLQA). In AMBIQUAL \cite{ambiqual}, researchers presented a full-reference objective audio quality metric designed to estimate the robustness of ambisonic audio compression in terms of perceived quality and localization accuracy. The metric is built on the same similarity measure as used in ViSQOL. However, unlike ViSQOL, a spectrogram of phase angles (rather than magnitudes) is used for signal similarity comparisons.
A recent more powerful alternative is provided using deep-learning-based concepts, which have offered solutions that are accurate, rapidly re-trainable, and easily expandable in many speech and audio-related tasks. In~\cite{smaq_netflix}, the authors leveraged the most important features computed by the perceptual frontend from several audio quality metrics. They trained a multi-task student model using unlabeled data and labels from multiple audio quality metrics; and showed improved generality of the student model in predicting the quality score. Their research follows a similar spirit to our previous InSE-NET~\cite{InSE-NET} research. However, neither of the models was designed for stereo. There exists a deep learning-based perceptual spatial audio quality metric~\cite{SAQAM} that evaluates the similarity of binaural presentations in terms of both localization and sound fidelity degradations between any pair of binaural signals. However, it does not predict an easy-to-interpret (e.g., MUSHRA) quality score. It can only predict (deep feature) distances from the quality-predictive features delivered by a DNN. Furthermore, the model is trained for speech signals at 16~kHz (information about the sampling rate is inferred from their previous paper~\cite{dplm}).
With InSE-NET~\cite{InSE-NET} we demonstrated mimicking a state-of-the-art coded audio quality metric with deep neural networks (DNN) and subsequently improving it – completely with programmatically generated data. We showed that with synthetic data augmentation, one can steer the model to predict accurately. Acknowledging that training with listening tests should further improve the prediction accuracy, in this contribution we: (1) present an extension for handling stereo signals, (2) transfer selected weights from the pre-trained InSE-NET and retrain with listening tests. The design considers stereo/spatial aspects, and we name our model Stereo InSE-NET. We are not aware of any other deep learning-based coded stereo audio quality prediction model handling general audio signals at 48~kHz. Furthermore, we directly predict the subjective quality scores of coded audio signals on a 0-100 MUSHRA~\cite{MUSHRA} quality scale; a well-established quality scale for assessing the quality of audio codecs. Our goal is to demonstrate consistent and reliable coded audio quality prediction, both for waveform preserving codecs (e.g., AAC) and for non-waveform preserving codecs that include parametric bandwidth extension tools (e.g., HE-AAC v1) and parametric stereo coding tools (e.g., HE-AAC v2). In such a use-case of predicting the overall audio coding quality, the interaction of spatial cue distortions along with monaural/timbral distortions (especially in non-waveform-preserving cases) presents a challenging scenario.
The most obvious application of our work is a stand-alone coded audio quality metric, which is useful e.g., for validating encoder improvements with new tunings. Furthermore, since it is based on a deep network, our proposed model can be developed into a learnable comparable loss function for coded audio enhancement~\cite{DCAE}.
The rest of the paper is organized as follows: In Section 2, the data used for model training and evaluation are introduced. Section 3 depicts the details of our proposed Stereo InSE-NET model. The related experimental results and analysis are given in Section 4, and finally, the conclusion is drawn in Section 5.
\section{Datasets}
\label{section:Datasets}
\subsection{Training set}
\label{subsection:trainingset}
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.1\linewidth]{MUSHRAdistribution2021.png}
\caption{Histogram of mean MUSHRA scores from the training database.}
\label{fig:mushraDistribution}
\end{center}
\end{figure}
When we started this study, we were aware that fine-tuning the mono InSE-NET with listening tests should further improve prediction accuracy. However, the time and expense of obtaining the “gold standard” human judgments limit the availability of such data. For this study, we consciously decided to make the best use of existing listening test data and pre-trained models. We realized we had significantly more stereo listening tests with which to train the model than mono listening tests. Therefore, we decided to enhance and extend the mono InSE-NET to Stereo InSE-NET by training with stereo listening tests.
We used our internal corpus of Multiple Stimuli with Hidden Reference and Anchor (MUSHRA)~\cite{MUSHRA} listening tests at a 48 kHz sample rate. Each of these tests included an unencoded hidden reference, the 3.5~kHz and 7~kHz low-pass filtered versions of the unencoded signals, and one or more coded signals. The codecs used in the training set were Advanced Audio Coding (AAC)~\cite{AAC}, High-Efficiency Advanced Audio Coding (HE-AAC v1 and v2)~\cite{HE-AAC}, and Dolby AC-4~\cite{ac4_IEEE} spanning a wide range of bitrates. We included both speech, music, and a mix of speech and music signals in the training set, but we did not use any listening tests that included a dedicated speech codec. Furthermore, from our audio codec listening experience, we are aware that listeners tend to find it difficult to weigh their preference between spatial coding and waveform coding artifacts when presented concurrently. From our initial experiments, we also observed similar confusion with the predicted quality score from Stereo InSE-NET. Therefore, we enlarged the training data with listening tests that include \emph{hybrid stereo coding}, i.e., codecs whose low bands are coded with waveform preserving stereo coding tools and high bands are coded with parametric stereo coding tools. For this purpose, we included listening tests which included a few experimental codecs that led to the development of the stereo coding tools in the AC-4 standard~\cite{ac-4Part1}. Figure 1 plots the histogram of the mean MUSHRA scores over the entire listening test database, showing a wide range of perceptual quality scores. Note that listening test excerpts can be of different lengths, ranging up to a maximum length of 56.48s in our training set. Therefore, we extended the smaller excerpts to the maximum length of 56.48s by zero-padding them on both sides of the excerpt.
\subsection{Test sets}
\label{subsection:testset}
\begin{figure}[t]
\begin{center}
\includegraphics[width=1.1\linewidth]{MUSHRAdistributionUSACstereo_testsets.png}
\caption{Histogram of mean MUSHRA scores from the stereo listening test database for testing our model.}
\label{fig:USAC-VT1and2_mushraDistribution}
\end{center}
\end{figure}
We benchmarked the prediction accuracy of Stereo InSE-NET against the same stereo listening tests presented in our previous work~\cite{InSE-NET}. We used the Unified Speech and Audio Coding (USAC)~\cite{MPEG-USAC} verification listening tests~\cite{usac_lt},\cite{USAC} to evaluate the performance of the Stereo InSE-NET against subjective listening scores.
USAC verification listening tests~\cite{usac_lt},\cite{USAC} contain 27 excerpts coded with USAC, HE-AAC, and AMR-WB+ with a wide range of bitrates from 8~kbps mono to 96~kbps stereo. These comprehensive verification tests were designed to provide information on the subjective performance of the USAC codec. It consists of three separate listening tests: mono at low bitrates and stereo at both low and high bitrates. All tests were MUSHRA tests, whose quality scale ranges from 0 to 100, where a higher score implies better quality. The histogram of the mean MUSHRA scores for the two stereo listening tests is shown in Figure 2. For the details of these MUSHRA listening tests, interested readers are referred to~\cite{usac_lt},\cite{USAC}. Note that we included the mono listening test because we also wanted to evaluate the accuracy of our stereo model when trained with only stereo listening tests.
\section{Stereo InSE-NET}
\label{section:StereoInSE-NET}
\subsection{Model architecture}
\label{subsection:model_arch}
As shown in Figure 3, our architecture is built upon our previous architecture, with only minimalistic change. The mono version of the InSE-NET operated on the Gammatone spectrograms of the reference-coded (ref.-cod.) signal pair. Gammatone filters are a popular approximation to the filtering performed by the ear. The Gammatone spectrogram can thus be considered as a more perceptually motivated representation than the traditional spectrogram. The most obvious way to extend the mono InSE-NET to stereo is to simply feed in Gammatone spectrograms of ref.-cod. pairs for left (L) and right (R) channels. However, to condition the model with stereo cues, we feed in additionally ref.-cod. pairs of mid (M) and side (S) channels, where $M = 0.5(L+R)$ and $S = 0.5(L-R)$. Thus, the proposed Stereo InSE-NET model operates on the Gammatone spectrograms of L, R, M, and S signals of reference and coded stereo signals. The Gammatone spectrogram of the audio signal is calculated with a window size of 80\,ms, hop size of 20\,ms, and 32 frequency bands ranging from 50~Hz up to 24~kHz. The four resulting Gammatone spectrograms of reference and coded signals are paired and stacked along the channel dimension, which results in an input size of 8$\times$32$\times$frames (channels$\times$bands$\times$frames) to the network. Since the signals in the training set are equalized to a length of 56.48s, the shape of the input layer is 8$\times$32$\times$2824. The core DNN backbone of Stereo InSE-NET is based on the same modules as the mono InSE-NET: a combination of modified Inception (In) and Squeeze-and-Excitation (SE) blocks followed by fully connected layers. Since the design considers stereo/spatial aspects, we name our model Stereo InSE-NET. Intuitively one would expect mono InSE-NET should have already learned some useful weights for the coded audio quality prediction task. So, we transferred the weights from its first two Inception blocks to the Stereo InSE-NET (except for the input layer dealing with the side-channel) and retrain with listening tests, i.e., transfer learning. The benefits of transfer learning will be shown in Section 3.2.
\begin{figure*}
\centering
\includegraphics[width=.74\linewidth]{StereoInSEnet.png}
\caption{Block diagrams of the mono InSE-NET~\cite{InSE-NET} (left) and the proposed Stereo InSE-NET model (right). Input to the Stereo InSE-NET is ref.-cod. pairs of Gammatone spectrograms (of L, R, M, and S signals) which are computed in a manner identical to ViSQOL-v3~\cite{ViSQOLgithub}. Except for the input layer dealing with the side-channel, Inception block A, and Inception blocks A and B are initialized from the pre-trained mono InSE-NET.}
\label{fig:StereoInSE-NET}
\end{figure*}
\begin{table}[t]
\setlength\tabcolsep{1.8pt}
\vspace{-0.4cm}
\caption{Architectures and parameters of the Stereo InSE-NET model. Note that compared to its mono counterpart~\cite{InSE-NET}, only the shape of the input layer has changed.}
\vspace{-0.4cm}
\begin{center}
\scriptsize{
\begin{tabular}{|l||r|r|r|r|r|}
\hline
\multicolumn{1}{|c||}{\textbf{\backslashbox{\textbf{Layer}}{\textbf{Shape}}}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}}Output\end{tabular}}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}}Horizontal \\ Conv.\\ (w$\times$h\,/\,ch)\end{tabular}}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}}Vertical \\ Conv. \\ (w$\times$h\,/\,ch)\end{tabular}}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}}Normal \\ Conv. \\ (w$\times$h\,/\,ch)\end{tabular}}} & \multicolumn{1}{c|}{\textbf{\begin{tabular}[c]{@{}c@{}} Average \\ Pooling \\ (w$\times$h) \end{tabular}}} \\ \hline\hline
\textbf{Input} & 8$\times$32$\times$2824 & & & & \\ \hline
\textbf{Inception A} & 208$\times$16$\times$180 & 3$\times$7 / 64 & 7$\times$3 / 64 & 1$\times$1 / 64 & 5$\times$5 \\ \hline
\textbf{Inception A} & 224$\times$16$\times$90 & 3$\times$7 / 64 & 7$\times$3 / 64 & 1$\times$1 / 64 & 5$\times$5 \\ \hline
\textbf{SE} & 224$\times$16$\times$90 & & & & \\ \hline
\textbf{Inception B} & 256$\times$16$\times$45 & 3$\times$5 / 64 & 5$\times$3 / 64 & 1$\times$1 / 64 & 5$\times$5 \\ \hline
\textbf{SE} & 256$\times$16$\times$45 & & & & \\ \hline
\textbf{Inception C} & 256$\times$14$\times$22 & 3$\times$3 / 64 & 5$\times$5 / 64 & 1$\times$1 / 64 & 3$\times$3 \\ \hline
\textbf{SE} & 256$\times$14$\times$22 & & & & \\ \hline
\textbf{\begin{tabular}[c]{@{}l@{}}Adaptive\\ AvgPool\end{tabular}} & 256$\times$4$\times$4 & & & & \\ \hline
\textbf{FCL 1} & 3200$\times$1 & & & & \\ \hline
\textbf{FCL 2} & 512$\times$1 & & & & \\ \hline
\textbf{FCL 3} & 1$\times$1 & & & & \\ \hline
\end{tabular}
}
\end{center}
\label{tbl:parameters}
\end{table}
Our proposed Stereo InSE-NET has 15.25M parameters and its architectural and parameter details are tabulated in Table 1. The number of parameters is only $0.022\%$ more than the mono version which is attributed to (six) additional input channels. As discussed for the mono model, the kernel sizes employed in the Inception blocks (A, B, and C) considered several possible kernel shapes and sizes (3$\times$3, 3$\times$5, up to 9$\times$9). A parametric grid search was performed to identify the optimal kernel sizes for each layer. In Table 1, we directly present the optimal kernel sizes of the Inception blocks as obtained for our mono model. Note that for the stereo model, we have not searched for optimal kernel sizes for each layer.
\begin{figure}[t]
\begin{center}
\includegraphics[width=1\linewidth]{training_losses_all.png}
\caption{(Top) Training losses with default PyTorch initializer~\cite{He_2015} (blue); Inception block A (red), and both Inception blocks A and B (yellow) initialized from pre-trained mono InSE-NET. Both blocks are initialized from a pre-trained mono InSE-NET, but without mid-channel (purple). (Bottom) Zoomed-in version of top plot.}
\label{fig:training_losses}
\end{center}
\end{figure}
It is interesting to note that in~\cite{dplm} the authors evaluated a variety of different convolutional network building blocks as the feature extraction block, and found Inception resulted in the best features for the spatial audio localization~\cite{dplm},\cite{SAQAM} and quality metric~\cite{SAQAM}. They used a 6-block Inception network operating on phase and magnitude spectrogram, however, unlike ours, they did not do any specific kernel size modification for handling audio signals.
\subsection{Training configuration}
\label{subsection:training_config}
We attempt to leverage the domain knowledge which has been learned during the training of mono InSE-Net with abundant synthetic mono data and transfer these weights of the first two layers of Inception blocks to Stereo InSE-NET rather than training from scratch. There are two advantages of such a strategy: first, diverse, well-structured, and carefully annotated stereo listening tests are still a rare resource, and second, initializing the stereo model with the weights from the mono model would be beneficial for optimization. We then retrained the stereo model with typical MUSHRA listening tests~\cite{MUSHRA}. For a ref.-cod. pair, our training target is the average of the MUSHRA scores rated by listeners. Similar to mono InSE-NET, we made use of the smooth L1-loss~\cite{SmoothL1LossPytorch} for training.
The training dataset was first normalized and partitioned randomly into $80\%$ for training and $20\%$ for validation. A 5-fold cross-validation is applied to ensure that the model could make full use of limited data. We keep the optimal kernel sizes and learning rates the same as we experimented in the mono InSE-Net and selected a learning rate of $10^{-4}$ and a batch size of 8. Batch normalization and ReLU as activation functions were applied after all convolutional layers. The setup is implemented with PyTorch and was trained for 10 epochs for each fold (i.e., 50 epochs in total) on an Nvidia GTX 1080 Ti GPU with Adam optimizer. The model is trained and evaluated under the following criteria: smooth L1 Loss, mean squared error (MSE), Spearman's correlation coefficient ($R_s$), and Pearson's correlation coefficient ($R_p$).
The decay of training losses is depicted in Figure 4. We plotted the training losses with the default PyTorch initializer, i.e., Kaiming uniform~\cite{He_2015}. Next, only the Inception block A, and both Inception blocks A and B are initialized from pre-trained mono InSE-NET. Note that we did not initialize the input layer dealing with the side-channel from the pre-trained mono InSE-NET. In the zoomed-in plot, we notice that the initialization from a mono model leads to starting from a lower loss and consequently leads to faster convergence. We also experimented with initializing subsequent blocks from the mono model, but we did not observe any benefit in going beyond initializing Inception block B. Next, we also experimented without mid-channel as the input to the Stereo InSE-NET. We can observe from the plot that for this dataset, it is possibly sufficient to exclude the mid-channel.
\section{Results and Discussion}
\label{section:Results}
We benchmark against ViSQOL because it has been reported in \cite{Fraunhofer}, that out of all objective measures designed to evaluate codecs, ViSQOL shows the best correlation with subjective scores and achieves high and stable performance for all content types. Similarly, it was reported in~\cite{smaq_netflix} that overall VISQOL performed very well across five different datasets (see, Table II in~\cite{smaq_netflix}, overall correlation with ViSQOL as the teacher). Thus, we will benchmark against the latest version of ViSQOL-v3 (operating in audio mode)~\cite{ViSQOLgithub}.
\begin{table}[t]
\setlength\tabcolsep{7pt}
\caption{Performance of ViSQOL-v3, mono InSE-NET, and Stereo InSE-NET on the two stereo listening tests. The table shows the correlation coefficients ($R_p$ and $R_s$) between predicted objective scores and subjective (MUSHRA) scores from the stereo listening tests. The Stereo InSE-NET (a)-(c) indicates training with different data.}
\vspace{-0.4cm}
\begin{center}
\scriptsize{
\begin{tabular}{|l||c|r|c|r|}
\hline
\multirow{2}{*}{\backslashbox{\textbf{Model}}{\textbf{Metric}}} & \multicolumn{2}{c|}{\textbf{Low Bitrates}} & \multicolumn{2}{c|}{\textbf{High Bitrates}} \\ \cline{2-5}
& $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} & $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} \\ \hline\hline
\textbf{ViSQOL-v3} & \multicolumn{1}{r|}{0.777} & 0.782 & \multicolumn{1}{r|}{0.825} & 0.906 \\ \hline
\textbf{Mono InSE-NET} & \multicolumn{1}{r|}{0.806} & 0.788 & \multicolumn{1}{r|}{0.847} & 0.895 \\ \hline
\textbf{Stereo InSE-NET (a)} & \multicolumn{1}{r|}{0.888} & 0.838 & \multicolumn{1}{r|}{0.892} & 0.874 \\ \hline
\textbf{Stereo InSE-NET (b)} & \multicolumn{1}{r|}{0.897} & 0.861 & \multicolumn{1}{r|}{0.907} & 0.899 \\ \hline
\textbf{Stereo InSE-NET (c)} & \multicolumn{1}{r|}{0.915} & 0.880 & \multicolumn{1}{r|}{\textbf{0.912}} & \textbf{0.911} \\ \hline
\textbf{Stereo InSE-NET w/o M (c)} & \multicolumn{1}{r|}{\textbf{0.922}} & \textbf{0.900} & \multicolumn{1}{r|}{0.910} & 0.910 \\ \hline
\end{tabular}}
\end{center}
\label{tbl:stereo_results}
\end{table}
We used the Spearman rank-order correlation coefficient ($R_s$) to measure the prediction monotonicity of the models and the Pearson linear correlation coefficient ($R_p$) to measure the prediction linearity. For both $R_p$ and $R_s$, larger values denote better performance. The prediction accuracy of the Stereo InSE-NET on two stereo listening tests is presented in Table 2. We benchmarked the prediction accuracy of Stereo InSE-NET against the same stereo listening tests presented in our previous work. Internally, ViSQOL downmixes stereo (or any multi-channel signal) to a mono mid-signal and then predicts the MOS. One can observe the prediction accuracy of ViSQOL in the first row. With mono InSE-NET, we compute the mid-signal and feed it into the mono model. Mono InSE-NET results in a slightly better $R_p$ than ViSQOL on the two stereo listening tests as shown in the second row. Moreover, the mono InSE-NET displays a higher accuracy in estimating the quality of excerpts encoded at high bitrates. In the subsequent rows, we demonstrate improving prediction accuracy by training the Stereo InSE-NET with the following data:
\begin{enumerate}[label=(\alph*)]
\item Stereo listening tests (as described in Section 2.1);
\item Synthetically augmenting (a) by swapping the left- and right-channel of all the audio signals, but with unchanged MUSHRA scores;
\item Additional hybrid stereo coding (as described in Section 2.1) listening tests along with swapping the left- and right-channel of all the audio signals, but keeping MUSHRA scores unchanged.
\end{enumerate}
In the third row in Table 2, we show the impact of training our stereo model with stereo listening tests (data: a). There is a boost in $R_p$. However, there is a slight degradation in $R_s$, suggesting probably the training data was insufficient. In the fourth row, we show the results of training with synthetic listening test data augmentation (data: b). Now both the correlation measures are improved. With the inclusion of hybrid stereo coding listening tests along with listening test data augmentation (data: c), we observe an additional improvement in the fifth row. Overall, we demonstrate a significant improvement ($12\%$ and $6\%$ improvement of Pearson's and Spearman's Rank correlation coefficient, respectively) over ViSQOL-v3 by training the Stereo InSE-NET with both stereo listening tests and synthetically augmented training data. Furthermore, in our ablation experiments, we found that quality prediction accuracy improves with the inclusion of mid- and side-channels. Particularly, the inclusion of the side-channel improves the prediction accuracy towards lower bitrates. In the sixth row, we show that to save computational complexity, it is perhaps sufficient to exclude the mid-channel. We observe a surprising boost in $R_s$ at lower stereo bitrates, but we attribute this result possibly (still) due to the limited amount of training data.
Next, we evaluate the performance of the stereo model with the mono listening test. Note that our stereo model is not trained with mono-listening tests. With our stereo model, we evaluate a dual-mono (i.e., stereo signal with L = R) ref-coded pair. In \cite{InSE-NET}, we examined the correlation between different objective scores (PEAQ-Advanced \cite{peaqOnline}, ViSQOL-v3 \cite{ViSQOLgithub}, and mono InSE-NET) against subjective scores from the mono USAC verification listening test. In Table 3, we list the results together with the Stereo InSE-NET. We can observe in the last row that even though our model was not trained with mono listening tests, the results display a strong correlation between the prediction of the Stereo InSE-NET and the subjective quality score. These results show, that programmatically generated training data, informed with domain-specific know-how is a powerful technique. But including listening tests as training data still provides a benefit in improving the quality prediction accuracy.
\begin{table}[t]
\setlength\tabcolsep{9pt}
\caption{Performance of PEAQ, ViSQOL-v3, mono InSE-NET, and Stereo InSE-NET with respect to the mono USAC verification listening test. The table shows the correlation coefficients ($R_p$ and $R_s$) between predicted objective scores and subjective (MUSHRA) scores from the mono listening test.}
\vspace{-0.4cm}
\begin{center}
\scriptsize{
\begin{tabular}{|l||r|r|}
\hline
\backslashbox{\textbf{Model}}{\textbf{Metric}} & \multicolumn{1}{c|}{\textbf{$\mathbf{R_{p}}$}} & \multicolumn{1}{c|}{\textbf{$\mathbf{R_{s}}$}} \\ \hline\hline
\textbf{PEAQ Advanced} & 0.650 & 0.700\\ \hline
\textbf{ViSQOL-v3} & 0.810 & 0.840\\ \hline
\textbf{Mono InSE-NET} & 0.830 & 0.835\\ \hline
\textbf{Stereo InSE-NET} & \textbf{0.905} & \textbf{0.903}\\ \hline
\end{tabular}}
\end{center}
\label{tbl:mono_results}
\end{table}
Furthermore, we examine the performance of our proposed stereo model along with the mono model on the individual codecs, namely AMR-WB+, HE-AAC, and USAC. The reference and two anchors are included in each examination, and corresponding $R_p$ and $R_s$ are listed in Tables 4a to 4c.
\begin{table}[t]
\setlength\tabcolsep{5pt}
\centering
\caption{Performance of mono and Stereo InSE-NET for various codecs in USAC verification listening tests. The table shows the correlation coefficients ($R_p$ and $R_s$) between predicted objective scores and subjective (MUSHRA) scores from (a) stereo low-bitrate, (b) stereo high-bitrate, and (c) mono listening tests.}
\vspace{-0.4cm}
\begin{center}
\scriptsize{
\begin{tabular}{|l||c|r|c|r|}
\hline
\multirow{2}{*}{\backslashbox{\textbf{Codec}}{\textbf{Metric}}} & \multicolumn{2}{c|}{\textbf{Mono InSE}} & \multicolumn{2}{c|}{\textbf{Stereo InSE}} \\ \cline{2-5}
& $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} & $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} \\ \hline\hline
\textbf{AMR-WB+} & 0.868 & 0.842 & \textbf{0.960} & \textbf{0.904} \\ \hline
\textbf{HE-AAC} & 0.830 & 0.790 & \textbf{0.945} & \textbf{0.877} \\ \hline
\textbf{USAC} & 0.891 & 0.860 & \textbf{0.976} & \textbf{0.943} \\ \hline
\end{tabular}}
\end{center}
\subfloat[\label{tbl:usac2}]
\newline
\vspace{-0.4cm}
\begin{center}
\scriptsize{
\begin{tabular}{|l||c|r|c|r|}
\hline
\multirow{2}{*}{\backslashbox{\textbf{Codec}}{\textbf{Metric}}} & \multicolumn{2}{c|}{\textbf{Mono InSE}} & \multicolumn{2}{c|}{\textbf{Stereo InSE}} \\ \cline{2-5}
& $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} & $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} \\ \hline\hline
\textbf{AMR-WB+} & 0.864 & 0.852 & \textbf{0.955} & \textbf{0.925} \\ \hline
\textbf{HE-AAC} & 0.871 & 0.925 & \textbf{0.946} & \textbf{0.949} \\ \hline
\textbf{USAC} & 0.909 & 0.920 & \textbf{0.964} & \textbf{0.942} \\ \hline
\end{tabular}}
\end{center}
\subfloat[\label{tbl:usac3}]
\newline
\vspace{-0.4cm}
\begin{center}
\scriptsize{
\begin{tabular}{|l||c|r|c|r|}
\hline
\multirow{2}{*}{\backslashbox{\textbf{Codec}}{\textbf{Metric}}} & \multicolumn{2}{c|}{\textbf{Mono InSE}} & \multicolumn{2}{c|}{\textbf{Stereo InSE}} \\ \cline{2-5}
& $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} & $\mathbf{R_p}$ & \multicolumn{1}{c|}{$\mathbf{R_s}$} \\ \hline\hline
\textbf{AMR-WB+} & 0.889 & 0.856 & \textbf{0.948} & \textbf{0.922} \\ \hline
\textbf{HE-AAC} & 0.853 & 0.791 & \textbf{0.945} & \textbf{0.887} \\ \hline
\textbf{USAC} & 0.873 & 0.881 & \textbf{0.950} & \textbf{0.939} \\ \hline
\end{tabular}}
\end{center}
\subfloat[\label{tbl:mono}]
\end{table}
The proposed stereo model results in a significantly better $R_p$ and $R_s$ than the mono model. Even though we have not trained our model with AMR-WB+ (a speech codec-based system) and USAC codec (which uses a dedicated speech codec for speech signals at low bitrates), the overall performance of our model on these unseen codecs has significantly improved. The proposed stereo model has performed homogeneously well on the experimented mono and stereo codecs. The estimation of coding quality (particularly $R_s$) of HE-AAC, under this comparison, is unexpectedly the worst among the three experimented codecs, even though our model is trained on the listening test excerpts encoded with HE-AAC and AAC. One possible explanation for this result is one of the lowest bitrates in the low bitrate mono and stereo listening tests is unseen during the training.
\section{Conclusion and Discussion}
\label{section:ConclusionDiscussion}
In this paper, we present Stereo InSE-NET: a novel DNN-based coded stereo audio quality prediction model. The model is an extension of our previous mono InSE-NET. The core DNN backbone of Stereo InSE-NET is based on the same modules as the mono InSE-NET: a combination of modified Inception and Squeeze-and-Excitation blocks followed by fully connected layers. The new design incorporates inter-channel aspects in the design by conditioning the DNN-backbone with Gammatone spectrograms of left, right, mid, and side channels. We demonstrate that weights learned from the mono quality prediction model fine-tuned for the stereo quality prediction task leads to faster convergence. Furthermore, we show that utilizing both real listening test data, and synthetically derived data from listening tests leads to an improvement in prediction accuracy.
In the future, we would like to extend our model to handle both coded stereo and binaural audio quality prediction. Assuming spatial audio content of more than two channels can be rendered to a binaural representation~\cite{peaq-mc},\cite{seo2013perceptual} we would then also investigate our model with multichannel subjective listening test data.
Finally, our model enables faster prediction. An un-optimized PyTorch implementation of our model (excluding the Gammatone spectrogram computation frontend) runs at 63.8x real-time on a CPU with the latest PyTorch version 1.12.1. Including the Gammatone spectrogram computation frontend in Python, our complete Stereo InSE-NET model runs at 1.2x real-time on a CPU, which is slightly faster than ViSQOL-v3 (which runs at 1.1x real-time on CPU). However, it is important to reiterate that ViSQOL-v3 consists of traditional signal processing-based blocks which are fully implemented in C++, whereas our proposed Stereo InSE-NET model is a unoptimized Python and PyTorch implementation of a deep CNN-based model. Furthermore, ViSQOL-v3 computes a pair of Gammatone spectrograms, whereas we compute four pairs of Gammatone spectrograms. Therefore, even the slight advantage in computational efficiency over ViSQOL-v3 is a positive outcome.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
Suitable engineering and control techniques for quantum systems are needed for the development of reliable quantum information processing devices \cite{altafini2012modeling}. Feedback control methods, which dominate classical applications, present unique challenges in the quantum domain, as measured systems exhibit stochastic evolutions: a rigorous formalism for their treatment has been developed hinging on stochastic differential equations that describe quantum conditional dynamics, or filtering equations
\cite{hudson1984quantum,belavkin1989nondemolition,bouten2007introduction,barchielli2009quantum}. A typical task is that of stabilizing a target state of interest, and different design techniques have been employed to this aim. These include output-feedback methods (also known as Markovian feedback) \cite{ticozzi2008quantum,ticozzi2009analysis}, and filtering-based feedback methods \cite{van2005feedback,mirrahimi2007stabilizing,ticozzi2012stabilization,liang2019exponential}. While the dynamics is that of a monitored open system, the control typically enters the dynamics as a perturbation of the system Hamiltonian (a so-called coherent control action). Dissipative resources have been so far mostly used in open-loop control strategy, or as an always-on control action complemented by a purely Hamiltonian feedback \cite{ticozzi2012stabilization}. While continuous feedback laws would be desirable in practice, due to the geometry and controllability properties of the systems at hand, most available global stability results need to employ some basic switching strategy to destabilize the undesired equilibria.
Instead of looking at switching as a necessary price to pay in order to attain the desired control task, the potential of full-fledged switching control has been investigated starting in \cite{scaramuzza2015switching,grigoletto2021stabilization}, considering not only coherent but also dissipative control resources, with the latter obtained as measurements and controlled interaction with suitably engineered quantum environments.
In \cite{scaramuzza2015switching}, the switching laws are based on semigroup (also referred to as Lindblad, or master equation) dynamics, that correspond to expected average state of a quantum stochastic filtering equation, and can thus be computed off-line. On the other hand, in \cite{grigoletto2021stabilization} the switching is based on a real-time estimate of the current state of the system, proposing a stabilizing switching feedback law with fixed, sufficiently small dwell times.
In this paper, we formalize and complete the analysis of feedback control strategy in both approaches, providing strategies for which we can also {\em global exponential stability} (GES), ensuring a convergence that both is fast and more robust \cite{liang2019exponential,liang2021robustness}.
We also generalize the approach from deterministic switching times with fixed dwell time to stochastic switching times with hysteresis technique. In doing this, it is crucial to prevent chattering and the so-called Zeno effect - where the switching intervals become infinitesimal in the asymptotic limit. To this aim, we employ a switching law with hysteresis inspired that follows the ideas of \cite{liberzon2003switching,wu2013stability,teel2014stability}.
In addition, a new control strategy is presented, where the possibility of modulating the intensity of the switched generators allows for a relaxation of the stringent invariance conditions needed in the existing methods. So far, invariance of the target for \emph{all} the control actions was a necessary condition for stochastic stabilizability, with only some weak result obtained in \cite{grigoletto2021stabilization} in the non-invariant case. Here instead, by allowing to slow down convergence close to the target, we can rely on simple switching Lyapunov-type conditions to guarantee stabilizability.
In order to clarify the contributions of the paper and allow for a quick comparison, we report in Table \ref{table} the existing and the new strategies, together with the type of stochastic stability we can guarantee.
\begin{table}
\caption{ Switching strategies for stabilization}
\label{table}
\begin{tabular}{l||l|l}
& Average state & Measurement-based \\
\hline\hline
Fixed dwell time & GAS: \cite[Thm 2]{scaramuzza2015switching} & GAS: \cite[Thm 2]{grigoletto2021stabilization}\\
& GES: Corollary~\ref{Cor:GES_Dwell} & GES: Theorem~\ref{Thm:GES_dwell} \\
\hline
Hysteresis switching & GES: Theorem~\ref{Thm:GES_average} & GES: Theorem~\ref{Thm:GES_state} \\
\hline
Modulating Lindbladian & - & GAS: Theorem \ref{Thm:GAS}
\\
\hline
\end{tabular}
\end{table}
The paper is organized as follows: Section \ref{sec:systemproblem} introduces the systems of interest, the stability notions and the problems we address. The last subsection collects some known results on the stability of sets for the semigroup generators, as well as an instrumental lemma. Section \ref{Sec:GES} presents the standard, stronger set of control assumptions and presents both average and measurement dependent switching strategy, for which exponential stability can be proven.
Section \ref{Sec:Modulate} relaxes the assumptions, and introduces the possibility of modulating the controlled dynamics amplitude. In this setting, we can show that the target can be made \emph{globally asymptotically stable} (GAS) even if not all the generators leave it invariant.
Lastly, Section \ref{Sec:examples} presents some numerical simulation that illustrate the effectiveness of the proposed strategies, followed by some comments and outlook in the conclusions.
\textit{Notations:}
The imaginary unit is denoted by $i$.
We denote by $\mathcal{B}(\mathcal{H})$ the set of all linear operators on a finite dimensional Hilbert space $\mathcal{H}$. We take $\mathbf{I}$ as the identity operator on $\mathcal{H}$ and $\mathbf{I}_R$ as the identity operator on the subspace $\mathcal{H}_R\subset\mathcal{H}$, and take $\mathds{1}$ as the indicator function. We denote the adjoint $A\in\mathcal{B}(\mathcal{H})$ by $A^*$. We denote by $\bar{\lambda}(A)$ and $\underline{\lambda}(A)$ the maximum and minimum eigenvalue of the Hermitian matrix $A$ respectively.
Define $\mathcal{B}_{*}(\mathcal{H}):=\{X\in\mathcal{B}(\mathcal{H})|X=X^*\}$, $\mathcal{B}_{\geq 0}(\mathcal{H}):=\{X\in\mathcal{B}(\mathcal{H})|X\geq 0\}$ and $\mathcal{B}_{>0}(\mathcal{H}):=\{X\in\mathcal{B}(\mathcal{H})|X> 0\}$, $a\wedge b=\min\{a,b\}$ and $a\vee b=\max\{a,b\}$ for any $a,b\in\mathcal{R}$.
The function $\mathrm{Tr}(A)$ corresponds to the trace of $A\in\mathcal{B}(\mathcal{H})$. The Hilbert-Schmidt norm of $A\in\mathcal{B}(\mathcal{H})$ is denoted by $\|A\|_{HS}:=\mathrm{Tr}(AA^*)^{1/2}$.
The commutator of two operators $A,B\in\mathcal{B}(\mathcal{H})$ is denoted by $[A,B]:=AB-BA.$ We denote by $\mathrm{int}(\mathcal{S})$ the interior of a subset of a topological space.
\section{System description and problem setting}
\label{sec:systemproblem}
\subsection{Quantum stochastic master equations} We consider quantum systems described on a finite dimensional Hilbert space $\mathcal{H}$. The state of the system is associated to a density matrix on $\mathcal{H}$,
$$
\mathcal{S}(\mathcal{H}):=\{\rho\in\mathcal{B}(\mathcal{H})|\,\rho=\rho^*\geq 0,\mathrm{Tr}(\rho)=1\}.
$$
Assume that the system is monitored continuously via homodyne/heterodyne measurements, which yield a diffusion observation process. In the quantum filtering regime~\cite{belavkin1989nondemolition,bouten2007introduction,van2007filtering}, the measurements record $Y(t)$ can be described by a continuous semimartingale with quadratic variation $\left\langle Y(t),Y(t)\right\rangle=t$ (see~\cite[Corollary 5.2.9]{van2007filtering} for the proof.). Let $\mathcal{F}^Y_t:=\sigma\{Y(s):\, 0\leq s\leq t\}$ be the filteration generated by the observation up to time $t$. The observation process satisfies
\begin{equation*}
dY(t)=\sqrt{\eta}{\rm Tr}((C+C^*)\rho(t))dt+dW(t),
\end{equation*}
where the innovation process $W(t)$ is an one-dimensional Wiener process and satisfies \cite[Proposition 5.2.14]{van2007filtering} $\mathcal{F}^Y_t=\mathcal{F}_t:=\sigma\{W(s):\, 0\leq s\leq t\}$, $\eta\in(0,1]$ describes the efficiency of the detector, and $C\in\mathcal{B}(\mathcal{H})$ is the noise operator induced by interactions with the probe.
Conditioning on the observation process $Y(t)$, we have that the conditional density matrix of the system evolves according to a stochastic master equation (SME):
\begin{equation}
d\rho(t)=(-i[H,\rho(t)]+\mathcal{D}_C(\rho(t)))dt+\mathcal{G}_C(\rho(t))dW(t),
\label{Eq:SME}
\end{equation}
where $\rho_0\in \mathcal{S}(\mathcal{H})$, $H_0\in\mathcal{B}_*(\mathcal{H})$ is the effective Hamiltonian, which is equal to the free Hamiltonian $H_S$ of the system plus a correction $H_C$ induced by the coupling to the probe, $\mathcal{G}_{C}(\rho):=\sqrt{\eta}(C \rho+\rho C^*-\mathrm{Tr}((C+C^*)\rho)\rho)$, and $\mathcal{D}_C(\rho):=C\rho C^*-C^*C \rho/2-\rho C^*C /2$ is the Lindblad generator associated to the noise operator $C\in\mathcal{B}(\mathcal{H})$. The existence and uniqueness of the solution of the stochastic master equation~\eqref{Eq:SME} as well as the strong Markov property of the solution are ensured by the results established in~\cite{mirrahimi2007stabilizing} or~\cite[Chapter 5]{barchielli2009quantum}. Thanks to the generator $\mathcal{D}_C(\rho)$, $\mathcal{S}(\mathcal{H})$ is almost surely invariant for the stochastic master equation~\eqref{Eq:SME}~\cite[Lemma 3.2, Proposition 3.3]{mirrahimi2007stabilizing}.
Now, we suppose that the monitored system~\eqref{Eq:SME} can be coupled to one of a finite set of external systems during assigned period of times. The effect of these couplings on the dynamics are our control resources, and which is active at which time is going to be determined by a switching law. Assuming that these external systems act as memory-less (Markov) environments~\cite{alicki2007quantum}, the time evolution of the system state is described by the following switched stochastic master equation,
\begin{equation}
d\rho(t)=\sum^m_{k=1}u^k_t\mathcal{L}_k(\rho(t))dt+\mathcal{G}_{C}(\rho(t))dW(t),
\label{Eq:SSME}
\end{equation}
where $\rho_0\in \mathcal{S}(\mathcal{H})$, for all $k\in\mathcal{M}:=\{1,\dots,m\}$, $\mathcal{L}_k(\rho):=-i[H_0+H_k,\rho]+\mathcal{D}_{L_k}(\rho)+\mathcal{D}_C(\rho)$ with $H_k\in\mathcal{B}_*(\mathcal{H})$ and $L_k\in\mathcal{B}(\mathcal{H})$ representing the Hamiltonian perturbation and the noise operator induced by interactions with $k$-th external systems, respectively, and $u^k_t\in\{0,1\}$ represents the switching law, which is a random variable adapted to $\mathcal{F}_t$. This will be designed so that is satisfies, for almost each sample path and $t\geq0$, $u^{i}_t u^{j}_t=0$ for any $i\neq j$ and $\sum^m_{k=1}u^{k}_t=1$, namely only one external system is coupled to the target one at any given time.
The drift term and the diffusion term of the switched stochastic master equation~\eqref{Eq:SSME} satisfy a global random Lipschitz condition, the existence and uniqueness of a global strong solution of the switched system~\eqref{Eq:SSME}
can be shown by combining the arguments as in~\cite[Theorem 1.2]{krylov1999kolmogorov} and~\cite[Proposition 3.3]{mirrahimi2007stabilizing} or~\cite[Chapter 5]{barchielli2009quantum}.
\subsection{Invariant and stable subspaces}
Let $\mathcal{H}_S\subset \mathcal{H}$ be the target subspace. Denote by $\Pi_{\mathfrak{S}}\notin\{0,\mathds{1}\}$ the orthogonal projection on $\mathcal{H}_S\subset \mathcal{H}$. Define the set of density matrices
\begin{equation*}
\mathcal{I}(\mathcal{H}_S):=\{\rho\in\mathcal{S}(\mathcal{H})| \mathrm{Tr}(\Pi_{\mathfrak{S}}\rho)=1\},
\end{equation*}
namely those whose support is contained in $\mathcal{H}_S$.
\begin{definition}
For the switched system~\eqref{Eq:SSME}, the subspace $\mathcal{H}_S$ is called invariant
\begin{itemize}
\item in mean if $\rho_0\in \mathcal{I}(\mathcal{H}_S)$, $\mathbb{E}(\rho(t))\in \mathcal{I}(\mathcal{H}_S)$ for all $t>0$.
\item almost surely if $\rho_0\in \mathcal{I}(\mathcal{H}_S)$, $\rho(t)\in \mathcal{I}(\mathcal{H}_S)$ for all $t>0$ almost surely.
\end{itemize}
\end{definition}
Based on the stochastic stability defined in~\cite{khasminskii2011stochastic,mao2007stochastic} and the definition used in~\cite{ticozzi2008quantum,benoist2017exponential}, we phrase the following definition on the stochastic stability of the invariant subspace for the switched system~\eqref{Eq:SSME}. In the following definition, $\|\cdot\|$ could be any matrix norm.
\begin{definition}
Let $\mathcal{H}_S\subset\mathcal{H}$ be the invariant subspace for the switched system~\eqref{Eq:SSME}, and denote by $\Pi_{\mathfrak{S}}$ the orthogonal projection on $\mathcal{H}_S$, $\mathbf{d}_{\mathfrak{S}}(\rho):=\|\rho-\Pi_{\mathfrak{S}}\rho\Pi_{\mathfrak{S}}\|$ and $\hat{\rho}(t):=\mathbb{E}(\rho(t))$, then $\mathcal{H}_S$ is said to be
\begin{enumerate}
\item
\emph{stable in mean}, if there exists class-$\mathcal{K}$ function $\gamma$ such that,
\begin{equation*}
\mathbf{d}_{\mathfrak{S}}(\hat{\rho}(t))\leq \gamma\big(\mathbf{d}_{\mathfrak{S}}(\rho_0)\big) \text{ for } t \geq 0,
\end{equation*}
for all $\rho_0\in \mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$.
\item
\emph{globally asymptotically stable (GAS) in mean}, if it is stable in mean and,
\begin{equation*}
\lim_{t\rightarrow\infty}\mathbf{d}_{\mathfrak{S}}(\hat{\rho}(t))=0, \quad \forall \rho_0\in\mathcal{S}(\mathcal{H}).
\end{equation*}
\item
\emph{globally exponentially stable (GES) in mean}, if there exist two positive constants $\lambda$ and $c$ such that
\begin{equation*}
\mathbf{d}_{\mathfrak{S}}(\hat{\rho}(t))\leq c \,\mathbf{d}_{\mathfrak{S}}(\rho_0) e^{-\lambda t},\quad \forall \rho_0\in\mathcal{S}(\mathcal{H}),
\end{equation*}
where $\lambda$ is called the \emph{average Lyapunov exponent}.
\item
\emph{stable in probability}, if for every $\varepsilon \in (0,1)$, there exists a class-$\mathcal{K}$ function $\gamma$ such that,
\begin{equation*}
\mathbb{P} \big( \mathbf{d}_{\mathfrak{S}}(\rho(t))\leq \gamma\big( \mathbf{d}_{\mathfrak{S}}(\rho_0) \big) \text{ for } t \geq 0 \big) \geq 1-\varepsilon,
\end{equation*}
for all $\rho_0\in \mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$.
\item
\emph{globally asymptotically stable (GAS) almost surely}, if it is stable in probability and,
\begin{equation*}
\mathbb{P} \big( \lim_{t\rightarrow\infty}\mathbf{d}_{\mathfrak{S}}(\rho(t))=0 \big) = 1, \quad \forall \rho_0\in\mathcal{S}(\mathcal{H}).
\end{equation*}
\item
\emph{globally exponentially stable (GES) almost surely}, if
\begin{equation*}
\limsup_{t \rightarrow \infty} \frac{1}{t} \log \big( \mathbf{d}_{\mathfrak{S}}(\rho(t))\big) < 0,\quad \forall \rho_0\in\mathcal{S}(\mathcal{H}), \quad a.s.
\end{equation*}
The left-hand side of the above inequality is called the \emph{sample Lyapunov exponent}.
\end{enumerate}
\end{definition}
The control problem we will be concerned with is the following:\\
\textbf{Switching stochastic stabilization of subspace:}
Given a target subspace $\mathcal{H}_S\subset \mathcal{H}$ and a finite set of generators $\{\mathcal{L}_k(\rho)\}_{k\in\mathcal{M}}$, construct switching laws $u_t$ that admits a set of non-Zeno switching sequence, which ensures that $\mathcal{H}_S$ is GAS and/or GES in mean and almost surely.
\subsection{Stability for semigroup dynamics }
In this section, we recall the invariance and GAS properties of the (Lindblad-Gorini-Kossakowski-Sudarshan) master equation,
\begin{equation}
\frac{d}{dt}\hat{\rho}(t)=\mathcal{L}(\hat{\rho}(t)),\quad \rho_0\in\mathcal{S}(\mathcal{H}),
\label{Eq:MME}
\end{equation}
where $\mathcal{L}(\rho):=i[H,\rho]+\mathcal{D}_L(\rho)+\mathcal{D}_C(\rho)$ is the Lindblad generator, and the unique solution is $\{e^{t\mathcal{L}}\}_{t\geq 0}$ which consists of completely positive trace-preserving map~\cite{ticozzi2008quantum}. This equation corresponds to the semigroup dynamics associated to the expectation of a time-invariant SME of the form \eqref{Eq:SSME}.
Let $\mathcal{H}=\mathcal{H}_S\oplus\mathcal{H}_R$ and $X\in\mathcal{B}(\mathcal{H})$, the matrix representation in an appropriately chosen basis can be written as
\begin{equation*}
X=\left[\begin{matrix}
X_S & X_P\\
X_Q & X_R
\end{matrix}\right],
\end{equation*}
where $X_S,X_R,X_P$ and $X_Q$ are matrices representing operators from $\mathcal{H}_S$ to $\mathcal{H}_S$, from $\mathcal{H}_R$ to $\mathcal{H}_R$, from $\mathcal{H}_R$ to $\mathcal{H}_S$, from $\mathcal{H}_S$ to $\mathcal{H}_R$, respectively. We denote by $\Pi_{\mathfrak{R}}$ the orthogonal projection on $\mathcal{H}_R\subset\mathcal{H}$. The invariance and GAS properties of the master equation~\eqref{Eq:MME} correspond directly to the structure of Lindblad generator. For the reader convenience, we summarize some useful results found in~\cite{ticozzi2008quantum,ticozzi2009analysis,benoist2017exponential}.
\begin{theorem}
For the system~\eqref{Eq:MME}, the subspace $\mathcal{H}_S$ is
\begin{enumerate}
\item invariant if and only if $L_Q=C_Q=0$ and $iH_P-\frac{1}{2}(L^*_SL_P+C^*_SC_P)=0$;
\item GAS if and only if it is invariant and no invariant subspaces are included in $\mathrm{ker}(L_{P})\cap\mathrm{ker}(C_{P})$.
\end{enumerate}
\label{Thm:InvarinceGAS}
\end{theorem}
We define the following map,
\begin{equation*}
\mathcal{L}_{R}(\rho_R):=-i[H_{R},\rho_R]+\mathsf{D}_{L}(\rho_R)+\mathsf{D}_{C}(\rho_R),
\end{equation*}
where $\mathsf{D}_{A}(\rho_R):=A_{R}\rho_R A^*_{R}-\frac{1}{2}\{A^*_{P}A_{P}+A^*_{R}A_{R},\rho_R\}$ and $\rho_R\in\mathcal{B}_{\geq 0}(\mathcal{H}_R)$.
\begin{lemma}
Suppose that $\mathcal{H}_S$ is invariant with respect to the system~\eqref{Eq:MME}. The family $\{e^{t\mathcal{L}_{R}}\}_{t\geq 0}$ is a semigroup of trace non-increasing completely positive maps. Moreover, for all $\rho_0\in\mathcal{S}(\mathcal{H})$, $\hat{\rho}_R(t)=e^{t\mathcal{L}_R}\rho_{0,R}.$
\label{Lemma:LR_TrNonIncreasing}
\end{lemma}
For $R$-block of any Lindblad generator $\mathcal{L}$, we denote by $\mathcal{L}^*_{R}$ the adjoint of $\mathcal{L}_{R}$ with respect to the Hilbert-Schmidt inner product on $\mathcal{B}(\mathcal{H}_R)$. we recall the following results on quantum Markovian dynamics~\cite{benoist2017exponential,evans1977spectral} concerning on the spectral property of $\mathcal{L}^*_R$ and the relation with the GAS in mean with respect to the Lindblad generator $\mathcal{L}$.
\begin{theorem}
Suppose that $\mathcal{H}_S$ is invariant with respect to the system~\eqref{Eq:MME}. For any $\epsilon>0$, there exists $K_R\in\mathcal{B}_{>0}(\mathcal{H}_R)$ such that
$
\mathcal{L}_R^*(K_R)\leq -(\alpha-\epsilon)K_R,
$
where $\alpha$ is the spectral abscissa of $\mathcal{L}_R$, i.e., $\alpha:=\min\{-\mathbf{Re}(\lambda)|\,\lambda\in\mathrm{sp}(\mathcal{L}_{R})\}$. Moreover, $\mathcal{H}_S$ is GAS for the system~\eqref{Eq:MME} if and only if $\alpha>0$.
\label{Thm:SpectralAbscissa}
\end{theorem}
\medskip
Based on the block-decomposition with respect to the orthogonal direct sum decomposition $\mathcal{H}=\mathcal{H}_S\oplus\mathcal{H}_R$, for any $X_R\in\mathcal{B}(\mathcal{H}_R)$, we call the following matrix the {\em extension of $X_R$ to} $\mathcal{B}(\mathcal{H})$,
\begin{equation*}
X=\left[\begin{matrix}
0 & 0\\
0 & X_R
\end{matrix}\right]\in\mathcal{B}(\mathcal{H}).
\end{equation*}
In order to quantify the distance between $\rho\in\mathcal{S}(\mathcal{H})$ and $\mathcal{I}(\mathcal{H}_S)$ we shall make use of linear functions associated to a positive $X_R\in\mathcal{B}_{>0}(\mathcal{H}_R)$, namely
$$
\mathrm{Tr}(X \rho)=\mathrm{Tr}(X_R \rho_R)\in[0,1],
$$
where $X$ is the extension in $\mathcal{B}(\mathcal{H})$ of $X_R$. Such function is used as an estimation of the distance $\mathbf{d}_{\mathfrak{S}}(\rho)$.
\begin{lemma}
For all $\rho\in\mathcal{S}(\mathcal{H})$ and the orthogonal projection $\Pi_{\mathfrak{S}}\in\mathcal{B}(\mathcal{H})$ on $\mathcal{H}_S$, there exist two constants $c_1>0$ and $c_2>0$ such that
\begin{equation}
c_1\mathrm{Tr}(X \rho)\leq \|\rho-\Pi_{\mathfrak{S}}\rho \Pi_{\mathfrak{S}}\| \leq c_2 \sqrt{\mathrm{Tr}(X \rho)},
\label{Eq:Relation_DisLya}
\end{equation}
where $X$ is the extension in $\mathcal{B}(\mathcal{H})$ of $X_R$.
\label{Lemma:Relation_DisLya}
\end{lemma}
\emph{Proof.}
Firstly, let us consider the special case $\mathrm{Tr}(\Pi_{\mathfrak{R}} \rho)=\mathrm{Tr}(\rho_R)$. By employing the arguments in the proof of~\cite[Lemma 4.8]{benoist2017exponential}, we have
\begin{equation*}
\|\rho-\Pi_{\mathfrak{S}} \rho \Pi_{\mathfrak{S}}\|_{1}\leq N\|\rho-\Pi_{\mathfrak{S}} \rho \Pi_{\mathfrak{S}}\|_{\max}\leq 3N \sqrt{\mathrm{Tr}(\Pi_{\mathfrak{R}} \rho)}
\end{equation*}
where $N=\mathrm{dim}(\mathcal{H})$, $\|\cdot\|_{1}$ and $\|\cdot\|_{\max}$ represents the trace norm and the max norm respectively. Moreover, since the trace norm is unitarily invariant, we have the following pinching inequality~\cite[Chapter 4.2]{bhatia2013matrix}
\begin{align*}
\|\rho-\Pi_{\mathfrak{S}} \rho \Pi_{\mathfrak{S}}\|_{1}&\geq\| \Pi_{\mathfrak{S}} (\rho-\Pi_{\mathfrak{S}} \rho \Pi_{\mathfrak{S}})\Pi_{\mathfrak{S}}+\Pi_{\mathfrak{R}}(\rho-\Pi_{\mathfrak{S}} \rho \Pi_{\mathfrak{S}})\Pi_{\mathfrak{R}}\|_{1}\\
&=\|\Pi_{\mathfrak{R}}\rho\Pi_{\mathfrak{R}}\|_{1}=\mathrm{Tr}(\Pi_{\mathfrak{R}} \rho).
\end{align*}
Then, the equivalence of matrix norms on finite dimensional vector spaces and the equivalence of $\mathrm{Tr}(\rho_R)$ and $\mathrm{Tr}(X_R \rho_R)$ conclude the proof.
\hfill$\square$
\section{Exponential stabilization of the target subspace}
\label{Sec:GES}
For practical implementations, it is crucial to avoid chattering and Zeno-type phenomena~\cite[Chapter 1.2]{liberzon2003switching} in the switching design. Hysteresis switching technique and dwell time are useful control designs to prevent such undesirable behaviors. In the following, we show switching algorithms ensuring GES of the target subspace $\mathcal{H}_S$ under the hysteresis switching and dwell time technique. We start algorithms whose switching sequence can be computed off-line, being based only on the average state evolution which does not depend on the measurement outcomes.
\subsection{ Switching strategy based on the average state}
\label{Subsec:Average}
Here, we reconsider the switching algorithm and control assumptions proposed in~\cite[Theorem 2]{scaramuzza2015switching} for the average dynamics, and show that the same strategy implies GES in mean and almost surely of $\mathcal{H}_S$ for the switching SME ~\eqref{Eq:SSME}. The strategy is based on the following invariance and exponential Lyapunov-like assumptions:
\begin{description}
\item[\textbf{A1.1}:] $\mathcal{H}_S$ is invariant with respect to $\mathcal{L}_{k}$ for all $k\in\mathcal{M}$.
\item[\textbf{A1.2}:]
There exist $K_R\in\mathcal{B}_{>0}(\mathcal{H}_R)$ and $c>0$ such that
for all $\rho\in\mathcal{S}(\mathcal{H})$,
$\mathbf{L}_K(\rho)\leq -c\mathrm{Tr}(K\rho)$, where $K$ is the extension in $\mathcal{B}(\mathcal{H})$ of $K_R$ and $\mathbf{L}_K(\rho):=\min_{k\in\mathcal{M}}\mathrm{Tr}(K\mathcal{L}_k(\rho))$.
\end{description}
The second assumption is a generalization of the typical assumption for switching stabilizing linear systems, namely the existence of a convex combination of generators that is stabilizing, as it will be made explicit in Corollary \ref{Cor:GES_convex} below. Now, we formulate the average state dependent switching law inspired by~\cite{scaramuzza2015switching} and~\cite[Chapter 3.4]{sun2006switched}. Suppose that \textbf{A1} hold.
For all $j\in\mathcal{M}$, we define the region
\begin{equation}
\Delta_j=\{\rho\in\mathcal{S}(\mathcal{H})|\,{\rm Tr}(K\mathcal{L}_j(\rho))\leq -\epsilon c {\rm Tr}(K\rho)\},
\label{Eq:Region_1}
\end{equation}
where the constant $\epsilon\in(0,1)$ are used to control the dwell-time and the lower bound of the convergence rate. Then, we have
$
\bigcup_{j\in\mathcal{M}}\mathrm{int}(\Delta_j)=\mathcal{S}(\mathcal{H})\setminus\mathcal{I}(\mathcal{H}_S).
$
Otherwise, there exists a $\rho\in\mathcal{S}(\mathcal{H})\setminus\mathcal{I}(\mathcal{H}_S)$ such that ${\rm Tr}(K\mathcal{L}_j(\rho))\geq -\epsilon c {\rm Tr}(K\rho)> -c {\rm Tr}(K\rho)$ for all $j\in\mathcal{M}$, which leads to a contradiction. Then, we define the switching algorithm $\sigma_1$ based on the average state $\hat{\rho}$.
\begin{definition}[Switching algorithm $\sigma_1$]
For any initial state $\rho_0\in\mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$, set $t_0=0$ and for all $n\in\mathbb{N}$,
\begin{align*}
&p_{n}=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}\big(K\mathcal{L}_k(\hat{\rho}(t_n))\big),\\
&u^{p_{n}}_{t_{n}}=1 \text{ and } u^{k}_{t_{n}}=0, \quad \forall\, k\neq p_{n},\\
&t_{n+1}=\inf\{t\geq t_n|\, \hat{\rho}(t)\notin \mathrm{int}(\Delta_{p_n})\},\\
&u_t \equiv u_{t_n}, \quad \forall\, t\in[t_n,t_{n+1}).
\end{align*}
where throughout this paper we set $\inf\{\emptyset\}=\infty$, and if several Lindbladians are active we choose the one with the minimum index.
\end{definition}
Note that the above constructed switching instants $\{t_n\}$, switching laws $u_t$ are non-random, and $t_{n+1}>t_n$ since the overlap of each adjacent open regions. Denote by $\bar{n}$ by the total number of switches, which may be infinity. In the following theorem, we show that $\mathcal{H}_S$ is GES in mean and almost surely under the switching algorithm. Moreover, we provide a lower bound of the dwell time of switching sequence.
Before stating the result, for any $K_R\in\mathcal{B}_{>0}(\mathcal{H}_R)$ and $j\in \mathcal{M}$, we define the following,
{\small
\begin{align*}
&\underline{l}_j(K_R):=\sup\{\lambda\in\mathbb{R}|\mathcal{L}^*_{j,R}(K_R)\geq \lambda K_R\}, \bar{l}_j(K_R):=\inf\{\lambda\in\mathbb{R}|\mathcal{L}^*_{j,R}(K_R)\leq \lambda K_R\},\\
&\underline{l}_{j,2}(K_R):=\sup\{\lambda\in\mathbb{R}|{\mathcal{L}^*}^2_{j,R}(K_R)\geq \lambda K_R\},\bar{l}_{j,2}(K_R):=\inf\{\lambda\in\mathbb{R}|{\mathcal{L}^*}^2_{j,R}(K_R)\leq \lambda K_R\},\\
&l_j(K_R):=\max\{|\bar{l}_j(K_R)|,|\underline{l}_j(K_R)|\},l_{j,2}(K_R):=\max\{|\bar{l}_{j,2}(K_R)|,|\underline{l}_{j,2}(K_R)|\}.
\end{align*}
}
\begin{theorem}
Suppose that \emph{\textbf{A1.1}} and \emph{\textbf{A1.2}} hold true. Then, for the switched system~\eqref{Eq:SSME} under the switching algorithm $\sigma_1$, $\mathcal{H}_S$ is GES in mean and almost surely with the Lyapunov exponent less than or equal to $-\epsilon c/2$.
\label{Thm:GES_average}
\end{theorem}
\emph{Proof.}
For all $\rho_0\in\mathcal{I}(\mathcal{H}_S)$, the results hold trivially since the invariance property ensured by \textbf{A1.1}. Let us suppose that $\rho_0\in\mathcal{S}(\mathcal{H})\setminus\mathcal{I}(\mathcal{H}_S)$. First, we show that the switching instants are well-defined, i.e., there exists a constant $t_D>0$ such that $t_{n+1}-t_n>t_D$ for all $n$. For an arbitrary $n$ such that $t_{n+1}<\infty$, due to the assumption \textbf{A1.2} and the definition of $\sigma_1$, we have
\begin{align*}
&{\rm Tr}\big(K(\mathcal{L}_{p_n}(\hat{\rho}(t_n))\big)\leq -c {\rm Tr}(K\hat{\rho}(t_{n})),\,{\rm Tr}\big(K\mathcal{L}_k(\hat{\rho}(t_{n+1}))\big)= -\epsilon c{\rm Tr}(K\hat{\rho}(t_{n+1})),
\end{align*}
where $\epsilon\in(0,1)$ and $c>0$ is defined in \textbf{A1.2}.
For all $t\in[t_n,t_{n+1}]$, we define
$$
g(t):={\rm Tr}\big(K\mathcal{L}_{p_n}(\hat{\rho}(t))\big)+\epsilon c {\rm Tr}(K\hat{\rho}(t)).
$$
One deduces $g(t_n)\leq -c(1-\epsilon){\rm Tr}(K\hat{\rho}(t_{n}))<0$ and $g(t_{n+1})=0$. Moreover, we have
\begin{align*}
|\dot{g}(t)|&\leq \big|\mathrm{Tr}\big((K_R\mathcal{L}^2_{p_n,R}(\hat{\rho}(t)))\big|+\epsilon c\big|\mathrm{Tr}\big(K_R\mathcal{L}_{p_n,R}(\hat{\rho}(t))\big)\big|\\
&\leq \big( l_{p_n,2}(K_R)+\epsilon c l_{p_n}(K_R)\big)\mathrm{Tr}(K\hat{\rho}(t)).
\end{align*}
Due to the mean value theorem, there exits $t^*\in[t_n,t_{n+1}]$ such that
\begin{equation*}
\dot{g}(t^*)(t_{n+1}-t_n)=g(t_{n+1})-g(t_n)\geq c(1-\epsilon){\rm Tr}(K\hat{\rho}(t_{n})).
\end{equation*}
By using Lemma~\ref{Lemma:LR_TrNonIncreasing}, we have
\begin{align*}
{\rm Tr}(K\hat{\rho}(t_{n}))&\geq \underline{\lambda}(K_R){\rm Tr}(\hat{\rho}_R(t_{n}))\geq \underline{\lambda}(K_R){\rm Tr}(\hat{\rho}_R(t^*))\geq \frac{\underline{\lambda}(K_R)}{\bar{\lambda}(K_R)}{\rm Tr}(K\hat{\rho}(t^*)),
\end{align*}
where $\bar{\lambda}(K_R)>0$ and $\underline{\lambda}(K_R)>0$ denote the maximum and minimum eigenvalue of $K_R$ respectively.
It implies
\begin{align*}
t_{n+1}-t_n&\geq \frac{c(1-\epsilon)}{l_{p_n,2}(K_R)+\epsilon c l_{p_n}(K_R)}\frac{{\rm Tr}(K\hat{\rho}(t_{n}))}{{\rm Tr}(K\hat{\rho}(t^*))}\\
&\geq \frac{c(1-\epsilon)}{l_{p_n,2}(K_R)+\epsilon c l_{p_n}(K_R)}\frac{\underline{\lambda}(K_R)}{\bar{\lambda}(K_R)}>0.
\end{align*}
Therefore, we have a lower bound of the dwell-time of the switching sequence given by
\begin{equation}
t_D= \min_{k\in\mathcal{M}} \Big\{ \frac{c(1-\epsilon)}{l_{k,2}(K_R)+\epsilon c l_{k}(K_R)} \frac{\underline{\lambda}(K_R)}{\bar{\lambda}(K_R)}\Big\}>0,
\label{Eq:DwellTime}
\end{equation}
which is negatively correlated to the value of $\epsilon$.
Now, we show $\mathcal{H}_S$ is GES in mean. Based on the definition of switching algorithm, we have,
\begin{align*}
{\rm Tr}(K\hat{\rho}(t))&={\rm Tr}(K\rho_0)+\sum^{\bar{n}}_{n=0}\int^{t\wedge t_{n+1}}_{t\wedge t_{n}}{\rm Tr}\big(K\mathcal{L}_{p_n}(\hat{\rho}(s))\big)ds\\
&\leq {\rm Tr}(K\rho_0)-\epsilon c\int^t_0{\rm Tr}(K\hat{\rho}(s))ds.
\end{align*}
By using the Gr\"onwall's inequality, it follows ${\rm Tr}(K\hat{\rho}(t))\leq {\rm Tr}(K\rho_0)e^{-\epsilon ct}$. Combining with the relation~\eqref{Eq:Relation_DisLya}, one deduces that $\mathcal{H}_S$ is GES in mean with average Lyapunov exponent less than or equal to $-\epsilon c/2$.
Set $Q_t=e^{\epsilon c t}{\rm Tr}(K\rho(t))$, for all $t\geq s\geq 0$, we have
\begin{align*}
\mathbb{E}\big(Q_t\big|\mathcal{F}_s\big)-Q_s=\sum^{\bar{n}}_{n=0}&\int^{t\wedge t_{n+1}}_{s\wedge t_{n}}e^{-\epsilon c u}\mathbb{E}\big( {\rm Tr}(K\mathcal{L}_{p_n}\rho(u))+\epsilon c{\rm Tr}(K\rho(u))\big|\mathcal{F}_s\big)du\leq 0.
\end{align*}
Thus, $Q_t$ is a positive supermartingale. Due to Doob's martingale convergence theorem~\cite[Corollary 2.11]{revuz2013continuous}, $\sup_{t\geq 0}Q_t=R(\omega)$ almost surely, where $R(\omega)$ is a finite random variable. Moreover, due to Lemma~\ref{Lemma:NeverReach}, it is straightforward to show that $\mathbb{P}\big( \mathrm{Tr}(K\rho(t))>0,\,\forall t\geq0\big)=1$ for all $\rho_0\in\mathcal{S}(\mathcal{H})\setminus\mathcal{I}(\mathcal{H}_S)$. Then, we can deduce $\limsup_{t\rightarrow\infty}\frac{1}{t}\log {\rm Tr}(K\rho(t))\leq -\epsilon c$ almost surely. Combining with the relation~\eqref{Eq:Relation_DisLya}, $\mathcal{H}_S$ is GES almost surely with sample Lyapunov exponent less than or equal to $-\epsilon c/2$.
\hfill$\square$
\begin{remark}
The division of the state space does not satisfy the classical hysteresis technique~\cite{morse1992applications}, since the target subset $\mathcal{I}(\mathcal{H}_S)$ is located at the boundary of all region $\Delta_j$. Due to the bounded convergence rate of $\hat{\rho}(t)$ in each region $\Delta_j$ and the invariance properties of the Lindblad generators (Lemma~\ref{Lemma:LR_TrNonIncreasing}), we can provide a lower bound of the dwell time. However, we cannot determine if the total number of switches $\bar{n}$ is finite or infinite.
\end{remark}
Based on lower bound of dwell time $t_D$ defined in Equation~\eqref{Eq:DwellTime}, we recall the following switching algorithm $\sigma_2$ with fixed dwell time, which is defined in~\cite[Theorem 2]{scaramuzza2015switching}
\begin{definition}[Switching algorithm $\sigma_2$]
For any initial state $\rho_0\in\mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$, fix the dwell time $\Delta t\in(0,t_D]$ where $t_D$ is defined in~\eqref{Eq:DwellTime}, set $t_0 = 0$ and for all $n\in \mathbb{N}$,
\begin{align*}
&p_{n}=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}\big(K\mathcal{L}_k(\hat{\rho}(t_n))\big),\\
&t_{n+1} = t_n+\Delta t,\\
&u^{p_{n}}_{t}=1 \text{ and } u^{k}_{t}=0, \quad \forall\, k\neq p_{n}, \, t\in[t_n,t_{n+1}).
\end{align*}
\end{definition}
\smallskip
\begin{corollary}
Suppose that \emph{\textbf{A1.1}} and \emph{\textbf{A1.2}} hold true. Then, for the switched system~\eqref{Eq:SSME} under the switching algorithm $\sigma_2$, $\mathcal{H}_S$ is GES in mean and almost surely with the Lyapunov exponent less than or equal to $-\epsilon c/2$.
\label{Cor:GES_Dwell}
\end{corollary}
\begin{corollary}
Suppose that \emph{\textbf{A1.1}} holds true and there exists $\gamma\in\{\gamma\in(0,1)^m| \sum_{j\in\mathcal{M}}\gamma_j=1\}$ such that $\mathcal{H}_S$ is GAS in mean for
$
\frac{d}{dt}\hat{\rho}(t)=\mathcal{L}_{\gamma}(\hat{\rho}(t))=\sum_{j\in\mathcal{M}}\gamma_j\mathcal{L}_{j}(\hat{\rho}(t)),
$
with $\rho_0\in\mathcal{S}(\mathcal{H})$.
Then, for the switched system~\eqref{Eq:SSME} under the switching algorithm $\sigma_1$ or $\sigma_2$, $\mathcal{H}_S$ is GES in mean and almost surely.
\label{Cor:GES_convex}
\end{corollary}
\emph{Proof.}
Due to Theorem~\ref{Thm:SpectralAbscissa}, the spectral abscissa of $\mathcal{L}_{\gamma}$ denoted by $\alpha_{\gamma}$ is strictly positive. For any $\delta\in(0,\alpha_{\gamma})$, there exists $K_R\in\mathcal{B}_{>0}(\mathcal{H}_R)$ such that $\mathcal{L}^*_{\gamma,R}(K_R)\leq-(\alpha_{\gamma}-\delta)K_R$. Then, we have
$
\min_{j\in\mathcal{M}}{\rm Tr}(K\mathcal{L}_j(\rho))\leq {\rm Tr}(K\mathcal{L}_{\gamma}(\rho))\leq -(\alpha_{\gamma}-\delta){\rm Tr}(K\rho),
$
where $K$ is the extension in $\mathcal{B}(\mathcal{H})$ of $K_R$.
Then \textbf{A1.2} is satisfied. By applying Theorem~\ref{Thm:GES_average} and Corollary~\ref{Cor:GES_Dwell}, $\mathcal{H}_S$ is GES in mean and almost surely.
\hfill$\square$
\subsection{Measurement-dependent switching strategies}
In subsection~\ref{Subsec:Average}, we provide a switching algorithms ensuring GES of the target subspace $\mathcal{H}_S$ based on the average state evolution. However, by doing so, we do not use the information made available from the measurement output to the fullest. Inspired by~\cite[Theorem 2]{grigoletto2021stabilization} and Theorem~\ref{Thm:GES_average}, we propose two state dependent switching algorithms to guarantee GES of the target subspace $\mathcal{H}_S$ in mean and almost surely. While both aim to select the fastest convergence rate, for each realization, the first one operates with a fixed dwell time, as obtained in the previous results, while the second one can reduce the number of switches by introducing the latter as suitably defined stochastic times.
We define the switching algorithm $\sigma_2$ with fixed dwell time $\Delta t\in(0,t_D]$ based on the state $\rho(t)$, where $t_D>0$ is defined in~\eqref{Eq:DwellTime}.
\begin{definition}[Switching algorithm $\sigma_3$]
For any initial state $\rho_0\in\mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$, fix the dwell time $\Delta t\in(0,t_D]$ and set $t_0=0$ and for all $n\in\mathbb{N}$,
\begin{align*}
&p_n(\omega)=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}(K\mathcal{L}_k(\rho(t_n))), \\
&t_{n+1}=t_n+\Delta t,\\
&u^{k}_{t}(\omega)=\mathds{1}_{\{p_n(\omega)=k\}}, \quad \forall\, k\in\mathcal{M},\,t\in[t_n,t_{n+1}).
\end{align*}
where for each $\omega$, if several $p_n(\omega)$ is active, we choose the minimum, and $u_t$ is $\mathcal{F}_{t_{n}}$-adapted for $t\in[t_n,t_{n+1})$.
\end{definition}
\begin{theorem}
Suppose that \emph{\textbf{A1.1}} and \emph{\textbf{A1.2}} hold true. Then, for the switched system~\eqref{Eq:SSME} under the switching algorithm $\sigma_3$, $\mathcal{H}_S$ is GES in mean and almost surely with the Lyapunov exponent less than or equal to $-\epsilon c/2$.
\label{Thm:GES_dwell}
\end{theorem}
\emph{Proof.}
For any $n\in\mathbb{N}$ and for all $t\in[t_n,t_{n+1})$, by the construction, the switching law $u_t$ is adapted to $\mathcal{F}_{t_n}$ and
\begin{equation*}
\mathbb{E}(\rho(t)|\mathcal{F}_{t_n})-\rho(t_n)=\mathbb{E}\Big( \int^t_{t_n} \sum^m_{k=1}u^k_s\mathcal{L}_k(\rho(s))ds \Big| \mathcal{F}_{t_n} \Big),\, a.s.
\end{equation*}
Set $\Lambda^{k}_{n}:=\{\omega\in\Omega|\, p_n(\omega)=k\}$ with $k\in \mathcal{M}$, then $\mathbb{P}(\bigcup_{k\in\mathcal{M}}\Lambda^{k}_{n})=1$ and $\mathbb{P}(\Lambda^{i}_{n}\cap \Lambda^{j}_{n})=0$ for $i\neq j$. Thus, for almost all $\omega \in \Lambda^{k}_{n}$, due to the linearity of $\mathcal{L}_k$, we have,
\begin{equation*}
\mathbb{E}(\rho(t)|\mathcal{F}_{t_n})-\rho(t_n)=\int^t_{t_n} \mathcal{L}_k\big( \mathbb{E}(\rho(s)|\mathcal{F}_{t_n}) \big)ds,
\end{equation*}
It follows that, for almost all $\omega \in \Lambda^{k}_{n}$ and $t\in[t_n,t_{n+1})$, $\mathbb{E}(\rho(t)|\mathcal{F}_{t_n})$ solves the master equation generated by $\mathcal{L}_k$.
By using the similar arguments as in the proof of Theorem~\ref{Thm:GES_average} and Corollary~\ref{Cor:GES_Dwell}, we have
$$
{\rm Tr}\big(K\mathcal{L}_{k}(\mathbb{E}(\rho(t)| \mathcal{F}_{t_n}))\big)\leq -\epsilon c {\rm Tr}\big(K\mathbb{E}(\rho(t)| \mathcal{F}_{t_n})\big)
$$
on the set $\Lambda^{k}_{n}$. Since $k$ is chosen arbitrarily, we have
\begin{align*}
&\mathbb{E}\big({\rm Tr}(K\rho(t)) \big| \mathcal{F}_{t_n} \big)-{\rm Tr}(K\rho(t_n))\leq -\epsilon c \,\mathbb{E}\Big(\int^t_{t_n}{\rm Tr}(K\rho(s))ds \Big| \mathcal{F}_{t_n} \Big), \quad a.s.
\end{align*}
Due to the law of total expectation, we have
\begin{equation*}
{\rm Tr}(K\hat{\rho}(t))-{\rm Tr}(K\hat{\rho}(t_n))\leq -\epsilon c \int^t_{t_n}{\rm Tr}(K\hat{\rho}(s))ds
\end{equation*}
Combine each switching interval together, then apply the Gr\"onwall's inequality and the relation~\eqref{Eq:Relation_DisLya}, one deduces that $\mathcal{H}_S$ is GES in mean with average Lyapunov exponent less than or equal to $-\epsilon c/2$.
Denote $Q_t=e^{\epsilon c t}{\rm Tr}(K\rho(t))$. For all $0\leq s \leq t<\infty$, there exist $l,d\in\mathbb{N}$ such that $t_{l}\leq s\leq t_{l+1}\leq \dots \leq t_{l+d} \leq t$. By It\^o formula~\cite[Theorem 2.32]{protter2004stochastic}, we have
\begin{equation*}
\mathbb{E}(Q_t|\mathcal{F}_s)-Q_s=\sum^{l+d}_{j=l}I_s(s\vee t_{j},t\wedge t_{j+1}),\quad a.s.
\end{equation*}
where $I_s(\alpha,\beta):=\mathbb{E}\big(\int^{\beta}_{\alpha} e^{\epsilon c r}\big(\sum^m_{k=1}u^k_r{\rm Tr}\big(K\mathcal{L}_k(\rho(r))\big)+\epsilon c {\rm Tr}(K\rho(r)) \big)dr \big|\mathcal{F}_s \big)$. Based on the above arguments, for almost each $\omega\in\Lambda^k_n$ with $k\in\mathcal{M}$ and $n\in\{l,\dots,l+d\}$, we have
$I_s(\alpha,\beta)\leq 0$. Thus, $Q_t$ is a positive supermartingale. By employing the similar arguments in the proof of Theorem~\ref{Thm:GES_average}, $\mathcal{H}_S$ is GES almost surely with sample Lyapunov exponent less than or equal to $-\epsilon c/2$.
\hfill$\square$
\medskip
Next, we define the switching algorithm $\sigma_4$ based on the state $\rho(t)$.
For any initial state $\rho_0\in\mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$, set $\tau_0=0$ and
\begin{align*}
&p_0=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}(K\mathcal{L}_k(\rho_0)), \\
&u^{p_0}_0=1 \text{ and } u^{k}_0=0, \quad \forall\, k\neq p_0.
\end{align*}
The solution of the switched stochastic master equation~\eqref{Eq:SSME} under the switching law $u_t\equiv u_0$ is well-defined on $t\in[\tau_0,\infty)$. Define the stopping time
\begin{equation*}
\tau_1:=\inf\{t\geq \tau_0|\, \rho(t)\notin \mathrm{int}(\Delta_{p_0}) \},
\end{equation*}
where $\Delta_j=\{\rho\in\mathcal{S}(\mathcal{H})|\,{\rm Tr}(K\mathcal{L}_j(\rho))\leq -\epsilon c {\rm Tr}(K\rho)\}$ with $j\in\mathcal{M}$.
Then, for any finite $T\geq 0$, we define
\begin{equation*}
\Omega^1_T:=\{\omega\in\Omega|\, \tau_1<T\}, \, \bar{\Omega}^1_T:=\{\omega\in\Omega|\, \tau_1\geq T\}.
\end{equation*}
Obviously, we have $\mathbb{P}(\Omega^1_T\cup \bar{\Omega}^1_T)=1$ and $\mathbb{P}(\Omega^1_T\cap \bar{\Omega}^1_T)=0$.
For almost each $\omega \in \bar{\Omega}^1_T$, the trajectory of the system stays in $\mathrm{int}(\Delta_{p_0})$ without switching till $T$. For all $\omega \in \Omega^1_T$, we define
\begin{align*}
&p_1(\omega)=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}\big(K \mathcal{L}_k (\rho(\tau_1))\big),\\
&u^{k}_t(\omega)=\mathds{1}_{\{p_1(\omega)=k\}}, \quad \forall\, k\in\mathcal{M},\, t\in[\tau_1,T).
\end{align*}
The second switching instant is defined as
\begin{equation*}
\tau_2(\omega):=\inf\{t\geq \tau_1|\, \rho(t)\notin \mathrm{int}(\Delta_{p_1(\omega)}) \}.
\end{equation*}
We denote
\begin{equation*}
\Omega^2_T:=\{\omega\in\Omega^1_T|\, \tau_2<T\}, \, \bar{\Omega}^2_T:=\{\omega\in\Omega^1_T|\, \tau_2\geq T\},
\end{equation*}
which follows $\mathbb{P}(\Omega^2_T\cup \bar{\Omega}^2_T)=\mathbb{P}(\Omega^1_T)$ and $\mathbb{P}(\Omega^2_T\cap \bar{\Omega}^2_T)=0$.
Then, we can define the switching laws at switching instants $\tau_n$ and $\Omega^n_T$ and $\bar{\Omega}^n_T$ for all $n\in\mathbb{N}$ recursively.
\begin{definition}[Switching algorithm $\sigma_4$]
For any initial state $\rho_0\in\mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$, set $\tau_0=0$ and for all $n\in\mathbb{N}$,
\begin{align*}
&p_{n}(\omega)=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}\big(K \mathcal{L}_k (\rho(\tau_{n}))\big),\\
&u^{k}_{\tau_n}(\omega)=\mathds{1}_{\{p_{n}(\omega)=k\}}, \quad \forall\, k\in\mathcal{M},\\
&\tau_{n+1}(\omega):=\inf\{t\geq \tau_n|\, \rho(t)\notin \mathrm{int}(\Delta_{p_{n}(\omega)}) \},\\
&u_t(\omega)\equiv u_{\tau_n}(\omega),\quad \forall\, t\in[\tau_n,\tau_{n+1}).
\end{align*}
\end{definition}
\smallskip
Due to the non-empty overlap of each adjacent open regions $\mathrm{int}(\Delta_j)$, the continuity of the solution $\rho(t)$ for almost each sample path on $t\in[0,T]$, and Lemma~\ref{Lemma:NeverReach} which implies the almost sure inaccessibility of $\mathcal{I}(\mathcal{H}_S)$ in finite time, one deduces that the number of switches before any finite time $T$ is finite almost surely, which is denoted by $\bar{n}_T(\omega)<\infty$.
For any positive constant $T<\infty$, the switching law $(u_t)_{t\in[0,T]}$ is adapted to $\mathcal{F}_T$, the solution of switching solution~\eqref{Eq:SSME} is well-defined.
\begin{theorem}
Suppose that \emph{\textbf{A1.1}} and \emph{\textbf{A1.2}} hold true. Then, for the switched system~\eqref{Eq:SSME} under the switching algorithm $\sigma_4$, $\mathcal{H}_S$ is GES in mean and almost surely with the Lyapunov exponent less than or equal to $-\epsilon c/2$.
\label{Thm:GES_state}
\end{theorem}
\emph{Proof.}
Fix an arbitrary positive constant $T<\infty$ and $n\in\mathbb{N}$. Suppose $\rho_0\in\mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$, for almost each $\omega\in\bar{\Omega}^n_T$, we have
\begin{align*}
\int^T_0 \sum^m_{k=1} u^k_{s}(\omega) {\rm Tr}(K\mathcal{L}_k \rho(s))ds &=\sum^{n-1}_{j=0}\int^{T\wedge \tau_{j+1}(\omega)}_{\tau_{j}(\omega)} \sum^m_{k=1} u^k_{s}(\omega) {\rm Tr}(K\mathcal{L}_k \rho(s))ds\\
&\leq -\epsilon c \int^T_0 {\rm Tr}(K\rho(s))ds.
\end{align*}
Since $\bar{n}_T<\infty$ almost surely, $\mathbb{P}(\lim_{n\rightarrow \infty} \bigcup^n_{i=1}\bar{\Omega}^i_T) =1$. Due to It\^o isometry, we obtain
\begin{align*}
&{\rm Tr}(K\hat{\rho}(T)) -{\rm Tr}(K\rho_0)=\mathbb{E}\Big( \int^T_0 \sum^m_{k=1} u^k_{s}(\omega) {\rm Tr}(K\mathcal{L}_k \rho(s))ds \Big)\\
&=\mathbb{E}\Big( \lim_{n\rightarrow \infty} \sum^n_{i=1}\mathds{1}_{\bar{\Omega}^i_T}\int^T_0 \sum^m_{k=1} u^k_{s}(\omega) {\rm Tr}(K\mathcal{L}_k \rho(s))ds \Big)\\
&\leq -\epsilon c\,\mathbb{E}\Big( \lim_{n\rightarrow \infty} \sum^n_{i=1}\mathds{1}_{\bar{\Omega}^i_T}\int^T_0 {\rm Tr}(K\rho(s))ds \Big)= -\epsilon c\, \int^T_0 {\rm Tr}(K\hat{\rho}(s))ds
\end{align*}
By applying the Gr\"onwall's inequality, in addition to the relation~\eqref{Eq:Relation_DisLya}, one deduces that $\mathcal{H}_S$ is GES in mean with average Lyapunov exponent less than or equal to $-\epsilon c/2$.
Denote $Q_t=e^{\epsilon c t}{\rm Tr}(K\rho(t))$. For all $0\leq s \leq t<\infty$, by It\^o formula, we have
\begin{equation*}
\mathbb{E}(Q_t|\mathcal{F}_s)-Q_s=I_s(s,t),\quad a.s.
\end{equation*}
where
\begin{align*}
I_s(s,t):=\mathbb{E}\big(\int^{t}_{s} e^{\epsilon c r}\big(&\sum^m_{k=1}u^k_r{\rm Tr}\big(K\mathcal{L}_k(\rho(r))\big)+\epsilon c {\rm Tr}(K\rho(r)) \big)dr \big|\mathcal{F}_s \big).
\end{align*}
Based on the above arguments, for almost all $\omega\in\bar{\Omega}^k_t$ with $k\in\mathcal{M}$, we have $I_s(s,t)\leq 0$. Thus, $Q_t$ is a supermartingale since $\mathbb{P}(\lim_{n\rightarrow \infty} \bigcup^n_{i=1}\bar{\Omega}^i_T)=1$. By employing the similar arguments in the proof of Theorem~\ref{Thm:GES_average}, $\mathcal{H}_S$ is GES almost surely with sample Lyapunov exponent less than or equal to $-\epsilon c/2$.
\hfill$\square$
\section{Asymptotic stabilization of target subspace by modulating Lindbladian}
\label{Sec:Modulate}
From the practical point of view, the assumption \textbf{A1.1} and \textbf{A1.2} might be too restrictive. In particular, the assumption \textbf{A1.1} on the invariance property of the target subspace $\mathcal{H}_S$ limits significantly the type of control actions that can be employed. Inspired by~\cite[Theorem 3]{grigoletto2021stabilization}, we relax such assumptions to the following:
\begin{description}
\item[\textbf{A2}:] There exists a $K_R\in\mathcal{B}_{>0}(\mathcal{H}_R)$ such that $\mathbf{L}_K(\rho)<0$ for all $\rho\in\mathcal{S}(\mathcal{H})\setminus\mathcal{I}(\mathcal{H}_S)$, where $K$ is the extension in $\mathcal{B}(\mathcal{H})$ of $K_R$.
\end{description}
Lemma~\ref{Lem:Continuity} shows that $\mathbf{L}_K(\rho)$ is continuous on $\mathcal{S}(\mathcal{H})$. In addition to the compactness of $\Theta_{\delta}:=\{\rho\in\mathcal{S}(\mathcal{H})|\,{\rm Tr}(K\rho)\geq \delta\}$ with $\delta>0$ and the assumption \textbf{A2}, there exists a constant $\gamma(\delta)>0$ such that $\mathbf{L}_K(\rho)\leq -\gamma(\delta)$ for all $\rho\in\Theta_{\delta}$. Then, we can deduce the average practical stability of the switched systems~\eqref{Eq:SSME} with a dwell time dependent on $\delta$ under \textbf{A2}. See~\cite[Theorem 3]{grigoletto2021stabilization} for the details. However, due to Lemma~\ref{Lem:Min=0}, $\gamma(\delta)$ converges to zero when $\delta$ tends to zero. Hence, we cannot fix a dwell time such that ${\rm Tr}(K\rho(t))$ decreases during each switching interval. However, by taking an appropriate gain for the active Lindbaldian during each switching interval, the speed of the derivative of ${\rm Tr}(K\rho(t))$ increasing to zero can be reduced, which can guarantee GAS of the target subspace $\mathcal{H}_S$ in mean and almost surely with an arbitrary non-zero dwell time and without any requirement on the invariance properties of Lindblad generators.
From now on, we suppose that the gain of Lindbladians are adjustable. The dynamics of the switched stochastic master equation is given by,
\begin{align}
d\rho(t)=\sum^m_{k=1}v^k_t\mathcal{L}_k(\rho(t))dt+\sum^m_{k=1}v^k_t\mathcal{G}_{C}(\rho(t))dW(t),
\label{Eq:SSME_modulate}
\end{align}
where $\rho_0\in \mathcal{S}(\mathcal{H})$, $v^k_t\in[0,V]$ with $V<\infty$ and $k\in\mathcal{M}$ represents the switching control laws, which is bounded random variable adapted to $\mathcal{F}_t$, and for almost every sample path and $t\geq 0$, $v^{i}_t v^{j}_t=0$ with $i\neq j$. The drift and diffusion term of the switched system~\eqref{Eq:SSME_modulate} obey a global random Lipschitz condition, the existence and uniqueness of a global strong solution of the switched system~\eqref{Eq:SSME_modulate} can be ensured by combining the arguments as in~\cite[Theorem 1.2]{krylov1999kolmogorov} and~\cite[Proposition 3.3]{mirrahimi2007stabilizing} or~\cite[Chapter 5]{barchielli2009quantum}.
Next, we introduce the switching strategy $\sigma_5$ with the fixed dwell time based on $\rho(t)$. Fix an arbitrary dwell time $\Delta t>0$ such that $t_{n+1}-t_n=\Delta t$ for any $n\in\mathbb{N}$. Set
\begin{align*}
&p_n(\omega)=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}(K\mathcal{L}_k(\rho(t_n))), \\
&v^{k}_t(\omega)=q_n(\omega)\mathds{1}_{\{p_n(\omega)=k\}}, \quad \forall\, k\in\mathcal{M},\, t\in[t_n,t_{n+1}),
\end{align*}
where $q_n$ is a bounded random variable adapted to $\mathcal{F}_{t_n}$. Set $q_n=0$ if ${\rm Tr}(K\rho(t_n))=0$. Due to the relation~\eqref{Eq:Relation_DisLya}, $\rho(t)$ stays in $\mathcal{I}(\mathcal{H}_S)$ afterwards in this case. For the case ${\rm Tr}(K\rho(t_n))>0$, $q_n$ will be determined later. Then, the solution of the switched system~\eqref{Eq:SSME_modulate} is well-defined on $[0,t_{n+1}]$. By It\^o formula, for $t\in[t_n,t_{n+1}]$, we have
\begin{equation*}
\mathbb{E}(\rho(t)|\mathcal{F}_{t_n})-\rho(t_n)=\mathbb{E}\Big( \int^t_{t_n} \sum^m_{k=1}v^k_s\mathcal{L}_k(\rho(s))ds \Big| \mathcal{F}_{t_n} \Big), \,a.s.
\end{equation*}
Set $\Lambda^k_n:=\{\omega\in\Omega|\, p_n(\omega)=k \text{ and }{\rm Tr}(K\rho(t_n))>0 \}$ with $k\in \mathcal{M}$, then $\mathbb{P}(\bigcup_{k\in\mathcal{M}}\Lambda^{k}_{n})=\mathbb{P}({\rm Tr}(K\rho(t_n))>0)$ and $\mathbb{P}(\Lambda^{i}_{n}\cap \Lambda^{j}_{n})=0$ for $i\neq j$.
For almost all $\omega\in\Lambda^k_n$, due to the linearity of $\mathcal{L}_k$, we have
\begin{equation*}
\mathbb{E}(\rho(t)|\mathcal{F}_{t_n})-\rho(t_n)=\int^t_{t_n} q_n(\omega)\mathcal{L}_k\big( \mathbb{E}(\rho(s)|\mathcal{F}_{t_n}) \big)ds,
\end{equation*}
where we used the fact that $q_n$ is adapted to $\mathcal{F}_{t_n}$. It implies that, for almost all $\omega\in\Lambda^k_n$, $\mathbb{E}(\rho(t)|\mathcal{F}_{t_n})=e^{(t-t_n)q_n(\omega)\mathcal{L}_k}\rho(t_n)$. Then, for almost every $\omega\in\Lambda^k_n$,
\begin{align*}
\mathbb{E}({\rm Tr}(K\mathcal{L}_k(\rho(t)))|\mathcal{F}_{t_n})-{\rm Tr}(K\mathcal{L}_k\rho(t_n))&=\mathbb{E}\Big( \int^{t}_{t_n} q_n(\omega){\rm Tr}(K\mathcal{L}^2_k(\rho(s)))ds \Big| \mathcal{F}_{t_n} \Big)\\
&\leq \bar{M}_K (t-t_n) q_n(\omega),
\end{align*}
where $\bar{M}_K:=\max_{k\in\mathcal{M}}\sup_{\rho\in\mathcal{S}(\mathcal{H})}{\rm Tr}(K\mathcal{L}^2_k(\rho))$. One deduces that, for all $t\in[t_n,t_{n+1}]$,
\begin{align*}
\mathbb{E}({\rm Tr}(K\mathcal{L}_k(\rho(t)))|\mathcal{F}_{t_n})&={\rm Tr}\big(K\mathcal{L}_k\big(e^{(t-t_n)q_n(\omega)\mathcal{L}_k}\rho(t_n)\big)\big)\\
&\leq \bar{M}_K (t-t_n) q_n(\omega)+{\rm Tr}(K\mathcal{L}_k\rho(t_n)),
\end{align*}
where ${\rm Tr}(K\mathcal{L}_k\rho(t_n))<0$ due to the assumption \textbf{A2}. Define
\begin{equation*}
q_n(\omega):=-\frac{{\rm Tr}(K\mathcal{L}_k\rho(t_n))}{\bar{M}_K \Delta t }>0, \quad \forall \, \omega\in\Lambda^k_n.
\end{equation*}
Hence, for almost all $\omega\in\Lambda^k_n$ and $t\in[t_n,t_{n+1})$,
\begin{equation*}
{\rm Tr}\big(K\mathcal{L}_k\big(e^{(t-t_n)q_n(\omega)\mathcal{L}_k}\rho(t_n)\big)\big)<0
\end{equation*}
Then, for almost each $\omega\in\Lambda^k_n$, we have
\begin{align*}
&\mathbb{E}({\rm Tr}(K\rho(t_{n+1}))|\mathcal{F}_{t_n})-{\rm Tr}(K\rho(t_n))=\mathbb{E}\Big( \int^{t_{n+1}}_{t_n} q_n(\omega){\rm Tr}(K\mathcal{L}_k(\rho(s)))ds \Big| \mathcal{F}_{t_n} \Big)\\
&=\int^{t_{n+1}}_{t_n} q_n(\omega){\rm Tr}\big(K\mathcal{L}_k\big(e^{(s-t_n)q_n(\omega)\mathcal{L}_k}\rho(t_n)\big)\big)ds<0.
\end{align*}
Due to the arbitrariness of $k$, the strict decrease of
$
\mathbb{E}({\rm Tr}(K\rho(t))|\mathcal{F}_{t_n})
$
on $[t_n,t_{n+1})$ when ${\rm Tr}(K\rho(t_n))>0$.
\begin{definition}[Switching strategy $\sigma_5$]
Fix the dwell time $\Delta t>0$ and set $t_0=0$,
\begin{align*}
&p_n(\omega)=\textstyle\operatorname*{arg\,min}_{k\in\mathcal{M}}{\rm Tr}(K\mathcal{L}_k(\rho(t_n))), \\
&q_n(\omega)=
\begin{cases}
-\frac{{\rm Tr}(K\mathcal{L}_k\rho(t_n))}{\bar{M}_K \Delta t },& \text{if } {\rm Tr}(K\rho(t_n))>0,\\
0,& \text{else},
\end{cases}\\
&t_{n+1}=t_n+\Delta t,\\
&v^{k}_t(\omega)=q_n(\omega)\mathds{1}_{\{p_n(\omega)=k\}}, \quad \forall\, k\in\mathcal{M},\, t\in[t_n,t_{n+1}).
\end{align*}
\end{definition}
\smallskip
\begin{theorem}
Suppose that \emph{\textbf{A2}} holds true. For any finite dwell time $\Delta t>0$, $\mathcal{H}_S$ is GAS in mean and almost surely for the switched system~\eqref{Eq:SSME_modulate} under the switching strategy $\sigma_5$.
\label{Thm:GAS}
\end{theorem}
\emph{Proof.}
If there exists $\bar{n}<\infty$ such that ${\rm Tr}(K\rho(t_{\bar{n}}))=0$ almost surely, it reduces to a trivial case. Suppose that, for all $\rho_0\in\mathcal{S}(\mathcal{H})\setminus \mathcal{I}(\mathcal{H}_S)$, $\mathbb{P}({\rm Tr}(K\rho(t_{n})>0,\, \forall n\in\mathbb{N})>0$.
Due to the law of total expectation, for any $n\in\mathbb{N}$,
\begin{align*}
&\mathbb{E}\big( {\rm Tr}(K\rho(t_{n+1})) \big)-\mathbb{E}\big( {\rm Tr}(K\rho(t_{n})) \big)=\mathbb{E}\big(\mathbb{E}({\rm Tr}(K\rho(t_{n+1}))|\mathcal{F}_{t_n})-{\rm Tr}(K\rho(t_n)) \big)\\
&=\mathbb{E}\big(\mathds{1}_{\{{\rm Tr}(K\rho(t_{n})>0\}}\big( \mathbb{E}({\rm Tr}(K\rho(t_{n+1}))|\mathcal{F}_{t_n})-{\rm Tr}(K\rho(t_n)) \big) \big)<0.
\end{align*}
Then, by employing the standard Lyapunov arguments~\cite[Theorem 10.1.1]{hale1980ordinary}, GAS in mean of $\mathcal{H}_S$ can be concluded. Moreover, due to the Chebyshev’s inequality, $\mathcal{H}_S$ is stable in probability~\cite[Theorem 5.3]{khasminskii2011stochastic}. By using the dominated convergence theorem, almost sure GAS of $\mathcal{H}_S$ can be concluded.
\hfill$\square$
\section{Numerical simulations} \label{Sec:examples}
In this section, we illustrate the performance of the proposed switching strategies in stabilizing a three-qubit system and a spin-1 system towards a pre-determined GHZ state and a pre-determined eigenstate, respectively. In order to ensure that the trajectories of switched stochastic master equation~\eqref{Eq:SSME} stay in $\mathcal{S}(\mathcal{H})$ during the simulation, we employ the numerical scheme proposed in~\cite{rouchon2015efficient}.
The results of simulations are shown as graphs of the trace norm distance of the actual state from the target state $\frac12\|\rho-\bar{\rho}\|_1$ against time.
\subsection{GHZ stabilization under under assumption \emph{\textbf{A1}}}
Here, we apply the switching strategies $\sigma_1$, $\sigma_2$, $\sigma_3$ and $\sigma_4$ on the three-qubit system proposed in~\cite{ticozzi2014steady,liang2022GHZ}. The target state is determined as
\begin{equation*}
\bar{\rho} = \frac{1}{2}(|000\rangle+|111\rangle)(\langle 000|+\langle 111|),
\end{equation*}
and the initial state is set as $\rho_0 = \frac{1}{4}|\Psi\rangle\langle\Psi|$ where
\begin{equation*}
|\Psi\rangle = |001\rangle+|010\rangle+|101\rangle+|110\rangle.
\end{equation*}
We construct two Lindbladian generators as follows
\begin{equation*}
\begin{split}
&H_1 = H_2 = \sigma_x\otimes \mathbf{I}\otimes \mathbf{I}-\mathbf{I}\otimes \sigma_x\otimes\sigma_x, \quad C = \sigma_z\otimes \mathbf{I}\otimes \sigma_z,\\
&L_1 = (|00\rangle\langle 01|+|11\rangle\langle10|)\otimes \mathbf{I}, \quad L_2 = \mathbf{I}\otimes (|00 \rangle \langle 01|+i|11\rangle\langle10|);\\
\end{split}
\end{equation*}
where $\sigma_x$ and $\sigma_z$ are Pauli matrices. Based on Theorem~\ref{Thm:InvarinceGAS}, it is easily to verify that $\bar{\rho}$ is invariant for two Lindbladian generators $\mathcal{L}_1$ and $\mathcal{L}_2$, not GAS under any single Lindbladian. However, the simultaneous action of $\mathcal{L}_1$ and $\mathcal{L}_2$, with arbitrary positive weights, leads to the GAS of $\bar{\rho}$. Then, the assumption \textbf{A1} is satisfied due to Corollary~\ref{Cor:GES_convex} and Theorem~\ref{Thm:InvarinceGAS}. Based on $K$ constructed by the effective numerical approach proposed in Appendix~\ref{App:Compute_K}, we can obtain the value of the constant $c=0.0232$ in the assumption \textbf{A1.2}.
The following simulations have been run with step length $\delta t = 0.002$, number of steps $N = 25000$, each with $200$ realizations.
Set $\epsilon=0.3$, we have the lower bound of the dwell time $t_D=0.016$ defined in the proof of Theorem~\ref{Thm:GES_average}. By taking the dwell time $\Delta t=5\delta t<t_D$, we observe that the behaviors of the trace norm distance along sample trajectories under the switching strategies $\sigma_2$ and $\sigma_3$ are bounded by the exponential reference $\epsilon c/2$ given in Corollary~\ref{Cor:GES_Dwell} and Theorem~\ref{Thm:GES_dwell}, respectively (see Fig.\ref{SIM:Fixed}).
For simulations of applying the hysteresis switching strategies, we set up to check the system state at each step time $\delta t$. We observe that the behaviors of the trace norm distance along sample trajectories under $\sigma_1$ and $\sigma_4$ are bounded by the exponential reference $\epsilon c/2$ given in Theorem~\ref{Thm:GES_average} and Theorem~\ref{Thm:GES_state}, respectively (see Fig.\ref{SIM:Hysteresis}). For both simulations, the measurement-based trajectories converge faster than the average-based one, which is consistent with our expectation.
\begin{figure}[!t]
\centerline{\includegraphics[width=\columnwidth]{fixed.pdf}}
\caption{Exponential stabilization of a three-qubit system toward $\bar{\rho}$ starting from $\rho_0$ with dwell time $t_D$. The blue curve represents the mean value of 200 measurement-based realizations and the light blue area shows the mean value plus or minus one standard deviation, the orange curve represents the mean value of 200 average-state realizations and the light orange area shows the mean value plus or minus one standard deviation.}
\label{SIM:Fixed}
\end{figure}
\begin{figure}[!t]
\centerline{\includegraphics[width=\columnwidth]{hysteresis.pdf}}
\caption{Exponential stabilization of a three-qubit system toward $\bar{\rho}$ starting from $\rho_0$ under hysteresis switching. The blue curve represents the mean value of 200 realizations under $\sigma_1$ and the light blue area shows the mean value plus or minus one standard deviation, the orange curve represents the mean value of 200 realizations under $\sigma_3$ and the light orange area shows the mean value plus or minus one standard deviation.}
\label{SIM:Hysteresis}
\end{figure}
In~\cite[Section V]{liang2022GHZ}, the authors discuss the Hamiltonian feedback control of multi-qubit systems toward the target GHZ state in presence of only $z$-type measurements, i.e., the measurement operators are in the form $\bigotimes^n_{i=1}\sigma^i$ with $\sigma^i\in\{\sigma_z,\mathbf{I}\}$. In this case, the GAS is not proved in~\cite{liang2022GHZ} since only Hamiltonian control may not drive the system through some invariant sets and approach the target state. In this paper, we combine the advantages of measurement-based feedback and dissipative control to easily overcome the above obstacles.
\subsection{Spin-1 system under assumption \emph{\textbf{A2}}}\label{sec:spin1}
Here, we apply the switching strategies $\sigma_5$ on the spin-1 system proposed in~\cite{liang2019exponential}. The target state is
$
\bar{\rho} = \mathrm{diag}(1,0,0),
$
and the initial state is
$
\rho_0 = \mathrm{diag}(0,0,1).
$
We construct three Lindbladian generators as follows
\begin{equation*}
\begin{split}
&\mathcal{L}_1(\rho) = -i[H,\rho]+\mathcal{D}_C(\rho),\,\mathcal{L}_2(\rho) = -i[-H,\rho]+\mathcal{D}_C(\rho),\,\mathcal{L}_3(\rho) = \mathcal{D}_L(\rho)+\mathcal{D}_C(\rho),
\end{split}
\end{equation*}
where $C=\mathrm{diag}(1,0,-1)$ and
\begin{equation*}
H = \left[
\begin{matrix}
0 & i & 0\\
-i & 0 & 0\\
0 & 0 & 0
\end{matrix}
\right], \quad
L = \left[
\begin{matrix}
1 & 1 & 0\\
0 & 0 & 1\\
0 & 0 & 0
\end{matrix}
\right].
\end{equation*}
It is easily to verify that $\bar{\rho}$ is not invariant for $\mathcal{L}_1$, $\mathcal{L}_2$ and $\mathcal{L}_3$ by Theorem~\ref{Thm:InvarinceGAS}. Thus, the switching strategies $\sigma_1$, $\sigma_2$, $\sigma_3$ and $\sigma_4$ cannot be applied. Consider $K=\mathrm{diag}(0,k_1,k_2)$ with $k_1,k_2>0$, by straightforward calculation, the assumption \textbf{A2} is satisfied if and only if $k_2>k_1$. In the switching strategy $\sigma_5$, the constant $\bar{M}_K$ can be estimated by
$
\bar{M}_K\leq \max_{i\in\{1,2,3\}} \|\mathrm{Tr}({\mathcal{L}^*_i}^2(K))\|,
$
where $\mathrm{Tr}(\rho^2)\leq 1$ and Cauchy-Schwarz inequality are used. The following simulations have been run with step length $\delta t = 0.1$, number of steps $N = 5000$, each with $1000$ realizations.
Set $k_1=1$, $k_2=2$ and the dwell time as $10\delta t$, the convergence of the spin-1 system toward $\bar{\rho}$ under $\sigma_5$, starting at $\rho_0$, is shown in Fig.\ref{SIM:Modulate}.
\begin{figure}[!t]
\centerline{\includegraphics[width=\columnwidth]{assumption_2_modulating.pdf}}
\caption{Asymptotic stabilization of a spin-1 system toward $\bar{\rho}$ starting from $\rho_0$ by modulating Lindbladians. The blue curve represents the mean value of 1000 realizations under $\sigma_5$ and the light blue area shows the mean value plus or minus one standard deviation.}
\label{SIM:Modulate}
\end{figure}
\section{Conclusion}
This work presents a thorough analysis of switching techniques for fast stabilization of pure states and subspaces for quantum filtering equations. We demonstrate how globally exponentially stabilizing control laws can be derived both based on the expected state dynamics, thus making them viable for an off-line computation and an open-loop implementation, and the current state estimate conditioned on the measurement record, which leads to more accurate switching and better performance in most cases \cite{grigoletto2021stabilization}. In proving GES, we propose ways to compute upper bounds for the convergence exponent. From the simulation results it is clear that these are not optimal, and the main reason is likely the fact that the stochastic back-action component of the dynamics is not directly taken into account in computing such bounds.
We also provide a novel way to derive feedback laws that ensure global asymptotic stability even in the case where the controlled dynamics do not necessarily maintain the target invariant, allowing us to address stabilization problems that were so far only considered with Hamiltonian control actions, see e.g. Section \ref{sec:spin1}.
This is possible by suitably modulating the amplitude of the switching controlled dynamics, avoiding chattering and stability problems that so far allowed to prove only weaker results. Because of the modulation, we can thus prove GAS but in general we lose exponential convergence. These results thus open new possibilities in the integration of coherent and dissipative resources for the control of quantum systems, and extend the applicability of feedback methods for effectively controlling quantum systems.
Further developments of the line of work shall focus on the derivation of better bounds for the exponential convergence speed and its optimization, the robustness of the filters and the proposed control law to initialization errors, simplified switching laws that do not need a full state reconstruction, and potential experimental applications.
|
\section*{Introduction}
Due to their vast and exciting phenomenology, as well as the rapid advances on their control protocols, many-body quantum systems have become key ingredients for the development of novel technologies at the nanoscale. In particular, in nonequilibrium regimes such systems feature properties which make them strongly appealing for applications in quantum circuits \cite{benenti2017physrep,bertini2021rmp,pekola2021rmp,landi2021arxiv}. This has been exemplified in several platforms where tunable transport of particles or heat can be induced and characterized, including setups based on electronic devices such as molecular electronics \cite{xin2019nat} and quantum dots \cite{josefsson2018nat,dutta2020prl}, or on quantum simulators such as cold atoms \cite{hausler2021prx,amico2022arxiv} and superconducting qubits \cite{ronzani2018nat,maillet2020nat}.
In parallel to these seminal experimental advances, there have been numerous recent theoretical proposals to use quantum systems of interacting particles as different types of circuit elements that efficiently perform specific tasks. This includes autonomous quantum thermal machines \cite{nosotros2020prx,khandelwal2021prx,BohrBrask2022operational}, transistors \cite{marchukov2016nat,wilsmann2018nat}, magnetoresistors \cite{poulsen2021prl}, and a quantum analogue of the Wheatstone bridge \cite{poulsen2022prl}. Much attention in this area has been directed towards quantum diodes \cite{pepino2009prl,balachandran2018prl,balachandran2019prl,pereira2019pre,balachandran2019pre,yamamoto2020prr,lee2020entropy,lee2021pre,lee2022pre,poulsen2022pra,chioquetta2021pre,upadhyay2021pre}. There, spatial asymmetries and non-linearities are engineered together to allow transport in one direction under a chemical potential, magnetic or thermal bias, and to suppress it in the opposite direction when the bias is inverted. These efforts include the proposal of giant rectification setups \cite{balachandran2018prl,lee2020entropy,lee2021pre,lee2022pre,poulsen2022pra}, which rely on complicated geometries and potential landscapes to induce rectification coefficients of several orders of magnitude using finely-tuned parameters.
Here we put forward a rectification scheme which naturally emerges over a broad range of parameters in tilted interacting quantum lattices. These systems have been the object of intense recent research as they feature disorder-free (Stark) many-body localization (MBL) \cite{schulz2019prl,nieuwenburg2910pnas,taylor2020prb,yao2021prb,zisling2022prb}, quantum scars \cite{khemani2020prb,desaules2021prl}, and counter-intuitive phenomena such as time crystals \cite{kshetrimayum2020prb} and transport opposite to an applied electric field \cite{klockner2020prl}. Moreover, they have been implemented experimentally in several quantum simulation platforms \cite{morong2021nat,guardado2020prx,scherg2021nat,hebbe2021prx,guo2021prl}, making them attainable for state-of-the-art applications in quantum technologies. Neither the spatial asymmetry of the tilted onsite-potential nor the non-linearity of inter-particle interactions can on their own induce rectification. However, the ability to simultaneously implement both ingredients in such platforms opens the possibility of rectification. Here we unveil the underlying mechanism of this interaction-tilt interplay and show how the vastly differing transport properties \cite{doggen2021prb} with the direction of bias give rise to giant rectification even with moderate interactions.
\section*{Results}
\begin{figure*}[t]
\begin{center}
\includegraphics[scale=0.45]{Fig1.eps}
\end{center}
\caption{Giant rectification in tilted strongly-interacting lattice. \textbf{a} For reverse transport, particles are mostly injected from a reservoir on the right boundary of the system, and mostly ejected to a reservoir on the left. For interactions or tilt alone, the resulting state is of high-energy, due to the adjacency of particles in the former case, or due to the particles filling the top of the potential in the latter. An enhanced insulator is induced when both coexist. \textbf{b} For forward transport, particles flow from the left to the right reservoir. The tilted noninteracting state is of low energy as particles now fill the bottom of the potential. Thus tilt now competes with interactions, which favors conduction. \textbf{c} Sketch of the eigenstructure of a tilted interacting system. (i) High-energy insulating eigenstate preferentially populated by reverse bias.
(ii) High-energy insulating eigenstate preferentially populated by forward bias and dominant interactions. (iii) Low-energy insulating eigenstate preferentially populated by forward bias and very large tilt. (iv) The path from the interaction- to the tilt-dominated insulator, indicated by the dashed line, features avoided crossings at intermediate regimes where occupation probability is transferred between eigenstates, enhancing the transport. \textbf{d} Forward (F) and reverse (R) particle currents in a lattice of $N=4$ and $\Delta=5$ as a function of the tilt. Currents have been rescaled by the maximum current $\mathcal{J}_0$ obtained with the same driving for a noninteracting homogeneous system (see Supplementary Methods). \textbf{e} The corresponding rectification coefficient $\mathcal{R}$.}
\label{setup}
\end{figure*}
\hspace{-0.35cm}\textbf{Giant rectification scheme.} Our quantum diode is based on a boundary-driven configuration, which can be biased from right to left ({\em reverse driving}) or left to right ({\em forward driving}). For reverse driving, sketched in Fig. \ref{setup}a, both interactions and tilt each support on their own the formation of a high-energy current-blocking particle domain at the right boundary.
When interactions and tilt coexist they thus cooperate to induce a high-energy state with an enhanced insulating nature. For forward driving, depicted in Fig. \ref{setup}b, both strong interactions and tilt favour on their own the formation of an insulating domain at the left boundary. For interactions alone this state remains high-energy, while for the tilt alone it is low-energy (see Fig. \ref{setup}c).
By being located at opposite ends of the energy spectrum interactions and tilt now conflict with each other when they coexist. When both are of similar order, this competition breaks the domains and thus allows particle conduction. This forward-reverse transport asymmetry makes the system behave as a quantum diode.
To unveil this phenomenon we consider a simple generic model of interacting particles: a one-dimensional lattice of spinless fermions in a tilted potential. The Hamiltonian for $N$ sites is
\begin{align} \label{hami_tilted}
H&=\sum_{j=1}^{N-1}\left[\frac{J}{2}\left(c_j^{\dagger}c_{j+1}+\text{H.c.}\right)+\Delta\left(n_j-\frac{1}{2}\right)\left(n_{j+1}-\frac{1}{2}\right)\right]\notag \\
&\qquad \qquad \qquad \qquad +\sum_{j=1}^N\left(\mu+\frac{E}{2}j\right)\left(n_j-\frac{1}{2}\right),
\end{align}
where $c_j^{\dagger}$ ($c_j$) creates (annihilates) a fermion on site $j$, $n_j=c_j^{\dagger}c_j$ is the corresponding particle number operator, $J$ is the hopping amplitude (we take $J=1$ to set the energy scale), $\Delta$ is the nearest-neighbor density-density interaction, and $\mu$ is the diode's chemical potential. We take the tilt strength $E>0$ so the potential always increases from left to right. For the noninteracting case $\Delta=0$, this model shows Wannier-Stark localization \cite{wannier1962rmp}. Furthermore, it was recently discovered that for strong enough tilt the interacting model still features nonergodic behavior in the absence of disorder, an effect known as Stark MBL \cite{nieuwenburg2910pnas,schulz2019prl}.
\begin{figure*}[t]
\begin{center}
\includegraphics[scale=0.55]{Fig2.eps}
\end{center}
\caption{Formation of particle domains. \textbf{a} Site population $\langle n_j\rangle$ for $N=4$, $\Delta=5$ and the resonant tilt values, for forward and reverse driving. \textbf{b} Site population $\langle n_j\rangle$ for a lattice of $N=8$, $E=0$ and different interactions $\Delta$ for reverse driving. \textbf{c} Site population $\langle n_x\rangle$ for $\Delta=0$, different system sizes and rescaled tilt $V$, with rescaled position $x = j/N \in[0,1]$.}
\label{fig_domains}
\end{figure*}
To drive the system to a nonequilibrium state, we incorporate reservoirs with differing chemical potentials at its boundaries. We assume a weak coupling to the system (Born approximation), high-temperature memory-less reservoirs (Markov approximation), and that the bandwidths of the reservoirs are much larger than those of the system, which leads to frequency-independent system-reservoir interactions (wide-band limit) \cite{mark2018njp}. Tracing out the reservoir degrees of freedom leads to a Lindblad master equation for the reduced density matrix $\rho$ of the system \cite{BreuerPetruccione},
\begin{align}\label{lindblad}
\frac{d\rho}{dt}=-i[H,\rho]+\sum_k\mathcal{L}_k(\rho),
\end{align}
with the dissipative superoperators $\mathcal{L}_k(\rho)$ defined by jump operators $L_k$ as
\begin{align}
\mathcal{L}_k(\rho)=L_k\rho L_k^{\dagger}-\frac{1}{2}\{L_k^{\dagger}L_k,\rho\}.
\end{align}
The driving is induced by boundary operators $L_{1,N}^+=\sqrt{\Gamma(1\pm f)/8}\ c_{1,N}^{\dagger}$ and $L_{1,N}^-=\sqrt{\Gamma(1\mp f)/8}\ c_{1,N}$, where $\Gamma$ is the coupling strength (taken as $\Gamma=1$) and the driving parameter $-1\leq f\leq1$ establishes the bias \cite{Benenti:2009,nosotros2013prb,landi2021arxiv}. Namely, when $f>0$, fermions are mostly created on site 1 and mostly annihilated on site $N$ giving forward driving, while inverting the sign of $f$ gives reverse driving. The particle current corresponds to the expectation value of the operator
\begin{align}
\mathcal{J}_j=i\frac{J}{2}\left(c_j^{\dagger}c_{j+1}-c_{j+1}^{\dagger}c_{j}\right),\qquad j=1,\ldots,N-1,
\end{align}
which directly arises from the particle number continuity equation \cite{landi2021arxiv}. We focus on the transport properties of the nonequilibrium steady state (NESS), where the current across the system is homogeneous, i.e. $\langle\mathcal{J}_j\rangle=\mathcal{J}$. In addition, we set $|f|=1$ to consider maximal forward/reverse driving, and take $\mu=-E(N+1)/4$ so the system features charge conjugation and parity (CP) symmetry. This choice will help make the rectification mechanism transparent. Nonetheless, under the assumed wide-band limit for the driving, the transport is independent of $\mu$ \cite{znidaric2011pre} (see Methods).
This simple nonequilibrium scenario allows us to observe large rectification even for very small lattices. For instance, in Fig. \ref{setup}d we illustrate the widely differing response of a system with $N=4$ and $\Delta=5$ when inverting the direction of the bias. For reverse driving of fermions down the tilted potential, the current decays rapidly and monotonically with $E$. On the other hand, for forward driving of fermions up the potential, the current features much larger values at well-defined resonances with $E$. As shown in Fig. \ref{setup}e, this leads to rectification coefficients of up to $O(10^3)$, a value that can be vastly enhanced in longer chains (see Fig. \ref{fig_recti}).
These properties are intimately linked to the presence or absence of a particular distribution of fermions across the system, as depicted in Fig. \ref{fig_domains}a. For reverse driving, the particles are pinned to the right boundary forming a domain half the size of the lattice, leaving the left half essentially unoccupied. The domain Pauli blocks more fermions from entering the system, resulting in an insulating state. For forward bias near the resonant values of $E$, the population profiles do not feature this constraint, so particles are allowed to move across the lattice.\\
\hspace{-0.35cm}\textbf{Reverse driving insulating NESS.} Domain formation is a common feature induced by strong interactions and tilt on their own, as depicted in Figs. \ref{fig_domains}b where $E=0$ and \ref{fig_domains}c where $\Delta = 0$. In both these scenarios there is no rectification; transport for forward and reverse drivings are identical up to a direction inversion, so we focus our attention to the latter.
For strongly-interacting homogeneous ($E=0$) systems the insulating domain-like NESS is a well-known phenomenon \cite{Benenti:2009,nosotros2013prb,nosotros2013jstat,landi2021arxiv}, reproduced in Fig. \ref{fig_domains}b with exact diagonalization calculations for small lattices (see Methods). The system features a flat profile in the noninteracting limit, typical of ballistic transport. As $\Delta$ increases so does the slope of the profile, until reaching extreme population values at the boundaries for $\Delta\gtrsim1$. Increasing $\Delta$ even further results in the creation of the particle domain.
\begin{figure*}[t]
\begin{center}
\includegraphics[scale=0.55]{Fig3.eps}
\end{center}
\caption{Transport resonance mechanism in interacting tilted lattice of $N=4$ and $\Delta=5$, illustrated in the ``even" CP symmetry sector of $n=2$ filling (sector $s=2\text{e}$). In all panels, the dotted vertical lines indicate the location of the forward current resonances. \textbf{a} Eigenenergies $E_{\nu}$ as a function of the tilt $E$. For low tilts, the eigenstate $\nu=4$ corresponds to the corrected left domain state. The inset amplifies the first avoided crossing. The dashed lines correspond to second order perturbative calculations of the eigenenergies (see Supplementary Methods). \textbf{b} Probabilities $\rho^{2\text{e}}_{\nu}$ of eigenstates $\nu$ on the NESS. The line of $\nu=5,\text{R}$ corresponds to the highest energy eigenstate and reverse transport; all the others are for forward driving. Also note that the curve of $\nu=1$ is peaked at the first two resonances, even though it is not involved in an avoided crossing. This is because it corresponds to a perturbative correction of $|1010\rangle$ (see Supplementary Methods), which is strongly coupled to the $f=1$ driving. \textbf{c} Coherences $\rho^{2\text{e}}_{\mu,\nu}$ of pairs of eigenstates $\mu,\nu$ for forward transport. At the avoided crossing of eigenstates $\mu$ and $\mu+1$, $\rho^{2\text{e}}_{\mu,\mu+1}$ is maximal, which breaks the insulating nature of the NESS. Also, when getting away from avoided crossings, the NESS is again mostly described by a single eigenstate suppressing transport.}
\label{fig_reso_4sites}
\end{figure*}
In contrast noninteracting ($\Delta=0$) boundary-driven tilted systems have received much less attention \cite{silva2022arxiv} and their manifestation of localization has remained unexplored. We perform such an analysis obtaining exactly the site populations $\langle n_j\rangle$ and particle current $\mathcal{J}$ for large systems by reducing the Lindblad master equation \eqref{lindblad} to a set of linear equations (see Methods). We use a rescaled tilt $V=EN$, to keep the on-site energy difference between boundaries approximately constant. This way the density profiles of different system sizes collapse for each $V$, evidencing the emergence of domains for small lattices and large tilt, as well as for long lattices and low tilt. In addition, for large systems the current decays exponential with $N$ and $V$ (See Supplementary Discussion).
While the physical origin of the interaction- and tilt-induced insulating states is different, they share the same underlying mechanism that remains applicable for tilted interacting lattices. Specifically, they arise from the interplay between the gapped eigenstructure of the system and the maximal driving. For small hopping $J\ll \Delta\text{ or }E$, the highest energy eigenstate $|\Psi^D_{n}\rangle$ of each $n$ particle sector (e.g. eigenstate $\nu=5$ of Fig. \ref{fig_reso_4sites}a) corresponds to a perturbative correction to the configuration state $|00\ldots0_{N-n}11\ldots1_n\rangle$ in which all $n$ particles form a domain pinned to the right. The energy gap between this eigenstate and the rest, which increases with $\Delta$ and $E$, ensures that the amplitudes of configuration states in $|\Psi^D_{n}\rangle$ that couple to the $f=-1$ driving (where a particle can be injected at site $N$ or ejected at site $1$) are exponentially suppressed (see Supplementary Methods). Thus the eigenstates $|\Psi^D_{n}\rangle$ become exponentially close to obeying $\sum_k\mathcal{L}_k(|\Psi^D_{n}\rangle\langle\Psi^D_{n}|)= 0$, making them {\em dark states} of the driving. The overall NESS density matrix $\rho$ of the system is very well captured by an analytical Ansatz built from a statistical mixture of these darks states (see Eq. \eqref{ness_ansatz} of Methods and Supplementary Discussion). From this, we show that $\rho$ is strongly dominated by the contribution $|\Psi^D_n\rangle\langle\Psi^D_n|$ with $n=N/2$, corresponding to a state with occupied and empty domains of equal size (e.g. near unit probability for state $\nu=5$ for reverse driving in Fig. \ref{fig_reso_4sites}b). These results directly lead to the insulating NESSs of equal empty and occupied domains for systems with strong interactions \cite{nosotros2013prb}, tilt or both, whose currents monotonically and exponentially decrease as they increase.\\
\hspace{-0.35cm}\textbf{Forward driving resonances.} Next we focus on the forward driving setup with $f=1$. Here, strong interactions and tilt on their own would favor a NESS largely dominated by perturbative corrections to the left domain $|11\ldots1_{\frac{N}{2}}00\ldots0_{\frac{N}{2}}\rangle$. However, this configuration is the highest energy state of the $n=N/2$ particle number sector (degenerate with the right domain) for an interaction-dominated scenario, while it is the lowest energy state for a strongly tilted lattice. Thus, increasing $\Delta$ raises the energy of this state, while increasing $E$ lowers it.
For instance, fixing $\Delta$ and sweeping through $E$ (as in Fig. Fig. \ref{fig_reso_4sites}a) closes the energy gap from the initial NESS dark state, eventually forcing it to cross through the bulk spectrum of the system (see sketch in Fig. \ref{setup}c). As a result, current resonances are induced. This effect is cleanly illustrated by the half-filled even CP symmetric sector with the same parameters of Fig. \ref{setup}d as shown in Fig. \ref{fig_reso_4sites}a. Crucially, since all the eigenstates have the same symmetry, the no-crossing theorem \cite{Landau} ensures that only avoided crossings occur. Their location accurately pinpoints the current resonances.
\begin{figure*}[t]
\begin{center}
\includegraphics[scale=0.52]{Fig4.eps}
\end{center}
\caption{Transport and rectification in larger interacting tilted systems. \textbf{a} Reverse particle current $\mathcal{J}_\text{R}$ as a function of tilt for $N=8$ and different interaction strengths. The solid lines correspond to exact diagonalization results, and the darker dashed lines to currents obtained from our analytical Ansatz (see Methods). \textbf{b} Exact forward particle current $\mathcal{J}_\text{F}$ for the same parameters. \textbf{c} Corresponding rectification coefficient $\mathcal{R}$. \textbf{d} Reverse particle current $\mathcal{J}_\text{R}$ as a function of tilt for $\Delta=2$ and different system sizes. \textbf{e} Forward particle current $\mathcal{J}_\text{F}$ for the same parameters. \textbf{f} Corresponding rectification coefficient $\mathcal{R}$. The legends of system sizes spread over \textbf{d} and \textbf{e} apply to the three bottom panels. Also, the currents have been rescaled by the maximal current $\mathcal{J}_0$ obtained with the same driving for a noninteracting homogeneous system (see Supplementary Methods).}
\label{fig_recti}
\end{figure*}
To understand the role of the avoided crossings we evaluate the contributions to the particle current in the energy eigenbasis. Since there are no coherences between different symmetry sectors, the current is given by $\mathcal{J}=\sum_{s}\sum_{\mu_s,\nu_s}\rho^s_{\mu_s,\nu_s}\mathcal{J}^{s}_{\nu_s,\mu_s}(q)$, where $s$ sums over all the sectors, $\rho^s_{\mu_s,\nu_s}=\langle\mu_s|\rho|\nu_s\rangle$ is the coherence between eigenstates $|\mu_s\rangle$ and $|\nu_s\rangle$ of sector $s$, and ${J}^{s}_{\nu_s,\mu_s}(q)=\langle\nu_s|\mathcal{J}_q|\mu_s\rangle$ is the corresponding current matrix element evaluated on site $q$. From the probabilities $\rho^s_{\nu_s}=\rho^s_{\nu_s,\nu_s}$ and coherences $\rho^s_{\mu_s,\nu_s}$, shown in Figs. \ref{fig_reso_4sites}b and \ref{fig_reso_4sites}c respectively, the following picture emerges. For very low tilts, the NESS is solely dominated by the high-energy left domain-like state, evidenced by its large probability. This eigenstate has an energy $\sim\frac{1}{4}\left[\Delta(N-3)–En(N-n)\right]+\mathcal{O}(J^2)$ to lowest perturbative order (see Supplementary Methods), so as $E$ increases it dives down towards the lower eigenenergies which have a weaker dependence on the tilt. As it approaches another eigenstate they mix and probability is transferred from the higher to the lower energy states. Thus, in the proximity of the avoided crossing the Hamiltonian no longer supports an eigenstate that is dark to the $f=1$ driving. This allows large coherences to develop, enhancing the current. Further avoided crossings occur for even larger $E$, with lower eigenstates also getting populated. Eventually for very large $E$ the whole spectrum is crossed and the lowest-energy eigenstate $\nu=1$ inherits the dominant role on the NESS. This leads again to an insulating left domain-like dark state, but now induced by the strong tilt.
The combination of this resonant behavior with the enhanced insulating state of reverse driving results in sizable rectification coefficients $\mathcal{R}=-\mathcal{J}_\text{F}/\mathcal{J}_\text{R}$ even for the small chain of the example (see Fig. \ref{setup}e). Crucially, rectification becomes more substantial and more robustly achievable for larger system sizes $N$. The enhancement in the maximum $\mathcal{R}$ with $N$ follows from the differing current scaling of reverse and forward drivings. Namely $\mathcal{J}_{\rm R} \sim \exp(-N)$, characteristic of an insulator, while at a resonance we expect $\mathcal{J}_{\rm F} \sim N^{-\alpha}$ with some exponent $\alpha>0$ characteristic of a conductor. As shown in Fig. \ref{fig_recti}c, already for $N=8$ with $\Delta=5$ we find coefficients of up to $O(10^7)$ compared to $O(10^3)$ for $N=4$. In addition, $\mathcal{R}$ is further enhanced by particle interactions. As depicted in Figs. \ref{fig_recti}a-c, while increasing $\Delta$ diminishes both reverse and forward currents, the former decrease is more prominent and thus their ratio features even higher resonances.
While in small systems achieving the maximum $\mathcal{R}$ requires fine-tuning of $E$ to locate a resonance, for larger $N$ this constraint is rapidly removed due to the exponentially increasing number of eigenstates crossed within the bulk of the spectrum. As a result the rectification is broadened over the many-body bandwidth of the chain. Moreover, the favourable scaling with $N$ means that giant rectification can be achieved with a reduced interaction strength. To evidence these effects, using the insulating Ansatz and tensor network simulations (see Methods) we calculated reverse and forward currents for $\Delta=2$ and systems of $N>8$, shown in Figs. \ref{fig_recti}d and \ref{fig_recti}e respectively. From these we find $\mathcal{R}\sim\mathcal{O}(10^{9})$ over a range of $E$, as seen in Fig. \ref{fig_recti}f. This establishes larger rectification coefficients than those previously obtained for different types of non-homogeneous lattices \cite{balachandran2018prl,lee2020entropy,lee2021pre,poulsen2022pra,lee2022pre}, with a more intuitive and easily implementable setup \cite{guardado2020prx,scherg2021nat,hebbe2021prx,guo2021prl}.
\section*{Discussion}
We have proposed a giant rectification scheme in correlated quantum systems, based on the natural asymmetry between forward and reverse transport in tilted interacting lattices. In our protocol, particles get locked into an insulating domain NESS when attempting to travel down the tilted potential (reverse bias). On the other hand, they propagate uphill resonantly as avoided crossings in the spectrum are approached (forward bias). Our calculations show rectification coefficients of up to $O(10^{9})$ for small chains, a number that can be pushed even further without fine tuning and at moderate interaction strengths by considering larger systems.
We emphasize that most of the reported rectification peaks are associated to sizeable values of the forward particle currents. This is evidenced in Figs. \ref{setup} and \ref{fig_recti} by rescaling $\mathcal{J}$ with the maximal current $\mathcal{J}_0$ achievable under the same driving scheme for a noninteracting lattice with no tilt, which is a well-known size-independent ballistic current (see Supplementary Methods). This way it is clear that at the resonances the system is truly conducting in the forward direction while it is insulating in reverse bias. Thus the observed giant rectification is not an artifact of dividing two very small currents associated to insulating states.
Our results pave the way for developing different types of giant rectification protocols (e.g. of spin, charge or heat) in a variety of driving schemes, including finite temperature reservoirs, and magnetic or thermal biases. For this purpose, recent approaches for accurately simulating more realistic reservoirs, even when they are strongly coupled to the system, might be exploited \cite{rams2020prl,nosotros2020prx,fux2021prl,riera2021prx}. Furthermore, our proposal could be realized with existing quantum simulators. On the one hand, several experiments have implemented trapped ion \cite{morong2021nat}, cold-atom \cite{guardado2020prx,scherg2021nat,hebbe2021prx} and superconducting qubit \cite{guo2021prl} setups with similar tilt and/or interactions to those used in the present work, and even in larger systems. In addition, exploiting boundary-driven architectures in cold-atom platforms \cite{pepino2009prl,krinner2017jpcm,amico2022arxiv} or reservoir engineering for irreversible particle injection in superconducting circuits \cite{ma2019nat}, similar versions of our diode could be incorporated in state-of-the-art nanoscale quantum circuits.
\section*{Methods}
\hspace{-0.35cm}\textbf{Exact solution of noninteracting model.} The relevant observables for our transport analysis can be obtained for large systems with $\Delta=0$. For this we propose the Ansatz for the density matrix \cite{znidaric2010jstat,znidaric2011pre}
\begin{align} \label{ansatz}
\begin{split}
\rho&=\frac{1}{2^N}I+\sum_{k=1}^N2a_k\left(n_k-\frac{1}{2}\right)\\
&+\sum_{l=2}^N\sum_{k=1}^{N-l+1}b_k^{(l)}B_k^{(l)}+\sum_{l=2}^N\sum_{k=1}^{N-l+1}h_k^{(l)}H_k^{(l)}+\cdots
\end{split}
\end{align}
where we define the $l-$site operators
\begin{align}
\begin{split}
&B_k^{(l)}=-2i\left(c_j^{\dagger}c_{j+l-1}-c_{j+l-1}^{\dagger}c_{j}\right)\\
&H_k^{(l)}=2\left(c_j^{\dagger}c_{j+l-1}+c_{j+l-1}^{\dagger}c_{j}\right),
\end{split}
\end{align}
which correspond to generalized current and energy density operators respectively; in particular, $B_k^{(2)}=-4\mathcal{J}_k$. Even though this is a first-order expansion in $f$ \cite{znidaric2010jstat,znidaric2011pre}, it leads to an exact solution of the parameters $a_k=\langle2n_k-1\rangle$, $b_k^{(l)}=\langle B_k^{(l)}\rangle/2$ and $h_k^{(l)}=\langle H_k^{(l)}\rangle/2$. For this, the Ansatz \eqref{ansatz} in inserted into the master equation \eqref{lindblad} with the stationary condition $d\rho/dt=0$, which results in a closed set of linear equations for the parameters when grouping the coefficients in front of each operator (see Supplementary Methods). Notably, through this linear set it is straightforward to link the boundary population and the current, since from the coefficients in front of $2n_1-1$ we get
\begin{align} \label{coeff_sigma_eq}
a_1+\frac{4b^{(2)}}{\Gamma}=f,\qquad\text{with }b^{(2)}=b_k^{(2)}\text{ homogeneous}.
\end{align}
Exploiting the symmetric boundary driving, it is found that several coefficients are equal to their counterpart from the other half of the lattice (e.g., $a_k=a_{N-k+1}$, so the population profile is symmetric around 0.5) or zero (e.g., $a_{\frac{N+1}{2}}=0$ for $N$ odd). This leads to a system of $\frac{N^2}{2}+1$ unknowns for $N$ even and $\frac{(N-1)^2}{2}+1$ unknowns for $N$ odd, which can be solved with standard linear algebra packages for lattices of several hundreds of sites. Analytical solutions for small systems are provided in the Supplementary Discussion.\\
\hspace{-0.35cm}\textbf{Calculation of transport properties.} The transport properties of small lattices ($N\leq8$) was obtained exactly by diagonalizing the full Lindblad superoperator, written as a $4^{N}\times 4^{N}$ matrix. The NESS is the eigenstate with zero eigenvalue, reshaped as the $2^N\times 2^N$ density matrix $\rho$. For larger systems and forward transport, we describe $\rho$ as a matrix product state \cite{prosen2009jstat}, and obtain the particle current using the time evolving block decimation method \cite{VidalTEBD2004,VerstraeteTEBD2004} implemented with the Tensor Network Theory (TNT) library \cite{tnt,tnt_review1}. Here, the evolution of an arbitrary state should be simulated until the current becomes homogeneous, which indicates convergence. However, strong tilts and interactions commonly result in a very slow approach to the NESS, making it impractical. Thus, in such cases we focused on converging expectation values on the first site only. This shortcut allowed us to reproduce very well the current of small lattices, so we extended its application to longer chains.\\
For $N>8$ and reverse transport, where even the previous shortcut is not enough due to exponentially slow convergence to the NESS \cite{Benenti:2009,nosotros2013prb}, we calculate the current from the analytical Ansatz for insulating states. This is given by (see Supplementary Methods)
\begin{align} \label{ness_ansatz}
\rho_{\text{Dom}}=\sum_{n=0}^Np_n |\Psi^D_n\rangle\langle\Psi^D_n|,\qquad\sum_{n=0}^Np_n=1,
\end{align}
which captures remarkably well the particle density profile. The current is calculated from the population of the first site, as Eq. \eqref{coeff_sigma_eq} is generally valid \cite{nosotros2013jstat,landi2021arxiv}. Since the considered system sizes can have very low currents $\mathcal{J}<10^{-15}$, exact calculations from this Ansatz were performed in MATLAB using the Variable-precision arithmetic (vpa) function to take 32 digits of precision.\\
\hspace{-0.35cm}\textbf{CP symmetry of tilted interacting model.} Hamiltonian \eqref{hami_tilted} is invariant under simultaneous charge conjugation and center-reflection operations, e.g. $Un_jU^{\dagger}=1-n_{N-j+1}$ for the CP symmetry operator $U$. Also, since $U^2=I$, its eigenvalues are $\pm1$, corresponding to even and odd symmetry. However, only at half filling $U$ commutes with the total particle number operator $\tilde{N}=\sum_jn_j$, since $U\tilde{N}U^{\dagger}=N-\tilde{N}$. Thus, only in this case we can consider particle number conservation and CP symmetry simultaneously. This property is used to build the Hamiltonian of the even CP half-filled symmetry sector (see Supplementary Methods) whose energy eigenspectrum features the avoided crossings that coincide with the forward current resonances (see Fig. \ref{fig_reso_4sites}a).\\
Crucially, this does not mean that our giant rectification mechanism relies on the presence of CP symmetry. In fact, the particle transport remains identical when this symmetry is broken by taking $\mu\neq-E(N+1)/4$. This occurs because within the assumed wide-band limit, the energy gap between different particle number sectors (controlled by $\mu$) is irrelevant. Only the energy differences between eigenstates within each sector, which are independent of $\mu$, are important for our device. The CP symmetry just allows us to isolate even and odd eigenstates, transparently evidencing the connection between the avoided crossings and current resonances.\\
In light of this discussion, also note that forward particle transport is identical (up to a direction change) to reverse transport with inverted tilt $-E$.
\bibliographystyle{unsrt}
|
\section{Introduction.}
During the last decades we have witnessed the rapidly growing fields of Artificial Intelligence (AI) and Quantum Computing (QC). The basis for AI and QC were developed in the past century. However, it is now that this knowledge is widely available for research, business, health, among others. AI aims to provide machines with human-like intelligence. From the very beginning, AI has been conceived in different ways, leading to the development of different branches, known as Machine Learning (ML)\cite{scikit-learn,Mehta2019,Carleo2019}, Deep Learning~\cite{Paszke2017} and Reinforcement Learning~\cite{Sutton1998}. ML is based on statistical learning, where the machine learns from data that has already been labelled (Supervised learning) or from unlabelled data (Unsupervised learning). In recent years, Supervised learning has undoubtedly impacted on physics~\cite{Mehta2019,Carleo2019,Dunjko2018}. In particular, it is known for unravelling patterns from datasets that yield quantum phase transitions~\cite{Carrasquilla2017,Canabarro2019aa}.
Quantum computing is also at the forefront of current technologies. Nowadays, research groups have delivered highly functional and fault-tolerant quantum algorithms encompassing a wide variety of systems including: superconducting qubits~\cite{Kandala2017,Arute2019}, trapped ions\cite{Wright2019}, cold atoms~\cite{Graham2022}, photonics~\cite{Peruzzo2014,Arrazola2019} and color centers in diamond~\cite{Abobeih2022}. In the last years, quantum computers have pushed further the boundaries of physics, chemistry, biology, and computing itself, with groundbreaking achievements in the simulation of novel materials~\cite{Babbush2018}, molecules~\cite{Peruzzo2014,OMalley2016,Kandala2017,Nakanishi2019}, in designing algorithms towards quantum supremacy~\cite{Peng2008,Arute2019} and quantum machine learning~\cite{Rebentrost2014,Wiebe2015,Cai2015,Li2015,Biamonte2017,Havlicek2019,Schuld2019,He2019,Mengoni2019,Bartkiewicz2020,Johri2021,Willsch2020,Zhang2020,Park2020,Khan2020,Schuld2021,Goto2021,Wang2021,Gyurik2021,Saeedi2021,Ding2021}.
Among the main obstacles to be overcome in the development of quantum technologies is the interaction of the quantum system with the environment. This interaction disturbs the quantum state and, in general, can be divided into two types of processes: Markovian and non-Markovian \cite{BreuerPet}. Non-Markovian processes are those in which memory effects are taken into account and their importance can be noted in several processes and protocols such as state teleportation \cite{laine2014}, quantum metrology \cite{chin2012} and even in current quantum computers \cite{white2020}. In this paper we use quantum machine learning to determine the degree of non-Markovianity of a quantum process. We focus on kernel-based machine learning models to learn from quantum states. Our results shows that the quantum computer can create the dataset, but also treat and learn from it, providing feedback on the very process in which it is involved.
The paper is organized as follows. In Sec.~\ref{Sec_QSVM} we introduce two quantum machine learning models based on kernels, namely: Quantum Support Vector Machine and Quantum Kernel Ridge models. The goal of these models is to estimate the degree of non-Markovianity from a dataset made of quantum states. Furthermore, we elaborate on the performance of the models based on three different kernel functions and four different kernel circuits to measure the overlap between two quantum states. All these possible combinations yield different Gram matrices. In Sec.~\ref{Sec_DQS}, we introduce the Digital Quantum Simulation approach that we followed to describe the evolution of the system in Amplitude Damping and Phase Damping channels. In Sec.~\ref{Sec_results}, we show our main results regarding the prediction of the degree of non-Markovianity. In Sec.~\ref{Sec_conclusions} we deliver the final remarks of this work.
\section{Kernel-based machine learning models.}\label{Sec_QSVM}
Quantum machine learning aims to perform machine learning tasks assisted by a quantum computer. In recent years, different implementations have been addressed, including Variational Quantum Circuits~\cite{Benedetti2019,Cerezo2021,Dinani2022}, quantum Nearest-Neighbor methods~\cite{Wiebe2015} and quantum Kernel Methods~\cite{Rebentrost2014,Li2015,Schuld2021}. The latter, naturally appears in models that support a kernel function to represent the data into a feature space. Two well-understood examples are the Support Vector Machine (SVM) and the Kernel Ridge Regressor (KRR) models. Their extension to the quantum domain via a precomputed kernel is straightforward. Next, we describe the SVM and KRR models and their connection with the kernel.
\subsection{Support Vector Machine}
One of the most broadly used models in ML is Support Vector Machines (SVM)~\cite{Vapnik1995}. This model can be used for classification~\cite{Burges1998,Opper2001} and regression~\cite{Vapnik1995,Scholkpf2000,Smola2004} tasks. The former, gives rise to an intuitive representation that relies on a hyperplane that splits the dataset into different classes. Therefore, predicting the label of unknown data only depends on where the data samples fall regarding the hyperplane. In general, other models also use a hyperplane. However, the SVM sets the maximum-margin, i.e. maximizing the distance between the hyperplane and some of the boundary training data, which are the data samples close to the edge of the class. These particular samples are known as support vectors (SVs). Since SVs are a subset of the training dataset, this model is suitable for situations where the number of training data samples is small compared to the feature vector's dimension. Once the model has fitted the training dataset, it can be used as a decision function that predicts new samples, without holding the training dataset (eager learning algorithm) in memory. In this work we will focus on a regression task, which predicts a real number rather than a class. In what follows, we briefly describe the mathematical formulation of the optimization problem. More details can be found in Ref.~\cite{Fanchini2021}.
SVM delivers the tools for finding a function $f(\vec{x})$ that fits the training dataset $\{\vec{x}_i,y_i \}$, where $\vec{x}_i\in \mathbb{R}^d$ are the feature vectors with dimension $d$, and $y_i\in\mathbb{R}$ are the corresponding labels. Note that $i$ runs over the number of training samples ($i=1,2,\dots,\mathit{l}$). We begin with the linear function $f(\vec{x})=\vec{w}\cdot\vec{x} +b$, with $\vec{w}\in \mathbb{R}^d$ and $b\in \mathbb{R}$ being fitting parameters. We shall discuss the case of nonlinear separable data later on. For $\epsilon$-SVM~\cite{Vapnik1995}, deviations of $f(\vec{x})$ from the labeled data ($y_i$) must be smaller than $\epsilon$, i.e. $\vert f(\vec{x})-y_i \vert\leq\epsilon$. Moreover, we must address the model complexity as given by the $l_2$-norm $\left\Vert\vec{w}\right\Vert^2$, and the tolerance for deviations $\xi_i,\xi_i^{\ast}$ (slack variables) larger than $\epsilon$, that are weighted by $C>0$. Therefore, the optimization problem can be stated as~\cite{scikit-learn,Vapnik1995,Smola2004},
\begin{eqnarray}
&\mbox{minimize} \hspace*{1cm} \frac{1}{2} \left\Vert \vec{w}\right\Vert ^2 + C\sum_i \left( \xi_i + \xi_i^\ast \right) \nonumber\\
&\mbox{subjected to} \hspace*{1cm} \left\lbrace \begin{array}{l}
y_i - \vec{w}\cdot\vec{x}_i -b \leq \epsilon + \xi_i \\
\vec{w}\cdot\vec{x}_i +b - y_i\leq \epsilon + \xi_i^\ast \\
\xi_i,\xi_i^\ast \geq 0
\end{array} \right.
\end{eqnarray}
One can solve this problem introducing the Lagrange multipliers $\alpha_i,\alpha^\ast_i,\eta_i,\eta^\ast_i \geq 0$, with the Lagrangian defined as~\cite{Vapnik1995,Smola2004,Scholkpf2000},
\begin{align}\label{lagrangian}
L &=\frac{1}{2}\left\Vert \vec{w}\right\Vert^2 + C\sum_i \left( \xi_i + \xi_i^\ast \right) - \sum_i (\eta_i\xi_i +\eta_i^\ast\xi_i^\ast)\nonumber \\
& -\sum_i\alpha_i(\epsilon + \xi_i -y_i + \vec{w}\cdot\vec{x}_i +b)\nonumber \\
& -\sum_i\alpha_i^\ast (\epsilon + \xi_i^\ast + y_i -\vec{w}\cdot\vec{x}_i -b).
\end{align}
From the vanishing partial derivatives $\partial_bL$, $\partial_wL$, $\partial_{\xi}L$ and $\partial_{\xi^\ast}L$ the optimization problem can be recast as,
\begin{align}
&\mbox{maximize} \hspace*{1cm} \left\lbrace \begin{array}{l}-\frac{1}{2}\sum_{i,j} (\alpha_i-\alpha_i^\ast)(\alpha_j-\alpha_j^\ast)\langle x_i,x_j\rangle \\
-\epsilon\sum_i (\alpha_i + \alpha_i^\ast) + \sum_i y_i(\alpha_i - \alpha_i^\ast) \end{array}\right. \nonumber \\
&\mbox{subjected to} \hspace*{1cm} \left\lbrace \begin{array}{l}
\sum_i (\alpha_i-\alpha_i^\ast)=0 \\
\alpha_i,\alpha_i^\ast\in[0,C]\end{array}\right.
\end{align}
For convenience, we have written the dot product as an inner product, $\langle x_i,x_j\rangle=\vec{x}_i\cdot\vec{x}_j$. From $\partial_wL=0$ we find $\vec{w}=\sum_i(\alpha_i - \alpha_i^\ast)\vec{x}_i$, that leads to the decision function
\begin{equation}\label{f_decision}
f(\vec{x})=\sum_i(\alpha_i - \alpha_i^\ast)\langle x_i,x\rangle + b,
\end{equation}
that depends on the inner product between the unlabeled data ($\vec{x}$) and the training data ($\vec{x}_i$). We can recover $b$ from the Karush-Kuhn-Tucker (KKT) condition, which states that at the solution point of the Lagrangian, the product between the Lagrange multipliers and the conditions vanishes. We remark that this calculation is computed internally in \textit{scikit-learn} library~\cite{scikit-learn}. We would like to stress that the decision function in Eq.~\eqref{f_decision} has a sparse representation in terms of $\alpha_i,\alpha_i^\ast$. Only a small subset of the training dataset (support vectors) contributes to the decision function. In Appendix~\ref{Appendix_svm}\textcolor{purple}{,} we show the arguments for the sparsity and the calculation of $b$.
We have introduced so far a linear decision function that can handle linearly separated data. For nonlinearly separated data, it is possible to define a clever kernel function $k(x_i,x)$ that generalizes $\langle x_i,x\rangle$ by taking the samples to a higher dimensional space, where they are linearly separable. We elaborate further on this idea later on.
\subsection{Kernel Ridge Regressor}
Kernel Ridge Regressor (KRR) is another important nonlinear machine learning model. It has been successfully used to predict the evolution of quantum systems~\cite{Rodriguez2022}. It combines Ridge Regression with the kernel trick \cite{scikit-learn,Hastie2009}. The former, provides a linear solution based on least squares with $l_2$ regularization that penalizes large coefficients. Like in SVM, the $l_2$-norm prevents model complexity, while the kernel allows the model to learn a nonlinear function in the original space. This model offers a straightforward optimization problem stated by \cite{scikit-learn}
\begin{equation}\label{rr}
\mbox{minimize} \hspace*{1cm} \sum_{i=1}^N \left\Vert \vec{w}\cdot\vec{x}_i-y_i\right\Vert ^2 + \alpha \left\Vert \vec{w} \right\Vert.
\end{equation}
The above problem can be written in an equivalent way as~\cite{Hastie2009},
\begin{align}\label{rr_dual}
&\mbox{minimize} \hspace*{1cm} \sum_{i=1}^N \left( y_i - \vec{w}\cdot\vec{x}_i - b \right)^2, \nonumber\\
&\mbox{subjected to} \hspace*{1cm} \left\Vert \vec{w} \right\Vert^2\leq \alpha_d,
\end{align}
where there is a one-to-one correspondence between the hyperparameters $\alpha$ and $\alpha_d$. Introducing the Lagrange multipliers as in the previous subsection the decision function can be found as,
\begin{equation}\label{f_decision_krr}
f(\vec{x})=\sum_i\beta_i k(x_i,x) + b,
\end{equation}
It is worth noting that SVM and KRR are similar in terms of the $l_2$ regularization and that both use the kernel trick, but the loss function is different. While SVM relies on a linear $\epsilon$-insensitive loss, KRR uses squared error loss. The former implies that all the training points that result in errors that fall inside the $\epsilon$-tube do not contribute in the solution, which originates sparseness. In contrast, KRR considers all the training points. This yields differences in the performance of these models.
Machine learning algorithms have greatly profited from kernel functions~\cite{Dunjko2018,Mengoni2019,Schuld2021}. Therefore, we now introduce a generalization of the decision function to learn from nonlinear data. The kernel can be understood as a measure of similarities between two vectors, and it supports representations ranging from polynomial to exponential functions~\cite{scikit-learn}. Along this paper we consider three different functions for the kernel $k(x_i,x_j)$, namely: linear $\langle x_i,x_j\rangle + c$, polynomial $\left(\langle x_i,x_j\rangle +c\right)^d$, and exponential $\exp(-\sigma\sqrt{1-\langle x_i,x_j\rangle})$.
We have so far addressed the classical part (optimization problem) of this hybrid quantum machine learning approach. In the next subsection we will focus on implementing the kernel through a quantum circuit.
\subsection{Quantum Kernels}
We have noted that the kernel provides efficient separability in nonlinear regions. The main idea behind the kernel is that it allows to map the data to a higher-dimensional space, termed as “featured space”~\cite{Fanchini2021}. In general lines, let’s consider a feature map $\phi:x\in\chi\rightarrow\phi(x)\in H$ that encodes information from a certain domain $\chi$ (commonly $\chi\in R^n$) to a feature space $H$. The advantages of using the map rely on the “kernel trick”~\cite{Dunjko2018}, which allows us to set the decision function without the explicit calculation of $\phi(x)$. This idea has encouraged researchers to bridge classical and quantum machine learning \cite{Havlicek2019,Schuld2019,Schuld2021}. Let’s consider a Hilbert space $H$ that contains the states of a quantum system. Now, instead of encoding the information of $\chi$ in a feature space given by functions $\phi(x)$, with $x\in\chi$, the information is encoded in quantum states $\ket{\phi(x)}\in H$~\cite{Schuld2021,LaRose2020,Weigold2021}, which is known as quantum embedding. Quantum embedding is a crucial step in the process and, in some cases, may lead to a disadvantage against classical models. To overcome this, we resorted to perform digital quantum simulation of the quantum dynamics rather than classical simulation~\cite{Fanchini2021}, which allows us to handle quantum states to build up the kernel. Thus, we train our model with a symmetric and semi-positive definite matrix (Gram matrix), rather than the data samples (quantum states).
The next step is to calculate the kernel from the training samples $\rho_i$. A natural choice is the pairwise trace distance between the quantum states ($\mbox{Tr}[\rho_i\rho_j]$), that is commonly carried by the Swap test~\cite{Smolin1996,Cincio2018}. In what follows we describe the circuit implementation. First, we encode the information into two different qubits. Each of these qubits undergoes a NM evolution (induced by independent ancilla qubits). Then the overlap between states $\rho_i$ and $\rho_j$ yields the matrix element $k(\theta_i,\theta_j) = \mbox{Tr}[\rho_i\rho_j]$, where $\theta_i$ is the parameter that control the NM evolution. We note that for the case of pure states, $\rho_i=\vert\psi_i\rangle\langle\psi_i\vert$ and $\rho_j=\vert\psi_j\rangle\langle\psi_j\vert$, the kernel simply reduces to $\vert\langle\psi_i\vert\psi_j\rangle\vert^2$.
We describe next different implementations for the overlapping.
\subsubsection{Swap test}
The Swap test is a high level sequence of quantum operations that involves two data qubits, an ancilla qubit, two-qubits (CNOT) gates, one-qubit gates and a final measurement on the ancilla~\cite{Smolin1996}, see Fig.~\ref{fig:kernel}. By measuring the probability of finding the ancilla in state $\ket{0}$ ($P_0$), one obtains the state overlapping by computing $\mbox{Tr}[\rho_i\rho_j] = 2P_0-1$.
\subsubsection{Inversion test}
Our second kernel considers the quantum state of a closed system (unitary evolution), that encompasses the system qubit and the environment ancilla qubit~\cite{Lloyd2020}. It begins with two different quantum states driven by an unitary evolution $U(\theta)$, such that $\ket{\Psi_{\theta}}=U(\theta)\vert00\rangle$, with $\vert00\rangle = \ket{0}_s\otimes\ket{0}_a$. The kernel is defined as the squared absolute value of the projection between these two states, that is equivalent to two subsequent evolutions--- assuming that the inverse evolution $U^{\dag}(\theta_i)$ can be implemented. The matrix elements reads,
\begin{eqnarray}
k(\theta_i,\theta_j) &=& |\bra{\Psi_{\theta_i}}\Psi_{\theta_j}\rangle|^2 = |\bra{00}U^{\dag}(\theta_i)U(\theta_j)\ket{00}|^2\nonumber \\
&=& |\bra{00}\Theta\rangle|^2,
\end{eqnarray}
\begin{figure*}[ht]
\centering
\includegraphics[width=0.7 \textwidth]{fig_circ.eps}
\caption{ Quantum circuits compute the overlap between two quantum states in the kernel function to calculate the Gram matrix. For the inversion test $U$ represents either the amplitude damping or phase damping channel depicted in Fig.~\ref{fig:channels_circ}. For the ancilla-based algorithm (ABA) $U=T^\dagger H$~\cite{Cincio2018}.}
\label{fig:kernel}
\end{figure*}
where $\ket{\Theta} = U^{\dag}(\theta_i)U(\theta_j)\ket{00}$. In contrast to the Swap test kernel, this one requires two measurements, which allows us to decrease the number of quantum registers (Fig.~\ref{fig:kernel}). We remark that this kernel is not experimentally feasible for the particular goal of detecting non-Markovianity. In general, one has no access to perform measurements upon the environment. In addition, it requieres reverse unitary interactions of the system-environment dynamics. Nevertheless, we consider it because it may be applied to other machine learning tasks~\cite{Lloyd2020} and it delivers the best accuracy we found in this paper.
\subsubsection{Ancilla-based algorithm}
The Ancilla-based algorithm (ABA) is a variation of the Swap test that conveniently reduces the number of gates. It was first discovered in the context of quantum optics~\cite{Garcia2013}, and rediscovered later with assistance of a neural network and introduced for quantum circuits~\cite{Cincio2018}. The circuit is depicted in Fig.~\ref{fig:kernel}.
\subsubsection{Bell-basis algorithm}
The Bell-basis algorithm (BBA) considers less resources than the previous one (ABA), but demands Bell-basis measurements on all the system qubits~\cite{Cincio2018}. The circuit is depicted in Fig.~\ref{fig:kernel}.
In this paper we do not intent to explicitly compare the accuracy of all these approaches for estimating the overlapping (for a comparison between Swap test, ABA and BBA see~\cite{Cincio2018}). We will compare them in terms of the accuracy of the decision function.
In the next section we describe the quantum circuits that account for the interaction between the system qubit with the environment ancilla that ultimately yields non-Markovianity.
\section{Digital Quantum Simulation of non-Markovian channels}\label{Sec_DQS}
The main purpose of this paper is to determine the degree of non-Markovianiaty of a quantum process using a quantum machine learning algorithm. We begin simulating two non-Markovian channels, amplitude damping and phase damping, whose degree of non-Markovianity can be controlled. For this purpose we simulate the processes using usual circuit routines, taking auxiliary qubits to represent the environment. In this section, we show how the degree of non-Markoviniaty is calculated and present how the non-Markovian amplitude damping and phase damping processes can be simulated using a quantum circuit.
\subsection{Calculating the degree of non-Markovianity}
There are different ways to measure the degree of non-Markovianity. The most popular measures are based on the trace distance dynamics~\cite{Breuer2009}, the dynamics of entanglement~\cite{Rivas2010,Chruscinski2011}, and mutual information \cite{Luo2012}, among others~\cite{Pollock2018}. In this paper we consider the measure based on entanglement dynamics of a bipartite quantum state that encompasses the system that interacts with the environment and an ancilla qubit that is isolated from it~\cite{Rivas2010}. Worthwhile noticing that this ancilla only serves the purpose of quantifying non-Markovianity and it is not implemented in the quantum circuits, in contrast to the ancilla used to simulate the effect of the environment for the amplitude damping and phase damping processes.
A monotonic decrease in the entanglement of the bipartite system implies that the dynamics is Markovian. An increase in the entanglement during the evolution is a result of memory effects and thus non-Markovianity. The measure can be calculated as
\begin{equation}
\mathcal{N}={\text {max}}\int_{dE(t)/dt>0}{\frac{dE(t)}{dt}dt},
\end{equation}
where the maximization is done over all initial states and $E$ is the measure of entanglement.
It has been found that the maximization is achieved for Bell states \cite{Neto2016}. Therefore, we consider a bipartite system in a Bell state and use concurrence as the measure of entanglement \cite{Hill1997}.
\subsection{Amplitude Damping}\label{sub_AD}
For the amplitude damping (AD) channel, we consider a qubit interacting with a bath of harmonic oscillators, given by the Hamiltonian ($\hbar=1$) \cite{Hakkika, Whalen}
\begin{eqnarray}
&&H=\omega_0 \sigma_{+} \sigma_{-}+\sum_{k}{\omega_k a^{\dagger}_k a_k} \nonumber\\
&&\qquad+\sum_k (g^{*}_k \sigma_{+}a_k+g_k \sigma_{-}a^{\dagger}_k).
\end{eqnarray}
Here, $\sigma_{+}=\sigma^{\dagger}_{-}=|1 \rangle\langle 0|$ with $|1\rangle$ ($|0\rangle$) corresponding to the excited (ground) state of the qubit with transition frequency $\omega_0$, $a_k (a^{\dagger}_k)$ is the annihilation (creation) operator of the $k$-th mode of the bath with frequency $\omega_k$, and $g_k$ is the coupling between the qubit and the $k$-th mode. We assume that the bath has a Lorentzian spectral density
\begin{equation}
J(\omega)=\frac{1}{2\pi}\frac{\gamma_0 \lambda^2}{(\omega_0-\omega)^2+\lambda^2},
\end{equation}
where $\lambda \approx 1/\tau_r$ with $\tau_r$ being the environment correlation time, $\gamma_0\approx 1/\tau_s$ where $\tau_s$ is the typical time scale of the system.
The dynamics of the qubit that is coupled resonantly with the environment can be expressed as
\begin{equation}
\rho(t)=\sum_{i=0}^{1}{M_i(t)\rho(0)M^{\dagger}_i(t)},
\end{equation}
where the Kraus operators are given by \cite{Nielsen} \cite{Garcia-Perez}
\begin{eqnarray}
&&M_0 (t)=|0\rangle\langle0|+\sqrt{p(t)}|1\rangle\langle 1|, \\
&&M_1 (t)=\sqrt{1-p(t)}|0\rangle \langle 1|,
\end{eqnarray}
in which
\begin{equation}\label{eq:ptAD}
p(t)=e^{-\lambda t}\left[\frac{\lambda}{d}\sinh(d t/2)+\cosh(d t/2)\right]^2,
\end{equation}
with $d=\sqrt{\lambda^2-2\gamma_0\lambda}$. The dynamics is known to be non-Markovian in the strong coupling regime $\lambda<2\gamma_0$ $(\tau_s<2\tau_r)$ \cite{Bellomo}.
\begin{figure}[h]
\centering
\includegraphics[width=0.3\textwidth]{fig_decoh_circ.eps}
\caption{Quantum circuits for simulating AD and PD channels.}
\label{fig:channels_circ}
\end{figure}
The AD process can be simulated for a general scenario with a quantum circuit via an ancilla qubit \cite{Nielsen,Garcia-Perez}. After tracing out the ancilla qubit we obtain the desired mixed state. Figure \ref{fig:channels_circ} shows the quantum circuit. The Hadamard gate prepares the qubit in the superposition state $\left(|0\rangle+|1\rangle\right)/\sqrt{2}$ while the controlled rotation and CNOT gates simulate the interaction of the qubit with the environment. In this circuit, the angle $\theta_a$ is given by \cite{Nielsen, Garcia-Perez}
\begin{equation}\label{eq:thetaad}
\theta_a=2\arccos\left(\sqrt{p(t)}\right),
\end{equation}
where $p(t)$ is given in Eq.~\eqref{eq:ptAD}.
\subsection{Phase Damping}
For the phase damping (PD) channel, following Ref.~\cite{daffer04}, we consider a qubit undergoing decoherence induced by a colored noise given by the time dependent Hamiltonian ($\hbar=1$)
\begin{equation}
H(t)=\Gamma(t) \sigma_z.
\end{equation}
Here, $\Gamma(t)$ is a random variable which obeys the statistics of a random telegraph signal defined as $\Gamma(t)=\alpha (-1)^{n(t)}$, where $\alpha$ is the coupling between the qubit and the external influences, $n(t)$ is a random variable with Poisson distribution with mean $t/(2\tau)$, and $\sigma_z$ is the Pauli $z$ operator. In this case, the dynamics of the qubit is given by the following Kraus operators \cite{daffer04}
\begin{eqnarray}
&&M_0(t)=\sqrt{\frac{1+\Lambda(t)}{2}}\mathbb{I}, \\ &&M_1(t)=\sqrt{\frac{1-\Lambda(t)}{2}}\sigma_z,
\end{eqnarray}
where
\begin{equation}\label{eq:LambdaPD}
\Lambda(t)=e^{-t/(2\tau)}\left[\cos(\frac{\mu t}{2\tau})+\frac{1}{\mu}\sin(\frac{\mu t}{ 2\tau})\right],
\end{equation}
with $\mu=\sqrt{(4 \alpha \tau)^2-1}$, and $\mathbb{I}$ being the identity matrix.
For $\alpha \tau>1/4$ the dynamics is non-Markovian, while for $\alpha \tau<1/4$ it is Markovian. The PD channel can be simulated using a quantum circuit, shown in Fig.~\ref{fig:channels_circ}~\cite{Nielsen}. In this circuit, the Hadamard gate prepares the qubit into the superposition state and the controlled rotation simulates the interaction with the environment. The angle $\theta_p$ is given by
\begin{equation}\label{eq:thetapd}
\theta_p=2\arccos\left(\Lambda(t)\right),
\end{equation}
where $\Lambda(t)$ is given in Eq.~\eqref{eq:LambdaPD}.
\section{Results}\label{Sec_results}
\begin{figure*}
\centering
\includegraphics[width=0.7 \textwidth]{exp_val.eps}
\caption{Expectations values delivered by the noisy \textit{qasm\_simulator} exhibit small dispersion given by the shot-noise, in contrast to the ideal statevector\_simulator. We only observe correlations in the plane defined by $\sigma_x$ and $\sigma_z$.}
\label{fig:exp_val}
\end{figure*}
We perform our simulations with the \textit{statevector\_simulator} and \textit{qasm\_simulator}, integrated in the Aer's package from IBM qiskit~\cite{qiskit}. For comparison, we also run simulations using Pennylane library~\cite{Bergholm2018}, obtaining similar outcomes. The \textit{statevector\_simulator} is an ideal simulator that considers the evolution of the wavefunction. In contrast, the \textit{qasm\_simulator} mimics the open dynamics of the IBM quantum computer. This means that it considers losses and shot-noise. However, it allows us to set all qubits equal and fully connected (not relying on a specific quantum hardware).
It is well-known that the quantum state of a qubit can be represented as a point in a sphere of radius one (Bloch's sphere). A generic state can be represented in the Bloch's sphere in terms of the expectations values as
\begin{equation}
\rho=\frac{1}{2}\left(\mathbb{I} + \sum_{i=x,y,z}\langle\sigma_i\rangle\sigma_i\right),
\end{equation}
where $\mathbb{I}$ is the $2\times2$ identity matrix.
For illustration we firstly focus on the amplitude damping channel. In Fig.~\ref{fig:exp_val} we show the expectation values calculated using the \textit{statevector\_simulator} and \textit{qasm\_simulator}. The former, provides outcomes with no dispersion (top), as expected from the ideal simulation. On the other hand, \textit{qasm\_simulator} delivers more realistic results that include dispersion (bottom). This dispersion will be pivotal for selecting the best algorithm that computes the overlap, since \textit{statevector\_simulator} brings no significant difference in the prediction. In order words, simulations on \textit{statevector\_simulator} may be misleading when selecting a machine learning model.
In Fig.~\ref{fig:kernels} we show the degree of NM for the amplitude damping channel as a function of the parameter $\theta$ (rotation angle that controls NM introduced in subsection~\ref{sub_AD}). For the calculations, we used \textit{qasm\_simulator} with the exponential kernel function ---that yields the best accuracy as shown in Appendix~\ref{Appendix_kernel_func}. For exploration of the algorithms we only focus on QSVM. We manually seek optimal hyperparameters and report the prediction on the training dataset. A more robust analysis will be given later on. We can observe that the inversion test leads to a feature space that allows better prediction of the degree of NM.
\begin{figure}
\centering
\includegraphics[width=0.45 \textwidth]{fig_kernels.eps}
\caption{QSVM prediction of non-Markovianity as a function of the rotation angle $\theta$ for different kernel circuits. The inversion test outperforms the others. We set the hyperparameters $\{C=0.5,\epsilon=0.01 \}$.}
\label{fig:kernels}
\end{figure}
We now compare the performance between QSVM and QKRR. Hereafter, we focus on simulations on the \textit{qasm\_simulator} for the inversion test with exponential function. To prevent overfitting, we use two steps for cross-validation. First, we use the \textit{train\_test\_split} function in scikit-learn~\cite{scikit-learn} to randomly split the training set from the test set. Then, we use the \textit{GridSearchCV} function to explore the best fitting hyperparameters for each model, and we use a five-fold cross-validation. Thus, \textit{GridSearchCV} provides the best estimator for the range of given parameters averaged over five different sampling of the training set. Finally, we used these estimators to predict the test set, which contains the data that the model has not seen. In Fig.~\ref{fig:krr} we show our predictions for amplitude damping and phase damping. One can observe that both models succeeded in predicting the degree of non-Markovianity, besides small differences in the score (mean squared error). However, there are important aspects that might be taken into account before selecting one over the other. First, we remark that QSVM requires less training data to deliver good fittings. This is known, and it results from the sparseness in the training samples (only SVs contribute). Therefore, QSVM provides a major advantage given that the most time consuming operation is the calculation of the Gram matrix. Thus, less training samples reduces the overall computation time. In contrast, we observe that as the number of data samples increases, QKRR improves.
\begin{figure}
\centering
\includegraphics[width=0.40 \textwidth]{fig_pred.eps}
\caption{Both QSVM and QKRR deliver accurate predictions of the degree of non-markovianity, based on the mean squared error score. For a small training dataset QSVM performs better (not shown here). For a sufficiently large number of points QKRR provides a smaller mean squared error. }
\label{fig:krr}
\end{figure}
For comparison, we estimate the degree of non-Markovianity using a classical kernel, i.e. the radial basis function (RBF). We follow the procedure reported in Ref.~\cite{Fanchini2021}, where the training is carried out with the expectation values ($\sigma_x,\sigma_y,\sigma_z$). Thus, instead of using quantum states to build up a kernel, we resort to use classical data, i.e. measurement outcomes. However, the process to obtain the states to be measured is the same we outlined in section~\ref{Sec_DQS}--- in Ref.\cite{Fanchini2021} the authors used a master equation approach instead of digital quantum simulation.
In Table~\ref{Table1} we show the mean squared errors for each model for the amplitude damping (AD) and phase damping (PD) channels. We remark that the quantum versions, those where the kernel is calculate from the overlap between quantum states, deliver accurate predictions that are comparable with the classical models, albeit we found that SVM with a RBF kernel provides the best accuracy, as evidenced in terms of the mean squared error and the coefficient of determination $R^2$ (not shown here). This particular problem illustrates that extending the kernel to be quantum provides interesting insights and contributes to concatenate quantum blocks of operations. It not necessarily outperforms a fully classical training process but delivers useful outcomes.
\begin{table}
\caption{\label{Table1}The Table shows the accuracy of the quantum and classical version of the studied machine learning models. The hyperparameters for AD (PD) are, QSVM: $ C=4\times10^{-1} (2\times10^{-1}),\epsilon=10^{-2}$; QKRR: $ \alpha=10^{-1} (2\times10^{-1})$; SVM: $ C=10^2,\epsilon=10^{-3}$; KRR: $\alpha=10^{-4} (10^{-5})$.}
\begin{ruledtabular}
\begin{tabular}{cccccc}
& \mbox{QSVM} & \mbox{QKRR} & \mbox{SVM} & \mbox{KRR}\\
\hline
AD & $6.0\times10^{-5}$ & $2.7\times10^{-5}$ & $2.6\times10^{-6}$ & $1.4\times10^{-5}$ \\
PD & $3.3\times10^{-4}$ & $1.6\times10^{-4}$ & $5.9\times10^{-5}$ & $1.8\times10^{-4}$
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Conclusions}\label{Sec_conclusions}
In this paper we have thoroughly studied kernel-based quantum machine learning models to predict the degree of non-Markovianity using quantum data (quantum states). Each state is obtained through digital quantum simulation, where an ancilla qubit originates the non-Markovian behavior. We focus on two different decoherence channels, amplitude damping and phase damping. These quantum states are mapped to a Gram matrix by calculating its overlap. We investigate different kernel functions, say: linear, polynomial and exponential, and different kernel circuits to compute the overlap, say: inversion test, bell-basis algorithm, ancilla-based algorithm and the Swap test. We found that the inversion test with the exponential function delivers the best results. We draw our attention to two well-known kernel based machine learning models, Support Vector Machine (SVM) and Kernel Ridge (KRR). Because of their used with a precomputed quantum kernel we dubbed them as quantum SVM (QSVM) and quantum KRR (QKRR), respectively. By optimizing the learning process through cross-validation steps and grid search we found a good accuracy in our models. We found QSVM to be slightly better than QKRR, not only in the prediction's accuracy, but also in requiring less training samples.
Finally, we compare our results with their classical counterpart, i.e. when using classical data (expectation values) to train the models. While there are not significant differences, we observe that SVM with an RBF kernel delivers the best performance. This means that in this particular case it is better to measure upon the system and then process the measurement outcomes with machine learning techniques.
\section{acknowledgments} D.T. acknowledges support from Universidad Mayor through the Doctoral fellowship. A.N. acknowledges financial support from Fondecyt Iniciaci\'on No. 11220266.
|
\section{ Symmetry Beyond the Planck Scale}
A lot of progress has been made in understanding quantum gravity at low
energies and small curvatures, but what happens at the Planck scale and beyond
remains something of a mystery. It is possible that a radically different
theory could emerge at the Planck scale and there have been many speculations
as to what might happen there.
One possibility is that, as one tries to probe distance scales smaller than
the Planck length, one could find that spacetime is no longer a continuum at
such scales and there is some discrete structure instead -- perhaps a matrix
model or some form of lattice theory.
Another speculation is that there could be a new highly symmetric phase
emerging at the Planck scale. An attractive proposal is that the fundamental
theory should be conformally invariant with no dimensionful parameters, with
the Planck scale arising via symmetry breaking -- see \cite{tHooft:2015vaz} for a recent
discussion of this idea. A major obstacle to this is that the conformal
theories of a graviton field that we know of are all higher derivative, so
that although they have good ultra-violet properties, they have ghosts and it
is difficult to recover Einstein's theory from them. In string theory, it has
been proposed that at trans-Planckian energies there is a highly symmetric
phase with an infinite number of symmetries associated with an infinite number
of gauge fields becoming massless in some tensionless string limit. Evidence
for such a picture was found in high-energy scattering amplitudes by Gross and
Mende \cite{Gross:1987kza,Gross:1987ar}. More recently, it has been shown that for string theory in anti-de
Sitter space, tensionless strings arise when the curvature approaches a
critical value at the string scale giving an infinite set of symmetries and
signalling a transition to a new phase \cite{Gaberdiel:2017oqg,Giribet:2018ada,Gaberdiel:2018rqv,Eberhardt:2018ouy,Gaberdiel:2021qbb}.
The standard picture of quantum gravity is that it arises from the
quantsation of a massless spin two field, and this field is then interpreted
as the metric governing the Riemannian geometry of spacetime. However, it is
interesting to ask whether a massless spin two field provides the only way of
formulating gravity. At low energies, we would require that the theory can be
viewed as a theory of a graviton. However, this theory could have an
ultra-violet origin that is a theory of another field, perhaps in higher
dimensions.
This paper will revisit the conjecture of \cite{Hull:2000zn} in which it was proposed that
gravity could be conformally invariant but formulated not in terms of the
familiar spin-two graviton field but in terms of something more exotic. In the
free limit, the theory of \cite{Hull:2000zn} is formulated in terms of a four index gauge
field $C_{M N P Q}$ (with the symmetries of the Riemann tensor), but on
compactification on a circle, a conventional ghost-free theory of a graviton
emerges at scales small compared to the compactification scale, which also
sets the Planck scale in the compactified theory. The outstanding question is
whether there could be an interacting version of this theory that reduces to
general relativity or supergravity.
This conjecture arose from supersymmetry. There is a remarkable supermultiplet
in 6 dimensions with (4,0) supersymmetry which is fact a superconformal
multiplet so that the theory has 32 normal supersymmetries plus 32 conformal
supersymmetries. It has the gauge field $C_{M N P Q}$ together with 27
self-dual 2-forms and 42 scalars. Reducing on a circle gives 5-dimensional N=8
supergravity. It exists as a free theory, but again it is not known whether
there is an interacting version. If there is, it would be expected to be a
non-lagrangian theory of the kind arising in the (2,0) theory that arises as
the world-volume theory of the M5-brane. An interacting theory would be very
interesting, giving a conformal theory of gravity that reduces to a
conventional theory of gravity at low energies. Little progress has been made
on this proposal since it was made in 2000, partly because it was waiting for
progress in understanding the (2,0) theory. Now much more is known about the
(2,0) theory, it is perhaps time to revisit this conjecture. In this paper the (4,0) theory will be reviewed following \cite{Hull:2000zn,Hull:2000rr,Hull:2000ih,Hull:2001iu}
and the conjectures concerning it \cite{Hull:2000zn} will be re-examined.
\section{The 6D (2,0) theory \& 5D Yang-Mills}
There is considerable evidence that there exists an interacting (2,0)
non-lagrangian superconformal field theory (SCFT) in six spacetime dimensions \cite{Berkooz:1997cq,Seiberg:1997ax,Aharony:1997th}.
Moreover, this provides a key to understanding strongly-coupled maximally
supersymmetric Yang-Mills (SYM) theory in D=5 \cite{Witten:1995zh,Rozali:1997cb,Lambert:2010iw,Douglas:2010iu} and S-duality in
D=4 \cite{Witten:1995zh},\cite{Witten:2009at}.
The free (2,0) supersymmetric theory in 6-D has a 2-form gauge field $B$
whose field strength $H = dB$ is self-dual, $H = \ast H $. The other fields are
5 scalars transforming as a $\bf{5}$ under the R-symmetry group
${Sp} (2)$ and 4 symplectic Majorana-Weyl fermions transforming as a
$\bf{4}$ under ${Sp} (2)$.
It is known that there is a quantum interacting (2,0) SCFT, but it is not
clear how to formulate the interacting theory. It is clearly not a
conventional theory of gauge fields -- it is said to be non-lagrangian -- but
nonetheless it is known that its dimensional reduction to 5 or less dimensions
does give a Yang-Mills gauge theory. Dimensional reduction of the free (2,0)
theory on a circle gives 5-D N=4 abelian gauge theory with $B$ giving rise to
the 5-D vector potential $A$. The interacting (2,0) SCFT is non-lagrangian but
it reduces to a conventional field theory, 5D N=4 SYM.
The (2,0) theory arises as the strong coupling limit of 5D SYM \cite{Witten:1995zh,Rozali:1997cb,Lambert:2010iw,Douglas:2010iu}. The
interacting (2,0) theory compactified on a circle of radius $R$ gives 5D SYM
with coupling constant $g_{YM}$ with $g_{YM}^2 = R$, so that as $g_{YM}
\rightarrow \infty$ the radius $R \rightarrow \infty$ and at strong coupling
the circle decompactifies to give a 6D theory. The 5D SYM theory is
non-renormalizable and is known to have divergences \cite{Bern:2012di} so here it is regarded
as being embedded in some UV complete theory, such as string theory. The (2,0)
theory also arises as a decoupling limit of the world-volume theory of a stack
of M5-branes and in IIB string theory compactified on K3.
The 5D SYM theory has BPS 0-branes given by lifting self-dual Yang-Mills
instantons in 4 Euclidean dimensions to soliton world-lines in 4+1 dimensions,
taking the product of the 4D instanton solution with a timelike line. These
solitons have mass $M \propto |n| / g_{YM}^2$ where $n$ is the instanton
number, so that these become light as the dimensionful Yang-Mills coupling
$g_{YM}$ becomes large. These solitons should be interpreted as Kaluza-Klein
modes for a $D = 6$ theory compactified on a circle of radius $R = g_{YM}^2$
so that in the strong coupling limit an extra dimension opens up to give a
6-dimensional (2,0) supersymmetric theory \cite{Witten:1995zh}. In the abelian theory, the 5D
vector gauge field is replaced by a 6D 2-form gauge field $B_{MN}$ with
self-dual field strength, and the 5 scalar fields are all promoted to scalar
fields in 6 dimensions
The $D = 6$ theory is believed to be a non-trivial superconformally invariant
quantum theory \cite{Seiberg:1997ax} and the $D = 5$ gauge coupling arises as the radius of the
compactification circle, $g_{YM}^2 = R$ \cite{Witten:1995zh}. The relationship between the $D = 5$
and $D = 6$ theories is straightforward to establish for the free case in
which the Yang-Mills gauge group is abelian, but in the interacting theory the
6-dimensional origin of the $D = 5$ non-abelian interactions is mysterious;
there are certainly no local covariant interactions that can be written down
that give Yang-Mills interactions when dimensionally reduced \cite{Bekaert:2000qx}. Nonetheless,
the fact that these $D = 5$ and $D = 6$ theories arise as the world volume
theories of D4 and M5 branes respectively gives strong support for the
existence of such a 6-dimensional origin for the gauge interactions.
The W-bosons and magnetic strings in $D = 5$ arise from self-dual strings in
$D = 6$. At the origin of moduli space the W-bosons become massless and the
tensions of the self-dual strings in $D = 6$ must also become zero. The nature
of the theory at such points is unclear. Nonetheless, given that mysterious
interactions with no conventional field theory formulation arise in the
M5-brane world-volume theory, it is natural to seek similar unconventional
interactions elsewhere in M-theory.
\section{The 6D (4,0) theory \& 5D Supergravity}
Given the close relationship between gauge theory and gravity, it is natural
to ask whether there could be a story for gravity similar to that of the
relation between the 6D (2,0) theory and 5D SYM. Remarkably, there is a free
SCFT with (4,0) supersymmetry in 6-D, and if this has an extension to an
interacting theory, it would provide an exotic conformal theory giving
supergravity in D<6 and could arise as a strong coupling limit of 5D
supergravity.
Five-dimensional $N = 8$ supergravity (ungauged) has a global $E_6$ symmetry
and a local ${Sp} (4) = {USp} (8)$ R-symmetry. It is
non-renormalisable, and will be regarded as arising as a massless sector of
some consistent theory, such as M-theory compactified on a 6-torus, in which
the global $E_6$ symmetry is broken to a discrete U-duality subgroup \cite{Hull:1994ys}.
The massless bosonic fields consist of a graviton, 27 abelian vector fields
and 42 scalars. The dimensional coupling constant is the 5-dimensional Planck
length $l$, so that strong gravitational coupling is the limit as $l \to
\infty$. For such a limit it is natural to look first for a theory with 32
supersymmetries and ${Sp} (4)$ R-symmetry and to expect that BPS states
are protected and survive as the coupling $l$ is increased.
The multiplet has (4,0) supersymmetry in $D = 6$ and was studied in \cite{Hull:2000zn}.
The particle content was identified in \cite{Strathdee:1986jr} and in \cite{Hull:2000zn} the covariant field content and gauge symmetry was found and further studied in \cite{Hull:2000rr,Hull:2000ih,Hull:2001iu}. Actions for the free theory have been considered in \cite{Henneaux:2017xsb,Bertrand:2020nob,Bertrand:2022pyi} and further discussion of the theory can be found in \cite{Schwarz:2000zg,Minasian:2020vxn,Gunaydin:2020mod,Lekeu:2021oti,Cederwall:2020dui}.
Instead of a graviton, it has an exotic fourth-rank tensor gauge field $C_{MN
PQ}$ with the algebraic properties of the Riemann tensor
\begin{equation}
C_{M N P Q} = - C_{N M P Q} = - C_{M N Q P} = C_{P Q M N}
\end{equation}
and the gauge symmetry
\begin{equation}
\delta C_{MN \hspace{0.17em} PQ} = \partial_{[M} \chi_{N] PQ} +
\partial_{[P} \chi_{Q] MN} - 2 \partial_{[M} \chi_{NPQ]} \label{delcis}
\end{equation}
with parameter $\chi_{MPQ} = - \chi_{MQP}$. The invariant field strength is
\begin{equation}
G_{MNP \hspace{0.17em} QRS} = \frac{1}{36} (\partial_M \partial_S C_{NP
\hspace{0.17em} RS} + \ldots) = \partial_{[M} C_{NP] \hspace{0.17em} [QR,
S]}
\end{equation}
and in the (4,0) multiplet it satisfies the self-duality constraint
\begin{equation}
G_{MNP \hspace{0.17em} QRS} = \frac{1}{6} \epsilon_{MNPTUV} G^{TUV} _{QRS}
\end{equation}
or $G = \ast G$. In addition, there are 27 2-form gauge-fields with self-dual 3-form
field strengths, 42 scalars, 48 symplectic Majorana-Weyl fermions and, instead
of gravitini, 8 spinor-valued 2-forms $\Psi_{MN}^a$ which satisfy a symplectic
Majorana-Weyl constraint and have self-dual field strengths. The fermionic
gauge symmetry is of the form
\begin{equation}
\delta \Psi_{MN}^a = \partial_{[M} \varepsilon_{N]}^a
\end{equation}
with parameter a spinor-vector $\varepsilon_N^{\alpha a}$. The free theory
based on this multiplet is a superconformally invariant theory, with conformal
supergroup ${OSp}^{\ast} (8 / 8)$ \cite{Hull:2000zn,Chiodaroli:2011pp}. This has bosonic subgroup
${USp} (8) \times {SO}^{\ast} (8) = {Sp} (4) \times {SO}
(6, 2)$ and 64 fermionic generators, consisting of the 32 supersymmetries of
the (4,0) superalgebra and 32 conformal supersymmetries.
It is remarkable that, in going from $D = 5$ to $D = 6$ in the free theory,
the vector gauge fields $A_{\mu}$ are lifted to 2-forms $B_{MN}$, the
gravitini $\psi_{\mu}$ are lifted to spinor-valued 2-forms $\Psi_{MN}$ and the
graviton $h_{\mu \nu}$ is lifted to the gauge field $C_{MN P Q}$, with these
$D = 6$ gauge-fields all satisfying self-duality constraints. Electrically
charged 0-branes and magnetic strings in $D = 5$ lift to BPS self-dual strings
in $D = 6$. In \cite{Hull:2000zn} it was shown that the dimensional reduction of the
free (4,0) theory on a circle indeed gives the linearised $D = 5, N = 8$
supergravity theory, with gravitational coupling (Planck length) given by the
circle radius $l = R$. However, there are no covariant local interactions in
$D = 6$ for this multiplet that could give rise to the $D = 5$ supergravity
interactions.
There is then a close analogy between the $D = 5, N = 4$ Yang-Mills theory and
the $D = 5, N = 8$ supergravity. The linearised versions of these theories
both arise from the dimensional reduction of a free superconformal field
theory in $D = 6$, with the dimensional $D = 5$ coupling constant arising from
the radius of compactification, so that the strong coupling limit of the $D =
5$ free theories is a decompactification to $D = 6$. For the interacting $D =
5$ Yang-Mills theory, there are a number of arguments to support the
conjecture that its strong coupling limit should be an interacting
superconformal theory in $D = 6$ with (2,0) supersymmetry, even though such a
theory has not been constructed directly and indeed cannot have a conventional
field theory formulation. This led to the conjecture of \cite{Hull:2000zn} that the
situation for $D = 5$ supergravity is similar to that of $D = 5$
super-Yang-Mills, and that a certain strong coupling limit of the interacting
supergravity theory should give an interacting theory whose free limit is the
(4,0) theory in $D = 6$. In this case there is no analogue of the M5-brane
argument to support this, although the M5-brane case does set a suggestive
precedent.
The limiting theory should have some novel form of interactions which give the
non-polynomial supergravity interactions on reduction. It could be that these
are some non-local or non-covariant self-interactions of the (4,0) multiplet,
or it could be that other degrees of freedom might be needed; one candidate
might be some form of string field theory. However, if a strong coupling limit
of the theory does exist that meets the requirements assumed here, then the
limit must be a (4,0) theory in six dimensions, and this would predict the
existence of interactions arising from the strong coupling limit of the
supergravity interactions. Although it has not been possible to prove the
existence of such a limit, it is remarkable that there is such a simple
candidate theory for the limit with so many properties in common with the
(2,0) limit of the $D = 5$ gauge theory. Conversely, if there is a
6-dimensional phase of M-theory which has (4,0) supersymmetry, then its circle
reduction to $D = 5$ must give an $N = 8$ supersymmetric theory and the
scenario described here should apply.
In the context in which the $D = 5$ supergravity is the massless sector of
M-theory on $T^6$, then the $D = 6$ superconformal theory would be a field
theory sector of a new 6-dimensional superconformal phase of M-theory. This
could tell us a great deal about M-theory: it would be perhaps the most
symmetric phase of M-theory so far found, with a huge amount of unbroken gauge
symmetry, and would be a phase that is not well-described by a conventional
field theory at low energies, so that it could give new insights into the
degrees of freedom of M-theory.
\section{The Superalgebra and BPS States}
The five-dimensional $N = 2 n$ superalgebra with automorphism group ${Sp}
(n)$ is
\begin{eqnarray}
\{Q_{\alpha}^a, Q_{\beta}^b \} & =& \Omega^{ab}
(\Gamma^{\mu} C)_{\alpha \beta} P_{\mu} + \Omega^{ab} C_{\alpha \beta} K
\nonumber\\
& +& (\Gamma^{\mu} C)_{\alpha \beta} Z_{\mu}^{ab} + C_{\alpha \beta} Z^{ab}
+ \frac{1}{2} (\Gamma^{\mu \nu} C)_{\alpha \beta} Z_{\mu \nu}^{ab} \label{salg}
\end{eqnarray}
where $\mu, \nu = 0, 1, \ldots, 4$ are spacetime indices, $\alpha = 1, \ldots,
4$ are spinor indices, $a = 1, \ldots, N$ are $Sp (n)$ indices, $C_{\alpha
\beta}$ is the charge conjugation matrix and $\Omega^{ab}$ is the symplectic
invariant of $Sp (n)$. The supercharges $Q_{\alpha}^a$ are symplectic Majorana
spinors satisfying
\[ (\bar{Q})^{\alpha}_a = C^{\alpha \beta} \Omega_{ab} Q_{\beta}^b \]
For the $D=5$, $N = 4$ super-Yang-Mills theory, the charges can be identified as follows \cite{Lambert:2010iw}.
The 5 central charges $Z^{ab} = -
Z^{ba}$ with $\Omega_{a b} Z^{ab} = 0$ are proportional to the five electric
charges $q^I \propto \int tr (\ast F \phi^I)$ and are carried by massive
vector multiplets. There is no vector field coupling to the singlet central
charge $K$, which is a topological charge carried by the instantonic 0-branes
and is proportional to the instanton number. There is a corresponding
conserved topological current
\begin{equation}
j = \ast {tr} (F \wedge F) \label{inst}
\end{equation}
The spatial components $Z_i^{ab}$ ($i, j = 1, ..., 4$) of $Z_{\mu}^{ab}$ are
the magnetic charges carried by the magnetically charged strings. The other
charges on the right-hand-side of (\ref{salg}) are 2,3 and 4-brane charges. For
example, a vortex solution in 2+1 dimensions lifts to a 2-brane in 4+1
dimensions with charge $Z_{ij}$.
The $N = 8$ superalgebra has automorphism group ${Sp} (4)$ and is given
by (\ref{salg}) with $a, b = 1, \ldots, 8$. The algebra with scalar central charges only
is
\begin{equation}
\{Q_{\alpha}^a, Q_{\beta}^b \} = \hspace{0.17em} \Omega^{ab} (\Gamma^{\mu}
C)_{\alpha \beta} P_{\mu} + C_{\alpha \beta} (Z^{ab} + \Omega^{ab} K)
\end{equation}
The 27 central charges $Z^{ab}$ satisfy $Z^{ab} = - Z^{ba}$, $\Omega_{a b}
Z^{ab} = 0$, and are the electric charges for the 27 vector fields (dressed
with scalars). There are 28 central charges but only 27 vector fields, and $K$
is not the conserved charge for any gauge field. However, on dimensional
reduction from 5 to 4 dimensions, $K$ becomes one of the 28 magnetic charges
of $D = 4, N = 8$ supergravity, and is the one coupling to the gravi-photon
field (i.e. the electromagnetic field from the dimensional reduction of the
metric) so that it is the Kaluza-Klein monopole charge. U-duality requires
that 1/2-supersymmetric states with $M = |K|$ occur in the $D = 4$ BPS
spectrum, and $K$ is quantized. A key question is whether there are
1/2-supersymmetric states in $D = 5$ carrying the central charge $K$ with $M =
|K|$. Presumably $K$ should be quantized, so that $K \propto n / l$ where $n$
is an integer. In any case, there is a conserved current analogous to (\ref{inst}) given
by
\begin{equation}
j = \ast {tr} (R \wedge R)
\end{equation}
The (4,0) theory has 27 self-dual 2-forms and these couple to 27 self-dual BPS
strings, or more precisely, to strings whose charges take values in a
27-dimensional lattice. In addition, there are 42 scalars and these can couple
to BPS 3-branes in 6 dimensions.
The (4,0) superalgebra in six dimensions with central charges is
\begin{eqnarray}
\{Q_{\alpha}^a, Q_{\beta}^b \} & = \hspace{0.17em} \Omega^{ab} (\Pi_+
\Gamma^M C)_{\alpha \beta} P_M + (\Pi_+ \Gamma^M C)_{\alpha \beta} Z_M^{ab}
\nonumber\\
& + \frac{1}{6} (\Pi_+ \Gamma^{MNP} C)_{\alpha \beta} Z_{MNP}^{ab}
\end{eqnarray}
where $\Pi_+$ is the chiral projector
\[ \Pi_{\pm} = \frac{1}{2} (1 \pm \Gamma^7) \]
The 27 one-form charges satisfy $Z_M^{ab} = - Z_M^{ba}$, $Z_M^{ab} \Omega ab =
0$, while $Z_{MNP}^{ab} = Z_{MNP}^{ba}$ and is a self-dual 3-form,
\begin{equation}
Z_{MNP}^{ab} = \frac{1}{6} \epsilon_{MNPQRS} Z^{ab \hspace{0.17em} QRS} \label{sdu}
\end{equation}
The 27 charges $Z_i^{ab}$ with spatial indices $i, j = 1, \ldots, 5$ are the
string charges, and $Z_{ijk}^{ab}$ are the 36 3-brane charges, while
$Z_0^{ab}$ are the charges for space-filling 5-branes. (Note that the
self-duality condition (\ref{sdu}) implies that $Z_{0 ij}$ are not independent charges.)
On dimensional reduction to 5 dimensions, the charges decompose as \cite{Hull:2000rr}
\begin{eqnarray}
Z_M^{ab} & \to & (Z_{\mu}^{ab}, Z_5^{ab} = Z^{ab}) \nonumber\\
P^M & \to & (P^{\mu}, P^5 = K) \nonumber\\
Z_{MNP}^{ab} & \to & Z_{\mu \nu 5}^{ab} = Z_{\mu \nu}^{ab}
\end{eqnarray}
The extra component of momentum becomes the $K$-charge, so that the
5-dimensional instantonic 0-brane is, from the $D = 6$ viewpoint, a wave
carrying momentum $P^5$ in the extra circular dimension. The electric 0-branes
in $D = 5$ arise from strings winding around the 6th dimension while the
magnetic strings arise from unwrapped $D = 6$ strings. The 2-branes and
3-branes in $D = 5$ arise from 3-branes in $D = 6$, while the space-filling $D
= 5$ 4-branes come from wrapped $D = 6$ 5-branes.
\section{Strong Coupling Limits and Solitons}
The 5D SYM theory has BPS solitons carrying the central charge $K$ in the
superalgebra (\ref{salg}). These arise from self-dual or anti-self-dual YM instantons in
4 Euclidean dimensions lifted to 0-branes in 5D by taking a product of the 4D
instanton with time. They have mass
\begin{equation}
M \propto \frac{|n|}{g_{YM}^2}
\end{equation}
$\text{}$where $n$ is the instanton number, so that these become light as the
dimensionful Yang-Mills coupling $g_{YM}$ becomes large. These are interpreted
as Kaluza-Klein modes for a $D = 6$ theory compactified on a circle of radius
$R = g_{YM}^2$, so that the strong coupling limit is a decompactification
limit $R \rightarrow \infty$. The current $j = \ast {tr} (F \wedge F)$
represents the density of these solitons. The BPS states carrying the charge
$K$ fit into massive 5D supermultiplets with precisely the right structure to
be the KK modes for a (2,0) theory \cite{Hull:2000cf}. In particular, they have massive
self-dual 2-form fields $B$ with mass $m$ satisfying
\begin{equation}
{dB} = \pm m \ast B
\end{equation}
together with massive fermions and scalars.
The D=5 SYM is non-renormalizable but a UV completion can be defined within
string theory e.g. the D4-brane theory. The strong coupling limit is then
defined within string theory e.g. the strong coupling limit of multiple D4
branes gives multiple M5 branes.
As the coupling $g_{Y M}$ is dimensionful, the limit is better expressed as a
high energy limit, going to energies $E$ large compared to the YM scale,
\begin{equation}
E \gg \frac{1}{g_{Y M}^2}
\end{equation}
so that at ultra-high energies the extra circle dimension becomes observable.
In the free case, the dimensional reduction can be done explicitly and the
free (2,0) theory reduced on a circle gives abelian SYM in 5D. However, in the
abelian case there are no non-singular instantons and so no solitonic
0-branes, but the Kaluza Klein modes can be associated with singular zero-size
instantons. A similar issue arises for the non-abelian SYM with the gauge
symmetry spontaneously broken to an abelian subgroup, with the instantonic
solitons shrinking to zero size as the Higgs expectation value is turned on.
However, in this case the soliton can be stabilized by turning on an electric
charge to give a regular BPS soliton solution \cite{Lambert:1999ua}, so that there should be BPS
0-branes in the 5D theory even when the symmetry becomes abelian.
The free (4,0) theory compactified on a circle of radius $R$ gives 5D
linearised supergravity with 5D Planck length $l = R$ plus Kaluza-Klein modes
with mass
\begin{equation}
M \propto \frac{|n|}{R} = \frac{|n|}{l} \label{spec}
\end{equation}
for integers $n$.
For the interacting 5D N=8 supergravity theory (embedded in string theory to
give a UV completion, as in M-theory compactified on a 6-torus), a necessary
condition for it to decompactify to a 6D (4,0) theory is that it should have
1/2-BPS states carrying the charge $K$ with a spectrum of the form (\ref{spec}). From \cite{Hull:2000cf},
the 1/2-BPS states carrying $K$ fit into short 5D multiplets whose highest
spin field is a massive self-dual field $C_{\mu \nu \rho \sigma}$ satisfying
\begin{equation}
\partial_{[\mu } C_{\nu \rho] \sigma \tau} = \pm
\frac{1}{2} m \epsilon_{\mu \nu \rho} ^{\alpha \beta} C_{\alpha \beta \sigma
\tau}
\end{equation}
as required for the KK modes of $C_{M N {PQ}}$.
The key question is then whether there are BPS states carrying the charge
$K$, and if so, what is their spectrum. This then provides a crucial test for the conjecture that there is a 6D (4,0) theory arising as a limit of M-theory.
M-theory toroidally compactified to 5D should have certain BPS states carrying the singlet central charge $K$. If there are no such states, then the conjecture is false, while the presence of such states would be strong indication that
the conjecture could be true.
Independent of this conjecture, it is interesting in
any case to ask whether there are BPS states carrying $K$. All other central
charges in the 5D superalgebra are carried by BPS states and so it would be
strange if there were none carrying $K$. The charge $K$ is a singlet under the
$E_6 (\mathbb{Z})$ U-duality and the ${Sp} (4)$ R-symmetry so such BPS
states would not be related to any of the standard BPS branes by duality.
The explicit form for $K$ can be found from the superalgebra and was given in
\cite{Hull:1997kt}. If the 5D theory is compactified on a circle to 4D, then there are BPS
states carrying $K$ which are the KK monopoles, with $K$ related to the NUT
charge and hence the 4D magnetic charge \cite{Hull:1997kt}.
Just as for 5D SYM, for supergravity taking the product of a self-dual
gravitational instanton in 4 Euclidean dimensions with a timelike line gives a
1/2-supersymmetric configuration associated with the charge $K$. It would be
desirable to have such a solution which is asymptotic to 5D Minkowski space
but unfortunately there are no such non-trivial non-singular solutions. There
are no asymptotically Euclidean gravitational instantons but there are ALE
ones which are asymptotic to a discrete quotient of 4D Euclidean space (See e.g. \cite{Eguchi:1980jx}). Thus
there are no regular solitons with the desired asymptotics. However, this does
not preclude the existence of BPS states of this kind: they could be
associated with singular solutions or zero-size gravitational instantons.
The role of gravitational instantons is further suggested by the topological symmetry associated with
the current $j = \ast {tr} (R \wedge R)$ associated with the gravitational instanton density.
It
is interesting that there remains a possibility of further BPS states in
M-theory that are not dual to any of the known branes and which correspond to
singular supergravity solutions.
\section{Gravity and Geometry}
Particularly intriguing are the consequences for gravity. In $D = 5$, gravity
is described geometrically in terms of a metric $g_{\mu \nu}$, but at strong
coupling it is described instead in terms of the gauge field $C_{MNPQ}$ in $D
= 6$ (at least in the free case). This suggests the possibility of some new
structure which reduces to Riemannian geometry but which is more general and
is the appropriate language for describing gravity beyond the Planck scale.
For example, while $g_{\mu \nu}$ provides a norm for vectors and a notion of
length, $C_{MNPQ}$ could provide a norm $C_{MNPQ} \omega^{MN} \omega^{PQ}$ for
2-forms $\omega^{MN}$ and hence gives a notion of area that is not derived
from a concept of length \cite{Hull:2000ih}. It may be that the interacting theory is not
described in conventional $D = 6$ spacetime at all, but in some other arena.
There seem to be three main possibilities. The first is that there is no
interacting version of the (4,0) theory, that it only exists as a free theory,
and that the limit proposed in\cite{Hull:2000zn} only exists for the free $D = 5$
theory. The second is that an interacting form of the theory does exist in 6
spacetime dimensions, with $D = 6$ diffeomorphism symmetry. The absence of a
spacetime metric means that such a generally covariant theory would be of an
unusual kind. The third and perhaps the most interesting possibility is that
the theory that reduces to the interacting supergravity in $D = 5$ is not a
diffeomorphism-invariant theory in six spacetime dimensions, but is something
more exotic.
For gravity in any dimension, the full non-linear gauge symmetry is
\begin{equation}
\delta g_{\mu \nu} = 2 \nabla_{(\mu} \xi_{\nu)} \label{dif}
\end{equation}
If the metric is written as
\begin{equation}
g_{\mu \nu} = \bar{g}_{\mu \nu} + h_{\mu \nu}
\end{equation}
in terms of a fluctuation $h_{\mu \nu}$ about some background metric
$\bar{g}_{\mu\nu}$ (e.g. a flat background metric) then two main types of
symmetry emerge. The first consists of \lq background reparameterizations'
\begin{equation}
\delta \bar{g}_{\mu \nu} = 2 \bar{\nabla}_{(\mu } \xi_{\nu)},
\qquad \delta h_{\mu \nu} =\mathcal{L}_{\xi} h_{\mu \nu} \label{bac}
\end{equation}
where $\bar{\nabla}$ is the background covariant derivative with connection
constructed from $\bar{g}_{\mu \nu}$, while $h_{\mu \nu}$ transforms as a
tensor ($\mathcal{L}_{\xi}$ is the Lie derivative with respect to the vector
field $\xi$), as do all other covariant fields. The second is the \lq gauge
symmetry' of the form
\begin{equation}
\delta \bar{g}_{\mu \nu} = 0, \qquad \delta h_{\mu \nu} = 2 \nabla_{(\mu}
\zeta_{\nu)} \label{gag}
\end{equation}
in which $h_{\mu \nu}$ transforms as a gauge field and the background is
invariant. There is in addition the standard shift symmetry under which
\begin{equation}
\delta \bar{g}_{\mu \nu} = \alpha_{\mu \nu}, \qquad \delta h_{\mu \nu} = -
\alpha_{\mu \nu}
\label{shift}
\end{equation}
In terms of the full metric $g_{\mu \nu}$, there is no shift symmetry and a
unique gauge symmetry (\ref{dif}); the various types of symmetry (\ref{bac}),(\ref{gag}),(\ref{shift}) are an artifice of
the background split. The shift symmetry is a signal of background
independence and plays an important role in the interacting theory.
The linearised $D = 5$ supergravity theory has both background
reparameterization and gauge invariances given by the linearised forms of (\ref{bac}),(\ref{gag})
respectively, and both of these have origins in $D = 6$ symmetries of the free
(4,0) theory. The background reparameterization invariance lifts to the
linearised $D = 6$ background reparameterization invariance
\begin{equation}
\delta \bar{g}_{MN} = 2 \partial_{(M} \xi_{N)}, \qquad \delta C_{MNPQ}
=\mathcal{L}_{\xi} C_{MNPQ}
\end{equation}
with the transformations leaving the flat background metric $\bar{g}_{MN}$
invariant forming the $D = 6$ Poincar{\'e} group. The $D = 5$ gauge symmetry
given by the linearised form of (\ref{gag}) arises from the $D = 6$ gauge symmetry (\ref{delcis})
with $\delta \bar{g}_{MN} = 0$ and the parameters related by $\zeta^{\mu} =
\chi^{55 \mu}$. The $D = 6$ theory has no analogue of the shift symmetry, and
the emergence of that symmetry on reduction to $D = 5$ and dualising to
formulate the theory in terms of a graviton $h_{\mu \nu}$ comes as a surprise
from this viewpoint.
The gravitational interactions of the full supergravity theory in $D = 5$ are
best expressed geometrically in terms of the total metric $g_{\mu \nu}$. If an
interacting form of the (4,0) theory exists that reduces to the $D = 5$
supergravity, it must be of an unusual kind. One possibility is that there is
no background metric of any kind in $D = 6$, and the full theory is formulated
in terms of a total field corresponding to $C$, with a spacetime metric
emerging only in a particular background $C$ field and a particular limit
corresponding to the free theory limit in $D = 5$.
It is not even clear that the interacting theory should be formulated in a $D
= 6$ spacetime. In $D = 5$, the diffeomorphisms act on the coordinates as
\begin{equation}
\delta x^{\mu} = \xi^{\mu}
\end{equation}
In the (4,0) theory, the parameter $\xi^{\mu}$ lifts to a parameter
$\chi^{MNP}$. If the coordinate transformations were to lift, it could be to
something like a manifold with coordinates $X^{MNP}$ transforming through
reparameterisations
\begin{equation}
\delta X^{MNP} = \chi^{MNP}
\end{equation}
with the $D = 5$ spacetime arising as a submanifold with $x^{\mu} = X^{55
\mu}$. Another possibility is as follows. The diffeomorphism $\delta x^{\mu} =
\xi^{\mu}$ gives a gauge transformation for the graviton
\begin{equation}
\delta g_{\mu \nu} = \partial_{(\mu } \xi_{ \nu)} +
\cdots
\end{equation}
with the index on the parameter lowered with the metric
\begin{equation}
\xi_{\mu} = \xi^{\nu} g_{\mu \nu}
\end{equation}
The parameter of the gauge transformation
\begin{equation}
\delta C_{MN \hspace{0.17em} PQ} = \partial_{[M} \chi_{N] PQ} +
\partial_{[P} \chi_{Q] MN} - 2 \partial_{[M} \chi_{NPQ]}
\end{equation}
could be related to that of a 6D diffeomorphism
\begin{equation}
\delta X^M = \xi^M
\end{equation}
by
\begin{equation}
\chi_{M N P} = \xi^Q C_{{MNPQ}}
\end{equation}
so that the gauge symmetry could after all be that of conventional 6D
diffeomorphisms.\footnote{I would like to thank Paul de Medeiros for discussions about this suggestion.}
Similar considerations apply to the local supersymmetry transformations. In
$D = 5$, the local supersymmetry transformations in a supergravity background
give rise to `background supersymmetry transformations' with symplectic
Majorana spinor parameters $\varepsilon^{\alpha a}$ (where $\alpha$ is a $D =
5$ spinor index and $a = 1, .., 8$ labels the 8 supersymmetries) in which the
gravitino fluctuation $\psi_{\mu}^a$ transforms without a derivative of
$\varepsilon^a$, and `gauge supersymmetries' with spinor parameter
$\varepsilon^{\alpha a}$ under which
\[ \delta \psi^a_{\mu} = \partial_{\mu} \varepsilon^a + \ldots \]
The background symmetries preserving a flat space background form the $D = 5$
super-Poincar{\'e} group. In the free theory, the $D = 5$ super-Poincar{\'e}
symmetry lifts to part of a $D = 6$ super-Poincar{\'e} symmetry with $D = 5$
translation parameters $\xi^{\mu}$ lifting to $D = 6$ ones $\Xi^M$ and
supersymmetry parameters $\varepsilon$ lifting to $D = 6$ spinor parameters
$\hat{\varepsilon}$. The corresponding $D = 6$ supersymmetry charges $Q$ and
momenta $P$ are generators of the (4,0) super-Poincar{\'e} algebra with
\begin{equation} \{Q_{\alpha}^a, Q_{\beta}^b \} = \hspace{0.17em} \Omega^{ab} (\Pi_+
\Gamma^M C)_{\alpha \beta} P_M
\label{fialgss}
\end{equation}
where $\Pi_{\pm}$ are the chiral projectors
\[ \Pi_{\pm} = \frac{1}{2} (1 \pm \Gamma^7) \]
$\alpha, \beta$ are $D = 6$ spinor indices and $a, b = 1, \ldots, 8$ are
${USp} (8)$ indices, with $\Omega^{ab}$ the ${USp} (8)$-invariant
anti-symmetric tensor. This is in turn part of the $D = 6$ superconformal
group ${OSp}^{\ast} (8 / 8)$.
The $D = 5$ gauge symmetries including those with parameters $\xi^{\mu},
\varepsilon^a$ satisfy a local algebra whose global limit is the $D = 5$
Poincar{\'e} algebra, but the $D = 6$ origin of this (at least in the free
theory) is an algebra including the generators $\mathcal{Q}^a_{\alpha M}$ of
the fermionic symmetries with parameter $\varepsilon_N^{\alpha}$ and the
generators $\mathcal{P}^{MNP}$ of the bosonic symmetries with parameter
$\chi^{MNP}$. The global algebra is of the form
\begin{equation}
\{\mathcal{Q}_{\alpha N}^a, \mathcal{Q}_{\beta P}^b \} = \hspace{0.17em}
\Omega^{ab} (\Pi_+ \Gamma^M C)_{\alpha \beta} \mathcal{P}_{(NP) M} \label{fialgsss}
\end{equation}
In the dimensional reduction, the $D = 5$ superalgebra has charges
$Q_{\alpha}^a =\mathcal{Q}_{\alpha 5}^a$, $P_{\mu} =\mathcal{P}_{55 \mu}$.
Supersymmetry provides a further argument against the possibility of a
background metric playing any role in an interacting (4,0) theory in $D = 6$.
The $D = 5$ supergravity can be formulated in an arbitrary supergravity
background, but these cannot be lifted to $D = 6$ (4,0) backgrounds involving
a background metric as there is no (4,0) multiplet including a metric or
graviton. The absence of a (4,0) supergravity multiplet makes problematic the
possibility of a background metric and the standard supersymmetry playing any
role in the $D = 6$ theory. Indeed, the interacting theory (if it exists)
should perhaps be a theory based on something like the algebra (\ref{fialgsss}) rather than
the super-Poincar{\'e} algebra (\ref{fialgss}).
\section{Double Copy}
It is intriguing that there is a sense in which gravity can be regarded as a
``square'' of Yang-Mills theory and supergravity can be regarded as a
``square'' of super-Yang-Mills theory. At the level of free theories, there is
a direct construction of supergravity as a square of SYM, with the
supergravity supermultiplet arising from the tensor product of two SYM
multiplets \cite{Borsten:2013bp,
Anastasiou:2013hba,Anastasiou:2014qba}. In the same way, the free (4,0) theory is a square of the free
(2,0) theory \cite{Chiodaroli:2011pp,Borsten:2017jpt}.
Remarkably, for the interacting theories 4D and 5D supergravity scattering
amplitudes can be constructed from ``squaring'' 4D and 5D SYM scattering
amplitudes in the double copy construction \cite{Bern:2010yg}. This suggests another approach to
seeking an interacting (4,0) theory. Given amplitudes for the (2,0) theory,
one could seek to apply the double copy construction to obtain (4,0)
amplitudes. After all, on circle compactification, the (2,0) amplitudes should
reduce to 5D SYM amplitudes and the (4,0) ones should reduce to 5D
supergravity ones, which should in turn be the double copy of the 5D SYM
amplitudes. This approach has been attempted in \cite{Cachazo:2018hqa,Geyer:2018xgb,Albonico:2020mge}.
A major problem with this approach is that we do not know the amplitudes for
the (2,0) theory even at tree level. In \cite{Cachazo:2018hqa,Geyer:2018xgb,Albonico:2020mge} formal expressions were derived
that had many of the properties that would be required of (2,0) amplitudes and
which reduced to 5D SYM amplitudes, but they had a number of problems. One was
that they could only be derived for even numbers of external states. Another
was that the poles had problematic non-local residues. As such, the
interpretation of these expressions is unclear. Nonetheless, the double copy
can be applied to these to give formal expressions for the (4,0) theory, but
again it is unclear as to how they might be interpreted and whether they can
be thought of as scattering amplitudes.
However, as the (2,0) theory exists then it should have well-defined
correlation functions and associated amplitudes, even though we do not yet
understand what these might be. Whatever they are, they should then give (4,0)
amplitudes. As these theories do not have conventional local covariant
interactions, it is to be expected that the corresponding amplitudes should
also involve non-standard features. It will be interesting to see whether
seeking amplitudes could be a useful way to unravel the mysteries of the
interacting theories.
\section{All Fo(u)r Nothing?}
There is a free (4,0) superconformal theory in 6D which provides a conformal
theory of an exotic ``graviton'' but which nevertheless reduces to linearised
5D N=8 supergravity on circle reduction, with a standard graviton. It then
gives a conformal theory of gravity with only two derivatives. Compactifying
on a 2-torus to 4D gives an ${SL} (2, \mathbb{Z})$ symmetry acting
through duality transformations, giving a gravitational duality for the free
theory \cite{Hull:2000zn,Hull:2000rr,Hull:2001iu}. The usual gauge symmetry of gravity is replaced by one with a
parameter which is a 3-tensor instead of a vector field.
The big question is whether there is an interacting version of this theory. It
is of course possible that there is no interacting theory, but it would be
disappointing if M-theory didn't avail itself of the existence of such an
interesting multiplet.
If an interacting theory exists, then it cannot be a conventional theory but
should be some ``non-lagrangian'' theory, presumably of a similar kind to the
(2,0) theory. If there is an interacting (4,0) theory, then the next question
is whether it arises in M-theory. It was argued in \cite{Hull:2000zn} that it should arise as
a strong coupling limit of M-theory toroidally compactified to 5D, understood
as a limit to energies much greater than the Planck scale.
A key question to ask here is whether M-theory compactified to 5D has BPS
states carrying the singlet central charge $K$ and, if so, what the spectrum of these is. The
conjecture that the (4,0) theory arises as a strong coupling limit requires
the 5D theory to have BPS states carrying $K$ with the right spectrum to be
Kaluza-Klein modes.
This then provides an important test for the conjecture that there could be a 6D (4,0) theory arising as a limit of M-theory:
if M-theory toroidally compactified to 5D does not have BPS states carrying $K$, then the conjecture is false.
On the other hand, the presence of $K$-charged states with a spectrum consistent with a KK tower would be strong evidence that
the conjecture could be true. Importantly, this is a test that can be examined within M-theory and would deepen our understanding of M-theory.
A promising approach to the (4,0) theory is to first understand better amplitudes and correlation
functions for the (2,0) theory and then to use these to construct (4,0)
amplitudes via the double copy mechanism. Such (4,0) tree amplitudes would
then define the classical theory.
There is a further exotic multiplet in 6D with (3,1) supersymmetry \cite{Hull:2000zn}. This
also gives 5D N=8 supergravity on circle reduction and involves a third rank
tensor gauge field with different symmetry properties to that arising in the
(4,0) theory, and the supermultiplet has both vector fields and self-dual
2-form gauge fields. This is not a superconformal supermultiplet and so
requires a dimensionful coupling constant in 6D. Again it is interesting to
ask whether an interacting theory exists and whether it has a role to play in
M-theory. While the (4,0) multiplet arises from the product of two (2,0) ones,
the (3,1) multiplet arises from the product of a (2,0) multiplet with a (1,1)
multiplet. Moreover, (3,1) amplitudes can be obtained from combining (2,0)
with (1,1) amplitudes and in \cite{Cachazo:2018hqa,Geyer:2018xgb,Albonico:2020mge} formal (3,1) amplitude-like expressions were
found using their (2,0) expressions for ``amplitudes". As in the (4,0) case, these have many
of the properties one might expect from an amplitude but have non-local poles
and other problematic features.
Whether or not the (4,0) theory turns out to have a role to play in M-theory,
it raises some interesting issues. Studying the multiplet led to the duality
between the graviton and dual graviton in linearised gravity \cite{Hull:2000zn,Hull:2000rr,Hull:2001iu} and to a
gravitational S-duality in 4D \cite{Hull:2000rr}. It raises the possibility that gravity could be
fundamentally conformal and that at high energies it could be described by
some field other than the usual symmetric 2-tensor graviton. More generally,
it is interesting to ask how M-theory behaves at trans-Planckian energies,
especially given the interesting structures found in string theory at
ultra-high energies or string-scale curvatures.
\section*{Acknowledgements}
{This work is supported by the STFC
Consolidated Grant ST/T000791/1 and a Royal Society Leverhulme Trust Senior Research
Fellowship.}
\vskip2pc
|
\section{Introduction}
In the era of big data, the explosion of distributed edge data exposes the weaknesses of centralized machine learning (ML), which requests edge devices to upload local data to a central server for training, leading to huge communication costs and high data leakage risks. Federated learning (FL) has recently arisen as a promising distributed learning paradigm to tackle these challenges by replacing data transmission with model/gradient exchange \cite{konevcny2016federated}. Specifically, in each FL training iteration, the PS first broadcasts the global model parameters to the devices, and the latter then calculates their local gradients separately based on the local datasets and uploads the local gradients to the PS. Subsequently, the PS aggregates the received local gradients and updates the global model.
However, due to the iterative nature, the huge communication cost is still a bottleneck for FL. A promising approach to reduce the communication cost of FL in the wireless edge is to introduce the over-the-air computation (AirComp) technique \cite{nazer2007computation} into FL uplink \cite{yang2020federated}. By utilizing the superposition property of wireless signals, AirComp allows the uplink cost to be unchanged with the growth of the device number (with fixed aggregation accuracy), thereby significantly reducing the communication cost of large FL systems. The corresponding FL paradigm is referred to as over-the-air FL (OA-FL).
Despite these advantages, the introduction of AirComp also brings some unique problems, such as the straggler issue \cite{zhu2019broadband, liu2021reconfigurable, zhong2021over}. Since AirComp requires local gradients to be aligned at the PS (i.e., the PS receives a linear combination of local gradients with desired coefficients), the devices with good channel conditions have to lower their transmitting powers to match the devices with relatively poor channel conditions (i.e., the stragglers) to ensure correct aggregation. Existing works like \cite{zhu2019broadband} and \cite{liu2021reconfigurable} discard the stragglers to relieve the straggler issue. However, such a coarse-grained discarding strategy will inevitably cast away some exploitable information, resulting in a learning performance loss. Ref. \cite{zhong2021over} suggests a soft aggregation approach to tackle the straggler issue without discarding devices, which, however, works well only in small service area cases (e.g., $100\times 100$ m$^2$). When it comes to a relatively large service area (e.g., $2\times 2$ km$^2$), it fails due to the large difference in path loss between nearby devices and those far from the PS (which results in the de facto discarding of remote devices).
Allowing PS mobility is a natural idea to perform OA-FL in a large service area since the PS can thus move close to the stragglers for better services. Therefore, in this paper, an unmanned aerial vehicle (UAV) is introduced as the PS. Due to the ground-to-air nature of UAV uplink communication, the device-UAV links are generally not blocked, potentially leading to good channel conditions \cite{lin2018sky,wu2018joint}. Besides, UAVs have extremely high mobility and can thus cover a large area quickly. There are already some works on UAV-assisted FL. For example, Ref. \cite{lim2021uav} employs a UAV as a mobile relay in the sky to assist FL; in \cite{donevski2021federated}, a UAV acts as an orchestrator, coordinating the transmissions and the learning schedule within a preset deadline.
However, these works consider orthogonal FL uplink and thus have various limitations in device number, service area, and communication condition, which means the state-of-art design is far from fully unleashing the potential of UAV-assisted FL.
\begin{figure}
[t]
\centering
\includegraphics[height=0.205\textheight]{modelscheme.eps}
\caption{UAV-assisted OA-FL in a large service area.}
\label{modelscheme}
\vspace{-1.0em}
\end{figure}
In this paper, we make the first attempt to employ the UAV to assist the OA-FL system by designing a hierarchical over-the-air aggregation scheme.
As shown in Fig. \ref{modelscheme}, we consider a UAV-assisted OA-FL system, where the UAV acts as the PS, denoted by UAV-PS, to serve devices distributed in a relatively large area. To avoid serious straggler issues, we restrict the UAV-PS to serve nearby devices, i.e., only aggregate the local gradients from nearby devices over the air. Thus the UAV-PS must fly across its large service area to serve more devices. After the UAV-PS completes one round, it further aggregates the received partially aggregated local gradients to obtain a noisy version of the desired global gradient for global model updating. We call this two-step aggregation hierarchical aggregation. Under this hierarchical gradient aggregation scheme, we carry out a gradient-correlation-aware FL convergence analysis, which leads us to formulate a mean squared error (MSE) minimization problem to tune the UAV trajectory and the global aggregation coefficients. An algorithm based on alternating optimization (AO) and successive convex approximation (SCA) is proposed to solve the formulated problem.
Simulation results demonstrate that our proposed scheme achieves robust performance improvement in complicated FL scenarios compared with existing solutions.
\section{System Model}
In this section, we detail the underlying models and the hierarchical aggregation scheme.
\subsection{Federated Learning Model}
We consider an FL system where $M$ devices collaboratively train a machine learning model assisted by a UAV-PS. Let $\mathbf{w} \in \mathbb{R}^D$ be the parameter vector of the machine learning model and $\mathcal{D}_m$ be the dataset on device $m$. Also let $Q_m \triangleq |\mathcal{D}_m|$ and $Q \triangleq \sum_{m=1}^M Q_m$. Then the global loss function can be written as
\begin{equation}
\label{LossFunc}
F(\mathbf{w}) = \sum_{m=1}^{M} b_m F_m \left( \mathbf{w}\right
\end{equation}
with the local loss function
\begin{equation}
\label{localLF}
F_m(\mathbf{w}) = \frac{1}{Q_m} \sum_{q=1}^{Q_m} f ( \mathbf{w}; \boldsymbol{\xi}_{m,q} )
\end{equation}
where $f ( \mathbf{w}; \boldsymbol{\xi}_{m,q} )$ denotes the sample-wise loss function based on the $q$-th data sample $\boldsymbol{\xi}_{m,q} \in \mathcal{D}_m$ and $b_m \triangleq Q_m/Q$.
FL requires, say, $T$ training rounds for convergence. The $t$-th training round consists of four steps:
\begin{itemize}
\item \textit{Global model broadcasting}: The UAV-PS broadcasts the current global model parameters, $\mathbf{w}^{(t)}$, to $M$ devices in the service area.
\item \textit{Local gradient computation}: Each device computes its local gradient by
\begin{gather}
\mathbf{g}_{m}^{(t)}=\nabla F_m(\mathbf{w}^{(t)}),
\end{gather}
where $\nabla F_m(\mathbf{w}^{(t)})$ denotes the gradient of the local loss function.
\item \textit{Hierarchical gradient aggregation}: The UAV-PS flies across the service area to hierarchically aggregate the local gradients to obtain $\mathbf{\hat{g}}^{(t)}$, which is a noisy version of the desired global gradient
\begin{gather}
\mathbf{g}^{(t)}\triangleq\sum_{m=1}^{M}b_m\mathbf{g}_{m}^{(t)}\label{idealtri}.
\end{gather}
\item \textit{Global model updating}: The UAV-PS updates the global model $\mathbf{w}^{(t)}$ by
\begin{gather}
\mathbf{w}^{(t+1)} = \mathbf{w}^{(t)} - \eta\mathbf{\hat{g}}^{(t)},
\end{gather}
where $\eta$ denotes the learning rate.
\end{itemize}
\subsection{UAV Channel Model}
\label{subsec_channel}
In this paper, we consider a three-dimensional Cartesian coordinate system. Assume that the $m$-th device is located on the ground with horizontal coordinate $\mathbf{v}_m=[x_m, y_m]^\top, m\in M$. Also assume that the UAV-PS flies at a fixed height $z$ above the ground with horizontal coordinate $\mathbf{u}(\tau)=[x(\tau),y(\tau)]^\top, 0\le\tau\le \Delta t$, at instant $\tau$, where $\Delta t$ is the duration required by the UAV-PS for flying one round. In each training round, the UAV trajectory satisfies the following constraints:
\begin{align}
\mathbf{u}(0)&=\mathbf{u}(\Delta t),\label{u0}\\
\lVert \mathbf{\dot{u}}(\tau) \rVert&\le V_{\mathrm{max}}, 0\le\tau\le \Delta t, \label{utv}
\end{align}
where $V_{\mathrm{max}}$ denotes the maximum speed of the UAV-PS. Note that (\ref{u0}) ensures that the UAV-PS starts and ends at the same point and that (\ref{utv}) imposes the practical maximal speed constraint. We discretize the duration $\Delta t$ into $N$ time slots indexed by $n$ with equal interval $\delta=\Delta t/N$. We assume that $\delta$ is sufficiently small such that the UAV-PS can be considered stationary in a time slot, even if it flies at the maximum speed $V_{\mathrm{max}}$. Therefore the trajectory constraints are rewritten as
\begin{align}
\mathbf{u}[0]&=\mathbf{u}[N],\label{un}\\
\lVert \mathbf{u}[n+1]-\mathbf{u}[n] \rVert ^2&\le (V_{\mathrm{max}}\delta)^2, n=0, 2, ..., N-1. \label{unv}
\end{align}
We assume the Doppler effect caused by the mobility of the UAV-PS can be well compensated at the receiver and perfect channel state information (CSI) is accessible at devices \cite{cao2020cooperative}. Further, due to the ground-to-air nature of UAV uplink communication, we consider the line-of-sight (LoS) channels for the device-UAV links. Thus the channel from the $m$-th device to the UAV-PS during the $n$-th time slot follows the free-space path loss model, expressed as
\begin{gather}
h_m[n]=\sqrt{\varrho d_m^{-2}[n]}\vartheta_m[n], \label{channel}
\end{gather}
where $\varrho$ denotes the channel power gain at the reference distance $d_{\mathrm{ref}}=1$ m, $\vartheta_m[n]\triangleq e^{j\theta_m[n]}$ with $\theta_m[n]$ denoting the phase of $h_m[n]$,
and $d_m[n]=\sqrt{z^2+\|\mathbf{u}[n]-\mathbf{v}_m\|^2}$ denotes the distance between the device $m$ and the UAV-PS at the $n$-th time slot.
\subsection{Transmitting Signal Model}
As mentioned before, the UAV-PS only serves nearby devices. We now define ``nearby'' precisely. At the $n$-th time slot, given a preset coverage distance threshold $d_{\mathrm{thr}}$, we define set $\mathcal{M}[n]\triangleq \{m \in [M] \ |\ \|\mathbf{u}[n]-\mathbf{v}_m\|\le d_{\mathrm{thr}}\}$. The devices with indexes in $\mathcal{M}[n]$ are referred to as the nearby devices.
At the $n$-th time slot, only the nearby devices need to transmit signals. Assume that device $m$ is a nearby device. We now introduce how device $m$ generates its transmitting signal.
It first normalizes the local gradient $\mathbf{g}_m^{(t)}$ for efficient transmission.
Specifically, each entry of the normalized gradient $\tilde{\mathbf{g}}_m^{(t)}$ is computed by
\begin{gather}
\tilde{g}_m^{(t)}(d)=\frac{g_m^{(t)}(d)-\bar{g}_m^{(t)}}{\sqrt{\upsilon_m^{(t)}}}, d\in [D],
\label{normar}
\end{gather}
where $g_m^{(t)}(d)$ denotes the $d$-th entry of $\mathbf{g}_m^{(t)}$, $\bar{g}_m^{(t)}$ and $\upsilon_m^{(t)}$ are computed by $\bar{g}_m^{(t)}=\sum_{d=1}^D g_m^{(t)}(d)/D$ and $\upsilon_m^{(t)}=\sum_{d=1}^D(g_m^{(t)}(d)-\bar{g}_m^{(t)})^2/D$ respectively. Since the cost of transmitting the scalars $\{\bar{g}_m^{(t)}, \upsilon_m^{(t)}\}$ is negligible compared to that of transmitting the gradients, we assume that these scalars are sent to the UAV-PS losslessly following \cite{lin2021deploying} and \cite{yang2022federated}.
The normalized gradient $\tilde{\mathbf{g}}_m^{(t)}$ is then modulated as a complex vector $\mathbf{r}_{m}^{(t)}$ by
\begin{gather}
\mathbf{r}_{m}^{(t)} \triangleq
\tilde{\mathbf{g}}_{m}^{(t)}\left (1:C \right ) +
j \tilde{\mathbf{g}}_{m}^{(t)}\left ((C+1) :2C\right ) \in \mathbb{C}^{C }, \label{pin}
\end{gather}
where $C=D/2$ and $\tilde{\mathbf{g}}_{m}^{(t)}\left (c_1:c_2 \right)$ denotes the sub-vector of $\tilde{\mathbf{g}}_{m}^{(t)}$ containing the entries with indexes from $c_1$ to $c_2$.\footnote{We assume an even $D$ for simplicity.} Then the transmitting signal is given by
\begin{gather}
\mathbf{x}_m^{(t)}[n]\triangleq \beta_m[n]\mathbf{r}_{m}^{(t)}, \label{fa}
\end{gather}
where $\beta_m[n]\in\mathbb{C}$ is the transmitting coefficient satisfying the power constraints $|\beta_m[n]|^2 = P_0$.\footnote{We assume that all the devices transmit with full power, which is reasonable since the mobility of the UAV-PS allows the signals to be approximately aligned with desired coefficients (by designing an appropriate trajectory).}
\subsection{Hierarchical Aggregation}
\label{Hierarchical_Over-the-Air_Model_Aggregation}
In our hierarchical aggregation scheme for the UAV-assisted OA-FL, the local gradients are aggregated in two steps, called \emph{over-the-air partial aggregation} and \emph{global aggregation}. In this subsection, we separately introduce the two steps.
\subsubsection{Over-the-air partial aggregation}
In this step, the UAV-PS flies around to receive the partially aggregated local gradients for $N$ times. Specifically, the received signal at the $n$-th time slot is given by
\begin{align}
\mathbf{y}^{(t)}[n]=\sum_{m\in \mathcal{M}[n]} h_m[n]\mathbf{x}_m^{(t)}[n]+\mathbf{n}^{(t)}[n], \label{yhx}
\end{align}
where $\mathbf{n}^{(t)}[n]\in \mathbb{C}^{C}$ is the additive white Gaussian noise (AWGN) with elements independently drawn from $\mathcal{CN}(0,\sigma^2)$.
For convenience, we introduce a binary indicator variable $\alpha_m[n] \triangleq \mathbbm{1}_{\|\mathbf{u}[n]-\mathbf{v}_m\|\le d_{\mathrm{thr}}}(\mathbf{u}[n])$, where the indicator function $\mathbbm{1}_{\|\mathbf{u}[n]-\mathbf{v}_m\|\le d_{\mathrm{thr}}}(\cdot)$ takes $1$ when $\mathbf{u}[n]$ satisfies $\|\mathbf{u}[n]-\mathbf{v}_m\|\le d_{\mathrm{thr}}$. Then the received signal can be rewritten as
\begin{align}
\label{received_signal}
\mathbf{y}^{(t)}[n]=\sum_{m=1}^{M} \alpha_m[n]h_m[n]\mathbf{x}_m^{(t)}[n]+\mathbf{n}^{(t)}[n].
\end{align}
\vspace{-0.1cm}
By substituting \eqref{channel} and \eqref{fa} into \eqref{received_signal}, we have
\begin{align}
\mathbf{y}^{(t)}[n]=\sum_{m=1}^ {M} \frac{\sqrt{\varrho} \alpha_m[n] \beta_{m}[n] \vartheta_m[n] }{\sqrt{z^2+\|\mathbf{u}[n]-\mathbf{v}_m\|^2}} \mathbf{r}_{m}^{(t)}[n] + \mathbf{n}^{(t)}[n].
\label{rhoppr}
\end{align}
Further, we set $\beta_m[n]=\overline{\vartheta_m[n]}\sqrt{P_0}$ to compensate for the phase shift of the channel. Thus the partially aggregated signal can finally be rewritten as
\vspace{-0.1cm}
\begin{gather}
\mathbf{y}^{(t)}[n]=\sum_{m=1}^ {M} \frac{\sqrt{\varrho P_0} \alpha_m[n]} {\sqrt{z^2+\|\mathbf{u}[n]-\mathbf{v}_m\|^2}} \mathbf{r}_{m}^{(t)}[n] + \mathbf{n}^{(t)}[n]. \label{yfinal}
\end{gather}
\subsubsection{Global aggregation}
After the UAV-PS completes one round, it aggregates $\{\mathbf{y}^{(t)}[n]\}_{n=1}^N$ using coefficients $\{\zeta[n]\in\mathbb{R}\}_{n=1}^N$ to obtain globally aggregated signal $\mathbf{a}^{(t)}$, i.e.,
\begin{gather}
\mathbf{a}^{(t)}=\sum_{n=1}^ {N}\zeta[n]\mathbf{y}^{(t)}[n]. \label{ay}
\end{gather}
\vspace{-0.1cm}
Then $\mathbf{\hat{g}}^{(t)}$ is computed by
\vspace{-0.08cm}
\begin{gather}
\mathbf{\hat{g}}^{(t)} = \left[ \Re\{\mathbf{a}^{(t)}\}^\top , \, \Im\{\mathbf{a}^{(t)}\}^\top \right]^\top + \bar{g}^{(t)}\mathbf{1}_{D }, \label{chai}
\end{gather}
where $ \bar{g}^{(t)}\triangleq\sum_{m=1}^M b_m\bar{g}_m^{(t)}$.
\section{Performance Analysis}
For the convenience of convergence analysis, we first introduce two standard assumptions \cite{friedlander2012hybrid, liu2021reconfigurable}:
\begin{assumption}
\label{assum1}
The global loss function $F$ is strongly convex with positive parameter $\mu$: $F(\mathbf{w})
\geq F(\mathbf{w^\prime}) + (\mathbf{w} - \mathbf{w^\prime})^{\top} \nabla F(\mathbf{w^\prime}) + \frac{\mu}{2} \| \mathbf{w} - \mathbf{w^\prime}\|^2, \forall \mathbf{w}, \mathbf{w^\prime} \in \mathbb{R}^D.$
\end{assumption}
\begin{assumption}
\label{assum2}
The
gradient $\nabla F(\cdot)$ is uniformly Lipschitz continuous with parameter $\omega$, i.e., $\| \nabla F(\mathbf{w}) - \nabla F(\mathbf{w^\prime})\| \leq \omega \|\mathbf{w} - \mathbf{w^\prime} \|, \forall \mathbf{w}, \mathbf{w^\prime} \in \mathbb{R}^D$.
\end{assumption}
Following \cite{zhong2021over}, we further assume the following gradient correlation model:
\begin{assumption}
\label{assum4}
(Gradient correlation model) Define $\tilde{\mathbf{G}}^{(t)} \triangleq [\tilde{\mathbf{g}}_{1}^{(t)}, \cdots, \tilde{\mathbf{g}}_{M} ^{(t)}] \in \mathbb{R}^{D \times M}$. Let $\mathbf{z}_{d}^{(t)\top} \in \mathbb{R}^{M}$ denote the $d$-th row of $\tilde{\mathbf{G}}^{(t)}$, $\forall d \in [D]$. We assume $\{\mathbf{z}_{d}^{(t)}\}_{d=1}^D$ to be independent and identically distributed (i.i.d.) with $\mathbb{E}[\mathbf{z}_{d}^{(t)}] = \mathbf{0}$ and $\mathbb{E}[\mathbf{z}_{d}^{(t)}(\mathbf{z}_{d}^{(t)})^\top] = \boldsymbol{\rho}^{(t)}$, $\forall d \in [D]$.
\end{assumption}
An upper bound of the convergence performance is given in the following theorem.
\vspace{-0.15cm}
\begin{theorem}
\label{Theorem:Convergence}
Set the learning rate $\eta=1/\omega$. Under Assumptions \ref{assum1} and \ref{assum2}, we have:
\vspace{-0.15cm}
\begin{gather}
\label{aaa}
\mathbb{E}\!\left[\!\mathcal{F}\!\left(\!\mathbf{w}^{(T+1)}\!\right)\!- \!\mathcal{F}\left(\mathbf{w}^*\right)\right]
\leq \sum_{t=0}^{T}\left(1-\frac{\mu}{\omega}\right)^{T-t}\mathbb{E}[\|\mathbf{e}^{(t)}\|^2] \notag\\
+\left(\mathcal{F}\left(\mathbf{w}^{(0)}\right) - \mathcal{F}\left(\mathbf{w}^*\right)\right) \left(1-\frac{\mu}{\omega}\right)^{T+1},
\end{gather}
where $\mathbf{e}^{(t)}\triangleq \mathbf{g}^{(t)} - \hat{\mathbf{g}}^{(t)}$ and the expectations are taken with respect to the gradients and the AWGN.
\end{theorem}
\begin{proof}
Take expectation of both sides of the recursion in [\citenum{friedlander2012hybrid}, Lemma~2.1] and iterate it for $t+1$ times.
\end{proof}
Clearly, by Theorem \ref{Theorem:Convergence}, to improve the learning performance, we only have to minimize the mean squared error (MSE) $\mathbb{E}[\|\mathbf{e}^{(t)}\|^2]$, which can be written as
\begin{align}
&\mathbb{E}[\|\mathbf{e}^{(t)}\|^2]
\notag\\
=&\mathbb{E}\left[\left\|\mathbf{g}^{(t)} - \hat{\mathbf{g}}^{(t)}\right\|^2\right]
\notag \\
\overset{\text{(a)}}{=}& \mathbb{E}\left[\left\|\sum_{m=1}^{M}b_m g_m^{(t)} - \sum_{m=1}^{M}b_m\bar{g}^{(t)}\mathbf{1}_{D} - \begin{bmatrix}
\Re\{\mathbf{a}^{(t)}\}\\
\Im\{\mathbf{a}^{(t)}\}
\end{bmatrix} \right\|^2\right]
\notag \\
\overset{\text{(b)}}{=}&\mathbb{E}\left[\left\|\sum_{m=1}^{M}\left( b_m\sqrt{\upsilon^{(t)}_m} - \sum_{n=1}^{N} \frac{\zeta[n]\sqrt{\varrho P_0} \alpha_m[n]} {\sqrt{z^2+\|\mathbf{u}[n]-\mathbf{v}_m\|^2}} \right) \tilde{\mathbf{g}}_m^{(t)}
\right.\right. \notag \\
&\left.\left. - \sum_{n=1}^{N}\zeta[n]\mathbf{n}_{r}^{(t)}[n] \right\|^2\right],
\label{Ee}
\end{align}
where $\mathbf{n}_{r}^{(t)}[n] \triangleq [\Re\{\mathbf{n}^{(t)}[n]\}^\top, \Im\{\mathbf{n}^{(t)}[n]\}^\top]^\top$, step ($a$) follows from \eqref{idealtri} and \eqref{chai}, and step ($b$) follows from \eqref{normar}, \eqref{pin}, \eqref{yfinal} and \eqref{ay}.
To proceed, define $\boldsymbol{\zeta}\triangleq\left[\zeta[1], ..., \zeta[N]\right]^\top$, $\boldsymbol{\upsilon}^{(t)}\triangleq\Big[b_{1}\sqrt{\upsilon^{(t)}_{1}} , ..., b_{M}\sqrt{\upsilon^{(t)}_{M}}\Big]^\top$, and $\mathbf{K}\in \mathbb{R}^{M\times N}$ with $K_{m,n}=\dfrac{\alpha_m[n]\sqrt{\varrho P_0}}{\sqrt{z^2+\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2}}$. Recall $\tilde{\mathbf{G}}^{(t)} = [\tilde{\mathbf{g}}_{1}^{(t)}, \cdots, \tilde{\mathbf{g}}_{M} ^{(t)}]$. Then \eqref{Ee} can be rewritten as
\begin{align}
&\mathbb{E}[\|\mathbf{e}^{(t)}\|^2]
\notag\\
=&\mathbb{E}\left[\left\|\tilde{\mathbf{G}}^{(t)} \left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right) - \sum_{n=1}^{N}\zeta[n]\mathbf{n}_{r}^{(t)}[n]\right\|^2\right]
\notag \\
\overset{\text{(a)}}{=}& D\left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right)^\top \boldsymbol{\rho}^{(t)} \left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right) +\frac{D\sigma^2}{2}\boldsymbol{\zeta}^\top\boldsymbol{\zeta},
\label{ana2}
\end{align}
where step ($a$) follows from Assumption \ref{assum4} and the fact that the AWGN is independent of the gradients.
%
\section{System optimization}
\subsection{Problem Formulation}
Define $\mathcal{U}=\{\mathbf{u}[n]\}_{n=1}^N$. The MSE minimization problem can be formulated as
\begin{subequations} \label{P0}
\begin{align}
\min_{\boldsymbol{\zeta}, \mathcal{U}} \quad &
\frac{\sigma^2}{2}\boldsymbol{\zeta}^\top\boldsymbol{\zeta}
+\left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right)^\top \boldsymbol{\rho}^{(t)} \left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right) \\
\operatorname{ s.t. } \quad
&\eqref{un}, \ \eqref{unv}.
\end{align}
\end{subequations}
To properly solve problem \eqref{P0}, we first introduce $\mathbf{K}$ as an optimization variable, and then slack the consequent equality constraints. By introducing $\mathbf{K}$, problem \eqref{P0} can be reformulated as
\begin{subequations} \label{P01}
\begin{align}
\min_{\boldsymbol{\zeta}, \mathbf{K}, \mathcal{U}} \quad &
\frac{\sigma^2}{2}\boldsymbol{\zeta}^\top\boldsymbol{\zeta}
+\left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right)^\top \boldsymbol{\rho}^{(t)} \left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right)
\label{p00} \\
\operatorname{ s.t. } \quad
&K_{m,n}\!=\!\dfrac{\alpha_m[n]\sqrt{\varrho P_0}}{\sqrt{z^2+\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2}}, \forall m \in [M], n\in [N],
\label{constraint_01}\\
&\eqref{un}, \ \eqref{unv}.
\end{align}
\end{subequations}
Note that the optimal $\boldsymbol{\zeta}^*$ and $\mathcal{U}^*$ for \eqref{P01} are also optimal for \eqref{P0}. We then relax constraints \eqref{constraint_01} by slacking $\alpha_m[n]$. Recall $\alpha_m[n]=\mathbbm{1}_{\|\mathbf{u}[n]-\mathbf{v}_m\|\le d_{\mathrm{thr}}}(\mathbf{u}[n])$, which is a binary variable. Thus a natural way to slack $\alpha_m[n]$ is to replace the constraint $\alpha_m[n]=\mathbbm{1}_{\|\mathbf{u}[n]-\mathbf{v}_m\|\le d_{\mathrm{thr}}}(\mathbf{u}[n])$ by
\begin{gather}
0\le\alpha_m[n]\le\frac{d_{\mathrm{thr}}^2}{d_{\mathrm{thr}}^2+\|\mathbf{u}[n]-\mathbf{v}_m\|^2}.
\end{gather}
Constraints \eqref{constraint_01} thus can be relaxed as
\begin{gather}\label{K_slacking}
0\!\le\! K_{m,n}\!\le\!\frac{d_{\mathrm{thr}}^2\sqrt{\varrho P_0}} {(d_{\mathrm{thr}}^2+\|\mathbf{u}[n]-\mathbf{v}_m\|^2)(z^2+\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2)^{\tfrac{1}{2}}},
\end{gather}
resulting in the following optimization problem:
\begin{subequations} \label{P1}
\begin{align}
\min_{\boldsymbol{\zeta}, \mathbf{K}, \mathcal{U}} \quad & \!\!\!
\frac{\sigma^2}{2}\boldsymbol{\zeta}^\top\boldsymbol{\zeta}
+\left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right)^\top\!\!\! \boldsymbol{\rho}^{(t)}\!\! \left(\boldsymbol{\upsilon}^{(t)}- \mathbf{K} \boldsymbol{\zeta}\right)
\label{std0}\\
\operatorname{ s.t. } \quad
&\eqref{un},\ \eqref{unv},\ \eqref{K_slacking}.
\end{align}
\end{subequations}
The objective of problem \eqref{P1} is a multi-convex function of $\boldsymbol{\zeta}$ and $\mathbf{K}$, which inspires us to apply the alternating optimization (AO) technique.
\subsection{Optimization of $\{\mathbf{K}, \mathcal{U}\}$ with a fixed $\boldsymbol{\zeta}$}
In this subsection, we optimize $\mathbf{K}$ and $\mathcal{U}$ with a fixed $\boldsymbol{\zeta}=\tilde{\boldsymbol{\zeta}}$. The corresponding optimization problem is
\begin{subequations} \label{P21}
\begin{align}
\min_{\mathbf{K}, \mathcal{U}} \quad & \tilde{\boldsymbol{\zeta}}^\top\mathbf{K}^\top \boldsymbol{\rho}^{(t)} \mathbf{K}\tilde{\boldsymbol{\zeta}} - 2 \boldsymbol{\upsilon}^{(t)\top} \boldsymbol{\rho}^{(t)} \mathbf{K}\tilde{\boldsymbol{\zeta}} \\
\operatorname{ s.t. } \quad
&\eqref{un}, \ \eqref{unv}, \ \eqref{K_slacking}.
\end{align}
\end{subequations}
Note that constraint \eqref{K_slacking} is non-convex for $\mathbf{u}[n]$. However, the right side of \eqref{K_slacking} is convex with respect to the whole expression $\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2$. Thus we can obtain a lower bound of the right side of \eqref{K_slacking} by expanding it as its first-order Taylor polynomial with respect to $\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2$ \cite{boyd2004convex} based on the idea of successive convex approximation (SCA), i.e.,
\begin{align}
\label{ss}
&\frac{d_{\mathrm{thr}}^2\sqrt{\varrho P_0}} {(d_{\mathrm{thr}}^2+\|\mathbf{u}[n]-\mathbf{v}_m\|^2)(z^2+\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2)^{\tfrac{1}{2}}}
\ge
\notag \\
&\ \ \ \ \ \Psi^{(t)} + \Psi'^{(t)}
\left(\|\mathbf{u}[n]-\mathbf{v}_m\|^2-\|\tilde{\mathbf{u}}[n]-\mathbf{v}_m\|^2\right),
\end{align}
where
\begin{align*}
\Psi^{(t)}=\frac{d_{\mathrm{thr}}^2 \sqrt{\varrho P_0}} {(d_{\mathrm{thr}}^2+\|\tilde{\mathbf{u}}[n]-\mathbf{v}_m\|^2)(z^2+\|\tilde{\mathbf{u}}[n]-\mathbf{v}_{m}\|^2)^{\tfrac{1}{2}}}, \\
\Psi'^{(t)}=\frac{-d_{\mathrm{thr}}^2 \left(1+\frac{1}{2}d_{\mathrm{thr}}^2 + \frac{3}{2} \|\tilde{\mathbf{u}}[n]-\mathbf{v}_{m}\|\right) \sqrt{\varrho P_0} } {(d_{\mathrm{thr}}^2+\|\tilde{\mathbf{u}}[n]-\mathbf{v}_m\|^2)^2(z^2+\|\tilde{\mathbf{u}}[n]-\mathbf{v}_{m}\|^2)^{\tfrac{3}{2}}},
\end{align*}
and $\|\tilde{\mathbf{u}}[n]-\mathbf{v}_{m}\|^2$ denotes the expand point (i.e., expanding at point $\|\tilde{\mathbf{u}}[n]-\mathbf{v}_{m}\|^2$). The equality in \eqref{ss} holds when $\mathbf{u}[n] = \tilde{\mathbf{u}}[n]$. Then \eqref{K_slacking} can be transformed into
\begin{gather}
0\le K_{m,n}\le\Psi^{(t)} + \Psi'^{(t)}\! \left(\|\mathbf{u}[n]-\mathbf{v}_m\|^2-\|\tilde{\mathbf{u}}[n]-\mathbf{v}_m\|^2\right).
\label{kfinal}
\end{gather}
Since $\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2$ is a convex function of $\mathbf{u}[n]$ and the right side is a linear function of $\|\mathbf{u}[n]-\mathbf{v}_{m}\|^2$, constraint \eqref{kfinal} is convex. Finally, problem \eqref{P21} can be reformulated into
\begin{subequations} \label{P2}
\begin{align}
\min_{\mathbf{K}, \mathcal{U}} \quad & \tilde{\boldsymbol{\zeta}}^\top\mathbf{K}^\top \boldsymbol{\rho}^{(t)} \mathbf{K}\tilde{\boldsymbol{\zeta}} - 2 \boldsymbol{\upsilon}^{(t)\top} \boldsymbol{\rho}^{(t)} \mathbf{K}\tilde{\boldsymbol{\zeta}} \\
\operatorname{ s.t. } \quad
&\eqref{un}, \ \eqref{unv}, \ \eqref{kfinal},
\end{align}
\end{subequations}
which is convex and can be efficiently solved by standard convex optimization solvers such as CVX.
\subsection{Optimization of $\boldsymbol{\zeta}$ with fixed $\mathbf{K}$ and $\mathcal{U}$}
\label{optizeta}
Since \eqref{std0} is convex with respect to $\boldsymbol{\zeta}$ with fixed $\mathbf{K} = \tilde{\mathbf{K}}$ and $\mathcal{U} = \tilde{\mathcal{U}}$, by letting $\dfrac{\partial\mathbb{E}[\|\mathbf{e}^{(t)}\|^2]}{\partial\boldsymbol{\zeta}}=0$, the optimal $\boldsymbol{\zeta}^*$ is given by
\begin{gather}
\boldsymbol{\zeta}^*= \left(\frac{\sigma^2}{2}\mathbf{I}_{N\times N} + \tilde{\mathbf{K}}^\top \boldsymbol{\rho}^{(t)} \tilde{\mathbf{K}} \right)^{-1} \tilde{\mathbf{K}}^\top \boldsymbol{\rho}^{(t)} \boldsymbol{\upsilon}^{(t)}. \label{zetaopt}
\end{gather}
\subsection{Overall Algorithm}
By alternatively optimizing $\{\mathbf{K}, \mathcal{U}\}$ and $\boldsymbol{\zeta}$, problem \eqref{P1} can be solved, and thus problem \eqref{P0}. We summarize the proposed algorithm in Algorithm \ref{algorithm}. This algorithm converges since the objective \eqref{std0} is monotonically non-decreasing in the iterative process.
\begin{algorithm} [t!]
\caption{Algorithm for Solving Problem \eqref{P0}}\label{algorithm}
\begin{algorithmic}[1]
\State Input: $\boldsymbol{\rho}^{(t)}$ and $\boldsymbol{\upsilon}^{(t)}$.
\State Initialize $\{\mathbf{u}^0[n]\}_{n=1}^N$ and $\boldsymbol{\zeta}^0$ to feasible values, and set $i=0$.
\Repeat
\State Set $\tilde{\boldsymbol{\zeta}} = \boldsymbol{\zeta}^i$ and $\tilde{\mathbf{u}}[n] = \mathbf{u}^i[n]$, $\forall n \in [N]$, then solve problem \eqref{P2} to obtain $\{\mathbf{u}^{i+1}[n]\}_{n=1}^N$ and $\mathbf{K}^{i+1}$.
\State Set $\tilde{\mathbf{K}} = \mathbf{K}^{i+1}$, then $\boldsymbol{\zeta}^{i+1}= (\frac{\sigma^2}{2}\mathbf{I}_{N\times N} + \tilde{\mathbf{K}}^\top \boldsymbol{\rho}^{(t)} \tilde{\mathbf{K}} )^{-1} \tilde{\mathbf{K}}^\top \boldsymbol{\rho}^{(t)} \boldsymbol{\upsilon}^{(t)}$.
\State Update $i \leftarrow i+1 $.
\Until The fractional decrease of the objective \eqref{std0} is below a threshold $\epsilon > 0$.
\State Set $\boldsymbol{\zeta}^* = \boldsymbol{\zeta}^{i}$ and $\mathbf{u}^*[n]=\mathbf{u}^{i}[n]$, $\forall n \in [N]$.
\end{algorithmic}
\end{algorithm}
\section{Simulation Results}
In this section, we validate the effectiveness of the proposed UAV-PS assisted OA-FL scheme. We consider an FL system with $M=20$ devices on the ground in a large service area of $2\times2 $ km$^2$. The devices are randomly distributed in groups in order to simulate the users in clusters as in the village. The UAV-PS flies in an assumed fixed height $z=50$ m above the area and has a fixed starting and ending horizontal location $\mathbf{u}[0]=\mathbf{u}[N]=[885,-10]^\top$. The channel power gain, the noise power, and the maximum transmitting power are set as $\varrho=-60$ dB, $\sigma^2=-90$ dBm, and $P_0=0.32$ W. The maximum speed and the flying time interval of the UAV-PS are assumed as $V_{\mathrm{max}}=50$ m/s and $\delta=1$ s. The coverage distance threshold is set as $d_{\mathrm{thr}}=158$ m and after optimization the reconstruction to the binary $\alpha_m[n]$ follows the rounded principle based on the optimized results. And we set the optimization threshold $\epsilon=10^{-4}$.
The learning task is conducted over the MNIST dataset
with $Q=60000$ training samples. We train a convolutional neural network (CNN) with two $5\times5$ convolution layers in which the first has $16$ channels, the second has $32$ channels, and both have a $2\times2$ max pooling, and with a fully connected layer of $50$ neurons and ReLu activation, and with a softmax output layer (total parameters $D=39408$). The loss function is the cross-entropy loss. The learning rate is set as $\eta=0.05$ with momentum $=0.5$ and the local updates include $5$ mini-batches of stochastic gradient descent (SGD). We conduct the experiments over $10$ Monte Carlo trials and $20$ Monte Carlo trials for i.i.d. datasets and non-i.i.d. datasets respectively. For the i.i.d. data condition, data samples are assigned evenly to all devices; for the non-i.i.d. data condition, each device randomly selects $5$ classes with $Q/5M$ samples for each selected class. The correlation matrices of the gradients are approximated by $\boldsymbol{\rho}^{(t)} = \frac{1}{D} \sum_{d=1}^D \mathbf{z}_{d}^{(t)} \mathbf{z}_{d}^{(t)}{}^\top$.
We first present the PS location/UAV-PS trajectory optimization results in the following simulation conditions: (1) Static PS without UAV assisted, which is located on the coordinate barycenter of all devices; (2) UAV-PS without optimization in a circular trajectory with the flying period $\Delta t=120$ s; (3) UAV-PS with optimization by Algorithm \ref{algorithm} and the flying period $\Delta t=120$ s; and (4) UAV-PS with optimization by Algorithm \ref{algorithm} and the flying period $\Delta t=80$ s. All the above simulations are on the i.i.d. datasets.
\begin{figure}
[!ht]
\centering
\vspace{-1.0em}
\includegraphics[scale=0.29]{traj6.eps}
\caption{PS location/UAV-PS trajectory with $20$ devices.}
\label{traj}
\vspace{-0.2em}
\end{figure}
In Fig. \ref{traj}, by appropriate trajectory optimization, the UAV-PS with sufficient time, like $\Delta t=120$ s, has no need to fly right above each device but can serve for all on account of the AirComp technique to aggregate in clusters, which definitely verifies the effectiveness of the UAV assisted hierarchical over-the-air aggregation. However, when the flying time is limited, for instance, the UAV-PS with $\Delta t=80$ s, it fails to fly across the whole service area and thus only be optimized to fly above nearby clusters.
\begin{figure}
[!ht]
\centering
\vspace{-1.0em}
\includegraphics[scale=0.317]{accuiidreal.eps}
\caption{OA-FL test accuracy versus training rounds, i.i.d. data.}
\label{accuiid}
\vspace{-0.2em}
\end{figure}
\begin{figure}
[!ht]
\centering
\vspace{-1.0em}
\includegraphics[scale=0.342]{accunonstxiu.eps}
\caption{OA-FL test accuracy versus training rounds, non-i.i.d. data.}
\label{accunon}
\vspace{-0.2em}
\end{figure}
Figs. \ref{accuiid} and \ref{accunon} show the test accuracy of the UAV-assisted OA-FL system with i.i.d. data and non-i.i.d. data respectively in the following schemes: (1) Error-free bound with the PS aggregating free of error; (2) Static PS without UAV assisted, as the green location in Fig. \ref{traj}, whose limitation of coverage area is waived for serving all devices, and which conducts the aggregation design as \cite{zhong2021over} to relieve the straggler issue; (3) UAV-PS in a circular trajectory, as the blue line in Fig. \ref{traj}, $\Delta t=120$ s, which conducts the aggregation design optimization following \ref{optizeta}, and the empirical optimization for the radius and the center of the circle; and (4) UAV-PS with joint optimization for the trajectory and aggregation design by Algorithm \ref{algorithm}, as the orange line in Fig. \ref{traj}, $\Delta t=120$ s.
For the static PS though located on the barycenter, large communication distances for remote devices as the stragglers severely limit the quality of gradient aggregation, thus leading to bad learning results. The UAV-PS with trajectory optimization by Algorithm \ref{algorithm} is closer to the error-free baseline when compared to the one with a circular trajectory. This indicates that the latter, without the trajectory optimization for the UAV-PS, still constrains the training effect.
And in Fig. \ref{accuiid} and \ref{accunon}, our scheme performs far better than any others on both i.i.d. and non-i.i.d. datasets, which demonstrates the effectiveness of the scheme. Further, we discover that on non-i.i.d. datasets, the learning accuracy is of higher dependence on communication and aggregation precision. Therefore our scheme is proven to have great potential on complex datasets and in complex learning scenarios.
\section{Conclusion}
In this paper, we proposed a UAV-PS assisted hierarchical aggregation for over-the-air FL in large service area scenarios and analyzed the convergence performance of the proposed system with gradient correlation in consideration. We formulated an MSE optimization problem to jointly optimize the UAV trajectory and the aggregation coefficients. An algorithm based on AO and SCA was proposed to solve the problem. Numerical results demonstrated the effectiveness of our scheme.
\bibliographystyle{IEEEtran
|
\section{Introduction}
An approach to understanding closed universes with big-bang/big-crunch cosmologies holographically was proposed in \cite{Cooper:2018cmb} (and further developed in \cite{Antonini:2019qkt,VanRaamsdonk:2020tlr,Sully:2020pza,VanRaamsdonk:2021qgv,Antonini:2022blk,Antonini:2022xzo}). The idea is to consider an asymptotically AdS$_{d+1}$ black hole spacetime with a $d$-dimensional dynamical end of the world (ETW) brane behind the horizon providing an inner boundary of the spacetime, as depicted in figure \ref{fig1}. Starting from the $t=0$ surface, the ETW brane falls into the black hole and terminates at the singularity, so its worldvolume geometry is a big-bang/big-crunch cosmology. The state in the bulk on the $t=0$ surface is dual to some state in the dual $d$-dimensional CFT on the asymptotic boundary on the right in figure \ref{fig1}, which therefore includes a description of the cosmology on the ETW brane worldvolume.
An appealing feature of this model is that the state in the bulk on the $t=0$ surface can be constructed from a Euclidean path integral. In the Euclidean section, the ETW brane moves outward away from $t=0$, and eventually meets the asymptotic boundary. Such solutions were proposed in \cite{Takayanagi:2011zk,Fujita:2011fp} as duals of boundary conformal field theories (BCFTs). The state in the $d$-dimensional CFT dual to the $t=0$ slice in the bulk is then constructed by starting with a $(d-1)$-dimensional boundary state specified by the BCFT and evolving through some period of Euclidean time. Such states have been extensively discussed in recent investigations of black holes, see e.g. \cite{Kourkoulou:2017zaj,Almheiri:2019hni,Penington:2019kki,Chen:2020tes}.
\begin{figure}[ht]
\centering
\includegraphics[width=.3\linewidth]{ETWbrane.pdf}
\caption{Penrose diagram of an AdS black hole with the left asymptotic region terminating in an ETW brane (shown in red). The worldvolume geometry of the ETW brane is a big-bang/big-crunch cosmology.}
\label{fig1}
\end{figure}
To have a controlled description of the cosmology on the ETW brane, we want to have a separation of scales between the scale that controls the curvature of the ETW brane and the bulk curvature, such that there is a good effective description of the dynamics in terms of ordinary Einstein gravity localised on the ETW brane. This can be achieved by taking the radial position $r_0$ of the brane at $t=0$ to be much larger than the horizon scale $r_H$, which can be achieved by increasing the tension $T$ of the ETW brane \cite{Randall:1999ee,Randall:1999vf,Karch:2000ct}.
However, in simple examples of this construction this separation of scales is incompatible with the path integral construction of the state. If we are in $d>2$, and we take the bulk solution to be an uncharged black hole with flat spatial slices, when we increase $T$ for fixed bulk black hole geometry, there is a critical value $T=T_*$ at which the ETW brane intersects the asymptotic boundary at $t=0$. Increasing $T$ beyond $T_*$, the ETW brane will self-intersect before it reaches the asymptotic boundary. Thus, for $T> T_*$, we lose the Euclidean path integral construction of the state dual to the $t=0$ slice.\footnote{The $d=2$ case, where the bulk black hole is a BTZ solution \cite{Banados:1992wn}, is special; there the ETW brane always intersects the boundary at a quarter the period of the Euclidean solution.} One resolution of this problem, proposed in \cite{Antonini:2019qkt}, is to take the bulk solution to be a charged black hole. In \cite{VanRaamsdonk:2021qgv}, it was suggested that the problem could be avoided by introducing an additional interface brane, but in \cite{Fallows:2022ioc} (see also \cite{Waddell:2022fbn}) we found that the self-intersection problem persists in the presence of the interface brane.
In this paper, we will see that another way to avoid this problem is to consider hyperbolic spatial slices, with constant negative curvature. Bulk black hole solutions with hyperbolic horizons were obtained in \cite{Birmingham:1998nr,Emparan:1998he,Emparan:1999gf}; these include a locally-AdS solution, which is a natural higher-dimensional analogue of the BTZ black hole. It is perhaps not surprising that considering ETW branes behind the horizon of these hyperbolic black hole then gives us a higher-dimensional scenario with behaviour similar to the $d=2$ case.
Another way to see why considering hyperbolic spatial slices is helpful is to consider the worldvolume of the ETW brane. In the Euclidean black hole this is a wormhole; for hyperbolic spatial sections this is $\Sigma_g \times \mathbb R$, and we will see that when we consider the locally-AdS ``topological'' black hole, the worldvolume geometry is precisely the Maldacena-Maoz wormhole of \cite{Maldacena:2004rf}. This is a solution of Einstein's equations with a negative cosmological constant. By contrast, there is no classical solution describing a wormhole with flat spatial sections; the bulk spacetime thus plays a more essential role in the construction of the solutions where the ETW brane is a wormhole with flat spatial sections, and it seems natural that it's harder to construct solutions in the limit where the effective dynamics is approximated by Einstein gravity localised on the ETW brane.
As in the charged black hole case, these hyperbolic black holes have an extremal limit, where the temperature goes to zero for fixed horizon area. But the more important effect from the ETW brane perspective is that the curvature term dominates over the mass term in the metric at large distances. As a result, in the limit of interest, the mass term is negligible and the dynamics of the ETW brane is as in the locally-AdS case, which gives a finite time for the ETW brane to reach the Euclidean boundary, as in the $d=2$ BTZ black hole example.
In the next section, we review some of the previous work on these models. In section \ref{hyp}, we introduce the generalization to consider hyperbolic spatial sections. Section \ref{disc} presents some concluding remarks. In appendix \ref{phase}, I include some speculations on the existence of other bulk saddle-points with the same boundary conditions as the hyperbolic black hole; this is a side comment independent of the main discussion.
\section{End of the world brane cosmology}
\label{rev}
The holographic model of cosmology we consider was first proposed in \cite{Cooper:2018cmb}. The model consists of an AdS black hole bulk with one asymptotic region, with a dynamical constant-tension ETW brane behind the horizon, as pictured in figure \ref{fig1}. The induced geometry on the ETW brane worldvolume is that of a closed FRW universe, with the radial position playing the role of the scale factor. The bulk action is
\begin{equation}
I = \frac{1}{16 \pi G}\left[ \int_\mathcal{M} d^{d+1} x \sqrt{-g}\, (R-2\Lambda) + 2\int_{\partial \mathcal{M}} d^d y \sqrt{-h}\, K - 2(d-1) \int_{\mathcal{Q}} d^d y \sqrt{-h}\, T \right] ,
\end{equation}
where $\Lambda=-\frac{d(d-1)}{2 L^2}$ is a cosmological constant, $K$ is the trace of the extrinsic curvature and $T$ is the tension of the ETW brane with worldvolume $\mathcal{Q}$, which we take to be one component of the boundary $\partial\mathcal{M}$ of the spacetime, the other component corresponding to the asymptotically AdS conformal boundary. We consider a $(d+1)$-dimensional bulk spacetime, dual to a $d$-dimensional CFT on the boundary.
Previous work considered an AdS-Schwarzchild black hole bulk solution
\begin{equation}
ds^2 = -f(r)dt^2+\frac{dr^2}{f(r)}+\frac{r^2}{L^2}dx^a dx_a, \quad \quad f(r)\equiv \frac{r^2}{L^2}-\frac{\mu}{r^{d-2}},
\end{equation}
where $a=1,\ldots,d-1$. This has a horizon at $r=r_H$, where $r_H^d = \mu L^2$. The brane has stress-energy tensor $8\pi G T_{ab}=(1-d)T h_{ab}$, and the action implies that the boundary condition for the bulk metric at $\mathcal{Q}$ is
\begin{equation}
K_{ab}-Kh_{ab}=(1-d)T h_{ab}
\end{equation}
The $tt$ component of this equation leads to the brane equation of motion
\begin{equation} \label{braneeom}
\left( \frac{d r}{d t}\right)^2=\frac{f^2(r)}{T^2 r^2}\left(T^2r^2 - f(r)\right).
\end{equation}
In the Lorentzian black hole geometry, the brane will reach a maximum radius $r_0$, with $(r_0)^d= \frac{r_H^d}{1 - T^2 L^2}$, which we take to occur at $t=0$. Note $r_0 > r_H$ for $T>0$, and $r_0 \to \infty$ as $T \to L^{-1}$. To the future and past of this, $r(t)$ decreases, as pictured in figure \ref{fig1}. The brane worldvolume geometry is thus a closed FRW big-bang/big-crunch cosmology,
where the brane radius $r(t)$ plays the role of the scale factor, and the brane equation of motion \eqref{braneeom} corresponds to the Friedmann equation in this worldvolume cosmology.
The state on the $t=0$ slice can be obtained by a Euclidean path integral. In the Euclidean black hole, the motion of the ETW brane is
\begin{equation} \label{Euclidean braneeom}
\left( \frac{d r}{d \tau}\right)^2=\frac{f^2(r)}{T^2 r^2}\left(f(r) - T^2 r^2 \right).
\end{equation}
This now has a minimum at $r=r_0$. Since the ETW brane is inside the black hole in the Lorentzian geometry, it is at its minimum radius in the Euclidean solution at $\tau= \beta/2$, where $\beta = 2\pi L^2/r_H$ is the periodicity in Euclidean time $\tau$. It reaches the AdS boundary at a time
\begin{equation} \label{tetw}
\tau^{ETW} = \frac{\beta}{2} - \int_{r_0}^\infty \frac{dr}{f(r)} \frac{Tr}{\sqrt{f(r) - T^2 r^2}}.
\end{equation}
To avoid self-intersections in the Euclidean solution, we need $\tau^{ETW}>0$. However, setting $r= r_0 x$, we have
\begin{equation}
\sigma^{ETW} = \frac{2 \tau^{ETW}}{\beta} = 1 - \frac{d}{2\pi} TL (y_0)^{\frac{d-2}{2}} \int_1^\infty \frac{dx}{x^2 (1- (y_0)^{-d} x^{-d}) \sqrt{1-x^{-d}}},
\end{equation}
where $y_0 = r_0/r_H$, so $(y_0)^{-d} = 1-T^2 L^2$. It is clear that for $d>2$ we can't take $r_0 \to \infty$ while keeping $\tau^{ETW} >0$. There must then be some critical value $T= T_* < L^{-1}$ such that $\tau^{ETW} = 0$, and if we consider $T > T_*$ we will have self-intersecting branes in the Euclidean solution.
\section{Hyperbolic spatial sections}
\label{hyp}
The previous discussion was for branes with flat spatial sections. Consider now the generalization to hyperbolic spatial sections. This changes relatively little in the previous analysis; the relevant black hole solution is now \cite{Birmingham:1998nr,Emparan:1998he,Emparan:1999gf}
\begin{equation}
ds^2 = -f(r)dt^2+\frac{dr^2}{f(r)}+\frac{r^2}{L^2}ds^2_{\Sigma_g}, \quad \quad f(r) = \frac{r^2}{L^2} -1 -\frac{\mu}{r^{d-2}},
\end{equation}
where $\Sigma_g$ is a $(d-1)$-dimensional compact manifold of constant negative curvature. For $\mu \in (\mu_{min}, \infty)$, where
\begin{equation}
\mu_{min} = - \frac{2}{d} \left( \frac{d-2}{d} \right)^{\frac{d-2}{2}} L^{d-2},
\end{equation}
the black hole has a horizon at $r=r_H>0$ where $f(r_H)=0$; we can write
\begin{equation}
\mu = r_H^{d-2} (\frac{r_H^2}{L^2} - 1 ).
\end{equation}
The inverse temperature of the black hole is
\begin{equation} \label{bhyp}
\beta = \frac{4\pi r_H L^2}{d r_H^2 - (d-2) L^2} .
\end{equation}
As $\mu \to \mu_{min}$, $r_H^2 \to \frac{d-2}{d} L^2$, and $f(r)$ develops a double root. The temperature goes to zero.\footnote{The possibility of having a vanishing temperature seems like it would be useful in constructing ETW brane solutions, but as we will see, the more important effect is the $-1$ in $f(r)$, which dominates over the mass term at large radial distances.}
For $\mu=0$, the solution is locally AdS$_{d+1}$. The metric with $\Sigma_g$ replaced by a non-compact hyperbolic space $H^{d-1}$ is simply AdS written in a Rindler-like coordinate system (the coordinate transformation between these coordinates and Poincar\'e coordinates basically reduces to the relation between Rindler coordinates and hyperbolic coordinates on the conformal boundary). The geometry of interest here is then obtained by quotienting by a discrete isometry group $\Gamma$ such that $H^{d-1}/\Gamma = \Sigma_g$. This $\mu=0$ solution is therefore referred to as the topological black hole. The $\mu=0$ solution has a finite temperature, which is related to the Rindler acceleration temperature in the boundary theory if we don't do the quotient.
We consider an ETW brane inside this black hole, with a turnaround at $r=r_0$ at $t=0$. Considering the Euclidean continuation, the brane equation of motion is \eqref{Euclidean braneeom}, and the brane reaches the AdS boundary at
\begin{equation}
\tau^{ETW} = \frac{\beta}{2} - \int_{r_0}^\infty \frac{dr}{f(r)} \frac{Tr}{\sqrt{f(r) - T^2 r^2}}.
\end{equation}
The key difference from the previous flat case is the behaviour of the square root factor in the denominator: if we set $r= r_0 x$, in both cases, we pull out a factor of $1/\sqrt{r_0^2 (1-T^2L^2)}$, but whereas in the flat case we had $1-T^2 L^2 \sim 1/r_0^d$, so this factor goes like $r_0^{\frac{d-2}{2}}$, which blows up as $r_0 \to \infty$, for the hyperbolic black hole, $r_0^2 (1-T^2 L^2) = L^2 - \mu \frac{L^2}{r_0^{d-2}}$. The last term becomes negligible if $\mu$ remains finite as $r_0 \to \infty$, so $1/\sqrt{r_0^2 (1-T^2L^2)}$ has a finite limit as $r_0 \to \infty$.
This can be simply illustrated by considering the case $\mu =0$. As noted above, we still have a finite temperature at $\mu=0$: \eqref{bhyp} gives $\beta = 2\pi L$. Setting $r=r_0 x$, we have
\begin{equation}
\tau^{ETW} = \pi L - TL^2 \int_{1}^\infty \frac{x dx}{\left( x^2 - \frac{L^2}{r_0^2} \right)\sqrt{x^2-1}}.
\end{equation}
As $TL \to 1$, $r_0 \to \infty$, and we can do the integral exactly; it's equal to $\frac{\pi}{2}$, so $\tau^{ETW} \to \frac{\pi L}{2}$. Thus, there's no self-intersection problem here; we can freely take the ETW brane to large radius in arbitrary dimensions.\footnote{The integrand is larger, and hence $\tau^{ETW}$ smaller, for finite $r_0$; but we are mostly interested in the region of large $r_0$, so we will not analyse in detail the positivity of $\tau^{ETW}$ for finite $r_0$.} It is interesting to note that for $\mu =0$, the bulk geometry is locally AdS$_{d+1}$, and the brane is locally AdS$_d$, so the solution we are considering here is just a quotient of the sub-critical braneworld introduced in \cite{Karch:2000ct}.
Generalizing to $\mu \neq 0$, so long as we hold $\mu$ fixed as we take the limit as $r_0 \to \infty$ the $\frac{\mu}{r^{d-2}}$ term in $f(r)$ becomes negligible in the limit, so the integral will converge to the above $\mu=0$ expression. The only difference for $\mu \neq 0$ is then that the black hole has a different temperature, so
\begin{equation}
\tau^{ETW} = \frac{\beta}{2} - \frac{\pi L}{2}.
\end{equation}
There is thus a good solution in the $r_0 \to \infty$ limit for all $\mu$ such that $\beta > \pi L$. This corresponds to $\mu < \mu_{max}$ where
\begin{equation}
\mu_{max} = (r_H^{max})^{d-2} \left( \frac{(r_H^{max})^2}{L^2} - 1 \right), \quad \frac{r_H^{max}}{L} = \frac{2 + \sqrt{4 + d(d-2)}}{d}.
\end{equation}
For fixed $T, L$, the boundary interval $\tau^{ETW}$ is determined by $\mu$; from the boundary perspective we would like to invert this relationship and fix $\tau^{ETW}$ and determine $\mu$. In the limit as $r_0 \to \infty$, it is easy to calculate that $\tau^{ETW}$ runs from $+\infty$ to $0$ monotonically as $\mu \in (\mu_{min}, \mu_{max})$. Thus, there is a one to one map from possible values of $\tau^{ETW}$ to $\mu$ in this range for large $r_0$.
Consider the induced geometry on the ETW brane: in general this is
\begin{equation}
ds_b^2 = \frac{dr^2}{f(r)- T^2 r^2} + r^2 d\Sigma_g^2.
\end{equation}
If $\mu=0$, setting $r=r_0 \cosh \rho$, we have
\begin{equation}
ds_b^2 = r_0^2 (d \rho^2 + \cosh^2 \rho d\Sigma_g^2).
\end{equation}
Thus, for the topological black hole, the induced geometry on the brane is precisely the Maldacena-Maoz wormhole \cite{Maldacena:2004rf}. In general, if we take $r_0 \to \infty$ at fixed $\mu$, the mass term in $f(r)$ is negligible at the brane position, and the induced geometry on the brane will be approximately the same. This is natural, as the effective theory on the brane in this limit reduces to Einstein gravity, and the Maldacena-Maoz wormhole is a classical solution of this theory. This offers another perspective on the advantages of considering the brane with hyperbolic cross-sections.
At finite $r_0$, the induced geometry on the brane is modified for $\mu \neq 0$; these modifications can be interpreted from the brane effective theory prespective as due to the effective stress tensor on the brane dual to the bulk geometry. Note that for $\mu <0$, this effective stress tensor will have a negative energy density. This can be interpreted as a Casimir energy due to putting the theory on a compact hyperbolic space. For $\mu=0$, the energy density on the brane vanishes due to a cancellation between the negative Casimir energy and the positive thermal energy associated with the finite temperature.
\section{Discussion}
\label{disc}
We have seen that considering spatial manifolds with negative curvature provides a good environment for implementing the holographic cosmology proposal of \cite{Cooper:2018cmb} in higher dimensions. The bulk solution is a hyperbolic black hole, whose structure in higher dimensions is more similar to the structure of the BTZ black hole in three dimensions. There are many questions about the holographic cosmology which it will be interesting to investigate in this context in the future: how the degrees of freedom on the brane are encoded in the boundary theory, what predictions these models make for the cosmological evolution on the brane, the role of the negative Casimir energy of the CFT, and others.
In top-down models, the CFT usually includes scalar fields; conformally invariant scalar fields have a coupling to the curvature of the background which leads to an instability when the background is negatively curved. In the bulk this is related to an instability for probe branes at large radial position to run away towards the boundary \cite{Seiberg:1999xz}. This would seem to be a significant issue with the models we are considering, and indeed if we consider the zero temperature black hole the bulk solution is unstable; branes at any radial position run away to the boundary. However, if we consider the finite temperature solutions (for example in AdS$_5$) there is a non-trivial radial potential for probe (D3)-branes, and branes close to the horizon will collapse into the black hole \cite{Landsteiner:1999up}. Thus, branes must tunnel through this potential barrier to escape to infinity, and the dual CFT at finite temperature is meta-stable. The model should provide a useful way of describing the cosmology holographically on timescales short compared to the decay time.
For flat spatial slices, the Euclidean ETW brane geometry can be analytically continued in one of the spatial directions to obtain an eternal traversable wormhole geometry. The relation between the problem of constructing ETW brane cosmologies with flat slices and the challenges in constructing eternal traversable wormholes was explored in \cite{VanRaamsdonk:2021qgv,Fallows:2022ioc,Waddell:2022fbn}, following \cite{Freivogel:2019lej}.\footnote{The construction of wormholes using a coupling between theories was earlier explored in \cite{Betzios:2019rds}. See also \cite{Betzios:2021fnm,Antonini:2022xzo,Antonini:2022opp} for more recent work on these wormholes.} It is therefore interesting to ask if there is a connection between the geometries with hyperbolic spatial slices studied here and Lorentzian wormholes. The metric on $\Sigma_g$ will not have a translation symmetry, but if it has a surface of time reflection symmetry, we can define a Lorentzian solution by analytic continuation. This Lorentzian geometry will be a quotient of AdS similar to those considered in \cite{Horowitz:1998xk}, which gives again a time-dependent cosmology with big bang big crunch ``singularities" where the quotient degenerates. The spatial slices of this brane geometry will be a wormhole; whether it is traversable or not depends on the ratio of the time needed to cross the radial spatial direction between the two boundaries and the finite lifetime of the cosmology. In the simple $\mu=0$ case, both times should be of order one in units of the AdS scale, so understanding whether these wormholes are traversable seems to require a detailed case by case analysis.
Finally, as we have emphasized connections to the $d=2$ case in this work, it is worth noting that for $d=2$ the Euclidean boundary conditions with an interval of Euclidean time between two brane boundaries can be filled in in two different ways: we can have a piece of the bulk BTZ black hole bounded by a connected ETW brane, or a piece of vacuum AdS bounded by disconnected ETW branes \cite{Fujita:2011fp}. We have discussed the analogue of the connected solution; is there an analogue of the disconnected one? This requires an analogue of thermal AdS for the hyperbolic black hole. No such solution is known, but in the appendix I suggest such solutions may exist. They are more difficult to construct analytically; it might be interesting to study them numerically.
\section*{Acknowledgements}
I thank Mark van Raamsdonk for helpful discussions. This work is supported in part by STFC through grant ST/T000708/1, and by a grant from the Simons Foundation. This work was performed in part at the Aspen Center for Physics, which is supported by National Science Foundation grant PHY-1607611.
|
\section{Introduction}
The observation of correlated insulator behavior and unconventional superconductivity in twisted bilayer graphene \cite{cao_correlated_2018,cao_unconventional_2018} has stimulated research on twisted van der Waals heterostructures. Those features appear close to a ``magic'' twist angle, $\theta\approx 1.1^\circ$, that had been predicted by previous theoretical work \cite{trambly_de_laissardiere_localization_2010,bistritzer_moire_2011}. Similar features have been observed in other graphene-based twisted systems \cite{park_tunable_2021,hao_electric_2021,cao_pauli-limit_2021}.
One of the issues with twisted bilayer graphene is the difficulty of reproducing the observations. A small deviation of the twist angle results in a large variation of the critical temperature, critical magnetic field, critical current and density range \cite{balents_superconductivity_2020}. This can be explained by the semi-metallic nature of graphene. Correlated features arise from the flat energy bands. Such bands can be observed near the $K$ valley of graphene, but only close to the magic angle. A small deviation from this magic angle has a big effect on the flatness of the bands, and then this leads to experimental features that are difficult to reproduce exactly.
Similar correlated states have been observed and predicted in other van der Waals systems that present a moir\'e pattern, like twisted bilayer boron nitride \cite{xian_multiflat_2019} and transition metal dichalcogenides (TMD). The latter can occur in a homobilayer ($MX{}_2/MX{}_2$) \cite{wang_correlated_2020,an_interaction_2020,zhang_flat_2020,venkateswarlu_electronic_2020,ghiotto_quantum_2021} or heterobilayer ($MX_2/M'X'_2$) \cite{phillips_commensurate_2019,ruiz-tijerina_interlayer_2019,regan_mott_2020,tang_simulation_2020,brotons-gisbert_moire-trapped_2021,huang_correlated_2021,li_imaging_2021,morales-duran_metal-insulator_2021,vitale_flat_2021}, where $M$ stands for the metal and $X$ for the chalcogen. One of the advantages of TMD is the existence of a continuum of magic angles with correlated insulating states. The semiconducting nature of TMD leads to a band structure from the monolayer that is flatter than that of graphene. It is then easier to observe flat bands for a larger spectrum of angles. This makes correlated behavior easier to observe and reproduce. \cite{balents_superconductivity_2020}. Another interesting feature of TMD is the valley-dependent spin splitting caused by spin-orbit coupling \cite{xiao_coupled_2012}. Depending on the strength of the spin-orbit splitting at the Brillouin zone corners, the valence band maximum can be at the $\pm K$ or $\Gamma$ point \cite{angeli__2021}. One TMD of particular interest is twisted WSe${}_2$, where correlated insulating states and evidence of superconductivity have been observed \cite{wang_correlated_2020,an_interaction_2020,zhang_flat_2020}.
A strategy has been proposed to obtain effective, low-energy models of twisted TMD \cite{jung_ab_2014}. It can be applied to specific materials like MoTe${}_2$ and WSe${}_2$ \cite{wu_topological_2019,pan_band_2020}.
Following this approach, twisted WSe${}_2$ can be described by a two-dimensional Hubbard model on a triangular lattice with complex-valued hopping amplitudes to the third nearest-neighbor. Moreover, the effect of a voltage bias across the bilayer is to change the relative phases of spin-up and spin-down hopping amplitudes \cite{pan_band_2020}.
A $120^\circ$ magnetic order is expected at half-filling in the simple, nearest-neighbor Hubbard model on a triangular lattice \cite{weber_magnetism_2006,laubach_phase_2015,misumi_mott_2017,wietek_mott_2021}. Away from half-filling, superconductivity of type $d+id$ has been predicted to exist \cite{weber_magnetism_2006,chen_unconventional_2013}. The effective moir\'e-band Hubbard model, although also based on the triangular lattice, has a very different dispersion relation from the simple nearest-neighbour model, but similar magnetic phases are predicted \cite{pan_band_2020,zang_hartree-fock_2021,zang_dynamical_2022}. Chiral superconductivity has been predicted using renormalization group analysis in the weak coupling regime for twisted bilayers TMD with only nearest-neighbour hopping \cite{wu_pair-density-wave_2022}.
In this paper, we use quantum cluster methods, namely the variational cluster approximation (VCA) and the cluster dynamical mean-field theory (CDMFT), to argue that superconductivity of type $d+id$ occurs in the effective Hubbard model describing twisted WSe${}_2$ with large values of interaction and hopping up to the third nearest neighbor. The superconducting domes can be tuned by modifying the twist angle and by using an external displacement field. The complex hopping terms introduced by the displacement field are a major factor in determining the shape and location of the superconducting dome. We did not take into consideration the $120^\circ$ antiferromagnetic magnetic phase predicted in Ref \cite{pan_band_2020}, since no competition with the superconducting phase is expected in the doping region considered (the magnetic phase is concentrated close to half-filling).
This paper is organized as follow. In Sect. \ref{sec:model}, we introduce the moir\'e Hubbard model. Then in Sect. \ref{sec:numerical} we present the two numerical methods used : VCA and CDMFT. In Sects. \ref{sec:electron} and \ref{sec:hole}, we present our results for the electron- and hole-doped sides, respectively. We discuss some implications of our results in section \ref{sec:discussion}.
\section{Model}
\label{sec:model}
The electronic states in twisted layered materials like bilayer graphene and hBN can be described by a moir\'e band model \cite{jung_ab_2014}, from which one can extract a number of bands equal to the number of atoms present in the moir\'e unit cell. The same ideas can be applied to TMD homobilayers like MoTe${}_2$ \cite{wu_topological_2019} and WSe${}_2$ \cite{pan_band_2020} with strong spin-orbit splitting at the $\pm K$ point of the Brillouin zone. A moir\'e model includes the effect of the interlayer tunneling between two layers of similar or identical lattice structures. A layer-dependent moir\'e potential describes the effect of the twist angle $\theta$ between the two layers. One can then tune this moir\'e potential and interlayer tunneling to reproduce the density of states from STM experiments \cite{zhang_flat_2020}. This also reproduces the local density of states around the high-symmetry positions where the atoms of the two layers sit on top of each other. It is also possible to introduce the effect of an external perpendicular displacement field that affects the moir\'e potential. From this moir\'e band model, one can construct an effective low-energy tight-binding model to study a few moir\'e valence bands.
To study the possibility of superconductivity in twisted WSe${}_2$, we use the model proposed in Ref. \cite{pan_band_2020}. In that work, the authors constructed an effective Hubbard model for the first moir\'e valence band. To do so, they extracted the moir\'e Wannier function from the moir\'e band model \cite{wu_topological_2019}. This method has been used to describe various homobilayer and heterobilayer TMDs \cite{wu_topological_2019,pan_band_2020}.
The Hamiltonian of this effective Hubbard model is expressed as
\begin{align}
H=\sum_{\sigma}\sum_{\mathbf{r},\mathbf{r'}}t_{\mathbf{r}\mathbf{r'},\sigma }c^\dagger_{\mathbf{r},\sigma}c_{\mathbf{r'},\sigma}+U\sum_{\mathbf{r}}n_{\mathbf{r},\uparrow}n_{\mathbf{r},\downarrow},
\label{eq:H}
\end{align}
where $\mathbf{r},\mathbf{r'}$ are sites on the triangular lattice and $\sigma=\uparrow,\downarrow$ represents the spins associated with the $+K$ and $-K$ valleys, respectively. $c_{\mathbf{r},\sigma}$ ($c^\dagger_{\mathbf{r},\sigma}$) destroys (creates) an electron at site $\mathbf{r}$ with spin $\sigma$. $n_{\mathbf{r},\sigma}=c^\dagger_{\mathbf{r},\sigma}c_{\mathbf{r},\sigma}$ is the number of electrons on site $\mathbf{r}$ with spin $\sigma$. $U$ is the on-site repulsion between electrons. The hopping matrix $t_{\mathbf{r}\mathbf{r'},\sigma}$ is Hermitian : $t_{\mathbf{r}\mathbf{r'},\sigma}=t^*_{\mathbf{r'}\mathbf{r},\sigma}$. The system has a threefold rotation symmetry and three reflections across mirror planes perpendicular to the $xy$ plane, corresponding to the $C_{3v}$ symmetry group. Finally, time-reversal symmetry $(\mathcal{T})$ requires that $t_{\mathbf{r}\mathbf{r'},\sigma}=t^*_{\mathbf{r}\mathbf{r'},-\sigma}$. Considering those relations, it is possible to express the hopping parameters between the $n$-th nearest neighbors as $t_{\sigma n}=|t_n|e^{i\phi_n^\sigma}$, where the amplitude $|t_n|$ and the phase $\phi_n^\sigma$ depend on the twist angle $\theta$ and on the inter layer potential $V_z$. The on-site interaction $U$ also depends on $\theta$, $V_z$ and $\epsilon$, the effective background dielectric constant, which can change depending on the experimental substrate. The value of the parameters considered here are extracted from Ref. \cite{pan_band_2020}.
The twist angle is defined in relation to an axis going through a AA stacking site where metal or chalcogen atoms are perfectly aligned with the same type of atom in the other layer. We consider hopping up to the third ($n=3$) nearest neighbor. We neglect interactions between different sites. Every parameter is expressed in units of $|t_1|$. We limit ourselves to twist angles $\theta\ge3^\circ$ since lattice relaxation effects, that are not taken into account in this model, become important for small twist angles ($\theta<2.5^\circ$) \cite{pan_band_2020,enaldiev_stacking_2020}.
When only nearest-neighbor hopping is present ($t_{\sigma2}=t_{\sigma3}=0$), there is a particle-hole symmetry under $n\rightarrow2-n$ and $\phi\rightarrow \pi-\phi$. This symmetry is lost with the addition of next-nearest-neighbor hopping. We then expect different properties on the electron- and hole-doped sides. This explains differences with some of the previous work on the triangular lattice Hubbard model \cite{wu_pair-density-wave_2022}.
Model (\ref{eq:H}) can be considered as a triangular Hubbard model where control over the different parameters is obtained by tuning $\theta$, $V_z$ and $\epsilon$. The simple nearest-neighbor triangular lattice Hubbard model has been studied previously with prediction of a $120^\circ$ antiferromagnetic order at half-filling \cite{weber_magnetism_2006,laubach_phase_2015,misumi_mott_2017,wietek_mott_2021}. This order as been predicted in model (\ref{eq:H}) for a range of $V_z$ \cite{pan_band_2020,zang_hartree-fock_2021,zang_dynamical_2022}. The simple nearest-neighbor triangular lattice Hubbard is also predicted to have a type $d+id$ superconductivity phase when doped away of half-filling \cite{weber_magnetism_2006,chen_unconventional_2013}. Chiral superconductivity, including type $d+id$, has been predicted using renormalization group analysis in the weak interaction limit for twisted bilayers TMD \cite{wu_pair-density-wave_2022}. We do expect to observe a similar phase in the twisted WSe${}_2$ in the strongly interacting case, modulated by the different control parameters.
The presence of complex hopping terms up to the third nearest neighbor in model (\ref{eq:H}) distinguishes our work from previous studies on the simple, nearest-neighbor triangular lattice Hubbard model. Another difference from previous work on superconductivity in twisted WSe${}_2$ is our ability to treat large values of the interaction $U$.
\begin{figure}[h]
\includegraphics[scale=0.75]{schema_model_new.pdf}
\caption{Cluster used in the VCA method. The different hopping parameters of model (\ref{eq:H}) are illustrated.
}
\label{fig:schema_model}
\end{figure}
\section{Numerical methods}
\label{sec:numerical}
In order to probe the possibility of superconductivity in model (\ref{eq:H}), we use quantum cluster methods with an exact diagonalization solver at zero temperature. Specifically, we use two of those methods: the variational cluster approximation (VCA) and cluster dynamical mean-field theory (CDMFT).
In those methods, the lattice is tiled into an infinite number of identical clusters. Two systems are then considered : the original system on the infinite lattice, described by the Hamiltonian $H$, and the ``reference'' system defined only on the cluster, described by the Hamiltonian $H'$. The only requirement is that both systems $H$ and $H'$ share the same interaction term.
Probing the various broken symmetries is done using an embedding scheme on $H'$. In the VCA, $H'$ is the restriction of $H$ on the cluster, augmented by one or more {\it Weiss fields} representing the broken symmetry operators being probed. In CDMFT, the effect of the environment of the cluster is added to $H'$ via a set of uncorrelated orbitals (the {\it bath}). In both cases, an optimal self-energy can be extracted from $H'$ and applied to $H$. In other words, the approximate Green function $\mathbf{G}$ of the original lattice Hamiltonian is constructed by inserting the self-energy $\mathbf{\Sigma}$ found from $H'$ in Dyson's equation:
\begin{align}
\mathbf{G}(\mathbf{k},\omega)=\frac{1}{\mathbf{G}_0^{-1}(\mathbf{k},\omega)-\mathbf{\Sigma}(\omega)}.
\label{eq:Dyson}
\end{align}
In this expression the wavevectors $\mathbf{k}$ are restricted to the reduced Brillouin zone (rBZ) of the superlattice of clusters. $\mathbf{G}_0$ is the non-interacting Green function of the infinite lattice. $\mathbf{G}$, $\mathbf{G}_0$ and $\mathbf{\Sigma}$ are $2L\times2L$ matrices, $L$ being the number of sites on the cluster (the factor of 2 because of spins). From the Green function, we can compute the average of any one-body operator, like the order parameter associated with superconductivity.
To explain the last point, let us consider a matrix $s_{\alpha\beta}$ that defines the one-body operator $\hat{S}$
\begin{align}
\hat{S}=\sum_{\alpha\beta,\mathbf{k}}s_{\alpha\beta}(\mathbf{k})c_\alpha^\dagger(\mathbf{k}) c_\beta(\mathbf{k}).
\end{align}
The indices $\alpha,\beta$ stand for a composite of cluster site and spin indices. The expectation value of $\hat{S}$ is
\begin{align}
\langle \hat{S}\rangle =\int\frac{d\omega}{2\pi}\int_{\text{rBZ}}\frac{d^2k}{(2\pi)^2}\text{tr}[\mathbf{s}(\mathbf{k})\mathbf{G}(\omega,\mathbf{k})]
\label{eq:expectation}
\end{align}
where the frequency integral is taken over a contour that circles the negative real axis, targeting only the occupied states.
The size of each cluster should allow one to compute the electron Green function numerically. The exact diagonalization method used to compute the cluster Green function limits the total number of orbitals used. The VCA procedure allows one to consider a larger cluster than CDMFT, because of the baths orbitals that need to be added in the latter (see section \ref{sec:cdmft} for details). This leads to a better account of spatial correlations in VCA, whereas CDMFT has a better treatment of temporal correlations. Both methods allow us to go beyond mean-field theory while keeping the fully correlated character of the model within each cluster. It is possible to apply those methods to strongly correlated systems to probe various broken symmetry phases \cite{dahnken_variational_2004,sahebsara_hubbard_2008,faye_interplay_2017,kancharla_anomalous_2008,dash_pseudogap_2019}.
Let us mention that the geometric frustration of the triangular lattice and the complex hopping amplitudes enhance the sign problem in Monte Carlo methods. This further justifies the use of the exact diagonalization solver in this work.
\subsection{The Variational Cluster Approximation}
The VCA is a variational method on the electron self-energy based on Potthoff's self-energy functional approach \cite{potthoff_variational_2003,dahnken_variational_2004}. This method as been used to predict magnetic phases \cite{dahnken_variational_2004,sahebsara_hubbard_2008} and superconductivity \cite{faye_interplay_2017}. For a detailed review of the method, see Refs \cite{potthoff_variational_2012,potthoff_cluster_2018}.
To determine the optimal one-body part of $H'$ in VCA, the electron self-energy $\mathbf{\Sigma}$ associated with $H'$ is used as a variational self-energy to construct the Potthoff self-energy functional \cite{dahnken_variational_2004}:
\begin{align}
\begin{split}
\Omega[\mathbf{\Sigma}(\xi)]=&\Omega'[\mathbf{\Sigma}(\xi)]+\text{Tr}\ln[-(\mathbf{G}_0^{-1}-\mathbf{\Sigma}(\xi))^{-1}]\\
&-\text{Tr}\ln(-\mathbf{G}'(\xi)),
\end{split}
\label{eq:potthoff}
\end{align}
with $\mathbf{G}'$ the physical Green function of the cluster. The symbol $\xi$ stands for a collection of parameters that define the one-body part of $H'$. Tr stand for a functional trace, which implies a sum over frequencies, momenta and bands. Finally, $\Omega'$ is the exact grand potential of $H'$, i.e., the ground state energy if the chemical potential $\mu$ is included in the Hamiltonian. We use the Lanczos method at zero temperature to compute numerically $\mathbf{G}'(\omega)$ and $\Omega'$. The Potthoff functional $\Omega[\mathbf{\Sigma}(\xi)]$ is computed exactly but only on a restricted space of $\mathbf{\Sigma}(\xi)$ that are the physical self-energy of the reference system $H'$.
The optimal value of the self-energy $\mathbf{\Sigma}$ corresponds to stationary value of $\Omega(\xi)$. We use a standard optimization method (Newton-Raphson) in the space of the one-body parameter $\xi$ to solve
\begin{align}
\frac{\partial\Omega(\xi)}{\partial \xi}=0.
\label{eq:min}
\end{align}
We then construct the approximate Green function $\mathbf{G}$ of the original lattice Hamiltonian by inserting the self-energy obtained here in Eq. (\ref{eq:Dyson}), and extract the order parameter using Eq. (\ref{eq:expectation}).
To study the superconducting phase in twisted WSe${}_2$, we use a 12-site cluster that respects the $C_{3v}$ symmetry of model (\ref{eq:H}) as shown in Fig. \ref{fig:schema_model}. We expect the superconducting order parameter to fall into one of the irreducible representations (irrep) of $C_{3v}$. Clusters that do not respect this symmetry do not allow us to probe the different irreps separately. Finally, the exact diagonalization used to find the Green function $\mathbf{G}'$ of the cluster limits the maximum number of sites we can use. All those reasons justify the use of the 12-site cluster shown in Fig \ref{fig:schema_model}.
The Weiss fields that define the set of one-body parameters $\xi$ are chosen to correspond to one of the irrep of $C_{3v}$. We first define the pairing operators between two adjacent sites as
\begin{align}
\text{singlet:}\quad S_{\mathbf{r},i}=c_{\mathbf{r},\uparrow}c_{\mathbf{r}+\mathbf{e}_i,\downarrow}-c_{\mathbf{r},\downarrow}c_{\mathbf{r}+\mathbf{e}_i,\uparrow}\\
\text{triplet:}\quad T_{\mathbf{r},i}=c_{\mathbf{r},\uparrow}c_{\mathbf{r}+\mathbf{e}_i,\downarrow}+c_{\mathbf{r},\downarrow}c_{\mathbf{r}+\mathbf{e}_i,\uparrow}
\end{align}
We then define pairing operators falling in the different irreps of $C_{3v}$ by the combinations given in Table \ref{table:pairing}. We note that the chiral pairings $d+id$ and $d-id$ are degenerate and so a VCA solution found with one of them would also exist with the other (same for $p+ip$ and $p-ip$).
\begin{table}[h]
\begin{ruledtabular}
\begin{tabular}{ccA}
Irrep&Symbol&\multicolumn{2}{c}{Operators}\\
\hline
$A_1$&$s$&\hat{\Delta}_s&=\sum_{\mathbf{r}}\left(S_{\mathbf{r},1}+S_{\mathbf{r},2}+S_{\mathbf{r},3}\right)\\
$A_2$&$f$&\hat{\Delta}_f&=\sum_{\mathbf{r}}\left(T_{\mathbf{r},1}+T_{\mathbf{r},2}+T_{\mathbf{r},3}\right)\\
$E_1$&$d+id$&\hat{\Delta}_{d+id}&=\sum_{\mathbf{r}}\left(S_{\mathbf{r},1}+\omega S_{\mathbf{r},2}+\bar{\omega}S_{\mathbf{r},3}\right)\\
&$d-id$&\hat{\Delta}_{d-id}&=\sum_{\mathbf{r}}\left(S_{\mathbf{r},1}+\bar{\omega}S_{\mathbf{r},2}+\omega S_{\mathbf{r},3}\right)\\
&$p+ip$&\hat{\Delta}_{p+ip}&=\sum_{\mathbf{r}}\left(T_{\mathbf{r},1}+\omega T_{\mathbf{r},2}+\bar{\omega}T_{\mathbf{r},3}\right)\\
&$p-ip$&\hat{\Delta}_{p-ip}&=\sum_{\mathbf{r}}\left(T_{\mathbf{r},1}+\bar{\omega}T_{\mathbf{r},2}+\omega T_{\mathbf{r},3}\right)
\end{tabular}
\end{ruledtabular}
\caption{Pairing operators used as Weiss field in the VCA procedure with $\omega=e^{\frac{2\pi i}{3}}$ and $\bar{\omega}=e^{-\frac{2\pi i}{3}}$. Each belongs to an irreducible representation (irrep) of the $C_{3v}$ point group. Chiral pairings belong to the $E_1$ irrep.\label{table:pairing}}
\end{table}
\subsection{Cluster dynamical mean-field theory}
\label{sec:cdmft}
In CDMFT, each cluster in $H'$ is augmented by an adjustable environment made of a set of uncorrelated orbitals (the {\it bath}). Each cluster then defines an Anderson impurity model that needs to be solved by an impurity solver to extract the Green function of the cluster $\mathbf{G}_c(\omega)$. Like before, we use exact diagonalization as our impurity solver at zero temperature. By using Dyson's equation we extract the cluster self-energy $\mathbf{\Sigma(\omega)}$
\begin{align}
\mathbf{G}_c^{-1}(\omega)=\omega-\mathbf{t}_c-\mathbf{\Gamma}(\omega)-\mathbf{\Sigma}(\omega)
\end{align}
where $\mathbf{t}_c$ is the hopping matrix on the cluster and $\mathbf{\Gamma}(\omega)$ is the hybridization function that contains the information about the uncorrelated bath orbitals. We then use the cluster self-energy as an approximation of the full lattice Green function in Eq. (\ref{eq:Dyson}). The bath parameters are chosen to minimize the difference between $\mathbf{G}_c(\omega)$ and the local version of $\mathbf{G}(\mathbf{k},\omega)$. We use a self-consistency procedure to do so. More details on this method can be found in Refs \cite{kancharla_anomalous_2008,senechal_bath_2010,senechal_quantum_2015}.
To probe superconductivity using CDMFT, we need to include anomalous hybridizations between bath orbitals. In this work we use a 3-site cluster with 9 bath orbitals defined on Fig. \ref{fig:schema_cdmft}. The bath parametrization is based on the $C_{3v}$ symmetry. To control the type of superconductivity probed, we change the value of the coefficient $\chi$: Type $s$ corresponding to $\chi=1$ and type $d+id$ to $\chi = e^{\frac{2\pi i}{3}}$. Using this method we can probe one by one every pairing symmetry listed in Table \ref{table:pairing}.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.5]{schema_cdmft_all.pdf}
\caption{Impurity model used in the CDMFT procedure with a 3-site cluster and 9 uncorrelated bath orbitals. The latter are separated in three groups (red, green, blue) with the same energy $\varepsilon_i$ and hybridization amplitude $\gamma_i$ within each group, by symmetry. The anomalous operators, $\Delta_i$, are represented by the dashed lines and can be modified to allow for different pairing symmetries.
}
\label{fig:schema_cdmft}
\end{center}
\end{figure}
\section{Electron doping}
\label{sec:electron}
\begin{figure*}[t!]
\subfigure{
\label{fig:supra_ep30}
\includegraphics[scale=0.49]{supra_ep30_new.pdf} }
\quad
\subfigure{
\includegraphics[scale=0.49]{supra_Vz_new.pdf}
\label{fig:supra_Vz}}
\caption{Amplitude of the $d+id$ order parameter on the electron doped-side obtained in the VCA for $\epsilon=30$. a) Effect of the twist angle on the order parameter function of density $n$. b) Effect of the interlayer potential $V_z$ (in meV) on the order parameter for $\theta=4^{\circ}$.}
\end{figure*}
We first start by the electron-doped side of the phase diagram. We use the VCA in this region in order to understand the effect of the twist angle $\theta$ and $V_z$. In all our VCA computations we used the 12-site cluster of Fig. \ref{fig:schema_model} with the Weiss field defined in Table \ref{table:pairing}.
We first look at the effect of the twist angle on the superconducting state in the absence of displacement field ($V_z=0$). We consider a typical value of the effective dielectric constant ($\epsilon=30$). The data from Ref. \cite{pan_band_2020}, gives us the parameters presented in Table \ref{table:hoppingVz0}, for the different twist angles considered. At $V_z=0$, the hopping parameters all have the same phase $\phi^s_n=-\pi$, which implies real-valued parameters.
\begin{table}[h]
\begin{ruledtabular}
\begin{tabular}{ccccc}
$\theta$&$t_1$&$t_2$&$t_3$&$U$\\
\hline
$4^\circ$&$-1$&$-0.053$&$-0.116$&$10.61$\\
$4.5^\circ$&$-1$&$-0.045$&$-0.135$&$7.78$\\
$4.92^\circ$&$-1$&$-0.037$&$-0.149$&$6.24$\\
\end{tabular}
\end{ruledtabular}
\caption{Hopping parameters for the different twist angles $\theta$ considered for $V_z=0$ and $\epsilon=30$. In this case, all hopping parameters are real.\label{table:hoppingVz0}}
\end{table}
We probed all the pairing symmetries enumerated in Table \ref{table:pairing} (the Weiss field is defined in the rightmost column). We found a non-trivial VCA solution with $d+id$ pairing, but none for the other pairing operators listed in Table \ref{table:pairing}, except $d-id$, which is degenerate with $d+id$.
In Fig. \ref{fig:supra_ep30}, we show the $d+id$ order parameter for $V_z=0$. A solution exists for all twist angles considered, for some interval of density $n$ ($n=1$ corresponding to half-filling). The order parameter is computed from the VCA Green function using Eq. (\ref{eq:expectation}). We also observe that the superconducting dome is greatly affected by the twist angle. Increasing the twist angle widens the range of density where superconductivity exists. The maximum order parameter also increases with twist angle. This is caused by the drop of $U$ as $\theta$ increases. Assuming a monotonous relation between the order parameter and the critical temperature $T_c$, we can infer that, in the range of angles considered, $T_c$ can be tuned by changing the twist angle. Another reason $T_c$ depends on the twist angle $\theta$ is that our energy scale is defined by $t_1$, which itself increases as a function of $\theta$ when expressed in meV \cite{pan_band_2020} (the relation between $T_c$ and the order parameter depends on this energy scale).
We then consider the effect of the external perpendicular displacement field. In Fig. \ref{fig:supra_Vz} we show the $d+id$ order parameter obtained from VCA for a twist angle $\theta=4^{\circ}$, for different values of $V_z$. Again, a nontrivial solution was found only for the $d\pm id$ order parameter. The doping range is greatly affected by the displacement field. It is noted in Ref. \cite{pan_band_2020} that the position of the van Hove singularity (VHS) depends on $V_z$ and crosses over the half-filling point at around $V_z=28$ meV. We expect that one of the factors affecting the doping range is this displacement of the VHS. This could explain why the maximum of the dome is closer to half-filling for values of $V_z$ close to $28$ meV. It could also explain the localization of the superconducting dome for higher values of $V_z$, when the VHS is beyond half-filling.
We also used CDMFT in this region, but the procedure did not converge for every twist angle and value of $V_z$.
\section{Hole doping}
\label{sec:hole}
\begin{figure*}[t!]
\subfigure{
\label{fig:supra_cdmft}
\includegraphics[scale=0.49]{sup_angle_ep15_plot_new.pdf} }
\quad
\subfigure{
\includegraphics[scale=0.49]{sup_Vz_plot_new.pdf}
\label{fig:supra_VZ_cdmft}}
\caption{Amplitude of the $d+id$ order parameter on the hole doped-side obtained from CDMFT for $\epsilon=15$. a) Effect of the twist angle on the order parameter function of density $n$. b) Effect of the interlayer potential $V_z$ (in meV) on the order parameter for $\theta=4^{\circ}$.}
\end{figure*}
We did not find any VCA solution on the hole-doped side. This was caused by a discontinuity in the Potthoff self-energy functional (\ref{eq:potthoff}) as a function of the Weiss field, leading to an invalid solution of Eq. (\ref{eq:min}) or no solution at all. This reflects the structure of the energy levels on the cluster : on the hole-doped side, the energy levels seem to be more closely spaced than on the electron-doped side, which may lead to those discontinuities. This does not mean that superconductivity is not present. In fact, by using CDMFT we are able to extract information on this region. To do so, we use the method and cluster from section \ref{sec:cdmft}. Note that the value of $U$ in the Hubbard model that gives access to the Mott physics depends on cluster size \cite{dang_mott_2015}. To make sure we are considering similar physical properties with a 3-site cluster, we set the dielectric constant $\epsilon$ to $15$, which doubles the value of $U$. As mentioned before, $\epsilon$ is determined by the sample's environment and $\epsilon=15$ is still in the range of values seen in the literature.
In Fig. \ref{fig:supra_cdmft}, we show the $d+id$ order parameter as a function of density, obtained from the CDMFT procedure at $V_z=0$. We observe that the maximum of the order parameter is affected by the twist angle. The increase in the order parameter with twist angle follows what was observed on the electron-doped side. The main difference here is that the width and position of the dome seem to be less affected by the twist angle.
For this set of parameters, we do not observe other types of superconductivity. We probed the other pairing symmetries by changing the value of $\chi$ in Fig. \ref{fig:schema_cdmft} (e.g. from $e^{\frac{2\pi i}{3}}$ to $1$ to probe extended s-wave pairing) and by changing the type of nearest-neighbor anomalous operator from $S_{\mathbf{r},i}$ to $T_{\mathbf{r},i}$ (singlet to triplet). In all cases, the order parameter vanishes for the range of doping accessible.
On figure \ref{fig:supra_VZ_cdmft}, we show the $d+id$ order parameter for different values of $V_z$ and a fixed twist angle of $4^{\circ}$. The effect of the displacement field is different from what was observed on the electron-doped side using VCA. The superconducting dome is greatly affected by the modulation of $V_z$. As observed in the figure, the maximum of the parameter decreases with $V_z$. We also observe that the superconducting dome is suppressed around half-filling when $V_z$ is increased. Here again, we did not find any solution for the other pairing symmetries.
\section{Discussion}
\label{sec:discussion}
Our results agree with the evidence of superconductivity presented in Ref. \cite{wang_correlated_2020}, were a zero resistivity state was observed when doping away from half-filling for a twist angle of $5.1^\circ$, in the presence of a displacement field. The two superconducting domes are separated by an insulating state around half-filling. We predict a superconducting phase on both electron- and hole-doped sides. We also predict that the two superconducting domes are separated from each other by a finite range of densities that includes an insulating state.
While looking at the effect of the twist angle, the order parameter seems to be mostly affected by the on-site interaction $U$. The shrinking of the superconducting dome seems to be related to increasing the interaction $U$. It has been shown before for the square-lattice Hubbard model, that the maximum of the superconducting dome decreases with $U$ \cite{senechal_competition_2005,dash_pseudogap_2019}. The values of $U$ considered in this work are in the strong interaction range. The $U$ dependence observed here agrees with the expected behavior.
The effect of $V_z$, is different on both sides of half-filling. As stated above, we expect that the behavior on the electron-doped side is linked to the shift in the VHS. However, we do not think the same explanation works on the hole-doped side, since the maximum of the dome does not shift according to the VHS. This difference might be explained by the choice of numerical method and by the difference in the cluster used.
The effect of the displacement field on the order parameter in the hole-doped region does not seem to come from the change in the interaction $U$. In fact, while we increase $V_z$, $U$ decreases but, contrary to what is expected, the superconducting dome also decreases, even though we are still in the strong interaction regime. This is likely related to the complex hopping parameter induced by $V_z$ : A phase factor is introduced on the bonds, which produces a non-zero staggered flux on each plaquette. In our CDMFT calculation, only one plaquette is considered, which implies a non-zero total flux. We expect that this flux, the equivalent of a magnetic field, pushes the superconducting phase down. To explore this, we tried a 4-site system with 8 baths to have two plaquettes with total flux of zero. Unfortunately, due to the loss of the $C_{3v}$ symmetry on that cluster, no meaningful results were found. This effect seem to be less present in the VCA procedure since the 12-site cluster is more balanced even if the total flux is still non-zero. It would be interesting to isolate the contribution of the complex phases and observe its effect on the superconducting dome.
In both methods used here we found a $d\pm id$ solution that was greatly affected by the perpendicular displacement field. We explored the possibility of other pairings, like those studied in Ref. \cite{wu_pair-density-wave_2022}, but did not detect them. This can be explained by the presence of second- and third-nearest neighbor hopping terms in Model (\ref{eq:H}), and by the numerical methods used here that allow us to probe larger values of the interaction $U$.
In both VCA and CDMFT, we tried different clusters with varying sizes and symmetries. To have a proper convergence of the algorithms, the cluster needs to respect a set of criteria. Like stated above, the most important condition is to respect the $C_{3v}$ group symmetry. This allows us to separate the different order parameters in each irreducible representation. This also provides stability to the procedure. Moreover, the use of exact diagonalization as impurity solver limits the total number of orbitals that can be realistically considered to 12, because of the computational resources needed. We also found a better convergence around low doping when the total number of orbitals was even. For example, a cluster of 7 sites gave rise to a discontinuity in the Potthoff functional close to half-filling, which translated into convergence issues in the VCA algorithm. This was likely due to the unpaired electron that shifted the cluster's energy levels. All those subtleties in the numerical procedure made it harder to probe the effect of cluster size and shape.
\section{Conclusion}
We have applied quantum cluster methods to an effective low-energy model for a twisted bilayer of WSe${}_2$, developed in Refs \cite{jung_ab_2014,pan_band_2020}, and found a superconducting dome in both electron- and hole-doped regions in the strong-interaction regime. In both cluster methods used, VCA and CDMFT, we only found superconductivity with $d\pm id$ pairing symmetry. The twist angle can be used to modify the maximum order parameter. If a perpendicular displacement field is applied, the superconducting dome is modified greatly because of the complex phase acquired by the hopping parameters. All those predictions are in agreement with previous experimental evidence \cite{wang_correlated_2020} of superconductivity in twisted bilayer WSe${}_2$.
\begin{acknowledgments}
This work has been supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) under grant RGPIN-2020-05060 and by NSERC postgraduate scholarships doctoral program.
Computational resources were provided by Digital Research Alliance of Canada and Calcul Qu\'ebec.
\end{acknowledgments}
|
\section{Introduction}
Traditionally, machine learning has been applied to areas where there are no known underlying laws that describe the observed data. For example, deep convolutional neural networks (CNNs) have been very successful in computer vision applications (see \cite{He2016,Ren2015,Krizhevsky2012} for some famous examples), even though it is unclear how the lower-dimensional manifold of ``valid'' images is parametrised. This manifold has to be discovered in a purely data-driven way by feeding the network vast amounts of data. Of course, also in this case domain specific prior knowledge is built into the structure of the network: deep CNNs are so successful since the network structure supports the hierarchical representation of higher level features and encapsulates the fact that the input images are not random, but nearby pixels are likely to have similar values. In contrast, other fields of science such as physics are governed by well known fundamental laws. There appears to be little value in applying machine learning in this case since these laws allow the construction of an exact mathematical description, which can then be translated into an algorithm for making predictions with traditional techniques from numerical analysis. However, often the dynamics of a system is only partially constrained by these underlying laws. For example, the motion of a particle in an external potential might conserve energy and angular momentum, while the exact form of the potential or the expression for the kinetic energy is still unknown and needs to be inferred from observations. In fact, a very successful approach to constructing new theories in modern physics is to write down the most general Lagrangian which is invariant under certain symmetry transformations and to then constrain the remaining parameters through a fit to experimental data. Famously, general relativity \cite{Einstein1923} is constructed by demanding that the theory is invariant under local coordinate transformations but this allows the existence of a cosmological constant which needs to be constrained by experimental evidence. In particle physics a similar method is used for the construction of Chiral Perturbation Theory \cite{Gasser1984,Scherer2003} and other effective theories for which the Lagrangian can be written as an infinite sum of terms that are invariant under certain symmetry transformations. The expansion coefficients have to be inferred from experiments.
In this work we study simpler systems arising in classical mechanics of point particles. The key idea is to represent the Lagrangian by a neural network in such a way that it exactly conserves quantities like linear and angular momentum. The weights of the network are then learned from training data which consists of trajectories that follow the true dynamics and are perturbed by random noise. While in the present paper we use synthetic training data that is obtained by integrating the exact equations of motion numerically for a set of model systems, it should be stressed that the data could also come from the observation of a real physical system. Drawing a historical analogy, our neural network can be seen as the virtual version of an astronomer who tries to derive the celestial equations of motion from observations of the planets in the night sky. In contrast to its medieval predecessors, our virtual astronomer has read Noether's paper \cite{Noether1918} and takes care not to violate fundamental physical laws such as angular momentum conservation.
\subsection{Related work}\label{sec:relatedwork}
Several approaches for solving time dependent differential with neural networks have been pursued in the literature. In \cite{Kadupitiya2022} the authors train an LSTM \cite{Hochreiter1997} based neural network to predict the solution at the next time step, given the solution at a number of $s$ previous times. Hence, this can be seen as an extension of classical linear multistep methods (see e.g. \cite{Hairer1993}) and the hope is that with a suitably trained network it is possible to substantially increase the timestep size without losing accuracy, thereby making the method more efficient than traditional methods. For Hamiltonian systems a completely different approach is pursued in \cite{Greydanus2019}: a neural network is trained to learn a scalar valued function $\mathcal{H}_{\text{NN}}(q,p)$ which approximates the true Hamiltonian $\mathcal{H}(q,p)$ as a function of the generalised coordinates $q\in\mathbb{R}^{d}$ and conjugate momenta $p\in\mathbb{R}^{d}$; here and in the following $d$ is the dimension of the dynamical system. It is well known that symplectic integrators preserve a so-called shadow Hamiltonian $\widetilde{\mathcal{H}}$, i.e. the numerical solution generated with the true Hamiltonian $H$ is the \textit{exact} solution of $\widetilde{\mathcal{H}}$. For this reason, the numerical solution approximately preserves energy. While the authors on \cite{Greydanus2019} do not exploit this since they use a non-symplectic fourth order Runge Kutta integrator, in \cite{Chen2019} it is argued that symplectic Neural Networks (SRNNs) show superior performance. SRNNs represent both the kinetic and potential energy by a neural network and then use a symplectic Verlet method \cite{Verlet1967} to propagate the solution. Instead of working in the Hamiltonian formulation, it might be more convenient to represent the Lagrangian by a neural network as in \cite{Cranmer2020}. As argued there, working in the Lagrangian framework is more flexible since the system can be formulated in \textit{any} coordinate system, not just canonical coordinates, which are often not easy to find.
The methods discussed so far only preserve some physical quantities such as the total energy. It can be shown that energy conservation directly follows from time invariance of the system, i.e. the fact that the Lagrangian does not explicitly depend on time. Noether's Theorem \cite{Noether1918} generalises this result and shows that invariance of the Lagrangian under \textit{any} infinitesimal transformation results in a corresponding conservation law. Hence, if the neural network that represents the system's Lagrangian can be constructed such that it is exactly invariant under a certain infinitesimal symmetry transformation, this will guarantee the conservation of a physical quantity. The role of symmetries in neural networks has been explored in \cite{Mattheakis2019}. The authors use so-called ``hub-neurons'' to construct a neural network which represents a symmetric function $f:\mathbb{R}\rightarrow \mathbb{R}$ that satisfies $f(x)=f(-x)$. They then show that this gives better fit to ground-truth training data which is polluted by non-symmetric noise. They use a similar technique to construct a symplectic neural network for the solution of time-dependent systems, but it is not clear whether their network is truly symplectic. The central role of symmetries is also exploited in \cite{Ling2016}. Here the aim is to use neural networks to represent closures for the Reynolds-averaged Navier Stokes equations. Such a closure allows computing the shear-stress in the fluid from the non-turbulent component of the velocity field. Crucially, if the flow is isotropic, the shear stress can only depend on combinations of the velocity field which are invariant under rotations. The authors of \cite{Ling2016} achieve this by systematically constructing rotationally invariant scalars from the vector field and restricting the input of their neural network to these scalars.
\subsection{Aim of this paper}\label{sec:achievements}
In this paper we show how neural networks can be used to solve Lagrangian dynamical systems while exactly preserving physical quantities. In contrast to \cite{Mattheakis2019} we consider invariance under \textit{continuous} transformations. For this, we follow \cite{Cranmer2020} and represent the Lagrangian as a neural network $\mathcal{L}_{\text{NN}}(q,\dot{q})$ which takes as input the coordinates $q$ and velocities $\dot{q}$. As in \cite{Ling2016}, the first layer of the network reduces $q,\dot{q}$ to a set of scalar variables which do not change under the symmetry transformations that correspond to the conserved quantities. Noether's Theorem then guarantees exact conservation and during training the network learns the finer details of dynamics of the system that are \textit{not} constrained by the conservation laws. Since the network has to find solutions in a smaller sub-manifold of the entire solution space, training is likely to be more efficient. The main achievements of this paper are as follows:
\begin{enumerate}
\item We show how invariance under continuous symmetries can be built into Lagrangian neural network models by passing the input through a so-called symmetry-enforcing layer; according to Noether's Theorem this leads to the exact conservation of a corresponding physical quantity.
\item To demonstrate this, we consider three example systems:
\begin{enumerate}
\item The motion of a single particle in three dimensions under the influence of a central force field, which is invariant under rotations and thus conserves angular momentum.
\item The motion of a massive particle in the rotationally invariant Schwarzschild metric, which is the relativistic pendant of the first problem. Again, three dimensional angular momentum is conserved.
\item The motion of two interacting particles in $D$-dimensional space, where the interaction potential is invariant under rotations and translations. This results in the conservation of the $D$-dimensional linear momentum vector and the $\frac{1}{2}D(D-1)$ independent components of the antisymmetric $D\times D$ angular momentum tensor.
\end{enumerate}
\item For all systems we demonstrate that the generated trajectories are more realistic and that they conserve angular momentum and (in the case of the two-particle system) linear momentum to a high degree of accuracy.
\item We further show numerically that enforcing invariance under continuous symmetry transformations makes the trajectories more stable with respect to small perturbations of the initial conditions.
\end{enumerate}
\paragraph{Structure} This paper is organised as follows: in Section \ref{sec:method} we review Noether's Theorem and explain the construction of Lagrangian neural networks that are invariant under continuous symmetry transformations. We write down the explicit form of the symmetry-enforcing input layers for the three model systems considered in this work. Numerical experiments for the model systems are described in Section \ref{sec:results}, where we present results that demonstrate the superior performance of our approach. Section \ref{sec:conclusion} contains our conclusions and ideas for future work. Some more technical details are relegated to the appendices: we discuss the construction of scalar invariants under the special orthogonal group $SO(D)$ in \ref{sec:rotationally_invariant_combinations} and present loss histories for the different neural networks in \ref{sec:loss_histories}.
\section{Methodology}\label{sec:method}
\subsection{Continuous symmetries and conservation laws}\label{sec:noether_theorem}
For completeness and further reference, we start by writing down Noether's Theorem here and refer the reader to \cite{Arnold2013} for a proof and further details.
\begin{theorem}[Noether \cite{Noether1918}]
Consider a dynamical system which is formulated on a manifold $M$ with tangent bundle $TM$; the Lagrangian of this system is a real-valued function $\mathcal{L}:TM\rightarrow \mathbb{R}$. Let $h^s:M\rightarrow M$ be a family of maps parametrised with the continuous parameter $s\in\mathbb{R}$ and let $h^s_{*,q}:TM_q\rightarrow TM_{h^s(q)}$ be the local derivative which maps between the tangent bundles at $q$ and $h^s(q)$. If the Lagrangian $\mathcal{L}$ is invariant under $h^s$, i.e.
\begin{equation}
\mathcal{L}(h^s(q),h^s_{*,q}(\dot{q})) = \mathcal{L}(q,\dot{q})\qquad\text{for all $q,\dot{q}\in TM$},\label{eqn:L_invariance}
\end{equation}
then the following quantity is a constant of motion in the sense that $dI/dt=0$:
\begin{equation}
I = \frac{\partial \mathcal{L}}{\partial \dot{q}} \frac{dh^s(q)}{ds}\Big|_{s=0}.\label{eqn:noether_conserved_I}
\end{equation}
\end{theorem}
As an example, consider a particle moving in three dimensional space $\mathbb{R}^3$ and assume that the Lagrangian is invariant under rotations, i.e. the action of the special orthogonal group $SO(3)$. In this case, $M=\mathbb{R}^3$, $TM=\mathbb{R}^3\times \mathbb{R}^3$ and the maps $h^{s}$, $h^{s}_{*,q}$ can be written as linear transformations
\begin{xalignat}{2}
h^{s}(q) &= \exp\left[s \Gamma \right]q, & h^{s}_{*,q}(v) &= \exp\left[s \Gamma \right]v
\qquad\text{for $q,v\in\mathbb{R}^3\times\mathbb{R}^3$}
\label{eqn:rotations}
\end{xalignat}
where $\Gamma$ is an element of the fundamental representation of the Lie-algebra $\mathfrak{so}(3)$ of $SO(3)$. The three matrices corresponding to rotations around the $x$-, $y$- and $z$- axis are
\begin{xalignat}{3}
\Gamma_x &= \begin{pmatrix}0 & 0 & 0\\0 & 0 & -1\\0 & 1 & 0\\\end{pmatrix}, &
\Gamma_y &= \begin{pmatrix}0 & 0 & 1\\0 & 0 & 0\\-1 & 0 & 0\\\end{pmatrix}, &
\Gamma_z &= \begin{pmatrix}0 & -1 & 0\\1 & 0 & 0\\0 & 0 & 0\\\end{pmatrix}.
\end{xalignat}
Further, we have that
\begin{equation}
\frac{dh^{s}}{ds}\Big|_{s=0} = \Gamma q
\end{equation}
and hence there are three conserved quantities, namely
\begin{xalignat}{3}
L_x &= \frac{\partial \mathcal{L}}{\partial \dot{q}_3}q_2 - \frac{\partial \mathcal{L}}{\partial \dot{q}_2}q_3, &
L_y &= \frac{\partial \mathcal{L}}{\partial \dot{q}_1}q_3 - \frac{\partial \mathcal{L}}{\partial \dot{q}_3}q_1, &
L_z &= \frac{\partial \mathcal{L}}{\partial \dot{q}_2}q_1 - \frac{\partial \mathcal{L}}{\partial \dot{q}_1}q_2.
\label{eqn:conserved_L_rotational}
\end{xalignat}
For a non-relativistic particle such as the one discussed in Section \ref{sec:kepler_problem}, the kinetic energy in the Lagrangian is $\frac{m}{2}(\dot{q}_1^2+\dot{q}_2^2+\dot{q}_3^2)$ and hence $\partial L/\partial \dot{q}_j = m\dot{q}_j$. In this case, the three conserved quantities are the components of the angular momentum vector
\begin{equation}
L = m\;q \times \dot{q}.\label{eqn:angular_momentum_3d}
\end{equation}
It is worth pointing out that the quantities in Eq. \eqref{eqn:conserved_L_rotational} are conserved for more general cases, such as for the Lagrangian that governs the motion of a massive relativistic particle in the rotationally invariant Schwarzschild metric discussed in Section \ref{sec:schwarzschild_problem}.
\subsection{Neural networks with built-in conservation laws}\label{sec:nn_with_symmetries}
The central idea of this paper is as follows: we wish to construct a mapping $\Phi:\mathbb{R}^{2d}\rightarrow \mathbb{R}^d$ which predicts the acceleration $\ddot{q}=\Phi(q,\dot{q})$ from the position $q$ and velocity $\dot{q}$ such that the dynamics of this system preserves certain quantities \textit{exactly}. To achieve this, we use the Lagrangian formalism and represent the Lagrangian $\mathcal{L}_{\text{NN}}$ by a neural network as in \cite{Greydanus2019}. The equations of motion are obtained by finding the stationary points of the action (= the time integral of the Lagrangian) which implies that
\begin{equation}
\frac{d}{dt}\frac{\partial \mathcal{L}_{\text{NN}}}{\partial\dot{q}} - \frac{\partial \mathcal{L}_{\text{NN}}}{\partial q} = 0.\label{eqn:lagrangian_stationary}
\end{equation}
Following \cite[Eq. (6)]{Cranmer2020} the acceleration can then be obtained by taking the total time derivative of Eq. \eqref{eqn:lagrangian_stationary}:
\begin{equation}
\ddot{q} = J_{\dot{q},\dot{q}}^{-T}\left( \frac{\partial \mathcal{L}_{\text{NN}}}{\partial q} - J^T_{q,\dot{q}}\dot{q}\right) \label{eqn:qddot_Lagrangian}
\end{equation}
where the components of the matrices $J_{q,\dot{q}}$, $J_{\dot{q},\dot{q}}$ are given by
\begin{xalignat}{2}
\left(J_{q,\dot{q}}\right)_{ij} &= \frac{\partial^2 \mathcal{L}_{\text{NN}}}{\partial q_i \partial \dot{q}_j}, &
\left(J_{\dot{q},\dot{q}}\right)_{ij} &= \frac{\partial^2 \mathcal{L}_{\text{NN}}}{\partial \dot{q}_i \partial \dot{q}_j}.\label{eqn:J_matrices}
\end{xalignat}
The crucial point is now that the Lagrangian is constructed such that it is exactly invariant under the symmetry transformations $h^s$ which correspond to the conserved quantities. To guarantee that $\mathcal{L}_{\text{NN}}$ satisfies Eq. \eqref{eqn:L_invariance}, we write $\mathcal{L}_{\text{NN}} = D_{L} \circ D_{L-1} \circ D_1 \circ S$, where $D_\ell$ are standard, non-linear dense layers. The symmetry-enforcing first layer $S:\mathbb{R}^{2d}\rightarrow \mathbb{R}^{M}$ is a function that combines the inputs $q,\dot{q}$ into a set of $M$ invariants such that
\begin{equation}
S_m(h^s(q),h^s_{*,q}(\dot{q})) = S_m(q,\dot{q})\qquad\text{for $m=1,2,\dots,M$}.\label{eqn:S_invariance}
\end{equation}
For example, if we consider a three-dimensional system which is invariant under the rotations in Eq. \eqref{eqn:rotations}, then the simplest three invariants that can be constructed from the position $q$ and the velocity $\dot{q}$ are $S_1(q,\dot{q})=q^2$, $S_2(q,\dot{q})=\dot{q}^2$ and $S_3(q,\dot{q})=q\cdot \dot{q}$.
The structure of the entire function $\Phi$ is shown in Fig. \ref{fig:nn_architecture}. Removing the symmetry-enforcing layer $S$ results in the standard architecture already introduced in \cite{Greydanus2019}. The explicit form of symmetry-enforcing layers for specific systems is discussed in Section \ref{sec:model_systems}.
\begin{figure}
\begin{center}
\includegraphics[width=0.8\linewidth]{.//network_architecture.pdf}
\caption{Structure of the function $\Phi$ which maps the inputs $q,\dot{q}$ to the predicted acceleration $\ddot{q}=\Phi(q,\dot{q})$. The symmetry-enforcing layers $T$, $R$ and $R\circ T$ for the specific systems studied in this work are discussed in the main text. A network without any symmetry constraints can be obtained by removing $S$, as represented by an undecorated downward arrow $\downarrow$.}
\label{fig:nn_architecture}
\end{center}
\end{figure}
The weights of the dense layers $D_\ell$ are learned by training the neural network on measurements of the physical system. As already pointed out above, the training data could consist of (noisy) observations of a real physical system. However, to demonstrate the principle and to not be restricted by a lack of training data, here we use synthetic, simulated data instead. More specifically, we compare the output of the neural network $\Phi_{\text{NN}}$ to (noisy) ground truth data. The ground truth predictions $\widehat{y}$ are generated by adding random noise to the true acceleration $\ddot{q} = \Phi_{\text{true}}(q,\dot{q})$ for some choice of $q$, $\dot{q}$; the function $\Phi_{\text{true}}$ is obtained by replacing $\mathcal{L}_{\text{NN}}$ with the true Lagrangian $\mathcal{L}_{\text{true}}$ in Eqs. \eqref{eqn:qddot_Lagrangian} and \eqref{eqn:J_matrices}. For each prediction $\widehat{y}$ the input $X$ of the neural network is obtained by also perturbing the corresponding $q,\dot{q}$ with random noise, thus simulating measurement errors for a real physical system.
During training we minimise the standard mean square error (MSE) loss function, which can be written as
\begin{equation}
\text{Loss}\left(\{X^{(j)},\widehat{y}^{(j)}\}_{j=1}^{N}\right) = \frac{1}{N}\sum_{j=1}^{N}\left(\Phi(X^{(j)})-\widehat{y}^{(j)}\right)^2\label{eqn:MSE_loss}
\end{equation}
for a set of $N$ training samples $(X^{(1)},\widehat{y}^{(1)}),(X^{(2)},\widehat{y}^{(2)}),\dots,(X^{(N)},\widehat{y}^{(N)})$ with $X^{(j)}:=(q^{(j)}+\nu_1^{(j)},\dot{q}^{(j)}+\nu_2^{(j)})$ and $\widehat{y}^{(j)}=\ddot{q}^{(j)}+\nu_3^{(j)}$ where $\nu_i^{(j)}$ represents the random noise on the data.
\subsection{Model systems}\label{sec:model_systems}
We now discuss the three physical model systems, their symmetries and the corresponding conserved quantities in more detail.
\subsubsection{Motion of single particle in a gravitational potential}
\label{sec:kepler_problem}
First, we consider the dynamics of a particle of mass $m$ moving in a central force field with a potential energy that is inversely proportional to the distance from the origin. The true Lagrangian is
\begin{equation}
\mathcal{L}_{\text{true}}(q,\dot{q}) = \frac{m}{2}\dot{q}^2 + \frac{\alpha}{|q|}
\qquad\text{with $q,\dot{q}\in\mathbb{R}^3$}
\label{eqn:true_lagrangian_single_particle}
\end{equation}
for some positive constant $\alpha>0$. It is well known that for negative total energy $E=\frac{m}{2}\dot{q}^2-\alpha/|q|<0$ the trajectory is an ellipse where the origin of the coordinate system coincides with one focal point. Kepler studied this kind of motion in the solar system and derived three laws from observations, without knowing the analytical relationship between position, velocity and acceleration. The Lagrangian in Eq. \eqref{eqn:true_lagrangian_single_particle} is invariant under the rotations in Eq. \eqref{eqn:rotations}, which leads to the conservation of the vector-valued angular momentum $L$ in Eq. \eqref{eqn:angular_momentum_3d}. As a consequence, the entire trajectory lies in the plane spanned by the initial position $q(0)$ and initial velocity $\dot{q}(0)$; this plane is orthogonal to the vector $L$.
Our objective is to learn the neural network Lagrangian $\mathcal{L}_{\text{NN}}(q,\dot{q})$ and hence the relationship between $q$, $\dot{q}$ and $\ddot{q}$ from data under the assumption that the dynamics is invariant under the action of the special orthogonal group $SO(3)$. To achieve this, recall that the simplest invariant scalars that can be constructed from position $q\in\mathbb{R}^3$ and velocity $\dot{q}\in\mathbb{R}^3$ are $q^2$, $\dot{q}^2$ and $q\cdot\dot{q}$. Hence, the symmetry-enforcing layer $S$ in Fig. \ref{fig:nn_architecture} should take as input $q$ and $\dot{q}$ and output these three invariant quantities; in the following we denote this layer as $R$ (for ``rotationally invariant''), see Fig. \ref{fig:rotational_invariant_layer_3d}.
\begin{figure}
\begin{center}
\includegraphics[width=0.3\linewidth]{.//symmetry_enforcing_layer_3d.pdf}
\caption{Layer $R$ enforcing invariance under the group $SO(3)$ of rotations for a single particle moving in three dimensional space as discussed in Section \ref{sec:kepler_problem}.}
\label{fig:rotational_invariant_layer_3d}
\end{center}
\end{figure}
Adopting the Einstein sum convention of summing over pairs of identical upper and lower indices, the three quantities that are conserved for the dynamics generated with neural network Lagrangian $\mathcal{L}_{\text{NN}}$ can be written as
\begin{equation}
L^{(\rho)}_{\text{NN}} = \varepsilon^{\rho\sigma\tau} q_\sigma\frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{q}_\tau}
\label{eqn:angular_momentum_NN}
\end{equation}
where $\varepsilon$ is the three-dimensional Levi Civita symbol defined in Eq. \eqref{eqn:levi_civita_definition}. We stress that by construction the angular momenta in Eq. \eqref{eqn:angular_momentum_NN} are \textit{exactly} conserved, independent of the weights of the neural network.
The quantity $L^{(\rho)}_{\text{NN}}$ can be likened to the shadow Hamiltonian, which is conserved for symplectic integrators of time-independent Hamiltonian systems. For the true Lagrangian in Eq. \eqref{eqn:true_lagrangian_single_particle} the ``neural network'' angular momentum $L^{(\rho)}_{\text{NN}}$ reduces to the three components of the ``true'' angular momentum vector in Eq. \eqref{eqn:angular_momentum_3d}, namely
\begin{equation}
L^{(\rho)}_{\text{true}} = m \varepsilon^{\rho\sigma\tau} q_\sigma\dot{q}_\tau.
\label{eqn:angular_momentum_true}
\end{equation}
For future reference we combine the quantities defined in Eqs. \eqref{eqn:angular_momentum_NN} and \eqref{eqn:angular_momentum_true} into two vectors:
\begin{xalignat}{2}
L_{\text{NN}} &= \left(L_{\text{NN}}^{(1)},L_{\text{NN}}^{(2)},L_{\text{NN}}^{(3)}\right),&
L_{\text{true}} &= \left(L_{\text{true}}^{(1)},L_{\text{true}}^{(2)},L_{\text{true}}^{(3)}\right).\label{eqn:angular_momentum_kepler_vector}
\end{xalignat}
\subsubsection{Motion of a massive relativistic particle in the Schwarzschild metric}\label{sec:schwarzschild_problem}
Next, we consider the motion of a massive particle in the Schwarzschild metric \cite{Schwarzschild1916,Droste1917}, which is the relativistic equivalent to the Kepler problem in Section \ref{sec:kepler_problem}. With the space-time dependent metric tensor $g(q)$ the Lagrangian can be written as\footnote{The geodesic line element is $ds=\sqrt{g_{\mu\nu}dq^\mu dq^\nu}$, which would imply that the Lagrangian is $ds/dt=\sqrt{g_{\mu\nu}(q)\dot{q}^\mu \dot{q}^\nu}$. However, since the square root is a monotonous function, the Lagrangian in Eq. \eqref{eqn:schwarzschild_lagrangian} has the same stationary points and thus generates the same dynamics.}
\begin{equation}
\mathcal{L} = g_{\mu\nu}(q)\dot{q}^\mu \dot{q}^\nu \qquad \text{for $\mu,\nu=1,2,3,4$},\label{eqn:schwarzschild_lagrangian}
\end{equation}
where $q^\mu$ are the components of the four-dimensional contravariant time-position vector $q=(x,\tau)=(x_1,x_2,x_3,\tau)$. $\dot{q}^\mu = dq^\mu/dt$ denotes the derivative with respect to the eigen-time $t$ experienced by a moving observer and $\tau$ is the time measured by a static observer far away from the origin, where $g \rightarrow \operatorname{diag}{(+1,+1,+1,-1)}$ tends to the constant metric of flat space-time.
Expressing the spatial coordinate $x$ in spherical coordinates, the true Lagrangian in Eq. \eqref{eqn:schwarzschild_lagrangian} can be written as
\begin{equation}
\mathcal{L}_{\text{true}} = -\left(1-\frac{r_s}{r}\right)\dot{\tau}^2 + \left(1-\frac{r_s}{r}\right)^{-1}\dot{r}^2 + r^2\left(\dot{\theta}^2 + \sin^2(\theta)\dot{\varphi}^2\right) \label{eqn:schwarzschild_lagrangian_II}
\end{equation}
where $r_s$ is the Schwarzschild radius and $r=|x|$ is the distance from the origin. It is easy to see\footnote{Observe that $r^2\left( \dot{\theta}^2 + \sin^2(\theta)\dot{\varphi}^2\right) = \dot{x}^2 - (x\cdot\dot{x})^2/x^2$.} that Eq. \eqref{eqn:schwarzschild_lagrangian_II} is invariant under rotations in three dimensional space and can be expressed entirely in terms of the three scalars $x^2$, $\dot{x}^2$, $x\cdot\dot{x}$ and the time derivative $\dot{\tau}$. As a consequence, the dynamics conserves the (specific) three-dimensional angular momentum which turns out to be
\begin{equation}
L = x\times \dot{x}.
\end{equation}
Note that this expression differs from the one in Eq. \eqref{eqn:angular_momentum_3d} only through scaling by the (constant) mass $m$. In analogy to Eq. \eqref{eqn:angular_momentum_true} we can also write down the three components of the angular momentum vector as
\begin{equation}
L^{(\rho)}_{\text{true}} = \varepsilon^{\rho\sigma\tau} x_\sigma\dot{x}_\tau.
\label{eqn:relativistic_angular_momentum_true}
\end{equation}
Again, we want to learn the Lagrangian $\mathcal{L}_{\text{NN}}(q,\dot{q})$ which is represented by a neural network. We drop the dependency on $\tau$, since we found that including it leads to instabilities during training. The reason for this is that for the true solution $\tau$ grows (approximately linearly) with time. While the neural network will learn that the coefficient that multiplies $\tau$ in the Lagrangian is small, it will always remain non-zero, thus leading to a large, unphysical contribution for $\tau\gg 1$. Note that dropping $\tau$ from the inputs is consistent with the true Lagrangian in Eq. \eqref{eqn:schwarzschild_lagrangian_II}, which only depends on $\dot{\tau}$ but not the $\tau$ itself. The symmetry enforcing layer takes as input the two vectors $q,\dot{q}\in\mathbb{R}^4$ and it will return the four scalars $\dot{\tau}$, $x^2$, $\dot{x}^2$ and $x\cdot\dot{x}$, as shown in Fig. \ref{fig:rotational_invariant_layer_schwarzschild}.
\begin{figure}
\begin{center}
\includegraphics[width=0.375\linewidth]{.//symmetry_enforcing_layer_schwarzschild.pdf}
\caption{Layer $R$ enforcing rotational invariance of the spatial part of the Schwarzschild problem discussed in Section \ref{sec:schwarzschild_problem}.}
\label{fig:rotational_invariant_layer_schwarzschild}
\end{center}
\end{figure}
With this symmetry enforcing layer the dynamics generated by the neural network Lagrangian conserves the three quantities
\begin{equation}
L^{(\rho)}_{\text{NN}} = \varepsilon^{\rho\sigma\tau}x_\sigma\frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}_\tau}
\qquad\text{for $\rho=1,2,3$}\label{eqn:relativistic_angular_momentum_NN}
\end{equation}
exactly.
\subsubsection{Two interacting particles in $D$ dimensions}\label{sec:two_particle_problem}
Finally, we consider a system of two non-relativistic particles with masses $m_1$ and $m_2$ that move in $D$-dimensional space and interact via a potential that is invariant under translations and rotations. Setting $d=2D$, the $d$-dimensional state vector $q=(x^{(1)},x^{(2)})\in\mathbb{R}^d$ contains the two particle positions $x_1,x_2\in\mathbb{R}^D$ with corresponding velocities $\dot{x}^{(2)},\dot{x}^{(2)}\in\mathbb{R}^D$. The true Lagrangian is
\begin{equation}
\mathcal{L}_{\text{true}}(q,\dot{q}) = \mathcal{L}_{\text{true}}(x^{(1)},x^{(2)},\dot{x}^{(1)},\dot{x}^{(2)})
= \frac{m_1}{2}||\dot{x}^{(1)}||_2^2 + \frac{m_2}{2}||\dot{x}^{(1)}||_2^2 - V(||x^{(1)}-x^{(2)}||_2)\label{eqn:true_lagrangian_two_particle}
\end{equation}
with the interaction potential $V(r)$ given by the double-well function
\begin{equation}
V(r) = \frac{\mu}{2}r^2 - \frac{\kappa}{4}r^4\label{eqn:double_well}
\end{equation}
for some positive constants $\mu,\kappa>0$. As usual, $||\cdot||_2$ denotes the Euclidean two-norm
\begin{equation}
||z||_2 := \left(\sum_{j=1}^{D}z_j^2\right)^{1/2}\qquad\text{for $\in\mathbb{R}^D$}.
\end{equation}
Again, the goal is to learn the neural network Lagrangian
\begin{equation}
\mathcal{L}_{\text{NN}}(q,\dot{q}) = \mathcal{L}_{\text{NN}}(x^{(1)},x^{(2)},\dot{x}^{(1)},\dot{x}^{(2)})
\label{eqn:nn_lagrangian_two_particle}
\end{equation}
which approximates the dynamics of the system, while taking into a account the symmetries of the problem. Although $\mathcal{L}_{\text{NN}}$ in Eq. \eqref{eqn:nn_lagrangian_two_particle} depends on $4D$ unknowns, it can be restricted considerably by assuming that the system is invariant under translations and/or rotations. For this, we consider the following two continuous transformations of the entire system:
\begin{description}
\item[Translations] $h^s_T$ with a constant vector $\Delta\in\mathbb{R}^D$
\begin{equation}
h^s_T: q = (x^{(1)}, x^{(2)})\mapsto (x^{(1)}+s\Delta,x^{(2)}+s\Delta)
\end{equation}
\item[Rotations] $h^s_R$ with a constant rotation matrix $R=\exp[s\Gamma]\in SO(D)$
\begin{equation}
h^s_R: q = (x^{(1)}, x^{(2)})\mapsto (R(s)x^{(1)},R(s)x^{(2)}),
\end{equation}
where the antisymmetric $D\times D$ matrix $\Gamma$ belongs to the fundamental representation of the Lie algebra $\mathfrak{so}(D)$ of the special orthogonal group $SO(D)$.
\end{description}
Assuming that the physics of the system is invariant under translations and/or rotations dramatically restricts the possible form of the Lagrangian $\mathcal{L}_{\text{NN}}$. For example, for a rotationally invariant (but not necessarily translationally invariant) system the Lagrangian can only depend on dot-products of pairs of the four dynamical variables $x^{(1)},x^{(2)},\dot{x}^{(1)},\dot{x}^{(2)}\in\mathbb{R}^D$ and contractions of these variables with the antisymmetric Levi-Civita symbol $\varepsilon$ in $D$ dimensions. More generally, for a set of $n$ vectors $A=\{a^{(1)},a^{(2)},\dots,a^{(n)}\}$ with $a^{(j)}\in\mathbb{R}^D$ we denote by $\mathscr{R}(A)$ the set of all rotationally invariant scalars; a detailed discussion of the construction of this set can be found in \ref{sec:rotationally_invariant_combinations}. For example we have for the three vectors $u,v,w\in\mathbb{R}^3$:
\begin{equation}
\mathscr{R}(\{u,v,w\}) = \{u^2,v^2,w^2,u\cdot v,u\cdot w,v\cdot w,u\cdot(v\times w)\}
\end{equation}
since there are six different scalar products and
\begin{equation}\varepsilon^{ijk}u_iv_jw_k = u\cdot (v\times w)
\end{equation}
is the only non-vanishing contraction with the three-dimensional Levi-Civita symbol in this case.
With this notation, the set of scalar invariants that are output by the symmetry-enforcing layer $S$ can be written down as in Tab. \ref{tab:lagrangian_restrictions}.
\begin{table}
\begin{center}
\begin{tabular}{|cc|c|c|}
\hline
& & \multicolumn{2}{c|}{rotationally invariant?} \\
& & no & yes \\
\hline
translationally & no & $\{x^{(1)},x^{(2)},\dot{x}^{(1)},\dot{x}^{(2)}\}$ & $R(q,\dot{q}) = \mathscr{R}(\{x^{(1)},x^{(2)},\dot{x}^{(1)},\dot{x}^{(2)}\})$ \\
\cline{2-4}
invariant? & yes & $T(q,\dot{q}) = \{x^{(1)}-x^{(2)},\dot{x}^{(1)},\dot{x}^{(2)}\}$ &
$(R\circ T)(q,\dot{q}) = \mathscr{R}(\{x^{(1)}-x^{(2)},\dot{x}^{(1)},\dot{x}^{(2)}\})$ \\
\hline
\end{tabular}
\caption{Output of the symmetry-enforcing layer $S$.}
\label{tab:lagrangian_restrictions}
\end{center}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[width=0.9\linewidth]{.//symmetry_enforcing_layers.pdf}
\caption{Pictorial representation of the symmetry enforcing layers.}
\label{fig:symmetry_enforcing_layers}
\end{center}
\end{figure}
While the most general Lagrangian is a function of $4D$ unknowns, translational invariance reduces it to a function which only depends on $3D$ variables. A Lagrangian which is invariant under rotations (but not necessarily translations) depends on $10+{4\choose D}$ unknowns; assuming both translational and rotational invariance reduces this further to only $6+{3\choose D}$ variables.
To derive the conserved quantities according to Eq. \eqref{eqn:noether_conserved_I}, observe that the $D$ generators of the translation group are the vectors $\Delta^{(\alpha)}\in\mathbb{R}^D$ with components
\begin{equation}
\Delta^{(\alpha)}_j = \delta_{\alpha j}=\begin{cases}
1 & \text{if $\alpha=j$} \\
0 & \text{otherwise}
\end{cases}
\qquad\text{for $1\le\alpha\le D$}
\end{equation}
where $\delta_{jk}$ is the Kronecker-$\delta$. The $\frac{1}{2}D(D-1)$ generators $\Gamma^{(\rho,\sigma)}$ of the rotation group are the antisymmetric $D\times D$ matrices with entries
\begin{equation}
\Gamma^{(\rho,\sigma)}_{jk} = \delta_{\rho j}\delta_{\sigma k} - \delta_{\rho k}\delta_{\sigma j}\qquad\text{for $1\le\rho<\sigma\le D$}.
\end{equation}
Using Noether's Theorem, this leads to the following $\frac{1}{2}D(D+1)$ conserved quantities $M^{(\alpha)}_{\text{NN}}$ and $J^{(\rho,\sigma)}_{\text{NN}}$:
\begin{equation}
\begin{aligned}
M^{(\alpha)}_{\text{NN}} & = \frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}^{(1)}_\alpha} + \frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}^{(2)}_\alpha}\qquad\text{for $1\le\alpha\le D$}, \\
L^{(\rho,\sigma)}_{\text{NN}} & = \frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}^{(1)}_\rho} x_\sigma^{(1)}-\frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}^{(1)}_\sigma} x_\rho^{(1)} + \frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}^{(2)}_\rho} x_\sigma^{(2)}-\frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}^{(2)}_\sigma} x_\rho^{(2)}\qquad\text{for $1\le\rho < \sigma\le D$}.
\end{aligned}
\label{eqn:two_particle_NN_momenta}
\end{equation}
For the true Lagrangian in Eq. \eqref{eqn:true_lagrangian_two_particle} this simplifies to the usual linear momenta $M^{(\alpha)}_{\text{true}}$ and (generalised) angular momenta $L^{(\rho,\sigma)}_{\text{true}}$ with
\begin{equation}
\begin{aligned}
M^{(\alpha)}_{\text{true}} & = m_1\dot{x}^{(1)}_\alpha + m_2\dot{x}^{(2)}_\alpha\qquad\text{for $1\le\alpha\le D$}, \\
L^{(\rho,\sigma)}_{\text{true}} & = m_1\left(\dot{x}^{(1)}_\rho x_\sigma^{(1)}-\dot{x}^{(1)}_\sigma x_\rho^{(1)}\right) + m_2\left(\dot{x}^{(2)}_\rho x_\sigma^{(2)}-\dot{x}^{(2)}_\sigma x_\rho^{(2)}\right)\qquad\text{for $1\le\rho < \sigma\le D$}.
\end{aligned}\label{eqn:momentum_true_twoparticle}
\end{equation}
For future reference we collect the conserved quantities in the following vectors
\begin{equation}
\begin{aligned}
M_{\text{NN}} & = \left(M_{\text{NN}}^{(1)},M_{\text{NN}}^{(2)},\dots,M_{\text{NN}}^{(D)}\right)^T\in\mathbb{R}^D, \\
L_{\text{NN}} & = \left(L_{\text{NN}}^{(1,2)},L_{\text{NN}}^{(1,3)},\dots,M_{\text{NN}}^{(D-1,D)}\right)^T\in\mathbb{R}^{\frac{1}{2}D(D-1)}, \\
M_{\text{true}} & = \left(M_{\text{true}}^{(1)},M_{\text{true}}^{(2)},\dots,M_{\text{true}}^{(D)}\right)^T\in\mathbb{R}^D, \\
L_{\text{true}} & = \left(L_{\text{true}}^{(1,2)},L_{\text{true}}^{(1,3)},\dots,M_{\text{true}}^{(D-1,D)}\right)^T\in\mathbb{R}^{\frac{1}{2}D(D-1)}.
\end{aligned}
\label{eqn:two_particle_momentum_vectors}
\end{equation}
\section{Results}\label{sec:results}
\subsection{Implementation and hyperparameters}\label{sec:implementation}
The code that was used to obtain the numerical results reported in this section has been implemented in tensorflow \cite{Tensorflow2015} and is freely available at the following URL, which also contains instructions on how to install and run the code:
\begin{center}
\url{https://github.com/eikehmueller/mlconservation_code}
\end{center}
The results reported in this paper were generated with the release published at \cite{code_release}. In all cases the Lagrangian is represented by two hidden dense layers $D_1$, $D_2$ with $n_{\text{h}}=128$ output units each and a final dense layer $D_3$ with a single output unit. This unit does not have a bias term since adding a constant to the Lagrangian does not change the equations of motion. A softplus activation function is used in all cases. The layer weights are initialised with a random normal distribution similarly to \cite{Greydanus2019}: the standard deviation of the normal distribution is set to $2/\sqrt{n_{\text{h}}}$ for the first hidden layer $D_1$, $1/\sqrt{n_{\text{h}}}$ for the second hidden layer $D_2$ and $\sqrt{n_{\text{h}}}$ for the output layer $D_3$; the biases of all layers are initialised to zero. Single precision arithmetic is used in all numerical experiments.
For all three systems discussed in Section \ref{sec:model_systems} the networks are trained over 2500 epochs with 100 steps per epoch and a batch size of $B=128$ using the Adam optimiser. The training schedule is a cosine decay, starting with a learning rate of $10^{-3}$ which is reduced to $10^{-5}$ at the end of the training cycle. We find that this reduces the MSE to $10^{-5}-10^{-6}$ for the single-particle problem described in Section \ref{sec:kepler_problem}, to $10^{-6}$ for the relativistic particle in Section \ref{sec:schwarzschild_problem} and to approximately $10^{-4}$ for the two-particle system in Section \ref{sec:two_particle_problem}. This is consistent with the chosen random noise on the training data, which is $\sigma=10^{-3}$ in all cases and therefore limits the minimal achievable MSE to the order of $\sigma^2\sim 10^{-6}$. The full loss histories can be found in \ref{sec:loss_histories}.
\subsubsection{Motion of a single particle in a gravitational potential}\label{sec:results_single_particle}
Both the mass of the particle and the strength of the gravitational potential in Eq. \eqref{eqn:true_lagrangian_single_particle} are set to $m=\alpha=1$. To train the model, we generate pairs of inputs $X^{(j)} = (q_{\text{exact}}(\varphi^{(j)})+\sigma \xi_1^{(j)},\dot{q}_{\text{exact}}(\varphi^{(j)})+\sigma\xi_2^{(j)})\in\mathbb{R}^6$ and ground truth $\widehat{y}^{(j)} = \ddot{q}_{\text{exact}}(\varphi^{(j)})+\sigma\xi_3^{(j)}\in\mathbb{R}^3$ where $q_{\text{true}}(\varphi)$ is the exact solution as a function of the angle azimuthal angle $\varphi$. The size of the noise is characterised by $\sigma=10^{-3}$ as discussed above. The angles $\varphi^{(j)}\sim\text{Uniform}(-\pi,+\pi)$ are uniformly distributed while $\xi^{(j)}_{1,2,3}\sim\mathcal{N}(0,1)$ are drawn from a normal distribution with mean zero and unit variance. We assume that the motion takes place entirely in the $x-y$ plane and the true trajectory is an ellipse with eccentricity $\varepsilon_{\text{ecc}}=0.8$. The initial conditions are chosen such that the vertical component of the angular momentum is $L_z=1$.
Fig. \ref{fig:trajectories_kepler} shows the trajectories predicted with the trained neural network Lagrangian. The equations of motion are integrated up to the final time $T=128$ with a fourth order Runge Kutta (RK4) method and a timestep size of $\Delta t=10^{-2}$. The figure also shows a trajectory for which both initial position and initial velocity are perturbed by normal random noise with a standard deviation of $10^{-3}$. As can be seen from the top figure, the trajectory obtained with the neural network Lagrangian deviates significantly from the true solution (the red ellipse) for the unconstrained Lagrangian and the two trajectories with perturbed initial conditions diverge. In fact, the neural network trajectories become unstable. Radically different behaviour is observed for the neural network with built-in rotational invariance (bottom plot in Fig. \ref{fig:trajectories_kepler}). Visually the trajectories obtained with the neural network Lagrangian can not be distinguished from the true solution. In particular, the trajectories appear to be confined to the $x-y$ plane, which implies that the two corresponding components of the angular momentum indeed remain close to zero.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//trajectories_kepler.pdf}\\
\includegraphics[width=\linewidth]{.//trajectories_kepler_rot.pdf}
\caption{Trajectories for the motion of a particle in a gravitational potential described in Section \ref{sec:kepler_problem} without (top) and with (bottom) constraints on the neural network to enforce rotational invariance of the Lagrangian. In each case, a trajectory that is obtained by perturbing the initial conditions by $\sim10^{-3}$ is also shown as a dashed curve. The ellipse that represents the true solution is marked with red dots.}
\label{fig:trajectories_kepler}
\end{center}
\end{figure}
To investigate the conservation of angular momentum further, we compute the time evolution of the deviation of the angular momentum from its original value. For this we consider the deviation
\begin{xalignat}{2}
\delta L_{\text{NN}}(t) & = \frac{\left|\left|L_{\text{NN}}(t)-L_{\text{NN}}(0)\right|\right|_2}{\left|\left|L_{\text{NN}}(0)\right|\right|_2}, &
\delta L_{\text{true}}(t) & =\frac{\left|\left|L_{\text{true}}(t)-L_{\text{true}}(0)\right|\right|_2}{\left|\left|L_{\text{true}}(0)\right|\right|_2}.
\label{eqn:deltaL_definition}
\end{xalignat}
of the angular momentum vectors defined by Eqs. \eqref{eqn:angular_momentum_NN}, \eqref{eqn:angular_momentum_true} and \eqref{eqn:angular_momentum_kepler_vector} from their initial values at time $t=0$. Exact conservation of angular momentum would correspond to $\delta L_{\text{true}}(t) = 0$ for all times $t$. As Fig. \ref{fig:conservation_kepler} (top) shows, for the unconstrained network $\delta L_{\text{true}}(t)$ is small initially, since the network has learned some degree of angular momentum conservation from the data, but then increases to around $1$ by time $t\approx 50$ and diverges shortly after that. As expected, for the rotationally invariant network shown in the bottom figure $\delta L_{\text{NN}}(t)$ remains zero within single precision rounding errors. More importantly, for the rotationally invariant neural network the deviation $\delta L_{\text{true}}(t)$ of the true angular momentum never exceeds $10^{-3}$. Hence, even though mathematically only the conservation of $L_{\text{NN}}$ can be guaranteed, numerically this appears to also help with the conservation of $L_{\text{true}}$. This behaviour is similar to the approximate conservation of the true energy observed for symplectic integrators of time-independent Hamiltonian systems, which is related to the exact conservation of the shadow Hamiltonian.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//invariants_kepler.pdf}
\caption{Conservation of angular momentum for the motion of a particle in a gravitational potential defined in Section \ref{sec:kepler_problem} without (top) and with (bottom) constraints on the neural network to enforce rotational invariance of the Lagrangian. In each case the deviation $\delta L_{\text{NN}}(t)$ and $\delta L_{\text{true}}(t)$ of the absolute value of the total three dimensional angular momentum vectors are shown as a function of time. Both the ``neural network'' angular momentum $L_{\text{NN}}(t)$ defined in Eq. \eqref{eqn:angular_momentum_NN} and the true angular momentum $L_{\text{true}}(t)$ defined in Eq. \eqref{eqn:angular_momentum_true} are considered.}
\label{fig:conservation_kepler}
\end{center}
\end{figure}
Finally, in Fig. \ref{fig:pertubation_kepler} we show the distance $\delta q_{\text{NN}}(t) = ||q_{\text{NN}}(t)-q_{\text{NN}}^{(\text{perturbed})}(t)||_2$ between the position vectors of the two neural network trajectories that only differ by a $\sim10^{-3}$ perturbation of the initial condition. While -- as expected -- the trajectories diverge over time, for the rotationally invariant neural network they stay much closer together and for the considered time interval their distance stays below $0.1$. Intuitively, the reason for this is that rotational invariance limits the allowed trajectories to a smaller sub-manifold, so there is less ``room'' for nearby trajectories to diverge.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//perturbation_kepler.pdf}
\caption{Evolution of the distance $\delta q_{\text{NN}}(t)= ||q_{\text{NN}}(t)-q_{\text{NN}}^{\text{(perturbed)}}(t)||_2$ between two trajectories obtained with slightly different initial conditions for the Kepler problem. The initial conditions that differ by $\delta q_{\text{NN}}(0)\sim 10^{-3}$.}
\label{fig:pertubation_kepler}
\end{center}
\end{figure}
\subsubsection{Motion of a massive relativistic particle in the Schwarzschild metric}
For the motion of a massive relativistic particle in the Schwarzschild metric as defined in Section \ref{sec:schwarzschild_problem} we set the Schwarzschild radius to $r_s=0.1$. In this case, the ``true'' trajectory $q_{\text{exact}}(t)$ is obtained by integrating the exact equations of motion, which can be obtained from the true Lagrangian in Eq. \eqref{eqn:schwarzschild_lagrangian_II}, with a RK4 integrator with a timestep size of $\Delta t=10^{-2}$. As above, normally distributed random noise with a standard deviation of $\sigma=10^{-3}$ is added to obtain training samples $(X^{(j)},\widehat{y}^{(j)})$ with $X^{(j)} = (q_{\text{exact}}(t^{(j)})+\sigma \xi_1^{(j)},\dot{q}_{\text{exact}}(t^{(j)})+\sigma\xi_2^{(j)})\in\mathbb{R}^8$ and ground truth $\widehat{y}^{(j)} = \ddot{q}_{\text{exact}}(t^{(j)})+\sigma\xi_3^{(j)}\in\mathbb{R}^4$, where $\xi_{1}^{(j)}, \xi_{2}^{(j)}, \xi_{3}^{(j)}\sim \mathcal{N}(0,1)$ and $t^{(j)}$ are sample times along the true trajectory.
Fig. \ref{fig:trajectories_schwarzschild} shows the trajectories predicted with the trained neural network Lagrangian up to the final time $T=1000$. Again, a RK4 integrator with a timestep size $\Delta t=10^{-2}$ is used. The trajectory obtained with a slightly perturbed initial condition is also shown as a dashed line. In contrast to the non-relativistic equivalent, Fig. \ref{fig:trajectories_schwarzschild} (top) shows that for the unconstrained network the trajectories do not become unstable. However, they still diverge strongly from the true solution (shown as a dotted red line) and they oscillate about the $x-y$ plane, which indicates that angular momentum is not conserved. Perturbing the initial conditions also leads to very different trajectories at later times. Fig. \ref{fig:trajectories_schwarzschild} (bottom) demonstrates that the picture is fundamentally different for the rotationally invariant Lagrangian neural network. Here the trajectory generated with the neural network stays very close to the true solution (which is indeed almost completely hidden beneath the solid blue curve) and the two trajectories with slightly differing initial conditions only diverge slowly at later times. Furthermore, the motion appears to be completely confined to the $x-y$ plane, which indicates that the two corresponding components of the angular momentum are conserved.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//trajectories_schwarzschild.pdf}\\
\includegraphics[width=\linewidth]{.//trajectories_schwarzschild_rot.pdf}
\caption{Trajectories for the motion of a relativistic massive particle in the Schwarzschild metric as defined in Section \ref{sec:schwarzschild_problem} without (top) and with (bottom) constraints on the neural network to enforce rotational invariance of the spatial part of the Lagrangian. In each case, a trajectory that is obtained by perturbing the initial conditions by $\sim10^{-3}$ is also shown as a dashed curve. The true solution is shown as a dotted red curve.}
\label{fig:trajectories_schwarzschild}
\end{center}
\end{figure}
Again, we investigate the conservation of angular momentum quantitatively by plotting the time evolution of the two-norms $\delta L_{\text{NN}}(t)$ and $\delta L_{\text{true}}(t)$ in Eq. \eqref{eqn:deltaL_definition}, now using the expressions for $L_{\text{NN}}$ and $L_{\text{true}}$ defined via Eqs. \eqref{eqn:relativistic_angular_momentum_true} and \eqref{eqn:relativistic_angular_momentum_NN}. As Fig. \ref{fig:conservation_schwarzschild} shows, the unconstrained network is able to learn the conservation of angular momentum to some degree, but the true angular momentum $L_{\text{true}}(t)$ deviates from its initial value by around $10\%$ at later times. This is consistent with the fact that the trajectories in Fig. \ref{fig:trajectories_schwarzschild} (top) are visibly not constrained to the $x-y$ plane, but do not stray too far from it either. The neural network with built-in rotational invariance, on the other hand, is able to reduce the relative deviation $\delta L_{\text{true}}(t)$ from the initial angular momentum to less than $10^{-3}$. As expected, the quantity $L_{\text{NN}}$, which would be zero in exact arithmetic and for an exact time integrator, is very small and never exceeds a value of around $10^{-5}$, which is consistent with (accumulated) rounding errors in single precision arithmetic.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//invariants_schwarzschild.pdf}
\caption{Conservation of angular momentum for the motion of a massive relativistic particle in the Schwarzschild metric as defined in Section \ref{sec:schwarzschild_problem} without (top) and with (bottom) constraints on the neural network to enforce rotational invariance of the Lagrangian. In each case the deviation $\delta L_{\text{NN}}(t)$ and $\delta L_{\text{true}}(t)$ of the absolute value of the total three dimensional angular momentum vectors are shown as a function of time. Both the ``neural network'' angular momentum $L_{\text{NN}}(t)$ defined in Eq. \eqref{eqn:relativistic_angular_momentum_NN} and the true angular momentum $L_{\text{true}}(t)$ defined in Eq. \eqref{eqn:relativistic_angular_momentum_true} are considered.}
\label{fig:conservation_schwarzschild}
\end{center}
\end{figure}
Fig. \ref{fig:pertubation_schwarzschild} shows the distance $\delta q_{\text{NN}}(t) = ||q_{\text{NN}}(t)-q_{\text{NN}}^{\text{(perturbed)}}(t)||_2$ between the two neural network trajectories with slightly perturbed initial conditions. The plot confirms that the rotationally invariant Lagrangian is much more robust under perturbations of the initial conditions. Compared to the unconstrained Lagrangian neural network, the distance between the perturbed and the unperturbed trajectories is around one order of magnitude smaller and grows only moderately over time.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//perturbation_schwarzschild.pdf}
\caption{Evolution of the distance $\delta q_{\text{NN}}(t)=||q_{\text{NN}}(t)-q_{\text{NN}}^{\text{(perturbed)}}(t)||_2$ between two trajectories obtained with slightly different initial conditions for the motion of a massive relativistic particle in the Schwarzschild metric. The initial conditions that differ by $\delta q_{\text{NN}}(0)\sim 10^{-3}$.}
\label{fig:pertubation_schwarzschild}
\end{center}
\end{figure}
\subsubsection{Two interacting particles in $D=4$ dimensions}
Finally, we consider the system of two interacting particles described in Section \ref{sec:two_particle_problem} for $D=4$ dimensions. The masses of the two particles are set to $m_1=1$ and $m_2=0.8$, while the parameters of the double well potential in Eq. \eqref{eqn:double_well} are fixed to $\mu=\kappa=1$. The initial condition was chosen such that at time $t=0$ the total linear momentum $M=m_1x^{(1)}+m_2x^{(2)}$ is zero. We consider four setups, corresponding to the constraints on the Lagrangian listed in the four quadrants of Tab. \ref{tab:lagrangian_restrictions}:
\begin{enumerate}
\item Unconstrained Lagrangian
\item Rotationally invariant Lagrangian
\item Translationally invariant Lagrangian
\item Rotationally and translationally invariant Lagrangian
\end{enumerate}
Since (except in the first case) the input to the neural network is first passed through a symmetry-enforcing layer, the number of inputs to the first hidden layer $D_1$ depends on which symmetries we assume for the Lagrangian, as shown in Tab. \ref{tab:n_inputs}. Note that as we do not enforce invariance under reflections (i.e. we only consider the $SO(4)$ subgroup instead of the full $O(4)$ group), in the case of a rotationally (but not necessarily translationally) invariant Lagrangian, the contraction $\varepsilon^{\alpha\beta\rho\sigma}x_{\alpha}^{(1)}x_{\beta}^{(2)}\dot{x}_{\rho}^{(1)}\dot{x}_{\sigma}^{(2)}$ with the Levi-Civita symbol is included in the set $\mathscr{R}\left(\{x^{(1)},x^{(2)},\dot{x}^{(2)},\dot{x}^{(2)}\}\right)$.
\begin{table}
\begin{center}
\begin{tabular}{|cc|>{\centering\arraybackslash}p{12ex}|>{\centering\arraybackslash}p{12ex}|}
\hline
& & \multicolumn{2}{c|}{rotationally invariant?} \\
& & no & yes \\
\hline
translationally & no & 16 & 11 \\
\cline{2-4}
invariant? & yes & 12 & 6 \\
\hline
\end{tabular}
\caption{Number of scalar inputs to the first dense layer $D_1$ of the Lagrangian neural network.}
\label{tab:n_inputs}
\end{center}
\end{table}
As in the other two cases, we generate synthetic training samples $(X^{(j)},\widehat{y}^{(j)})\in \mathbb{R}^{16}\times \mathbb{R}^{8}$ by integrating the true equations of motion obtained from the Lagrangian in Eq. \eqref{eqn:true_lagrangian_two_particle} with a RK4 integrator ($\Delta t=10^{-2}$) and adding normally distributed noise with a standard deviation of $\sigma=10^{-3}$.
Fig. \ref{fig:trajectories_two_particle} shows a projection of the four-dimensional trajectories onto the first two dimensions for each of the four considered setups. The true solution (dashed) and the neural network solution (solid) are shown for both particles with the final positions at time $T=8$ marked by circles ($\medblackcircle$/$\medcircle$). Visually, the completely unconstrained network gives the worst solution. While enforcing rotational or translation invariance improves this somewhat, the best qualitative agreement is achieved with the neural network Lagrangian that is both rotationally and translationally invariant. Although even in this case the final positions have a distance of order 1, this appears to be mainly attributed to the fact that the neural network solution lags behind the true solution. Ignoring this phase error, the trajectories generated with the neural network Lagrangian show good agreement with the true solution.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//trajectories_two_particle_4d.pdf}\\
\caption{Trajectories for the two-particle problem assuming different constraints on the neural network Lagrangian as laid out in Tab. \ref{tab:lagrangian_restrictions}. In each case, both the true trajectories (dashed lines) and those obtained with the neural network Lagrangian (solid lines) are shown for the two particles. The starting positions are denoted with a star ($\medblackstar$) and the final positions are marked by circles ($\medblackcircle$/$\medcircle$). The figures shows a projection on the first two coordinates of four dimensional space.}
\label{fig:trajectories_two_particle}
\end{center}
\end{figure}
To quantify the conservation of the linear- and angular momentum, we again plot the time evolution of $\delta L_{\text{NN}}(t)$ and $\delta L_{\text{true}}$ as defined in Eq. \eqref{eqn:deltaL_definition}, but now with $L_{\text{NN}}$ and $L_{\text{true}}$ being the six-dimensional vectors defined in Eq. \eqref{eqn:two_particle_momentum_vectors}. In analogy to Eq. \eqref{eqn:deltaL_definition}, we define the time dependent deviation of the linear momentum as
\begin{xalignat}{2}
\delta M_{\text{NN}}(t) & =\frac{\left|\left|M_{\text{NN}}(t)-M_{\text{NN}}(0)\right|\right|_2}{\left|\left||\widetilde{M}_{\text{NN}}(0)\right|\right|_2}, &
\delta M_{\text{true}}(t) & =\frac{\left|\left|M_{\text{true}}(t)-M_{\text{true}}(0)\right|\right|_2}{\left|\left|\widetilde{M}_{\text{true}}(0)\right|\right|_2}.
\end{xalignat}
where $M_{\text{NN}}$ and $M_{\text{true}}$ are the four-dimensional linear momentum vectors given in Eq. \eqref{eqn:two_particle_momentum_vectors}. Since the initial conditions are chosen such that the total linear momentum is zero, we cannot normalise by $\left|\left|M_{\text{NN}}(0)\right|\right|_2$ and $\left|\left|M_{\text{true}}(0)\right|\right|_2$ but instead divide by the two-norm of the linear momentum of the first particle defined as
\begin{xalignat}{2}
\widetilde{M}_{\text{NN}} &= \frac{\partial \mathcal{L}_{\text{NN}}}{\partial \dot{x}^{(1)}}, &
\widetilde{M}_{\text{true}} &= m^{(1)}\dot{x}^{(1)}.
\end{xalignat}
While the unconstrained network is able to learn the conservation of $L_{\text{true}}$ and $M_{\text{true}}$ to some degree, at the final time both quantities have deviated from their initial values by $10\%-100\%$. The picture is somewhat better once either translational or rotational invariance is enforced. In particular rotational invariance limits both the relative deviation of both the linear and angular momentum to below $\approx 10\%$. As expected, the linear and angular momenta $M_{\text{NN}}$ and $L_{\text{NN}}$ derived from the neural network Lagrangian (see Eq. \eqref{eqn:two_particle_NN_momenta}) are conserved up to rounding errors caused by single precision arithmetic. Enforcing both rotational and translation invariance improves the conservation of the true linear and angular momentum substantially, with both now not deviating by more than $1\%$ from their initial values. In contrast to the other cases, there also does not appear to be any upwards trend for $\delta M_{\text{true}}(t)$ or $\delta L_{\text{NN}}(t)$ as time increases.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//invariants_two_particle_4d.pdf}
\caption{Conservation of linear and angular momentum for the four-dimensional two particle problem described in Section \ref{sec:two_particle_problem} with different constraints on the neural network to enforce rotational and/or translational invariance of the Lagrangian. In each case the relative deviations $\delta L_{\text{NN}}(t)$, $\delta M_{\text{NN}}(t)$, $\delta L_{\text{true}}(t)$ and $\delta M_{\text{true}}(t)$ of the linear and angular momentum vectors in Eq. \eqref{eqn:two_particle_momentum_vectors} are shown as a function of time. Both the ``neural network'' momenta defined in Eq. \eqref{eqn:two_particle_NN_momenta} and the true momenta in Eq. \eqref{eqn:momentum_true_twoparticle} are considered.}
\label{fig:conservation_two_particle}
\end{center}
\end{figure}
Finally, we study the stability of the neural network trajectories under small perturbations of the initial conditions. For this, in each of the four cases the initial position and velocity were perturbed by the same normal noise with standard deviation $10^{-3}$. The distance $\delta q_{\text{NN}}(t) = ||q_{\text{NN}}(t)-q_{\text{NN}}^{(\text{perturbed})}(t)||_2$ between the trajectories with these two different initial conditions is visualised in Fig. \ref{fig:pertubation_two_particle}. As can be seen from this plot, enforcing rotational and translational invariance reduces the growth of the perturbation significantly. However, only enforcing either rotational or translational invariance seems to be almost as good, with the latter performing even slightly better at large times $t$.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{.//perturbation_two_particle_4d.pdf}
\caption{Evolution of the distance $\delta q_{\text{NN}}(t)= ||q_{\text{NN}}(t)-q_{\text{NN}}^{\text{(perturbed)}}(t)||_2$ between two trajectories obtained with slightly different initial conditions for the two-particle problem. The initial conditions that differ by $\delta q_{\text{NN}}(0)\sim 10^{-3}$.}
\label{fig:pertubation_two_particle}
\end{center}
\end{figure}
\section{Conclusion}\label{sec:conclusion}
In this paper we showed how the Lagrangian neural network approach in \cite{Cranmer2020} can be combined with Noether's Theorem to enforce conservation laws while learning the dynamics of a physical system from noisy data. As the numerical results demonstrate, this generates significantly more realistic trajectories since the network does not have to learn the underlying conservation laws from data. As expected, for each symmetry of the Lagrangian the neural network integrators conserve a quantity up to rounding errors from single precision arithmetic. While the corresponding invariant quantity of the true Lagrangian is not conserved exactly, its deviations from the initial value are significantly reduced. Using a Lagrangian with built-in symmetries also improves the stability of the solution in the sense that the system is less sensitive to small perturbations of the initial conditions.
There are several avenues for extending the work presented here. To demonstrate the principle ideas behind our approach we assumed that essentially unlimited (synthetic) data with relatively small errors is available for training. It would be interesting to instead train the Lagrangian neural networks on measured experimental data for real physical systems. While we already assumed that the training data is perturbed by random noise, in real-life situations measured data might be corrupted or only projections of the state space vectors might be available. For example, in celestial dynamics usually only motion perpendicular to the line of sight of the observer can be measured. More generally, it would be interesting to further explore the how the symmetry enforcing layers improve the solution for limited data with larger errors. So far, we considered Lagrangians that describe the motion of point particles. However, the Lagrangian formalism can also be extended to continuous systems such as (classical) field theories and it would be interesting to see how our approach works in this case. For example, in \cite{Cranmer2020} a discretised one-dimensional wave equation is studied and the authors show that the Lagrangian approach leads to the approximate conservation of energy.
\section*{Acknowledgements}
The author would like to thank Patrick Lavelle and Tony Shardlow for useful discussions during the early stages of this project.
\FloatBarrier
|
\section{Kinetic energy length-scale} \label{appendixB}
We compute the dominant scale of convection as the peak of the time-averaged poloidal kinetic energy spectra \citep[][]{SGA19,SGA21}, defined as
\begin{equation} \label{eq:lpeak}
\ell_\mathrm{peak} = \text{argmax}(E_{K,P}(\ell)).
\end{equation}
Figure~\ref{fig:ekinspec} shows examples of poloidal kinetic energy spectra for one dipolar case (red line) and one multipolar case (purple line). The degree at which the spectra is maximum, $\ell_\mathrm{peak}$, is indicated by a dashed vertical line. These reference dipole and multipole models feature convective flows with similar dominant length scale. Considering the entire set of simulations, we find $\ell_\mathrm{peak}$ ranging from 14 to 45 with a median value of 30.
\begin{figure}
\centering
{\includegraphics[width=\columnwidth]{figs/ekinspec.pdf}}
\caption{Time-averaged dimensionless poloidal kinetic energy spectra for the dipolar (red solid line) and multipolar (purple solid line) cases given in Fig.~\ref{fig:FBspec}. Shaded areas correspond to one standard deviation about the time-averaged spectra and the dashed vertical lines mark the location of the peak. }
\label{fig:ekinspec}
\end{figure}
\section{Averaging strategy} \label{appendix:dipolarity}
Figure~\ref{fig:tsFdip} illustrates the time dependence of the dipolarity (Eq.~\ref{eq:fdip}) and the dipole tilt angle ($\theta_\text{dip}$) for two simulations with $N_\rho = 1.5$ in our sample. The simulation FC10 (top panel) shows an axial-dipole that is anti-aligned with the rotation axis ($\theta_\text{dip} \sim 180\degr$) and whose field strength is stable through out the time span of the simulation. For this simulation, we find $f_\mathrm{dip} = 0.62 \pm 0.04$ when using an averaging interval $\tau_\mathrm{avg}$ that is defined as the difference between the time at the end of the run ($\tau_\text{end}$) minus a predefined initial time (represented by the blue dashed line in top plot). The bottom panel of Figure~\ref{fig:tsFdip} corresponds to the simulation FC11. The evolution of $\theta_\text{dip}$ evidences a reversing dipole with periodic switches in polarity that occur at irregular intervals of time. We find $f_\mathrm{dip} = 0.41 \pm 0.12$ when considering a large number of reversals to compute the time average (achieved after setting $\tau_\mathrm{avg} = 2.5 \tau_\lambda$).
We remind the reader that all of our simulations were initialized with a dipole of strength $\Lambda = 0.44$ and the solutions we obtained may depend on the initial conditions.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{TS_fdip_N15_dS250.pdf}
\includegraphics[width=\columnwidth]{TS_fdip_N15_dS200.pdf}
\caption{Dipolarity (black line) and tilt angle of the total dipole (purple line) as a function of time (given in units of magnetic diffusion time $\tau_\lambda$). The top plot corresponds to the simulation FC10 and the bottom one to FC11. The vertical blue line indicates the initial time used to compute the time-averaged dipolarity in the top panel. For illustrative purposes only the time-averaged window is shown in the bottom plot.
}
\label{fig:tsFdip}
\end{figure}
\section{Flow configuration} \label{appendix:flow}
It was proposed in the literature that the dipole collapse is directly linked to a arrangement in the convective flow. Two main quantities characterising the structure of convective flows in the simulations were explored:
\begin{enumerate}
\item the \textit{columnarity} $\mathcal{C}_{\omega\mathrm{z}}$ which offers a quantitative way to define columnar flows and is expressed by \begin{equation} \label{eq:columnarity}
\mathcal{C}_{\omega\mathrm{z}} = \frac{\sum_{s,\phi} \abs{\expval{\va{\omega'}\vdot\vu{e}_\mathrm{z}}_z}}{\sum_{s,\phi} \expval{\abs{\va{\omega'}}}_z} \qc
\end{equation} where $\va{\omega'}$ is the vorticity generated by the non-axisymmetric velocity field \citep{SKA12}. The summation occurs in the equatorial plane and $\avg{\cdot}_z$ represents an average in the axial direction $\vu{e}_\mathrm{z}$;
\item the relative axial helicity of the flow $\abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}}$ computed as the average of the absolute contribution from the Northern and Southern hemispheres: $\abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}} = \qty(\abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}_{\mathrm{NH}}} + \abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}_{\mathrm{SH}}})/2$, where each hemispheric contribution is given by \begin{equation} \label{eq:rhelicity}
{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}_{\mathrm{NH/SH}} = \frac{\expval{\large u_z\omega_z}_{\mathrm{NH/SH}}}{\sqrt{\expval{u_z^2}_{\mathrm{NH/SH}}\expval{\omega_z^2}_{\mathrm{NH/SH}}}}.
\end{equation}
\end{enumerate}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Columnarity.pdf}
\caption{Dipolarity as a function of the flow columnarity (top) and relative axial helicity (bottom). Symbols are defined as in Fig.~\ref{fig:fdip}.
}
\label{fig:columnarity}
\end{figure}
The top panel of Fig.~\ref{fig:columnarity} shows $f_\mathrm{dip}$ as a function of $\mathcal{C}_{\omega\mathrm{z}}$ for our data set. The overall result shows a homogeneous distribution of dipole-dominated and complex multipolar surface fields for the explored range of $\mathcal{C}_{\omega\mathrm{z}}$ (going from 0.4 to 1). It also evidences the lack of correlation between $f_\mathrm{dip}$ and $\mathcal{C}_{\omega\mathrm{z}}$. A possible explanation for this might be the high values of columnarity attained in this work. Prior Boussinesq simulations of \citet{SKA12} found that columnar flows with $\mathcal{C}_{\omega\mathrm{z}} > 0.5$ can generate either dipolar or multipolar surface magnetic fields, while flows with $\mathcal{C}_{\omega\mathrm{z}} \lesssim 0.5$ only results in multipolar fields. Indeed if we restrain ourselves to the runs with columnarity around the threshold of $0.5$, we identify three runs FC08, FC15, and FC23, giving hints of a transition to a multipolar branch (all three with $f_\mathrm{dip} < 0.25$). Nevertheless, the diversity of magnetic field complexities obtained at high–$\mathcal{C}_{\omega\mathrm{z}}$ makes the columnarity a poor proxy to describe the dipolar collapse.
Often associated with the magnetic field amplification in the dynamo framework (through the so-called $\alpha$–effect), the decrease in the flow's relative axial helicity has also been suggested to cause the dipole breakdown \citep{SKA12}. The bottom panel of Fig.~\ref{fig:columnarity} shows the dependency of $\abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}}$ with the different magnetic morphologies. The simulations yield weak to moderate relative helicity values, $\abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}} < 0.6$, that are consistent with the values obtained in previous works \citep{T14,GOD17}. It is apparent from Fig.~\ref{fig:columnarity} that the only case displaying $f_\mathrm{dip} \approx 0$ features the highest helicity in our sample. On the other hand, the strongest dipoles possess weak helicity values with $\abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}}$ spread around $0.28$ (corresponding to five dipolar dynamos obtained for $N_\rho = 1.0$ and the two strongest dipoles for $N_\rho = 3.0$). These results suggest that the magnetic morphology is unaffected by $\abs{{\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}}$ for the parameter space we explored. Although these findings differ from some published studies \citep[e.g.,][]{SKA12}, they are consistent with mean-field simulations of \citet{LHT07} and the 3D simulations of \citet{B08} mimicking the interior of a fully convective M~dwarf. The likely cause for these differences is that the mean-helicity becomes a poor approximation for the $\alpha$–effect in some cases \citep{SRS07,WRT18}.
These results corroborate earlier suggestions of \citet{GOD17}, who argued that hydrodynamic transitions in the flow (e.g. measured by $\mathcal{C}_{\omega\mathrm{z}}$ or ${\mathcal{H}^{\mathrm{rel}}_{\mathrm{z}}}$) would only capture the dipole collapse in systems where the Lorentz force plays a minor role in the flow dynamics.
\section*{Acknowledgements}
The authors wish to thank Pascal Petit and Claire Moutou for very fruitful discussions. We also thank the anonymous referees for their insightful comments and suggestions that helped to improve this manuscript. BZ and JFD acknowledge funding from the European Research Council (ERC) under the H2020 research $\&$ innovation programme (grant agreement $\#740651$ New-Worlds). LJ acknowledges funding by the Institut Universitaire de France. NL acknowledges funding by CNPq, CAPES and FAPEMIG. Numerical simulations were performed using HPC resources from GENCI-CINES (Grants 2020-A0070410970 and 2021-A0090410970) and CALMIP (Grant P19031).
\section*{Data Availability}
The authors confirm that the data supporting the findings of this study are available within the article or through requests to the corresponding author. Numerical simulations were performed with the fluid dynamics code \verb|MagIC| using the open-source library \verb|SHTns|, both publicly available at \url{https://github.com/magic-sph/magic} and \url{https://bitbucket.org/nschaeff/shtns}, respectively.
\bibliographystyle{mnras}
\section{Introduction}
Over the last decade, spectropolarimetric observations coupled to tomographic inversion techniques enabled the reconstruction of the large-scale magnetic topology that stars host at their surfaces. Cool stars with significant convective envelopes (with spectral types later than G0) revealed a large diversity of magnetic morphologies \citep{DMP08,MDP10,FPB16,FBP18}. Fully convective stars typically are found to harbour strong poloidal fields with a significant dipolar component, while partly convective stars host more complex magnetic topologies, consisting of non-axisymmetric multipolar poloidal fields and significant toroidal fields \citep{DL09}. Although there is now a consensus that the magnetism of cool stars are generated through dynamo processes occurring within the outer convective zones \citep[see][for a recent review on the subject]{BB17}, the physical mechanism driving such a variety of large-scale field topologies is still debated.
The fact that both rotation and convection play a major role in the stellar dynamo process is, however, well established \citep[see e.g. activity proxy studies of][]{MP84,NHB84,PMM03,WND11,WNW18}. Their joint effect on the magnetic field generation becomes obvious when considering observational measurements of the large-scale fields of low-mass stars as a function of the non-dimensional Rossby number (defined as the ratio of inertial to Coriolis forces and traditionally computed as $Ro = P_\mathrm{rot}/\tau$, where $\tau$ is the convective turnover time and $P_\mathrm{rot}$ is the rotation period of the star). The averaged surface field strength $\langle B\rangle$ shows two clear trends with the Rossby number. For $Ro > 0.1$, spectropolarimetric observations show that the large-scale magnetic field of cool stars weakens with increasing Rossby number \citep{VGJ14,FPB16}. This parameter region is often called "the unsaturated regime" and follows $\langle B\rangle \propto Ro^{-1.40 \pm 0.10}$ \citep{SMF19}, where the toroidal component of the large-scale field is reported to weaken faster than the poloidal component \citep{PDS08,SJV15}. As the Rossby number decreases below the $Ro \sim 0.1$ threshold, cool stars enter the "saturated regime" in which the large-scale field strength is roughly constant \citep{DMP08}.
The Rossby number has also proved to be quite successful at distinguishing various magnetic field morphologies in stellar observations \citep{MDP10,FBP18}. Stars with masses lower than $0.5$\,M$_{\sun}$ and $Ro \lesssim 0.1$ happen to have simple (dipole dominated) surface magnetic fields, whereas most stars featuring more complex surface fields tend to have larger Rossby numbers. Based on these observational results, it has been argued that stellar magnetic fields increase in complexity for stars with higher Rossby numbers. However, counterexamples that include stars harbouring complex field structures at low $Ro$ and others hosting dipole-dominated magnetic morphologies at large $Ro$ \citep[with Rossby numbers ranging from 0.2 to 0.3 --][]{DMP08,FPB16,FBP18} question the idea of magnetic fields getting more complex for stars with higher Rossby numbers. These results indicate that although the Rossby number may help at distinguishing between various generation mechanisms for the stellar magnetic fields, other proxies need to be invoked to clearly understand the transition between dipole-dominated and more complex field structures.
In the last two decades, numerical simulations mimicking the interior of planets (and, to a lesser extent, stars) have focused on understanding the origins of the magnetic morphology produced by convective dynamos. Parametric studies were conducted, using the relative strength of the axial dipole as a topological diagnostic to characterize the large-scale magnetic field. Geodynamo simulations with a constant density across the convective zone \citep[e.g.,][]{CA06,OC06,SJ06,SKA12} advocated that the Rossby number is indeed a key factor regulating the magnetic morphology. These initial numerical experiments suggested that dipole dominated morphologies only occur when $Ro \lesssim 0.1$ (commonly referred to as "the dipolar branch"), while complex surface fields could exist at both low and high Rossby numbers. Nevertheless, very recently \citet{MPG20} and \citet{TGF21} performed geodynamo simulations to explore the influence of the Lorentz force on the dipole breakdown. The authors found that strong dipoles can be recovered at high-Rossby numbers (up to $Ro = 0.18$) provided that a significant Lorentz force is acting on the fluid, challenging the canonical use of the Rossby number to distinguish between dipolar and multipolar field geometries. They suggested the ratio of inertial over Lorentz forces as an alternative proxy to capture the dipolar-multipolar transition. We propose to test this appealing hypothesis when the effect of a density contrast is introduced in the system.
Similar to what was initially found in geodynamo studies, stellar dynamo simulations showed a dipolar-multipolar transition with the Rossby number when considering weak density contrasts \citep{GDW12,J14}. However, these studies found that the dipolar branch disappeared for increasing density contrast. The apparent disagreement between the magnetic morphology observed in stars and those obtained in simulations of stratified flows raised the important question of why numerical experiments were apparently preventing dipoles from existing when the density contrast is more realistic \citep{PR19}. Further explorations of stratified flows with different physical properties showed that dipoles could be recovered at $Ro \lesssim 0.1$ when modifying the relative importance of the forces acting on the flow \citep{SPR14,RPD15}. To our knowledge, the simulation of \citet{YCM15} with $Ro = 0.04$ corresponds to the highest density contrast in which dipolar dynamos are reported to date. The authors obtained a strong dipole after considering a reduced influence of the inertial force by adopting a high ratio of viscous to thermal diffusions in a simulation with a density contrast of $N_\rho = \ln{\rho_i/\rho_o} = 5$ (where $\rho_i$ and $\rho_o$ are the density at the bottom and top of the convective zone, respectively). These various numerical experiments suggest that the dipole collapse could be an artificial bias of the parameter space explored with simulations. Thus, a close look at the force balance is needed to assess if the chosen parameter regime is indeed relevant for stars.
In this work, we attempt at reproducing for the first time the dipole-dominated field morphologies observed in some stars with $Ro > 0.1$. To do so, we perform a systematic parametric study of 3D convective dynamo simulations with different Rossby numbers and density contrasts, both of which are important ingredients in the stellar dynamo context. Guided by previous geodynamo studies, we focus on regime where the Lorentz force is dynamically active on the flow. The paper is organized as follows: we discuss our dynamo model and the selected control parameters in Sec.~\ref{sec:dynmodel}. The magnetic field morphology obtained in our simulations is presented in Sec.~\ref{sec:magmorph}, while the physical mechanisms controlling it are explored in Sec.~\ref{sec:diptran}. In Sec.~\ref{sec:dyngen}, we examine more closely the magnetic field generation in our simulations. Finally, we compare our results with previous stellar and geodynamo simulations and explore their implications in light of stellar observations in Sec.~\ref{sec:conclusions}.
\section{Dynamo model} \label{sec:dynmodel}
\subsection{Governing equations} \label{gov_eqs}
We model a stratified fluid in a spherical shell with inner radius $r_\mathrm{i}$ and outer radius $r_\mathrm{o}$ that rotates with angular velocity $\Omega_o$ about the axis $\vu{e}_\mathrm{z}$. We solve the non-dimensional magneto-hydrodynamics (MHD) equations under the anelastic formulation of \citet{BR95} and \citet{LF99}, expressed by
\begin{equation} \label{eq:mom_conserv}
\begin{split}
E\left[\pdv{\va{u}}{t} + (\va{u} \cdot \grad) \va{u} \right] &+ 2\vu{e}_z\cross\va{u} = - \grad(\frac{p'}{\tilde{\rho}}) + \frac{RaE}{Pr} gs'\vu{e}_\mathrm{r} \\
&+ \frac{1}{Pm \tilde{\rho}}(\curl \va{B} )\cross\va{B} + \frac{E}{\tilde{\rho}}\div S \qq*{,}
\end{split}
\end{equation}
\begin{equation} \label{eq:induction}
\begin{split}
\pdv{\va{B}}{t} = \curl( \va{u}\cross\va{B} ) -\frac{1}{Pm}\curl(\curl\va{B}) \qq*{,}
\end{split}
\end{equation}
\begin{equation} \label{eq:energy_conserv}
\begin{split}
\tilde{\rho}\tilde{T}\left[\pdv{s'}{t} + (\va{u} \cdot \grad) s' +u_\mathrm{r}\dv{\tilde{s}}{r} \right] = &\frac{1}{Pr} \div(\tilde{\rho} \tilde{T} \grad s') + \frac{PrDi}{Ra}Q_\nu \\
&+ \frac{PrDi}{Pm^2\ekRa}(\curl\va{B})^2 ,
\end{split} \\
\end{equation}
\begin{eqnarray}
\div(\tilde{\rho} \va{u}) &=& 0 \qq*{,} \\
\div \va{B} &=& 0, \label{eq:div}
\end{eqnarray}
where $\va{u}$ is the velocity field, $\va{B}$ is the magnetic field, $S$ represents the strain-rate tensor given by $$ S = \pdv{x_j} \left[\tilde{\rho}\left( \pdv{u_i}{x_j} + \pdv{u_j}{x_i}\right)\right] - \frac{2}{3} \pdv{x_i} \left(\tilde{\rho}\pdv{u_j}{x_j}\right),
$$ and $Q_\nu$ is the viscous heating expressed as $$ Q_\nu = \tilde{\rho}\left( \pdv{u_i}{x_j} + \pdv{u_j}{x_i} - \frac{2}{3} \delta_{ij} \div \va{u}\right) \pdv{u_i}{x_j}.
$$ Pressure and entropy fluctuations ($p'$ and $s'$, respectively) are defined with respect to the reference state (see Subsec.~\ref{sec:refstate}). We adopt a dimensionless formulation where the reference length scale is $r_\mathrm{o}$ and the time is given in units of $\tau_\nu = r_\mathrm{o}^2/\nu$, where $\nu$ is the fluid viscosity. The entropy scale is set to $r_\mathrm{o}|\mathrm{d}{\tilde{s}}/\mathrm{d}{r}|_{r_o}$, where $|\mathrm{d}{\tilde{s}}/\mathrm{d}{r}|_{r_o}$ is the normalized background entropy gradient at the outer boundary (see Sec.~\ref{sec:refstate}). The magnetic field is given in units of $\sqrt{\rho_{o}\mu\lambda\Omega_o}$, where $\mu$ is the magnetic permeability and $\lambda$ is the magnetic diffusivity. The gravity, density, and temperature are normalised by their outer radius values given by $g_{o}$, $\rho_{o}$, and $T_{o}$, respectively.
The dimensionless control parameters that appear in the equations above are the Ekman number $(E)$, Rayleigh number $(Ra)$, Prandtl number $(Pr)$, magnetic Prandtl number $(Pm)$, and dissipation number $(Di)$. They are defined as
\begin{displaymath}
E = \frac{\nu}{\Omega_o r_\mathrm{o}^2}\, , \;
Ra = \frac{g_{o} r_\mathrm{o}^4}{c_\mathrm{p}\kappa\nu} \left| \dv{\tilde{s}}{r} \right|_{r_\mathrm{o}} \, , \;
Pr = \frac{\nu}{\kappa} \, , \;
Pm = \frac{\nu}{\lambda} \, , \;
Di = \frac{g_{o} r_\mathrm{o}}{c_\mathrm{p} T_{o}},
\end{displaymath}
where $\kappa$ is the thermal diffusivity and $c_\mathrm{p}$ is the specific heat at constant pressure. We note that in the anelastic formulation adopted here, a non-adiabatic reference state is used. This translates into the appearance of a non-zero background entropy gradient $\dv{\tilde{s}}{r}$ in the entropy equation (Eq.~\ref{eq:energy_conserv}). The details of this reference state are discussed below.
\subsection{Reference state} \label{sec:refstate}
Thermodynamical quantities in Eqs.~\ref{eq:mom_conserv} to \ref{eq:energy_conserv} are expressed in terms of a reference (static) state and fluctuations around it. We adopt as reference state a nearly adiabatic ideal gas for which we prescribe the background entropy gradient $\dv{\tilde{s}}{r}$. We then deduce the reference temperature and density by solving the following equations:
\begin{equation}
\frac{1}{\tilde{T}}\pdv{\tilde{T}}{r} = \epsilon_\mathrm{s} \dv{\tilde{s}}{r} - \frac{Di}{T_\mathrm{o}} g(r)
\end{equation}
and
\begin{equation}
\frac{1}{\tilde{\rho}}\pdv{\tilde{\rho}}{r} = \epsilon_\mathrm{s} \dv{\tilde{s}}{r} - \frac{Di c_\mathrm{v}}{ (c_\mathrm{p} - c_\mathrm{v})T_\mathrm{o}}g(r),
\end{equation}
where we set the control parameter $\epsilon_\mathrm{s} = 10^{-4} \ll 1$, which is a necessary condition to ensure that we are still close to an adiabatic state. This formulation with a prescribed non-adiabaticity $\mathrm{d}\tilde{s}/\mathrm{d}r$ allows us to control the energy transport inside the star (notice its presence in Eq.~\ref{eq:energy_conserv}) and has been previously adopted in numerical models of gas giant planets \citep{DW18,GW21}. The background entropy sets radiative regions whenever $\mathrm{d}\tilde{s}/\mathrm{d}r > 0 $, while convectively-unstable regions occur when $ \mathrm{d}\tilde{s}/\mathrm{d}r < 0 $.
In the present work, we simulate convective shells with $r_\mathrm{i}/r_\mathrm{o} = 0.6$ and a fixed background entropy gradient $\mathrm{d}\tilde{s}/\mathrm{d}r = - 1$. We note that this choice is motivated by the fact that the entropy gradient calculated from 1D stellar evolution models of Sun-like stars is indeed approximately constant in the bulk of the convection zone (i.e., excluding the outer $5\%$ of the star in radius), which is the region we aim at modelling in this work. Our background entropy profile thus differs from previous anelastic studies, like the ones presented in the anelastic benchmark of \citet{JBB11}, where the reference state entropy is the solution of a conduction equation on which conditions of fixed entropy are applied. This leads to a solution with a gradient varying with radius, the maximal values of which being located in the outer part of the spherical shell. In our case, the gradient is constant throughout the shell, leading to a more homogeneous forcing of convection. This difference is illustrated in Figure~\ref{fig:unstable}, where the structure of the most unstable mode at the onset of convection is shown for our present work (left) and for an adiabatic reference state as used in \citet{JKM09} (right) with the same values of $N_\rho$, $E$ and $Pr$. At onset, our forcing of convection results in unstable modes located close to the bottom boundary \citep[see also][for similar results with an adiabatic reference state but different boundary conditions]{CH18}. When the Rayleigh number is increased however, strong convective velocities build close to the outer shell, as expected in stratified systems. To be more specific, we now give in Table~\ref{tab:critical} the values of the critical Rayleigh number and the critical azimuthal wavenumber in our setup, determined numerically at the different density contrasts used in our simulations and for the values of $E$ and $Pr$ adopted in all our calculations and which are specified in the next Subsection~\ref{subsec:param}.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figs/test_unstable.png}
\caption{Structure of the most unstable mode for convection forced through our background entropy profile (left) and through a more traditional entropy profile (right) for a density contrast $N_\rho=3$. Represented on the figure is an equatorial cut of the radial velocity close to the onset of convection at the values of $E=1.6\times 10^{-5}$ and $Pr=1$.}
\label{fig:unstable}
\end{figure}
\begin{table}
\caption{Critical Rayleigh numbers and azimuthal wavenumbers for our setup, for the three different density contrasts used in our simulations. These numbers are determined without taking into account the presence of a magnetic field.}
\label{tab:critical}
\centering
\begin{tabular}{ccc}
\hline
$N_\rho$ & ${\mathrm{Ra_\mathrm{c}}}$ & $m_c$ \\
\hline
1 & $1.92\times 10^7$& 32 \\
1.5 & $2.40\times 10^7$& 37 \\
3 & $3.56\times 10^7$& 39\\
\hline
\end{tabular}
\end{table}
We adopt a physically-motivated gravity based on the reference state of a main-sequence cool star that reads
\begin{equation} \label{eq:grav}
g(r) = -\frac{7.36~r}{r_\mathrm{o}} + \frac{4.99~r^2}{r_\mathrm{o}^2} + \frac{3.71~r_\mathrm{o}}{r} - \frac{0.34~r_\mathrm{o}^2}{r^2}.
\end{equation}
For the radial domain explored in this paper (with radius ratio $r_\mathrm{i}/r_\mathrm{o} = 0.6$), this gravity profile is virtually identical to the point mass approximation used in many parametric studies investigating dynamo action in planets and stars. We expect thus that any differences between our simulations and other similar ones in the literature with $r_\mathrm{i}/r_\mathrm{o} = 0.6$ are most likely caused by differences in the background entropy profile or control parameters (see Sec.~\ref{subsec:param}) rather than in the gravity profile.
\subsection{Numerical model and boundary conditions}
We use the anelastic version of the open-source code \verb|MagIC| \citep[][freely available at \url{https://github.com/magic-sph/magic}]{GW12} to solve Eqs.~\ref{eq:mom_conserv} to \ref{eq:div} in spherical coordinates. \verb|MagIC| has been validated through several anelastic benchmarks \citep{JBB11}. To evolve the Eqs.~\ref{eq:mom_conserv}-\ref{eq:energy_conserv} in time a mixed algorithm is adopted, where linear terms (except for the Coriolis one) are treated implicitly and non-linear terms are handled explicitly. Spherical harmonics are used as basis functions of the angular coordinates $(\theta, \phi)$ and are handled using the \verb|SHTns| library \citep[][freely available at \url{https://bitbucket.org/nschaeff/shtns}]{S13}. These functions are truncated at a maximum degree $\ell_\mathrm{max}$, sufficient to capture physical processes at play (typically ranging from $213$ to $341$ in our simulations). Chebyshev polynomials are used in the radial direction along with the mapping proposed by \citet{KT93}, which alleviates the grid refinement created near inner and outer boundaries in the standard formulation of the Chebyshev-collocation points. We refer to \citet{GW21} for additional details of this implementation in \verb|MagIC|.
In the full set of simulations, we adopt stress-free boundary conditions on the velocity field,
\begin{equation}
u_r = \pdv{r}(\frac{u_\theta}{r}) = \pdv{r}(\frac{u_\phi}{r}) = 0 \qq{on} r=r_\mathrm{i} \qand r = r_\mathrm{o},
\end{equation}
potential field boundaries on the magnetic field,
\begin{equation}
\va{J} = \curl\va{B} = 0 \qq{on} r=r_\mathrm{i} \qand r = r_\mathrm{o},
\end{equation}
and fixed entropy values, set to 0, at both boundaries. We initialize the velocity field with a small-amplitude random perturbation. The initial magnetic field is set to a dipole of strength $\Lambda = 0.44$ at the bottom of the convective zone (i.e., at $r = r_\mathrm{i}$), where $\Lambda = \mean{\expval{B^2}}$ is the Elsasser number expressed in terms of the dimensionless magnetic field.
\subsection{Choice of parameters}
\label{subsec:param}
In order to perform stellar dynamo simulations, a crucial ingredient to take into account is the density stratification. In the main-sequence, cool stars show a density contrast between the bottom ($\rho_i$) and the top ($\rho_o$) of the convective zone that can reach $N_\rho \sim 11$ \citep[according to models generated with the ATON code,][]{LVD06}. However, density contrasts as high as those seen in stars cannot be attained by numerical simulations as it drives fast small-scale motions that are too computationally demanding. In order to bypass this limitation, some authors chose to exclude from the numerical domain the outer few per cent of the stellar radii where the sharpest density gradients exist \citep{DSB06,B08,BMB11,ZGK16,EBS17,GZS19}.
We here also exclude this sharp gradient region from our domain and study the effect of varying $N_\rho$ from 1 to 3 to assess the influence of an increase of the density contrast on the magnetic field generation and flow dynamics.
We consider three different setups with $N_\rho = 1$, $1.5$, and $3$. These density contrasts are practically achieved in our formulation after fixing the dissipation number $Di = 1.53$, $2.7$, and $10$, respectively. Following previous studies, we adopt moderate values of $E = 1.6 \times 10^{-5}$ and $Pr = 1$ that reduce the numerical cost of each simulation, allowing us to perform a parametric study varying the Rayleigh number for the three different density contrasts. We increase the Rayleigh number from $1.3$ to $32.7~Ra_\mathrm{c}$ to explore the implications of distinct turbulence levels on the magnetic field morphology, where the convective onset $Ra_\mathrm{c}$ varies depending on the density contrast over the convective zone (see Table \ref{tab:critical}).
We are thus left with the choice of the magnetic Prandtl number $Pm$. Recent studies \citep[e.g.,][]{D16,DOP18,SGA19} have advocated that pushing a single parameter closer to the values observed in astrophysical objects may not represent the correct force balance at stake \citep[e.g., $E \approx 10^{-13}$, $Pr \approx 10^{-7}$, and $Pm \approx 10^{-3}$ at the bottom of the Solar convective zone;][]{O03}. There is considerable evidence from numerical simulations with/without density contrast that there is a critical magnetic Prandtl number $Pm_c$ below which dipolar dynamo solutions cannot be achieved for a fixed Ekman number. This brings some concerns as strong dipoles are observed in stars \citep[e.g.,][]{DL09}. One potential way to overcome this limitation is to adopt $Pm > Pm_c$. However, previous works showed that $Pm_c$ varies with $E$ and $N_\rho$. For the value adopted in this work of $E = 1.6 \times 10^{-5}$, it was shown that the critical magnetic Prandtl number obeys the relation $Pm_c = 2 N_\rho - 2$ \citep{SPR14}. Therefore, we choose to fix $Pm = 5$ for the entire set of simulations, which is greater than the critical value obtained for the highest stratified setup $N_\rho=3$. Moreover, we initialize our simulations with a dipole of strength $\Lambda = 0.44$, which has the same order of magnitude of typical stellar strengths \citep[e.g.,][]{MDP08,GMD13}.
\section{Results}
We performed altogether 23 dynamo simulations with different density contrasts and Rayleigh numbers. We ran numerical models for a few magnetic diffusion times to achieve meaningful dynamo steady-states, which resulted here in rather costly simulations. The journal of simulations is summarized in Table~\ref{tab:runs}.
We provide the total simulation time $\tau_\text{end}$ in units of magnetic diffusion time, which we defined as
\begin{equation} \label{eq:taulambda}
\tau_\lambda = \frac{D_\mathrm{cz}^2}{\lambda} = Pm \left(\frac{D_\mathrm{cz}}{r_\mathrm{o}}\right)^2 \tau_\nu
\end{equation}
using the convective shell size $D_\mathrm{cz} = r_\mathrm{o} - r_\mathrm{i}$ as the relevant length scale. Throughout this work, we employ overbars $\mean{\; \cdot \;}$ to represent averages over time, brackets $\expval{\cdot}$ to represent volume averages, and $\expval{\cdot}_i$ to represent spatial averages in the direction $\vu{e}_\mathrm{i}$. Time averages are performed only after the solutions have reached a well-established steady-state and typically cover a few magnetic diffusion times (for more information see Appendix~\ref{appendix:dipolarity}).
\begin{table*}
\caption{Journal of simulations. First column yields the run ID. Columns 2-6 indicate the parameters imposed in each simulation (see Sec.~\ref{subsec:param}). Column 7 and 8 show the total simulation time $\tau_\text{end}$ and the averaging time $\tau_\mathrm{avg}$ (both expressed in units of magnetic diffusion time $\tau_\lambda$ as defined in Eq.~\ref{eq:taulambda}), respectively. Column 9 gives the dominant scale of convection $\ell_\text{peak}$ (Eq.~\ref{eq:lpeak}). Column 10 displays the local Rossby number (Eq.~\ref{eq:rossby}). Column 11 shows the Inertia over Lorentz force ratio (see Sec.~\ref{sec:fb}) and column 12 the kinetic over magnetic energy ratio (Eq.~\ref{eq:ekem}). Column 13 shows the dipolarity computed using Eq.~\ref{eq:fdip}, whereas column 14 gives a variation of the dipolarity measure based on the total dipole $f_\mathrm{dip,Tot.}$ (see discussion in Sec.~\ref{sec:magmorph}). }
\setlength{\tabcolsep}{3.2pt}
\label{tab:runs}
\centering
\begin{tabular}{lccccccccccccc}
\hline
Run ID & $N_\rho$ & $\rho_i/\rho_o$ & $Ra$ & $Ra$/$Ra_\mathrm{c}$ & ($N_r, N_\theta, N_\phi$) & $\tau_\text{end}$ & $\tau_\text{avg}$ & $\ell_\text{peak}$ & $Ro_\ell$ & $\mathcal{F}_\mathrm{I}/\mathcal{F}_\mathrm{L}$ & $E_{K}/E_{M}$ & $f_\mathrm{dip}$ & $f_\mathrm{dip,Tot.}$ \\
& & & & & & ($\tau_\lambda$) & ($\tau_\lambda$) & & & & & & \\
\hline
FC01 & 1.0 & 2.7 & $4.77 \times 10^7 $ & 2.5 & $(73 ,320,640)$ & 5.1 & 2.0 & 17 & $ 0.023 \pm 0.004$ & $ 0.03 \pm 0.25$ & $ 0.11 \pm 0.17$ & $ 0.84 \pm 0.06$ & $ 0.84 \pm 0.06$ \\
FC02 & 1.0 & 2.7 & $6.25 \times 10^7 $ & 3.3 & $(73 ,320,640)$ & 4.3 & 2.0 & 18 & $ 0.036 \pm 0.007$ & $ 0.05 \pm 0.39$ & $ 0.05 \pm 0.01$ & $ 0.87 \pm 0.02$ & $ 0.87 \pm 0.02$ \\
FC03 & 1.0 & 2.7 & $7.81 \times 10^7 $ & 4.1 & $(73 ,320,640)$ & 4.1 & 2.0 & 19 & $ 0.05 \pm 0.01$ & $ 0.07 \pm 0.30$ & $ 0.08 \pm 0.01$ & $ 0.87 \pm 0.02$ & $ 0.87 \pm 0.01$ \\
FC04 & 1.0 & 2.7 & $1.04 \times 10^8 $ & 5.5 & $(73 ,512,1024)$ & 4.1 & 2.0 & 22 & $ 0.09 \pm 0.02$ & $ 0.12 \pm 0.11$ & $ 0.15 \pm 0.02$ & $ 0.87 \pm 0.01$ & $ 0.87 \pm 0.01$ \\
FC05 & 1.0 & 2.7 & $1.25 \times 10^8 $ & 6.5 & $(73 ,512,1024)$ & 0.9 & 0.3 & 28 & $ 0.12 \pm 0.03$ & $ 0.16 \pm 0.12$ & $ 0.26 \pm 0.03$ & $ 0.87 \pm 0.02$ & $ 0.88 \pm 0.02$ \\
FC06 & 1.0 & 2.7 & $1.56 \times 10^8 $ & 8.2 & $(73 ,512,1024)$ & 4.6 & 1.5 & 24 & $ 0.18 \pm 0.05$ & $ 0.49 \pm 0.09$ & $ 1.05 \pm 0.08$ & $ 0.12 \pm 0.03$ & $ 0.13 \pm 0.03$ \\
FC07 & 1.0 & 2.7 & $3.12 \times 10^8 $ & 16.3 & $(73 ,512,1024)$ & 3.1 & 1.0 & 19 & $ 0.32 \pm 0.09$ & $ 0.57 \pm 0.09$ & $ 1.10 \pm 0.09$ & $ 0.11 \pm 0.03$ & $ 0.12 \pm 0.02$ \\
FC08 & 1.0 & 2.7 & $6.25 \times 10^8 $ & 32.7 & $(73 ,1024,2048)$ & 1.3 & 0.5 & 14 & $ 0.53 \pm 0.12$ & $ 0.58 \pm 0.08$ & $ 1.36 \pm 0.10$ & $ 0.12 \pm 0.03$ & $ 0.18 \pm 0.03$ \\
\hline
FC09 & 1.5 & 4.4 & $4.77 \times 10^7 $ & 2.0 & $(73 ,320,640)$ & 4.3 & 1.0 & 45 & $ 0.031 \pm 0.007$ & $ 0.25 \pm 0.13$ & $ 0.59 \pm 0.30$ & $ 0.71 \pm 0.06$ & $ 0.71 \pm 0.06$ \\
FC10 & 1.5 & 4.4 & $6.25 \times 10^7 $ & 2.6 & $(73 ,320,640)$ & 4.5 & 1.4 & 38 & $ 0.05 \pm 0.01$ & $ 0.19 \pm 0.19$ & $ 0.58 \pm 0.25$ & $ 0.62 \pm 0.04$ & $ 0.62 \pm 0.04$ \\
FC11 & 1.5 & 4.4 & $7.81 \times 10^7 $ & 3.3 & $(73 ,320,640)$ & 6.5 & 2.5 & 38 & $ 0.07 \pm 0.02$ & $ 0.33 \pm 0.13$ & $ 0.70 \pm 0.19$ & $ 0.44 \pm 0.11$ & $ 0.45 \pm 0.11$ \\
FC12 & 1.5 & 4.4 & $1.04 \times 10^8 $ & 4.3 & $(73 ,512,1024)$ & 4.9 & 1.9 & 35 & $ 0.11 \pm 0.04$ & $ 0.28 \pm 0.11$ & $ 0.59 \pm 0.08$ & $ 0.15 \pm 0.04$ & $ 0.45 \pm 0.04$ \\
FC13 & 1.5 & 4.4 & $1.56 \times 10^8 $ & 6.5 & $(73 ,512,1024)$ & 5.1 & 1.5 & 35 & $ 0.17 \pm 0.06$ & $ 0.34 \pm 0.11$ & $ 0.70 \pm 0.09$ & $ 0.46 \pm 0.14$ & $ 0.56 \pm 0.08$ \\
FC14 & 1.5 & 4.4 & $3.12 \times 10^8 $ & 13.0 & $(73 ,512,1024)$ & 3.8 & 1.0 & 25 & $ 0.31 \pm 0.12$ & $ 0.57 \pm 0.09$ & $ 1.05 \pm 0.09$ & $ 0.12 \pm 0.03$ & $ 0.14 \pm 0.03$ \\
FC15 & 1.5 & 4.4 & $6.25 \times 10^8 $ & 26.0 & $(73 ,512,1024)$ & 1.5 & 0.5 & 20 & $ 0.52 \pm 0.19$ & $ 0.64 \pm 0.08$ & $ 1.35 \pm 0.09$ & $ 0.13 \pm 0.03$ & $ 0.17 \pm 0.03$ \\
\hline
FC16 & 3.0 & 19.3 & $4.77 \times 10^7 $ & 1.3 & $(73 ,320,640)$ & 8.1 & 1.9 & 42 & $ 0.013 \pm 0.003$ & $ 1.78 \pm 0.54$ & $ 1.17 \pm 0.18$ & $ 0.04 \pm 0.02$ & $ 0.12 \pm 0.04$ \\
FC17 & 3.0 & 19.3 & $7.81 \times 10^7 $ & 2.2 & $(73 ,320,640)$ & 5.4 & 1.5 & 38 & $ 0.037 \pm 0.008$ & $ 0.33 \pm 0.19$ & $ 0.41 \pm 0.10$ & $ 0.63 \pm 0.03$ & $ 0.63 \pm 0.03$ \\
FC18 & 3.0 & 19.3 & $1.56 \times 10^8 $ & 4.4 & $(73 ,512,1024)$ & 6.0 & 1.4 & 36 & $ 0.11 \pm 0.05$ & $ 0.36 \pm 0.11$ & $ 0.54 \pm 0.05$ & $ 0.54 \pm 0.03$ & $ 0.55 \pm 0.03$ \\
FC19 & 3.0 & 19.3 & $2.08 \times 10^8 $ & 5.8 & $(73 ,512,1024)$ & 2.4 & 1.0 & 39 & $ 0.15 \pm 0.08$ & $ 0.35 \pm 0.10$ & $ 0.58 \pm 0.06$ & $ 0.53 \pm 0.03$ & $ 0.54 \pm 0.03$ \\
FC20 & 3.0 & 19.3 & $3.12 \times 10^8 $ & 8.8 & $(73 ,512,1024)$ & 4.1 & 1.3 & 34 & $ 0.21 \pm 0.13$ & $ 0.36 \pm 0.09$ & $ 0.52 \pm 0.04$ & $ 0.63 \pm 0.05$ & $ 0.63 \pm 0.05$ \\
FC21 & 3.0 & 19.3 & $6.25 \times 10^8 $ & 17.6 & $(73 ,512,1024)$ & 1.6 & 0.5 & 31 & $ 0.38 \pm 0.25$ & $ 0.47 \pm 0.07$ & $ 0.64 \pm 0.06$ & $ 0.75 \pm 0.03$ & $ 0.75 \pm 0.03$ \\
FC22 & 3.0 & 19.3 & $7.44 \times 10^8 $ & 20.9 & $(73 ,1024,2048)$ & 1.1 & 0.4 & 30 & $ 0.41 \pm 0.26$ & $ 0.45 \pm 0.05$ & $ 0.68 \pm 0.05$ & $ 0.77 \pm 0.02$ & $ 0.77 \pm 0.02$ \\
FC23 & 3.0 & 19.3 & $9.20 \times 10^8 $ & 25.8 & $(73 ,1024,2048)$ & 1.3 & 0.4 & 29 & $ 0.51 \pm 0.32$ & $ 0.70 \pm 0.07$ & $ 1.20 \pm 0.08$ & $ 0.23 \pm 0.05$ & $ 0.25 \pm 0.05$ \\
\hline
\end{tabular}
\end{table*}
\subsection{Magnetic morphology} \label{sec:magmorph}
Since the physical origin of the various magnetic field morphologies observed in cool stars is still debated, in this study we particularly focus on the field topology achieved in our simulations. Traditionally, the magnetic field morphology has been assessed by measuring the relative importance of the axial-dipole at the stellar surface. This quantity, named \textit{dipolarity}, is defined as the relative strength of the axial-dipole\footnote{A different definition of `dipolarity' based on the relative energy of the axial dipole also appears in the literature, in which the right-hand-side of Eq.~\ref{eq:fdip} is squared.} \citep{CA06}:
\begin{equation} \label{eq:fdip}
f_\mathrm{dip} = \mean{\sqrt{\frac{ \iint \va{B}^2_{\ell=1, m=0} (r =r_\mathrm{o}, \theta, \phi) \sin{\theta}\dd{\theta}\dd{\phi}}{\sum_{\ell=1}^{11} \sum_{m=0}^{\ell} \iint \va{B}^2_{\ell,m} (r =r_\mathrm{o}, \theta, \phi) \sin{\theta}\dd{\theta}\dd{\phi}}}}.
\end{equation}
Here, the normalization factor corresponds to the square root of the total surface magnetic energy stored in the largest spatial scales, i.e. in modes with order $\ell < 12$. It thus matches the typical resolution achieved in the surface magnetic field reconstruction of stars other than the Sun \citep[e.g., ][]{DMP08,MDP10,FPB16,FBP18}. We recall the reader that toroidal fields vanish at the outer boundary because of our magnetic boundary condition (and, therefore, only poloidal fields contribute in Eq.~\ref{eq:fdip}). Following previous authors \citep[e.g.,][]{OLD14,MPG20,TGF21}, we define simulations with $f_\mathrm{dip} \ge 0.5$ (or equivalently, with an axial-dipole containing $25\%$ of the magnetic energy stored at modes up to $\ell = 11$) as dipolar dynamos. Conversely, simulations in which $f_\mathrm{dip} < 0.5$ are defined as ``multipolar'' dynamos. The dipolarity measurements are given in Table~\ref{tab:runs} along with an alternative estimate based on the total dipole $f_\mathrm{dip,Tot.}$ (i.e., including the equatorial dipole contribution in the summation at the numerator of Eq.~\ref{eq:fdip}). We note that none of our simulations would change their classification as dipolar or multipolar dynamos if considering a dipolarity based on the total dipole. We thus stick to the dipolarity definition given by Eq.~\ref{eq:fdip} throughout this work.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{DipolarityRac.pdf}
\caption{Surface dipolar fraction as a function of the Rayleigh number for the 23 runs listed in Table~\ref{tab:runs} (grey symbols). The shape of the symbols distinguishes between dipolar dynamos (circle) and multipolar dynamos (cross). Simulations with density contrast N$_\rho = \log{\rho_i/\rho_o} = 1$, $1.5$, and $3$, are separated respectively in panels (a), (b), and (c). Error bars represent one standard deviation about the time averaged dipolarity. Stratified dynamos with the same radius ratio ($r_\mathrm{i}/r_\mathrm{o} = 0.6$) and density contrasts, but $Pm=1$ are included for comparison \citep[purple symbols;][]{GDW12,GMD13}.
}
\label{fig:fdipRa}
\end{figure}
Figure~\ref{fig:fdipRa} shows how the dipolarity varies with the Rayleigh number. This figure shows three panels with $f_\mathrm{dip}$ as a function of $Ra$, each at a particular $N_\rho$. Starting from the set of simulations with $N_\rho = 1$ (Figure~\ref{fig:fdipRa} a), we identify dipolar dynamos at low Rayleigh numbers followed by a sharp transition to multipolar dynamos as $Ra$ increases. This finding is in line with earlier simulations of \citet{GDW12,GMD13}\footnote{The control parameters adopted by \citet{GDW12,GMD13} coincide with those employed in this work with the exception of $Pm$. However, with also different formulations of convective forcing (similar to what has been described in Figure~\ref{fig:unstable}), caution must be applied when attributing possible differences between the models to $Pm$.} using $Pm =1$ (purple symbols), which showed that the morphology transitions to a more complex configuration around $Ra = 7 Ra_\mathrm{c}$. It also extends Rayleigh's parameter space coverage by about a factor of three when compared to \citet{GDW12,GMD13}, corroborating the hypothesis that only multipolar dynamos exist for forcings above the threshold leading to the dipole collapse (i.e., $Ra \gtrsim 7 Ra_\mathrm{c}$ for $N_\rho = 1$).
The dipolarity trend, however, changes for the models with $N_\rho = 1.5$ (Figure~\ref{fig:fdipRa} b). While the \textit{plateau} with strong dipolar dynamos seen for the runs with $N_\rho = 1$ no longer exists, intermediate values of $f_\mathrm{dip}$ appear, defining a rather continuous transition to the multipolar branch. We highlight that two of our multipolar cases are compatible with a dipole within error bars (estimated as one standard deviation over the time averaged value). An inspection of the simulations around $5~Ra_\mathrm{c}$ reveals one case with polarity reversals (FC11) and two with excursions (FC12 and FC13) of the dipole field, thus explaining why large error bars are found in those cases where the dipolar field strongly varies in time. This finding is in accordance with previous studies evaluating reversing dipoles, which observed a tendency for its occurrence at Rayleigh numbers close to the transition between dipolar and multipolar dynamos \citep{KC02,OC06,WT10}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Br_Nr3_21rac.png}
\includegraphics[width=\columnwidth]{figs/Br_Nr3_26rac.png}
\caption{Mollweide projections of the surface radial magnetic field for a dipolar (top) and a multipolar (bottom) case with $N_\rho = 3$, corresponding to the run IDs FC22 and FC23, respectively. Red shades correspond to radial fields point outward and blue shades inward.
}
\label{fig:surfBr2}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=1.8\columnwidth]{figs/Vr_Br.png}
\caption{Snapshot of the radial velocity (left) and radial magnetic field (right) in the dipolar run shown in Figure\ref{fig:surfBr2} (FC22).}
\label{fig:surfBr3}
\end{figure*}
The most striking result to emerge from the data is seen for the density contrast $N_\rho = 3$ (Figure~\ref{fig:fdipRa} c). Contrary to the other setups considered in this work, a multipolar dynamo is found close to the dynamo onset ($Ra = 1.3 Ra_\mathrm{c}$). The dipolarity then shows a marked rise going from almost $0$ to $0.62$ as the forcing reaches about two times the critical Rayleigh number. Dipolar dynamos are then consistently sustained for a wide range of supercriticality until the morphology finally transitions to a multipolar configuration at $Ra\sim 25Ra_\mathrm{c}$. Compared to the previous simulations of \citet{GDW12,GMD13} with $Pm = 1$ and covering a parameter space of ${Ra < 5 Ra_\mathrm{c}}$, we note that dipolar dynamos are kept for a much wider range of forcing.
Comparing $N_\rho=1$ and $N_\rho=1.5$ simulations, we see that the range of $Ra$ numbers where the dipolar branch can be obtained shrinks as the density contrast increases. Although this result seems to reflect those of \citet{GDW12,GMD13} and \citet{J14}, who pointed out that dipolar dynamos would ultimately disappear for $N_\rho \gtrsim 2$, the strong dipoles obtained for $N_\rho = 3$ do not support this early conclusion. In fact, these results substantiate the previously unique simulation of \citet{YCM15}, which yielded a strong dipole ($f_\mathrm{dip} \approx 0.55$) despite the high density contrast of $N_\rho = 5$. As argued by \citet{PR19}, one possibility is that the dipolarity loss found in previous works resulted from the restricted parameter space explored rather than being caused by a real modification of the dynamo mechanisms taking place in stars with different density contrasts. Indeed as we shall explore in Sec.~\ref{sec:fb}, our setup with $Pm = 5$ increases the contribution of the Lorentz force to the force balance, sustaining dipolar dynamos even for stratification as high as $N_\rho = 3$. If anything, our simulations reinforce the idea that the regime of stability of dipolar dynamos depends on the parameter space explored \citep{RPD15} and provides evidence that sometimes higher stratification helps to sustain dipolar fields.
Figure~\ref{fig:surfBr2} shows the surface radial magnetic field for the last dipole before the transition (FC22) and the multipolar case after the collapse (FC23). Compared to the runs with $N_\rho=1$ (not shown here), smaller scales dominate the structure of the surface radial magnetic field in both cases. Indeed, a well-known effect of increasing the density stratification is to decrease the typical flow length scale, which in turn decreases the typical size of magnetic structures. We come back to this point when we discuss the scale at which the kinetic energy peaks in our simulations (see Sec.~\ref{sec:fb}). It is rather clear from this figure that a large-scale dipolar structure is present in the upper panel, with a positive North pole and negative South pole. On the contrary, in the bottom panel, the magnetic field is dominated by a salt and pepper like structure with the strongest field concentrations located in narrow bands more or less extended in latitude. Figure~\ref{fig:surfBr3} enables us to proceed to a closer inspection of the relationship between the flow and field morphologies. This figure shows a 3D rendering of the radial velocity field (left panel) and of the radial magnetic field (right panel) in the dipolar run shown in the top panel of Figure~\ref{fig:surfBr2}. It is rather clear from these 3D snapshots that narrow downwelling flows create intense magnetic flux concentrations, while broad upwelling flows diffuse the magnetic field. We also note that in this strongly stratified case and at this level of supercriticality $(Ra = 20.9 \, Ra_\mathrm{c})$, the amplitude of the convective velocities is strongest at the outer shell, as expected for strongly stratified systems.
\subsection{The dipolar-multipolar transition} \label{sec:diptran}
Many studies interpreted the transition from the dipolar dynamos to multipolar dynamos in terms of the balance between inertia and Coriolis forces in the Navier-Stokes equation (Eq.~\ref{eq:mom_conserv}). A proxy to estimate this force ratio is the local Rossby number $Ro_\ell$ introduced by \citet{CA06}. They suggested that the dipole-multipole transition is well captured by
\begin{equation} \label{eq:rossby}
Ro_\ell = \mean{\avg{\frac{u_\mathrm{rms}}{\Omega_oD_\mathrm{cz}}\frac{\ell_u}{\pi}}}, \qq{where}
\ell_u = \frac{\sum_\ell \ell u_\ell^2}{\sum_\ell u_\ell^2}
\end{equation}
is the mean spherical harmonic degree of the flow. The global picture suggested that axial-dipole dominated solutions could only exist at low-Rossby numbers because of the ordering role played by the Coriolis force \citep[with typically $Ro_\ell \lesssim 0.12$,][]{CA06}. Beyond this limit, the increased importance of inertia compared to Coriolis would cause the dipole collapse (with the star thus joining the multipolar branch).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Dipolarity.pdf}
\caption{Surface dipolar fraction as a function of the local Rossby number $Ro_\ell$ (Eq.~\ref{eq:rossby}). Colours group different levels of stratification (see legend), whereas symbols distinguish dipolar dynamos (circle) from multipolar dynamos (cross). The horizontal dashed black line marks the dipolar-multipolar transition, and the vertical one indicates the standard dipolar collapse predicted from geodynamo simulations \citep{CA06}.
}
\label{fig:fdip}
\end{figure}
We plot $f_\mathrm{dip}$ as a function of $Ro_\ell$ in Figure~\ref{fig:fdip}. Simulations with $N_\rho = 1$ display a dipolar-multipolar transition at $Ro_\ell \sim 0.12$ (vertical dashed line), in agreement with Boussinesq results and arguments of \citet{CA06}. However, if we now turn to the runs with $N_\rho = 1.5$ or $3$, there is no clear evidence that $Ro_\ell$ influences the dipole collapse. For these density contrasts, multipolar solutions are identified in the Rossby regime where mainly dipolar fields are predicted and vice-versa.
Perhaps one of the most interesting aspect evidenced by our simulations is that axial-dipole dominated simulations might display similar values of $f_\mathrm{dip}$ regardless of whether it falls in the dipolar or multipolar branch as initially advised from Boussinesq simulations \citep{CA06}. Another key aspect is that dipolar solutions persist for large Rossby numbers precisely for the setup of highest density contrast ($N_\rho = 3$), which corresponds to the most realistic model in the stellar context.
In an attempt to create a more general description for the dipolar transition, other proxies besides the Rossby number were explored in the literature to explain the possible causes for the dipole breakdown. As we discuss in Appendix~\ref{appendix:flow}, the change on the flow structure \citep{SKA12,GOD17} is not enough to explain the transition from dipoles to multipoles in our numerical simulations. In particular, it seems that the magnetic morphology can only be described by a change on the flow arrangement when considering systems where the magnetic feedback on the flow is small/nonexistent (essentially behaving as a hydrodynamic flow).
Recently, Boussinesq simulations have shown that for systems in which the magnetic feedback is significant the relative importance of the Lorentz force in the Navier Stokes equation can describe the dipole breakdown \citep{MPG20,TGF21}. However, it is not clear whether those analyses still hold in anelastic dynamos. We explore next whether the balance between the forces entering the Navier-Stokes equation control the magnetic morphology in stratified systems.
\subsection{Force balance: inertia \textit{vs} Lorentz force} \label{sec:fb}
Following previous studies \citep{AGF17,SGA19,TGF21,GW21}, we compute the time-averaged root-mean-square (RMS) force spectra of the individual forces identified below
\begin{equation*} \label{eq:mom_conserv2}
\begin{split}
&\underbrace{E\left[\pdv{\va{u}}{t}+(\va{u}\cdot\grad)\va{u}\right]}_{\text{Inertia}} +\underbrace{2\vu{e}_z\cross\va{u}}_{\text{Coriolis}} = \\ &-\underbrace{\grad(\frac{p}{\tilde{\rho}})}_{\text{Pressure}} + \underbrace{\frac{RaE}{Pr} gs'\vu{e}_\mathrm{r}}_{\text{Buoyancy}} + \underbrace{\frac{1}{Pm \tilde{\rho}}(\curl \va{B} )\cross\va{B}}_{\text{Lorentz}} + \underbrace{\frac{E}{\tilde{\rho}}\div S}_{\text{Viscous}} \qq*{.}
\end{split}
\end{equation*}
Here, time-averaged RMS force spectra are given by
\begin{equation} \label{eq:forcelpeak}
\mathcal{F}_\mathrm{RMS}(\ell) = \mean{\sqrt{\expval{\sum_{m=-\ell}^{\ell} \abs{\vec{F}_{\ell,m}(r,\theta,\phi,t)}^2}}}.
\end{equation}
where $\vec{F}_{\ell,m}$ is the vector spherical harmonic transform of the force at stake.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/FBspec.pdf}
\caption{Force balance spectra for the same dipolar (top) and multipolar (bottom) models shown in Figure~\ref{fig:surfBr2}. Solid lines correspond to time-averaged force spectra, with colours representing the different forces entering the Navier-Stokes equation. Shaded regions represent one standard deviation from the time-averaged value. The vertical dashed line marks the integral scale $\ell_\mathrm{peak}$ defined in Appendix~\ref{appendixB}.}
\label{fig:FBspec}
\end{figure}
Figure~\ref{fig:FBspec} illustrates the force balance spectra for a dipolar and a multipolar run with $N_\rho = 3$ (corresponding to the same runs shown in Figure~\ref{fig:surfBr2}). Both models display forces whose respective contributions vary depending on the spatial scale. At scales up to $\ell \sim 40$, the Coriolis (black) and pressure (blue) forces balance each other at first order resulting in a quasi-geostrophic balance \citep[QG, for further details, see][]{C18}, whereas buoyancy (green), Lorentz (red), and inertial (yellow) forces show a marginal contribution at second-order. On the other hand, at small scales ($\ell \gtrsim 40$) the Lorentz force becomes dominant and starts to balance the pressure force in the place of the Coriolis force. Comparing both models, we can identify an increase in the inertial contribution from the dipolar to the multipolar case, with the inertial force reaching values comparable to the Lorentz force in the latter.
To track the relative contribution of each force in our parametric study, we look for a particular length scale $\ell_\text{peak}$ defined as the dominant scale of the convective flow \citep[for more details on its calculation, see appendix~\ref{appendixB} and][]{SGA21}. The values of $\ell_\text{peak}$ are given in Table~\ref{tab:runs} for each simulation. We note here that the impact of the density stratification is reflected in the strong increase of $\ell_\text{peak}$ with $N_\rho$. Indeed, from $N_\rho=1$ to $N_\rho=3$, $\ell_\text{peak}$ is typically multiplied by a factor 2. We now compute the RMS forces at the integral scale $\ell_\text{peak}$, namely, Coriolis force $\mathcal{F}_{C}$, pressure gradient force $\mathcal{F}_{P}$, buoyancy (or Archimedes) force $\mathcal{F}_{B}$, Lorentz force $\mathcal{F}_{L}$, inertial force $\mathcal{F}_{I}$, and the viscous force $\mathcal{F}_{V}$.
Figure~\ref{fig:FBra} shows these forces as a function of $Ra/Ra_\mathrm{c}$ for models with $N_\rho = 1$ and 3. While the entire data set features a QG balance at first order, the ageostrophic part of the Coriolis force, defined as $\mathcal{F}_\mathrm{Ageo} = \abs{\mathcal{F}_{C} - \mathcal{F}_{P}}$, enters a second-order force balance that varies depending on $N_\rho$ and $Ra$.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/FBLpolRac.pdf}
\caption{Force contributions at the integral scale $\ell_\text{peak}$ (Eq.~\ref{eq:forcelpeak}) as a function of $Ra/Ra_\mathrm{c}$. Top and bottom panels show runs with N$_\rho = 1$ and $3$, respectively}.
\label{fig:FBra}
\end{figure}
For $N_\rho = 1$ (top panel), we identify two kinds of second-order balance depending on the Rayleigh number. At $Ra < 7 Ra_\mathrm{c}$, the ageostrophic Coriolis force is balanced by $\mathcal{F}_{L}$ and $\mathcal{F}_{B}$ forces, which dominate over $\mathcal{F}_{I}$ and $\mathcal{F}_{V}$ by roughly an order of magnitude. This flow state, devised by \citet{D13}, is frequently referred to as the quasi-geostrophic Magneto-Archimedean-Coriolis (QG-MAC) balance, and it has been obtained in geodynamo models \citep{YGC16,AGF17,SJN17} and in anelastic models of gas giant planets \citep{GW21}. At $Ra > 7 Ra_\mathrm{c}$, inertial forces become important and contribute to the second-order balance of the Navier-Stokes equation. We observe that the breakdown of the dipole occurs at this point. The role played by inertia in destabilizing dipoles was likewise found before in Boussinesq simulations \citep[e.g.,][]{SJ06,CA06}.
Similar conclusions can be drawn for the $N_\rho = 3$ data set (bottom panel), with the main difference relying on the isolated multipolar solution at $Ra = 1.3 Ra_\mathrm{c}$, i.e., very close to the convective onset. Among the entire set of simulations performed, this case is the only one that does not display a dominant Lorentz contribution to the flow dynamics. Instead, it yields a strong contribution of $\mathcal{F}_{B}$ and a marginal one of $\mathcal{F}_{V}$. This flow adjustment is often called quasi-geostrophic Viscous-Archimedean-Coriolis (QG-VAC) balance \citep{YGC16,SGA21}. The QG-VAC balance is quickly destroyed as turbulence builds-up due to a sharp rise in the $\mathcal{F}_{L}$ with $Ra$. One of the main conclusions we can extract from Figure~\ref{fig:FBra} is that with this stratification, dipolar dynamos prevail for much higher $Ra/Ra_\mathrm{c}$ than for the less stratified cases. The transition in the surface field morphology is indeed seen at $Ra = 25.8 Ra_\mathrm{c}$. Akin to what has been described for $N_\rho = 1$, the morphology transition occurs as the gap between $\mathcal{F}_{L}$ and $\mathcal{F}_{I}$ decreases. This finding suggests that, in the Lorentz force dominated regime, the effect of the density stratification is to increase the level of turbulence at which inertial forces become comparable to the Lorentz forces.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Dipolarity_FB.pdf}
\caption{Surface dipolar fraction as a function of the ratio between inertia and Lorentz force at the integral scale. Symbols are defined as in Figure~\ref{fig:fdip}. The vertical dashed black line indicates the tentative threshold ${\mathcal{F}_{I}/\mathcal{F}_{L} = 0.4}$ for the dipole breakdown. The error bars correspond to one standard deviation about the time-averaged quantities. Shaded areas indicate the dipolar (cyan) and multipolar (coral) branches proposed in this work.
}
\label{fig:FBdip}
\end{figure}
To test the hypothesis that the importance of inertia in the 2nd-order force balance is the main factor responsible for destabilising dipolar solutions, we plot in Figure~\ref{fig:FBdip} the dependence between $f_\mathrm{dip}$ and $\mathcal{F}_{I}/\mathcal{F}_{L}$ for the three setups considered in this work. Dipolar and multipolar branches are identified using this proxy. We find that simulations with ${\mathcal{F}_{L} \gg \mathcal{F}_{I}}$ develop strong dipolar dynamos, while a sharp transition to multipolar dynamos is obtained as inertia increases in intensity. A tentative description for the dipolar-multipolar transition gives ${\mathcal{F}_{I}/\mathcal{F}_{L} \simeq 0.4}$ (vertical dashed line). It follows that $\mathcal{F}_{I}/\mathcal{F}_{L}$ provides a more unified view of the dipolar-multipolar transition than $Ro_\ell$ (Figure~\ref{fig:fdip}), independently of the density contrast $N_\rho$. This result agrees with those of \citet{MPG20} and \citet{TGF21}, who also found that the competition between inertial and Lorentz forces can capture the dipole collapse in Boussinesq simulations. We thus confirm that these results still hold in stratified systems, and even argue that the transition may occur at larger levels of turbulence for strongly stratified cases, opening the possibility that stars harbouring strong dipoles may indeed operate in this Lorentz force-dominated regime.
\subsection{Possible proxies for stellar observations}
\label{sec:criteriaObs}
\subsubsection{Energy distribution}\label{energy_proxy}
Following \cite{TGF21}, we now try to look for an alternative quantity to the ratio ${\mathcal{F}_{I}/\mathcal{F}_{L}}$ that is more accessible to observations and yet incorporates the physics behind the dipole collapse. To establish this new measure, we use the kinetic energy stored in the convective motions ($E_K$) as a proxy of the inertial force and the magnetic energy ($E_M$) as a proxy of the Lorentz force. The rough approximation of ${\mathcal{F}_{I}/\mathcal{F}_{L}}$ is then given by the time and volume-averaged energy ratio
\begin{equation} \label{eq:ekem}
\frac{E_K}{E_M} = EPm \mean{\frac{\avg{\tilde{\rho}\va{u}^2}}{\avg{\va{B}^2}}}.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Dipolarity_ekem.pdf}
\caption{Surface dipolar fraction in terms of the ratio of the time and volume-integrated kinetic energy stored in the convective motions and magnetic energy. The vertical dashed black line indicates the tentative threshold ${E_K/E_M = 0.7}$ for the dipole collapse. Shaded areas indicate the dipolar (cyan) and multipolar (coral) branches proposed in this work.
}
\label{fig:Energydip}
\end{figure}
Figure~\ref{fig:Energydip} shows the dipolarity in our simulations as a function of this new proxy $E_K/E_M$. We find dipolar morphologies at low-$E_K/E_M$ and complex multipolar morphologies below equipartition (i.e., $E_K/E_M > 1$). These findings suggest that the energy ratio can likewise capture the dipolar-multipolar transition. It stands out that the energy ratio $E_K/E_M$ in the dipolar cases with $N_\rho = 1$ are significantly smaller than those obtained for the other density contrasts. This behaviour reflects what was already seen in Figure~\ref{fig:FBdip} using the force ratio, providing further evidence that $\mathcal{F}_{I}/\mathcal{F}_{L}$ and $E_K/E_M$ are indeed correlated. This occurs because the magnetic energy generated in these models is 2-6 times larger than the ones reached by other dipolar simulations in the same range of supercriticality (and hence with similar $E_K$). The shaded areas in Figure~\ref{fig:Energydip} show the tentative dipolar (cyan) and multipolar (coral) branches, along with a transitional region (grey) set to match the uncertainties of $E_K/E_M$ in the runs falling in the transition. From the data, we derive that the dipole breakdown occurs around ${E_K/E_M \simeq 0.7}$ (vertical dashed line).
As we will discuss in Section~\ref{sec:conclusions}, one advantage of the energy ratio is that we can use stellar observations to estimate $E_K/E_M$ at the stellar surface. Such an observational quantity is not strictly speaking identical to the definition in Eq.~\ref{eq:ekem}, and we could instead compute $E_K/E_M$ at the surface of our numerical simulations. However, the surface of numerical simulations differs from the surface of stars because boundary conditions constrain the field and flow. Moreover, 3D anelastic dynamo simulations better reflect the physics of the stellar convective envelope when excluding the outer few per cent of the radial domain as the anelastic approximation loses its validity at the stellar surface. For those reasons, we believe that volume-averaged energies are more adequate when drawing a parallel between numerical simulations and observations in Section~\ref{sec:conclusions}.
\subsubsection{Differential rotation} \label{sec:dyngen}
Stellar observations can give access not only to the surface magnetic fields in stars but also on some flow characteristics, like the surface differential rotation \citep[e.g.,][]{DMP08,MDP08}. Since we can measure in detail the differential rotation obtained in our calculations, we propose here to determine the amplitude and sign of the latitudinal differential rotation obtained in our dipolar and multipolar dynamo simulations. This will be used mostly for a comparison to the observations discussed in the following section.
Although numerical studies usually compute the latitudinal shear as the difference between the angular velocity at the equator minus an arbitrary latitude close to the poles, this parameter strongly depends on the chosen polar latitude as fast zonal flow variations may exist. Therefore, we compute the relative surface shear using a less dependent definition based on the difference between the angular velocity averaged on the near-surface layer (NSL) at equatorial regions and polar regions:
\begin{equation}\label{eq:shear}
\chi_\Omega = \frac{\langle\mean{\Omega}\rangle_{\mathrm{NSL}, |\theta|<40^o} - \langle\mean{\Omega}\rangle_{\mathrm{NSL}, 40^o<|\theta|<80^o}}{\Omega_o}.
\end{equation}
Here, we define as NSL the outer shell with thickness 0.05 $r_\mathrm{o}$ and we exclude high latitudes with $|\theta|>80^o$ from our computations (where small scale features are observed but should likely average out if considering longer time averages).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/SurfShearRol.pdf}
\caption{Dipolarity as a function of the differential rotation measured at the surface. The dashed vertical line represents a solid body rotation using our shear definition in Eq.~\ref{eq:shear}. Simulations with negative (positive) $\chi_\Omega$ display antisolar differential rotation profiles, while those with positive $\chi_\Omega$ have solar-like differential rotation profiles.}
\label{fig:DRsurf}
\end{figure}
Figure~\ref{fig:DRsurf} shows the dipolarity as a function of the relative latitudinal shear at the near-surface layer (cf. Eq.~\ref{eq:shear}). We want to emphasize that a non-dimensional quantity is used here to quantify the shear since the value of $\Omega_0$ in physical units is not set \emph{a priori} in our simulations. However, note that the same Ekman number (and thus the same $\Omega_0$) is used in all simulations so that the trend would be similar if only the numerator of Eq.~\ref{eq:shear} was used as the x-axis of Figure~\ref{fig:DRsurf}. The first striking feature is that all simulations exhibit a rather weak level of differential rotation with $\chi_\Omega < 2\%$. This quenching on the differential rotation can be understood because magnetic stresses are always active in our calculations as Lorentz forces significantly impact the flow \citep{COG99,B02}. Another important result is that the level of surface differential rotation is not negligible in dipolar cases, especially at $N_\rho=3$, compared to the multipolar ones. However, an important difference between dipolar and multipolar simulations is the differential rotation sign. Figure \ref{fig:DRsurf} indeed reveals that all simulations with dipole dominated morphology build an antisolar differential rotation profile. We find that non-negligible relative shears exist in our dipolar cases, with $\chi_\Omega$ ranging from $-0.57$ to $-0.03\%$. We note that these antisolar profiles were also observed in the geodynamo simulations of \citet{A05} and only illustrate the fact that the Lorentz force plays a significant role here in the angular momentum transport. On the other hand, solar-like differential rotation profiles only show up in the multipolar simulations. The only three multipolar cases developing antisolar profiles are those with $N_\rho = 1.5$, whose dipoles are either reversing or excursioning. The equatorial acceleration seen in the multipolar cases is consistent with the fact that it is only in this situation that inertia becomes comparable to Lorentz forces, as discussed in Section~\ref{sec:diptran}. This finding is in line with the non-magnetic simulations of \cite{GYM14}, where solar-like profiles are found when Reynolds stresses are enhanced. Indeed, the Reynolds stresses, associated with inertial forces, are known to be responsible for the equatorial acceleration of the flow \citep{M05}. They thus need to be significant enough to counteract the angular momentum transport by magnetic fields. When considering the multipolar simulations with solar-like differential rotation, we find that equatorial regions indeed accelerate, with values going up to $1.5\%$.
\section{Discussion and Conclusions} \label{sec:conclusions}
This paper explored through 3D dynamo simulations the physical mechanisms responsible for controlling the magnetic morphology of large-scale fields in partly convective cool stars. To address this point, we carried out 23 simulations of a spherical convective rotating shell with a radius ratio of 0.6 between the bottom and the top of the shell. Our modelling strategy follows recent geodynamo studies of \citet{MPG20} and \citet{TGF21}, who suggested that having a significant Lorentz force contribution in the force balance when simulating convective dynamos could modify conclusions about the magnetic morphology. However, unlike their study, we considered a fluid layer with a density contrast between the top and bottom of the convective zone to model conditions applicable to stellar interiors.
Our simulations demonstrate for the first time that axial dipole dominated solutions can be achieved at large Rossby numbers in stratified systems (up to $Ro_\ell=0.4$). Even more important maybe is the fact that these dipoles at high $Ro_\ell$ are obtained for simulations with a large density contrast between the top and bottom of the convective zone, at $N_\rho=3$. This finding differs from previous numerical studies suggesting that dipolar dynamos would only exist at low-Rossby numbers \citep[e.g.,][]{CA06,GDW12} and that strong stratification may make it more difficult for dipoles to survive.
In the same vein, \citet{RPD15} have also suggested that dynamos may be obtained for strong stratification, but we here extend the validity of their result to $Ro > 0.1$. In particular, it represents an important step towards the understanding of the magnetic morphology of stars, as strong axial dipoles have been likewise observed in some stars with $Ro_\ell > 0.1$, e.g., \mbox{TYC~5164-567-1} \citep[$f_\mathrm{dip}=0.77$; ][]{FPB16}, V439~And \citep[$f_\mathrm{dip}=0.60$; ][]{FPB16}, HD~6569 \citep[$f_\mathrm{dip}=0.53$;][]{FBP18}, and CE~Boo \citep[$f_\mathrm{dip}=0.76$;][]{DMP08}. We note that we also find solutions at $N_\rho=1.5$ with flipping or excursioning dipoles, producing measures of the dipolar fraction which can significantly vary in time. This could potentially be reminiscent to the strong variations in the dipolar and quadrupolar modes observed in the Sun \citep{DBH12} or other solar-like stars over their magnetic cycle \citep[e.g.,][]{PDS08,BLJ18}, all falling under the high Rossby regime.
Taken together, our parameter survey evidenced that the Rossby number cannot capture the transition in the surface field morphology when the Lorentz force is strong. We explored the possible mechanisms causing the axial dipole collapse using the relative amplitude of the axial dipole at the surface to measure the magnetic morphology in our simulations (cf. Eq.~\ref{eq:fdip}). From the investigation of the flow configuration, there was no evidence of its influence on the magnetic morphology. These findings can be understood by the significant back reaction of the magnetic field on the flow through the Lorentz force. As argued in the early study of \citet{GOD17}, the flow configuration only emerges as a good proxy of the magnetic morphology when the flow transitions are similar to those observed in hydrodynamical simulations. Indeed the force balance analysis shows a significant Lorentz force contribution to the flow dynamics in our calculations.
An important finding that emerged from the force balance study is that the ratio between the inertial and magnetic forces can describe the dipole-multipole transition of dynamo models with a background density contrast. We found that the dipole branch is recovered when the Lorentz force dominates over the initial force, with the transition to multipolar branch occurring around $\mathcal{F}_{I}/\mathcal{F}_{L} \simeq 0.4$. Similar to the conclusions obtained in past anelastic studies, it remains valid that the increased influence of inertia on the flow is responsible for destabilizing the axial dipoles. However, our work shows that instead of the traditional comparison with the Coriolis force (through the Rossby number), it is the relative importance of inertia compared to the Lorentz force that controls the transition if the magnetic back reaction on the flow is strong. With similar conclusions drawn by recent geodynamo simulations with $N_\rho= 0$ \citep{MPG20}, $\mathcal{F}_{I}/\mathcal{F}_{L}$ seems to emerge as a reliable predictor of the magnetic field morphology of stars and planets.
Because a direct estimate of the actual forces at play is not practical in stellar interiors, we explored an alternative proxy based on the ratio of kinetic to magnetic energies \citep{TGF21}. The investigation of $E_K/E_M$ revealed dipolar and multipolar branches confirming the ability of $E_K/E_M$ to describe the dipole collapse in numerical simulations \citep[early proposed by Boussinesq simulations;][]{KC02,TGF21}. From our data set, we found that stratified systems emerge as multipolar dynamos whenever $E_K/E_M \gtrsim 0.7$.
To tentatively test this proxy with observations, we gathered from the literature partly-convective stars with large-scale surface magnetic fields reconstructed using the Zeeman-Doppler imaging technique \citep[for details of the technique see, e.g.,][]{DSC97,DB97,DHJ06}. Given that our simulations correspond to a convective shell spanning the outer $40\%$ of the radial domain, we focused on partly convective M dwarfs with masses ranging from 0.38 to 0.60\,$M_\odot$, whose convective zones are expected to feature radius ratios (between the bottom and top of the convective zone) ranging from 0.50 to 0.66 \citep[estimated with the ATON code, described in][]{LVD06}, i.e., with roughly the same extension as those modeled in our simulations. We consider for consistency the homogeneous sample of stars published by \citet{DMP08} and \citet{MDP08}, which had their surface magnetic maps reconstructed with the same Zeeman-Doppler imaging code. We find eight stars obeying the mass condition described above: GJ~182, DT~Vir, DS~Leo, GJ~49, OT~Ser, CE~Boo, AD~Leo, and EQ~Peg~A. We also take into account multiple magnetic field reconstructions existent for DT~Vir, DS~Leo, and OT~Ser (with each star being observed at two different epochs).
From their magnetic surface maps, we directly derive $E_M$ based on the averaged surface magnetic field ($B_\mathrm{rms}$) and a modified dipolarity that is comparable to our definition in Eq.~\ref{eq:fdip} but with a maximum spherical harmonic degree that varies depending on the spatial resolution achieved for each star (typically $\ell_\mathrm{max}$ ranged from 6 to 10). We find that under our morphology classification CE Boo, AD Leo, and EQ~Peg~A fall in the criteria of dipolar dynamos ($f_\mathrm{dip}$ = 0.76, 0.57, and 0.57, respectively), while the other stars harbour a multipolar dynamo. Because observations only have access to the magnetic energy at the surface, we accordingly estimate the surface kinetic energy $E_K$ to compute the energy ratio of each star. We use published values of mass $M_\star$ and radius $R_\star$ present in the original Zeeman-Doppler imaging study. We adopt a rough approximation for the turbulent velocity $u_\mathrm{rms} = R_\star/\tau_c$ and photospheric density $\rho_{\star,\mathrm{pho}} = \frac{\bar{\rho}_\star}{\bar{\rho}_\odot}\rho_{\odot, \mathrm{pho}}$, where $\tau_c$ is the convective turnover time derived with the empirical relationships based in the stellar mass $M_\star$ \citep{WNW18},
$\bar{\rho}_{\odot, \star} = M_{\odot, \star}/(4\pi R_{\odot, \star}^3/3)$ is the mean density, and $\rho_{\odot,\mathrm{pho}}\approx 10^{-6}\,g\,cm^{-3}$ is the Sun's photospheric density \citep{BS05}. We thus estimate
\begin{equation} \label{eq:empiralEkEm}
\frac{E_K}{E_M} = \frac{\rho_{\star,\mathrm{pho}} u_\mathrm{rms}^2}{2} \frac{8\pi}{B_\mathrm{rms}^2} \approx \frac{4\pi}{B_\mathrm{rms}^2} \left(\frac{M_\star}{M_\odot} \right) \left(\frac{R_\odot}{R_\star}\right)^3 \left(\frac{R_\star}{\tau_c} \right)^2 \rho_{\odot,\mathrm{pho}}.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Dipolarity_ekem_OBS.pdf}
\caption{Observational counterpart of Figure~\ref{fig:Energydip}. Symbols show the magnetic properties of the M dwarfs derived with the Zeeman-Doppler imaging technique \citep{DMP08,MDP08}. The symbol size correspond to the field strength at the surface $\langle B \rangle$, the shape corresponds to the degree of axisymmetry of the magnetic field, and colors represent the amount of energy stored in the poloidal field. Shaded areas are similar to Figure~\ref{fig:Energydip}, with cyan representing strong dipoles axisymmetric fields (top left) and coral the multipolar non-axisymmetric fields (bottom right). However, we use a dipole-multipole transition of $E_K/E_M = 0.35$ (vertical dashed line) that is lower than the one obtained with simulations ($E_K/E_M = 0.7$). }
\label{fig:dipOBS}
\end{figure}
Figure~\ref{fig:dipOBS} illustrates the magnetic properties of M~dwarfs as a function of the energy ratio computed with Eq.~\ref{eq:empiralEkEm}. The sharp transition in the magnetic morphology is apparent from this plot. We find that M dwarfs with $E_K/E_M \lesssim 0.35$ have surface large-scale magnetic fields that are mostly poloidal and with strong axisymmetric dipoles. In contrast, M dwarf stars with higher energy ratios $E_K/E_M$ host large-scale fields with strong toroidal fields and weak axial dipoles. We infer a dipolar-multipolar transition around $E_K/E_M \simeq 0.35$ (dashed vertical line) from the observational data. As we considered volume-averaged energies instead of surface-averaged energies in our simulations (see details in Section~\ref{energy_proxy}), it is not surprising that observations show a dipole collapse at a different value than the one predicted from our simulations. Despite that, it is encouraging to see that an energy ratio proxy also seems to describe the transition in the magnetic morphology of M~dwarfs. Future simulations with different sizes of the convective envelope will help assess whether the dipole collapse is sensitive to this parameter and, therefore, if it is a potential source of uncertainties when determining the $E_K/E_M$ threshold.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{figs/Dipolarity_DR_OBS.pdf}
\caption{Dipolarity as a function of the surface differential rotation $d\Omega$ measured for a sample of M~dwarfs \citep{DMP08,MDP08}. The surface differential rotation is defined as $d\Omega = \Omega_\mathrm{eq} - \Omega_\mathrm{pol}$, where $\Omega_\mathrm{eq}$ is the angular velocity at the equator and $\Omega_\mathrm{pol}$ at the pole. Symbols are defined as in Figure~\ref{fig:dipOBS}.}
\label{fig:dipOBSshear}
\end{figure}
Finally, we explored the surface shear achieved in our simulations. We identified that, although quite weak, simulations with multipolar surface magnetic fields favour solar-like differential rotation profiles. In contrast, all dipole dominated simulations yield antisolar differential rotation \citep[similar to][]{A05,DSB06}. Here, we can also draw an observational parallel as surface shears have been measured for some of the stars in Figure~\ref{fig:dipOBS} \citep{DMP08,MDP08}. Because our numerical simulations have a constant rotation period, the transition in the magnetic field morphology with the relative shear reflects the change in the surface shear (Figure~\ref{fig:DRsurf}). Therefore, we use the latitudinal surface shear $d\Omega$ rather than the relative shear as the relevant parameter to consider for observations when the rotation period varies from star to star (from 1 to 9\,d in our sample). Figure~\ref{fig:dipOBSshear} shows the link between the axial dipole contribution to the large-scale magnetic morphology and the measured latitudinal surface shear for M~dwarf stars. The data in Figure~\ref{fig:dipOBSshear} give hints of a sharp transition in the magnetic complexity of M~dwarfs with the increase of $d\Omega$, with strong dipoles preventing significant latitudinal differential rotation at the surface and multipoles co-existing with large latitudinal surface shears. We note that this observational trend also extends to fully convective stars, with those harboring strong dipoles almost rotating as solid bodies, i.e., $d\Omega \sim 0$ \citep{DFC06,MDP08}. However, contrary to the trend in our simulations, we find that the dipole collapses at positive shears for M~dwarfs ($d\Omega \sim 55$\,mrad\,d$^{-1}$). Moreover, none of the stars from \citet{DMP08} or \citet{MDP08} had an antisolar differential rotation \citep[akin to other shear detection in M dwarfs, e.g.,][]{HDD16,ZVC20}. The direct comparison between observations and simulations is thus slightly less straightforward when shear profiles are concerned. It remains therefore to be investigated whether lowering the viscosity and magnetic diffusivity in our simulations can modify the differential rotation profile. For instance, it would be important to test if the antisolar regime found in the present calculations survives in more realistic parameter ranges. Further research is thus necessary to investigate how smaller Ekman numbers and/or larger magnetic Reynolds numbers can impact the transition seen in the differential rotation profile and amplitude.
The parameter space explored in this study offers new insights into the mechanisms controlling the magnetic morphology of stars. Our 3D dynamo simulations show that the magnetic morphology of the large-scale field depends on how much the Lorentz force is able to impact the flow. Although we cannot exclude the possibility that stronger anelastic effects in stars may modify this conclusion, we found that the energy ratio proxy proposed in our work to describe the transition in the magnetic morphology indeed succeeds at describing the varying large-scale magnetic topology of a small sample of M dwarfs featuring similar convective zone geometries, and for which a homogeneous collection of ZDI measurements is available in the literature. This first result leaves room for further numerical explorations aimed at studying the impact of more parameters, such as the size of the convective zone and the rotation rate. These simulations will broaden potential comparisons with stars of different spectral types than the ones considered here, and therefore to further investigate whether the proxy that we propose can be used in a more general context. We also leave for a forthcoming paper the study of whether a radiative interior in the numerical domain is also able to impact the magnetic morphology of the large-scale field and its transition from a mainly dipolar to a mainly multipolar structure, and to modify the conclusions reached here regarding the proxies that best describe where this transition occurs in the parameter space.
|
\section{Introduction}
Randomised clinical trials (RCTs) are regarded as the gold standard for evaluating the effectiveness of a new treatment. In RCTs with time to event outcomes, the primary outcome is to measure the differences in the survival curves of the different arms and report the treatment effect to quantify treatment differences. The analysis of such trials are routinely accomplished by the log rank test and the Cox proportional hazard (PH) model. The hazard ratio is obtained from the Cox PH model to measure the effectiveness of the new treatment compared to the control group. The Cox PH model assumes that the hazard ratio between the treatment and control group stays constant from the start of study period until the end of the follow up period. Under PH, the log rank test is the most powerful test to detect differences in the survival curves \cite{FH}. However, non proportional hazards (NPH) are commonly observed in RCTs and have been reported in many published trials e.g. the IPASS trial in lung cancer \cite{mok2009gefitinib} and the ICON7 trial in ovarian cancer \cite{kristensen2011result} . When the PH assumption does not hold, the log rank test is still valid but may suffer substantial power loss and the interpretation of the hazard ratio becomes invalid and challenging \cite{hernan2010hazards}. There may be many reasons for non PH to exist in clinical trials, for instance due to delayed clinical effect in the case of immuno-oncology trials.
Various statistical methods have been proposed to analyse time to event outcomes in clinical trials under non proportional hazards scenarios including restricted mean survival time (RMST) \cite{royston2013restricted}, weighted log rank tests \cite{fleming1981class} and maximum combination test (MaxCombo) \cite{lin2020alternative}. The RMST measures the average survival time from time $t= 0$ to some pre-specified time horizon $t=\tau$ and it may be estimated as the area under the survival curve up to that time point. The weighted log rank test based on the Fleming-Harrington class of weights $G^{\rho,\gamma}$ with weight $S(t)^{\rho}$ $(1-S(t))^{\gamma}$ where $S(t)$ is the survival function of the pooled patients from control and treatment group; the parameters $\rho$, $\gamma$ allows one to down weight early, late or middle events. The MaxCombo test uses the Fleming-Harrington weight function to analyse time to event outcomes in the presence of non PH. The test avoids having to pre-specify the weight function and considers a procedure which selects the best test from the set of test statistics. The procedure consist of selecting the best combination of weighted log rank test statistics with different choices of $\rho$ and $\gamma$. When the MaxCombo approach is used, Lin et al \cite{lin2020alternative} suggest using the weighted hazard ratio for effect estimation. The estimated effect is obtained from a Cox model with a time-varying effect, where the weights are those associated with the weighted log rank test \cite{lin2017estimation}.
This paper focuses on the estimation of treatment effect under non proportional hazards, particularly the weighted hazard ratio method proposed by Lin and León \cite{lin2017estimation}. They propose fitting a particular Cox model with a time-varying covariate/effect for which the score test corresponds to a weighted log rank test. The model proposes an effect adjustment factor $A(t)$ = $\frac{w(t)}{max(w(t))}$ where $w(t)$ is the weight function from the chosen weighted log rank test. This is incorporated in to the Cox PH model to provide time varying effect which can be viewed as the treatment coefficient $\beta$ weighted by the adjustment factor $A(t)$. The hazard ratio obtained from the model is expressed as $HR(t) =e^{\beta A(t)}$ = $[HR^{F}]^{A(t)}$ where $HR^{F}$ represents the maximal effect at time $t$ with $A(t)=1$. In this paper, we are interested in investigating whether the estimated treatment effect is unbiased under scenarios in which the associated weighted log rank test is the most powerful.
In Section 2, we describe the standard weighted log rank test and the proposed weighted hazard ratio estimation \cite{lin2017estimation}
and its estimation of the time-varying HR. In Section 3, we compare the hazard ratio functions obtained from the proposed model with the true hazard ratio functions analytically. In Section 4, we assess, through simulations, the performance of the proposed model under the two non PH scenario with different choices of $\rho$ and $\gamma$. Section 5 is a discussion.
\section{Weighted log rank tests and effect estimation}
In this section, we will provide a brief overview of the log rank test and the weighted log rank test, as well as introduce weighted hazard ratio for effect estimation proposed by Lin and León \cite{lin2017estimation}.
\subsection{Weighted log rank test}
Let $t_{1}$,$t_{2}$,\ldots,$t_{r}$ be the ordered failure times across both treatment arms, $d_{1j}$ and $d_{2j}$ be the number of deaths at $t_{j}$ in treatment and control arm respectively with $d_{j}$= $d_{1j}$ +$d_{2j}$ and $n_{1j}$ and $n_{2j}$ be the number of individuals at risk before $t_{j}$ with $n_{j}$ = $n_{1j}$ + $n_{2j}$. Let $X_{1}, X_{2},\dots,X_{r}$ denote the treatment assignment where $X_{j}$=1 if the j-th subject is assigned to treatment arm and $X_{j}$=0 if assigned to control arm. The log-rank test statistic is defined as,
\begin{align*}
\begin{split}
Z= \frac{\sum_{j=1}^{r}(d_{1j} - e_{1j})}{(\sum_{j=1}^{r}V_{L})^{1/2}}
\end{split},
\end{align*}
where $e_{1j}$= $\frac{n_{1j} d_{j}}{n_{j}}$ is the expected number of failure at time $t_{(j)}$ under the null hypothesis of equal survival curves. $V_{L}$ is the variance of the observed number of failures, $V_{L}$ = $ \sum_{j=1}^{r}\frac{n_{1j}n_{2j}d_{j}(n_{j}-d_{j})}{n_{j}^{2}(n_{j}-1)}$. Under the null hypothesis of equal survival functions in the two treatment groups, asymptotically the test statistic has a chi-squared distribution on one degree of freedom. Although the log rank test is still a valid test under non proportional hazards it is not the most powerful test \cite{royston2013restricted}. Therefore, the weighted version of log rank test is sometimes utilised when the ratio of the two hazard functions are not constant over time \cite{fleming1981class}.
The weighted log rank test (WLRT) assigns a weight function to different time points depending on the expected type of non-proportional hazard scenario. For instance, in immuno-oncology trials there may be a delayed treatment effect and therefore the survival curve for the treatment group will only emerge to separate from the control survival curve after a certain period of time. In this case, higher weights can be allocated to later time points \cite{lin2020alternative}. Formally, the WLRT is defined as,
\begin{equation}
\begin{split}
\label{WLRTST}
W_{K}=
\frac{\sum_{j=1}^{r}w(t_{j})(d_{1j} - e_{1j})}{(\sum_{j=1}^{r}V_{L})^{1/2}}
\end{split},
\end{equation}
with variance
\begin{align*}
\begin{split}
V_{K} = \sum_{j=1}^{r}w(t(j))^{2}\bigg(\frac{n_{1j}n_{2j}d_{j}(n_{j}-d_{j})}{n_{j}^{2}(n_{j}-1)}\bigg)
\end{split}
\end{align*}
where $w(\cdot)$ is the non-negative weight function. We will consider the Fleming-Harrington family of weighted log rank test, commonly denoted as $G^{\rho,\gamma}$ with weight,
\begin{equation}
\label{weight}
w(t) = \hat{S}(t)^{\rho} (1-\hat{S}(t))^{\gamma}
\end{equation}
where $\hat{S}(t)$ is the Kaplan- Meier estimate of the survival function of the pooled patients from the treatment and control arm. The parameters $\rho$ and $\gamma$ control the shape of the weight function. $G^{0,0}$ represents the log-rank test with $w(t)=1$ that has constant weight over time; delayed treatment effects can be tested using $G^{0,\gamma}$ with $w(t)=(1-\hat{S}(t))^{\gamma}$ that allocates higher weight at later time points to detect late survival difference, $G^{1,0}$ represents diminishing effect with $w(t)=\hat{S}(t)$ that gives more weight the earlier time points to detect early separation; $G^{1,1}$ represents mid separation with $w(t)=\hat{S}(t) (1-\hat{S}(t))$ that puts more weight at the middle of the follow-up period than at the ends.
\subsection{Weighted hazard ratios}
Lin and León \cite{lin2017estimation} proposed fitting a Cox model to provide a time varying treatment effect estimate to complement the weighted log rank test. Lin and León propose fitting a Cox model with,
$$
\lambda(t;X) = \lambda_{0}(t)e^{A(t)\beta X}
$$
where $A(t)$ is defined by,
$$
A(t) = \frac{w(t)}{\textrm{max}(w(s))}
$$
$\textrm{max}(w(s))$ is evaluated over the values of $t$ in the observed dataset and $A(t)$ thus has a maximum value of 1. The time varying covariate $X^{*}(t) = A(t)X$ represents the treatment assignment weighted by the adjustment factor. Once $X^{*}(t)$ has been determined, the constant coefficient $\beta$ can be estimated.
It is shown in the paper that the score test of the null hypothesis that $\beta=0$ in the proposed model above is equivalent to the weighted log rank test with the corresponding choice of weight function \cite{lin2017estimation}. Lin and León used the fact that the score test from this Cox model corresponds to a particular weighted log rank to motivate estimating the time-varying treatment effect using this Cox model. The coefficient $\beta$ can be interpreted as the maximal effect where the model assumes the patients experience the most benefit ($A(t)=1$). To differentiate the time varying hazard ratio derived from the proposed method, we will denote this as $HR_{LL}$ and it can be expressed as,
\begin{equation}
HR_{LL}(t)= \frac{\lambda_{0}(t)e^{A(t)\beta\times1}}{\lambda_{0}(t)e^{A(t)\beta\times0}} = e^{\beta A(t)} = \big[HR^{F}\big]^{A(t)}
\label{HRlin}
\end{equation}
where $HR^{F}$=$e^{\beta}$ represents the maximal effect (i.e., at time $t$ with $A(t)$=1). The time profile of the treatment effect can be observed as the treatment coefficient $\beta$ weighted by the adjustment factor $A(t)$. The estimate derived from the model provides a consistent time profile of the hazard ratio over time given that the true hazard ratio also varies over time as specified by $A(t)$. If this does not hold then the treatment effect measure will in general no longer be unbiased for the true value of the hazard ratio at a given time.
\section{Evaluation of weighted hazard ratio estimation method}
In this section we give expressions for the hazard ratio function for which the Fleming-Harrington \cite{FH} $G^{\rho,\gamma}$ test statistics are most efficient.
We will also be exploring whether, for a given weight function, the Lin and León estimation method \cite{lin2017estimation} gives you the correct treatment effect profile if the true data generating mechanism is such that the corresponding weighted log rank test is the most powerful.
\subsection{Diminishing effect}
Fleming and Harrington \cite{FH} showed that the $G^{\rho}$ test statistic is the most powerful test statistic when the hazard ratio at time $t$ is of the following form,
\begin{equation}
\label{eqn: efficient}
\frac{\lambda_{2}(t)}{\lambda_{1}(t)}=\frac{e^{\Delta}}{\{S_{1}(t)\}^{\rho} + e^{\Delta}[1-\{S_{1}(t)\}^{\rho}]}
\end{equation}
where $S_1(t)=\exp\left(-\int_{0}^{t}\lambda_{1}(x) \textrm{dx}\right)$ is the survival function for the control arm. The parameter $e^{\Delta}$ represents the hazard ratio at $t=0$. The parameter $\rho$ controls how quickly the effect diminishes. The above hazard ratio function \eqref{eqn: efficient} has its full treatment effect initially $e^{\beta}$ and decreases monotonically to 1. Later, we will demonstrate various examples with $\rho =0.5,1,2$ to further develop our understanding about the behaviour of the weighted logrank test and the proposed model.
We now calculate the weighted hazard ratio function proposed by Lin and León \cite{lin2017estimation} under the set up where the $G^{\rho}$ test statistic would be optimal. In order to calculate the proposed weighted hazard ratio function under the set up where WLRT is optimal, we will need to derive the true pooled survival function which is determined by finding the survival function of the two treatment arms. The hazard function for the treatment arm $\lambda_{2}(t)$ can be derived, assuming that the true data generating mechanism is such that that WLRT is optimal, by using equation \eqref{eqn: efficient}
\begin{align*}
\begin{split}
\lambda_{2}(t)& =\frac{\lambda_{1}(t)e^{\Delta}}{\{S_{1}(t)\}^{\rho} + e^{\Delta}[1-\{S_{1}(t)\}^{\rho}]} \\
\end{split}
\end{align*}
Thus, the survival function for the treatment arm is,
\begin{equation}
\begin{split}
\label{survivalTRT}
S_{2}(t) &= \exp\left(-\int_{0}^{t}\lambda_{2}(x) \textrm{dx}\right) \\
&= \exp\left(-\int_{0}^{t} \frac{\lambda_{1}(x)e^{\Delta}}{\{S_{1}(x)\}^{\rho} + e^{\Delta}[1-\{S_{1}(x)\}^{\rho}]} dx\right)
\end{split}
\end{equation}
We can replace the adjustment factor $A(t)$ with the weight function that the Lin and León method converges to (as $n$ tends to $\infty$) \eqref{HRlin}, in this case $S(t)$. For the Lin and León \cite{lin2017estimation} method, we estimate the pooled survival function $S(t)^{\rho}$ by the Kaplan-Meier estimator for the pooled sample which is consistent for the true pooled survival function. Therefore the Lin and León method consistently estimates,
\begin{equation}
\label{eqn:hr}
\begin{split}
HR_{LL}(t) = e^{\beta A(t)} = e^{\beta\big(0.5S_{1}(t) + 0.5S_{2}(t)\big)^{\rho}}
\end{split}
\end{equation}
Substituting $S_{2}(t)$ in the hazard ratio function \eqref{eqn:hr} for which this WLRT is optimal for gives,
\begin{equation}
\begin{split}
\label{LLeff}
HR_{LL}(t) = \textrm{exp}\left[\beta \left(\frac{1}{2}S_{1}(t)\right) +\frac{1}{2}\exp\left(-\int_{0}^{t} \frac{\lambda_{1}(x)e^{\Delta}}{\{S_{1}(x)\}^{\rho} + e^{\Delta}[1-\{S_{1}(x)\}^{\rho}]} dx\right)\Bigg)^{\rho}\right]
\end{split}
\end{equation}
When fitting the proposed model, the $\beta$ is estimated from the Cox PH model once the adjustment factor $A(t)$ is determined.
Comparing the true hazard ratio function \eqref{eqn: efficient} with the proposed hazard ratio function \eqref{LLeff}, we see that the two expressions do not appear to be equal to each other. In Appendix B, we demonstrate that the two hazard ratio functions are approximately equal to each other when $\Delta$ is close to 0. We also demonstrate this equality graphically in the next section where we choose the value of $e^{\Delta}$ to compare the hazard ratio functions.
We also compare the hazard ratio expressions analytically for some particular choices of $\lambda_1(t)$ and $e^{\Delta}$ to illustrate in Appendix A.
\subsection{Delayed treatment effect}
The paper by Garès \cite{gares2017fleming} has provided an expression for the hazard function for which the Fleming-Harrington test $G^{\gamma}$ for detecting late effects is optimal. Assume, the survival time in the control arm is exponentially distributed with some constant hazard $\lambda_{1}$. Then Garès et al \cite{gares2017fleming} show that the $G^{\gamma}$ test is optimal for testing the hypothesis,
\begin{equation}
\begin{split}
H_{0}: \lambda_{CT}&= \lambda_{TRT}= \lambda_{1}\\
H_{1}: \lambda_{TRT}(t)&=\lambda_{CT} \frac{L^{\gamma}((\mathcal{L^{\gamma}})^{-1}(\mathcal{L^{\gamma}}(e^{-\lambda_{1}t})+\varphi))}{L^{\gamma}(e^{-\lambda_{1}t})}
\label{haz2}
\end{split}
\end{equation}
where $L(x)=\int_{x}^{1}\frac{(1-s)}{s} ds= 1
+\ln(x)-x$ and $\mathcal{L}(x)=\int_{0.5}^{x} \frac{1}{s\log(s)+s-s^{2}} ds$. The expression for the hazard ratio has no closed form and needs numerical integration techniques to compute. Garès et al \cite{gares2017fleming} has provided the definition for the parameter $\varphi$ which is evaluated at some chosen time point $\tau$ and is of the following form,
\begin{equation}
\begin{split}
\label{delta}
\varphi&= \mathcal{L}^{\gamma}(r(1-S_{1}(\tau))+S_{1}(\tau))-\mathcal{L}^{\gamma}(S_{1}(\tau)) \\
&= \mathcal{L}^{\gamma}(S_{2}(\tau))- \mathcal{L}^{\gamma}(S_{1}(\tau))
\end{split}
\end{equation}
where $r$ is the discrepancy rate, $r=\frac{S_{2}(\tau)-S_{1}(\tau)}{1-S_{1}(\tau)}$.
Similar to the diminishing effect scenario, we can check whether the true hazard ratio for which the $G^{\gamma}$ test is optimal, as given in equation \eqref{haz2}, agrees with the hazard ratio estimation proposed by Lin and León \cite{lin2017estimation}.
The true hazard ratio under which the $G^{\gamma}$ test statistic is optimal is shown to be the second equation in \eqref{haz2}. Now the the weighted hazard ratio model proposed by Lin and León corresponds to,
\begin{align*}
\begin{split}
HR_{LL}(t) =e^{\beta A(t)} &= e^{\beta \frac{w(t)}{max(w(s))}}\\
&= e^{\beta \left(\frac{\left(1-S(t)\right)^{\gamma}}{max(w(s))}\right)}=e^{\beta\left(\frac{(1-0.5S_{1}(t)-0.5S_{2}(t))^{\gamma}}{max(w(s))}\right)}
\end{split}
\end{align*}
From Garès et al, the survival in the treatment and control arm is defined as $S_{2}(t)=(\mathcal{L})^{-1}(\mathcal{L}(e^{-\lambda_{1}t}) +\mathcal{L}(S_{2}(\tau))-\mathcal{L}(S_{1}(\tau)))= (\mathcal{L})^{-1}(\mathcal{L}(e^{-\lambda_{1}t})+ \varphi)$ and $S_{1}(t)=e^{-\lambda_{1}}t$ \cite{gares2017fleming}. We now substitute in for $S_{1}(t)$ and $S_{2}(t)$,
\begin{equation}
\begin{split}
HR_{LL}(t)&=e^{\beta \left( \frac{\left(1-0.5e^{-\lambda_{1}t}-0.5\mathcal{L}^{-1}(\mathcal{L}(e^{-\lambda_{1}t})+\varphi)\right)^{\gamma}}{\left(1-0.5e^{-\lambda_{1}\tau}-0.5\mathcal{L}^{-1}(\mathcal{L}(e^{-\lambda_{1}\tau})+\varphi)\right)^{\gamma}}\right)}
\label{LLHR}
\end{split}
\end{equation}
We assume that the Lin and León method estimates the correct hazard ratio ($e^{\beta}$) at the time point in the study where it is the biggest effect i.e. at $t=\tau$. Like in Section 3.1, we show that if the Lin and León method estimates the correct hazard ratio at the time point when it is largest, then it will estimate the HR at all other times correctly too. In Appendix C, we see that the expression for the hazard ratio function from Lin and León model \eqref{LLHR} does approximately match the true HR assumed as per equation \eqref{haz2} under the condition that $\varphi$ is small.
\section{Simulation studies}
This section investigates the weighted hazard ratio method under diminishing and delayed effect using simulations \cite{lin2017estimation}. We will focus on investigating the method proposed by Lin and León and its properties when non proportional hazards are present in the data. Specifically, we will empirically check the analytical results of Section 3 that showed that the Lin and León approach is approximately unbiased when the treatment effect is small, and how large the bias might be when the treatment effect is larger.
\subsection{Scenario I: Diminishing effect}
We simulated the survival data where the treatment is assumed to have its maximal effect initially and decrease monotonically to 1. The survival function for the treatment arm is given in equation \eqref{survivalTRT}. Three different scenarios $\rho=0.5,1,2$ were considered to understand the behaviour of the proposed model under various rates at which the effect diminishes. We also chose three different values of the maximum effect $e^{\Delta}=1.4,4,8$ to investigate whether substantial bias occurs when the treatment effect is larger. We fitted the Cox model proposed by \cite{lin2017estimation} where the weight function $w(t)=\hat{S}(t)^{\rho}$ for the non PH data we simulated is specified for three scenarios $\rho=0.5,1,2$. The study enrolls 200 patients, equally allocated in the two arms. The survival of the control arm follows an exponential distribution with hazard rate $\lambda_{1}=0.5$. The survival in the treatment arm is as per equation \eqref{survivalTRT} . The follow up time is assumed to be 3 years; patients whose event time exceeds 3 years are censored. We conduced 5000 runs for each scenario.
Figure \ref{fig:hazard} shows the true hazard ratio over time for which the $G^{\rho}$ statistic is optimal and the average hazard ratio profile over time estimated by the Lin and Leòn method. The plot displays $e^{\bar{\hat{\beta}}S(t)^{\rho}}$ where $\bar{\hat{\beta}}$ is the average of the hazard ratio at $t=0$ estimated from the Lin and Leòn method and $S(t)$ is the true pooled survival function. The proposed model overestimates the maximal effect when the true $HR$ at $t=0$ i.e. $e^{\Delta}$ is large, and slightly underestimates immediately after 1 month of follow up. Note that when the treatment effect is small, the proposed model estimates the maximal effect very close to the true one, in line with our analytical results in Appendix B. We observe through our work that when the treatment effect is large, the proposed model is not able to estimate the maximal effect at $t=0$ correctly which has an impact on the treatment effect profile over time, in that case the implied profile of treatment effect over time would not match the true one (as given by equation \eqref{eqn: efficient}).
\vspace{3cm}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.7\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{Rplot3.eps}
\caption{$e^{\Delta}=1.4$}
\label{fig:y equals x}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.7\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{Rplot4.eps}
\caption{$e^{\Delta}=4$}
\label{fig:three sin x}
\end{subfigure}
\end{figure}
\begin{figure}[ht]\ContinuedFloat
\centering
\begin{subfigure}[b]{0.7\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{Rplot5.eps}
\caption{$e^{\Delta}=8$}
\label{fig:five over x}
\end{subfigure}
\caption{Hazard ratio over time targeted by weighted hazard ratio model superimposed true hazard ratio over time under a data generating mechanism where the $G^{\rho}$ test is optimal. The solid line shows the true hazard ratio over time for $\rho=0.5,1,2$. The true hazard ratio for the three scenarios $\rho=0.5,1,2$ at $t=0$ are $e^{\Delta}=1.4,4,8$ respectively and the hazard rate in the control arm stays the same in all the scenarios $\lambda_{1}(t)=0.5$. The dotted lines show the proposed hazard ratio targeted by the weighted hazard ratio model. The hazard ratio for the dotted lines is the average of the hazard ratios estimated from fitting the Lin and León method and is as follows; the average of the log hazard ratio across 5000 simulations is $e^{\bar{\hat{\beta}}}=1.41,4.07,8.15$,$e^{\bar{\hat{\beta}}}=1.41,4.07,8.40$, $e^{\bar{\hat{\beta}}}=1.39,4.16,9.06$ for $\rho=0.5,1,2$.}
\label{fig:hazard}
\end{figure}
\newpage
\subsection{Scenario II: Delayed treatment effect}
In this scenario, the data were simulated based on the assumption that the treatment will have no effect at the start of the follow up and gradually increases to have its maximal effect at the final time point of follow up period. The study enrolls the same number of patients as in the diminishing effect scenario i.e. 200. The maximal treatment effect occurs at the final time point $\tau=2$.
The survival times in the control arm are exponentially distributed with hazard rate 0.5. The survival times in the active arm can then be simulated using the expression for the hazard function given in \eqref{haz2}. To simulate the survival times in the active arm, we calculated the value of $S_{2}(t)$ for $t$ from 0 to $\tau$ in increments of 0.0005. We then simulated $U \sim U(0,1)$ and found the value of $t$ such that $S_{2}(t)$ was closest to $U$. We consider two scenarios of $S_{2}(\tau)=0.25$ and 0.1 which provides the values of discrepancy rate $r=-0.19$ and $-0.42$ for $\gamma=0.5,1,2$ to verify our results obtained in Section 3. The mechanism behind simulating the non proportional data in this way would ensure that the weight function $(1-S(t))^{\gamma}$ would provide the most powerful test statistic. Similar to the previous scenario, we applied the Lin and León Cox PH model to our simulated datasets and compared the estimated HR profile with the true HR profile over time. We again considered three scenarios of $\gamma$ where it is varied to control the rate of how quickly the treatment effect reaches its maximum effect.
The true hazard ratio over time for which the $G^{\gamma}$ statistic is optimal and the average hazard ratio profile over time estimated by the Lin and Leòn method is displayed in Figure \ref{fig:hazard_ratioplot}. The plot displays the three HR profiles over time for $\gamma=0.5,1,2$ for two different values of $S_{2}(\tau)$.
The solid lines represents the true HR profile and the dashed lines represents $e^{\bar{\hat{\beta}}(1-S(t)^{\gamma}}$. Note that for both values of $S_{2}(\tau)$, the estimated average hazard ratio is overestimated for $\gamma=1,2$ and this deviation is reflected on the HR time profile. Though for other scenarios, we see that the Lin and León method estimates the treatment effect profile over time with minimal bias, in line with our analytical results in Appendix C.The proposed model closely estimates the maximal treatment effect for both scenarios where $\gamma=0.5$ and $r$ is relatively small.
\begin{figure}[hb!]
\centering
\begin{subfigure}[b]{0.7\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{Rplot1.eps}
\caption{$S_{2}(\tau)=0.25$}
\label{fig:sc2}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.7\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{Rplot2.eps}
\caption{$S_{2}(\tau)=0.1$}
\label{fig:sc3}
\end{subfigure}
\caption{Hazard ratio over time targeted by weighted hazard ratio model superimposed true hazard ratio over time under a data generating mechanism where the $G^{\gamma}$ test is optimal. The solid line shows the true hazard ratio over time for $\gamma=0.5,1,2$. The true HR for $r=-0.19$ at $\tau=2$ is $e^{\beta}=1.555,1.73,2.112$ and for $r=-0.42$ is $e^{\beta}=2.999,3.813,5.935$ for $\gamma=0.5,1,2$ respectively. The dashed lines show the proposed hazard ratio targeted by the weighted hazard ratio model. The hazard ratio for the dashed lines is the average of the hazard ratios across 5000 simulations estimated from fitting the Lin and León method and is as follows; $e^{\bar{\hat{\beta}}}=1.558,1.732,2.138$, $e^{\bar{\hat{\beta}}}=2.998,3.803,6.070$ for $r=-0.19,-0.42$ and $\gamma=0.5,1,2$ respectively}
\label{fig:hazard_ratioplot}
\end{figure}
\newpage
\section{Discussion}
In this paper, we considered the Fleming-Harrington class of weights to simulate non-proportional hazard data for the which the same weight function would give the most optimal test. The Lin and León method consists of fitting a Cox model based on a time-varying treatment covariate. The score test from the proposed model gives a $p$-value which is equivalent to the weighted log rank test and the estimate
derived from the model provides a time-profile of the treatment effect. We considered two non proportional hazard scenarios to investigate its ability to correctly estimate the maximal treatment effect and the corresponding treatment effect profile over time when the data were simulated such that the corresponding weighted log rank test would be most powerful.
We showed analytically that the Lin and Leòn method is approximately unbiased in both scenarios when the treatment effect is small. Our simulation results for the diminishing treatment effect scenario showed that the Lin and León method overestimates the maximal treatment effect when the true treatment effect at $t=0$ is large as well as when $\rho$ is large. For scenarios where the treatment effect is small, the proposed model estimates the maximal effect very close to the true one. For the delayed treatment effect scenario, the Lin and León method overestimates the maximal effect for $\gamma=1,2$ but in our simulations not by very much. The Lin and León method gave essentially unbiased estimates of the full effect and treatment effect profile over time for $\gamma=0.5$.
We have shown through simulations that under the delayed treatment effect scenarios, the proposed model is giving close to the correct profile over time when treatment effect is small. However, this does not guarantee that the proposed model would provide unbiased results had the data generating mechanism been different, for example if the the hazard in the control arm was not constant and changes over time.
The choice of $\rho$ and $\gamma$ in estimating the treatment effect requires extensive knowledge of how the true hazard ratio changes over time. From a testing perspective, choosing the correct value of $\rho$ or $\gamma$ is less crucial since the test still controls the type I error rate if the incorrect value is used. However, for estimation of the treatment effect profile over time, use of the incorrect value would lead to biased estimates. Given the difficulty of knowing the correct value of $\rho$ and $\gamma$, it may be preferable to use methods such as restricted mean survival time which do not rely on how quickly or slowly the treatment effect changes over time, and for which the interpretation of the treatment effect is clinically meaningful to the clinicians and patients \cite{royston2013restricted}.
\section*{Acknowledgement}
Bharati Kumar’s research was funded by a UK EPSRC studentship (2281147). Jonathan Bartlett was supported by the UK MRC (grant MR/T023953/1).
|
\section{Introduction}
In recent, a number of domain adaptation (DA) methods~\cite{cao2020unsupervised,wang2019adversarial,lewis2019unsupervised} for machine reading comprehension (MRC) have been proposed, which usually pre-train an MRC model in a high-resource domain and then transfer it to a low-resource domain.
Specifically, most existing methods~\cite{cao2020unsupervised,wang2019adversarial,lewis2019unsupervised} consist of two steps. First, they construct some pseudo QA pairs by using a pre-trained QA construction model from the available documents in the target domain. Then they fine-tune the pre-trained MRC model by using the constructed pairs.
Although these methods have achieved promising results, almost all of them ignore the mismatched QA pairs, i.e., the irrelevant QA pairs which are wrongly treated as positive. In the scenario of MRC, such a so-called noisy correspondence (NC) issue~\cite{huang2021learning} is caused by the following reasons. First, the domain adaptation methods for MRC often construct pseudo QA pairs by using the documents in the target domain which does not contain natural questions. As a result, the generated questions will be probably irrelevant to the answers. Second, the QA construction model is pre-trained in the source domain and then directly applied to the target domain without fine-tuning. In consequence, such a domain shift issue will lead to noisy correspondence. A toy example of NC is shown in Fig.~\ref{fig:msn}, and more real-world samples generated by the existing works refer to Fig.~\ref{fig:case_study}. Notably, the NC problem is remarkably different from the well-studied noisy labels. To be specific, noisy labels generally refer to the category-level annotation errors of a given data point, whereas here NC refers to the mismatched relationship between two data points. Undoubtedly, NC will degenerate the performance of the MRC model, which is however neglected so far as we know.
\begin{figure*}[!hbt]
\centering
\subfigure[Noisy Labels vs. Noisy Correspondence] {
\label{fig:msn}
\includegraphics[width=0.46\textwidth]{figures/MSN.pdf}
}
\hspace{0.6cm}
\subfigure[Existing Works vs. Our Work] {
\label{fig:motivation}
\includegraphics[width=0.43\textwidth]{figures/difference.pdf}
}
\vskip -0.2cm
\caption{(a) Noisy Labels vs. Noisy Correspondence. The noisy labels refer to the errors in the category annotation of data samples caused by human annotation. The noisy correspondence here refers to the misalignment between two data points generated by the model itself. (b) An example of the document along with dialogue, and the difference between our method and the existing ones. The dialogue is the conversations between the questioner and the answerer (the customer and customer service here) about the document. In other words, the questioner raises a question about the document, and the answerer answers it by referring to the document. Hence, the dialogues have natural QA-form conversations that is helpful in QA pair construction. }
\vskip -0.4cm
\end{figure*}
To sum up, domain adaptation methods for MRC will face the NC challenge caused by i) the unavailable QA pairs in the target documents, and ii) domain shift during applying the QA construction model to the target domain. In this paper, we propose to construct more credible QA pairs by additionally using document-associated dialogues to fill the vacancy of the natural question information in the document. In addition, our method will fine-tune the QA construction model in the target domain with the help of the MRC model feedback. For clarity, we summarize the major differences between our MRC method and the existing ones in Fig.~\ref{fig:motivation} by taking the customer service as a showcase without loss of generality. As shown, in real-world applications, the customer service usually answers customers' questions by referring to the documents, forming an associated dialogue. Existing works only use the sole document for QA construction and do not further fine-tune the QA construction model in the target domain. In contrast, our work leverages both documents and the associated dialogues for QA construction. Dialogues are the conversation corpus between the questioners and answerers, which naturally preserves QA-similar chats and thus are more credible for QA construction. Moreover, our method will use the feedback of the MRC model on constructed QA pairs to optimize the QA construction model, thus alleviating the domain shift issue and improving the QA quality.
In practice, however, difficulties arise when an attempt is made to apply the above approaches. Specifically, although dialogues are more credible than documents for the QA construction, they still contain a huge number of irrelevant and discontinuous conversations. In other words, the above QA construction method could partially alleviate but still cannot solve the NC problem. As shown in Fig.~\ref{fig:motivation}, the irrelevant conversation is about the irrelevant conversations to the document, e.g., there are some greetings or other chats which are irrelevant w.r.t the document. The discontinuous conversation is that the question and answer are not exactly aligned in a single round of chat due to the complexity in interaction, e.g., the customer may raise a new question before receiving the answer for the last ones. Besides the challenges rooted in the dialogue data, another difficulty is how to fine-tune the non-differentiable QA construction model in the target domain to alleviate the domain shift issue. In brief, the existing works often generate QA pairs by resorting to the discrete $\mathop{\arg \max}$ sampling operator which hinders the optimization of the QA construction model.
To overcome the above challenges in data quality and model optimization, we propose a novel domain adaptation method for MRC, dubbed Robust Domain Adaptation for Machine Reading Comprehension (RMRC) which consists of an answer extractor (AE), a question selector (QS), and an MRC model. In brief, RMRC leverages both documents and the associated dialogue for MRC transfer by i) constructing QA pairs via the AE and the QS, and ii) alternative training the MRC model and the QS.
In the first stage, for a given document, RMRC extracts the candidate answers from dialogues and filters out unrelated ones in terms of the estimated relevance using the pre-trained AE, thus alleviating the NC issue caused by irrelevant chats. After that, for the extracted answer, RMRC seeks the most related questions that appeared in multiple rounds of chats via the pre-trained QS, thus tackling the NC problem caused by discontinuous conversations.
In the second stage, RMRC optimizes the MRC model and the QS in an alternative fashion, which will favor MRC transfer. In detail, the domain shift could be alleviated and accordingly the NC problem is tackled by optimizing the QS using the feedback of the MRC model.
Note that, as the QS is non-differentiable and cannot be directly optimized by back-propagation, we propose a novel reinforced self-training optimization method by recasting the model evaluation on the constructed QA pairs as a training reward.
The main contributions and novelty of this paper could be summarized as below.
\begin{itemize}
\item This work could be the first successful attempt to study the NC problem that is common but ignored in existing domain adaptation methods for MRC.
\item To solve the NC problem, we propose to leverage both the document and associated dialogue for MRC model training. To the best of our knowledge, this is the first method on study how to leverage the dialogue in the domain adaptation for MRC.
\item To implement robust domain adaptation method for MRC, RMRC consists of the AE, QS, and MRC model. Thanks to the reinforced self-training optimization method, the QS could be fine-tuned with the MRC model feedback, thus further alleviating the influence of NC.
\end{itemize}
\section{Related Work}
In this section, we will briefly introduce some recent developments in machine reading comprehension, the domain adaptation for MRC, and noisy label learning.
\subsection{Machine Reading Comprehension}
Machine reading comprehension aims to read the documents and answer questions about a document. Thanks to the collected benchmark datasets like SQuAD~\cite{rajpurkar2016squad}, CoQA~\cite{reddy2019coqa}, and QuAC~\cite{choi2018quac}, MRC has made great progress in recent years, and even surpasses the human level~\cite{seo2016bidirectional,yu2018qanet,devlin2018bert,zhang2020retrospective,jiang2018stackreader,gao2019product}.
\cite{seo2016bidirectional} proposed BiDAF which leverages RNN and bi-directional attention mechanism between question and document to achieve promising performance in machine reading comprehension. QANet~\cite{yu2018qanet} used CNN rather than RNN to better capture local information in question and documents.
BERT~\cite{devlin2018bert} proposes to use a large amount of unsupervised corpus for pre-training and successfully improve performance on many downstream NLP tasks including MRC.
\subsection{Domain Adaptation}
Domain adaptation (DA) is a well-developed technique that aims to transfer knowledge in the presence of the domain gap, i.e. making a model pre-trained in the source domain generalizable to the target domain~\cite{zhang2018collaborative,zhang2021meta,ding2018graph,hu2018duplex,kan2015bi,li2013learning}.
The existing domain adaptation methods for MRC could be roughly grouped into two categories:
1) \textbf{Model generalization.} It aims to improve the generalization capability of the model trained on the source domain to the target domain~\cite{li2019d,su2019generalizing,baradaran2021ensemble}.
For example, \cite{su2019generalizing} propose to use a pre-trained model trained on multiple MRC datasets simultaneously to improve the generalization ability on the new domains.
\cite{baradaran2021ensemble} propose to ensemble the models individually trained on different datasets to improve the generalization ability.
However, this kind of methods do not leverage the available information on the target domain, thus obtaining less satisfactory performance.
2) \textbf{QA generation.} It often utilizes the target-domain documents to construct QA pairs to fine-tune the MRC models~\cite{du2017learning,wang2019adversarial, lewis2019unsupervised,cao2020unsupervised}.
For example, \cite{du2017learning} propose to generate questions on the target domain using a Seq2seq question generator that trained on the source domain, and then fine-tune the MRC model with the pseudo QA pairs.
\cite{cao2020unsupervised} propose to add a self-training step to filter out low quality QA pairs using the estimated scores of each constructed pairs through the MRC model.
Though these methods have made great progress, they still suffer from the noisy correspondence problem as discussed in the introduction, thus leading to sub-optimal performance in real-world applications.
Unlike these works, to address the NC problem, this paper proposes to construct more credible QA pairs by additionally using the dialogue, and optimize the MRC model and the question selector in an alternative fashion to alleviate the domain shift of the QS to the target domain.
\subsection{Learning with Noisy Labels}
To alleviate even eliminate the influence of noisy labels, many methods have been proposed to achieve robust classification results~\citep{arazo2019unsupervised,song2020learning,liu2015classification,xia2020robust,bai2021understanding,luo2021bi}. Currently, the existing works often resort to sample selection for achieving noise robustness.
To be specific, sample selection methods seek to select the clean samples from the noisy dataset for training. For example, \cite{arazo2019unsupervised} proposes to select the samples with small loss as the clean samples.
Moreover, to further enhance the clean sample selection capacity, Co-teaching methods~\citep{han2018co,yu2019does} leverage two individual trained networks to filter out noises in an alternative manner.
In recent, PES~\citep{bai2021understanding} proposes a progressive early stopping strategy in the semi-supervised learning framework by treating the clean and noisy samples as the labeled and unlabeled data.
In addition, some very recent works~\cite{yang2021partially,huang2021learning} study the paradigm of noisy correspondence like this paper. Although the works share some similarity with this work, they are significantly different in motivation, application, and method.
\begin{figure*}[t]
\centering
\includegraphics[width=0.92\textwidth]{figures/Pipeline.pdf}
\vskip -0.1cm
\caption{Overview of our method. RMRC consists of an answer extractor (AE), a question selector (AS), and an MRC model. In the first stage, AE and QS construct the QA pairs in the target domain by leveraging both the documents and dialogues. In the second stage, RMRC optimizes the MRC model and QS in an alternative fashion. }
\label{RMRC_pipeline}
\vskip -0.4cm
\end{figure*}
\section{Method}
In this section, we elaborate on the proposed Robust Domain Adaptation for Machine Reading Comprehension (RMRC).
\subsection{Problem Definition}
\label{sec:problem_definition}
For a given document and a related question, MRC aims to find a text span from the document as the corresponding answer. To overcome the data scarcity issue in MRC, some domain adaptation methods are proposed to transfer an MRC model pre-trained in the source domain $\mathcal{S}$ to the target domain $\mathcal{T}$.
Different from existing domain adaptation methods for MRC, we utilize the documents along with the associated dialogues in $\mathcal{T}$ to construct QA pairs for model fine-tuning.
Formally, given a pre-trained MRC model $M$ and $B$ documents $\mathcal{D} = \{D_1, D_2, \cdots, D_B\}$ in the target domain, where each document $D_b$ is associated with a dialogue set generated by the questioners and answerers, we aim to transfer $M$ from $\mathcal{S}$ to $\mathcal{T}$. For each dialogue $ C =\{C^{r_1}_1, C^{r_2}_2, \cdots, C^{r_{N_c}}_{N_c} | r_i \in (a, q)\}$, it contains $N_c$ chats, where $r_i$ denotes the speaker of the $i$-th chat, $a$ and $q$ indicate the role of answerer and the questioner, respectively.
For clarity, we first provide an overview of the proposed RMRC and then introduce its components one by one. As shown in Fig.~\ref{RMRC_pipeline}, RMRC consists of an answer extractor (AE), a question selector (QS), and an MRC model which are applied to the following two stages: 1) \textbf{QA construction with AE and QS}. For each given document and the associated dialogues, the AE first splits the document into a set of candidate answers. Then, the AE outputs the pseudo answers by filtering out the irrelevant answers in terms of the relevance to the corresponding chat. After that, for each given pseudo answer and the related chat, QS selects multiple chats located before the answer-related chat in the dialogue as the candidate questions. Finally, the most related candidate questions are concatenated as pseudo questions. 2) \textbf{Alternative training of the MRC model and QS}. For a given pseudo question and document, RMRC uses the MRC model to obtain the corresponding answer. After that, the obtained answer is used to optimize the MRC model via a cross-entropy loss and the QS via our novel reinforce self-training optimizer.
\subsection{Answer Extractor}
\label{sec:answer_extractor}
The AE is designed to extract answers by finding the most similar text span from the document in terms of the answerer's response contained in the dialogue. Formally, with the document set $\mathcal{D}$ and associated dialogue $C$, AE extracts the corresponding answer $\{\mathbf{a}_i\}_{i=1}^{N} \in \mathcal{D}$ by filtering out the unrelated answers.
Specifically, we first split all documents in $\mathcal{D}$ into $\mathrm{n\text{-}gram}$ tokens $G$ via
\begin{equation}
\label{eq:ngram}
\begin{aligned}
G &= \{G_1 \cup \cdots \cup G_B\} \\
G_b &= \{G_b^k\}_{k=1}^K, G_b^k = \mathrm{n\text{-}gram}(D_b, k),
\end{aligned}
\end{equation}
where $\mathrm{n\text{-}gram}$ is the operator to extract the candidate answers from $\mathcal{D}$, $K$ denotes the maximum token number of the candidate answers, $G_b^k$ denotes the $k$-gram token set from document $G_b$. In other words, we will extract $\mathrm{1\text{-}gram}, \mathrm{2\text{-}gram}, \cdots, \mathrm{K\text{-}gram}$ tokens for answer extraction.
With the candidate answers, a text matching function is developed to find the best matched answer $\mathbf{g}_i \in G$ to $C^{a}_t$ which denoted as $\mathbf{a}_t$, i.e.,
\begin{equation}
\label{eq:matching}
\mathbf{a}_t, D_t^* = \mathop{\arg \max}_{\mathbf{g}_i \in G_b, G_b\in G} S(C^{a}_t, \textbf{g}_i),
\end{equation}
where $\mathbf{g}_i$ is the extracted n-gram token in $G$, $(\mathbf{a}_t, D_t^*)$ are the best matched answer and the corresponding document to $C^{a}_t$.
In the equation, we compute the cosine similarity $S(C^{a}_t, \mathbf{g}_i)$ between the extracted features of $C^{a}_t$ and $\mathbf{g}_i$ from a pre-trained BERT~\cite{devlin2018bert}.
However in the real-world dialogues, the conversation contains not only the document-related answers but also some unrelated answers like greetings as shown in Fig.~\ref{fig:motivation}. Hence, to filter out these unrelated answers, we select the answer $\mathbf{a}_t$ whose matching score is larger than a given threshold $\gamma$. Formally,
\begin{equation}
\label{eq:a_pick}
A = \{(\mathbf{a}_t, D_t^*, C_t^a) | \gamma \leq S(C^{a}_t, \mathbf{a}_t), \forall C^{a}_t \in C\}.
\end{equation}
\subsection{Question Selector}
\label{sec:QA_selector}
For a given answer $\mathbf{a}_t$, we pass it and its associated document $D_t^*$ along with the corresponding chat $C^a_t$ through the QS to find the corresponding question $\mathbf{q}_t$ from the dialogue.
In detail, we first obtain the candidate questions by selecting multiple chats located before $C^{a}_t$ in the dialogue, i.e.,
\begin{equation}
\label{eq:q_candidate}
\tilde{Q} = \{C^q_{t'}\vert C^q_{t'} \in \mathcal{N}^{\tau}(C^{a}_t)\},
\end{equation}
where $\mathcal{N}^{\tau}(C^{a}_t)$ denotes $\tau$ closest chats located before $C^{a}_t$ from questioner, $\tau$ is the maximal question selection range which is fixed to $16$ throughout our experiments. Such a question selection strategy is based on the observation that the corresponding questions only exist before the answers and are more irrelevant as they are further away from the answer.
For each $C^{q}_{t'} \in \tilde{Q}$, we then compute the relevance score $R(C^{a}_t, C^{q}_{t'})$ between the answer $C^{a}_t$ and the question $C^{q}_{t'}$ by
\begin{equation}
\begin{aligned}
R(C^{a}_t, C^{q}_{t'}) & =\mathrm{Sigmoid}(h^{qa}_{t'}), \quad
h^{qa}_{t'} & = \mathrm{Encoder}_Q([C^{a}_t; C^{q}_{t'}])
\label{eq:qa_relevance}
\end{aligned}
\end{equation}
where $\mathrm{Encoder}_Q$ transforms $[C^{a}_t; C^{q}_{t'}]$ into a hidden vector for relevance prediction, $[;]$ denotes the concatenation operator. Note that, the relevance score $R(C^{a}_t, C^{q}_{t'})$ could be regarded as the conditional probability of $C^{q}_{t'}$ to be the corresponding question of the given answer $C^{a}_t$, i.e., $P(C^{q}_{t'}|C^{a}_t) \Leftrightarrow R(C^{a}_t, C^{q}_{t'})$.
To further alleviate the aforementioned discontinuous conversation issue in the real world (e.g., the questioner may raise a new question before the last one is answered.), we concatenate the most related questions in terms of the relevance score w.r.t. $C^{a}_t$, i.e.,
\begin{equation}
\label{eq:q_pick}
\begin{aligned}
\mathbf{q}_t = & \mathrm{Concat}(\{ C^{q}_{t'} | C^{q}_{t'} \in \mathcal{N}^{\kappa}(C_t^a)\}),
\end{aligned}
\end{equation}
where $\mathcal{N}^{\kappa}(C_i^q)$ denotes $\kappa$-nearest neighbors of $C^a_t$ in $\tilde{Q}$ in terms of the relevance score. Accordingly, the probability of $\mathbf{q}_t$ to be the corresponding question of $C_t^a$ is formulated as,
\begin{equation}
P(\mathbf{q}_t|C_t^a)=\prod_{C_{t'}^q \in \mathcal{N}^{\kappa}(C_t^a)}P(C_{t'}^q|C_t^a).
\end{equation}
\subsection{MRC Model Training}
\label{sec:training}
With the constructed pseudo QA pairs and the corresponding document $\{(\mathbf{a}_i, \mathbf{q}_i, D_i)\}_{i=1}^{N}$, we fine-tune the pre-trained MRC model to improve its generalization to the target domain. In detail, we first embed the question and associated document into a hidden space with a BERT encoder, i.e.,
\begin{equation}
\mathbf{v}_i = \mathrm{Encoder_{MRC}}([D_i; \mathbf{q}_i]).
\label{mrc_encode}
\end{equation}
Then, $\mathbf{v}_i$ is utilized to predict the positions of the corresponding answer in the document. Specifically,
\begin{equation}
\begin{aligned}
\label{mrc_prob}
P^{s}_i & = \mathrm{Softmax}(\mathrm{FFN}(\mathbf{W}^{s}; \mathbf{v}_i)),\\
P^{e}_i & = \mathrm{Softmax}(\mathrm{FFN}(\mathbf{W}^{e}; \mathbf{v}_i)),
\end{aligned}
\end{equation}
where $P_i^{s}$ and $P_i^{e}$ denote the probability of each token being the start and end positions of the answer in $D_i$, respectively. $\mathrm{FFN}(\mathbf{W}^{s}; \mathbf{v})$ and $\mathrm{FFN}(\mathbf{W}^{e}; \mathbf{v})$ are two one-layer feed-forward networks with parameters $\mathbf{W}^{s}$ and $\mathbf{W}^{e}$, respectively.
Finally, we use the cross entropy between the predicted probability $(P_i^{s}, P_i^{e})$ and the ground truth as the training loss for our MRC model, i.e.,
\begin{equation}
\label{eq:loss_mrc}
\mathcal{L}_{MRC} = \frac{1}{N}\sum_i^{N} \mathrm{CE}(P_i^{s}, Y^{s}_i) + \mathrm{CE}(P_i^{e}, Y^{e}_i),
\end{equation}
where $Y^{s}_i$ and $Y^{e}_i$ are two one-hot vectors which denote the start and end positions of $\mathbf{a}_i$ in $D_i$.
\subsection{Reinforced Self-training for QS}
\label{sec:training_for_qa}
As the QS is pre-trained on the source domain, the noisy correspondence will be inevitably introduced when the QS is applied to the target domain for constructing QA pairs.
To address this domain shift issue, we propose to optimize the QS in the target domain by recasting the model evaluation on the constructed pseudo QA pairs as a training reward.
Specifically, for a given pair $(\mathbf{a}_i, \mathbf{q}_i, D_i)$, we first obtain the predicted answer by
\begin{equation}
\begin{aligned}
\tilde{\mathbf{a}}_i &= [t_{\tilde{s}}:t_{\tilde{e}}], t \in D, \\
\tilde{s} &= \mathop{\arg \max} P_i^{s}, \\
\tilde{e} &= \mathop{\arg \max} P_i^{e},
\label{mrc_answer_predict}
\end{aligned}
\end{equation}
where $P_i^{s}$ and $P_i^{e}$ are the model outputs defined in Eq.~\ref{mrc_prob}, $t_{\tilde{s}}$ and $t_{\tilde{e}}$ denote the start and end token of the predicted answer, respectively. By using the f1-score as the quality evaluation score of the constructed QA pairs, it is expected to generate QA pairs with high f1-score, i.e.,
\begin{equation}
\theta^*=\mathop{\arg \max}_{\theta} \mathbb{E} f1(\mathbf{a}, \tilde{\mathbf{a}}),
\label{eq:ev_reward}
\end{equation}
where $\theta$ denotes the parameters of QS. As shown in Eq.~\ref{eq:q_pick}, the QA construction adopts the selection and concatenate operators via $\mathop{\arg \max}$ sampling technique which is discrete.
To overcome this non-differentiable problem, we adopt policy gradient based reinforcement learning (REINFORCE)~\cite{williams1992simple} to approximate the gradients w.r.t. $\theta$. Specifically, let $\mathcal{J}(\theta)$ denote the objective function (Eq.~\ref{eq:ev_reward}), its gradient can be approximated as below:
\begin{equation}
\begin{aligned}
\nabla_{\theta}\mathcal{J}(\theta)
&\approx\frac{1}{N}\sum_{i=1}^N [\nabla_{\theta}\sum_{C_j^q\in\mathcal{N}^{\kappa}_t}\mathrm{log} P_{\theta}(C_j^q|C_t^a) f1(\mathbf{\mathbf{a}_i}, \mathbf{f}(\mathbf{q}_i))],
\end{aligned}
\end{equation}
where $\mathbf{f}(\mathbf{q})$ transforms the question $\mathbf{q}$ into the predicted answer $\tilde{\mathbf{a}}$ via the MRC model, and $\mathcal{N}^{\kappa}_t$ denotes $\mathcal{N}^{\kappa}(C_t^a)$ for simplicity, and $P_{\theta}(C_j^q|C_t^a)$ is conditional probability of $C_j^q$ to be the corresponding question of $C_t^a$. The detail of the gradient approximation is provided in the Supplementary Material. With the approximated gradients $\nabla_{\theta}\mathcal{J}(\theta)$, the loss function for QA could be rewritten by,
\begin{equation}
\label{eq:loss_q}
\begin{aligned}
\mathcal{L}_{Q} &= -\frac{1}{N}\sum_{i=1}^N\sum_{C_j^q\in \mathcal{N}^{\kappa}_t}\mathrm{log} R_{\theta}(C_t^a, C_j^q) [f1(\mathbf{\mathbf{a}_i}, \mathbf{f}(\mathbf{q}_i))-r_b],
\end{aligned}
\end{equation}
where $R_{\theta}(C_i^a, C_j^q) $ denotes the relevance defined in Eq.~\ref{eq:q_pick}. As only one sample per reward estimation is used, we use $r_b$ as the baseline score for reducing variance like~\cite{williams1992simple}.
\section{Experiments}
In this section, we evaluate the RMRC on three datasets by comparing it with three MRC domain adaptation methods. The code and used datasets will be released soon.
\begin{table*}[!hbt]
\vskip -0.2cm
\renewcommand\arraystretch{1.3}
\centering
\caption{Results on d-QuAC and d-CoQA. \textbf{Bold} values indicate the best performance.}
\vskip -0.2cm
\scalebox{0.9}{
\begin{tabular}{c|c|cc|cc|cc|cc|cc|cc} \hline
\multirow{2}{*}{\textbf{Data}} & \multirow{2}{*}{\textbf{Size}} & \multicolumn{2}{c|}{\textbf{BERT-S}} & \multicolumn{2}{c|}{\textbf{UQACT}} & \multicolumn{2}{c|}{\textbf{CASe}} & \multicolumn{2}{c|}{\textbf{AdaMRC}} & \multicolumn{2}{c|}{\textbf{RMRC-fix}} & \multicolumn{2}{c}{\textbf{RMRC}} \\
\cline{3-14}
& & EM & F1 & EM & F1 & EM & F1 & EM & F1 & EM & F1 & EM & F1 \\
\hline
\hline
\multirow{3}{*}{d-QuAC} & ALL & 3.36 & 16.36 & 6.81 & 19.05 & 6.79 & 19.33 & 5.44 & 18.87 & 7.45 & 23.13 & \textbf{8.27} & \textbf{24.46} \\
& 5000 & - & - & 4.26 & 17.88 & 4.85 & 17.92 & 4.57 & 18.01 & 5.85 & 21.25 & \textbf{6.37} & \textbf{22.27} \\
& 1000 & - & - & 3.62 & 16.84 & 3.91 & 17.55 & 4.12 & 17.34 & 4.42 & 19.99 & \textbf{5.67} & \textbf{21.28} \\
\hline
\multirow{3}{*}{\text{d-CoQA}} & ALL & 12.50 & 37.80 & 13.92 & 40.18 & 17.01 & 43.27 & 14.82 & 39.92 & 20.85 & 49.09 & \textbf{20.97} & \textbf{49.35} \\
& 5000 & - & - & 12.86 & 38.27 & 14.09 & 39.52 & 13.21 & 38.88 & \textbf{18.30} & \textbf{45.49} & 17.63 & 44.94 \\
& 1000 & - & - & 11.45 & 35.43 & 12.63 & 38.92 & 12.39 & 38.11 & 15.43 & 42.75 & \textbf{16.08} & \textbf{42.27} \\
\hline
\end{tabular}
}
\label{tab:pub_dataset_result}
\vskip -0.4cm
\end{table*}
\begin{table}[!hbt]
\centering
\caption{Results on Alipay dataset.}
\vskip -0.2cm
\scalebox{0.95}{
\begin{tabular}{l|cc}
\hline
\textbf{Model} & EM & F1 \\
\hline\hline
\textbf{BERT-S}~\cite{devlin2018bert} & 1.81 & 32.62 \\
\textbf{UQACT}~\cite{lewis2019unsupervised} & 5.32 & 42.23 \\
\textbf{CASe}~\cite{cao2020unsupervised} & 5.75 & 44.28 \\
\textbf{AdaMRC}~\cite{wang2019adversarial} & 7.20 & 43.11 \\
\hline
\textbf{RMRC-fix} & 9.25 & 52.08 \\
\textbf{RMRC} & \textbf{9.28} & \textbf{55.67} \\
\hline
\end{tabular}
}
\label{tab:alipay_dataset_result}
\vskip -0.1cm
\end{table}
\begin{table}[!hbt]
\centering
\caption{Performance of different variants of RMRC.}
\vskip -0.2cm
\begin{tabular}{l|cc}
\hline
\textbf{Model} & EM & F1 \\
\hline\hline
BERT-S & 1.81 & 32.62 \\
RMRC w/o $r_b$ & 8.44 & 54.13 \\
RMRC w/o Answer Filtering & 2.10 & 35.05 \\
RMRC w/o Question Fusing & 8.03 & 53.71 \\
RMRC w/o QS Training & 8.52 & 51.72 \\
RMRC w/ Confidence-based Selector & 8.76 & 52.34 \\
RMRC w/ CE Reward & 7.71 & 55.52 \\
\hline
\textbf{RMRC} & \textbf{9.28} & \textbf{55.67} \\
\hline
\end{tabular}
\label{tab:ablation_study}
\vskip -0.4cm
\end{table}
\subsection{Datasets}
We pretrain the MRC model on the SQUAD dataset~\cite{rajpurkar2016squad} and fine-tune it on three target datasets including two public datasets (QuAC~\cite{choi2018quac} and CoQA~\cite{reddy2019coqa}) and one real-world dataset from Alipay. Note that, as the real-world Alipay data is in Chinese, we use another Chinese corpus instead of SQUAD for pre-training MRC model, i.e., a collection of CMRC~\cite{cui2018span}, DRCD~\cite{shao2018drcd} and DUREADER~\cite{he2017dureader}.
\textbf{QuAC and CoQA with Synthetic Noises: }
QuAC consists of 12,567 documents, including 11,567 training documents and 1,000 testing documents. Each document is affiliated with a related dialogue. In each round of chat in the dialogue, one user asks a question about the document, and the other user answers it. In total, there are 69,109 QA pairs for training and 5,868 for testing.
Similar to QuAC, CoQA is composed of a training set of 7,199 documents along with 107,285 QA pairs, and a testing set of 500 documents along with 7,918 QA pairs.
As the QA pairs of QuAC and CoQA are well-matched in the dialogue, we simulate the real conversation with noisy correspondence by randomly shuffling the questions in each dialogue.
Specifically, for each dialogue, we randomly move each question ahead of the corresponding answer up to $1\sim5$ rounds. We denote the shuffled datasets of QuAC and CoQA as d-QuAC and d-CoQA, respectively in the following.
\textbf{Alipay Dataset with Real Noises: }
Alipay Dataset is collected in real-world scenarios, which contains the conversations about the marketing activities from the customer and customer service.
In total, the dataset consists of 1,526 dialogues and 3,813 human-annotated QA pairs for testing.
\subsection{Implementation}
In our experiments, we take the widely-used BERT as the base encoder for the QS and the MRC encoder. The BERT network contains 12 hidden layers, each of which consists of 12 attention heads. For all experiments, we generate n-grams for each document by setting $K=7$ for Eq.~\ref{eq:ngram} and set the threshold for answer filtering $\gamma$ to $0.7$ and select the question by fixing $\kappa$ in Eq.~\ref{eq:q_pick} to $5$. The optimal parameters are determined by the grid search in the Alipay dataset and used for all experiments. We set the baseline score of the reward $r_b$ to $0.7$ for Eq.~\ref{eq:loss_q} in all experiments.
For network training, we use the Adam optimizer whose learning rate is set to $2e-5$ and $1e-5$ for pre-training and fine-tuning, respectively. More training details are provided in the supplemental material.
For evaluations, we take the \textit{Exact Match} (\textbf{EM}) and \textit{F1-score} (\textbf{F1}) as the performance measurements.
Both the metrics are the higher the better.
\begin{figure*}[!hbt]
\vskip -0.1cm
\centering
\subfigure[Influence of $\gamma$] {
\label{fig:influence_of_gamma}
\includegraphics[width=0.31\textwidth]{figures/influence_of_gamma.pdf}
}
\subfigure[Influence of $\kappa$] {
\label{fig:influence_of_topk}
\includegraphics[width=0.31\textwidth]{figures/influence_of_topk.pdf}
}
\subfigure[Influence of $r_b$] {
\label{fig:influence_of_baseline}
\includegraphics[width=0.31\textwidth]{figures/influence_of_baseline.pdf}
}
\vskip -0.3cm
\caption{(a-c) Influence of answer filtering threshold ($\gamma$), the selected question number ($\kappa$), and the baseline reward ($r_b$). }\label{fig:paras}
\vskip -0.2cm
\end{figure*}
\begin{figure*}[!hbt]
\centering
\includegraphics[width=1.0\textwidth]{figures/case.pdf}
\caption{Comparisons between the questions constructed by our method and existing work~\cite{lewis2019unsupervised}. The \textcolor{red}{red number} and \textcolor{green}{green number} denote the negative and positive reward from MRC model, respectively.}
\label{fig:case_study}
\vskip -0.3cm
\end{figure*}
\subsection{Comparison Experiments}
To show the effectiveness of our method, we compare it with the baselines including a vanilla BERT model trained in the source domain (denoted by BERT-S) and three QA generation based MRC domain adaptation methods, i.e., UQACT~\cite{lewis2019unsupervised}, AdaMRC~\cite{wang2019adversarial} and CASe~\cite{cao2020unsupervised}. Note that CASe~\cite{cao2020unsupervised} uses the annotated questions in the target domain for fine-tuning, while the questions are unavailable in our experiment settings. As a remedy, we use a pre-trained question generator to generate questions in the target domain for CASe. All the baselines and our method follow the same training pipeline, i.e., first pretraining the MRC model on the SQUAD data, then fine-tuning and evaluating it on the target datasets (d-QuAC, d-CoQA and Alipay). For all baselines, we conduct the experiments with the recommended parameters and report the best one. For our method, we repeat the experiments three times with different seeds and report the average.
Moreover, we additionally use a variant of RMRC with a fixed QS (denoted as RMRC-fix) that keep the weights of QS fixed instead of optimizing with proposed reinforce self-training optimizer. Such a baseline could help us to understand the influence of the proposed reinforce self-training strategy.
\textbf{Results on d-QuAC and d-CoQA}: Table~\ref{tab:pub_dataset_result} shows the quantitative results on d-QuAC and d-CoQA.
From the results, one could observe that all domain adaptation methods outperform BERT-S.
Moreover, the proposed methods (i.e., RMRC and RMRC-fix) significantly outperform the other three domain adaptation methods. Especially, RMRC achieves performance gains of \textbf{1.48/5.13} and \textbf{3.96/6.08} in d-QuAC-ALL and d-CoQA-ALL compared with the best baseline, respectively. From the comparisons between RMRC and RMRC-fix, one could easily find the effectiveness of our QS training strategy.
Furthermore, to investigate the performance of the proposed method with small training datasets, we conduct experiments on QuAC and CoQA with $5,000$ and $1,000$ training samples. In other words, only 1.45\% of original training QuAC samples are used in the evaluation with d-QuAC-1000. As shown, with the data size decreasing, the performance of all methods decreases while RMRC still outperforms all baselines.
\textbf{Results on Alipay Dataset}: Table~\ref{tab:alipay_dataset_result} shows the quantitative results on the Alipay dataset. From the results, one could find that RMRC outperforms all baselines by a considerable performance margin, demonstrating its effectiveness in real-world dataset. Specifically, RMRC achieves performance gains of \textbf{2.08} and \textbf{12.56} over the best baseline (CASe) in terms of EM and F1, respectively.
\subsection{Ablation Study}
In this section, we carry out ablation study on the Alipay dataset. In the experiments, we report the performance by removing or replacing some modules, including removing answer filtering (RMRC w/o Answer Filtering), question fusing (RMRC w/o Question Fusing), QS training (RMRC w/o QS Training); replacing f1-score reward by cross-entropy reward (RMRC w/ CE Reward), and reinforced self-training mechanism by confidence-based self-training mechanism (RMRC w/ Confidence-based Selector).
As shown in Table~\ref{tab:ablation_study}, all the proposed modules are crucial to achieving encouraging results and the following conclusions could be obtained: i) All the RMRC variants outperform BERT-S, demonstrating the value of the dialogue-based QA pair construction; ii) RMRC and RMRC w/ CE Reward significantly outperform both RMRC w/o QS Training and RMRC w/ Confidence-based Selector, which shows the effectiveness of the proposed reinforced self-training algorithm for QS training; iii) RMRC that using F1 as the reward achieves better performance CE as F1 score directly measures the precision of the predicted answer.
\subsection{Parameter Sensitivity Analysis}
In this section, we carry out experiments on the Alipay dataset to investigate the influence of different parameters in RMRC including the answer filtering threshold $\gamma$ in Eq.~\ref{eq:a_pick}, the number of nearest questions $\kappa$ in Eq.~\ref{eq:q_pick}, and the baseline reward $r_b$ in Eq.~\ref{eq:loss_q}.
As shown in Fig.~\ref{fig:influence_of_gamma}, the performance keeps continuously increasing with increasing $\gamma$ and gets the best result when $\gamma=0.7$. As shown in Fig.~\ref{fig:influence_of_topk}, the performance of the MRC model keeps continuously increasing with increasing $\kappa$ from $1$ to $5$, showing that the correct question was more likely selected with more selected questions. As $\kappa$ increases from $5$ to $7$, the performance significantly decreases. The reason is that too many selected questions inevitably introduce more noisy pairs. As shown in Fig.~\ref{fig:influence_of_baseline}, the performance of the MRC model keeps improving when $r_b$ increases from $0.1$ to $0.7$, and then decreases when $r_b$ is larger than $0.7$. The reason is that an over-high $r_b$ will discourage almost all the predictions for QS training, thus leading to incorrect feedback for QS.
\subsection{Case Study}
To investigate the effectiveness of the QS and reinforced self-training optimizer, we conduct the case study in Fig.~\ref{fig:case_study}. As shown, the previous domain adaptation works will suffer from the NC problem, thus yielding irrelevant questions to the answer.
In contrast, RMRC explicitly solves the NC problem by training QS with the MRC model feedback. Specifically, in the first epoch, QS would select an irrelevant question with a negative reward from the MRC model. After training with several epochs, the optimized QS could find the desirable question to the answer. Such an example intuitively demonstrates the effectiveness of our QS and the reinforced self-training algorithm.
\section{Conclusion}
This paper could be the first successful attempt to solve the noisy correspondence problem in the domain adaptation methods for MRC. Different from the well-studied noisy labels, the noisy correspondence refers to the errors in alignment instead of category-level annotation. To overcome this challenge, we propose a robust domain adaptation method for MRC with a novel reinforced self-training optimizer. Extensive experiments verify the effectiveness of the proposed method in leveraging synthesized and real-world dialogue data for MRC.
|
\section{\def\@secnumfont{\mdseries}\@startsection{section}{1}%
\z@{.7\linespacing\@plus\linespacing}{.5\linespacing}%
{\normalfont\scshape\centering}}
\def\subsection{\def\@secnumfont{\bfseries}\@startsection{subsection}{2}%
{\parindent}{.5\linespacing\@plus.7\linespacing}{-.5em}%
{\normalfont\bfseries}}
\makeatother
\begin{document}
\title[]{On symmetric closed subsets of real affine root systems}
\author{Dipnit Biswas}\address{Department of Mathematics, Indian Institute of Science, Bangalore 560012}
\email{<EMAIL>}
\thanks{}
\author{Irfan Habib}\address{Department of Mathematics, Indian Institute of Science, Bangalore 560012}
\email{<EMAIL>}
\thanks{}
\author{R. Venkatesh}
\address{Department of Mathematics, Indian Institute of Science, Bangalore 560012}
\email{<EMAIL>}
\thanks{}
\subjclass[2010]{}
\begin{abstract}
Any symmetric closed subset of a finite crystallographic root system must be a closed subroot system. This is not, in general, true for real affine root systems. In this paper, we determine when this is true and also give a very explicit description of symmetric closed subsets of real affine root systems. At the end, using our results, we study the correspondence between symmetric closed subsets of real affine root systems and the regular subalgebras generated by them.
\end{abstract}
\maketitle
\section{Introduction}
\iffalse
The problem of classifying closed subsets of a given finite crystallographic root system has a very long history and it appears in many contexts for example they appear
in the classification of maximal closed connected subgroups of maximal rank of a connected compact Lie group \cite{BdS},
in the classification of reflection subgroups of finite and affine Weyl groups \cite{DL11} and
in the theory Chevalley groups \cite{Harebov}, etc. We are mainly interested in this problem because of their connection with the regular subalgebras of the corresponding semi-simple Lie algebras (see \cite{Dynkin, DG20}).\fi
One of the main motivations to study closed subsets of a given finite crystallographic root system is that they are closely related to the regular subalgebras of the corresponding semi-simple Lie algebra (see \cite{Dynkin, DG20}).
They also appear in various other contexts. For example, they appear in
\begin{itemize}
\item the classification of maximal closed connected subgroups of maximal rank of a connected compact Lie group \cite{BdS}
\item the theory of abelian/ad-nilpotent ideals of Borel subalgebras \cite{Suter, Papi}
\item the theory Chevalley groups \cite{Harebov}
\item the classification of reflection subgroups of finite and affine Weyl groups \cite{DL11}.
\end{itemize}
Various special classes of closed subsets of a finite root system were classified by many authors. For example,
\begin{itemize}
\item A. Borel and J. De Siebenthal classified maximal closed subroot systems of finite root systems in \cite{BdS}
\item E. B. Dynkin came up with an algorithm to classify all subroot systems (upto the Weyl group conjugacy) of a finite root system in \cite{Dynkin}
\item Z. I. Borevi{\v{c}} used topological methods in \cite{Bo97} to determine all closed subsets of root systems of type $A_n, 1\le n\le 8$, and this work was subsequently taken forward by others, see for e.g. \cite{BM02}
\item invertible closed subsets were classified in \cite{DCH94}
\item recently, A. Douglas and W.A. de Graaf have given an algorithm for classifying all closed subsets of a finite root system, up to conjugation by the associated Weyl group, in \cite{DG20}.
\end{itemize}
The problem of classifying closed subsets of (real) affine root systems is wide open. Again some very particular classes were studied in this setting.
The invertible closed subsets of affine root systems were classified in \cite{CKS98} and the parabolic subsets were classified in \cite{Futorny}.
Anna Felikson et al. started the classification regular subalgebras of affine Kac-Moody algebras (which are related to the closed subroot systems of real affine root systems) in \cite{FRT08} and it was completed in \cite{RV19}, see also \cite{DV21}.
The classification of the reflection subgroups of finite and affine Weyl groups has been achieved in \cite{DL11} by classifying
subroot systems finite and real affine root systems.
It is well-known that the classification of all subroot systems may be deduced from that of the closed subroot systems in the finite setting, see \cite{Carter72}. A combinatorial description of biclosed sets of real affine root systems has been given very recently in \cite{BS22}.
We are mainly interested in the closed subsets of real affine root systems in this paper, as they are closely related to the Cartan invariant subalgebras of corresponding affine Kac-Moody algebras and this will be discussed elsewhere.
Suppose $\Psi$ is a closed subset of a real affine root system $\Phi$, i.e., if $\alpha, \beta\in \Psi$ and $\alpha+\beta\in \Phi$ implies that $\alpha+\beta\in \Psi$, then it is easy see that $\Psi$ can be written as a union of its symmetric part $\Psi^r=\Psi\cap -\Psi$ and special part $\Psi^s=\Psi\backslash \Psi^r.$
Both symmetric and special parts of $\Psi$ are closed in $\Phi$. So the classification of closed subsets of $\Phi$ reduces to the classification of
symmetric and special closed subsets of $\Phi$. In general the problem of classifying special closed subsets is very hard and it will be discussed in \cite{BIV23}.
In this paper we only focus on symmetric closed subsets of real affine root systems. The main motivation for this work comes from the following two assertions about the finite root system:
\begin{itemize}
\item each symmetric closed subset of a finite root system must be a closed subroot system and
\item there is a one to one correspondence between closed subsets of a finite root system and the Cartan invariant subalgebras of corresponding semi-simple Lie algebra (see \cite[Proposition 4.1]{DG20} for the precise statement).
\end{itemize}
Both these statements are not true in general for real affine root systems and this naturally motivates us to ask the following questions:
\begin{enumerate}
\item When a given symmetric closed subset of a real affine root system is a closed subroot system?
\item Is it possible to classify all symmetric closed subsets of a real affine root system?
\item Consider the map $\Psi\mapsto \lie g(\Psi)$ where $\Psi$ is a symmetric closed subset of a real affine root system and $\lie g(\Psi)$ is the subalgebra generated by $\lie g_\alpha, \,\alpha\in \Psi.$ We will see that this map is not injective for any affine Kac-Moody algebras (see Section \ref{regularsect}). Is it possible to determine the preimage of given $\lie g(\Psi)$?
\end{enumerate}
We will address all these questions in this paper. The paper is organized as follows: we recall the definitions and set up all the notations in Section \ref{prem}.
The symmetric closed subsets of real affine root systems whose gradient is closed are studied in Section \ref{closedsect}, and
the symmetric closed subsets of real affine root systems whose gradient is semi-closed are studied in Section \ref{semiclosedsect}.
The correspondence between symmetric closed subsets of real affine root systems and regular subalgebras of corresponding affine Lie algebras is discussed in Section \ref{regularsect}.
We summarize all our results at the end, in Section \ref{summarysect}.
\section{Preliminaries}\label{prem}
Throughout this paper, we denote by $\mathbb{C}$ (resp. $\mathbb{R}$), the field of complex numbers (resp. real numbers) and by $\bz$ (resp. $\bz_+$), the set of integers (resp. non-negative integers).
\subsection{}\label{section21}Let $E$ be an Euclidean space over $\mathbb{R}$ endowed with a positive definite symmetric bilinear form $(\cdot,\cdot)$. A finite (crystallographic) root system
$\mathring{\Phi}$ is a finite subset of $E$ satisfying the following properties (see \cite[Chapter VI]{Bou46} or \cite[Section 9.2]{Hu80}):
$$0\notin \mathring{\Phi},\ \ \text{Span}_{\mathbb{R}}\mathring{\Phi}=E,\ \ s_{\alpha}(\mathring{\Phi})=\mathring{\Phi},\ \forall\, \alpha\in\mathring{\Phi},\ \ (\beta,\alpha^{\vee})\in\bz,\ \forall\, \alpha,\beta \in \mathring{\Phi},$$
where $\alpha^{\vee}:=2\alpha/(\alpha,\alpha)$ and $s_{\alpha}$ is the reflection acting on $E$ defined by $s_{\alpha}(x)=x-(x,\alpha^{\vee})\alpha$, $x\in E$. For the rest of this paper, we denote $\mathring{\Phi}$ by a finite root system in $E$. In addition, if $\mathring{\Phi}$ satisfies $\br\alpha\cap \mathring{\Phi}=\{\pm \alpha\}$ for $\alpha\in \mathring{\Phi}$, then we call $\mathring{\Phi}$ reduced. Moreover, we call any subset $\Psi\subseteq \mathring{\Phi}$ irreducible whenever $\Psi=\Psi'\cup \Psi''$ with $(\Psi',\Psi'')=0$ implies $\Psi'=\emptyset$ or $\Psi''=\emptyset$. Any root system can be written as a direct sum of irreducible root systems (see \cite[Chapter VI]{Bou46}) and the reduced irreducible root systems were classified in terms of their Dynkin diagrams (see \cite[Theorem 11.4]{Hu80}).
They are the classical types $A_n\, (n\geq 1), B_n\, (n\geq 2), C_n\, (n\geq 3), D_n\, (n\geq 4)$ and the exceptional types $E_{6,7,8}, F_4$ and $G_2$. For a direct construction of these root systems we refer to \cite[Section 12.1]{Hu80}. Moreover, there is only one non--reduced irreducible root system of rank $n$, namely
\begin{equation*}BC_n=B_n\cup C_n=\{\pm\epsilon_i: 1\leq i\leq n\}\cup\{\pm\epsilon_i\pm \epsilon_j: 1\leq i\neq j\leq n\}\cup\{\pm 2\epsilon_i: 1\leq i\leq n\},\end{equation*}
where $\epsilon_1,\dots,\epsilon_n$ denotes an orthonormal basis of $E$ with respect to $(\cdot,\cdot)$.
Recall that a subset $\mathring{\Psi}$ of $\mathring{\Phi}$ is said to be symmetric if $\mathring{\Psi}=-\mathring{\Psi}$, and it is called
closed in $\mathring{\Phi}$ if $\alpha, \beta\in \mathring{\Psi}$ and $\alpha+\beta\in \mathring{\Phi}$ implies $\alpha+\beta\in \mathring{\Psi}$.
We record the following simple and important fact on finite root systems, and we include a proof for the reader's convenience.
\begin{lem}\label{keylem:finite}
Let $\mathring{\Psi}$ be a symmetric closed subset of $\mathring{\Phi}$, then $\mathring{\Psi}$ is a closed subroot system of $\mathring{\Phi}$.
\end{lem}
\begin{proof}
Let $\alpha,\beta\in \mathring{\Psi}$. Suppose $(\beta,\alpha^\vee)=0$, then $s_\alpha(\beta)=\beta\in \mathring{\Psi}$. So assume that $(\beta,\alpha^\vee)\neq 0$.
If $(\beta,\alpha^\vee)< 0$, then $\beta,\beta+\alpha,\ldots,\beta+(-(\beta,\alpha^\vee))\alpha$ are elements of $\mathring{\Phi}$ by \cite[Proposition 8.4, Page 89]{Hu80}. Since $\mathring{\Psi}$ is closed, we have $\beta,\beta+\alpha,\ldots,\beta+(-(\beta,\alpha^\vee))\alpha\in \mathring{\Psi}$ which implies $s_\alpha(\beta)=\beta-(\beta,\alpha^\vee)\alpha\in \mathring{\Psi}$.
Since $\mathring{\Psi}$ is symmetric, the case $(\beta,\alpha^\vee)> 0$ can be done similarly.
\end{proof}
\subsection{}
The Weyl group $W$ of a finite root system $\mathring{\Phi}$ is defined to be the subgroup of $GL(E)$ generated by $s_{\alpha},\alpha\in\mathring{\Phi}$. At most two root lengths occur in any reduced irreducible finite root system $\mathring{\Phi}$ and all roots of a given length are conjugate under the Weyl group $W$ of $\mathring{\Phi}$ (see for example \cite[Section 10.4]{Hu80}). We denote the set of short roots (resp. long roots) by $\mathring{\Phi}_s$ (resp. $\mathring{\Phi}_\ell$) and if there is only one root length then we say that every root is short by convention. If $\mathring{\Phi}$ is non--reduced irreducible finite root system, we define:
$$\mathring{\Phi}_s=\{\pm\epsilon_i,\ 1\leq i\leq n\},\ \ \mathring{\Phi}_\ell=\{\pm\epsilon_i\pm\epsilon_j: 1\leq i\neq j\leq n\},\ \ \mathring{\Phi}_d=\{\pm2\epsilon_i,\ 1\leq i\leq n\}.$$
Note that $\mathring{\Phi}_d=\{\alpha\in \mathring{\Phi}: \alpha/2\in \mathring{\Phi}\}$ is the set of divisible roots in $BC_n$ and define the non--divisible roots of $BC_n$ by $\mathring{\Phi}_{nd}=\mathring{\Phi}_{}\backslash \mathring{\Phi}_{d}$. Further, set by $\mathring{\Pi}$ a base of $\mathring{\Phi}$ and set
$$ m_{\mathring{\Phi}}=\begin{cases}
1, & \text{$\mathring{\Phi}$ is of type $A_n$, $D_n$ or $E_n$}\\
2, & \text{$\mathring{\Phi}$ is of type $B_n$, $C_n$ or $F_4$}\\
3, & \text{$\mathring{\Phi}$ is of type $G_2$}\\
\end{cases}
$$
We often use $m$ instead of $m_{\mathring{\Phi}}$ if the underlying $\mathring{\Phi}$ is understood.
We end with the following fact (see \cite[Proposition 8.17]{Carter}).
\begin{lem}\label{longroot}
Let $\mathring{\Phi}$ be a finite irreducible root system. Let $\beta\in \mathring{\Phi}$ and write $\beta=\sum_{\alpha\in \mathring{\Pi}}k_\alpha \alpha$. Then $\beta$ is a long root if and only if $m$ divides $k_\alpha$ for each short root $\alpha\in \mathring{\Pi}$. \qed
\end{lem}
\medskip
\subsection{}\label{phi}
Let $\mathring{\Phi}$ be an irreducible reduced finite root system. Let $\mathring{\lie g}$ be a finite--dimensional semi-simple Lie algebra over $\mathbb{C}$ and $\mathring{\lie h}$ a Cartan subalgebra of $\mathring{\lie g}$ such that the root system corresponding to the pair $(\mathring{\lie g}, \mathring{\lie h})$ is
$\mathring{\Phi}$. Let $\sigma$ be a Dynkin diagram automorphism
of $\mathring{\lie h}$ with respect to $\mathring{\Phi}$ and denote $m$ by the order of $\sigma$. We know that $m\in\{1,2,3\}$. Let $\xi$ be a primitive $m$--th root of unity. We have
$$\mathring{\lie g}=\bigoplus_{j\in \bz/m\bz} \lie g_j,\ \ \lie g_{j}=\{x\in \lie g: \sigma(x)=\xi^j x\}.$$
It is known that $\lie g_0$ is again a finite-dimensional simple Lie algebra over $\mathbb{C}$ with a Cartan subalgebra $\lie h_0=\mathring{\lie h}\cap \lie g_0$. Moreover, $\lie g_j$ is a $\lie g_0$--module and we denote the set of non--zero weights of $\lie g_j$ with respect to $\lie h_0$ by $\mathring{\Phi}_j$.
Then we have $$\mathring{\Phi}:=\mathring{\Phi}_0 \cup\cdots\cup \mathring{\Phi}_{m-1}.$$
The types of $\mathring{\Phi},\mathring{\Phi}_0, \ldots, \mathring{\Phi}_{m-1}$ can be extracted from \cite[Section 7.8, 7.9, 8.3]{K90}.
The corresponding affine Kac--Moody algebra $\lie g=\widehat{\mathcal{L}}(\mathring{\lie g},\sigma)$ is defined by
$$\lie g=\mathcal{L}(\mathring{\lie g},\sigma)\oplus \bc c\oplus \bc d,\ \ \mathcal{L}(\mathring{\lie g},\sigma)=\bigoplus_{j\in \bz/m\bz} \lie g_{j}\otimes \bc[t^{\pm m}]t^{j},$$ where $\mathcal{L}(\mathring{\lie g},\sigma)$ is called the loop algebra, $\mathcal{L}(\mathring{\lie g},\sigma)\oplus \bc c$ is the universal central extension of the loop algebra and $d=t\frac{d}{dt}$ is the degree derivation. For more details we refer the reader to \cite[Section 7,8]{K90}. The set of roots of $\lie g$ with respect to the Cartan subalgebra $\lie h=\lie h_0\oplus \bc c\oplus \bc d$ is exactly $\Delta\backslash\{0\}$, where
$$\Delta:=\{\alpha+ r\delta: \alpha\in \mathring{\Phi}_j\cup\{0\},\ \ r\in j+m\bz ,\ 0\leq j< m\},$$
Denote $\Phi$ (resp. $\Phi^{\text{im}}$) by the set of real (resp. imaginary) roots of $\lie g$. Then we have $$\Phi=\bigcup_{\alpha\in \mathring{\Phi}}\big(\alpha+ \Lambda_\alpha\big),\ \Phi^{\text{im}}=\bz\delta.$$
where $\Lambda_\alpha=\Lambda_\beta$ if $\alpha$ and $\beta$ have same length and common $\Lambda_s, \Lambda_\ell, \Lambda_d$ can be found in the following table for each case.
\begin{equation*} \label{n:table1} {\renewcommand{\arraystretch}{1.5}
\begin{array}{|c |c| c| c| c | c| c| c|} \hline
(\lie g, m) & \mathring{\Phi}_{0} & \mathring{\Phi}_{1} & \mathring{\Phi}_{2} &\mathring{\Phi} & \Lambda_s &\Lambda_\ell & \Lambda_d \\ \hline
( \lie g(\mathring{\Phi}), 1) & \mathring{\Phi} & / & / & \mathring{\Phi} & \bz & \bz & / \\ \hline
(A_{2n}, 2) & B_n
& \mathring{\Phi}_0 \cup \{\pm 2\epsilon_i : 1 \le i \le n \} & / & BC_n &\frac{1}{2}+\bz& \bz & 2\bz \\ \hline
(A_{2n-1}, 2) & C_n & (C_{n})_ {s} & / &C_n & \bz & 2\bz & / \\ \hline
(D_{n+1}, 2) & B_n & (B_{n})_ {s} & /& B_n & \bz & 2\bz & / \\
\hline
(E_6,2) & F_4 & (F_{4})_ {s} &/ & F_4 & \bz & 2\bz & / \\ \hline
(D_4, 3) & G_2 & (G_{2})_ {s} & (G_{2})_ {s} & G_2 & \bz & 3\bz & / \\ \hline
\end{array}}
\end{equation*}
\medskip
We end this section with the following definitions.
\iffalse
\begin{defn}
Let $\Psi$ be a subset of $\Delta.$ An imaginary root $\alpha\in\Psi$ is called \textit{undesirable} if $\alpha$ can not be expressed as $\alpha=\alpha'+\alpha''$ where $\alpha',\alpha''$ are real roots of
\end{defn}
\begin{defn}
Let $\Psi$ be a subset of $\Delta.$ $\Psi$ is called closed in $\Delta$ if
\begin{enumerate}
\item if $\alpha,\beta$ are real roots in $\Psi$ and $\alpha+\beta\in \Delta,$ then $\alpha+\beta\in\Psi.$
\item Let $\alpha$ and $\beta\in \Psi$ be respectively imaginary and real roots in $\Psi.$ Write $\alpha=\alpha'+\alpha''$ with $\alpha,\alpha''\in\Phi.$ If $(\beta,\alpha')\neq 0$ and $\beta+\alpha\in\Delta,$ then $\beta+\alpha\in \Psi.$
\end{enumerate}
\end{defn}
The motivation of this definition comes from the next lemma.
\begin{lem}
Let $H$ be an $\lie{h}$-invariant subalgebra of $\lie{g}.$ Define $\Delta(H):\{\alpha\in\Delta:H\cap \lie{g}_{\alpha}\neq \{0\}\}.$ Then $\Delta(H)$ is a closed subset in $\Delta.$
\end{lem}
\begin{proof}
if $\alpha,\beta\in\Psi$ are real and $\alpha+\beta\in \Delta,$ we have $[\lie{g}_\alpha,\lie{g}_\beta]\neq 0$ if Hence $\alpha+\beta\in \Delta(H).$
\end{proof}
\fi
\begin{defn}
Let $\Psi$ be a subset of $\Delta$ (resp. $\Phi$). Set $\Psi_{\text{re}}:=\Psi\cap\Phi\ \text{and} \ \Psi_{\text{im}}:=\{r\in\mathbb{Z}:r\delta\in\Psi\}\cup \{0\}.$
\begin{enumerate}
\item We say $\Psi$ is \textit{symmetric} if $\Psi=-\Psi$, where $-\Psi=\{-\alpha : \alpha\in \Psi\}$.
\item We say $\Psi$ is \textit{real closed} or \textit{closed} in $\Phi$ if
\begin{enumerate}
\item $\Psi$ is a non-empty subset of $\Phi$
\item if $\alpha, \beta\in \Psi$ such that $\alpha+\beta\in \Phi$, then we have $\alpha+\beta\in \Psi$.
\end{enumerate}
\item We say $\Psi$ is a \textit{subroot system} if for any $\alpha\in \Psi_{\text{re}}$, $\beta\in \Psi$ we have $s_\alpha(\beta)\in \Psi.$
\end{enumerate}
\end{defn}
\iffalse
\begin{lem}\label{deltaclosed}
Let $\Psi$ be a symmetric subset of $\Delta.$ Then we have $\Psi$ is a closed in $\Delta$ if and only if $\Psi$ is a closed subroot system in $\Delta.$
\end{lem}
\begin{proof}
The if part is immediate. To prove the only if part, let $\Psi$ is a symmetric closed subset of $\Delta$
and let $\alpha,\beta\in\Psi$ such that $\alpha\in \Psi_{\text{re}}$. Since $\Psi$ is symmetric we have $s_\alpha(\beta)\in\Psi$ if $\beta=\pm\alpha$, so assume that $\beta\neq\pm\alpha.$
By \cite[Proposition 5.1]{K90}, there exists $p,q\in\bz_+$ such that $p-q=\langle\beta,\alpha^\vee\rangle$ and we have $\beta+k\alpha\in\Delta$ for all $-p\le k\le q.$ Since $\Psi$ is closed in $\Delta,$ we get $\beta+k\alpha\in \Psi$ for all such $k.$ Since $-p\le -p+q\le q,$ we have $s_\alpha(\beta)\in \Psi.$ This completes the proof.
\end{proof}
The following is immediate from \cref{deltaclosed}.
\begin{cor}\label{deltaclosedcor}
Let $\Psi$ be a symmetric closed subset of $\Delta$. Then $\Psi_{\text{re}}$ must be a real closed subroot system of $\Phi.$
\end{cor}
\begin{proof}
If $\Psi_{\text{re}}=\emptyset$, then there is nothing to prove. So assume that $\Psi_{\text{re}}\neq \emptyset$, then
it is clear that $\Psi_\text{re}$ is real closed and symmetric. So we need to prove that it is a subroot system. For this let $\alpha+r\delta, \beta+s\delta\in \Psi_\text{re}$, then we have
$$s_{\alpha+r\delta}(\beta+s\delta)=s_\alpha(\beta)+(s-(\beta, \alpha^\vee)r)\delta\in \Psi$$ as $\Psi$ is a closed subroot system of $\Delta.$ Now since
$s_\alpha(\beta)$ cannot be zero, we must have $s_{\alpha+r\delta}(\beta+s\delta)\in \Psi_{\text{re}}$. This proves $\Psi_{\text{re}}$ is a real closed subroot system of $\Phi.$
\end{proof}
\cref{deltaclosed} and \cref{deltaclosedcor} imply that any symmetric closed subset $\Psi$ of $\Delta$ must be symmetric closed subroot system in $\Delta$ and the real part of $\Psi$ is also a symmetric closed subroot system of $\Phi.$
\begin{rem}
A symmetric closed subset of $\Phi$ need not be in general a subroot system. This motivates us to determine when a given symmetric closed subset of $\Phi$ is a subroot system.
This will occupy the rest of the paper.
\end{rem}
\subsection{} Now we will see that any symmetric closed subset $\Psi$ of $\Delta$ is completely determined by its real part $\Psi_{\text{re}}$.
Since $\Psi_{\text{re}}$ is a real closed subroot system of $\Phi$, we can determine its structure very explicitly using the results of \cite[Section 2]{RV19}.
We summarize the results of \cite[Section 2]{RV19} here for the reader's convenience.
\begin{thm}
Let $\Delta$ be an irreducible affine root system which is not of type $A_{2n}^{(2)}$ and $\Psi$ be a symmetric closed subset of $\Delta.$
Let $\Psi_{\text{re}}=\Psi_1\sqcup\dots\sqcup \Psi_r$ be the decomposition of $\Psi_{\text{re}}$ into irreducible components. Then there exists a $\bz$-linear function $p:Gr(\Psi_{\text{re}})\to \bz$ and $n_i\in \bz, 1\le i\le r,$ such that, we have either
\begin{enumerate}
\item $\Psi_i=\{\alpha+(p_\alpha+n_i\bz)\delta:\alpha\in Gr(\Psi_i)\}$, or
\item $\Psi_i=\{\alpha+(p_\alpha+n_i\bz)\delta:\alpha\in Gr(\Psi_i)_s\}\cup\{\alpha+(p_\alpha+mn_i\bz)\delta:\alpha\in Gr(\Psi_i)_\ell\}$.
\end{enumerate}
for each $1\le i\le r.$ \qed
\end{thm}
Similarly, we can determine explicit structure of $\Psi_{\text{re}}$ when $\Delta$ is of type $A_{2n}^{(2)}$ using the results of \cite[Section 2]{RV19}.
\begin{prop}\label{realpartdeterm}
Let $\Psi$ be a subset $\Delta$.
Then the following statements are equivalent:
\begin{enumerate}
\item $\Psi$ is symmetric closed in $\Delta$
\item $\Psi$ is symmetric closed subroot system of $\Delta$
\item \begin{enumerate}
\item $\Psi_\text{re}$ is a real closed subroot system in $\Phi$,
\item For each $\alpha\in Gr(\Psi_{\text{re}}),$ $Z_\alpha(\Psi_{\text{re}})+Z_{-\alpha}(\Psi_{\text{re}})$ is a subgroup of $\bz$ and $Z_\alpha(\Psi_{\text{re}})+Z_{-\alpha}(\Psi_{\text{re}})=Z_{\beta}(\Psi_{\text{re}})+Z_{-\beta}(\Psi_{\text{re}})$ for all $\alpha,\beta\in Gr(\Psi_{\text{re}})$ with the same length.
\item If $n_s\bz$ (resp. $n_\ell\bz$) denotes the subgroup corresponding to short (resp. long) roots, then $n_s\mathrel{|}n_\ell,\ \ n_\ell\mathrel{|}\text{lcm}(m,n_s).\ ($ In particular, if $m\mathrel{|}n_s,$ then $n_\ell=n_s).$
\item We have $$\Psi_{\text{im}}=\begin{cases}
n_s\bz & \text{ if } Gr(\Psi_\text{re})\cap\mathring{\Phi}_s\neq \emptyset\\
n_\ell\bz & \text{ if } \Phi\text{ is untwisted and } Gr(\Psi_\text{re})\cap\mathring{\Phi}_s= \emptyset\\
k\bz & \text{ if } \Phi\text{ is twisted and } Gr(\Psi_\text{re})\cap\mathring{\Phi}_s=\emptyset
\end{cases}$$
where $k=n_\ell$ or $k=\frac{n_\ell}{m}$ depending on
\end{enumerate}
\end{enumerate}
\end{prop}
\begin{proof}
One direction of the first two equivalences follow from \cref{deltaclosed} and the other direction is trivial. Suppose that $\Psi$ is a symmetric closed subroot system in $\Delta.$ Then $\Psi_{\text{re}}$ is a real closed subroot system by \cref{deltaclosedcor}. Let $Z_\alpha:=Z_\alpha(\Psi_{\text{re}})$ for $\alpha\in\Psi_{\text{re}}.$
$(2)\Longrightarrow(3):$ We always have $Z_{\alpha}+Z_{-\alpha}$ is a subgroup of $\bz$ and $Z_{\alpha}+Z_{-\alpha}\subseteq \Psi_{\text{im}}$ since $\Psi$ is closed. If $\alpha$ is a short root and $r\in \Psi_{\text{im}}\backslash\{0\}, s\in Z_\alpha,$ we also have $-\alpha+(r-s)\delta=r\delta-(\alpha+s\delta)\in \Psi.$ So $r-s\in Z_{-\alpha}$ and hence $\Psi_{\text{im}}\backslash\{0\}\subseteq Z_\alpha+Z_{-\alpha}.$ Since $0\in Z_\alpha+Z_{-\alpha}$ it follows that $\Psi_{\text{im}}\subseteq Z_\alpha+Z_{-\alpha}.$ Since $\alpha$ is an arbitrary short root, $3(b)$ follows for short roots. Now let $\alpha,\beta$ be two long roots and $Z_\alpha=p_\alpha+n_\alpha\bz,Z_\beta=p_\beta+n_\beta\bz.$ Then $m\mathrel{|} n_\alpha$ and $m\mathrel{|}n_\beta.$ Since $n_\alpha\bz,n_\beta\bz\subseteq \Psi_{\text{im}}$ and $\Psi$ is closed, we have the following relations:
$$n_\alpha\bz+n_\beta\bz\subseteq n_\beta\bz,\ \ n_\beta\bz+n_\alpha\bz\subseteq n_\alpha\bz$$ Hence we have $\text{gcd}(n_\alpha,n_\beta)\bz\subseteq \text{lcm}(n_\alpha,n_\beta)\bz.$ Hence we have $n_\alpha=n_\beta.$ This proves $3(b).$
\noindent
$3(c)$ follows immediately using the following relations and using the fact that $\Psi_{\text{im}}=n_s\bz$:
$$n_\ell\bz\subseteq \Psi_{\text{im}},\ \ (\Psi_{\text{im}}+n_\ell\bz)\cap m\bz\subseteq n_\ell\bz.$$
\noindent
$3(d)$ is clear in the first two cases. Let $\Phi$ be twisted and $Gr(\Psi_\text{re})\cap\mathring{\Phi}_s=\emptyset.$ Of course we have $n_\ell\bz\subseteq \Psi_\text{im}.$ Since $\Psi$ is closed, $a\in \Psi_\text{im}$ implies that $a+n_\ell\bz\subseteq \Psi_\text{im}.$ It follows that $\Psi_\text{im}$ is a union of cosets modulo $n_\ell\bz.$ Moreover since $\Psi_\text{im}$ is closed under addition, the result follows.
\medskip
$(3)\Longrightarrow (1):$ Note that $3(b)$ is equivalent to \begin{equation}\label{stable}
n_\ell\bz\subseteq n_s\bz,\ \ (n_\ell\bz+n_s\bz)\cap m\bz\subseteq n_\ell\bz.
\end{equation} Now it is easy to check that $\Psi:=\Psi_{\text{re}}\cup (\Psi_{\text{im}}\backslash\{0\})\delta$ is a closed subset of $\Delta.$
\end{proof}
\subsection{} Let $\lie{g}'$ be an $\lie{h}$-invariant subalgebra of $\lie{g}$ i.e. $\lie{g}'=\oplus_{\alpha\in\Delta}\lie{g}_\alpha\cap \lie{g}',$
define $\Delta(\lie{g}'):=\{\alpha\in\Delta:\lie{g}_\alpha\cap \lie{g}'\neq \{0\}\}.$ For $\alpha\in \Delta(\lie{g}'),$ define $\lie{g}'_\alpha:=\lie{g}_\alpha\cap \lie{g}'$ We shall describe the $\Delta(\lie{g}')$ for all $\lie{h}$ invariant subalgebras such that $\Delta(\lie{g}')=-\Delta(\lie{g}').$
\begin{rem}
$\Delta(\lie{g}')$ need not be a closed subset of $\Delta.$ Let $S$ be any symmetric subset of $\bz$ not containing $0.$ Let $\lie{g}'_S:=\oplus_{k\in S}\lie{g}_{k\delta}.$ Since $[\lie{g}_{k\delta},\lie{g}_{\ell\delta}]=0$ for all $k,\ell\in \bz,$ we have that $\lie{g}'_S$ is a subalgebra of $\lie{g}.$ But $S\delta$ need not be closed in $\Delta.$
\end{rem}
\iffalse
\medskip
For an affine Kac Moody algebra $\lie{g},$ the imaginary root spaces are given by (see \cref{phi})
$$\lie{g}_{\text{im}}= \left(\bigoplus_{j\in\bz/m\bz,k\in\bz}(\mathring{\lie{h}})_j\otimes t^{mk+j}\right)\setminus \mathring{\lie{h}}.$$
\noindent
Define a $\mathbb{C}$- linear map $\Pi: \lie{g}_{\text{im}}\to \mathring{\lie{h}}$ by defining on basis elements by $h\otimes t^s\mapsto h.$ \fi
\begin{prop}
Let $\lie{g}'$ be an $\lie{h}$ invariant subalgebra of $\lie{g}$ with $\Delta(\lie{g}')=-\Delta(\lie{g}').$ Then there are irreducible real closed subroot systems $\Psi_1,\dots,\Psi_r\ (r\in \bz_+)$ of $\Phi$ and a symmetric subset $A\subseteq \bz\backslash\{0\}$ such that
$$\Psi=\Delta(\Psi_1)\sqcup \dots \sqcup \Delta(\Psi_r)\sqcup A\delta$$
and $\lie{g}'=\lie{g}(\Psi_1)\oplus \dots \lie{g}(\Psi_2)\oplus (\sum_{k\in A}\lie{g}'_{k\delta}).$
\iffalse
and the following are satisfied.
\begin{enumerate}
\item $\alpha+\beta\notin \Psi$ for all real roots $\alpha\in \Delta(\Psi_i),\beta\in \Delta(\Psi_j)$ with $i\neq j.$
\item $\Delta(\Psi_i)$ is closed in $\Delta$ for each $i.$
\item
\end{enumerate}with the property that if $\alpha,\beta$ are in different components in the decomposition, then $[\lie{g}'_\alpha,\lie{g}'_\beta]=0.$\fi
\end{prop}
\begin{proof}
For simplicity let $\Psi:=\Delta(\lie{g}')\cap\Phi.$ If $\Psi=\emptyset,$ we have $r=0$ and there is nothing to prove. If $\Psi\neq \emptyset,$ then $\Psi$ is a real closed subroot system of $\Phi.$ Let $\Psi=\Psi_1\sqcup \dots\sqcup\Psi_r$ be the decomposition of $\Psi$ into irreducible components. Clearly we have $\Delta(\Psi)=\Delta(\Psi_1)\sqcup \dots \sqcup \Delta(\Psi_r)\subseteq \Delta(\lie{g}')$ and $[\lie{g}'_\alpha,\lie{g}'_\beta]=0$ if $\alpha,\beta$ lie in different components by \cref{prop:subroot}. Since we have $\Psi=\Delta(\Psi)\cap \Phi,$ it follows that $\Delta(\lie{g}')\backslash \Delta(\Psi)\subseteq \bz\delta.$ If we let $A\delta=\Delta(\lie{g}')\backslash \Delta(\Psi),$ the result follows if we can show that $\alpha\in A\delta,\beta\in \Delta(\Psi_i)$ implies that $[\lie{g}'_\alpha,\lie{g}'_\beta]=0.$ Suppose that $[\lie{g}'_\alpha,\lie{g}'_\beta]\neq 0.$ Then $\alpha+\beta\in\Delta(\lie{g}').$ Ler $\alpha=r\delta,\beta=\beta_1+s\delta.$ We have $\beta_1+(r+s)\delta\in \Delta(\lie{g}').$ Since $\Delta(\lie{g}',$ $-\beta=-\beta_1-s\delta\in \Delta(\lie{g}'.$ Of course $[\lie{g}'_\beta,\lie{g}'_{-\beta}]\neq 0.$ Hence we have $r\delta=(\beta_1+(r+s)\delta)+(-\beta_1-s\delta)\in \Delta(\Psi_i)$ which is a contradiction and hence the result.
\end{proof}
\fi
\subsection{}
Let $\Phi$ be a real irreducible affine root system as in \Cref{phi} and let $\Psi$ be a closed subset of $\Phi$.
The symmetric part of $\Psi$ defined to be $\Psi^r:=\{\alpha\in\Psi\mid -\alpha\in \Psi\}$ and the special part of $\Psi$ defined to be $\Psi^s:=\{\alpha\in \Psi\mid -\alpha \notin \Psi\}$. It is clear that $\Psi^r$ is a symmetric closed subset of $\Phi$ and $\Psi^s$ is a closed subset of $\Phi$, and we have
$$\Psi=\Psi^r\sqcup \Psi^s.$$
The gradient of $\Psi$ is $Gr(\Psi)\,:=\{\alpha\in \mathring{\Phi} \mid \alpha+k\delta\,\in \Psi $ for some $k\in \bz\}$.
For given $\alpha \in Gr(\Psi)$, we define $Z_\alpha(\Psi)\,:=\{k\in\bz \mid \alpha+k\delta \in \Psi\}$.
We will simply use $Z_{\alpha}$ for $Z_\alpha(\Psi)$ if the dependence of the underlying $\Psi$ is understood.
Clearly we have $$\Psi =\bigcup_{\alpha\in Gr(\Psi)}\{\alpha+k\delta \mid k\in Z_{\alpha}\}.$$
For a given $(p_\alpha)_{\alpha\in Gr(\Psi)}$, where $p_\alpha\in Z_\alpha$, define
$Z_{\alpha}':=Z_\alpha-p_\alpha$ for $\alpha\in Gr(\Psi)$. We often make specific choices of $(p_\alpha)_{\alpha\in Gr(\Psi)}$ and make sure that the map $\alpha\mapsto p_\alpha$ gives us a $\mathbb{Z}$-linear function from $p: Gr(\Psi)\to \mathbb{Z}.$ For example, we have (see \cite[Lemma 13]{DL11a} and \cite[Lemma 2.1.1]{RV19}):
\begin{lem}\label{existp}
Let $\Psi$ be a symmetric closed subset of $\Phi$. Suppose
$Gr(\Psi)$ is a closed, reduced subroot system with a base $B$ and assume that we have
\begin{equation}\label{eq:z}
Z_\alpha+Z_\beta\subseteq Z_{\alpha+\beta}\ \ \text{for all}\
\ (\alpha,\beta,\alpha+\beta)\in Gr(\Psi)^{\times 3}.
\end{equation}
Choose $p_\alpha\in Z_\alpha$ arbitrarily for $\alpha\in B$ and extend the map
$\alpha\mapsto p_\alpha, \alpha\in B$ to $Gr(\Psi)$ $\mathbb{Z}$-linearly. Then we have $p_\alpha\in Z_\alpha$, for all $\alpha\in Gr(\Psi)$.
\end{lem}
\begin{proof}
Define a $\bz$-linear function $p: Gr(\Psi)\to \bz$ extending $\alpha\mapsto p_\alpha, \alpha\in B$.
We claim that $p_\alpha\in Z_\alpha$ for all $\alpha\in Gr(\Psi).$ Since $p_{-\alpha}=-p_{\alpha}$ and $Z_{-\alpha}=-Z_\alpha$, it is enough to prove that $p_\alpha\in Z_\alpha$ for all positive roots $\alpha\in Gr(\Psi).$ Let $\alpha\in Gr(\Psi)$ be a positive root, then we can write $\alpha=\alpha_1+\cdots +\alpha_r$ with
each partial sum $\alpha_1+\cdots+\alpha_i$ is again a root in $Gr(\Psi)$ for $1\le i\le r.$ Since $\Psi$ is closed and using the condition \eqref{eq:z}, we have
$\alpha_1+\cdots+\alpha_i+(p_{\alpha_1}+\cdots +p_{\alpha_i})\delta\in \Psi$ for $1\le i\le r.$ In particular, we have $p_\alpha=p_{\alpha_1}+\cdots +p_{\alpha_r}\in Z_\alpha.$
This completes the proof.
\end{proof}
\iffalse
Note that if $\Psi$ is irreducible (resp. symmetric) then $Gr(\Psi)$ is also irreducible (resp. symmetric). Let $\Psi$ be a symmetric subset of $\Phi$ and $\Psi=\Psi_1\sqcup\dots\sqcup\Psi_r$ be the decomposition of $\Psi$ into irreducible subsets. Then we have $Gr(\Psi)=Gr(\Psi_1)\sqcup\dots \sqcup Gr(\Psi_r)$ is the decomposition of $Gr(\Psi)$ into irreducible symmetric subsets.\fi
\iffalse
\begin{rem}
We do not have the following property as in the finite case: suppose $\mathring{\Psi}$ is a special closed subset of $\mathring{\Phi}$, then there is a $\mathring{W}$--conjugate of $\mathring{\Psi}$ which is contained in $\mathring{\Phi}^+$. For example let us consider $\Psi=\Phi^-$, then $\Psi^r=\emptyset$ and $\Psi^s=\Psi=\Phi^-$. Suppose that there is a $w\in W$ such that $w(\Psi^s)\subseteq \Phi^+$, then $w^{-1}(w\Psi^s)\subseteq(\Phi)^-$ which is not possible since $w\Psi^s$ is an infinite subset of $\Phi^+$.
\end{rem}
\fi
\subsection{} We collect here some basic facts in this subsection. Let $\Psi$ be a subset of $\Phi$ and let
\begin{equation}\label{decomppsi}
\Psi=\Psi_1\sqcup\dots\sqcup\Psi_r
\end{equation}
be the decomposition of $\Psi$ into irreducible subsets.
Then we have
\begin{equation}\label{decompgrpsi}
Gr(\Psi)=Gr(\Psi_1)\sqcup\dots \sqcup Gr(\Psi_r)
\end{equation}
is the decomposition of $Gr(\Psi)$ into irreducible subsets.
\begin{prop}\label{prop:subroot}
Let $\Psi$ be a subset of $\Phi.$
\begin{enumerate}
\item Suppose $\Psi$ is symmetric, then each $\Psi_i$ in \eqref{decomppsi} and each $Gr(\Psi_i)$ in \eqref{decompgrpsi} are symmetric.
\item Suppose $\Psi$ is closed in $\Phi$, then each $\Psi_i$ in \eqref{decomppsi} is closed in $\Phi$. In addition if $Gr(\Psi) $ is closed in $\mathring{\Phi}$, then each $Gr(\Psi_i)$ in \eqref{decompgrpsi} is closed in $\mathring{\Phi}$.
\item Suppose $\Psi$ is symmetric closed in $\Phi$. Then $\alpha+\beta\notin \Phi$ if $\alpha\in \Psi_i \text{ and } \beta\in \Psi_j$ for $i\neq j$.
\item Assume that $Gr(\Psi)$ is closed in $\mathring{\Phi}$. Then
\begin{enumerate}
\item $\Psi$ is symmetric closed in $\Phi$ if and only if each $\Psi_i$ is symmetric closed in $\Phi$.
\item $\Psi$ is a closed subroot system of $\Phi$ if and only if each $\Psi_i$ is a closed subroot system of $\Phi$.
\end{enumerate}
\end{enumerate}
\end{prop}
\begin{proof}
We only prove the statement $4(a)$ as all other statements are easy to check. The forward direction follows from $(3)$. For the converse part, we assume that each $\Psi_i$ is symmetric closed in $\Phi.$
Since $Gr(\Psi)$ is symmetric closed in $\mathring{\Phi}$, it must be a closed subroot system. Hence each
$Gr(\Psi_i)$ is a closed subroot system of $\mathring{\Phi}$ and it has a base say $B_i$. Then $B:=\cup_i B_i$ is a base for $Gr(\Psi)$. Let $\alpha+r\delta\in \Psi_i$ and $\beta+s\delta\in \Psi_j$ with $i\neq j$. If $\alpha+\beta+(r+s)\delta\in \Phi$, then $\alpha+\beta+(r+s)\delta\in \Psi$ since $\Psi$ is closed. But this implies that $\alpha+\beta\in Gr(\Psi)$, which is impossible because any root of a finite root system has connected support, see \cite[Proposition 16.21]{Carter}. This completes the proof.
\end{proof}
\begin{rem}\label{remark1}
The fourth statement in \Cref{prop:subroot} is false in general if we drop the condition $Gr(\Psi)$ is closed. For example, let us consider $\Phi$ be of type $D_{n+1}^{(2)}, n
\ge 2$. Set $I_n=\{1,2,\dots,n\}$. For $I\subseteq I_n$, let $$\Psi_{I}:=\{\pm \epsilon_k+(2\bz+1)\delta:k\in I\}\cup \{\pm (\epsilon_k\pm \epsilon_\ell)+2\bz\delta:k\neq \ell\in I\}.$$
It is clear that $\Psi_I$ is a symmetric closed subset of $\Phi$ for each $I.$ Now take $\Psi=\Psi_I\cup \Psi_J$ where $I, J$ form a partition of $I_n$, then $\Psi$ is not a closed subset of $\Phi.$
\iffalse \begin{enumerate}
\item If $\Psi$ is closed in $\Phi$ then we may not always have $Gr(\Psi)$ is closed in $\mathring{\Phi}$.
\item Let $\Phi$ be of type $D_{2n+1}^{(2)}$. Let $\Psi$ be defined by $$\Psi=\{\pm\epsilon_i+(2\bz+i)\delta:1\le i\le 2n\}\cup \{\pm (\epsilon_i\pm\epsilon_j)+2\bz\delta: 1\le i\neq j\le 2n,\ i\equiv j (\mathrm{mod}\ 2)\}$$
Then $\Psi$ is a symmetric closed subset of $\Phi$ with $\Psi=\Psi_1\sqcup \Psi_2$ where $$\Psi_k=\bigcup _{i=1}^{2n}\{\pm\epsilon_i+(2\bz+i)\delta:i\equiv k (\mathrm{mod}\ 2)\}\cup \{\pm (\epsilon_i\pm\epsilon_j)+2\bz\delta: 1\le i\neq j\le 2n,\ i\equiv j\equiv k (\mathrm{mod}\ 2)\}$$
We have $\epsilon_1+\epsilon_2\in \mathring{\Phi}$ but not in $Gr(\Psi)$.
\end{enumerate}
\fi
\end{rem}
\subsection{}\label{regularsubalgebras}
Given $S\subseteq \Phi$, we define $\mathfrak{g}(S)$ by the subalgebra of $\mathfrak{g}$ generated by $\mathfrak{g}_\alpha, \alpha\in S.$ Let $\Delta(S)$ be the set of roots of
$\mathfrak{g}(S)$.
We end this section with the following proposition.
\begin{prop}\label{smallestclosed}
Let $\Psi$ be a symmetric subset of $\Phi$. Then $\Delta(\Psi)\cap \Phi$ is a minimal closed subroot system of $\Phi$ containing $\Psi.$
\end{prop}
\begin{proof}
By \cite[Lemma 11.1.2, Page 1301]{RV19}, it is enough to prove that $\Delta(\Psi)=-\Delta(\Psi)$. Suppose $\beta\in \Delta(\Psi)$, then there exists $\beta_1, \ldots, \beta_r\in \Psi$ such that $\beta=\beta_1+\cdots +\beta_r$.
Since the Chevalley involution \cite[Chapter 1, Page 7]{K90} of $\mathfrak{g}$ takes $\mathfrak{g}_\alpha$ to $\mathfrak{g}_{-\alpha}$ for all $\alpha\in \Delta$, we have
$$[\mathfrak{g}_{\alpha_1}, \cdots [\mathfrak{g}_{\alpha_{r-1}}, \mathfrak{g}_{\alpha_r}]] \neq 0 \iff [\mathfrak{g}_{-\alpha_1}, \cdots [\mathfrak{g}_{-\alpha_{r-1}}, \mathfrak{g}_{-\alpha_r}]] \neq 0.$$
Since $\Psi$ is symmetric, we have $-\beta_1, \ldots, -\beta_r\in \Psi$
and $0\neq [\mathfrak{g}_{-\alpha_1}, \cdots [\mathfrak{g}_{-\alpha_{r-1}}, \mathfrak{g}_{-\alpha_r}]]\subseteq \mathfrak{g}(\Psi)$.
This implies $-\beta\in \Delta$ . Hence we have $\Delta(\Psi)=-\Delta(\Psi).$
Suppose $\Psi\subseteq S$ is a closed subroot system of $\Phi.$ Then we have
$\mathfrak{g}(\Psi)\subseteq \mathfrak{g}(S)$ and the real root of $\mathfrak{g}(S)$ is equal to $S$ by \cite[Corollary 11.1.5, Page 1304]{RV19}. Since the real roots of $\mathfrak{g}(\Psi)$ are also real roots of $\mathfrak{g}(S)$, we must have $\Delta(\Psi)\cap \Phi\subseteq S.$
\end{proof}
\section{Symmetric closed subsets of affine root systems}\label{closedsect}
In this section, we fix a real affine root system $\Phi$ which is not of type $A_{2n}^{(2)}$. We need the following notation: for $k\in \bz_+,$ denote by $\pi_k:\mathbb{Z}\to \mathbb{Z}/k\mathbb{Z}$ the qutioent map $x\to x\ (\mathrm{mod}\ k)$. Recall that $m$ is the lacing number associated with $\mathring{\Phi}.$
\subsection{} We define semi-closed subset of finite root systems as in \cite[Definition 4.4.1]{RV19}.
\begin{defn}
A symmetric subset $\mathring{\Psi}$ of $\mathring{\Phi}$ is called semi-closed subset of $\mathring{\Phi}$ if
\begin{enumerate}
\item $\mathring{\Psi}$ is not closed in $\mathring{\Phi}$.
\item If $\alpha,\beta\in \mathring{\Psi}$ such that $\alpha+\beta\in \mathring{\Phi}\setminus \mathring{\Psi}$, then
$(\alpha, \beta, \alpha+\beta)$ is of type $(s,s,\ell)$.
\end{enumerate}
\end{defn}
We need the following simple lemma, see \cite[Proposition 4.1.2]{RV19}.
\begin{lem}\label{keylem:affine}
Let $\Phi$ be a real twisted affine root system not of type $A_{2n}^{(2)}$. Let $\Psi\leq \Phi$ be a closed subset. Then $Gr(\Psi)$ is either closed or semi-closed subset of $\mathring{\Phi}$. \qed
\end{lem}
\subsection{}
We now consider the case when $Gr(\Psi)$ is a closed subroot system of $\mathring{\Phi}$.
The following proposition will be used as a primary tool to prove our main theorem in most of the cases.
\begin{prop}\label{prop:1}
Let $\Phi$ be a real affine root system not of type $A_{2n}^{(2)}$.
Let $\Psi$ be a symmetric closed subset of $\Phi$ such that $Gr(\Psi)$ is closed in $\mathring{\Phi}$ and no irreducible component of $Gr(\Psi)$ is of type $A_1$. Suppose that there is a $\bz$-linear function $p:Gr(\Psi)\to\bz$, $\alpha\mapsto p_\alpha$, such that $p_\alpha\in Z_\alpha,$\ and $|\pi_m(Z_\alpha)|=1, \text{for all} \ \alpha\in Gr(\Psi)$
Then we have
\begin{enumerate}
\item $Z_\alpha'=Z_\beta'=Z_{\alpha+\beta}':=A$ for
$(\alpha,\beta, \alpha+\beta)\in Gr(\Psi)^{\times 3}$ and $A$ is a subgroup of $m\bz$. \item For any $\alpha,\beta\in\Psi$, we have $s_\alpha(\beta)\in\Psi$. In particular, $\Psi$ is a closed subroot system of $\Phi$.
\end{enumerate}
\end{prop}
\begin{proof}
Recall that $Z_\alpha'=Z_\alpha-p_\alpha$ for $\alpha\in Gr(\Psi).$
Let $(\alpha, \beta, \alpha+\beta)\in Gr(\Psi)^{\times 3}$. We claim that $$Z_\alpha+Z_\beta\subseteq Z_{\alpha+\beta}.$$Let $r\in Z_\alpha$ and $s\in Z_\beta$. If $\alpha+\beta$ is short or $\Phi$ is untwisted then the result is immediate. So assume that $\alpha+\beta$ is long. Then we have
$p_{\alpha+\beta}\equiv 0 \,(\mathrm{mod}\,m)$.
Since $r\equiv p_\alpha\,(\mathrm{mod}\, m)\text{ and } s\equiv p_\beta\,(\mathrm{mod}\, m)$ and the map $p$ is $\bz$-linear we have $r+s\equiv 0 \,(\mathrm{mod}\,m)$.
This implies $\alpha+\beta+(r+s)\delta\in \Phi$. Since $\Psi$ is closed, it follows that $\alpha+\beta+(r+s)\delta\in \Psi$ and so $r+s\in Z_{\alpha+\beta}$.
Hence $Z_\alpha+Z_\beta\subseteq Z_{\alpha+\beta}.$ Now using the linearity of $p$, we get
\begin{equation}\label{keyfact1}
Z_\alpha'+Z_\beta'\subseteq Z_{\alpha+\beta}'
\end{equation}
for all $(\alpha, \beta, \alpha+\beta)\in Gr(\Psi)^{\times 3}$.
Since \eqref{keyfact1} is true for all
tuples $(\alpha,\beta, \alpha+\beta)$ such that $\alpha, \beta, \alpha+\beta\in Gr(\Psi)$.
We have
$$ Z_\alpha'+Z_\beta' \subseteq Z_{\alpha+\beta}',\ \ Z_{\alpha+\beta}'+Z_{-\alpha}'\subseteq Z_\beta',\ \text{and} \
Z_{\alpha+\beta}'+Z_{-\beta}' \subseteq Z_\alpha'.
$$ This implies $A=Z_\alpha'=Z_\beta'=Z_{\alpha+\beta}'$.
It is easy to see that $Z_{-\mu}'=-Z_{\mu}'$ for all $\mu\in$ Gr$(\Psi)$. Thus $A\subseteq m\bz$ satisfies $A=-A$ and $A+A\subseteq A$, so it must be a subgroup of $m\bz$.
\medskip
To prove $(2)$, let $\alpha, \beta \in \Psi.$ Write $\alpha=\alpha'+s\delta,\, \beta=\beta'+r\delta$. We claim that
$s_{\alpha}(\beta)=\beta-(\beta,\alpha^\vee)\alpha\in \Psi$. Suppose $\beta'\neq \pm \alpha'$, then using \say{unbroken string property} of $\Delta$ we get
$s_{\alpha}(\beta)\in \Psi$ as $\Psi$ is closed and any root of the form $\beta+p\alpha, p\in \bz,$ must be real.
Now assume that $\beta'= \pm \alpha'$. Since no irreducible component of $Gr(\Psi)$ is of type $A_1$, there is a $\gamma'\in Gr(\Psi)$ such that $\alpha'+\gamma'\in$ Gr$(\Psi)$.
Applying the Part (1) for the triple $(\alpha',\gamma',\alpha'+\gamma')$, we get that $Z_{\alpha'}'$ is a subgroup of $m\bz$.
We have
$$s_{\alpha'+s\delta}(\pm\alpha'+r\delta)=\mp\alpha'+(r\mp 2s)\delta$$
Since $s\in Z_{\alpha'},\,r\in Z_{\pm\alpha'}$, there are $z_1,z_2\in Z_{\alpha'}'$ such that $s=z_1+p_{\alpha'}$ and $r=z_2+ p_{\pm \alpha'}$.
Now we have $r\mp 2s=z_2\pm p_{\alpha'}\mp 2(z_1+p_{\alpha'})=(z_2\mp 2z_1)\mp p_{\alpha'}\in Z_{\alpha'}'+p_{\mp \alpha'}$ as $Z_{\alpha'}'$ is a group.
Hence $s_{\alpha}(\beta)\in \Psi$ and this completes the proof.
\end{proof}
\iffalse
\subsection{} We need the following lemma.
\begin{lem}\label{lem:1}
Let $\Psi$ be a symmetric closed subset of the untwisted affine root system $\Phi$. Let $\alpha,\beta\in Gr(\Psi)$ such that $\alpha+\beta\in \mathring{\Phi}$. Then $\alpha+\beta\in Gr(\Psi)$ and each of $Z_\gamma, \gamma\in\{\pm\alpha, \pm\beta, \pm(\alpha+\beta)\}$
correspond to cosets of a fixed subgroup of $\bz$.
\end{lem}
\begin{proof}
Fix $\alpha,\beta\in Gr(\Psi)$ such that $\alpha+\beta\in \mathring{\Phi}$.
Let $p_\alpha \in Z_\alpha$ and $p_\beta \in Z_\beta$. Set $Z_\alpha'\,:=Z_\alpha-p_\alpha$ and $Z_\beta'\,:=Z_\beta-p_\beta$. Since $\Psi$ is closed and $\Phi$ is untwisted, we have
$$\alpha+\beta\in Gr(\Psi)\,\, \text{and}\,\, Z_\alpha+Z_\beta \subseteq Z_{\alpha+\beta}.$$
In particular, we have $p_\alpha+p_\beta\in Z_{\alpha+\beta}.$ Now set $Z_{\alpha+\beta}'=Z_{\alpha+\beta}-(p_\alpha+p_\beta).$ We immediately have
\begin{equation}\label{eq:1}
Z_\alpha'+Z_\beta' \subseteq Z_{\alpha+\beta}'
\end{equation}
Since $0\in Z_\alpha'$, we have $Z_\beta' \subseteq Z_{\alpha+\beta}'$ from ~\Cref{eq:1}.
Since $\Psi$ is symmetric, we have $Z_{-\gamma}=-Z_\gamma$ for all $\gamma\in Gr(\Psi).$ Set $Z_{-\gamma}'=Z_{-\gamma}+p_{\gamma}$ for all $\gamma\in \{\alpha, \beta, \alpha+\beta\}$.
Then it is immediate that $Z_{-\gamma}'=-Z_{\gamma}'$ for all $\gamma\in \{\alpha, \beta, \alpha+\beta\}$.
Using ~\Cref{eq:1} for the tuple $(\alpha+\beta, -\alpha, \beta)$, we have
$$Z_{\alpha+\beta}'+Z_{-\alpha}'\subseteq Z_{\beta}'.$$
This gives $Z_{\alpha+\beta}'\subseteq Z_\beta'$ and hence we have $Z_\beta'=Z_{\alpha+\beta}'$. Similarly considering the tuple $(-\beta, \alpha+\beta, \alpha)$, we get $Z_{\alpha+\beta}'=Z_{\alpha}'$. Thus we have $$Z_{\alpha}'=Z_\beta'=Z_{\alpha+\beta}'.$$ Suppose $A=Z_\alpha'$, then $A$ satisfies $0\in A$, $A+A\subseteq A$ and $A=-A$.
This implies $A$ is a subgroup of $\bz$. This completes the proof.
\end{proof}
\begin{cor}\label{cor:1}
Let $\Psi$ be an irreducible symmetric closed subset of the untwisted affine root system $\Phi$.
Then there exists $n\in \bz$ and a $\bz-$linear function $p: Gr(\Psi)\to \bz$ such that $Z_\alpha=p_\alpha+n\bz$ for all
$\alpha\in Gr(\Psi)$.
\end{cor}
\begin{proof}
Since $\Psi$ is irreducible and symmetric, we have $Gr(\Psi)$ is irreducible and symmetric. Since
$\Phi$ is untwisted, we have $Gr(\Psi)$ is closed. Then by Lemma \Cref{keylem:finite}, we have $Gr(\Psi)$ is a closed subroot system of $\mathring{\Phi}.$ In particular, it has a base $\Pi$ and choose $p_\alpha\in Z_{\alpha}$ for $\alpha\in \Pi$. Define a $\bz$--linear function $p: Gr(\Psi)\to \bz$ extending $\alpha\mapsto p_\alpha, \alpha\in \Pi$.
We claim that $p_\alpha$ (the image of $\alpha$ under the map $p$) belongs to $Z_\alpha$ for all $\alpha\in Gr(\Psi).$ Since $p_{-\alpha}=-p_{\alpha}$ and $Z_{-\alpha}=-Z_\alpha$, it is enough to prove that $p_\alpha\in Z_\alpha$ for all positive roots $\alpha\in Gr(\Psi).$ Let $\alpha\in Gr(\Psi)$ be a positive root, then we can write $\alpha=\alpha_1+\cdots +\alpha_r$ with
each partial sum $\alpha_1+\cdots+\alpha_i$ is again a root in $Gr(\Psi)$ for $1\le i\le r.$ Since $\Psi$ is closed and $\Phi$ is untwisted, we have
$\alpha_1+\cdots+\alpha_i+(p_{\alpha_1}+\cdots +p_{\alpha_i})\delta\in \Psi$ for $1\le i\le r.$ In particular, we have $p_\alpha=p_{\alpha_1}+\cdots +p_{\alpha_r}\in Z_\alpha.$
\medskip
Since $Gr(\Psi)$ is not of type $A_1$, there exists $\beta, \gamma\in \Pi$ that are connected by an edge in the Dynkin diagram of $Gr(\Psi)$. Thus
$\beta+\gamma$ must be a root in $Gr(\Psi)$. By Lemma \Cref{lem:1} we have $A=Z_{\beta}^{'}=Z_{\gamma}^{'}$ and $A$ is a subgroup of $\bz$. By connectedness of Dynkin diagram of $Gr(\Psi)$,
we get $A=Z_{\beta}'$ for all simple roots $\beta$ of $Gr(\Psi)$. Suppose $\gamma \in Gr(\Psi)$ is a positive root, then $\gamma$ can be written as $\alpha+\beta$ where $\alpha$ is a positive root in $Gr(\Psi)$ and
$\beta\in \Pi$. Using induction on the height of $\alpha$ (with respect to $\Pi$), we have $Z_{\alpha}'=Z_{\beta}'=A$. Now again by Lemma \Cref{lem:1}, we have $Z_{\gamma}'=Z_{\alpha}'=Z_{\beta}'=A$. Hence we have $A=Z_{\alpha}'$ for all $\alpha \in Gr(\Psi)$. This completes the proof.
\end{proof}\fi
\iffalse
\subsection{} We now record our main theorem in the untwisted type.
\begin{cor}\label{thm:1}
Let $\Phi$ be a real untwisted irreducible affine root system.
Let $\Psi$ be a symmetric closed subset of $\Phi$ such that none of the irreducible components of $Gr(\Psi)$ is of type $A_1$. Then $\Psi$ is a closed subroot system of $\Phi$. In particular,
there exists a $n\in\bz_+$ and $\bz$-linear function $p:Gr(\Psi)\to\bz$, $\alpha\mapsto p_\alpha$, with $p_\alpha\in Z_\alpha,\, \text{for all} \ \alpha\in Gr(\Psi)$ such that $$\Psi=\{\alpha+(p_\alpha+n\bz)\delta:\alpha\in Gr(\Psi)\}.$$
\end{cor}
\begin{proof}
With out loss of generality, we assume that $\Psi$ is irreducible. It is easy to see that the condition $Z_\alpha+Z_\beta\subseteq Z_{\alpha+\beta}$ for all $(\alpha,\beta,\alpha+\beta)\in Gr(\Psi)^{\times 3}$
is always satisfied in this case. Now from \Cref{existp}, we get a $\bz$-linear map $p:Gr(\Psi)\to \bz$ such that
$p_\alpha\in Z_\alpha$ and \Cref{prop:1} completes the proof.
\end{proof}
\begin{proof}
With out loss of generality, we assume that $\Psi$ is irreducible. From Proposition \Cref{prop:1},
it is clear that we only need to verify the existence of $\bz$-linear map $p:Gr(\Psi)\to \bz$. We will verify it now.
Since $\Psi$ is irreducible and symmetric, we have $Gr(\Psi)$ is irreducible and symmetric. Since
$\Phi$ is untwisted, we have $Gr(\Psi)$ is closed. Then by Lemma \Cref{keylem:finite}, we have $Gr(\Psi)$ is a closed subroot system of $\mathring{\Phi}.$ In particular, it has a base $\Pi$ and choose $p_\alpha\in Z_{\alpha}$ for $\alpha\in \Pi$. Define a $\bz$--linear function $p: Gr(\Psi)\to \bz$ extending $\alpha\mapsto p_\alpha, \alpha\in \Pi$.
We claim that $p_\alpha\in Z_\alpha$ for all $\alpha\in Gr(\Psi).$ Since $p_{-\alpha}=-p_{\alpha}$ and $Z_{-\alpha}=-Z_\alpha$, it is enough to prove that $p_\alpha\in Z_\alpha$ for all positive roots $\alpha\in Gr(\Psi).$ Let $\alpha\in Gr(\Psi)$ be a positive root, then we can write $\alpha=\alpha_1+\cdots +\alpha_r$ with
each partial sum $\alpha_1+\cdots+\alpha_i$ is again a root in $Gr(\Psi)$ for $1\le i\le r.$ Since $\Psi$ is closed and $\Phi$ is untwisted, we have
$\alpha_1+\cdots+\alpha_i+(p_{\alpha_1}+\cdots +p_{\alpha_i})\delta\in \Psi$ for $1\le i\le r.$ In particular, we have $p_\alpha=p_{\alpha_1}+\cdots +p_{\alpha_r}\in Z_\alpha.$
This completes the proof. \end{proof}\fi
\iffalse \begin{proof}
Let $\alpha,\,\beta\,\in Gr(\Psi)$ be such that $\alpha+\beta\,\in\Phi^0\setminus Gr(\Psi)$. Then first of all $\alpha+\beta$ is a long root. Hence either both $\alpha,\beta$ are short roots or both are long roots by lemma ~\Cref{lem:7}. If both are long, then $\exists r,s\in\bz$ such that $\alpha+mr\delta$ and $\alpha+ms\delta$ are in $\Psi$. Hence their sum is a root and hence in $\Psi$ which contradicts that $\alpha+\beta\notin Gr(\Psi)$.
\end{proof}
We shall now investigate the properties of the sets $Z_\alpha\,$'s for $\alpha\in\,Gr(\Psi)$. First note that if $\Psi$ is symmetric and closed, then $Gr(\Psi)$ is symmetric and almost closed.
\begin{defn}
Let $\alpha,\beta\in Gr(\Psi)$ with $\beta\neq \pm \alpha$. Define $Z_\alpha\dotplus Z_\beta \,:=\{s'\,\mid\, s'=r+s$ with $r\in Z_\alpha,s\in Z_\beta$ and $\alpha+\beta +(r+s)\,\delta\in \Phi$\}.
\end{defn}
Note that the operation $\dotplus$ extends the usual $+$ for the untwisted affine root system.
\begin{prop}\label{prop:8}
Let $\Psi$ be a symmetric closed subset of $\Phi$. For $\alpha,\beta\in\,Gr(\Psi)$ if $\alpha+\beta$ is a root, then we have
\begin{equation}\label{eq:5}
Z_\alpha+Z_\beta\supseteq Z_{\alpha+\beta}
\end{equation}
Furthermore, if $\alpha+\beta$ is a short root or if $\alpha+\beta$ is long with at least one of $\alpha$ or $\beta$ is a long root ( hence both by lemma ~\Cref{lem:7} ), then we have
\begin{equation}\label{eq:6}
Z_\alpha+Z_\beta\subseteq Z_{\alpha+\beta}
\end{equation}
as well.
\end{prop}
\begin{proof}
Let $r\in Z_{\alpha+\beta}$. If one of the roots $\alpha,\,\beta$ is short, then it's very easy to see that the equation \Cref{eq:5} holds. Otherwise let both $\alpha,\beta$ be long. Then $\alpha+\beta$ is long, then $r$ is a multiple of $m$ . Let $s\in Z_\alpha$. Then $\beta+(s-r)\delta$ is a root in $\Psi$ [$\beta+(s-r)\delta\,=(\alpha+\beta+r\delta)-(\alpha+s\delta)$] since $s-r$ is a multiple of $m$. This proves the first part of the proposition.
Now Let $\alpha+\beta$ be a short root or $\alpha+\beta$ be long with both $\alpha$ or $\beta$ be long roots. Let $r\in\, Z_\alpha$ and $s\in \,Z_\beta$. Then we have $\alpha+r\delta\,\in \Psi$ and $\beta+s\delta\,\in \Psi$. For the second case both $r$ and $s$ are multiples of $m$. Since $\alpha+\beta$ is a short root ( long for the second case) we have $\alpha+\beta+\,(r+s)\delta\,\in \Phi$ ( holds for case $(2)$ because $r+s$ is a multiple of $m$). Since $\Psi$ is closed, we have $\alpha+\beta+\,(r+s)\delta\,\in \Psi$. Hence $r+s\,\in Z_{\alpha+\beta}$.
\end{proof}
\begin{example}
Let the type of $\Phi$ be $D_{n+1}^{(2)}$. Let $p_1,p_2,p_1',p_2',p_{12}^+,p_{12}^-,s$ be integers such that
\begin{enumerate}
\item $p_1,p_2,s$ are even.
\item $p_1',p_2'$ are odd.
\item $p_{12}^+=p_1+p_2=p_1'+p_2'$.
\item $p_{12}^-=p_1-p_2=p_1'-p_2'$.
\end{enumerate}
Let
\begin{center}
$\Psi^+:=\{\epsilon_1+(p_1+2s\bz)\delta\}\cup \{\epsilon_1+(p_1'+2s\bz)\delta\}\cup \{\epsilon_2+(p_2+2s\bz)\delta\}\cup \{\epsilon_2+(p_2'+2s\bz)\delta\}\cup\{\epsilon_1+\epsilon_2+(p_{12}^++2s\bz)\delta\}\cup \{\epsilon_1-\epsilon_2+(p_{12}^-+2s\bz)\delta\}$.
\end{center}
Let $\Psi:=\Psi^+\cup (-\Psi^+)$.
It's easy to see that $\Psi$ is a symmetric closed subset of $\Phi$ with gradient $\{\pm\epsilon_1,\pm\epsilon_2,\pm(\epsilon_1\pm\epsilon_2)\}$.
we shall show that $\Psi$ is not a closed subroot system.
Let $\alpha=\epsilon_1+(p_1+2sr)\delta$ and $\beta=\epsilon_1+(p_1'+2sr_1)\delta$. We have $s_\alpha(\beta)=-\epsilon_1+[p_1'-2p_1+2s(r_1-2r)]\delta$.
If $s_\alpha(\beta)\in\Psi$, then either $p_1'-2p_1+2s(r_1-2r)\delta=-p_1'+2sk$ or $p_1'-2p_1+2s(r_1-2r)\delta=-p_1+2sk_1$ for some integers $k,k_1$.
Hence either $2s\mathrel{|}2(p_1'-p_1)$ or $2s\mathrel{|}(p_1'-p_1)$. Since $s$ is even and $p_1'-p_1$ is odd, none of them can be true, a contradiction. Hence $s_\alpha(\beta)\not\in\Psi$. Hence $\Psi$ is not a closed subroot system.
\end{example}
\fi
\subsection{}\label{closedcase} When $Gr(\Psi)$ is closed in $\mathring{\Phi}$, using \Cref{prop:subroot} 4(a), we see that to classify a symmetric closed subsets of $\Phi$ we only need to classify irreducible symmetric closed subsets of $\Phi.$
So without loss of any generality we assume that $\Psi$ is irreducible symmetric and closed in $\Phi$ in what follows.
\begin{prop}\label{prop:equal1}
Let $\Phi$ be a real affine root system not of type $A_{2n}^{(2)}$.
Let $\Psi$ be an irreducible symmetric closed subset of $\Phi$ such that $Gr(\Psi)$ is closed in $\mathring{\Phi}$ and it is not of type $A_1$. If there exists a short root $\beta$ such that
$|\pi_m(Z_\beta)|=1$, then
there exists $n\in \bz_+$ and a $\bz$-linear function $p:Gr(\Psi)\to\bz$, $\alpha\mapsto p_\alpha$, with $p_\alpha\in Z_\alpha,\, \text{for all} \ \alpha\in Gr(\Psi)$ such that $$\Psi=\{\alpha+(p_\alpha+n\bz)\delta:\alpha\in Gr(\Psi)\}$$
In particular $\Psi$ is a closed subroot system of $\Phi$. \end{prop}
\begin{proof}
Note that $Gr(\Psi)$ must be a closed subroot system of $\mathring{\Phi}.$
Fix a short simple root $\beta\in Gr(\Psi)$ such that $|\pi_m(Z_\beta)|=1$. We first show that $|\pi_m(Z_\alpha)|=1$ for all simple roots $\alpha$. We prove this by induction on $k$ where $\beta=\alpha_0,\cdots,\alpha_k=\alpha$ is the unique path from $\beta$ to $\alpha$.
Set $\gamma_j=\sum_{i=0}^{j}\alpha_{i}$ for $0\le j\le k$, and note that $\gamma_j\in Gr(\Psi)$, for all $0\le j\le k$.
We claim that $|\pi_m(Z_{\alpha_j})|=1$ and $|\pi_m(Z_{\gamma_j})|=1$, for all $0\le j\le k$.
For $k=1$, we have $\alpha$ is adjacent to
$\beta$ and $\alpha+\beta\in Gr(\Psi)$.
Consider $Z_{\alpha+\beta}+Z_{-\alpha}\subseteq Z_\beta$ which is always true as $\beta$ is short. This gives $Z_{\alpha+\beta}$ and
$Z_{-\alpha}$ must contain only one congruence class modulo $m.$ Since $Z_{-\alpha}=-Z_\alpha$, we have $|\pi_m(Z_\gamma)|=1$ for $\gamma=\alpha, \alpha+\beta.$
For general $k$, we must have $\gamma_{k-1}$ is short by \Cref{longroot} since $\beta$ is short. Since $\gamma_k=\gamma_{k-1}+\alpha$, we have $Z_{\gamma_k}+Z_{-\alpha}\subseteq Z_{\gamma_{k-1}}$. By induction hypothesis, we have $|\pi_m(Z_{\gamma_{k-1}})|=1$, and this implies the desired claim.
We shall now show that $|\pi_m(Z_\alpha)|=1$ for every positive root $\alpha.$ We proceed by induction on ht$(\alpha)$.
If $\alpha$ is long or ht$(\alpha)=1$, then there is nothing to prove.
Assume that ht$(\alpha)>1$, then we can write $\alpha=\gamma+\alpha_k$ where $\gamma\in \mathring{\Phi}^+ $ and $\alpha_k$ is a simple root.
If $\alpha_k$ is short, we have $Z_{\gamma}+Z_{-\alpha}\subseteq Z_{-\alpha_k}$ and this immediately implies that $|\pi_m(Z_\alpha)|=1$.
If $\alpha_k$ is long, then $\gamma$ is short and we have $Z_{\alpha}+Z_{-\alpha_k}\subseteq Z_{\gamma}$. By induction hypothesis, we have $|\pi_m(Z_\gamma)|=1$, hence we get $|\pi_m(Z_\alpha)|=1$. We now check the condition \eqref{eq:z}
is satisfied. Since $\Psi$ is closed, the only non-trivial case is when $\Phi$ is twisted and
$(\alpha_1,\alpha_2,\alpha_1+\alpha_2)\in Gr(\Psi)^{\times 3}$ is of type $(s,s,\ell)$.
Since $Z_{\alpha_1+\alpha_2}+Z_{-\alpha_1}\subseteq Z_{\alpha_2}$ holds, we must have $r+s\equiv 0\ (\mathrm{mod}\ m)$ for all $r\in Z_{\alpha_1}$ and $s\in Z_{\alpha_2}$.
This implies $Z_{\alpha_1}+Z_{\alpha_2}\subseteq Z_{\alpha_1+\alpha_2}$. Now \Cref{prop:1} completes the proof.
\end{proof}
\begin{cor}\label{coruntwisted}
Let $\Phi$ be a real untwisted irreducible affine root system.
Let $\Psi$ be a symmetric closed subset of $\Phi$ such that none of the irreducible components of $Gr(\Psi)$ is of type $A_1$. Then $\Psi$ is a closed subroot system of $\Phi$.
\end{cor}
\begin{proof}
Since $m=1,$ we get the result from \cref{prop:equal1}.
\end{proof}
\begin{rem}
The assumptions on the $Gr(\Psi)$ in \cref{prop:1} and \ref{prop:equal1} and \Cref{coruntwisted} are very sharp, i.e., the conclusions of \cref{prop:1} and \ref{prop:equal1} and \Cref{coruntwisted} are not valid when one of the components of $Gr(\Psi)$ is of type $A_1$. For example, we can take $$\Psi=\{\alpha+\delta,-\alpha+\delta,-\alpha-\delta,\alpha-\delta\}$$ in $A_1^{(1)}$, it is symmetric closed but not subroot system.
Take $\alpha_1=\alpha+\delta,\,\beta_1=-\alpha+\delta$, then $s_{\alpha_1}(\beta_1)=\alpha+3\delta\notin \Psi$.
Note that $Z_\alpha(\Psi)=\{\pm 1\}$ which is far from being a coset.
\end{rem}
\iffalse
\begin{prop}
Let $\Phi$ be a real affine root system not of type $A_{2n}^{(2)}$. Let $\Psi$ be a symmetric closed subset of $\Phi$ such that $Gr(\Psi)$ is a closed subroot system of $\mathring{\Phi}$.
If there exists
a short root $\alpha\in Gr(\Psi)$ such that $|\{r(\mathrm{mod}\ m): r\in Z_{\alpha}\}|=1$, then we have $\Psi$ is a closed subroot system of $\Phi$.
\end{prop}
\begin{proof} By Lemma \Cref{keylem:2}, we have
$|\{r(\mathrm{mod}\ m): r\in Z_\alpha\}|=1$ for all $\alpha\in Gr(\Psi)$.
We claim that the condition \eqref{eq:z} is satisfied. Let $\alpha,\beta\in Gr(\Psi)$ such that $\alpha+\beta\in Gr(\Psi).$
Only non-trivial case is $(\alpha,\beta,\alpha+\beta)$ is of type $(s,s,\ell)$. In this case, we have
$Z_{\alpha+\beta}+Z_{-\alpha}\subseteq Z_\beta$ for sure. Let $r\in Z_\alpha$, $r'\in Z_{\alpha+\beta}$ and $s\in Z_\beta$.
Since $-Z_\alpha=Z_{-\alpha}$, we must have
$r'+(-r)\equiv s (\rm{mod}\ m)$ as $|\{r(\mathrm{mod}\ m): r\in Z_\beta\}|=1$. This implies $r+s\equiv 0\ (\rm{mod}\ m)$ as $Z_{\alpha+\beta}\subseteq m\bz$, and hence the condition
\eqref{eq:z} is satisfied. Now Lemma \Cref{existp} and Proposition \Cref{prop:1} completes the proof.
\end{proof}
\fi
\subsubsection{}
Now we assume that there is a short root $\beta\in Gr(\Psi)$ such that $|\pi_m(Z_\beta)|>1$. In this case, we have to deal with the case $Gr(\Psi)$ is of type $B_2$ separately. First we assume that $Gr(\Psi)$ is not of type $B_2$, then we have:
\begin{prop}\label{prop:notb2}
Let $\Phi$, $\Psi$, $Gr(\Psi)$ be as before in \Cref{prop:equal1}.
Further assume that $Gr(\Psi)$ is not of type $B_2$. Suppose that there exists a short root $\beta\in Gr(\Psi)$ such that $|\pi_m(Z_\beta)|>1$,
then we have,
\begin{equation}\label{notb2}
\Psi=\{\alpha+(p_\alpha+n_s\bz)\delta:\alpha\in Gr(\Psi)_s\}\cup \{\gamma+(p_\gamma+mn_s\bz)\delta:\gamma\in Gr(\Psi)_\ell\}
\end{equation}
for some $n_s\in \bz_+$. In particular $\Psi$ is a closed subroot system.
\end{prop}
\begin{proof}
Clearly $\Phi$ is twisted and we have $|\pi_m(Z_\alpha)|>1$ for all short roots $\alpha\in \Pi$ by the proof of \Cref{prop:equal1}. Let $Z_{\gamma, 0}=Z_\gamma\cap m\bz,$ for $\gamma\in Gr(\Psi)$ and note that $Z_{\gamma,0}=Z_\gamma$ for long roots $\gamma\in Gr(\Psi)$. Define $\Psi_0$ as follows
$$\Psi_0:=\{\gamma+Z_{\gamma, 0}\delta:\gamma\in Gr(\Psi)_s\}\cup \{\gamma+Z_{\gamma,0}\delta:\gamma\in Gr(\Psi)_\ell\}$$
If $m=2$, clearly $Z_{\gamma, 0}\neq \emptyset$ for all $\gamma\in Gr(\Psi)$. If $m=3$, then $\Phi=D_4^{(3)}$. Since $Gr(\Psi)$ is not of type $A_1$ and it is irreducible, we must have $Gr(\Psi)=\mathring{\Phi}$ (which is of type $G_2$). If $\{\alpha_1,\alpha_2\}$ is a basis of $\mathring{\Phi}$ with $\alpha_1$ is short, then
the short roots of $\mathring{\Phi}$ are $\alpha_1, \alpha_2+\alpha_1$ and $\alpha_2+2\alpha_1$.
Since $Z_{\alpha_2}\subseteq m\bz$ and $Z_{\alpha_2}+Z_{\alpha_1}\subseteq Z_{\alpha_2+\alpha_1}$, we have $$\pi_m(Z_{\alpha_1})\subseteq \pi_m(Z_{\alpha_2+\alpha_1}).$$ Let $x, y\in \pi_m(Z_{\alpha_1})$ such that $x\neq y.$
Since $Z_{\alpha_2+\alpha_1}+Z_{\alpha_1}\subseteq Z_{\alpha_2+2\alpha_1}$, we have
$2x, x+y, 2y\in \pi_m(Z_{\alpha_2+2\alpha_1}).$ Since $x\neq y$, we have
$|\pi_m(Z_{\alpha_2+2\alpha_1})| = 3$. Now since
$Z_{\alpha_2+2\alpha_1}+Z_{-\alpha_1}\subseteq Z_{\alpha_2+\alpha_1}$, we have
$|\pi_m(Z_{\alpha_2+\alpha_1})| = 3$.
Again since $Z_{\alpha_2+\alpha_1}+Z_{-\alpha_2}\subseteq Z_{\alpha_1}$, we must have
$|\pi_m(Z_{\alpha_1})| = 3$.
This proves that $Z_{\alpha,0}\neq \emptyset$ for $\alpha\in \mathring{\Phi}_s$. Hence $Z_{\gamma,0}\neq \emptyset$ for all $\gamma\in Gr(\Psi)$.
\medskip
Clearly the condition \eqref{eq:z} is satisfied for $\Psi_0$. Hence we have a $\bz$-linear function $p:Gr(\Psi_0)\to \bz$ with $p_\gamma\in Z_{\gamma,0}$ for all $\gamma\in Gr(\Psi_0)=Gr(\Psi)$. All the necessary conditions of \Cref{prop:1} are satisfied for $\Psi_0$. Since $Gr(\Psi)$ is irreducible, there exists $n_\ell\in m\bz$ such that
$Z'_{\gamma, 0}=n_\ell \bz$ for all $\gamma\in Gr(\Psi)$. In particular, we have
$Z'_{\gamma}=n_\ell \bz$ for all long roots $\gamma\in Gr(\Psi)$.
\medskip
Now let $Gr(\Psi)$ be of type $B_n$ with $n\geq 2$, in particular $m=2$. We claim that $Z_\alpha'$ is a union on cosets modulo $n_\ell\bz$ for all $\alpha\in Gr(\Psi)_s$. Let $\alpha\in Gr(\Psi)$ be a short root. Since $Gr(\Psi)$ is irreducible, there exists a long root $\gamma$ such that $(\alpha,\gamma)\neq 0$. Without loss of generality let $(\alpha,\gamma)<0$. Since $\Psi$ is closed, we have the following relations
$$Z_\gamma+Z_\alpha\subseteq Z_{\alpha+\gamma},\ Z_{\alpha+\gamma}+Z_{-\gamma}\subseteq Z_{\alpha}.$$
Since $p$ is $\bz$-linear, we have
$$Z_\gamma'+Z_\alpha'\subseteq Z_{\alpha+\gamma}',\ \ Z_{\alpha+\gamma}'+Z_{-\gamma}'\subseteq Z_{\alpha}'.$$
Since $Z_\gamma'=n_\ell \bz$, the above relations show that $Z_\alpha'=Z_{\alpha+\gamma}'$ and $Z_\alpha'+n_\ell\bz\subseteq Z_\alpha'$. Thus we have
$Z_\alpha'$ is a union of cosets modulo $n_\ell\bz$ for all short roots $\alpha\in Gr(\Psi)$.
Now if $\alpha$ and $\gamma$ are short roots such that $\alpha+\gamma$ is long, then we have
\begin{equation}\label{keyeq2}
(Z_\alpha'+Z_\gamma')\cap m\bz\subseteq Z_{\alpha+\gamma}',\ \ Z_{\alpha+\gamma}'+Z_{-\gamma}'\subseteq Z_{\alpha}'
\end{equation}
Fix a short root $\alpha$ and write $Z_\alpha'=\cup_i(a^i_{\alpha}+n_\ell\bz)$ with $0\le a^i_{\alpha}\neq a^j_{\alpha}< n_\ell,\ i\neq j$. Since $Gr(\Psi)$ is of type $B_n\ (n\geq 2)$, there is a short root $\gamma$ such that $\alpha+\gamma$ is long. Equations in \eqref{keyeq2} imply that $n_\ell\bz\subseteq Z_{\alpha}'$ and hence $a^k_\alpha=0$ for some $k$ and $2\nmid a^j_\alpha$ for all $j\neq k$. Since $|\pi_m(Z_\alpha)|>1,$ there exists $i$ such that $a_\alpha^i$ is odd. Let $a_\gamma^r$ be one such odd coset for $\gamma$. If $a^i_\alpha,a^j_\alpha$ are both odd, then by equation \eqref{keyeq2}, we get that $a^i_\alpha+a^r_\gamma\equiv 0\ (\mathrm{mod}\ n_\ell)$ and $a^j_\alpha+a^r_\gamma\equiv 0\ (\mathrm{mod}\ n_\ell)$. Hence it follows that $a^i_\alpha=a^j_\alpha$ and $Z_\alpha'$ is a union of exactly two cosets $n_\ell\bz\cup (a_\alpha+n_\ell\bz)$ with $a_\alpha$ odd. Similarly $Z_\gamma'=n_\ell\bz\cup (a_\gamma+n_\ell\bz)$. Furthermore $a_\alpha+a_\gamma\equiv 0(\mathrm{mod}\ n_\ell)$. Till this point the argument is valid for $B_2$ as well.
Now assume that $n\geq 3$. Let $\beta_1,\beta_2$ be short roots. Then there is a short root $\beta_3$ such that $\beta_i+\beta_j$ is a long root for $1\le i\neq j\le 3$. Applying equation \eqref{keyeq2} for the triples $(\beta_1,\beta_2,\beta_1+\beta_2),\ (\beta_2,\beta_3,\beta_2+\beta_3),\ (\beta_1,\beta_3,\beta_1+\beta_3)$ successively we get that $$a_{\beta_1}+a_{\beta_2}\equiv 0\ (\mathrm{mod}\ n_\ell),\ \ a_{\beta_2}+a_{\beta_3}\equiv 0\ (\mathrm{mod}\ n_\ell),\ \ a_{\beta_1}+a_{\beta_3}\equiv 0\ (\mathrm{mod}\ n_\ell)$$
Hence we have $a_{\beta_1}=a_{\beta_2}=a_{\beta_3}=n_\ell/2$. This proves that $Z_{\beta_1}'=Z_{\beta_2}'=n_\ell\bz\cup (n_\ell/2+n_\ell\bz)=\frac{n_\ell}{2}\bz.$ Hence if $Gr(\Psi)$ is of type $B_n(n\geq 3),$ then $\Psi$ is of the form \eqref{notb2}.
\medskip
Now let $Gr(\Psi)$ be not of type $B_n$. Let $\Psi'=\bigcup_{\alpha\in \mathring{\Phi}_s}\{\alpha+r\delta\in \Psi:r\in \bz\}$. Since $Gr(\Psi)_s=Gr(\Psi')$, we have
$\Psi'$ is symmetric closed subset of $\mathring{\Phi}_s^{(1)}$. Note that $Z_\alpha(\Psi)=Z_\alpha(\Psi')$ for all $\alpha\in Gr(\Psi)_s.$
Since $Gr(\Psi)_s$ is irreducible, by \Cref{prop:1}, there exists $n_s\in \bz_+$ such that $Z_\alpha'(\Psi)=Z_\gamma'(\Psi)=n_s\bz$ for all $\alpha,\gamma\in Gr(\Psi)_s$.
Note that given $\alpha\in Gr(\Psi)_s$, there exists a $\gamma\in Gr(\Psi)_s$ such that $\alpha+\gamma\in Gr(\Psi)_\ell$. Applying equation \eqref{keyeq2}, we get that $n_\ell\bz\subseteq n_s\bz$ and $mn_s\bz=n_s\bz\cap m\bz\subseteq n_\ell\bz$. We have $n_s\mid n_\ell$ and $n_\ell\mathrel{|} mn_s$. Let $n_\ell=rn_s$. Then $r\mid m$. Since $m$ is prime, either $r=1$ or $r=m$. If $r=1,$ then $n_\ell=n_s$, and since $m\mid n_\ell$, we have that $|\pi_m(Z_\alpha)|=1$ for all short roots $\alpha$ which is a contradiction.
Hence $r=m$ and consequently, $\Psi$ is of the form \eqref{notb2}. This completes the proof.
\end{proof}
\begin{prop}\label{prop:b2}
Let $\Phi$, $\Psi$, $Gr(\Psi)$ as before in \Cref{prop:equal1}.
Assume that there exists a short root $\beta\in Gr(\Psi)$ such that $|\pi_m( Z_\beta)|>1$. Suppose $Gr(\Psi)$ is of type $B_2$, then we have $\Psi=\Psi^+\cup (-\Psi^+)$ where $\Psi^+$ is of the form
\begin{equation}\label{b2}
\Psi^+=\{\epsilon_i+ (p_{\epsilon_i}+A_i)\delta: i=1, 2\}\cup\{\alpha+(p_\alpha+n_\ell\bz)\delta:\alpha\in \mathring{\Phi}^+_\ell\}
\end{equation}
where
\begin{enumerate}
\item $p:Gr(\Psi)\to \bz$ is a $\bz$-linear function such that $p_{\epsilon_i}\in 2\bz$ for $i=1,2$,
\item $A_i=n_\ell\bz \cup (a_i+n_\ell\bz),$
$n_\ell\in 2\bz_+$, $0\le a_1,a_2< n_\ell$ such that $a_1+a_2\equiv 0\ (\mathrm{mod}\ n_\ell)$ and both $a_1, a_2$ are odd.
\end{enumerate}
Moreover, we have $\Psi$ is a closed subroot system if and only if $a_1=a_2=n_\ell/2$. In this case $\Psi$ is of the form
\begin{equation}\label{b2subroot}
\Psi=\{\alpha+(p_\alpha+n_s\bz)\delta:\alpha\in Gr(\Psi)_s\}\cup \{\beta+(p_\beta+mn_s\bz)\delta:\beta\in Gr(\Psi)_\ell\}
\end{equation}
\end{prop}
\begin{proof}
Note that in this case, from the proof of \Cref{prop:notb2}, $Gr(\Psi)$ is of the form \eqref{b2}. If $a_1\equiv a_2(\mathrm{mod}\ n_\ell),$ then along with $a_1+a_2\equiv 0(\mathrm{mod}\ n_\ell)$, we get that $a_1=a_2=n_\ell/2$. Hence $(n_\ell\bz)\cup (a_2+n_\ell\bz)=(n_\ell/2)\bz$ and hence $\Psi$ is of the form \eqref{notb2}. In particular, $\Psi$ is a closed subroot system.
Conversely, suppose that $Gr(\Psi)$ is of type $B_2$ and $\Psi$ given in \eqref{b2} is a closed subroot system of $\Phi$. Let $\alpha=\epsilon_i+(p_{\epsilon_i}+a_i)\delta, \ \alpha'=\epsilon_i+p_{\epsilon_i}\delta$. Then $s_{\alpha'}(\alpha)=-\epsilon_i+(p_{-\epsilon_i}+a_i)\delta$. Since $p_{-\epsilon_i}+a_i$ is odd and $p_{-\epsilon_i}+(-a_i)+n_\ell\bz$ is the coset in $Z_{-\epsilon_i}$ which contains odd integers, we must have $a_i\equiv -a_i(\mathrm{mod}\ n_\ell)$. Hence $a_i=n_\ell/2$ for $i=1,2.$
\end{proof}
\section{Semi-closed gradient case}\label{semiclosedsect}
As before, we assume that $\Phi$ is a real twisted affine root system which is not of type $A^{(2)}_{2n}$. In this section,
we consider the case when $\Psi$ is a symmetric closed subset of $\Phi$ with $Gr(\Psi)$ is semi-closed in $\mathring{\Phi}.$ We cannot assume that $\Psi$ is irreducible as in \cref{closedcase}, because we may miss the interaction between two roots coming from the different orthogonal components of \eqref{decompgrpsi} (see \Cref{remark1}). Moreover, we have to deal with each individual affine root system separately due to their gradient behaviour.
\subsection{Twisted real affine root system of type \texorpdfstring{$D_{n+1}^{(2)}$}{D(2)n+1}}\label{Dn+1} In this case, the real affine roots are given by
$$\Phi=\{\pm \epsilon_i\,+r\,\delta\,\mid r\in\mathbb{Z},1\le i\le n\}\cup\,\{\pm\epsilon_i\pm\epsilon_j\,+2r\delta\,\mid r\in \mathbb{Z},\,1\le i\neq j\le n\}$$
and $\mathring{\Phi}$ is of type $B_n, n\ge 2$. We assume that $\Psi$ is a symmetric closed subset of $\Phi$ such that $Gr(\Psi)$ is semi-closed in $\mathring{\Phi}$. For $J\subseteq I_n$, we define $$B_{J}:=\{\pm\epsilon_{j}:j\in J\}\cup \{\pm\epsilon_j\pm \epsilon_{j'}: j\neq j'\in J\}.$$
\iffalse
To describe explicitly the root systems of type $A$ which occur in type $D$, we need the following definition.
\begin{defn}
Let $S$ be a totally ordered subset of $I_n$ with a total order $\le$ which need not be the usual total order on $\bz$. Write $S=\{{p_0},{p_1},{p_2},\dots,{p_r}\}$ where ${p_i}\le {p_j}$ if and only if $i\le j$.
Given a function $f:S\to \{\pm 1\}$ such that $f(p_0)=-1$, define $$\alpha_i:=-f({p_{i-1}})\epsilon_{p_{i-1}} +f({p_{i}})\epsilon_{p_{i}}\ \text{for }1\le i\le r.$$
Set $A(S,f):=A(S,f)^+\cup (-A(S,f)^+)$ where $A(S,f)^+:=\bz_{+}\{\alpha_i:1\le i\le r\}\cap D_{n+1}.$
\end{defn}
\medskip
To understand the closed subroot systems in $D_{n}$, we need the following lemmas.
\begin{lem}\label{typeAD1}
Let $\mathring{\Psi}$ be a closed subroot system of the finite root system of type $D_n$ such that no irreducible component of $\mathring{\Psi}$ is of type $A_1$.
\begin{enumerate}
\item If $\epsilon_p\pm\epsilon_q\in \mathring{\Psi}$, then both lie in the same irreducible component of $\mathring{\Psi}$.
\item Suppose that $\mathring{\Psi}'$ is an irreducible component of $\mathring{\Psi}$ such that $\epsilon_p\pm\epsilon_q\in \mathring{\Psi}'$ for some $p,q\in I_n.$ If $\epsilon_p+\epsilon_r\in \mathring{\Psi}$ (resp. $\epsilon_p-\epsilon_r\in\mathring{\Psi}$), then $\epsilon_p-\epsilon_r\in \mathring{\Psi}'$ (resp. $\epsilon_p+\epsilon_r\in\mathring{\Psi}'$).
\end{enumerate}
\end{lem}
\begin{proof}
Since the component containing $\epsilon_p+\epsilon_q$ is not of type $A_1$, there exists $\epsilon_q\pm \epsilon_r\in\mathring{\Psi}$ in the irreducible component of $\epsilon_p+\epsilon_q$ with $r\neq p$. Since $\epsilon_p-\epsilon_q$ is not orthogonal to $\epsilon_q\pm \epsilon_r$, it follows that all the three roots lie in the same irreducible component. This proves $(1)$.
To prove $(2)$ without loss of generality let $\epsilon_p+\epsilon_r\in\mathring{\Psi}$ and $r\neq q$. Since $\mathring{\Psi}$ is closed, we have $$\epsilon_q+\epsilon_r=(\epsilon_p+\epsilon_r)-(\epsilon_p-\epsilon_q)\in\mathring{\Psi}, \ \ \epsilon_p-\epsilon_r=(\epsilon_p+\epsilon_q)-(\epsilon_q+\epsilon_r)\in \mathring{\Psi}$$
Since $(\epsilon_p-\epsilon_r,\epsilon_p+\epsilon_q)\neq 0$, it follows that $\epsilon_p-\epsilon_r\in\mathring{\Psi}'$.
\end{proof}
\begin{lem}\label{typeAD2}
Suppose $\mathring{\Psi}'$ is an irreducible component of $\mathring{\Psi}$ which has the property that: for all $k,\ell\in I_n$, we have
$$|\{\epsilon_k\pm \epsilon_\ell\}\cap \mathring{\Psi}'|\le 1$$ then $\mathring{\Psi}'$ is a root system of type $A$ in $\mathring{\Psi}$.
\end{lem}
\begin{proof}
From the given condition we can write each positive root $\epsilon_i\pm\epsilon_j\in Gr(\Psi)$ simply as a pair $(i,j).$ Note that for two such pair $(i,j)$ and $(k,\ell), |\{i,j\}\cap \{k,\ell\}|=0$ if and only if the corresponding roots are orthogonal. If $\mathring{\Psi}'$ is a root system of type $D$, then the Dynkin diagram of $\mathring{\Psi}'$ looks like
\begin{figure}[h]
\centering
\begin{tikzpicture}[scale=0.6,every node/.style={transform shape}]
\tikzstyle{S}=[circle,draw=red,fill=red]
\node[S] (3) at (-2,0) [label=below:$\alpha_3$] {};
\node[S] (1) at (0,0) [label=below:$\alpha_1$] {};
\node[S] (2) at (2,1) [label=above:$\alpha_2$] {};
\node[S] (4) at (2,-1) [label=below:$\alpha_4$] {};
\node[S] (5) at (-5,0) {};
\node[S] (7) at (-7,0) {};
\draw[-,blue] (3) to node {} (1);
\draw[-,blue] (1) to node {} (2);
\draw[-,blue] (1) to node {} (4);
\draw[-,blue] (7) to node {} (5);
\draw[dashed,blue] (5) to node {} (3);
\end{tikzpicture}
\end{figure}
Let $\alpha_1=(p,q), \ \alpha_2=(a,b), \ \alpha_3=(c,d).$ Since $(\alpha_2,\alpha_1)=(\alpha_3,\alpha_1)=-1, \ \ (\alpha_2,\alpha_3)=0,$ we have that $|\{p,q\}\cap \{a,b\}|=1, |\{p,q\}\cap \{c,d\}|=1, |\{a,b\}\cap \{c,d\}|=0.$ Hence exactly one of $p$ and $q$ occurs in one of $\{a,b\}$ and $\{c,d\}$ and the other occurs in the other set. Let $\alpha_4=(e,f).$ Since $(\alpha_4,\alpha_1)=-1,$ either $p$ or $q$ will be in $\{e,f\}$ which is impossible since $(\alpha_4,\alpha_2)=(\alpha_4,\alpha_3)=0.$
We have $(\alpha_4,\alpha_1)=(\alpha_2,\alpha_1)=(\alpha_3,\alpha_1)=-1, \ \ (\alpha_4,\alpha_2)=(\alpha_4,\alpha_3)=(\alpha_2,\alpha_3)=0.$ We shall prove that no such $\alpha_1,\alpha_2,\alpha_3,\alpha_4$ exist in one particular case. Other cases are very similar. Let $\alpha_2=\epsilon_p+\epsilon_q$ and $\alpha_3=\epsilon_s-\epsilon_t$ with $\{p,q\}\cap \{s,t\}=\emptyset$. Since $(\alpha_2,\alpha_1)=-1$, without loss of generality we can assume that $\alpha_1=\epsilon_r-\epsilon_q$ where $r\neq p$. Again since $(\alpha_1,\alpha_3)=-1,$ we must have $r=t$ or $s=q$. Let $r=t$. Suppose that $\alpha_4=\epsilon_a+\epsilon_b$. Since $(\alpha_4,\alpha_1)=-1$, we have either $a=q$ or $b=q$. In any case $(\alpha_4,\alpha_2)\neq 0$ which is a contradiction. If $\alpha_4=\epsilon_a-\epsilon_b$, then either $b=r$ or $a=q$. If $a=q$ we have $(\alpha_4,\alpha_2)\neq 0$ (note that $b\neq p$). If $b=r$, then $(\alpha_4,\alpha_3)\neq 0$ (since $a\neq s$), contradiction in both the cases. This completes the proof.
\end{proof}
\begin{rem}
$A(S,f)$ is a root system of type $A$.
\end{rem}
Recall that by definition, a semi-closed subset is symmetric. Classification of all semi-closed subset of root system of type $B_n$ is a cumbersome task. Maximal semi-closed subsets were determined in \cite{DV21}. For our analysis, it is sufficient to determine the semi-closed subsets which occur as the gradient of some symmetric closed subset of $D_{n+1}^{(2)}$. These sets are determined in the next lemma. It generalises \cite[Theorem 2(2)]{DV21} for $B_n$ type.
\medskip
Throughout the discussion we let $\Gamma=\mathbb{A}\cup \mathbb{D}\cup \{o,e\}$ be an index set.
\fi
\noindent
Recall that by definition, a semi-closed subset is symmetric.
First we will determine the semi-closed subsets of $\mathring{\Phi}$ which occur as the gradient of some symmetric closed subset of $D_{n+1}^{(2)}$. Maximal semi-closed subroot systems of $\mathring{\Phi}$ were determined in \cite{DV21}. The following result generalizes \cite[Theorem 2(2)]{DV21} for $B_n$ type.
\begin{lem}\label{semiclosedD}
Let $\Phi$ be of type $D_{n+1}^{(2)}$ and $\Psi$ a symmetric closed subset of $\Phi$. Suppose that $\mathring{
\Psi}$ is a semi-closed subset of $\mathring{\Phi}$ such that $\mathring{\Psi}=Gr(\Psi)$, then there exist non-empty subsets $J_o$ and $J_e$ of $I_n$ such that $J_e\cap J_o=\emptyset$ and $\mathring{\Psi}\backslash (B_{J_e}\cup B_{J_o})$ is a closed subroot system of long roots of $B_n$.
\end{lem}
\begin{proof}
Assume that $\mathring{\Psi}=Gr(\Psi)$ for some symmetric closed subset $\Psi$ of $D_{n+1}^{(2)}$. Set $Z_p=Z_{\epsilon_p}$ for each $p\in I$ where $$I:=\{p\in I_n\mid \epsilon_p\in \mathring{\Psi}\}.$$ Note that $\epsilon_p+\epsilon_q\in \mathring{\Psi}$ for $p, q\in I$ if and only if $(Z_p+Z_q)\cap 2\bz\neq \emptyset$.
Since $Z_{-\epsilon_p}=-Z_p$, we have $\epsilon_p+\epsilon_q\in \mathring{\Psi}$ for $p, q\in I$ if and only if $\epsilon_p-\epsilon_q\in \mathring{\Psi}$ for $p, q\in I$.
Since $\mathring{\Psi}$ is semi-closed, there exist $\epsilon_i,\epsilon_j\in \mathring{\Psi}$ such that $\epsilon_i+\epsilon_j\notin \mathring{\Psi}$, this implies $I\neq \emptyset$ and
$(Z_i+Z_j)\cap 2\bz=\emptyset$. Without loss of generality, assume that $Z_i\subseteq 2\mathbb{Z}$ and $Z_j\subseteq 2\mathbb{Z}+1$.
We claim that for any $k\in I$, we have either $Z_k\subseteq 2\bz$ or $Z_k\subseteq 2\bz+1$. Suppose that for some $k\in I$, $Z_k$ contains both even and odd elements. Then $\exists \,r,s\in\bz$ such that $\epsilon_k+2r\delta,\epsilon_k+(2s+1)\delta\in\Psi$. Let $2q\in Z_i$ and $2p+1\in Z_j$ be arbitrary. Then we have
$\epsilon_k+\epsilon_i+2(r+q)\delta, \epsilon_k-\epsilon_j+2(s-p)\delta\in \Psi$ and
$$\epsilon_k+\epsilon_i+2(r+q)\delta-(\epsilon_k-\epsilon_j+2(s-p)\delta)= \epsilon_i+\epsilon_j+2(r+q+p-s)\delta\in \Psi$$ as $\Psi$ is closed in $\Phi$. This implies
$\epsilon_i+\epsilon_j\in \mathring{\Psi}$, a contradiction. Hence the claim.
Now define $J_e:=\{ p\in I\,\mid Z_p\subseteq 2\bz\}$ and $J_o:=I\backslash J_e$.
We claim that $(s, t), (t, s)\in J_e\times J_o$ implies that $\pm\epsilon_s\pm \epsilon_t\notin \mathring{\Psi}$.
Suppose that $\pm \epsilon_s\pm \epsilon_t+2k\delta\in \Psi$ with $(s, t)\in J_e\times J_o$. Then we have $$\pm\epsilon_t+2(k+r^{\mp})\delta\,=(\pm \epsilon_s\pm \epsilon_t+2k\delta)+(\mp \epsilon_s+2r^{\mp}\delta)\in \Psi,$$
where $r^{\mp}\in\bz$ such that $\mp \epsilon_s+2r^{\mp}\delta\in \Psi$. This is a contradiction since $t\in J_o$. By symmetry we get $(t, s)\in J_e\times J_o$ implies that $\pm\epsilon_s\pm \epsilon_t\notin \mathring{\Psi}$.
Note that $\pm\epsilon_i\pm \epsilon_j\in \mathring{\Psi}$ for all $i,j\in J_e\ $(or $J_o$). Hence $\{\pm \epsilon_i: i\in I\}$ generates a root system of type $B_{J_e}\sqcup B_{J_o}$ in $\mathring{\Psi}$. If $\mathring{\Psi}=B_{J_e}\sqcup B_{J_o}$, then there is nothing to prove. So assume that
$\mathring{\Psi}\neq B_{J_e}\sqcup B_{J_o}$. It is clear that $\mathring{\Psi}\backslash (B_{J_e}\sqcup B_{J_o})$ is a symmetric closed subset of $(B_n)_\ell$ since
$\mathring{\Psi}$ is semi-closed in $\mathring{\Phi}.$ Hence it is a closed subroot system of $(B_n)_\ell$ by \Cref{keylem:finite}.
\end{proof}
\begin{rem}
\begin{enumerate}
\item Since the long root of $B_n$ forms a root system of type $D_n$,
it follows that $\mathring{\Psi}\backslash (B_{J_e}\sqcup B_{J_o})$ is a union of root systems of type $A$ and $D$ (\cite{BdS}).
\item The converse of \Cref{semiclosedD} is also true.
\end{enumerate}
\end{rem}
\iffalse
$$J':=\{(p,q)\in (I_n\setminus I)^{\times 2}: \epsilon_p\pm\epsilon_q\in \mathring{\Psi}\}, \ \
J'':=\{(p,q)\in (I_n\backslash I)^{\times 2}:|\{\epsilon_p\pm \epsilon_q\}\cap\mathring{\Psi}|=1\}$$
Note that $(p,q)\in J'\ (\text{resp. }J'')\implies (q,p)\in J'\ (\text{resp. }J'')$. Let $\mathbb{D}':=\{p\in I_n\backslash I:\exists\ q\in I_n\backslash I\text{ such that } (p,q)\in J'\}$. \Cref{typeAD1} ensures that we have a disjoint decomposition $\mathbb{D}'=\sqcup_{\gamma\in \mathbb{D}}J_\gamma$ for some index set $\mathbb{D}$ such that $$\{\pm(\epsilon_p\pm \epsilon_q):(p,q)\in J'\}=\sqcup_{\gamma\in \mathbb{D}}D_{J_\gamma}$$ This takes care of the second type of subroot systems in the decomposition.
Let $\mathbb{A}':=\{p\in I_n\backslash I:(p,q)\in J''\}$. Let $\Psi_{\mathbb{A}'}:=\{\pm(\epsilon_p\pm \epsilon_q):(p,q)\in J''\}\cap \mathring{\Psi}$. Let $\Psi_{\mathbb{A}'}=\sqcup_{\gamma\in \mathbb{A}}(\Psi_{\mathbb{A}'})_\gamma$ be the decomposition of $\Psi_{\mathbb{A}'}$ into irreducible components for some index set $\mathbb{A}$. \Cref{typeAD2} ensures that each $(\Psi_{\mathbb{A}'})_\gamma$ is a root system of type $A$. We shall show that there is a disjoint decomposition $\mathbb{A}'=\sqcup_{\gamma\in\mathbb{A}}J_\gamma$ such that for each $\gamma\in \mathbb{A}$ there exists a function $f:(J_\gamma)_{min}\to \{\pm 1\}$ so that $(\Psi_{\mathbb{A}'})_\gamma=A(J_\gamma,f_\gamma)$.
Fix $\gamma\in \mathbb{A}$. Let $\Delta_\gamma=\{\alpha_1,\alpha_2,\dots,\alpha_r\}$ be a basis of $(\Psi_{\mathbb{A}'})_\gamma$. We can choose $\alpha_1=\epsilon_{p_0}+ \epsilon_{p_1}$ or $\epsilon_{p_0}-\epsilon_{p_1}$ for some $(p_0,p_1)\in J''$; if $\alpha_1=-\epsilon_{r_1}\pm \epsilon_{r_2}$ for some $(r_1,r_2)\in J''$, then $s_{\alpha_1}(\Delta_\gamma)$ is another basis of $(\Psi_{\mathbb{A}'})_\gamma$ with the desired property. Since $(\alpha_i,\alpha_{i+1})=-1$ for $1\le i\le r-1$, it follows that there exists a $J_{\gamma}$-minimum function $f_\gamma$ such that $\alpha_i:=-f_\gamma({p_{i-1}})\epsilon_{p_{i-1}} +f_\gamma({p_{i}})\epsilon_{p_{i}}, \ \ \forall 1\le i\le r.$ We fix the total order on $\{p_i:0\le i\le r\}$ so that $p_i\le p_j$ if and only if $i\le j$. Now it follows that $(\Psi_{\mathbb{A}'})_\gamma=A(J_\gamma,f_\gamma)$.
Let $\mathring{\Psi}$ be of the given form. Choose $p_{\epsilon_i}\in 2\bz$ for $i\in J_e$ and $p_{\epsilon_j}\in 2\bz+1$ for $j\in J_o$ arbitrarily. Define $p:B_{J_e}\cup B_{J_o}\to \bz$ extending $\bz$--linearly. Let $$\Psi:=\{\alpha+(p_\alpha+2\bz)\delta:\alpha\in B_{J_e}\cup B_{J_o} \}\cup\{\beta+2\bz\delta:\beta\in \mathring{\Psi}\backslash (B_{J_e}\cup B_{J_o})\}$$
It is easy to see that $\Psi$ is a symmetric closed subset with $Gr(\Psi)=\mathring{\Psi}.$ This completes the proof.
\fi
\begin{prop}\label{propdn+1}
Let $\Psi$ be a symmetric closed subset of $\Phi$ such that $Gr(\Psi)$ is semi-closed. If \eqref{decomppsi} and \eqref{decompgrpsi} are the decomposition of $\Psi$ and $Gr(\Psi)$ respectively into irreducible components, then for each $i,$ there exists a $\bz$-linear function $p_i:Gr(\Psi_i)\to \bz,\ \alpha\mapsto p_i(\alpha)\in Z_\alpha,$ and
$n_i\in 2\bz_+$ such that $$\Psi_i=\{\alpha+(p_i(\alpha)+n_i\bz)\delta:\alpha\in Gr(\Psi_i)\}.$$
In particular, $\Psi$ is a closed subroot system of $\Phi.$ Moreover, we have $r\ge 2$, and $Gr(\Psi_1)$, $Gr(\Psi_2)$ are of type $B$, and the rest of $Gr(\Psi_i)$'s are of either type $A$ or $D$, and $p_1(\alpha)\in 2\bz+1\,(\text{resp.}\ p_2(\alpha)\in 2\bz)$ for a short root $\alpha\in Gr(\Psi_1)\,(\text{resp.}\ Gr(\Psi_2)).$
\iffalse
\begin{enumerate}
\item there exists a $\bz$-linear function $p:Gr(\Psi)\to \bz$ such that $p_\alpha\in Z_\alpha$ for all $\alpha\in Gr(\Psi)$.\label{first}
\item There is a pairwise disjoint subsets $\{J_\gamma\}_{\gamma\in \Gamma}$ of $I_n$ where $\Gamma=\mathbb{A}\cup\mathbb{D}\cup \{o,e\}$ is an index set with $J_e\neq \emptyset,J_o\neq \emptyset$.
\item There are non-negative integers $\{n_\gamma\}$ for $\gamma\in \Gamma$ and functions $f_\gamma:(J_\gamma)_{\text{min}}\to \{\pm 1\}$ for $\gamma\in \mathbb{A}$.\label{third}
\end{enumerate}
such that
\begin{align*}
\Psi=&\bigcup\limits_{i\in\{o,e\}}\{\alpha+(p_\alpha+2n_i\bz)\delta: \alpha\in B_{J_i}\}\cup \bigcup\limits_{\gamma\in \mathbb{D}}\{\alpha+(p_\alpha+2n_\gamma\bz)\delta: \alpha\in D_{J_\gamma}\}\cup\\ & \bigcup\limits_{\gamma\in \mathbb{A}}\{\alpha+(p_\alpha+2n_\gamma\bz)\delta: \alpha\in A(J_\gamma,f_\gamma)\}
\end{align*}
Furthermore if $J_\gamma\neq \emptyset$ for $\gamma\neq o,e$, then $|J_\gamma|>2$ and $p_{\epsilon_i}\in 2\bz\ \forall \ i\in J_e,\ p_{\epsilon_i}\in 2\bz+1\ \forall \ i\in J_o$. In particular, $\Psi$ is a closed subroot system.
Conversely, suppose that $\mathring{\Psi}$ is a semi-closed subset of $\mathring{\Phi}$. Then by lemma \Cref{lem:4} we have $$\mathring{\Psi}=B_{J_e}\cup B_{J_o}\cup\ \bigcup_{\gamma\in \mathbb{A}}A(J_\gamma,f_\gamma)\cup \bigcup_{\gamma\in \mathbb{D}}D_{J_\gamma}$$. Let $p_{\epsilon_i}\in 2\bz$(resp. $\in 2\bz+1)$ for $i\in J_e$(resp. $i\in J_o)$. Let $p:B_{J_e}\cup B_{J_o}\to \bz$ be its unique $\bz$-linear extension. Let $p:\mathring{\Psi}\to \bz$ be any arbitrary $\bz$-linear extension of it satisfying \Cref{first}. Let $\{n_\gamma\}_{\gamma\in \Gamma}$ and $\{f_\gamma\}_{\gamma\in \mathbb{A}}$ be as in \Cref{third}. Then the subset $\Psi$ defined by \begin{align*}
\Psi=&\bigcup\limits_{i\in\{o,e\}}\{\alpha+(p_\alpha+2n_i\bz)\delta: \alpha\in B_{J_i}\}\cup \bigcup\limits_{\gamma\in \mathbb{D}}\{\alpha+(p_\alpha+2n_\gamma\bz)\delta: \alpha\in D_{J_\gamma}\}\cup\\ & \bigcup\limits_{\gamma\in \mathbb{A}}\{\alpha+(p_\alpha+2n_\gamma\bz)\delta: \alpha\in A(J_\gamma,f_\gamma)\}
\end{align*} is a symmetric closed subset with gradient $\mathring{\Psi}$.
\fi
\end{prop}
\begin{proof}
Since $Gr(\Psi)$ is semi-closed, by \Cref{semiclosedD}, we have $Gr(\Psi)= \mathring{\Psi}_1\sqcup \mathring{\Psi}_2\sqcup \cdots \sqcup \mathring{\Psi}_r$ where
$\mathring{\Psi}_1$ and $\mathring{\Psi}_2$ are of type $B.$ For $1\le i\le r,$ let $\Psi_i:=\{\alpha+r\delta\in \Psi\mid \alpha\in \mathring{\Psi}_i\}.$ Then $\Psi=\Psi_1\sqcup \cdots \sqcup \Psi_r$ is the decomposition of $\Psi$ into irreducible components. It is easy to see that each $\Psi_i$ is a symmetric closed subset of $\Phi$ with $Gr(\Psi_i)$ is closed in $\mathring{\Phi}$. From the proof of \cref{semiclosedD}, we have $|\pi_m(Z_\alpha)|=1$ for all $\alpha\in Gr(\Psi).$ Now \cref{prop:equal1} gives the desired result for the first part of the proof. Again the second part is clear from \cref{semiclosedD}.
\end{proof}
\subsection{Twisted real affine root system of type \texorpdfstring{$A_{2n-1}^{(2)}$}{A{2n-1}(2)}} In this case, the real affine roots are given by
$$\Phi=\{\pm 2\epsilon_i+2r\delta\,\mid r\in\bz,\,1\le i\le n\}\cup \{\pm \epsilon_i\pm\epsilon_j+r\delta,\mid r\in\bz,\,1\le i\neq j\le n\}.$$
The next proposition is the main result of this subsection.
\iffalse \begin{lem}\label{lem:16}
Let $\mathring{\Phi}$ be a finite root system of type $C_n$. Let $\{\alpha_1,\alpha_2,\dots,\alpha_n\}$ be the standard base of $\mathring{\Phi}$. Suppose that there is a $\mathbb{Z}$--linear function $\tau:\mathring{\Phi}\to \mathbb{Z}$. Then $\tau(\epsilon_i+\epsilon_j)\equiv \tau(\epsilon_i-\epsilon_j)(\ \mathrm{mod}\ 2)$ for all $i\neq j$ if and only if $\tau(\alpha_n)\equiv 0\ (\mathrm{mod}\ 2)$. \qed
\end{lem}\fi
\iffalse
In \cite{RV19}, $\Psi_p(A_{2n-1}^{(2)})$ are defined and is shown that they are the maximal closed subroot systems of $\Phi$. We shall investigate the symmetric closed subsets with semi-closed gradient and show that maximal closed subroot systems with semi-closed gradient are of the form $\Psi_p(A_{2n-1}^{(2)})$. Let $\Psi$ be a closed symmetric subset of $\Psi$ with semi-closed gradient. Let $I_\Psi=\{i\in I_n: 2\epsilon_i\in Gr(\Psi)\}$. Since $\Psi_p(A_{2n-1}^{(2)})$ is a maximal closed subroot system with semi-closed gradient it follows that if $\Psi$ is a closed subroot system, then $I_\Psi=\emptyset$.
\fi
\begin{prop}\label{a22n-1}
Let $\Psi$ be a symmetric closed subset of $\Phi$ such that $Gr(\Psi)$ is semi-closed in $\mathring{\Phi}$.
Let $\Psi=\Psi_1\sqcup \cdots \sqcup \Psi_r$ be the decomposition of $\Psi$ into irreducible components.
Assume that $Gr(\Psi_i)$ is not of type $A_1$ or $B_2$ for each $1\le i\le r.$
Then we have $\Psi$ is a closed subroot system.
\iffalse
\begin{enumerate}
\item $\Upsilon:=Gr(\Omega)$ is a root system of type $C_n$ and there are two $\mathbb{Z}$-linear functions $p^e:\Upsilon\to \mathbb{Z}$ and $p^o:\Upsilon\to \mathbb{Z}$ with $p^e(\epsilon_i\pm\epsilon_j)\in 2\mathbb{Z}$ and $p^o(\epsilon_i\pm \epsilon_j)\in 2\mathbb{Z}+1$ for all $i\neq j$, a non-negative integer $k$ such that
\begin{equation*}
\begin{split}
\Omega=&\{\alpha+\lceil(p^e(\alpha)+2k\mathbb{Z})\cup(p^o(\alpha)+2k\mathbb{Z})\rceil \delta:\alpha\in(\Psi_i^{C_n})_s \}\\ &\cup \{\alpha+\lceil p^e(\alpha)+2k\mathbb{Z}\rceil \delta:\alpha\in (\Psi_i^{C_n})_\ell\}
\end{split}
\end{equation*}
\item $\Upsilon=Gr(\Omega)$ is of type $D_{n+1}$ and there is a $\mathbb{Z}$--linear function $p:\omega\to \mathbb{Z}$ satisfying $p_{\epsilon_i+\epsilon_j}\not\equiv p_{\epsilon_i-\epsilon_j}\ (\mathrm{mod}\ 2)$ and a non-negative integer $k$ such that $$\Omega=\{\alpha+(p_\alpha+2k\mathbb{Z})\}.$$ In this case $\Omega$ is a closed subroot system.
\end{enumerate}
\fi
\end{prop}
\begin{proof}
Define $I:=\{i\in I_n\mid 2\epsilon_i\in Gr(\Psi)\}$. Since $Gr(\Psi)$ is semi-closed, there exist $\epsilon_s-\epsilon_t,\epsilon_s+\epsilon_t\in Gr(\Psi)$ but $2\epsilon_s\notin Gr(\Psi)$. Hence $I\neq I_n$. Let
$$\Psi':=\left(\bigcup_{i\in I} (\pm2\epsilon_i+2\bz\delta) \cup \bigcup_{i\neq j \in I} (\pm\epsilon_i\pm\epsilon_j+\bz\delta) \right)\cap \Psi,\ \ \Psi'':=\Psi\backslash \Psi'.$$
If $\epsilon_k\pm \epsilon_\ell\in Gr(\Psi)$ for some $k\in I$, but $\ell\in I_n\backslash I$. Then we have
$2\epsilon_k+2p\delta, \epsilon_k\pm\epsilon_\ell+r\delta \in \Psi$ for some $p, r\in \mathbb{Z}.$ This implies that
$$\epsilon_k\mp \epsilon_\ell+(2p-r)\delta=\,(2\epsilon_k+2p\delta)-(\epsilon_k\pm\epsilon_\ell+r\delta)\in\Psi,$$ and hence $\mp2\epsilon_\ell+2(p-r)\delta=(\epsilon_k\mp \epsilon_\ell+(2p-r))-(\epsilon_k\pm\epsilon_\ell+r\delta)\in\Psi$, which is a contradiction as $\ell\notin I$. So if $\epsilon_i\pm \epsilon_j\in Gr(\Psi)$, then either both $i,j\in I$ or both $i,j\notin I.$ In particular, we have
$$\Psi''= \left( \bigcup_{i\neq j \in I_n\backslash I} (\pm\epsilon_i\pm\epsilon_j+\bz\delta)\right)\cap \Psi$$ is a closed subroot system of $\Phi$,
and $\Psi=\Psi'\sqcup \Psi''$ is an orthogonal decomposition.
\medskip
It is easy to see that $\Psi'$ is a symmetric closed subset of $\Phi$ with $Gr(\Psi')$ is a symmetric closed subset of $\mathring{\Phi}.$
Then by \cref{prop:notb2}, $\Psi'$ is also a closed subroot systems of $\Phi$, and hence $\Psi$ is a closed subroot system of $\Phi$.
\iffalse In particular, there is a $\bz$-linear function $p:Gr(\Psi)\to \bz$ and non-negative integers $n_i, 1\le i\le r$, such that $\Psi_i=\{\alpha+(p_\alpha+n_i\bz)\delta:\alpha\in Gr(\Psi_i)\}$ for $1\le i\le r.$
If $\epsilon_i\pm \epsilon_j\in Gr(\Psi)$ such that $2\epsilon_i\notin Gr(\Psi),$ then $\epsilon_i\pm \epsilon_j\in Gr(\Psi'').$ In this case, we can not have $p_{\epsilon_i+\epsilon_j}\equiv p_{\epsilon_i-\epsilon_j}\ (\mathrm{mod}\ 2),$ otherwise we will have $2\epsilon_i+(p_{\epsilon_i+\epsilon_j}+p_{\epsilon_i-\epsilon_j})\delta=(\epsilon_i+\epsilon_j+p_{\epsilon_i+\epsilon_j}\delta)+(\epsilon_i-\epsilon_j+p_{\epsilon_i-\epsilon_j}\delta)\in \Psi$ as $\Psi$ is closed, a contradiction. \fi
\end{proof}
\begin{rem}
Let $\Psi_+=(2\epsilon_1+4\bz\delta)\cup (2\epsilon_2+4\bz\delta)\cup (\epsilon_1\pm\epsilon_2+ (4\bz\cup \pm 1+4\bz)\delta)\cup \{\epsilon_i+\epsilon_j+ (1+4\bz)\delta,\epsilon_i-\epsilon_j+ 4\bz\delta: 3\le i\neq j\le 5\}$
and set $\Psi=\Psi_+\cup -\Psi_+$. Then $\Psi$ is not a closed subroot system. So the condition that any irreducible component of $Gr(\Psi)$ is not of type $A_1$ or $B_2$ cannot be dropped from \cref{a22n-1}.
\end{rem}
\subsection{Twisted real affine root system of type \texorpdfstring{$D_{4}^{(3)}$}{D{4}(3)}}
In this case, the real roots are given by
$$\Phi=\{\pm \,(\epsilon_i-\epsilon_j)+r\delta,\,\pm\,(2\epsilon_i-\epsilon_j-\epsilon_k)+3r\delta\,\mid i,j,k\in I_3, i\neq j,\,r\in\bz\}$$
Note that $\mathring{\Phi}$ is of type $G_2$ and both $\mathring{\Phi}_s$ and $\mathring{\Phi}_\ell$ are of type $A_2.$ The next lemma is a generalization of \cite[Theorem 2]{DV21} and \cite[Lemma 7.1.1]{RV19}.
\iffalse We need the following easy lemma.
\begin{lem}\label{containsshort}
Let $\Psi$ be a closed symmetric subset of $\Phi$ such that $Gr(\Psi)$ is semi-closed. Then $\mathring{\Phi}_s\subseteq Gr(\Psi)$.
\end{lem}
\begin{proof}
Since $Gr(\Psi)$ is semi-closed, there exist short roots $\alpha,\beta\in Gr(\Psi)$ such that the long root $\alpha+\beta\notin Gr(\Psi)$. Now since $\alpha+\beta$ is long in type $G_2$, $\alpha-\beta$ is short and hence $\alpha-\beta\in Gr(\Psi)$. Since $\alpha-\beta\not=\alpha$ or $\beta$, since $Gr(\Psi)$ is symmetric we have $\mathring{\Phi}_s\subseteq Gr(\Psi)$.
\end{proof}\fi
\begin{lem}\label{g2short}
Let $\Psi$ be a symmetric closed subset of $\Phi$ with $Gr(\Psi)$ semi-closed. Then $Gr(\Psi)$ is $\mathring{\Phi}_s$.
\end{lem}
\begin{proof}
Since $Gr(\Psi)$ is semi-closed, there exist short roots $\alpha,\beta\in Gr(\Psi)$ such that $\alpha+\beta\in \mathring{\Phi}_\ell$ and $\alpha+\beta\notin Gr(\Psi)$.
Now since $\alpha+\beta$ is long in type $G_2$, $\alpha-\beta$ must be short and hence $\alpha-\beta\in Gr(\Psi)$. Since $\alpha-\beta\not=\alpha, \beta$ and $Gr(\Psi)$ is symmetric, we have $\mathring{\Phi}_s\subseteq Gr(\Psi)$.
Note that $Gr(\Psi)$ can not contain two long positive roots as $Gr(\Psi)\subsetneq \mathring{\Phi}$. We claim that $Gr(\Psi)$ can not contain any long root. On the contrary assume that $Gr(\Psi)$ contains a long root, say $2\epsilon_i-\epsilon_j-\epsilon_k\in Gr(\Psi)$. Let $S=\pi_m(Z_{\epsilon_j-\epsilon_k}).$ It is clear that $|S|\neq 3.$ Suppose that $|S|=2$. Let $\epsilon_i-\epsilon_j+p_1\delta,\epsilon_j-\epsilon_k+p_2\delta,\epsilon_i-\epsilon_k+p_3\delta\in \Psi$. Since $Gr(\Psi)=\mathring{\Phi}_s\cup \{\pm (2\epsilon_i-\epsilon_j-\epsilon_k)\}$ we have
\begin{equation}\label{rel1}
p_1\not\equiv p_2\ (\mathrm{mod}\ 3), \ \ p_3\not\equiv -p_2\ (\mathrm{mod}\ 3)
\end{equation}
Let $\ell$ be the element which is not in $S$. Then $p_1\equiv \ell(\mathrm{mod}\ 3)$ and $p_3\equiv -\ell(\mathrm{mod}\ 3)$ which gives $|\pi_m(Z_{\epsilon_i-\epsilon_k})|=1,$ but this is impossible since we have $Z_{\epsilon_i-\epsilon_j}+Z_{\epsilon_j-\epsilon_k}\subseteq Z_{\epsilon_i-\epsilon_k}.$
\medskip
Now assume that $|S|=1$ and let $S=\{\ell'\}$. In this case, we have one more relation along with \eqref{rel1} namely
\begin{equation}\label{rel2}
p_3-p_1\equiv \ell'\ (\mathrm{mod}\ 3)
\end{equation}
which implies that $|\pi_m(Z_{\epsilon_i-\epsilon_j})|=1$ and $|\pi_m(Z_{\epsilon_i-\epsilon_k})|=1.$ Suppose that $\pi_m(Z_{\epsilon_i-\epsilon_j})=\{r\}$. Then we have $\pi_m(Z_{\epsilon_i-\epsilon_k})=\{r+\ell'\}$. Since $\Psi$ is closed, for any $2\epsilon_i-\epsilon_j-\epsilon_k+3a\delta\in\Psi,$ we have $$\epsilon_i-\epsilon_k+(3a-p_1)\delta\,=\,(2\epsilon_i-\epsilon_j-\epsilon_k+3a\delta)+(\epsilon_j-\epsilon_i-p_1\delta)\in\Psi.$$ This implies $3a-p_1\equiv r+\ell'\ (\mathrm{mod}\ 3),$ which in turn implies that $\ell'\equiv -2r\equiv r\ (\mathrm{mod}\ 3).$ This contradicts \cref{rel1}. So $|S|\neq 1$ and hence the result.
\end{proof}
\begin{prop}\label{prop: D43}
Let $\Psi$ be a closed symmetric subset of $\Phi$ with $Gr(\Psi)$ semi-closed in $\mathring{\Phi}$. Then $\Psi$ is of the form
$$\Psi=\pm (\epsilon_i-\epsilon_j+r\bz\delta)\cup\,\pm(\epsilon_j-\epsilon_k+(r\bz+\ell)\delta)\cup\,\pm (\epsilon_i-\epsilon_k+(r\bz+\ell)\delta)$$ where $r\in 3\bz_+$ and $\{i,j,k\}$ is a permutation of $\{1,2,3\}$. In particular, $\Psi$ is a closed subroot system of $\Phi$.
\end{prop}
\begin{proof}
By \cref{g2short}, it follows that $Gr(\Psi)=\mathring{\Phi}_s$. Note that we can consider $\Psi$ as an irreducible symmetric closed subset of the untwisted root system of type $A_2^{(1)}$. By \cref{coruntwisted}, there is a $\bz$-linear function $p:Gr(\Psi)\to \bz, \alpha\mapsto p_\alpha\in Z_\alpha$ and $r\in \bz_+$ such that $$\Psi=\{\alpha+(p_\alpha+r\bz)\delta:\alpha\in Gr(\Psi)\}$$ In particular, $\Psi$ is a closed subroot system of $\Phi.$ Next we determine $r$ and $p_\alpha$ explicitly.
At first, we claim that $0\in \pi_m(Z_\alpha)$ for some $\alpha\in Gr(\Psi)\cap \mathring{\Phi}^+$. Suppose that $\epsilon_1-\epsilon_2+r_1\delta,\,\epsilon_2-\epsilon_3+r_2\delta,\,\epsilon_1-\epsilon_3+r_3\delta\in \Psi$. We then have the following relations:
\begin{equation}\label{d34}
r_1\not\equiv r_2\ (\mathrm{mod}\ 3),\ \ -r_3\not\equiv r_2\ (\mathrm{mod}\ 3),\ \ -r_3\not\equiv r_1\ (\mathrm{mod}\ 3).
\end{equation}
This implies that the integers $r_1,r_2,-r_3\ (\mathrm{mod}\ 3)$ are all distinct, so one of them must be $0\ (\mathrm{mod}\ 3)$.
With out loss of generality we assume that $r_1\equiv 0\ (\mathrm{mod}\ 3).$ By similar argument we get
$r_1',r_2,-r_3\ (\mathrm{mod}\ 3)$ are all distinct for any $r'\in Z_{\epsilon_1-\epsilon_2}$ and using \ref{d34} we get
$r_1'\equiv 0\ (\mathrm{mod}\ 3)$ and $r_2\equiv r_3\ (\mathrm{mod}\ 3).$ This implies
$\Psi$ is of the form
$$\Psi=\pm (\epsilon_i-\epsilon_j+r\bz\delta)\cup\,\pm(\epsilon_j-\epsilon_k+(r\bz+\ell)\delta)\cup\,\pm (\epsilon_i-\epsilon_k+(r\bz+\ell)\delta)$$ where $r\in3\bz_+$ and $\{i,j,k\}$ is a permutation of $\{1,2,3\}$.
\end{proof}
\subsection{Twisted real affine root system of type \texorpdfstring{$E_6^{(2)}$}{E{6}(2)}}
In this case, the real roots
are given by $$\Phi=\big\{\pm\epsilon_i+r\delta, \pm\epsilon_i\pm\epsilon_j+2s\delta : 1\le i\neq j\le 4, r, s\in \bz\big\}\cup\{\sum_{i=1}^{4} \pm\epsilon_i/2+r\delta : r\in \bz\}.$$
The gradient root system is of type $F_4,$ and the short roots $\mathring{\Phi}_s$ form a root system of type $D_4$.
For convenience, we call a short root of the form $\frac{1}{2}\left(\sum_{i=1}^4\pm \epsilon_{i}\right)$ as a \textit{special short root}.
For $I\neq\emptyset$ and a special short root $\mu=\frac{1}{2}\,(\sum_{i=1}^4\nu_{i}\epsilon_{i})$ define
$D_\mu^+(\Psi,I):=\{\epsilon_i\mid i\in I\}\cup\{\sum_{i=1}^4 \lambda_i\epsilon_i/2\,\mid\, \lambda_i\in \{\pm 1\}, i\in I, \lambda_j=\nu_j\text{ for } j\notin I\}$
and set $$D_\mu(\Psi,I)=D_\mu^+(\Psi,I)\cup -D_\mu^+(\Psi,I).$$
Let $\Psi$ be a symmetric closed subset of $\Phi$ such that Gr$(\Psi)$ is semi-closed. If $Gr(\Psi)$ does not contain any special short root, then $\Psi$ can be realized as a symmetric, closed subset of the real affine root system of type $D_5^{(2)}$ and hence it is a closed subroot system (see \cref{propdn+1}). So without loss of generality, we can assume that there is a special short root in $Gr(\Psi)$, say $\alpha=\frac{1}{2}\,(\sum_{i=1}^4\nu_{i}\epsilon_{i})\in Gr(\Psi).$
Let $I=\{i\in I_4\mid \epsilon_i\in Gr(\Psi)\}.$ Since $Gr(\Psi)$ is semi-closed, there exist short roots $\beta,\gamma\in Gr(\Psi)$ such that $\beta+\gamma$ is a long root and $\beta+\gamma\notin Gr(\Psi).$ We claim that we can always choose $\beta,\gamma$ to be special short roots. For $|I|\le 1,$ it is obvious. Suppose $|I|\geq 2,$ and there are
$\mu_k\epsilon_{k}, \mu_r\epsilon_{r}\in Gr(\Psi)$ with $k\neq r$, but $\mu_k\epsilon_{k}+\mu_r\epsilon_{r}\notin Gr(\Psi)$. Then since $\alpha=\frac{1}{2}\,(\sum_{i=1}^4\nu_i\epsilon_i)\in Gr(\Psi)$ and $Gr(\Psi)$ is semi-closed, we have that $\beta=\frac{1}{2}\left(\mu_k\epsilon_k+\mu_r\epsilon_r+\sum\limits_{i\notin \{k, r\}}\nu_i\epsilon_i\right)$ and
$\gamma=\frac{1}{2}\left(\mu_k\epsilon_k+\mu_r\epsilon_r+\sum\limits_{i\notin \{k, r\}}-\nu_i\epsilon_i\right)$ are in $Gr(\Psi).$ Clearly
$\beta+\gamma=\mu_k\epsilon_k+\mu_r\epsilon_r.$
\iffalse
Note that if $\epsilon_j,\,\frac{1}{2}\,(\sum_{i=1}^4\lambda_i\epsilon_i)\in Gr(\Psi)$, since Gr$(\Psi)$ is semi-closed, we also have
\begin{equation}\label{changesign}
\frac{1}{2}\,(\sum_{i=1}^4\mu_i\lambda_i\epsilon_i)\in Gr(\Psi)\ \ \text{where } \mu_i=\begin{cases} 1 & \text{if } i\neq j \\ -1 & \text{if } i=j \end{cases}
\end{equation}
i.e. we can change sign before $\epsilon_j$ arbitrarily. This observation will be used more than once and \cref{changesign} will be referred for that.
Let $\mathcal{D}_4=\mathring{\Phi}_s$. Note that there is a twisted affine root subsystem of type $D_5^{(2)}$ contained in $E_6^{(2)}$. We simply call it root system of type $D_5^{(2)}$.
If $\Psi\subseteq D_5^{(2)}$, then it is a closed subroot system by \textcolor{blue}{Case $D_{n+1}^{(2)}$} case. Hence it's enough to classify $\Psi\not\subseteq D_5^{(2)}$. Using the next lemma, (\textcolor{blue}{ Or straightaway}) we assume that Gr$(\Psi)$ contains a root $\alpha=\frac{1}{2}\,(\sum_{i=1}^4\nu_{i}\epsilon_{i})$. For convenience we call a short root of this form as a \textit{special short root}. Note that if $\epsilon_j,\,\frac{1}{2}\,(\sum_{i=1}^4\lambda_i\epsilon_i)\in Gr(\Psi)$, since Gr$(\Psi)$ is semi-closed, we also have $$\frac{1}{2}\,(\sum_{i=1}^4\mu_i\lambda_i\epsilon_i)\in Gr(\Psi)\ \ \text{where } \mu_i=\left\{\begin{array}{cl} 1, & \text{if } i\neq j \\ -1, & \text{if } i=j \end{array}\right.$$ i.e. we can change the sign arbitrarily before $\epsilon_j$. This observation will be used several times in what follows $(*)$.
\begin{lem}
Let $\Psi$ be a symmetric closed subset of $\Phi$ with $Gr(\Psi)$ is semi-closed. Then $Gr(\Psi)$ contains a short root of the form $\frac{1}{2}\sum\limits_{i=1}^4\lambda_i\epsilon_i$.
\end{lem}
\begin{proof}
Suppose that there is no short root in $Gr(\Psi)$ of the required form. Since $Gr(\Psi)$ is semi-closed, there are short roots $\epsilon_i,\epsilon_j\in$ $Gr(\Psi)$ such that $\epsilon_i+\epsilon_j\notin$ $Gr(\Psi)$. Note that $\epsilon_i+\epsilon_j\notin$ Gr$(\Psi)$ if and only if $\epsilon_i-\epsilon_j\notin$ $Gr(\Psi)$. Hence we must have that each pair of elements from $Z_{\epsilon_i}$ and $Z_{\epsilon_j}$ have different parity. In other words $(Z_{\epsilon_i}+Z_{\epsilon_j})\cap 2\Z=\emptyset$. Hence we have $Z_{\epsilon_i}\subseteq 2\Z \text{ or } 2\Z+1$ and similarly for $Z_{\epsilon_j}$.
Since $Gr(\Psi)$ has no component of type $A_1$, there is a long root of the form $\epsilon_i\pm \epsilon_k\,(k\neq j)$ in $Gr(\Psi)$. We have $\mp\epsilon_k=\epsilon_i-(\epsilon_i\pm \epsilon_k)\in Gr(\Psi)$. Since $Z_{\epsilon_i}\subseteq 2\Z \text{ or } 2\Z+1$, it follows that $Z_{\epsilon_k}$ will also have the same property and elements of $Z_{\epsilon_k}(\Psi)$ and $Z_{\epsilon_i}(\Psi) $ have same parities i.e. $(Z_{\epsilon_k}(\Psi)+ Z_{\epsilon_i}(\Psi))\cap (2\Z+1)=\emptyset$. As $\epsilon_i +\epsilon_j \notin Gr(\Psi)$, we have $(Z_{\epsilon_k}(\Psi)+Z_{\epsilon_j}(\Psi))\cap 2\Z=\emptyset$. Again $Gr(\Psi)$ has no irreducible component of the form $\{\pm \epsilon_j\},$ there is a long root of the form $\epsilon_j\pm\epsilon_s$ with $s\neq k$ ($s=k$ will imply $\epsilon_i\pm\epsilon_j\in Gr(\Psi)$). Consequently $\epsilon_s\in $ $Gr(\Psi)$ and $(Z_{\epsilon_j}\cap Z_{\epsilon_s})\cap (2\Z+1)=\emptyset$. Now if $Gr(\Psi)=\{\pm\epsilon_i,\pm\epsilon_k,\pm\epsilon_i\pm\epsilon_k\}\cup\{\pm\epsilon_j,\pm\epsilon_s,\pm\epsilon_j\pm\epsilon_s\}$, $Gr(\Psi)$ is of type $B{_2}\times B_{2}$ which is not irreducible. So $Gr(\Psi)\neq \{\pm\epsilon_i,\pm\epsilon_k,\pm\epsilon_i\pm\epsilon_k\}\cup\{\pm\epsilon_j,\pm\epsilon_s,\pm\epsilon_j\pm\epsilon_s\}$. As $Gr(\Psi)$ is irreducible, there exists a short root in $\{\pm\epsilon_i,\pm\epsilon_k,\pm\epsilon_i\pm\epsilon_k\}$ and $\{\pm\epsilon_j,\pm\epsilon_s,\pm\epsilon_j\pm\epsilon_s\}$ or a pair of sets of this form inside $Gr(\Psi)$ such that their sum is in $Gr(\Psi)$. Hence some other long root is in Gr$(\Psi)$ but this will contradict one of the relations
\begin{center}
$(Z_{\epsilon_i}(\Psi)+Z_{\epsilon_j}(\Psi))\cap 2\Z=\emptyset,\,(Z_{\epsilon_k}(\Psi)+ Z_{\epsilon_i}(\Psi))\cap (2\Z+1)=\emptyset,\,(Z_{\epsilon_j}(\Psi)+ Z_{\epsilon_s}(\Psi))\cap (2\Z+1)=\emptyset$
\end{center}
For example if $\epsilon_k+\epsilon_s \in Gr(\Psi)$, then as sum of long roots is long $\epsilon_i+\epsilon_s \in Gr(\Psi)$, which in turns imply $\epsilon_i+\epsilon_j\in Gr(\Psi)$, which gives us a contradiction.
\end{proof}
\fi
\iffalse
If $I=\emptyset$, suppose that $\text{pr}_i:\mathbb{R}^4\to \mathbb{R}$ are the projection maps onto the $i^{th}$ coordinate. For a special short root $\gamma\in Gr(\Psi)$ and for two subsets $J,J'$ of $I_4$ with $|J|=|J'|=2,\ |J\cap J'|=1$, define
\begin{align*}
E(J,\gamma):&=\left\{\beta=\frac{1}{2}\sum_{i=1}^4 \lambda_i\epsilon_i\in Gr(\Psi):\prod_{j\in J}\text{pr}_j(\beta)\text{pr}_j(\gamma)=\prod_{j\notin J}\text{pr}_j(\beta)\text{pr}_j(\gamma)\right\}\\
E(J,J',\gamma):&=E(J,\gamma)\cup \left\{\mu\pm\nu:\mu,\nu\in E(J,\gamma),\ \prod_{j\in J'}\text{pr}_j(\mu)\text{pr}_j(\gamma)=\prod_{j\in J'}\text{pr}_j(\nu)\text{pr}_j(\gamma)\right\}\\
\text{Write }E(J,J',\gamma)&=E(J,J',\gamma)_{+}\cup E(J,J',\gamma)_{-}\ \text{ where }\\
E(J,J',\gamma)_{\pm}:&=\left\{\mu,\nu,\mu\pm\nu:\mu,\nu\in E(J,\gamma), \prod_{j\in J'}\text{pr}_j(\mu)\text{pr}_j(\gamma)=\pm 4\right\}
\end{align*}
In next lemma we shall show that $E(J,J',\gamma)$ is a direct product of root systems of type $B_2$ and $E(J,J',\gamma)_{\pm 1}$ are the irreducible components. $E(J,J',\gamma)$ occurs naturally in the main Proposition of this section.
\begin{lem}\label{b2b2}
$E(J,J',\gamma)$ is a root system of type $B_2\times B_2$.
\end{lem}
\begin{proof}
Let $\gamma_1\in E(J,J',\gamma)_{+}, \gamma_2\in E(J,J',\gamma)_{-}$. We claim that $(\gamma_1,\gamma_2)=0$. To prove the claim it is sufficient to show that the result holds for special short roots $\gamma_1,\gamma_2.$ Let $\gamma=\frac{1}{2}\sum\limits_{i=1}^4 r_i\epsilon_i,\ \gamma_1=\frac{1}{2}\sum\limits_{i=1}^4 s_i\epsilon_i,\ \gamma_2=\frac{1}{2}\sum\limits_{i=1}^4 t_i\epsilon_i$. Since $\prod\limits_{j\in J'}\text{pr}_j(\gamma_1)\text{pr}_j(\gamma)=+4,$ it follows that $\{s_j:j\in J'\}=\{r_j:j\in J'\}$ or $\{-r_j:j\in J'\}$. Replacing $\gamma_1$ to $-\gamma_1$ if necessary, we assume that $\{s_j:j\in J'\}=\{r_j:j\in J'\}$ i.e. $\pi_{j}(\gamma_1)=\text{pr}_j(\gamma)\ \forall j\in J'$. With this condition, the condition $\prod_{j\in J}\text{pr}_j(\gamma_1)\text{pr}_j(\gamma)=\prod_{j\notin J}\text{pr}_j(\gamma_1)\text{pr}_j(\gamma)$ implies
\begin{equation*}\label{eq1}
\prod_{j\in J\setminus J'}\text{pr}_j(\gamma_1)\text{pr}_j(\gamma)=\prod_{j\in J^c\setminus J'}\text{pr}_j(\gamma_1)\text{pr}_j(\gamma)
\end{equation*}
Similarly for $\gamma_2$ we get the following
$$\sum\limits_{j\in J'}\text{pr}_j(\gamma_2)\text{pr}_j(\gamma)=0,\ \ \ \ \prod_{j\in J\setminus J'}\text{pr}_j(\gamma_2)\text{pr}_j(\gamma)=-\prod_{j\in J^c\setminus J'}\text{pr}_j(\gamma_2)\text{pr}_j(\gamma)$$
Using these two equations we get that $$(\gamma_1,\gamma_2)=\sum\limits_{i\in I_4}\text{pr}_i(\gamma_1)\text{pr}_i(\gamma_2)=\sum\limits_{i\in (J\setminus J')\cup (J^c\setminus J')}\text{pr}_i(\gamma_1)\text{pr}_i(\gamma_2)=0$$ Hence the claim.
We shall determine all the roots in $E(J,J',\gamma)$. Note that $\gamma\in E(J,J',\gamma)$. We assume the following form of $\gamma$ to determine all the roots very explicitly. Let $\gamma=\frac{1}{2}\,(\lambda_{s_1}\epsilon_{s_1}+\lambda_{s_2}\epsilon_{s_2}+\lambda_{s_3}\epsilon_{s_3}+\lambda_{s_4}\epsilon_{s_4})$. From the definition it is clear that $E(J,\gamma)=\{\pm \alpha_i:1\le i\le 4\}$ where $\alpha_1=\gamma$ and \begin{align*}
\alpha_2=\frac{1}{2}\,(\lambda_{s_1}\epsilon_{s_1}+\lambda_{s_2}&\epsilon_{s_2}-\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4}),\ \alpha_3=\frac{1}{2}\ (\lambda_{s_1}\epsilon_{s_1}-\lambda_{s_2}\epsilon_{s_2}+\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4})\\ \alpha_4=& \frac{1}{2}\ (-\lambda_{s_1}\epsilon_{s_1}+\lambda_{s_2}\epsilon_{s_2}+\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4}).
\end{align*}
Let $J=\{s_1,s_2\} \text{ and }J'=\{s_1,s_3\}$. Then $$E(J,J',\gamma)_{+}=\{\pm \alpha_1,\pm \alpha_3,\pm(\alpha_1\pm \alpha_3)\}\text{ and }E(J,J',\gamma)_{-}=\{\pm \alpha_2,\pm \alpha_4,\pm(\alpha_2\pm \alpha_4)\}$$ both are root systems of type $B_2$. Other cases are similar. Hence the lemma.
\end{proof}\fi
\begin{prop}\label{prop:10}
Let $\Psi$ be a symmetric, closed subset of $\Phi$ such that no irreducible component of Gr$(\Psi)$ is of type $A_1$ and $Gr(\Psi)$ is semi-closed. Set $I=\{i\in I_4:\epsilon_i\in Gr(\Psi)\}$, then we have $|I|\equiv 0\ (\mathrm{mod}\ 2$) and exactly one of the following holds.
\begin{enumerate}
\item If $|I|=2,$ then $\mathring{\Phi}_s\supseteq Gr(\Psi)=D_\beta(\Psi,I)$ where $\beta$ is a special short root in $Gr(\Psi).$ There exists a $\bz$-linear function $p:D_\beta(\Psi,I)\to \bz,\ p_\alpha\in Z_\alpha,\alpha\in D_\beta(\Psi,I)$ and $n\in 2\bz_{+}$ such that
$$\Psi=\{\alpha+(p_\alpha+n\bz)\delta\mid \alpha\in D_\beta(\Psi,I)\}.$$
\item If $|I|=4,$ then there exists a partition $J_1,J_2$ of $I_4$ such that $|J_1|=|J_2|=2$ and a $\bz$--linear function $p:Gr(\Psi)\to \bz$ with $p_{\epsilon_{i}}\in 2\bz$
(resp. $p_{\epsilon_{i}}\in 2\bz+1$) if $i\in J_1$ (resp. $i\in J_2$) such that
$$\Psi:=\bigcup_{\alpha\in \mathring{\Phi}_s}\{\alpha+(p_\alpha+n\bz)\delta\}\cup \bigcup_{J\in \{J_1,J_2\}}\{\pm(\epsilon_i\pm \epsilon_j)+(p_{\pm(\epsilon_i\pm \epsilon_j)}+n\bz)\delta:i\neq j\in J\}.$$
\item If $I=\emptyset,$ there exists a $\bz$-linear function $p:Gr(\Psi)\to \bz$ and $m_0,m_1\in 2\bz_{+}$ such that $\Psi=\Psi_0\cup \Psi_1$ where $\Psi_i=\{\alpha+(p_\alpha+m_i\bz)\delta\mid \alpha\in Gr(\Psi_i)\}.$ Moreover, $Gr(\Psi)$ is a root system of type $B_2\times B_2$ and we have $p_\alpha\in 2\bz\ (resp.\ 2\bz+1)$ if $\alpha\in Gr(\Psi_0)$ (resp. $Gr(\Psi_1)$ and $\alpha$ is short).
\end{enumerate}
\end{prop}
\begin{proof}
Since $Gr(\Psi)$ is semi-closed, there exist special short roots $\beta,\gamma\in Gr(\Psi)$ such that the $\beta+\gamma$ is a long root and $\beta+\gamma\notin Gr(\Psi)$. Let $\beta=\frac{1}{2}\,(\lambda_{s_1}\epsilon_{s_1}+\lambda_{s_2}\epsilon_{s_2}+\lambda_{s_3}\epsilon_{s_3}+\lambda_{s_4}\epsilon_{s_4})$ and $\gamma=\frac{1}{2}\,(\lambda_{s_1}\epsilon_{s_1}+\lambda_{s_2}\epsilon_{s_2}-\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4})$. Note that we have $(Z_\beta+Z_\gamma)\cap 2\bz=\emptyset$, so we have $\beta-\gamma\notin Gr(\Psi)$ and $|\pi_m(Z_\beta)|=1$.
First we assume that $I\neq \emptyset$. We claim that $\epsilon_{s_1}\,($resp. $\epsilon_{s_3})\in Gr(\Psi)$ if and only if $\epsilon_{s_2}\,($resp. $\epsilon_{s_4})\in Gr(\Psi)$.
Suppose $s_1\in I$, then we have $\gamma'=\gamma-\lambda_{s_1}\epsilon_{s_1}\in Gr(\Psi).$ This implies that $\lambda_{s_2}\epsilon_{s_2}=\gamma'+\beta\in Gr(\Psi)$ and the case
$s_3\in I$ is done similarly.
This argument also shows that for any $i\in I$ and a special short root
$\gamma\in Gr(\Psi)$, we can always change the sign of the co-efficient of $\epsilon_i$ in $\gamma$ and that resulted element is again in $Gr(\Psi).$
It is easy to see that we always have $D_\beta(\Psi,I)\subseteq Gr(\Psi)\cap \mathring{\Phi}_s.$
We claim that $Gr(\Psi)\cap \mathring{\Phi}_s=D_\beta(\Psi,I).$ Suppose $|I|=4,$ then it is clear as $D_\beta(\Psi,I)=\mathring{\Phi}_s$. So assume that $|I|=2.$
If possible let $Gr(\Psi)\cap \mathring{\Phi}_s\backslash D_\beta(\Psi,I)\neq \emptyset$, then
there exists $\mu\in Gr(\Psi)\cap \mathring{\Phi}_s\backslash D_\beta(\Psi,I)$ and unique
$j\notin I$ such that the co-efficient of $\epsilon_j$ in both $\mu$ and $\beta$ is different.
But this implies $\beta-\mu=\pm \epsilon_j\in Gr(\Psi)\cap \mathring{\Phi}_s$, which is a contradiction.
Thus we get $$Gr(\Psi)\cap \mathring{\Phi}_s = D_\beta(\Psi,I).$$
Also since sum of a long root and a short root is short, it follows that $\epsilon_s\pm\epsilon_t\in Gr(\Psi)$ only if $\{s,t\}\subseteq I$ or $\{s,t\}\subseteq I_4\backslash I$. Note that every special short root $\mu\in D_\beta(\Psi,I)$ can be written as $\mu=\beta+\sum \pm \epsilon_i$ with $\epsilon_i\in Gr(\Psi)$ such that each partial sum is a root. Since all the roots appearing are short and $|\pi_m(Z_\beta)|=1,$ it follows that $|\pi_m(Z_\mu)|=1$ for each $\mu\in D_\beta(\Psi,I).$
So we have $|\pi_m(Z_\alpha)|=1$ for all $\alpha\in Gr(\Psi).$
\medskip
For $|I|=2,$ we show that no long root can occur in $Gr(\Psi)$. We shall show the case for $I=\{s_1,s_2\}$ and the remaining cases are similar. Since $\lambda_{s_1}\epsilon_{s_1}+\lambda_{s_2}\epsilon_{s_2}\notin Gr(\Psi)$, it follows that $(Z_{\epsilon_{s_1}}+Z_{\epsilon_{s_2}})\cap 2\bz=\emptyset$. Hence $\lambda_{s_1}\epsilon_{s_1}-\lambda_{s_2}\epsilon_{s_2}\notin Gr(\Psi)$ as well. Recall that $\epsilon_s\pm\epsilon_t\in Gr(\Psi)$ only if $\{s,t\}\subseteq I$ or $\{s,t\}\subseteq I_4\backslash I$. So we only can have $\lambda_{s_3}\epsilon_{s_3}\pm \lambda_{s_4}\epsilon_{s_4}\in Gr(\Psi)$.
Since $Gr(\Psi)$ has no component of type $A_1$ and $\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4}$ is orthogonal to $D_\beta(\Psi,I)$ and
$\lambda_{s_3}\epsilon_{s_3}+\lambda_{s_4}\epsilon_{s_4}$, we have
$$\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4}\notin Gr(\Psi).$$
Suppose $\lambda_{s_3}\epsilon_{s_3}+ \lambda_{s_4}\epsilon_{s_4}\in Gr(\Psi)$. We have $(Z_\beta+Z_\gamma)\cap2\bz\not=\emptyset$
since $\beta=\gamma+\lambda_{s_3}\epsilon_{s_3}+\lambda_{s_4}\epsilon_{s_4} \in Gr(\Psi)$, which in turn implies that $\beta+\gamma\in Gr(\Psi)$, a contradiction.
Hence no long root can occur in $Gr(\Psi)$ if $|I|=2.$
In this case we have $Gr(\Psi)=D_\beta(\Psi,I)$ which is irreducible.
When $|I|=4$ we have $D_\beta(\Psi,I)=\mathring{\Phi}_s\subseteq \Psi$. So we must have $Gr(\Psi)$ irreducible in this case.
Now using \cref{prop:1}, we get a $\bz$-linear function $p$ and $n\in 2\mathbb{Z}_+$ such that
$\Psi=\{\alpha+(p_\alpha+n\bz)\delta:\alpha\in Gr(\Psi)\}.$ This completes the proof for $I\neq \emptyset$.
Moreover when $I=I_4$ we have the following restrictions:
$$|\pi_m(Z_{\epsilon_{s_i}})|=1,\ \ \forall i \in I_4,\ \ (Z_{\epsilon_{s_1}}+Z_{\epsilon_{s_2}})\cap 2\bz=\emptyset, \ \ (Z_{\epsilon_{s_3}}+Z_{\epsilon_{s_4}})\cap 2\bz=\emptyset.$$
Set $J_1:=\{s_i\in I_4:Z_{\epsilon_{s_i}}\subseteq 2\bz\}$ and $J_2:=\{s_i\in I_4:Z_{\epsilon_{s_i}}\subseteq 2\bz+1\}.$ From the relations above, we have $|J_1|=|J_2|=2.$ Then there exists a $\bz$--linear function $p:Gr(\Psi)\to \bz$ such that $p_{\epsilon_{s_i}}\in 2\bz$ (resp. in $2\bz+1$) if $s_i\in J_1$ (resp. in $J_2$) and $\Psi$ is of the form
$$\Psi:=\bigcup_{\alpha\in \mathring{\Phi}_s}\{\alpha+(p_\alpha+n\bz)\delta\}\cup \bigcup_{J\in \{J_1,J_2\}}\{\pm(\epsilon_i\pm \epsilon_j)+(p_{\pm(\epsilon_i\pm \epsilon_j)}+n\bz)\delta:i\neq j\in J\}.$$
\medskip
Assume that $I=\emptyset$. It is clear that the only short roots can appear in $Gr(\Psi)$ are $\alpha_1:=\beta,\alpha_2:=\gamma,\alpha_3=\frac{1}{2}\,(\lambda_{s_1}\epsilon_{s_1}-\lambda_{s_2}\epsilon_{s_2}+\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4}),\alpha_4=\frac{1}{2}\,(-\lambda_{s_1}\epsilon_{s_1}+\lambda_{s_2}\epsilon_{s_2}+\lambda_{s_3}\epsilon_{s_3}-\lambda_{s_4}\epsilon_{s_4})$ and their negatives (see \cite[Proposition 8.1.3]{RV19}). We claim that all of these roots occur in $Gr(\Psi)$. As $(\alpha_i,\alpha_j)=0,\ 1\le i\neq j \le 4,$ $Gr(\Psi)\not=\{\pm\alpha_i\mid i=1,2\}.$ Hence there is a long root in $Gr(\Psi)$. Since $\alpha_1\pm\alpha_2\notin Gr(\Psi)$, there is a long root in Gr$(\Psi)$ of the form $\lambda_{s_i}\epsilon_{s_i}\pm \lambda_{s_j}\epsilon_{s_j}$ with $i\in\{1,2\}$ and $j\in\{3,4\}$. We shall prove the claim for one case, other cases are similar. Let $\lambda_{s_1}\epsilon_{s_1}+ \lambda_{s_3}\epsilon_{s_3}\in Gr(\Psi)$. Then $-\alpha_3=\alpha_1-(\lambda_{s_1}\epsilon_{s_1}+ \lambda_{s_3}\epsilon_{s_3})\in Gr(\Psi)$. Furthermore since $Z_{\alpha_1}$ contains elements of same parity, it follows that $(Z_{\alpha_1}+Z_{\alpha_3})\cap (2\bz+1)=\emptyset.$ Hence $\alpha_1+\alpha_3=\lambda_{s_2}\epsilon_{s_2}+ \lambda_{s_4}\epsilon_{s_4}\in Gr(\Psi)$. Since $Gr(\Psi)$ has no irreducible component of type $A_1$ and $\alpha_2$ is orthogonal to $\alpha_1,\alpha_3,\lambda_{s_1}\epsilon_{s_1}+ \lambda_{s_3}\epsilon_{s_3}$, it follows that $\lambda_{s_2}\epsilon_{s_2}- \lambda_{s_4}\epsilon_{s_4}\in Gr(\Psi)$. Consequently $-\alpha_4=\alpha_2-(\lambda_{s_2}\epsilon_{s_2}- \lambda_{s_4}\epsilon_{s_4})\in Gr(\Psi)$. By a similar argument we get that $\lambda_{s_1}\epsilon_{s_1}- \lambda_{s_3}\epsilon_{s_3}\in Gr(\Psi)$.
Hence $|\pi_m(Z_\gamma)|=1$ for all $\gamma\in Gr(\Psi).$ Moreover there exist a partition $J_0\sqcup J_1$ of $I_4$ with $|J_i|=2,\ i=0,1$ such that $Gr(\Psi)=\mathring{\Psi}_0\cup \mathring{\Psi}_1$ where $\mathring{\Psi}_i=\{\pm \alpha_k,\pm \alpha_\ell,\pm(\alpha_k\pm \alpha_\ell):J_i=\{k,\ell\}\}.$ We also have $\alpha_k\pm\alpha_\ell\notin Gr(\Psi)$ if $k\in J_i$ but $\ell\notin J_i.$ We can choose $\mathring{\Psi}_0$ to be the component so that $Z_\gamma\subseteq 2\bz$ for all $\gamma\in \mathring{\Psi}_0.$ In each component, all the hypothesis of \cref{prop:1} is satisfied. Hence $\Psi$ is of the required form
with exactly two of $p_{\alpha_i}$'s being even.
\end{proof}
\section{Twisted real affine root system with non-reduced gradient}
In this section, we shall consider the twisted affine root system of type $A_{2n}^{(2)}.$
The real affine roots are $$\Phi=\left\{\pm\epsilon_i+(r+1/2)\delta,\,\pm 2\epsilon_i+2r\delta,\,\pm \epsilon_i\pm\epsilon_j+r\delta\mid 1\le i\neq j\le n,r\in\bz \right\}.$$
The gradient root system is an irreducible reduced root system of type $BC_n$. We set
$$\mathring{\Phi}_d := \{\pm2\epsilon_i\mid 1\le i\le n\}, \ \mathring{\Phi}_s := \{\pm \epsilon_i\mid 1\le i\le n\}, \ \mathring{\Phi}_\ell := \{\pm\epsilon_i\pm\epsilon_j\mid 1\le i\neq j\le n\}$$ and
$$\widehat{C_n}=\left\{\pm 2\epsilon_i+2r\delta,\,\pm \epsilon_i\pm\epsilon_j+r\delta\mid 1\le i\neq j\le n,r\in\bz \right\}\subseteq \Phi.$$
Note that $\widehat{C_n}$ is a closed subroot system of type $A_{2n-1}^{(2)}$. Let $\Psi\subseteq \widehat{C_n}$, then we have $\Psi$ is a symmetric closed in $\widehat{C_n}$ if and only if
$\Psi$ is a symmetric closed in $\Phi.$ We will fix some notations before proceeding further.
For any $I\subseteq I_n$, set $$B_I=\{\pm \epsilon_i, \pm \epsilon_i\pm\epsilon_j \mid i\neq j\in I\},\, \, C_I=\{\pm 2\epsilon_i, \pm \epsilon_i\pm\epsilon_j \mid i\neq j\in I\}$$ and
$BC_I=\{\pm \epsilon_i, \pm 2\epsilon_i, \pm \epsilon_i\pm\epsilon_j \mid i\neq j\in I\}$.
\subsection{}
Let $\Psi$ be a symmetric closed subset of $\Phi$. The next lemma shows that it is enough to classify irreducible symmetric, closed subsets of $\Phi$ which contains at least one short root.
\begin{lem}\label{irrpsi}
Let $\Psi$ be a symmetric closed subset of $\Phi$ and \cref{decomppsi} be the decomposition of $\Psi$ into irreducible components. Then there exists at most one $\Psi_i, 1\le i\le r$ such that $Gr(\Psi_i)\cap \mathring{\Phi}_s\neq \emptyset.$
Conversely, suppose that $\Psi'$ is an irreducible symmetric, closed subset of $\Phi$ such that $Gr(\Psi')\cap \mathring{\Phi}_s\neq \emptyset$ and $\Psi''$ is a symmetric, closed subset of $\wh{C_n}$ such that $(\Psi',\Psi'')=0.$ Then $\Psi'\cup \Psi''$ is a symmetric, closed subset of $\Phi.$
\end{lem}
\begin{proof}
If possible, assume that there exists $i,j$ with $1\le i\neq j\le r$ such that $\epsilon_k\in Gr(\Psi_i), \epsilon_\ell\in Gr(\Psi_j).$ Then we have $\epsilon_k+q\delta\in \Psi, \epsilon_\ell+s\delta\in \Psi$ for some $q,s\in \bz+1/2.$ Since $\Psi$ is closed, we have $\epsilon_k+\epsilon_\ell+(q+s)\delta\in \Psi.$ But the root $\epsilon_k+\epsilon_\ell+(q+s)\delta$ is non-orthogonal to both $\Psi_i$ and $\Psi_j$, which is impossible. For the converse part, it is easy to see that $\Psi'\cup \Psi''$ is a symmetric subset of $\Phi.$ Now we claim if $\alpha\in \Psi'$ and $\beta\in \Psi''$ then $\alpha+\beta\notin \Phi.$ This clearly implies $\Psi'\cup \Psi''$ is closed in $\Phi$ as both $\Psi'$, $\Psi''$ are closed in $\Phi.$
Set $$I'=\{i\in I_n\mid \epsilon_i\in Gr(\Psi')\}.$$ Since $\Psi$ is closed, we have $\epsilon_i\pm \epsilon_j\in Gr(\Psi')$ for all $i\neq j\in I'$ and $\epsilon_i\pm \epsilon_j\notin Gr(\Psi')$ for all $i\in I',j\in I_n\backslash I'.$ Moreover, since $\Psi'$ is irreducible, we have $B_{I'}\subseteq Gr(\Psi')\subseteq BC_{I'}.$
Also we have that $Gr(\Psi'')\subseteq \mathring{\Phi}_\ell\cup \mathring{\Phi}_d$ and it is orthogonal to $Gr(\Psi').$ So it follows that
$Gr(\Psi'')\subseteq C_{I_n\backslash I'}.$ Hence if $\alpha\in Gr(\Psi'),\beta\in Gr(\Psi''),$ then $\alpha+\beta\notin \mathring{\Phi}$. This gives our claim and completes the proof.
\end{proof}
\subsection{}Now onwards, we assume that $\Psi$ is an irreducible symmetric, closed subset of $\Phi$ such that $Gr(\Psi)\cap \mathring{\Phi}_s\neq \emptyset.$ Let
$I:=\{i\in I_n\mid \epsilon_i\in$ Gr$(\Psi)\}$, and we have $I\neq \emptyset.$
Let $q_i\in Z_{\epsilon_i}$ for each $i\in I$ and set $\tilde{Z}_{\epsilon_i}:=Z_{\epsilon_i}-q_i$. Then we have $\tilde{Z}_{\epsilon_i}\subseteq \bz$ for $i\in I.$
The next two lemmas are important and we give proof for the first one and skip the details of the second one as it is similar to the first one. Recall that $m=2.$
\begin{lem}\label{a2nequal1}
Let $\Psi$ be an irreducible symmetric closed subset of $\Phi$ and $I$ be defined as above. Suppose that $|\pi_m(\tilde{Z}_{\epsilon_i})|=1$ for some $i\in I.$ Then
\begin{enumerate}
\item $|\pi_m(\tilde{Z}_{\epsilon_j})|=1$ for all $j\in I$.
\item $Gr(\Psi)=B_I$ and $|\pi_m(Z_{\epsilon_i\pm \epsilon_j})|=1.$
\end{enumerate}
\end{lem}
\begin{proof}
Let $j\in I$ such that $j\neq i$. Since $\Psi$ is closed, we have $\pm(\epsilon_i\pm \epsilon_j)\in Gr(\Psi)$ and $Z_{\epsilon_i\pm\epsilon_j}+Z_{\mp\epsilon_j}= Z_{\epsilon_i}.$ Hence we have $\tilde{Z}_{\epsilon_i\pm\epsilon_j}+\tilde{Z}_{\mp\epsilon_j}= \tilde{Z}_{\epsilon_i}$, $\tilde{Z}_{\epsilon_i\pm\epsilon_j}\subseteq \tilde{Z}_{\epsilon_i}$ and $\tilde{Z}_{\mp\epsilon_j}\subseteq \tilde{Z}_{\epsilon_i}$. This implies $(1)$ and second part of $(2)$.
Since $\Psi$ is closed, we also get $\epsilon_r\pm \epsilon_s\notin Gr(\Psi)$ if $r\in I$ and $s\notin I.$ As $Gr(\Psi)$ is irreducible, we must have
$B_I\subseteq Gr(\Psi')\subseteq BC_I$. If possible assume that $2\epsilon_i\in Gr(\Psi)$ for some $i\in I.$ Since $\Psi$ is closed, we have $Z_{2\epsilon_i}+Z_{-\epsilon_i}\subseteq Z_{\epsilon_i}.$
For $s\in \tilde{Z_{i}}$ and $2k\in Z_{2\epsilon_i},$ we have $2k-s-q_i=2k-2q_i-s+q_i\in Z_{\epsilon_i}.$ This implies both $s$ and $2k-2q_i-s\in \tilde{Z}_{\epsilon_j}$
which is a contradiction as they have different parities as $q_i\in \bz+1/2$. Thus we must have $Gr(\Psi)=\{\pm \epsilon_i,\pm(\epsilon_i\pm \epsilon_j):i\neq j\in I\}$.
\end{proof}
\begin{lem}\label{a2nneq1}
Let $\Psi$ be an irreducible symmetric closed subset of $\Phi$ and $I$ be defined as above. Suppose that $|\pi_m(\tilde{Z}_{\epsilon_i})|=2$ for some $i\in I.$ Then
\begin{enumerate}
\item $|\pi_m(\tilde{Z}_{\epsilon_j})|=2$ for all $j\in I$.
\item $Gr(\Psi)=BC_I$ and $|\pi_m(Z_{\epsilon_i\pm \epsilon_j})|=2.$
\end{enumerate}\qed
\end{lem}
\subsection{}
Generalizing \cite[Section 9.1]{RV19}, we now define certain symmetric, closed subsets of $\Phi$. For
$I\subseteq I_n$, $J\subseteq I, \tau\in 2\bz_{+}$, and a $\bz$-linear function
$p:B_I\to \bz+1/2$ such that $p_{\epsilon_i}-\frac{1}{2}\in 2\bz\ (\text{resp.}\in 2\bz+1)$ for $i\in J\ (\text{resp.}\in I\backslash J),$ define
$$\Psi_\tau^+(p,I,J):=\bigcup\limits_{\alpha\in B_I}(\alpha + (p_\alpha +\tau \bz)\delta),$$
and $\Psi_\tau(p,I,J):=\Psi_\tau^+(p,I,J)\cup (-\Psi_\tau^+(p,I,J)).$ We first consider the situation that appears in \cref{a2nequal1}. In this case, we have
\iffalse
$$\Psi_n^+(p,I,J,\frac{1}{2}):=\{\epsilon_i+(n\mathbb{Z}+p_i)\delta,\,(\epsilon_k\pm\epsilon_\ell)+(n\mathbb{Z}+p_{\epsilon_k\pm\epsilon_\ell})\delta\mid i,k\neq \ell\in J\},$$
$$\Psi_n^+(p,I,J,0,1):=\{(\epsilon_k\pm\epsilon_\ell)+(n\mathbb{Z}+p_{\epsilon_k\pm\epsilon_\ell})\delta\mid k\in J,\ell\in I\backslash J\},$$
$$\Psi_n^+(p,I,I\backslash J,\frac{3}{2}):=\{\epsilon_i+(n\mathbb{Z}+p_i)\delta,(\epsilon_k\pm\epsilon_\ell)+(n\mathbb{Z}+p_{\epsilon_k\pm\epsilon_\ell})\delta\mid i,k\neq \ell\in I\backslash J\},$$
$$\Psi_n(p,I,J):=\pm \Psi_n^+(p,I,J,\frac{1}{2}) \cup \pm\Psi_n^+(p,I,J,0,1) \cup \pm \Psi_n^+(p,I,I\setminus J,3/2)$$\fi
\iffalse
\begin{lem}\label{lem:24}
Let $\Psi$ be an irreducible closed symmetric subset of $\Phi$ where some $Z_{\epsilon_i}(\Psi)-\frac{1}{2}$ contains elements of same parity. Then $\Psi$ is a subroot system of $\Phi$ of the form
\begin{center}
$\Psi=\Psi_m(I,J)$ for some $I\subseteq I_n$, $J\subseteq I$, $m\in\Z_{\geq 0}$.
\end{center}
\end{lem}
\begin{proof}
Using \cref{lem:23}\, we get that $\Psi\subseteq \Psi_m(I,J)$ for some $I,J,m$. Note that any two $\epsilon_i$ with $i\in I$ are summable in Gr$(\Psi)$ and each indeterminate root occurs in this way. We define $p:$ $Gr(\Psi) \to \Z$ by defining $p_{\epsilon_i}$ for $i\in I$ and extending linearly. Then \cref{lem:31} gives the required.
\end{proof}
\fi
\begin{prop}\label{a2n21}
Let $\Psi$ be an irreducible symmetric closed subset of $\Phi$ such that $Gr(\Psi)\cap \mathring{\Phi}_s\neq \emptyset.$
Let $I=\{i \in I_n : \epsilon_i\in Gr(\Psi)\}$ and assume that $|\pi_m(\tilde{Z}_{\epsilon_i})|=1$ for some $i\in I.$ Then
there exist $J\subseteq I\subseteq I_n$, $\tau\in 2\bz_{+}$, and a $\bz$-linear function $p:B_I\to \bz+1/2$ satisfying $p_{\epsilon_i}\in 2\bz+\frac{1}{2}\ (\text{resp.}\, p_{\epsilon_i}\in 2\bz+\frac{3}{2})$ for $i\in J\ (\text{resp.}\, i\in I\backslash J),$ such that $$\Psi=\Psi_\tau(p,I,J).$$
In particular, $\Psi$ is a closed subroot system of $\Phi.$
\end{prop}
\begin{proof}
We have that $Gr(\Psi)=B_I$ by \cref{a2nequal1}$(2).$ It is easy to check that $Z_\alpha+Z_\beta\subseteq Z_{\alpha+\beta}$ holds for all $\alpha,\beta,\alpha+\beta\in Gr(\Psi).$ Define $J:=\{j\in I\mid Z_{\epsilon_j}\subseteq 2\bz+\frac{1}{2}\}.$
Define a $\bz$-linear function $p:Gr(\Psi)\to \bz+1/2$ by choosing $p_i\in Z_{\epsilon_i},\ i\in I$ and extending $\bz$-linearly.
Then we have that $p_{\epsilon_i}\in 2\bz+\frac{1}{2}\ (\text{resp.}\, p_{\epsilon_i}\in 2\bz+\frac{3}{2})$ for $i\in J\ (\text{resp.}\, i\in I\backslash J).$ As before, set $Z_\alpha':=Z_\alpha-p_\alpha, \ \alpha\in Gr(\Psi).$ Note that the argument of \cref{prop:1} goes through in this case, so we get that $A=Z_\alpha'=Z_\beta' \ \forall\, \alpha,\beta\in Gr(\Psi)$ and $A$ is a subgroup of $\bz,$ say $A=\tau\bz.$ Since $|\pi_m(Z_{\epsilon_i\pm \epsilon_j})|=1,$ we must have $\tau\in 2\bz_{+}.$ Thus $\Psi=\Psi_\tau(p,I,J).$
\end{proof}
\subsection{} For $k\in \bz_{+}, I\subseteq I_n$ and a $\bz$-linear function $p:BC_I\to \bz+1/2,$ define $\Psi_k(p,I):=\Psi^+_k(p,I)\cup (-\Psi^+_k(p,I))$ where $\Psi^+_k(p,I)$ is given by
$$\Psi^+_k(p,I):=\bigcup\limits_{\alpha\in B_I}(\alpha+(p_\alpha+k\bz)\delta)\cup \bigcup\limits_{i\in I}(2\epsilon_i+(2p_{\epsilon_i}+k(2\bz+1))\delta).$$ We now consider the case that appears in \cref{a2nneq1}, i.e., when some $Z_{\epsilon_i}$ contains elements of different parities. In this case, we have
\begin{prop}\label{a2n22}
Let $\Psi$ be an irreducible symmetric closed subset of $\Phi.$ Let $I=\{i \in I_n : \epsilon_i\in Gr(\Psi)\}$ and assume that $|\pi_m(\tilde{Z_i})|=2$ for some $i\in I.$
Then there is an odd integer $k\in\bz_+$ and a $\bz$-linear function $p:Gr(\Psi)\to \bz+1/2$ such that $\Psi=\Psi_k(p,I).$ In particular, $\Psi$ is a closed subroot system of $\Phi$.
\end{prop}
\begin{proof}
By \cref{a2nequal1}, we have $Gr(\Psi)=BC_I.$ It is easy to check that $Z_\alpha+Z_\beta\subseteq Z_{\alpha+\beta}$ whenever $\alpha\neq\beta\in \{\pm \epsilon_i:i\in I\}.$
We define $p:B_I\to\bz+1/2$ by choosing $p_i\in Z_{\epsilon_i}, \ i\in I$ and extending $\mathbb{Z}$-linearly. As before,
$Z_{\alpha}'=Z_\alpha-p_\alpha$ are equal and subgroups of $\bz$ for all $\alpha\in B_I.$
Since $|\pi_m(Z_{\epsilon_i\pm \epsilon_j})|=2$ holds by \cref{a2nequal1}, so we must have that $k$ is an odd integer. Write $Z_{\epsilon_i}=A\cup B$ where $A=p_{\epsilon_i}+2k\mathbb{Z},\ B=p_{\epsilon_i}+k(1+2\mathbb{Z})$. Since $A+B\subseteq 2\bz$ and $\Psi$ is closed, we have $A+B=k+2p_{\epsilon_i}+2k\mathbb{Z}\subseteq Z_{2\epsilon_i}.$
We see that $Z_{2\epsilon_i}= k+2p_i+2k\mathbb{Z}$, since $\Psi$ is closed and we have $Z_{2\epsilon_i}+Z_{-\epsilon_i}\subseteq Z_{\epsilon_i}.$
This gives us that $\Psi=\Psi_k(p,I).$
\end{proof}
\begin{rem}
As in the reduced case, we define $Gr(\Psi)$ is semi-closed if $\alpha,\beta\in Gr(\Psi)$ such that $\alpha+\beta\in \mathring{\Phi}$ but $\alpha+\beta\notin Gr(\Psi)$ implies that $(\alpha,\beta,\alpha+\beta)$ is of type $(s,s,d)$ or $(\ell,\ell,d).$ Note that the gradient of $\Psi_\tau(p,I,J)$ is semi-closed subset of $\mathring{\Phi}$, while the gradient of $\Psi_k(p,I)$ is closed in $\mathring{\Phi}.$
\end{rem}
\begin{rem}
Maximal closed subroot systems in this case are classified in \cite[Section 9]{RV19}. Note that
\begin{enumerate}
\item $\Psi_I(A_{2n}^{(2)})=\Psi_n(p,I,J)$ with $I=I_n,J=I,p_{\epsilon_i}=\frac{1}{2}, i\in I.$
\item $\Psi(p,n_s)=\Psi_n(p,I).$
\item $\widehat{A_J}$ is a union of a symmetric closed subset of $A_{2n-1}^{(2)}$ and $\Psi_n(p,I).$
\end{enumerate}
\end{rem}
\iffalse \section{Closed subsets of an affine root system}
\iffalse
In this section, we will see the connection between real closed subsets of $\Phi$ and the closed subsets of $\Delta.$
It is easy to see that the real part of the closed subset of $\Delta$ must be real closed in $\Phi.$
We prove that any closed subset of $\Delta$ is completely determined its real part.
More precisely, let $\Psi'$ be a closed subset of $\Delta$ and define $$\Psi'_{\text{re}}:=\Psi'\cap\Phi\ \text{and} \ \Psi'_{\text{im}}:=\{r\in\mathbb{Z}:r\delta\in\Psi'\}\cup \{0\}.$$
We are mainly interested in closed subsets $\Psi'$ such that $\Psi'_{\text{re}}\neq \emptyset.$ In that case $\Psi'_{\text{re}}$ is a real closed subset of $\Phi.$
\begin{lem}
Let $\Psi'$ be a symmetric subset of $\Delta.$ $\Psi'$ is a closed subroot system in $\Delta$ if and only if $\Psi'$ is closed in $\Delta.$
\end{lem}
\begin{proof}
The only thing is to prove that each symmetric closed subset is a closed subroot system of $\Delta.$ Let $\alpha,\beta\in\Psi'$ such that $\alpha$ is a real root. Since $\Psi'$ is symmetric we have $s_\alpha(\beta)\in\Psi'$ if $\beta=\pm\alpha.$ Let $\beta\neq\pm\alpha.$ By \cite[Proposition 5.1]{K90} there exist $p,q\in\bz_+$ such that $p-q=\langle\beta,\alpha^\vee\rangle$ and we have $\beta+k\alpha\in\Delta$ for all $-p\le k\le q.$ Since $\Psi'$ is closed in $\Delta,$ $\beta+k\alpha\in \Psi'$ for all such $k.$ Since $-p\le -p+q\le q,$ we have $s_\alpha(\beta)\in \Psi'.$
\end{proof}
\begin{cor}
Let $\Psi'$ be a symmetric closed subset of $\Delta.$ If $\Psi'_{\text{re}}\neq \emptyset,$ $\Psi'_{\text{re}}$ is a real closed subroot system of $\Phi.$
\end{cor}
\begin{cor}
Let $\Psi'$ be a subset $\Delta$ such that $\Psi'_{\text{re}}\neq \emptyset.$ Then $\Psi'$ is symmetric closed in $\Delta$ if and only if $\Psi'$ is a closed subroot system of $\Delta$ if and only if
\begin{enumerate}
\item $\Psi'_\text{re}$ is a real closed subroot system.
\item $Z_\alpha(\Psi'_\text{re})+Z_{-\alpha}(\Psi'_\text{re})=Z_\beta(\Psi'_\text{re})+Z_{-\beta}(\Psi'_\text{re})$ for all short roots $\alpha,\beta\in Gr(\Psi'_\text{re}).$
\item $\mathcal{I}_{\Psi'}=Z_\alpha(\Psi'_\text{re})+Z_{-\alpha}(\Psi'_\text{re})$ for some short root $\alpha.$
\end{enumerate}
\end{cor}
\begin{proof}
\end{proof}\fi
Recall $\Delta(S)$ from \cref{regularsubalgebras}. For a real closed subset $\Psi$ of $\Phi,$ we define $$\Psi^e:=\Delta(\Psi), \ \ S(\Psi):=\{\Psi':\Psi' \text{ is a symmetric closed subset of }\Delta\text{ with } \Psi'_{\text{re}}=\Psi\}.$$
There is no guarantee that $S(\Psi)$ is non-empty even if $\Psi$ is a closed subroot system. In general it can be empty, singleton or even infinite. But for a closed subroot system $\Psi,$ if $S(\Psi)$ is non-empty if and only if $S(\Psi)$
is singleton. Which is our next proposition.
\begin{prop}
Let $\Psi$ be a symmetric, real closed subset of $\Phi.$ Then the following are equivalent:
\begin{enumerate}
\item $\Psi^e\in S(\Psi).$
\item $S(\Psi)\neq \emptyset.$
\item $\Psi$ is a closed subroot system in $\Phi$ satisfying $(3)$ in \cref{realpartdeterm}.
\end{enumerate}
Moreover we have $S(\Psi)=\{\Psi^e\}$ if and only if either $\Phi$ is untwisted or $\Phi$ is twisted and $Gr(\Psi)\cap\mathring{\Phi}_s\neq \emptyset.$
\end{prop}
\begin{proof}
Obviously $(1)\Longrightarrow (2)$ and $(2)\Longrightarrow (3)$ from \cref{realpartdeterm}. To prove $(3)\Longrightarrow (1),$ note that from \cref{realpartdeterm} we have $\Psi^e=\Psi\cup (\Psi_{\text{im}}^e\backslash\{0\})\delta$ is closed in $\Delta$ where $$\Psi_{\text{im}}^e=\begin{cases}
n_s\bz & \text{ if } Gr(\Psi_\text{re})\cap\mathring{\Phi}_s\neq \emptyset\\
n_\ell\bz & \text{ otherwise }
\end{cases}$$
Moreover $S(\Psi)=\{\Psi^e\}$ holds in all cases except for \cref{realpartdeterm}$3(d).$
\end{proof}\fi
\section{Real closed Subsets and Regular subalgebras}\label{regularsect}
In this section, we will study the correspondence between the symmetric closed subsets of $\Phi$ and the regular subalgebras of the affine Lie algebra $\lie g$ generated by them.
Let us denote by $\mathcal{C}_{\text{sym}}(\lie g)$ the set of symmetric closed subsets of $\Phi$ and recall that
$\lie g(\Psi)$ denotes the subalgebra of $\lie g$ generated by $\bigcup_{\alpha\in \Psi}\lie g_\alpha$, for $\Psi\in \mathcal{C}_{\text{sym}}(\lie g)$.
Denote by $\mathcal{R}(\lie g)=\{\lie g(\Psi) : \Psi\in \mathcal{C}_{\text{sym}}(\lie g)\}$ the set of all regular subalgebras of $\lie g$ generated by the symmetric closed subsets of $\Phi,$ and
define a map $$\iota_{\lie g}:\mathcal{C}_{\text{sym}}(\lie g)\to \mathcal{R}(\lie g)$$ by $\iota_{\lie g}(\Psi)=\lie g(\Psi)$.
The following result is well-known in the finite setting (see, for example \cite[Proposition 4.1]{DG20}):
\begin{prop}
Let $(\mathring{\lie g}, \mathring{\lie h})$ be the pair of finite-dimensional simple Lie algebra and its Cartan subalgebra corresponding to the root system $\mathring{\Phi}$. Then
the map $\mathring{\Psi}\mapsto \lie g(\mathring{\Psi})$ from the set of closed subsets of $\mathring{\Phi}$ to
the set of $\mathring{\lie h}$-invariant subalgebras of $\mathring{\lie g}$ is bijective. \qed
\end{prop}
It is easy to see that unlike in the finite case, the map $\iota_{\lie g}$ is not injective for any $\lie g$, even if we restrict it to
symmetric closed subsets of $\Phi.$
\iffalse We consider only real closed subsets of of $\Delta$ which contains at least one real root for the following reason; $S\subseteq \bz\backslash \{0\},$ $H:=\oplus_{k\in S}\lie{g}_{k\delta}$ is a resular subalgebra of $\lie g$ since $[\lie g_{r\delta},\lie g_{s\delta}]=0.$ Obviously not every subset of $(\bz\backslash \{0\})\delta$ is closed in $\Delta.$ Define $f$ by
\begin{center}
$\begin{array}{ccc}
\left\{\begin{array}{c}
\text{ Real closed} \\
\text{subsets of } \Phi
\end{array}\right\}& \stackrel{f}{\longrightarrow} & \left\{\begin{array}{c}
\lie h \text{ invariant subalgebras} \\
\text{ of } \lie g
\end{array}\right\}\\
& & \\
\Psi & \longrightarrow & \lie g(\Psi)
\end{array}$
\end{center}\fi
\begin{example}
Let $\lie{g}$ be any affine Lie algebra not of type $A_{2n}^{(2)}$ and $\alpha$ be a short root in $\mathring{\Phi}.$ Consider two symmetric, real closed subsets $\Psi_1,\Psi_2$ of $\Phi$ defined by $$\Psi_1:=\{\alpha+\delta,-\alpha+\delta,-\alpha-\delta,\alpha-\delta\},\ \ \Psi_2:=\{\alpha+3\delta,\alpha+\delta,-\alpha-3\delta,-\alpha-\delta\}.$$ The following relations show that $\pm 2\delta\in\Delta(\Psi_i),\ i=1,2:$
\begin{equation}\label{getimroots}
\begin{aligned}
\mathbb{C}\alpha^\vee\otimes t^{\pm 2}&=[\lie{g}_{\pm\alpha}\otimes t^{\pm 1},\lie{g}_{\mp \alpha}\otimes t^{\pm 1}]\subseteq \lie{g}(\Psi_1),\\
\mathbb{C}\alpha^\vee\otimes t^{\pm 2}&=[\lie{g}_{\pm\alpha}\otimes t^{\pm 3},\lie{g}_{\mp \alpha}\otimes t^{\mp 1}]\subseteq \lie{g}(\Psi_2).
\end{aligned}
\end{equation}
Since $[\alpha^\vee\otimes t^r,\lie{g}_{\pm\alpha}\otimes t^s]\neq 0$ for any $r,s\in \bz$, it follows that $\lie{g}_{\beta\pm 2\delta}\subseteq \lie{g}(\Psi_i)$ for all $\beta\in \Psi_i,\ i=1,2.$ Again using the commutation relations as in \cref{getimroots} we get $\mathbb{C}\alpha^\vee\otimes t^{\pm 4}\subseteq \lie{g}(\Psi_i), \ i=1,2.$ Proceeding in this way we get that $\lie{g}_{\beta\pm 2\bz\delta}\subseteq \lie{g}(\Psi_i)$ for all $\beta\in \Psi_i, \ i=1,2.$ Hence we have $$\lie{g}(\Psi_1)=\bigoplus_{r\in \bz}\lie{g}_{\pm \alpha+(2r+1)\delta}\oplus \bigoplus_{r\in\bz}\mathbb{C}\alpha^\vee\otimes t^{2r},\ \ \lie{g}(\Psi_2)=\bigoplus_{r\in \bz}\lie{g}_{\pm \alpha+(2r+3)\delta}\oplus \bigoplus_{r\in\bz}\mathbb{C}\alpha^\vee\otimes t^{2r}$$
So $\Psi_1\neq \Psi_2$ but $\iota_{\lie{g}}(\Psi_1)=\iota_{\lie{g}}(\Psi_2).$
\end{example}
\begin{rem}
It is possible to get similar examples using \cref{b2} when $Gr(\Psi)$ is of type $B_2$, e.g. $\Psi_1$ be given by \cref{b2} and $\Psi_2$ be defined by switching the role of $a_1$ and $a_2.$ We will leave to the details, one can prove that $\Psi_1\neq \Psi_2$, but $\iota_{\lie{g}}(\Psi_1)=\iota_{\lie{g}}(\Psi_2).$
\end{rem}
\subsection{} The previous discussion motivates us to look for the best possible subclass of $\mathcal{C}_{\text{sym}}(\lie g)$ for which the map $\iota_{\lie g}$ restricted to that class
is injective. Recall a result from \cite[Corollary 11.1.5]{RV19}, the map $\Psi\mapsto \lie g(\Psi)$ is injective if we restrict to the
set of closed subroot systems of $\Phi$. That means the map $\iota_{\lie g}$ restricted to all closed subroot systems of $\Phi$ is injective.
We indeed prove that this is only the best possible subclass of symmetric closed subsets of $\Phi$ for which we can have injective:
\begin{prop}\label{injectivemaps}
Suppose $\mathcal{S}$ is a subclass of the set of symmetric closed subsets of $\Phi$ such that the restriction of $\iota_{\lie g}$ is injective, then $\mathcal{S}$ must be the set of all closed subroot systems of $\Phi$.
\end{prop}
\begin{proof}
Let $\Psi$ be a symmetric closed subset of $\Phi,$ which is not a subroot system of $\Phi.$ Then by \cref{smallestclosed}, we know that
$\Psi'=\Delta(\Psi)\cap \Phi$ is the minimal closed subroot system of $\Phi$ that contains $\Psi.$ It is clear that $\Psi\neq \Psi'$, but
$\lie{g}(\Psi)=\lie{g}(\Psi'),$ i.e.,
$\iota_{\lie{g}}(\Psi)=\iota_{\lie{g}}(\Psi').$ So the largest subclass $\mathcal{S}$ of symmetric closed subsets of $\Phi$ for which the restriction
$\iota|_{\mathcal{S}}$ is injective is the set of all closed subroot systems of $\Phi$.
\end{proof}
\begin{cor}
Let $\Psi$ be a symmetric closed subset of $\Phi.$ Then $\Delta(\lie g(\Psi))\cap \Phi=\Psi$ if and only if $\Psi$ is a closed subroot system of $\Phi.$
\end{cor}
\subsection{} Fix $\Psi\in \mathcal{C}_{\text{sym}}(\lie g)$, now we will determine the preimage $\iota^{-1}_{\lie g}(\lie g(\Psi))$ using our results, i.e., we will determine all possible
$\Psi'\in \mathcal{C}_{\text{sym}}(\lie g)$ such that $\lie g(\Psi')=\lie g(\Psi).$ Note that $\Delta(\Psi)\cap\Phi$ is the unique real closed subroot system in $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$, so it is enough to determine $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ when $\Psi$ is a real closed subroot system. If $\Psi'\in \iota_{\lie{g}}^{-1}(\lie{g}(\Psi)),$ then from the definition it is clear that $Gr(\Psi)=Gr(\Psi')$ and by \cref{smallestclosed} we have $$\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))=\{\Psi'\in \mathcal{C}_{\text{sym}}(\lie g): \Delta(\Psi')\cap \Phi=\Psi\}.$$ To determine the preimage we need the following example.
\begin{example}\label{rootb2}
Let $\Psi$ be a symmetric closed subset of $\Phi$ such that $Gr(\Psi)$ is of type $B_2.$ Then by \cref{prop:b2} there is a $\bz$-linear function $p:Gr(\Psi)\to \bz,n_\ell\in 2\bz_+$ and odd integers $1\le a_1,a_2\le n_\ell$ with $a_1+a_2\equiv 0(\mathrm{mod}\ n_\ell)$ such that $$\Psi^+=\{\alpha_i+ (p_i+A_i)\delta: i=1, 2\}\cup\{\pm(\alpha_1\pm \alpha_2)+(\pm p_1\pm p_2+n_\ell\bz)\delta:\alpha\in \mathring{\Phi}^+_\ell\}$$ where $A_i=n_\ell\bz\cup (a_i+n_\ell\bz), i=1,2.$ We shall determine $\Delta(\Psi)$ in this case and we shall show that $\Delta(\Psi)\cap \Phi$ is of the form \cref{b2subroot} with $n_s=\mathrm{gcd}(a_1,n)=\mathrm{gcd}(a_2,n).$ Since $a_1$ is odd, we have $2\nmid\ n_s.$ Since $a_1+a_2\equiv 0\ (\mathrm{mod}\ n_\ell)$ we have
\begin{align*}
&Z_{\alpha_1}(\Psi)=n_\ell\bz\cup (a_1+n_\ell\bz), \ \ Z_{-\alpha_1}(\Psi)=n_\ell\bz\cup (a_2+n_\ell\bz)\ \text{ and }\\
&Z_{\alpha_2}(\Psi)=n_\ell\bz\cup (a_2+n_\ell\bz), \ \ Z_{-\alpha_2}(\Psi)=n_\ell\bz\cup (a_1+n_\ell\bz)
\end{align*}
Hence $Z_{\alpha_i}(\Psi)+Z_{-\alpha_i}(\Psi)=n_\ell\bz\cup (a_1+n_\ell\bz)\cup (a_2+n_\ell\bz)$ for $i=1,2.$ We shall show that $Z_{\pm\alpha_i}(\Delta(\Psi))\supseteq\pm p_i+\mathrm{gcd}(a_i,n_\ell)\bz.$ Since for any $r\in \bz,$ $[\alpha_1^\vee,\lie{g}_{\pm\alpha_1+r\delta}]\neq 0$ holds, for any $s\in \bz_+$ we have $p_1+sa_1+n_\ell\bz\subseteq Z_{\alpha_1}(\Delta(\Psi)).$ Also since $a_1+a_2\equiv 0\ (\mathrm{mod}\ n_\ell),$ $a_2+n_\ell\bz$ lies in the set $\sum_{k\in \bz}ka_1+n_\ell\bz.$ Exchanging the role of $a_1$ and $a_2$ in the argument we get $$Z_{\alpha_i}(\Delta(\Psi))\supseteq p_i+\sum_{k\in \bz}(ka_i+n_\ell\bz)=p_i+\mathrm{gcd}(a_1,n_\ell)\bz, \ \ i=1,2.$$
Although $[\alpha_i^\vee,\lie{g}_{\alpha_1\pm\alpha_2+r\delta}]\neq 0, i=1,2,$ we have $p_1\pm p_2+sa_1+n_\ell\bz\subseteq Z_{\alpha_1\pm\alpha_2}(\Delta(\Psi))$ if and only if $s$ is even. Hence it follows that $Z_{\alpha_1\pm\alpha_2}(\Delta(\Psi))\supseteq p_1\pm p_2+2\mathrm{gcd}(a_1,n_\ell)\bz.$ But $$\Psi':=\{\pm \alpha_i+\lceil\pm p_i+\mathrm{gcd}(a_1,n_\ell)\bz\rceil\delta\}\cup \{\pm(\alpha_1\pm \alpha_2)+\lceil\pm p_1\pm p_2+2\mathrm{gcd}(a_1,n_\ell)\bz\rceil\delta\}$$ is a real closed subroot system of $\Phi$ containing $\Psi$ and contained in $\Delta(\Psi)\cap \Phi.$ By \cref{smallestclosed} we have $\Psi'=\Delta(\Psi)\cap \Phi$ and hence the result.
\end{example}
\subsection{}
We first determine the preimage for $\Psi$, which is irreducible and of the form \cref{b2subroot}.
\begin{lem}\label{preimageb2}
Let $\Psi$ be an irreducible closed subroot system of $\Phi$ of the form \cref{b2subroot} and $n_s$ as given in \cref{b2subroot}.
For any given positive integer $r$, there are exactly $\varphi(2r)$ symmetric, real closed subsets $\Psi'$ in $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ such that
$n_\ell(\Psi')=2rn_s$, where $\varphi$ is the Euler's totient function.
In particular, $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ is infinite in this case. For a fixed $n_s$ and $r,$ all $\Psi'\in \iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ with $n_\ell(\Psi')=2rn_s$ are given by \cref{b2} where $a_1$ is a cyclic generator of the group $\langle n_s \rangle$ in $\mathbb{Z}/(2rn_s)\bz.$
\iffalse For any positive even multiple $n_\ell$ of $n_s,$ there are exactly $\varphi(\frac{n_\ell}{n_s})$ symmetric, real closed subsets $\Psi'$ in $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ such that $Z_{\alpha}(\Psi')$ is a union of cosets modulo $n_\ell\bz$ for all $\alpha\in Gr(\Psi'),$ where $\varphi$ is the Euler's totient function. In particular, $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ is infinite in this case.\fi
\end{lem}
\begin{proof}
If $\Psi'\in \iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$, then $\Psi'$ must be of the form \cref{b2}. For a fixed $r,$ $\Psi'\in \iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ such that $n_\ell(\Psi')=2rn_s$ are in one to one correspondence with all $a_1$ such that $ 1\le a_1\le 2rn_s$ and $\mathrm{gcd}(a_1,2rn_s)=n_s$ by \cref{rootb2}. \iffalse for a given odd $n_s\in\bz_+$ and even $n_\ell\in 2\bz$\fi It is elementary to check that $\mathrm{gcd}(a_1,2rn_s)=n_s$ if and only if $a_1$ is a generator of the cyclic group $\langle n_s\rangle$ in $\bz/({2rn_s})\bz.$ Hence the number of such $a_1$ is equal to $\varphi(2rn_s/n_s)=\varphi(2r).$
\end{proof}
We are now ready to prove our main result of this section.
\begin{prop}\label{preimage}
For any irreducible real closed subroot system $\Psi$ of $\Phi,$ we have
$$\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))=\{\Psi\}$$ if either $Gr(\Psi) \text{ is not of type } A_1,B_2$ or $\text{ if } Gr(\Psi)=B_2 \text{ and }
\Psi\text{ is as in } \cref{prop:equal1}.$
And $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ is infinite for all other cases.
\iffalse
$$\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))=\begin{cases}
\Psi & \text{ if } Gr(\Psi) \text{ is not of type } A_1,B_2,C_2\\
\Psi &
\begin{array}{l}
\text{ if } Gr(\Psi)=B_2\text{ or }C_2 \text{ and }\\
\Psi\text{ is as in } \cref{prop:equal1}.
\end{array}
\\
\cref{preimageb2} & \begin{array}{l}
\text{ if }Gr(\Psi)=B_2\text{ or }C_2 \text{ and } \\
\Psi\text{ is of the form } \cref{b2subroot}.
\end{array}
\end{cases}$$
\fi
\end{prop}
\begin{proof}
Suppose that no irreducible component of $Gr(\Psi)$ is of type $A_1$ or $B_2$, then any $\Psi'\in \iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ is a closed subroot system by our previous results and in this case we have $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))=\{\Psi\}.$
Now suppose that $Gr(\Psi)$ is of type $B_2$. From \cref{rootb2}, a symmetric, real closed subset $\Psi'$ of the form \cref{b2} generates a closed subroot system only of the form \cref{b2subroot}. Hence if $Gr(\Psi)$ is of type $B_2$ and $\Psi$ is of the form of \cref{prop:equal1}, then any $\Psi'\in \iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ is a closed subroot system of $\Phi$ and hence in this case $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))=\{\Psi\}.$
Now let $\Psi$ be of the form \cref{b2subroot} and let $n_s$ as in \cref{b2subroot}.
If $\Psi'\in \iota_{\lie{g}}^{-1}(\lie{g}(\Psi)),$ then $\Psi'$ is of the form \cref{b2} where $n_\ell(\Psi')$ is an even multiple of $n_s.$ Remaining follows from \cref{preimageb2}.
\begin{rem}
One of the drastic differences between the finite and affine root system theory is that even maximal symmetric closed subsets are not necessarily closed subroot systems.
For example, let $n_\ell$ be any even integer $\geq 4$ and $a_1$ be any integer which is $\le n_\ell$, $a_1\neq n_\ell/2$ and $\mathrm{gcd}(a_, n_\ell)=1.$ Then define $\Psi$ with this $n_\ell$ and $a_1$ as in \cref{prop:b2}. Then $\Psi$ is a proper symmetric closed subset which is not a closed subroot system and by \cref{rootb2}, $\Delta(\Psi)\cap \Phi=\Phi.$
Suppose $\Psi'\supseteq \Psi$ is a symmetric closed subset that is not a subroot system, then
we have that $Z_{\epsilon_i}(\Psi')$ is a union of two cosets of $n_\ell\bz$. This implies that $\Psi'=\Psi.$ Hence $\Psi$ is a maximal symmetric closed subset of $\Phi$, but not a closed subroot system.
\end{rem}
\iffalse
Let $a_1=a$ and $a_2=-a.$ By \cref{rootb2}, it is enough to find $n_\ell\in 2\bz$ and an odd $a, 1\le a\le n_\ell$ for a given odd $n_s\in\bz_+$ such that $\mathrm{gcd}(a,n_\ell)=n_s.$
Fix an $n_\ell\in\bz_+$ such that $n_s|n_\ell.$ Let an odd $a, 1\le a\le n_\ell$ satisfy $a\bz+n_\ell\bz=n_s\bz.$ Since $n_s\in a\bz+n_\ell\bz=\{ka+n_\ell\bz:k\in\bz\}$ and $1 \le n_s\le n_\ell,$ there exists an integer $k\in \bz_+$ such that $ka=n_s.$ Hence we have $a|n_s$ and so $n_s=a.$ Hence for a fix $\Psi$ $\iota_{\lie{g}}^{-1}(\lie{g}(\Psi))$ is in one one correspondence with $2n_s\bz$ and the correspondence is given by $\Psi'\mapsto 2n_sk:=n_\ell$ {\color{red} write more clearly}.\fi
\end{proof}
\iffalse
\begin{example}\label{oneone}
Let $\Psi_1$ be given by \cref{b2} and $\Psi_2$ be defined by switching the role of $a_1$ and $a_2.$ Since $a_1\neq a_2,$ we have $\Psi_1\neq \Psi_2.$ Note that $Z_{-\epsilon_i}=Z_{\epsilon_j}$ with $1\le i\neq j\le 2$ in both $\Psi_1$ and $\Psi_2.$ We shall show that $(\Psi_1^e)_{\text{im}}=\cup_{k\in\bz}\{ka_1+n_\ell\bz\}.$ Since $[\lie{g}_{\epsilon_i}\otimes t^r,\lie{g}_{-\epsilon_i}\otimes t^s]= \mathbb{C}\epsilon_i^\vee\otimes t^{r+s}\neq 0,$ we always have $n_\ell\bz\cup (a_1+n_\ell\bz)\cup (a_2+n_\ell\bz) \subseteq (\Psi_1^e)_{\text{im}}.$ Again since $[\mathbb{C}\epsilon_i^\vee\otimes t^{r},\lie{g}_{\pm\epsilon_i}\otimes t^s]=\lie{g}_{\pm\epsilon_i}\otimes t^s\neq 0,$ we have
\begin{align*}
n_\ell\bz\cup (a_1+n_\ell\bz)\cup (2a_1+n_\ell\bz)\cup (a_2+n_\ell\bz) &\subseteq Z_{\epsilon_1}(\Psi_1^e)\\
n_\ell\bz\cup (a_1+n_\ell\bz)\cup (a_2+n_\ell\bz)\cup (2a_2+n_\ell\bz) &\subseteq Z_{-\epsilon_1}(\Psi_1^e)
\end{align*}
Again $Z_{\epsilon_1}(\Psi_1^e)+Z_{-\epsilon_1}(\Psi_1^e)\subseteq (\Psi_1^e)_{\text{im}}.$ Proceeding in this way we get that $\cup_{k\in\bz}\{ka_1+n_\ell\bz\}\subseteq Z_{\epsilon_1}(\Psi_1^e).$ Since $(\Psi_1^e)_{\text{im}}+Z_{\epsilon_1}(\Psi_1^e)\subseteq Z_{\epsilon_1}(\Psi_1^e),$ it follows that $(\Psi_1^e)_{\text{im}}=Z_{\epsilon_1}(\Psi_1^e)+Z_{-\epsilon_1}(\Psi_1^e).$ Hence it follows that $\Psi_1^e=(\Psi_1^e)_{\text{re}}\cup (\Psi_1^e)_{\text{re}}$ where $$Z_\alpha=\begin{cases}
\cup_{k\in\bz}\{ka_1+n_\ell\bz\}& \text{ if } \alpha \text{ is short.}\\
\cup_{k\in2\bz}\{ka_1+n_\ell\bz\}& \text{ if } \alpha \text{ is long.}
\end{cases}$$
\end{example}
\begin{defn}
We define a class $\mathcal{S}$ in $\Phi$ to be any one of the set of roots in the following flow chart.
\end{defn}
\begin{tikzpicture}[scale=0.8]
\draw[->] (-1,4.5)--(-1,4);
\draw (-6,4)--(4,4);
\draw[->] (-6,4)--(-6,3.5);
\draw[->] (4,4)--(4,3.5);
\draw[->] (-6,2.5)--(-6,2);
\draw[->] (4,2.5)--(4,2);
\draw (-8,2)--(-4,2);
\draw (6,2)--(2,2);
\draw[->] (-8,2)--(-8,1.5);
\draw[->] (-4,2)--(-4,1.5);
\draw[->] (6,2)--(6,1.5);
\draw[->] (2,2)--(2,1.5);
\draw[->] (6,0.5)-- (6,0);
\draw (4,0)--(8,0);
\draw[->] (4,0)--(4,-0.5);
\draw[->] (8,0)--(8,-0.5);
\node (C) at (-1,5) {Closed};
\node (SC) at (-6,3) {Symmetric closed};
\node (NSC) at (4,3) {Non-symmetric closed};
\node (SRS) at (-8,1) {Subroot system};
\node (NSRS) at (-4,1) {Non-subroot system};
\node (SP) at (2,1) {Non-Special Closed};
\node (SP) at (6,1) {Special Closed};
\node (SGR) at (4,-1) {Special gradient};
\node (SGR) at (8,-1) {Non-special gradient};
\end{tikzpicture}
The restriction of $f$ to the class `Closed subroot systems of $\Phi$' is injective by \cite[Corollary 11.1.5]{RV19}. \cref{oneone} shows that restriction of $f$ to the class of `symmetric closed subsets of $\Phi$' is not injective. Next proposition shows that inside `symmetric closed subsets of $\Phi$', closed subroot systems are the `best' subclass where $f$ is injective.
Recall the definition of $\Delta(\Psi)$ for a symmetric closed subset $\Psi$ of $\Phi$. They are the roots of the regular subalgebra $\lie g(\Psi)$ generated by $\lie g_\alpha$ with $\alpha\in \Psi$. furthermore $\Delta(\Psi)\cap \Phi$ is a real closed subset of $\Phi$ (section 1). We define a map $f$ as follows
$\begin{array}{ccc}
\left\{\begin{array}{c}
\text{ Symmetric real closed} \\
\text{subsets of } \Phi
\end{array}\right\}& \stackrel{f}{\longrightarrow} & \left\{\begin{array}{c}
\text{ Regular subalgebras} \\
\text{ of } \lie g
\end{array}\right\}\\
\Psi & \longrightarrow & \lie g(\Psi)
\end{array}$
\medskip
Similarly we define a map $g$ in the opposite direction
$\begin{array}{ccc}
\left\{\begin{array}{c}
\text{ Regular subalgebras} \\
\text{ of } \lie g
\end{array}\right\}& \stackrel{g}{\longrightarrow} & \left\{\begin{array}{c}
\text{ Symmetric real closed} \\
\text{subsets of } \Phi
\end{array}\right\}\\
H & \longrightarrow & \Delta(H)\cap \Phi
\end{array}$
Note that $g\circ f(\Psi)=\Delta(\Psi)\cap \Phi$ for all symmetric real closed subsets $\Psi$ of $\Phi$. To understand the map for the twisted case well we need the following proposition.
\fi
\iffalse
Now we consider the other extreme situation: let $\Psi$ be a special closed subset of $\Phi$ such that $Gr(\Psi)$ is special. In this case, we prove that the set of roots of $\lie g(\Psi)$ is equal to $\Psi$, i.e., $\Delta(\Psi)=\Psi$.
\begin{lem}\label{specialclosedlem}
Let $\Psi$ be a special closed subset of $\Phi$ such that $Gr(\Psi)$ is special. Then we have $\Delta(\Psi)=\Psi$.
\end{lem}
\begin{proof}
Let $\beta\in \Delta(\Psi)$. Then there exists $\alpha_1, \cdots, \alpha_k\in \Psi$ such that $\beta=\alpha_1+\cdots +\alpha_k,$ and each partial sum
$\alpha_1+\cdots +\alpha_i$ is a root for $1\le i\le k.$ Since $\Psi$ consists of real roots and $Gr(\Psi)$ is special, each $\alpha_1+\cdots +\alpha_i$ must be real root. Since $\Psi$ is closed in $\Phi$, we must have $\alpha_1+\cdots +\alpha_i\in \Psi$ for each $1\le i\le k.$ In particular, we have $\beta\in \Psi$, and hence $\Delta(\Psi)=\Psi.$
\end{proof}
So \cref{specialclosedlem} allows us to include the set of special closed subsets of $\Phi$ whose gradients are special in $\mathcal{S}$.
Consider the following example.
\begin{example}
Let $a,b_1,b_2\in \bz$ such that $b_1\neq b_2.$ Let $\Phi=A_2^{(1)}$ with $\{\alpha,\beta\}$ as a base of $\mathring{\Phi}.$ Define
\begin{align*}
\Psi_1:=&\{\alpha+a\delta,\beta+\lceil b_2+(b_2-b_1)\bz\rceil\delta,\alpha+\beta+\lceil a+b_2+(b_2-b_1)\bz\rceil\delta,-\alpha+(b_1-b_2-a)\delta\}\\
\Psi_2:=&\{\gamma+\lceil Z_\gamma(\Psi_1)+(b_2-b_1)\bz\rceil\delta:\gamma\in Gr(\Psi_1)\}
\end{align*}
Then both $\Psi_1$ and $\Psi_2$ are real closed subset of $\Phi$ such that $$\lie g(\Psi_1)=\lie g(\Psi_2)=\bigoplus_{\gamma\in \Psi_2}\lie g_{\gamma}\bigoplus _{k\in \bz\backslash\{0\}}\lie g_{k(b_2-b_1)\delta}.$$
\end{example}
Hence $f$ is not injective on the class of `non-symmetric closed subset of $\Phi$' and `all special closed subset of $\Phi$' because by the previous example (if $b_1-b_2\neq 2a$) the statement is not true on a smaller class `all non-symmetric closed subset of $\Phi$ with $Gr(\Psi)$ is not special in $\mathring{\Phi}$'. Now if $b_1-b_2=2a,$ the same example shows that $f$ is not one to one on `non-symmetric and non-special closed subset of $\Phi$.' If $\Psi$ is a closed subset of $\Phi$ such that $Gr(\Psi)$ is special closed in $\mathring{\Phi}$ then $\Psi$ is non-symmetric and clearly $\Delta(\lie g(\Psi))=\Psi\subseteq \Phi.$ Hence we have the next lemma.
\begin{lem}
The restriction of $f$ is injective on the class of all `closed subset of $\Phi$ such that $Gr(\Psi)$ is special closed in $\mathring{\Phi}$.'
\end{lem}
Let $\mathcal{B}$ be the class of all closed sets in $\Phi$ consisting of closed subroot systems in $\Phi$ and all closed subsets of $\Phi$ such that $Gr(\Psi)$ is special closed in $\mathring{\Phi}.$ We shall classify the class $\mathcal{B}$ in the next section by classifying all closed subsets of $\Phi$ such that $Gr(\Psi)$ is special closed in $\mathring{\Phi}.$
\begin{cor}
The class $\mathcal{B}$ is the `best' class of closed subsets of $\Phi$ where $f$ is one to one.
\end{cor}
\begin{cor}
Let $\mathcal{S}$ be a class of closed subsets of $\Phi.$ Then the restriction of $f$ to the class $\mathcal{S}$ is one to one if and only if $\Delta(\lie g(\Psi))\cap \Phi=\Psi$ for all $\Psi\in \mathcal{S}.$
\end{cor}
\fi
\section{Summary}\label{summarysect}
Let $\Phi$ be a real affine root system and $\Psi$ be a symmetric closed subset of $\Phi$. In this section we shall summarize all the results in the following table. We assume that \cref{decomppsi} and \cref{decompgrpsi} are the decomposition of $\Psi$ and $Gr(\Psi)$ into irreducible components respectively. We assume that $p:Gr(\Psi)\to \bz$ is a $\bz$-linear function such that $p_\alpha\in Z_\alpha,\ \forall \alpha\in Gr(\Psi).$ We have three different type of forms for $\Psi.$
\begin{align}
\Psi_i&=\{\alpha+(p_\alpha+n\bz)\delta:\alpha\in Gr(\Psi_i)\}\label{typeI}\\
\Psi&=\{\alpha+(p_\alpha+n_s\bz)\delta:\alpha\in Gr(\Psi)_s\}\cup \{\gamma+(p_\gamma+mn_s\bz)\delta:\gamma\in Gr(\Psi)_\ell\}\label{typeII}\\
\Psi_i^+&=\{\epsilon_i+ (p_{\epsilon_i}+A_i)\delta: i=1, 2\}\cup\{\alpha+(p_\alpha+n_\ell\bz)\delta:\alpha\in \mathring{\Phi}^+_\ell\}\text{ as in } \cref{prop:b2}\label{typeIII}
\end{align}
\cmark\, and \xmark\, in remark imply that $\Psi_i$ is a closed subroot system and is not a closed subroot system respectively.
\medskip
\begin{center}
\begin{NiceTabular}{cccc}[hvlines]
$Gr(\Psi_i)$ & Case & Form of $\Psi_i$ & Remark\\
\Block{3-1}{Closed} & $|\pi_m(Z_\alpha)|=1,\ \forall \alpha\in Gr(\Psi)$ & \cref{typeI} & \cmark\\
& \Block{2-1}{\begin{tabular}{c}
$\exists \alpha\in Gr(\Psi)_s$ \\
such that $|\pi_m(Z_\alpha)|>1$
\end{tabular}}& \cref{typeII} & \cmark\\
& & \cref{typeIII} & \xmark \\
\Block{6-1}{Semi-closed} & \Block{2-1}{$D_{n+1}^{(2)}$} & \cref{typeII} & \cmark \\
& & \cref{typeII} & \cmark \\
& \Block{2-1}{$A_{2n-1}^{(2)}$} & \cref{typeIII} & \xmark \\
& & \cref{a22n-1} & \cmark \\
& $D_{4}^{(3)}$ & \cref{typeI} & \cmark \\
& $E_6^{(2)}$ & \cref{typeI} & \cmark\\
\Block{2-1}{} & \Block{2-1}{$A_{2n}^{(2)}$} & \cref{a2n21} & \cmark \\
& & \cref{a2n22} & \cmark \\
\end{NiceTabular}
\end{center}
\bibliographystyle{plain}
|
\section{Introduction}
Some of the most challenging control problems are characterized by nonlinear uncertain dynamics and safety critical constraints.
Optimization-based control techniques, such as model predictive control (MPC)~\cite{rawlings2017model}, can offer high performance control for nonlinear systems, while respecting general state and input constraints.
However, the desired closed-loop properties (stability, performance, constraint satisfaction) are in general not preserved under model uncertainty.
In this paper, we develop a robust adaptive MPC framework that robustly ensures such closed-loop properties for a large class of uncertain nonlinear systems, while incorporating online model updates to reduce conservatism.
This is achieved by introducing a novel \textit{robust} MPC design, including a general method to incorporate \textit{adaptive} model updates.
\subsubsection*{Robust MPC methods for nonlinear uncertain systems}
Robust MPC (RMPC) methods typically ensure constraint satisfaction despite model uncertainty by predicting a tube/funnel over the prediction horizon that contains all uncertain trajectories.
Corresponding nonlinear RMPC approaches can, e.g., be found in~\cite{lucia2015robust,villanueva2017robust,houska2019robust}, however, most formulations result in a drastic increase of the computational complexity that may impede real-time implementation.
In case of linear systems, the trade-offs between computational complexity and conservatism are well-understood~\cite{kouvaritakis2016model} and the tube-based RMPC schemes in~\cite{mayne2005robust,rakovic2012homothetic,parsi2022scalable} are particularly appealing considering computational complexity.
Therein, computational efficiency is ensured by performing more expensive computations related to reachability analysis and invariant sets offline, resulting in a parameterized tube, e.g., in the form of a polytope or ellipsoid, that is simple to predict over the prediction horizon.
Correspondingly, utilizing similar concepts for nonlinear systems requires an offline computation of (parameterized) robust positive invariant sets, which can be done using simple ellipsoidal sets~\cite{yu2013tube,wabersich2021nonlinear}, general incremental Lyapunov functions~\cite{bayer2013discrete,Koehler2020Robust}, and more recently control contraction metrics (CCMs)~\cite{singh2017robust,singh2019robust,zhao2021tube}.
These invariant sets can then be used as a \textit{rigid} tube around a nominal prediction to obtain a simple nonlinear RMPC scheme~\cite{yu2013tube,wabersich2021nonlinear,bayer2013discrete,singh2017robust,singh2019robust,zhao2021tube}, compare also~\cite{mayne2005robust}.
However, such a rigid tube is conservative if the model mismatch depends on the state or input, e.g., due to parametric uncertainty.
Such effects can be better captured using a \textit{homothetic} tube~\cite{lopez2019adaptive,Koehler2020Robust,koehler2021output,kohler2020RAMPC,rakovic2022homothetic}, i.e., the tube scaling depends on the nominal trajectory, compare also~\cite{rakovic2012homothetic,lorenzen2019robust}.
In this paper, we extend and unify these concepts by deriving a homothetic tube MPC framework for nonlinear uncertain systems based on general CCMs.
\subsubsection*{Parameter updates in robust \textit{adaptive} MPC}
Model mismatch is often partially due to inaccurate system knowledge and performance may be enhanced by incorporating online model updates in RMPC frameworks~\cite{hewing2020learning}.
In case of linearly parameterized models with bounded disturbance, we can efficiently update parameter sets using set-membership estimation~\cite{lorenzen2019robust,kohler2020RAMPC} or Lyapunov-type arguments~\cite{Adetola2009adaptive,guay2015robust}.
This reduced uncertainty can be directly utilized in some RMPC methods, resulting in robust adaptive MPC (RAMPC) schemes~\cite{lorenzen2019robust,lu2021robust}.
However, since most (nonlinear) RMPC frameworks explicitly depend on a nominal trajectory with some nominal parameter, most resulting RAMPC schemes (cf. \cite{kohler2019linear,didier2021robust,Adetola2009adaptive,guay2015robust,lopez2019adaptive,kohler2020RAMPC}) require restrictive additional assumptions to ensure recursive feasibility under model updates, such as the restriction of the parameter set to balls/hypercubes instead of general ellipsoids/polytopes.
In this paper, we show how to incorporate general parameter updates in nonlinear RMPC schemes without posing such additional restrictions, resulting in a flexible RAMPC scheme.
\subsubsection*{Contribution}
In this paper, we present a RAMPC framework for nonlinear continuous-time systems that uses a homothetic tube based on general CCMs and naturally incorporates online model updates.
As one of the main technical contributions, we extend existing results on (robust) CCMs (cf. \cite{lohmiller1998contraction,manchester2017control,singh2017robust,wang2020virtual,zhao2021tube}) by deriving a differential equation for the tube scaling, resulting in a homothetic tube that contains all possible nonlinear uncertain trajectories.
In particular, the derived tube dynamics directly use state and input dependent bounds on the model mismatch, thus avoiding the conservatism of constant bounds~\cite{singh2017robust,zhao2021tube}.
To the best knowledge of the authors, this is also the first result on tube MPC schemes that rigorously shows that the CCM conditions are only needed on the constraint set, without imposing additional assumptions.
Compared to existing homothetic tube MPC formulations (cf.~\cite{lopez2019adaptive,Koehler2020Robust,koehler2021output,rakovic2022homothetic}), the reliance on established CCM synthesis tools (cf.~\cite{manchester2017control,singh2017robust,zhao2021tube}) ensures the practical applicability to a large class of uncertain nonlinear systems.
We incorporate online model updates by using set-membership estimation for the parametric uncertainty.
In contrast to existing approaches, we propose to additionally optimize over the nominal parameter that characterizes the nominal trajectory.
This seemingly trivial modification retains the desired flexibility of the model updates without requiring the additional restrictions needed in related work~\cite{kohler2019linear,didier2021robust,Adetola2009adaptive,guay2015robust,lopez2019adaptive,kohler2020RAMPC}.
As a result, the derived RAMPC scheme can also directly track steady-states which explicitly depend on the unknown parameter.
Overall, the resulting homothetic tube RAMPC framework guarantees recursive feasibility, constraint satisfaction, and convergence to a neighborhood of the desired steady-state.
We demonstrate the benefits of the homothetic tube MPC and online model updates using the planar quadrotor example from~\cite{zhao2021tube}.
\subsubsection*{Outline}
Section~\ref{sec:prelim} introduces the problem setup and outlines the general RAMPC approach.
Section~\ref{sec:main} presents the main results, including the novel homothetic tube propagation based on CCMs, the proposed RAMPC scheme, and the theoretical analysis.
Finally, Section~\ref{sec:example} presents a numerical example and Section~\ref{sec:conclusion} provides a conclusion.
\subsubsection*{Notation}
The set of non-negative real numbers is ${\mathbb{R}}_{\geq 0}$, while $\I{a,b}$ is the set of integers in the interval $[a,b]$. The $i$-th element of a vector $x$ is denoted by $[x]_i$. Furthermore, $[X]_{:,k}$ is the $k$-th column of $X$.
The zero matrix and the identity matrix of appropriate dimensions are denoted by $0$ and $I$, respectively.
We use the notation $\langle A \rangle = A + A^\top$ for square matrices $A\in{\mathbb{R}}^{n\times n}$.
The set of positive definite matrices of size $n \times n$ is $\mathbb{S}^+_n$, and $A \preceq B$ indicates that $B-A$ is positive semi-definite.
The quadratic norm of a vector $x$ w.r.t. a positive definite matrix $M$ is denoted by $\|x\|_{M} := \sqrt{x^\top M x}$.
The Cholesky decomposition of a positive definite matrix is denoted by $\left (M^{\frac{1}{2}} \right)^\top M^{\frac{1}{2}} = M$, yielding $\left\|x\right\|_M = \left\|M^{\frac{1}{2}}x \right\|$. Throughout the paper, we indicate the time dependence by $x(t)$ for actual trajectories or $z_{t|t_k}$ for predicted trajectories at time $t_k$.
For a continuously differentiable function $f:{\mathbb{R}}^a \times {\mathbb{R}}^b \to {\mathbb{R}}^c$, $f(x,u)$, the partial and total derivative w.r.t. $x$ evaluated at some point $(x,u) = (z,v)$ is defined as $\left.\dfrac{\partial f}{\partial x}\right|_{(z,v)} \in {\mathbb{R}}^{c \times a}$ and $\left.\dfrac{\mathrm{d}f}{\mathrm{d} x}\right|_{(z,v)}$, respectively. Similarly, we abbreviate the total time derivative by $\dot{f}(x,u):=\left.\dfrac{\mathrm{d}f}{\mathrm{d}t}\right|_{(x,u)}$.
\section{Preliminaries} \label{sec:prelim}
We first present the considered problem setup (Sec.~\ref{sec:setup}).
Then, we outline how the problem is addressed using the proposed homothetic tube RAMPC scheme (Sec.~\ref{sec:problem-statement}).
\subsection{Problem setup} \label{sec:setup}
We consider a continuous-time nonlinear system
\begin{align} \label{eq:system}
\dot{x}(t)=f_\mathrm{w}(x(t),u(t),\theta,d(t)), \quad x(0) = x_0, \quad t \geq 0,
\end{align}
with state $x(t)\in\mathbb{R}^n$, initial state $x_0\in\mathbb{R}^n$, input $u(t)\in\mathbb{R}^m$, disturbance $d(t) \in \mathbb{R}^q$, and constant model parameter $\theta \in {\mathbb{R}}^p$.
The disturbance $d(t)$ and the model parameter $\theta$ are unknown and satisfy a known bound $d(t)\in \mathbb{D}$, $t\geq0$, $\theta \in \Theta_0$ with polytopes $\Theta_0 \subseteq {\mathbb{R}}^p$, $\mathbb{D} \subseteq \mathbb{R}^q$ and $0\in\mathbb{D}$, and the state $x(t)$, $t\geq 0$ can be measured.
Furthermore, we assume that the dynamics $f_{\mathrm{w}}$ are affine in $u$, $d$, $\theta$, i.e., we have
\begin{align*}
f_\mathrm{w}(x,u,\theta,d) = f(x) + B(x)u + G(x,u)\theta+{E(x)}d,\nonumber
\end{align*}
with $G$ affine in $u$, i.e., $f_{\mathrm{w}}$ is affine in $u$, $d$ and $\theta$, respectively.
Throughout the paper, we assume that $f_\mathrm{w}$ is continuously differentiable and Lipschitz continuous
and $u(t)$ and $d(t)$ are piecewise continuous, which implies existence of a unique solution $x(\cdot)$ to the system \eqref{eq:system}, see \cite[Thm. 3.2]{khalil2002nonlinear}.\footnote{Similarly to \eqref{eq:constraint-tight1}--\eqref{eq:c_j-def}, one can show that the feedback $\kappa(x,z,v)$ in Proposition~\ref{prop:V-delta} is Lipschitz continuous w.r.t. $x$. Hence, existence and uniqueness of the solution to \eqref{eq:system} remain valid under the feedback $\kappa$.}
The system is subject to state and input constraints
\begin{equation} \label{eq:constr}
(x(t),u(t))\in{\mathbb{Z}}, \quad t \geq 0,
\end{equation}
where ${\mathbb{Z}}$ is a compact set defined by
\begin{equation} \label{eq:Z_set}
{\mathbb{Z}}:=\left\{(x,u)\in\mathbb{R}^{n+m}|h_j(x,u)\leq 0,~j \in \I{1,r}\right\}.
\end{equation}
We denote the projection of ${\mathbb{Z}}$ on $x$ by ${\mathbb{Z}}_\mathrm{x}$.
The overall goal can be summarized in the following conceptual optimization problem:
\begin{equation} \label{eq:conceptual}
\begin{split}
&\min_{\pi(\cdot)} \; \max_{d(\cdot) \in \mathbb{D},~\theta \in \Theta_0} \mathcal{J}(x(0),\pi(\cdot),d(\cdot),\theta) \\
\textnormal{s.t.} \quad & u(t) = \pi(x(0:t),t), \\
&(x(t),u(t)) \in {\mathbb{Z}},\\
&\dot{x}(t) = f_\mathrm{w}(x(t),u(t),\theta,d(t)), \quad t \geq 0.
\end{split}
\end{equation}
The goal is to find a policy $\pi$, that minimizes some cost $\mathcal{J}$ under the worst-case disturbance and parameter, while robustly satisfying the state and input constraints.
\subsection{General RAMPC approach} \label{sec:problem-statement}
Problem~\eqref{eq:conceptual} is intractable in this form due to the optimization over a general policy $\pi$ that depends on all past data and the difficulty to compute the exact (disturbance) reachable set for nonlinear systems.
In the following, we provide a high-level description how we
compute a feasible solution to Problem~\eqref{eq:conceptual} by using a finitely parametrized policy $\pi$ and an over-approximation of the possible state trajectories $x(t) \in \mathbb{T}_t$, called tube.
As common in tube-based robust MPC formulations, this is achieved by defining a nominal state and input trajectory $z_t$, $v_t$ based on some nominal parameter $\overline{\theta}\in\Theta_0$ with the nominal dynamics $\dot{z}_t = \overline{f}(z_t,v_t,\overline{\theta}) := f_\mathrm{w}(z_t,v_t,\overline{\theta},0)$.
The policy is then given as a feedback law $\kappa:{\mathbb{R}}^n \times {\mathbb{Z}} \rightarrow {\mathbb{R}}^m$ parametrized by the nominal trajectories, i.e., $\pi(x(0:t),t) = \kappa(x(t),z_t,v_t)$. Furthermore, the tube is parametrized as
\begin{align} \label{eq:tube-def}
\mathbb{T}_t = \{x \in {\mathbb{R}}^n|V_\delta(x,z_t) \leq \delta_t \},\quad t\geq 0,
\end{align} where $V_\delta: {\mathbb{R}}^n \times {\mathbb{R}}^n \rightarrow {\mathbb{R}}_{\geq 0}$ is a later specified Lyapunov-type function and $\delta_t \geq 0$ is a suitable scaling.
Given this parametrization, we can provide a feasible solution to Problem~\eqref{eq:conceptual} with
\begin{subequations}
\label{eq:tube_conceptual}
\begin{align}
\label{eq:tube_conceptual_1}
&\dot{z}_t=\overline{f}(z_t,v_t,\overline{\theta}),\quad \dot{\delta}_t=f_\delta(z_t,v_t,\delta_t,\overline{\theta},\Theta_0),\\
\label{eq:tube_conceptual_2}
&(x(t),\kappa(x(t),z_t,v_t))\in\mathbb{Z},\quad \forall x(t)\in\mathbb{T}_t,\quad t\geq 0,
\end{align}
\end{subequations}
where $f_\delta$ corresponds to a later derived dynamics for the tube scaling (cf. Sec.~\ref{sec:tube-design}) ensuring $x(t)\in\mathbb{T}_t$, $t\geq 0$.
Since the parametrization~\eqref{eq:tube-def} contains a (not necessarily linear) translation $z_t$ and a scaling $\delta_t$,
we refer to it as a homothetic tube MPC due to its similarity to existing approaches for linear systems~\cite{lorenzen2019robust,rakovic2012homothetic}.
In particular, in case $V_\delta$ is a polytopic or quadratic function of the error $x-z_t$, we recover the homothetic tube parametrization in~\cite{didier2021robust,kohler2019linear,lorenzen2019robust,rakovic2012homothetic,parsi2022scalable,rakovic2022homothetic}.
As one of the main technical contributions, in Sections~\ref{sec:CCM}--\ref{sec:tube-design} we show how suitable functions $\kappa$, $V_\delta$ can be constructed offline using CCMs and derive dynamics for the tube scaling $\delta_t$.
For simplicity, we assume that the cost $\mathcal{J}$ in~\eqref{eq:conceptual} corresponds to a quadratic tracking stage cost of the form
\begin{equation} \label{eq:stage-cost-def}
\ell(z,v,\overline{\theta}) = \|z-z_\mathrm{ref}(\overline{\theta})\|_Q^2 + \|v - v_\mathrm{ref}(\overline{\theta})\|_R^2,
\end{equation}
with positive definite matrices $Q \in {\mathbb{R}}^{n \times n}$, $R \in {\mathbb{R}}^{m \times m}$ and a feasible steady-state and input $(z_{\mathrm{ref}}(\overline{\theta}),v_{\mathrm{ref}}(\overline{\theta}))\in {\mathbb{Z}}$, $\overline{f}(z_{\mathrm{ref}}(\overline{\theta}),v_{\mathrm{ref}}(\overline{\theta}),\overline{\theta})=0$, $\overline{\theta}\in\Theta_0$, which may depend (continuously) on the nominal parameter $\overline{\theta}$ (cf. also Sec.~\ref{sec:discussion}).
To approximate the infinite-horizon problem~\eqref{eq:conceptual}, we repeatedly solve a finite-horizon problem at every sampling time $t_k = kT_\mathrm{s}$, $k \in \mathbb{I}_{\geq0}$ with sampling period $T_\mathrm{s} > 0$, using the standard receding horizon principle from MPC~\cite{rawlings2017model}.
We incorporate model adaptation by updating the set $\Theta_0$ online at every sampling time $t_k$ using set-membership estimation.
\section{Robust adaptive MPC framework} \label{sec:main}
This section derives the proposed RAMPC framework, which is the main result of the paper. First, we use CCMs to construct the homothetic tube (Sec.~\ref{sec:CCM}--\ref{sec:tube-design}). Next, we present how the initial parameter set $\Theta_0$ can be updated using set-membership estimation (Sec.~\ref{sec:set-update}), followed by the design of the terminal ingredients (Sec.~\ref{sec:terminal}). Then, the overall algorithm, including the MPC problem, is described and the theoretical properties are derived (Sec.~\ref{sec:framework}). Finally, we provide some discussion, including a qualitative comparison to existing approaches (Sec.~\ref{sec:discussion}).
\subsection{Control contraction metrics} \label{sec:CCM}
In the following, we show that CCMs provide a natural way to derive the ingredients needed in the homothetic tube design ($V_\delta,\kappa,f_\delta$). In particular, CCMs use the Jacobian of the dynamics $f_{\mathrm{w}}$ to synthesize controllers ensuring (incremental) stability for the nonlinear dynamics~\cite{manchester2017control,singh2017robust,singh2019robust,wang2020virtual,zhao2021tube}.
\begin{assum} \label{ass:CCM}
There exists a smooth function $M: {\mathbb{R}}^n \to \mathbb{S}^+_n$, a continuous function $K:{\mathbb{R}}^n \to {\mathbb{R}}^{m \times n}$, a contraction rate $\rho_\mathrm{c} > 0$, and matrices $\underline{M},~\overline{M} \in \mathbb{S}^+_n$, such that for all $(x,u) \in {\mathbb{Z}}$, and all $\theta \in \Theta_0$, $d \in \mathbb{D}$, the following properties hold:
\begin{subequations}
\label{eq:CCM}
\begin{align}
\dot{M}(x)+ \langle M(x) A_\mathrm{cl}(x,u,\theta,d) \rangle \preceq & -2\rho_\mathrm{c} M(x), \label{eq:CCM-contraction} \\
\underline{M} \preceq M(x) \preceq & \overline{M}, \label{eq:CCM-bounded-M}
\end{align}
\end{subequations}
with $\dot{x} = f_\mathrm{w}(x,u,\theta,d)$, and
\begin{equation*}
A_\mathrm{cl}(x,u,\theta,d) := \left.\dfrac{\partial f_\mathrm{w}}{\partial x}\right|_{(x,u,\theta,d)} + \left.\dfrac{\partial f_\mathrm{w}}{\partial u}\right|_{(x,u,\theta,d)}K(x).
\end{equation*}
\end{assum}
Conditions \eqref{eq:CCM-contraction}--\eqref{eq:CCM-bounded-M} imply stability of the linearized dynamics with the differential feedback $K$ using the contraction metric $M$. By integrating this differential form we can obtain stability properties for the nonlinear system. To this end, we denote the set of piece-wise smooth curves $\gamma : [0,1] \rightarrow {\mathbb{R}}^n$ with $\gamma(0) = z$ and $\gamma(1) = x$ by $\Gamma(z,x)$, and define $\gamma_{\mathrm{s}}(s) := \left.\dfrac{\partial \gamma}{\partial s}\right|_s$. Given a CCM satisfying Assumption~\ref{ass:CCM}, we define a corresponding incremental Lyapunov function $V_\delta$ (cf.~\cite{manchester2017control}) as follows:
\begin{equation} \label{eq:V-delta-def}
V_\delta(x,z) := \min_{\gamma \in \Gamma(z,x)} \int_{0}^{1} \|\gamma_{\mathrm{s}}(s)\|_{M(\gamma(s))} \mathrm{d}s.
\end{equation}
A minimizer is called a geodesic $\gamma^\star$ (with $\gamma_{\mathrm{s}}^\star$), which we assume to exist (cf.~\cite{manchester2017control} for corresponding conditions). Integration of the local feedback $K(x)$ along the geodesic $\gamma^\star$ yields the input $\gamma^{\mathrm{u}}$, and the feedback law $\kappa$:
\begin{align} \label{eq:kappa-def}
\begin{split}
\gamma^{\mathrm{u}}(s) :=& v + \int_0^{s} K(\gamma^\star(\tilde{s}))\gamma_{\mathrm{s}}^\star(\tilde{s}) \mathrm{d}\tilde{s}, \\
\kappa(x,z,v) :=& \gamma^{\mathrm{u}}(1).
\end{split}
\end{align}
The following proposition summarizes the properties of $V_\delta$ and $\kappa$.
\begin{prop} \label{prop:V-delta}
Suppose Assumption~\ref{ass:CCM} holds. Consider some $(x,z,v)\in {\mathbb{R}}^n \times {\mathbb{R}}^n \times {\mathbb{R}}^m$ such that for all $s\in[0,1]$ it holds that $(\gamma^\star(s),\gamma^{\mathrm{u}}(s))\in{\mathbb{Z}}$. Then, for all $\theta, \overline{\theta} \in \Theta_0$, and all $d \in \mathbb{D}$, the following inequalities hold:
\begin{subequations}
\begin{align}
\|x-z\|_{\underline{M}} \leq& V_\delta(x,z) \leq \|x-z\|_{\overline{M}}, \label{eq:Vd-bound} \\
\dot{V_\delta}(x,z) =& \left.\dfrac{\partial V_\delta}{\partial x}\right|_{(x,z)} \dot{x}
+ \left.\dfrac{\partial V_\delta}{\partial z}\right|_{(x,z)} \dot{z} \label{eq:Vd-contraction}\\
\leq& - \left(\rho_\mathrm{c} - L_\mathbb{D} - \sum_{k=1}^{p}L_{\mathrm{G},k}\left|[\theta - \overline{\theta}]_k\right| \right) V_\delta(x,z) \nonumber\\
&+ \|G(z,v) (\theta-\overline{\theta}) + E(z)d\|_{M(z)}, \nonumber
\end{align}
with $\dot{x} = f_\mathrm{w}(x,\kappa(x,z,v),\theta,d)$, $\dot{z} = \overline{f}(z,v,\overline{\theta})$, $V_\delta$, $\kappa$ according to
Equations~\eqref{eq:V-delta-def}, \eqref{eq:kappa-def}, and constants $L_{\mathrm{G},k} \geq 0$, $k \in \I{1,p}$, $L_\mathbb{D} \geq 0$ according to \eqref{eq:app-L-G} and \eqref{eq:app-L-D} in the appendix.
\end{subequations}
\end{prop}
The proof of this proposition is given in Appendix~\ref{app:proof-V-delta}. In the nominal setting, i.e., $\mathbb{D}=\{0\}$, $\theta=\overline{\theta}$, Proposition~\ref{prop:V-delta} ensures that $V_\delta$ is an incremental Lyapunov function for $\dot{x}=\overline{f}(x,u,\overline{\theta})$ with feedback $\kappa(x,z,v)$, analogous to~\cite{manchester2017control}.
The effect of the disturbance $d$ and the parametric uncertainty $\theta-\overline{\theta}$ is suitably bounded with the constants $L_\mathbb{D}$, $L_{\mathrm{G},k}$ modifying the contraction rate $\rho_\mathrm{c}$, and the model mismatch $\|G(z,v) (\theta-\overline{\theta}) + E(z)d\|_{M(z)}$ evaluated at the nominal trajectory.
In the next section, we demonstrate that the condition $(\gamma^\star(s),\gamma^{\mathrm{u}}(s)) \in {\mathbb{Z}}$, $s \in [0,1]$ is intrinsically satisfied by the proposed tube MPC and derive the tube dynamics based on Inequality~\eqref{eq:Vd-contraction}.
\begin{rem} \label{rem:CCM-computation}
The offline computation of matrices $M$, $K$ satisfying Assumption~\ref{ass:CCM} can be cast as linear matrix inequalities (LMIs), or a sum of squares (SOS) problem, using the convex re-parametrization $W=M^{-1}$ and $Y = KW$, compare~\cite{manchester2017control,singh2019robust,zhao2021tube,wang2020virtual} for details.
The geodesic $\gamma^\star$ is used in Equations~\eqref{eq:V-delta-def} and \eqref{eq:kappa-def} and can be computed using the Chebyshev global pseudospectral method with a suitable discretization~\cite{leung2017nonlinear}.
\end{rem}
\begin{rem}
In contrast to Proposition~\ref{prop:V-delta}, most of the CCM literature only considers a nominal setting~\cite{lohmiller1998contraction,manchester2017control}.
Closer to our work, the disturbance $d$ is explicitly considered in the CCM conditions in~\cite{zhao2021tube}.
While~\cite{zhao2021tube} results in a robust positive invariant set, we derived a state and input dependent bound, which will be exploited in the homothetic tube MPC formulation.
We note that the CCM parametrization in Assumption~\ref{ass:CCM} can also be generalized to allow for a dependence of $M$, $K$ on the nominal parameters $\overline{\theta}$, as suggested in~\cite{lopez2021universal} for adaptive control.
\end{rem}
\subsection{Homothetic tube} \label{sec:tube-design}
For the homothetic tube MPC formulation, we need conditions on the tube scaling $\delta$ to ensure that all uncertain trajectories are confined in the tube $\mathbb{T}_t$~\eqref{eq:tube-def} and satisfy the constraints.
First, the following proposition provides sufficient conditions for the constraint satisfaction inside the tube (cf. Eq.~\eqref{eq:tube_conceptual_2}).
\begin{prop} \label{prop:c_j}
Suppose Assumption~\ref{ass:CCM} holds. Then, for any $x,z \in {\mathbb{R}}^n$, $v \in {\mathbb{R}}^m$ satisfying
\begin{subequations}
\begin{align}
&h_j(z,v) + c_j V_\delta(x,z) \leq 0, \quad j \in \I{1,r}, \label{eq:constraint-tight1}\\
&c_j := \max_{(z,v) \in {\mathbb{Z}}} \left \| \left(\left.\dfrac{\partial h_j}{\partial x}\right|_{(z,v)} + \left.\dfrac{\partial h_j}{\partial u}\right|_{(z,v)} K(z) \right) M(z)^{-\frac{1}{2}} \right \| \label{eq:c_j-def}
\end{align}
it holds
\begin{align} \label{eq:constraint-sat}
(\gamma^\star(s),\gamma^\mathrm{u}(s)) \in {\mathbb{Z}}, \quad s \in [0,1],
\end{align}
\end{subequations}
with $\gamma^\star$ and $\gamma^\mathrm{u}$ according to Equations \eqref{eq:V-delta-def} and \eqref{eq:kappa-def}.
\end{prop}
\begin{pf}
Note that~\eqref{eq:constraint-tight1} yields $h_j(z,v) \leq 0$, $j \in \I{1,r}$.
We only consider $c_j>0$, as $c_j=0$ implies $\left. \dfrac{\mathrm{d} h_j}{\mathrm{d} s}\right|_{(\gamma^\star(s),\gamma^\mathrm{u}(s))} = 0$ for all $s \in [0,1]$, and thus $h_j(\gamma^\star(s),\gamma^\mathrm{u}(s)) = h_j(z,v) \leq 0$, $s \in [0,1]$, $j \in \I{1,r}$ follows trivially. For contradiction, suppose that $h_{j}(\gamma^\star(\hat{s}),\gamma^\mathrm{u}(\hat{s})) > 0$ for some $\hat{s} \in (0,1]$, and some $j \in \I{1,r}$. Since $h_{j}(\gamma^\star(0),\gamma^\mathrm{u}(0)) \leq 0$, and $h_{j}(\gamma^\star(s),\gamma^\mathrm{u}(s))$ is continuous w.r.t. $s$, there exists a $\overline{s}\in [0,\hat{s})$, such that $\max_{j \in \I{1,r}} h_j(\gamma^\star(\overline{s}),\gamma^\mathrm{u}(\overline{s})) =0$, and $(\gamma^\star(s),\gamma^\mathrm{u}(s)) \in {\mathbb{Z}}$, $s \in [0,\overline{s}]$. The gradient theorem yields
\begin{align*}
&h_j(\gamma^\star(\overline{s}),\gamma^\mathrm{u}(\overline{s})) - h_j(z,v) = \int_0^{\overline{s}} \left.\dfrac{\mathrm{d} h_j}{\mathrm{d} s}\right|_{(\gamma^\star(s),\gamma^\mathrm{u}(s))} \mathrm{d}s \\
=& \int_0^{\overline{s}} \left (\left.\dfrac{\partial h_j}{\partial x}\right|_{(\gamma^\star(s),\gamma^\mathrm{u}(s))} \right.\\
& + \left. \left.\dfrac{\partial h_j}{\partial u}\right|_{(\gamma^\star(s), \gamma^\mathrm{u}(s))} K(\gamma^\star(s)) \right) \cdot \gamma_{\mathrm{s}}^\star(s) \mathrm{d} s \\
\stackrel{\eqref{eq:c_j-def}}{\leq} & c_j \int_0^{\overline{s}} \left \| M(\gamma^\star(s))^{\frac{1}{2}} \gamma^\star_s(s) \right \| \mathrm{d} s \stackrel{\eqref{eq:V-delta-def}}{<} c_j V_\delta(x,z) \\
\stackrel{\eqref{eq:constraint-tight1}}{\leq}& -h_j(z,v), \quad j \in \I{1,r},
\end{align*}
which contradicts $\max_{j \in \I{1,r}} h_j(\gamma^\star(\overline{s}),\gamma^\mathrm{u}(\overline{s})) =0$. $~\hfill \square$
\end{pf}
This proposition provides sufficient conditions on the nominal trajectories and the scaling $\delta \geq V_\delta(x,z)$, such that the state/input and the corresponding geodesic satisfy the constraints. Hence, in Assumption~\ref{ass:CCM} it suffices to impose Inequalities~\eqref{eq:CCM} only on the constraint set, while most related CCM results require global assumptions~\cite{manchester2017control,singh2017robust,zhao2021tube} or additional geodesic convexity~\cite{schiller2022lyapunov}.
The following result uses Propositions~\ref{prop:V-delta} and \ref{prop:c_j} to derive a differential equation for the tube scaling $\delta$ that accounts for the disturbance $d$ and the parametric uncertainty in $\theta$.
\begin{prop} \label{prop:tube-dyn}
Suppose Assumption~\ref{ass:CCM} holds. Consider any initial state $x_0 \in {\mathbb{R}}^n$, nominal parameter $\overline{\theta}\in\Theta_0$, and trajectories $z_t$, $v_t$, $\delta_t$, $t\geq 0$ that satisfy
\begin{subequations}
\begin{align}
&h_j(z_t,v_t) + c_j \delta_t \leq 0, \quad j \in \I{1,r}, \label{eq:prop-constraint-tight} \\
&\dot{z}_t = \overline{f}(z_t,v_t,\overline{\theta}), \label{eq:prop-z-dot}\\
&V_\delta(x_0,z_0) \leq \delta_0, \label{eq:prop-delta_0}\\
&\dot{\delta}_t = -\left (\rho_\mathrm{c} - L_\mathbb{D}\right)\delta_t + \max_{\substack{i \in \I{1,n_{\Theta_0}} \\ j \in \I{1,n_\mathbb{D}}}} \left ( \sum_{k=1}^p L_{\mathrm{G},k}|[\theta^i - \overline{\theta}]_k| \delta_t \right. \nonumber \\
& \qquad \underbrace{\left. \vphantom{\sum_0^p}+\|G(z_t,v_t)(\theta^i - \overline{\theta}) + E(z_t)d^j\|_{M(z_t)} \right), \qquad \quad }_{:=f_\delta(\delta_t,z_t,v_t,\Theta_0,\overline{\theta})} \label{eq:f-delta-def}
\end{align}
for all $t \geq 0$, with $\theta^i$, $i \in \I{1,n_{\Theta_0}}$ and $d^j$, $j \in \I{1,n_\mathbb{D}}$ denoting the vertices of $\Theta_0$ and $\mathbb{D}$. Then, for any $\overline{\theta}, \theta \in \Theta_0$, $d(t) \in \mathbb{D}$, $t\geq 0$, the trajectory $x(0) = x_0$, $\dot{x}(t) = f_\mathrm{w}(x(t),\kappa(x(t),z_t,v_t),\theta,d(t))$, $t\geq 0$, satisfies
\begin{align}
V_\delta(x(t),z_t) \leq & \delta_t, \quad t\geq 0, \label{eq:delta-larger} \\
(x(t),\kappa(x(t),z_t,v_t)) \in & {\mathbb{Z}}, \quad t \geq 0. \label{eq:tube-constraint-sat}
\end{align}
\end{subequations}
\end{prop}
\begin{pf}
First, note that \eqref{eq:delta-larger} holds at time $t=0$ due to Condition~\eqref{eq:prop-delta_0}. For contradiction, suppose that there exists a time $\tau \geq 0$, such that \eqref{eq:delta-larger} holds for all $t \in [0,\tau]$, but is violated for $t>\tau$, i.e., $V_\delta(x(\tau),z_\tau) = \delta_\tau$ and $\dot{V}_\delta(x(\tau),z_\tau) > \dot{\delta}_\tau$. Given that \eqref{eq:delta-larger} holds for $t =\tau$, then Proposition \ref{prop:c_j} in combination with \eqref{eq:prop-constraint-tight} implies $(\gamma^\star(s),\gamma^\mathrm{u}(s))\in {\mathbb{Z}}$, $s \in [0,1]$, with $\gamma^\star \in \Gamma(z_\tau,x(\tau)), \gamma^\mathrm{u}$ based on Equations~\eqref{eq:V-delta-def} and \eqref{eq:kappa-def}. Hence, we can invoke Inequality~\eqref{eq:Vd-contraction} in Proposition~\ref{prop:V-delta}, yielding
\begin{align*}
\dot{V}_\delta(x(\tau),z_\tau) \stackrel{\eqref{eq:Vd-contraction},\eqref{eq:f-delta-def}}{\leq}& f_\delta(V_\delta(x(\tau),z_\tau),z_\tau,v_\tau,\Theta_0,\overline{\theta}) \\
=& f_\delta(\delta_\tau,z_\tau,v_\tau,\Theta_0,\overline{\theta}) \stackrel{\eqref{eq:f-delta-def}}{=} \dot{\delta}_\tau,
\end{align*}
which yields a contradiction. Hence, \eqref{eq:delta-larger} holds for $t\geq 0$. Furthermore, Proposition~\ref{prop:c_j} and Inequality~\eqref{eq:prop-constraint-tight} imply Inequality~\eqref{eq:tube-constraint-sat}. $\hfill \square$
\end{pf}
Proposition~\ref{prop:tube-dyn} is applicable if the nominal trajectories $z_t,v_t$ and the tube scaling $\delta_t$ are computed according to Conditions~\eqref{eq:prop-constraint-tight}--\eqref{eq:f-delta-def}. In that case, all uncertain trajectories $x(t)$, with feedback $\kappa(x(t),z_t,v_t)$ are confined within the tube $\mathbb{T}_t$ given by \eqref{eq:tube-def} (cf. Eq.~\eqref{eq:delta-larger}) and satisfy the constraints (cf. Eq.~\eqref{eq:tube-constraint-sat}).
Thus, we obtained simpler sufficient conditions (cf. Eq.~\eqref{eq:prop-constraint-tight}--\eqref{eq:f-delta-def}) to compute a feasible solution to Problem~\eqref{eq:conceptual}.
\subsection{Parameter update using set-membership estimation} \label{sec:set-update}
We use set-membership estimation to update the polytopic parameter set $\Theta_0$ online. This method is widely used in related RAMPC schemes~\cite{kohler2019linear,kohler2020RAMPC,lu2021robust,lorenzen2019robust,lopez2019adaptive}, since it allows for reduced conservatism.
For the considered continuous-time setting, we assume that the state $x(t)$ and input $u(t)$ can be measured exactly, and a noisy measurement of the state derivative is available, i.e., $\dot{\tilde{x}}(t) = \dot{x}(t) + \epsilon(t)$, where $\epsilon(t) \in \mathbb{D}_{\epsilon}$, $t\geq 0$, with some polytope $\mathbb{D}_\epsilon$. Then, the non-falsified parameter set given $(x,u,\dot{\tilde{x}})$ at time t is computed as
\begin{align*}
\Delta_{t} := \{ \theta \in {\mathbb{R}}^p \big| \exists (d,\epsilon) \in \mathbb{D} \times \mathbb{D}_{\epsilon}: \dot{\tilde{x}}(t) = \dot{x}(t) + \epsilon, &\\
\dot{x}(t) = f_\mathrm{w}(x(t),u(t),\theta,d)& \}.
\end{align*}
We update the parameter set at each sampling time $t_k$, $k \in \mathbb{I}_{\geq 1}$ using $n_\mathrm{m} \in \mathbb{I}_{\geq 1}$ measurements between $t_{k-1}$ and $t_{k}$, using the following polytopic set-intersection:
\begin{equation} \label{eq:Theta-update}
\Theta_{t_{k}} = \Theta_{t_{k-1}} \cap_{j \in \I{0,n_\mathrm{m}-1}} \Delta_{t_{k-1} + j \frac{T_\mathrm{s}}{n_\mathrm{m}}}, \quad k\in\mathbb{I}_{\geq 1}.
\end{equation}
\begin{prop} \label{prop:Theta-update}
The set $\Theta_{t_{k}}$ computed according to \eqref{eq:Theta-update} is a polytope, which satisfies
\begin{equation} \label{eq:Theta_t-prop}
\theta \in \Theta_{t_{k}} \subseteq \Theta_{t_{k-1}} \subseteq \Theta_0, \quad k \in \mathbb{I}_{\geq 1}.
\end{equation}
\end{prop}
\begin{pf}
Since $\Theta_{t_{k-1}}$, $\mathbb{D}$, and $\mathbb{D}_\epsilon$ are ploytopes, the sets $\Delta_{t}$, $t \geq 0$ are polyhedrons, and hence $\Theta_{t_{k}}$, $k \in \mathbb{I}_{\geq 1}$ are also polytopes. The condition $\theta \in \Theta_{t_{k}} \subseteq \Theta_{t_{k-1}}$ follows for all $k \in \mathbb{I}_{\geq 1}$ recursively, using the fact that $\theta \in \Theta_{0}$, and the definition of the non-falsified sets $\Delta_{t}$. $\hfill \qed$
\end{pf}
\begin{rem}
Computing the intersection of finitely many polytopes in \eqref{eq:Theta-update} is straightforward. However, the number of facets and vertices of the parameter set $\Theta_{t_k}$ might grow with the updates leading to increasing computational complexity. This issue can be addressed by considering fixed-complexity polytopes of the form $\Theta_{t_k} = \{\theta \in {\mathbb{R}}^p | H \theta \leq h_{t_k}\}$, with some fixed matrix $H \in {\mathbb{R}}^{n_\mathrm{p} \times p}$, and only update the vector $h_{t_k} \in {\mathbb{R}}^{n_p}$. This update retains the theoretical properties \eqref{eq:Theta_t-prop} and can be formulated as a linear program, see~\cite{lorenzen2019robust}.
\end{rem}
\begin{rem}
A noisy derivative $\dot{\tilde{x}}$ can be obtained using suitable filters~ \cite{Adetola2009adaptive,guay2015robust,lopez2019adaptive}.
Equivalently, the non-falsified set can be written using an integral form~\cite[Eq.~(17)]{cohen2022integralSetID}.
Note that instead of set-membership estimation, the parameter sets $\Theta_t$ can, e.g., also be updated using Lyapunov-type arguments~\cite{Adetola2009adaptive,guay2015robust} as long as Condition~\eqref{eq:Theta_t-prop} remains valid.
\end{rem}
\subsection{Terminal set constraint} \label{sec:terminal}
In order to derive recursive feasibility and convergence guarantees of the proposed MPC scheme, we require the following standard conditions on the terminal ingredients (terminal set, terminal control law, terminal cost).
\begin{assum} \label{ass:terminal}
There exists a terminal set ${\mathbb{X}_{\mathrm{f}}} \subseteq {\mathbb{R}}^n \times {\mathbb{R}}_{\geq 0} \times \Theta_0 \times 2^{{\mathbb{R}}^p}$, a terminal control law $k_{\mathrm{f}}:{\mathbb{R}}^n \times \Theta_0 \rightarrow {\mathbb{R}}^m$, and a continuous terminal cost $\ell_{\mathrm{f}}: {\mathbb{R}}^n \times \Theta_0 \rightarrow {\mathbb{R}}_{\geq 0}$, such that for any $(z,\delta,\overline{\theta},\Theta) \in {\mathbb{X}_{\mathrm{f}}}$,
the trajectories $\dot{z}_\tau = \overline{f}(z_\tau,k_\mathrm{f}(z_0,\overline{\theta}),\overline{\theta})$, $\dot{\delta}_\tau = f_\delta(\delta_\tau,z_\tau,k_\mathrm{f}(z_0,\overline{\theta}),\Theta,\overline{\theta})$, $\tau\in[0,T_{\mathrm{s}}]$, with $\delta_0 = \delta$, $z_0 = z$ satisfy
\begin{subequations}
\begin{align}
(z_{T_\mathrm{s}}, \delta_{T_\mathrm{s}}, \overline{\theta}, \Theta) & \in {\mathbb{X}_{\mathrm{f}}}, \label{eq:ass-term-PI} \\
h_j(z_\tau,k_{\mathrm{f}}(z_0,\overline{\theta})) + c_j \delta_\tau & \leq 0, \quad j \in \I{1,r}, \quad \tau \in [0,T_\mathrm{s}], \label{eq:ass-term-constraint} \\
\int_{0}^{T_\mathrm{s}} \ell(z_{\tau},k_{\mathrm{f}}(z_0,\overline{\theta}),\overline{\theta}) \mathrm{d}\tau & \leq \ell_{\mathrm{f}}(z_0, \overline{\theta}) - \ell_{\mathrm{f}}(z_{T_\mathrm{s}},\overline{\theta}) \label{eq:ass-term-convergence}.
\end{align}
Furthermore, for any $\hat{\delta} \in [0,\delta]$, $\hat{\Theta} \subseteq \Theta$, it holds
\begin{align} \label{eq:ass-term-monoton}
(z,\delta,\overline{\theta},\Theta) \in {\mathbb{X}_{\mathrm{f}}} \Rightarrow (z,\hat{\delta},\overline{\theta},\hat{\Theta}) \in {\mathbb{X}_{\mathrm{f}}}.
\end{align}
\end{subequations}
\end{assum}
This assumption is comparable to the standard conditions in MPC (cf.~\cite{rawlings2017model}), requiring positive invariance~\eqref{eq:ass-term-PI}, constraint satisfaction~\eqref{eq:ass-term-constraint}, and a local control Lyapunov function~\eqref{eq:ass-term-convergence}. Furthermore, we pose the monotonicity property~\eqref{eq:ass-term-monoton}. The terminal control input is chosen to be constant in the interval $[0,T_\mathrm{s}]$, which ensures that a piece-wise constant input parametrization can be used in the MPC.
The following proposition shows that these conditions hold using a simple terminal equality constraint w.r.t. the desired steady-state (cf. Sec.~\ref{sec:problem-statement}).
\begin{prop} \label{prop:term-eq}
Suppose Assumption~\ref{ass:CCM} holds.
Then, Assumption~\ref{ass:terminal} is satisfied with $k_\mathrm{f}(z,\overline{\theta}) = v_\mathrm{ref}(\overline{\theta})$, $\ell_\mathrm{f}(z,\overline{\theta}) \equiv 0$, and the following terminal set:
\begin{subequations}
\begin{align}
{\mathbb{X}_{\mathrm{f}}} = \left \{ \vphantom{2^{{\mathbb{R}}^p}} \right. & (z,\delta,\overline{\theta},\Theta) \in {\mathbb{R}}^n \times {\mathbb{R}}_{\geq 0} \times \Theta_0 \times \left. 2^{{\mathbb{R}}^p} \right |\\
& z = z_\mathrm{ref}(\overline{\theta}), \label{eq:term-eq-z}\\
& \exists \overline{\delta}_\mathrm{f} \in {\mathbb{R}}_{\geq 0}:~\delta \in [0,\overline{\delta}_\mathrm{f}], \label{eq:term-eq-delta-bar} \\
& f_\delta(\overline{\delta}_\mathrm{f},z,k_\mathrm{f}(z,\overline{\theta}),\Theta,\overline{\theta}) \leq 0, \label{eq:term-eq-delta-PI} \\
& \left. h_j(z,k_\mathrm{f}(z,\overline{\theta})) + c_j \overline{\delta}_\mathrm{f} \leq 0,~\forall j \in \I{1,r} \vphantom{2^{{\mathbb{R}}^p}}\right \}. \label{eq:term-eq-constraint}
\end{align}
\end{subequations}
\end{prop}
\begin{pf}
Condition \eqref{eq:term-eq-z} along with the choice of $k_\mathrm{f}$ yields $z_\tau = z_\mathrm{ref}(\overline{\theta})$ for $\tau \in [0,T_\mathrm{s}]$. Furthermore, $\delta_0\in[0,\overline{\delta}_\mathrm{f}]$ (cf.~\eqref{eq:term-eq-delta-bar}) and Inequality~\eqref{eq:term-eq-delta-PI} yields $\delta_\tau \in [0,\overline{\delta}_\mathrm{f}]$, $\tau \in [0,T_\mathrm{s}]$. Hence, positive invariance \eqref{eq:ass-term-PI} follows by considering the same $\overline{\delta}_\mathrm{f}$. Moreover, Condition~\eqref{eq:term-eq-constraint} and $\delta_\tau \in [0,\overline{\delta}_\mathrm{f}]$, $\tau\in[0,T_{\mathrm{s}}]$ yield Inequality \eqref{eq:ass-term-constraint}.
Condition~\eqref{eq:ass-term-convergence} follows with $\ell(z_\tau,k_\mathrm{f}(z_0,\overline{\theta}),\overline{\theta}) = 0$ using~\eqref{eq:stage-cost-def}. The monotonicity property~\eqref{eq:ass-term-monoton} follows using Condition~\eqref{eq:term-eq-delta-bar} and the definition of $f_\delta$ \eqref{eq:f-delta-def} yielding $f_\delta(\overline{\delta}_\mathrm{f},z,k_\mathrm{f}(z,\overline{\theta}),\hat{\Theta},\overline{\theta}) \leq f_\delta(\overline{\delta}_\mathrm{f},z,k_\mathrm{f}(z,\overline{\theta}),\Theta,\overline{\theta}) \leq 0$, $\hat{\Theta}\subseteq\Theta$. $\hfill \qed$
\end{pf}
Invariance of the terminal set w.r.t. the nominal dynamics $\overline{f}$ is guaranteed by requiring the terminal state to coincide with the steady-state of the nominal dynamics $z_\mathrm{ref}$~\eqref{eq:term-eq-z} and using the corresponding input $v_\mathrm{ref}$ as the terminal control law.
Note that the terminal set constraint includes not only the nominal state $z$, but also the tube scaling $\delta$, the nominal parameter $\overline{\theta}$, and the parameter set $\Theta$, since $\Theta$ is updated online and $\delta$, $\overline{\theta}$ are optimization variables in the proposed RAMPC framework (Sec.~\ref{sec:framework}). Positive invariance (cf.~\eqref{eq:term-eq-delta-PI}) and constraint satisfaction (cf.~\eqref{eq:term-eq-constraint}) are shown by additionally computing a constant $\overline{\delta}_\mathrm{f}\geq 0$ that upper bounds the tube scaling $\delta$ in the terminal set. For implementation, the constant $\overline{\delta}_\mathrm{f}\geq 0$ needs to be included as a decision variable in the MPC optimization problem introduced below.
\subsection{RAMPC framework and theoretical analysis} \label{sec:framework}
In the following, we present the overall RAMPC algorithm (Alg.~\ref{alg:offline}--\ref{alg:online}), including a corresponding theoretical analysis of the closed-loop properties (Thm.~\ref{thm:main}). The RAMPC scheme approximates the solution to the infinite-horizon Problem~\eqref{eq:conceptual} by a finite prediction horizon $T_\mathrm{f}=T_\mathrm{s} N$, $N\in\mathbb{I}_{\geq 1}$ using a receding horizon implementation with a sampling period of $T_\mathrm{s}$. In particular, at time $t$, given the measured state $x(t)$ and the updated parameter set $\Theta_t$, the proposed RAMPC scheme is characterized by the following optimization problem:
\begin{subequations} \label{eq:MPC}
\begin{align}
V_{T_\mathrm{f}}^\star(x(t),\Theta_t) =& \min_{\substack{v_{\cdot|t},z_{\cdot|t} \\ \delta_{\cdot|t}, \overline{\theta}_t}} \int_{0}^{T_\mathrm{f}} \ell(z_{\tau|t},v_{\tau|t},\overline{\theta}_t) + \ell_{\mathrm{f}}(z_{T_\mathrm{f}|t},\overline{\theta}_t) \mathrm{d}\tau \nonumber \\
\text{s.t.} \quad & \dot{z}_{\tau|t}=\overline{f}(z_{\tau|t},v_{\tau|t},\overline{\theta}_t), \label{eq:MPC-dynamics}\\
&\dot{\delta}_{\tau|t}=f_\delta(\delta_{\tau|t},z_{\tau|t},v_{\tau|t},\Theta_t,\overline{\theta}_t), \label{eq:MPC-delta}\\
& \overline{\theta}_t \in \Theta_0, \label{eq:MPC-theta-in-Theta}\\
&h_j(z_{\tau|t},v_{\tau|t}) + c_j \delta_{\tau|t} \leq 0, \label{eq:MPC-constraints}\\
&V_\delta(x(t),z_{0|t}) \leq \delta_{0|t}, \label{eq:MPC-delta0} \\
&(z_{T_\mathrm{f}|t},\delta_{T_\mathrm{f}|t},\overline{\theta}_t,\Theta_t) \in {\mathbb{X}_{\mathrm{f}}}, \label{eq:MPC-terminal}\\
&\tau \in [0,T_\mathrm{f}],\quad j \in \I{1,r}. \nonumber
\end{align}
\end{subequations}
The decision variables are the nominal state and input trajectories $z_{\cdot|t}$, $v_{\cdot|t}$, the tube scaling $\delta_{\cdot|t}$, and the nominal parameter $\overline{\theta}_t$. The tube scaling $\delta_{\tau|t}$ is propagated according to Proposition~\ref{prop:tube-dyn} (cf.~\eqref{eq:MPC-delta}, \eqref{eq:MPC-delta0}) and the nominal trajectory $z_{\tau|t}$ satisfies the nominal dynamics $\overline{f}$ \eqref{eq:MPC-dynamics}. The constraint tightening~\eqref{eq:MPC-constraints} ensures that the true state and input trajectory satisfy the constraints~\eqref{eq:constr} (cf. Prop.~\ref{prop:c_j}/\ref{prop:tube-dyn}). Condition~\eqref{eq:MPC-theta-in-Theta} allows to additionally optimize over the nominal parameter that influences the dynamics in~\eqref{eq:MPC-dynamics}--\eqref{eq:MPC-delta}. Furthermore, Condition~\eqref{eq:MPC-terminal} captures the terminal constraint.
For simplicity, we parameterize the nominal input $v_{\tau|t}$ as piece-wise constant in each interval $\tau \in [jT_\mathrm{s},(j+1)T_\mathrm{s})$, $j \in \I{0,N-1}$.
Appropriate conditions ensuring that a minimizer to Problem~\eqref{eq:MPC} exists can be found in \cite[Prop. 2]{fontes2001general}. We assume that a corresponding unique minimizer exists, which is denoted by the superscript $\star$. We solve the problem at each sampling time $t_k$, $k \in \mathbb{I}_{\geq 0}$, and the resulting closed-loop system is given by
\begin{align} \label{eq:closed-loop}
\begin{split}
u(t) =& \kappa(x(t),z_{t-t_k|t_k}^*,v_{t-t_k|t_k}^*), \quad t \in [t_k,t_{k+1}), \\
\dot{x}(t) =& f_{\mathrm{w}}(x(t),u(t),\theta,d(t)).
\end{split}
\end{align}
The offline design and the online operation are captured in the algorithms below.
\begin{algorithm}[H]
\caption{Offline design}\label{alg:offline}
\begin{algorithmic}[1]
\Statex Given model with uncertainty characterization $\mathbb{D}$, $\Theta_0$, $\mathbb{D}_{\epsilon}$, constraints~\eqref{eq:constr}, desired setpoint $z_\mathrm{ref}(\overline{\theta})$, $v_\mathrm{ref}(\overline{\theta})$ and stage cost weighting $Q,R \succeq 0$~\eqref{eq:stage-cost-def}.
\State Compute $M(x)$, $K(x)$ and $\rho_\mathrm{c}$ (Ass.~\ref{ass:CCM}, Rk.~\ref{rem:CCM-computation}).
\State Compute constants $L_\mathbb{D}$~\eqref{eq:app-L-D}, $L_{\mathrm{G},k}$~\eqref{eq:app-L-G}, $c_j$~\eqref{eq:c_j-def}.
\State Set sampling period $T_\mathrm{s}$ and prediction horizon $T_\mathrm{f}$.
\State Design terminal ingredients (Ass.~\ref{ass:terminal}, Prop.~\ref{prop:term-eq}).
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[H]
\caption{Online operation}\label{alg:online}
\begin{algorithmic}
\For{each sampling time $t_k = k T_\mathrm{s}$, $k \in \mathbb{I}_{\geq 0}$}
\State Update $\Theta_{t_k}$ \eqref{eq:Theta-update} using measurements ($x,\dot{\tilde{x}},u$).
\State Solve Problem \eqref{eq:MPC}.
\State Apply feedback $\kappa$ over period $t \in [t_k,t_{k+1})$~\eqref{eq:closed-loop}.
\EndFor
\end{algorithmic}
\end{algorithm}
The following theorem summarizes the theoretical properties of the proposed RAMPC algorithm.
\begin{thm} \label{thm:main}
Let Assumptions~\ref{ass:CCM} and \ref{ass:terminal} hold. Suppose that Problem \eqref{eq:MPC} is feasible at time $t=0$ with initial state $x_0$ and parameter set $\Theta_{0}$. Then, Problem \eqref{eq:MPC} is feasible for all sampling times $t_k$, $k \in \mathbb{I}_{\geq 0}$, and the closed-loop system~\eqref{eq:closed-loop} resulting from Algorithm~\ref{alg:online} satisfies the constraints \eqref{eq:constr} for all $t \geq 0$. Furthermore, the resulting nominal trajectories converge to the reference state and input, i.e., $\lim_{k \to \infty}\left\|(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star) - (z_\mathrm{ref}(\overline{\theta}_{t_k}^\star),v_\mathrm{ref}(\overline{\theta}_{t_k}^\star)) \right\| = 0$, $\tau \in [0,T_\mathrm{s})$.
\end{thm}
\begin{pf}
\textbf{Part I:} Assume that Problem~\eqref{eq:MPC} is feasible at some time $t_k$, $k\in \mathbb{I}_{\geq 0}$. For simplicity, we define $v_{\tau|t_k}^\star := k_{\mathrm{f}}(z_{T|t_k}^\star,\overline{\theta}_{t_k}^\star)$ and $z_{\tau|t_k}^\star$, $\delta_{\tau|t_k}^\star$ according to \eqref{eq:MPC-dynamics} and \eqref{eq:MPC-delta} for $\tau \in [T_\mathrm{f},T_\mathrm{f}+T_\mathrm{s}]$. At time $t_{k+1}$, consider the candidate solution
\begin{align*}
\overline{\theta}_{t_{k+1}} &= \overline{\theta}_{t_k}^\star \in \Theta_0, \quad v_{\tau|t_{k+1}} = v_{\tau + T_\mathrm{s}|t_k}^\star, \quad \tau \in [0,T_\mathrm{f}], \\
\quad z_{0|t_{k+1}} & = z_{T_\mathrm{s}|t_k}^\star, \quad \delta_{0|t_{k+1}} = V_\delta(x(t_{k+1}),z_{T_\mathrm{s}|t_k}^\star),
\end{align*}
with trajectories $z_{\tau|t_{k+1}}$, $\delta_{\tau|t_{k+1}}$, $\tau \in (0,T_\mathrm{f}]$ according to the constraints \eqref{eq:MPC-dynamics}, \eqref{eq:MPC-delta}, which implies $z_{\tau|t_{k+1}} = z_{\tau+T_\mathrm{s}|t_k}^\star$, $\tau \in~[0,T_\mathrm{f}]$. In the following, we show that the tube scaling satisfies the following nestedness property:
\begin{equation} \label{eq:T-cont-proof-delta-nested}
\delta_{\tau|t_{k+1}} \leq \delta_{\tau+T_\mathrm{s}|t_k}^\star, \quad \tau \in [0,T_\mathrm{f}].
\end{equation}
First, note that Condition~\eqref{eq:delta-larger} in Proposition~\ref{prop:tube-dyn} also holds for any $\Theta \subseteq \Theta_0$ with $\dot{x} = f_\mathrm{w}(x,u,\theta,d)$, $\dot{\delta}=f_\delta(\delta,z,v,\Theta,\overline{\theta})$, $\overline{\theta}\in\Theta_0$, $\theta \in \Theta$, $d \in \mathbb{D}$ yielding
\begin{equation*}
V_\delta(x(t_k+\tau),z_{\tau|t_k}^\star) \leq \delta_{\tau|t_k}^\star, \quad \tau \in [0,T_\mathrm{s}].
\end{equation*}
Hence, the initial value of the tube scaling satisfies
\begin{equation*}
\delta_{0|t_{k+1}} = V_\delta(x(t_{k+1}),z_{T_\mathrm{s}|t_k}^\star) \leq \delta_{T_\mathrm{s}|t_k}^\star.
\end{equation*}
For contradiction, suppose that there exists a time $\hat{\tau} \in [0,T_\mathrm{f})$, such that \eqref{eq:T-cont-proof-delta-nested} holds for all $\tau \in [0,\hat{\tau}]$, but is violated for $\tau > \hat{\tau}$, i.e., $\delta_{\hat{\tau}|t_{k+1}} = \delta_{\hat{\tau}+T_\mathrm{s}|t_k}^\star$ and $\dot{\delta}_{\hat{\tau}|t_{k+1}} > \dot{\delta}_{\hat{\tau}+T_\mathrm{s}|t_k}^\star$. The dynamics of the tube scaling~\eqref{eq:f-delta-def} in combination with $\Theta_{t_{k+1}} \subseteq \Theta_{t_k}$ yield
\begin{align*}
\dot{\delta}_{\hat{\tau}|t_{k+1}}=& f_\delta(\delta_{\hat{\tau}|t_{k+1}},z_{\hat{\tau}+T_\mathrm{s}|t_k}^\star,v_{\hat{\tau}+T_\mathrm{s}|t_k}^\star,\Theta_{t_{k+1}},\overline{\theta}_{t_k}^\star) \\
\leq & f_\delta(\delta_{\hat{\tau}+T_\mathrm{s}|t_k}^\star,z_{\hat{\tau}+T_\mathrm{s}|t_k}^\star,v_{\hat{\tau}+T_\mathrm{s}|t_k}^\star,\Theta_{t_k},\overline{\theta}_{t_k}^\star) \\
= & \dot{\delta}_{\hat{\tau}+T_\mathrm{s}|t_k}^\star,
\end{align*}
which is a contradiction, and hence~\eqref{eq:T-cont-proof-delta-nested} holds. \\
The candidate solution satisfies Condition~\eqref{eq:MPC-terminal} due to the invariance~\eqref{eq:ass-term-PI} and monotonicity property~\eqref{eq:ass-term-monoton} of ${\mathbb{X}_{\mathrm{f}}}$, using Inequality~\eqref{eq:T-cont-proof-delta-nested} and $\Theta_{t_{k+1}}\subseteq\Theta_{t_k}$. Inequality~\eqref{eq:MPC-constraints} for $\tau \in [T_\mathrm{f}-T_\mathrm{s},T_\mathrm{f}]$ follows analogously using Condition~\eqref{eq:ass-term-constraint}. For $\tau \in [0,T_\mathrm{f} - T_\mathrm{s}]$, Condition~\eqref{eq:MPC-constraints} holds with:
\begin{align*}
& h_j(z_{\tau|t_{k+1}}, v_{\tau|t_{k+1}}) + c_j \delta_{\tau|t_{k+1}} \\
& \stackrel{\eqref{eq:T-cont-proof-delta-nested}}{\leq} h_j(z_{\tau+T_\mathrm{s}|t_k}^\star, v_{\tau+T_\mathrm{s}|t_k}^\star) + c_j \delta_{\tau+T_\mathrm{s}|t_k}^\star \stackrel{\eqref{eq:MPC-constraints}}{\leq} 0,
\end{align*}
where we used feasibility of Problem~\eqref{eq:MPC} at time $t_k$ and the fact that $c_j \geq 0$. \\
\textbf{Part II:} Constraint satisfaction~\eqref{eq:constr} for the closed-loop system follows from Proposition~\ref{prop:tube-dyn} using the tightened constraints~\eqref{eq:MPC-constraints} for $\tau\in[0,T_\mathrm{s}]$ and the tube dynamics~\eqref{eq:MPC-delta0}, \eqref{eq:MPC-delta}. \\
\textbf{Part III:}
Convergence is established using standard inequalities for the optimal cost $V_{T_\mathrm{f}}^\star$. In particular, the feasible candidate solution provides an upper bound to the optimal cost, i.e.,
\begin{equation*}
\begin{split}
&V_{T_\mathrm{f}}^\star(x(t_{k+1}),\Theta_{t_{k+1}})\\
\leq & \int_{0}^{T_\mathrm{f}} \ell(z_{\tau+T_\mathrm{s}|t_k}^\star, v_{\tau+T_\mathrm{s}|t_k}^\star,\overline{\theta}_{t_k}^\star) + \ell_{\mathrm{f}}(z_{T_\mathrm{f}+T_\mathrm{s}|t_k}^\star,\overline{\theta}_{t_k}^\star) \mathrm{d}\tau\\
= & V_{T_\mathrm{f}}^\star(x(t_k),\Theta_{t_k}) - \int_0^{T_\mathrm{s}} \ell(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star,\overline{\theta}_{t_k}^\star)\mathrm{d}\tau \\
&- \ell_{\mathrm{f}}(z_{T_\mathrm{f}|t_k}^\star,\overline{\theta}_{t_k}^\star) + \int_{T_\mathrm{f}}^{T_\mathrm{f}+T_\mathrm{s}} \ell(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star,\overline{\theta}_{t_k}^\star)\mathrm{d}\tau \\
&+ \ell_{\mathrm{f}}(z_{T_\mathrm{f}+T_\mathrm{s}|t_k}^\star,\overline{\theta}_{t_k}^\star) \\
\stackrel{\eqref{eq:ass-term-convergence}}{\leq} & V_{T_\mathrm{f}}^\star(x(t_k),\Theta_{t_k}) - \int_0^{T_\mathrm{s}} \ell(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star,\overline{\theta}_{t_k}^\star)\mathrm{d}\tau.
\end{split}
\end{equation*}
Applying the previous inequality recursively yields
\begin{align*}
&V_{T_\mathrm{f}}^\star(x_0,\Theta_0) - \limsup_{k \rightarrow \infty}V_{T_\mathrm{f}}^\star(x(t_k),\Theta_{t_k}) \\
\geq &\sum_{k=0}^\infty \int_0^{T_\mathrm{s}} \ell(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star,\overline{\theta}_{t_k}^\star)\mathrm{d}\tau.
\end{align*}
Compact constraints in combination with a continuous cost imply a uniform bound on $V_{T_\mathrm{f}}^\star$, and hence the right hand side of the above inequality is bounded, yielding
\begin{align*}
\lim_{k\to \infty} \int_0^{T_\mathrm{s}} \ell(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star,\overline{\theta}_{t_k}^\star)\mathrm{d}\tau = 0.
\end{align*}
Note that $z_{\tau|t_k}^\star$ is uniformly continuous in $\tau$, given $\overline{f}$ Lipschitz and $(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star)$, $\overline{\theta}^\star_{t_k}$ subject to compact constraints~\eqref{eq:MPC-constraints}, \eqref{eq:MPC-theta-in-Theta}. Furthermore, $v_{\tau|t_k}^\star$ and $\overline{\theta}^\star_{t_k}$ are constant on $\tau \in [0,T_\mathrm{s})$, thus $\ell(z_{\tau|t_k}^\star,v_{\tau|t_k}^\star,\overline{\theta}_{t_k}^\star)$ is uniformly continuous on $\tau \in [0,T_\mathrm{s})$. Since $\ell$ is also positive definite, Barbalat's Lemma \cite{khalil2002nonlinear} can be invoked yielding:
\begin{align*}
\lim_{k \to \infty} \left \|(z_{\tau|t_k}^\star, v_{\tau|t_k}^\star) - (z_\mathrm{ref}(\overline{\theta}^\star_{t_k}), v_\mathrm{ref}(\overline{\theta}^\star_{t_k})) \right\| = 0. \qed
\end{align*}
\end{pf}
Theorem~\ref{thm:main} provides all the desired properties of a robust MPC scheme, while incorporating online updates of the parameter set $\Theta_t$ to reduce conservatism.
\begin{rem} \label{rem:implementation}
The dynamics~\eqref{eq:MPC-delta} involve the maximization operator from $f_\delta$~\eqref{eq:f-delta-def}, which may complicate integration. This issue is circumvented by introducing an auxiliary variable $\overline{w}_{\tau|t} \geq 0$ and using the following constraints:
\begin{align}
\dot{\delta}_{\tau|t}= &-(\rho_\mathrm{c} - L_\mathbb{D})\delta_{\tau|t}+\overline{w}_{\tau|t}, \label{eq:w_bar} \nonumber\\
\overline{w}_{\tau|t} \geq& \sum_{k=1}^p L_{\mathrm{G},k}|[\theta^i - \overline{\theta}_t]_k| \delta_{\tau|t} +\|G(z_{\tau|t},v_{\tau|t})(\theta^i - \overline{\theta}_t) \nonumber \\
& + E(z_{\tau|t})d^j\|_{M(z_{\tau|t})}, \quad i \in \I{1,n_{\Theta_t}}, \quad j \in \I{1,n_\mathbb{D}}.
\end{align}
Hence, we can equivalently write the dynamics~\eqref{eq:MPC-dynamics}, \eqref{eq:MPC-delta} and the constraints~\eqref{eq:MPC-constraints} using smooth continuous-time dynamics with state $(z,\delta) \in {\mathbb{R}}^{n+1}$, input $(v,\overline{w})\in{\mathbb{R}}^{m+1}$, and continuous-time constraints~\eqref{eq:MPC-constraints}, \eqref{eq:w_bar}, analogous to standard continuous-time MPC schemes~\cite[Fig. 12.10]{grune2017nonlinear}. Correspondingly, we use standard methods to integrate the dynamics. Furthermore, we relax the continuous-time constraint satisfaction by only enforcing constraints on some discrete points~\cite[Sec. 8]{rawlings2017model}.
However, one could also use additional modifications to rigorously guarantee continuous-time constraint satisfaction (cf. \cite[Thm.~3]{magni2004model}, \cite{fontes2018guaranteed}). \\
Note that constraint~\eqref{eq:MPC-delta0} on the initial state can be equivalently written as $\int_0^1\|\gamma_{\mathrm{s}}(s)\|_{M(\gamma(s))}\mathrm{d}s \leq \delta_{0|t}$ (cf. Eq.~\eqref{eq:V-delta-def}) by using the curve $\gamma \in \Gamma(z_{0|t},x(t))$ as an additional optimization variable. The integral is approximated with numerical quadrature, and the curve $\gamma$ is finitely parametrized using a pseudospectral method (cf., Remark~\ref{rem:CCM-computation} and \cite{leung2017nonlinear,zhao2021tube}). Thus, Problem~\eqref{eq:MPC} is approximated with a finite dimensional nonlinear program.
\end{rem}
\subsection{Discussion} \label{sec:discussion}
In the following, we discuss the results and benefits of the proposed RAMPC framework in comparison to existing robust and robust adaptive MPC schemes.
\textit{Stability properties}:
Suppose for simplicity that $z_\mathrm{ref}(\overline{\theta})$, $v_\mathrm{ref}(\overline{\theta})$ is unique. Then, the convergence result in Theorem~\ref{thm:main} yields that the uncertain closed-loop trajectory $x(t)$ converges to an RPI set corresponding to the stabilizing feedback $u(t)=\kappa\left(x(t),z_\mathrm{ref}(\overline{\theta}),v_\mathrm{ref}(\overline{\theta})\right)$, analogous to standard tube MPC results~\cite{mayne2005robust,bayer2013discrete}. To improve performance, the cost function can also be based on a state trajectory predicted according to the true measured state $x(t)$ and some other parameter estimate $\hat{\theta}_{t_k}$. Specifically, by using least mean square estimate, finite-gain $\mathcal{L}_2$-stability w.r.t. disturbance $d$ is shown in \cite{kohler2020RAMPC,kohler2019linear,lorenzen2019robust}.
\textit{Tube propagation}:
The tube propagation derived in Proposition~\ref{prop:tube-dyn} is similar to the RMPC formulations presented in~\cite[App. B]{Koehler2020Robust} and \cite{kohler2020RAMPC}, which use the same tube parametrization~\eqref{eq:tube-def} with $V_\delta$, $z$, $\delta$, compare also~\cite{rakovic2022homothetic}.
However, in~\cite{Koehler2020Robust,kohler2020RAMPC} explicit design formulas for the tube propagation where only established for the special case of ellipsoidal sets (cf.~\cite[Sec.~3.5]{kohler2020RAMPC}) and \cite{rakovic2022homothetic} only studies the special case of a shift invariant parametrization.\footnote{The controller $\kappa$, the incremental Lyapunov function $V_\delta$, the tube $\mathbb{T}$, and the bound on the model mismatch may only depend on the error $x-z$, which excludes general CCMs.}
Furthermore, the simple tube propagation in~\cite{Adetola2009adaptive,guay2015robust,pin2009robust} based on Lipschitz continuity is a further special case of the proposed framework with $M(z)=I$ and $K(z)=0$, compare also~\cite[Rk. 4]{Koehler2020Robust} and \cite[Sec.~3.6]{kohler2020RAMPC} for a detailed discussion.
In contrast to these existing RMPC approaches, we utilized general CCMs to provide simple explicit formulas for the tube dynamics and constraint tightening using the gradient theorem, compare~\eqref{eq:app-L-G}, \eqref{eq:app-L-D}, \eqref{eq:c_j-def}. \\
By using the upper bound $M(z)\preceq\overline{M}$ in Inequality~\eqref{eq:w_bar}, we can also obtain a computationally cheaper, yet more conservative result.
Moreover, in the robust case ($\overline{\theta}$, $\Theta$ constant), we can compute a constant $L \leq L_\mathbb{D} + \max_{i \in \I{1,n_{\Theta_0}}} \sum_{k=1}^{p} L_{\mathrm{G},k} \left | [\theta^i - \overline{\theta}]_k \right|$, which is also less conservative compared to the formulation in~\cite{Koehler2020Robust}.
\textit{Homothetic vs. rigid tube}:
Many robust MPC formulations use a fixed (rigid) tube scaling $\delta_{\tau|t}=\overline{\delta}>0$ for all $\tau\geq0$ instead of implementing a tube propagation of the form Eq.~\eqref{eq:f-delta-def}, compare~\cite{singh2019robust,singh2017robust,wabersich2021nonlinear,zhao2021tube,bayer2013discrete,yu2013tube}.
This special case can be recovered by computing a constant $\overline{\delta}>0$, such that $\dot{\delta}=f_\delta(\overline{\delta},z,v,\overline{\theta},\Theta)\leq 0$ for all $(z,v)\in\mathbb{Z}$.
Similarly, an arbitrary constant $\delta_{\tau|t}=\overline{\delta}_{\max}>0$, $\tau\geq 0$, can be fixed offline by adding the constraint $\overline{w}_{\tau|t}\leq \overline{w}_{\max}:=(\rho_c-L_{\mathbb{D}})\overline{\delta}_{\max}$ to upper bound the right hand side of Inequality~\eqref{eq:w_bar} to the optimization problem.
This restricts the nominal state and input to regions with small uncertainty, compare~\cite{wabersich2021nonlinear}.
On the other hand, the proposed homothetic tube MPC overcomes this conservatism by systematically changing the scaling $\delta$ along the horizon, compare also the numerical example in Section~\ref{sec:example}.
\textit{Initial state constraint}:
The proposed homothetic tube MPC framework allows for an optimization over the initial state $z_{0|t}$. However, this requires an evaluation of $V_\delta$ in Condition~\eqref{eq:MPC-delta0}, which increases the computational complexity, compare Remark~\ref{rem:implementation}.
This can be avoided by considering the more restrictive initial state constraint $z_{0|t}=x(t)$, $\delta_{0|t}=0$ (cf. Eq.~\eqref{eq:MPC-delta0}). Such an initial state constraint is also suggested in the homothetic tube MPC schemes in~\cite{Koehler2020Robust,kohler2020RAMPC}, where recursive feasibility is demonstrated using a candidate input $v$ that tracks the previous optimal solution $z^\star$.\footnote{Accounting for the piece-wise constant input parametrization requires additional modifications (cf., e.g.,~\cite[Rk.~14]{Koehler2020Robust}).}
We note that this approach necessitates a stronger (robust) positive invariance condition on the terminal set, which excludes the simple terminal equality constraints (cf. Prop.~\ref{prop:term-eq}).
\textit{Adaptive MPC}:
Compared to existing adaptive MPC frameworks (e.g.,~\cite{kohler2020RAMPC,Adetola2009adaptive,guay2015robust,lopez2019adaptive}), one main novelty is the optimization over the nominal parameter $\overline{\theta}_t$, yielding a number of benefits.
By choosing a nominal parameter $\overline{\theta}_t$ close to the center of the updated parameter set $\Theta_t$, the tube scaling $\delta_t$ can be kept small (cf. Eq.~\eqref{eq:f-delta-def}). Intuitively, one might be inclined to simply fix $\overline{\theta}_t$ at the center of $\Theta_t$, as done in \cite{kohler2020RAMPC,Adetola2009adaptive,guay2015robust,lopez2019adaptive,kohler2019linear,didier2021robust}. However, this complicates the recursive feasibility analysis, limiting the applicability to restrictive parameter sets, such as hypercubes~\cite{kohler2020RAMPC,kohler2019linear,didier2021robust,lopez2019adaptive} or balls~\cite{guay2015robust,Adetola2009adaptive}. On the other hand, optimizing over $\overline{\theta}_t$ enables a simple recursive feasibility proof by shifting the previous optimal solution.
As a result, the proposed RAMPC scheme is applicable to general CCMs (Ass.~\ref{ass:CCM}), and requires no restrictions on the parametrization in the set-membership estimation (Sec.~\ref{sec:set-update}). \\
Finally, in contrast to existing adaptive methods (cf.~\cite[App. B]{kohler2019linear}, \cite[Chap. II.D]{didier2021robust}), tracking parameter dependent set-points $(z_{\mathrm{ref}}(\overline{\theta}),v_{\mathrm{ref}}(\overline{\theta}))$ is straightforward.
In case a desired output $y^\mathrm{d} \stackrel{!}{=} C x+D u$ is available, such a set-point can be implicitly specified using:
\begin{equation*}
\begin{split}
(z_\mathrm{ref}(\overline{\theta}),v_\mathrm{ref}(\overline{\theta})) \in& \arg \min_{\substack{z\in {\mathbb{R}}^n \\ v \in {\mathbb{R}}^m}} \|C z + D v - y^\mathrm{d}\|^2 \\
\textnormal{s.t.} \quad & \overline{f}(z,v,\overline{\theta}) = 0, \quad (z,v) \in {\mathbb{Z}},
\end{split}
\end{equation*}
which can also be included in the online computations, similar to~\cite{limon2018nonlinear}.
\section{Numerical example} \label{sec:example}
The following example demonstrates the benefits of the proposed homothetic tube MPC framework and the online model updates. We consider the following planar quadrotor from \cite{singh2017robust, zhao2021tube}:
\begin{equation*}
{\scriptsize
\begin{bmatrix}
\dot{p}_1 \\ \dot{p}_2 \\ \dot{\phi} \\ \dot{v}_1 \\ \dot{v}_2 \\ \ddot{\phi}
\end{bmatrix}
=
\begin{bmatrix}
v_1 \cos(\phi) - v_2 \sin(\phi) \\
v_1 \sin (\phi) + v_2 \cos(\phi) \\
\dot{\phi} \\
v_2 \dot{\phi} - g \sin(\phi) \\
-v_1 \dot{\phi} - g \cos(\phi) \\
0
\end{bmatrix}
+
\begin{bmatrix}
0 & 0 \\
0 & 0 \\
0 & 0 \\
0 & 0 \\
\frac{1}{m} & \frac{1}{m} \\
\frac{l}{J} & \frac{-l}{J} \\
\end{bmatrix}
u +
\begin{bmatrix}
0\\0\\0\\\cos(\phi)\\ -\sin(\phi) \\ 0
\end{bmatrix}
d,}
\end{equation*}
where $p_1, p_2$ are horizontal and vertical positions, $v_1, v_2$ are velocities in the body frame, and $\phi,\dot{\phi}$ denote the angle/angular velocity. The control input $u=[u_1,u_2]^\top$ is the thrust force produced by the propellers and $d$ is a wind disturbance. Furthermore, $g$, $m$, and $l$ denote the gravitational acceleration, mass, and the distance between each of the propellers and the vehicle center, respectively. Compared to the setting in~\cite{singh2017robust,zhao2021tube}, we additionally consider the mass as an uncertain parameter $\theta=\frac{1}{m}$ with initial estimate $\overline{\theta}_0 = 2.058$, uncertainty set $\Theta_0 = [0.99\overline{\theta}_0, 1.01\overline{\theta}_0]$, and true parameter $\theta=1.01\overline{\theta}_0$. The disturbance satisfies $d(t)\in \mathbb{D} := [-0.1,0.1]$. The state and input constraint set~\eqref{eq:Z_set} are given by:
$|\phi| \leq \pi/3$, $|\dot{\phi}| \leq \pi$, $|v_1| \leq 2$, $|v_2| \leq 1$, $u_i\in[-1,3.5]$, $i\in\{1,2\}$.
The goal is to reach position $p=0$, which corresponds to $z_\mathrm{ref}(\overline{\theta}) = 0$, $v_\mathrm{ref}(\overline{\theta}) = \frac{g}{2\overline{\theta}} [1, 1]^\top$.
We have placed circular obstacles between the initial position and the goal in order to provide a challenging control problem, compare Figures~\ref{fig:A-R}--\ref{fig:R-CB}.
\subsubsection*{Implementation details}
The CCM (Ass.~\ref{ass:CCM}) is computed using the convex re-parametrization $Y$, $W$ (cf. Remark~\ref{rem:CCM-computation}) by adjusting the SOS code from~\cite{zhao2021tube}.
We optimize for a CCM that results in a smaller tube by adding the following LMI
\begin{equation} \label{eq:extra-LMI}
\begin{bmatrix}
W & G\tilde{\theta} + Ed \\ (G\tilde{\theta} + Ed)^\top & \tilde{w}^2_{\max}
\end{bmatrix}
\succeq 0, \quad \forall (d,\tilde{\theta}+\overline{\theta}_0)\in\mathbb{D}\times\Theta_0
\end{equation}
and minimizing $\tilde{w}_{\max}$.\footnote{Applying the Schur complement to~\eqref{eq:extra-LMI} yields $\tilde{w}_{\max}^2 \geq \|G\tilde{\theta} + Ed\|_M^2$ which provides an upper bound to the last term in $f_\delta$~\eqref{eq:f-delta-def}.}
We note that simpler RMPC schemes based on ellipsoidal/polytopic RPI sets (cf., e.g.,~\cite{yu2013tube,wabersich2021nonlinear,rakovic2022homothetic}, corresponding to constant matrices $Y,W$) or feedback linearization (cf.~\cite{lopez2019adaptive}) are not feasible for this nonlinear problem.\\
The dynamics~\eqref{eq:MPC-dynamics}, \eqref{eq:MPC-delta} as well as the true dynamics~\eqref{eq:system} are discretized using explicit Runge-Kutta 4 discretization with the sampling interval $T_\mathrm{s} = 150$ ms, and we only enforce the state and input constraints at the sampling times $t_k$ (cf. Remark~\ref{rem:implementation}). We choose a prediction horizon of $T_{\mathrm{f}}=N \cdot T_{\mathrm{s}}=3.75$ s based on $N=25$ steps. The disturbance $d(t)$ and the measurement noise $\epsilon(t)$ are sampled uniformly from the sets $\mathbb{D}$ and $\mathbb{D}_\epsilon = \mathbb{D}$. The geodesic in~\eqref{eq:MPC-delta0} is approximated using Chebyshev interpolating polynomials up to degree 2 (cf. Remark~\ref{rem:implementation}) with the code from~\cite{zhao2021tube}. \\
All computations were carried out using Matlab, on a Dell Inspiron 15-3567 laptop with Intel i7-7500U CPU and 8 GB RAM running Windows 10. The MPC problems are solved with IPOPT \cite{wachter2006implementation} formulated in CasADi \cite{andersson2019casadi}.\footnote{To avoid local minima, we solve the resulting NLP with initial guesses based on different paths around the obstacles.} The corresponding code is available online.\footnote{gitlab.ethz.ch/ics/RAMPC-CCM}
\subsubsection*{Benefits of online model updates}
We demonstrate the benefits of online model updates using set-membership estimation (cf. Sec.~\ref{sec:set-update}) by comparing the proposed robust adaptive MPC to a purely robust MPC variation that uses $\Theta_{t_k}=\Theta_0$. The results are shown in Figure~\ref{fig:A-R}. Given the initial uncertainty, we cannot plan a safe trajectory through the smaller gap between the obstacles. Hence, both formulations plan a more conservative trajectory passing through the larger gap. After the first sampling period $t=T_\mathrm{s}$, we perform the set-membership update~\eqref{eq:Theta-update} using $n_m=1$, resulting in a $42\%$ reduction of the size of the parameter set $\Theta_{t_k}$. As a result, the adaptive formulation can plan through the smaller gap.
As standard in (robust) MPC, the closed-loop trajectories can be significantly more aggressive than the conservative open-loop predictions due to the re-planning inherent in MPC.
\vspace{2mm}
\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{Figures/AMPC-RMPC.eps}
\caption{Comparison of the proposed RMPC scheme with (blue) and without (red) adaptation. The open-loop solutions at $t=T_\mathrm{s}$ are shown with nominal trajectories (solid) and ellipsoidal over-approximations of the tube based on $\underline{M}$. The closed-loop trajectories are dashed.} \label{fig:A-R}
\end{figure}
\subsubsection*{Comparison to rigid tube MPC}
We also show the advantages of the state and input dependent disturbance bound~\eqref{eq:f-delta-def} in the proposed homothetic tube MPC scheme compared to a rigid tube approach, which is widely used in literature~\cite{mayne2005robust,yu2013tube,zhao2021tube,singh2017robust,singh2019robust}. To this end, we focus on the purely robust setting and use no online model updates. We compute a constant scaling $\overline{\delta}$ to recover the rigid tube, compare Section~\ref{sec:discussion}.
To ensure initial feasibility of this more conservative approach, we decrease the parametric uncertainty by 50\% and increase the prediction horizon to $N=35$ steps for both approaches. The resulting open-loop predictions and closed-loop trajectories can be seen in Figure~\ref{fig:R-CB}.
The rigid tube (red) is significantly larger and correspondingly the planned trajectory cannot pass between the obstacles, but bypasses them from the right.
On the other hand, the proposed homothetic tube MPC formulation finds a feasible solution through the smaller gap between the obstacles, demonstrating improved flexibility.
\begin{figure}[ht]
\centering
\includegraphics[width=0.4\textwidth]{Figures/rigid_tube.eps}
\caption{Comparison of rigid tube MPC (red) and the proposed homothetic tube RMPC (blue). The open-loop solutions at $t=0$ are shown with nominal trajectories (solid) and ellipsoidal over-approximations of the tube based on $\underline{M}$. The closed-loop trajectories are dashed.} \label{fig:R-CB}
\end{figure}
\subsubsection*{Computational complexity}
The offline computations to obtain the CCM and corresponding constants (Alg.~\ref{alg:offline}) took 10 minutes, while the just-in-time compilation (jit, cf.~\cite{andersson2019casadi}) of the RAMPC scheme took another 20 minutes.
The following table lists the online computation times for one iteration in IPOPT for the considered MPC schemes with a prediction horizon of $N=25$ steps. \\
\begin{tabular}{p{13mm}|p{13mm}|p{13mm}|p{12mm}|p{12mm}}
RAMPC & RMPC & RMPC - Euler & Rigid tube & Nominal MPC \\
\hline
60.46 ms & 54.64 ms & 19.52 ms & 8.79 ms & 4.32 ms\\
\end{tabular}
As one can see, adaptation leads to a relatively small increase in computation time (around 10\%) due to the additional variables $\overline{\theta}_t$, $\overline{\delta}_\mathrm{f}$ that appear in constraints \eqref{eq:MPC-dynamics}--\eqref{eq:MPC-theta-in-Theta}, \eqref{eq:MPC-terminal}. We also included an RMPC scheme that uses an explicit Euler discretization for the tube propagation~\eqref{eq:MPC-delta}. This simplification yields a significant reduction in computation times, yet the error in the tube propagation is below 5\% over the full horizon.
Finally, the rigid tube approach has the lowest computational complexity among the RMPC frameworks, due to the absence of the tube dynamics~\eqref{eq:MPC-delta} leading to simpler constraints.\footnote{Among others, the tube dynamics~\eqref{eq:f-delta-def} require the expensive evaluation of $M(z)=W(z)^{-1}$. This inverse can also be avoided in the homothetic tube MPC by using the over-approximation $\overline{M}\succeq M(z)$ (cf. discussion Sec.~\ref{sec:discussion}).}
Compared to a nominal MPC, the additional complexity of the rigid tube MPC is mainly due to the geodesic computation in the initial condition~\eqref{eq:MPC-delta0}.
\section{Conclusion}
\label{sec:conclusion}
We have presented a robust adaptive MPC framework that uses a homothetic tube based on general CCMs and incorporates recursive model updates using set-membership estimation.
We demonstrated the improved flexibility using a nonlinear example.
Open-issues include the consideration of non-parametric model uncertainty and noisy output measurements.
\bibliographystyle{plain}
|
\section{Introduction}
\label{intro}
In this talk, I would like to discuss various aspects of the
phases of QCD at zero temperature, but finite isospin and strangeness density.
However, before I do that, I would like to briefly comment on the
QCD phase diagram as it is normally presented, namely in the $\mu_B$--$T$ plane.
It is shown in Fig.~\ref{fase0}, borrowed from Ref.~\cite{hatsuda}.
Few of the results for the phase diagram are
rigorous in the sense that they are obtained from first principles, rather they
are obtained by model calculations. For asymptotically high temperatures and zero
baryon chemical potential, we know that QCD is in a quark-gluon plasma phase
consisting of weakly interacting deconfined quarks and gluons.
Similarly, we know that at asymptotically high baryon density and zero temperature,
QCD is in the color-flavor locked phase arising from an attracting channel of
one-gluon exchange and the resulting instability of the Fermi surface.
From lattice simulations, we know that there is a cross-over transition for
$\mu_B=0$ at a temperature of around 155 MeV.
For low temperatures and large chemical potentials, the infamous sign problem,
prohibits the use of standard Monte Carlo techniques to study the properties
of QCD. One must therefore resort to low-energy models
such as the NJL model and the quark-meson model. Over the past two decades,
a huge amount of work has been done to map out the phase diagram.
The situation is even more complex than this since, instead of using a common quark chemical
potential for all quarks, one can introduce a separate chemical potential $\mu_f$ for
each flavor. For three flavors, we use either $\mu_u$, $\mu_d$, and $\mu_s$
or the baryon chemical potential $\mu_B$, the isospin chemical potential $\mu_I$,
and strangeness chemical potential $\mu_S$ defined as
\begin{eqnarray}
\mu_B={3\over2}(\mu_u+\mu_d)\;,
\hspace{1cm}
\mu_I={1\over2}(\mu_u-\mu_d)\;,
\hspace{1cm}
\mu_S={1\over2}(\mu_u+\mu_d-2\mu_s)\;.
\end{eqnarray}
For $\mu_B=\mu_S=0$ but nonzero $\mu_I$,
one can carry out Monte Carlo simulations using standard techniques
since the fermion determinant is real in this case and consequently there is no
sign problem. This opens up the possibility to study charged pion condensation
on the lattice and confront it with results from low-energy effective theories.
In this talk, I will discuss pion condensation for two and three
flavors using chiral perturbation theory ($\chi$PT) as a low-energy effective theory
for QCD and show results for the light quark and pion condensates as a function of $\mu_I$ with $\mu_B=\mu_S=0$. The results will be compared
to recent high-precision lattice simulations~\cite{gergo,gergo1,gergo2,gergo3}.
I will also discuss the phase
diagram and meson condensation for three flavors
in the $\mu_I$--$\mu_S$ plane at zero temperature.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.6]{qcd.jpg}
\caption{Phase diagram of QCD in the $\mu_B$--$T$ plane. Figure from Ref.~\cite{hatsuda}.}
\label{fase0}
\end{figure}
In Fig.~\ref{fase}, we sketch the phase diagram of two-flavor QCD in the
the $T$--$\mu_I$ plane. In the lower left region, we have the hadronic phase
where chiral symmetry is broken and quarks are confined.
As the temperature increases, one enters the
quark-gluon plasma phase. Along the $\mu_I$ axis, there is a transition from the
hadronic phase to a Bose-condensed phase of charged pions. In this phase, the
$U(1)_{I_3}$ symmetry is broken giving rise to a massless Goldstone boson, which
is a mixture of $\pi^+$ and $\pi^-$. For large isospin chemical and low temperature,
one expects that quarks are the relevant degrees of freedom rather than pions~\cite{son} .
The Fermi surface that exists when the interactions are turned off, is rendered
unstable once they are turned on, since they are attractive. The system is then described
in terms of loosely bound Cooper pairs instead of tightly bound pions. Since the
symmetry breaking pattern is the same, there is a cross-over transition rather
than a true phase transition between the BEC and the BCS phases~\cite{son}.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.3]{FasediagramInkscape.pdf}
\caption{Phase diagram of QCD in the $\mu_I$--$T$ plane.}
\label{fase}
\end{figure}
\section{$\chi$PT at finite isospin chemical potential $\mu_I$}
We will be using chiral perturbation theory to describe the pion-condensed phase
of QCD at finite $\mu_I$ and zero temperature. $\chi$PT is a low-energy effective theory
for QCD based on the (global) symmetries and degrees of freedom. It provides
a correct model-independent description as long as one is inside its
domain of validity~\cite{weinberg,gasser1,gasser2}. The effective Lagrangian that describes the low-energy degrees of freedom of QCD (pions, kaons, eta) can be written in a low-energy expansion,
\begin{eqnarray}
{\cal L}&=&{\cal L}_2+{\cal L}_4+...\;,
\end{eqnarray}
where the subscript $n$ denoted the order in the expansion. The expansion parameter
can be written as ${M\over4\pi f}$ with $M$ being a relevant mass or momentum scale
and $f$ is the pion-decay constant. The leading-order Lagrangian is
the nonlinear sigma model, which for two flavors reads
\begin{eqnarray}
{\cal L}_{2}&=&{1\over4}f^2\langle\nabla^{\mu} \Sigma^{\dagger} \nabla_{\mu}\Sigma\rangle
+{1\over4}f^2\langle\chi^{\dagger}\Sigma+\Sigma^{\dagger}\chi\rangle\; ,
\end{eqnarray}
where $\langle\rangle$ means trace in flavor space and
\begin{eqnarray}
\Sigma=e^{i{\phi_a\tau_a\over f}}\;
\hspace{2cm}
\chi=2B_0{\rm diag}(m_u,m_d)\;,
\end{eqnarray}
with $\phi_a$ being the Goldstone bosons fields, $m_{u,d}$ are the quark masses, and
$B_0$ is the related to the quark condensate in the vacuum via $\langle\bar{\psi}\psi\rangle=-f^2B_0$.
The covariant derivative is
\begin{eqnarray}
\label{cov}
\nabla_{\mu} \Sigma\equiv
\partial_{\mu}\Sigma-i\left [v_{\mu},\Sigma \right]\;,
\end{eqnarray}
with
$v_{\mu}=\delta_{\mu0}(\mbox{$1\over3$}\mu_B\mathbb{I}+\mbox{$1\over2$}\mu_I\tau_3)$.
Note that the results will be independent of $\mu_B$ since the unit matrix
$\mathbb{I}$ commutes with everything, this reflects that the
mesons have zero baryon number.
The most general ansatz for the (normalized) ground state can after some simplifcations
be written as~\cite{son}
\begin{eqnarray}
\Sigma_{\alpha}&=&
\mathbb{I}\cos\alpha+i\tau_2\sin\alpha
=e^{i\alpha\tau_2}
\;,
\end{eqnarray}
which simply is a rotation in flavor space of the vacuum state $\mathbb{I}$
by an angle $\alpha$. The leading-order thermodynamic potential is
\begin{eqnarray}
\Omega_0&=&-f^2B_0(m_u+m_d)\cos\alpha-{1\over2}f^2\mu_I^2\sin^2\alpha\;.
\label{oo}
\end{eqnarray}
We see that there is a competition between the two terms in Eq.~(\ref{oo}), the first term prefers the vacuum state $\alpha=0$, while the second term prefers $\alpha={1\over2}{\pi}$.
The optimum is found by balancing these two terms in the
thermodynamic potential,
\begin{eqnarray}
\cos\alpha&=&{m_{\pi,0}^2\over\mu_I^2}\;,
\hspace{1cm}\mu_I^2\geq m_{\pi,0}^2\;,
\\
\alpha&=&0\;,\hspace{1cm}\mu_I^2<m_{\pi,0}^2
\;,
\end{eqnarray}
where $m_{\pi,0}^2=B_0(m_u+m_d)$ is the tree-level pion mass.
Thus there is phase transition from the vacuum to a pion-condensed phase
at a critical $\mu_I$, $\mu_I^c=\pm m_{\pi,0}$. To determine the order of the transition,
one can construct a Ginzburg-Landau energy functional by expanding
the thermodynamic potential $\Omega_0$ around $\alpha=0$,
\begin{eqnarray}
\Omega_0&=&-f^2m_{\pi,0}^2+{1\over2}f^2[m_{\pi,0}^2-\mu_I^2]\alpha^2-{1\over24}f^2
\left[m_{\pi,0}^2-4\mu_I^2
\right]\alpha^4+{\cal O}(\alpha^6)\;.
\end{eqnarray}
We define a critical chemical isospin potential $\mu_I^c$
when the order-$\alpha^2$ term
vanishes, i.e. $\mu_I^c=\pm m_{\pi,0}$. Since the coefficient of
the order-$\alpha^4$ term is positive for $\mu_I=\mu_I^c$, the transition is second order. Note that all thermodynamic quantities are independent of $\mu_I$
for all $\mu_I^2<m_{\pi,0}^2$ implying that e.g. the isospin density vanishes in the same region
and not only for $\mu_I=0$. This is an example of the Silver-Blaze property~\cite{cohen0}.
\section{Phase diagram for three flavors}
We next consider the phase diagram for three-flavor QCD at finite $\mu_I$ and $\mu_S$
including electromagnetic effects.
If we couple $\chi$PT to dynamical photons, the Lagrangian contains a few
extra terms at leading order in the low-energy expansion~\cite{gasser89},
\begin{eqnarray}
\nonumber
{\cal L}_2&=&-{1\over4}F_{\mu\nu}F^{\mu\nu}+
{1\over4}f^2\langle\nabla_{\mu}\Sigma\nabla^{\mu}\Sigma^{\dagger}
\rangle
+{1\over4}f^2
\langle\chi^{\dagger}\Sigma+\Sigma^{\dagger}
\chi\rangle
+C\langle Q\Sigma Q\Sigma^{\dagger}\rangle
\\ &&
+{\cal L}_{\rm gf}+{\cal L}_{\rm ghost}
\;,
\label{lagel}
\end{eqnarray}
with $\chi=2B_0{\rm diag}(m_u,m_d,m_s)$ and $\Sigma=e^{i{\phi_a\lambda_a\over f}}$.
The new term $C\langle Q\Sigma Q\Sigma^{\dagger}\rangle$ is responsible for the
tree-level mass splitting between the charged and neutral pions, and it also contributes
to the tree-level mass splitting between the charged and neutral kaons.
The inclusion of electromagnetic effects also implies that the phases with charged
meson condensates are superconductors and that the massless degree of freedom
(the Goldstone bosons) is eaten up by the now massive photon via the Higgs mechanism.
The term $v_{\mu}$ in the covariant derivative Eq.~(\ref{cov}) is replaced by
\begin{eqnarray}
v_0&=&\mbox{$1\over3$}(\mu_B-\mu_S)\mathbb{I}+\mbox{$1\over2$}\mu_{K^{\pm}}
\lambda_Q+{1\over2}\mu_{K^0}\lambda_K\;,
\hspace{1cm}
v_i=0
\;,
\end{eqnarray}
where
\begin{eqnarray}
\mu_{K^{\pm}}&=&{1\over2}\mu_I+\mu_S\;,
\hspace{2cm}
\mu_{K^0}=-{1\over2}\mu_I+\mu_S\;,\\
\lambda_Q&=&\lambda_3+{1\over\sqrt{3}}\lambda_8\;,
\hspace{1.8cm}
\lambda_K=-\lambda_3+{1\over\sqrt{3}}\lambda_8\;.
\end{eqnarray}
In analogy with the two-flavor case, we expect onset of charged kaon condensation
when $\mu_{K^{\pm}}^2=m_{K^{\pm}}^2$ and neutral kaon condensation when
$\mu_{K^{0}}^2=m_{K^{0}}^2$.
The corresponding ans\"atze for the ground states are~\footnote{One could imagine
multiple condensates in parts of the $\mu_I$--$\mu_S$ plane, but that is ruled
out by actual calculations. The angles $\beta$ and $\gamma$ are the rotation angles
of the quark condensate into a charged or neutral kaon condensate, respectively.}
\begin{eqnarray}
\Sigma_{\beta}=e^{i\beta\lambda_5}\;,
\hspace{3cm}
\Sigma_{\gamma}=e^{i\gamma\lambda_7}\;.
\end{eqnarray}
The thermodynamic potential in the different phases can then be computed
as functions of the chemical potentials.
For example, in the charged kaon condensed phase, the thermodynamic potential is
\begin{eqnarray}
\nonumber
\Omega_0&=&
-f^2B_0(m_u+m_s)\cos\alpha
-{1\over2}f^2
\left[\mu_{K^{\pm}}^2-\Delta m^2_{\rm EM} \right]\sin^2\alpha
\;,
\end{eqnarray}
where $\Delta m^2_{\rm EM}={2Ce^2\over f^2}$ is the splitting between the
charged and neutral kaons due to electromagnetism. It follows that
the transition takes place exactly at $\mu_{K^{\pm}}^2=m_{K^{\pm}}^2=B_0(m_u+m_d)+{2Ce^2\over f^2}$ as expected.
For each value of $(\mu_I,\mu_S)$, we find the
phase with the lowest value of $\Omega_0$ (or largest pressure).
This phase wins and we can map out the phase diagram
in this manner. The result is shown in the left panel of Fig.~\ref{diagram}.
The black lines are the transition lines without electromagnetic
interactions and the red lines are with electromagnetic interactions.
The former was first obtained by Kogut and Toublan~\cite{kogut33} in the isospin limit.
The phases with charged meson condensation become Higgs phases upon including
electromagnetic effects. The tree-level mass of the photon is $m_A=ef\sin\alpha$.
The transitions from the
normal phase to a meson-condensed phase is always second order
with mean-field exponents in the $O(2)$ universality class. The transitions
between the various condensed phases are always first order
and involve the competition between the order parameters
of the different phases. As we cross the transition lines, the order parameters as well as the isospin and strangeness densities, $n_I$ and $n_S$
jump discontinuously.
The small offset of the dashed vertical lines is due to
the mass difference between the charged and neutral
kaons, which is both due to $\Delta m_\text{EM}\neq0$, and $m_u\neq m_d$.
These contributions, however, pull in opposite directions,
as we see in the phase diagram. The contribution due to the difference in
quark masses adds to the mass of the $K^0/\bar{K}^0$ meson, which is why the black transition
line between the kaon condensate is to the left of the $\mu_I = 0$ line,
while the electromagnetic contribution adds to the mass of the charged kaon,
which is why the red line is between these two lines.
The partition function in the normal phase is independent of the two chemical
potentials $\mu_I$ and $\mu_S$, which again is the Silver Blaze
property~\cite{cohen0}.
In the right panels, we have zoomed in on the triple points.
Upper panel shows the intersection of the normal, neutral kaon condensed and charged kaon
condensed phases, while the lower shows the intersection of the normal, pion condensed, and charged kaon condensed phases.
\begin{figure}[htb!]
\centering
\includegraphics[width=\columnwidth]{phase_diagram_EM.pdf}
\caption{Left panel shows the phase diagram as predicted by $\chi$PT
in the $\mu_I$--$\mu_S$-plane.
In the right panel, we have zoomed in on the triple points. See main text for details.
Fig. from Ref.~\cite{johnsrud}.}
\label{diagram}
\end{figure}
\section{${\cal O}(p^4)$ calculation of thermodynamic potential}
I will next sketch the NLO calculation of the thermodynamic potential. For simplicity
I consider two flavors in the isospin limit, $m_u=m_d=m$.
The thermodynamic potential can be calculated in a low-energy expansion,
\begin{eqnarray}\nonumber
\Omega&=&\Omega_0+\Omega_1+...,
\end{eqnarray}
where $\Omega_n$ is the order-${\cal O}(p^{2n+2})$ contribution.
The term $\Omega_1$ receives contributions from the one-loop graphs of
${\cal L}_2^{\rm quadratic}$ and counterterms coming from
${\cal L}_4^{\rm static}$.
The relevant terms are
\begin{eqnarray}
\nonumber
\mathcal{L}_2^{\text{quadratic}
&=&\mbox{$1\over2$}
(\partial_{\mu}\phi_{a})
(\partial^{\mu}\phi_{a})+\mu_{I}\cos\alpha(\phi_{1}\partial_{0}\phi_{2}-\phi_{2}
\partial_{0}\phi_{1})\\ \nonumber
&&-\mbox{$1\over2$}\left [(m^{2}_{\pi,0}\cos\alpha-\mu_{I}^{2}\cos2\alpha)\phi_{1}^{2}
\right.
+(m^{2}_{\pi,0}\cos\alpha-\mu_{I}^{2}\cos^{2}\alpha)\phi_{2}^{2}\\
&&\left.+(m^{2}_{\pi,0}\cos\alpha+\mu_{I}^{2}\sin^{2}\alpha)\phi_{3}^{2}\right]
\label{l2}
\;,\\ \nonumber
\mathcal{L}_{4}^{\text{static}}&=&(l_{1}+l_{2})\mu_{I}^{4}\sin^4\alpha
+l_{4}m^{2}_{\pi,0}\mu_{I}^{2}\cos\alpha\sin^{2}\alpha
+(l_{3}+l_{4})m^{4}_{\pi,0}\cos^{2}\alpha
+h_1m_{\pi,0}^4\;,
\\ &&
\;,
\label{l4}
\end{eqnarray}
where $l_1$--$l_4$ and $h_1$ are bare couplings.
They are related to the renormalized couplings ${l_i^r}$ and $h_i^r$
via
$l_i=l_i^r(\Lambda)+{\gamma_i\Lambda^{-2\epsilon}\over2(4\pi)^2}\left[{1\over\epsilon}+1\right]$ and
$h_i=h_i^r(\Lambda)+{\delta_i\Lambda^{-2\epsilon}\over2(4\pi)^2}\left[{1\over\epsilon}+1\right]$, where
$\gamma_i$, $\delta_i$ are coefficients and $\Lambda$ is the renormalization scale
in the $\overline{\rm MS}$ scheme. Since $\delta_1=0$, $h_1=h_1^r$ and does not run.
Performing the Gaussian integral over the quantum fields $\phi_a$
in dimensionsal regularization using
Eq.~(\ref{l2}), we obtain a divergent contribution to $\Omega_1$.
The divergences are cancelled by adding the quartic terms from Eq.~(\ref{l4})
and renormalizing the couplings by making replacing the bare couplings with the
renormalized ones.
The final result is
\begin{eqnarray}\nonumber
\Omega_0+\Omega_1&=&-f^{2}m^{2}_{\pi,0}\cos\alpha-\frac{1}{2}f^{2}\mu_{I}^{2}
\sin^{2}\alpha
\\ && \nonumber
-\frac{1}{4(4\pi)^{2}}\left [\frac{3}{2}-\bar{l}_{3}
+4\bar{l}_{4}
+\log\left({m_{\pi,0}^2\over \tilde{m}_2^2}\right)
+ 2\log\left({m_{\pi,0}^2\over m_3^2}\right)
\right ]
m_{\pi,0}^{4}\cos^{2}\alpha
\\&& \nonumber
-\frac{1}{(4\pi)^{2}}
\left [{1\over2}+\bar{l}_{4}
+ \log\left({m_{\pi,0}^2\over m_3^2}\right)
\right ]
m^{2}_{\pi,0}\mu_{I}^{2}\cos\alpha\sin^{2}\alpha
\\ &&
\nonumber
-\frac{1}{4(4\pi)^{2}}\left [1
+\frac{2}{3}\bar{l}_{1}+\frac{4}{3}\bar{l}_{2}
+ 2\log\left({m^2_{\pi,0}\over m_3^2}\right)
\right ]\mu_{I}^{4}
\sin^{4}\alpha
\\ &&
-{1\over(4\pi)^2}\bar{h}_1m_{\pi,0}^4
+V_{{1},\pi^+}^{\text{fin}}+V_{{1},\pi^-}^{\text{fin}}
\;,
\label{effpotnlo}
\end{eqnarray}
where $\tilde{m}_2^2=m_{\pi,0}^2\cos\alpha$,
$m_3^2=m_{\pi,0}^2+\mu_I^2\sin^2\alpha$, and
$V_{{1},\pi^+}^{\text{fin}}+V_{{1},\pi^-}^{\text{fin}}$ are two complicated
finite terms that must be evaluated numerically.
Finally, $\bar{l}_i$ and $\bar{h}_i$ are, up to a prefactor, equal to $l_i^r$
and $h_i^r$ at the scale $\Lambda=m_{\pi,0}$.
Using Eq.~(\ref{effpotnlo}) one can show that the phase transition takes place
at $\mu_I^c=m_{\pi}$, where the physical pion mass now includes
radiative corrections~\cite{gasser1}, see Eq.~(\ref{mpi}) below.
The parameters $\bar{l}_i$ are determined by experiment and $\bar{h}_1$
estimated by model calculations.
and the parameters $m_{\pi,0}^2=2B_0m$ and $f$ can be found by inverting the one-loop
relations using the experimental values for the pion mass and the pion decay constant,
\begin{eqnarray}
\label{mpi}
m_{\pi}^2=m^2_{\pi,0}\left[1-{m^2_{\pi,0}\over2(4\pi)^2f^2}\bar{l}_3\right]\;,
\hspace{1cm}
f_{\pi}^2
=f^2\left[1+{2m^2_{\pi,0}\over(4\pi)^2f^2}\bar{l}_4\right]\;.
\end{eqnarray}
\section{Condensates}
In order to obtain the light quark and pion condensates, we need to calculate the
thermodynamic potential $\Omega$ with sources $m$ and $j$, where the latter is a
pionic source.
The former has already been
included in the calculations I have shown and it is also straightforward to
include a pionic source $j$ in the calculations. For example, in the two-flavor
expression for the thermodynamic potential, one simply makes the replacements
$m\cos\alpha\rightarrow m\cos\alpha+j\sin\alpha$
and
$\bar{h}_1m_{\pi,0}^4=\bar{h}_1(2B_0m)^2\rightarrow\bar{h}_1[(2B_0m)^2+(2B_0j)^2]$~\cite{martin}.
Once these replacements are made, the condensates are given by
\begin{eqnarray}
\label{c1}
\langle \bar{\psi}\psi\rangle_{\mu_I}
={1\over2}{\partial \Omega\over \partial m}
=-f^2B_0\cos\alpha+...\;,
\hspace{1cm}
\langle \pi^+\rangle_{\mu_I}={1\over2}{\partial \Omega\over \partial j}
=-f^2B_0\sin\alpha+...\;,
\label{c2}
\end{eqnarray}
where I on the the right-hand side have written explicitly the tree-level contributions.
The subscript $\mu_I$ on the expectation values indicate that they depend on the
chemical potential. Instead of plotting the condensates directly, we define
the normalized deviations as
\begin{eqnarray}
\label{deviation}
\Sigma_{\bar{\psi}\psi}=-\frac{2m}{m_{\pi}^{2}f_{\pi}^{2}}\left[\langle\bar{\psi}\psi\rangle_{\mu_{I}}^{}-\langle\bar{\psi}\psi\rangle_{0}^{j=0}\right]+1\;,
\hspace{1cm}
\label{deviation2}
\Sigma_{\pi}=-\frac{2m}{m_{\pi}^{2}f_{\pi}^{2}}\langle\pi^{+}\rangle^{}_{\mu_{I}}\;.
\end{eqnarray}
At tree level, Eq.~(\ref{c2}) shows the rotation of the
quark condensate into a pion condensate.
Equivalently, from Eq.~(\ref{deviation2}), the deviations at tree level satisfy
$\Sigma_{\bar{\psi}\psi,\rm tree}+\Sigma^2_{\pi,\rm tree}=1$.
This interpretation no longer holds beyond ${\cal O}(p^2)$.
In the left panel of Fig.~\ref{condie}, we show $\Sigma_{\bar{\psi}\psi}$
as a function of ${\mu_I\over m_{\pi}}$ at leading order (red line)~\footnote{The leading order result is the same for two and three flavors.}
and next-to-leading order
for two flavors (blue line) and three flavor (green line).
The data points are from the lattice simulations of Ref.~\cite{gergo,gergo1,gergo2,gergo3}.
In the right panel, we show $\Sigma_{\pi}$ in the same approximations.
We note that the difference between $\Sigma_{\bar{\psi}\psi}$ in the various approximations is very
small and they all agree very well with the lattice data points. Regarding
$\Sigma_{\pi}$, we notice that it is nonzero
for $\mu_I<m_{\pi}$, which simply reflects that the curves shown are for nonzero
pion source, $j=0.00517054m_{\pi}$
The $U(1)_{I_3}$-symmetry is therefore broken explicitly
for all values of $\mu_I$. Comparing the various approximations and lattice data,
it is evident that including the ${\cal O}(p^4)$ corrections results in a substantially better agreement between $\chi$PT and simulations.
All the numerical results have been obtained by using the same physical meson masses
as well as $f_{\pi}$
as in the lattice simulations. This requires that we invert the relations
Eq.~(\ref{mpi}) to obtain the values for the bare mass $m$ and the bare pion decay constant $f$ using the experimental values of $\bar{l}_i$.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.5]{ccj1.pdf}
\includegraphics[scale=0.5]{pcj1.pdf}
\caption{$\Sigma_{\bar{\psi}\psi}$ (left panel) and $\Sigma_{\pi}$ (right panel) as functions of
$\mu_I/m_{\pi}$ at zero temperature and finite source $j=0.00517054m_{\pi}$. Fig. taken from Ref.~\cite{martin}.}
\label{condie}
\end{figure}
\section{Acknowledgements}
I would like to thank Prabal Adhikari, Martin Mojahed, and Martin Kj{\o}llesdal
Johnsrud for collaboration. I would also like to thank the organizers of
the XVth Quark confinement and the Hadron spectrum conference for a very interesting meeting.
|
\section{Introduction}
Pulsars were first discovered in the 1960s, and knowledge about their nature and physical properties has been growing relentlessly ever since. In particular, the last two decades have witnessed remarkable progress both in theoretical understanding and observational capabilities. However, despite great advances, some key aspects are yet to be fully understood. The neutron star (NS) interior compositions and equations of state, the magnetic field configuration, and the structure of the surrounding plasma remain fundamental open questions in pulsar astrophysics.
Rotation-powered millisecond pulsars (MSPs) are a peculiar class of pulsars with highly stable spin rates. Their periods are around $1-30\,$ms, and their spin-down rates are of the order of $\dot{P} \sim 10^{-20}$. According to the usual magnetic braking model, they are old objects with characteristic ages $\tau_c = P/2\dot{P} \sim 0.1 - 10\,$ Gyr, and they have relatively low surface magnetic fields $B \sim 10^{8-9}\,$ G. Outside of the radio band, the greatest amount of information about these objects can be found in the soft X-ray range (∼0.1–10 keV). The emission in this range is mainly of thermal origin, although non-thermal components (usually modeled by power laws) often exist as well. The thermal component is produced in small hot regions on the surface of the star, commonly called ``hot spots''. These regions are arguably heated to temperatures up to $T \sim 10^6\,$ K by the constant bombardment of particles that accelerate in magnetospheric gaps up to very high energies \cite{ruderman1975}.
A good method for indirectly measuring several important properties of MSPs is the precise observation and modelling of their X-ray light curves \cite{pechenick1983hot, strohmayer1992light, page1994, miller1998effects, braje2000light, harding2001, harding2002, beloborodov2002gravitational, cadeau2007light, morsink2007oblate, zavlin2007studying, bogdanov2007, bogdanov2008, ozel2012, psaltis2014pulse, psaltis2014prospects, miller2015determining, nattila2018radiation, baubock2019, lockhart2019x, bilous2019, riley2019, bogdanov2019constrainingI, bogdanov2019constrainingII, miller2019, salmi2020, chen2020, kalapotharakos2021}. MSPs are especially suited for this approach because of their highly stable spin rates and pulse shapes.
These techniques have been applied, with increasing level of refinement, for four decades already. However, despite the sophistication achieved, pulsar X-ray light curve fitting continues to be a highly nontrivial task, and the problem is in general quite degenerate unless very good observational data is present.
One important application of pulsar pulse profile modelling to the inference of physical properties is in constraining their mass-radius relations and so the fundamental equation of state. The latter closes the system of stellar structure equations: it is therefore a necessary ingredient to have the solution of the star and the spacetime around it, once the central energy density and the rotation rate have been fixed. In this manner, a mapping is obtained which translates equations of state into mass-radius relations. Given that thermal X rays are emitted over the surface of the star, the stellar compactness has a strong impact on the luminosity profiles through the gravitational field in which the rays propagate. It is for that reason that pulse profile modelling can be used to pose constraints on compactness, and, via the inverse mapping, on the equations of state. Many recent theoretical and observational efforts have been directed towards that end. Likewise, the technique can be used to determine the geometry of the magnetic field close to the surface, because the location and shape of the hot spots depend on it, which in turn determines the observed luminosity profiles.
As mentioned before, the pulse profiles are subject to various physical effects. The impact of the gravitational field occurs through the deflection of light rays, and the gravitational redshift of the photon frequencies. Additionally, the spin of the star also affects the frequency of emitted light through local Doppler boosting according to the tangential speed of the emitting portion of the surface. These effects need to be taken into account by relativistic ray tracing and radiative transport calculations. For rapidly rotating neutron stars, the gravitational field has to be numerically computed as part of the solution of the stellar structure equations. However, for rotation frequencies below 400 Hz, the Kerr spacetime is a sufficient approximation of the exterior gravitational field of a NS \cite{miller1998effects, braje2000light}.
The Neutron Star Interior Composition Explorer (NICER; \cite{gendreau2016neutron}) began operating in June 2017, and has since measured the soft X-ray light curves of a few nearby rotation-powered MSPs with unprecedented sensitivity (e.g., \cite{bogdanov2019constrainingI, guillot2019nicer}). One of its main purposes is, precisely, to obtain high quality data for inferring the mass and radii of the observed neutron stars, and the geometry of their surface magnetic fields, via pulse profile modelling.
A Bayesian modelling of the X-ray light curve has been performed for PSR J0030+0451 \cite{miller2019, riley2019, bilous2019}, one of the pulsars observed by NICER. They constructed multi-parametric families of models with two circular or annular hot spots of uniform temperature, allowing for non-antipodal configurations. Then, they did a statistical sampling over the parameters of the families, searching for the best fits to the light curve. They found that the sampling favoured the non-antipodal configurations, suggesting that the surface magnetic field of PSR J0030+0451 should have a strong non-dipolar component. Posterior works \cite{chen2020, kalapotharakos2021} searched for vacuum magnetic field configurations such that the foot-prints on the stellar surface of those
field lines crossing the light cylinder would closely resemble the hot spots suggested by the observations; this included offset dipole-plus-quadrupole components. Then, they seeded these vacuum fields to force-free simulations to obtain the polar caps, defined as the regions on the surface where the magnetic field lines are open.
Using this information on the global magnetospheric structure, and assuming a uniform temperature over these polar caps, they were able to construct multi-wavelength emission models, including radio and $\gamma$-rays, that were consistent --to some extent-- with observations.
In this paper, we pursue a slightly different route. We start from general relativistic force-free (GRFF) simulations of a rotating NS, endowed with a standard centered dipolar magnetic field configuration (as done in \cite{Pulsar}); and then, we link the resulting electric currents on the surrounding plasma with the surface thermal emission from the star, essentially following the model proposed in \cite{lockhart2019x}. Once we have the emissivity map from the NS surface, we perform the relativistic ray tracing with the recently developed code {\sc Skylight}~\cite{Skylight}, to compute the corresponding light curves and spectra.
We observe that the additional input from the GRFF simulations in modelling the emissions results in more complex ERs than the traditional circular polar caps associated to the centered dipolar magnetic field. Here, we explore the possibility that such new structures for the emitting regions could help in accounting for the observational X-ray data of MSPs, without the need
of invoking strong non-dipolar magnetic field components and/or offsets from the NS center. We concentrate our exploration on a set of four target pulsars, PSR J0437–4715, PSR J1231−1411, PSR J2124−3358 and PSR J0030+0451, for which there are good quality X-ray data publicly available (see \cite{bogdanov2019constrainingI} and references therein). We find that our approach allow to reproduce the observations with surprising accuracy.
The article is organized as follows. In Section \ref{sec:setup} we describe the methods employed, from the modelling of the pulsar magnetosphere up to the fitting of our modelled light curves and spectra to the observational data. In Section \ref{sec:results},
we first introduce the emission regions derived from the GRFF simulations, and then present and analyze our best-fits to the X-ray data of the four target MSPs. Additional information regarding the fitting procedure and relevant posterior probabilities distributions can be found in Appendix \ref{sec:appendix}. Finally, we summarize our work and conclude in Section \ref{sec:conclusions}.
\section{Methods}\label{sec:setup}
We shall tackle the problem by four separate instances. First, we numerically solve the magnetosphere surrounding a rotating pulsar endowed with a simple dipolar magnetic field. We use the force-free approximation, which provides with a global solution for the electromagnetic field. Then, we model a temperature map at the stellar surface,
by linking the magnetospheric currents to the energy deposited at the surface (or in the cm-thick atmosphere) by the bombardment of relativistic particles.
We then perform the ray tracing to simulate the detection of emitted photon by different distant observers and compute their corresponding light curves and spectra. Finally, we compare our models with the observations by NICER and XMM-Newton, searching for those configurations that better reproduce the X-rays light curves and spectra.
\subsection{Pulsar magnetosphere}
We employ the numerical code {\sc Onion} to solve for the exterior force-free (FF) magnetosphere surrounding a neutron star. This code has been used to model several astrophysical scenarios involving black holes and neutron star magnetospheres (see \cite{FFE2, Pulsar, Boost, carrasco2019, carrasco2020, carrasco2021}) and, in particular, it has already been applied to study misaligned pulsars in general relativity \cite{Pulsar}. Hence, we briefly describe here the most relevant aspects of our setup for the purposes of the present work, while referring the reader to \cite{FFE2, Pulsar} for technical details about the numerical implementation.
We approximate the exterior of a rotating neutron star by the Kerr metric, parametrized by the mass $M$ and spin $a$.
Assuming a spherically symmetric neutron star of radius $R$,
we fix the dimensionless moment of inertia $\mathcal{I} := I/MR^2$ to the value $2/5$ (as in, e.g., \cite{petri2015general, gralla2016pulsar}).
This allow to relate the star angular velocity $\Omega$ with the spin parameter in units of the stellar radius, namely:
\begin{equation}\label{eq:spin}
\frac{a}{R}=\frac{2}{5} \, \Omega \, R
\end{equation}
Note that realistic stars are expected to have a slightly lower value $\mathcal{I}$, since the mass is concentrated towards the center. However, the value $2/5$ remains a good approximation within the range of typical NS compactness \cite{bejger02}.
In the present work, we consider a standard (centered) magnetic dipole field configuration, as we would like to test, in our model, the need for significant higher-multipoles components and/or off-centered fields to account for (some of) the observed light curves by NICER.
Within this setup, there are three dimensionless parameters which completely characterize a given GRFF simulation, namely: (i) the misalignment angle $\chi$ between the magnetic moment and the rotational axis of the pulsar; (ii) the surface rotation velocity $v_s := \frac{R}{R_{LC}} $, being $R_{LC}$ the light-cylinder radius; and (iii) the stellar compactness,
\begin{equation}\label{eq:compactness}
\mathcal{C} := \frac{M}{R} \left(\frac{G}{c^2} \right)
\end{equation}
where $G$ is Newton's constant and $c$ the speed of light.
Notice that, once these dimensionless quantities are fixed, choosing the physical value of one of the three parameters (i.e., $R$, $M$ or $\Omega$) completely determine the other two.
On the other hand, we note that the magnetic field strength can be freely re-scaled\footnote{This is a well known property of the FF equations (e.g., \cite{komissarov2002}).} and, thus, can be given \textit{a-posteriori}.
Meaning, in practice, that we can use the same FF run to represent pulsars of different magnetic field strengths, whose values are inferred from the accurate measurements of the timing properties (spin period $P$ and its time derivative $\dot{P}$), via the popular --and rough-- vacuum spin-down formula estimate, $B_{\rm pole} [G]=6.4 \times 10^{19} (P[s] \dot{P})^{1/2}$ (where $B_{\rm pole}$ is the surface magnetic field strength at the pole).
For the present study, we perform an initial set of 25 GRFF simulations at fixed rate $v_s = 0.05$ (which can represent, for instance, a pulsar with a $4.8\,$ms period and NS radius of $11.5\,$km), exploring various misalignments $\chi= \{ 15\degree, 30\degree, 45\degree, 60\degree, 75\degree \}$ and stellar compactness $\mathcal{C} = \{ 0.15, 0.18, 0.2, 0.22, 0.25 \}$ (which correspond to masses $M \sim \{ 1.2, 1.4, 1.6, 1.7, 1.9 \} \, M_{\odot}$ in the above example).
We solve the system in a region between the NS surface at radius $R$, and an exterior spherical surface located at $64 R \gtrsim 3 R_{\rm LC}$.
The domain is represented by a total of $6 \times 6$ sub-domains, with $6$ patches to cover the angular directions, times $6$ spherical shells expanding in radius. These spherical shells do not overlap each other, and have more resolution in the inner regions: from layer to layer, the radial resolution is decreased by a factor $2$.
We adopt a resolution with grid numbers $N_{\theta} \times N_{\phi} \times N_{r}$ with $N_{\phi} = 2 N_{\theta} = 320$, while $N_{r} = 360$ to span the whole computational domain.
In all cases, we evolve the system until it relaxes and reaches a stationary state (i.e., after roughly two stellar rotations) before extracting the FF field configuration. In particular, we look at the relevant information contained in the FF charge and current densities\footnote{As standard in FF electrodynamics, one can always recover partial information about the plasma from its four-current $J^a$, by means of the co-variant Maxwell equations, i.e.: $J^a = \nabla_b F^{ab}$.}, $\rho$ and $\vec{j}$, as seen by a co-moving observer rotating with the NS.
\subsection{Emission model} \label{sec:emission_model}
It is generally accepted that the thermal surface emission in millisecond pulsars arises from bombardment of relativistic particles hitting at the stellar surface. Such energetic particles are assumed to originate within the magnetosphere at the so-called \textit{gaps} where $E\cdot B \neq 0$ (i.e., regions where the force-free condition breaks-down). We shall follow here the approach presented in \cite{lockhart2019x},
in which pair-creation and accelerations are assumed to take place along a subset of the magnetospheric currents where the four-current $J^a$ is spacelike. The argument supporting this choice is that spacelike currents develop high voltage drops with counter-streaming flow of electrons and positrons, thus favoring intense bursts of pair-production \cite{timokhin2013, harding2022}.
Once the emitting region at the stellar surface is established by the above condition (i.e., $J_a J^a > 0$), the kinetic energy deposition is estimated by assuming that a fixed fraction $ \kappa $ of the spatial current $|\vec{j}|$ (defined in the local co-moving frame of the NS) is carried by relativistic particles with averaged Lorentz factor $\bar{\gamma}$, traveling towards the stellar surface (see \cite{baubock2019} for a detailed discussion). The rate at which this energy is deposited is balanced to the power radiated as a black body, yielding the temperature:
\begin{equation}\label{eq:Teff}
T = \left( \frac{c^2 m_e \kappa (\bar{\gamma}-1)}{e \sigma} \right)^{1/4} |\vec{j}|^{1/4}
\end{equation}
where $e$, $m_e$ and $\sigma$ represent the electron charge, electron mass and the Stefan-Boltzmann constant, respectively.
The factor $\kappa$ relates to the pair multiplicity $\mathcal{M}$ through $\kappa = \zeta \mathcal{M}$, being $\zeta$ the ratio of in-going to total pairs \cite{salmi2020}.
We note that since we are assuming relativistic particles, $\kappa (\bar{\gamma} -1) \approx \kappa \bar{\gamma} $, and it acts as a single effective parameter controlling the local temperature in the model.
Both $\kappa$ and $\bar{\gamma}$ (and, therefore, their product $\kappa \bar{\gamma}$), are largely unconstrained by the current theoretical understanding.
In order to convert this effective temperature to a model for the emerging specific intensity of the radiation field, one needs a detailed calculation of the thermal structure of the NS atmosphere, as it was done for instance in \cite{baubock2019, salmi2020}. Instead, we use a simplified approximation of blackbody emission at the effective temperature $T_{eff}$ (e.g., \cite{lockhart2019x}),
\begin{equation}
I_{\nu} = \frac{\nu^3 \, \mathcal{B}(\Theta)}{\exp [\nu/T_{eff}] -1 } \label{eq:bb}
\end{equation}
where $\nu$ is the frequency of the electromagnetic radiation and we have considered a non-isotropic beaming distribution given by,
\begin{equation}
\mathcal{B}(\Theta) = \cos^b (\Theta)
\end{equation}
where $b$ is the anisotropy index and $\Theta$ the angle between the photon propagation direction and the surface normal. In previous works (e.g., \cite{kalapotharakos2021}), exponents in the range $b \sim 0.5-1.0$ have been used as a decent approximation to the NS atmosphere. This assumption is known to be needed to explain high pulsed fractions observed in the thermal X-ray light curves of many pulsars.
Note that the simple analytical and energy-independent beaming correction is highly simplified, compared to the (energy-dependent) anisotropy obtained by detailed radiative models \cite{ozel13} (their Fig. 5, in particular). Moreover, besides introducing anisotropy, the atmosphere modifies the spectrum, so that, when it is modeled by a blackbody with an effective temperature $T_{eff}$ like in \eqref{eq:bb}, the latter looks higher than the physical temperature $T$. The ratio $T_{eff}/T$, called color-correction factor, is of the order $\sim 2$ for light-elements atmosphere \cite{sulemainov11}. Such factor, needed to convert the value of $T$, eq.~\eqref{eq:Teff}, to $T_{eff}$, is not explicitly indicated here, but it is implicitly reabsorbed in the free effective parameter $\kappa \bar\gamma$.
We will discuss the limitations of the simple emission model adopted here in Sections \ref{sec:normalization} and \ref{sec:conclusions}.
\subsection{Ray tracing and radiative transfer}
To transport the emission from the stellar surface to the observers, and thereby calculate light curves and spectra, we use the general-relativistic ray tracing and radiative transfer code {\sc Skylight}~\cite{Skylight}, which supports arbitrary spacetime geometries. We adopt the full Kerr spacetime, with its spin and mass linked to the NS properties through equations \eqref{eq:spin}-\eqref{eq:compactness}. For this work, we use the observer-to-emitter scheme of the code, where we set a virtual image plane at the observer, and we take a bundle of vectors perpendicular to the image plane as initial tangent vectors to the geodesics. Once the initial data is set, the geodesic equations are integrated backwards in time, until the trajectories intersect the stellar surface.
Then, for each observer we obtain a mapping from the image plane to the stellar surface, including the angles that the rays make with the surface normal. Since the spacetime is stationary, these mappings contain all the necessary information to compute the observed time-dependent fluxes. We compute these mappings for a single observer time. Then, for different times, we rotate the stellar surface to the corresponding phase and identify the rays intersecting the hot spots. For each ray, we consider the gravitational redshift and the Doppler boosting to the co-rotating frame of the star to calculate the specific intensity at the camera from the local emission model. Finally, we integrate over the image plane to obtain light curves and spectra.
In all the ray tracing simulations, the image planes are located at $750 r_g$, where $r_g = GM/c^2$, and we take grids of $600 \times 600$ rays over the image planes. These values are appropriate for our purposes, since locating the image plane further away, or increasing the resolution, do not significantly affect the results. For each pulsar, the simulations take the stellar compactness, $\mathcal{C}$, the spin parameter, $a$, and the viewing angle, $\xi$, as inputs: $\mathcal{C}$ and $a$ determine the spacetime geometry (we take units such that the stellar radius is fixed), and $\xi$ determines the position and orientation of the image plane. The rest of the parameters, as those listed in Table~\ref{table1}, only enter the calculation at the end, when retrieving the physical units of the calculated fluxes.
\subsection{Searching for the best-fits} \label{sec:search}
We want to find the set of parameters in our model that produces the best-fit to the observational data for the four target millisecond pulsars in \cite{bogdanov2019constrainingI}.
Ideally, one should directly compare the full phase-resolved spectral information against the observational data, as done in \cite{riley2019} for PSR J0030+0451. Our main aim here is to reproduce the light curves, for which NICER provides high-accuracy data. At the same time, we shall vary the parameters of our idealized emission model to reproduce the spectral shape as well.
We proceed as follows: we first estimate an appropriate value for the parameter $\kappa \bar{\gamma}$, which controls the effective temperature of the emitting regions in our model, by looking at the phase-averaged spectra. Then, we vary the remaining parameters --which only affect the spectrum very weakly-- in order to search for the light curve best-fits.
Once an approximated value for $\kappa \bar{\gamma}$ is found, it is fixed. Next, for each compactness considered by the GRFF simulations, we sample on observer orientation within the range $\xi \in (0, \pi/2)$ at intervals $\Delta \xi = 5\degree$, performing the ray tracing with {\sc Skylight}. Thus, we get a discrete sample in $(\mathcal{C}, \chi, \xi)$, for which we are ready to compute their light curves, provided the value of the last parameter is given, i.e.: the anisotropy index $b$.
In order to find the best fits, we compare the curves normalized by their maximum flux values. That is, we compare $F/F_{max}$ and $F^{N}/F^{N}_{max}$, being $F$ and $F^N$ the modelled and measured (by NICER) fluxes, respectively.
Notice that for the comparison one needs to account for the possible phase-shift $\varphi_o$ among the two curves, as well as a linear factor $\lambda_o$ arising from the fact that the peak values do not need to coincide (i.e., $F_{max} = \lambda_o F^{N}_{max}$). Thus, what we really compare is $f(\varphi; \lambda_o, \varphi_o, b) := (\lambda_o F/F_{max}) (\varphi+\varphi_o; b)$ against $ f^N (\varphi) := (F^{N}/F^{N}_{max}) (\varphi)$ for the parameters $\{ \lambda_o, \varphi_o, b \}$.
The performance in the above comparison can be captured by the classical \textit{reduced chi-square} factor,
\begin{equation}\label{eq:dist}
\chi_{r}^2 := \chi^2 / \varpi = \frac{1}{\varpi} \sum_{i}^{n} \frac{\left( f(\varphi_i; \lambda_o, \varphi_o, b) - f^N (\varphi_i)\right)^2}{\varepsilon_{iN}^2}
\end{equation}
where the sum runs over the $n$ phase-bins on the given NICER dataset; $\varpi = n - k$ are the estimated degrees of freedom, with $k = \# \{ \mathcal{C}, \chi, \xi, b, \kappa \bar{\gamma}, \varphi_o , \lambda_o \} = 7$ being the number of parameters of the model; and with $\varepsilon_{iN}$ being the statistical errors on the flux for each bin, as published in the cited NICER papers.
We use an L-BFGS-B algorithm from the Python \textit{SciPy optimization libraries} \cite{2020SciPy} to minimize $\chi_{r}^2$ over $\{ \lambda_o, \varphi_o, b \}$ for each configuration of the discrete sample in $(\mathcal{C}, \chi, \xi)$, restricting the beaming parameter to the range $b\in [0.5, 1.1]$. Finally, we refine the search in $\chi$ and $\xi$
simply by taking smaller intervals ($\Delta \chi=5 \degree$ and $\Delta \xi=1\degree$, respectively) at the most promising regions of the parameter space with the smallest \textit{reduced chi-square} values.
By applying this procedure to each of the four target pulsars, we obtain combined approximate best-fits for their light curves and phase-averaged spectra (Sec.\ref{sec:LCandS}).
\section{Results} \label{sec:results}
We focus on a NS of radius $\sim$ $11.5\,$km that rotates with period $4.8\,$ms, which is close to the estimated periods of our four targets PSR J0437–4715, PSR J1231−1411, PSR J2124−3358 and PSR J0030+0451 (see Table \ref{table1})~\footnote{Recalling that, what is actually being fixed at the set of FF simulations, is the relation $R = 0.05 \, c \, P/2\pi$. (As well as $\chi$ \& $\mathcal{C}$). }.
We first show the emission regions (ERs) arising from the GRFF simulations, defined as the regions at the stellar surface for which the four-current is spacelike (see Sec.~\ref{sec:emission_model}). In particular, we discuss the existence of an additional ER within the closed-zone of the pulsar, i.e., outside of the traditional polar caps along open field lines.
Then, we present our simultaneous best-fits for the light curves and spectra of each target pulsar, following the methodology described in Sec.~\ref{sec:setup}. Part of the procedure is illustrated at Appx.~\ref{sec:appendix}, where the posterior probability density distributions are also shown. Finally, we analyze the relevance of these extra ERs on fitting the light curves and discuss the main limitations of the adopted emission model, specially regarding the flux normalization.
\begin{table
\caption{ Relevant physical parameters (priors) of our four target millisecond pulsars. The third column is the magnetic field at the polar surface assuming the standard spin-down formula for a dipole in vacuum. Data are taken from the Australia Telescope National Facility pulsar catalogs \href{http://www.atnf.csiro.au/people/pulsar/psrcat/}{http://www.atnf.csiro.au/people/pulsar/psrcat/}.}
\vspace{0.2cm}
\centering {
\begin{tabular}{ c | c c c } \hline \hline
~ PSR ~ & ~ period [ms] ~ & ~ $B_{\rm pole}$ [$10^8$ G] ~ & ~ distance [pc] ~ \\ \hline
J0437–4715 & $~ 5.75 ~$ & $~ 5.80 ~$ & $~ 157 ~$ \\
J1231−1411 & $~ 3.68 ~$ & $~ 2.93 ~$ & $~ 420 ~$ \\
J2124−3358 & $~ 4.93 ~$ & $~ 3.22 ~$ & $~ 410 ~$ \\
J0030+0451 & $~ 4.87 ~$ & $~ 2.25 ~$ & $~ 324 ~$ \\ \hline \hline
\end{tabular}
}
\label{table1}
\end{table}
\subsection{Emitting regions} \label{sec:ERs}
In Fig.~\ref{fig:ER-chi} we display the regions of spacelike four-current over the stellar surface of a NS with intermediate compactness $\mathcal{C}=0.2$, for different misalignments $\chi$. The 3D plots illustrates one hemisphere of the NS showing (typically) a first ER, being a subset of the traditional polar caps (i.e. the regions demarcated by the footprints of all the open magnetic field lines). Interestingly, a second ER lies within the closed-zone of the pulsar, coincident with the intersection of the NS and the null charge surface (e.g., \cite{hirotani2001}).
We analyze the origin of such additional spacelike currents by comparing these configurations to their flat spacetime analogues (i.e., coming from FF simulations with Minkowski metric). In Figure \ref{fig:ER-metric}, we display the resulting ERs, together with the distribution of parallel current and electric charge densities at the meridional plane, both in flat and curved spacetimes.
The comparison indicates that these additional spacelike currents arise from the existence of electric currents along closed field lines for GR pulsar magnetospheres (which are absent in flat spacetime), combined with the well known null charge surface that is present in both backgrounds.
We checked that these electric currents are not just a transient effect, by extending one representative GRFF simulation to longer timescales. Additionally, we verified that they do not depend significantly on numerical resolution either. Our pulsar solutions are essentially converged at the resolution employed here, which can be further confirmed by looking at the luminosity within the light cylinder (as it was demonstrated in Fig.~2 of \cite{Pulsar}).
If not stated otherwise, we include these extra "non-standard" ERs as part of our emission model. Moreover, we consider a single parameter $\kappa \bar{\gamma}$ to describe both emitting regions (i.e., over open/closed zones). With this choice, we would typically find that the open ER has an averaged effective temperature which is roughly twice the averaged temperature over the additional (closed) ER.
Since it seems plausible that the heating and emission from these two quite different regions could or should be modelled separately (see Sec.~\ref{sec:normalization} for more discussion),
we shall examine the possibility of decoupling the parameter $\kappa \bar{\gamma}$ for the particular case of PSR J0030+0451 (see Fig.\ref{fig:PSR0451-alt}). Additionally, we will also analyze what happens if we exclude these extra (closed) ERs, later in Sec.~\ref{sec:extraER}.
\begin{figure*
\begin{center}
\includegraphics[scale=0.18]{cap_kerr01_C02_chi15.png}
\includegraphics[scale=0.18]{cap_kerr01_C02_chi30.png}
\includegraphics[scale=0.18]{cap_kerr01_C02_chi45.png}
\\
\vspace{0.4cm}
\includegraphics[scale=0.18]{cap_kerr01_C02_chi60.png}
\includegraphics[scale=0.18]{cap_kerr01_C02_chi75.png}
\includegraphics[scale=0.18]{cap_kerr01_C02_chi90.png}
\end{center}
\caption{ Emission regions for a pulsar with period $P\sim4.8\,$ms and compactness $\mathcal{C}=0.2$, for several misalignment angles $\chi= \{ 15\degree, 30\degree, 45\degree, 60\degree, 75\degree, 90\degree \}$ (from left-to-right and top-to-bottom). The colormaps show the temperature distribution at the NS surface (normalized by its maximum value $T_{max}$). That is, regions of spacelike current with temperature given by eq.~\eqref{eq:Teff}.
}
\label{fig:ER-chi}
\end{figure*}
\begin{figure*
\begin{center}
\includegraphics[scale=0.18]{cap_flat_chi30.png}
\includegraphics[scale=0.13]{pulsar005_deg30_JP-plane_flat.png}
\includegraphics[scale=0.13]{pulsar005_deg30_RHO-plane_flat.png}
\\
\vspace{0.4cm}
\includegraphics[scale=0.18]{cap_kerr01_C02_chi30.png}
\includegraphics[scale=0.13]{pulsar005_deg30_JP-plane_kerr.png}
\includegraphics[scale=0.13]{pulsar005_deg30_RHO-plane_kerr.png}
\end{center}
\caption{ \textit{Emission region analysis for pulsar with period $P\sim4.8\,$ms and misalignment angle $\chi = 30\degree$}. FF configurations in flat (top panel)
and Kerr (bottom panel) spacetimes. The colormap shows the
temperature map for spacelike currents at the stellar surface (left); electric current along the magnetic field $J_{\parallel}/ \Omega B $ (middle) and charge distribution $\rho$ (right) on the meridional plane. The main difference is the presence of electric currents at the inner region of the pulsar closed-zone in curved spacetime.
}
\label{fig:ER-metric}
\end{figure*}
\subsection{Light curves and spectra} \label{sec:LCandS}
We present our best-fits to the pulsar light curves and spectral distributions in Figures \ref{fig:PSR4715-best} to \ref{fig:PSR0451-alt}, and summarize relevant information about the parameters in Table \ref{table:best-fits}.
Even though the exploration of the parameter space is not fully exhaustive\footnote{Meaning that our coverage for some of the parameters, like the stellar compactness $\mathcal{C}$ and the pulsar misalignment $\chi$, is severely limited by the computational cost of the GRFF simulations.}, we achieve very good simultaneous fits to both the light curves and spectra for all the target pulsars.
In particular, for PSR J0437–4715 and PSR J2124−3358, we obtain really good light curve fits for which the \textit{reduced chi-square} factors are close to unity.
Whereas for the other two, PSR J1231−1411 and PSR J0030+0451, we find minimum values around $\chi_{r}^2 \sim 3$.
It is worth emphasizing again that we are assuming a centered dipole magnetic field configuration (i.e., we do not include any centered nor off-centered higher multipoles).
Our model has only few parameters, and the observational X-ray data for these pulsars is very accurate. The values of $\chi_{r}^2$ attained here are, hence, quite encouraging.
All the light curve fittings are accompanied by a consistent phase-averaged spectral distribution on their relevant energy windows (bottom panels). The only mild exception being the spectrum of PSR J0437–4715, which presents a visible deviation at high energies. On the other hand, the reason for the apparent deviation for the PSR J1231−1411 spectrum at low energies in Fig.~\ref{fig:PSR1411-best} is simply that we are comparing
it against the absorbed fitted observational data, an absorption that mainly affects at the lower energies. This fact imply a larger uncertainty in the determination of $\kappa \bar{\gamma}$ for this pulsar, although we notice that it does not affect significantly the light curve fitting; that is, for a wide range of $\kappa \bar{\gamma}$ values, we find best-fits with almost identical \textit{reduced chi-square} factors at the same values of $\mathcal{C}$, $\chi$ and $\xi$.
Notice that we are comparing our spectra with respect to some given fitted observational data, taken from different papers. There is, thus, an inherent uncertainty in the data which is not indicated on the plots.
As apparent from Table \ref{table:best-fits}, there are small degeneracies on the best fits among the geometrical parameters of the system, like the angles $\chi$ and $\xi$ (e.g., in the case of PSR J0437–4715), or $\mathcal{C}$ (as the case of PSR J2124−3358, which presents various configurations with $\chi_{r}^2 \sim 1$). This kind of degeneracies are commonly encountered,
and they could potentially be resolved by simultaneously contrasting (via an additional modelling) with electromagnetic observations in the radio and/or gamma-ray bands (e.g., \cite{chen2020, kalapotharakos2021}).
In the light curve of PSR J1231−1411, there is a small (but clear) secondary peak in the signal, which turns out to be quite challenging to accurately capture. In particular, we find that the configurations with minimum values of $\chi_{r}^2$ fail to reproduce this feature of the signal.
However, we do find other configurations with slightly larger $\chi_{r}^2$ values that provide a better qualitative fit (see middle panel of Fig.\ref{fig:PSR1411-best}).
The most challenging X-ray light curve is, as perhaps expected, that of PSR J0030+0451. For this case,
we get a more modest fitting with a factor $\chi_{r}^2 = 10$ in the general search, although the result is still qualitatively good, in the sense that we are able to capture the significant difference in the amplitude of profiles and the depths of valleys. Moreover, we note that it is possible to significantly improve the light curve fitting (achieving $\chi_{r}^2 =3$) if allowing a decoupling of the effective temperature coefficient $\kappa \bar{\gamma}$ for the two kinds of ERs (i.e., over open/closed regions).
This is illustrated on Fig.~\ref{fig:PSR0451-alt}, where the quality of this alternative light curve fit can be fully appreciated. It is interesting to note how the inclusion of a single additional parameter in our model, in this case, can considerably decrease the \textit{reduced chi-square} factor.
In doing so, on the other hand, we observe there is a negative impact on its spectral distribution (see bottom panel of Fig.~\ref{fig:PSR0451-alt}).
We shall come back to this point below, when discussing the limitations of the emission model.
\begin{table
\caption{ Best-fits parameters for the X-rays signals of the four target millisecond pulsars. We fixed a single approximate rotation period of $4.8\,$ms for all cases, while taking individual priors for magnetic field strength and distance from Table~\ref{table1}.\\
{\footnotesize (*) The effective temperature parameter $\kappa \bar{\gamma}$ is decoupled here for the two different kind of emission regions, i.e. over the polar cap (open field lines) and the extra ER over the closed zone of the pulsar. Being $(\kappa \bar{\gamma})_{\rm open} = 3.8 \times 10^7$ and $(\kappa \bar{\gamma})_{\rm closed} \approx 4 \, (\kappa \bar{\gamma})_{\rm open}$.}
}
\vspace{0.2cm}
\centering {
\begin{tabular}{ c | c c c c c | c} \hline \hline
~ PSR ~ & ~ $\mathcal{C}$ ~ & ~ $\chi$ ~ & ~ $\xi$ ~ & ~ $b$ ~ & ~ $\kappa \bar{\gamma}$ ~ & $\chi_{r}^2$\\ \hline
J0437–4715 & $~ 0.25 ~$ & $~ 20 \degree ~$ & $~ 53 \degree ~$ & $~ 0.94 ~$ & $~ 7.0 \times 10^6 ~$ & $~ ~ 1.2~$ \\
& $~ 0.25 ~$ & $~ 15 \degree ~$ & $~ 61 \degree ~$ & $~ 1.00 ~$ & $~ 7.0 \times 10^6 ~$ & $~ ~ 1.2~$ \\
J1231−1411 & $~ 0.18 ~$ & $~ 75 \degree ~$ & $~ 12 \degree ~$ & $~ 0.81 ~$ & $~ 2.5 \times 10^7 ~$ & $~ ~ 3.0~$ \\
& $~ 0.22 ~$ & $~ 43 \degree ~$ & $~ 30 \degree ~$ & $~ 0.60 ~$ & $~ 2.5 \times 10^7 ~$ & $~ ~ 5.3~$ \\
J2124−3358 & $~ 0.22 ~$ & $~ 60 \degree ~$ & $~ 24 \degree ~$ & $~ 0.67 ~$ & $~ 1.3 \times 10^8 ~$ & $~ ~ 0.8~$ \\
& $~ 0.20 ~$ & $~ 75 \degree ~$ & $~ 17 \degree ~$ & $~ 0.67 ~$ & $~ 1.3 \times 10^8 ~$ & $~ ~ 0.9~$ \\
J0030+0451 & $~ 0.22 ~$ & $~ 25 \degree ~$ & $~ 85 \degree ~$ & $~ 1.09 ~$ & $~ 3.8 \times 10^7 ~$ & $~ 10.0~$ \\
& $~ 0.22 ~$ & $~ 25 \degree ~$ & $~ 80 \degree ~$ & $~ 1.02 ~$ & (*) & $~ 3.0~$ \\ \hline \hline
\end{tabular}
}
\label{table:best-fits}
\end{table}
\begin{figure
\begin{center}
\includegraphics[scale=0.55]{lightcurve_J0437-4715_C025_chi_deg20_xi_deg53.png}
\\
\includegraphics[scale=0.55]{spectrum_J0437-4715_C025_chi_deg20_xi_deg53.png}
\end{center}
\caption{ \textit{Best-fit to NICER X-ray data of PSR J0437-4715.}
Comparison of our best-fit configuration (first row in Table~\ref{table:best-fits}) against the observed light curve (top) and spectrum (bottom). The light curve shown in black, with its error bars, corresponds to NICER data taken from \cite{bogdanov2019constrainingI}.
Our modelled light curve (in blue) was integrated on the same energy window, as indicated in the plot.
The spectral data, on the other hand, is from XMM-Newton X-ray EPIC MOS1/2, fitted with a 2Hatm+BB+PL model \cite{bogdanov2012}, and we plotted (in black) the unabsorbed component. Our modelled spectrum (in blue) was re-scaled by a factor of $8.2 \times 10^{-3}$ for the comparison.
}
\label{fig:PSR4715-best}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=0.55]{lightcurve_J1231-1411_C018_chi_deg75_xi_deg12.png}
\\
\includegraphics[scale=0.55]{lightcurve_J1231-1411_C022_chi_deg43_xi_deg30.png}\\
\includegraphics[scale=0.55]{spectrum_J1231-1411_C022_chi_deg60_xi_deg24.png}
\end{center}
\caption{ \textit{Best-fit to NICER X-ray data of PSR J1231-1411.}
Comparison of our best-fit configuration (first row in Table~\ref{table:best-fits}) against the observed light curve (top) and spectrum (bottom). The light curve shown in black, with its error bars, correspond to NICER data taken from \cite{bogdanov2019constrainingI}.
Our modelled light curve (in blue) was integrated on the same energy window, as indicated in the plot.
The spectral data is from NICER, fitted with an Hatm (\textit{nsatmos}) model \cite{ray2019discovery}, and we plotted (in black) the unabsorbed component. Our modelled spectrum (in blue) was re-scaled by a factor of $4.1 \times 10^{-3}$ for the comparison. The middle panel corresponds to an alternative light curve fit (second row in Table~\ref{table:best-fits}) that better captures the qualitative features of the signal.
}
\label{fig:PSR1411-best}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=0.55]{lightcurve_J2124-3358_C022_chi_deg60_xi_deg24.png}
\\
\includegraphics[scale=0.55]{spectrum_J2124-3358_C022_chi_deg60_xi_deg24.png}
\end{center}
\caption{ \textit{Best-fit to NICER X-ray data of PSR J2124-3358.}
Comparison of our best-fit configuration (first row in Table~\ref{table:best-fits}) against the observed light curve (top) and spectrum (bottom). The light curve shown in black, with its error bars, correspond to NICER data taken from \cite{bogdanov2019constrainingI}.
Our modelled light curve (in blue) was integrated on the same energy window, as indicated in the plot.
The spectral data is taken from XMM-Newton observations, fitted with an Hatm+PL model \cite{zavlin2006xmm}, and we plotted (in black) the --only published-- absorbed component. Our modelled spectrum (in blue) was re-scaled by a factor of $2.1 \times 10^{-4}$ for the comparison.
}
\label{fig:PSR3358-best}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=0.55]{lightcurve_J0030+0451_C022_chi_deg25_xi_deg85.png}
\\
\includegraphics[scale=0.55]{spectrum_J0030+0451_C022_chi_deg25_xi_deg85.png}
\end{center}
\caption{ \textit{Best-fit to NICER X-ray data of PSR J0030+0451.}
Comparison of our best-fit configuration (first row in Table~\ref{table:best-fits}) against the observed light curve (top) and spectrum (bottom). The light curve shown in black, with its error bars, correspond to NICER data taken from \cite{bogdanov2019constrainingI}.
Our modelled light curve (in blue) was integrated on the same energy window, as indicated in the plot.
The spectral data is taken from the unabsorbed posterior-expected phase-averaged spectrum for ST+PST (i.e., Hatm model) \cite{riley2019}, consistent with both XMM-Newton and NICER data.
Our modelled spectrum (in blue) was re-scaled by a factor of $5 \times 10^{-3}$ for the comparison.
}
\label{fig:PSR0451-best}
\end{figure}
\begin{figure
\begin{center}
\includegraphics[scale=0.55]{lightcurve_J0030+0451_C022_chi_deg25_xi_deg80.png}
\\
\includegraphics[scale=0.55]{spectrum_J0030+0451_C022_chi_deg25_xi_deg80.png}
\end{center}
\caption{ \textit{Alternative best-fit for PSR J0030+0451 light curve, by decoupling the temperature parameter $\kappa \bar{\gamma}$.} The effective temperature parameter $\kappa \bar{\gamma}$ is decoupled here for the two different kind of emission regions, i.e. over the polar cap (open field lines) and the extra ER over the closed zone of the pulsar. Being $(\kappa \bar{\gamma})_{\rm open} = 3.8 \times 10^7$ and $(\kappa \bar{\gamma})_{\rm closed} \approx 4 \, (\kappa \bar{\gamma})_{\rm open}$.
All the other specifications of the plots are identical to Fig.~\ref{fig:PSR0451-best}.
}
\label{fig:PSR0451-alt}
\end{figure}
\subsection{Relevance of the additional emitting region} \label{sec:extraER}
We want to elucidate now the importance of the non-standard ER found over the closed-zone of the pulsar. To this end, we first examine the impact of such additional ER on the light curves best-fits from the previous section. In Figure \ref{fig:LC-extra} we present a splitting of each light curve into the contributions arising from the open/closed ERs.
It can be seen that the signals from the open ER, sub-regions of the antipodal polar caps, tend to be dominant and highly symmetric. While those arising from the ER in the closed-zone, also antipodal and quite symmetric, are slightly out of phase with respect to the open ER curve. This is what allows to produce a combined signal with an asymmetric interpulse between the peaks as in PSR J0437–4715 and PSR J2124−3358. Or, in the case of PSR J0030+0451, to approximate the observed imbalance between the two valleys on its light curve.
\begin{figure*
\subfloat[ PSR J0437-4715 \label{split:a}]{%
\includegraphics[scale=0.56]{ER-split_C025_chi_deg20_xi_deg53.png}
}
\subfloat[PSR J1231-1411 \label{split:b}]{%
\includegraphics[scale=0.56]{ER-split_C018_chi_deg75_xi_deg12.png}
}
\vspace{0.25cm}
\subfloat[PSR J2124-3358 \label{split:c}]{%
\includegraphics[scale=0.56]{ER-split_C022_chi_deg60_xi_deg24.png}
}
\subfloat[PSR J0030+0451 \label{split:d}]{%
\includegraphics[scale=0.56]{ER-split_C022_chi_deg25_xi_deg85.png}
}
\vspace{0.5cm}
\caption{\textit{Contribution of the non-standard (closed) ERs to the light curves for each of the target pulsars (a), (b), (c) and (d).} We split the total ERs (i.e., spacelike currents) on a piece contained within the usual polar caps (i.e., open field lines) region, and an extra piece arising from a spacetime curvature effect at the closed-zone of the pulsar. The plots correspond to the first configurations listed in Table~\ref{table:best-fits} for each of the four target pulsars.
}
\label{fig:LC-extra}
\end{figure*}
However, one might ask what happens if the closed ER is excluded, while allowing for a full exploration of the parameter space. Is it possible to achieve fits equally good as those of Table \ref{table:best-fits} where both ERs were included?
The answer is generally NO, with the exception of the PSR J1231−1411 signal. We filtered-out the closed ER and repeated the whole search procedure of Sec.\ref{sec:search} for all our target pulsars. The results of such exploration are summarized on Table \ref{table:filt-fits}, where one clearly sees worse quantitative fits, reflected on $\chi_{r}^2$ values increasing by a factor up to $\sim 8$. The reason is, as anticipated above, that without the additional (closed) ER, the model is unable to capture the observed features between the main pulses in the signals.
In PSR J1231−1411, such interpulse is quite small and symmetric, explaining why this case can be fitted almost equally well with both (open/closed) and a single (open) emitting regions. This is consistent with previous results by other collaborations for PSR J0030+0451, according to which a single pair of antipodal regions is unable to properly capture the asymmetries of the pulse profiles \cite{miller2019, riley2019, chen2020, kalapotharakos2021}. The previous works resolve the tension by introducing off-centered higher magnetic multipoles, thus giving place to more complex polar caps. In contrast, in our case the tension is resolved by having another pair of antipodal regions (the closed ERs, arising from the GRFF simulations), that allow us to capture the asymmetries of the pulses while still maintaining purely dipolar magnetic fields.
\begin{table
\caption{ Best-fits parameters for the X-rays signals of the four target millisecond pulsars, when filtering-out the additional ERs over the close-region.
}
\vspace{0.2cm}
\centering {
\begin{tabular}{ c | c c c c c | c} \hline \hline
~ PSR ~ & ~ $\mathcal{C}$ ~ & ~ $\chi$ ~ & ~ $\xi$ ~ & ~ $b$ ~ & ~ $\kappa \bar{\gamma}$ ~ & $\chi_{r}^2$\\ \hline
J0437–4715 & $~ 0.22 ~$ & $~ 60 \degree ~$ & $~ 26 \degree ~$ & $~ 0.52 ~$ & $~ 7.0 \times 10^6 ~$ & $~ ~ 8.7~$ \\
J1231−1411 & $~ 0.25 ~$ & $~ 15 \degree ~$ & $~ 69 \degree ~$ & $~ 0.64 ~$ & $~ 1.3 \times 10^7 ~$ & $~ ~ 3.6~$ \\
& $~ 0.25 ~$ & $~ 15 \degree ~$ & $~ 76 \degree ~$ & $~ 0.78 ~$ & $~ 1.3 \times 10^7 ~$ & $~ ~ 5.6~$ \\
J2124−3358 & $~ 0.25 ~$ & $~ 15 \degree ~$ & $~ 78 \degree ~$ & $~ 0.58 ~$ & $~ 3.5 \times 10^7 ~$ & $~ ~ 3.7~$ \\
J0030+0451 & $~ 0.25 ~$ & $~ 45 \degree ~$ & $~ 85 \degree ~$ & $~ 0.78 ~$ & $~ 3.2 \times 10^7 ~$ & $~ 24.4~$ \\ \hline \hline
\end{tabular}
}
\label{table:filt-fits}
\end{table}
Excluding the closed ER --as well as bringing its averaged temperature closer to the one of the open ER (like the case shown in Fig.\ref{fig:PSR0451-alt})-- implies a much narrower distribution around a single effective temperature. We notice that, within the emission model employed, this would typically lead to poorer fits of the spectral shapes.
On the other hand, we find that the spectrum of the best-fit configuration for PSR J1231−1411 excluding the closed ER (from Table \ref{table:filt-fits}) remains well described by the model.
\subsection{Normalization issues} \label{sec:normalization}
Arguably, the main limitation of our current modelling is reflected on the overall flux intensities, which --as indicated in figures \ref{fig:PSR4715-best} to \ref{fig:PSR0451-alt}-- are about $2$--$4$ orders of magnitude higher as compared to the (fitted) observational data.
We shall examine here in some detail the case of PSR J0030+0451, for which there is more data analysis available. To this end, we take its best-fit configuration from Table~\ref{table:best-fits} (the first row for this pulsar) and estimate the effective emitting areas and the averaged effective temperatures for the two kinds of emitting regions; namely, the open/closed ERs. The resulting values are summarized on Table~\ref{table:estimations}, together with analogous parameters of a few relevant spectral fits from \cite{bogdanov2009deep} for blackbody (BB) and Hydrogen (Hatm) atmospheric models, with and without additional power-law (PL) components.
\begin{table
\caption{ Estimated effective averaged temperatures and emitting areas in our model, compared with analogous parameters from different spectral fittings for PSR J0030+0451.\\
{\footnotesize (*) i.e.: BB with beaming. $T_{\rm eff, 1}$ and $T_{\rm eff, 2}$ represent in this case, the averaged effective temperatures over the open and closed ERs, respectively. Whereas $R_{\rm eff, 1/2}$ are estimated from the corresponding net emitting areas of each region, as $A_{\rm eff, 1/2} = 4 \pi R_{\rm eff, 1/2}^2$. }
}
\vspace{0.2cm}
\centering {
\begin{tabular}{ l | c c c c } \hline \hline
~ Spectral model ~ & ~ $T_{\rm eff, 1}$ ~ & ~ $R_{\rm eff, 1}$~ & ~ $T_{\rm eff, 2}$ ~ & ~ $R_{\rm eff, 2}$ ~ \\
~ ~& ~[$10^6$ K] ~ & ~[km] ~ & ~ [$10^6$ K] ~ & ~ [km] ~ \\ \hline
This work (*) & $~ 3.2 ~$ & $~1.17~$ & $~ 1.5 ~$ & $~ 3.55 ~$ \\
BB(x2) & $~ 3.1 ~$ & $~0.05~$ & $~ 1.5 ~$ & $~ 0.28 ~$ \\
Hatm(x2) & $~ 1.7 ~$ & $~0.25~$ & $~ 0.7 ~$ & $~ 2.2 ~$ \\
BB(x2) + PL & $~ 2.5 ~$ & $~0.08~$ & $~ 1.2 ~$ & $~ 0.34 ~$ \\
Hatm(x2) + PL & $~ 1.4 ~$ & $~0.37~$ & $~ 0.7 ~$ & $~ 1.9 ~$ \\ \hline \hline
\end{tabular}
}
\label{table:estimations}
\end{table}
We observe that our averaged temperatures are similar to those from the BB fits, but our emitting areas are a factor of $\gtrsim 100$ larger than the corresponding BB ones, thus explaining the roughly two orders of magnitude excess in our fluxes.
However, we notice that --since the spectrum of a Hydrogen atmosphere is harder than that of a BB for the same effective temperature-- the emitting areas inferred for PSR J0030+0451 with the Hatm model are much closer to ours (although still smaller, by a factor $\sim 3$). On the other hand, we note that the effective temperatures are lower in the Hatm model, but still maintaining the same proportion, $T_{\rm eff, 1} \approx 2 T_{\rm eff, 2}$.
Therefore, we might expect that replacing the BB (with beaming) in our emission model by an Hydrogen atmosphere, and by re-adjusting the $\kappa \bar{\gamma}$ values accordingly (i.e., reducing the effective temperatures by a factor $\sim 1/2$), we should get much more consistent spectra in absolute physical units for all our target pulsars, while keeping the good quality of their light curves fits. Testing this hypothesis is, however, beyond the scope of the present article and we defer it to a future work.
Following the above reasoning, we could look back to the spectrum found in Fig.~\ref{fig:PSR0451-alt} for the alternative best-fit of PSR J0030+0451. We note that by increasing the value of $(\kappa \bar{\gamma})_{\rm closed}$ on a factor $\sim 4$, one is increasing the effective temperature of this ER by a factor $\sim 1.4$; thus, bringing it closer to the averaged values of the hotter (open) region.
This has shown a negative impact for fitting the spectrum with our BB $+$ beaming model, but it would not necessarily be the case with the Hatm model. For instance, one of the most favoured synthetic ER models constructed in \cite{riley2019}, ST+PST, consist on a single effective temperature along with the Hatm atmospheric model.
A more thorough exploration of such decoupling for $\kappa \bar{\gamma}$ in the emission model, both from a physical perspective and from its implications for fitting the light curves, as well as shifting to an Hydrogen atmospheric model, will be a clear continuation route for a future study.
\section{Conclusions}\label{sec:conclusions}
In this article, we modelled thermal X-ray signals produced by MSPs and fitted them to observational data from several target pulsars: PSR J0437–4715, PSR J1231−1411, PSR J2124−3358 and PSR J0030+0451.
Our approach starts by numerically solving the pulsar magnetospheres under the FF approximation,
assuming a simple centered dipolar magnetic field. The solutions are then linked to an effective emissivity over the stellar surface trough a simple emission model; followed by the propagation of the emitted photons (via ray tracing) to compute light curves and spectra. These are finally compared against the observational data of the target MSPs.
We found non-standard emission regions (ERs) on the stellar surface, associated to spacelike electric currents over the closed-zone of the pulsar. Such ERs appear as a consequence of having included the gravitational curvature of the NS in the FF simulations, and they are also related to the null charge surfaces in the magnetosphere.
With these closed ERs included in our emission model, we were able to achieve excellent fits to the light curves for all the four MSPs considered. Simultaneously, we have obtained a good agreement with the spectral shape of their thermal emission. This is quite remarkable, given that these X-rays light curves have been typically attributed to the presence of strong higher-multipole moments and/or off-centered dipolar components in the magnetic field (e.g., \cite{bogdanov2007, bogdanov2008, bogdanov2012, lockhart2019x}), especially for the widely studied case of PSR J0030+0451 \cite{bilous2019, chen2020,kalapotharakos2021}.
We have observed that without these additional (closed) ERs, the light curves fitting become significantly worse, even at the qualitative level. The reason for this, is that the closed ERs provides an additional contribution to the signal, capable of producing the asymmetries found in between the main pulses (interpulses) of the observed light curves. That is, even when the individual signals arising from the (antipodal) open/closed ERs are quite symmetric, the slight dephasing among them lead to the necessary interpulses modulations when combined.
The spirit of this work is not to claim we have found the ``correct" model, but rather to emphasize the need for a better connection between the global magnetospheric properties and the determination of the emitting regions, as well as the appropriate modelling of such emissions.
We have shown that including the spacetime curvature on the magnetosphere's description results in more complex ERs, that might account for the observed light curves without the need of invoking large non-dipolar magnetic field components and/or offsets from the NS center. In particular, we showed that (regardless of the particular emission model employed here) antipodal ERs can indeed produce X-ray light curves approaching those detected by NICER. This also raises interesting questions regarding the physical nature of these additional spacelike currents and their entailment to the bombardment mechanism responsible for heating the stellar surface. Will these currents contribute in the same way to the bombardment as those reaching the polar caps along open field lines? Should their thermal emissions be modelled separately?
We have recognized that the emission model considered is the main limitation of our present X-ray modelling, which considerably overestimates the observational fluxes. For the BB+beaming model assumed here, the effective emitting areas of our ERs turn-out to be too large. However, we noticed that for an Hydrogen atmospheric model --which has a much harder spectrum as compared to BB-- the situation would improve significantly. We expect that an Hatm model should lead to more consistent spectra in absolute physical units, while keeping the good quality of the light curve fits found here for all the target MSPs.
A natural next step to pursue this research further, will be to improve the emission model by considering an Hatm model with a more realistic (self-consistent) angular distribution (i.e., beaming pattern), as well as a more refined modelling of the bombardment mechanism (based on the FF currents) along the lines of \cite{baubock2019, salmi2020}. Such refinement, could incorporate the idea of decoupling the description of the two type of ERs, namely: open/closed ERs.
Moreover, a more sophisticated simultaneous fitting based on the full phase-resolved spectral data of PSR J0030+0451 would be an interesting improvement of this study. As well as expanding the application of our model to other MSPs, if sufficiently accurate X-ray data is available.
\section{Acknowledgments}
The authors would like to thank Rosalba Perna for helpful suggestions. F.C., J.P. and O.R acknowledge financial support from CONICET, SeCyT-UNC, and MinCyT-Argentina. D.V. is funded by the European Research Council
(ERC) under the European Union’s Horizon 2020 research
and innovation programme (ERC Starting Grant IMAGINE,
No. 948582) and his institution is supported by the Spanish
program Unidad de Excelencia Mar{\i}a de Maeztu, CEX2020-
001058-M. C.P. is supported by the Grant PID2019-110301GB-I00 funded by MCIN/AEI/10.13039/501100011033 and by "ERDF A way of making Europe".
Numerical computations were performed on the Sakura cluster at Max-Planck Computing and Data Facility, and on the Serafin Cluster at Centro de Computación de Alto Desempeño, Universidad Nacional de Cordoba, which is part of the Sistema Nacional de Computación de Alto Desempeño, MinCyT-Argentina.
|
\section{Maximal Independent Set}
\label{sec:MIS}
In this section, we describe a deterministic distributed algorithm that computes an MIS in $O(\log^2 \Delta \,\cdot \,\log n)$ rounds of the $\mathsf{LOCAL}$\xspace model. Furthermore, we explain how a variant of this MIS algorithm can be implemented in $\tilde{O}(\log^2 \Delta \,\cdot \, \log n)$ rounds of the $\mathsf{CONGEST}$\xspace model.
\begin{theorem}
\label{thm:mis}
There is a deterministic distributed algorithm that computes a maximal independent set in $O(\log^2 \Delta \,\cdot \,\log n)$ rounds of the $\mathsf{LOCAL}$\xspace model, and in $O(\log^2 \Delta \,\cdot \, \log\log \Delta \,\cdot \,\log n)$ rounds of the $\mathsf{CONGEST}$\xspace model, in any $n$-node graph with maximum degree at most $\Delta$.
\end{theorem}
The rest of the section is devoted to the proof of \cref{thm:mis}.
We first recall Luby's classic algorithm\cite{luby86}, which in each iteration chooses an independent set of nodes such that, when we add them to the output and remove them from the graph along with their neighbors, in expectation half of the edges of the graph are removed.
We discuss the randomized analysis of this algorithm and explain how we can formulate it in the framework of the rounding procedure of \cref{sec:genericrounding}, such that we can derandomize each iteration in $O(\log^2 \Delta)$ rounds of the $\mathsf{LOCAL}$\xspace model, and $O(\log^2 \Delta \,\cdot \,\log\log \Delta \,+\, \log\Delta \,\cdot \, \log^* n)$ rounds of the $\mathsf{CONGEST}$\xspace model. In the latter bound, the second term is upper bounded by the first term in all cases of interest where $\Delta>\log n$ and thus $\log \Delta \gg \log^* n$. This is because for $\Delta<\log n$, the theorem statement already follows from the classic $O(\Delta+\log^* n)$-round algorithm of Barenboim, Elkin, and Kuhn~\cite{BEK15}.
\paragraph{Luby's Randomized MIS Algorithm.} The starting point is to recall Luby's classic randomized algorithm from \cite{luby1993removing}.
Each iteration of it works as follows.
We mark each node $v$ with probability $1/(20 \deg(v))$. Then, for each edge $\{u,v\}$, let us orient this edges as $u\rightarrow v$ if and only if $\deg(u)<\deg(v)$ or $\deg(u)=\deg(v)$ and $ID(u)<ID(v)$. For each marked node $v$, we add $v$ to the independent set if and only if $v$ does not have a marked out-neighbor. Finally, as a clean-up step at the end of this iteration, we remove all nodes that have been added to the independent set along with their neighbors. We then proceed to the next iteration.
\paragraph{Derandomizing Luby via Local Rounding.} It is well-known that in each iteration of Luby's algorithm a constant fraction of the edges of the remaining graph gets removed, in expectation. Hence, the process terminates in $O(\log n)$ iterations with probability $1 - 1/\operatorname{\text{{\rm poly}}}(n)$.
We explain how to derandomize each iteration of the algorithm in $O(\log^2 \Delta)$ rounds, such that we still remove a constant fraction of the edges per iteration.
For the rest of this proof, we focus on an arbitrary iteration, and we assume that $G=(V, E)$ is the graph induced by the remaining vertices at the beginning of this iteration. Let $\vec{x} \in \{0, 1\}^{|V|}$ be the indicator vector of whether different nodes are marked, that is, we have $x_v=1$ if $v$ is marked and $x_v=0$ otherwise. Let $R_v(\vec{x})$ be the indicator variable of the event that $v$ gets removed, for the marking vector $\vec{x}$. Let $Z(\vec{x})$ be the corresponding number of removed edges. Luby's algorithm determines the markings $\vec{x}$ randomly. Our task is to derandomize this and select the marked nodes in a deterministic way such that when we remove nodes added to the independent set (those marked nodes that do not have a marked out-neighbor) and their neighbors, along with all the edges incident on these nodes, at least a constant fraction of edges $E$ get removed.
\paragraph{Good and bad nodes and prevalence of edges incident on good nodes.} We call any node $v$ \emph{good} if and only if it has at least $\deg(v)/3$ incoming edges. A node $v$ that is not good is called bad. It can be proven \cite{luby1993removing} that
\begin{align}
\label{eq:luby0}
\sum_{\textit{good vertex\;} v} \deg(v) \geq |E|/2.
\end{align}
Even though the reader may skip this paragraph, for completeness we include the reason as it is a simple and intuitive charging argument.
Recall that by definition any bad node has less than $1/3$ of its edges incoming.
Thus any edge incoming to a bad node $v$ can be charged to two unique edges going out of $v$, in such a manner that each edge of the graph is charged at most once.
Hence, the number of edges incoming to bad nodes is at most $|E|/2$.
Thus, the number of edges that have at least one good endpoint is at least $|E|/2$, which implies the desired bound $\sum_{\textit{good vertex\;} v} \deg(v) \geq |E|/2$.
\paragraph{Lower bounding removed edges.}
We can lower bound the number of removed edges as $$Z(\vec{x}) \geq \sum_{\textit{good vertex\;} v} \deg(v) \cdot R_{v}(\vec{x})/2.$$
Here, the $2$ factor in the denominator is because for an edge, both endpoints might be good nodes and get removed.
Given that $\sum_{\textit{good vertex\;} v} \deg(v) \geq |E|/2$, to prove that $\mathbb{E}[Z(\vec{x})] = \Omega(|E|)$, it suffices to show that each good vertex $v$ has $Pr[R_v(\vec{x})] =\Omega(1)$.
This fact can be proven via elementary probability calculations.
Next, we discuss how to prove it using only pairwise independence in the analysis and then how this nicely fit our deterministic local rounding framework.
\paragraph{Pessimistic estimator of removed edges via pairwise-independent analysis.} Let us use $IN(u)$ and $OUT(u)$ to denote in-neighbor and out-neighbors of a vertex $u$. Consider a good node $v$ and consider all its incoming neighbors $u$, i.e., neighbors $u$ such that $(\deg(u), ID(u))<(\deg(v), ID(v))$. We know that for each such neighbor $u$, we have $Pr[x_u] \geq \frac{1}{20 \deg(v)}$. Furthermore, since $v$ is good, it has at least $\deg(v)/3$ such neighbors.
Hence, we have $\sum_{\textit{incoming neighbor}\; u} Pr[x_u] \geq 1/60$.
Choose a subset $IN^*(v)\subseteq IN(v)$ of incoming neighbors such that
\begin{align}
\label{eq:luby1}
\sum_{u \in IN^{*}(v)} Pr[x_u] \in [1/60, 4/60].
\end{align}
Notice that such a subset $IN^*(v)$ exists since the summation over all incoming neighbors is at least $1/60$ and each neighbor's probability is at most $1/20$.
On the other hand, notice that for any node $u$, we have
\begin{align}
\label{eq:luby2}
\sum_{w \in OUT(u)} Pr[x_{w}] \leq 1/20.
\end{align}
This is because $|OUT(u)| \leq \deg(u)$ and for each ${w}\in OUT(u)$, we have $(\deg({w}), ID({w}))>(\deg(u), ID(u))$ and thus $Pr[x_{w}] \leq 1/(20\deg(u))$.
A sufficient event $\mathcal{E}(v,u)$ that causes $v$ to be removed is if some $u\in IN^*(v)$ is marked and no other node in $IN^{*}(v)\cup OUT(u)$ is marked.
By union bound, this event's indicator is lower bounded by $$x_u - \sum_{u' \in IN^*(v), u\neq u'} x_{u}\cdot x_{u'} - \sum_{w \in OUT(u)} x_{u}\cdot x_{w}.$$
Furthermore, the events $\mathcal{E}(v,u_1), \mathcal{E}(v,u_2), \dots, \mathcal{E}(v,u_{|IN^*(v)|})$ are mutually disjoint for different $u_1, u_2, \dots,$ $u_{|IN^*(v)|} \in IN^*(v)$.
Hence, we can sum over these events for different $u\in IN^*(v)$ and conclude that
\begin{align*}
R_v(\vec{x}) & \geq \sum_{u\in IN^*(v)} \bigg(x_u - \sum_{u' \in IN^*(v), u\neq u'} x_{u}\cdot x_{u'} - \sum_{w \in OUT(u)} x_{u}\cdot x_{w}\bigg) \\
& =
\sum_{u\in IN^*(v)} x_u - \sum_{u, u' \in IN^*(v)} x_{u}\cdot x_{u'} - \sum_{u\in IN^*(v)}\sum_{w \in OUT(u)} x_{u}\cdot x_{w}
\end{align*}
Therefore, our overall pessimistic estimator for the number of removed edges gives that
\begin{align*}
Z(\vec{x}) \geq & \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot R_v(\vec{x}) \\
\geq & \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot \bigg(
\sum_{u\in IN^*(v)} x_u - \sum_{u, u' \in IN^*(v)} x_{u}\cdot x_{u'} - \sum_{u\in IN^*(v)}\sum_{w \in OUT(u)} x_{u}\cdot x_{w}\bigg)
\end{align*}
\paragraph{Formulating the pessimistic estimator in our rounding framework.} We next describe how we can formulate the above pessimistic estimator in the framework of our rounding procedure described in \Cref{sec:basicrounding}. The labeling space is whether each node is marked or not, i.e., each node takes simply one of two possible labels $\{0,1\}$ where $1$ indicates that the node is marked. For a given label assignment $\vec{x} \in \{0,1\}^{|V|}$, we define the utility function as $$\mathbf{u}(\vec{x}) = \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot \big(\sum_{u\in IN^*(v)} x_u\big),$$ and the cost function as
$$\mathbf{c}(\vec{x}) = \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot \bigg(
\sum_{u, u' \in IN^*(v)} x_{u}\cdot x_{u'} + \sum_{u\in IN^*(v)}\sum_{w \in OUT(u)} x_{u}\cdot x_{w}\bigg).$$
If the label assignment is relaxed to be a fractional assignment $\vec{x}\in [0,1]^{|V|}$, where intuitively now $x_v$ is the probability of $v$ being marked, the same definitions apply for the utility and cost of this fractional assignment.
The utility function is simply a summation of functions, each of which depends on the label of only one vertex. Hence, it directly fits out rounding framework as discussed in \Cref{sec:nodeutilitiesandcosts}.
To capture the cost function as a summation of costs over edges, we next define an auxiliary multi-graph $H$ as follows: For each good node $v$, for every two vertices $u, u'\in IN^{*}(v)$, we add an auxiliary edge between $u$ and $u'$, with a cost function which is equal $\deg(v)/2$ when both $u$ and $u'$ are marked, and zero otherwise. Furthermore, for each $u \in IN^{*}(v)$ and each $w\in OUT(u)$, we add to the edge $(u, w)$ a cost function which is equal to $\deg(v)/2$ when both $u$ and $w$ are marked and zero otherwise. Notice that $H$ is a $d2$-multigraph of $G$ according to the definition in \Cref{sec:d2communication}.
For the fractional label assignment $\vec{x}\in [0,1]^{|V|}$ in Luby's algorithm --- i.e., where $x_v=1/(20\deg(v))$--- these utility and cost functions are clearly polynomially bounded in $\Delta$, simply because each term is at least $1/{20\Delta}$ and there are no more than $\operatorname{\text{{\rm poly}}}(\Delta)$ terms per node. We next argue that these utility and cost functions also satisfy the key requirement of \Cref{lemma:d2rounding} with $\mu=1/2$:
\begin{claim}\label[claim]{clm:MIS1} For the fractional label assignment $\vec{x}\in [0,1]^{|V|}$ in Luby's algorithm --- i.e., where $x_v=1/(20\deg(v))$--- we have $\mathbf{u}(\vec{x})-\mathbf{c}(\vec{x}) \geq \mathbf{u}(\vec{x})/2$.
\end{claim}
\begin{proof}
We have \begin{align*}
\mathbf{u}(\vec{x})-\mathbf{c}(\vec{x}) = & \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot \bigg(\sum_{u\in IN^*(v)} x_u - \sum_{u, u' \in IN^*(v)} x_{u}\cdot x_{u'} - \sum_{u\in IN^*(v)}\sum_{w \in OUT(u)} x_{u}\cdot x_{w}\bigg) \\
= & \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot \bigg(\sum_{u\in IN^*(v)} x_u \cdot \big(1- \sum_{u' \in IN^*(v)} x_{u'} - \sum_{w \in OUT(u)} x_{w}\big)\bigg) \\
\geq & \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot \bigg(\sum_{u\in IN^*(v)} x_u \big(1- 4/60 - 1/20 \big)\bigg) \\
\geq & \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot \big(\sum_{u\in IN^*(v)} x_u/2\big) = \mathbf{u}(\vec{x})/2
\end{align*}
where we have used \cref{eq:luby1,eq:luby2}.
\end{proof}
Then, in the $\mathsf{CONGEST}$\xspace model, we apply the rounding procedure of \Cref{lemma:d2rounding} with $\varepsilon=1/2$ on the fractional label assignment $\vec{x}\in [0,1]^{|V|}$ in Luby's algorithm. The algorithm runs in $O(\log^2 \Delta \cdot \log\log \Delta + \log \Delta \log^* n)$ rounds of the $\mathsf{CONGEST}$\xspace model. If we were in the more relaxed $\mathsf{LOCAL}$\xspace model which allows unbounded message sizes, then auxiliary graph $H$ can be directly simulated in graph $G$ with no asymptotic round complexity loss, and we can already invoke \Cref{lemma:distributedonG} which performs a rounding with the same guarantees in $O(\log^2 \Delta + \log \Delta \log^* n)$ rounds. The $\log^* n$ term can be replaced by a $O(\log^* \Delta)$, which thus makes the entire second additive term negligible, by computing a $\operatorname{\text{{\rm poly}}}(\Delta)$-color distance-$2$ coloring of $G$ at the beginning of the first iteration. The one-time $O(\log^* n)$ additive round complexity of computing this distance-$2$ coloring of $G$, which happens at the very beginning of the algorithm, is subsumed by our $O(\log n)$ number of iterations.
Suppose that we get an integral marking assignment $\vec{y}\in \{0,1\}^{|V|}$. From \Cref{lemma:d2rounding}, we know that $Z(\vec{y})=\mathbf{u}(\vec{y})-\mathbf{c}(\vec{y}) \geq (1/2) \cdot (\mathbf{u}(\vec{x})-\mathbf{c}(\vec{x})).$ Next, we argue that this implies $Z(\vec{y}) \geq |E|/500$.
\begin{claim}\label[claim]{clm:MIS2} For the fractional label assignment $\vec{x}\in [0,1]^{|V|}$ in Luby's algorithm --- i.e., where $x_v=1/(20\deg(v))$--- we have $Z(\vec{x}) = \mathbf{u}(\vec{x})-\mathbf{c}(\vec{x}) \geq |E| /240$. Hence, for the integral marking assignment we obtain from rounding $\vec{x}$ by invoking \Cref{lemma:d2rounding} in the $\mathsf{CONGEST}$\xspace model or \Cref{lemma:distributedonG} in the $\mathsf{LOCAL}$\xspace model, we have $Z(\vec{y})=\mathbf{u}(\vec{y})-\mathbf{c}(\vec{y}) \geq (1/2) \cdot (\mathbf{u}(\vec{x})-\mathbf{c}(\vec{x})) \geq |E|/500$.
\end{claim}
\begin{proof}From \Cref{clm:MIS1}, we have $Z(\vec{x}) \geq \mathbf{u}(\vec{x})-\mathbf{c}(\vec{x}) \geq \mathbf{u}(\vec{x})/2$. Hence,
\begin{align*} Z(\vec{x}) \geq \mathbf{u}(\vec{x})/2 &=
\sum_{\textit{good vertex \,} v} (\deg(v)/2) \bigg(\sum_{u\in IN^*(v)} x_u/2\bigg) \\
&\ge \sum_{\textit{good vertex \,} v} (\deg(v)/2) \cdot (1/120) \geq |E|/240,
\end{align*}
where we first used \cref{eq:luby1} that says that $\sum_{u \in IN^*(v)} x_u \ge 1/60$ and then we used \cref{eq:luby0} that bounds $\sum_{\textit{good vertex \,} v} \deg(v) \geq |E|/2$.
Since $Z(\vec{y})=\mathbf{u}(\vec{y})-\mathbf{c}(\vec{y}) \geq (1/2) \cdot (\mathbf{u}(\vec{x})-\mathbf{c}(\vec{x}))$, the claim follows.
\end{proof}
Hence, from the rounding procedure described above, which runs in $O(\log^2 \Delta \cdot \log\log \Delta + \log \Delta \cdot \log^* n)$ rounds of the $\mathsf{CONGEST}$\xspace model or $O(\log^2 \Delta + \log \Delta \log^* \Delta) = O(\log^2 \Delta)$ rounds of $\mathsf{LOCAL}$\xspace model, we get an integral marking assignment $\vec{y}$ with the following guarantee: if we add marked nodes $u$ that have no marked out-neighbor to the independent set and remove them along with their neighbors, we remove at least a $1/500$ fraction of the remaining edges. Hence, $O(\log n)$ such iterations suffice to complete the computation and have a maximal independent set. This completes the proof of \cref{thm:mis}.
\section{Generic Distributed Rounding Algorithm}
\label{sec:genericrounding}
\subsection{Setup: Multigraph with Edge Utilities and Costs}
\label{sec:roundingsetup}
Let $H=(V,E)$ be an undirected multigraph with no self-loops and let $\Sigma$ be a finite set of possible node labels. A label assignment for $H$ is a function $\ell: V\to \Sigma$ that assigns a label $\alpha=\ell(v)\in \Sigma$ to each node $v\in V$. A fractional label assignment $\lambda:V\to [0,1]^{|\Sigma|}$ for $H$ is an assignment of a distribution (i.e., convex combination) of labels to each node. That is, for each node $v\in V$, we have $\sum_{\alpha\in \Sigma}\lambda_\alpha(v)=1$, where we use $\lambda_\alpha(v)\in [0,1]$ to refer to the component corresponding to label $\alpha$ in the vector $\lambda(v)$. For an integer $K\geq 1$, a fractional label assignment $\lambda$ is called $1/K$-integral if for every $v\in V$ and every $\alpha\in \Sigma$, $\lambda_\alpha(v)=a/K$ for some integer $a\geq 0$. For a set of nodes $S\subseteq V$ and a label assignment $\ell$ (or a fractional label assignment $\lambda$), we use $\ell(S)$ (or $\lambda(S)$) to denote the (fractional) label assignment to the set of nodes in $S$. Similarly for an edge $e\in E$, we use $\ell(e)$ and $\lambda(e)$ to denote $\ell(V(e))$ and $\lambda(V(e))$. Further, for each node $v$, we use $\mathcal{L}(v)$ (and $\Lambda(v)$) to denote the set of possible (fractional) label assignments to $v$ and we similarly use $\mathcal{L}(S)$, $\Lambda(S)$, $\mathcal{L}(e)$, $\Lambda(e)$ for node sets $S$ and edges $e$.
For a multi-graph $H=(V,E)$, we define a non-negative utility function $\mathbf{u}:E\times \mathcal{L}(V) \to \mathbb{R}_{\geq 0}$ and a non-negative cost function $\mathbf{c}:E\times \mathcal{L}(V) \to \mathbb{R}_{\geq 0}$.
That is, for a given label assignment $\ell\in \mathcal{L}(V)$, $\mathbf{u}$ and $\mathbf{c}$ assign non-negative utility and cost values $\mathbf{u}(e,\ell)$ and $\mathbf{c}(e,\ell)$ to every edge $e\in E$. Although utility and cost of an edge $e$ are defined as a function of the label assignment to all nodes in $V$, they must only depend on the label assignment to the two nodes $V(e)$ of $e$. That is, for any two label assignments $\ell,\ell'\in \mathcal{L}(V)$ such that $\ell(e)=\ell'(e)$ (i.e., $\ell$ and $\ell'$ assign the same labels to the nodes in $V(e)$), it must hold that $\mathbf{u}(e,\ell)=\mathbf{u}(e,\ell')$ and $\mathbf{c}(e,\ell)=\mathbf{c}(e,\ell')$.
We slightly overload notation and also define $\mathbf{u}$ and $\mathbf{c}$ for a fractional label assignment $\lambda\in\Lambda(V)$. In this case, utility and cost of $e$ are defined as the expected values of $\mathbf{u}$ and $\mathbf{c}$ if the labels of the two nodes $V(e)$ of $e$ are chosen independently at random from the distributions given by the fractional label assignment. That is, if $V(e)=\set{u,v}$ and if for each $(\alpha,\beta)\in \Sigma^2$, $\ell_{\alpha,\beta}(e)$ is a label assignment that assigns label $\alpha$ to $u$ and label $\beta$ to $v$, then for $\lambda\in \Lambda(V)$, we have
\begin{equation}\label{eq:fractionalutilitycost}
\mathbf{u}(e, \lambda) := \!\!\sum_{(\alpha,\beta)\in \Sigma^2}\!\!
\lambda_\alpha(u)\cdot\lambda_\beta(v)\cdot
\mathbf{u}(e,\ell_{\alpha,\beta}(e))
\quad\text{and}\quad
\mathbf{c}(e, \lambda) := \!\!\sum_{(\alpha,\beta)\in \Sigma^2}\!\!
\lambda_\alpha(u)\cdot\lambda_\beta(v)\cdot
\mathbf{c}(e, \ell_{\alpha,\beta}(e)).
\end{equation}
Finally, for a set of edges $F\subseteq E$ and a label assignment $\ell$, we use $\mathbf{u}(F,\ell)=\sum_{e\in F}\mathbf{u}(e,\ell)$, and $\mathbf{c}(F,\ell)=\sum_{e\in F}\mathbf{c}(e,\ell)$. The definitions extend analogously if the label assignment $\ell$ is replaced with a fractional label assignment $\lambda$. Finally, to denote the total cost and utility of a label assignment $\ell$, we define $\mathbf{u}(\ell):=\mathbf{u}(E,\ell)$ and $\mathbf{c}(\ell):=\mathbf{c}(E,\ell)$. Again, the definition extends analogously if $\ell$ is replaced with a fractional label assignment $\lambda$.
\paragraph{Polynomially Bounded Instances.} In some cases, the message sizes (or the time complexities in case message sizes are bounded) of our algorithms depend on the range of values that utilities and costs can have. We say that a given instance (i.e., multigraph with corresponding utility and cost functions and with a fractional label assignment $\lambda$) is \emph{polynomially bounded in $q$} for some parameter $q>1$ if there is a value $Q\leq q^d$ for some constant $d>0$ such that every non-zero edge utility or cost is lower bounded by a value $1/Q$ and upper bounded by $Q$ and such that any non-zero fractional value $\lambda_\alpha(v)\geq 1/Q$. Most of our instances will be polynomially bounded in the maximum degree $\Delta$ of the graph.
\subsection{Abstract Basic Rounding Algorithm}
\label{sec:basicrounding}
In the following, we first describe our local rounding algorithm in an abstract form, which is somewhat independent of the specific communication model. In the concrete distributed implementation of the algorithm, we will assume that in the multigraph $G=(V,E)$ on which we run the rounding, the nodes $V$ are active entities that locally perform their part of the rounding algorithm. In some applications, some edges of $G$ might however not be physical communication links, so the communication between nodes in $V$ depends on the relation between $G$ and the underlying physical communication network. In \Cref{alg:basicroundingstep}, we first describe the algorithm for a single rounding step in detail. The full rounding procedure is then given by \Cref{alg:basicrounding}.
\begin{algorithm}[ht]
\caption{Basic Rounding Step, Parameters $\delta\in[0,1]$, $\eta\geq 1$}\label{alg:basicroundingstep}
\textbf{Given:} multigraph $G=(V,E)$ with utility and cost functions $\mathbf{u}$ and $\mathbf{c}$ and with a $1/(2K)$-integral fractional label assignment $\lambda$ (for a given integer $K\geq 1$).
\begin{algorithmic}[1]
\State For each $e\in E$, define weight $w_e := \mathbf{u}(e,\lambda)+ \eta\mathbf{c}(e,\lambda)$
\State Compute a weighted $\delta/6$-relative average defective $p$-coloring of $G$ w.r.t.\ the edge weights $w_e$
\State Define $E_b\subseteq E$ to be the set of bichromatic edges (w.r.t.\ the defective coloring)
\ForAll {colors $\gamma\in \set{1,\dots,p}$}
\ForAll {nodes $v\in V$ of color $\gamma$ (in the defective coloring) \textbf{in parallel}}
\State ${\Sigma}_v:=\set{\alpha\in \Sigma : \lambda_\alpha(v)=\frac{2i+1}{2K}\text{ for some integer }i\geq 0}$
\ForAll {labels $\alpha\in{\Sigma}_v$}
\State Define $\lambda^{(v,\alpha)}$ s.t.\ $\lambda^{(v,\alpha)}_\alpha(v)=1$, $\lambda^{(v,\alpha)}_\beta(v)=0$ for $\beta\neq \alpha$, and $\lambda^{(v,\alpha)}=\lambda$ otherwise
\State Define $\phi_{v,\alpha}:=\mathbf{u}\big(E_b(v),\lambda^{(v,\alpha)}\big)-\eta\cdot\mathbf{c}\big(E_b(v),\lambda^{(v,\alpha)}\big)$
\State Define $\theta_{v,\alpha}:=\mathbf{u}\big(E_b(v),\lambda^{(v,\alpha)}\big)+\eta\cdot\mathbf{c}\big(E_b(v),\lambda^{(v,\alpha)}\big)$.
\State $v$ obtains estimate $\hat{\phi}_{v,\alpha}$ satisfying $\phi_{v,\alpha}\geq \hat{\phi}_{v,\alpha}\geq
\phi_{v,\alpha} - \frac{\delta}{6}\cdot\theta_{v,\alpha}$
\EndFor
\State Partition ${\Sigma}_v$ into ${\Sigma}_v^+$ and ${\Sigma}_v^-$ s.t.\ $|{\Sigma}_v^+|=|{\Sigma}_v^-|$ and s.t.\ $\forall(\alpha,\beta)\in {\Sigma}_v^+\times {\Sigma}_v^-$, $\hat{\phi}_{v,\alpha}\geq \hat{\phi}_{v,\beta}$
\State For $\alpha\in {\Sigma}_v^+$, set $\lambda_\alpha(v):=\lambda_{\alpha}(v) + \frac{1}{2K}$ and for $\alpha\in {\Sigma}_v^-$, set $\lambda_\alpha(v):=\lambda_{\alpha}(v) - \frac{1}{2K}$
\EndFor
\EndFor
\end{algorithmic}
\textit{For line 12 of the algorithm, note that the set $\Sigma_v$ must have an even size. Therefore, it can be split into $\Sigma_v^+$ and $\Sigma_v^-$ such that $|{\Sigma}_v^+|=|{\Sigma}_v^-|$.}
\end{algorithm}
\begin{lemma}\label{lemma:basicroundingstep}
Let $G=(V,E)$ be a multigraph, which is equipped with utility and cost functions $\mathbf{u}$ and $\mathbf{c}$ and with a $1/(2K)$-integral fractional label assignment $\lambda$ for a given integer $K\geq 1$. Let $\delta\in[0,1]$ and $\eta\geq 1$ be two parameters. Assume that \Cref{alg:basicroundingstep} is run on $G$ with the given parameters and let $\lambda'$ be the fractional label assignment after running the algorithm. Then, $\lambda'$ is a $1/K$-integral fractional label assignment and
\[
\mathbf{u}(\lambda') - \eta\mathbf{c}(\lambda') \geq \mathbf{u}(\lambda) - \eta\mathbf{c}(\lambda) -
\delta\cdot\big(\mathbf{u}(\lambda)+\eta\mathbf{c}(\lambda)\big).
\]
\end{lemma}
\begin{proof}
First, the algorithm only changes fractional values $\lambda_\alpha(v)$ of the form $(2i+1)/(2K)$ and all those values are either decremented or incremented by exactly $1/(2K)$. The fractional label assignment at the end of the algorithm is thus $1/K$-integral.
For each color $\gamma\in\set{1,\dots,p}$ of the average defective coloring, we define $\lambda_\gamma$ to be the fractional label assignment after processing the nodes of color $\gamma$ in \Cref{alg:basicroundingstep}. We further define $\lambda_0$ to be the fractional label assignment at the beginning of the algorithm. Note that $\lambda_0=\lambda$ and $\lambda_p=\lambda'$.
We first upper bound how the utility and cost of an edge change during the algorithm. Consider an edge $e$ between two nodes $u$ and $v$ (i.e., $V(e)=\set{u,v}$). Assume that in the average defective coloring $u$ has color $\gamma_u$ and $v$ has color $\gamma_v$ and w.l.o.g., assume that $\gamma_u\leq \gamma_v$. Note that when rounding the fractional label assignment, the fractional value of a node for a given label can at most double. When rounding a single node, the utility and cost of its incident edges can therefore also at most double and when rounding both nodes of an edge, the utility and cost can grow by a factor at most $4$. For any $\gamma\in\set{0,\dots,p}$ and edge $e$ between $u$ and $v$, we therefore have
\begin{equation}\label{eq:edgeroundingupper}
\mathbf{u}(e,\lambda_\gamma) \leq
\begin{cases}
\mathbf{u}(e, \lambda) & \text{if } \gamma < \gamma_u\\
2\mathbf{u}(e, \lambda) & \text{if } \gamma_u\leq \gamma < \gamma_v\\
4\mathbf{u}(e, \lambda) & \text{if } \gamma \geq \gamma_v
\end{cases}
\qquad\text{and}\qquad
\mathbf{c}(e,\lambda_\gamma) \leq
\begin{cases}
\mathbf{c}(e, \lambda) & \text{if } \gamma < \gamma_u\\
2\mathbf{c}(e, \lambda) & \text{if } \gamma_u\leq \gamma < \gamma_v\\
4\mathbf{c}(e, \lambda) & \text{if } \gamma \geq \gamma_v.
\end{cases}
\end{equation}
As defined in \Cref{alg:basicroundingstep}, let $E_b$ the the set of bichromatic edges w.r.t.\ the computed average defective $p$-coloring and let $E_m:=E\setminus E_b$ be the set of monochromatic edges w.r.t.\ the same coloring. Note that we have $\mathbf{u}(\lambda)=\mathbf{u}(E,\lambda)=\mathbf{u}(E_b,\lambda)+\mathbf{u}(E_m,\lambda)$ and analogously also for the cost function $\mathbf{c}(\cdot)$ and the fractional assignment $\lambda'$. We first consider the contribution of the monochromatic edges to $\mathbf{u}(\lambda')-\eta\mathbf{c}(\lambda')$. We have
\begin{eqnarray}
\mathbf{u}(E_m, \lambda')-\eta\mathbf{c}(E_m, \lambda') & \stackrel{\mathbf{u}(\cdot)\geq 0, \eqref{eq:edgeroundingupper}}{\geq} & -4\eta\mathbf{c}(E_m, \lambda)\nonumber\\
& = & \mathbf{u}(E_m, \lambda) - \eta\mathbf{c}(E_m, \lambda) - \mathbf{u}(E_m, \lambda) - 3\eta\mathbf{c}(E_m, \lambda)\nonumber\\
& \geq & \mathbf{u}(E_m, \lambda) - \eta\mathbf{c}(E_m, \lambda) - 3\big(\mathbf{u}(E_m, \lambda) + \eta\mathbf{c}(E_m, \lambda)\big)\nonumber\\
& \geq & \mathbf{u}(E_m, \lambda) - \eta\mathbf{c}(E_m, \lambda) -\frac{\delta}{2}\cdot \big(\mathbf{u}(\lambda) + \eta\mathbf{c}(\lambda)\big).\label{eq:boundmonochromatic}
\end{eqnarray}
The last inequality follows because the given vertex coloring is a weighted $\delta/6$-average defective coloring w.r.t. edge weights $\mathbf{u}(e, \lambda) + \eta\mathbf{c}(\lambda)$ and thus $\mathbf{u}(E_m, \lambda) + \eta\mathbf{c}(E_m, \lambda)\leq \delta/6\cdot(\mathbf{u}(\lambda) + \eta\mathbf{c}(\lambda))$.
We next look at the contribution of the bichromatic edges to $\mathbf{u}(\lambda')-\eta\mathbf{c}(\lambda')$. For this, we first consider the effect of the rounding of a single node $v$. Assume that node $v$ has color $\gamma_v$ in the average defective coloring. Note that before iteration $\gamma_v$ of the for loop of \Cref{alg:basicroundingstep} in which the fractional assignment of $v$ is changed, the fractional label assignment is $\lambda_{\gamma_v-1}$ and directly after changing the fractional assignment of $v$, the fractional label assignment is $\lambda_{\gamma_v}$. Also note that by the definition of $\phi_{v,\alpha}$ in line 9 of \Cref{alg:basicroundingstep}, we have
\begin{equation}\label{eq:vpotentialbefore}
\mathbf{u}(E_b(v), \lambda_{\gamma_v-1}) - \eta\mathbf{c}(E_b(v), \lambda_{\gamma_v-1}) =
\sum_{\alpha\in\Sigma} \lambda_{\gamma_v-1, \alpha}(v)\cdot \phi_{v,\alpha},
\end{equation}
where $\lambda_{\gamma_v-1,\alpha}(v)$ denotes the fractional value that $v$ has for label $\alpha$ in the fractional label assignment $\lambda_{\gamma_v-1}$. When rounding the fractional values of $v$, the fractional values are incremented by $1/(2K)$ for all labels $\alpha\in \Sigma_v^+$ and the fractional values are decremented by $1/(2K)$ for all labels $\alpha\in\Sigma_v^-$. Because the neighbors of a color different from $\gamma_v$ do not change their fractional values at the same time, after rounding the values of $v$ (i.e., at the end of iteration $\gamma_v$ of the for loop), we have
\begin{eqnarray}
\mathbf{u}(E_b(v), \lambda_{\gamma_v}) - \eta\mathbf{c}(E_b(v), \lambda_{\gamma_v})
& = &
\sum_{\alpha\in\Sigma} \lambda_{\gamma_v, \alpha}\cdot \phi_{v,\alpha}\nonumber\\
& \stackrel{\eqref{eq:vpotentialbefore}}{=} &
\mathbf{u}(E_b(v), \lambda_{\gamma_v-1}) - \eta\mathbf{c}(E_b(v), \lambda_{\gamma_v-1})
+ \varphi_v.\label{eq:bichromaticchange}
\end{eqnarray}
where
\begin{equation}
\label{eq:vpotentialchange}
\varphi_v = \frac{1}{2K}\cdot \left[\sum_{\alpha\in\Sigma_v^+} \phi_{v,\alpha} - \sum_{\alpha\in\Sigma_v^-}\phi_{v,\alpha}\right].
\end{equation}
As in line 10 of \Cref{alg:basicroundingstep}, define $\theta_{v,\alpha}:=\mathbf{u}\big(E_b(v), \lambda_{\gamma_v-1}^{(v,\alpha)}\big)+\eta\mathbf{c}\big(E_b(v), \lambda_{\gamma_v-1}^{(v,\alpha)}\big)$, where $\lambda_{\gamma_v-1}^{(v,\alpha)}$ is a fractional label assignment that is equal to $\lambda_{\gamma_v-1}$, except that at node $v$, we have $\lambda_{\gamma_v-1,\alpha}(v)=1$ and $\lambda_{\gamma_v,\beta}(v)=0$ for $\beta\neq \alpha$. Note that similarly to \eqref{eq:vpotentialbefore} we have
\begin{equation}\label{eq:utilitycostsum}
\mathbf{u}\big(E_b(v), \lambda_{\gamma_v-1}\big)+\eta\mathbf{c}\big(E_b(v), \lambda_{\gamma_v-1}\big) =
\sum_{\alpha\in \Sigma} \lambda_{\gamma_v-1,\alpha}(v)\cdot \theta_{v,\alpha}.
\end{equation}
By line 11 of \Cref{alg:basicroundingstep}, for every $\alpha\in \Sigma_v^+\cup\Sigma_v^-$, $v$ obtains an estimate $\hat{\phi}_{v,\alpha}$ such that $\phi_{v,\alpha}\geq \hat{\phi}_{v,\alpha}\geq \phi_{v,\alpha}-\frac{\delta}{6}\cdot\theta_{v,\alpha}$. We can therefore rewrite $\varphi_v$ (cf.\ \Cref{eq:vpotentialchange}) as
\begin{eqnarray}
\varphi_v
& \geq &
\frac{1}{2K}\cdot \underbrace{\left[\sum_{\alpha\in\Sigma_v^+} \hat{\phi}_{v,\alpha} - \sum_{\alpha\in\Sigma_v^-}\hat{\phi}_{v,\alpha}\right]}_{\geq 0\text{ (by def.\ of $\Sigma_v^+$ and $\Sigma_v^-$)}}
- \frac{1}{2K}\cdot\sum_{\alpha\in\Sigma_v^-}\frac{\delta}{6}\cdot\theta_{v,\alpha}\nonumber\\
& \stackrel{(\lambda_{\gamma_v-1,\alpha}(v)\geq \frac{1}{2K})}{\geq} &
- \frac{\delta}{6}\cdot \sum_{\alpha\in\Sigma_v^-}
\lambda_{\gamma_v-1,\alpha}(v)\cdot \theta_{v,\alpha}\nonumber\\
& \stackrel{\eqref{eq:utilitycostsum}}{\geq} &
-\frac{\delta}{6}\cdot\left(
\mathbf{u}\big(E_b(v), \lambda_{\gamma_v-1}\big)+\eta\mathbf{c}\big(E_b(v), \lambda_{\gamma_v-1}\big)
\right).\label{eq:varphibound}
\end{eqnarray}
By \eqref{eq:bichromaticchange}, we have
\begin{equation}\label{eq:totalpotentialchange}
\mathbf{u}(E_b, \lambda') - \eta\mathbf{c}(E_b, \lambda') =
\mathbf{u}(E_b, \lambda) - \eta\mathbf{c}(E_b, \lambda) + \sum_{v\in V} \varphi_v
\end{equation}
By using \eqref{eq:varphibound}, the term $\sum_{v\in V}\varphi_v$ can be lower bounded as
\begin{eqnarray}
\sum_{v\in V}\varphi_v
& \geq &
-\frac{\delta}{6}\cdot\sum_{v\in V}
\left(\mathbf{u}\big(E_b(v), \lambda_{\gamma_v-1}\big)+\eta\mathbf{c}\big(E_b(v), \lambda_{\gamma_v-1}\big)\right)\nonumber\\
& = &
-\frac{\delta}{6}\cdot\sum_{e\in E_b}\sum_{v\in V(e)}
\left(\mathbf{u}\big(e, \lambda_{\gamma_v-1}\big)+\eta\mathbf{c}\big(e, \lambda_{\gamma_v-1}\big)\right)\nonumber\\
& \geq &
-\frac{\delta}{6}\cdot\sum_{e\in E_b}
3\big(\mathbf{u}(e, \lambda)+\eta\mathbf{c}(e,\lambda)\big)\label{eq:boundonincrease}\\
& = & -\frac{\delta}{2}\cdot\big(\mathbf{u}(E_b,\lambda)+\eta\mathbf{c}(E_b,\lambda)\big)
\ \geq\ -\frac{\delta}{2}\cdot\big(\mathbf{u}(\lambda)+\eta\mathbf{c}(\lambda)\big).\nonumber
\end{eqnarray}
Inequality \eqref{eq:boundonincrease} follows because by \eqref{eq:edgeroundingupper} for a bichromatic edge $e$ with $V(e)=\set{u,v}$ and $\gamma_u<\gamma_v$, we have $\mathbf{u}(e, \lambda_{\gamma_u-1})\leq \mathbf{u}(e,\lambda)$ and $\mathbf{u}(e,\lambda_{\gamma_v-1})\leq 2 \mathbf{u}(e,\lambda)$ and analogously $\mathbf{c}(e, \lambda_{\gamma_u-1})\leq \mathbf{c}(e,\lambda)$ and $\mathbf{c}(e,\lambda_{\gamma_v-1})\leq 2 \mathbf{c}(e,\lambda)$. The last inequality follows because utility and cost are non-negative functions. In combination with Equations \eqref{eq:boundmonochromatic} and \eqref{eq:totalpotentialchange}, we obtain
\[
\mathbf{u}(\lambda') - \eta\mathbf{c}(\lambda') \geq \mathbf{u}(\lambda) -\eta\mathbf{c}(\lambda) -
\delta\cdot\big(\mathbf{u}(\lambda)+\eta\mathbf{c}(\lambda)\big)
\]
as required by the claim of the lemma.
\end{proof}
\Cref{alg:basicroundingstep} describes a single rounding step of our abstract generic rounding algorithm. One application of this basic step doubles the integrality of a given fractional label assignment. Given a $1/K$-integral fractional label assignment, we therefore have to invoke \Cref{alg:basicroundingstep} $\log K$ times in order to obtain an integral label assignment. The details of this are given by \Cref{alg:basicrounding}.
\begin{algorithm}
\caption{Basic Rounding Algorithm, Parameters $\varepsilon\in[0,1]$, $\mu\in(0,1]$}\label{alg:basicrounding}
\textbf{Given:} multigraph $G=(V,E)$ with utility and cost functions $\mathbf{u}$ and $\mathbf{c}$ and with a $1/2^k$-integral fractional label assignment $\lambda$ (for a given integer $k\geq 1$) such that $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\geq \mu\mathbf{u}(\lambda)$.
\begin{algorithmic}[1]
\State Define $\delta:=\frac{\varepsilon\cdot\mu}{6k}$
\For{$i\in\set{1,\dots,k}$}
\State $\eta_i := 1+\left(1-\frac{i}{k}\right)\frac{\varepsilon\cdot\mu}{2}$
\State Run \Cref{alg:basicroundingstep} with param.\ $\delta$ and $\eta_i$ on the current $\frac{1}{2^{k-i+1}}$-integral fract.\ assignment $\lambda$
\EndFor
\State Define integral label assignment $\ell$ s.t.\ $\forall v\in V, \alpha\in \Sigma : \ell(v)=\alpha\ \Leftrightarrow\ \lambda_\alpha(v)=1$
\end{algorithmic}
\end{algorithm}
\begin{lemma}\label{lemma:basicrounding}
Let $G=(V,E)$ be a multigraph, which is equipped with utility and cost functions $\mathbf{u}$ and $\mathbf{c}$ and with a $1/2^k$-integral fractional label assignment $\lambda$ for a given integer $k\geq 1$. Let $\varepsilon\in [0,1]$ and $\mu\in(0,1]$ be two parameters. If $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\geq \mu\mathbf{u}(\lambda)$, \Cref{alg:basicrounding} returns an integral label assignment $\ell$ with
\[
\mathbf{u}(\ell)-\mathbf{c}(\ell) \geq (1-\varepsilon)\cdot \big(\mathbf{u}(\lambda) - \mathbf{c}(\lambda)\big).
\]
\end{lemma}
\begin{proof}
For the proof, we define $\lambda_0$ to be the $1/2^k$-integral fractional label assignment with which \Cref{alg:basicrounding} is started and for $i\in\set{1,\dots,k}$, we define $\lambda_i$ to be the fractional label assignment at the end of the $i^{\mathrm{th}}$ iteration of the for loop in \Cref{alg:basicrounding}. For all $i\in\set{0,\dots,k}$, we further define a potential
\[
\Phi_i := \mathbf{u}(\lambda_i) - \eta_i\cdot \mathbf{c}(\lambda_i),
\]
where $\eta_i =1 + \big(1-\frac{i}{k}\big)\frac{\varepsilon\cdot\mu}{2}$ for all $i\in\set{0,\dots,k}$.
By induction on $i$, we first show that a) $\lambda_i$ is a $2^i/2^k$-integral fractional label assignment and b)
\begin{equation}\label{eq:potentialinduction}
\Phi_i \geq (1-\delta)^i\cdot \Phi_0.
\end{equation}
\noindent\textit{Base Case $i=0$:} We know that $\lambda_0$ is a $1/2^k$-integral fractional label assignment and \Cref{eq:potentialinduction} certainly also holds for $i=0$. Thus, both a) and b) hold.
\medskip
\noindent\textit{Induction Step $i>0$:} The induction hypothesis implies that a) $\lambda_i$ is a $2^i/2^k$-integral fractional label assignment and b) $\Phi_i \geq (1-\delta)^i\cdot \Phi_0$. We need to show that a) $\lambda_{i+1}$ is a $2^{i+1}/2^k$-integral fractional label assignment and b) $\Phi_{i+1} \geq (1-\delta)^{i+1}\cdot \Phi_0$.
For a), notice that at the beginning of the ${i+1}^{\mathrm{st}}$ iteration of the for loop of \Cref{alg:basicrounding}, the current fractional label assignment is $\lambda_i$, which by the induction hypothesis is a $2^i/2^k$-integral fractional label assignment. By \Cref{lemma:basicroundingstep}, \Cref{alg:basicroundingstep} doubles the integrality of the given fractional label assignment. Hence, when applying \Cref{alg:basicroundingstep} in iteration $i+1$ of \Cref{alg:basicrounding}, we obtain a $2^{i+1}/2^k$-integral fractional label assignment and thus a) follows.
For b), by \Cref{lemma:basicroundingstep}, at the end of the $i+1^{\mathrm{st}}$ iteration of the for loop in \Cref{alg:basicrounding}, we have
\begin{eqnarray*}
\Phi_{i+1} & = & \mathbf{u}(\lambda_{i+1}) - \eta_{i+1} \mathbf{c}(\lambda_{i+1})\\
& \stackrel{(L.\ \ref{lemma:basicroundingstep})}{\geq} & \mathbf{u}(\lambda_{i}) - \eta_{i+1} \mathbf{c}(\lambda_{i})- \delta \cdot ( \mathbf{u}(\lambda_{i}) + \eta_{i+1} \mathbf{c}(\lambda_{i}))\\
& = & \mathbf{u}(\lambda_{i}) - \eta_{i}\cdot \mathbf{c}(\lambda_{i})+ (\eta_{i}-\eta_{i+1}) \cdot \mathbf{c}(\lambda_{i}) -\delta\cdot ( \mathbf{u}(\lambda_{i}) + \eta_{i+1}\cdot \mathbf{c}(\lambda_{i}))\\
& = & (1-\delta)\cdot\underbrace{\big(\mathbf{u}(\lambda_{i}) - \eta_{i}\cdot \mathbf{c}(\lambda_{i})\big)}_{=\Phi_i} + \underbrace{(\eta_{i}-\eta_{i+1})}_{=\frac{\varepsilon\cdot\mu}{2k}} \cdot \mathbf{c}(\lambda_{i}) -\delta \cdot \underbrace{(\eta_i + \eta_{i+1})}_{\leq 3}\cdot\mathbf{c}(\lambda_i)\\
& \geq & (1-\delta) \cdot \Phi_i
+ \mathbf{c}(\lambda_{i}) \cdot \left( \frac{\varepsilon \cdot \mu}{2 k} - 3\cdot \delta \right)\\
& \stackrel{\big(\delta=\frac{\varepsilon\cdot\mu}{6k}\big)}{=} & (1-\delta) \cdot \Phi_i \\
& \stackrel{(\text{I.H.})}{\geq} & (1-\delta)^{i+1} \cdot \Phi_0.
\end{eqnarray*}
In the fourth line, $\eta_i+\eta_{i+1}\leq 3$ holds because by definition, $\eta_j\leq 1 + \frac{\varepsilon\cdot\mu}{2}\leq \frac{3}{2}$ for all $j\in \set{0,\dots,k}$. Hence, \Cref{eq:potentialinduction} holds and thus the induction step is complete.
We next show that for the initial potential $\Phi_0$, it holds that
\begin{equation}\label{eq:initialpotential}
\Phi_0 \geq \left(1-\frac{\varepsilon}{2}\right)\cdot\big(\mathbf{u}(\lambda_0)-\mathbf{c}(\lambda_0)\big).
\end{equation}
We need to show that $\Phi_0 = \mathbf{u}(\lambda_0) - \eta_0\cdot \mathbf{c}(\lambda_0) \geq \left(1-\frac{\varepsilon}{2}\right)\cdot\big(\mathbf{u}(\lambda_0)-\mathbf{c}(\lambda_0)\big)$, which is equivalent to showing that $ \eta_0\cdot \mathbf{c}(\lambda_0) \leq \mathbf{u}(\lambda_0) -\left(1-\frac{\varepsilon}{2}\right)\cdot\big(\mathbf{u}(\lambda_0)-\mathbf{c}(\lambda_0)\big)
= \frac{\varepsilon}{2} \cdot (\mathbf{u}(\lambda_0)- \mathbf{c}(\lambda_0))+\mathbf{c}(\lambda_0)$.
This is equivalent to showing that
$(\eta_0-1) \cdot \mathbf{c}(\lambda_0) \leq \frac{\varepsilon}{2} \cdot (\mathbf{u}(\lambda_0)- \mathbf{c}(\lambda_0))$, which follows because
\begin{eqnarray*}
(\eta_0-1) \cdot \mathbf{c}(\lambda_0)
& \leq & (\eta_0-1) \cdot (1-\mu) \cdot \mathbf{u}(\lambda_0)\\
& = & \left(\left(1+\frac{\varepsilon \cdot \mu}{2}\right)-1\right) \cdot (1-\mu) \cdot \mathbf{u}(\lambda_0)\\
& \leq & \frac{\varepsilon \cdot \mu}{2}\cdot \mathbf{u}(\lambda_0)\\
& \leq & \frac{\varepsilon}{2}\cdot \big(\mathbf{u}(\lambda_0)- \mathbf{c}(\lambda_0)\big).
\end{eqnarray*}
The first and last inequality follow since we have that $\mathbf{u}(\lambda_0)-\mathbf{c}(\lambda_0)\geq \mu\mathbf{u}(\lambda_0)$.
The claim of the lemma now follows because
\begin{eqnarray*}
\mathbf{u}(\lambda_k) - \mathbf{c}(\lambda_k) & = & \Phi_k\ \stackrel{\eqref{eq:potentialinduction}}{\geq}\
(1-\delta)^k\cdot \Phi_0\\
& \stackrel{\eqref{eq:initialpotential}}{\geq} &
(1-\delta)^k\cdot\left(1-\frac{\varepsilon}{2}\right)\cdot \big(\mathbf{u}(\lambda_0)-\mathbf{c}(\lambda_0)\big)\\ & \geq & (1-\varepsilon)\cdot\big(\mathbf{u}(\lambda_0)-\mathbf{c}(\lambda_0)\big).
\end{eqnarray*}
The last inequality follows because $\delta=\frac{\varepsilon\mu}{8k}\leq \frac{\varepsilon}{2k}$ and thus $(1-\delta)^k\geq 1-\frac{\varepsilon}{2}$.
\end{proof}
\subsubsection{Preprocessing of Fractional Label Assignments}
\label{sec:preprocessingassignment}
\Cref{alg:basicrounding} requires the given fractional solution to be $1/2^k$-integral for some positive integer $k$. It is often more natural to write down a fractional solution by using general real fractional values. The following simple technical lemma shows how a given fractional label assignment can be rounded to a $1/2^k$-fractional label assignment.
\begin{lemma}\label{lemma:fractionalrounding}
Let $G=(V,E)$ be a multigraph, which is equipped with utility and cost functions $\mathbf{u}$ and $\mathbf{c}$ and with a fractional label assignment $\lambda$ such that for all $v\in V$ and all $\alpha\in \Sigma$, $\lambda_\alpha(v)=0$ or $\lambda_\alpha(v)\geq \lambda_{\min}$ for some given value $\lambda_{\min}$. Let $\varepsilon\in [0,1]$ and $\mu\in(0,1]$ be two parameters. If $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\geq \mu\mathbf{u}(\lambda)$, the nodes can internally (i.e., without communication) compute a new $1/2^k$-integral fractional label assignment $\lambda'$ for some integer $k$ such that $2^k=O\big(\frac{1}{\varepsilon\cdot\mu\cdot \lambda_{\min}}\big)$,
\[
\mathbf{u}(\lambda')-\mathbf{c}(\lambda') \geq (1-\varepsilon)\cdot \big(\mathbf{u}(\lambda) - \mathbf{c}(\lambda)\big),
\quad\text{and}\quad
\mathbf{u}(\lambda')-\mathbf{c}(\lambda') \geq \frac{\mu}{2}\cdot\mathbf{u}(\lambda').
\]
\end{lemma}
\begin{proof}
Clearly, for a given integer $k\geq 0$, each node $v\in V$ can compute new values $\lambda_\alpha'(v)$ such that for every $\alpha\in\Sigma$, $\lambda_\alpha'(v)=0$ if $\lambda_\alpha(v)=0$ and otherwise, $\lambda_\alpha'(v)$ is an integer multiple of $2^{-k}$ such that $|\lambda_\alpha(v) - \lambda_\alpha'(v)|\leq 2^{-k}$. Node $v$ can just round up or down each $\lambda_\alpha(v)$ to the nearest larger or smaller integer multiple of $2^{-k}$ such that the total sum of all fractional values of $v$ remains $1$. For all $v\in V$ and $\alpha \in \Sigma$, we then have
\begin{equation}\label{eq:roundedfractvalues}
\left(1-\frac{1}{\lambda_{\min}\cdot 2^k}\right)\cdot \lambda_\alpha(v)
\ \leq\ \lambda_\alpha'(v)\ \leq\
\left(1+\frac{1}{\lambda_{\min}\cdot 2^k}\right)\cdot \lambda_\alpha(v).
\end{equation}
Recall that utility and cost are computed as a sum over the individual edge utilities and costs. By the definition of the utility and cost of an edge for a fractional assignment (\Cref{eq:fractionalutilitycost}), for an edge $e\in E$ with $V(e)=\set{u,v}$, we have
\begin{eqnarray}
\mathbf{u}(e, \lambda')
& \stackrel{\eqref{eq:fractionalutilitycost}}{\geq} &
\min_{(\alpha,\beta)\in\Sigma^2}\frac{\lambda_{\alpha}'(u)\cdot\lambda_{\beta}'(v)}{\lambda_{\alpha}(u)\cdot\lambda_{\beta}(v)}\cdot
\mathbf{u}(e, \lambda)
\ \stackrel{\eqref{eq:roundedfractvalues}}{\geq}
\left(1-\frac{1}{\lambda_{\min}\cdot 2^k}\right)^2\cdot \mathbf{u}(e, \lambda),\label{eq:edgeutilitylower}\\
\mathbf{u}(e, \lambda')
& \stackrel{\eqref{eq:fractionalutilitycost}}{\leq} &
\max_{(\alpha,\beta)\in\Sigma^2}\frac{\lambda_{\alpha}'(u)\cdot\lambda_{\beta}'(v)}{\lambda_{\alpha}(u)\cdot\lambda_{\beta}(v)}\cdot
\mathbf{u}(e, \lambda)
\ \stackrel{\eqref{eq:roundedfractvalues}}{\leq}
\left(1+\frac{1}{\lambda_{\min}\cdot 2^k}\right)^2\cdot \mathbf{u}(e, \lambda). \label{eq:edgeutilityupper}
\end{eqnarray}
We choose $k$ as the smallest integer such that $2^k\geq \frac{9}{\varepsilon\cdot\mu\cdot\lambda_{\min}}$. We then get
\begin{eqnarray}
\mathbf{u}(\lambda')
& = &
\sum_{e\in E} \mathbf{u}(e, \lambda')
\ \geq\
\left(1-\frac{\varepsilon\cdot\mu}{9}\right)^2\cdot\sum_{e\in E} \mathbf{u}(e, \lambda)
\ \geq\ \left(1-\frac{\varepsilon\cdot\mu}{4}\right)\mathbf{u}(\lambda), \label{eq:utilitylower}\\
\mathbf{u}(\lambda')
& = &
\sum_{e\in E} \mathbf{u}(e, \lambda')
\ \leq\
\left(1+\frac{\varepsilon\cdot\mu}{9}\right)^2\cdot\sum_{e\in E} \mathbf{u}(e, \lambda)
\stackrel{(\varepsilon\mu\leq 1)}{\leq} \left(1+\frac{\varepsilon\cdot\mu}{4}\right)\mathbf{u}(\lambda).\label{eq:utilityupper}
\end{eqnarray}
Clearly, Equations \eqref{eq:edgeutilitylower}, \eqref{eq:edgeutilityupper}, \eqref{eq:utilitylower}, and \eqref{eq:utilityupper} also hold for the cost function $\mathbf{c}(\cdot)$ in exactly the same way. We therefore obtain
\begin{eqnarray*}
\mathbf{u}(\lambda') - \mathbf{c}(\lambda')
& \stackrel{(\eqref{eq:utilityupper}, \eqref{eq:utilitylower}\text{ for }\mathbf{c}(\cdot))}{\geq} &
\mathbf{u}(\lambda) - \mathbf{c}(\lambda) - \frac{\varepsilon \mu}{4}\cdot\big(\mathbf{u}(\lambda) + \mathbf{c}(\lambda)\big)\\
& \stackrel{(\mathbf{c}(\lambda)\leq\mathbf{u}(\lambda))}{\geq} &
\mathbf{u}(\lambda) - \mathbf{c}(\lambda) - \frac{\varepsilon \mu}{2}\cdot\mathbf{u}(\lambda)\\
& \stackrel{(\mathbf{u}(\lambda)\leq \frac{1}{\mu}\cdot(\mathbf{u}(\lambda)-\mathbf{c}(\lambda))}{\geq} &
(1-\varepsilon)\cdot\big(\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\big).
\end{eqnarray*}
Hence, the first lower bound on $\mathbf{u}(\lambda')-\mathbf{c}(\lambda')$ claimed by the lemma holds. We further get
\begin{eqnarray*}
\mathbf{u}(\lambda') - \mathbf{c}(\lambda')
& \stackrel{(\eqref{eq:utilityupper}\text{ for }\mathbf{c}(\cdot))}{\geq} &
\mathbf{u}(\lambda') - \left(1+\frac{\varepsilon\mu}{4}\right)\cdot\mathbf{c}(\lambda)\\
& \stackrel{(\mathbf{c}(\lambda)\leq (1-\mu)\mathbf{u}(\lambda))}{\geq} &
\mathbf{u}(\lambda') - \left(1+\frac{\varepsilon\mu}{4}\right)\cdot(1-\mu)\cdot\mathbf{u}(\lambda)\\
& \stackrel{\eqref{eq:utilityupper}}{\geq} &
\mathbf{u}(\lambda') - \left(1+\frac{\varepsilon\mu}{4}\right)^2\cdot(1-\mu)\cdot\mathbf{u}(\lambda')\\
& = &
\left[\left(1-\frac{\varepsilon}{2}\right)\cdot\mu + \left(\frac{\varepsilon}{2}-\frac{\varepsilon^2}{16}\right)\cdot\mu^2 + \frac{\varepsilon^2\mu^3}{16}\right]
\cdot \mathbf{u}(\lambda')
\ \stackrel{(\varepsilon\leq 1)}{\geq} \
\frac{\mu}{2}\cdot \mathbf{u}(\lambda').
\end{eqnarray*}
Hence, also the second claim of the lemma holds.
\end{proof}
\subsubsection{Node Utilities and Costs}
\label{sec:nodeutilitiesandcosts}
In some applications, a part of the cost and/or utility only depends on the labeling of individual nodes and not on the combined labeling of neighboring nodes. In such cases, it is natural to extend the cost and utility functions and also define the utility/cost of single nodes. The following lemma shows that node utilities and costs can be easily incorporated into our rounding framework.
\begin{lemma}\label{lemma:nodeutilitycost}
Let $G=(V,E)$ be a multigraph, which is equipped with utility and cost functions $\mathbf{u}$ and $\mathbf{c}$. Assume that, the functions $\mathbf{u}$ and $\mathbf{c}$ are extended to also assign a non-negative utility and cost to each node $v\in V$, where the utility and cost of a node is a function of the node's label. \Cref{alg:basicroundingstep,alg:basicrounding} can be adapted to incorporate node utilities and costs so that \Cref{lemma:basicroundingstep,lemma:basicrounding} hold accordingly.
\end{lemma}
\begin{proof}
Essentially, a node $v\in V$ can distribute its utility and cost arbitrarily among its edges and then use \Cref{alg:basicroundingstep,alg:basicrounding} to do the rounding. To make sure that handling node utilities and costs is independent of the concrete communication model, we do this as follows. For every node $v\in V$, we define a virtual dummy node $\tilde{v}$, which is simulated by node $v$. Nodes $v$ and $\tilde{v}$ are connected by a virtual edge $\tilde{e}_v$ and $\tilde{e}_v$ is the only edge of node $\tilde{v}$. The utility and cost of edge $\tilde{e}_v$ is always equal to the utility and cost of node $v$. The edge utility and cost of $\tilde{e}_v$ therefore only depend on the (fractional) label assignment to $v$ and are independent of the label of $\tilde{v}$. We can therefore initially assign an arbitrary label $\alpha$ to node $\tilde{v}$ (i.e., if we start with a fractional label assignment $\lambda$, then $\lambda_\alpha(\tilde{v})=1$ and $\lambda_\beta(\tilde{v})=0$ for all $\beta\neq \alpha$. In this way, the fractional label assignment of $\tilde{v}$ is never changed during the rounding process. When computing the defective coloring, we always assign $\tilde{v}$ a color that is different from $v$'s color so that the edge $\tilde{e}_v$ is bichromatic. Note that we can clearly assume that the defective coloring always has at least $2$ colors as otherwise all edges would be monochromatic and the rounding could not satisfy any non-trivial guarantees.
\end{proof}
\subsection{Distributed Implementation of Basic Rounding Algorithm}
\label{sec:distributedrounding}
As long as every edge $e\in E$ in the multigraph $G=(V,E)$ is between nodes that are close in the underlying communication graph, implementing \Cref{alg:basicroundingstep,alg:basicrounding} is straightforward in the $\mathsf{LOCAL}$\xspace model. We however need to be more careful if the communication links in the underlying communication graph have limited capacity (e.g., when implementing the algorithms in the $\mathsf{CONGEST}$\xspace model).
\subsubsection[Direct Communication on $G$]{Direct Communication on \boldmath$G$}
\label{sec:directcommunication}
We first assume that communication is done directly between neighbors in $G$ and we explicitly analyze the necessary maximum message size for implementing our rounding algorithm.
\begin{lemma}\label{lemma:distributedonG}
Let $G=(V,E)$ be a multigraph, which is equipped with utility and cost functions $\mathbf{u}$ and $\mathbf{c}$ and with a fractional label assignment $\lambda$ such that for every label $\alpha\in \Sigma$ and every $v\in V$, if $\lambda_\alpha(v)>0$, then $\lambda_\alpha(v)\geq \lambda_{\min}$ for some given value $\lambda_{\min}\in(0,1]$. Further assume that $G$ is equipped with a proper $\zeta$-vertex coloring. Let $\varepsilon\in [0,1]$ and $\mu\in(0,1]$ be two parameters. If $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\geq \mu\mathbf{u}(\lambda)$ and if each node knows the utility and cost functions of its incident edges, there is a deterministic $O\big(\frac{1}{\varepsilon \mu}\cdot \log^2\big(\frac{1}{\varepsilon\mu\lambda_{\min}}\big)+ \log\big(\frac{1}{\varepsilon\mu\lambda_{\min}}\big)\cdot \log^* \zeta\big)$-round distributed message passing algorithm on $G$ that computes an integral label assignment $\ell$ for which
\[
\mathbf{u}(\ell)-\mathbf{c}(\ell) \geq (1-\varepsilon)\cdot \big(\mathbf{u}(\lambda) - \mathbf{c}(\lambda)\big).
\]
The algorithm uses messages of at size most $O\big(\min\set{|\Sigma|, \log|\Sigma|/(\varepsilon\mu\lambda_{\min})}+\log\zeta\big)$ bits. If the given fractional label assignment $\lambda$ is $1/2^k$-integral for some integer $k\geq 1$, the round complexity of the algorithm is $O\big(\frac{k^2}{\varepsilon\mu} + k \log^*\zeta\big)$ and the maximum message size is $O\big(\min\set{|\Sigma|, 2^k\log|\Sigma|}+\log\zeta\big)$ bits.
\end{lemma}
\begin{proof} The algorithm is \Cref{alg:basicrounding} which itself is simply a number of invocations of \Cref{alg:basicroundingstep}.
In the beginning, each node sends its initial fractional label assignments to all of its neighbors. This can be done using messages of size $O(|\Sigma| k)$, where we use $O(k)$ bits to encode the fractional value of each label. Since the runtime of the overall algorithm exceeds $\Theta(k)$, without any loss in round complexity, we can perform this communication by using $k$ rounds and sending one message of size $O(|\Sigma|)$ per round. When $2^k<|\Sigma|$, a more efficient method would be to send $O(\log|\Sigma|)$ bits for each of the at most $2^k$ non-zero fractional values, for a total of $O(2^{k} log |\Sigma|)$ bits. These share the initial fractional label assignments in at most $k$ rounds using messages of size $O\big(\min\set{|\Sigma|, 2^k\log|\Sigma|}\big)$ bits.
Then, we have $k$ rounding steps, where we invoke \Cref{alg:basicroundingstep}. Per step, in line 2 of \Cref{alg:basicroundingstep}, we first compute an average defective coloring with defect $\delta/6=\frac{\varepsilon \mu}{36k}$, in time $O\big(\frac{k}{\varepsilon\mu} + \log^*\zeta\big)$ using messages of $O(\log\zeta)$ bits, via the average weighted defective coloring algorithm of Ghaffari and Kuhn~\cite[Lemma 2.3]{GhaffariK21}. The produced coloring has $O\big(\frac{k}{\varepsilon\mu}\big)$ colors and therefore the inner loop in line 4 of \Cref{alg:basicroundingstep} has $O\big(\frac{k}{\varepsilon\mu}\big)$ iterations.
In each iteration, it is needed that each node $v$ knows the fractional label assignment of all its neighbors. In the first rounding step, this is given as nodes shared their initial fractional label assignments at the beginning, as we discussed above. After that, in each rounding step, the value of each label either decreases by a $ 2$ factor, increases by a $2$ factor, or remains as it was. Hence, each node can inform its neighbors about the update to its fractional label assignment using messages with $O(1)$ bits per label, i.e., $O(|\Sigma|)$. Alternatively, as above, when $2^k<|\Sigma|$, a more efficient method would be to send $O(\log|\Sigma|)$ bits for each of the at most $2^k$ non-zero fractional values, for a total of $O(2^{k} \log |\Sigma|)$ bits. Hence, the fractional value updates can be performed using messages of size at most $O\big(\min\set{|\Sigma|, 2^k\log|\Sigma|}\big)$ bits. Then, given that each node $v$ knows the fractional label assignment of its neighbors, it can compute $\phi_{v,\alpha}$ and $\theta_{v,\alpha}$, in lines 9 and 10 of \Cref{alg:basicroundingstep}. It thus can perform the rest of the computations of this rounding step locally, after which it can inform its neighbors about its fractional label assignment update using a message of size $O\big(\min\set{|\Sigma|, 2^k\log|\Sigma|}\big)$ bits, as discussed above.
Since there are $k$ rounding steps and each takes $O\big(\frac{k}{\varepsilon\mu} + \log^*\zeta\big)$ rounds, the claimed complexity follows.
\end{proof}
\subsubsection[Communication on $G^2$]{Communication on \boldmath$G^2$}
\label{sec:d2communication}
In some cases, the graph on which one runs the rounding algorithm is not equal to the communication graph $G$. For example, for computing an MIS in \Cref{sec:MIS} or a set cover in \Cref{sec:setcover}, our cost functions has contributions by pairs of nodes that are not direct neighbors in $G$, but which have a common neighbor in $G$. We generically analyze such a communication setting here.
\paragraph{Communication Model.} Formally, we assume that $G=(V,E)$ is the communication graph and that the rounding is done on a virtual multigraph $H=(V_H,E_H)$ with the following properties.
\begin{definition}[d2-Multigraph]\label{def:d2multigraph}
A \emph{d2-multigraph} is a multigraph $H=(V_H,E_H)$ that is simulated on top of an underlying communication graph $G=(V,E)$ by a distributed message-passing algorithm on $G$. The nodes of $H$ are a subset of the nodes of $G$, i.e., $V_H\subseteq V$. The edge set $E_H$ consists of two kinds of edges, \emph{physical edges} and \emph{virtual edges}. Physical edges in $E_H$ are edges between direct neighbors in $G$. For each physical edge in $e\in E_H$ with $V(e)=\set{u,v}$, both nodes $u$ and $v$ know about $e$. Virtual edges in $E_H$ are edges between two nodes $u,v\in V_H$ for which there is a common neighbor in $G$. Each virtual edge between two nodes $u$ and $v$ is known and managed by a common neighbor $u$ and $v$. We define a function $\xi:E_H\to V$ to refer to the node managing the virtual edge.
\end{definition}
We will always assume that the nodes know all relevant information about the edges of $H$ they manage.
For example, if we run an algorithm on a graph $H$ with edge weights $w:E_H\to\mathbb{R}_{\geq 0}$, then for every physical edge $e$ between $u$ and $v$, we assume that $u$ and $v$ know $w(e)$ and if $e$ is a virtual edge between $u$ and $v$, the common neighbor $\xi(e)$ of $u$ and $v$ knows $w(e)$. Similarly, when running an instance of the rounding algorithm of \Cref{sec:basicrounding} on $H$, the node(s) simulating an edge $e\in E_H$ know the utility and cost functions for edge $e$. Note also that nodes $u$ and $v$ might not be aware of virtual edges between them and that $H$ might have several edges between two nodes $u$ and $v$. If $u$ and $v$ are neighbors in $G$, there can be one physical edge between $u$ and $v$. If $u$ and $v$ have common neighbors in $G$, there can potentially be a virtual edge between $u$ and $v$ for each common neighbor of $u$ and $v$ in $G$. We could also allow multiple physical edges or multiple virtual edges for the same common neighbor in $G$. However such edges can typically also easily be aggregated into a single physical edge and one virtual edge per common neighbor in $G$. In the following, we assume standard synchronous, distributed message-passing algorithms on $G$, where in each round, a message of a certain size can be exchanged over each edge in $G$ and we show how to utilize this communication to implement \Cref{alg:basicroundingstep,alg:basicrounding} on the virtual multigraph $H$.
\paragraph{Computation of a Defective Coloring.}
As an important part of \Cref{alg:basicroundingstep}, the nodes have to compute a weighted average defective coloring of $H$. We first analyze how efficiently such a defective coloring can be computed. In the following, we say that an edge weight function $w$ is polynomially bounded in $q$ if there is a value $Q\leq \operatorname{\text{{\rm poly}}}(m)$ such that for every edge $e$, $w(e)=0$ or $w(e)\in[1/Q, Q]$. Note that the following lemma statement contains a somewhat artificial technical condition $\delta > 2^{-O(\sqrt{\log n})}$ on the relative defect parameter $\delta$. This condition is added to make the bounds look slightly nicer and cleaner. At the cost of a minor additional term in the round complexity, the condition could also be dropped. Also, the condition could also be strengthened to $\delta>2^{-O(\log^\vartheta n)}$ for any constant $\vartheta<1$ and in fact even for some $\vartheta=1-o(1)$. Note than in our applications the value of $\delta$ is typically $1/\operatorname{\text{{\rm poly}}}\log \Delta$ or $1/\operatorname{\text{{\rm poly}}}\log n$.
\begin{lemma}\label{lemma:d2defective}
Let $H=(V_H,E_H)$ be an $n$-node d2-multigraph with edge weights $w:E_H\to \mathbb{R}_{\geq 0}$, assume that $G=(V,E)$ is the underlying communication graph, and let $\Delta$ denote the maximum degree of $G$. Assume that the edge weights $w(e)$ are polynomially bounded in $q\geq \Delta$ and let $\delta > 2^{-O(\sqrt{\log n})}$ be a parameter. A weighted average $\delta$-relative defective $O(1/\delta)$-coloring of $H$ can be computed deterministically $O\big(\frac{\log\log q}{\delta}+\log^* n\big)$ $\mathsf{CONGEST}$\xspace rounds on $G$.
\end{lemma}
\begin{proof}
As a first step, we describe how to compute a weighted $\delta$-relative defective $O(1/\delta^2)$-coloring. For this, we adapt an algorithm that has been described in \cite{Kuhn2009WeakColoring} and that has been adapted to the weighted case in \cite{KawarabayashiS18}. The algorithm is based on a classic $O(\log^* n)$-round $O(\Delta^2)$-coloring algorithm of Linial~\cite{linial1987LOCAL}. The algorithm starts with an initial (defective) vertex coloring (in our case given by the unique $O(\log n)$-bit IDs) and consists of $O(\log^* n)$ consecutive steps. In each step, the coloring is improved in the following way. As a function of its current color (and without communicating), each node computes a set of candidate colors for the next coloring. Each node then picks a candidate color that (approximately) minimizes the total weight to neighbors (of different initial colors) that also have this candidate color. If in each step of the algorithm, a node chooses a candidate color for which the total weight to neighbors with the same candidate color is within a factor $2$ of the minimum, the relative defect in the end also only grows by a factor $2$. In order to determine its new color, in each step, each node, therefore for each of its candidate colors needs to approximately learn the total weight of its edges to neighboring nodes that also have this candidate color.
Before making the algorithm more concrete, let us focus on the communication required from a single step. First note that since node colors can always be represented by $O(\log n)$ bits (we start with $O(\log n)$-bit IDs), in a single round on the communication graph $G$, each node can learn the colors of all its $G$-neighbors. This implies that each node knows the candidate colors of all its $G$-neighbors. Assume now that in a given step, each node $v$ has at most $p$ candidate colors. Therefore, for each of its at most $p$ candidate colors $z$, each node $v\in V_H$ needs to learn an estimate of the total weight of its edges in $E_H$ to nodes that also have $z$ as a candidate color. Because $v$ knows the candidate colors of its $G$-neighbors, it exactly knows the total weight of the physical edges to neighbors with $z$ as a candidate color. Let us consider a virtual edge $e$ between node $v$ and some other node $u$. The common $G$-neighbor $\xi(e)$ of $u$ and $v$ knows the weight of $e$ and $\xi(e)$ also knows that candidate sets of both $v$ and $u$ and potentially also of other virtual edges of $v$ for which it is responsible. For each candidate color $z$ of $v$, node $\xi(e)$ can aggregate the total weight of $v$'s virtual edges for which $\xi(e)$ is responsible and which go to nodes that also have candidate color $z$. Because the weights are polynomially bounded in $q\geq \Delta$, the total weight of those virtual edges is either $0$ or a value between $1/\operatorname{\text{{\rm poly}}}(q)$ and $\operatorname{\text{{\rm poly}}}(q)$. Since $v$ only needs to learn the weight to other nodes with candidate color $z$ up to a factor $2$, it suffices to communicate one of the $O(\log q)$ different possible values over each edge of $G$. Therefore, for each candidate color, the algorithm has to send $O(\log\log q)$ bits over each $G$-edge. A defective coloring step with at most $p$ candidate colors can therefore be carried out in $O\big(1+\frac{p\log\log q}{\log n}\big)$ rounds in the $\mathsf{CONGEST}$\xspace model.
Let us now look at the details of the individual steps of the defective $O(1/\delta^2)$-coloring algorithm. In \cite{linial1987LOCAL} (also cf.\ \cite{erdos82,MausTonoyan20}), it is shown that for positive integer parameters $s$ and $N$ and a sufficiently large constant $c>0$, there are $N$ sets $S_1,\dots,S_N$ such that
\[
\forall i\in [N]: S_i\in [s^2\cdot \tau],\
|S_i| \leq s\cdot \tau,\ \text{and}\
\forall i\neq j: |S_i\cap S_j| \leq \tau,\ \text{where}\
\tau = c\cdot \min\set{\log N, \log_s^2 N}.
\]
Note that if we have a proper $N$-coloring of the nodes, and when choosing $s=1/\delta$, we can compute a weighted $\delta$-defective $\tau/\delta^2$-coloring as follows. Each node $v$ of initial color $i\in [N]$, picks set $S_i$ as its set of candidate colors. Because the candidate sets of nodes of different colors (and thus of neighbors) intersect in at most $\tau$ colors, each neighbor of $v$ only shares at most a $\delta$-fraction of $v$'s candidate colors. Therefore, there must be a color among $v$'s candidate colors such that the total weight of $v$'s edges to neighbors that also have this candidate color is at most a $\delta$-fraction of the total weight of all of $v$'s edges. If $v$ only knows a constant factor approximation to the total weight of its edges for each candidate color, taking the color with the best estimate still guarantees a relative defect of $\delta$ if we choose $s=c'/\delta$ for a sufficiently large constant $c'$. In the following, we will implicitly assume that all $s$-values are multiplied by a sufficiently large constant factor to make sure that a constant-factor estimate for the total edge weight per candidate color is sufficient. If the initial $N$-coloring is already a defective coloring with some relative defect $\delta_0$, then the argument holds for the bichromatic edges (w.r.t.\ to the initial $N$-coloring) and the new defective coloring becomes $(\delta_0+\delta)$-relative defective.
In \cite{Kuhn2009WeakColoring,KawarabayashiS18}, it is shown that when starting with an $N$-coloring and running $t=O(\log^* n)$ steps $i=1,\dots,t$ of the above algorithm with $s_i=2^{t-i+1}/\delta$, the overall relative defect in the end is $1/s_1+\cdots+s_{t}<\delta$ and the final number of colors is $O(s_{t-1}^2)=O(1/\delta^2)$. In the very first step, when starting with the initial $O(\log n)$-bit IDs, we have $\tau=\Theta(\log n)$ and in subsequent steps, $\tau$ will be exponentially smaller. As $\tau$ goes linearly into the number of candidate colors and thus the number of bits that have to be transmitted, we would like $s$ to be as small as possible in the very first step. We therefore add a step $0$, where we set $s_0=4/\delta$ and we multiply all the $s_i$ for $i>0$ by $4$. Let $\tau_i$ be the value of $\tau$ in step $i$ and let $p_i$ be the maximum number of candidate colors per node in step $i$. The overall relative defect is then less than $\delta/2$. In step $0$, the number of candidate colors per node is then at most $p_0=s_0\cdot\tau_0=O\big(\frac{\log n}{\delta}\big)$. In subsequent steps $i=1,\dots,t$, the number of candidate colors is at most $p_i \leq s_i\cdot \tau_i \leq s_1\cdot \tau_1$. After the step $0$, the number of colors is $O\big(\frac{\log n}{\delta^2}\big)$ and we therefore have $\tau_1 = O(\log\log n + \log(1/\delta))$. We further have $s_1=2^{O(\log^* n)}/\delta$. The total number of $\mathsf{CONGEST}$\xspace rounds on $G$ to run all $O(\log^* n)$ steps is therefore
\[
O\left(\log^* n + \frac{(s_0\tau_0 + s_1\tau_1\cdot\log^* n)\cdot \log\log q}{\log n}\right)
= O\left(\log^* n + \frac{\log\log q}{\delta}\right).
\]
In the above equation, we use that $s_1\tau_1\log^* n = O(1/\delta)\cdot 2^{O(\log^* n)}\cdot O(\log^* n)\cdot O(\log\log n + \log(1/\delta)) = O(\log(n)/\delta)$. Note that we have $\delta \geq 2^{-O(\sqrt{\log n})}$ and thus $\log(1/\delta)=O(\sqrt{\log n})$.
We now have a weighted $\delta/2$-relative defective coloring with $O(1/\delta^2)$ colors and we still need to turn this into a $\delta$-relative average defective coloring with $O(1/\delta)$ colors. For this, we adapt a distributed coloring algorithm that was presented in \cite{BEG18}. At the core of the algorithm of \cite{BEG18} is the following idea. For any prime $p$, there are $p^2$ orderings of the numbers $1,\dots,p$ such that any two orderings coincide in at most $1$ place. We now choose $p\geq 8/\delta$ and such that $p^2$ is larger than the initial number of $O(1/\delta^2)$ colors. In this way, nodes of a different initial color pick different orderings of the numbers $1,\dots,p$. The algorithm now consists of $p$ steps. For a node $v$, let $z_{v,1},\dots,z_{v,p}$ be the ordering of colors $1,\dots,p$ of node $v$. In step $i$, $v$ tries to take color $z_{v,i}$ and it takes the color if less than a $\delta/4$-fraction of the weight of its edges go to nodes that have already committed to taking color $z_{v,i}$ in previous steps or that are currently also trying to take color $z_{v,i}$. Note that if every node takes a color, the weighted average relative defect is at most $\delta/2$. This can be seen by orienting every edge towards the node that first commits to taking a color and orienting arbitrarily in case both nodes take the color in the same step. Because every node $v$ can conflict with each neighbor at most twice (once when trying the same color and once when $u$ has already committed to a color and $v$ tries this color), the total weight of the monochromatic outgoing edges of a node $v$ is at most at $\delta/4$-factor of the total total weight of all edges of $v$. The total weight of all monochromatic edges is therefore a $\delta/4$-factor of $2$ times the total weight of all edges. For further details, we refer to \cite{BEG18}.
It remains to show that each step of this algorithm can be implemented efficiently on the d2-multigraph $H$ in the $\mathsf{CONGEST}$\xspace model on $G$. To implement a step, a node $v$ needs to know the total weight of all its edges to neighbors that have already committed to the color that $v$ currently tries or that are currently trying the same color. Again, the algorithm can easily be adapted such that it suffices for each node $v$ to learn a constant approximation of this total edge weight. This now is the same communication problem that we had in the first part of the algorithm for a single candidate color. We therefore need to communicate $O(\log\log q)$ bits over each edge of $G$. This can clearly be done in $O\big(1+\frac{\log\log q}{\log n}\big)=O(\log\log q)$ rounds in the $\mathsf{CONGEST}$\xspace model. The total time for the second step is therefore $O(p\log\log q)=O\big(\frac{\log\log q}{\delta}\big)$ rounds. This concludes the proof.
\end{proof}
\paragraph{Implementing the Rounding Algorithm.} We next show how efficiently \Cref{alg:basicroundingstep,alg:basicrounding} can be implemented on a d2-multigraph of an underlying communication graph, when using $\mathsf{CONGEST}$\xspace algorithms on the communication graph. For the following lemma, recall that a rounding instance is called polynomially bounded in some value $q$ if all utility and cost values are either $0$ or bounded between $1/\operatorname{\text{{\rm poly}}}(q)$ and $\operatorname{\text{{\rm poly}}}(q)$ and if all fractional node values are either $0$ or lower bounded by $1/\operatorname{\text{{\rm poly}}}(q)$.
\begin{lemma}\label{lemma:d2rounding}
Let $H=(V_H,E_H)$ be a d2-multigraph of an underlying communication graph $G=(V,E)$ of maximum degree $\Delta$. Assume that $H$ is equipped with utility and cost functions $\mathbf{u}(\cdot)$ and $\mathbf{c}(\cdot)$ (with label set $\Sigma$) and with a fractional label assignment $\lambda$. Further assume that the given rounding instance is polynomially bounded in a parameter $q \leq n$. Then for every constant $c>0$ and every $\varepsilon,\mu>\max\set{1/q^c, 2^{-c\sqrt{\log n}}}$, if $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)>\mu\mathbf{u}(\lambda)$, there is a deterministic $\mathsf{CONGEST}$\xspace algorithm on $G$ to compute an integral label assignment $\ell$ for which $\mathbf{u}(\ell)-\mathbf{c}(\ell)\geq (1-\varepsilon)\cdot\big(\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\big)$ and such that the round complexity of the algorithm is
\[
O\left(\frac{\log^2 q}{\varepsilon\cdot\mu}\cdot\left(\frac{|\Sigma| \log(q\Delta)}{\log n} + \log\log q \right)+\log q\cdot\log^* n \right).
\]
\end{lemma}
\begin{proof}
First note that by applying \Cref{lemma:fractionalrounding}, we can turn the initial fractional label assignment into a $2^{-k}$-integral fractional label assignment with $k = O\big(\frac{\log q}{\varepsilon\cdot\mu}\big)$ (recall that the initial fractional assignment is polynomially bounded in $q$). Since we assume that $\varepsilon,\mu\geq 1/\operatorname{\text{{\rm poly}}}(q)$, we have $O\big(\frac{\log q}{\varepsilon\cdot\mu}\big)=O(\log q)$.
We assume that throughout the algorithm, each node of the communication graph $G$ keeps track of the fractional assignment of all neighbors in $G$. To achieve this, initially, each node of $G$ needs to learn $O(\log q)$ bits per label. This requires $O\big(1+ \frac{|\Sigma|\log q}{\log n}\big)$ rounds in the $\mathsf{CONGEST}$\xspace model on $G$. To maintain the fractional values of all $G$-neighbors, each node has to learn $O(1)$ bits per label from each neighbor after each change to the fractional assignment (i.e., after each iteration of the outer for loop in \Cref{alg:basicroundingstep}). We will account for the time for doing this, when analyzing the cost of such a step in the following.
\Cref{alg:basicrounding} consists of $O\big(\frac{\log q}{\varepsilon\cdot\mu}\big)=O(\log q)$ runs of \Cref{alg:basicroundingstep}. We therefore need to understand the time required to run a single instance of \Cref{alg:basicroundingstep}. At the beginning of \Cref{alg:basicroundingstep}, we compute a weighted $\delta$-relative defective coloring. Note that since all utilities, costs, and fractional values are polynomially bounded in $q$, also the edge weights of this defective coloring instance is polynomially bounded in $q$. Further, because nodes in $G$ know the fractional assignments of their $G$-neighbors, each node knows the weights of all edges in $H$ for which it is responsible (i.e., nodes know the weights of their physical edges and of the virtual edges for which they are the responsible middle node). By \Cref{lemma:d2defective}, we can therefore compute such a weighted $\delta$-relative defective coloring in $O\big(\frac{\log\log q}{\delta} + \log^* n\big)$ rounds and with $O(1/\delta)$ colors. By the description of \Cref{alg:basicrounding}, we have $\delta=\Theta\big(\frac{\varepsilon \mu}{\log q}\big)$ and thus, the time for computing the defective coloring is $O\big(\frac{\log q \cdot \log\log q}{\varepsilon\cdot \mu} + \log^* n\big)$.
Now, the algorithm iterates over the $O(1/\delta)$ colors. In each of these iterations, each node $v$ for each label $\alpha\in \Sigma$ needs to learn an estimate of $\phi_{v,\alpha}$ as defined in lines 9--11 of \Cref{alg:basicroundingstep}. By knowing the fractional assignment of all $G$-neighbors, each node $v$ exactly know the contribution of its physical edges to $\phi_{v,\alpha}$. However, this information is not sufficient to know the contribution of the virtual edges to $\phi_{v,\alpha}$. Node $v$ has to learn this information from its neigbors that are responsible for those edges. Consider some neighbor $w$ of $v$ and let $E_{w,v}$ be the set of virtual edges of $v$ for which $w$ is the responsible middle node. Further, let $E'_{w,v}$ be the subset of those edge that are bichromatic w.r.t.\ to the current defective coloring. Node $w$ knows the fractional assignment of both nodes for all edges in $E'_{w,v}$ and it can therefore compute the exact contribution of the edges in $E'_{w,v}$ to $\phi_{v,\alpha}$. Note that the number of edges in $E'_{w,v}$ is at most $O(\Delta)$. The contribution of edges in $E'_{w,v}$ to $\phi_{v,\alpha}$ is therefore upper bounded by $O(\Delta)\cdot\operatorname{\text{{\rm poly}}}(q)=\operatorname{\text{{\rm poly}}}(q\Delta)$. Node $v$ however only needs to learn an estimate of $\phi_{v,\alpha}$ that is accurate up $\Theta(\delta\cdot\theta_{v,\alpha})$. The value of $\theta_{v,\alpha}$ is either $0$ (in this case $\phi_{v,\alpha}$ and also the contribution of $E_{w,v}'$ to $\phi_{v,\alpha}$ are also $0$) or it is lower bounded by $1/\operatorname{\text{{\rm poly}}}(q)$. If the contribution of $E_{w,v}'$ to $\phi_{v,\alpha}$ is non-zero, $v$ therefore has to learn an estimate of it that is accurate up to an additive term that is lower bounded by $1/\operatorname{\text{{\rm poly}}}(q)$. We can therefore discretize so that the approximate contribution of edges in $E_{w,v}'$ to $\phi_{v,\alpha}$ can only take $\operatorname{\text{{\rm poly}}}(q\Delta)$ different values. Hence, $w$ only has to send $O(\log (q\Delta))$ bits for each possible label $\alpha\in \Sigma$ to $v$. To accomplish line 11 of \Cref{alg:basicroundingstep}, we therefore have to send at most $O(|\Sigma|\cdot\log (q\Delta))$ bits over each edge of $G$. Updating the knowledge of the fractional assignment of $G$-neighbors can then be done by exchanging another $O(|\Sigma|)$ bits over each edge of $G$. The total time for running one loop (iterating over one color of the defective coloring) in \Cref{alg:basicroundingstep} is therefore $O\big(1 + \frac{|\Sigma|\cdot\log (q\Delta)}{\log n}\big)$ in the $\mathsf{CONGEST}$\xspace model on $G$. Combining everything, we therefore obtain a time complexity of
\[
O\left(\frac{\log q\cdot\log\log q}{\varepsilon\cdot \mu} + \log^* n + \frac{\log q}{\varepsilon\cdot \mu}\cdot\left(1+\frac{|\Sigma|\cdot\log (q\Delta)}{\log n}\right)\right)
\]
for each instance \Cref{alg:basicroundingstep}. By multiplying with $O(\log q)$, i.e., with the number of times we run \Cref{alg:basicroundingstep}, we obtain the bound that is claimed by the lemma.
\end{proof}
\paragraph{Remark on a slight generalization} For some applications, we use a slight generalization of the d2-multigraph definition, provided above in \Cref{def:d2multigraph}, where we allow the managers of virtual edges to be some node further away, so long as we can provide some communication primitives, as we describe next.
\begin{definition}[long-range d2-Multigraph]\label{def:long-range-d2multigraph}
A long-range \emph{d2-multigraph} is a multigraph $H=(V_H,E_H)$ that is simulated on top of an underlying communication graph $G=(V,E)$ by a distributed message-passing algorithm on $G$. The nodes of $H$ are a subset of the nodes of $G$, i.e., $V_H\subseteq V$. The edge set $E_H$ consists of two kinds of edges, \emph{physical edges} and \emph{virtual edges}. Physical edges in $E_H$ are edges between direct neighbors in $G$. For each physical edge in $e\in E_H$ with $V(e)=\set{u,v}$, both nodes $u$ and $v$ know about $e$. Virtual edges in $E_H$ are edges between two nodes $u,v\in V_H$, and for each such virtual edge, there is a manager node $w$ which knows about this edge. We define a function $\xi:E_H\to V$ to refer to the node managing the virtual edge.
We next describe the assumed communication primitives. Let $M(v)$ be the set of nodes $w$ who manage virtual edges that include $v$. We assume $T$-round primitives that provide the following: (1) each node $v$ can send one $O(\log n)$-bit message that is delivered to all nodes in $M(v)$ in $T$ rounds; (2) given $O(\log n)$-bit messages prepared at nodes $M(v)$ specific for node $v$, node $v$ can receive an aggregation of these messages, e.g., the summation of the values, in $T$ rounds.
\end{definition}
We get an analogue of \Cref{lemma:d2rounding} for rounding in long-range d2-multigraphs:
\begin{lemma}\label{lemma:long-range-d2rounding}
Let $H=(V_H,E_H)$ be a long-range-d2-multigraph of an underlying communication graph $G=(V,E)$ of maximum degree $\Delta$, where the communication primitives have round complexity $T$. Assume that $H$ is equipped with utility and cost functions $\mathbf{u}(\cdot)$ and $\mathbf{c}(\cdot)$ (with label set $\Sigma$) and with a fractional label assignment $\lambda$. Further assume that the given rounding instance is polynomially bounded in a parameter $q \leq n$. Then for every constant $c>0$ and every $\varepsilon,\mu>\max\set{1/q^c, 2^{-c\sqrt{\log n}}}$, if $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)>\mu\mathbf{u}(\lambda)$, there is a deterministic $\mathsf{CONGEST}$\xspace algorithm on $G$ to compute an integral label assignment $\ell$ for which $\mathbf{u}(\ell)-\mathbf{c}(\ell)\geq (1-\varepsilon)\cdot\big(\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\big)$ and such that the round complexity of the algorithm is
\[
T \cdot O\left(\frac{\log^2 q}{\varepsilon\cdot\mu}\cdot\left(\frac{|\Sigma| \log(q\Delta)}{\log n} + \log\log q \right)+\log q\cdot\log^* n \right).
\]
\end{lemma}
\begin{proof}[Proof Sketch]
One change is that we need a version of \Cref{lemma:d2defective} for long-range d2-multigraph, but that follows by the same proof, with only a $T$ factor slow down, because each virtual edge manager needs to learn the current color of its endpoints, and then informs each endpoint about the total weight of edges to endpoints that have chosen each particular color (in the collection this endpoint is interested in). The former clearly fits the first communication primitive of long-range d2 multi-graphs, sending from $v$ to $M(v)$, and the latter fits the second primitive, $v$ receiving an aggregation of messages prepared by $M(v)$ for $v$.
Then, the rest of the proof is identical to that of \Cref{lemma:d2rounding}, again with only a $T$ factor slow-down, where $T$ denotes the round complexity of the corresponding two communication primitives. Namely, in our rounding, each virtual edge manager needs to receive the current fractional label of each of the endpoints. Then it prepares a message for each endpoint, and each endpoint needs to receive the summation of these messages. Again, the former fits the first communication primitive of long-range d2 multi-graphs, sending from $v$ to $M(v)$, and the latter fits the second primitive, $v$ receiving an aggregation of messages prepared by $M(v)$ for $v$.
\end{proof}
\section{Introduction \& Related Work}
In this paper, we present a \emph{local rounding method}
that can be used to transform probabilistic/fractional assignments in a certain class of randomized distributed graph algorithms into deterministic/integral assignments. Roughly speaking, the method applies to randomized distributed algorithms that can be analyzed assuming pairwise independence. This allows us to transform such randomized algorithms into deterministic distributed algorithms. This unified approach leads to novel and/or improved algorithms for a number of the central problems of the area---e.g., maximal independent set, $(\Delta+1)$-coloring, maximal (hypergraph) matching, and set cover. Several of these problems had remained open for over three decades until a recent breakthrough on network decomposition~\cite{rozhonghaffari20}. Our approach is completely independent and yields faster algorithms (besides providing a new and more systematic construction for the network decomposition problem itself).
\subsection{Background and State of the Art}
\noindent\textbf{Model.} We work with the standard synchronous message-passing model of distributed computing~\cite{peleg00}. The network is abstracted as an undirected graph $G=(V, E)$, with $n:=|V|$. Initially, each node knows only its own unique $O(\log n)$-bit identifier, and nothing else about the graph $G$, except for potentially some upper bounds on global parameters such as $n$, the number of nodes, and $\Delta$, the maximum degree. Per round, each node can send one $B$-bit message to each of its neighbors. We usually assume $B=O(\log n)$ and refer to this model variant as $\mathsf{CONGEST}$\xspace. The variant that allows unbounded message sizes is referred to as $\mathsf{LOCAL}$\xspace. At the end of the computation, each node should know its own part of the output, e.g., whether it is in the computed independent set.
\paragraph{Linial's MIS question.}
Four of the best studied problems in distributed graph algorithms are maximal independent set (MIS), maximal matching, $(\Delta+1)$-vertex coloring, and $(2\Delta-1)$-edge coloring. MIS is the hardest of these four, as all others can be reduced to it~\cite{luby86,linial1987LOCAL}. Luby's classic algorithm from 1986 provides an $O(\log n)$-round randomized algorithm for MIS, and thus for all the others~\cite{luby86} (see also \cite{alon86}). It however remained open for over three decades whether there also is a deterministic $\operatorname{\text{{\rm poly}}}\log n$-time algorithm for MIS, and this came to be known as \emph{Linial's open question}, first raised in~\cite{linial1987LOCAL, linial92}. See also the book of Barenboim and Elkin~\cite{barenboimelkin_book} for further discussions on the significance of this and other related problems.
\paragraph{The first solution, via network decomposition. } The first solution to the MIS question was provided recently by Rozho\v{n} and Ghaffari~\cite{rozhonghaffari20}. Their algorithm actually solves the network decomposition problem in $\operatorname{\text{{\rm poly}}}\log n$ time, which roughly speaking partitions the vertices into $\operatorname{\text{{\rm poly}}}\log n$ colors, such that the components in the subgraph induced by each color have $\operatorname{\text{{\rm poly}}}\log n$ diameter. This directly leads to a $\operatorname{\text{{\rm poly}}}\log n$-time deterministic algorithm for MIS in the $\mathsf{LOCAL}$\xspace model. In the $\mathsf{CONGEST}$\xspace model, a more complex $\operatorname{\text{{\rm poly}}}\log n$-time MIS algorithm follows from combining network decomposition with a deterministic $\mathsf{CONGEST}$\xspace algorithm of \cite{censor2017derandomizing} for low-diameter networks. The combination works by leveraging the pairwise analysis of the randomized MIS algorithms (Luby~\cite{luby1993removing} or Ghaffari~\cite{gmis}) and using the low-diameter components of the decomposition to fix the bits of the randomized algorithm one by one. Currently, the fastest known deterministic MIS algorithm both in $\mathsf{LOCAL}$\xspace and $\mathsf{CONGEST}$\xspace runs in $O(\log^5 n)$ rounds, and it is based on an improved variant of the Rozho\v{n}-Ghaffari decomposition presented in \cite{GGR20}.
As a side note, we remark that the decomposition result, in combination with \cite{ghaffari2017complexity, ghaffari2018derandomizing}, yields a $\mathsf{LOCAL}$\xspace-model derandomization with $\operatorname{\text{{\rm poly}}}\log n$ round slow down, for $\operatorname{\text{{\rm poly}}}\log n$-time checkable problems. However, there is no such general derandomization for the $\mathsf{CONGEST}$\xspace model.
\paragraph{A different line of attack, via rounding.}
A more direct line of attack toward developing $\operatorname{\text{{\rm poly}}}\log n$-time deterministic algorithms for the four problems mentioned above is via \emph{local rounding}. Roughly speaking, this approach starts with a suitable fractional solution to a certain problem and gradually and deterministically rounds this fractional solution to obtain an integral solution with similar quality. One can trace this local rounding approach back to the work of Hanckowiak, Karonski, and Panconesi~\cite{hanckowiak01} who gave the first $\operatorname{\text{{\rm poly}}}\log n$-time deterministic distributed algorithm for the maximal matching problem; though they never talked about rounding. The \emph{local rounding} nature of the approach was made explicit and used to improve the complexity for maximal matching in the work of Fischer\cite{fischer2020improved}. Fischer observed that for the matching problem, it is easy to obtain an $O(1)$-approximation of maximum fractional matching in a deterministic way, and that the real challenge is in rounding this fractional matching into an integral one without much loss in the size. Fischer developed a specific $O(\log^2 \Delta)$-time rounding procedure for the matching problem (in bipartite graphs). Repeating this $O(\log n)$ times (in a suitable bipartite version of the graph) adds in each repetition a matching within a constant factor of the remaining maximum, hence reducing the maximum remaining matching by a constant factor, and therefore yields a maximal matching in $O(\log^2 \Delta \cdot \log n)$ time. That remains the state of the art for the maximal matching problem. It however should be noted that Fischer's rounding was very specific to matching in graphs, and it did not generalize to the other three problems. In fact, the method did not extend even to matching in hypergraphs of rank three, i.e., where each edge has three endpoints.
A follow-up work by Fischer, Ghaffari, and Kuhn\cite{FischerGK17} developed a different rounding method for matching that extended to low-rank hypergraphs. Along with a reduction that they gave from $(2\Delta-1)$-edge coloring in graphs to maximal matching in hypergraphs of rank $3$, this led to a $\operatorname{\text{{\rm poly}}}\log n$-time deterministic algorithm for $(2\Delta-1)$-edge coloring. This result put the second of the four problems in the $\operatorname{\text{{\rm poly}}}\log n$-time deterministic regime. The exact complexity was notably higher. Harris\cite{harris2019distributed} improved the complexity of this hypergraph matching rounding. Among other speed-ups, that led the complexity of deterministic $(2\Delta-1)$-edge coloring to reach $\tilde{O}(\log^2 \Delta \cdot \log n)$, which nearly matches that of maximal matching.
The above lines of work on local rounding appeared limited to rounding for matching in graphs or hypergraphs. Then (and shortly after the network decomposition result~\cite{rozhonghaffari20}), Ghaffari and Kuhn~\cite{GhaffariK21} developed an efficient rounding method for $\Delta+1$ coloring. Roughly speaking, their approach starts with a simple fractional assignment of the colors (where each node takes an equal portion of each of the colors in its palette) and gradually rounds this assignment until reaching (a potentially improper) integral color assignments. This is done in a manner that, the rounding approximately maintains a certain potential function, such that in the end, once given the final integral color assignments, we can efficiently turn it into a proper coloring of a constant fraction of the remaining node. Interestingly enough, this rounding also led to the $O(\log^2 \Delta \cdot \log n)$ time complexity for vertex coloring. This put the third of the four problems in the $\operatorname{\text{{\rm poly}}}\log\Delta\cdot \log n$-time, and concretely $O(\log^2 \Delta \cdot \log n)$-time, deterministic regime. We note that, at that time, $\operatorname{\text{{\rm poly}}}\log n$-round complexity coloring was already known from the decomposition result~\cite{rozhonghaffari20}, but the rounding-based method~\cite{GhaffariK21} gave a more direct and faster coloring algorithm.
Despite this exciting progress on local rounding, the known methods appeared ad hoc, specifically tailored to matching or coloring problems. For instance, they did not extend to the hardest of the four classic problems, the maximal independent set problem.
\subsection{Our Contributions}
\label{sec:contributions}
In this paper, we vastly generalize the local rounding method, in such a way that, roughly speaking, we can now derandomize randomized local algorithms that can be analyzed with pairwise independence. This rounding works in a local and efficient manner, and without relying on network decompositions. We show that this generalized local rounding method yields new and/or improved deterministic distributed algorithms for a range of graph problems of interest. As a notable example, we obtain an $O(\log^2 \Delta \cdot \log n)$ round $\mathsf{LOCAL}$\xspace model algorithm for MIS, and a $O(\log^2 \Delta \cdot \log\log \Delta \cdot \log n)$-round $\mathsf{CONGEST}$\xspace model variant.
\begin{theorem} There is a deterministic distributed algorithm that computes an MIS in time $O(\log^2 \Delta \cdot \log n)$ in the $\mathsf{LOCAL}$\xspace model. A variant of this algorithm computes an MIS deterministically in $O(\log^2 \Delta\cdot \log\log \Delta \cdot \log n)$ rounds of the $\mathsf{CONGEST}$\xspace model.
\end{theorem}
Hence, now all the four classic problems are in the $O(\log^2 \Delta \cdot \log n)$ deterministic round complexity regime, and in a unified way. This result provides a second solution to Linial's famous open question\cite{linial1987LOCAL}, which had remained open for over three decades. The new solution is completely independent of the first decomposition-based solution and is also more efficient. In contrast, the fastest previously known MIS algorithm, in either $\mathsf{LOCAL}$\xspace or $\mathsf{CONGEST}$\xspace models, required $O(\log^5 n)$ rounds\cite{GGR20} and was based on network decomposition.
We also note for maximal independent set and maximal matching, a celebrated recent result of Balliu et al.~\cite{balliu2019LB} gives a lower bound of $\Omega(\min\{\Delta, \log n/\log\log n\})$ for deterministic algorithms in the $\mathsf{LOCAL}$\xspace model. This almost justifies the need for one $\log n$ factor in the upper bound. In particular, in the regime where $\Delta=\operatorname{\text{{\rm poly}}}(\log n)$, the lower bound becomes $\Omega(\log n/\log\log n)$ and our $\mathsf{LOCAL}$\xspace-model upper bound is $O(\log n\cdot \log^2\log n)$, which means that the upper and lower bounds are matching up to exponentially lower order factors.
\paragraph{Other applications, set cover.} As another prominent application, we obtain an improved approximation algorithm for the minimum set cover problem, with an approximation factor roughly matching that of the best-known centralized algorithm (below which the approximation problem becomes $\mathsf{NP}$-hard):
\begin{theorem} There is a deterministic distributed algorithm that computes an $O(\log s)$ approximation of the minimum set cover problem in $\tilde{O}(\log s \log^2 t+ \log^* n)$ rounds of the $\mathsf{CONGEST}$\xspace model, where $s$ denotes the maximum set cardinality, and $t$ denotes the maximum number of sets that contain a given element.
\end{theorem}
Notably, the complexity of this algorithm depends logarithmically only on the local parameters of the problem (degrees in a bipartite formulation) and has only an additive $O(\log^* n)$ dependency on the network size, which is known to be necessary~\cite{lenzen08}. In contrast, the best previously known algorithm has a polylogarithmic dependency on the global network size $n$~\cite{Deurer2019, GGR20}. This is an important qualitative difference and has implications for other computational settings\footnote{In particular, in the \emph{local computations algorithms} (LCA) model of sublinear-time centralized computation \cite{rubinfeld2011LCA, alon2012LCA}, our result provides a deterministic set cover approximation with query complexity $d^{\tilde{O}(\log^3 d)} \operatorname{\text{{\rm poly}}}(\log n)$, where $d=st$. This follows from a direct stimulation via gathering the relevant local neighborhood\cite{parnas2007approximating} (with an appropriate prior coloring~\cite{chang2019complexity}). This LCA can answer whether each asked set is in the selected cover or not, using only $d^{\tilde{O}(\log^3 d)} \operatorname{\text{{\rm poly}}}(\log n)$ query accesses to the adjacency lists of the graph. In contrast, a simulation of the decomposition-based approach would imply query complexity $d^{\operatorname{\text{{\rm poly}}}(\log n)}$, which is even super-polynomial in $n$ and thus completely useless/uninteresting. The best known LCA for set cover has query complexity $d^{O(\log d \log\log d)} \operatorname{\text{{\rm poly}}}(\log n)$~\cite{grunau2020improved}, but that heavily relies on randomness. No deterministic LCA with query complexity $\operatorname{\text{{\rm quasipoly}}}(d) \cdot \operatorname{\text{{\rm poly}}}(\log n)$ was known previously.}.
We also note that Deuer, Kuhn, and Maus~\cite{Deurer2019} previously used a deterministic rounding method to obtain an $O(\log \Delta)$ approximation for the minimum dominating set problem in $O(\Delta \operatorname{\text{{\rm poly}}}(\log \Delta) + \operatorname{\text{{\rm poly}}}(\log \Delta) \log^* n)$ rounds of the $\mathsf{CONGEST}$\xspace model, in any graph with maximum degree at most $\Delta$. The minimum dominating set problem is equivalent to the set cover problem in the setting where $s=t=\Delta$, and one can provide a simple approximation preserving reduction in both directions. Using the straightforward direction of this connection (reducing dominating set to set cover by viewing each node as a set that includes all of its neighbors) and invoking our minimum set cover approximation algorithm, we obtain an $O(\log \Delta)$ approximation for minimum dominating set in $\tilde{O}(\log^3 \Delta + \log^* n)$ rounds of the $\mathsf{CONGEST}$\xspace model. This improves on the $O(\Delta \operatorname{\text{{\rm poly}}}(\log \Delta) + \operatorname{\text{{\rm poly}}}(\log \Delta) \log^* n)$ complexity of ~\cite{Deurer2019} nearly exponentially.
\paragraph{Other applications, network decomposition.} In fact, since the network decomposition problem has been known to be reducible to the set cover problem\cite{ghaffari2018derandomizing}, our method yields a novel $\operatorname{\text{{\rm poly}}}(\log n)$-round algorithm for the network decomposition problem. One may view this as a more systematic solution for network decomposition, in contrast to the specific combinatorial approach of \cite{rozhonghaffari20}. This new construction reduces the decomposition problem to its pairwise analyzable core (captured in set cover) and then derandomizes the corresponding natural randomized algorithm via local rounding. This is more systematic in the sense that it is similar to other problems that are derandomized by going through pairwise analysis.
\paragraph{Other applications, maximum independent set and hypergraph matching.} As another applications, we get efficient algorithms for computing large cardinality or weight independent sets. In the following, the neighborhood independence $\beta$ of a graph $G=(V,E)$ is the size of the largest independent set of any subgraph $G[N(v)]$ induced by the set of neighbors $N(v)$ of some node $v\in V$. Further, for a node weight function $w:V\to\mathbb{N}$ and a node set $S\subseteq V$, $w(S)=\sum_{v\in S} w(v)$.
\begin{theorem}\label{thm:ISthm}
Let $G=(V,E)$ be a node-weighted $n$-graph of maximum degree $\Delta$, neighborhood independence $\beta$, node weights $w:V\to \mathbb{N}$, and maximum weight $W$. Further, let $\mathsf{OPT}$ be the weight of a maximum weight independent set and assume that $G$ is equipped with an proper $\xi$-coloring. Then, for every $\varepsilon>0$, there are deterministic $\mathsf{CONGEST}$\xspace algorithms to compute independent sets
\begin{align}
\text{of weight } & \frac{1-\varepsilon}{\beta}\cdot\mathsf{OPT} & \text{ in } & O\big(\log^2(\Delta W)\cdot\log(1/\varepsilon) + \log^* \xi\big)\text{ rounds,}\label{eq:ISthm1}\\
\text{of weight } & (1-\varepsilon)\cdot \frac{w(V)}{\Delta+1} & \text{ in } & O\big(\log^2\Delta\cdot\log(1/\varepsilon) + \log^* \xi\big)\text{ rounds, and}\label{eq:ISthm2}\\
\text{of weight } & \left(\frac{1}{2}-\varepsilon\right)\cdot \sum_{v\in V}\frac{(w(v))^2}{w(v) + w(N(v))} & \text{ in } & O\left(\frac{\log^2(\Delta/\varepsilon)}{\varepsilon} + \log^* \xi\right)\text{ rounds.}\label{eq:ISthm3}
\end{align}
\end{theorem}
{\it \noindent\Cref{thm:ISthm} has several implications:}
\begin{itemize}
\item Since $\beta\leq \Delta$, result \eqref{eq:ISthm1} implies a $(1-\varepsilon)/\Delta$-approximation in the same round complexity. For small $\Delta$ (and $W$), this is a significant improvement over an $O(T_{\mathrm{MIS}}/\varepsilon)$-round algorithm for this problem in \cite{KawarabayashiKS20} ($T_{\mathrm{MIS}}$ denotes the time for computing an MIS).
\item Line graphs of hypergraphs of rank $r$ have neighborhood independence $r$. Thus, result \eqref{eq:ISthm1} implies the same result for computing a $(1-\varepsilon)/r$-approximation for maximum weight matching in hypergraphs of rank $r$. For small $W$, this is an improvement over an algorithm in \cite{harris2019distributed}.
\item Result \eqref{eq:ISthm1} also implies an $O(\log^2\Delta + \log^*\xi)$-round algorithm for computing a constant maximum matching approximation in graphs. Repeating $O(\log n)$ times yields an $O(\log^2\Delta\cdot\log n)$-round maximal matching algorithm and in \Cref{sec:boundedindep} (\Cref{thm:maximalmatching}), we show that this also works in the $\mathsf{CONGEST}$\xspace model. This gives an alternative to the deterministic $O(\log^2\Delta\cdot \log n)$-round $\mathsf{CONGEST}$\xspace algorithm of Fischer~\cite{fischer2020improved} and it implies that our generic rounding framework implies the current best deterministic $\mathsf{LOCAL}$\xspace and $\mathsf{CONGEST}$\xspace algorithms for the four classic distributed symmetry breaking problems.
\item In \cite{KawarabayashiKS20}, Kawarabayashi, Khoury, Schild, and Schwartzman showed that with randomization, an independent set of size $\Omega(n/\Delta)$ can be computed exponentially faster than an MIS and they raised the question whether the same is also true for deterministic algorithms. Result \eqref{eq:ISthm2} answers this question in the affirmative, because for deterministically computing an MIS, there is an $\Omega\big(\min\set{\Delta,\frac{\log n}{\log\log n}}\big)$-round lower bound even in the $\mathsf{LOCAL}$\xspace model~\cite{balliu2019LB}.
\item By a simple application of the Cauchy-Schwarz inequality, the independent set weight in \eqref{eq:ISthm3} can be lower bounded by $\big(\frac{1}{2}-\varepsilon\big)\cdot\frac{(w(V))^2}{w(V) + \sum_{v\in V} w(N(v))}$ (for details, see \cite{kako2005approximation}). Without the factor $\frac{1}{2}-\varepsilon$, the two bounds are natural weighted generalizations of the well-known lower bound $\sum_{v\in V}\frac{1}{\deg(v)+1}\geq \frac{n}{\deg_{\mathrm{avg}}+1}$ on the size of a maximum cardinality independent set. The bound is sometimes also known as the Caro-Wei-Tur\'an bound~\cite{turan1941external,Wei1981generalizedturan,griggs1983lower}.
\end{itemize}
\subsection{Overview of Our Method}
In the following, we give an extended high-level description of our general rounding method and we also discuss the main novel ideas that are necessary to obtain the results given in \Cref{sec:contributions}.
\subsubsection[The Rounding of Ghaffari and Kuhn\cite{GhaffariK21} for $(\Delta+1)$-Coloring]{The Rounding of Ghaffari and Kuhn\cite{GhaffariK21} for \boldmath$(\Delta+1)$-Coloring}
Our generic rounding algorithm is a generalization of a recent deterministic, distributed algorithm by Ghaffari and Kuhn~\cite{GhaffariK21} for solving the $(\Delta+1)$-coloring problem. We therefore first review (a slightly adapted version of) the algorithm of \cite{GhaffariK21}. At the core of the algorithm of \cite{GhaffariK21} is a method to color a constant fraction of the nodes of the graph in time $O(\log^2\Delta)$. Repeating this $O(\log n)$ times then colors the whole graph in time $O(\log^2\Delta\cdot\log n)$.\footnote{Formally, this requires the method for coloring a constant fraction of the nodes to work for the more general $(\mathit{degree}+1)$-list coloring problem. We however ignore this in this high-level overview.} The algorithm for coloring a constant fraction of the nodes is based on derandomizing the following trivial randomized algorithm for the same problem: Every node $v\in V$ chooses one of the $\Delta+1$ colors uniformly at random and $v$ keeps the color if no higher-ID neighbor chooses the same color. When doing this, in expectation, a constant fraction of the nodes can keep their color. One way to see this is by the following analysis. For every edge $\set{u,v}$ of the graph, the probability that both nodes $u$ and $v$ choose the same color is $1/(\Delta+1)$. Hence, the total expected number of such monochromatic edges is at most $|E|/(\Delta+1)<n/2$. For each such edge, we uncolor the lower ID node. The expected number of nodes that do not keep their color is thus less than $n/2$.
\paragraph{Rounding a Fractional Solution.} In \cite{GhaffariK21}, the above step is derandomized by considering the assignment of a uniformly random color to each node as a fractional assignment of colors to nodes, and picking a single color per node is then considered as rounding this fractional color assignment to an integral color assignment. To do this, the algorithm \cite{GhaffariK21} assigns a cost to each edge, which is equal to the probability of both endpoints picking the same color when independently choosing the color at random from the current fractional color assignment. We can then define the total cost of a given fractional color assignment as the sum over the individual edge costs. Note that by the linearity of expectation, the total cost is equal to the expected number of monochromatic edges if each node picks its color randomly according to its current fractional distribution over colors. As discussed above for the initial fractional assignment (i.e., if each node has a uniform distribution over all $\Delta+1$ colors), the total cost is at most $n/2$. The goal now is to find an integral assignment of colors for which the total cost (i.e., the total number of monochromatic edges) is not much larger (say, at most $3n/4$) such that still, a constant fraction of the nodes can keep their colors. First note that if we sequentially iterate over the nodes, there is a simple way to achieve this. When rounding the fractional assignment of a node $v$, $v$ just picks the color that minimizes the sum of the costs of its edges (i.e., the color that minimizes the total number of monochromatic edges of $v$). In this way, we can obtain an integral color assignment for which the total cost is at most the total cost of the initial fractional color assignment and thus at most $n/2$. Because nodes that are non-adjacent do not share edges, they can do this rounding in parallel. If we are given a proper vertex coloring with $C$ colors, the rounding can therefore be done in $O(C)$ rounds in the distributed setting.
\paragraph{Fast Rounding of a Fractional Solution.}
While the described rounding algorithm can perfectly preserve the cost of a given fractional color assignment, having to iterate through the colors of a proper vertex coloring of the graph will be extremely slow. In this way, we can at best hope for an algorithm with a round complexity that is linear in $\Delta$, which is exponentially slower than what we aim to achieve. In order to obtain a faster rounding algorithm, \cite{GhaffariK21} applies two main ideas. First, instead of iterating over the colors of a proper vertex coloring, the algorithm first computes a defective $C$-coloring and then iterates over the $C$ colors of this coloring. In a defective coloring, each node is allowed to have some neighbors of the same color. By extending ideas of \cite{Kuhn2009WeakColoring,BEG18,KawarabayashiS18}, it is shown in \cite{GhaffariK21} that on a graph $G=(V,E)$ with edge weights $w(e)\geq0$, for a given parameter $\delta>0$, one can compute a vertex coloring with only $O(1/\delta)$ colors such that the total weight of the monochromatic edges is at most a $\delta$-fraction of the total weight of all edges. We call such a coloring a \emph{weighted $\delta$-relative average defective coloring}. If one starts with an initial proper $O(\Delta^2)$-coloring of $G$ (which can be computed in time $O(\log^* n)$ by using an algorithm of \cite{linial92}), a weighted $\delta$-relative average defective $O(1/\delta)$-coloring can be computed deterministically in time $O(1/\delta + \log^*\Delta)$. The idea now is to compute a weighted $\delta$-relative average defective $O(1/\delta)$-coloring where the edge weights are equal to the edge costs before rounding and to run the above simple distributed rounding algorithm on the graph induced by the bichromatic edges (w.r.t.\ the defective coloring). For this, we have to iterate over $O(1/\delta)$ colors, which even when choosing $\delta=1/\operatorname{\text{{\rm poly}}}\log \Delta$ is much faster than iterating over $\Omega(\Delta)$ colors. The problem with the approach is that the cost of monochromatic edges (w.r.t.\ the defective coloring) can now grow arbitrarily. Those edges are not considered during rounding and their cost can increase from initially $1/(\Delta+1)$ to $1$ if both nodes of an edge choose the same color. In order to avoid such drastic cost increases of the monochromatic edges, we have to do the rounding in several small steps. If in each iteration, each node at most doubles the fractional value for each color, the probability of an edge becoming monochromatic can increase by at most a factor $4$, even if the nodes are rounded in a worst-case way. As a consequence, in such a rounding step, the total cost over all edges can grow by at most a factor of $1+O(\delta)$. In the algorithm of \cite{GhaffariK21}, the step-wise rounding is done in such a way that with every iteration, the minimal non-zero fractional color value is at least doubled. One can then get from fractional values $1/O(\Delta)$ to integral values in $O(\log\Delta)$ steps. By choosing $\delta=c/\log\Delta$ for a sufficiently small constant $c>0$, the total cost can grow by at most a constant factor that can be made arbitrarily close to $1$. Since in each rounding step, we have to compute a weighted $\delta$-relative average defective coloring with $O(1/\delta)$ colors and iterate over $O(1/\delta)=O(\log\Delta)$ colors of this defective coloring, each rounding step requires $O(\log\Delta)$ rounds. We can therefore obtain a coloring with at most $3n/4$ monochromatic edges in $O(\log^2\Delta)$ rounds and we can thus properly color at least $n/4$ nodes in $O(\log^2\Delta)$ rounds.
\subsubsection{Extending the Algorithm to Obtain Large Independent Sets}
In the present paper, we generalize the rounding method of \cite{GhaffariK21} in a significant way to make it applicable to a much wider family of problems. Our simplest application of the more general rounding algorithm is a deterministic algorithm for obtaining large (or heavy in the case of node-weighted graphs) independent sets. As this algorithm already allows to highlight some of the key challenges and ideas, we discuss it here first.
For this high-level discussion, assume that we want to compute an independent set of size $\Omega(n/\Delta)$. There is a simple randomized algorithm to achieve this (in expectation). First, each node marks independently itself with probability $1/\Delta$ and afterward, for every edge $\set{u,v}$, if both $u$ and $v$ are marked, then we unmark the lower-ID node. The set of nodes that are still marked then clearly form an independent set of $G$. The expected number of marked nodes is $n/\Delta$ and for every edge, both nodes are marked with probability $1/\Delta^2$. The expected number of edges for which both nodes are marked is therefore at most $|E|/\Delta^2 \leq n/(2\Delta)$. Therefore, in expectation, we obtain an independent set of size at least $n/(2\Delta)$. In order to obtain a deterministic variant of this algorithm, we can try to adapt the idea of \cite{GhaffariK21}. A fractional independent set is an assignment of fractional values $x_v\in [0,1]$ to each node and we can define a potential function $\Phi(\vec{x})$ that measures the expected size of the resulting independent set when running the randomized algorithm described above, i.e., if each node $v$ gets marked independently with probability $x_v$ and if for every edge $\set{u,v}$, if both $u$ and $v$ are marked, the lower-ID node unmarks itself. When doing this, we obtain
\[
\Phi(\vec{x}) = \sum_{v\in V} x_v - \sum_{\set{u,v}\in E} \!\!x_u\cdot x_v
= \sum_{\set{u,v}\in E} \!\!\bigg(\underbrace{\frac{x_u}{\deg(u)}+\frac{x_v}{\deg(v)} - x_u\cdot x_v}_{=: \phi_{\set{u,v}}(\vec{x})}\bigg).
\]
That is, the expected size of the resulting independent set can be expressed by a potential function $\Phi(\vec{x})$ that can be written as a sum over individual edge potentials $\phi_e(\vec{x})$. By iterating over the colors of a proper coloring of the graph $G$, we can therefore again round the fractional value of each node such that the independent set in the resulting rounded solution is at least as large as the expected independent set size initially. However, unlike in the case of the coloring algorithm of \cite{GhaffariK21}, when speeding up defective coloring and rounding gradually, a direct generalization of the algorithm of \cite{GhaffariK21} does not work. We next discuss the reason and explain what we do instead.
\paragraph{Fast Rounding of Large Fractional Independent Sets.} Note that the potential $\Phi(\vec{x})$ can be split into a positive \emph{utility} $\mathbf{u}(\vec{x})$ and a negative \emph{cost} $\mathbf{c}(\vec{x})$. We have $\Phi(\vec{x})=\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}) =\sum_{e\in E} \phi_e(\vec{x}) = \sum_{e\in E} \big(\mathbf{u}_e(\vec{x}) - \mathbf{c}_e(\vec{x})\big)$, where
\[
\mathbf{u}(\vec{x}) = \sum_{v\in V} x_v,\
\mathbf{c}(\vec{x}) = \sum_{e\in E} x_u x_v,\text{ and for }
e=\set{u,v}\in E:
\mathbf{u}_{e}(\vec{x}) = \frac{x_u}{\deg(u)}+\frac{x_v}{\deg(v)},\
\mathbf{c}_{e}(\vec{x}) = x_u x_v.
\]
Assume again that we round gradually such that in a single rounding step each fractional $x_v$ value is at most doubled and that we use a defective coloring to obtain a fast implementation of such a rounding step. We then again have to assume that the rounding on monochromatic edges happens in a worst-case way. However, while gradual rounding guarantees that $\mathbf{c}_{e}(\vec{x})$ can increase by at most a factor $4$, we cannot lower bound $\mathbf{u}_{e}(\vec{x})$ and even if we could, the difference $\phi_e(\vec{x})=\mathbf{u}_e(\vec{x})-\mathbf{c}_e(\vec{x})$ can change by an arbitrary factor. Note that before the rounding step we could even have $\phi_e(\vec{x})=0$ so that the multiplicative change in the edge potential even becomes unbounded.
It is however possible to efficiently implement a gradual rounding step if for the fractional assignment $\vec{x}$ before the rounding, $\mathbf{u}(\vec{x})$ is by a constant factor larger than $\mathbf{c}(\vec{x})$, i.e., if for example $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$. In this case, one can compute a weighted $\delta$-relative defective coloring w.r.t.\ edge weights $\mathbf{u}_e(\vec{x})+\mathbf{c}_e(\vec{x})$. One can then show that even doing a worst-case gradual rounding step for all the monochromatic edges (w.r.t.\ the defective coloring), the potential decreases by at most $O\big(\delta\cdot(\mathbf{u}(\vec{x}) +\mathbf{c}(\vec{x}))\big)=O\big(\delta\cdot\Phi(\vec{x})\big)$. For our initial fractional independent set, the condition $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$ is satisfied and we can therefore efficiently carry out the first rounding step. However, in a single rounding step, we can only guarantee that the difference $\mathbf{u}(\vec{x})-\mathbf{c}(\vec{x})$ is approximately preserved and we cannot guarantee that the terms $\mathbf{u}(\vec{x})$ and $\mathbf{c}(\vec{x})$ are approximately preserved individually. We therefore cannot guarantee that the condition $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$ is also approximately preserved. To cope with this challenge, we introduce one additional idea, which we describe next.
\paragraph{Dynamically Adapting the Potential Function.}
Instead of using a single potential function $\Phi(\vec{x})$ for the whole rounding process, we use a sequence $\Phi_0(\vec{x}), \Phi_1(\vec{x}), \dots, \Phi_{T}(\vec{x})$ of potential functions for the $T=O(\log\Delta)$ gradual rounding steps. Assume that initially, we have $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$ and that we are satisfied if we can maintain the value of $\mathbf{u}(\vec{x}) -\mathbf{c}(\vec{x})$ within some constant factor. We can then define
\begin{equation}\label{eq:potentialoverview}
\Phi_i(\vec{x}) := \mathbf{u}(\vec{x}) - \eta_i\cdot \mathbf{c}(\vec{x})\text{, where }\eta_i = \frac{1}{2}\left(3 - \frac{i}{T}\right).
\end{equation}
That is, we have $\Phi_0(\vec{x}) = \mathbf{u}(\vec{x}) - \frac{3}{2}\cdot \mathbf{c}(\vec{x})$, $\Phi_T(\vec{x}) = \mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x})$, and $\Phi_{i+1}(\vec{x}) - \Phi_i(\vec{x}) = \frac{3}{2T}\cdot\mathbf{c}(\vec{x})$. In rounding step $i$, we use the potential function $\Phi_i$ for the rounding. By slightly increasing the gap between the positive and the negative term in the potential function for each rounding step, we can make sure that $\mathbf{u}(\vec{x})$ is always sufficiently larger than $\mathbf{c}(\vec{x})$ in order to efficiently perform the rounding step and to make sure that after the rounding step $i$, $\Phi_i$ is within a $(1-O(\delta))$-factor of $\Phi_{i-1}$. That is, in the end, we obtain an integral assignment $\vec{x}'$ for which $\mathbf{u}(\vec{x}')- \mathbf{c}(\vec{x}')$ is within a constant factor of $\mathbf{u}(\vec{x}) - \frac{3}{2}\cdot\mathbf{c}(\vec{x})$ for the original fractional assignment $\vec{x}$. Overall, we obtain an algorithm to obtain an independent set of size $\Omega(n/\Delta)$ in $O(\log^2\Delta + \log^* n)$ rounds, where the $O(\log^* n)$-term comes from computing a proper $O(\Delta^2)$-coloring of $G$ at the very beginning of the algorithm. The technical details of the rounding procedure appear in \Cref{sec:basicrounding}.
The described rounding method for independent sets works for all fractional independent set assignments $\vec{x}$ with $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$. In this way, we can directly obtain an independent of size $n/\deg_{\mathrm{avg}}$, where $\deg_{\mathrm{avg}}$ denotes the average degree, and analogously in weighted graphs, we can obtain an independent set of total weight asymptotically at least a weighted average degree fractional of the total weight. For the corresponding definitions and details, we refer to \Cref{sec:turan}. In the same way, in graphs of neighborhood independence at most $\beta$, we can obtain a $1/O(\beta)$-approximation to the maximum weight independent set problem in time $O(\log^2(\Delta W)+\log^* n)$ (where $W$ is the ratio between largest and smallest weight). By adapting an idea of \cite{KawarabayashiKS20} and repeating the $1/O(\beta)$-approximation algorithm in an appropriate way, at the cost of an additional $\log(1/\varepsilon)$ factor in the round complexity, we can even obtain a $(1-\varepsilon)/\beta$-approximation for maximum weight independent set.
\subsubsection{Our Generic Rounding Algorithm}
\label{sec:genericroundingsummary}
In \Cref{sec:genericrounding}, we describe a general rounding algorithm that is based on the ideas discussed for the maximum independent set problem above. We consider local graph problems where each node has to choose one label from a finite alphabet $\Sigma$ of labels. In the $(\Delta+1)$-coloring problem, the set of labels is $\Sigma=\set{1,\dots,\Delta+1}$ or more generally the colors from some given color space if we consider the list coloring variant of the problem. In the independent set problem discussed above, there are only two possible labels, either a node decides to be in the independent set or it decides not to be in the independent set. In fact, in all problems for which is provide a new algorithm in this paper, we only have two labels. However, to cover the larger range of possible problems, the generic algorithm in \Cref{sec:genericrounding} is written for an arbitrary set $\Sigma$ of labels.
As in the maximum independent set example above, the quality of an assignment of labels to the nodes is measured by utility function $\mathbf{u}(\cdot)$ and a cost function $\mathbf{c}(\cdot)$. Both functions can be expressed as the sum of pairwise functions and the set of node pairs that contribute to the overall utility and/or cost naturally define a graph $H=(V_H,E_H)$ on the set of nodes. That is, utility and cost are defined as the sum of individual edge utilities and costs over the edges in $E_H$. Sometimes this graph $H$ is equal to the communication graph (such as in the discussed vertex coloring and maximum independent set examples) and sometimes $H$ is a graph that can be simulated efficiently by a distributed algorithm on the network graph (such as in the examples of MIS and set cover discussed below). In the latter case, the graph $H$ might also contain multiple edges between the same pair of nodes and those parallel edges might be simulated in different ways on the underlying communication graph. In our formal set-up, we therefore think of $H$ as a multigraph. A fractional label assignment $\lambda$ assigns a fractional value $\lambda_\alpha(v)\in[0,1]$ for every label $\alpha\in \Sigma$ to every node $v\in V$ such that for all $v\in V$, $\sum_{\alpha\in \Sigma} \lambda_\alpha(v)=1$, i.e., the fractional assignment defines a probability distribution over the labels to each node. The utility and cost of a fractional assignment are given as the expected utility and cost values when each node $v$ picks its label independently according to the distribution given by its fractional values $\lambda_\alpha(v)$.
Given some fractional label assignment $\lambda$, the goal of the rounding algorithm is to obtain an integral label assignment $\lambda'$ for which $\mathbf{u}(\lambda')-\mathbf{c}(\lambda')\geq (1-\varepsilon)\cdot\big(\mathbf{u}(\lambda)-\mathbf{c}(\lambda\big)$ for some given parameter $\varepsilon>0$. If initially $\mathbf{u}(\lambda)-\mathbf{c}(\lambda)\geq \mu \mathbf{u}(\lambda)$ for some $\mu\in(0,1]$, our rounding algorithm achieves this in time linear proportional $1/(\varepsilon \mu)$. Our algorithm is based on gradually rounding the solution as described above for rounding fractional independent set solutions. More specifically, for some integer $k\geq 0$, we call a fractional label assignment $\lambda$ $2^{-k}$-integral if all fractional values $\lambda_\alpha(v)$ are integer multiples of $2^{-k}$. In a single rounding step, we turn a $2^{-k}$-integral solution into a $2^{-(k-1)}$-integral one. In the general case, the rounding is implemented in exactly the same way as sketched above for the case of independent sets. in particular, we have to use a dynamically changing potential function analogous to \eqref{eq:potentialoverview} in order to keep utility and cost sufficiently separated to make each rounding step efficient. In each step, we have to use a weighted average defective coloring with $O(k/(\varepsilon \mu))$ colors so that an implementation of a single rounding step would take $O(k/(\varepsilon \mu))$ rounds when running it on $H$ and with potentially large messages. Below, we discuss some of the ideas that are necessary to implement the rounding algorithm with small messages. The precise formal setup, the detailed generic rounding algorithm, its analysis, and implementation in the distributed setting appear in \Cref{sec:genericrounding}.
\subsubsection{Maximal Independent Set}
The standard randomized distributed algorithm to compute an MIS of a graph $G=(V,E)$ is an algorithm that is usually referred to as Luby's algorithm~\cite{alon86,luby86}. One version of this algorithm consists of iteratively applying the following basic step. Every node $v\in V$ marks itself with probability $1/(2\deg(v))$. Each marked node $u$ joins the independent set unless $u$ has a higher priority marked neighbor $v$. A neighbor $v$ has higher priority if either $\deg(v)>\deg(u)$ or if $\deg(v)=\deg(u)$ and if $v$ has a larger ID than $u$. After that, the new independent set nodes and their neighbors are removed from the graph and the step is again applied on the remaining graph. The standard analysis shows that, in expectation, a constant fraction of the edges of the graph are removed in each step. This is shown by defining a set of good nodes, showing that a constant fraction of all edges are incident to at least one good node, and showing that good nodes are removed with constant probability. In \Cref{sec:MIS}, we show that this algorithm and analysis can be adapted to the setting to which we can apply our generic rounding method. This allows to deterministically find an independent set that removes a constant fraction of the good nodes and thus a constant fraction of all the edges.
To illustrate some of the key ideas, in this overview, we consider a significantly simpler setting. We assume that $G$ is $\Delta$-regular and we discuss how to find an independent set that removes a constant fraction of all the nodes in this case. In \Cref{sec:MIS}, we do a careful per-node analysis to count the number of edges incident on good nodes that are removed. Here, where things are more symmetric, we can directly analyze the total number of nodes that are removed.
For the initial fractional solution, we set $x_v=1/(2\Delta)$ for each node $v\in V$. After rounding to an integral solution, we proceed as in the independent set algorithm above. For every edge $\set{u,v}$ on which both nodes are marked, we unmark the lower-ID node. The set of marked nodes then forms an independent set. In order for this independent set to remove a constant fraction of the nodes, we have to build a utility and a cost function that measures the total number of nodes that get removed. Each node $v\in V$ can remove all its $\Delta$ neighbors if $v$ joins the independent set. We therefore define the overall utility as $\mathbf{u}(\vec{x})=\Delta\cdot\sum_{v\in V} x_v$. In order to lower bound the number of nodes that get removed, we have to (a) deduct the number of neighbors of nodes that get unmarked because they have a higher-ID marked neighbor and (b) make sure that we do not double count the removal of nodes. For (a), we can proceed in the same way as when computing large independent sets. For each edge $\set{u,v}$ for which both nodes are marked, the lower-ID node gets unmarked and we have to subtract $\Delta$ from the number of removed nodes. Therefore, the cost function contains a term of the form $\Delta\cdot\sum_{\set{u,v}\in E}x_u x_v$. For (b), note that the removal of a node $v$ is counted $\kappa$ times if $\kappa$ neighbors of $v$ join the independent set. Hence, the removal of node $v$ is counted at most $\kappa$ times if $\kappa$ neighbors of $v$ are initially marked. Further, the number of marked neighbors of $v$ minus $1$ (i.e., the number of times $v$'s removal is overcounted) can clearly be upper bounded by the total number of node pairs among the neighbors of $v$ such that both nodes in the pair are marked. The cost function therefore also contains a second term that upper bounds the number of overcounted node removals as $\sum_{u\in V}\sum_{\set{v,w}\in \binom{N(u)}{2}}x_v x_w$, where $\binom{N(u)}{2}$ denotes the set of $2$-element subsets of $N(u)$. We can therefore lower bound the expected number of removed nodes as
\[
\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}) = \Delta\cdot\sum_{u\in V} x_u\ -\ \Delta\cdot\!\!\sum_{\set{u,v}\in E} x_u\cdot x_v - \sum_{u\in V}\sum_{\set{v,w}\in \binom{N(u)}{2}} x_v\cdot x_w.
\]
Note that the utility and cost functions can be expressed as sums over individual edge utilities and costs, if we do this on a graph $H$ for which we add an edge between any two nodes at distance $2$ in $G$. By using our generic rounding algorithm, we can therefore round a fractional solution to an integral one of approximately the same quality (i.e., an independent set that removes approximately $\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x})$ nodes, where $\vec{x}$ is the initial fractional assignment). For the initial fractional assignment $x_v=1/(2\Delta)$ for all $v\in V$, the expected number of removed nodes can be lower bounded by
\[
\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x}) = \frac{n}{2} - \Delta\cdot\frac{n\Delta}{2}\cdot\frac{1}{4\Delta^2} -
n\cdot \binom{\Delta}{2} \cdot \frac{1}{4\Delta^2}\geq \frac{n}{4}.
\]
Since we start with a $\Theta(1/\Delta)$-fractional assignment, the rounding will require $O(\log\Delta)$ steps and each step requires $O(\log\Delta)$ rounds (when communicating on $H$). In time $O(\log^2\Delta)$ we can therefore compute an independent set that removes a constant fraction of the nodes (after computing an initial $O(\Delta^2)$-coloring in time $O(\log^* n)$ at the very beginning of the algorithm). We show in \Cref{sec:MIS} that, in the same time complexity and with a natural generalization of the same cost function, we can compute an independent set that removes a constant fraction of the edges in time $O(\log^2\Delta)$ in the $\mathsf{LOCAL}$\xspace model. Therefore, we obtain an $O(\log^2\Delta\cdot\log n)$-round deterministic MIS algorithm for the $\mathsf{LOCAL}$\xspace model.
\subsubsection{Set Cover Approximation}
We do not provide a complete overview of our set cover approximation algorithm here, however, we remark one key difference: In the set cover problem, we have $O(\log s)$ iterations of rounding the fractional solution (where $s$ is the maximum set size). To avoid the need for gathering the parameters of the instance remaining after each iteration (which would require global communication), we define a potential function (in the format of utility minus cost) for the entirety of the iterations. This is unlike the maximal independent set solution, where each iteration is handled separately, with its own potential function. The particular potential function we use and our gradual rounding ensure that, while per iteration we lose a constant factor in comparison to what the randomized algorithm would achieve, over the entire span of the $O(\log s)$ iterations, we still manage to cover all but a $1/poly(s)$ fraction of elements. After that, a simple greedy step of taking a set for each remaining element finishes the problem at a negligible cost. We refer to \Cref{sec:setcover} for the details and the algorithm.
\subsubsection{Implementation With Small Messages}
Note that both for computing an MIS and for computing a small set cover, the rounding graph includes edges that are between nodes at distance of $2$ in the communication graph $G$. While such edges are trivial to simulate in the $\mathsf{LOCAL}$\xspace model with only a factor $2$ overhead, the same is not true for the $\mathsf{CONGEST}$\xspace model. That is, when running a $\mathsf{CONGEST}$\xspace algorithm on the rounding graph $H$ on the underlying communication graph $G$, the necessary message size on $G$ can potentially grow by a factor linear in $\Delta$, which would not be feasible within $\operatorname{\text{{\rm poly}}}(\log n)$-time. Therefore, to obtain efficient $\mathsf{CONGEST}$\xspace algorithms on $G$ that perform rounding algorithms on $H$, we cannot work with a black-box interface; we have to instead examine the actual communication needed for the rounding.
For each virtual edge of $H$ which is between two nodes $v$ and $u$ that are at distance $2$ and have a common neighbor $w$, we make $w$ the manager of this virtual edge and let $w$ simulate the computations and communications relevant for this edge. Notice that a node $v$ might not even know all of its virtual edges. However, the managers of all those virtual edges are direct neighbors of $v$. Since $v$ can share its state (e.g., its current color, or its current fractional label assignment) efficiently with its neighbors, the managers are able to simulate the computations and communications of the virtual edges. The only possible difficulty is that this information has to be delivered back to node $v$, who is the endpoint of those virtual edges and will use this information to take some action, e.g., set its new color or increase/decrease its fractional value. Fortunately, in all cases, simply aggregating the information that needs to be sent back to $v$ suffices, so each manager $w$ can send to its direct neighbor $v$ some aggregate information about all the virtual edges that $w$ is managing for $v$. Furthermore, we note that doing this naively would result in an additional $\log \Delta$ factor in the round complexity, due to the message sizes in a part of the defective coloring procedure. However, fortunately, in that case, we can make do with a $2$ approximate aggregation of the values, instead of the exact values, and this reduces the round complexity loss exponentially, to only a $\log\log \Delta$ factor (which is essentially the only difference between the round complexity in the $\mathsf{LOCAL}$\xspace model and that of the $\mathsf{CONGEST}$\xspace model. Please see \Cref{sec:d2communication} for details.
\subsection{Mathematical Preliminaries and Notation}
We use $\mathbb{R}_{\geq 0}:=\set{x\in \mathbb{R} : x\geq 0}$ to denote the non-negative reals. Further, for a positive integer $N$, we use $[N]$ as a shortcut for $[N]:=\set{1,\dots,N}$. For an undirected multigraph $G=(V,E)$ and a node $v\in V$, we use the following notation. For every edge $e$, we use $V(e)$ to denote the set of nodes of $e\in E$ (note that if $G$ has no self-loops, $|V(e)|=2$ for all $e\in E$). Further, for every node $v\in V$, we use $E(v)\subseteq E$ to denote the set of edges that contain node $v$. For convenience, for any subsets $F\subseteq E$ of the edges and any node $v\in V$, we also define the shortcut $F(v):=E(v)\cap F$. If $G$ is equipped with edge (or node) weights $w:E\to \mathbb{R}$ (or $w:V\to \mathbb{R}$), for any set $F\subseteq E$ (or $U\subseteq V$), we use $w(F)$ (or $w(U)$) to denote the sum of the weights of all edges in $F$ (or nodes in $U$). When dealing with simple graphs, we also use the more standard definition and consider the edges $E$ as a subset of $\binom{V}{2}$, where $\binom{V}{2}$ denotes the set of $2$-element subsets of $V$. For a node $v\in V$, we use $N(v)$ to denote the set of neighbors of $v$ and we use $N^+(v):=\set{v}\cup N(v)$ to denote the inclusive neighborhood of $v$. For a graph $G=(V,E)$, we will typically use $n$ and $\Delta$ to refer to the number of nodes and the maximum degree of $G$, respectively. Finally, we say that a graph $G=(V,E)$ has \emph{neighborhood independence} $\beta\geq 1$ if for every node $v\in V$, the the largest independent set of the subgraph induced by $N(v)$ (or equivalently by $N^+(v)$) is of size at most $\beta$. Note that in particular, line graphs of hypergraphs of rank $r$ have neighborhood independence at most $r$.
As a technical tool to obtain fast deterministic rounding algorithms, we need to compute \emph{defective colorings} of the graph. More concretely, we need an edge-weighted version of standard defective coloring as it was introduced in \cite{KawarabayashiS18}. In order to obtain efficient defective coloring algorithms, we use a relaxed variant of defective coloring in which the (relative) defect per node only needs to be small on average.
\begin{definition}[Weighted Defective and Weighted Average Defective Coloring]\label{def:avgdefcoloring}
Given a weighted undirected multigraph $G=(V,E,w)$ with non-negative edge weights $w(e)\geq 0$ for all $e\in E$, a parameter $\varepsilon>0$, and an integer $C\geq 1$, a \emph{weighted $\varepsilon$-relative defective $C$-coloring} of $G$ is an assignment $\varphi:V\to [C]$ of colors in $[C]$ to the nodes of $G$ such that
\[
\forall v\in V\,:\,\sum_{e \in E(v), \, V(e)=\{v, u\}}\mathds{1}_{\set{\varphi(u)=\varphi(v)}}\cdot
w(e) \ \leq\ \varepsilon\cdot\sum_{e \in E(v)} w(e).
\]
The coloring is called a \emph{weighted average $\varepsilon$-relative defective $C$-coloring} if the above condition only holds on average over all nodes:
\[
\sum_{v\in V}\,\,\sum_{e \in E(v), \, V(e)=\{v, u\}}\mathds{1}_{\set{\varphi(u)=\varphi(v)}}\cdot
w(e) \ \leq\ \varepsilon\cdot\sum_{v\in V}\sum_{e \in E(v)} w(e).
\]
\end{definition}
As shown in \cite{KawarabayashiS18}, by slightly adapting an algorithm of \cite{Kuhn2009WeakColoring}, an $\varepsilon$-relative defective $O(1/\varepsilon^2)$-coloring can be computed in time $O(\log^* \xi)$ in the $\mathsf{CONGEST}$\xspace model, if an initial proper vertex coloring with $\xi$ colors is given. Further, by using algorithm of \cite{GhaffariK21} (or by adapting an coloring algorithm of \cite{BEG18}), an $\varepsilon$-relative average defective $O(1/\varepsilon)$-coloring can be computed in time $O(1/\varepsilon + \log^* \xi)$. We also note that the notion of average defect is a relaxation of the notion of arbdefect that was introduced in \cite{barenboim10}.
\section*{Acknowledgments}
M.G., C.G. and V.R. were supported in part by the European Research Council (ERC) under the European Unions Horizon 2020 research and innovation program (grant agreement No.~853109) and the Swiss National Science Foundation (project grant 200021\_184735).
\bibliographystyle{alpha}
\section{Minimum Set Cover}
\label{sec:setcover}
In this section we use our framework to give a fast deterministic algorithm for the set cover problem.
In this problem, the input consists of a universe of elements and a family of their subsets. We want to find the smallest subfamily covering all elements. We will use the following bipartite-graph formulation of the problem.
\begin{definition}
\label{def:set_cover}
In the \emph{set cover} problem, the input consists of a bipartite graph $G$ with $V(G) = U \sqcup V$.
The goal is to find the smallest possible subfamily $V_{out} \subseteq V$ such that $N(V_{out}) = U$.
\end{definition}
The section is dedicated to the proof of the following theorem.
\begin{theorem}
\label{thm:set_cover}
Consider an instance of a set cover problem where $\max_{u \in U} \deg(u) \le t$ (i.e., each element $u \in U$ is present in at most $t$ sets) and $\max_{v \in V} \deg(v) \le s$ (i.e., each set $v \in V$ contains at most $s$ elements). There is a distributed $O(\log s)$-approximation algorithm for the problem with round complexity of
\begin{enumerate}
\item $O(\log s \cdot \log ^2 t + \log^* n)$ in the $\mathsf{LOCAL}$\xspace model,
\item $O\left(\log s \ \cdot \left( \log^2 t\ \cdot \log\log t + \log t \cdot \log^* n\right)\right)$ in the $\mathsf{CONGEST}$\xspace model.
\end{enumerate}
\end{theorem}
The rest of this section is devoted to the proof of \cref{thm:set_cover}.
The outline of the proof is similar to the MIS algorithm from \cref{sec:MIS}:
We first discuss how this problem can be solved using a randomized algorithm, then we construct a natural pessimistic estimator for the randomized algorithm, and finally we verify that the estimator satisfies the requirements of \cref{lemma:d2rounding}.
\paragraph{A Randomized Set Cover Algorithm}
We start with an informal discussion of a randomized algorithm that we derandomize to get \cref{thm:set_cover}.
We rely on the following result of \cite{nearsighted} for the \emph{fractional} variant of the set cover problem.
There, we ask for a function $x : V \rightarrow [0,1]$ that assigns a fractional value to each set $v \in V$ and we require each element $u \in U$ to satisfy that $\sum_{v \in N(u)} x_v \ge 1$. Note that a solution to the set cover problem directly implies a solution to its fractional variant.
\begin{theorem}[Kuhn, Moscibroda, and Wattenhofer~\cite{nearsighted}]
There exists a distributed algorithm that computes a $2$-approximate solution to the fractional set cover problem in $O(\log s \cdot \log t)$ rounds of the \CONGEST model.
\end{theorem}
Consider the following randomized algorithm that has $\tau = O(\log s)$ rounds and in each round it simply constructs a subset $V'_i \subseteq V$ by including each $v \in V'_i$ with probability $x_v$.
Note that if we define $V' = \bigcup_{i = 1}^{\tau} V'_i$, we have for the expected size of $V'$ that $\E[x_{V'}] = O(\log s) \cdot OPT$.
On the other hand, we observe that whenever we consider some node $u$ not covered by $V'_1 \cup \dots \cup V'_i$, i.e., $u \not\in N(V'_1 \cup \dots \cup V'_i)$, we have $u \in N(V'_{i+1})$ with constant probability because $\sum_{v \in N(u)} x_v \ge 1$.
Hence, choosing $\tau = O(\log s)$ large enough, we get that the expected number of uncovered elements is $\E\left[ U \setminus N(V')\right] \le |U|/s$.
This means that although $V'$ is not expected to be a solution to the set cover problem, we can fix this issue as follows. Observe that for the size of the optimum solution $OPT$ we have
\begin{align}
\label{eq:opt_lb}
OPT \ge |U| / s
\end{align}
because every subset of $V$ smaller than $|U|/s$ would cover less than $|U|/s \cdot s = |U|$ elements.
This means that we can define $V''$ by letting each uncovered element of $U \setminus N(V')$ choose an arbitrary neighboring node $v \in V$ that we add to $V''$.
The set $V' \cup V''$ clearly covers all nodes of $U$ and its expected size is at most $|V'| + |U \setminus N(V')| = O(\log s) \cdot OPT + |U|/s = \left( 1 + O(\log s)\right) \cdot OPT$.
\paragraph{The Deterministic Algorithm}
In \cref{alg:set_cover}, we present the derandomized version of the above randomized procedure.
In the rest of the section, we first explain the algorithm, then verify that we can use \cref{lemma:distributedonG}, prove that the algorithm is $O(\log s)$-approximate, and finally discuss its implementation in the \LOCAL and \CONGEST models.
Let us explain the intuition behind \cref{alg:set_cover}. The issue with the derandomization is that we want to optimize two conflicting goals.
On one hand, we want to cover as many elements of the remaining set $U_i$ as possible. This means that we want to round our fractional solution $x$ to maximize the value of the following pessimistic lower bound for $|U_i| \setminus |U_{i+1}|$ coming from the pairwise analysis:
\[
\sum_{u \in U_i} \sum_{v \in V} \left( x_v - \sum_{v' \not= v \in V} x_{v'} \right) .
\]
On the other hand, we want to minimize the number of selected nodes of $V$, i.e., \[
\sum_{v \in V} x_v.
\]
To optimize these two conflicting expressions, we simply subtract the second one from the first one and invoke \cref{lemma:distributedonG} that can deal with summing terms of different signs.
An issue with this approach is that to optimize both expressions at once, they should be of the same order of magnitude. The value of $|U_i|$ however gradually decreases from $|U|$ to $|U|/s^{O(1)}$ until we can finish by constructing the set $V''$ as in the randomized algorithm. This is why in \cref{alg:set_cover} we slowly increase the relative weight of the first optimized expression with respect to the second one (see \cref{line:u,line:c}).
We also add a constant term of $10\sum_{v \in V}x_v$ to the utility function ${\bf u}(\vec{\tilde{x}})$ so that our functions satisfy ${\bf u}(\vec{x}) - {\bf c}(\vec{x}) \ge {\bf u}(\vec{x})/2$ which is needed to apply \cref{lemma:distributedonG}.
\begin{algorithm}[ht]
\caption{Deterministic Set Cover}
\label{alg:set_cover}
Input: A bipartite graph $G$ with $V(G) = U \sqcup V$.\\
Output: A set $V_{out} \subseteq V$ with $N(V_{out}) = U$
\begin{algorithmic}[1]
\State $\tau \leftarrow O(\log s)$
\State $\vec{x}_0: V \rightarrow [0,1]$ is a $2$-approximate fractional set cover of \cite{nearsighted
\State For every $v \in V$, define $x_v \leftarrow x_0(v)/10$ unless $x_0(v) \le 1/(2t)$ in which case $x_v \leftarrow 0$.
\State For every $u \in U_i$ define $N^*(u) \subseteq N(u)$ such that $\frac{1}{20} \le \sum_{v \in N^*(u)}x_v \le \frac{1}{5}$ (see \cref{eq:frac2}) \label{line:n*}
\State Compute a $2$-hop coloring of $G$ with $\zeta = (st)^2$ colors using Linial's algorithm \cite{linial1987LOCAL}
\For{$i \leftarrow 1, \dots, \tau$}
\State $U_i \leftarrow U \setminus \left( V'_1 \cup \dots \cup V'_{i-1} \right)$
\State Define ${\bf u}(\vec{\tilde{x}}) \leftarrow \frac{1}{1.01^{\tau-i}} \cdot \sum_{u \in U_i} \sum_{v \in N^*(u)} \tilde{x}_v + 10\sum_{v \in V} x_v$ \label{line:u}
\State Define ${\bf c}(\vec{\tilde{x}}) \leftarrow \frac{1}{1.01^{\tau-i}} \cdot \sum_{u \in U_i} \sum_{v \in N^*(u)}\sum_{v'\not=v \in N^*(u)} \tilde{x}_v \tilde{x}_{v'}
+ \sum_{v \in V} \tilde{x}_v$\label{line:c}
\State By plugging $\vec{\tilde{x}} = \vec{x}$, use \cref{lemma:distributedonG} with ${\bf u}, {\bf c}, \zeta, \varepsilon = \frac{1}{100}, \mu = \frac12, \lambda_{\min} = \frac{1}{2t}$ to round $\vec{x}$ to $\vec{x_i'}$
\State Define $V'_i \leftarrow \{v \in V: x_i'(v) = 1\}$
\EndFor
\State Define $V''$ by letting each element of $U \setminus N\left( V'_1 \cup \dots \cup V'_{\tau} \right)$ choose an arbitrary neighbor.
\State \Return {$V_{out} = V'_1 \cup \dots \cup V'_\tau \cup V''$}
\end{algorithmic}
\end{algorithm}
\paragraph{Properties of Fractional Weights}
Let us first discuss the properties of the fractional solution $\vec{x}$ that we repeatedly round inside \cref{alg:set_cover}. By definition, $\vec{x}$ satisfies for every $u \in U$ that
\begin{align}
\label{eq:frac0}
x_v \in \{0\} \cup [1/(20t), 1/10],
\end{align}
so in particular $\vec{x}$ is $t$-fractional. Next, we note that for every $u \in U$ we have
\begin{align}
\label{eq:frac1}
\sum_{v \in N(u)} x_v \ge 1/20.
\end{align}
To see this, let $N_{small}(u) \subseteq N(u)$ be the subset of nodes $v \in N(u)$ such that $x_0(v)\le 1/(2t)$. We have $\sum_{v \in N_{small}(u)} x_v \le t \cdot 1/(2t) \le 1/2$ where we used $|N_{small}(u)| \le |N(u)| \le t$.
Hence $\sum_{v \in N(u)} x_v \ge \sum_{v \in N(u) \setminus N_{small}(u)} x_v \ge (1/2) / 10 = 1/20$.
In view of \cref{eq:frac1}, in \cref{line:n*} we can define $N^*(u)$ for each $u \in U$ by repeatedly adding neighbors of $u$ to $N^*(u)$ until $\sum_{v \in N^*(u)} x_v \ge 1/20$.
By \cref{eq:frac0} we then get $\sum_{v \in N^*(u)} x_v \le 1/20 + 1/10 \le 1/5$. Hence, we have
\begin{align}
\label{eq:frac2}
1/20 \le \sum_{v \in N^*(u)} x_v \le 1/5
\end{align}
as required in \cref{line:n*}.
Finally, observe that
\begin{align}
\label{eq:frac3}
\sum_{v \in V} x_v \le 2\cdot OPT/10
\end{align}
where we use that $x_0$ is $2$-approximation of $OPT$.
\paragraph{Checking Assumptions for \cref{lemma:distributedonG}}
We continue by verifying the assumptions of \cref{lemma:distributedonG}.
\cref{eq:frac0} implies that $\lambda_{min} = 1/(20t)$ is the smallest nonzero value to round. Next, we need to prove that ${\bf u}(\vec{x}) - {\bf c}(\vec{x}) \ge {\bf u}(\vec{x})/2$.
To see this, note that for every $u$ we have
\begin{align}
\label{eq:ass}
\sum_{v \in N^*(u)} \sum_{v' \not= v \in N^*(u)} x_v x_{v'}
\le \sum_{v \in N^*(u)}\left( x_v \cdot \left( \sum_{v'\in N^*(u)} x_{v'}\right)\right)
\le \sum_{v \in N^*(u)} x_v/5
\end{align}
using \cref{eq:frac2}.
That is, the first term of ${\bf c}(\vec{x})$ is dominated by the first term of ${\bf u}(\vec{x})$. Similarly, by definition the second term of ${\bf c}(\vec{x})$ is dominated by the second term in ${\bf u}(\vec{x})$ and we thus have
\begin{align}
{\bf c}(\vec{x}) \le {\bf u}(\vec{x})/5.
\end{align}
\paragraph{Analyzing One Step With Pessimistic Estimators}
The utility and cost functions ${\bf u}(\vec{x})$ and ${\bf c}(\vec{x})$ correspond to a pessimistic estimator for the original randomized procedure.
Namely, consider sampling each $v \in S_i$ with probability $x_v$ that yields a binary vector $\vec{x'}$. For a fixed $u \in U_i$ let $\mathcal{E}_{u,v}$ be the event that $v$ is the only neighbor of $u$ in $N^*(u)$ that is sampled. By union bound the indicator of the event $R_u(\vec{x})$ that $u \in U_i$ gets covered by a set selected in $\tilde{x}'$ can be bounded for any $\vec{x'}$ as
\begin{align}
\label{eq:domination}
R_u(\vec{x'}) \ge \sum_{v \in N^*(u)} \left( x'_v - \sum_{v'\not= v \in N^*(u)} x'_v x'_{v'} \right)
\end{align}
For our rounded weights $x'$, this fact implies that after we plug in the definition of ${\bf u}(\vec{x'}), {\bf c}(\vec{x'})$ we have
\begin{align}
\label{eq:fingers}
{\bf u}(\vec{x'}) - {\bf c}(\vec{x'}) - 10\sum_{v \in V} x_v
&\le \frac{1}{1.01^{\tau - i}}\sum_{u \in U_i} R_u(\vec{x'}) - \sum_{v \in V}x'(v)
= \frac{1}{1.01^{\tau - i}}\left( |U_i| - |U_{i+1}|\right) - |V'_i|
\end{align}
On the other hand, let us now bound the value of $ {\bf u}(\vec{x'}) - {\bf c}(\vec{x'})$ from the other side by comparing it with the ideal randomized process corresponding to fractional weights $\vec{x}$. For those we have
\begin{align}
\label{eq:above}
{\bf u}(\vec{x}) - {\bf c}(\vec{x}) - 10\sum_{v \in V}x_v
&= \frac{1}{1.01^{\tau - i}} \sum_{u \in U_i} \sum_{v \in N^*(u)} \left( x_v - \sum_{v'\not= v \in N^*(u)} x_v x_{v'} \right) - 10\sum_{v \in V}x_v\\
&\ge \frac{1}{1.01^{\tau - i}} \cdot |U_{i}|\cdot \frac{4}{5} \cdot \frac{1}{20} - 2OPT\nonumber
\end{align}
where we first used \cref{eq:ass} to get rid of the quadratic term and then used the lower bound on $\sum_{v \in N(u)} x_u \ge 1/20$ from \cref{eq:frac2}. We also used \cref{eq:frac3}.
We now use \cref{lemma:distributedonG} to relate $x'$ with $x$. Our choice of $\varepsilon_{\text{L\ref{lemma:distributedonG}}} = 1/100$ implies that
\begin{align}
\label{eq:toes}
{\bf u}(\vec{x'}) - {\bf c}(\vec{x'}) - 10\sum_{v \in V}x_v
&\ge \frac{99}{100} \cdot \left( {\bf u}(\vec{x}) - {\bf c}(\vec{x})\right) - 10\sum_{v \in V}x_v\\
&= \frac{99}{100} \cdot \left( {\bf u}(\vec{x}) - {\bf c}(\vec{x}) - 10\sum_{v \in V}x_v\right) - 0.1\sum_{v \in V}x_v\nonumber\\
&\ge 0.02 \cdot \frac{1}{1.01^{\tau - i}} |U_i| - 3OPT\nonumber
\end{align}
where the last inequality follows from \cref{eq:above} together with the fact that $0.1 \sum_{v \in V} x_v < OPT$ by \cref{eq:frac3}.
By comparing \cref{eq:fingers} with \cref{eq:toes}, we conclude that
\begin{align}
\label{eq:pot_dec}
\frac{1}{1.01^{\tau - i}}\left( |U_i| - |U_{i+1}|\right) - |V'_i|
&\ge 0.02\cdot \frac{1}{1.01^{\tau - i}} |U_i| - 3OPT
\end{align}
\paragraph{Finishing the analysis}
To analyze the progress of the algorithm, we define the following potential function $\Phi$ in every step.
\begin{definition}
For every $0 \le i \le \tau = O(\log s)$, we define
\[
\Phi_i = \frac{1}{1.01^{\tau-i}} \cdot |U_i| + |V'_1| + |V'_2| + \dots + |V'_{i-1}| + 3(\tau-i)\cdot OPT
\]
\end{definition}
At the very beginning before the first step of the algorithm, we have $|U_0| = |U|$ and $|S_0| = 0$, hence $\Phi_0 = \frac{1}{1.01^{\tau}} \cdot |U| + 2\tau \cdot OPT$. Choosing $\tau = O(\log s)$ large enough, we have $\Phi_0 \le |U|/s + 2\tau \cdot OPT \le 3\tau \cdot OPT$ using \cref{eq:opt_lb}.
We will now prove that the potential $\Phi$ is monotone, that is $\Phi_{i+1} \le \Phi_{i}$ for every $1 \le i \le \tau$.
Let us first see why the proof of monotonicity of $\Phi$ also finishes the analysis.
It implies that $\Phi_\tau \le 3\tau \cdot OPT$. Using the definition of $\Phi$, this means that $|U_\tau| + |V'_1| + \dots + |V'_\tau| \le 3\tau \cdot OPT$.
Since $|U_\tau| = |V''|$, we conclude that $|V_{out}| \le 3\tau\cdot OPT = O(\log s) \cdot OPT$ as needed.
To see that $\Phi$ is monotone, we write the difference of two consecutive potentials as
\begin{align}
\Phi_{i} - \Phi_{i+1}
&= \frac{1}{1.01^{\tau - i}} \left( |U_{i}| - 1.01|U_{i+1}| \right) - |V'_{i}| + 3OPT\\
& = \frac{1}{1.01^{\tau - i}} \left( |U_{i}| - |U_{i+1}| \right) - |V'_{i}| + 3OPT - 0.01 |U_{i+1}|/1.01^{\tau - i}\\
\end{align}
Applying \cref{eq:pot_dec}, we conclude that
\begin{align}
\Phi_{i} - \Phi_{i+1}
&\ge 0.02 \cdot \frac{1}{1.01^{\tau - i}} |U_i| - 3OPT + 3OPT - 0.01 |U_{i+1}|/1.01^{\tau - i}
\ge 0
\end{align}
and we are done.
\paragraph{Implementation and Round Complexity}
We begin by discussing the implementation in the \LOCAL model.
The algorithm of \cite{nearsighted} needs $O(\log s \cdot \log t)$ rounds and we construct the 2-hop coloring of $G$ with $\zeta = (st)^2$ colors in $O(\log^* n)$ rounds using Linial's $\Delta^2$-coloring algorithm \cite{linial1987LOCAL}.
The complexity of every subsequent step is dominated by the call to \cref{lemma:distributedonG} with round complexity $O(\log^2 (1/\lambda_{\min}) + \log^*(\zeta)) = O(\log^2 t + \log^* (st))$.
We note that if the first term in this expression does not dominate the second one, we have in particular $t < \log s$. In that case, we can however achieve $O(\log s)$-approximation by simply taking all nodes $v \in V$ with $x_0(v) \ge 1/t$ to $V_{out}$. This solution clearly covers all elements of $U$ since every $u \in U$ has necessarily a neighbor with $x_0(v) \ge 1/|N(u)| \ge 1/t$. Moreover, its weight is at most $t$-times larger than the fractional weight $x$. Hence, we may assume that $O(\log^2 t + \log^*(st)) = O(\log^2 t)$ and we are getting an algorithm with round complexity $O(\log s \cdot \log^2 t + \log^* n)$ as needed.
We continue by discussing the implementation in the \CONGEST model. There, we simply replace all calls of \cref{lemma:distributedonG} by calls to \cref{lemma:d2rounding}.
Another change that we do is that we do not construct the $2$-hop coloring with $\zeta = (st)^2$ colors at the beginning, but simply use the unique identifiers instead, i.e., we have $\zeta = n^{O(1)}$.
We are using the $d2$-multigraph $H$ where each node $u \in U$ simulates a virtual edge between every pair of its neighbors $v, v'$.
The round complexity of one round of the algorithm becomes $O\left(\log s \ \cdot \left( \log^2 t\ \cdot \log\log t + \log t \cdot \log^* n\right)\right)$.
\begin{remark}
We believe that one can directly generalize \cref{thm:set_cover} to give an $O(\log (sW))$-approximation for the more general \emph{min cost set cover} problem. There, each node $v \in V$ comes with a cost $w(v)$ such that $1 \le w(v) \le W$. The goal is to find a subset $V_{out} \subseteq V$ of smallest total cost such that $N(V_{out}) = U$.
We discuss the changes that need to be done in \cref{alg:set_cover}. We use a min-cost version of the fractional algorithm of \cite{nearsighted} that needs $O(\log (sW) \cdot \log t)$ rounds. Then, on \cref{line:u,line:c} in \cref{alg:set_cover} we change the utility $10\sum_{v \in V} x_v$ to $10\sum_{v \in V} w(v) \cdot x_v$ and the cost $\sum_{v \in V)} \tilde{x}_v$ to $\sum_{v \in V)} w(v) \cdot \tilde{x}_v$.
Finally, we run the algorithm for $\tau = O(\log (sW))$ steps. This way, we have $|U \setminus N(V_1 \cup \dots \cup V_\tau)| \le |U|/(sW)$ which implies that $w(V'') \le OPT$.
The distributed complexity of this algorithm is
\begin{enumerate}
\item $O(\log (sW) \cdot \log ^2 t + \log^* n)$ in the $\mathsf{LOCAL}$\xspace model,
\item $O\left(\log (sW)\ \cdot \left( \log^2 t\ \cdot \log\log t + \log t \cdot \log^* n\right)\right)$ in the $\mathsf{CONGEST}$\xspace model, assuming $W \le n^{O(1)}$.
\end{enumerate}
\end{remark}
\section{Maximum Weight Independent Set}
\label{sec:weightedIS}
In this section, we apply our generic rounding scheme of \Cref{sec:basicrounding} to obtain fast deterministic algorithms for computing a large weight independent set of a graph $G=(V,E)$ with node weights $w:V\to\mathbb{N}$.\footnote{We assume that the node weights are positive integers to keep the algorithms and their analyses simpler. The algorithms can however all be generalized to handle arbitrary positive weights.} Note that we here assume that $G$ is a simple graph and we can therefore define $E$ as a subset of $\binom{V}{2}$. Further, when discussing distributed algorithms on $G$, for simplicity, we will generally assume that initially all nodes $u\in V$ know the weights of all their neighbors. Equivalently we can assume that we have enough bandwidth to exchange this information in a single round (e.g., that node weights can be encoded using $O(\log n)$ bits when using the $\mathsf{CONGEST}$\xspace model). Further, for a subset $U\subseteq V$ of nodes, we use $w(U):=\sum_{u\in U}w(u)$ as a shortcut for the total weight of nodes in $U$. We will also assume that all nodes know an upper bound $W$ on the maximum node weight.
While our rounding method allows us to study problems with a potentially large alphabet of node labels, as in \Cref{sec:MIS}, we will only need $2$ labels: either a node is in the independent set or the node is not in the independent set. We can thus again characterize a fractional solution by a single fractional value $x_v\in[0,1]$ per node $v\in V$, where $x_v$ is the fractional value for label $1$, i.e., the fractional value for being in the set. In an integral solution, we have $x_v\in \set{0,1}$ and the computed subset of nodes is defined by the nodes $v$ with $x_v=1$.
\paragraph{Definition of Utility and Cost.} We next define the utility and cost function for computing heavy independent sets. For simplicity, we define the utility as a sum over nodes instead of edges. It is however straightforward to get an equivalent definition that fits the framework of \Cref{sec:roundingsetup}. Let $\vec{x}\in[0,1]^{|V|}$ be the vector of fractional values $x_v$. Utility $\mathbf{u}(\vec{x})$ and cost $\mathbf{c}(\vec{x})$ are then defined as
\begin{equation}
\label{eq:ISutilityandcost}
\mathbf{u}(\vec{x}) := \sum_{v\in V} w(v)\cdot x_v
\qquad\text{and}\qquad
\mathbf{c}(\vec{x}) := \sum_{\set{u,v}\in E} \min\set{w(u), w(v)}\cdot x_u\cdot x_v.
\end{equation}
\paragraph{From an Integral Solution to an Independent Set.} Given integral node value $x_v\in\set{0,1}$, it is straightforward to obtain an independent set of total value $\mathbf{u}(\vec{x}) - \mathbf{c}(\vec{x})$: we start with the set $I_0$ of nodes $v$ for which $x_v = 1$ and for every two neighbors $u,v$ in $I_0$ we simply remove the one with smaller weight. This is formalized as follows.
\begin{lemma}\label{lemma:computingIS}
Let $G=(V,E)$ be a graph with node weights $w:V\to \mathbb{N}$ and let $\vec{x}$ be a vector of (integral) node values $x_v\in\set{0,1}$. Then, there is a deterministic $1$-round $\mathsf{CONGEST}$\xspace algorithm to compute an independent set $I$ of $G$ of total weight at least $\mathbf{u}(\vec{x}) -\mathbf{c}(\vec{x})$.
\end{lemma}
\begin{proof}
Let $I_0\subseteq V$ be the set of nodes $v\in V$ with $x_v=1$. We define a set $I'\subseteq V$ of nodes to be removed as follows. A node $u\in I_0$ is in $I'$ iff there is a neighbor $v\in N(u)\cap I_0$ such that either $w(v)>w(u)$ or $w(v)=w(u)$ and the ID of $v$ is larger than the ID of $u$. The independent set $I$ is then defined as $I:=I_0 \setminus I'$. Clearly, this algorithm can be implemented in a single $\mathsf{CONGEST}$\xspace algorithm. To determine if a node $u\in I_0$ is in the set $I$, it only needs to know which of its neighbors are in $I_0$ and it needs to know the weights and IDs of those neighbors (recall that we assume that node weights can be communicated in a single round or that they the node weights of neighbors are known initially). It is further not hard to see that $I$ is an independent set. For every edge $\set{u,v}\in E$, if $u\in I_0$ and $v\in I_0$, at least one of the two nodes is added to $I'$. If $w(u)\neq w(v)$, then the node of smaller weight is added to $I'$, otherwise, the node of smaller ID is added to $I'$. It remains to show that the weight $w(I)$ of the independent set $I$ is as claimed. We clearly have
\[
w(I) = w(I_0) - w(I')\qquad\text{and}\qquad w(I_0) = \mathbf{u}(\vec{x}).
\]
We therefore need to show that $w(I')\leq \mathbf{c}(\vec{x})$. To see this, note that by the construction of $I'$, the inclusion of a node $v$ in $I'$ can be blamed on an incident edge $\set{u,v}$ and every edge with $x_u=x_v=1$. Further, every such edge can be blamed by at most one node $v\in I'$ and we also know that if $v\in I'$ is blamed on edge $\set{u,v}$, then $w(u)\geq w(v)$. The contribution of edge $\set{u,v}$ to $\mathbf{c}(\vec{x})$ is equal to $\min\set{w(u), w(v)}\cdot x_u\cdot x_v = w(v)$ and we therefore have $w(I')\leq \mathbf{c}(\vec{x})$ as required.
\end{proof}
\paragraph{Basic Independent Set Rounding Algorithm.}
We next show that given a fractional solution $\vec{x}$ for which the utility is sufficiently larger than the cost, there is an efficient $\mathsf{CONGEST}$\xspace algorithm to compute a large independent set. Specifically, we prove the following lemma, which we will use in the remaining algorithms of this section.
\begin{lemma} \label{lemma:basicISalgorithm}
Let $G=(V,E)$ be a graph with maximum degree $\Delta$ and node weight function $w:V\to \mathbb{N}$. Further assume that we are given a fractional independent set solution $\vec{x}$ with $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$. If $G$ is equipped with a $\zeta$-vertex coloring for some $\zeta\leq\operatorname{\text{{\rm poly}}}(n)$, for every $\varepsilon\in(0,1]$, an independent set of total weight at least $(1/2 - \varepsilon)\mathbf{u}(\vec{x})$ can be computed in $O(\log^2(\Delta/\varepsilon) + \log^* \zeta)$ deterministic rounds in the $\mathsf{CONGEST}$\xspace model.
\end{lemma}
\begin{proof}
We first show that the fractional solution $\vec{x}$ can be transformed into a fractional solution $\vec{x}'$ with no small positive fractional values. Specifically, for all $v\in V$, we set $x_v' := x_v + \frac{\varepsilon}{2\Delta}$. We then have
\begin{eqnarray}
\mathbf{u}(\vec{x}')
& = &
\sum_{v \in V} w(v)\cdot \left (x _v+ \frac{\varepsilon}{2\Delta}\right)
\ =\ \mathbf{u}(\vec{x}) + w(V)\cdot\frac{\varepsilon}{2\Delta},\label{eq:utilityprime}\\
\mathbf{c}(\vec{x}')
& = &
\!\!\!\!\sum_{\set{u,v}\in E}\!\!\! \min\{w(u),w(v)\} \cdot
\left(x_u + \frac{\varepsilon}{2\Delta}\right)\cdot
\left(x_v + \frac{\varepsilon}{2\Delta}\right)\nonumber\\
& \le &
\!\!\!\!\sum_{\set{u,v}\in E}\!\!\! \left(
\min\{w(u),w(v)\} \cdot x_u x_v
+ w(u) x_u \frac{\varepsilon}{2\Delta} + w(v) x_v \frac{\varepsilon}{2\Delta} + \frac{w(u) + w(v) }{2} \cdot \left( \frac{\varepsilon}{2\Delta} \right)^2 \right)
\nonumber\\
& \leq &
\!\!\!\!\sum_{\set{u,v}\in E}\!\!\! \min\{w(u),w(v)\} \cdot x_u x_v +
\sum_{v\in V} \frac{\deg_G(v)\cdot w(v)\cdot x_v \cdot\varepsilon}{2\Delta} +
\frac{1}{2} \cdot \sum_{v \in V} \frac{\deg_{G}(v)\cdot w(v)\cdot\varepsilon^2}{4\Delta^2} \nonumber\\
& \leq &
\mathbf{c}(\vec{x}) + \frac{\varepsilon}{2}\cdot\mathbf{u}(\vec{x}) + w(V)\cdot\frac{\varepsilon^2}{8\Delta}\nonumber\\
& \leq &
\left(\frac{1}{2}+\frac{\varepsilon}{2}\right)\cdot\mathbf{u}(\vec{x}) + \frac{\varepsilon}{4}\cdot\big(\mathbf{u}(\vec{x}')-\mathbf{u}(\vec{x})\big).\label{eq:costprime}
\end{eqnarray}
In the last inequality, we use that $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$ and we use \eqref{eq:utilityprime}. We now have
\begin{eqnarray*}
\mathbf{u}(\vec{x}') - \mathbf{c}(\vec{x}')
& = &
\mathbf{u}(\vec{x}) + \big(\mathbf{u}(\vec{x}') - \mathbf{u}(\vec{x})\big) - \mathbf{c}(\vec{x}')\\
& \stackrel{\eqref{eq:costprime}}{\geq} &
\left(\frac{1}{2}-\frac{\varepsilon}{2}\right)\cdot\mathbf{u}(\vec{x}) +
\left(1-\frac{\varepsilon}{4}\right)\cdot \big(\mathbf{u}(\vec{x}') - \mathbf{u}(\vec{x})\big)
\geq
\left(\frac{1}{2}-\frac{\varepsilon}{2}\right)\mathbf{u}(\vec{x}').
\end{eqnarray*}
In the last inequality, we use that $\mathbf{u}(\vec{x}')\geq \mathbf{u}(\vec{x})$ and that $1-\varepsilon/4\geq 0$. By using \Cref{lemma:distributedonG} and the fact that for all $v\in V$, $x_v'\geq \frac{\varepsilon}{2\Delta}$, we directly get that an independent set of size $(1-\varepsilon)\cdot \big(\frac{1}{2}-\frac{\varepsilon}{2}\big)\cdot\mathbf{u}(\vec{x}')\geq \big(\frac{1}{2}-\varepsilon\big)\cdot\mathbf{u}(\vec{x})$ can be computed deterministically in $O(\log^2(\Delta/\varepsilon) + \log^*\zeta)$ rounds in the $\mathsf{CONGEST}$\xspace model. Here, we also use that the necessary message size of \cref{lemma:distributedonG} is $O(\log n)$, since $|\Sigma|=2$ and $\log\zeta=O(\log n)$).
\end{proof}
\subsection{Approximation for Graphs with Bounded Neighborhood Independence}
\label{sec:boundedindep}
We next show how to efficiently compute a $(1-\varepsilon)/\beta$-approximation to the maximum weight independent set (MWIS) problem of a graph $G=(V,E)$, where $\beta$ denotes the neighborhood independence of $G$ (i.e., the maximum number of pairwise non-adjacent neighbors of a node). We use the following linear program (LP) to define the family of fractional solutions that we use.
\begin{equation}\label{eq:fractionISLP}
\max \sum_{v\in V} w(v)\cdot x_v\qquad\text{s.t.}\quad
\forall v\in V\, :\!\!\! \sum_{u\in N^+(v)}\!\!\! x_u \leq 1\quad\text{and}\quad
x_v \geq 0.
\end{equation}
In the following, let $S^*(w)$ be the value of an optimal solution of LP \eqref{eq:fractionISLP} on graph $G$ with weight function $w$. Note that in a graph of neighborhood independence $\beta$, for each node $u\in V$, the number of nodes in $N^+(u)$ in any independent set is at most $\beta$. Hence, the weight of a maximum weight independent set is at most $\beta\cdot S^*(w)$. We therefore need to show that we can efficiently compute an independent set of weight at least $(1-\varepsilon)S^*(w)$. We first show that we can efficiently compute an independent set of weight at least $S^*(w)/4$.
\begin{lemma}\label{lemma:LPapproxIS}
Let $G=(V,E)$ be a graph with maximum degree $\Delta$ and node weight function $w:V\to \mathbb{N}$. If $G$ is equipped with an $O(\Delta^2)$-vertex coloring, an independent set of total weight at least $S^*(w)/4$ can be computed in $O\big(\log^2(\Delta W)\big)$ deterministic rounds in the $\mathsf{CONGEST}$\xspace model.
\end{lemma}
\begin{proof}
First, observe that a $(2/3)$-approximation of the LP \eqref{eq:fractionISLP} can be computed in $O\big(\log^2(\Delta W)\big)$ rounds by using the distributed covering and packing LP algorithm of \cite{nearsighted}.
Next, note that any feasible solution of \eqref{eq:fractionISLP} satisfies $\mathbf{u}(\vec{x})\geq 2 \mathbf{c}(\vec{x})$:
\begin{eqnarray*}
\mathbf{c}(\vec{x})
&=&\sum_{ \{u,v\} \in E}x_u\cdot x_v \cdot \min\{w(u),w(v)\}\\
& = & \frac{1}{2}\sum_{u \in V} \sum_{v \in N(u)} x_u\cdot x_v \cdot \min \{w(u),w(v)\}\\
& \leq &\frac{1}{2}\sum_{u \in V} x_u \cdot w(u) \cdot \sum_{v \in N(u)} x_v\\
& \stackrel{\eqref{eq:fractionISLP}}{\leq} & \frac{1}{2}\sum_{u \in V} x_u \cdot w(u) = \frac{1}{2}\ \mathbf{u}(\vec{x}),
\end{eqnarray*}
The claim of the lemma now directly follows from \cref{lemma:basicISalgorithm}.
\end{proof}
Note that in \cref{lemma:LPapproxIS}, the constant $4$ could be replaced by any constant larger than $2$. In order to obtain an independent set of weight $(1-\varepsilon)S^*(w)$, we however have to combine the lemma with some additional ideas. To this end, we adapt a technique that has been used in \cite{KawarabayashiKS20} (and which is based on the local-ratio technique described in \cite{localratio}). We first describe and analyze an abstract iterative process to compute an independent set. Consider a sequence of independent sets $I_1, I_2, \dots, I_T$ and node weight functions $w_1, w_2, \dots, w_{T+1}$ and $w_1',w_2'\dots,w_T'$, which are constructed as follows. The weight function $w_1$ is equal to the original weight function $w$ (i.e., $w_1:=w$) and $I_1$ is an arbitrary independent set of $G$. For each $i\in \set{1,\dots,T}$, $w_{i+1}$ and $w_{i}'$ are defined as follows:
\[
\forall u\in V\,:\, w_{i+1}(u) = \max\set{0, w_{i}(u) - \sum_{v\in N^+(u)\cap I_{i}} w_i(v)}
\quad\text{and}\quad
w_{i}'(u) := w_{i}(u) - w_{i+1}(u).
\]
Further, for $i>1$, $I_i$ is an arbitrary independent set of the subgraph of $G$ induced by the nodes $v$ with $w_i(v)>0$. Hence, the process runs in phases. In phase $i$, we start with edge weights $w_i$ and we determine some independent set $I_i$ of the nodes of $G$ with positive weight. We then adjust the weights as follows. For each node in $I_i$, we set the weight $w_{i+1}$ to $0$ and for each node that has neighbors in $I_i$, we deduce the sum of the weights of the neighboring nodes in $I_i$ (or we set the weight to $0$ in case it would become negative otherwise). The weights $w_i'$ are just defined as the amount by which the weights are reduced when going from weights $w_i$ to weights $w_{i+1}$. Because for nodes $v\in I_i$, we set $w_{i+1}(v)=0$ and for such nodes, we therefore have $w_i(v)=w_i'(v)$. Also note that since the weights $w_{i+1}$ of nodes in independent set $I_i$ are set to $0$, the independent sets $I_i$ for different $i$ are disjoint.
We now define a combined independent set $I$ as follows. We start with $I=\emptyset$ and go through the sets $I_1, \dots, I_T$ in reverse order (i.e., starting with $I_T$). When considering set $I_j$, we add all nodes of $I_j$ to $I$, which do not already have a neighbor in $I$. That is, more formally, we consider $j=T,T-1,\dots,1$ and in step $j$, we first define $I_j' := I_j \setminus N^+(I)$ and we then set $I:=I\cup I_j'$. The following lemma shows that the total weight $w(I)$ of the independent set $I$ can be lower bounded by the sum of the weights of the independent sets $I_i$ w.r.t.\ the weight function $w_i$ that is used when computing $I_i$.
\begin{lemma}\label{lemma:totalindsetsize}
\[
w(I) \geq \sum_{i=1}^T w_i(I_i) = \sum_{i=1}^T w_i'(I_i).
\]
\end{lemma}
\begin{proof}
Let $\bar{I}:= I_1\cup\dots\cup I_T$. Note that we have $I\subseteq \bar{I}$. For each node $u\in \bar{I}$, we define $i_u$ as the index $i\in \set{1,\dots,T}$ for which $v\in I_i$. We further define a function $\nu:\bar{I}\to I$ as follows. For $u\in I$, we set $\nu(u)=u$. Now, consider a node $u\in\bar{I}\setminus I$. The fact that $u$ was not added to $I$ implies that there is at least one neighbor $v\in N(u)$ such that $v$ was added to $I$ before considering set $I_{i_u}$, i.e., $i_v>i_u$. If there are several such nodes, we pick $v$ arbitrarily among them and we set $\nu(u)=v$. For each node $v\in I$, we further define $\nu^{-1}(v):=\set{u\in \bar{I} : \nu(u)=v}$. We next show that
\begin{equation}\label{eq:indsetcharging}
\forall v\in I\,:\,w(v)\geq \sum_{u\in \nu^{-1}(v)} w_{i_u}(u)
\end{equation}
To see \eqref{eq:indsetcharging}, recall the above process. Also recall that $i_v>i_u$ for all $u\in \nu^{-1}(v)$. Whenever a neighbor $u$ of $v$ is added to independent set $I_{j}$ for some $j<i_u$, the weight of $v$ is reduced by $w_{i_u}(u)$. Note that since $v\in I_{i_v}$, we have $w_{i_v}(v)>0$ and thus when $u$ is added to $I_{j}$, the full weight $w_{j}(u)$ is reduced from the weight of $v$. We therefore have
\[
w_{i_u}(v) = w(v) - \sum_{j=1}^{i_u-1} w_j(I_j\cap N(v))
\]
\cref{eq:indsetcharging} now follows because $\nu^{-1}(v)\subseteq \set{v}\cup\bigcup_{j<i_u} (I_j\cap N(v))$. From \cref{eq:indsetcharging}, we directly get the claim of the lemma by summing over all nodes in $I$. For the second part of the claim, note that for every $v\in I_i$, we have $w_i'(v)=w_i(v)$.
\end{proof}
Recall that $w_i'$ is the difference between $w_i$ and $w_{i+1}$, i.e., it is the amount by which the weights $w_i$ are decreased after adding independent set $I_i$ to the sequence. The next technical lemma shows that w.r.t.\ the weights $w_i'$, independent set $I_i$ achieves the bound given by the LP \eqref{eq:fractionISLP}.
\begin{lemma}\label{lemma:piecewiseindset}
\[
\forall i \in \set{1,\dots,T}, w_i(I_i) = w_i'(I_i) \geq S^*(w'_i).
\]
\end{lemma}
\begin{proof}
We prove this by considering the dual LP of LP \eqref{eq:fractionISLP} on graph $G$ with weight function $w'_i$:
\begin{equation}\label{eq:dualLP}
\min \sum_{v\in V} y_v\qquad\text{s.t.}\quad
\forall v\in V\, :\!\!\! \sum_{u\in N^+(v)}\!\!\! y_u \geq w'_i(v)\quad\text{and}\quad
y_v \geq 0.
\end{equation}
Note that by the strong duality theorem for linear programs, the optimal value of the dual LP \eqref{eq:dualLP} is equal to the optimal value of LP \eqref{eq:fractionISLP} and thus equal to $S^*(w_i')$. We next show that a feasible solution to \eqref{eq:dualLP} is achieved by setting $y_v= w'_i(v)$ if $v \in I_i$, $y_v=0$ otherwise. We need to show that for all $v\in V$, $\sum_{u\in N^+(v)} y_u \geq w_i'(v)$. This is clearly true for $v\in I_i$. It is also true if $v\not\in I_i$ and $v$ has no neighbor in $I_i$ because in this case $w_i'(v)=0$. Let us therefore assume that $N(v)\cap I_i\neq \emptyset$. Note that in this case, $w_i'(v) = \min\big\{w_i(v), \sum_{u\in N(v)\cap I_i} w_i(u) = \sum_{u\in N(v)\cap I_i} w_i'(u)\big\}$. We therefore have $\sum_{u\in N^+(v)} y_u \geq w_i'(v)$ and thus $y_v=w_i'(v)$ gives a feasible solution for \eqref{eq:dualLP}. The objective value of this feasible solution is exactly $w_i'(I_i)=w_i(I_i)$, which is therefore an upper bound on the optimal value $S^*(w_i')$ of LP \eqref{eq:dualLP}.
\end{proof}
The following theorem shows that if each independent set $I_i$ is chosen such that $w_i(I_i)=\Omega(S^*(w_i))$, then after computing a sequence of $O(\log(1/\varepsilon))$ independent sets, the resulting independent set $I$ has weight at least $(1-\varepsilon)\cdot S^*(w)$.
\begin{lemma}\label{lemma:boostingalgo}
Let $G=(V,E)$ be a graph with node weights $w:V\to\mathbb{N}$ s.t.\ $w(v)\leq W$ for all $v\in V$. Assume that in the above iterative process, for every $i\geq 1$, independent set $I_i$ is chosen such that $w_i(I_i)\geq\rho\cdot S^*(w_i)$ for some given $\rho\in(0,1)$. Then after computing independent sets $I_1,\dots,I_T$ for $T\geq \ln(1/\varepsilon)/\rho$, the resulting independent set $I$ has a total weight of $w(I)\geq (1-\varepsilon)S^*(w)$.
\end{lemma}
\begin{proof}
Let $I^{(t)}$ be the independent set that we obtain when running the above process for $t\leq T$ steps. That is, $I^{(t)}$ is the independent set that we get when greedily going through the independent sets $I_t, I_{t-1}, \dots, I_1$. Note that by \Cref{lemma:totalindsetsize}, we have $w(I^{(t)})\geq \sum_{i=1}^t w_i(I_i)$. For $t\geq 0$, we define a potential $\Upsilon_t$ as
\[
\Upsilon_t := S^*(w) - \sum_{i=1}^t w_i(I_i)
\]
to measure the progress towards obtaining an independent set of weight $w(I)$ approaching $S^*(w)$. For all $t\geq 0$, we inductively prove the following properties:
\[
\text{(I)}\,:\, \Upsilon_t \leq S^*(w_{t+1})\quad\text{and}\quad
\text{(II)}\,:\, \Upsilon_t \leq (1-\rho)^{t}\cdot \Upsilon_0.
\]
Note that both statements are clearly true for $t=0$ (recall that for all $v$, $w_1(v)=w(v)$). For the induction step, note first that by assumption, we have that $w_t(I_t)\geq \rho S^*(w_t)$ and thus by the induction hypothesis (property (I) for $t-1$), $w_t(I_t)\geq \rho \Upsilon_{t-1}$. Since we have $\Upsilon_t = \max\set{0, \Upsilon_{t-1}-w_t(I_t)}$, this implies that $\Upsilon_t \leq \Upsilon_{t-1} - \rho\Upsilon_{t-1}$, which together with the induction hypothesis (property (II) for $t-1$) proves the induction step for property (II). To do the induction step for property (I), first consider three weight functions $w_a$, $w_b$, and $w_{ab}$ such that for all $v\in V$, $w_{ab}(v)= w_a(v) + w_b(v)$. Clearly, for any feasible solution for LP \eqref{eq:fractionISLP}, the objective value for weights $w_{ab}$ is equal to the sum of the objective values for weights $w_a$ and $w_b$. For the optimal objective values w.r.t.\ the three weight functions, we therefore have $S^*(w_a)+S^*(w_b)\geq S^*(w_{ab})$. Now recall that we have $w_t'=w_t - w_{t+1}$ and we therefore have $S^*(w_{t+1}) + S^*(w_t')\geq S^*(w_t)$. From \Cref{lemma:piecewiseindset}, we know that $S^*(w_t')\leq w_t(I_t)$. We therefore have
\begin{eqnarray*}
S^*(w_{t+1}) & \geq & S^*(w_t) - S^*(w_t')\ \geq\ S^*(w_t) - w_t(I_t)\\
& \geq & S^*(w_t) - (\Upsilon_{t-1} - \Upsilon_t)
\ \geq\ \Upsilon_{t-1} - (\Upsilon_{t-1}-\Upsilon_t) = \Upsilon_t.
\end{eqnarray*}
The last inequality follows from the induction hypothesis (property (I) for $t-1$). This also completes the induction step for property (I) for $t$.
By \Cref{lemma:totalindsetsize} and the definition of $\Upsilon_T$, we have $w(I)\geq \Upsilon_0 - \Upsilon_T$. Note that we have $T\geq \ln(1/\varepsilon)/\rho\geq \ln(1/\varepsilon)/\ln(1/(1-\rho))$. This holds because $\ln(1/(1-\rho)) = \sum_{i=1}^\infty \rho^i/i!\geq \rho$. By property (II) for $t=T$, we therefore have
\[
\Upsilon_T\leq (1-\rho)^T\cdot\Upsilon_0 = (1-\rho)^{\ln(1/\varepsilon)/\ln(1/(1-\rho))}\cdot\Upsilon_0 = \varepsilon\cdot\Upsilon_0.
\]
We therefore have $w(I) \geq (1-\varepsilon)\cdot\Upsilon_0 = (1-\varepsilon)\cdot S^*(w)$ as required.
\end{proof}
\paragraph{Proof of \Cref{thm:ISthm}, Part 1:} The following lemma proves result \eqref{eq:ISthm1} of \Cref{thm:ISthm}.
\begin{lemma}\label{lemma:ISbetaapprox}
Let $G=(V,E)$ be a graph with maximum degree $\Delta$, node weight function $w:V\to \mathbb{R}_{\geq 0}$, neighborhood independence $\beta$, and assume that $G$ is equipped with a proper $\xi$-vertex coloring. For any $\varepsilon>0$, there is deterministic $\mathsf{CONGEST}$\xspace algorithm that computes a $(1-\varepsilon)/\beta$-approximation to the maximum weight independent set problem on $G$. The round complexity of the algorithm is $O\big(\log^2(\Delta W) \cdot \log(1/\varepsilon)+\log^* \xi\big)$.
\end{lemma}
\begin{proof}
At the start, we compute a proper $O(\Delta^2)$-vertex coloring of $G$. This can be done in $O(\log^*\xi)$ deterministic rounds in the $\mathsf{CONGEST}$\xspace model by using a classic algorithm of \cite{linial92}. We then use the above framework, where in each step $i$, we use the algorithm of \Cref{lemma:LPapproxIS} to compute an independent set $I_i$ of weight $w_i(I_i)\geq S^*(w_i)/4$. By \Cref{lemma:LPapproxIS} and by using the initial $O(\Delta^2)$-coloring, the round complexity for computing each such independent set $I_i$ is $O\big(\log^2(\Delta W)\big)$. By \Cref{lemma:boostingalgo}, we have to do $O(\log(1/\varepsilon))$ steps to obtain an independent set of weight $(1-\varepsilon)S^*(w)$. To see that this is a $(1-\varepsilon)/\beta$-approximation, note that in a graph of neighborhood independence $\beta$, clearly the neighborhood $N^+(v)$ of every node $v$ can contain at most $\beta$ nodes of any independent set. Hence, if we replace the right-hand sides of all constraints in LP \eqref{eq:fractionISLP} by $\beta$, the LP is a relaxation of the maximum weight independent set problem (any independent set satisfies the constraints). Clearly, increasing the right-hand sides of all constraints to $\beta$ increases the optimal objective value by exactly a factor $\beta$. Consequently, the weight of an optimal weighted independent set in a graph of neighborhood independence $\beta$ is at most $\beta\cdot S^*(w)$. This completes the proof.
\end{proof}
\paragraph{Proof of \Cref{thm:ISthm}, Part 2:} Result \eqref{eq:ISthm2} of \Cref{thm:ISthm} can be proven in a similar way. The following lemma proves \eqref{eq:ISthm2} by adapting the inductive argument in \Cref{lemma:boostingalgo}.
\begin{lemma}\label{lemma:ISresult2}
Let $G=(V,E)$ be a graph with maximum degree $\Delta$ and node weight function $w:V\to \mathbb{R}_{\geq 0}$, and assume that $G$ is equipped with a proper $\xi$-vertex coloring. Then, for any $\varepsilon>0$, there is deterministic $\mathsf{CONGEST}$\xspace algorithm that computes an independent set of $G$ of total weight at least $(1-\varepsilon)\cdot\frac{w(V)}{\Delta+1}$ in $O(\log^2\Delta\cdot\log(1/\varepsilon) + \log^*\xi)$ rounds.
\end{lemma}
\begin{proof}
As a initial step, we compute a proper $O(\Delta^2)$-vertex coloring of $G$ in time $O(\log^*\xi)$ by using an algorithm of \cite{linial92}. Now, we again use the method of iteratively computing independent sets $I_1,\dots,I_T$ as above. First note that for a given weight function $w:V\to\mathbb{N}$, we can efficiently compute an independent set of weight $w(V)/(4(\Delta+1))$ in $O(\log^2\Delta +\log^*(\Delta^2))=O(\log^2\Delta)$ rounds by using \Cref{lemma:LPapproxIS}. To see this, observe that setting the fractional value of all nodes $v$ to $x_v=1/(\Delta+1)$ gives a feasible solution to LP \eqref{eq:fractionISLP} with objective value $w(V)/(\Delta+1)$. We therefore in particular have $S^*(w)\geq w(V)/(\Delta+1)$ and for $W\leq\operatorname{\text{{\rm poly}}}(\Delta)$, the claim of the lemma directly follows from \Cref{lemma:boostingalgo}. In the following, we show that by using a similar argument as in the proof of \Cref{lemma:boostingalgo}, we can also get a round complexity of $O(\log^2\Delta + \log^* \xi)$.
As discussed, we use the iterative framework from above. In each iteration $t$, we choose the independent set $I_t$ such that $w_t(I_t)\geq w_t(V)/(4(\Delta+1))$. As discussed above, we can find such an independent set $I_t$ in $O(\log^2\Delta)$ rounds. We again define a potential $\Upsilon_t$ that measures how close we are to achieving our goal:
\[
\forall t\geq 0\,:\, \Upsilon_t := \frac{w(V)}{\Delta+1} - \sum_{i=1}^t w_i(I_i).
\]
For all $t\geq 0$, we inductively prove the following two properties:
\[
\text{(i)}\,:\, \Upsilon_t \leq \frac{w_{t+1}(V)}{\Delta+1}\quad\text{and}\quad
\text{(ii)}\,:\, \Upsilon_t \leq \left(1-\frac{1}{4}\right)^{t}\cdot \Upsilon_0.
\]
Both statements are clearly true for $t=0$. Let us therefore consider the induction step from $t-1$ to $t$. We have
\[
\Upsilon_t = \Upsilon_{t-1} - w_t(I_t) \leq
\Upsilon_{t-1} - \frac{w_t(V)}{4(\Delta+1)} \leq
\Upsilon_{t-1}\cdot\left(1-\frac{1}{4}\right).
\]
The first inequality follows from $w_t(I_t)\geq w_t(V)/(4(\Delta+1))$ and the second inequality follows from the induction hypothesis for $t-1$ (part (i)). Together with the induction hypothesis (part (ii)), the above inequality directly implies statement (ii) for $t$. To also prove statement (i) for $t$, recall that $w_t'(V) = w_t(V) - w_{t+1}(V)$. Also note that we have $w_t'(V) \leq (\Delta+1)\cdot w_t(I_t)$ because for every node $v\in I_t$, the weight $w_t(v)$ is deducted (and thus added to $w_t'$) for at most $\Delta+1$ nodes. We therefore have $w_t(I_t) \geq \frac{w_t(V)}{\Delta+1}-\frac{w_{t+1}(V)}{\Delta+1}$. We thus obtain
\[
\Upsilon_t = \Upsilon_{t-1} - w_t(I_t) \leq
\Upsilon_{t-1} + \left(\frac{w_{t+1}(V)}{\Delta+1}-\frac{w_{t}(V)}{\Delta+1}\right) \leq
\frac{w_{t+1}(V)}{\Delta+1}.
\]
The last step follows from the induction hypothesis (part (i)). For $T=\log_{4/3}(1/\varepsilon)$, statement (ii) implies $\Upsilon_T\leq \varepsilon \Upsilon_0 =\varepsilon\cdot\frac{w(V)}{\Delta+1}$ and \Cref{lemma:totalindsetsize} then gives $w(I)\geq (1-\varepsilon)\cdot \frac{w(V)}{\Delta+1}$ as required. The overall time complexity is $O(\log^*\xi)$ for the initial $O(\Delta^2)$-coloring and $O(\log^2\Delta\cdot\log(1/\varepsilon))$ for computing independent sets $I_1,\dots,I_T$ and then independent set $I$.
\end{proof}
We conclude \Cref{sec:boundedindep} by showing that our method yields an alternative to Fischer's $O(\log^2\Delta\cdot\log n)$-round deterministic $\mathsf{CONGEST}$\xspace algorithm~\cite{fischer2020improved} for computing a maximal matching of a graph. This show that all the four classic symmetry breaking problems (MIS, maximal matching, $(\Delta+1)$-vertex coloring, and $(2\Delta-1)$-edge coloring) can be solved by the same general method and in all cases, the method yields the best current distributed algorithms.
\begin{theorem}\label{thm:maximalmatching}
A maximal matching in an $n$-node graph $G=(V,E)$ of maximum degree $\Delta$ can be computed deterministically in $O(\log^2\Delta\cdot\log n)$ rounds in the $\mathsf{CONGEST}$\xspace model.
\end{theorem}
\begin{proof}[Proof Sketch]
We start by computing an $O(\Delta^2)$-edge coloring. A simple way to do this in $O(\log^* n)$ rounds in the $\mathsf{CONGEST}$\xspace model is the following (also see \cite{Kuhn2009WeakColoring}). Each node $v$ first uniquely labels its $\deg(v)$ edges with the number $1,\dots,\deg(v)$ in an arbitrary way. In this way, each edge obtains two numbers in $[\Delta]$, and we thus have a coloring of the edges with $O(\Delta^2)$ colors. The coloring is however not proper. For each of the colors, each node can however have at most $2$ incident edges. Therefore, the color classes induce paths and cycles. Those paths and cycles can be properly edge-colored with $3$ colors in $O(\log^* n)$ rounds by using a standard algorithm~\cite{cole86,linial92}.
Given an $O(\Delta^2)$-edge coloring, we next show that a constant factor approximation for the maximum cardinality matching problem can be computed in $O(\log^2\Delta)$ rounds. One can then obtain a maximal matching by repeating $O(\log n)$ times. The matching approximation is computed by using the rounding approach for graphs with bounded neighborhood independence.
As a first step, we need to be able to compute a constant factor approximate feasible solution for LP \eqref{eq:fractionISLP} on the line graph of $G$ (and with all weights being equal to $1$). This can be done by slightly adapting a standard fractional matching approximation algorithm (see, e.g., \cite{fischer2020improved}). We first assign a fractional value $y_e=1/(2\Delta)$ to each edge. We call a node saturated if the sum of the fractional values of its edges is at least $1/4$ and we call an edge frozen if at least one of its nodes is saturated. We now proceed in $O(\log \Delta)$ phases, where in each phase, we double the fractional value of all non-frozen edges. At the end all edges are frozen and it is not hard to see that this gives a $1/4$ approximation for LP \eqref{eq:fractionISLP} (multiplying each fractional edge value with $4$ gives a feasible solution of the dual LP and in the unweighted case, LP \eqref{eq:fractionISLP} and its dual LP both have the same objective function).
We now just have to show that the rounding of this fractional solution to an integral solution can be done in $O(\log^2\Delta)$ rounds. For this, observe that we can use the algorithm for rounding in d2-multigraphs (\Cref{def:d2multigraph}): An edge of the line graph of $G$ is defined by a pair of edges that share a common node and we can thus use this node as the managing node of the edge. We therefore obtain at least the same round complexity as for d2-multigraphs. For the case of rounding independent sets, we have $|\Sigma|=2$ and as a consequence, the only expensive step is the computation of the weighted relative average defective coloring. Once the defective coloring is given, doing one rounding step for the edges of a single color can be done in $O(1)$ $\mathsf{CONGEST}$\xspace rounds (cf.\ the proof of \Cref{lemma:d2rounding}). The computation of the defective coloring in \Cref{lemma:d2defective} consists of two parts. First, one computes a weighted $\delta$-relative defective coloring with $O(1/\delta^2)$ colors and one then reduces the number of colors to $O(1/\delta)$ by adapting the algorithm of \cite{BEG18}. The second step requires only $O(1/\delta)$ rounds in d2-multigraphs (cf.\ the proof of \Cref{lemma:d2defective}) and only the first step is expensive and requires the additional $\log\log \Delta$ factor (note that in our case $q=O(\Delta)$). However, on line graphs, a weighted $\delta$-relative defective coloring can be computed fast by using a similar approach as for computing a proper $O(\Delta^2)$-edge coloring at the beginning of the algorithm.
We proceed as follows. Each node $v$ first marks all its edges of weight more than a $\delta/4$-fraction of the total weight of all of $v$'edges. As every node has at most $4/\delta$ such edges, those edges can be properly colored with $O(1/\delta^2)$ colors by using the same algorithm as above. An edge might only be marked from one of its nodes, however, the algorithm also works if an edge only gets a number in $[\Delta]$ from one side. We now proceed with all the edges that are not yet colored and we use a new set of $O(1/\delta^2)$ colors for those. Each node $v$ now labels its edges with numbers between $1$ and $4/\delta$ in such a way that each label is given to edges of total weight at most a $\delta/2$-fraction of the initial total weight of all of $v$'s edges. This is possible by greedily assigning the labels. Now for each combination of two labels, at most at $\delta$-fraction of the total weight of $v$'s edges have this combination of labels (because $v$ must assign one of the two labels). The algorithm can clearly be implemented in $O(\log^*\Delta)$ rounds in the $\mathsf{CONGEST}$\xspace model (with a given initial proper $O(\Delta^2)$-edge coloring.
\end{proof}
\subsection{Approximating a Generalized Caro-Wei-Tur\'{a}n Bound}
\label{sec:turan}
It is well-known that there exist independent sets of size at least $\sum_{v\in V}\frac{1}{\deg(v)+1}$ (sometimes known as the Caro-Wei bound~\cite{Wei1981generalizedturan,griggs1983lower}). This can be lower bounded by $n/\deg_{\mathrm{avg}}$, which is known as the Tur\'{a}n bound~\cite{turan1941external}.
Moreover, the Caro-Wei independent set can be computed sequentially in polynomial time by a simple greedy algorithm (see, e.g., \cite{halldorsson1997greed}) and a simple randomized one-round algorithm (one round of the classic Luby algorithm~\cite{luby86}) achieves an independent set of this size in expectation.
For node-weighted graphs, a simple generalization of this greedy algorithm~\cite{kako2005approximation} gives an independent set of weight $\sum_{v\in V}\frac{(w(v))^2}{w(N^+(v))} \geq \frac{w(V)}{\mathrm{wdeg}_{\mathrm{avg}}+1}$, where $\mathrm{wdeg}_{\mathrm{avg}}= \frac{ \sum_{v\in V} w(N(v))}{w(V)}=\frac{\sum_{v \in V} w(v)\deg(v)}{w(V)}$.\footnote{The inequality $\sum_{v\in V}\frac{(w(v))^2}{w(N^+(v))} \geq \frac{w(V)}{\mathrm{wdeg}_{\mathrm{avg}}+1}$ can be obtained by a simple application of the Cauchy-Schwarz inequality $ (\sum_i x^2_i)( \sum_i y^2_i) \geq (\sum_i x_i y_i )^2$ by assigning $x_v=\sqrt{w(N^+(v))}$ and $y_v= \frac{w(v)} { \sqrt{w(N^+(v))} }$.} We next show that there is an efficient deterministic $\mathsf{CONGEST}$\xspace algorithm to get within an $\big(\frac{1}{2}-\varepsilon\big)$-factor of those bounds.
\paragraph{Proof of \Cref{thm:ISthm}, Part 3:} The following lemma proves \eqref{eq:ISthm3}, the third claim of \Cref{thm:ISthm}.
\begin{lemma}
Let $G=(V,E)$ be a graph with node weights $w:V\to \mathbb{N}$ and let $\vec{x}$ be a vector of fractional node values $x_v= \frac{w(v)}{W_v}$, where $W_v:=w(N^+(v))$. Then, for any $\varepsilon>0$, we can deterministically compute an independent set of total weight $\big(\frac{1}{2}- \varepsilon\big)\cdot \sum_{v\in V}\frac{(w(v))^2}{W_v}$ in time $O\big(\frac{\log^2(\Delta/\varepsilon)}{\varepsilon} + \log^* n\big)$ in the $\mathsf{CONGEST}$\xspace model.
\end{lemma}
\begin{proof}
Recall that $\mathbf{u}(\vec{x}) = \sum_{v\in V} w(v)\cdot x_v$ and $\mathbf{c}(\vec{x})=\sum_{\set{u,v}\in E} \min\set{w(u), w(v)}\cdot x_u\cdot x_v$. We first show that for the fractional values $x_v$ as given, we have $\mathbf{u}(\vec{x})\geq 2\mathbf{c}(\vec{x})$. For this, we first give an upper bound on the cost $\mathbf{c}(e,\vec{x})=\min\set{w(u), w(v)}\cdot x_u\cdot x_v$ of a single edges $e=\set{u,v}$. We have
\begin{eqnarray}
\mathbf{c}(e,\vec{x}) & = & w(u) \cdot w(v)\cdot\min\set{w(u),w(v)}\cdot
\frac{1}{W_u\cdot W_v}\nonumber\\
& \leq &
w(u)\cdot w(v)\cdot \min\set{w(u),w(v)}\cdot \frac{1}{2}\cdot \left[
\frac{1}{W_u^2} + \frac{1}{W_v^2}\right]\label{eq:123}\\
& \leq &
\frac{1}{2}\cdot\left[
\frac{(w(u))^2\cdot w(v)}{W_u^2} + \frac{(w(v))^2\cdot w(u)}{W_v^2}
\right].\label{eq:edgecostupper}
\end{eqnarray}
Inequality \eqref{eq:123} follows because $\big(\frac{1}{W_u}-\frac{1}{W_v}\big)^2=\frac{1}{W_u^2} + \frac{1}{W_v^2} -
\frac{2}{W_u\cdot W_v}\geq 0$. We can now use this to obtain an upper bound on the total cost $\mathbf{c}(\vec{x})$:
\begin{eqnarray*}
\mathbf{c}(\vec{x}) & = & \sum_{e=\set{u,v}\in E} \mathbf{c}(e,\vec{x})\\
& \stackrel{\eqref{eq:edgecostupper}}{\leq} &
\sum_{\set{u,v}\in E} \frac{1}{2}\cdot\left[
\frac{(w(u))^2\cdot w(v)}{W_u^2} + \frac{(w(v))^2\cdot w(u)}{W_v^2}
\right]\\
& = &
\frac{1}{2}\cdot\sum_{u\in V}\sum_{v\in N(u)} \frac{(w(u))^2\cdot w(v)}{W_u^2}\\
& = & \frac{1}{2}\cdot\underbrace{\sum_{u\in V}\frac{(w(u))^2}{W_u}}_{=\mathbf{u}(\vec{x})}\cdot
\underbrace{\sum_{v\in N(u)}\frac{w(v)}{W_u}}_{\leq 1}\ \leq\ \frac{1}{2}\cdot\mathbf{u}(\vec{x}).
\end{eqnarray*}
The claim of the lemma now follows directly from \Cref{lemma:basicISalgorithm}.
\end{proof}
|
\section{Introduction}
In the conventional interpretation of quantum gravity \cite{Kuchar,Isham} without the holographic principle \cite{Hol1,Hol2,Hol3}, the quantum state is that of the whole Universe.
A typical application of the Born rule in this interpretation is seen in the inflationary multiverse scenario \cite{CC1,CC2,Vilenkin}.
Taking a different approach, the author has proposed a novel interpretation of quantum gravity \cite{EPL1,EPL2} based on the holographic principle \cite{Hol1,Hol2,Hol3} in the context of the three-dimensional anti-de Sitter spacetime/two-dimensional conformal field theory (AdS$_3$/CFT$_2$) correspondence \cite{AdSCFT1,AdSCFT2,AdSCFT3,AdSCFT4} at the strong-coupling limit of the boundary CFT$_2$ \cite{RT1,RT2,HRT,RT3,Swingle,Matsueda,Review1,Review2,Review3}.
In this interpretation of quantum gravity, non-selective quantum measurement \cite{dEspagnat} of the ground state or a purified quantum thermal equilibrium state of space, that is, a holographic tensor network (HTN) \cite{Swingle,Matsueda,Review1}, is done in the ensemble interpretation of quantum mechanics by decohering quantum coherence in this quantum state completely.
The decoherence (i.e., loss of quantum interference with respect to the observables) is exactly done by introducing a superselection rule operator and then restricting the set of observables acting on the Hilbert space of the HTN to the Abelian set whose elements commute with the superselection rule operator \cite{JHAP1}.
The author refers to this decoherence as {\it classicalization}.
The classicalization of quantum gravity is {\it not} classical gravity; indeed, the classicalized state of the HTN is still a quantum state but a highly non-trivial mixed state.
Since this quantum state is a statistical mixture of the product quantum eigenstates, there are {\it negative} local degrees of freedom \cite{EPL2,JHAP1}.
So far, we have classicalized space in the Euclidean regime of the HTN, that is, the purified quantum thermal equilibrium states of the boundary CFT$_2$ including the ground state \cite{EPL1,EPL2,JHAP1,JHAP2}.
Then, to formulate the time-dependent HTN in the Lorentzian regime, how do we incorporate real time $t$ into this interpretation of quantum gravity?
The answer proposed in this article is to {\it classicalize real time, too} \cite{JPCO}.
Namely, we treat real time as a classical observable {\it $\grave{a}$ la} von Neumann \cite{dEspagnat,Neumann} in the exact sense.
Here, we treat real time continuously.
Then, the Hilbert space of the HTN can be decomposed into a direct integral of the continuous coherent subspaces with an absolutely continuous temporal measure $d\mu(t)$ of the density matrix \cite{AnnMath}.
In this article, we propose a generally covariant Lorentzian action of the classicalized holographic tensor network (cHTN) as the holographic reduction of the Einstein--Hilbert action of gravity in three spacetime dimensions
\begin{equation}
I_{\rm EH}[g_{\mu\nu}]=\frac{1}{16 \pi G_N}\int (R-2\Lambda)\sqrt{-g}d^2x dt\label{eq:EH}
\end{equation}
in the presence of a negative cosmological constant $\Lambda$.
Here, we choose $(-,+,+)$ as the signature of the Lorentzian spacetime metric $g_{\mu\nu}$, and $G_N$ and $R$ denote the three-dimensional Newtonian gravitational constant and the scalar curvature of the Lorentzian spacetime metric $g_{\mu\nu}$, respectively.
Our Lorentzian action of the cHTN is defined for a generic quantum pure state $|\psi(t)\ra$ of the boundary CFT$_2$ by
\begin{equation}
I_L[|\psi\rra_L]=-\hbar H[|\psi\rra_L]\;,\label{eq:H}
\end{equation}
where we define the quantum state of the HTN in the representation of the Lorentzian boundary conformal symmetry
\begin{equation}
|\psi\rra_L\equiv\int^\bigoplus|\psi(t)\ra \sqrt{d\mu(t)}\;.\label{eq:rra}
\end{equation}
$H[|\psi\rra_L]$ is the von Neumann entropy (here, the measurement entropy) of the classicalized state of $|\psi\rra_L$ in nats.
In addition to the Euclidean action \cite{EPL2}, this Lorentzian action accords with the holographic principle \cite{Hol1,Hol2,Hol3} and asserts that the negative number of the local {\it spin} degrees of freedom in the bulk spacetime is given by the amount of boundary CFT$_2$ state information in nats that is lost by the classicalization \cite{EPL2}.
Note that, if $|\psi\rra_L$ is the ground state, it is time independent (i.e., a pure state with respect to real time), and thus $|\psi\rra_L {}_L\lla \psi|=|\psi\ra\la \psi|\otimes \widehat{1}$ and $H[|\psi\rra_L]=H[|\psi\ra]$ hold.
In this case, our Lorentzian action (\ref{eq:H}) of the cHTN is reduced to the Euclidean action of the cHTN
\begin{equation}
I_E[|\psi\ra]=-\hbar H[|\psi\ra]\;,\label{eq:H20}
\end{equation}
which was used by the author in Refs. \cite{EPL2,JHAP1,JHAP2}.
In the Euclidean action (\ref{eq:H20}) of the cHTN, $|\psi\ra$ is the ground state or a purified quantum thermal equilibrium state of the HTN \cite{EPL2}.
Subject to a given average energy, the HTN in the Euclidean regime is the most probable statistical mixture with respect to {\it energy}, and the HTN in the Lorentzian regime is now the most probable statistical mixture with respect to {\it real time}.
Here, in the HTN, there is a quantum uncertainty relation between energy and real time because energy and real time are conjugate to each other.
However, after the classicalization of the HTN, there is no quantum uncertainty relation between them.
So, in the presence or absence of matter, the cHTN in the Lorentzian regime is the most probable statistical mixture with respect to energy {\it and} real time {\it simultaneously}.
Based on this fact, we introduce the imaginary time $\tau \equiv it$ and extend the Euclidean action of the cHTN from Eq. (\ref{eq:H20}) to
\begin{equation}
I_E[|\psi\rra_E]=-\hbar H[|\psi\rra_E]\;,\label{eq:H2}
\end{equation}
where we define the quantum state of the HTN in the representation of the Euclidean boundary conformal symmetry
\begin{equation}
|\psi\rra_E\equiv \int^\bigoplus |\psi(\tau)\ra \sqrt{d\mu (\tau)}\label{eq:rra2}
\end{equation}
for an absolutely continuous temporal measure $d\mu(\tau)$ of the density matrix, and $H[|\psi\rra_E]$ is the measurement entropy of the classicalized state of $|\psi\rra_E$ in nats.
Here, this imaginary time $\tau$ is the real-valued time coordinate in the Euclidean spacetime and is distinguished from the inverse temperature of the quantum thermal equilibrium states of the HTN, which is the Lagrange multiplier for the fixed average energy of the HTN, except for the identification of the period of the imaginary time with the inverse temperature \cite{GH,Harlow}.\footnote{For the issue of allowing complex-valued spacetime metrics in Euclidean quantum gravity, see Ref. \cite{Witten} and the references therein.}
Then, in the presence or absence of matter, the cHTN in the Euclidean regime is the most probable statistical mixture with respect to energy {\it and} imaginary time {\it simultaneously}.
From the results in Ref. \cite{JHAP1}, in the cHTN of the HTN in the ground state, the Euclidean regime is more fundamental than the Lorentzian regime because bulk quantum mechanics of a non-relativistic free particle in the Lorentzian regime follows from the bulk classical stochastic process of this particle (i.e., the readout process of local spin events by the classicalized hologram) in the Euclidean regime via the inverse Wick rotation
\begin{equation}
t=-i \tau\;.\label{eq:Wick}
\end{equation}
Here, in the Lorentzian regime, quantum measuring systems with the ability to read out events \cite{EPL3} exist only in the bulk spacetime; in the Euclidean regime, on the other hand, the classicalized hologram on the boundary spacetime is the only quantum measuring system.
In the rest of this article, we study the properties of the Euclidean and Lorentzian actions of the cHTN.
In Sec. II, we investigate the properties of the proposed Lorentzian action (\ref{eq:H}) of the cHTN in the ground state.
In Sec. III, we derive the gravity perturbation induced by a massive particle at rest in the cHTN from the Euclidean action (\ref{eq:H2}) of the cHTN as the Unruh effect \cite{Davies,Unruh1,Unruh2}.
In Sec. IV, we conclude the article by arguing that our holographic formulation of spacetime can be viewed as a non-equilibrium second law subject to general covariance.
\section{Ground-state Properties of the Lorentzian Action}
In this section, we show three properties of the Lorentzian action (\ref{eq:H}) of the cHTN in the ground state:
\begin{enumerate}
\item[(I)] In the ground state, the Lorentzian action (\ref{eq:H}) of the cHTN is the holographic reduction of the Einstein--Hilbert action (\ref{eq:EH}).
\item[(II)] In the ground state, the proposed Lorentzian action (\ref{eq:H}) of the cHTN is generally covariant.
\item[(III)] The Lorentzian AdS$_3$ spacetime metric can be recovered from the ground state of the boundary CFT$_2$.
\end{enumerate}
Here, the {\it ground state} refers to that of the Einstein--Hilbert action (\ref{eq:EH}) and that of the Lorentzian action (\ref{eq:H}) of the cHTN when there are no additional actions (i.e., there is gravity and a negative cosmological constant only).
\subsection{Holographic reduction}
We consider the ground state, which is a pure state with respect to real time, in the Hilbert space of the boundary CFT$_2$.
The quantum entanglement folded in the boundary ground state is unfolded to the multi-scale entanglement renormalization ansatz (MERA) of this state \cite{Vidal1,Vidal2} along the extra spatial dimension in the bulk space \cite{Swingle}.
We unfold the Lorentzian action (\ref{eq:H}) of the cHTN also into the bulk space from the boundary.
Then, the measurement entropy $H[|\psi\rra_L]$ of the cHTN in bits is given by the discretized area of the MERA \cite{EPL1,EPL2}.
Because the pixel of the MERA is given by $R_{\rm AdS}^2$ for the curvature radius $R_{\rm AdS}$ of the AdS$_3$, the on-shell local information density, $\eta$, is given by
\begin{equation}
\eta=-\frac{1}{R_{\rm AdS}^2}\;.\label{eq:eta}
\end{equation}
Here, the number $1$ represents the spatial dimensions of the boundary.
Note that, in the flat spacetime limit, $\eta$ converges to zero.
On the other hand, the on-shell solution, that is, the Lorentzian AdS$_3$ spacetime of the Einstein--Hilbert action (\ref{eq:EH}), is a maximally symmetric spacetime and has a negative constant scalar curvature.
Because its Ricci tensor is $R_{\mu\nu}=(1-3)g_{\mu\nu}/R_{\rm AdS}^2$ \cite{Nastase}, the on-shell local information density of the gravity part of the Einstein--Hilbert action (\ref{eq:EH}) is given by
\begin{equation}
\eta_{\rm EH}=\frac{R}{2}=-\frac{3}{R_{\rm AdS}^2}\;.\label{eq:eta2}
\end{equation}
Here, the number $3$ represents half the number of off-diagonal elements (i.e., the number of plane combinations) in a square matrix of order $3$ (i.e., the spacetime dimensions).
From this and Eq. (\ref{eq:eta}), in the ground state, the Lorentzian action (\ref{eq:H}) of the cHTN is the holographic reduction of the Einstein--Hilbert action (\ref{eq:EH}).
\subsection{General covariance}
First, the inner product ${}_L\lla \psi_2|\psi_1\rra_L$ between two generic states $|\psi_1\rra_L$ and $|\psi_2\rra_L$ of Eq. (\ref{eq:rra}) is invariant under the change of the temporal measure $d\mu(t)$ to another temporal measure $d\nu(t)$ used in the direct integral decomposition to which the same temporal resolution of unity $\widehat{1}$ belongs \cite{AnnMath}.
This means that the unitary equivalence class of the Hilbert space of $|\psi\rra_L$ and thus the measurement entropy $H[|\psi\rra_L]$ are independent from the choice of temporal measure in the direct integral decomposition, and thus the Lorentzian action (\ref{eq:H}) of the cHTN is well-defined for a generic state $|\psi\rra_L$ of Eq. (\ref{eq:rra}).
Next, in the ground state, from Eqs. (\ref{eq:eta}) and (\ref{eq:eta2}), the Lorentzian action (\ref{eq:H}) of the cHTN has one locally independent negative degree of freedom per pixel $R_{\rm AdS}^2$, the same as the Einstein--Hilbert action (\ref{eq:EH}), to gauge the symmetry spatial coordinate transformations on the boundary CFT$_2$ to the spatial diffeomorphisms in the bulk.
In the cHTN, the bulk spatial diffeomorphisms are unitary transformations (i.e., classical gauge transformations) on the diagonal classicalized state of $|\psi(t)\ra$, and these are enhanced to the bulk spacetime diffeomorphisms as unitary transformations on the diagonal classicalized state of $|\psi\rra_L$.
Then, the Lorentzian action (\ref{eq:H}) of the cHTN is invariant under the bulk spacetime diffeomorphisms (i.e., generally covariant) because the von Neumann entropy is invariant under the unitary transformation.
\subsection{Recovery of the Lorentzian spacetime metric}
We recover the Lorentzian AdS$_3$ spacetime metric $g_{\mu\nu}$ from the ground state of the boundary CFT$_2$.\footnote{In this subsection, we set $R_{\rm AdS}=1$.}
In Ref. \cite{RINP}, after averaging with respect to the local spin degree of freedom over the statistical mixture of its two eigenstates at each site of the cHTN, we recovered the spatial metric of a real-time slice of the Lorentzian AdS$_3$ spacetime
\begin{equation}
ds^2|_{dt=0}=\frac{dx^2+dr^2}{r^2}\;,\label{eq:metric1}
\end{equation}
where $x$ and $r$ are the rescaled horizontal and redefined radial coordinates of the sites of the MERA, respectively.
Now, we regard $x$ and $r$ as {\it spatial} coordinates and incorporate real time $t$ into this previous result.
Due to the conformal $SO(2,2)$ symmetry of the Lorentzian boundary CFT$_2$, the Lorentzian bulk spacetime has the $SO(2,2)$ isometry group.
From this spacetime symmetry and Eq. (\ref{eq:metric1}), we obtain the static Lorentzian spacetime metric
\begin{equation}
ds^2=-\frac{f(r)dt^2}{r^2}+ds^2|_{dt=0}\label{eq:metric2}
\end{equation}
for a dimensionless function $f(r)$.
Here, note that the ground state is the thermal equilibrium state at zero temperature and thus has no length-scale variable in natural units.
As a result of this fact and Eq. (\ref{eq:metric2}), we recover the Lorentzian AdS$_3$ spacetime metric
\begin{equation}
ds^2=\frac{-dt^2+dx^2+dr^2}{r^2}\;,\label{eq:metric3}
\end{equation}
where the coordinates $t$ and $x$ in Eq. (\ref{eq:metric3}) on the $r=0$ slice without the conformal factor match those of the boundary spacetime \cite{Nastase}.
\section{Gravity as the Unruh Effect}
In this section, in the Euclidean regime of the HTN, we derive the gravity perturbation that is the Euclidean proper {\it acceleration} induced by a massive particle at rest in the cHTN as the Unruh effect.
We assume a particle with non-zero mass $M$ located at the {\it top tensor} of the cHTN \cite{Vidal2} and study its effect over the cHTN.
After an infinitesimal imaginary proper time interval $d \tau_0$ at the top tensor, this mass $M$ of the particle creates an infinitesimal information reading in nats with an amount
\begin{equation}
dI_0=\frac{d_\tau S_E}{\hbar}\;,\ \ d_\tau S_E=Mc^2 d \tau_0\;,
\end{equation}
where $S_E$ is the relativistic Euclidean action of the particle $M$ and is added to the Euclidean action (\ref{eq:H2}) of the cHTN.
This infinitesimal information reading $dI_0$ at the top tensor would be fine-grained in the cHTN along the inverse renormalization group (RG) direction and equally divided per site at each deeper inverse RG step $n$ counted from the top tensor.
Then, at the inverse RG step $n$, the infinitesimal information reading $dI_0$ is fine-grained to a smaller amount of information reading per site
\begin{equation}
di_{0\to n}=\frac{dI_0}{N_n}\;,
\end{equation}
where $N_n$ is the number of sites in the cHTN at the inverse RG step $n$.
This smaller amount of infinitesimal information reading $di_{0 \to n}$ per site is equivalent to a finite energy per site
\begin{equation}
\varepsilon_n=\hbar \frac{di_{0\to n}}{d \tau_n}
\end{equation}
for the infinitesimal imaginary proper time interval $d \tau_n$ at the inverse RG step $n$.
Now, we note two facts: there is local von Neumann entropy $\sigma$ of $1$ nat at every site of the cHTN \cite{RINP}, and the cHTN is originally in the ground state.
Because of these two facts, per site, this lost energy $\varepsilon_n$ defines the physical Unruh temperature $T^U_n$ by \cite{Sagawa}
\begin{equation}
\sigma k_BT^U_n \equiv \varepsilon_n\;,\ \ \sigma=1\;.
\end{equation}
As the Unruh effect \cite{Davies,Unruh1,Unruh2}, this physical Unruh temperature $T_n^U$ is created by the physical Euclidean proper acceleration $a_n$ of the observational frame of reference as \footnote{Here, the Fock space is spatiotemporally locally defined and thus differs from the Fock space which is used in Refs. \cite{DL1,DL2}.}
\begin{eqnarray}
\frac{\hbar a_n}{2\pi c}=k_BT^U_n=\frac{Mc^2}{N_n}\frac{d \tau_0}{d \tau_n}\;.\label{eq:Unruh}
\end{eqnarray}
From this, we arrive at the final formula
\begin{equation}
a_n=\frac{2\pi c^3}{\hbar}\frac{M}{N_n}\frac{d \tau_0}{d \tau_n}\;,\label{eq:an}
\end{equation}
where the direction of the Euclidean proper acceleration maximally increases the coarse grain of the infinitesimal information reading $dI_0$ toward the top tensor, where the particle $M$ is located, and matches the forward RG direction.
Now, $a_n$ is the Euclidean proper acceleration, induced by the particle $M$ located at the top tensor, in the cHTN at the inverse RG step $n$.
Note that, in the context of general relativity, the Lorentzian gravitational proper acceleration in the cHTN at the inverse RG step $n$ is identically zero, since gravity is not a real force but a curved spacetime.
However, since we fix the Lorentzian spacetime metric $g_{\mu\nu}$ to the background static Lorentzian spacetime metric (\ref{eq:metric3}) recovered from the ground state of the boundary CFT$_2$, we can interpret this Euclidean proper acceleration $a_n$ as a gravity perturbation in the cHTN in the Euclidean regime of the HTN.
Next, the Euclidean proper acceleration, induced by a particle with non-zero mass $m$ located at a site in the cHTN at the inverse RG step $n$, in the cHTN at the top tensor is given by
\begin{equation}
a_0=\frac{2\pi c^3}{\hbar}\frac{m}{N_0}\frac{d \tau_n}{d \tau_0}\;.\label{eq:an2}
\end{equation}
Here, the direction of the Euclidean proper acceleration maximally increases the coarse grain (i.e., maximally decreases the inverse RG step $n$) of the infinitesimal information reading $dI_n$ toward the site where the particle $m$ is located; $dI_n$ is created by the particle $m$ after the infinitesimal imaginary proper time interval $d \tau_n$ and is coarse-grained to $di_{n\to 0}=dI_n/N_0$ at the top tensor.
Then, from Eqs. (\ref{eq:an}) and (\ref{eq:an2}), the consistency of this interpretation, that is, the conservation of Euclidean three-momentum, $p^\mu$, in the system of the particle $M$ with $p_0^\mu$ and the particle $m$ with $p_n^\mu$, holds as
\begin{equation}
p_0^r\sqrt{\gamma|_0}+p_n^r\sqrt{\gamma|_n}=0
\end{equation}
for the Wick-rotated metric $\gamma_{\mu\nu}|_n$ of the Euclidean spacetime at the inverse RG step $n$.\footnote{Here, $a_0^r=-a_0/\sqrt{\gamma_{rr}|_0}$ and $a_n^r=a_n/\sqrt{\gamma_{rr}|_n}$ hold.}
In this derivation of the gravity perturbation induced by a massive particle located at the top tensor of the cHTN from the Euclidean action (\ref{eq:H2}) of the cHTN, three points are essential:
\begin{enumerate}
\item[(i)] The relativistic Euclidean action of a massive particle at rest generates information reading after an imaginary time interval.
\item[(ii)] Information reading at the top tensor would be fine-grained in the cHTN along the inverse RG direction.
\item[(iii)] There is local von Neumann entropy of $1$ nat at every site of the cHTN \cite{RINP}.
\end{enumerate}
Because of this, the physical Unruh temperature $T^U_n$ is defined from the rest energy $Mc^2$ of the particle.
Then, in the Euclidean regime of the HTN, a gravity perturbation is induced by the particle $M$ in the cHTN as the Unruh effect.
\section{Conclusion}
In this article, we studied three subjects.
First, we proposed the generally covariant Lorentzian action of the cHTN by classicalizing real time in addition to the HTN.
Second, we investigated the properties of this Lorentzian action of the cHTN in the ground state.
Third, based on the Euclidean action of the cHTN, we derived the gravity perturbation induced by a massive particle at rest in the cHTN as the Unruh effect.
Our Euclidean and Lorentzian actions of the cHTN do not require the minimum action principle of classical mechanics but require the principal argument that the most probable configuration of the cHTN (i.e., the highest measurement entropy $H[|\psi\rra]$ of the cHTN) is likely realizable \cite{JHAP1,JHAP2}.
To conclude this article, we qualify the most probable real-time evolution of $|\psi(t)\ra$ in a generic state $|\psi\rra_L$ of Eq. (\ref{eq:rra}) in the Lorentzian regime of the cHTN in the presence or absence of matter beyond the gravity perturbation and consider its physical meanings.
(Here, the same qualification is applicable to the most probable imaginary-time evolution in the Euclidean regime of the cHTN.)
When we extremize the generally covariant Lorentzian action (\ref{eq:H}) of the cHTN with respect to $|\psi\rra_L$, there are two distinct tendencies.
First, the measurement entropy $H[|\psi(t)\ra]$ of the cHTN tends to be maximized at every real-time instance $t$.
Second, with respect to real time, the cHTNs tend to diversify into distinct ones with equal statistical weight: the more non-trivial real-time evolution is, the more entropy of the temporal part of $|\psi\rra_L$ is generated.
In the most probable real-time evolution of $|\psi(t)\ra$, these two distinct tendencies of the measurement entropy $H[|\psi\rra_L]$ stem from the equilibrium Boltzmann principle in the Euclidean regime of the HTN and the non-equilibrium second law in the Lorentzian regime of the HTN, respectively, in a generally covariant manner.
Here, the general covariance of $I_L[|\psi\rra_L]$ with respect to the bulk isometries is the first law.
Finally, we note that, in general relativity governed by the Einstein field equations, the counterpart of our holographic formulation of the Lorentzian spacetime exists in the real-time evolution of black holes, where the second law of gravity is the area theorem of the event horizons of black holes \cite{Laws1,Laws2}.
|
\section{Introduction and earlier results}
We consider the following transport-Stokes system:
\begin{equation}
\label{TS}
\left\{
\begin{array}{rcll}
\partial_t \rho+{\div}( u \rho ) &=& 0\,,& \text{on $\mathbb{R}^+ \times \mathbb{R}^3,$}\\
- \Delta u+ \nabla q &=& - \rho e_3 \,,& \text{on $\mathbb{R}^+ \times \mathbb{R}^3$},\\
\div u & =&0\,,& \text{on $\mathbb{R}^+ \times \mathbb{R}^3$},\\
\underset{|x|\to \infty}{\lim}|u| &=& 0,& \text{on } \mathbb{R}^+,\\
\rho(0,\cdot) &= & \rho_0 \,,&\text{on $ \mathbb{R}^3$}.
\end{array}
\right.
\end{equation}
The system \eqref{TS} has been derived in \cite{Hofer, Mecherbet} as a model for the sedimentation of a cloud of rigid particles in a viscous fluid in the case where inertia of both fluid and particles is neglected. Let us also mention that, recently, some systems coupling a transport equation and the Navier-Stokes equations have been obtained as
hydrodynamic limits of the Vlasov-Navier-Stokes system in \cite{HKM}.
Existence and uniqueness of solutions to \eqref{TS}
has been proved in \cite{Hofer} for regular initial data density, and extended to the case where the initial density is in $L^1(\R^3) \cap L^\infty(\R^3)$ with finite first moment in \cite{Mecherbet2}. A similar result, without the assumption on the finite first moment, has been proved in the parallel contribution \cite{Hofer&Schubert}. In the former paper, the authors also present a well-posedness result for a transport-Stokes like model taking into account the correction to the effective viscosity, the initial probability density is assumed to be $W^{1,\infty}(\R^3)\cap W^{1,1}(\R^3) $ in this case. In the paper \cite{Leblond}, the author proves global existence and uniqueness for $L^\infty(\R^3)$ initial densities in the case of bounded domain in $\R^3$ and $\R^2$ and also in the case of an infinite strip $\Omega= (0,1)\times \R$ with a flux condition. We refer also to the recent paper \cite{GrayerII} where the author proves global existence and uniqueness in $\R^2$ for $L^1(\R^2)\cap L^\infty(\R^2)$ compactly supported initial densities as well as propagation of H\"older regularity of the boundary of the patch.
Let us finally mention the work in progress \cite{Cobb} regarding the
wellposedness of the transport-fractional Stokes system.
Let us emphasize that the transport-Stokes system can be seen as a particular case of a large family of transport equations with a velocity given as the convolution of the density $\rho$ with a kernel $K$. The main key in order to ensure uniqueness is to establish suitable stability estimates. The first result relying on the 1-Wasserstein distance was obtained by Dobrushin \cite{Dobrushin} in the case where $\rho \in \mathcal{P}_1(\R^3)$ and $K\in Lip(\R^3,\R^3)$. This result has been generalized since for more singular kernels but for which, in counterpart, the probability density $\rho$ lies in $L^\infty(\R^3)$; we refer to the stability result by Hauray \cite{Hauray} in the case where the singular kernel (at origin) satisfies a compatibility condition that mainly ensures local integrability of $K$, $\nabla K$. This result has been adapted from the 2-Wasserstein stability estimate obtained by Loeper \cite{Loeper} for the Vlasov-Poisson system with $L^\infty$ probability measures. We finally refer to the paper by Duerinckx and Serfaty \cite{Serfaty&Duerinckx} where $K$ is the gradient of Coulomb, logarithmic or Riesz kernel. Authors prove the mean field convergence for a large system of interacting particles by introducing a suitable distance built as Coulomb (or Riesz) metric which overcomes the stability issue encountered when considering the Coulomb case with Wasserstein metric.
\section{Statement of the main results}
\subsection{Global existence and uniqueness result for \texorpdfstring{$L^1 \cap L^p$}{L1 n Lp} initial densities where \texorpdfstring{$p \geq 3$}{p>=3}.}
In this section, we state the first main result of this paper regarding the global existence and uniqueness of solutions to the
transport-stokes system with lower regularity than in the previous results, that is for $L^1 \cap L^p$ initial densities where $p \geq 3$.
Let us fix a notation before to state precisely the statement.
For $p\geq 1$, let $E_p$ be the subspace of probability measures
in $ L^p (\R^3) $.
In addition to the global existence and uniqueness issue, next result also establishes the validity of some well-known conservation laws and the existence of a flow map in a classical sense.
\begin{thm} \label{thm0}
Let $p\geq 3$ and $\rho_0$ in $E_p$. Then there is a unique corresponding solution
$$(u,q,\rho)\in
\left \{
\begin{array}{lcl}
C([0,+\infty),W^{2,p} (\R^3) \times W^{1,p} (\R^3)\times E_p) & \text {if} & p>3 ,\\ C([0,+\infty),\dot{W}^{2,3} (\R^3) \cap \dot{W}^{1,3} (\R^3)\cap \underset{q\in(3,+\infty]}{\bigcap} W^{1,q}(\R^3) \times W^{1,3} (\R^3)\times E_p) & \text {if} & p=3,
\end{array}
\right. $$
of the transport-Stokes equation \eqref{TS} in the sense of distributions, the density function satisfies the following conservation law: for any $t\geq 0$,
%
\begin{eqnarray}\label{conserLp}
\|\rho(t)\|_{ L^1 (\R^3) } =\|\rho_0\|_{ L^1 (\R^3) } \, \text{ and } \, \|\rho(t)\|_{L^p} =\|\rho_0\|_{ L^p} .
\end{eqnarray}
The velocity and pressure $(u,q)$ satisfy for $p>3$ and all $t\geq 0$
\begin{equation}\label{reg_u,p}
\|u(t)\|_{W^{2,p}(\R^3)}+ \|q(t)\|_{W^{1,p}(\R^3)} \leq C_p \| \rho(t) \|_{L^1(\R^3)\cap L^p(\R^3)},
\end{equation}
and for $p=3$, $q\in(3,+\infty]$
\begin{equation}\label{reg_u,p_bis}
\|\nabla u(t)\|_{W^{1,3}(\R^3)}+\|u(t)\|_{W^{1,q}(\R^3)}+ \|q(t)\|_{W^{1,3}(\R^3)} \leq C_q \| \rho(t) \|_{L^1(\R^3)\cap L^3(\R^3)}.
\end{equation}
Moreover,
\begin{equation}\label{ulipoullip}
u \in
\begin{cases}
C([0,+\infty), C^{1,\mu} (\R^3), & \text{ if } \, p > 3,
\\ C([0,+\infty), \text{Log-Lip} (\R^3) ), & \text{ if } \, p = 3.
\end{cases}
\end{equation}
where $\mu=1-3/p$,
see Definition \ref{def:LLspace} for more details on Log-Lipschitz spaces. Finally, for all $s\in[0,+\infty)$, there exists a unique
\begin{equation}\label{uflox}
X(\cdot,s,\cdot)-\operatorname{Id} \in
\begin{cases}
C([0,+\infty), C^{1,\mu} (\R^3) ), & \text{ if } \, p > 3,
\\ {C([0,+\infty), C^{0,r_t} (\R^3) ),} & \text{ if } \, p =3,
\end{cases}
\end{equation}
where $r_t = e^{-Ct} $, with $C>0$ a constant depending only on $\|\rho_0\|_{L^1(\R^3)\cap L^3(\R^3)}$,
such that
\begin{equation} \label{cflow-def}
\left \{
\begin{array}{rcll}
\partial_t Xt,s,x) & = & u(t, X(t,s,x)),&\: \: \forall \, t,s\in[0,+\infty) , \\
X(s,s,x) & = & x,& \: \: \forall \, s \in [0,+\infty) ,
\end{array}
\right.
\end{equation}
and for any $t$ in $[0,+\infty)$,
\begin{equation}
\label{pousse-th}
\rho(t,\cdot) =X(t,0,\cdot) \#\rho_0 .
\end{equation}
%
\end{thm}
%
In \eqref{pousse-th} the notation
$\#$ stands for the push-forward of the measure which follows the symbol by the mapping which precedes the symbol.
\begin{rem}\label{analog}
The critical case where $p=3$ can be seen as a counterpart of the famous result by Yudovich in
\cite{Yudovich63} on the existence and uniqueness of the weak solutions with bounded vorticity of the 2D incompressible Euler equations.
In particular, this result uses in a key manner that the velocity field is log-Lipschitz. A related result, with a conditional flavour, also holds in the case of the Vlasov-Poisson equations, see \cite{Loeper}.
\end{rem}
The most delicate part of the proof of Theorem \ref{thm0} is the uniqueness issue.
Indeed Theorem \ref{thm0} is obtained by classical arguments from the following stability estimates where $W_1$ denotes the first Wasserstein distance defined for two probability measures $\rho_1,\rho_2 $
by
$$ W_1 (\rho_1,\rho_2) :=
\underset{\pi \in \Pi(\rho_1,\rho_2)} \inf \int |x-y| d\pi(x,y) , $$
where $\Pi(\rho_1,\rho_2)$ is the set of admissible transport plans $\pi $ in $\mathcal{P}(\mathbb{R}^3\times \mathbb{R}^3)$
having $\rho_1$ as first marginal and $\rho_2$ as second marginal.
\begin{thm} \label{thm}
Let $p\geq 3$.
For any $R>0$, there is $C>0$ such that for any couple of initial densities
$ (\rho_1^0 , \rho_2^0 )$ in $E_p $ with
$ \max_{i=1,2} (\|\rho_i^0\|_{ L^1 (\R^3) },\|\rho_i^0\|_{ L^p (\R^3) }) \leq R$, if $ (u_1,\rho_1) $ and $ (u_2 ,\rho_2)$ satisfy the transport-Stokes equation \eqref{TS}
for any $t\geq 0$, then
%
\begin{align}
\label{sta-geq-3}
W_1(\rho_1(t),\rho_2(t)) &\leq W_1(\rho_1^0,\rho_2^0)e^{C t}, \quad \text{if } p > 3,
\\ \label{sta-geq-critik}
W_1 (\rho_1(t),\rho_2(t)) &\leq W_1(\rho_1^0,\rho_2^0)^{e^{-C t}} , \quad \text{if } p = 3 .
\end{align}
%
\end{thm}
%
The proof of Theorem \ref{thm} is given in Section
\ref{sec-stabb}
and
Theorem \ref{thm0} is deduced from Theorem \ref{thm} in Section
\ref{sec-proof-princi}.
%
\begin{rem}\label{rem:TS_avec_terme_source}
We emphasize that the existence and uniqueness result can be extented to the case of transport-Stokes system with an additional source term $f$ in the right hand side of the Stokes equation provided that $f \in C([0,+\infty), L^1(\R^3) \cap L^p(\R^3))$.
\end{rem}
\subsection{Analyticity of the flow map when \texorpdfstring{$p > 3$}{p>3}.}
Next, we prove that, in the case where $p>3$,
the flow map which describes the trajectories of these solutions is analytic with respect to time. More precisely the following result holds true.
\begin{thm}
\label{analytic}
Let $p>3$, $\rho_0$ in $E_p$ and
$ (u,\rho)$ in $C([0,+\infty),W^{2,p} \times E_p)$
the unique corresponding solution of the transport-Stokes equation \eqref{TS}.
Let $\mu:=1-3/p$.
Let $X$ the flow associated with $u$ by \eqref{uflox} and
\eqref{cflow-def}.
Then $ X$ is analytic from $[0,+\infty)$ to $ C^{1,\mu}(\R^3)$.
\end{thm}
This result extends to the transport-Stokes system a result which was obtained for classical solutions of the incompressible Euler system by Chemin in \cite{Chemin}, and reproved by various other methods see by, among others \cite{CVW,FZ,Gamblin,Inci,Kato,ogfstt,Serfati,Shnirelman}.
To prove Theorem \ref{analytic}, the key point is to prove the following result.
%
\begin{prop} \label{apana}
There exists $T>0$ depending only on $\|\rho_0\|_{L^1 \cap L^p}$ and two constants $C_1 $ and $C_0$ depending only on $\|\rho_0\|$, $T$ and $\mu$ such that for all $t\in[0,T]$ and for all $n\geq1$,
\begin{equation}\label{rec_formule}
\|\partial_t^{n} X(t,\cdot)\|_{1,\mu} \leq (-1)^{n-1} n! \begin{pmatrix}1/2 \\ n \end{pmatrix}C_0^n C_1^{n-1}.
\end{equation}
\end{prop}
Above we have used, for any $n\geq 1$, the notation
%
\begin{equation}\label{binomial_1/2}
\begin{pmatrix}
1/2\\
n
\end{pmatrix}= \frac{(1/2)(1/2-1) \cdots (1/2-n+1)}{n!};
\end{equation}
which is extended by convention by
$\begin{pmatrix}
1/2\\0
\end{pmatrix}=-1$.
With this definition we observe that
$$
(-1)^{n-1}\begin{pmatrix}
1/2\\n
\end{pmatrix}\geq 0.
$$
Moreover, for all $n\geq 2$,
\begin{equation}\label{binomial_1/2_borne}
(-1)^{n-1} n!\begin{pmatrix}
1/2\\n
\end{pmatrix} \leq C \frac{n!}{2^n}.
\end{equation}
Therefore to deduce Theorem \ref{analytic} from Proposition \ref{apana},
it is sufficient to combine \eqref{rec_formule} and \eqref{binomial_1/2_borne}
on any time interval of the form $[t,t+T] $. Indeed, since the ODE equation for the flow is autonomous, we have stated Proposition \ref{apana} in the case where $t=0$ for simplicity.
The proof of Proposition \ref{apana} is given in Section
\ref{sec-apana} and is based on the approach developed in \cite{CVW} where authors prove analyticity of Lagrangian trajectories for several incompressible inviscid models including the 2D surface quasi-geostrophic equation, the 2D incompressible porous medium equation, the 2D and the 3D incompressible Euler equations and the 2D Boussinesq equations. It turns out that, for the transport-Stokes model, the Oseen kernel benefits from better integrability properties, compared to the kernels considered in \cite{CVW}, which allows us to consider an initial density function $\rho_0$ which is less regular than the initial data considered in \cite{CVW}.
\begin{rem}\label{G3}
In the case where $p=3$, it is likely that
the flow map $X$ still benefits from Gevrey regularity from $[0,+\infty)$ to $C^{r_t}(\R^3)$.
Such a result is proved in the case of the $2D$ incompressible Euler equations with bounded vorticity in \cite{Gamblin} in the case where the fluid occupies the full space and in
\cite{Sueur} in the case where the fluid occupies a bounded domain.
\end{rem}
\subsection{Exact controllability of the transport-Stokes system}
Our third main result establishes the small-time global exact controllability of the transport-Stokes system
when distributed forces are allowed in an arbitrary non-empty open subset of $\R^3$, in the case where the initial and final densities are in $L^p_c$ with $p \geq 3$.
Here the index $c$ means that these densities are compactly supported.
\begin{thm}\label{TS-control}
Let $T>0$, $p\geq 3$. Let $\omega$ a non-empty open subset of $\R^3$.
Let $\rho_0$ and $\rho_f$ in $L^p_c$.
There exists $ (u,\rho)$ in $C([0,T),W^{2,p}(\R^3) \times L^p(\R^3))$ and analogously for $p=3$, $u \in C([0,T],\dot{W}^{1,3}(\R^3)\cap \dot{W}^{2,3}(\R^3) \cap \bigcap_{q>3} L^q(\R^3)$, there are $f \in C_c(0,T, L^\infty(\R^3))$ and $g \in C_c(0,T,L^1(\R^3)\cap L^p(\R^3)) $, both compactly supported in $(0,T)\times \omega $ such that
\begin{equation*}
\left \{
\begin{array}{rcll}
\partial_t \rho +\div(\rho u)&=& g,& \text{ on } \R^3 ,\\
-\Delta u+\nabla p &=& - \rho e_3+f ,& \text{ on } \R^3 ,\\
\div u&=& 0,& \text{ on } \R^3 ,\\
\rho(0)=\rho_0&,& \rho(T)=\rho_f,& \text{ on } \R^3 .
\end{array}
\right.
\end{equation*}
\end{thm}
Theorem \ref{TS-control} adapts to the case of the transport-Stokes system some earlier results obtained for the incompressible Euler equations in \cite{Coron96,G20,G}.
The proof follows closely a now well-known strategy, relying on Coron's return method and the existence of peculiar vector fields, which solves the Stokes system in the uncontrolled zone $\R^3 \setminus \omega$. For sake of completeness, we give a sketch of proof in Section
\ref{sec-cont}.
Let us also mention that results regarding the Lagrangian controllability could also be obtained, in the spirit of the results \cite{GH1,GH2,HK} in the case of the Euler equations and the result \cite{GH3} in the case of the steady Stokes equations.
The proof of Theorem \ref{TS-control} is given in Section
\ref{sec-control}.
\subsection{A few open questions on the transport-Stokes system}
\begin{itemize}
\item
It would be interesting to investigate whether the results obtained in this paper can be extended to the case of $\R^2$ and the case of domains with boundaries. We have already mentioned the recent paper \cite{GrayerII} for the case of compactly supported initial densities in $L^1(\R^2)\cap L^\infty(\R^2)$ and the paper \cite{Leblond} where existence and uniqueness is proved in the case of bounded domain in $\R^3$ and $\R^2$ and also in the case of an infinite strip $\Omega= (0,1)\times \R$ with a flux condition. In both cases the density is assumed to be in $ L^{\infty}$. Regarding the issue of the regularity in time of the flow map, let us mention \cite{Kato,ogfstt,GS,Sueur} where the case
of the incompressible Euler equations in domains with boundaries is tackled.
\item Another question is whether or not the well-posedness part of our results could be extended to lower regularity by making use of the theory of renormalized or Lagrangian solutions in the spirit, for example, of
\cite{Lions,CCS}. One natural goal in this direction would be to be able to deal with densities $\rho$ which are merely measures, with the aim to tackle
the mean field limit.
\item It has been shown in \cite{Mecherbet2} that if $\rho_0=1_{B_0}$ with $B_0$ the unit ball then the solution $\rho=1_{B}$ remains a spherical patch for all time, more precisely there exists a constant $v\in \R^3$ such that $(\rho,u)$ satisfies
\begin{equation}\label{eq:hadamard_Rybczynski}
\begin{array}{rcl}
(u(t,x),\rho(t,x))&=&(u_0(x-vt),\rho_0(x-vt)),\\
(u-v)\cdot n& =& 0 \text{ on } \partial B ,
\end{array}
\end{equation}
where $n$ is the unit outer normal vector. A natural question is then to investigate the reciprocal property, that is, if there exists a constant velocity $v$ and a bounded domain $B_0$ for which \eqref{eq:hadamard_Rybczynski} is satisfied then $B_0$ is necessarily a ball. Note that this type of characterisation of the ball has been proved in any dimension by \cite[Theorem 1.1]{Fraenkel} in the case where $u_0$ is a Newtonian potential housed in a bounded open set $G$. The author shows that if $u_0$ is constant on $\partial G$ then necessarily $G$ is a ball.
A related issue would be to investigate the spherical case in the two dimensional setting as well since well posedness of $2$d patchs were established in \cite{GrayerII}.
\end{itemize}
\section{A few preliminary reminders}
In this section we gather some classical material which is useful later on.
\subsection{Multivariable calculus tools}
We denote by $\mathbb{N}_0^3$ the set of three dimensional multi-indices $\alpha=(\alpha_1,\alpha_2,\alpha_3)$ endowed with the component-wise sum and difference and with a partial order
$$
\alpha \leq \beta \Leftrightarrow \Big( \alpha_i \leq \beta_i ,\quad \forall 1 \leq i \leq 3 \Big).
$$
We introduce the following notations: for any $\alpha,\beta \in \mathbb{N}_0^3$, for any $y\in\R^3$, we set
\begin{gather*}
|\alpha|:= \alpha_1+\alpha_2+\alpha_3 ,\quad
\alpha! := \alpha_1! \alpha_2!\alpha_3! ,\quad
\partial^\alpha :=\partial_{x_1}^{\alpha_1}\partial_{x_2}^{\alpha_2} \partial^{\alpha_3} ,\quad
y^\alpha :=\left(y_1^{\alpha_1}\right) \left( y_2^{\alpha_2} \right) \left(y_3^{\alpha_3}\right) ,\\
\begin{pmatrix}
\alpha \\
\beta
\end{pmatrix}= \begin{pmatrix}
\alpha_1 \\
\beta_1
\end{pmatrix}\begin{pmatrix}
\alpha_2 \\
\beta_2
\end{pmatrix}\begin{pmatrix}
\alpha_3 \\
\beta_3
\end{pmatrix}= \frac{\alpha!}{\beta!(\alpha-\beta)!},
\end{gather*}
and we recall the classical Leibniz formula
$$
\partial^\alpha (fg) = \underset{\beta\leq\alpha}{\sum} (\partial^\beta f) (\partial^{\alpha-\beta} g).
$$
We introduce for all $n\geq1$, $1\leq s \leq n$ and $\alpha \in \mathbb{N}_0^3$ with $1 \leq |\alpha|\leq n$ the set
$P_s(n,\alpha)$ of the
$$(k_1,\cdots,k_s;l_1,\cdots,l_s)\in \N_0^3\times \cdots\times \N_0^3 \times \N \times \cdots \N, $$
such that
$$0<|k_j|,\quad 0<l_1<\cdots<l_s,\quad \underset{j=1}{\overset{s}{\sum}}k_j=\alpha,\quad \underset{j=1}{\overset{s}{\sum}}|k_j| l_j=n .$$
We now recall the following formula from \cite[Lemma 3.2]{CVW}, which is useful for the proof of Proposition \ref{apana}.
\begin{lemme}[Multivariate Fa\`a di Bruno formula]\label{Lemme_FDB}
Let $g:\R \mapsto \R^3$ be a vector function $ C^\infty$ in the neighborhood of $x_0 \in \R$.
Let $h : \R^3 \to \R$ be a scalar function $ C^\infty$ in the neighborhood of $y_0=g(x_0)$. Define $f(x):=h(g(x)): \R \to \R$. Then
\begin{equation}\label{FDB_formula}
f^{(n)}(x_0)= n! \underset{1\leq |\alpha|\leq n}{\sum} (\partial^\alpha h)(g(x_0)) \underset{s=1}{\overset{n}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{(\left(\partial^{l_j} g)(x_0)\right)^{k_j}}{(k_j!) (l_j!)^{|k_j|}} ,
\end{equation}
holds for any $n\geq 1$, with the convention that $0^0:=1$.
\end{lemme}
\subsection{H\"older, Lipschitz and log-Lipschitz spaces}
We introduce the following notations for Lipschitz, log-Lipschitz and H\"older spaces.
\begin{defi}[H\"older spaces]
Let $ \mu \in ]0,1]$ and $n\in \mathbb{N} $
$$
C^{n,\mu}(\R^3):= \{ \psi \in C^n(\mathbb{R}^3), \sum_{|\alpha|\leq n}\underset{x \in \R^3}{\sup}|\partial^\alpha \psi(x)|+
\sum_{|\alpha|= n}[\partial^\alpha \psi]_{0,\mu}<+\infty \},
$$
\end{defi}
where
$$ [\psi]_{0,\mu} := \underset{x \neq y}{\sup}\, \frac{|\psi (x)-\psi (y)|}{|x-y|^\mu}.$$
The space $ C^{n,\mu}(\R^3)$ is endowed with the following norm
$$
\|\psi\|_{n,\mu}:= \underset{i=0}{\overset{n}{\sum}}\underset{x \in \R^3}{\sup}|\partial_i \psi(x)|+ [\partial_n \psi]_{0,\mu}.
$$
In the case of the space of Lipschitz functions $ C^{0,1}(\R^3) $ we use the notation
$$
\text{Lip}(\psi):= \underset{x \neq y}{\sup}\, \frac{|\psi(x)-\psi(y)|}{|x-y|},
$$
and set
$$
\text{Lip}(\R^3):= C^{0,1}(\R^3).
$$
We introduce the following real function $\ln_{-}$
\begin{equation}\label{def_ln-}
\ln_{-}(s):=\max(0,-\ln(s))=\left\{
\begin{array}{rcl}
-\ln(s) & if & s\in[0,1], \\
0 & if & s \geq 1 .
\end{array}
\right.
\end{equation}
\begin{defi}[Log-Lipschitz space]\label{def:LLspace}
We introduce the following notation for any function $\psi$
$$
\text{Log-Lip}(\psi):= \underset{x \neq y}{\sup} \frac{|\psi(x)-\psi(y)|}{|x-y|(1+\ln_{-}|x-y|)}.
$$
We say that a bounded continuous function $\psi$ is Log-Lipschitz if $\text{Log-Lip}(\psi)<+\infty$ and we set
$$
\text{Log-Lip}(\R^3):=\{\psi \in C(\R^3), \|\psi\|_\infty+ \text{Log-Lip}(\psi)<+\infty \}.
$$
\end{defi}
\subsection{Steady Stokes equations}
\label{SSE}
Let $(\mathcal{U},\mathcal{P})$ the Stokes Green function, also called the Oseen tensor, given by
\begin{equation}\label{Oseen}
\mathcal{U}(x)= \frac{1}{8\pi |x|} \left(Id + \frac{x \otimes x}{|x|^2} \right) \ \text{ and } \ \mathcal{P}(x)=- \frac{1}{4\pi} \frac{x}{|x|^2} .
\end{equation}
Let us observe that the Oseen tensor
$\mathcal{U}$ satisfies the following bound:
\begin{equation}
\label{B-prop}
\left|\mathcal{U}(x) \right|\leq C \frac{1}{|x|}, \: \: \forall x \in \R^3 \setminus \{0\} ,
\end{equation}
and
\begin{equation}
\label{Lip-prop}
\left|\mathcal{U}(x)-\mathcal{U}(y) \right|\leq C |x-y| \left(\frac{1}{|x|^2}+ \frac{1}{|y|^2} \right), \: \: \forall x, y \in \R^3 \setminus \{0\} .
\end{equation}
Let us recall the following classical result on the Stokes equations, see for example \cite[Theorem IV.2.1]{Galdi}.
\begin{prop}\label{class-Stokes}
Let $f $ in $L^p (\R^3) \cap L^1 (\R^3) $ with $p\geq 3$.
Then
there is a unique solution of the Stokes equation
\begin{equation}
\label{cstokes}
\left\{
\begin{array}{rcll}
- \Delta u+ \nabla \pi &=& f \,,& \text{in $ \mathbb{R}^3$},\\
\div u & =&0\,,& \text{in $ \mathbb{R}^3$},\\
\underset{|x|\to \infty}{\lim}|u| &=& 0 .
\end{array}
\right.
\end{equation}
such that $(u,\pi)\in W^{2,p} (\R^3) \times W^{1,p} (\R^3) $ if $p>3$ and $(\nabla u,\pi)\in W^{1,3} (\R^3) \times W^{1,3} (\R^3) $, $u\in W^{1,q}(\R^3)$ for any $q>3$ in the case $p=3$.
Moreover,
\begin{equation}
\label{formu-conv}
u=\mathcal{U} \star f \text{ and } \pi=\mathcal{P} \star f .
\end{equation}
\end{prop}
\subsection{Sobolev embedding}
The reason why the value $p=3$ appears in our analysis is due to the following result, which establishes that the unique velocity solution of the Stokes equation
with data in $ L^1\cap L^p$ is Lipschitz when $p>3$ and Log Lipschitz in
the critical case $p=3$.
\begin{prop}\label{prop_log_lip_u}
Let $\rho$ in $L^p (\R^3) \cap L^1 (\R^3) $ with $p\geq 3$, and
$(u,p) $ the unique corresponding solution in $W^{2,p} (\R^3) \times W^{1,p} (\R^3) $ (resp. in $\dot{W}^{2,3} (\R^3)\cap \dot{W}^{1,3}(\R^3)\cap W^{1,q}(\R^3) \times W^{1,3} (\R^3) $ for $p=3$, $q\in(3,+\infty]$)
of the Stokes equation \eqref{cstokes}.
Then
\begin{equation}\label{uli}
u \in
\begin{cases}
\text{Lip} (\R^3) , & \text{ if } \, p > 3,
\\ \text{Log-Lip} (\R^3) , & \text{ if } \, p = 3,
\end{cases}
\end{equation}
with the following estimates:
\begin{align} \label{esti-uli}
\text{Lip}(u) \leq C \| \rho\|_{L^1 (\R^3)\cap L^p (\R^3)} , & \text{ if } \, p > 3,
\\ \label{esti-ulli}
\text{Log-Lip} (u) \leq C \| \rho\|_{L^1 (\R^3)\cap L^3 (\R^3)} & \text{ if } \, p = 3 .
\end{align}
\end{prop}
\begin{proof}
In the case where $p>3$ we have the embedding $W^{2,p} (\mathbb{R}^3) \hookrightarrow W^{1,\infty} (\mathbb{R}^3)$, so that
\begin{eqnarray} \label{cotestokes}
\|u\|_{W^{1,\infty} (\R^3)}\leq C\|u\|_{W^{2,p} (\R^3)} \leq C \| \rho\|_{L^1 (\R^3)\cap L^p (\R^3)}.
\end{eqnarray}
Let us now deal with the case where $p=3$.
We provide here a self-contained proof which we will adapt later on in the proof of
Theorem \ref{thm}.
First, by the convolution formula \eqref{formu-conv}, we have
\begin{align*}
|u(x)| &\leq C \int_{B(x,1)} \frac{\rho(z)}{|x-z|} dz + \int_{B(x,1)^c}\frac{\rho(z)}{|x-z|} dz
\\ &\leq \left(\int_{B(x,1)}\frac{dz}{|x-z|^2} \right)^{1/2} \|\rho\|_{L^2(\R^3)}+\|\rho\|_{L^1(\R^3)}
\\ &\leq C \|\rho\|_{L^1(\R^3) \cap L^3(\R^3)},
\end{align*}
where the notation $\cdot^c$ stands for the complementary set in $\R^3$.
Second, by the convolution formula \eqref{formu-conv}, we obtain that for
any $x\neq x'$ in $\mathbb{R}^3$,
\begin{align*}
|u(x)-u(x') |&\leq
I_1 +I_2 ,
\end{align*}
where
\begin{align*}
I_1 &:= \int_{B(x,\epsilon) \cup B(x',\epsilon) } \left|\mathcal{U}(x-u)-\mathcal{U}(x'-u)\right|\rho(u)du , \\
I_2 &:= \int_{B(x,\epsilon)^c \cap B(x',\epsilon)^c} \left|\mathcal{U}(x-u)-\mathcal{U}(x'-u)\right|\rho(u)du ,
\end{align*}
with $\epsilon :=|x-x'|$.
For $I_1$, by \eqref{B-prop},
we have
\begin{align*}
I_1& \leq C \Big( \int_{B(x,\epsilon)} \frac{1}{|x-u|}\rho(u)du + \int_{B(x',\epsilon)} \frac{1}{|x-u|}\rho(u)du
\\ &\quad + \int_{B(x,\epsilon)} \frac{1}{|x'-u|}\rho(u)du + \int_{B(x',\epsilon)} \frac{1}{|x'-u|}\rho(u)du \Big) .
\end{align*}
Since
$B(x,\epsilon) \subset B(x',2\epsilon) $ and $ B(x',\epsilon) \subset B(x,2\epsilon)$ by the triangle inequality and the definition of $\epsilon$,
we have that
$$
I_1 \leq C \left( \int_{B(x,2\epsilon) } \frac{1}{|x-u|}\rho(u)du + \int_{B(x',2\epsilon)} \frac{1}{|x'-u|}\rho(u)du \right) \leq C \epsilon \|\rho\|_{L^3 (\R^3)} ,
$$
by H\"older's inequality.
For $I_2$, by \eqref{B-prop},
we have
\begin{align*}
I_2 &\leq C |x-x'| \int_{B(x,\epsilon)^c \cap B(x',\epsilon)^c} \left(\frac{1}{|x-u|^2}+\frac{1}{|x'-u|^2} \right)\rho(u) du\\
&\leq C |x-x'|\left( \int_{B(x,\epsilon)^c} \frac{1}{|x-u|^2} \rho(u) du+ \int_{B(x',\epsilon)^c} \frac{1}{|x'-u|^2} \rho(u) du\right) .
\end{align*}
If $\epsilon=|x-x'|\geq 1$ we have
$I_2\leq C |x-x'|\| \rho\|_1$ and this yields the desired result.
On the other hand, if $\epsilon<1$, then
\begin{align*}
\int_{B(x,\epsilon)^c} \frac{1}{|x-u|^2} \rho(u) du &\leq \int_{B(x,2)\setminus B(x,\epsilon) } \frac{1}{|x-u|^2} \rho(u) du + \int_{B(x,2)^c} \frac{1}{|x-u|^2} \rho(u) du\\
&\leq C\|\rho\|_{L^3 (\R^3)} | \log(\epsilon)|+ \|\rho\|_{L^1 (\R^3)} ,
\end{align*}
by H\"older's inequality.
By combining the estimates we conclude the proof of \eqref{esti-ulli}.
\end{proof}
\subsection{Flow map and Osgood's lemma}
The following result on the existence and uniqueness of a classical flow in the case of Lipschitz or Log-Lipschitz vector fiels is quite classical, let us refer for example to \cite{BCD}.
\begin{prop}\label{prop_flow}
Let $u$ a vector field in \eqref{ulipoullip}.
Then there exists a unique corresponding flow
$X$ such as in \eqref{uflox} and satisfying \eqref{cflow}.
\end{prop}
In order to recall Osgood Lemma we introduce first the notion of Osgood modulus of continuity, see \cite[Definition 3.1]{BCD} for more details.
\begin{defi}
Let $a>0$. A modulus of continuity $ \eta: [0,a] \to [0,+\infty[$ is any continuous non-decreasing function vanishing at $0$ and continuous at $0$.
We say that $\eta$ is an Osgood modulus of continuity if in addition
$$\int_0^a \frac{dz}{\eta(z)} = \infty.
$$
\end{defi}
We recall Osgood lemma which will be useful for the proof of Theorems \ref{thm0} and \ref{thm} in the critical case $p=3$.
\begin{lemme}[Osgood Lemma]\label{Osgood}
Let $f$ a measurable function from $[t_0,T]$ to $[0,a]$, $v$ a locally integrable function from $[t_0,T]$ to $\R^+$ and $ \eta: [0,a] \to [0,+\infty[$ an Osgood modulus of continuity. Assume that there exists $c \geq0$ such that for a.e $t\in[t_0,T]$
$$
f(t) \leq c+\int_{t_0}^t v(s)\, \eta(f(s)) ds.
$$
\begin{enumerate}
\item If $c>0$ then we have
$$
\mathcal{M}(c) \leq \mathcal{M}(f(t)) + \int_{t_0}^t v(s) ds \text{ with } \mathcal{M}(x)=\int_x^a \frac{dz}{\eta(z)}.
$$
\item If $c=0$ then $f=0$ a.e on $[t_0,T]$.
\end{enumerate}
\end{lemme}
\section{Proof of Theorem \ref{thm} }
\label{sec-stabb}
%
Let $p\geq 3$, $R>0$, a couple of initial densities
$ (\rho_1^0 , \rho_2^0 )$ in $E_p $ with $ \max_{i=1,2} (\|\rho_i^0\|_{ L^1 (\R^3) },\|\rho_i^0\|_{ L^p (\R^3) }) \leq R$,
and $ (u_1,\rho_1) $ and $ (u_2 ,\rho_2) $ in $C([0,+\infty),W^{2,p} \times E_p)$
satisfying the transport-Stokes equation \eqref{TS}
for any $t\geq 0$, with, respectively, $ \rho_1^0 $ and $\rho_2^0 $ as initial condition.
%
Let us recall that
if $\rho_2$ is absolutely continuous with respect to the Lebesgue measure then
\begin{equation}
\label{W-infi}
W_1 (\rho_1,\rho_2)= \underset{T\# \rho_2=\rho_1} \inf \int_{\R^3} |T(x)-x|d\rho_2(x) ,
\end{equation}
%
where the infimum runs over all measurable transport maps $T : \mathbb{R}^3 \to \mathbb{R}^3$ and there exists an optimal transport map $T^*$ for which the infimum is reached, see \cite[Theorem 1.5]{Santambrogio}. We emphasize that we use transport maps in the whole analysis for the sake of lightness and that one can adapt all the arguments using only optimal transport plans which are known to exist, see \cite[Theorem 1.7]{Santambrogio_book}.
We introduce $T_0$ the optimal transport map such that
\begin{equation}
\label{enplu}
\rho_1^0=T_0\#\rho_2^0 ,
\end{equation}
%
and
\begin{equation}
\label{azero}
W_1(\rho_1^0,\rho_2^0)=\int_{\R^3} |T_0(x)-x| \rho_2^0(x) dx.
\end{equation}
%
For $i=1,2$, let
$X_i$ associated the characteristic flow with $u_i$ by
\begin{equation} \label{cflow}
\left \{
\begin{array}{rcll}
\partial_t X_i(t,s,x) & = & u_i(s, X_i(t,s,x)),&\: \: \forall \, t,s\in[0,+\infty) , \\
X_i(s,s,x) & = & x,& \: \: \forall \, s \in [0,+\infty) ,
\end{array}
\right.
\end{equation}
For $i=1,2$,
for any $t$ in $[0,+\infty)$,
\begin{equation}
\label{pousse}
\rho_i (t,\cdot) =X_i (t,0,\cdot) \#\rho_i^0 .
\end{equation}
%
We set, for any $t\geq 0$,
\begin{equation}
\label{RG}
T_t :=X_1(t,0,\cdot) \circ T_0 \circ X_2(0,t,\cdot) .
\end{equation}
Let us define, for any $t\geq 0$,
\begin{equation*}
Q(t) := \int_{\R^3} |T_t(x) -x| \rho_2^t(dx) .
\end{equation*}
By \eqref{azero}, we observe that
\begin{equation}
\label{prem-zero}
W_1 (\rho_1^0,\rho_2^0) =Q(0) .
\end{equation}
In what follows we use the shortcut $\rho_i^t:=\rho_i(t,\cdot)$ for $t\geq 0 $ and $i=1,2$.
From \eqref{pousse}, \eqref{enplu} and \eqref{RG} we deduce that for any $t\geq 0$,
\begin{equation}
\label{enplusse}
\rho_1^t=T_t \#\rho_2^t .
\end{equation}
%
We therefore deduce from \eqref{W-infi} and \eqref{enplusse}
that for any $t\geq 0$,
\begin{equation}
\label{prem}
W_1 (\rho_1^t,\rho_2^t) \leq Q(t) .
\end{equation}
Now, using \eqref{pousse} (with $i=2$) and \eqref{RG} (which implies that for any $t\geq 0$,
$T_t \circ X_2(t,0,\cdot)=X_1(t,0,\cdot) \circ T_0 $),
we have for any $t\geq 0$,
\begin{equation}
\label{reexpQ}
Q(t)
= \int_{\R^3} | X_1(t,0,T_0 (x)) -X_2(t,0,x)| \rho_2^0(dx).
\end{equation}
From \eqref{cflow}
we deduce that for any $x$ in $\R^3$, for any $t\geq 0$,
\begin{align} \nonumber
| X_1(t,0,T_0 (x)) -X_2(t,0,x)| & \leq |T_0 (x)- x| + \int_0^t |u_1(s,{X}_1(s,0,T_0 (x))) -u_2(s,{X}_2(s,0,x))| ds
\\ \nonumber & \leq |T_0 (x)- x| + \int_0^t |u_1(s,{X}_1(s,0,T_0 (x))) -u_1(s,{X}_2(s,0,x))| ds
\\ \label{esti-arep}
& \quad + \int_0^t |u_1(s,{X}_2(s,0,x)) -u_2(s,{X}_2(s,0,x))| ds .
\end{align}
To go further we need to distinguish the case where $p> 3$ and the case where $p=3$.
\subsection{Case where \texorpdfstring{$p> 3$}{p>3}}
Let us first deal with the case where $p> 3$.
Then by Proposition \ref{prop_log_lip_u}, the velocity field $u_1$ is Lipshitz and therefore for any $t\geq 0$,
\begin{align*}
| X_1(t,0,T_0 (x)) -X_2(t,0,x)|
& \leq |T_0 (x)- x| + \int_0^t \text{Lip}(u_1(s)) |{X}_1(s,0,T_0 (x)) -{X}_2(s,0,x)| ds\\ \quad
&+ \int_0^t |u_1(s,{X}_2(s,0,x)) -u_2(s,{X}_2(s,0,x))| ds .
\end{align*}
By integration over $\R^3$ with respect to the measure $\rho_2^0(dx)$, we deduce by
using \eqref{pousse} and \eqref{reexpQ} that for any $t\geq 0$,
%
\begin{equation} \label{ineq-Q}
Q(t) \leq Q(0) + \int_0^t \text{Lip}(u_1(s)) Q(s) ds+ \int_0^t \int |u_1(s,x) -u_2(s,x)| \rho_2^s(dx) ds .
\end{equation}
Let us now estimate the last term in the right hand side of \eqref{ineq-Q}.
By using the convolution formula \eqref{formu-conv}
and \eqref{enplusse}
we obtain that for any $s\geq 0$,
\begin{equation} \label{demen}
\int_{\mathbb{R}^3} |u_1(s,x)-u_2(s,x)| \rho^s_2(x)dx
= \int_{\mathbb{R}^3} \left| \int_{\mathbb{R}^3}( \mathcal{U}( x-T_s (y)) - \mathcal{U}( x-y) )e_3\rho_2^s(y)dy \right| \rho_2^s(dx) .
\end{equation}
Thus we deduce from \eqref{demen} and \eqref{Lip-prop} that for any $s\geq 0$,
\begin{align*}
&\int_{\mathbb{R}^3} |u_1(s,x)-u_2(s,x)| \rho^s_2(x)dx
\\ &\quad \quad \leq C \int_{\mathbb{R}^3} \left| \int_{\mathbb{R}^3}|T_s (y)-y| \left(\frac{1}{|x-y|^2}+ \frac{1}{|x-T_s (y)|^2} \right) \rho_2^s(y)dy \right| \rho_2^s(dx),
\end{align*}
Therefore, by Fubini's principle, for any $s\geq 0$,
%
\begin{align} \label{triple}
& \int_{\mathbb{R}^3} |u_1(s,x)-u_2(s,x)| \rho_2^s(x)dx \leq I_1 (s) Q(s) ,
\end{align}
%
where
%
\begin{equation*}
I_1 (s):= \sup_y \int_{\mathbb{R}^3} \left(\frac{1}{|x-y|^2}+ \frac{1}{|x-T_s (y)|^2} \right) )\rho_2^s (x)dx .
\end{equation*}
%
The integral corresponding to the first term in $I_1 (s)$ can be estimated for all $y$ in $\R^3$ and for all $s\geq 0$ as follows:
\begin{align*}
\int_{\mathbb{R}^3} \frac{1}{|x-y|^2}\rho_2^s(x)dx &\leq \int_{B(y,1)}\frac{1}{|x-y|^2} \rho_2^s(x) dx + \int_{B(y,1)^c} \rho_2^s(dx) \\
& \leq \left(\int_{B(y,1)} \frac{1}{|x-y|^{2p'}}\right)^{1/p'} \| \rho_2^s\|_{L^p(\R^3)} + \|\rho_2^s\|_{L^1(\R^3)} ,
\end{align*}
%
by H\"older's inequality, with $p'= p/(p-1)<3/2$, so that $z\mapsto z^{2-2p'}$ is integrable near the origin, and therefore
%
\begin{align*}
\int_{\mathbb{R}^3} \frac{1}{|x-y|^2}\rho_2^s(x)dx
&\leq C\| \rho_2^s\|_{L^p(\R^3)} \int_0^1 z^{2-2p'} dz + \|\rho_2^s\|_{L^1(\R^3)} \leq C(\| \rho_2^s\|_{L^p(\R^3)}+ \|\rho_2^s\|_{L^1(\R^3)}).
\end{align*}
Proceeding similarly for the second term in $I_1 (s)$, we arrive at
\begin{equation}
\label{borneI1}
I_1 (s) \leq C (\| \rho_2^s\|_{L^p(\R^3)}+ \|\rho_2^s\|_{L^1(\R^3)}),
\end{equation}
for any $s\geq 0$.
Combining \eqref{ineq-Q}, \eqref{triple} and \eqref{borneI1} we get that for any $t\geq 0$,
\begin{align*}
Q(t) \leq Q(0) + \int_0^t \text{Lip}(u_1(s)) Q(s) ds+ \int_0^t C (\| \rho_2^s\|_{L^p(\R^3)}+ \|\rho_2^s\|_{L^1(\R^3)})Q(s) ds .
\end{align*}
By using $\|\rho_i^t\|_{L^q(\R^3)}=\|\rho_i^0\|_{L^q(\R^3)}$ for $q=1,p$ and $i=1,2$, \eqref{cotestokes} and a Gronwall-type argument, and combining this with \eqref{prem} and \eqref{prem-zero} we arrive at \eqref{sta-geq-3}.
This concludes the proof of
Theorem \ref{thm} in the case where $p> 3$.
\subsection{Case where \texorpdfstring{$p= 3$}{p=3}}
Let us now deal with the critical case where $p= 3$.
By using \eqref{esti-arep} and Proposition \ref{prop_log_lip_u},
we obtain that for any $x$ in $\R^3$, for any $t\geq 0$,
%
\begin{align*}
& |X_1(t,0,T_0 (x))-X_2(t,0,x)| \leq |T_0 (x)-x|
\\ & \quad + C \int_0^t |X_1(s,0,T_0 (x)) - X_2(s,0,x) |(1+\ln_{-}(|X_1(s,0,T_0 (x)) - X_2(s,0,x) |) ds\\
&\quad +
\int_0^t |u_1(s,X_2(s,0,x) - u_2(s,X_2(s,0,x) | ds ,
\end{align*}
where $C=C(\|\rho_1^0\|_{L^3 (\R^3) \cap L^1 (\R^3)} )>0$.
By integration over $\R^3$ with respect to the measure $\rho_2^0(dx)$, we deduce by
using \eqref{pousse} and \eqref{reexpQ} that for any $t\geq 0$,
%
\begin{gather*}
Q(t) \leq Q(0)
\\ + C \int_0^t\int_{\mathbb{R}^3} |X_1(s,0,T_0 (x)) - X_2(s,0,x) |(1+\ln_{-}(|X_1(s,0,T_0 (x)) - X_2(s,0,x) |) \rho_2^0(dx) ds
\\ + \int_0^t \int_{\mathbb{R}^3}|u_1(s,X_2(s,0,x) - u_2(s,X_2(s,0,x) |\rho_2^0(dx) ds .
\end{gather*}
%
Since $z\mapsto z(1+\ln_{-}z)$ is concave for $z\geq 0$, it follows from Jensen's inequality
and \eqref{reexpQ} that for any $s\geq 0$,
%
\begin{align*}
\int_{\mathbb{R}^3} |X_1(s,0,T_0 (x)) - X_2(s,0,x) |(1+\ln_{-}(|X_1(s,0,T_0 (x)) - X_2(s,0,x) |) \rho_2^0(dx)
\\ \quad \leq Q(s)(1+\ln_{-}Q(s))
.
\end{align*}
%
Therefore, for any $t\geq 0$,
%
\begin{align}\label{ineq-Q_p=3}
Q(t)
&\leq Q(0)+ C \int_0^t Q(s)(1+\ln_{-}Q(s)) ds+ \int_0^t \int_{\mathbb{R}^3}|u_1(s,x) - u_2(s,x) |\rho_2^s(x)dx ds .
\end{align}
To bound the last term above, we establish the following result.
\begin{lemme}\label{LLint}
For any $t\geq 0$,
%
\begin{equation}
\label{esti-LLint}
\int|u_1(t,x)-u_2(t,x)|\rho_2(t,x)dx \leq C(\|\rho_2^t\|_{L^1\cap L^3}) Q(t) (1+\ln_{-} Q(t)).
\end{equation}
\end{lemme}
\begin{proof}
By \eqref{demen} and Fubini's principle, we obtain that
for any $t\geq 0$,
\begin{align*}
&\int|u_1(t,x)-u_2(t,x)|\rho_2(t,x)dx \leq \int \Big( \int |\mathcal{U}(x-T_t (y))-\mathcal{U}(x-y)| \rho_2(t,x) dx \Big) \rho_2^t(y) dy .
\end{align*}
%
Now we split the inner integral into three parts, corresponding to the domains
$B(T_t (y),\epsilon_y^t)$, $B(T_t (y),\epsilon_y^t)$ and $B(y,\epsilon_y^t)^c \cap B(T_t (y),\epsilon_y^t)^c$, where $\epsilon_y^t =|T_t (y) - y|$.
Observing that the Oseen tensor
$\mathcal{U}$ satisfies that there exists $C>0$ such that for all $x $ in $\R^3 \setminus \{ 0\}$,
$ \left|\mathcal{U}(x) \right|\leq C\frac{1}{|x|} $,
and using \eqref{Lip-prop} we deduce from \eqref{demen} that for any $t\geq 0$,
\begin{align*}
&\int|u_1(t,x)-u_2(t,x)|\rho_2(t,x)dx
\leq C \int \int_{B(T_t (y),\epsilon_y^t)} \left(\frac{1}{|x-T_t (y) |}+\frac{1}{|x-y|} \right)\rho_2(t,x) dx \rho_2^t(y) dy\\
&\quad + C \int \int_{B(y,\epsilon_y^t)} \left(\frac{1}{|x-T_t (y)|}+\frac{1}{|x-y|} \right)\rho_2(t,x) dx \rho_2^t(y) dy\\
&\quad +C \int \epsilon_y^t \int_{B(y,\epsilon_y^t)^c \cap B(T_t (y),\epsilon_y^t)^c} \left(\frac{1}{|x-T_t (y)|^2}+\frac{1}{|x-y|^2} \right)\rho_2(t,x) dx \rho_2^t(y) dy\\
&=: I_1+I_2+I_3 .
\end{align*}
Regarding $I_1$ and $I_2$ we can proceed in the same way as in the proof of Proposition \ref{prop_log_lip_u} and obtain
\begin{equation*}
I_1 +I_2\leq C(\|\rho_2\|_{L^3 (\R^3)}) Q(t) .
\end{equation*}
%
For $I_3$ we distinguish two cases. First if $\epsilon_y^t=|y-T_t (y)|\geq 1 $ we get $$
\int_{B(y,\epsilon_y^t)^c \cap B(T_t (y),\epsilon_y^t)^c} \left(\frac{1}{|x-T_t (y)|^2}+\frac{1}{|x-y|^2} \right)\rho_2(t,x) dx \leq \|\rho_2^t\|_1.
$$
On the other hand, if $\epsilon_y^t=|y-T_t (y)|\leq 1 $, then
\begin{align*}
&\int_{B(y,\epsilon_y^t)^c \cap B(T_t (y),\epsilon_y^t)^c} \left(\frac{1}{|x-T_t (y)|^2}+\frac{1}{|x-y|^2} \right)\rho_2(t,x) dx\\
&\leq \int_{B(T_t (y),\epsilon_y^t)^c } \frac{1}{|x-T_t (y)|^2}\rho_2(t,x)dx + \int_{B(y,\epsilon_y^t)^c } \frac{1}{|x-y|^2}\rho_2(t,x)dx\\
& \leq \int_{B(T_t (y),2)\setminus B(T_t (y),\epsilon_y^t) } \frac{1}{|x-T_t (y)|^2}\rho_2(t,x)dx + \int_{B(y,2)\setminus B(y,\epsilon_y^t) } \frac{1}{|x-y|^2}\rho_2(t,x)dx+ \|\rho_2\|_1\\
&\leq C(\|\rho_2\|_{L^1\cap L^3})(1+ |\log|T_t (y) -y||) .
\end{align*}
Gathering all the estimates we get
$$
\int|u_1(t,x)-u_2(t,x)|\rho_2(t,x)dx \leq C(\|\rho_2\|_{L^1\cap L^3}) \int |T_t (y) - y |(1+ \ln_{-}|T_t (y) - y |) \rho_2^t(y) dy.
$$
Using the fact that $z\mapsto z(1+\ln_{-} z)$ is concave for $z\geq 0$ we get
\eqref{esti-LLint} and this concludes the proof of Lemma \ref{LLint}.
\end{proof}
We have for any $t\geq 0$,
\begin{equation*}
Q(t) \leq Q(0)+ C(\|\rho_i\|_{L^\infty(0,t, L^1\cap L^3)})\int_0^t Q(s)(1+\ln_{-} Q(s)) ds.
\end{equation*}
Thanks to the Osgood theorem, see Lemma \ref{Osgood} with $\eta(z)=z(1+\ln_{-}(z)) $, we deduce \eqref{sta-geq-critik}.
This concludes the proof of Theorem \ref{thm}.
\section{Proof of Theorem \ref{thm0}}
\label{sec-proof-princi}
The proof of existence can be done by a density argument using the well known results for the existence of a solution in the case of regular data, see \cite{Hofer} and \cite{Mecherbet2}, however we propose below a self-contained proof based on a fixed-point argument.
\subsection{Iterative scheme}
Let $p\geq 3$, $\rho_0 \in E_p$ and consider the sequence $(u^n,\rho^n)$ where
$$ u^n(t) \in
\left\{
\begin{array}{lcl}
W^{2,p}(\R^3) & \text{ if } & p>3 ,\\
\dot{W}^{2,3}(\R^3)\cap \dot{W}^{1,3}(\R^3) \cap \underset{q\in(3,+\infty]}{\bigcap} W^{1,q}(\R^3) & \text{ if } & p=3 ,
\end{array}
\right.
$$
the unique solution to the Stokes equation
\begin{equation*}
\begin{array}{c}
-\Delta u^n+ \nabla p^n= - \rho^n e_3, \:\div(u^n)=0, \text{ on } \R^3 ,\\
\underset{|x|\to \infty}{\lim} |u^n| = 0 ,
\end{array}
\end{equation*}
and $\rho^{n+1}$ the unique solution to the transport equation
\begin{equation*}
\left\{
\begin{array}{rcl}
\partial_t \rho^{n+1}+\div(\rho^{n+1} u^n)&=& 0, \text{ on } \R^3 ,\\
\rho^{n+1}(0,\cdot)&=& \rho_0,
\end{array}\right.
\end{equation*}
with $\rho^0=\rho_0$. Since $u^n$ is divergence free and the associated flow $Xn$ is well defined, classical considerations for the transport equation ensure that $\rho^n \in C([0,+\infty), L^1(\R^3)\cap L^p(\R^3)) $ for each $n$ and we have for any $t$,
\begin{equation}\label{borne_rho_n}
\|\rho^{n+1}(t)\|_{L^1(\R^3)}=\|\rho_0\|_{L^1(\R^3)}, \: \|\rho^{n+1}(t)\|_{L^p(\R^3)}=\|\rho_0\|_{L^p(\R^3)},
\end{equation}
Hence, we get the following uniform bounds of the velocity:
%
\begin{equation}\label{borne_u_n}
\|u^n(t)\|_{W^{2,p}(\R^3)}\leq C \|\rho^n(t)\|_{L^1(\R^3)\cap L^p(\R^3)} \leq C \| \rho_0\|_{L^1(\R^3)\cap L^p(\R^3)}
\end{equation}
in the case where $p>3$
and
\begin{equation}\label{borne_u_n_bis}
\text{Log-Lip}(u^n(t))+ \|\nabla u^n(t)\|_{W^{1,3}(\R^3)}+ \| u^n(t)\|_{W^{1,q}(\R^3)}\leq C \|\rho^n(t)\|_{L^1(\R^3)\cap L^3(\R^3)} \leq C \| \rho_0\|_{L^1(\R^3)\cap L^3(\R^3)} ,
\end{equation}
in the case $p=3$ with $q\in(3,+\infty]$.
Moreover for $p\geq3$, we have, for all $0 \leq t \leq t' \leq T$,
\begin{align}\label{continuite_rho_n}
W_1(\rho^{n+1}(t),\rho^{n+1}(t')) &\leq \int_{\R^3}|X^n(t',0,x)-X^n(t,0,x)| \rho_0(dx) \notag\\
&\leq (t'-t) \|u^n\|_{L^\infty(0,T;L^\infty(\R^3))} \|\rho_0\|_{L^1(\R^3)} .
\end{align}
This shows that
$$\rho^{n} \in C([0,T],\mathcal{P}(\R^3)\cap L^1(\R^3)\cap L^p(\R^3)) \text{ uniformly in n for any }T>0.$$
Using the same arguments as in \eqref{demen}, We emphasize the following estimates for any compact set $K$ of $\R^3$
\begin{eqnarray}
\|u_n(t)-u_n(t')\|_{W^{1,\infty}(K)}\leq C_K W_1(\rho^n(t),\rho^n(t')),& \text{ if } p>3,\label{equiconinuity_u_n(t),p>3}\\
\|u_n(t)-u_n(t')\|_{L^{\infty}(K)}\leq C_K W_1(\rho^n(t),\rho^n(t')),& \text{ if } p=3 \label{equiconinuity_u_n(t),p=3} .
\end{eqnarray}
\subsection{Convergence of \texorpdfstring{$(\rho^n)_{n}$}{(rhon)n}}
We introduce the following space for $M>0$
\begin{equation*}
B_{T,M}=\{\rho \in C([0,T], \mathcal{P}(\R^3)), \quad
\underset{[0,T]}{\sup}W_1(\rho(t),\rho_0) \leq M \} .
\end{equation*}
which is complete for the metric $d(\rho_1,\rho_2):=\underset{[0,T]}{\sup}W_1(\rho_1(t),\rho_2(t)) $.
Since $\rho^{n+1}(t):=X_n(t,0,\cdot)\#\rho_0 $ we have
\begin{align*}
W_1(\rho^n(t),\rho_0) \leq \int_{\R}|X^n(t,0,x)-x|\rho_0(dy) \leq T \|\rho_0\|_{L^1} \|u^n \|_{L^\infty(0,T;L^\infty)} ,
\end{align*}
which shows that $(\rho_n)_n$ lies in $B_{T,M}$ for an adequate choice of $M>0$. We aim to show that $(\rho_n)$ is a Cauchy sequence in $B_{T,M}$ for $T>0$ small enough. We present below the argument by distinguishing the case $p=3$ from the case $p>3$.
\subsection*{Case where $p>3$.}
For $p>3$, the stability estimates \eqref{ineq-Q} ensure that
$$
W_1(\rho^n(t),\rho^{n+1}(t))\leq \left(\int_0^t \int_{\R^3}|u^{n-1}(s,x)-u^{n}(s,x)|\rho^{n}(s,x)dx ds \right)e^{M_n t} ,
$$
with $M_n=\left(\underset{0 \leq s \leq t }{\sup}\text{Lip}(u^n(s))\right) $. On the other hand, performing an estimate as in \eqref{triple} yields
$$
\int_{\R^3}|u^{n-1}(s,x)-u^{n}(s,x)|\rho^{n}(s,x)dx\leq C(\|\rho^n\|_{L^1(\R^3)\cap L^p(\R^3)})W_1(\rho^{n-1}(s),\rho^n(s)) .
$$
Hence for all $t\in[0,T]$ we get using \eqref{borne_rho_n} and $M_n=\underset{0 \leq s \leq T }{\sup}\text{Lip}(u^n(s))\leq M$ thanks to \eqref{borne_u_n}
\begin{equation*}
\underset{t\in[0,T]}{\sup} W_1(\rho^n,\rho^{n+1})\leq C Te^{MT} \underset{t\in[0,T]}{\sup} W_1(\rho^n,\rho^{n-1}) .
\end{equation*}
This shows that for $T>0$ small enough, $\rho_n$ is a Cauchy sequence.
\subsection*{Case where $p=3$.}
For $p=3$ we use estimates \eqref{ineq-Q_p=3} to get for all $n,k\in \mathbb{N}$
\begin{multline}\label{eq:stab1_n,k}
W_1(\rho^{n+k+1}(t),\rho^{k+1}(t)) \leq C^{n,k} \int_0^t \eta \left(W_1(\rho^{n+k+1}(s),\rho^{k+1}(s)) \right) ds\\
+ \int_0^t \int_{\R^3} |u^{n+k}(s,x)-u^{k}(s,x)|\rho^{n+k}(s,x) dx ds ,
\end{multline}
with $\eta(z)=z(1+\ln_{-}(z)) $ for $z\geq 0$ and $$C^{n,k}=\underset{s\in[0,T]}{\sup}\text{Log-Lip}(u^{n+k}(s)) \leq C \underset{s\in[0,T]}{\sup}\| \rho^{n+k}(s)\|_{L^1(\R^3)\cap L^3(\R^3)} \leq C ,$$ thanks to \eqref{esti-uli} and the uniform bounds on $(\rho^n)_{n}$.
On the one hand, using the uniform bounds of $(u_n)_n$ in $ C([0,T],L^\infty)$ and $(\rho_n)_n$ in $ C([0,T],L^{1})$ and the fact that $\eta(z)\leq \max(z,1) $ we have from $\eqref{eq:stab1_n,k} $
$$
\underset{0 \leq s \leq t}{\sup} W_1(\rho^{n+k+1}(t),\rho^{k+1}(t)) \leq C t \left(1+ \underset{0 \leq s \leq t}{\sup}W_1(\rho^{n+k+1}(t),\rho^{k+1}(t)) \right) + t C ,
$$
which shows that, for $t$ small enough, the sequence $\left(\underset{0 \leq s \leq t}{\sup}W_1(\rho^{n+k}(t),\rho^{k}(t))\right)_{n,k}$ is uniformly bounded.
On the other hand, thanks to Lemma \ref{LLint} we have
\begin{equation}\label{eq:stab2_n,k}
\int_0^t \int_{\R^3} |u^{n+k}(s,x)-u^{k}(s,x)|\rho^{n+k}(s,x) dx ds\leq C \int_0^t \eta \left(W_1(\rho^{n+k}(s),\rho^{k}(s)) \right)ds,
\end{equation}
where we used again \eqref{borne_rho_n}, \eqref{borne_u_n} to get a constant independent of $n,k$. Hence gathering inequalities \eqref{eq:stab1_n,k} and \eqref{eq:stab2_n,k} and setting $$f_{k}(t):=\underset{n}{\sup} \underset{0 \leq s \leq t}{\sup} W_1(\rho^{n+k}(s),\rho^{k}(s)) , $$
we get for all $T\geq 0$
$$
f_{k+1}(T) \leq C \int_0^T \eta(f_{k+1}(s))ds+ C\int_0^T \eta(f_k(s)) ds,
$$
where we used the fact that $z \mapsto \eta(z)$ is a non-decreasing function. By setting $\tilde{f}(t):=\underset{k\to \infty}{\limsup }f_k(t) $ we have
$$
\tilde{f}(T) \leq 2 C \int_0^T \eta(\tilde{f}(s)) ds.
$$
We conclude thanks to Lemma \ref{Osgood} that $\tilde{f}=0 $ a.e on $[0,T]$, which shows that $ \rho^n$ is a Cauchy sequence.
\subsection*{Conclusion.}
Hence for $p\geq 3$, there exists a limit measure $\rho \in B_{T,M}\subset C([0,T],\mathcal{P}(\R^3))$ such that $$\underset{t\in[0,T]}{\sup} W_1(\rho^n,\rho)\to 0. $$
\subsection{Convergence of \texorpdfstring{$(u^n)_n$}{(un)n} and \texorpdfstring{$(X^n)_n$}{(Xn)n}} The second step is to extract a converging subsequence of $(u^n)_n$ by distinguishing again the case where $p>3$ from the case where $p=3$.
\subsection*{Case where $p>3$.}
Since for each compact set $K$ of $\R^3$ the sequence $ t \mapsto u_n(t)$ is equicontinuous in $W^{1,1}(K) $ thanks to \eqref{equiconinuity_u_n(t),p>3}, \eqref{continuite_rho_n} and $x\mapsto u_n(t,x)$ relatively compact in $ C^1(K) $ (using Ascoli with the equicontinuity of $x\mapsto u_n(t,x)$ in $ C^1(K)$), we get using again Ascoli the existence of $u\in C([0,T], C^1(\R^3))$ such that $u_n \to u $ in $C([0,T], C^1(K))$ up to a subsequence for each compact $K$.
Regarding the flow $X^n$, one can show that for each $0 \leq s\leq T $, $X^n(\cdot,s,\cdot) $ is uniformly bounded in $ C([0,T], C^{1}(K))$ on each compact set $K\in \R^3$, equicontinuous in $ C([0,T], C^{1}(\R^3))$ and that for each $t\in[0,T]$, $X^n(t,s,\cdot)$ is relatively compact in $ C^{1}(K)$. Indeed we have the following bounds for each $x,y \in \R^3 $, $t,t'\in[0,T]$, $|t-t'|<1$.
\begin{gather}\label{borne_uniforme_x_t}
|X_n(t,s,x)| \leq |x|+ |t-s| \|u^n\|_{ C(0,T; C(\R^3))},\: |\nabla X_n(t,s,x)|\leq e^{L(t-s)} ,
\\ \label{equicontinuite_x}
|\nabla X^n(t,s,x)-\nabla X^n(t,s,y)| \leq |x-y|^\mu e^{L |t-s|} ,
\\ \label{equicontinuite_t}
\|X^n(t,s,\cdot)-X^n(t',s,\cdot)\|_{ C^{1}(\R^3)} \leq Me^{LT}|t-t'|^\mu .
\end{gather}
From \eqref{equicontinuite_t} we get that $t\mapsto X^n(t,s,\cdot) $ is equicontinuous with values in $ C^{1,\mu}$,\eqref{borne_uniforme_x_t} and \eqref{equicontinuite_x} ensure that for each $t$, $X^n(t,s,\cdot) $ is relatively compact in $ C^1(K)$. Hence, since $X^n$ is uniformly bounded in $ C([0,T], C^{1,\mu}(K))$ thanks to \eqref{borne_uniforme_x_t}, then using the Ascoli theorem, there exists $X(\cdot,s,\cdot)\in C([0,T], C^{1,\mu}(\R^3)) $ such that $X^n(\cdot,s,\cdot) \to X(\cdot,s,\cdot) $ in $ C([0,T], C^1(K)) $ for each compact $K$. In particular passing in the limit we have for all $s,t\in[0,T]$,
$X(t,s,\cdot) \circ X(s,t,\cdot)= Id$,
and using the convergence of $u^n$ to $u$ in $ C([0,T], C^1(K))$ for each compact $K$ we get that $X$ satisfies \eqref{cflow-def}.
\subsection*{Case where $p=3$.}
We proceed analogously using \eqref{equiconinuity_u_n(t),p=3} and show that there exists $u\in C([0,T], C(\R^3)) $ such that for each compact set $K$ we have $u^n\to u$ in $ C([0,T]; C(K))$.
Regarding the flow $X^n$,
$X^n(\cdot,s,\cdot)-Id$ is uniformly bounded in $ C([0,T], C(\R^3))$ and in particular in $ C([0,T], C(K))$ for each compact set. Now, let $x,y \in \R^3$, $|x-y|\leq 1/2$ and $T^n$ the maximal time such that $|X^n(t,s,x)-X^n(t,s,y)|<3/4 $
for $t,s \in[0,T]$, we have
\begin{align*}
|X^n(t,s,x)-X^n(t,s,y)|&\leq |x-y|+\left(\underset{\tau\in[0,T]}{\sup}\text{Log-Lip}(u^n(\tau)) \right) \\
&\times \int_s^t |X^n(t,s,x)-X^n(t,s,y)|(1+\ln_{-} |X^n(t,s,x)-X^n(t,s,y)|),\end{align*}
where $\left(\underset{\tau\in[0,T]}{\sup}\text{Log-Lip}(u^n(\tau)) \right) $ is uniformly bounded thanks to \eqref{borne_u_n_bis}. Using Lemma \ref{Osgood} with $\eta(z):=z(1+\ln_{-}(z))$, we get for some constant $\bar C$ for any $t\leq T^n$, $|x-y|\leq1/2$
\begin{equation}\label{equicontinuite_x_3}
|X^n(t,s,x)-X^n(t,s,y)| \leq \bar{C}|x-y|^{e^{-CT}},
\end{equation}
which shows that the maximal time $T^n$ can be taken independent of $n$ and $T^n=T$ for $T$ small enough. hence, $X_n(t,s,\cdot)$ is relatively compact in $ C(K)$ for each $t,s \in [0,T]$. Moreover
$$
\|X^n(t,s,\cdot)-X^n(t',s,\cdot)\|_{ C(K)}\leq |t-t'| \| u^n\|_{ C(0,T;L^\infty(\R^3))},
$$
which ensures that $t\mapsto X^n(t,s,\cdot) $ is equicontinuous for each $s\in[0,T]$. Using again the Ascoli theorem this allows us to extract a converging subsequence in $ C([0,T], C(K))$. This allows us to construct $X\in C([0,T], C(\R^3))$ and by passing in the limit in \eqref{equicontinuite_x_3} pointwisely we get for all $x,y \in \R^3$, for all $t,s$ in $[0,T]$,
$$
|X(t,s,x)-X(t,s,y)| \leq \bar{C}|x-y|^{e^{-CT}},
$$
and analogously to the case $p>3$ we have
for $s,t\in[0,T]$,
$
X(t,s,\cdot) \circ X(s,t,\cdot)= Id.
$
\subsection{Existence and uniqueness of the solution}
Convergence of $\rho^n$ and $u^n$ (up to a subsequence) allows to pass in the limit in both the Stokes and transport equations and show that $(u,\rho)$ satisfies the transport-Stokes system weakly. Moreover, using the fact that $X^n $ satisfies
$$
\rho^{n+1}(t,\cdot)= X^n(t,0,\cdot)\#\rho_0,
$$
one can pass in the limit in both sides thanks to the strong convergence of $X^n$ up to a subsequence and the convergence of $\rho^n$ using the Wasserstein metric. Indeed, we get for all $\psi\in C_b(\R^3)$
$$
\int \psi(x) \rho(t,x) dx =\int \psi(X(0,t,x)) \rho_0(x) dx,
$$
which means
$$
\rho(t,\cdot)= X(t,0,\cdot)\# \rho_0,
$$
and thanks to the fact that $X\in C([0,T], C(\R^3))$ we recover that $\rho \in C([0,T], L^p(\R^3) \cap L^1(\R^3)) $.
From this it follows that $u \in C([0,T], W^{2,p}(\R^3)) $ with the bounds \eqref{reg_u,p} (resp. $\nabla u \in C([0,T], W^{1,3}(\R^3))$, $u\in C([0,T],W^{1,q}(\R^3)) $ for $q\in]3,+\infty]$ with the bounds \eqref{reg_u,p_bis} ).
Global in time existence is ensured thanks to the uniform bounds of $\rho(t)$ in $L^p(\R^3) \cap L^1(\R^3)$ while uniqueness is ensured thanks to Theorem \ref{thm}. This concludes the proof of Theorem \ref{thm0}.
\section{Proof of Proposition \ref{apana}}\label{sec-apana}
This section is devoted to the proof of Proposition \ref{apana}.
In the course of the proof we shall use a technical lemma, see Lemma \ref{lemme_Miot_general}, which is proved in the next section.
\subsection{A first few basic estimates}
We gather below some useful estimates.
First, by \eqref{cflow-def}, for all $x$ and $y$ in $\R^3$,
%
\begin{equation}
\label{coince}
|x-y|e^{-\int_0^t \text{Lip}(u(s,\cdot))ds}\leq |X(t,x)-X(t,y)| \leq |x-y| e^{\int_0^t \text{Lip}(u(s,\cdot))ds}.
\end{equation}
Thus, for $T>0$ small enough,
this implies that there exists
$\lambda \in (1,3/2]$ such that
%
\begin{equation}\label{aveclambda}
\frac{1}{\lambda} \leq \frac{|X(t,x)-X(t,y)}{|x-y|} \leq \lambda,\: \forall \, x \neq y,\: t\in[0,T] .
\end{equation}
Next, by splitting the integral into two parts, distinguishing the cases where $|x-z|\leq 1$ and where $|x-z|> 1$, and using H\"older's inequality for the first case with the observation that $p>3$, we obtain the existence of
$L=L(p)>0$ such that for all $0\leq s\leq 2$,
\begin{equation}\label{fors}
\underset{x \in \R^3}{\sup} \int \frac{\rho_0(dz)}{|x-z|^s} \leq L \|\rho_0\|_{L^1\cap L^p} .
\end{equation}
Finally, from the definition of the the Oseen tensor in
\eqref{Oseen} we deduce the following property.
\begin{lemme}\label{lemme_oseen_rec}
There exists
$K>0$ such that for all $x$ in $\R^3\setminus \{0\}$ and all multi-index $\alpha \in \mathbb{N}^3$
%
\begin{equation}
\label{alpha-U}
|\partial^\alpha \mathcal{U}(x)| \leq \frac{K^{|\alpha|} |\alpha|!}{|x|^{1+|\alpha|}} .
\end{equation}
\end{lemme}
\begin{proof}
Let $n\geq 1$ and $\alpha \in \mathbb{N}^3$ a multi-index such that $|\alpha|=n $. We consider first the term $u(x):=\frac{1}{|x|}$ in the Oseen tensor. Since \begin{equation} \label{marre}
\partial_{x_i} \frac{1}{|x|^p}=-p \frac{x_i}{|x|^{p+2}} ,
\end{equation}
one can show by induction that for all $|\alpha|=n$ there exists polynomials $P_k^\alpha$, $k=\lceil\frac{n}{2}\rceil,\cdots, n$ such that
\begin{equation}\label{eq:reccurence}
\partial^\alpha u(x)=\underset{k=\lceil \frac{n}{2}\rceil }{\overset{n}{\sum}} \frac{P_k^\alpha(x)}{|x|^{1+2k}},
\end{equation}
where $P_k^\alpha$ is a a sum of monomials in $x_1,x_2,x_3$ such that each monomial is of degree
$ -n+2k \geq 0 $ and $\lceil \frac{n}{2}\rceil$ denotes the ceiling function of $\frac{n}{2}$.
Indeed the induction formula is satisfied for $n=1$ as a consequence
of \eqref{marre} with $p=1$.
Now we assume that \eqref{eq:reccurence} holds true up to $n$, and we observe that, for any $1 \leq i \leq 3$ and for any $\alpha \in \mathbb{N}^3$ with $|\alpha|=n$, by the induction assumption \eqref{eq:reccurence}, Leibniz' rule and \eqref{marre},
\begin{align}\label{eq:reccurence_derivation}
\partial_i \partial^\alpha u(x)&= - \underset{k=\lceil \frac{n}{2}\rceil}{\overset{n}{\sum}} (1+2k)\frac{x_i P_k^\alpha(x)}{|x|^{3+2k}}+\underset{k=\lceil\frac{n}{2}\rceil}{\overset{n}{\sum}} \frac{\partial_i P_k^\alpha(x)}{|x|^{1+2k}} \notag \\
&= \frac{\partial_i P^\alpha_{\lceil\frac{n}{2}\rceil}(x)}{|x|^{1+2\lceil\frac{n}{2}\rceil}}
+ \underset{k=\lceil\frac{n}{2}\rceil+1}{\overset{n}{\sum}} \frac{\partial_i P^\alpha_k(x)-(2k-1)x_i P^\alpha_{k-1}(x)}{|x|^{1+2k}}
- (1+2n)\frac{x_i P_n^\alpha(x)}{|x|^{3+2n}} .
\end{align}
Since $\lceil \frac{n}{2} \rceil+1= \lceil \frac{n+1}{2} \rceil$ for $n\geq0$, all the terms in \eqref{eq:reccurence_derivation} except the first one correspond to a term of the form $\frac{P_k^\beta}{|x|^{1+2k}} $ with $\beta=e_i+\alpha$ and $k\in \{\lceil \frac{n+1}{2} \rceil,\dots,n+1 \}$. For the first term we distinguish two cases
\begin{itemize}
\item If $n=2p$, the polynomial $P_{\lceil \frac{n}{2} \rceil}^\alpha$ is of degree $-n+ 2\lceil \frac{n}{2} \rceil=-n+2p=0 $ hence $\partial_i P_{\lceil \frac{n}{2} \rceil}^\alpha=0 $.
\item If $n=2p+1$, then $\lceil \frac{n}{2} \rceil=p+1=\lceil \frac{2p+2}{2} \rceil=\lceil \frac{n+1}{2} \rceil$ and hence the polynomial $\partial_i P_{\lceil \frac{n}{2} \rceil}^\alpha$ corresponds to a term of the form $P_k^\beta$ with $k=\lceil \frac{n+1}{2} \rceil$.
\end{itemize}
Now if we set $D_n$ the { maximal number of monomials appearing in} \eqref{eq:reccurence} for all $|\alpha|=n$ we have using \eqref{eq:reccurence_derivation}
$$
D_{n+1} \leq D_n + 2 D_n = 3 D_n \leq 3^{n} D_1 .
$$
If we set $C_n$ the largest coefficient (in absolute value) appearing in \eqref{eq:reccurence} for all $\alpha \in \mathbb{N}^3$ with $|\alpha|=n$, we have, by using \eqref{eq:reccurence_derivation}, that
$$
C_{n+1} \leq (1+2n) C_{n} \leq \underset{k=1}{\overset{n}{\prod}} (2k+1) C_1 .
$$
We observe that $C_1=1$, that $D_1=1$ and that
$$
|\partial^\alpha \frac{1}{|x|} |\leq \frac{D_n C_n}{|x|^{1+n}} \leq \frac{3^{n} 2^{3n+2}(n!)} {|x|^{1+n}},
$$
with $n=|\alpha|$. Above we used that $$(2n+1)!=\left(\underset{k=1}{\overset{n}{\prod}} (2k+1)\right)\left(\underset{k=1}{\overset{n}{\prod}} (2k)\right)=2^n n! \left(\underset{k=1}{\overset{n}{\prod}} (2k+1)\right),
$$
and
$$
(2n+1)! = (2n+1) (2n)! \leq n(2n+1) 2^{n+1} (n!)^2 \leq 2^{3n+1} 2^{n+1} (n!)^2 .
$$
For the second term in the Oseen tensor $\frac{x_ix_j}{|x|^3} $ one can show analogously that the derivatives of $u(x)=\frac{1}{|x|^3} $ are of the form
$$
\partial^\alpha u(x)= \underset{k=\lceil \frac{n}{2}\rceil}{\overset{n}{\sum}}\frac{P_k^\alpha(x)}{|x|^{2k+3}},
$$
where in $P_k^\alpha$ is a sum of monomials of degree $ -n+2k $.
Using the same arguments as above,
we get that for $\alpha \in \mathbb{N}^3$
with $n=|\alpha|$
$$
\left|\partial^\alpha \frac{1}{|x|^3}\right|\leq \frac{3^{n} 2^{3n+5} }{|x|^{3+n}} (n+1)! \leq \frac{3^{11n}}{|x|^{3+n}} n! .
$$
we conclude by using Leibniz formula
\begin{align*}
\left|\partial^\alpha \frac{x_i x_j}{|x|^3} \right|&= \underset{\beta+\gamma=\alpha}{\sum} \begin{pmatrix} \beta \\ \alpha \end{pmatrix} \partial^\beta \left( \frac{1}{|x|^3}\right) \partial^{\gamma} (x_ix_j) \\
&\leq \underset{|\alpha|-2\leq |\beta| \leq |\alpha|}{\sum} \begin{pmatrix} \beta \\ \alpha \end{pmatrix} \frac{3^{11|\beta|}}{|x|^{3+|\beta|}}|\beta|! 2|x|^{2-|\alpha|+|\beta|}\\
&\leq \frac{3^{13|\alpha|}}{|x|^{1+|\alpha|}}|\alpha|!.
\end{align*}
where we used $ \underset{0\leq |\beta| \leq |\alpha|}{\sum} \begin{pmatrix} \beta \\ \alpha \end{pmatrix}=2^{|\alpha|} $.
This concludes the proof of Lemma \ref{lemme_oseen_rec}.
\end{proof}
\subsection{The case where \texorpdfstring{$n=1$}{n=1}}
From \eqref{cflow-def}, we deduce that
\begin{align*}
\|\partial_t X(t,\cdot)\|_\infty \leq \|u\|_\infty \leq C \|\rho_0\| \quad \text{ and }
\quad
\|\nabla \partial_t X(t,\cdot)\|_\infty \leq \|\nabla u\|_\infty \|\nabla X \|_\infty \leq C \|\rho_0\| \lambda ,
\\ \quad \text{ and }
\quad
|\nabla \partial_t X(t,x)-\nabla \partial_t X(t,y)| \leq [\nabla u]_{0,\mu}|X(t,x)-X(t,y)|^\mu \|\nabla X\|_\infty +
|\nabla X(t,x)-\nabla X(t,y)| \|\nabla u\|_\infty .
\end{align*}
Therefore by the
Gronwall lemma, we arrive at
$$
[\nabla \partial_t X]_{0,\mu} \leq C\|u\|_{2,p}\lambda^\mu \|\nabla X\|_\infty e^{\|\nabla u\|_\infty t}.
$$
This entails that the induction assumption is satisfied
for $n=1$, with some $C_0=C_0(\|\rho_0\|, \lambda, T,\mu)$ such that
\begin{equation}
\|\partial_t X(t,\cdot)\|_{1,\mu} \leq \frac{C_0}{2}.
\end{equation}
\subsection{Beginning of the iteration}
Assume that \eqref{rec_formule} is satisfied up to $n \in \N^*$, We need to prove by induction the bound \eqref{rec_formule} for the $L^\infty$ estimate $\|\partial_t^{n+1} X(t,\cdot)\|_\infty$, the Lipschitz estimate $\|\nabla \partial_t^{n+1} X(t,\cdot)\|_\infty$ and the H\"older estimate $[\nabla \partial_t^{n+1} X(t,\cdot)]_{0,\mu}$. Hence we split the proof into three parts.
From \eqref{cflow-def} and \eqref{formu-conv}, we deduce the key formula
\begin{equation}
\label{mainformula}
\partial_t X(t,x)=\int \mathcal{U}_3(X(t,x)-X(t,y)) \rho_0(y) dy,
\end{equation}
where $$\mathcal{U}_3(x):=- \mathcal{U}(x)e_3 .$$
\subsection{\texorpdfstring{$L^\infty$}{L infty} Estimates}
By \eqref{mainformula} and
the Multivariate Fa\`a di Bruno formula \eqref{FDB_formula},
\begin{align*}
\partial_t^{n+1} X(t,x)&= \partial_t^n \partial_t X(t,x)\\
&=n! \underset{1\leq |\alpha|\leq n}{\sum} \int (\partial^\alpha \mathcal{U}_3(X(t,x)-X(t,z))\underset{s=1}{\overset{n}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\partial_t^{l_j} (X(t,x)-X(t,z))^{k_j}}{(k_j!) (l_j!)^{|k_j|}}\rho_0(dz),
\end{align*}
Since the integers $l_j$ in the formula above satisfy $l_j\leq n$, we get by using the (Lipschitz part of the) induction hypothesis \eqref{rec_formule} and \eqref{alpha-U},
\begin{align*}
\left|\partial_t^{n+1} X(t,x)\right|
&\leq n! \underset{1\leq |\alpha|\leq n}{\sum} \int_{\R^3} \frac{K^{|\alpha|}|\alpha|!}{|X(t,x)-X(t,z)|^{1+|\alpha|}}\\
&\quad \times \underset{s=1}{\overset{n}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\left| (-1)^{l_j-1}(l_j)!C_0^{l_j}C_1^{l_j-1} \begin{pmatrix}1/2\\ l_j \end{pmatrix}\right|^{|k_j|} |x-z|^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\rho_0(dz) ,\end{align*}
and therefore
\begin{align*}\left|\partial_t^{n+1} X(t,x)\right|
&\leq n!(-1)^n C_0^n C_1^n \underset{1\leq |\alpha|\leq n}{\sum} K^{|\alpha|}|\alpha|!(-1)^{|\alpha|}C_1^{-|\alpha|}\\
&\quad \times \left(\int_{\R^3}
\frac{|x-z|^{|\alpha|}}{|X(t,x)-X(t,z)|^{1+|\alpha|}}\rho_0(dz)\right) \underset{s=1}{\overset{n}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}} \frac{\begin{pmatrix}1/2\\ l_j \end{pmatrix}^{|k_j|}}{k_j!},\end{align*}
%
by using the definition of $P_s(n,\alpha)$.
Then, by \eqref{aveclambda} and \eqref{fors}, we deduce that
%
\begin{align*}
\left|\partial_t^{n+1} X(t,x)\right|
&\leq n!(-1)^n C_0^nC_1^n\lambda L\|\rho_0\|\underset{1\leq |\alpha|\leq n}{\sum}|\alpha|!(-1)^{|\alpha|} (K C_1^{-1} \lambda)^{|\alpha|}\underset{s=1}{\overset{n}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}} \frac{\begin{pmatrix}1/2\\ l_j \end{pmatrix}^{|k_j|}}{k_j!} .
\end{align*}
Hence, considering $C_0>0$ and $C_1>0$ such that
$K C_1^{-1} \lambda<1 $ and $L\|\rho_0\|\lambda<C_0/2 $, and
using that, according to \cite[Lemma 3.3]{CVW}, there holds:
%
\begin{equation}\label{formule_magique1}
\underset{1\leq |\alpha|\leq n}{\sum}|\alpha|!(-1)^{|\alpha|} \underset{s=1}{\overset{n}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}} \frac{\begin{pmatrix}1/2\\ l_j \end{pmatrix}^{|k_j|}}{k_j!}=2(n+1)\begin{pmatrix}1/2\\ n+1\end{pmatrix},
\end{equation}
we get that
$$
\left|\partial_t^{n+1} X(t,x)\right|\leq (n+1)!(-1)^n C_0^{n+1} C_1^n \begin{pmatrix}1/2\\ n+1\end{pmatrix} .
$$
\subsection{Lipschitz estimates}
First, from \eqref{mainformula} we deduce, by derivation in space, that
\begin{align} \label{lip-e}
\partial_t \nabla X(t,x)= \int_{\R^3} \nabla \mathcal{U}_3(X(t,x)-X(t,z))\nabla X(t,x) \rho_0(dz).
\end{align}
Moreover, by applying $\partial_t^{n}$
and applying the Leibniz formula, we arrive at
\begin{align*}
\partial_t^{n+1} \nabla X(t,x)&=\underset{k=0}{\overset{n}{\sum}} \begin{pmatrix}n \\k \end{pmatrix} \int_{\R^3} \partial_t^k \left( \nabla \mathcal{U}_3(X(t,x)-X(t,z))\right) \partial_t^{n-k}\nabla X(t,x) \rho_0(dz).
\end{align*}
A subtlety here is that it is necessary to distinguish between the case where $0 \leq k<n$ and the case where $k=n$.
First, for $0 \leq k<n$, it follows from the induction hypothesis \eqref{rec_formule} that
\begin{align}\label{formule_reccurence_gradX}
\left| \partial_t^{n-k}\nabla X(t,x) \right|&\leq (-1)^{n-k-1}(n-k)! \begin{pmatrix}1/2 \\ n-k \end{pmatrix} C_0^{n-k} C_1^{n-k-1}.
\end{align}
Second using for all $0 \leq k\leq n$, by the multivariate Fa\`a di Bruno formula \eqref{FDB_formula},
\begin{align*}
\partial_t^k \left( \nabla \mathcal{U}_3(X(t,x)-X(t,z))\right)&=k! \underset{1\leq |\alpha|\leq n}{\sum} \left(\partial^\alpha \nabla \mathcal{U}_3 \right)(X(t,x)-X(t,z))\\
&\quad \times \underset{s=1}{\overset{n}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\partial_t^{l_j} (X(t,x)-X(t,z))^{k_j}}{(k_j!) (l_j!)^{|k_j|}} .
\end{align*}
Therefore, for $0 \leq k <n$, by \eqref{alpha-U} and the induction assumption
\eqref{rec_formule}, which is assumed to be satisfied up to $n \in \N^*$,
\begin{align*}
&\int_{\R^3} \left|\partial_t^k \left( \nabla \mathcal{U}_3(X(t,x)-X(t,z))\right) \right| \rho_0(dz) \leq k! \underset{1\leq |\alpha|\leq k}{\sum} \int_{\R^3} \frac{K^{|\alpha|+1}(|\alpha|+1)!}{|X(t,x)-X(t,z)|^{2+|\alpha|} }\\
&\times \underset{s=1}{\overset{k}{\sum}}\underset{P_s(n,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\left| (-1)^{l_j-1}(l_j)!C_0^{l_j}C_1^{l_j-1} \begin{pmatrix}1/2\\ l_j \end{pmatrix}\right|^{|k_j|} |x-z|^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\rho_0(dz)\\
& \leq (-1)^{k} k!C_0^kC_1^k\underset{1\leq |\alpha|\leq k}{\sum} C_1^{-|\alpha|}K^{1+|\alpha|}(1+|\alpha|)\lambda^{2+|\alpha|} \left(\int_{\R^3}\frac{\rho_0(dz)}{|x-z|^2} \right) \\
&\quad \times (-1)^{|\alpha|}|\alpha|!\underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{ \begin{pmatrix}1/2\\ l_j \end{pmatrix}^{|k_j|} }{(k_j!) },
\end{align*}
by \eqref{aveclambda}.
Therefore,
\begin{align*}
\int_{\R^3} \left|\partial_t^k \left( \nabla \mathcal{U}_3(X(t,x)-X(t,z))\right) \right| \rho_0(dz)
& \leq (-1)^{k} k!C_0^kC_1^k \underset{1\leq |\alpha|\leq k}{\sum} (2KL \|\rho_0\| \lambda^2) (2 C_1^{-1}K\lambda)^{|\alpha|} \\
&\quad \times (-1)^{|\alpha|}|\alpha|!\underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{ \begin{pmatrix}1/2\\ l_j \end{pmatrix}^{|k_j|} }{(k_j!) }\\
&\leq (-1)^{k} (k+1)!C_0^{k+1}C_1^{k+1}/4 \begin{pmatrix}1/2\\ k+1 \end{pmatrix},
\end{align*}
where we used \eqref{formule_magique1} and assumed that $2 C_1^{-1}K\lambda<1 $ and $2KL \|\rho_0\| \lambda^2<\frac{C_0}{2}\frac{C_1}{4} $.
We treat the case $n=k$ slightly differently by assuming that $2KL \|\rho_0\| \lambda^2<\frac{C_0}{2}\frac{1}{4}\frac{1}{\|\nabla X\|_\infty} $ instead.
Thus,
\begin{align*}
\int_{\R^3} \left|\partial_t^n \left( \nabla \mathcal{U}_3(X(t,x)-X(t,z))\right) \right| \rho_0(dz)
&\leq (-1)^{n} (n+1)!C_0^{n+1}C_1^{n}\frac{1}{4 \|\nabla X\|_\infty} \begin{pmatrix}1/2\\ n+1 \end{pmatrix} .
\end{align*}
Gathering the estimates we get
\begin{align*}
&\left|\partial_t^{n+1} \nabla X(t,x)\right|\\
&\leq \underset{k=0}{\overset{n}{\sum}} \begin{pmatrix}n \\k \end{pmatrix} (-1)^{n-k-1}(n-k)! \begin{pmatrix}1/2 \\n-k \end{pmatrix}C_0^{n-k}C_1^{n-k-1} (-1)^k(k+1)!C_0^{k+1}\frac{C_1^{k+1}}{4} \begin{pmatrix}1/2 \\k+1 \end{pmatrix} \\
&\leq n! C_1^{n} C_0^{n+1}\underset{k=0}{\overset{n}{\sum}} (k+1) (-1)^{n-k-1} \begin{pmatrix}1/2 \\n-k \end{pmatrix} (-1)^k\frac{1}{4} \begin{pmatrix}1/2 \\k+1 \end{pmatrix} \\
&\leq (n+1)! C_1^{n} C_0^{n+1}(-1)^n\begin{pmatrix}1/2 \\n+1 \end{pmatrix},
\end{align*}
where we used \cite[Lemma 4.2]{CVW} for $r=n$ to get
\begin{equation}\label{formule_magique2}
\underset{k=0}{\overset{n}{\sum}} (k+1) (-1)^k \begin{pmatrix}1/2 \\k+1 \end{pmatrix} (-1)^{n-k-1}\begin{pmatrix}1/2 \\n-k \end{pmatrix} \leq 4(n+1)(-1)^n \begin{pmatrix}1/2 \\n+1 \end{pmatrix}.
\end{equation}
Indeed we emphasize that $(-1)^n \begin{pmatrix}1/2 \\n+1 \end{pmatrix}\geq 0 $ which allows us to bound the sum on $m=0 \dots n$ and $r=n$ by the double sum on $m,r=0 \cdots n$ in \cite[Lemma 4.2]{CVW}.
\subsection{H\"older estimates}
By \eqref{lip-e} and Leibniz' rule,
\begin{align*}
&\partial_t^{n+1}(\nabla X(t,x)-\nabla X(t,y))=
\\
&= \partial_t^n \int_{\R^3} \left[\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right]\nabla X(t,x) \rho_0(dz)\\
&\quad+ \partial_t^n \int_{\R^3} \left[\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right] \left( \nabla X(t,x)-\nabla X(t,y) \right) \rho_0(dz)\\
&=\underset{k=0}{\overset{n}{\sum}} \begin{pmatrix} n\\ k \end{pmatrix} (A_k+ B_k),
\end{align*}
with
\begin{align*}
A_k &:= \int_{\R^3} \partial_t^k \left(\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right)\rho_0(dz) \left( \partial_t^{n-k}\nabla X(t,x)\right) ,
\\ B_k &:=\int_{\R^3} \partial_t^k \left[\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right] \rho_0(dz) \partial_t^{n-k}\left( \nabla X(t,x)-\nabla X(t,y) \right) .
\end{align*}
For $A_k$, by the Multivariate Fa\`a di Bruno formula \eqref{FDB_formula},
\begin{align*}
& \partial_t^k \left(\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right)\\
&=k! \underset{1\leq |\alpha|\leq k}{\sum} \Big\{ \left(\partial^\alpha \nabla \mathcal{U}_3 \right)(X(t,x)-X(t,z)) \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\partial_t^{l_j} (X(t,x)-X(t,z))^{k_j}}{(k_j!) (l_j!)^{|k_j|}}\\
&\quad - \left(\partial^\alpha \nabla \mathcal{U}_3 \right)(X(t,y)-X(t,z)) \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\partial_t^{l_j} (X(t,y)-X(t,z))^{k_j}}{(k_j!) (l_j!)^{|k_j|}} \Big\},
\end{align*}
and therefore
\begin{equation}
\label{tard}
\partial_t^k \left(\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right)
=k! \underset{1\leq |\alpha|\leq k}{\sum} G_\alpha(t,x,z)-G_\alpha(t,y,z),
\end{equation}
with, for $1\leq |\alpha|\leq k$,
$$
G_\alpha(t,x,z):=\left(\partial^\alpha \nabla \mathcal{U}_3 \right)(X(t,x)-X(t,z)) \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\partial_t^{l_j} (X(t,x)-X(t,z))^{k_j}}{(k_j!) (l_j!)^{|k_j|}}.
$$
We shall use the following Lemma.
\begin{lemme}\label{lemme_Miot_general}
Let $k\geq 1$ and
$1\leq |\alpha|\leq k$. Then there exists a constant $C_p\geq 2L$ independent of $k$ such that
\begin{equation}\label{Rapp}
\int \left|G_\alpha(t,y,z) \right|\rho_0(dz) \leq L \|\rho_0\|_{L^1\cap L^p} \Lambda ,
\end{equation}
and
\begin{equation}\label{Rapp-diff}
\int \left|G_\alpha(t,x,z)-G_\alpha(t,y,z)\right| \rho_0(dz) \leq \Lambda |x-y|^\mu C_p\|\rho_0\|_{L^1\cap L^p},
\end{equation}
with
\begin{equation}
\label{def-Lamb}
\Lambda:= \lambda^{2+|\alpha|} K^{1+|\alpha|}(2+|\alpha|)! \left(K\lambda+1 \right)\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\|\partial_t^{l_j}X(t,\cdot) \|_{1,\mu}^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right).
\end{equation}
\end{lemme}
\begin{proof}
First, by \eqref{alpha-U},
\begin{eqnarray*}
|G_\alpha(t,x,z)|&\leq \frac{K^{2+|\alpha|}(2+|\alpha|)!}{|X(t,x)-X(t,z)|^{2+|\alpha|}}\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\|\partial_t^{l_j}X(t,\cdot) \|_{1,\mu}^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right)|x-z|^{|\alpha|} ,
\end{eqnarray*}
and therefore, by \eqref{aveclambda},
\begin{eqnarray}\label{borne_G}
|G_\alpha(t,x,z)| \leq \Lambda_1 \frac{1}{|x-z|^2} ,
\end{eqnarray}
with
$$
\Lambda_1= \lambda^{2+|\alpha|}K^{2+|\alpha|}(2+|\alpha|)!\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\|\partial_t^{l_j}X(t,\cdot) \|_{1,\mu}^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right).
$$
Observing that, since $\lambda>1$, $ \Lambda_1 \leq \Lambda$, where we recall that $\Lambda$ is defined in \eqref{def-Lamb}, we conclude the proof of
\eqref{Rapp} using \eqref{fors}.
Next, to prove \eqref{Rapp-diff},
we first observe that, for any distinct $x$, $y$ and $z$ in $\R^3$, we have that
\begin{align}\label{app1}
|G_\alpha(t,x,z)-G_\alpha(t,y,z)| \leq \underset{u\in S_{x,y}}{\sup}|\nabla_u G_\alpha(t,u,z)| |x-y|,
\end{align}
where $S_{x,y}$ is a $C^1$ path from $x$ to $y$
such that for all $u \in S_{x,y}$,
\begin{align}\label{app2}
|u-z| \geq \min(|z-x|,|z-y|).
\end{align}
We refer to \cite[Lemma 3.15]{Hofer} for the proof of existence of such a path.
Now, using \eqref{alpha-U} and \eqref{aveclambda},
we obtain that
\begin{align*}
|\partial_{u} G_\alpha(t,u,z)|&\leq \frac{K^{2+|\alpha|}(2+|\alpha|)!}{|X(t,u)-X(t,z)|^{3+|\alpha|}}|u-z|^{|\alpha|}\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\|\partial_t^{l_j}X(t,\cdot) \|_{1,\mu}^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right)\\
&+ \frac{K^{1+|\alpha|}(1+|\alpha|)!}{| X(t,u)-X(t,z)|^{2+|\alpha|} }|\alpha||u-z|^{|\alpha|-1}\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\|\partial_t^{l_j}X(t,\cdot) \|_{1,\mu}^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right)\\
&\leq \lambda^{2+|\alpha|} K^{1+|\alpha|}(2+|\alpha|)! \left(K\lambda+1 \right) \frac{1}{|u-z|^3}\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\|\partial_t^{l_j}X(t,\cdot) \|_{1,\mu}^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right).
\end{align*}
By combining with \eqref{app1} and \eqref{app2},
this ensures that
\begin{equation}\label{lipschitz_type}
|G_\alpha(t,x,z)-G_\alpha(t,y,z)|\leq \Lambda |x-y| \left(\frac{1}{|x-z|^3}+\frac{1}{|y-z|^3} \right),
\end{equation}
recalling the definition of
$
\Lambda
$ in \eqref{def-Lamb}.
Now, following the proof of \cite[Lemma 2]{Miot} with the kernel $G_\alpha(t,x,z)$, we
set $d:=|x-y|$ and observe that, thanks to \eqref{borne_G}, it is enough to show the result for $d<1 $. Indeed if $d=|x-y|\geq 1$ then
\begin{equation}\label{eq:Holder_d>1}
\int_{\R^3} |G_\alpha(t,x,z)-G_\alpha(t,y,z)|\rho_0(dz)\leq 2 \underset{x}{\sup}\int_{\R^3} |G_\alpha(t,x,z)|\rho_0(dz) \leq 2\Lambda_1 L \|\rho_0\|_{L^1\cap L^p} \leq 2\Lambda_1 L \|\rho_0\|_{L^1\cap L^p} |x-y|^\mu.
\end{equation}
Now assume $d<1$, we
set $A:=\frac{x+y}{2}$, and we split the following integral into three parts:
\begin{align}\label{trich}
&\int_{\R^3} |G_\alpha(t,x,z)-G_\alpha(t,y,z)|\rho_0(dz) =J_1 +J_2+ J_3,
\end{align}
where
\begin{align*}
J_1 :=&\int_{\R^3\setminus B(A,1)}|G_\alpha(t,x,z)-G_\alpha(t,y,z)|\rho_0(dz) ,
\\ J_2 :=&\int_{B(A,1)\setminus B(1,d)}|G_\alpha(t,x,z)-G_\alpha(t,y,z)|\rho_0(dz) ,
\\ J_3 :=& \int_{B(A,d)}|G_\alpha(t,x,z)-G_\alpha(t,y,z)|\rho_0(dz) .
\end{align*}
\begin{itemize}
\item For $J_1$, we have for all $z\in B(A,d)$, $|z-x|\leq |z-1|+d/2 \leq 3d/2$. This yields, by \eqref{borne_G}, that
\begin{align*}
J_1&\leq 2 \Lambda_1 \int_{B(x,3d/2)} \frac{1}{|x-z|^2}\rho_0(dz)\\
&\leq C\Lambda_1 \left(\int_0^{3d/2} s^{2-2p'} \right)^{1/{p'}} \|\rho_0\|_{L^p} ,
\end{align*}
and therefore
\begin{align}\label{J1}
J_1 &\leq \tilde{C}_p \Lambda_1 \|\rho_0\|_{L^p} d^{3/{p'}-2}\leq \tilde{C}_p \Lambda \|\rho_0\|_{L^p} d^{3/{p'}-2}.
\end{align}
\item For $J_2$ we use that for $ z\in B(A,1)\setminus B(A,d)$,
$$
|z-x|>|z-A|-d/2>d/2
\text{ and } |z-x|\leq |z-A|+d/2 \leq 1+d/2 ,$$
and \eqref{lipschitz_type} to arrive at
\begin{align*}
J_2&\leq |x-y|\Lambda \int_{ B(A,1)\setminus B(1,d)}\left(\frac{1 }{|x-z|^3}+\frac{1}{|y-z|^3} \right)\rho_0(dz)\\
&\leq C |x-y|\Lambda \|\rho_0\|_{L^p} \left(\int_{d/2}^{1+d/2} s^{2-3p'}\right)^{1/{p'}} ds ,
\end{align*}
and therefore
\begin{align}\label{J2}
J_2&\leq \tilde{C}_p |x-y|\Lambda \|\rho_0\|_{L^p}d^{3/{p'}-3}=\tilde{C}_p \Lambda \|\rho_0\|_{L^p}d^{3/{p'}-2} .
\end{align}
\item For $J_3$ we have that any $z$ such that $|z-A|>1$ satisfies
$$
|z-x|>|z-A|- \frac{1}{2} d>|z-A|- \frac{1}{2} > \frac{1}{2} |z-A|,
$$
which yields using \eqref{lipschitz_type},
\begin{align}\label{J3}
J_3 &\leq \Lambda |x-y| \int_{\R^3\setminus B(A,1)}\left(\frac{1}{|x-z|^3}+\frac{1}{|y-z|^3} \right)\rho_0(dz)\leq 16|x-y|\Lambda \|\rho_0\|_{L^1} .
\end{align}
\end{itemize}
By combining \eqref{eq:Holder_d>1}, \eqref{trich}, \eqref{J1}, \eqref{J2} and \eqref{J3}, we establish the inequality \eqref{Rapp-diff} by choosing $C_p\geq \max(2L, 2\tilde{C}_p+16)$.
This concludes the proof of Lemma \ref{lemme_Miot_general}.
\end{proof}
We start with the treatment of the term $A_k$ which is the most difficult.
By combining \eqref{tard} and \eqref{Rapp-diff}
we first deduce that
\begin{align*}
&\frac{1}{|x-y|^\mu}\int_{\R^3}
\left| \partial_t^k \left(\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right)\rho_0(dz) \right| \\
& \leq k! \underset{1 \leq |\alpha|\leq k}{\sum} \lambda^{2+|\alpha|} K^{1+|\alpha|}C_p \|\rho_0\|_{L^1\cap L^p}(2+|\alpha|)! \left(K\lambda+1 \right)\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\|\partial_t^{l_j}X(t,\cdot) \|_{1,\mu}^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right) .
\end{align*}
Then we use the induction assumption
\eqref{rec_formule}, which is assumed to be satisfied up to $n \in \N^*$, to infer that
%
\begin{align*}
&\frac{1}{|x-y|^\mu}\int_{\R^3} \left| \partial_t^k \left(\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right)\rho_0(dz) \right| \\
& \leq k! \underset{1 \leq |\alpha|\leq k}{\sum} \lambda^{2+|\alpha|} K^{1+|\alpha|}C_p \|\rho_0\|_{L^1\cap L^p}(2+|\alpha|)! \left(K\lambda+1 \right)\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{\left( (-1)^{l_j-1} (l_j!) \begin{pmatrix}1/2\\ l_j \end{pmatrix}C_0^{l_j}C_1^{l_j-1}\right)^{|k_j|}}{(k_j!) (l_j!)^{|k_j|}}\right)\\
& \leq k!C_0^k C_1^k (-1)^k \underset{1 \leq |\alpha|\leq k}{\sum} \lambda^{2+|\alpha|} C_1^{-|\alpha|}(-1)^{|\alpha|}K^{1+|\alpha|}C_p \|\rho_0\|_{L^1\cap L^p}(2+|\alpha|)! \left(K\lambda+1 \right)\left( \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{ \begin{pmatrix}1/2\\ l_j \end{pmatrix})^{|k_j|}}{(k_j!)}\right)\\
&= k!C_0^k C_1^k (-1)^k \underset{1 \leq |\alpha|\leq k}{\sum} \lambda^{2+|\alpha|} C_1^{-|\alpha|}K^{1+|\alpha|}C_p \|\rho_0\|_{L^1\cap L^p}2^{2+|\alpha|}2^{1+|\alpha|} \left(K\lambda+1 \right)\left( (-1)^{|\alpha|} |\alpha|! \underset{s=1}{\overset{k}{\sum}}\underset{P_s(k,\alpha)}{\sum}\underset{j=1}{\overset{s}{\prod}}\frac{ \begin{pmatrix}1/2\\ l_j \end{pmatrix})^{|k_j|}}{(k_j!)}\right).
\end{align*}
We now assume that $4\lambda C_1^{-1}K<1$ and $ \lambda^2 K C_p \|\rho_0\|_{L^1\cap L^p}2^3(K\lambda+1)\leq \frac{C_0}{2} \frac{C_1}{4}$, in order to get for $n\neq k$
\begin{align*}
&\frac{1}{|x-y|^\mu}
\left| \int \partial_t^k \left(\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right)\rho_0(dz) \right| \\
&\leq \frac{1}{4}(k+1)!C_0^{k+1} C_1^{k+1} (-1)^k \begin{pmatrix}1/2\\ k+1 \end{pmatrix},
\end{align*}
by using again \eqref{formule_magique1}.
Again, we treat the case $n=k$ differently by assuming $$ \lambda^2 K C_p \|\rho_0\|_{L^1\cap L^p}2^3(K\lambda+1)\leq \frac{C_0}{2} \frac{1}{4\|\nabla X\|_{0,\mu}},$$ which yields
\begin{align*}
&\frac{1}{|x-y|^\mu}\int_{\R^3} \partial_t^n \left(\nabla \mathcal{U}_3(X(t,x)-X(t,z))-\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right)\rho_0(dz)\\
&\leq \frac{1}{4 \|\nabla X\|_{0,\mu}}(n+1)!C_0^{n+1} C_1^{n} (-1)^n \begin{pmatrix}1/2\\ n+1 \end{pmatrix}.
\end{align*}
We get, by using again \eqref{formule_magique1} and \eqref{formule_reccurence_gradX}, that
\begin{align*}
\underset{k=0}{\overset{n}{\sum}}\begin{pmatrix}n \\k \end{pmatrix}\frac{|A_k|}{|x-y|^\mu}&\leq
\underset{k=0}{\overset{n}{\sum}}\begin{pmatrix}n \\k \end{pmatrix}(k+1)!C_0^{k+1} C_1^{k+1} (-1)^k \frac{1}{4} \begin{pmatrix}1/2\\ k+1 \end{pmatrix}(-1)^{n-k-1}(n-k)! \begin{pmatrix}1/2 \\ n-k \end{pmatrix} C_0^{n-k} C_1^{n-k-1}\\
&\leq n! \underset{k=0}{\overset{n}{\sum}}\frac{1}{4} C_0^{n+1} C_1^{n} (k+1)(-1)^k\begin{pmatrix}1/2\\ k+1 \end{pmatrix} (-1)^{n-k-1}\begin{pmatrix}1/2 \\ n-k \end{pmatrix} \\
&\leq C_0^{n+1} C_1^{n} (-1)^n (n+1)!\begin{pmatrix}1/2 \\ n+1 \end{pmatrix},
\end{align*}
where we used \eqref{formule_magique2}.
Analogously for $B_k$ we emphasize that
$$
\partial_t^k \left[\nabla \mathcal{U}_3(X(t,y)-X(t,z)) \right]=k! \underset{1\leq |\alpha|\leq k}{\sum} G_\alpha(t,y,z),
$$
and that, analogously to \eqref{formule_reccurence_gradX} we have for $n \neq k$
$$
\left| \partial_t^{n-k}\left(\nabla X(t,x)-\nabla X(t,y)\right) \right|\leq (-1)^{n-k-1}(n-k)! \begin{pmatrix}1/2 \\ n-k \end{pmatrix} C_0^{n-k} C_1^{n-k-1} |x-y|^\mu .
$$
Using \eqref{Rapp} and
the two estimates above, we can treat the term $B_k$ as the term $A_k$ and conclude the proof.
\section{Proof of Theorem \ref{TS-control}}\label{sec-cont}
\label{sec-control}
This section is devoted to the proof of Theorem \ref{TS-control}.
By a reversibility argument, it is enough to find a control which sends the initial density $\rho_0$ to $0$ on the time interval $[0,T/2]$, indeed one can then proceed in a similar way and send $0$ to the final density $\rho_f$ in the time interval $[T/2,T]$ by switching the time variable $t$ into $T-t$.
We emphasize that it is actually sufficient to prove that it is possible to prove that there exists $\epsilon\in (0,\frac12)$ and a control which steers the initial density $\rho_0$ to $0$ on the time interval $[0,\epsilon T]$, since then one may turn off the control and let the density stays at zero during the rest of the interval $]\epsilon T, T/2[$.
Without loss of generality we assume that there exists $\delta>0$ small enough such that $B(0,4\delta) \subset \omega$.
\subsection{Construction of a controlled auxiliary solution starting from $0$ and returning at $0$}
Following \cite{G20}, in this subsection we construct, for any $T>0$, a vector field $u$ with zero initial and final values and such that during the imparted time the corresponding flow map sends the support of $\rho_0$ to the ball $B(0,2 \delta) $ in several time steps and then back to the support of $\rho_0$. More precisely we have the following result.
\begin{prop}\label{lemme1}
Let $T>0$, $\omega$ and $\rho_0$ as above. There exists $f\in C_c((0,T);L^\infty(\R^3))$ compactly supported in $(0,T)\times \omega$ and $(u_\text{aux},p_{\text{aux}}) \in C_c((0,T);\bigcup_{q\geq3} W^{2,q}(\R^3) \times \bigcup_{q\geq3}W^{1,q}(\R^3))$ satisfying
$$-\Delta u_\text{aux} +\nabla p_\text{aux} = f \quad \text{ and } \quad \div u_\text{aux}=0,
\quad \text{ on } \quad \R^3 , $$
and
there exists a covering $\underset{i=1}{\overset{L}{\bigcup}} B(x_i,r_i)$ of $\text{supp} \rho_0$ and a sequence $0\leq t_i<t_{i+1/4}<t_{i+1/2}<t_{i+1}\leq T/2$, $0\leq i \leq L $ with $t_0=T/4$, $t_{L+1}=T/2$ such that
\begin{equation}\label{forumule_controle_zone}
\forall 1 \leq i \leq L,\: \forall x\in B(x_i, r_i), \quad \forall \, t \in ]t_{i+1/4},t_{i+1/2}[ \,\Rightarrow
X_\text{aux}(t,0,x)\in B(0,2\delta) ,
\end{equation}
$$
\forall 1 \leq i \leq L, \forall\, x \in \R^3, X_{\text{aux}}(t_i,0,x)=x .
$$
\end{prop}
\begin{proof}[Proof of Proposition \ref{lemme1}]
By a scaling argument we may choose $T=1$. Let $\alpha \in \omega $ and $r>0$ such that $B(\alpha,r)\subset \omega \setminus \overline{}{B(0,4 \delta)}$.
The proof relies on the following Lemma that we prove later.
\begin{lemme}\label{lemme2}
Let $ \gamma : [0,1] \to \R^3\setminus \overline{B(\alpha,r)}$ a curve of class $C^\infty$ such that
$\gamma(t)=\gamma(0)$ for $t \in [0,\zeta]$ and
$\gamma(t)=\gamma(1)$, for $ t \in [1-\zeta,1]$,
for a given small $\zeta>0$. Then there exists
$f\in C(0,1,L^\infty(\R^3))$ compactly supported in $(0,1)\times B(\alpha,r)$ and the unique associated solution $u\in C_c(0,1,\bigcap_{q\geq 3}W^{2,q}(\R^3) )$ of
$$-\Delta u +\nabla p = f,\quad \div u =0,\quad \text{ on } \R^3$$
such that the associated characteristic flow $X$ satisfies $X(t,0,\gamma(0))=\gamma(t)$.
\end{lemme}
Now let $a \in \text{supp} \rho_0 \setminus \overline{B(\alpha,r)}$.
There exists a curve $\gamma_a:[0,1] \to \R^3 \setminus \overline{B(\alpha,r)}$ of class $C^\infty$ such that $$\gamma_a(t)=a \text{ for } t\in[0,1/4], \: \gamma_a(t)=0 \text{ for }t\in[1/2, 1]$$
Denote by $u_a$ the velocity and $f_a$ the right hand side obtained by Lemma \ref{lemme2} associated to the curve $\gamma_a$.
By continuity of the flow $X_a$, there exists $r_a>0$ and $\eta_a>0$ small enough such that for all $x \in B(a,r_a)$ and for all $t \in ]1/2-\eta_a,1/2+\eta_a[$ we have $X_a(t,0,x)\in B(0,2\delta)$. We emphasize that the stationarity of the curve near $t=0$ and $t=1$ is crucial in order to have a compactly supported velocity field $u_a$.
Now for $ a \in \text{supp} \rho_0\cap \overline{B(\alpha,r)} $ we consider $B(\beta,r')$ such that $B(\beta,r') \subset \omega \setminus \overline{\left( B(0,4\delta)\cup B(\alpha,r)\right )}$ and apply Lemma \ref{lemme2} for $B(\beta,r')$ instead of $B(\alpha,r)$ in order to drive each point $a$ of $\text{supp} \rho_0\cap \overline{B(\alpha,r)}$ to $0$. This yields the existence of a control $u_a$ satisfying a Stokes equation with a right hand side $f_a$ compactly supported in $B(\beta,r')$ and we may apply the same continuity argument as above to get the existence of $r_a$ and $\eta_a$ such that $X_a(t,0,x)\in B(0,2\delta)$ for $x\in B(a,r_a)$, $t\in ]1/2-\eta_a,1/2+\eta_a[$.
Hence, to summarize, for all $a\in \text{supp} \rho_0 $, there exists $r_a>0$ and $u_a$ compactly supported in $(0,1)$ which solves the Stokes equation with a right hand side $f_a$ compactly supported in $(0,1)\times \omega$ and such that for all $x \in B(a,r_a)$, in $X_a(t,0,x) \in B(0,2\delta)$ for $t\in]1/4,1/2[$.
By compactness one can extract a finite subcover
$$\underset{i=1}{\overset{L}{\bigcup}}B(x_i,r_i),$$
of $\text{supp} \rho_0$.
Let $(u_i)_i$ the associated velocity fields.
We split the time segment $[1/4,1/2]$ by considering the subdivision%
\begin{gather*}
t_0=1/4, \\
t_i=t_0+ \frac{i}{4(L+1)}, \quad \forall\: i \in \{1, \cdots, L+1\},\\
t_{i+1/4}=t_i +\frac{1}{4} \frac{1}{4(L+1)}, \quad \forall\: i \in \{0, \cdots, L\}, \\
t_{i+1/2}=t_i +\frac{1}{2} \frac{1}{4(L+1)},\quad \forall\: i \in \{0, \cdots, L\}.
\end{gather*}
We set then
%
\begin{align*}
u_\text{aux}(t,x)&=8(L+1)u_i(8(L+1)(t-t_{i}),x), \: 0 \leq i \leq L, \: t \in [t_{i},t_{i+1/2}],
\\ u_\text{aux}(t,x)&=-8(L+1)u_i(8(L+1)(t_{i}-t),x), \: 0 \leq i \leq L, \: t \in [t_{i+1/2},t_{i+1}].
\end{align*}
$u_\text{aux}$ is then well defined and compactly supported in $(0,T)$ since each $u_i$ is compactly supported. Moreover, $u_\text{aux}$ solves the Stokes equation on $\R^3 \setminus \omega$.
Moreover, on each time segment $]t_{i},t_{i+1}[ $ the domain $B(x_i,r_i)$ is transported inside $B(0,2\delta)$ during the time interval $]t_{i+1/4},t_{i+1/2}[$ and then again to $B(x_i,r_i)$ during the time interval $]t_{i+1/2},t_{i+1}[$, more precisely
the characteristic flow $X_\text{aux}$ associated with $u_\text{aux}$ satisfies
\begin{equation*}
\forall 1 \leq i \leq L,\: \forall x\in B(x_i, r_i), \quad \forall \, t \in ]t_{i+1/4},t_{i+1/2}[ \,\Rightarrow
X_\text{aux}(t,0,x)\in B(0,2\delta).
\end{equation*}
$$
\forall 1 \leq i \leq L,\forall \, x \in \R^3, \: X_\text{aux}(t_{i},0,x)=x .$$
\end{proof}
Following \cite[Proposition 3.2.8]{G}, the proof of Lemma \ref{lemme2} relies on the following result which establishes the pointwise reachability of any value in $\R^3$ by the velocity vector fields satisfying controlled steady Stokes systems.
\begin{lemme}\label{lemme3}
Let $ x \in \R^3 \setminus B(\alpha,r)$, then we have
$$
\{ u(x),\exists \, f\in L^\infty(\R^3) \text{ compactly supported in } B(\alpha,r) \text{st } -\Delta u+\nabla p =f \text{ and } \div u= 0 \text{ on } \R^3\}=\R^3.
$$
\end{lemme}
\begin{proof}[Proof of Lemma \ref{lemme3}]
We set for a given $x \in \R^3 \setminus B(\alpha,r)$ and any $a \in \R^3$
$$
F:=8\pi|x-\alpha | \left(I-\frac{1}{2}\frac{(x-\alpha) \otimes (x-\alpha)}{|x-\alpha|^2} \right)a \in \R^3 ,
$$
we consider then the velocity given by $u=\mathcal{U}(\cdot-\alpha)F$
where we recall that $\mathcal{U} $
is the fundamental solution of the Stokes system in $\R^3$, see Section \ref{SSE}. We define as well the associated pressure as $p:=\mathcal{P}(\cdot-\alpha)F $.
Thus
$(u,p)$ satisfies $\div u= 0 $ and the steady Stokes equation $-\Delta u+\nabla p =\delta_\alpha F $ on on $\R^3$.
Moreover,
%
\begin{align*}
u(x) = \mathcal{U} (x-a) F
= \frac{1}{8\pi} \left(\frac{I}{|x-\alpha|}+\frac{(x-\alpha) \otimes (x-\alpha)}{|x-a|^3} \right) 8\pi|x-\alpha| \left(I-\frac{1}{2}\frac{(x-\alpha) \otimes (x-\alpha)}{|x-\alpha|^2} \right)a = a .
\end{align*}
In order to regularise the right hand side $\delta_{\alpha}F$ we set $\tilde{u} =\mathcal{U} \star \frac{\chi_{B(\delta,\varepsilon)}}{|B(\delta,\varepsilon)|}F$ with $\varepsilon<r$. One can show that $ |a-\tilde{u}(x)|=|u(x)-\tilde{u}(x)|\leq C_{x,\alpha} \varepsilon$ for $\varepsilon$ small enough and this shows that the set
$$
\{ u(x),\exists \, f\in L^\infty(\R^3) \text{ compactly supported in } B(\alpha,r) \text{st } -\Delta u+\nabla p =f, \quad \div u= 0 \text{ on } \R^3\}
$$
is dense in $\R^3$, since it is a finite dimensional vector space and hence closed, we get the desired equality.
\end{proof}
%
In the sequel we will look for a solution $ (u,\rho)$ to the transport-Stokes system
in $\R^3 \setminus \omega$ starting with $\rho_0$ at time $0$, and exactly reaching the null state at time $T$, whose main part is given by a rescaling of the auxiliary solution $u_\text{aux}$, while the influence of the initial data will be considered as an error term.
More precisely, let $\epsilon>0$, set
$$u^\epsilon_\text{aux}(t,x):=\frac{1}{\epsilon}u_\text{aux}(\frac{t}{\epsilon},x) \quad \text{ and } \quad
p^\epsilon_\text{aux}(t,x):=\frac{1}{\epsilon} p_\text{aux}(\frac{t}{\epsilon},x),
$$
which satisfy
\begin{equation}
\label{s*}
-\Delta u^\epsilon_\text{aux} +\nabla p^\epsilon_\text{aux} = 0 \quad \text{ in } \quad \R^3 \setminus w, \quad
\div u^\epsilon_\text{aux}=0 \quad \text{ in }
\R^3 .
\end{equation}
Let us set accordingly
$$X_\epsilon(t,s,x):=X(t/\epsilon,s,x),$$
the associated flow defined on $[0,\epsilon T]$.
\subsection{Construction of an error term due to the initial density and reaching equilibrium}
This section is devoted to the proof of the following result, where, roughly speaking, it is proved that for $\epsilon >0$ small enough, one may construct an error term, due to the initial density, which reaches equilibrium before the imparted time.
%
%
\begin{prop}\label{propRn}
There exists $(u^* , \rho^*) $ in $C([0,+\infty),W^{2,p}(\R^3) \times L_p(\R^3))$
so that, in $\R^3$, for $\epsilon >0$ small enough,
\begin{equation}
\label{ts*}
\left\{
\begin{array}{rcl}
&\partial_t \rho^* + \div ((u^\epsilon_\text{aux}+u^*)\rho)=g^* ,\\
&-\Delta u^* + \nabla p = - \rho^* e_3 \quad \text{ and } \div u^*= 0 ,\text{ in } \R^3\\
&\rho^*(0,\cdot)=\rho_0 \quad \text{ and } \quad \rho^*(\epsilon T ,\cdot)=0 .
\end{array}
\right.
\end{equation}
with
$g^*$ compactly supported in $(0,\epsilon T)\times \omega$.
\end{prop}
\begin{proof}[Proof of Proposition \ref{propRn}]
We introduce the following sequence $(\rho^n,u^n)_{n \geq 0}$ defined on $[0,\epsilon T]\times \R^3$ such that
%
\begin{equation}\label{formule1}
\|\rho^n\|_{L^1\cap L^p} \leq \|\rho_0\|_{L^1\cap L^p} , \|u^n\|_{W^{2,p}} \leq K ,
\end{equation}
defined as follows.
\begin{itemize}
\item We set $(u^0,\rho^0):=(u_0,\rho_0)$ where $u_0$ the solution to the Stokes equation on $\R^3$ with a right hand side given by $-\rho_0e_3$.
\item Given $(u^n,\rho^n)$ satisfying \eqref{formule1}, we set $\rho^{n+1/2}$ as
$$
\rho^{n+1/2}:=X_\epsilon^n \# \rho_0
$$
where $X^n_\varepsilon$ the characteristic flow associated to $u^\epsilon_\text{aux}+u^n $. Then $\rho^{n+1/2}$ satisfies weakly
\begin{eqnarray*}
\partial_t \rho^{n+1/2}+ \div ((u^\epsilon_\text{aux}+u^n)\rho^{n+1/2})=0 , {\text{ on } \R^3 }\\
\rho^{n+1/2}(0,\cdot)=\rho_0.
\end{eqnarray*}
\item Using \eqref{forumule_controle_zone} together with \eqref{formule1}, there exists $\epsilon_0=C(T,K,\delta, \text{Lip}(u_\text{aux}))>0$ independent of $n$ such that for all $\varepsilon$ in $(0, \varepsilon_0)$,
\begin{equation}\label{formule2}
\forall t \in [0,\epsilon T],\: \forall x\in B(x_i,r_i), \quad t \in ]t_{i+1/4},t_{i+1/2}[ \Rightarrow
X^n_\epsilon(t,0,x)\in B(0,3 \delta).
\end{equation}
With the same notations as above, let $\{ \chi_i \}_{1 \leq i \leq L}$ a partition of unity such that $\text{supp} \chi_i \subset B(x_i, r_i)$.
We set then for $t\in[0,\epsilon T]$ and $x\in \R^3$,
$$
\rho^{n+1}(t,x):= \underset{i=1}{\overset{L}{\sum}}\beta_i(t) \rho_0(X^n_\varepsilon(0,t,x)) \chi_i(X^n_\varepsilon(0,t,x))\\
=\underset{i=1}{\overset{L}{\sum}}\beta_i(t) \rho^{n+1/2}(t,x) \chi_i(X^n_\varepsilon(0,t,x)),
$$
with $\beta_i(t) :=\beta\left(8(L+1)(t-t_i) \right) $ with $\beta$ a cutoff function satisfying $\beta_i=1 $ for $t\leq 1/2$ and $\beta_i=0$ for $t\geq 1$.
We observe that
$\rho^{n+1}$ satisfies weakly
%
\begin{equation} \label{n-den}
\partial_t \rho^{n+1}+ \div ((u^\epsilon_\text{aux}+u^n)\rho^{n+1})=g^n \text{ on } \R^3 , \quad
\rho^{n+1}(0,\cdot)=\rho_0 \quad \text{ and } \quad
\rho^{n+1}(\epsilon T ,\cdot)=0 ,
\end{equation}
with
$$
g^n(t,x) := \underset{i=1}{\overset{L}{\sum}}\beta'_i(t) \rho_0(X^n_\varepsilon(0,t,x)) \chi_i(X^n_\varepsilon(0,t,x)) . $$
\item Finally we define $u^{n+1}$ as the unique solution to the Stokes equation:
\begin{equation} \label{n-sto}
-\Delta u^{n+1} + \nabla p^{n+1} =-\rho^{n+1}e_3 \quad \text{ and } \quad \div u^{n+1}= 0 \quad \text{ in } \quad \R^3.
\end{equation}
\end{itemize}
%
\begin{lemme}\label{lemmeRn}
The vector fields $g^n$ are compactly supported in $(0,\epsilon T)\times \omega$.
\end{lemme}
\begin{proof}[Proof of Lemma \ref{lemmeRn}]
If $g^n(t,x)\neq 0$ there exists $1 \leq i \leq L$ such that $\beta'_i(t) \rho_0(X^n_\varepsilon(0,t,x)) \chi_i(X^n_\varepsilon(0,t,x))\neq 0$ hence $t\in \epsilon ]t_{i+1/4},t_{i+1/2}[$ with $X_\epsilon^n(0,t,x)\in \text{supp} \rho_0 \cap B(x_i, r_i)$ which yields $x \in X_\epsilon^n(t,0,B(x_i,r_i)) \subset \omega $ according to \eqref{formule2}.
\end{proof}
The sequences $\rho^n$ and $g^n$ satisfy the following uniform bounds for all $q\in [0,+\infty]$,
$$
\|g^n\|_{L^q(0,\epsilon T;L^1\cap L^p)}+\|\rho^n\|_{L^q(0,\epsilon T;L^1\cap L^p)}+ \|\partial_t\rho^n\|_{L^q(0,\epsilon T;L^p+\dot{W}^{-1,p})} \leq M
$$
which yields for $u^n$ using the fact that $-\Delta \partial_t u^n+ \nabla \partial_t p^n= - \partial_t \rho^n e_3$, $\div \partial_t u^n=0$,
\begin{eqnarray*}
\|u^n\|_{L^q(0,\epsilon T;W^{2,p})}+ \|\partial_t u^n\|_{L^q(0,\epsilon T;\dot{W}^{1,p})} \leq M, & \text{ if } p>3,\\
\|u^n\|_{L^q(0,\epsilon T;W_{\operatorname{loc}}^{2,3})}
+ \|\partial_t u^n\|_{L^q(0,\epsilon T;\dot{W}^{-1,3})} \leq M, & \text{ if } p=3.
\end{eqnarray*}
Hence using the Aubin-Lions Theorem we get the convergence of a subsequence $u^n \to u^* $ in $L^q(0,\epsilon T; C^{1,\mu}_{\operatorname{loc}}\cap W^{1,p}_{\operatorname{loc}} (\R^3))$ if $p>3$. This allows us to pass in the limit in the weak formulations of the Stokes equations \eqref{n-sto} and the transport equation \eqref{n-den}, using the weak convergence $\rho^n \rightharpoonup \rho^* $ and $g^n \rightharpoonup g^* $ in $L^q(0,\epsilon T;L^p) $. For $p=3$ we get analogously $u^n \to u^* $ in $L^q(0,\epsilon T,W^{1,3}_{\operatorname{loc}}(\R^3)) $ together with $\rho^n \rightharpoonup \rho^* $ weakly in $L^q(0,\epsilon T;L^s) $ for any $1 \leq s \leq 3$ which allows us to pass in the transport equation for $s=3/2$, so that, in $\R^3$, \eqref{ts*} holds true.
We emphasize that an Ascoli-type argument similar to the one used in the proof of Theorem 1.1 still holds true for the sequence of flows $X_\epsilon^n $ and we get $X^n_\epsilon(\cdot,s,\cdot) \to X^* (\cdot,s,\cdot) $ in $ C[0,\epsilon T], C^1_{\operatorname{loc}}) $ (resp. in $ C[0,\epsilon T], C_{\operatorname{loc}}) $) if $p>3$ (resp. $p=3$).
This allows us to get a pointwise convergence of $g^n$ to
$$g^* =\underset{i=1}{\overset{L}{\sum}} \partial_t \beta_i(t) \rho_0(X^*(0,t,x))\chi_i(X^*(0,t,x)),$$
in $(0,\epsilon T)\times \R^3$ and ensures that $g^*$ is compactly supported in $(0,\epsilon T)\times \omega$.
Moreover, passing in the limit, pointwise, in the identity:
$$
\rho^{n+1}(t,x):= \underset{i=1}{\overset{L}{\sum}}\beta_i(t) \rho_0(X^n_\varepsilon(0,t,x)) \chi_i(X^n_\varepsilon(0,t,x)),
$$
we obtain that $\rho^* $ is in $ C(0,T; L^p) $ and the continuity for the velocity $u^*$ as well.
This concludes the proof of Proposition \ref{propRn}.
\end{proof}
\subsection{Conclusion of the proof of Theorem \ref{TS-control}}
Combining Proposition \ref{lemme1}
and
Proposition \ref{lemmeRn}, in particular by \eqref{s*} and \eqref{ts*}, and since $g^*$ is compactly supported in $(0,\epsilon T)\times \omega$, the vector field
$(u:= u^\epsilon_\text{aux}+u^*,\rho :=\rho^*)$ satisfies
%
\begin{gather*}
\partial_t \rho+ \div (\rho u)= 0 \quad \text{ in } \quad \R^3 \setminus \omega ,
\\ -\Delta u + \nabla p = - \rho e_3 \quad \text{ in } \R^3 \setminus \omega ,\\
\div u = 0
\quad \text{ in } \quad \R^3 ,\\
\rho(0,\cdot)=\rho_0 \quad \text{ and } \quad
\rho(T\epsilon ,\cdot)=0 .
\end{gather*}
Recalling the preliminary remark at the beginning of this section, this concludes the proof of Theorem \ref{TS-control}.
\smallskip
\smallskip
\smallskip
\noindent
\subsection*{Acknowledgment}
A.M. and F.S. are supported by the SingFlows project, grant ANR-18-CE40-0027 of the French National Research Agency (ANR).
This work was partly accomplished while F.S. was participating in a program hosted by the Mathematical Sciences Research Institute in Berkeley, California, during the Spring 2021 semester, and supported by the National Science Foundation under Grant No. DMS-1928930.
|
\section{System Model and Problem Formulation}
\subsection{System Model}
We consider a heterogeneous status update system consisting of two independent sources\footnote{Two sources are considered for simplicity of presentation.
However, the
formulated problem and solutions can be extended for an arbitrary number of sources.}: one random arrival (Source 1) and one generate-at-will (Source 2), a buffer-aided transmitter, and a monitor, as depicted in Fig. \ref{SM}.
A discrete-time system with unit time slots $t \in \{0, 1, \dots\}$ is considered.
The sources are indexed
by $i \in \{1, 2\}$.
The status update packets of Source 1 are generated according to the Bernoulli distribution with parameter $\mu$.
To maintain the freshest information currently available at the transmitter,
the most recently arrived update of Source 1 and the most recent sampled update of Source 2 are stored in the transmitter's buffers of size one packet per source.
Notice that considering a buffer of size one packet for each source is sufficient in our system, because storing and transmitting the outdated packets does not improve the AoI.
\begin{figure}[t!]
\centering
\includegraphics[width=.5\textwidth]{Figures_2/SM_MarkovQV5.eps}\vspace{-3mm}
\caption{
The considered heterogeneous status update system. The binary Markov chain models the query flag associated with source $i$ and is generated by the monitor.
}
\label{SM}
\vspace{-2 em }
\end{figure}
\\\indent
To capture the randomness of the wireless link between the transmitter and the monitor, we assume the link is error-prone, i.e., the reception of updates by
the monitor are successful
with a constant (over slots) probability ${ p\in(0, 1]}$.
Moreover, we assume that at each slot, at most one packet can be transmitted over the link, which stands for the \textit{per-slot transmission constraint} in the system.
Besides, we assume that each transmission occupies one slot, and perfect feedback (i.e., instantaneous and error-free) is
available for the link.
\\\indent
\textit{Decision Variables:}
We define two decision variables: one for the scheduling at the transmitter and the other for the sampling decision for Source 2. Let $\alpha_i(t) \in \{0,1\}$, $i\in \{ 1 , 2\} $, denote the transmission
decision of source $i$ at the transmitter in slot $t$, where $\alpha_i(t) = 1$ means that the
transmitter sends the packet of source $i$, and $\alpha_i(t) = 0$ otherwise.
Moreover, let $\beta(t)\in\{0,1\}$ denote the sampling decision of Source 2: $\beta(t)=1$ indicates that the transmitter takes a new sample of Source 2 in slot $t$, and $\beta(t)=0$ otherwise.
\\\indent
\textit{Age of Information:}
First, we make a common assumption (see e.g., \cite{hatami, Marian_Scheduling_Cost} and references therein) that the AoI values
are upper-bounded by a sufficiently large value $N$.
Besides tractability, this accounts for the fact that when the status information
becomes excessively stale by reaching $N$, the time-critical end application would not be affected if counting further.
Let $\theta_i(t)$ be the AoI of source $i$ at the transmitter in slot $t$, and $\delta_i(t)$ be the AoI of source $i$ at the monitor in slot $t$.
Let
$q_i(t) \in\{0,1\},\,i\in\{1,2\}, $ be an indicator for source $i$, where $q_i(t) = 1$ means that the packet of source $i$ is transmitted and successfully received in slot $t$, and ${q_i(t) = 0}$ otherwise.
Assuming that the packet arrivals occur at the beginning of slots,
the evolution of the ages for $t=1,2,\dots$, with the initial values $\theta_i(0) = 0$ and $\delta_i(0)=0$, are given as
\begin{equation}
\begin{array}{ll}\label{Eq_AoIDy}
& \theta_1(t+1)=\begin{cases}%
0, & \text{if a packet arrives in $t+1$}, \\
\minXc{\theta_1(t)+1}, & \text{otherwise},
\end{cases}
\\&
\theta_2(t+1)=\begin{cases}%
0, & \text{if}~~ \beta(t+1)=1, \\
\minXc{\theta_2(t)+1}, & \text{otherwise}.
\end{cases}
\\ &
\delta_1(t+1)=\begin{cases}
\minXc{\theta_1(t)+1}, & ~\text{if}~
q_1(t) = 1,
\\
\minXc{\delta_1(t)+1}, &
~\text{otherwise},
\end{cases}
\\ &
\delta_2(t+1)=\begin{cases}
1, & \text{if}~
q_2(t) = 1 , \beta(t) = 1,
\\
\minXc{\theta_2(t)+1}, & \text{if}~
q_2(t) = 1 , \beta(t) \neq 1,
\\
\minXc{\delta_2(t)+1 }, &\text{otherwise},
\end{cases}
\end{array}
\end{equation}
where ${\left[.\right]^N } \triangleq \min(.\,, N)$.
\\\indent
We consider that the AoI at the monitor becomes important (only) in the instants when the updates are actually needed/used by the time-critical end application.
We capture this by the QAoI \cite{Petar_QAoI_J} by defining
a per-source query flag denoted by $ { r_i(t)\in \{0,1\},\,i\in\{1,2\} },$ where $r_i(t) = 1$ means that the monitor queries the information of source $i$ in slot $t$, and $r_i(t) = 0$ otherwise.
Accordingly, the QAoI for source $i$ in slot $t$ is defined as $r_i(t) \delta_i(t)$.
We assume that the query flags are generated according to the binary Markov chains, as depicted in Fig. \ref{SM}, where $\rho_i$ and $\bar{\rho}_i$ are, respectively, the self-transition probabilities of State $1$ and State $0$.
\vspace{- 3.9 mm}
\subsection{Problem Formulation}
We denote the sum (discounted) average QAoI
at the monitor by $\bar{\delta}$, defined by
\begin{equation}
\begin{array}{ll}
& \displaystyle \bar{\delta}\triangleq
\textstyle
(1-\lambda)
\sum_{t=1}^{\infty}
{\lambda^{t-1}}
\sum_i
\mathbb{E}\left\{r_i(t)\delta_i(t)
\right\} ,
\end{array}
\end{equation}
where $\lambda\in(0,1)$ is the discount factor, and $\mathbb{E}\{\cdot\}$ is the expectation with respect to the system randomness (i.e., the random wireless channel and the packet arrival process) and the (possibly random) decision variables $\alpha_i(t)$ and $\beta(t)$ made in reaction to the past and current AoI values and query values.
Moreover,
we define the (discounted) average number of transmissions, denoted by $\bar{\alpha}$, and the (discounted) average number of sampling actions, denoted by $\bar{\beta}$, as
\begin{align}
& \displaystyle \bar{\alpha} \triangleq
(1-\lambda)
\textstyle
\sum_{t=1}^{\infty}
\lambda^{t-1}
\mathbb{E}\left\{\alpha_1(t)+\alpha_2(t)
\right\},
\\
& \displaystyle \bar{\beta} \triangleq
(1-\lambda)
\textstyle
\sum_{t=1}^{\infty}
\lambda^{t-1}
\mathbb{E}\left\{\beta(t)
\right\}.
\vspace{-1 em}
\end{align}
\vspace{-6 mm}
\\\indent
In this letter, we are interested in solving the following stochastic control problem\footnote{
As proposing the CMDP approach to the main problem \eqref{Org_P1},
the provided solutions can be extended to the average (expected) case (i.e., $\lambda = 1$) if the unichain structure \cite[Sec. 8.3.1]{Puterman_Book} exits.
However, one can show that the unichain structure does not exist for our problem, and thus,
for the average case, finding a solution could be, in general, difficult \cite[Sec. 9.2.6]{Kallenberg_MDPBook}.
}
\allowdisplaybreaks
\begin{subequations}
\label{Org_P1}
\begin{align}
\underset{\{\{\alpha_i(t)\}_{i=1}^{2},\beta(t)\}_{t=1,2,\ldots}}{\mbox{minimize}}~~~ &
\bar{\delta}
\vspace{-1em}
\\
\mbox{subject to}~~~ &
\bar{\alpha}
\le \Gamma^{\mathrm{tr}},
\label{Con_Tra1}
\\&
\bar{\beta}
\le \Gamma^{\mathrm{sm}},
\label{Con_Sam1}
\\&
\label{Cons_capacity_1}
\textstyle\sum_i \alpha_i(t)\le 1,~\forall\,t,
\\&
\label{Cons_smp_trans_S2}
\beta(t) \le \alpha_2(t), ~\forall\,t,
\vspace{-0.5 em}
\end{align}
\end{subequations}
with the decision variables ${ \{\{\alpha_i(t)\}_{i=1}^{2},\beta(t)\}_{t=1,2,\ldots} }$,
where ${ \Gamma^{\mathrm{tr}}\in(0,1] }$ and $ { \Gamma^{\mathrm{sm}} \in(0,1] } $
are limits on the
average
number of transmissions and sampling actions in the system.
The per-slot constraint \eqref{Cons_capacity_1} limits that at most one packet can be transmitted over the link in each slot. The per-slot constraint \eqref{Cons_smp_trans_S2} ensures the sampled update in slot $t$ is transmitted at the same slot.
Note that
constraint \eqref{Cons_smp_trans_S2} is not obligatory for problem \eqref{Org_P1}, but
it eliminates a sub-optimal action, i.e., taking a new sample without its concurrent transmission,
which simplifies solving the CMDP problem (introduced next).
The next section presents a solution (i.e., an optimal policy) to the main problem \eqref{Org_P1}.
\section{Introduction}
The \textit{age of information} (AoI) has been proposed to characterize the information freshness in status update systems
\cite{Roy_2012}. The AoI is defined as the time elapsed since the latest received status update packet was generated \cite{Roy_2012, AoI_Monograph_Modiano}.
\\\indent
Minimizing the AoI is mainly associated with the sampling and (transmission) scheduling optimization.
However, in some status update systems, the sampling cannot be controlled \cite{Anthony_CSMA_2021},
e.g., a sensor takes a sample whenever it harvests enough energy.
At the same time, all different types of sources (i.e., controllable or uncontrollable\footnote{Here, the controllable source refers to the generate-at-will source (i.e., new updates can be generated at any time), and the uncontrollable source refers to the stochastic arrival source whose updates arrive at random times.
}) can coexist, and they are served
using shared resources.
Besides, in practice,
sampling and transmission of a status update incur different costs,
arising a trade-off between the resource usage and freshness.
These motivate us to
address
the total cost and freshness trade-off
in a so-called \textit{heterogeneous} status update system, consisting of stochastic arrival and generate-at-will sources,
under joint sampling and transmission limitations.
\\\indent
We consider two sources,
one generate-at-will and one stochastic arrival,
with a buffer-aided transmitter communicating with a remote monitor over an error-prone channel, as depicted in Fig. \ref{SM}.
The transmitter is in charge to sample (for the generate-at-will source) and transmit updates to the monitor.
The freshness at the monitor is captured by
a generalized form of the AoI, the \textit{query AoI} \cite{Petar_QAoI_J} (QAoI),
where the query process of each source is modeled by a binary Markov chain.
We formulate a (stochastic control) optimization
problem aiming to minimize the sum average QAoI at the monitor subject to average sampling, average transmission, and per-slot transmission constraints.
We then develop optimal
polices (i.e., joint sampling and scheduling policies) by solving the optimization problem. To do so, the problem is cast as
a constrained Markov decision process (CMDP) problem and subsequently as a linear program (LP). To strike a balance between the complexity and the performance, a heuristic near-optimal policy is also developed.
The following summarizes the main contributions of the letter:\vspace{-1mm}
\begin{itemize}
\item We study the QAoI in a heterogeneous status update system under average
sampling and transmission constraints.
We provide
a query-age-optimal
policy by solving
a CMDP problem with \textit{multiple constraints}.
\item We devise a \textit{near-optimal low-complexity} policy. To do so,
we relax the per-slot transmission constraint,
cast and solve a \textit{weakly coupled} CMDP problem \cite{WC_CMDP_Mukul}, and propose a dynamic truncation algorithm.
\item Finally, simulation evaluations are conducted
revealing the effectiveness of the proposed policies.
\end{itemize}\vspace{-1mm}
\textit{Related work:}
Recently, the information
freshness has been studied from the sampling and/or scheduling optimization perspectives by using the AoI metric in, e.g., \cite{Mohammad_DC, Jnt_Smpl_NonSlot_Sched_May2022, Marian_Scheduling_Cost, Shroff_EnergyCons_Dec2020, Elif_ARQ, Shroff_sampling_2022, Zakeri_Journal_Relay, Eytan_Sch2, Yin_Sun_2, Walid_Sam_Updat1}, and using the query/request-based AoI metric in, e.g., \cite{Waiting_But_not_Agin_ACM2021,Petar_QAoI_J,hatami, Elif_QAoI_2021}.
Assuming the generate-at-will model, the sampling times in \cite{Mohammad_DC, Jnt_Smpl_NonSlot_Sched_May2022, Marian_Scheduling_Cost, Elif_ARQ, Shroff_sampling_2022, Eytan_Sch2, Yin_Sun_2, Walid_Sam_Updat1, Petar_QAoI_J, Elif_QAoI_2021}, and the sampling rate in \cite{Jnt_Smpl_NonSlot_Sched_May2022} were optimized.
The work \cite{Shroff_sampling_2022} considered a limit on the sampling frequency. The works \cite{Elif_ARQ, Zakeri_Journal_Relay, Mohammad_DC, Petar_QAoI_J} considered the transmission cost, which limits the average number of transmissions \cite{Elif_ARQ, Zakeri_Journal_Relay}.
Nevertheless, because of the scarcity of various resources (e.g., computing, power, and spectrum), sampling of the monitored process and transmission of an update \textit{both} incur significant, yet different costs.
To this end, only a few works, e.g., \cite{Jnt_Smpl_NonSlot_Sched_May2022, Marian_Scheduling_Cost, Walid_Sam_Updat1}, studied joint sampling and transmission costs. However,
all these models incorporated a \textit{weighted sum} of the costs
which has two main shortcomings: 1) setting a proper weight can be a non-trivial task in practice, and 2) there is less freedom in exclusively controlling the impact of either the sampling cost or the transmission cost.
It is worth noting that if the sampling and transmission costs are considered as two distinct constraints, as done in this letter,
casting the underlying optimization problem as a CMDP inevitably leads to a CMDP problem with multiple constraints.
Solving such a problem
is, in general, more challenging than solving the most studied single-constraint CMDP problems in, e.g., \cite{Elif_ARQ, hatami, Zakeri_Journal_Relay, Yin_Sun_2, Jnt_Smpl_NonSlot_Sched_May2022, Walid_Sam_Updat1}.
In summary,
the main distinctive feature of this letter is to consider exclusive sampling cost and transmission cost constraints in a heterogeneous status update system,
whereas the prior works have considered a single constraint to account for the both costs.
Moreover, most of the discussed works considered the AoI metric, whereas we consider the more general QAoI metric.
\section{Numerical Results}\label{Sec_Numerical_Res}
Here, we investigate the performance of the proposed policies. We set $N = 20$, $\rho_1=\rho_2=0.7$, $\bar{\rho}_1=\bar{\rho}_2=0.4$, and $\lambda = 0.99999$, and
initialize the query flags by the steady-state probabilities of the Markov chains.
For a benchmark, we consider a greedy-based baseline policy with the following decision rule:
\textit{If $\bar{D}^{\mathrm{tr}}(t)\le \Gamma^{\mathrm{tr}}$, then find ${ \underline{i}=\argmax_{i\in\{1,2\}}\{h_i(t)\} }$; otherwise, $a(t) =0$. If $\underline{i} = 1$, then $a(t) = 1$; otherwise, do the following. If $\bar{D}^{\mathrm{sm}}(t) \le \Gamma^{\mathrm{sm}}$, then $a(t) = 2_{\mathrm{n}}$; otherwise, $a(t) = 2_{\mathrm{r}}$}.
Above, $\bar{D}^{\mathrm{tr}}(t)$ and $\bar{D}^{\mathrm{sm}}(t)$ denote, respectively, the discounted average number of transmissions and sampling actions until slot $t$, and $h_i(t)$ was given by \eqref{Eq_h_Slec_Rul}.
\\\indent
Fig. \ref{Fig_tranBudget} shows the sum (discounted) average QAoI performance of the proposed policies (the optimal and the heuristic policies) and the baseline policy as a function of the transmission budget $\Gamma^{\mathrm{tr}}$ (i.e., the limit on the average number of transmissions). First, Fig. \ref{Fig_tranBudget} (as well as Fig. \ref{Fig_chnreliab}) reveals that the heuristic policy has near-optimal performance, i.e., it nearly coincides with the optimal policy (almost in all settings).
The figure shows the importance of an optimal trade-off design between freshness and resource usage, as the performance gap between the baseline policy and the proposed policies is significant when the transmission budget is small.
Moreover, the figure reveals that the QAoI is more sensitive to the transmission budget than the sampling budget, as expected.
\\\indent
Fig. \ref{Fig_chnreliab} examines the impact
of the channel reliability $p$ on the sum average QAoI performance of the
different policies for different packet arrival rate $\mu$ for Source 1.
First, the figure shows that the performance gap between the proposed policies
and the baseline policy is large, especially when the channel reliability is small.
The figure also shows that the QAoI considerably decreases as the channel reliability increases because it increases the chance that the monitor receives updates more frequently.
Moreover, it can be seen that the QAoI would not change much by increasing the arrival rate under a good channel condition.
\vspace{-1 em}
\begin{figure}[h]
\centering
\includegraphics[width=.4\textwidth]{Figures_2/QADisc_GammaTr_Gamasm-1-5_p-5_mu-6_V2.eps}
\vspace{-.5 em}
\caption{ The sum (discounted) average QAoI versus the transmission budget $\Gamma^{\mathrm{tr}} $, where $p=0.5$ and $\mu=0.6$.
\label{Fig_tranBudget}
}
\vspace{- 2 em }
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=.4\textwidth]{Figures_2/QADisc_ChRel_Gamasm-3_Gamatr-5_V1.eps}
\label{-2 em}
\caption{ The sum (discounted) average QAoI versus the channel reliability, where $\Gamma^{\mathrm{tr}} = 0.5$ and $\Gamma^{\mathrm{sm}} = 0.3$.
}
\label{Fig_chnreliab}
\vspace{-2 em }
\end{figure}
\section{ Optimal Policy }
To solve the main problem \eqref{Org_P1}, we recast it into a CMDP problem.
Then, we find an optimal policy for the CMDP problem by solving its equivalent LP \cite{Eitam_CMDP}.
\vspace{-1 em }
\subsection{CMDP Formulation of Problem \eqref{Org_P1}}\label{Sec_CMDP}
The CMDP is specified by the following elements:
\\$\bullet $
\textit{State}:
We define the state in slot $t$ by
${ \bold{s}(t)\triangleq \big (r_1(t),\theta_1(t),\delta_1(t),r_2(t),\theta_2(t),\delta_2(t)\big ) }$.
We denote the state space by $\mathcal{S}$ which is a finite set.
\\$\bullet $ \textit{Action}:
We define
the action taken in slot $t$ by $ { a(t)=(0,1,2_\mathrm{r},2_\mathrm{n}) }$, where $a(t)=0$ means that the transmitter stays idle, $a(t)=1$ means the transmitter transmits/re-transmits the packet of Source 1, $a(t)=2_\mathrm{r}$ means that the transmitter re-transmits the packet of Source 2, and $a(t)=2_\mathrm{n}$ means that the transmitter takes a new sample of Source 2 and transmits the packet.
Also, $\mathcal{A}$
is the action space of size $|\mathcal{A}|=4$.
Actions are determined by a policy, denoted by $\pi$, which is a (possibly randomized) mapping from $\mathcal{S}$ to $\mathcal{A}$.
\\$\bullet $ \textit{State Transition Probabilities}:
We denote the state transition probability from state $\bold{s}$ to next state $\bold{s}'$ under an action ${a}$ by $\mathcal{P}_{\bold{s}\bold{s}'}(a)$.
Since the evolution of the AoIs in \eqref{Eq_AoIDy} and the query processes are independent among the sources, the transition probability can be written as
$
{
\mathcal{P}_{\bold{s}\bold{s}'}(a)=\prod_{i}
\Pr\{r_i'\,|\,r_i\}
\Pr\{\bold{\underline{s}}_i'
\,\big|\,\bold{\underline{s}}_i,a\} }$,
where
$\Pr\{r_i'\,|\,r_i\}$ is given by \eqref{Eq_TP_query}, and
$\Pr\{\bold{\underline{s}}_i'\,\big|\,\bold{\underline{s}}_i,a\}$
is given by \eqref{Eq_TranPro_Unr},
in which
$\bold{\underline{s}}_i \triangleq (\theta_i,\delta_i)$ contains the current age values, as a part of the current state, associated with source $i$, and $\bold{\underline{s}}_i' \triangleq(\theta_i',\delta'_i)$ contains the next age values, as a part of the next state, associated with source $i$.
\allowdisplaybreaks
\begin{align}
\label{Eq_TP_query}
\Pr\{r_i'\,|\, r_i\} =
\begin{cases}
\rho_i, & r_i = 1;~ r'_i =1,
\\
1-\rho_i, & r_i = 1;~ r'_i =0,
\\
\bar{ \rho}_i, & r_i = 0;~ r'_i =0,
\\
1- \bar{ \rho}_i, & r_i = 0;~ r'_i =1.
\end{cases}
\end{align}
\begin{align}
\label{Eq_TranPro_Unr}
& \Pr\{\bold{\underline{s}}_1'\,\big|\,\bold{\underline{s}}_1,a\}=
\\
&
\nonumber
\begin{cases}
\mu, & a \neq 1;~ \theta'_1=0,~\delta'_1=\minXc{\delta_1+1},
\\
1-\mu, & a \neq 1;~ \theta'_1 = \minXc{\theta_1+1},\, \delta'_1=\minXc{\delta_1+1},
\\
\mu p, & a = 1;~ \theta'_1=0,~\delta'_1=\minXc{\theta_1+1},
\\
\mu (1-p), & a = 1;~ \theta'_1=0,~\delta'_1=\minXc{\delta_1+1},
\\
(1-\mu) p, & a = 1;~ \theta'_1=\minXc{\theta_1+1},~\delta'_1=\minXc{\theta_1+1},
\\
( 1-\mu)(1-p),
& a = 1;~ \theta'_1=\minXc{\theta_1+1},~\delta'_1=\minXc{\delta_1+1},
\\ 0& \text{otherwise}.
\end{cases}
\\ & \Pr\{\bold{\underline{s}}_2'~\big|~\bold{\underline{s}}_2,a\}=
\nonumber
\\ &
\nonumber
\begin{cases}
1, & a \neq \{2_\mathrm{r},2_\mathrm{n}\};\,
\theta'_2 = \minXc{\theta_2+1},\, \delta'_2=\minXc{\delta_2+1},
\\
p, & a = 2_\mathrm{r};~ \theta'_2= \minXc{\theta_2+1},~\delta'_2=\minXc{\theta_2+1},
\\
1-p, & a = 2_\mathrm{r};~ \theta'_2=\minXc{\theta_2+1},~\delta'_2=\minXc{\delta_2+1},
\\
p, & a = 2_\mathrm{n};~ \theta'_2=1,~\delta'_2=1,
\\
1-p, & a = 2_\mathrm{n};~ \theta'_2=1,~\delta'_2=\minXc{\delta_2+1},
\\ 0& \text{otherwise}.
\end{cases}
\end{align}
\\
$\bullet $ \textit{Cost Functions}:
The (immediate) cost functions include: 1) the QAoI cost defined by $ c(\bold{s}(t))= \sum_i r_i(t) \delta_i(t) $,
2) the transmission cost defined by $d^{\mathrm{tr}}(a(t))=\mathds{1}_{\{a\neq 0\}}$, and 3) the sampling cost defined by $d^{\mathrm{sm}}(a(t))=\mathds{1}_{\{a= 2_\mathrm{n}\}}$,
where
$\mathds{1}_{\{\cdot\}}$ is an indicator function which equals to $1$ when the condition in $\{\cdot\}$ holds.
\\\indent
By the above definitions,
problem \eqref{Org_P1} can
be recast as the following CMDP problem
\begin{subequations}
\label{Pro_CMDP_1}
\begin{align}
\underset{\pi\in\Pi}{\text{minimize}}~ &
(1-\lambda)
\textstyle \sum_{t=1}^{\infty}
\lambda^{t-1}
\Bbb{E}^{\pi, \eta} \left\{
c(\bold{s}(t))
\right\},
\\
\mbox{subject to}~ &
\label{Eq_Avr_Tra}
(1-\lambda)
\textstyle \sum_{t=1}^{\infty}
\lambda^{t-1}
\Bbb{E}^{\pi, \eta} \left\{ d^{\mathrm{tr}}(a(t))
\right\} \le \Gamma^{\mathrm{tr}},
\\&
\label{Eq_Avr_Smp}
(1-\lambda)
\textstyle \sum_{t=1}^{\infty}
\lambda^{t-1}
\Bbb{E}^{\pi, \eta} \left\{ d^{\mathrm{sm}}(a(t))
\right\} \le \Gamma^{\mathrm{sm}},
\end{align}
\end{subequations}
where $\Pi$ is the set of all stationary randomized policies, and $\Bbb{E}^{\pi, \eta}$ denotes the expectation when following a policy $\pi$ for a given initial distribution (over the state space) $\eta$.
\\\indent
Next, we provide an optimal policy for problem \eqref{Pro_CMDP_1}.
\vspace{-1 em}
\subsection{Linear Programming of the CMDP Problem \eqref{Pro_CMDP_1}} \label{Subsec_LP}
Here, we transform problem \eqref{Pro_CMDP_1} into an LP.
First, for a notional simplicity, we denote state $\bold{s}$ by $s$.
Let ${ x(s,a)\in\Bbb{R}^{|\mathcal{S}||\mathcal{A}|} ,~\forall\,s\in\mathcal{S},~\forall\,a\in\mathcal{A}}$, be defined as \cite[Ch. 3]{Eitam_CMDP}
\vspace{- 1 em }
\begin{align}
\nonumber
\begin{array}{ll}
x(s,a) \triangleq (1-\lambda)\sum_{t=1}^{\infty}
\lambda^{t-1}
\Pr^{\eta}{\{s(t)=s,\,a(t)=a\}},
\end{array}
\end{align}
where $\Pr^{\eta}{\{s(t)=s,\,a(t)=a\}}$ denotes the probability of taking action $a$ at state $s$ in slot $t$ given the initial distribution $\eta$.
It is noteworthy that $x(s,a),~\forall\,s\in\mathcal{S},~\forall\,a\in\mathcal{A},$ can be interpreted as the long
run discounted time that the system is in state $s$ and action
$a$ is chosen.
Then, the CMDP problem \eqref{Pro_CMDP_1} can be transformed into the following LP
\cite[Ch. 3]{Eitam_CMDP}:
\begin{subequations}\label{Prob_LP_1}
\begin{align}
\underset{\bold{X}}{\text{minimize}}~~
&
\textstyle\sum_{s\in\mathcal{S}}\sum_{a\in\mathcal{A}} x(s,a)c(s)
\\
\text{subject to}~~ &
\label{Eq_LP_Targ}
\textstyle\sum_{s\in\mathcal{S}}\sum_{a\in\mathcal{A}} x(s,a)d^{\mathrm{tr}}(a) \le \Gamma^{\mathrm{tr}},
\\& \label{Eq_LP_sm}
\textstyle\sum_{s\in\mathcal{S}}\sum_{a\in\mathcal{A}} x(s,a)d^{\mathrm{sm}}(a) \le \Gamma^{\mathrm{sm}},
\\& \nonumber
\textstyle \sum_{s'\in\mathcal{S}} \sum_{a\in\mathcal{A}} x(s',a)\big(\mathds{1}_{\{s=s'\}} - \lambda \mathcal{P}_{s's}(a) \big)
\\ & \label{Eq_LP_balance}
= (1-\lambda)\eta(s),
\forall\, s\in\mathcal{S},
\\
& \label{Eq_LP_positive}
x(s,a)\ge 0,~\forall\, s\in\mathcal{S},~\forall\, a\in\mathcal{A},
\end{align}
\end{subequations}
with variables $\bold{X}=[x(s,a)]$, where $\eta(s),\,\forall\, s\in\mathcal{S}$, is the probability that the initial state is state $s$.
By \cite[Theorem 3.3]{Eitam_CMDP},
the following theorem relates a solution of the LP \eqref{Prob_LP_1} to a solution of
problem \eqref{Pro_CMDP_1}.
\begin{theorem}
First, the optimal value of the CMDP problem \eqref{Pro_CMDP_1} equals to the optimal value of the LP \eqref{Prob_LP_1}.
Moreover,
let ${ \bold{X}^*=[x^*(s,a)] },$ be a solution of the LP \eqref{Prob_LP_1}. Then, a stationary randomized policy $\pi\triangleq\{f(s,a)\}$ with
\begin{align}
\label{Eq_Optimal_Policy}
\begin{array}{cc}
f(s,a)=
\begin{cases}
x^*(s,a)/\bar{x}_s, & \text{if}~\bar{x}_s>0, \\
\text{arbitrary}, & \text{otherwise},
\end{cases}
\end{array}
\end{align}
is an optimal policy to the CMDP problem \eqref{Pro_CMDP_1},
where $ { \bar{x}_s=\textstyle\sum_{a'\in\mathcal{A}}x^*(s,a') } $, and $f(s, a),~\forall\, s\in\mathcal{S},~\forall\, a\in\mathcal{A},$ is the probability that an action $a$ is chosen at state $s$.
\end{theorem}
\textbf{Complexity analysis:}
The complexity of finding the optimal policy given by \eqref{Eq_Optimal_Policy} comes from
the complexity of the LP \eqref{Prob_LP_1}, which is exponential in the number of sources,
as the number of variables and constraints grow exponentially with them.
Thus, the optimal policy becomes computationally inefficient when applied for large numbers of sources.
To strike a balance between the performance and complexity, in the next section, a heuristic low-complexity policy is provided.
\vspace{-0.8 em}
\section{Low-Complexity Policy }
In this section, we develop a low-complexity policy to the main problem \eqref{Org_P1} via the following procedure: 1) we relax the per-slot transmission constraint \eqref{Cons_capacity_1}, 2) transform the relaxed problem to a weakly coupled CMDP problem which is then solved by its equivalent LP
\cite{WC_CMDP_Mukul}, and 3) propose a dynamic truncation algorithm to satisfy the per-slot transmission constraint \eqref{Cons_capacity_1}.
\\\indent
We relax the per-slot constraint \eqref{Cons_capacity_1} to the (discounted) time average constraint
$ (1-\lambda)\sum_{t=1}^{\infty}
\lambda^{t-1}\mathbb{E}\{\textstyle\sum_i \alpha_i(t)\}\le 1 $.
This constraint then becomes inactive due to constraint \eqref{Con_Tra1}.
Then, the main problem \eqref{Org_P1} without constraint \eqref{Cons_capacity_1}
can be cast as a weakly coupled CMDP problem.
This is because the transition probabilities, the action space, and the immediate cost functions become independent among the sources; now, the average transmission constraint \eqref{Con_Tra1} is the only constraint coupling the sources.
\\\indent
In order to proceed, we shall define the state, action, transition probabilities, and cost functions associated with each source $i$. The state of source $i$ in slot $t$
is
${ s_i(t) \triangleq \big( r_i(t), \theta_i(t),\delta_i(t) \big) }$, and its state space is denoted by $\mathcal{S}_i$ which is a finite set.
With abuse of notation,
the action of source $i $ in slot $t$ is denoted by $a_i(t)\in\mathcal{A}_i$ with action spaces $\mathcal{A}_1=\{0,1\}$ and $\mathcal{A}_2=\{0, 2_\mathrm{r}, 2_\mathrm{n}\}$.
The state transition probabilities of source $i$, denoted
as ${ \mathcal{\tilde{P}}_{s_is_i'}^{i}(a),~\forall\, s_i,\,s_i'\in\mathcal{S}_i }$,
is given by
$ \Pr\{r_i' \,|\,r_i \}
\Pr\{\bold{\underline{s}}_i'
\,\big|\,\bold{\underline{s}}_i,a\}
$, where $\Pr\{r_i' \,|\,r_i\}$ and $
\Pr\{\bold{\underline{s}}_i'
\,\big|\,\bold{\underline{s}}_i,a\}$ were defined in
\eqref{Eq_TP_query} and \eqref{Eq_TranPro_Unr}. The QAoI cost of source $i$ is ${ c_i({s}_i(t))=r_i(t)\delta_i(t) }$.
The transmission cost of source $i$ is ${ d_i^{\mathrm{tr}}(a_i(t))=\mathds{1}_{\{a_i(t)\neq 0 \}} }$. The sampling cost of Source 2 is $ { d^{\mathrm{sm}}(a_2(t))=\mathds{1}_{\{a_2(t)= 2_n \}} }$.
\\\indent
First, for a presentation simplicity, we remove the subscript $i$ from $s_i(t)$ and $a_i(t)$.
Then, for each source $i\in\{1,2\}$, let $x_i(s,a)\in\Bbb{R}^{|\mathcal{S}_i||\mathcal{A}_i|}$ be defined as
\begin{align}
\nonumber
\begin{array}{ll}
x_i(s,a) & \triangleq (1-\lambda)\sum_{t=1}^{\infty}
\lambda^{t-1}
\Pr_i^{\eta_i}{\{s(t)=s,\,a(t)=a\}},
\\&
\nonumber
~\forall\,s\in\mathcal{S}_i,~\forall\,a\in\mathcal{A}_i,
\end{array}
\end{align}
where $\Pr_i^{\eta_i}{\{s(t)=s,\,a(t)=a\}}$ denotes the probability of taking action $a\in\mathcal{A}_i$ at state $s\in\mathcal{S}_i$ in slot $t$ given the initial distribution $\eta_i$.
Then, the corresponding LP of the weakly coupled CMDP problem is given by
\begin{subequations}\label{Prob_LP_2}
\begin{align}
\underset{\bold{X}_i}{\mbox{minimize}}~~
&\textstyle\sum_i \textstyle\sum_{s\in\mathcal{S}_i}\sum_{a\in\mathcal{A}_i} x_i(s,a)c_i(s)
\\
\mbox{subject to}~~ &
\textstyle\sum_i \textstyle\sum_{s\in\mathcal{S}_i}\sum_{a\in\mathcal{A}_i} x_i(s,a)d_i^{\mathrm{tr}}(a) \le \Gamma^{\mathrm{tr}},
\\&
\textstyle\sum_{s\in\mathcal{S}_2}\sum_{a\in\mathcal{A}_2} x_2(s,a)d^{\mathrm{sm}}(a) \le \Gamma^{\mathrm{sm}},
\\& \nonumber
\textstyle \sum_{s'\in\mathcal{S}_i} \sum_{a\in\mathcal{A}_i} x_i(s',a)\big(\mathds{1}_{\{s=s'\}} - \lambda \mathcal{\tilde{P}}^i_{s's}(a) \big)
\\ &
= (1-\lambda)\eta_i(s),
\forall\,i,~\forall\, s\in\mathcal{S}_i,
\\
& x_i(s,a)\ge 0,\forall\,i,~\forall\, s\in\mathcal{S}_i,~\forall\, a\in\mathcal{A}_i,
\end{align}
\end{subequations}
with variables $\bold{X}_i=[x_i(s,a)],~i\in\{1,2\}$.
Let $\bold{X}_i^*=[x_i^*(s,a)]$ be a solution of the LP \eqref{Prob_LP_2}.
By results of \cite{WC_CMDP_Mukul},
an optimal stationary randomized policy for each source $i$, $\pi^*_i\triangleq\{f_i(s,a)\}$, is given by
\begin{align}
\label{Eq_Dec_Optm_polcy}
\begin{array}{cc}
f_i(s,a)=
\begin{cases}
x_i^*(s,a)/\bar{x}_{i,s}, & \text{if}~\bar{x}_{i,s}>0, \\
\text{arbitrary}, & \text{otherwise},
\end{cases}
\end{array}
\end{align}
where ${ \bar{x}_{i,s}=\textstyle\sum_{a'\in\mathcal{A}_i}x_i^*(s,a')}$, and $ { f_i(s, a),~\forall\, s\in\mathcal{S}_i,~\forall\, a\in\mathcal{A}_i, }$ is the probability that an action $a\in\mathcal{A}_i$ is chosen at state $s\in\mathcal{S}_i$ for source $i$.
\\\indent
Having constructed the above weakly coupled CMDP and derived its solution,
we now propose a near-optimal heuristic policy for the main problem \eqref{Org_P1}
that operates as follows:
(i) it determines an action of each source $i$ in each slot $t$ according to $\pi^*_{i}$, and
(ii) having the actions determined, if there is more than one packet to be transmitted,
the packet of source $\underline{i}$ will be transmitted, where
${ \underline{i}=\argmax_{i\in\{1,2\}}\{h_i(t)\} }$ in which
$h_1(t)$ and $h_2(t)$ are given as
\begin{align}
\label{Eq_h_Slec_Rul}
\begin{array}{cc}
h_1(t) \triangleq r_1(t) (\delta_1(t)-\theta_1(t)),~h_2(t) \triangleq r_2(t)\delta_2 (t).
\end{array}
\end{align}
We term the above Step (ii) as a \textit{dynamic truncation algorithm}.
The main idea of the truncation algorithm is to transmit the packet that is more fresh from the monitor's perspective, as it has greater potential to decrease the AoI at the monitor.
To capture such relative freshness, we have used the AoI difference between the monitor and the transmitter for Source 1, and the AoI at the monitor itself for Source 2.
\\\textbf{Complexity analysis:}
The main complexity of finding the heuristic policy comes from solving the LP
\eqref{Prob_LP_2}, which is (only) \textit{linear} in the number of sources. Thus, the proposed policy has low complexity,
and as numerically shown in the next section, obtains
near-optimal performance.
\section{Query AoI}
\begin{figure}[h]
\centering
\includegraphics[width=.5\textwidth]{Figures_2/SM_Query_V3.png}
\caption{Basic query AoI setup}
\label{Fig_SM_Query}
\end{figure}
\subsection{Previous Models of the Query Process}
\begin{itemize}
\item Memoryless model: considering, e.g., i.i.d. (over slots and across sensors in the multiple sensor setup) Bernoulli distribution. Let $r(t)\in\{0,1\}$ indicates the query flag, then ${\Pr\{r(t) = 1\}=\eta,~\forall \,t}$.
\item The query process is represented by a finite Markov chain with state space $\mathcal{S}_q$, then the query flag $r(t)$ equals to $1$ if at time $t$ the state of the Markov chain is in a predetermined subset ${\mathcal{Q}\subseteq \mathcal{S}_q}$. An example of this case is the case in which queries
arrive at stochastic i.i.d. intervals with a known distribution.
\end{itemize}
The above models:
\begin{itemize}
\item did not consider the effect of the underlying system's randomness, e.g., the channel reliability, and even the control actions.
\\\indent
To make this be clear, consider an example in which the Bernoulli model is used with $\eta=1/2$ and the perfect query link. But the transmission link is error-prone with reliability $p$ (see Fig. \ref{Fig_SM_Query}). Now consider two different values of $p$: (1) $p=0.1$ and (2) $p=0.9$. For both cases, the query mode is the same. However, it is more logical to increase the query rate (i.e., $\eta$) for the first case $p=0.1$ since in this case (even without any resource limitations) more updates are unsuccessful; hence in most of times, the AoI at the monitor is large, but we care all of them with probability $1/2$, e.g., if the AoI at the monitor is $100$, the QAoI is 0 with the half probability meaning that the monitor is fresh which is not true.
\\\indent More precisely, the above models do not have a closed-loop feedback from the environment (and even control actions) to customize/modify/update/adopt the query process (in an online fashion) in order to improve the awareness of the monitor.
\item \textit{The query link is assumed to be perfect}, which is not always the case. Consider a case in which the queries are experienced random delays. This situation can be another reason for considering an efficient model of the query process to deal such cases.
\end{itemize}
Besides the query models, in the QAoI-oriented papers, the model of the status updates are the generate-at-will (without sampling constraints). Also, it seems the query process should be revisited for the stochastic arrival models.
Next describes the proposed model.
\subsection{Modulated Markov Process to model the query process (proposed)}
We model the query process as follows. The random query process $r(t)$ is modulated by a Markov chain (MC) with a finite state space $\mathcal{S}$ as follows: whenever the state of MC is an arbitrary state $s\in\mathcal{S}$, the query process follows independently with some distribution. This is realized by the following idea.
\\\indent
Let denotes the AoI at the monitor at slot $t$ by $\delta(t)$.
First, we divide the AoI at the monitor into 3 main regions each represents a state in the Markov chain as follows:
\begin{enumerate}
\item $\delta(t)\le C_{1}$ representing low-AoI state,
\item $C_{1}< \delta(t)\le C_{2}$ representing medium-AoI state,
\item $C_{2}< \delta(t)$ representing high-AoI state,
\end{enumerate}
where $C_1 < C_2$ are some positive integer constants and could depend on applications.
Thus the Markov chain has three states, i.e., $\mathcal{S}=\{1,2,3\}$. Now we can use any distribution for the query process at each of these three states. For example, if $s=1$ we use the Bernoulli distribution with constant parameter $\eta$ and if $s=2$ we use the Bernoulli distribution with time-varying parameter $\eta(t) = 1- \frac{1}{\delta(t)}$ and so on.
\\\indent
As it can be understood, we have used the AoI at the monitor to update the model of query process, which can be interpreted as a close-loop feedback from the environment since the AoI is affected by the environment and our control actions.
Notice that the model is quite general than the existing models since the others are a special case of this by setting that $C_1=\infty$ (means that the MC has one state) and then following the adopted query models.
\section{Application-Level Freshness: A Model for Correlated Information}
\begin{figure}[h]
\centering
\includegraphics[width=.5\textwidth]{Figures_2/SM_AL_V1.png}
\caption{Basic application-level AoI setup}
\label{Fig_SM_Query}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=.5\textwidth]{Figures_2/SM_AL_QAoI_V1.png}
\caption{Basic application-level QAoI setup}
\label{Fig_SM_Query}
\end{figure}
\section{Separate Look at the Source 2 (Just in Case)
}
\begin{figure}[h]
\centering
\includegraphics[width=.5\textwidth]{Figures_2/SM_AoII_V1.png}
\caption{A two states Morkovian source}
\label{Fig_SM_2}
\end{figure}
\subsection{Main Assumptions of System Model Fig. \ref{Fig_SM_2}}
\begin{itemize}
\item The sensor can only observe the source at the sampling times
\item The transition probability matrix of the source is known
\item The control actions could be $\{0,n,r\}$, where $n$ means new sample and transmission, $r$ means re-transmission (we can keep the last sample).
\item Each sampling takes unit of the sampling cost
\item Each transmission incurs unit of the transmission cost
\item The metric is the age of incorrect information
\end{itemize}
\subsection{The Main Difference Between the System Model in Fig. \ref{Fig_SM_2} and [R]}
\begin{itemize}
\item Here we have a sampling cost, which can limits the average number of sampling
\item In contrast of just having two actions $\{0, 1\}$ in [R], we have $\{0,n,r\}$ which enables retransmission mechanisms and possibly better performance.
\end{itemize}
\section{What can we consider on top of [R]}
\section{Proposed System Model}
We can consider system shown in Fig. \ref{Fig_SM_1} with the following assumptions:
\begin{itemize}
\item The controller can only observe the process of Source 2 at the sampling times
\item The transition probability matrix of the sources are known
\item The most recent information based on the last received packet by the transmitter is the current estimate of the process at the transmitter
\item The most recent information based on the last received packet by the monitor is the current estimate of the process at the monitor
\item Each sampling takes unit of the sampling cost
\item Each transmission incurs unit of the transmission cost
\item The metric is the age of incorrect information
\item The control actions for Source 1 is being idle or transmitting the packet
\item The control actions of Source 2 are: 1) being idle, 2) taking a new sample, and 3) (re)transmission.
\item The goal is minimizing sum age of incorrect information subject to average transmission and average sampling constraint and per-slot capacity constraint
\end{itemize}
\begin{figure}[t]
\centering
\includegraphics[width=.6\textwidth]{Figures_2/SM_AoII_V3.png}
\caption{Proposed system model.}
\label{Fig_SM_1}
\end{figure}
\section{The Main Contributions}
\begin{itemize}
\item Heterogeneous status update system
\item For each source we have different setup; the stochastic arrival model was not studied yet and for the generate-at-will source we have sampling cost
\item Average transmission and average sampling constraints
\end{itemize}
\section{Proposed Problem Formulation For Setup in Fig. \ref{Fig_SM_1}}
\begin{itemize}
\item
Let $\alpha_i(t)\in\{0,1\}$, $i\in\{1,2\}$, denote the transmission decision of source $i$, where $\alpha_i(t)=1$ means that the transmitter sends the packet of source $i$, and $\alpha_i(t)=0$ means that the transmitter is idle.
\item
Let $\beta(t)\in\{0,1\}$, denote the sampling decision of source 2, where $\beta(t)=1$ means that the transmitter takes a new sample of source 2, and $\beta(t)=0$ otherwise.
\item
Let $\delta_i(t)$ denote the age of incorrect information (AoII) of source $i$.
\end{itemize}
Let
$\mathcal{D}=\left\{\alpha_i(t), \beta(t)\right\}_{t=1,2,3,\ldots}$
be a sequence of decision variables in the system, then define the following:
\begin{itemize}
\item
The \textit{sum average AoII} is denoted by $\bar{\delta}(\mathcal{D})$ and given by
\begin{equation}
\begin{array}{ll}
\nonumber
& \displaystyle \bar{\delta}(\mathcal{D}) \triangleq\limsup_{T\rightarrow \infty} ~\frac{1}{T}
\textstyle
\sum_{t=1}^{T}
\mathbb{E}\left\{\delta_1(t)+\delta_2(t)
\right\}.
\end{array}
\end{equation}
\item
The average number of transmissions is given by
\begin{equation}
\begin{array}{ll}
\nonumber
& \displaystyle \bar{\alpha}(\mathcal{D}) \triangleq\limsup_{T\rightarrow \infty} ~\frac{1}{T}
\textstyle
\sum_{t=1}^{T}
\mathbb{E}\left\{\alpha_1(t)+\alpha_2(t)
\right\}.
\end{array}
\end{equation}
\item
The average number of sampling (for Source 2) is given by
\begin{equation}
\begin{array}{ll}
\nonumber
& \displaystyle \bar{\beta}(\mathcal{D}) \triangleq\limsup_{T\rightarrow \infty} ~\frac{1}{T}
\textstyle
\sum_{t=1}^{T}
\mathbb{E}\left\{\beta(t)
\right\}.
\end{array}
\end{equation}
\end{itemize}
We now formulate the following problem
\begin{tcolorbox}
\begin{subequations}
\label{Org_P1}
\begin{align}
\underset{\mathcal{D}}{\mbox{minimize}}~~~ &
\bar{\delta}(\mathcal{D})
\vspace{-1em}
\\
\mbox{subject to}~~~ &
\bar{\alpha}(\mathcal{D})
\le \Gamma^{\mathrm{tr}},
\label{Con_Sam1}
\\&
\bar{\beta}(\mathcal{D})
\le \Gamma^{\mathrm{sm}},
\label{Con_Sam1}
\\&
\textstyle\sum_i \alpha_i(t)\le 1.
\end{align}
\end{subequations}
\end{tcolorbox}
\section{\textcolor{blue}{Separate Look at the Stochastic Arrival Source}}
\begin{figure}
\centering
\includegraphics[width=.8\textwidth]{Figures_2/SM_AoII_SingleS_SA_V3.png}
\caption{Stochastic arrival source}
\label{Fig_SS_1}
\end{figure}
Let define $\rho(t)\in\{0,1\}$ be an indicator denoting the state of being $X(t) = \hat{X}(t)$, which is defined by
\begin{align}
\begin{array}{cc}
\rho(t) =
\begin{cases}
0 & \text{if}~~ X(t) = \hat{X}(t),
\\
1 & \text{if}~~ X(t) \neq \hat{X}(t).
\end{cases}
\end{array}
\end{align}
The indicator $\rho(t)$ is a stochastic process and its dynamic can be shown by a two-state Markov chain.
\\
If we observe $X(t)$ then $\rho(t)$ becomes deterministic. Thus, to model the belief we need to define probabilities for all possible values of $\rho(t)$.
\\
Let us define the following for the setup shown in Fig. \ref{Fig_SS_1}:
\begin{itemize}
\item $a(t)\in\{0,1\}$: is the transmission action at the transmitter
\item $u(t)\in\{0,1\}$: be the successful packet reception by the monitor and $\Pr\{u(t)=1\}=q$
\item $\pi(t)$: the probability of $\hat{X}(t) = X(t)$, i.e., $\pi(t) = \Pr\{\hat{X}(t) = X(t)\}$
\item $\tilde{\pi}(t)$: the probability of $\tilde{X}(t) = X(t)$, i.e., $\tilde{\pi}(t) = \Pr\{\tilde{X}(t) = X(t)\}$
\item $\delta(t)$: is the belief value of AoII, which can be characterized based on $\pi(t)$
\end{itemize}
We have the following:
\begin{align}
\label{Eq_pi_t}
\begin{array}{cc}
\pi(t +1) =
\begin{cases}
p\tilde{\pi}(t) + (1-p) (1-\tilde{\pi}(t))& \text{if}~ a(t) = 1, u(t)=1,
\\
p{\pi}(t) + (1-p)(1-{\pi}(t)) & \text{otherwise},
\end{cases}
\end{array}
\end{align}
\begin{align}
\begin{array}{ll}
\tilde{\pi}(t + 1 ) & = p\mu + p(1-\mu)\tilde{\pi}(t) + (1-p)\mu + (1-p) (1-\mu) (1-\tilde{\pi}(t))
\\
& = \mu + (1-\mu)(p\tilde{\pi}(t) + (1-p)(1-\tilde{\pi}(t))
\\
&= \mu + (1-\mu)(1-p+(2p-1)\tilde{\pi}(t)).
\end{array}
\end{align}
***\textcolor{blue}{Some numerical tests show that the sequence $x(t)_{\{t=0,1,2,\dots\}}$ converge to $\frac{1-p(1-\mu)}{1-(2p-1)(1-\mu)}$, for any $p,\mu,\tilde{\pi}(0)$. *** }
Moreover, the dynamic of $\delta(t)$ can be formulated as
\begin{align}
\label{Eq_delta_blf}
\begin{array}{cc}
\delta(t + 1) =
\begin{cases}
0 & \text{w.p.}~~ \pi(t+1),
\\
\delta(t) + 1 & \text{w.p.}~~ 1- \pi(t+1).
\end{cases}
\end{array}
\end{align}
with $\pi(0) = \tilde{\pi}(0) = 1$ and $ \delta(0)=0$.
Note that, for the fully observable case, $\pi(t)$ and $\tilde{\pi}(t)$ takes values 0 or 1 certainly.
Based on these definitions, we formulate the following problem
\begin{tcolorbox}
\begin{subequations}
\label{Org_P12}
\begin{align}
\underset{\mathcal{D}}{\mbox{minimize}}~~~ &
\limsup_{T\rightarrow \infty} ~\frac{1}{T}
\textstyle
\sum_{t=1}^{T}
\mathbb{E}\left\{\delta(t)
\right\}
\vspace{-1em}
\\
\mbox{subject to}~~~ &
\limsup_{T\rightarrow \infty} ~\frac{1}{T}
\textstyle
\sum_{t=1}^{T}
\mathbb{E}\left\{a(t)
\right\}
\le \Gamma^{\mathrm{tr}},
\label{Con_Trans}
\end{align}
\end{subequations}
\end{tcolorbox}
where
$\mathcal{D}=\left\{a(t)\right\}_{t=1,2,3,\ldots}$ be a sequence of the transmission decision variables at the transmitter.
\section{System Model and Formulation}
As shown in Fig. \ref{SM}, our considered system consists of a sensor (source), a monitor (destination), and a dual-band communication channel, i.e., channel between sensor-to-monitor and monitor-to-sensor\footnote{Commonly sensor-to-monitor channel is data channel and monitor-to-sensor channel is feedback channel.}.
\begin{figure}
\centering
\includegraphics{SystemModel.png}
\caption{System model}
\label{SM}
\end{figure}
We assume the continuous-time is slotted with the same and fixed slot time duration denoted by $\delta$. Each slot indexed by $t\in\Bbb{N}$. The sensor observes underlaying process which is modeled as discrete-time stochastic process as follows:
\begin{align}
\begin{split}
\label{Proc_Mod}
&X(t+1)=f(X(t))+W(t),
\end{split}
\end{align}
where $X(t)$ is the process and $W(t)$ is a stationary stochastic noise. $f(.)$ is an unknown function and can be a linear or non-linear function.
The model considered in \eqref{Proc_Mod} is more general and practical in comparison to linear model \cite{RE_twc}
and can be applied for various systems and applications such as autonomous driving. The monitor has not direct access to the process and only has some samples of the process\footnote{Since each of two channels incurs a transmission delay, the sample's time on the monitor side is different from the reception time.}.
The monitor interest to fresh against to the process as possible. Let us, define that the monitor is at the maximum freshness, if it can reproduces the process $X(t)$.
\\\indnet
Let $S_{i}(t)\in\{0,1\}$ be the sampling action, if $i$-th sample is generated and transmitted
to the monitor at slot $t$, $S_i(t)=1$, otherwise $S_i(t)=0$. Notable $S(t)=0$ means the sensor is idle (or waiting) at slot $t$. Let $Y_{i}(t')$ denotes $i$-th received sample at slot $t'$ and is given by
\begin{align}
Y_i(t')=g(X(t))+V_i(t),~t'=t+D_i,
\end{align}
where $V_i(t)$ is a stationary stochastic noise which models the noise of channel and $D_i$ is discrete time variable captures the delay of channel \cite{Star2}. $g(X(t))=\mathds{1}_{\{S(t)=1\}}X(t)$, where $\mathds{1}_{\{\cdot\}}$ is the indicator function. As mentioned earlier, if the monitor can reproduce the process $X(t)$, meaning that all update information of the process in the sensor side is also available in the monitor side.
\\\indnet
Let $\bold{W}_{t}=\{(Y_{i}(t'),M): t'\le t\}$, indicates $M$ most recently samples that have been existed (both received and estimated) to the monitor until time $t$. By this, if a algorithm exits that uses these sample to estimate the next value of the process, i.e., $X(t+1)$ without getting its value from the sensor, we can say system is still fresh.
Let $\hat{X}(t)$ denotes the estimation of $X(t)$ at time $t$. This estimation is based on the samples of $Y_i(t)$, generally $\bold{W}_{t}$.
Hence, we assume that $\hat{X}(t')=H(\bold{W}_{t}),t\le t'$, where $H(.)$ is a estimator function. Since in our model each transmitted sample experience a unknown random delay, it is impossible to have $\hat{X}(t')=X(t), \forall~ t'=t$, even the the value of $X(t)$ is transmitted, i.e., $S(t)=1$.
One typical estimator design is based on the minimizing the mean of square error defined as $(e(t))^2$, where $e(t)=\tilde{X}(t)-X(t)$. The minimum value of mean square errors implies that the monitor can accurately reproduce the process, hence the system is fresh. Estimation error commonly depends on the quality and quantity of available samples and the performance of the estimator.
It is worthwhile to mention that we do not make any further assumptions about the process model \eqref{Proc_Mod} as well as $V_i(t)$, e.g., Gaussian distribution. The main limitation of the mathematical solutions is considering some further assumptions on the model. This motivates us to (i) transform a general freshness management problem into a reinforcement learning framework and (ii) apply advance deep RL algorithm to handle the problem. We discuss these in the following.
\textcolor{blue}{Therefore, we intend to design a scheduling policy in order to minimizing the average estimation error under a maximum allowable sampling cost.
Formally, we interest in the following optimization problem:
\begin{align}
\min_{\bold{S}}
\end{align}
}
\section{The Main Questions and Discussions }
\input{Chapters/SystemModel+Problem}
\input{Chapters/Solution}
\input{Chapters/Simulation}
\section{Conclusion}
In this letter, we provided a query-age-optimal joint transmission and sampling scheduling policy for a heterogeneous status update system under average sampling, average transmission, and per-slot transmission constraints.
To this end, a CMDP problem with two average constraints was cast and solved via its equivalent LP.
We also provided a low-complexity near-optimal heuristic policy using the notion of the weakly coupled CMDP problem and its equivalent LP.
Numerical results showed the effectiveness of
the proposed policies, revealing that age-optimal sampling and scheduling is crucial for
resource-constrained status update systems, where
a greedy-based policy is inefficient.
Moreover, the near-optimal performance of the proposed heuristic policy substantiates that the LP approach provides a viable solution for practical systems having many sources/sensors.
\vspace{-1.5 em }
\bibliographystyle{ieeetr}
|
\section{{Introduction\label{sec:int}}}
Consider a sequence of observations $x_{1},x_{2},...$ whose distribution is
governed by a parameter $\theta,$ in a standard Bayesian setup with a prior
distribution $\pi$ on the space of parameters $\Theta$. For simplicity assume
for now that $\Theta$ and the space of observations are finite (the results
extend readily to general observations and parameter spaces, as we will see later).
Let ${{\mathbb{P}}}_{\theta}$ denote the probability distribution under the
parameter $\theta,$ and let ${{\mathbb{P}}}=\sum_{\theta\in\Theta}\pi
(\theta){{\mathbb{P}}}_{\theta}$ denote the marginal probability; the
corresponding expectations are denoted by ${{\mathbb{E}}}_{\theta}$ and
${{\mathbb{E}}},$ respectively (thus ${{\mathbb{P}}}_{\theta}({{\mathbb{\cdot
)\equiv P(}}}\cdot|\theta)$ and ${{\mathbb{{{\mathbb{E}}}}}}_{\theta
}\mathbb{[}{{\mathbb{\cdot]\equiv{{\mathbb{E}}}}}}[\cdot|\theta]).$
Let $\theta_{0}$ in $\Theta$ be a fixed value of the parameter. We are
interested in the way the belief in $\theta_{0}$ varies as one gets more and
more observations, i.e., as $n$ increases. We denote by $q_{n}^{\theta_{0}}$
the \emph{posterior} probability of $\theta_{0}$ at time $n;$ i.e., for every
sequence $s_{n}=(x_{1},...,x_{n})$ of observations up to time $n,$
\[
q_{n}^{\theta_{0}}\equiv q_{n}^{\theta_{0}}(s_{n})
{\;:=\;}
{{\mathbb{P}}}(\theta_{0}|s_{n})=\frac{{{\mathbb{P}}}_{\theta_{0}}(s_{n}%
)\pi(\theta_0)}{{{\mathbb{P(}}}s_{n})}.
\]
As is well known, the sequence $q_{n}^{\theta_{0}}$ of posteriors is a
martingale with respect to the marginal probability ${{\mathbb{P}}}$, i.e.,
\begin{equation}
{{\mathbb{E[}}}q_{n+1}^{\theta_{0}}|s_{n}]=q_{n}^{\theta_{0}}(s_{n}%
)\label{eq:martingale}%
\end{equation}
for every $n$ and $s_{n}.$ Thus, given $s_{n},$ a new observation $x_{n+1}$
distributed according to ${{\mathbb{P}}}$ may increase or decrease the
posterior of $\theta_{0},$ but on average this posterior does not change.
Posterior probabilities are used in Bayesian hypothesis testing, where the
decision between hypotheses depends on their posterior probabilities (see
\cite{Berg}, Section 4.3.3). In sequential Bayes decision rules, monotonicity
properties of posteriors determine situations where more data lead to better
decisions vs. ones where more data could at times be misleading. This question
was raised and discussed in \cite{FK}. For a discussion of sequential Bayesian
inference and references see, e.g., \cite{Ferg} and \cite{Berg}. Setups where
more data may be harmful according to certain criteria of the
statistician are given, for example, in \cite{Barron} and \cite{Gru}. In
addition to the Bayesian setup, the questions addressed here are pertinent to
setups with agents that have different prior beliefs on the parameters. See
\cite{HR} for details, including applications to models with informed agents
vs. an uninformed market, and reputation-building models.
Suppose that the true parameter is $\theta_{0}$ (but this is, of course,
unknown to the observer). What can one say about the sequence of $\theta_{0}%
$-posteriors under ${{\mathbb{P}}}_{\theta_{0}}$? When the observations
$x_{n}$ are iid and the distributions ${{\mathbb{P}}}_{\theta}$ are distinct
(i.e., ${{\mathbb{P}}}_{\theta}\neq{{\mathbb{P}}}_{\theta^{\prime}}$ for
$\theta\neq\theta^{\prime},$ referred to as \textquotedblleft identifiable
parameters"), the Doob consistency theorem (see, e.g., \cite{VaV} and
\cite{Mill}) says that under ${{\mathbb{P}}}_{\theta_{0}}$ the sequence of
$\theta_{0}$-posteriors $q_{n}^{\theta_{0}}$ converges to $1$ almost surely
(i.e., except in a ${{\mathbb{P}}}_{\theta_{0}}$-null set). In fact, this
happens monotonically (see \cite{FK} and \cite{HR}, and the references
therein), in the sense that under ${{\mathbb{P}}}_{\theta_{0}}$ the posterior
of $\theta_{0}$ always increases on average; that is,
\begin{equation}
{{\mathbb{E}}}_{\theta_{0}}[q_{n+1}^{\theta_{0}}|s_{n}]\geq q_{n}^{\theta_{0}%
}(s_{n})\label{eq:ge-mart}%
\end{equation}
for every $n$ and $s_{n}$ (recall that ${{\mathbb{E}}}_{\theta_{0}}%
\mathbb{[}{{\mathbb{\cdot]\equiv{{\mathbb{E}}}}}}[\cdot|\theta_{0}]$ stands
for the expectation with respect to ${{\mathbb{P}}}_{\theta_{0}}$). This
submartingale inequality means that \emph{each additional observation
increases on average the posterior of the true parameter, under the
probability law of the true parameter.}
Taking the posterior probability of $\theta_{0}$ as one's belief in the model
determined by $\theta_{0}$, the expected belief in $\theta_{0}$ increases with
more data generated under $\theta_{0}$. Thus, as stated in \cite{FK}, Bayesian
inference does not lead one astray on average. In \cite{HR} we show that this
is in fact a consequence of an even stronger result, which holds for any
observation (and thus, in particular, for $x_{n+1}$ after $s_{n}$): the
distribution of the $\theta_{0}$-posterior under ${{\mathbb{P}}}_{\theta_{0}}$
dominates the distribution of that same $\theta_{0}$-posterior under the
marginal probability ${{\mathbb{P}}},$ where the domination is in the
likelihood ratio order, which is a strengthening of the usual stochastic order; see
Sections \ref{sec:sing} and \ref{sec:monot}. We thus get the submartingale
inequality (\ref{eq:ge-mart}), from which it follows that the overall
expectation ${{\mathbb{E}}}_{\theta_{0}}[q_{n}^{\theta_{0}}]\equiv
{{\mathbb{E}}}_{\theta_{0}}[{{\mathbb{P}}}({\theta_{0}}|s_{n})]$ of the
$\theta_{0}$-posterior $q_{n}^{\theta_{0}}$ under the probability
$\mathbb{P}_{\theta_{0}}$ is an increasing function of the number of
observations $n.$
Now suppose that the true parameter, $\theta_{1},$ is different from
$\theta_{0}.$ In the above case of iid observations and identifiable
parameters, the Doob consistency theorem now says that under ${{\mathbb{P}}%
}_{\theta_{1}}$ the sequence of $\theta_{0}$-posteriors $q_{n}^{\theta_{0}}$
converges to $0$ almost surely. When $\theta_{0}$ and $\theta_{1}$ are the
only possible parameter values (i.e., $\Theta=\{\theta_{0},\theta_{1}\}),$
by (\ref{eq:martingale}), (\ref{eq:ge-mart}), and ${{\mathbb{P}}}$ being the
average of ${{\mathbb{P}}}_{\theta_{0}}$ and ${{\mathbb{P}}}_{\theta_{1}},$ it
immediately follows that ${{\mathbb{E}}}_{\theta_{1}}[q_{n+1}^{\theta_{0}%
}|s_{n}]\leq q_{n}^{\theta_{0}}(s_{n});$ thus, each additional observation
decreases on average the posterior of a \textquotedblleft false" parameter,
under the probability law of the true parameter. However, this seemingly
natural property need \emph{not} hold when there are more than two possible
values of $\theta$ (see \cite{HR} for a simple example, and Section
\ref{sec:nonmonmult} below).
We next turn to consider monotonicity as a function of $n$ of $\psi(n)%
{\;:=\;}%
{{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]\equiv{{\mathbb{E}}}%
_{\theta_{1}}[{{\mathbb{P}}}({\theta_{0}}|s_{n})]$, the expectation over
$s_{n}$ of the $\theta_{0}$-posterior $q_{n}^{\theta_{0}}$ under the
probability ${{\mathbb{P}}}_{\theta_{1}},$ where $\theta_{1}\neq\theta_{0}.$
Consider for concreteness the simple setup of iid coin tosses. By Doob's
consistency result, $\psi(n)\rightarrow0$ as $n\rightarrow\infty$. This
convergence to zero need not however be monotonic. Indeed, when the true
parameter $\theta_{1}$ is \textquotedblleft close" to $\theta_{0}$ in a
suitable sense (which we will quantify in Proposition \ref{cor:uni}), it is
natural for data under $\theta_{1}$ to strengthen the belief in $\theta_{0}$
at first (i.e., for small $n),$ and so for $\psi(n)$ to increase for small $n$
(as is indeed the case when $\theta_{1}=\theta_{0}$; see (\ref{eq:ge-mart})).
Eventually, however, $\psi(n)$ must approach zero. Once $\psi(n)$ starts
decreasing in $n$, suggesting that evidence against $\theta_{0}$ is mounting,
can $\psi(n)$ increase again? While there may well be particular
realizations after which the $\theta_{0}$-posterior goes up (i.e.,
$q_{n+1}^{\theta_{0}}>q_{n}^{\theta_{0}}),$ and perhaps even particular data
$s_{n}$ after which the $\theta_{0}$-posterior is expected to go up (i.e.,
${{\mathbb{E}}}_{\theta_{1}}[q_{n+1}^{\theta_{0}}|s_{n}]>q_{n}^{\theta_{0}%
}(s_{n})),$ the \emph{overall expectation} is expected to be well behaved and
continue to go down (i.e., ${{\mathbb{E}}}_{\theta_{1}}[q_{n+1}^{\theta_{0}%
}]\leq{{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]).$ Perhaps surprisingly,
that is \emph{not} the case: we provide simple examples where $\psi(n)$ is
\emph{not} unimodal in $n$ and may have multiple local maxima; that is, it can
go down and then up many times. Thus, after $\psi(n)$ starts decreasing and
the statistician may begin to doubt that $\theta_{0}$ is the true parameter,
the increase in $\psi(n)$ with more observations strengthens the
statistician's wrong belief that $\theta_{0}$ is the true parameter; the new
observations \emph{do} \textquotedblleft lead one astray on average." While
this is not a knife-edge phenomenon and may indeed happen, we show that
certain natural assumptions rule it out: the expected posterior is eventually
decreasing, and even log-concave and thus unimodal.
We now summarize the results on the behavior of $\psi(n)\equiv{{\mathbb{E}}%
}_{\theta_{1}}[q_{n}^{\theta_{0}}]$:
\begin{enumerate}
\item When $\theta_{1}=\theta_{0},$ the sequence $\psi(n)$ is increasing in
$n$ (as stated above, this is a consequence of the likelihood ratio dominance,
which in turn implies the inequality (\ref{eq:ge-mart}); see Sections
\ref{sec:sing} and \ref{sec:monot}).
\item When $\theta_{1}\neq\theta_{0},$ the sequence $\psi(n)$ may increase in
some range of values of $n$ and decrease in others, and may, for example,
decrease first, then increase, and then decrease again (which, as we will show in
Section \ref{sec:normnorm}, can happen already in the case of iid normal
observations with a normal prior). Moreover, a large number of modes (i.e.,
local maxima) can occur in the simple case of iid Bernoulli coin tosses (see
Section \ref{sec:nonmonmult}).
\item The sequence $\psi(n)$ is asymptotically equivalent to $C\sqrt{n}%
\,w^{n}$ as $n\rightarrow\infty$ (for constants $C>0$ and $0<w\leq1)$ in the
case of exponential families of distributions (discrete and continuous) with a
continuous prior, where $w=1$ when $\theta_{1}=\theta_{0},$ and $w<1$ when
$\theta_{1}\neq\theta_{0};$ in the latter case $\psi(n)$ is therefore strictly
decreasing from some $n$ on (see Section \ref{sec:asympt}).
\item The sequence $\psi(n)$ is log-concave in $n,$ and thus unimodal, in a
number of scenarios: iid coin tosses with a uniform prior, iid normal
observations with a normal prior (in a wide region of parameters), and iid
exponential observations with an exponential prior (see Section
\ref{sec:loggcon}).
\end{enumerate}
The paper is organized as follows. In Section \ref{sec:sing} we discuss
various order relations between posteriors under different distributions. In
Section \ref{sec:monot} we extend these results to sequences of observations
and see that $\psi(n)$ is increasing when $\theta_{1}=\theta_{0}.$ In Section
\ref{sec:nonmonmult} we exhibit situations in which $\psi(n)$ for $\theta
_{1}\neq\theta_{0}$ is not unimodal in $n$, and quantify, for Bernoulli
observations, the notion of $\theta_{0}$ and $\theta_{1}$ being close together such that
an observation under $\theta_{1}$ increases the belief in $\theta_{0}$ on
average. Section \ref{sec:asympt} provides the asymptotic analysis of
$\psi(n),$ and Section \ref{sec:loggcon} deals with cases where the sequence
$\psi(n)$ is unimodal and even log-concave. For Bernoulli observations this is
obtained by proving in Section \ref{sec:revtur} a reversal of the reverse
Tur\'{a}n inequality for orthogonal polynomials, which is of independent
interest. The Appendix contains discussions on possible extensions of the
results as well as some technical details.
\section{Preliminaries: A single signal}
\label{sec:sing}
In this section we summarize results on various order
relations for posterior distributions that appeared with a somewhat different
emphasis in \cite{HR}. We consider a signal $s$, which can be a single
observation as well as multiple ones. For simplicity we now consider discrete
random variables $s$ and finite parameter spaces $\Theta$. The extension to
continuous random variables and prior distributions is straightforward; see
Appendix \ref{sec:general} and Section \ref{sec:asympt}.
We use the following standard notions and notation. The random variables $x$
and $y$ satisfy $y\geq_{{{\mathrm{st}}}}x$ (\emph{stochastic order}; also
known as \emph{first-order stochastic dominance}) or $y\geq_{{{\mathrm{icx}}}%
}x$ (\emph{increasing convex order}) if ${{\mathbb{E[}}}f(y)]\geq
{{\mathbb{E[}}}f(x)]$ for every increasing function or increasing convex
function $f$, respectively. By ``increasing" or ``convex" we do not mean
\textquotedblleft strictly" unless otherwise stated. The probability law (distribution) of a random
variable $x$ is denoted by ${\cal{L}}(x)$ (formally, ${\cal{L}%
}(x)=\mathbb{P}\circ x^{-1})$. Instead of $y\geq_{{{\mathrm{st}}}}x$ we may
write ${\cal{L}}(y)\geq_{{{\mathrm{st}}}}{\cal{L}}(x)$.
The \emph{likelihood ratio order}, denoted by $y\geq_{{{\mathrm{lr}}}}%
x$\thinspace or ${\cal{L}}(y)\geq_{{{\mathrm{lr}}}}{\cal{L}}(x)$,
is said to hold if ${{\mathbb{P}}}(y=t)/{{\mathbb{P}}}(x=t)$ is increasing in
$t$, or equivalently, ${{\mathbb{P}}}(y=t^{\prime}){{\mathbb{P}}}%
(x=t)\geq{{\mathbb{P}}}(y=t){{\mathbb{P}}}(x=t^{\prime})$ for all $t^{\prime
}>t$. This is stronger than the stochastic order: ${\cal{L}}%
(y)\geq_{{{\mathrm{lr}}}}{\cal{L}}(x)$ implies ${\cal{L}}%
(y)\geq_{{{\mathrm{st}}}}{\cal{L}}(x)$. Moreover, ${\cal{L}%
}(y)\geq_{{{\mathrm{lr}}}}{\cal{L}}(x)$ implies ${\cal{L}}(y|y\in
A)\geq_{{{\mathrm{lr}}}}{\cal{L}}(x|x\in A),$ and hence ${\cal{L}%
}(y|y\in A)\geq_{{{\mathrm{st}}}}{\cal{L}}(x|x\in A),$ for any
measurable subset $A$ of the real line. In fact, the latter condition of
stochastic dominance for every $A$ is equivalent to ${\cal{L}}%
(y)\geq_{{{\mathrm{lr}}}}{\cal{L}}(x)$; see \cite{Shak}.
In the standard Bayesian setup we have a prior $\pi$ with support $\Theta,$
and so $\pi(\theta)>0$ for every $\theta\in\Theta,$ and a random variable $s$
whose distribution depends on $\theta$. The conditional distribution
${{\mathbb{P}}}(s|\theta)$, namely, the distribution of $s$ given $\theta$,
is denoted by ${{\mathbb{P}}}_{\theta},$ and its probability law by
${\cal{L}}_{\theta}$. We also consider the marginal probability (also called the \textquotedblleft
prior predictive probability")
${{\mathbb{P(}}}s)%
{\;:=\;}%
\sum_{\theta\in\Theta}{{\mathbb{P}}}_{\theta}(s)\pi(\theta)$, and denote its
law by ${\cal{L}}$. Expectations with respect to ${{\mathbb{P}}}$,
${{\mathbb{P}}}_{\theta},$ and ${{\mathbb{P}}}_{\Gamma}(s)%
{\;:=\;}%
{{\mathbb{P}}}(s|\Gamma)=\sum_{\theta\in\Gamma}\pi(\theta){{\mathbb{P}}%
}_{\theta}(s)/\sum_{\theta\in\Gamma}\pi(\theta)$, for a set of parameters
$\Gamma\subset\Theta,$ are denoted by ${{\mathbb{E}}}$, ${{\mathbb{E}}%
}_{\theta}$, and ${{\mathbb{E}}}_{\Gamma}$, respectively. We use the notation
\[
q^{\theta}\equiv q^{\theta}(s)%
{\;:=\;}%
{{\mathbb{P}}}(\theta|s)
\]
for the \emph{posterior} probability of $\theta$ (the \textquotedblleft%
$\theta$\emph{-posterior}" for short) given $s.$ We will compare random
variables like $q^{\theta}$ under different distributions, such as
${{\mathbb{P}}}$ and ${{\mathbb{P}}}_{\theta}$.
We now summarize several simple results with short proofs. They are given in
\cite{HR} with more details, interpretations, and related references.
\begin{proposition}
\label{prop:p1p2}(i)\textbf{\ }Let $P_{1}$ and $P_{2}$ be two probability
measures on a measure space ${\mathcal{S}}$ such that $P_{1}\ll P_{2}$ (i.e.,
$P_{2}(s)=0$ implies $P_{1}(s)=0),$ and let $r(s)=P_{1}(s)/P_{2}(s)$ be the
likelihood ratio.\footnote{When
the ratio is $0/0$ define $r(s)$ arbitrarily; this will not matter since it
occurs on a null event for both $P_{1}$ and $P_{2}.$} Then
\[
{\cal{L}}_{P_{1}}(r)\geq_{{{\mathrm{lr}}}}{\cal{L}}_{P_{2}}(r),
\]
where ${\cal{L}}_{P_{i}}$ denotes the probability law with respect to
$P_{i},$ and for any increasing function $f$,
\begin{equation}
{\cal{L}}_{P_{1}}(f(r))\geq_{{{\mathrm{lr}}}}{\cal{L}}_{P_{2}%
}(f(r)).\label{eq:f(r)}%
\end{equation}
(ii) In the Bayesian setup, for every $\theta$ the posterior $q^{\theta
}(s)\equiv{{\mathbb{P}}}(\theta|s)$ of $\theta$ satisfies
\begin{equation}
{\cal{L}}_{\theta}(q^{\theta})\geq_{{{\mathrm{lr}}}}{\cal{L}%
}(q^{\theta}).\label{eq:iii}%
\end{equation}
\end{proposition}
\begin{proof}
(i) For every value $t$ of $r$ let $B:=\{s:\frac{P_{1}(s)}{P_{2}(s)}=t\}$;
then $P_{1}(r=t)=\sum_{s\in B}P_{1}(s)=t\sum_{s\in B}P_{2}(s)=tP_{2}(r=t)$. It
follows that $\frac{P_{1}(r=t)}{P_{2}(r=t)}=t$, which is an increasing
function of $t$, and so we have the result for $r$ by definition. The result
for increasing $f$ then follows readily (see \cite{Shak}, Theorem 1.C.8).
(ii) Setting $P_{1}={{\mathbb{P}}}_{\theta}$ and $P_{2}={{\mathbb{P}}}$ we
have $q^{\theta}(s)=\pi(\theta)\frac{P_{1}(s)}{P_{2}(s)}=\pi(\theta)r(s)$, and
the result follows from (i) since $\mathbb{P}_{\theta}\ll\mathbb{P}.$
\end{proof}
\noindent\textbf{Remark. }In (i) the ratio $P_{1}(r=t)/P_{2}(r=t)=t$ is a
\emph{strictly} increasing function of $t$---unless $r$ is constant, which
happens only when $P_{1}\equiv P_{2}$ (and then $r\equiv1)$---and the
domination is therefore \emph{strict}. In (ii) this happens except when the
signal $s$ is completely uniformative, i.e., when the posterior is identical
to the prior: $q^{\theta}(s)=\pi(\theta)$ for all $s.$ The strict domination
implies that the resulting inequalities, such as (\ref{eq:E-up}) below, are
strict (this is pointed out in \cite{HR}).
Since likelihood ratio order implies stochastic order, (\ref{eq:f(r)}%
)\textbf{\ }implies ${\cal{L}}_{P_{1}}(f(r))\geq_{{{\mathrm{st}}}%
}{\cal{L}}_{P_{2}}(f(r)),$ and so for any increasing function $f$ we
have
\[
\sum_{s}P_{1}(s)f\left( \frac{P_{1}(s)}{P_{2}(s)}\right) \geq\sum_{s}%
P_{2}(s)f\left( \frac{P_{1}(s)}{P_{2}(s)}\right) .
\]
The quantity on the right-hand side is known (for convex $f$) as
$f$\emph{-divergence}. Similarly, the likelihood ratio order relation
\eqref {eq:iii} implies stochastic order; that is, ${{\mathbb{E}}}_{\theta
}[f(q^{\theta}(s))]\geq{{\mathbb{E[}}}f(q^{\theta}(s))]$ for increasing $f$.
Moreover, this holds also when
conditioning on a set of values of the posterior (such as being, say, more
than $1/2)$:%
\begin{equation}
{{{\mathbb{E}}}}_{\theta}[f(q^{\theta}(s))|q^{\theta}\in A]\geq{{{\mathbb{E[}%
}}}f(q^{\theta}(s))|q^{\theta}\in A]\label{eq:st-f-A}%
\end{equation}
for increasing $f$ and $A\subseteq\lbrack0,1].$ Thus the posterior of $\theta$ when the data $s$ are generated according to
${{\mathbb{P}}}_{\theta}(s)$ is stochastically larger than the posterior when
the data are generated under ${{\mathbb{P}}}(s)$. Taking $f(x)=x$ we obtain
\begin{equation}
{{\mathbb{E}}}_{\theta}[q^{\theta}]\geq{{\mathbb{E[}}}q^{\theta}]=\pi
(\theta),\label{eq:E-up}%
\end{equation}
where the inequality is strict for any informative signal $s$ (see the above remark), and the equality is the \emph{martingale property} of posteriors under
${{\mathbb{P}}}$ (see (\ref{eq:martingale})). Replacing the single parameter
$\theta_{0}$ with a set $\Gamma\subset\Theta$ of parameter values, the result
of \eqref {eq:iii} readily implies that
\begin{equation}
{\cal{L}}_{\Gamma}\left( {{\mathbb{P}}}(\Gamma|s)\right)
\geq_{{{\mathrm{lr}}}}{\cal{L}}\left( {{\mathbb{P}}}(\Gamma|s)\right)
\geq_{{{\mathrm{lr}}}}{\cal{L}}_{\Gamma^{c}}\left( {{\mathbb{P}}%
}(\Gamma|s)\right) \label{eq:iii2G}%
\end{equation}
(for the second $\geq_{{{\mathrm{lr}}}}$ use ${{\mathbb{P}}}=\pi
(\Gamma){{\mathbb{P}}}_{\Gamma}+\pi(\Gamma^{c}){{\mathbb{P}}}_{\Gamma^{c}}),$
and thus
\begin{equation}
{{\mathbb{E}}}_{\Gamma}[{{\mathbb{P}}}(\Gamma|s)]\geq{{\mathbb{E[P}}}%
(\Gamma|s)]=\pi(\Gamma)\geq{{\mathbb{E}}}_{\Gamma^{c}}[{{\mathbb{P}}}%
(\Gamma|s)].\label{eq:submartin1G}%
\end{equation}
Result (\ref{eq:iii2G}) is given in \cite{HR}, and \eqref {eq:submartin1G} is
given in \cite{FK}; see these papers for further results, references, and history.
When there are only two parameter values, say $\Theta=\{\theta_{0},\theta
_{1}\},$ (\ref{eq:iii2G}) becomes ${\cal{L}}_{\theta_{0}}(q^{\theta_{0}%
})\geq_{{{\mathrm{lr}}}}{\cal{L}}(q^{\theta_{0}})\geq_{{{\mathrm{lr}}}%
}{\cal{L}}_{\theta_{1}}(q^{\theta_{0}}).$ However, the latter dominance
relation need \emph{not }hold when there are additional parameter values in
$\Theta$; see, for instance, the example at the end of Section 1 in
\cite{HR}.\footnote{Where $\Theta=\{\alpha,\beta,\gamma\}$ and the dominance
is reversed: ${\cal{L}}_{\gamma}(q^{\alpha})>_{{{\mathrm{lr}}}%
}{{\mathcal{L(}}}q^{\alpha})$ (in the notation of the present paper, \cite{HR} shows that ${\cal{L}}_{\gamma}(1-q^{\alpha})<_{{{\mathrm{lr}%
}}}{{\mathcal{L(}}}1{{\mathcal{-}}}q^{\alpha})).$} A case where it does hold
is provided in the proposition below, which applies, for instance, to the
Bernoulli and normal distributions, and many other exponential families discussed later; see, e.g., \cite{Karlin} or \cite{Jew}. The parameter space is now an
interval on the real line, $\theta_{0}$ and $\theta_{1}$ are the two interval
ends, and the family of distributions ${{\mathbb{P}}}_{\theta}$ satisfies the
\emph{monotone likelihood ratio property} (\emph{MLRP}), i.e., ${{\mathbb{P}}%
}_{\theta^{\prime}}(s)/{{\mathbb{P}}}_{\theta}(s)$ is increasing in
$s\in{{\mathbb{R}}}$ for $\theta^{\prime}>\theta.$ The order $\geq
_{{{\mathrm{lr}}}}$ below can of course be replaced by the weaker
$\geq_{{{\mathrm{st}}}}$\thinspace or by comparisons of expectations.
\begin{proposition}
\label{prop:sto}Let $\Theta=[\theta_{0},\theta_{1}]\subset{{\mathbb{R}}}$ and
assume that ${{\mathbb{P}}}_{\theta}$ is a monotone likelihood ratio (MLRP)
family. Then
\[
{\cal{L}}_{\theta_{0}}(q^{\theta_{0}})\geq_{{{\mathrm{lr}}}%
}{\cal{L}}(q^{\theta_{0}})\geq_{{{\mathrm{lr}}}}{\cal{L}}%
_{\theta_{1}}(q^{\theta_{0}}).
\]
\end{proposition}
\begin{proof}
The only new part is ${\cal{L}}(q^{\theta_{0}})\geq_{{{\mathrm{lr}}}%
}{\cal{L}}_{\theta_{1}}(q^{\theta_{0}})$. First, $\frac{{{\mathbb{P}}%
}_{\theta_{1}}(s)}{{{\mathbb{P}}}(s)}$ is increasing in $s$ by the MLRP
assumption, because the denominator is a mixture of ${{\mathbb{P}}}_{\theta}%
$'s over $\theta\leq\theta_{1}$, and so ${\cal{L}}_{\theta_{1}}%
(s)\geq_{{{\mathrm{lr}}}}{\cal{L}}(s).$ Similarly, $q^{\theta_{0}%
}(s)=\pi(\theta_{0})\frac{{{\mathbb{P}}}_{\theta_{0}}(s)}{{{\mathbb{P}}}(s)}$
is decreasing in $s,$ and the argument in the proof of Proposition
\ref{prop:p1p2} (i), now applied to a decreasing rather than increasing
function and thus reversing the order, implies ${\cal{L}}(q^{\theta_{0}%
})\geq_{{{\mathrm{lr}}}}{\cal{L}}_{\theta_{1}}(q^{\theta_{0}})$.
\end{proof}
We conclude with a simple symmetry between the $\theta_{0}$-posterior under
$\theta_{1}$ and the $\theta_{1}$-posterior under $\theta_{0}.$
\begin{proposition}
\label{prop:symm}Let $\theta_{0}$ and $\theta_{1}$ be in $\Theta$. Then
\[
\frac{1}{\pi(\theta_{0})}{{\mathbb{E}}}_{\theta_{1}}[q^{\theta_{0}}]=\frac
{1}{\pi(\theta_{1})}{{\mathbb{E}}}_{\theta_{0}}[q^{\theta_{1}}]=\frac{1}%
{\pi(\theta_{0})\pi(\theta_{1})}{{\mathbb{E[}}}q^{\theta_{0}}\cdot
q^{\theta_{1}}].
\]
\end{proposition}
\begin{proof}
We have
\begin{align*}
&\frac{1}{\pi(\theta_{0})}{{\mathbb{E}}}_{\theta_{1}}[q^{\theta_{0}}]
=\frac{1}{\pi(\theta_{0})}\sum_{s}\frac{{{\mathbb{P}}}_{\theta_{0}}%
(s)\pi(\theta_{0})}{{{\mathbb{P}}}(s)}{{\mathbb{P}}}_{\theta_{1}}(s)\\
&=\frac{1}{\pi(\theta_{0})\pi(\theta_{1})}\sum_{s}\frac{{{\mathbb{P}}%
}_{\theta_{0}}(s)\pi(\theta_{0})}{{{\mathbb{P}}}(s)}\frac{{{\mathbb{P}}%
}_{\theta_{1}}(s)\pi(\theta_{1})}{{{\mathbb{P}}}(s)}{{\mathbb{P}}}(s)
=\frac{1}{\pi(\theta_{0})\pi(\theta_{1})}{{\mathbb{E[}}}q^{\theta_{0}}\cdot
q^{\theta_{1}}].
\end{align*}
The last expression is symmetric in $\theta_{0}$ and $\theta_{1},$ and so it is
equal to $\frac{1}{\pi(\theta_{1})}{{\mathbb{E}}}_{\theta_{0}}[q^{\theta_{1}%
}]$ as well.
\end{proof}
The same symmetry applies of course when we consider sequences of observations
(see for instance Corollary \ref{cor:uninn} below).
\section{Increasing posterior of the true state\label{sec:monot}}
We now consider observations that arrive sequentially and apply the results of
Section \ref{sec:sing} to obtain monotonicity and order relations as a
function of the sample size $n$.
The data consist of a process of observations $x_{1},x_{2},\ldots$ whose
distribution is ${{\mathbb{P}}}_{\theta}$, where $\theta$ lies in the
parameter space $\Theta$. At this point we make no assumptions about the
distributions of the observation process and the dependence structure (over
$n).$ We assume the standard Bayesian framework given in Section
\ref{sec:int}. Given the vector of observations up to stage $n,$ which we
denote by $s_{n}=(x_{1},\ldots,x_{n})$, the \emph{posterior} of $\theta$ at
time $n$ is $q_{n}^{\theta}\equiv{{\mathbb{P}}}\left( \theta|s_{n}\right) $.
Viewing $q_{n}^{\theta}$ as an $s_{n}$-measurable random variable, we obtain
that the sequence $q_{n}^{\theta}$ is a martingale with respect to the
probability ${{\mathbb{P}}},$ i.e., ${{\mathbb{E}}}[q_{n+1}^{\theta}%
|s_{n}]=q_{n}^{\theta}.$
Proposition \ref{prop:p1p2} (ii) applied to $x_{n+1}|s_{n}$ yields
\begin{proposition}
\label{cor:nlr} For every $\theta,$ the posterior $q_{n+1}^{\theta}$ of
$\theta$ at time $n+1$ satisfies
\[
{\cal{L}}_{\theta}\left( q_{n+1}^{\theta}|s_{n}\right) \geq
_{{{\mathrm{lr}}}}{\cal{L}}\left( q_{n+1}^{\theta}|s_{n}\right)
.\label{eq:seqq}
\]
\end{proposition}
Thus, given $s_{n},$ the $\theta$-posterior $q_{n+1}^{\theta}\equiv
{{\mathbb{P}}}(\theta|x_{n+1},s_{n})$ under the probability ${{\mathbb{P}}%
}_{\theta}(\cdot|s_{n})$ likelihood-ratio dominates that same $\theta
$-posterior under the probability ${{\mathbb{P}}}(\cdot|s_{n})$. Since, again,
$\geq_{{{\mathrm{lr}}}}$ implies $\geq_{{{\mathrm{st}}}},$ by Proposition
\ref{cor:nlr} and the martingale property of $q_{n}^{\theta}\equiv
{{\mathbb{P}}}\left( \theta|s_{n}\right) $ we get
\begin{equation}
{{\mathbb{E}}}_{\theta}[q_{n+1}^{\theta}|s_{n}]\geq{{\mathbb{E}}}%
[q_{n+1}^{\theta}|s_{n}]=q_{n}^{\theta}\label{eq:subsub}%
\end{equation}
(as in (\ref{eq:st-f-A}), one may
also condition on $q_{n+1}^{\theta}$ lying in a certain
set).
Proposition \ref{cor:nlr} generalizes Proposition \ref{prop:p1p2} (ii): given
any past data $s_{n}$, the posterior belief in $\theta$ with an additional
observation $x_{n+1}$ distributed according to ${{\mathbb{{{\mathbb{P}}}}}%
}_{\theta}$ likelihood-ratio dominates the same posterior when the additional
observation is distributed according to ${{\mathbb{P}}}$; \eqref {eq:subsub}
is then the consequent expectation comparison.
Inequality \eqref {eq:subsub} means that under ${{\mathbb{P}}}_{\theta}$ the
process $q_{n}^{\theta}$ is a submartingale. Since every increasing convex
(integrable) function of a submartingale is a submartingale, for any
increasing convex $f$ we have
\begin{equation}
{{\mathbb{E}}}_{\theta}[f(q_{n+1}^{\theta})|s_{n}]\geq f(q_{n}^{\theta}).
\label{eq:gencor3}%
\end{equation}
Taking expectations on the two sides of \eqref {eq:subsub} and
\eqref {eq:gencor3} with respect to $s_{n}$ distributed under $\theta$ we obtain
\begin{corollary}
\label{cor:mon}The expectation ${{\mathbb{E}}}_{\theta}[q_{n}^{\theta}]$ of
the posterior probability of $\theta$ with respect to ${{\mathbb{P}}}_{\theta
}$ is increasing in $n$, and, more generally, so is ${{\mathbb{E}}}_{\theta
}[f(q_{n}^{\theta})]$ for any convex increasing function $f$.
\end{corollary}
Thus, with more data generated under $\theta$, the $\theta$-posterior
increases in the increasing convex order, and in particular in expectation.
The convergence of the expected posterior to $1$ (by Doob's theorem) is thus
monotone. We have obtained this result starting from a strong ordering: the
likelihood ratio order. That $q_{n}^{\theta}$ is a submartingale under
${{\mathbb{P}}}_{\theta}$ is shown in \cite{MS} and \cite{FK} (see also the
references therein), where other relevant results are given.
\section{Nonmonotonic and multimodal expected posteriors}
\label{sec:nonmonmult}
Assume now that the observations are generated under $\theta_{1}$, which is
different from $\theta_{0}$; then the posterior probability of $\theta_{0}$
given $s_{n}$ converges to zero as $n\rightarrow\infty$ by Doob's theorem.
This convergence may not be monotone, and in fact, if $\theta_{1}$ and
$\theta_{0}$ are close together, the expected posterior $\psi(n)\equiv\psi_{\theta
_{0},\theta_{1}}(n)={{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ may
increase as a function of $n$ for small $n$ before it starts decreasing. The
question that we address in this section is whether it is possible for the
expected $\theta_{0}$-posterior, with data generated under $\theta_{1}$, to
increase again after it starts decreasing. If some observations distributed
under $\theta_{1}$ cause the expected $\theta_{0}$-posterior to decrease, the
average belief in $\theta_{0}$ decreases, as it should under $\theta_{1}$, but
as we will show below it is possible for further observations generated under the
same $\theta_{1}$ to cause $\psi(n)$ to increase before it eventually
decreases to zero. Such an increase leads to an erroneous upturn of the
Bayesian statistician's degree of belief in $\theta_{0}$. Furthermore,
${{\mathbb{\psi(}}}n)$ need not be unimodal, and may fluctuate many times.
We present two examples of this behavior in the simplest case of iid coin
tosses, i.e., ${{\mathrm{Bernoulli}}}(\theta)$ observations. In Figure 1 the
sequence $\psi(n)$ decreases, then increases, and then decreases again to $0$.
Figure 2 provides a further counterexample to the unimodality of the sequence
${{\mathbb{\psi(}}}n)$, showing that it may have many
modes, and thus may alternate several times between increasing and decreasing. In both examples
the priors concentrate on three points $\theta_{j}$ (for $j=0,1,2)$ with
probabilities $\pi(\theta_{j})=\alpha_{j}$, and so the distribution of the
sufficient statistic $u_{n}:=\sum_{i=1}^{n}x_{i}$ is a mixture of three
${{\mathrm{Binomial}}}(n,\theta_{i})$ distributions: ${{\mathbb{P}}}%
(u_{n}=k)=\sum_{j=0}^{2}\alpha_{j}{{\mathbb{P}}}_{\theta_{j}}(u_{n}=k)$.
By continuity, it is clear that such examples are robust to small changes in the parameters and their associated probabilities, and priors having a larger or even continuous support with a similar behavior can
be constructed. The function $\psi_{\theta_{0},\theta_{1}%
}(n)={{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}}],$ which is given by the
formula
\begin{equation}
\psi_{\theta_{0},\theta_{1}}(n)={{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}%
}]=\sum_{k=0}^{n}\frac{{\binom{n}{k}}^{2}\theta_{0}^{k}(1-\theta_{0}%
)^{n-k}\theta_{1}^{k}(1-\theta_{1})^{n-k}\pi(\theta_{0})}{{{\mathbb{P}}}%
(u_{n}=k)}\label{eq:gg}%
\end{equation}
(and thus equals $\frac{\pi(\theta_{0})}{\pi(\theta_{1})}{{\mathbb{E}}%
}_{\theta_{0}}[q_{n}^{\theta_{1}}],$ as in Proposition \ref{prop:symm}), is
depicted in Figures 1 and 2. In Figure 3 we provide an example of iid
observations with a normal prior, which will be analyzed in Section
\ref{sec:normnorm}.
For some intuitive explanations, take Figure 1 first, where we consider the
posterior of $\theta_{0}=0.5,$ the data are generated under $\theta_{1}=0.65,$
and there is another possible parameter, $\theta_{2}=0.85.$ Initially the
expected posterior belief in $\theta_{0}$ decreases, as the observations under
$\theta_{1}$ make both $\theta_{1}$ and $\theta_{2}$ seem more likely (on
average). After $11$ observations the expected belief in $\theta_{0}$ starts
to increase, as $\theta_{2},$ whose distance from $\theta_1$ (under which the data are generated) is greater than that of $\theta_{0}$, begins to seem less likely.
Eventually, after a further $70$ observations, the expected $\theta_{0}%
$-posterior begins its final descent to $0.$
Turning to Figure 2, we see that, in addition to long-term fluctuations similar to those
of Figure 1 (see also Figure 3), there are many short-term up and down
fluctuations\footnote{We have generated other examples where the number of
modes exceeds $8$ by far.} that are likely due to the discreteness of the data
(cf. Figure 3) and of the time steps. In Proposition \ref{cor:uni} and
Corollary \ref{cor:uninn} below we try to shed some light on these
fluctuations. Interestingly, the fluctuations of the expected posterior
beliefs are rather sensitive to the values of the parameters; for example,
changing $\theta_{1}$ from $0.5$ to $0.45$, or $\theta_{2}$ from $0.85$ to
$0.8,$ yields a strictly decreasing sequence $\psi(n),$ with no up and down fluctuations.
\begin{figure}[!tbp]
\centering
\begin{minipage}[b]{0.48\textwidth}
\includegraphics[width=\textwidth,height=6.5cm]{Plot_Binom_Smooth.pdf}
\caption{The sequence $\psi(n)=\mathbb{E}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ for iid
Bernoulli observations: $\Theta=\{\theta_{0},\theta_{1},\theta_{2}\}$ with
$\theta_{j}=0.5,~0.65,~0.85,$ and prior probabilities $\alpha_{j}=\pi
(\theta_{j})=4100/5001,~1/5001,~900/5001$ (for $j=0,1,2)$.}
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\textwidth}
\includegraphics[width=\textwidth,height=6.5cm]{eightmodes.pdf}
\caption{The sequence $\psi(n)=\mathbb{E}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ for iid
Bernoulli observations: $\Theta=\{\theta_{0},\theta_{1},\theta_{2}\}$ with
$\theta_{j}=0.2,~0.5,~0.85,$ and prior probabilities $\alpha_{j}=\pi(\theta
_{j})=2000/3001,~1/3001,~1000/3001$ (for $j=0,1,2);$ we see $8$ modes.}
\end{minipage}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=0.7\textwidth,height=6.5cm]{Normal_for_Paper.pdf}
\caption{The sequence $\psi(n)=\mathbb{E}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ for iid
$\mathcal{N}(\theta,\sigma^{2})$ normal observations with a standard normal
prior on $\theta$ (see (\ref{eq:nicenorm})): $\theta_{0}=-1/3,$~$\theta
_{1}=1/3$ and $\sigma=100.$}
\end{figure}
In Figure 3 we have iid normal $\mathcal{N}(\theta,\sigma^{2})$ observations
with a large variance $(\sigma=100)$ and a standard normal prior (see Section
\ref{sec:normnorm}). For data generated under $\theta_{1}=1/3,$ the expected
belief in $\theta_{0}=-1/3$ starts by decreasing for nearly 2,000 periods,
following which it goes in the \textquotedblleft wrong direction" for a very
long time, increasing for about 30,000 additional observations; only then it
starts decreasing monotonically to its limit of $0.$ As we will show in Theorem
\ref{th:normal} (iv), additional up and down turns are not possible for normal distributions.
Consider now the effect of a single observation $x$ on the $\theta_{0}$-belief
when $x$ is generated under a different $\theta_{1}.$ As seen above, while in
many cases the belief in $\theta_{0}$ is reduced, this need not be so when
$\theta_{1}$ is close to $\theta_{0}$ and there are other points in $\Theta;$
indeed, the $\theta_{0}$-posterior strictly increases on average when
$\theta_{1}=\theta_{0}$ (see the remark after Proposition \ref{prop:p1p2}),
and thus, by continuity, also when $\theta_{1}$ and $\theta_{0}$ are close
enough together. We quantify precisely this \textquotedblleft closeness" of parameters
in the simplest case of a Bernoulli observation, as follows.
\begin{proposition}
\label{cor:uni} Let $x$ be a ${{\mathrm{Bernoulli}}}(\theta)$ observation, and
let $\pi$ be the prior probability of $\theta$ for $\theta\in\Theta
\subseteq\lbrack0,1].$ Let $\overline{\theta}:={{\mathbb{E}}}[\theta
]=\sum_{\theta\in\Theta}\theta\pi(\theta)$ be the (prior) average parameter.
For any $\theta_{0}$ and $\theta_{1}$ in $\Theta,$ the inequality
${{\mathbb{E}}}_{\theta_{1}}[q^{\theta_{0}}]\leq\pi(\theta_{0})$ holds if and
only if $\overline{\theta}$ lies between $\theta_{0}$ and $\theta_{1},$ with
equality if and only if $\overline{\theta}=\theta_{0}$ or $\overline{\theta
}=\theta_{1}.$
\end{proposition}
The condition that $\overline{\theta}$ lies between $\theta_{0}$ and
$\theta_{1}$ is symmetric in $\theta_{0}$ and $\theta_{1},$ and so it is also
equivalent to ${{\mathbb{E}}}_{\theta_{0}}[q^{\theta_{1}}]\leq\pi(\theta_{1})$
(this equivalence follows from Proposition \ref{prop:symm} as well). Assume
therefore without loss of generality that $\theta_{0}\leq\theta_{1}.$
Proposition \ref{cor:uni} says that data under $\theta_{1}$ make the expected
posterior of $\theta_{0}$ lower than the prior of $\theta_{0}$ if and only if
$\theta_{0}\leq\overline{\theta}\leq\theta_{1},$ and make it higher than the
prior if and only if either $\theta_{0},\theta_{1}\leq\overline{\theta}$ or
$\theta_{0},\theta_{1}\geq\overline{\theta}$. Thus $\theta_{1}$ being
\textquotedblleft close" to $\theta_{0}$ in the sense that an observation
under $\theta_{1}$ increases the expected belief in $\theta_{0}$ (i.e.,
${{\mathbb{E}}}_{\theta_{1}}[q^{\theta_{0}}]\geq\pi(\theta_{0})$ holds) is
equivalent to $\theta_{1}$ being on the same side of $\overline{\theta}$ as
$\theta_{0}.$ In fact, this notion of \textquotedblleft belief-close" is
related to \textquotedblleft metric-close" as follows: the shorter the
distance $|\theta_{1}-\theta_{0}|$ between $\theta_{1}$ and $\theta_{0}$ is,
the fewer the priors $\pi$ yielding $\overline{\theta}$ in the interval
$[\theta_{0},\theta_{1}]$ there are, and so the fewer the priors $\pi$
yielding ${{\mathbb{E}}}_{\theta_{1}}[q^{\theta_{0}}]\geq\pi(\theta_{0})$
there are. Finally, when $\theta_{1}=\theta_{0},$ Proposition \ref{cor:uni}
gives (\ref{eq:E-up}), with strict inequality unless $\overline{\theta}%
=\theta_{0}.$
\begin{proof}
Using ${{\mathbb{P}}}(x=1)=\overline{\theta}$ we have
\begin{align}
{{\mathbb{E}}}_{\theta_{1}}[q^{\theta_{0}}] & =\theta_{1}q^{\theta_{0}%
}(1)+(1-\theta_{1})q^{\theta_{0}}(0)=\theta_{1}\frac{\theta_{0}\pi(\theta
_{0})}{\overline{\theta}}+(1-\theta_{1})\frac{(1-\theta_{0})\pi(\theta_{0}%
)}{1-\overline{\theta}}\nonumber\\
& =\pi(\theta_{0})V(\overline{\theta}), \label{eq:V}%
\end{align}
where
\begin{equation}
V(y){\;:=\;}\frac{\theta_{0}\theta_{1}}{y}+\frac{(1-\theta_{0})(1-\theta_{1}%
)}{1-y}. \label{eq:pphhii}%
\end{equation}
The function $V$ is strictly convex and satisfies $V(\theta_{0})=V(\theta
_{1})=1,$ and so $V(\overline{\theta})\leq1$ if and only if $\theta_{0}%
\leq\overline{\theta}\leq\theta_{1}.$
\end{proof}
This extends to a sequence $x_{1},x_{2},\dots$ of iid ${{\mathrm{Bernoulli}}%
}(\theta)$ observations, where we obtain the condition for the expected
posterior to increase after the next observation, given the past data. Indeed,
using (\ref{eq:V}) for the observation $x_{n+1}$ that comes after data
$s_{n}=(x_{1},...,x_{n})$ yields
\begin{corollary}
\label{cor:uninn} Let $x_{1},x_{2},\ldots$ be iid ${{\mathrm{Bernoulli}}%
}(\theta)$ observations, let $\pi$ be the prior probability of $\theta$ for
$\theta\in\Theta\subseteq\lbrack0,1],$ and denote by $\overline{\theta}%
_{n}:={{\mathbb{E}}}[\theta|s_{n}]=\sum_{\theta\in\Theta}\theta{{\mathbb{P}}%
}(\theta|s_{n})$ the average parameter conditional on $s_{n}.$ For any
$\theta_{0}$ and $\theta_{1}$ in $\Theta$ we have
\begin{equation}
{{\mathbb{E}}}_{\theta_{1}}[q_{n+1}^{\theta_{0}}|s_{n}]=q_{n}^{\theta_{0}%
}V(\overline{\theta}_{n})\;\;{{\text{and\ \ \ }}}{{\mathbb{E}}}_{\theta_{0}%
}[q_{n+1}^{\theta_{1}}|s_{n}]=q_{n}^{\theta_{1}}V(\overline{\theta}%
_{n}),\label{eq:V-n}%
\end{equation}
where the function $V$ is given by (\ref{eq:pphhii}), and so each one of the
inequalities ${{\mathbb{E}}}_{\theta_{1}}[q_{n+1}^{\theta_{0}}|s_{n}]\leq
q_{n}^{\theta_{0}}$ and ${{\mathbb{E}}}_{\theta_{0}}[q_{n+1}^{\theta_{1}%
}|s_{n}]\leq q_{n}^{\theta_{1}}$ holds if and only if $\overline{\theta}_{n}$
lies between $\theta_{0}$ and $\theta_{1}$.
\end{corollary}
Thus $\theta_{0}$ and $\theta_{1}$ are \textquotedblleft not close together" if they
are separated by $\overline{\theta}_{n}$, in which case an additional
observation under $\theta_{1}$ decreases the expected belief in $\theta_{0},$
and vice versa. The notion of closeness depends, of course, on $n$ and $s_{n}%
$. Since for data generated under $\theta_{1}$ (in the support of $\pi)$ we
have $\overline{\theta}_{n}\rightarrow\theta_{1}$, as $n$ increases
$\theta_{0}$ and $\theta_{1}$ may well fluctuate between being and not being close together given
$s_{n}$. Despite these fluctuations, we will see a large class of
natural setups (Theorem \ref{th:asymp} and Corollary \ref{c:asymp-monot})
where the overall expectation ${{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}%
}]$ is strictly decreasing from some $n$ on.
\section{Asymptotic rates and eventual monotonicity}
\label{sec:asympt}
In this section we obtain the precise asymptotic behavior of ${{{\mathbb{E}}}%
}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ as $n\rightarrow\infty$ in the commonly
used rich class of \emph{exponential families} of distributions; for clarity
we focus on one-dimensional families (see Appendix \ref{sec-a:asymp} for
extensions). The prior is now assumed to be continuous (which allows the use
of analysis tools), while the data may be discrete or continuous. All the
results of the previous sections are clearly seen to extend here, with
densities replacing probabilities as needed.
The general setup is as follows. Let ${{{\mathcal{X}}}}\subseteq{{{\mathbb{R}%
}}}$ be the space of observations, and let $\nu$ be a $\sigma$-finite measure on
the Borel sets of ${{{\mathcal{X}}}}$; for instance, take $\nu$ to be the
counting measure in the discrete case where ${{{\mathcal{X}}}}$ is a finite or
countable set, and the Lebesgue measure in the continuous case where
${{{\mathcal{X}}}}$ is a bounded or unbounded interval. Let $\Theta
\subseteq{{{\mathbb{R}}}}$ be the space of parameters, and let $\Pi,$ the
prior, be a probability measure on $\Theta;$ we assume that $\Theta$ is a
convex set and that $\Pi$ has a density $\pi(\theta)$ that is continuous and
strictly positive on $\Theta.$ Finally, the conditional-on-$\theta$
probability ${{\mathbb{P}}}_{\theta}(\cdot)\equiv{{\mathbb{P}}}(\cdot|\theta)$
on ${{\mathcal{X}}}$ has a density $p_{\theta}$ with respect to $\nu,$ given
by\footnote{Formally, $p_{\theta}$ is the Radon--Nikodym derivative
${{\mathrm{d}}}{{{\mathbb{P}}}}_{\theta}/{{\mathrm{d}}}\nu$. In the discrete
case where $\nu$ is the counting measure, $p_{\theta}(x)={{{\mathbb{P}}}%
}_{\theta}(x)$ for all $x,$ and $\int_{Y}p_{\theta}(x)\,{{{\mathrm{d}}}}%
\nu(x)=\sum_{x\in Y}p_{\theta}(x)=\sum_{x\in Y}{{\mathbb{P}}}_{\theta}(x)$ for
every $Y\subseteq\mathcal{X}.$}
\begin{equation}
p_{\theta}(x)=\operatorname{exp}(\eta(\theta)T(x)-A(\eta(\theta
))-B(x))\label{eq:exp-fam}%
\end{equation}
for some functions $\eta,T,A,$ and $B,$ where $\eta$ is differentiable and
$\eta^{\prime}(\theta)>0$ for all $\theta\in\Theta.$ The following are well
known (see, e.g., \cite{Berk}): the function $A(\eta)$ is determined by the
functions $B$ and $T$ (use the condition $\int_{{{{\mathcal{X}}}}}p_{\theta
}(x)\,{{{\mathrm{d}}}}\nu(x)=1$ for every $\theta\in\Theta),$ it is infinitely
differentiable, $A^{\prime}(\eta(\theta))={{{\mathbb{E}}}}_{\theta}[T(x)],$
and $A^{\prime\prime}(\eta(\theta))={{{\mathbb{V}}}}ar_{\theta}(T(x)).$ We
assume that $A^{\prime\prime}(\eta(\theta))>0$ (that is, $T(x)$ is not
${{\mathbb{P}}}_{\theta}$-a.s. constant) for all $\theta\in\Theta.$ Let
$I(\theta):={{{\mathbb{E}}}}_{\theta}[(\partial/\partial\theta\log p_{\theta
}(x))^{2}]=-\mathbb{E}_{\theta}[\partial^{2}/\partial\theta^{2}\log p_{\theta
}(x)]$ be the \emph{Fisher information} at $\theta;$ for exponential families
(\ref{eq:exp-fam}) we have $I(\theta)=A^{\prime\prime}(\eta(\theta))\cdot
(\eta^{\prime}(\theta))^{2}={{{\mathbb{V}}}}ar_{\theta}(T(x))\cdot
(\eta^{\prime}(\theta))^{2}>0$.
Consider a sequence of iid observations $x_{1},x_{2},...$ and set
$s_{n}=(x_{1},...,x_{n})$ $\in{{{\mathcal{X}}}}^{n}.$ The density of the
$\theta_{0}$-posterior (for $\theta_{0}\in\Theta)$ is\footnote{We use the
notation $p_{\theta}(\cdot)$ for the conditional-on-$\theta$ density of any
variable; thus, $p_{\theta}(s_{n})=\prod_{i=1}^{n}p_{\theta}(x_{i})$.}
\[
q_{n}^{\theta_{0}}\equiv q_{n}^{\theta_{0}}(s_{n})=\frac{p_{\theta_{0}}%
(s_{n})\pi(\theta_{0})}{p(s_{n})},
\]
where $p$ denotes the marginal density, i.e., $p(s_{n})=\int_{\Theta}%
p_{\theta}(s_{n})\pi(\theta)\,{{{\mathrm{d}}}}\theta.$ The $\theta_{1}%
$-expectation (for $\theta_{1}\in\Theta)$ of the $\theta_{0}$-posterior
is\footnote{The measure $\nu^{n}$ on ${{{\mathcal{X}}}}^{n}$ is the $n$-fold
product of the measure $\nu$ on ${{{\mathcal{X}}}}.$}
\begin{equation}
\psi_{\theta_{0},\theta_{1}}(n)\equiv{{{\mathbb{E}}}}_{\theta_{1}}%
[q_{n}^{\theta_{0}}]=\int_{{{{\mathcal{X}}}}^{n}}\frac{p_{\theta_{0}}%
(s_{n})\pi(\theta_{0})}{p(s_{n})}p_{\theta_{1}}(s_{n})\,{{{\mathrm{d}}}}%
\nu^{n}(s_{n}).\label{eq:psi}%
\end{equation}
In exponential families there is a simple relation between $\psi_{\theta
_{0},\theta_{1}}$ and $\psi_{\theta,\theta}$ for an appropriate $\theta
\in\Theta.$
\begin{proposition}
\label{p:theta-bar}Let $(p_{\theta})_{\theta\in\Theta}$ be a family of
densities (\ref{eq:exp-fam}), and let $\pi$ be a positive density on the
convex set $\Theta\subseteq\mathbb{R}.$ Let $\theta_{0},\theta_{1}\in\Theta.$
Then
\[
\psi_{\theta_{0},\theta_{1}}(n)=\frac{\pi(\theta_{0})}{\pi(\theta_{2})}%
\psi_{\theta_{2},\theta_{2}}(n)\,w^{n}%
\]
for every $n\geq1,$ where
\begin{equation}
\theta_{2}=\eta^{-1}\left( \frac{\eta(\theta_{0})+\eta(\theta_{1})}%
{2}\right) \in\Theta\label{eq:theta2}%
\end{equation}
and\footnote{The constant $\sqrt{w}$ is known as the Bhattacharrya
coefficient, and also as the Chernoff $1/2$-coefficient (see \cite{Bhat}, \cite{Cher}, or \cite{Nielsen} for a convenient reference).}
\begin{equation}
w=\left( \int_{{{{\mathcal{X}}}}}\sqrt{p_{\theta_{0}}(x)\,p_{\theta_{1}}%
(x)}\,{{{\mathrm{d}}}}\nu(x)\right) ^{2}\leq1,\label{eq:w}%
\end{equation}
with equality (i.e., $w=1)$ if and only if $\theta_{0}=\theta_{1}.$
\end{proposition}
\begin{proof}
The function $\eta$ is strictly increasing and continuous, and so it attains
the value $(\eta(\theta_{0})+\eta(\theta_{1}))/2$ at a (unique) point between
$\theta_{0}$ and $\theta_{1}$, and thus in $\Theta$ (which is a convex set);
this is the point $\theta_{2}$ given by (\ref{eq:theta2}). From
(\ref{eq:exp-fam}) we have
\begin{equation}
p_{\theta_{0}}(x)\,p_{\theta_{1}}(x)=w\,p_{\theta_{2}}(x)^{2},\label{eq:p0*p1}%
\end{equation}
for every $x$, where $w:=\operatorname{exp}(2A(\eta_{2})-A(\eta_{0}%
)-A(\eta_{1}))$ and $\eta_{i}:=\eta(\theta_{i})$ (and so $\eta_{2}=(\eta
_{0}+\eta_{1})/2).$ Therefore $p_{\theta_{0}}(s_{n})\,p_{\theta_{1}}%
(s_{n})=w^{n}p_{\theta_{2}}(s_{n})^{2},$ yielding the result by (\ref{eq:psi}%
). Formula (\ref{eq:w}) follows from (\ref{eq:p0*p1}): $\int_{{{{\mathcal{X}}%
}}}\sqrt{p_{\theta_{0}}(x)\,p_{\theta_{1}}(x)}=\int_{{{{\mathcal{X}}}}}%
\sqrt{w}\,p_{\theta_{2}}(x)=\sqrt{w}.$ Since $\int_{{{{\mathcal{X}}}}}%
\sqrt{p_{\theta_{0}}(x)\,p_{\theta_{1}}(x)}\leq\int_{{{{\mathcal{X}}}}%
}(p_{\theta_{0}}(x)+p_{\theta_{1}}(x))/2=1$ we get $w\leq1,$ with equality if
and only if $p_{\theta_{0}}(x)=p_{\theta_{1}}(x)$ for all $x\in{{{\mathcal{X}%
}}},$ which occurs if and only if $\theta_{0}=\theta_{1}$ (because $\eta$ is
one-to-one and $T\ $is not constant).
\end{proof}
Proposition \ref{p:theta-bar} thus reduces the analysis of the $\theta_{1}%
\neq\theta_{0}$ case to that of the $\theta_{1}=\theta_{0}$ case. The relation
(\ref{eq:p0*p1}), or, equivalently, $p_{\theta_{2}}=c\sqrt{p_{\theta_{0}%
}\,p_{\theta_{1}}}$ (for the constant $c=1/\sqrt{w}$), says that the density
$p_{\theta_{2}}$ is proportional to the geometric average of the densities
$p_{\theta_{1}}$ and $p_{\theta_{2}};$ it is their \textquotedblleft
normalized geometric average." The exponential family (\ref{eq:exp-fam}) is
closed under this averaging operation by the convexity of $\Theta$. For
further discussions and extensions, see Appendix \ref{sec-a:asymp}.
When $\theta_{1}=\theta_{0}$ the sequence ${{{\mathbb{E}}}}_{\theta_{0}}%
[q_{n}^{\theta_{0}}]$ is monotonically increasing (by Corollary \ref{cor:mon}%
), whereas when $\theta_{1}\neq\theta_{0}$ the sequence ${{{\mathbb{E}}}%
}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ converges to $0$ (by Doob's theorem).
Theorem \ref{th:asymp} strengthens these by yielding the precise asymptotic
rates.
The standard notation \textquotedblleft$f(n)\sim g(n)$ as $n\rightarrow\infty
$\textquotedblright\ means \textquotedblleft asymptotic equivalence," i.e.,
$\lim_{n\rightarrow\infty}f(n)/g(n)=1$.
\begin{theorem}
\label{th:asymp} Let $x_{1},x_{2},\ldots$ be iid observations distributed
according to an exponential family with density $p_{\theta}$ given by
(\ref{eq:exp-fam}), and let $\theta$ be distributed according to a prior
having a continuous strictly positive density $\pi$ on a convex set
$\Theta\subseteq{{\mathbb{R}}}.$
(i) Let $\theta_{0}$ be an interior point of $\Theta$; then
\begin{equation}
{{{\mathbb{E}}}}_{\theta_{0}}[q_{n}^{\theta_{0}}]\sim\frac{\sqrt{I(\theta
_{0})}}{2\sqrt{\uppi }}\sqrt{n}{{\text{\ \ as }}}n\rightarrow\infty.
\label{eq:asymp-theta0}%
\end{equation}
(ii) Let $\theta_{1}\neq\theta_{0}$ be in $\Theta$; then
\begin{equation}
{{{\mathbb{E}}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]=\frac{\pi(\theta_{0})}%
{\pi(\theta_{2})}{{{\mathbb{E}}}}_{\theta_{2}}[q_{n}^{\theta_{2}}]\,w^{n}%
\sim\frac{\pi(\theta_{0})}{\pi(\theta_{2})}\frac{\sqrt{I(\theta_{2})}}%
{2\sqrt{\uppi}}\sqrt{n}\,w^{n}{{\text{\ \ as }}}n\rightarrow\infty
,\label{eq:asymp}%
\end{equation}
where $\theta_{2}$ and $w$ are given by (\ref{eq:theta2})--(\ref{eq:w}).
\end{theorem}
\begin{corollary}
\label{c:asymp-monot}When $\theta_{1}\neq\theta_{0}$ the sequence
${{{\mathbb{E}}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ is eventually strictly decreasing.
\end{corollary}
\begin{proof}
As $n\rightarrow\infty$ we have ${{{\mathbb{E}}}}_{\theta_{1}}[q_{n+1}%
^{\theta_{0}}]/{{{\mathbb{E}}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]\sim
(\sqrt{n+1}/\sqrt{n})w\rightarrow w<1.$
\end{proof}
Note that the prior $\pi$ does not appear in (\ref{eq:asymp-theta0}); as in
the Bernstein--von Mises theorem (which is used in the proof below), the prior
is \textquotedblleft overwhelmed" by the data as the number of observations
increases; this is however \emph{not} the case in (\ref{eq:asymp}) when
$\theta_{1}\neq\theta_{0}$.
Before proving the theorem, we apply it to a number of classical examples:
\begin{itemize}
\item \textbf{Bernoulli}$(\theta)$ for $\theta\in\Theta=(0,1).$ Here
${{{\mathcal{X}}}}=\{0,1\},$ $\eta(\theta)=\log(\theta/(1-\theta)),$ and
$T(x)=x,$ and so ${{\mathbb{V}}}ar_{\theta}(T(x))=\theta(1-\theta)$ and
$\eta^{\prime}(\theta)=1/(\theta(1-\theta)),$ yielding $I(\theta
)=1/(\theta(1-\theta))$ and
\begin{align*}
{{{\mathbb{E}}}}_{\theta_{0}}[q_{n}^{\theta_{0}}] & \sim\frac{\sqrt{n}%
}{2\sqrt{\uppi\,\theta_{0}(1-\theta_{0})}}{{{\text{\ \ \ (for }}}}0<\theta
_{0}<1),\\
{{{\mathbb{E}}}}_{\theta_{1}}[q_{n}^{\theta_{0}}] & \sim\frac{\pi(\theta
_{0})\sqrt{n}\,w^{n}}{2\pi(\theta_{2})\sqrt{\uppi\,\theta_{2}(1-\theta_{2})}%
}{{{\text{\ \ (for }}}}\theta_{0}\neq\theta_{1}),
\end{align*}
where
\[
\theta_{2}=\frac{\sqrt{\theta_{0}\theta_{1}}}{\sqrt{\theta_{0}\theta_{1}%
}+\sqrt{(1-\theta_{0})(1-\theta_{1})}}{{{\text{\ and\ \ }}}}w=\left(
\sqrt{\theta_{0}\theta_{1}}+\sqrt{(1-\theta_{0})(1-\theta_{1})}\right) ^{2}.
\]
\item \textbf{Normal}$(\theta,\sigma^{2})$ for $\theta\in\Theta={{{\mathbb{R}%
}}}$ and fixed $\sigma>0.$ Here ${{{\mathcal{X}}}}={{{\mathbb{R}}}}$,
$\eta(\theta)=\theta,$ and $T(x)=x/\sigma^{2},$ and so $I(\theta
)={{\mathbb{V}}}ar_{\theta}(T(x))=1/\sigma^{2}$, yielding
\begin{align*}
{{{\mathbb{E}}}}_{\theta_{0}}[q_{n}^{\theta_{0}}] & \sim\frac{\sqrt{n}%
}{2\sigma\sqrt{\uppi}},\nonumber\\
{{{\mathbb{E}}}}_{\theta_{1}}[q_{n}^{\theta_{0}}] & \sim\frac{\pi(\theta
_{0})\sqrt{n}}{2\pi(\theta_{2})\sigma\sqrt{\uppi}}\operatorname{exp}\left(
-\frac{n(\theta_{0}-\theta_{1})^{2}}{4\sigma^{2}}\right) ,
\end{align*}
where $\theta_{2}=(\theta_{0}+\theta_{1})/2.$
\item \textbf{Exponential}$(\theta)$ for $\theta\in\Theta=(0,\infty)$ (i.e.,
$p_{\theta}(x)=\theta e^{-\theta x}$ for $x>0)$. Here ${{{\mathcal{X}}}%
}=(0,\infty),$ $\eta(\theta)=\theta,$ and $T(x)=-x,$ and so $I(\theta
)={{\mathbb{V}}}ar_{\theta}(T(x))=1/\theta^{2}$, yielding
\begin{align}\label{eq:expp}
{{{\mathbb{E}}}}_{\theta_{0}}[q_{n}^{\theta_{0}}] & \sim\frac{\sqrt{n}%
}{2\theta_{0}\sqrt{\uppi}},\nonumber\\
{{{\mathbb{E}}}}_{\theta_{1}}[q_{n}^{\theta_{0}}] & \sim\frac{\pi(\theta
_{0})\sqrt{n}}{2\pi(\theta_{2})\theta_{2}\sqrt{\uppi}}\left( \frac{\theta
_{0}\theta_{1}}{\theta_{2}^{2}}\right) ^{n},
\end{align}
where $\theta_{2}=(\theta_{0}+\theta_{1})/2.$
\item \textbf{Poisson}$(\theta)$ for $\theta\in\Theta=(0,\infty).$ Here
${{{\mathcal{X}}}}={{{\mathbb{N}}}}$, $\eta(\theta)=\log\theta,$ and $T(x)=x,$
and so ${{\mathbb{V}}}ar_{\theta}(T(x))=\theta$ and $\eta^{\prime}%
(\theta)=1/\theta,$ yielding $I(\theta)=1/\theta$ and
\begin{align*}
{{{\mathbb{E}}}}_{\theta_{0}}[q_{n}^{\theta_{0}}] & \sim\frac{\sqrt{n}%
}{2\sqrt{\uppi \,\theta_{0}}},\\
{{{\mathbb{E}}}}_{\theta_{1}}[q_{n}^{\theta_{0}}] & \sim\frac{\pi(\theta
_{0})\sqrt{n}}{2\pi(\theta_{2})\sqrt{\uppi \,\theta_{2}}}\operatorname{exp}%
\left( -n(\theta_{0}+\theta_{1}-2\theta_{2}\right) ),
\end{align*}
where $\theta_{2}=\sqrt{\theta_{0}\theta_{1}}.$
\end{itemize}
\begin{proof}
[Proof of Theorem \ref{th:asymp}]Part (ii) follows from part (i) by
Proposition \ref{p:theta-bar} (when $\theta_{1}\neq\theta_{0}$ the point
$\theta_{2}$ lies strictly between $\theta_{0}$ and $\theta_{1},$ and so is an
interior point of $\Theta).$ We will thus prove (i). It is convenient to
assume without loss of generality that $\eta(\theta)\equiv\theta$ (this is
called the \textquotedblleft canonical" representation); indeed, since
$\eta^{\prime}>0,$ the transformation $\tilde{\theta}:=\eta(\theta)$ (which
preserves the convexity of the parameter space and maps interior points to
interior points) yields: $\tilde{p}_{\tilde{\theta}}(x)=\operatorname{exp}%
(\tilde{\theta}\cdot T(x)-A(\tilde{\theta})-B(x));\;$ $\tilde{\pi}%
(\tilde{\theta})=\pi(\theta)/\eta^{\prime}(\theta);$\ $\tilde{q}_{n}%
^{\tilde{\theta}}=q_{n}^{\theta}/\eta^{\prime}(\theta);$ and $\tilde{I}%
(\tilde{\theta})=I(\theta)/(\eta^{\prime}(\theta))^{2};$ hence $\tilde{q}%
_{n}^{\tilde{\theta}}/\sqrt{\tilde{I}(\tilde{\theta})}=q_{n}^{\theta}%
/\sqrt{I(\theta)},$ and so (\ref{eq:asymp-theta0}) for $\theta$ is equivalent
to (\ref{eq:asymp-theta0}) for $\tilde{\theta}.$ From now on we thus have
\begin{equation}
p_{\theta}(x)=\operatorname{exp}(\theta\,T(x)-A(\theta
)-B(x))\label{eq:exp-fam-canon}%
\end{equation}
for all $x\in{{{\mathcal{X}}}}$ and $\theta\in\Theta,$ and so $I(\theta
)=A^{\prime\prime}(\theta)$. For $s_{n}=(x_{1},...,x_{n})\in{{{\mathcal{X}}}%
}^{n},$ let $\widehat{\theta}_{n}\equiv\widehat{\theta}_{n}(s_{n}):=\arg
\max_{\theta\in\Theta}\sum_{i=1}^{n}\log p_{\theta}(x_{i})$ denote the maximum
likelihood estimator (MLE); thus $\widehat{\theta}_{n}$ minimizes the strictly
convex function $h_{n}(\theta):=A(\theta)-\theta\cdot\overline{t}_{n},$ where
$\overline{t}_{n}\equiv\overline{t}_{n}(s_{n}):=(1/n)\sum_{i=1}^{n}T(x_{i});$
if $\widehat{\theta}_{n}$ is an interior point of $\Theta$ then $h_{n}%
^{\prime}(\widehat{\theta}_{n})=0,$ i.e., $A^{\prime}(\widehat{\theta}%
_{n})=\overline{t}_{n}.$ Put ${{{\mathbb{P}}}}_{0}\equiv{{{\mathbb{P}}}%
}_{\theta_{0}},$ $p_{0}\equiv p_{\theta_{0}},$ and ${{{\mathbb{E}}}}_{0}%
\equiv{{{\mathbb{E}}}}_{\theta_{0}},$ respectively, for the probability, density, and
expectation under $\theta_{0}$. Given iid observations $x_{1},x_{2},...,$
under ${{\mathbb{P}}}_{0},$ we have (see, e.g., \cite{Schervish}, Theorem 7.57
or \cite{FergLS}, Theorem 18)\footnote{Notation: $\overset{{{\cal{L}}%
}}{\longrightarrow}$ means convergence in law (or distribution), and
$\overset{{{{\mathbb{P}}}}_{0}}{\longrightarrow}$ means convergence in
probability with respect to the probability ${{{\mathbb{P}}}}_{0}.$}
$\sqrt{nI(\theta_{0})}(\widehat{\theta}_{n}-\theta_{0}%
)\overset{{{\cal{L}}}}{\longrightarrow}$ $\mathcal{N}(0,1)$, which
implies $\widehat{\theta}_{n}\overset{{{{\mathbb{P}}}}_{0}%
}{\longrightarrow}\theta_{0}$ (the \textquotedblleft consistency" of the MLE).
For convenience we divide the proof of (\ref{eq:asymp-theta0}) (for
$\theta_{0}\in{{{\mathrm{int}}}}\Theta)$ into a number of steps, as follows.
First, we show that with high ${\mathbb{P}}_{0}$-probability the posterior
$q_{n}^{\widehat{\theta}_{n}}$ at $\widehat{\theta}_{n}$ converges to the
appropriate limit by the Bernstein--von Mises theorem (Step 1). Second, since
$\widehat{\theta}_{n}-\theta_{0}$ is approximately normal (as seen above), we
show that replacing $\widehat{\theta}_{n}$ with $\theta_{0}$ requires a factor
of $1/\sqrt{2}$ on expectation (Steps 2 and 3). We then prove that the
$\theta_{0}$-posterior $q_{n}^{\theta_{0}}$ is ${\mathrm{O}}(\sqrt{n})$ (Step
4), and so sets of small ${\mathbb{P}}_{0}$-probability can be ignored (Step
5); this completes the proof.
\noindent$\bullet$ \emph{Step 1:} Let $J(\theta):=\sqrt{I(\theta)/(2\uppi)};$
then
\[
\frac{1}{\sqrt{n}}q_{n}^{\widehat{\theta}_{n}}(s_{n})\overset{{{{\mathbb{P}}}%
}_{0}}{\longrightarrow}J(\theta_{0}){{{\text{ as }}}}n\rightarrow\infty.
\]
\begin{proof}
The Bernstein--von Mises theorem says that under $\theta_{0}$ the posterior
density of $\vartheta:=\sqrt{nI(\widehat{\theta}_{n})}(\theta-\widehat{\theta
}_{n})$ converges as $n\rightarrow\infty$ to the standard normal density
$\varphi$; specifically, Theorem 7.89 in \cite{Schervish} (in Appendix
\ref{sec-a:step1} we show that all the \textquotedblleft general regularity
conditions" of the theorem hold) applied at $\vartheta=0,$ i.e.,
$\theta=\widehat{\theta}_{n}$, yields $q_{n}^{\widehat{\theta}_{n}}%
(s_{n})/\sqrt{nI(\widehat{\theta}_{n})}\overset{{{{\mathbb{P}}}}%
_{0}}{\longrightarrow}\varphi(0)=1/\sqrt{2\uppi}$ . Now use $\widehat{\theta
}_{n}\overset{{{{\mathbb{P}}}}_{0}}{\longrightarrow}\theta_{0}$ and the
continuity of $I$.
\end{proof}
Given $0<\varepsilon<1,$ let $\Omega_{n}^{1}\equiv\Omega_{n}^{1}(\varepsilon)$
be the event that $\left\vert (1/\sqrt{n})q_{n}^{\widehat{\theta}_{n}}%
(s_{n})/J(\theta_{0})-1\right\vert \leq\varepsilon;$ thus ${{{\mathbb{P}}}%
}_{0}(\Omega_{n}^{1})\rightarrow1$ (as $n\rightarrow\infty)$ by Step 1. Let
$\Omega_{n}^{2}\equiv\Omega_{n}^{2}(\varepsilon)$ be the event that
$\widehat{\theta}_{n}$ is an interior point of $\Theta$ and $|\pi(\theta
_{0})/\pi(\widehat{\theta}_{n})-1|\leq\varepsilon;$ since $\widehat{\theta
}_{n}\overset{{{{\mathbb{P}}}}_{0}}{\longrightarrow}\theta_{0}\in
{{{\mathrm{int}}}}\Theta$ and $\pi$ is continuous and positive, ${{{\mathbb{P}%
}}}_{0}(\Omega_{n}^{2})\rightarrow1.$ Put $\Omega_{n}:=\Omega_{n}^{1}%
\cap\Omega_{n}^{2};$ then ${{{\mathbb{P}}}}_{0}(\Omega_{n})\rightarrow1.$
\noindent$\bullet$ \emph{Step 2:}
\[
{{{\mathbb{E}}}}_{\theta_{0}}\left[ \frac{p_{0}(s_{n})}{p_{\widehat{\theta
}_{n}}(s_{n})}{{{\mathbf{1}}}}_{\Omega_{n}}\right] \rightarrow\frac{1}%
{\sqrt{2}}{{{\text{ as }}}}n\rightarrow\infty.
\]
\begin{proof}
Recall that $h_{n}(\theta):=A(\theta)-\theta\cdot\overline{t}_{n};$ then
\[
Y_{n}:=\frac{p_{0}(s_{n})}{p_{\widehat{\theta}_{n}}(s_{n})}{{{\mathbf{1}}}%
}_{\Omega_{n}}=\operatorname{exp}\left( -n[h(\theta_{0})-h(\widehat{\theta
}_{n})]\right) {{{\mathbf{1}}}}_{\Omega_{n}}.
\]
In $\Omega_{n}\subseteq\Omega_{n}^{2}$ the point $\widehat{\theta}_{n}$ is an
interior point of $\Theta$ and so the Taylor series of $h$ around
$\widehat{\theta}_{n},$ where $h^{\prime}(\widehat{\theta}_{n})=0$ and
$h^{\prime\prime}(\widehat{\theta}_{n})=A^{\prime\prime}(\widehat{\theta}%
_{n}),$ yields an intermediate point $\theta_{n}$ between $\theta_{0}$ and
$\widehat{\theta}_{n}$ such that
\[
Y_{n}=\operatorname{exp}\left( -\frac{1}{2}nA^{\prime\prime}(\theta
_{n})(\theta_{0}-\widehat{\theta}_{n})^{2}\right) {{{\mathbf{1}}}}%
_{\Omega_{n}}.
\]
Now $\widehat{\theta}_{n}\overset{{{{\mathbb{P}}}}_{0}}{\longrightarrow}%
\theta_{0}$ implies $\theta_{n}\overset{{{{\mathbb{P}}}}_{0}}{\longrightarrow
}\theta_{0}$ and thus $A^{\prime\prime}(\theta_{n})\overset{{{{\mathbb{P}}}%
}_{0}}{\longrightarrow}A^{\prime\prime}(\theta_{0})=I(\theta_{0});$ also,
${{{\mathbf{1}}}}_{\Omega_{n}}\overset{{{{\mathbb{P}}}}_{0}}{\longrightarrow
}1$ (because ${{{\mathbb{P}}}}_{0}(\Omega_{n})\rightarrow1).$ Under $p_{0}$ we
have $\sqrt{nI(\theta_{0})}(\widehat{\theta}_{n}-\theta_{0}%
)\overset{{{\cal{L}}}}{\longrightarrow}$ $Z\equiv\mathcal{N}(0,1),$ as
mentioned before Step 1; altogether, $Y_{n}\overset{{{\cal{L}}%
}}{\longrightarrow}\operatorname{exp}(-Z^{2}/2)$. The $Y_{n}$ are uniformly
bounded ($0\leq Y_{n}\leq1,$ because $A^{\prime\prime}>0),$ and so
${{{\mathbb{E}}}}_{0}[Y_{n}]\rightarrow{{{\mathbb{E[}}}}\operatorname{exp}%
(-Z^{2}/2)]=1/\sqrt{2}.$
\end{proof}
\noindent$\bullet$ \emph{Step 3:}
\[
\limsup_{n\rightarrow\infty}\left\vert \frac{1}{\sqrt{n}}{{{\mathbb{E}}}%
}_{\theta_{0}}[q_{n}^{\theta_{0}}{{{\mathbf{1}}}}_{\Omega_{n}}]-\frac
{J(\theta_{0})}{\sqrt{2}}\right\vert \leq\varepsilon^{\prime},
\]
where $\varepsilon^{\prime}:=(3J(\theta_{0})/\sqrt{2})\varepsilon.$
\begin{proof}
For every $s_{n}$ we have
\[
\frac{1}{\sqrt{n}}q_{n}^{\theta_{0}}(s_{n})=\frac{q_{n}^{\widehat{\theta}_{n}%
}(s_{n})}{\sqrt{n}}\cdot\frac{\pi(\theta_{0})}{\pi(\widehat{\theta}_{n})}%
\cdot\frac{p_{0}(s_{n})}{p_{\widehat{\theta}_{n}}(s_{n})}.
\]
In $\Omega_{n}^{1}$ the first factor is at most $(1+\varepsilon)J(\theta
_{0}),$ and in $\Omega_{n}^{2}$ the second factor is at most $1+\varepsilon$;
hence
\[
\frac{1}{\sqrt{n}}q_{n}^{\theta_{0}}(s_{n}){{{\mathbf{1}}}}_{\Omega_{n}}%
\leq(1+\varepsilon)^{2}J(\theta_{0})\frac{p_{0}(s_{n})}{p_{\widehat{\theta
}_{n}}(s_{n})}{{{\mathbf{1}}}}_{\Omega_{n}}.
\]
Taking expectation under $\mathbb{P}_{0}$ yields by Step 2
\[
\limsup_{n\rightarrow\infty}\frac{1}{\sqrt{n}}{{{\mathbb{E}}}}_{\theta_{0}%
}[q_{n}^{\theta_{0}}{{{\mathbf{1}}}}_{\Omega_{n}}]\leq(1+\varepsilon)^{2}%
\frac{J(\theta_{0})}{\sqrt{2}}\leq\frac{J(\theta_{0})}{\sqrt{2}}%
+\varepsilon^{\prime}.
\]
Similarly,
\[
\liminf_{n\rightarrow\infty}\frac{1}{\sqrt{n}}{{{\mathbb{E}}}}_{\theta_{0}%
}[q_{n}^{\theta_{0}}{{{\mathbf{1}}}}_{\Omega_{n}}]\geq(1-\varepsilon)^{2}%
\frac{J(\theta_{0})}{\sqrt{2}}\geq\frac{J(\theta_{0})}{\sqrt{2}}%
-\varepsilon^{\prime},
\]
completing the proof.
\end{proof}
\noindent$\bullet$ \emph{Step 4:} There is a constant $C<\infty$ such that
\[
q_{n}^{\theta_{0}}(s_{n})\leq C\sqrt{n}
\]
for all $s_{n}$ and all $n\geq1.$
\begin{proof}
We will show that there is a constant $c>0$ such that
\[
\frac{\pi(\theta_{0})}{q_{n}^{\theta_{0}}(s_{n})}=\frac{p(s_{n})}{p_{0}%
(s_{n})}\geq\frac{c}{\sqrt{n}}%
\]
for all $n\geq1.$ Take $\delta>0$ such that $[\theta_{0}-\delta,\theta
_{0}+\delta]\subset{{{\mathrm{int}}}}\Theta;$ then
\[
\frac{p(s_{n})}{p_{0}(s_{n})}=\int_{\Theta}\operatorname{exp}(-n[h_{n}%
(\theta)-h_{n}(\theta_{0})])\pi(\theta)\,{{{\mathrm{d}}}}\theta\geq\rho
\int_{\theta_{0}-\delta}^{\theta_{0}+\delta}H_{n}(\theta)\,{{{\mathrm{d}}}%
}\theta,
\]
where $H_{n}(\theta):=\operatorname{exp}(-n[h_{n}(\theta)-h_{n}(\theta_{0})])$
and $\rho:=\min_{\theta\in\lbrack\theta_{0}-\delta,\theta_{0}+\delta]}%
\pi(\theta)>0$ (recall that the density $\pi$ is continuous and strictly
positive on $\Theta).$ The strictly convex function $h_{n}(\theta
)=A(\theta)-\theta\cdot\overline{t}_{n}$ has a unique minimizer in
$[\theta_{0}-\delta,\theta_{0}+\delta],$ call it $\xi;$ without loss of
generality assume that $\xi\geq\theta_{0}.$ When $\xi\geq\theta_{0}+\delta$
the function $h_{n}$ is decreasing for $\theta\leq\xi,$ and thus for all
$\theta\in\lbrack\theta_{0},\theta_{0}+\delta]$ we have $h_{n}(\theta)\leq
h_{n}(\theta_{0}),$ and hence $H_{n}(\theta)\geq1,$ which gives $p(s_{n}%
)/p_{0}(s_{n})\geq\rho\delta\geq\rho\delta/\sqrt{n}$ for all $n\geq1.$ When
$\theta_{0}\leq\xi<\theta_{0}+\delta$ we have\footnote{In this case, where
$\xi$ is an interior point, $\xi$ is the minimum over all $\Theta,$ and thus
$\xi=\widehat{\theta}_{n}.$ The argument here is an instance of the Laplace
method.} $h_{n}^{\prime}(\xi)=0$ and so $h_{n}(\theta)-h_{n}(\theta_{0})\leq
h_{n}(\theta)-h_{n}(\xi)=h_{n}^{\prime\prime}(\zeta)(\theta-\xi)^{2}/2$ for
every $\theta\in\lbrack\theta_{0}-\delta,\theta_{0}+\delta]$ (by the
second-order Taylor expansion), where $\zeta\equiv\zeta_{\theta}$ is an
intermediate point between $\theta$ and $\xi,$ and so $\zeta\in\lbrack
\theta_{0}-\delta,\theta_{0}+\delta].$ Since $h_{n}^{\prime\prime}%
=A^{\prime\prime}$ we get $0<h_{n}^{\prime\prime}(\zeta)\leq\alpha
:=\max_{\theta\in\lbrack\theta_{0}-\delta,\theta_{0}+\delta]}A^{\prime\prime
}(\theta),$ and so $h_{n}(\theta)-h_{n}(\theta_{0})\leq\alpha(\theta-\xi
)^{2}/2$ and
\[
\rho\int_{\theta_{0}-\delta}^{\theta_{0}+\delta}H(\theta)\,{{{\mathrm{d}}}%
}\theta\geq\rho\int_{\xi-\delta}^{\xi}\operatorname{exp}\left( -\frac
{n\alpha}{2}(\theta-\xi)^{2}\right) {{\mathrm{d}}}\theta=\frac{\rho
\sqrt{2\uppi}}{\sqrt{n\alpha}}\left( \frac{1}{2}-\Phi(-\delta\sqrt{n\alpha
})\right)
\]
(because $[\xi-\delta,\xi]\subset\lbrack\theta_{0}-\delta,\theta_{0}+\delta];$
here $\Phi$ denotes the cumulative standard normal distribution). Since
$\Phi(-\delta\sqrt{n\alpha})\rightarrow0$ as $n\rightarrow\infty,$ the final
expression is $\sim c/\sqrt{n}$ for some $c>0,$ which completes the proof.
\end{proof}
\noindent$\bullet$ \emph{Step 5:} Let $\Omega_{n}^{c}$ denote the complement of $\Omega_{n}$. We have
\[
\frac{1}{\sqrt{n}}{{{\mathbb{E}}}}_{\theta_{0}}[q_{n}^{\theta_{0}%
}{{{\mathbf{1}}}}_{\Omega_{n}^{c}}]\rightarrow0{{{\text{ \ as }}}}%
n\rightarrow\infty.
\]
\begin{proof}
Use the uniform boundedness of $(1/\sqrt{n})q_{n}^{\theta_{0}}$ by Step 4 and
${{{\mathbb{P}}}}_{0}(\Omega_{n})\rightarrow1$.
\end{proof}
Adding the results of Steps 3 and 5 and noting that $\varepsilon>0$ is
arbitrary yields (\ref{eq:asymp-theta0}), and thus completes the proof of Theorem \ref{th:asymp}.
\end{proof}
See Appendix \ref{sec-a:asymp} for additional comments, extensions, and
technical details.
\section{Log-concavity}
\label{sec:loggcon} In this section we discuss setups in which the sequence
${{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ is log-concave in $n$. A
sequence of positive numbers $\psi(n)$ for $n\geq1$ is (\emph{strictly}%
)\emph{\ log-concave} if $\log\psi(n)$ is a (strictly) concave function of
$n$; this is equivalent to $\psi(n)^{2}\geq\psi(n-1)\psi(n+1)$ for every
$n\geq2$, with $>$ for the strict version. The sequence $\psi(n)$
is\emph{\ unimodal} if there exists $0\leq n_{0}\leq\infty$ (possibly equal to
$0$ or $\infty)$ such that $\psi(n)$ is increasing for $n\leq n_{0}$ and
decreasing for $n\geq n_{0}.$ Log-concavity clearly implies unimodality (with
$n_{0}$ that maximizes $\psi(n)$).
In Corollary \ref{c:asymp-monot} we saw that for $\theta_{1}\neq\theta_{0}$
the sequence $\psi_{\theta_{0},\theta_{1}}(n)={{\mathbb{E}}}_{\theta_{1}%
}[q_{n}^{\theta_{0}}]$ is eventually strictly decreasing. We will now show
that in certain natural setups, with conjugate priors, this can be
strengthened to unimodality, and, in fact, to log-concavity, with respect to
the time period $n.$ We do this in three setups. The first one consists of iid
Bernoulli observations with a uniform prior; the second, of iid normal
observations with a normal prior; and the third, of iid exponential observations
with an exponential prior. In the normal case log-concavity is obtained only
from some $n_{0}$ on, where $n_{0}$ may be $1$ or arbitrarily large, depending
on the parameters; Figure 3 in Section \ref{sec:nonmonmult} is typical of the
latter case. The analysis suggests that general log-concavity results may be
hard to obtain, as indicated by the different proofs in the three cases, as
well as by the dependence of the result on the specific prior (see the normal
case, or take a Beta prior in the Bernoulli case).
In Proposition \ref{p:theta-bar} of Section \ref{sec:asympt} we get a
log-linear relation between $\psi_{\theta_{0},\theta_{1}}(n)$ and
$\psi_{\theta_{2},\theta_{2}}(n),$ and so it suffices to consider only the
case where $\theta_{1}=\theta_{0}.$
\begin{corollary}
\label{cor:suffsame} Under the assumptions of Proposition \ref{p:theta-bar},
the sequence $\psi_{\theta_{0},\theta_{1}}(n)$ is log-concave/convex in $n$ if
and only if the sequence $\psi_{\theta_{2},\theta_{2}}(n)$ is
log-concave/convex in $n.$
\end{corollary}
\subsection{Bernoulli observations with uniform prior}
\label{sec:uniprior}
This section deals with sequences of iid Bernoulli observations with a
parameter $\theta$ that is uniformly distributed in $(0,1).$ We show that in
this case $\psi_{\theta_{0},\theta_{1}}(n)={{\mathbb{E}}}_{\theta_{1}}%
[q_{n}^{\theta_{0}}]$ is a log-concave function of the time period $n$, and so
unimodal in $n.$ We obtain this result by proving in Section \ref{sec:revtur}
a \textquotedblleft reversal" of the reverse Tur\'{a}n inequality for Legendre
polynomials, which may be of independent interest.
\begin{theorem}
\label{th:binunprlog} Let $x_{1},x_{2},\ldots$ be iid ${{\mathrm{Bernoulli}}%
}(\theta)$ observations, and let the prior distribution of $\theta$ be the
uniform distribution on $\Theta=(0,1).$ For every $\theta_{0}$ and $\theta
_{1}$ in $\Theta$ the sequence $\psi_{\theta_{0},\theta_{1}}(n)={{\mathbb{E}}%
}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ is log-concave for $n\geq1$ \textbf{(}and
strictly log-concave for $n\geq2$\textbf{)}, and hence unimodal.
\end{theorem}
\begin{proof}
As in Section \ref{sec:nonmonmult}, we work with the sufficient statistic
$u_{n}:=\sum_{i=1}^{n}x_{i}$, whose distribution given $\theta$ is
\textrm{Binomial}$(n,\theta).$ The marginal distribution of $u_{n}$ when the
prior is uniform on $(0,1)$ is then the uniform distribution on the set
$\{0,1,...,n\};$ i.e., ${{\mathbb{P}}}(u_{n}=k)=1/(n+1)$ for $0\leq k\leq n$
(this is a well-known result; see, e.g., \cite{GG}, page 155). Therefore
\[
\psi_{\theta,\theta}(n)=\sum_{k=0}^{n}\frac{\left( {{\mathbb{P}}}_{\theta
}(u_{n}=k)\right) ^{2}}{{{\mathbb{P}}}(u_{n}=k)}=(n+1)\sum_{k=0}^{n}%
{\binom{n}{k}}^{2}\theta^{2k}(1-\theta)^{2(n-k)}.
\]
The log-concavity of the last expression follows from Corollary \ref{c:PSI} in
the next section, with $y=\theta^{2}$ and $z=(1-\theta)^{2}.$ Finally, for
$\theta_{1}\neq\theta_{0}$ apply Corollary \ref{cor:suffsame}.
\end{proof}
It is natural to try to generalize from the uniform prior to other priors,
such as Beta distributions (the class of conjugate priors here). However,
numerical calculations show that ${{\mathbb{E}}}_{\theta_{1}}[q_{n}%
^{\theta_{0}}]$ need \emph{not} be log-concave: take, for example, the
${{\mathrm{Beta}}}(7,1)$ prior, $\theta_{0}=3/4,$ $\theta_{1}=9/10,$ and
$n=2,3,4$ (this is clearly robust to small changes in the parameters).
\subsection{Reversing the reverse Tur\'{a}n inequality for Legendre
polynomials}
\label{sec:revtur} This section proves an interesting reversal of the reverse
Tur\'{a}n inequality for Legendre polynomials for $|x|>1;$ it yields in
particular the log-concavity of the previous section.
The Legendre polynomial of degree $n\geq0$ is defined as
\[
P_{n}(x){\;:=\;}{\frac{1}{2^{n}}}\sum_{k=0}^{n}{\binom{n}{k}}^{2}%
(x-1)^{n-k}(x+1)^{k};
\]
see, e.g., \cite{szw} and the references therein. The well-known Tur\'{a}n
inequality for Legendre polynomials, first published in \cite{sze}, states
that
\begin{equation}
P_{n}^{2}(x)\geq P_{n-1}(x)P_{n+1}(x)\,\,{{\text{ for all }}}\,|x|\leq1
\label{eq:turan}%
\end{equation}
holds for every $n\geq1,$ with equality if and only if\footnote{Put $0^{0}=1$;
then $P_{n}(1)=1$ and $P_{n}(-1)=(-1)^{n}.$} $|x|=1$. Its reverse version,
\begin{equation}
P_{n}^{2}(x)<P_{n-1}(x)P_{n+1}(x)\,\,{{\text{ for all }}}\,|x|>1,
\label{eq:anti-turan}%
\end{equation}
holds for every $n\ge1;$ see, e.g., \cite{szw}, Theorem 1.
We next show that multiplying $P_{n}$ by $n+1$ reverses (\ref{eq:anti-turan})
for all\footnote{Starting with $P_{1}$ rather than $P_{0}$ (which is what is
needed for our Theorem \ref{th:binunprlog}).} $|x|>1$.
\begin{theorem}
\label{th:turan} The inequality
\begin{equation}
\frac{P_{n-1}(x)P_{n+1}(x)}{P_{n}^{2}(x)}\leq\frac{(n+1)^{2}}{n(n+2)}%
\,\,{{\text{ for all }}}\,|x|>1 \label{eq:beta}%
\end{equation}
holds for every $n\geq2,$ with equality if and only if $n=2$ and $|x|=\sqrt
{3}$.
\end{theorem}
\begin{proof}
Putting
\[
R_{n}{\;:=\;}\frac{P_{n-1}P_{n+1}}{P_{n}^{2}}\;\;{{\text{and\ \ }}}%
a_{n}{\;:=\;}\frac{(n+1)^{2}}{n(n+2)}
\]
we will prove that
\[
1<R_{n}(x)\leq a_{n}
\]
for all $|x|>1$ and $n\geq2;$ the first inequality (which also holds for
$n=1)$ is the reverse Tur\'{a}n inequality (\ref{eq:anti-turan}), for which we
provide a simple proof as well. Because $P_{n}(-x)=(-1)^{n}P_{n}(x)$ it
suffices to consider the range $x>1,$ where $P_{n}(x)>0$ for all $n\geq1$. Let
$b_{n}$ be the leading coefficient of $P_{n}$, i.e., the coefficient of its
highest power, $x^{n}$; then
\[
b_{n}=\frac{1}{2^{n}}\sum_{k=0}^{n}\binom{n}{k}^{2}=\frac{1}{2^{n}}\binom
{2n}{n},
\]
and thus
\begin{equation}
R_{n}(\infty){\;:=\;}\lim_{x\rightarrow\infty}R_{n}(x)=\frac{b_{n-1}b_{n+1}%
}{b_{n}^{2}}=\frac{n(2n+1)}{(n+1)(2n-1)}. \label{eq:nnn}%
\end{equation}
It is straightforward to verify that
\begin{equation}
1<R_{n}(\infty)<a_{n} \label{eq:limit}%
\end{equation}
for every $n\geq2$ (the first inequality is equivalent to $(2n^{2}%
+n)/(2n^{2}+n-1)>1,$ and the second to $n^{2}-n-1>0).$ Next, differentiating
$\log R_{n}(x)$ with respect to $x$ yields
\[
(\log R_{n})^{\prime}=\frac{R_{n}^{\prime}}{R_{n}}=\frac{P_{n-1}^{\prime}%
}{P_{n-1}}+\frac{P_{n+1}^{\prime}}{P_{n+1}}-\frac{2P_{n}^{\prime}}{P_{n}}.
\]
Using the following well-known formula (see, e.g., \cite{Ar}, Equation (12.26)
for a convenient reference)
\[
(x^{2}-1)P_{n}^{\prime}(x)=nxP_{n}(x)-nP_{n-1}(x)
\]
for every $n\geq1$ and every $x$ we then obtain
\begin{multline}
(x^{2}-1)\frac{R_{n}^{\prime}}{R_{n}
=(n-1)x-(n-1)\frac{P_{n-2}}{P_{n-1}}+(n+1)x-(n+1)\frac{P_{n}}{P_{n+1}%
}-2nx+2n\frac{P_{n-1}}{P_{n}}\\
=2n\frac{P_{n-1}}{P_{n}}-(n-1)\frac{P_{n-2}}{P_{n-1}}-(n+1)\frac{P_{n}%
}{P_{n+1}}
=(n+1)\frac{P_{n-1}}{P_{n}}\left( \frac{2n}{n+1}-\frac{n-1}{n+1}R_{n-1}%
-\frac{1}{R_{n}}\right) . \label{eq:diff=0}%
\end{multline}
The proof of \eqref {eq:beta} is by induction on $n,$ separately for each one
of the two inequalities. For the first inequality, assume by induction that
$R_{n-1}(x)>1$ for every $x>1.$ If $R_{n}(x)\leq1$ for some $x>1,$ then, since
$R_{n}(1)=1$ and $R_{n}(\infty)>1$ (see \eqref {eq:nnn}), there is $x_{\ast
}>1$ where $R_{n}$ attains its minimum, and so $R_{n}(x_{\ast})\leq1$ and
$R_{n}^{\prime}(x_{\ast})=0.$ Using (\ref{eq:diff=0}) and then $R_{n-1}%
(x_{\ast})>1$ (by the induction hypothesis) yields
\[
\frac{1}{R_{n}(x_{\ast})}=\frac{2n}{n+1}-\frac{n-1}{n+1}R_{n-1}(x_{\ast
})<\frac{2n}{n+1}-\frac{n-1}{n+1}=1,
\]
in contradiction to $R_{n}(x_{\ast})\leq1.$ The induction starts with $n=1,$
where we have
\[
P_{0}P_{2}-P_{1}^{2}=1\cdot\frac{1}{2}(3x^{2}-1)-\left( x\right) ^{2}%
=\frac{1}{2}(x^{2}-1)>0{{\text{ for }}}x>1,
\]
and so $R_{1}(x)>1$ for every $x>1.$ For the second inequality, assume by
induction that $R_{n-1}(x)<a_{n-1}$ for every $x>1.$ If $R_{n}(x)\geq a_{n}$
for some $x>1,$ then, because $R_{n}(\infty)<a_{n}$ (see \eqref {eq:nnn} and
(\ref{eq:limit})), there is $x^{\ast}>1$ where $R_{n}$ attains its maximum,
and so $R_{n}(x^{\ast})\geq a_{n}$ and $R_{n}^{\prime}(x^{\ast})=0.$ Using
\eqref {eq:diff=0} whose left-hand side vanishes at $x^{\ast}$, and then
$R_{n-1}(x^{\ast})<a_{n-1}$ (by the induction hypothesis) yields
\begin{align*}
\frac{1}{R_{n}(x^{\ast})} & =\frac{2n}{n+1}-\frac{n-1}{n+1}R_{n-1}(x^{\ast
})>\frac{2n}{n+1}-\frac{n-1}{n+1}a_{n-1}\\
& =\frac{2n}{n+1}-\frac{n-1}{n+1}\frac{n^{2}}{(n-1)(n+1)}=\frac
{n(n+2)}{(n+1)^{2}}=\frac{1}{a_{n}},
\end{align*}
in contradiction to $R_{n}(x^{\ast})\geq a_{n}.$ The induction now starts with
$n=2,$ where we have
\[
P_{1}P_{3}-\frac{9}{8}P_{2}^{2}=x\cdot\frac{1}{2}(5x^{3}-3x)-\frac{9}%
{8}\left( \frac{1}{2}(3x^{2}-1)\right) ^{2}=-\frac{(x^{2}-3)^{2}}{32}%
\leq0{{\text{ for }}}x>1,
\]
and so $R_{2}(x)\leq a_{2}$ for every $x>1,$ with equality only for
$x=\sqrt{3};$ for $n=3$ we have $R_{3}(\sqrt{3})=19/18<16/15=a_{3},$ and so
$x^{\ast}$ cannot be $\sqrt{3},$ and thus $R_{2}(x^{\ast})<a_{2}$ and the
induction argument above gives $R_{3}(x)<a_{3}$ for every $x>1,$ and then
$R_{n}(x)<a_{n}$ for every $x>1$ and $n\geq4.$
\end{proof}
\begin{corollary}
\label{c:PHI}For fixed $x>1,$ the sequence $Q_{n}(x){\;:=\;}(n+1)P_{n}(x)$ is
log-concave in $n$ for $n\geq1,$ and strictly log-concave for $n\geq2.$
\end{corollary}
Thus, $Q_{n}^{2}(x)\geq Q_{n-1}(x)Q_{n+1}(x)$ holds for every $n\geq2,$ with
strict inequality for $n\geq3.$ The result holds for $x=1$ as well, where
$P_{n}(1)=1$ for all $n.$
The sequence $P_{n}(x),$ for fixed $x\geq1,$ is log-convex in $n$ by the
reverse Tur\'{a}n inequality (\ref{eq:anti-turan}); Theorem \ref{th:turan}
says that multiplying $P_{n}(x)$ by $n+1$ makes the sequence log-concave
instead. Moreover, $(n+1)P_{n}(x)$ is the \textquotedblleft right" multiple
for which this reversal occurs: taking any smaller multiple, such as
$nP_{n}(x),$ also reverses the inequality (by (\ref{eq:beta}) and
$\frac{(n+1)^{2}}{n(n+2)}\leq\frac{n^{2}}{(n-1)(n+1)}$), whereas any larger
multiple, such as $(n+2)P_{n}(x),$ does not (consider $n=2$ and $x=\sqrt{3}).$
\begin{corollary}
\label{c:PSI}For fixed $y,z\geq0$ and not both $0,$ the sequence
\[
S_{n}(y,z){\;:=\;}(n+1)\sum_{k=0}^{n}{\binom{n}{k}}^{2}y^{k}z^{n-k}
\]
is log-concave in $n$ for $n\geq1,$ and strictly log-concave for $n\geq2$ .
\end{corollary}
\begin{proof}
When $y\neq z,$ say $y>z,$ put $x=(y+z)/(y-z),$ and then $x\geq1$ and
$S_{n}(y,z)=(y-z)^{n}(n+1)P_{n}\left( x\right) =(y-z)^{n}Q_{n}(x)$, and we
use Corollary \ref{c:PHI} for $x>1,$ and $P_{n}(1)=1$ for $x=1.$ When $y=z>0$
we have $S_{n}(y,y)=y^{n}(n+1)\binom{2n}{n},$ and then $S_{n}^{2}%
>S_{n-1}S_{n+1}$ is obtained from (\ref{eq:nnn})--(\ref{eq:limit}) or by
direct calculation.
\end{proof}
\subsection{Normal observations with normal prior}
\label{sec:normnorm}
We now consider normal observations whose mean is normally distributed;
specifically, $x_{1},x_{2},\ldots$ are iid ${{\mathcal{N}}}(\theta,\sigma
^{2})$ observations (with $\sigma>0$ fixed), and the prior on $\theta
{{\mathbb{\ }}}$is the standard normal distribution ${{\mathcal{N}}}(0,1)$. We
show that the expected posterior $\psi_{\theta_{0},\theta_{1}}(n)\equiv
{{\mathbb{E}}}_{\theta_{1}}[q_{n}^{\theta_{0}}]$ is either a log-concave
function of $n,$ or a log-convex function up to some point and a log-concave
function thereafter (as in Figure 3); which case it is depends on $\theta
_{0},\theta_{1},$ and $\sigma$.
\begin{theorem}
\label{th:normal} Let $x_{1},x_{2},\ldots$ be iid ${{\mathcal{N}}}%
(\theta,\sigma^{2})$ observations (with $\sigma>0$ fixed), and let the prior
distribution of $\theta$ be the normal standard distribution ${{\mathcal{N}}%
}(0,1)$ on $\Theta={{\mathbb{R}}}.$
(i) For every real $\theta$ and $n\geq1$ we have
\begin{equation}
\psi_{\theta,\theta}(n)=\frac{(n+\sigma^{2})}{\sigma\sqrt{2\uppi \,(2n+\sigma
^{2})}}\operatorname{exp}\left( -\frac{\theta^{2}\sigma^{2}}{4n+2{\sigma}%
^{2}}\right) , \label{eq:normal}%
\end{equation}
and thus
\begin{equation}
\psi_{\theta_{0},\theta_{1}}(n)=\operatorname{exp}\left( \frac{\theta
^{2}-\theta_{0}^{2}}{2}\right) \psi_{\theta,\theta}(n)\,{\operatorname{exp}%
}\left( -\frac{n(\theta_{0}-\theta_{1})^{2}}{4\sigma^{2}}\right)
\label{eq:nicenorm}%
\end{equation}
for every real $\theta_{0}$ and $\theta_{1}$ with $(\theta_{0}+\theta
_{1})/2=\theta.$
(ii) There is $n_{0}\geq0$ that depends on $\sigma$ and $\theta$ such that the
sequence $\psi_{\theta,\theta}(n)$ is strictly log-convex for $n<n_{0}$ and
strictly log-concave for $n>n_{0},$ and thus so are the sequences
$\psi_{\theta_{0},\theta_{1}}(n)$ for every $\theta_{0},\theta_{1}$ with
$(\theta_{0}+\theta_{1})/2=\theta.$
(iii) When $\sigma^{2}\leq\sqrt{2}$ or $|\theta|\geq1/2$ the sequence
$\psi_{\theta,\theta}(n)$ is strictly log-concave for $n\geq1.$
(iv) For every $\theta_{1}\neq\theta_{0}$ the sequence $\psi_{\theta
_{0},\theta_{1}}(n)$ has at most two critical points, and so it is of one of
three types: always decreasing; increasing and then decreasing; decreasing,
increasing, and then decreasing.
\end{theorem}
\begin{proof}
(i) Take the sufficient statistic $u_{n}:=\sum_{i=1}^{n}x_{i}.$ The
distribution of $u_{n}$ given $\theta$ is ${{\mathcal{N}}}(n\theta,n\sigma
^{2}),$ and the marginal distribution of $u_{n}$ is ${{\mathcal{N}}}%
(0,n^{2}+n\sigma^{2})$ (express $u_{n}$ as the sum of $n\theta$ and
$u_{n}-n\theta).$ The result (\ref{eq:normal}) is then obtained by a standard
computation, which we relegate to Appendix \ref{sec-a:normnorm}; as for
(\ref{eq:nicenorm}), it then follows from Proposition \ref{p:theta-bar}.
(ii) Let $\xi(n):=\log\psi_{\theta,\theta}(n).$ Taking derivatives with
respect to $n$ (which we view as a continuous variable in (\ref{eq:normal}))
yields $\xi^{\prime\prime}(n)=\gamma(n)/(2n+\sigma^{2})^{3},$ where
\[
\gamma(n)={\frac{({\sigma}^{4}-2n^{2})\left( 2n+{\sigma}^{2}\right)
}{\left( n+{\sigma}^{2}\right) ^{2}}}-4\,{\theta}^{2}{\sigma}^{2}.
\]
The function $\gamma(n)$ is strictly decreasing for $n>0$ (because
$\gamma^{\prime}(n)=-2n(2n^{2}+6n\sigma^{2}+3\sigma^{4})/(n+\sigma^{2}%
)^{3}<0),$ and is negative for large enough $n$ (for sure when $2n^{2}%
\geq\sigma^{4}).$ Thus either $\xi^{\prime\prime}$ is always negative, or it
changes sign once from positive to negative, which means that either $\xi$ is
always concave, or it is first convex and then concave.
(iii) If $\sigma^{2}\leq\sqrt{2}$ then $\sigma^{4}\leq2n^{2}$ for all
$n\geq1,$ and so $\gamma(n)<0$ for all $n\geq1.$ If $|\theta|\geq1/2$ then
$\gamma(0)\leq\sigma^{2}-4\theta^{2}\sigma^{2}\leq0,$ and so $\gamma(n)<0$ for
all $n>0.$
(iv) Let $\tilde{\xi}(n):=\log\psi_{\theta_{0},\theta_{1}}(n)$ and $\xi(n):=$
$\log\psi_{\theta,\theta}(n)$ for $\theta=(\theta_{0}+\theta_{1})/2;$ then
$\tilde{\xi}^{\prime}=\xi^{\prime}+\log w$ (for the appropriate $w)$ and
$\tilde{\xi}^{\prime\prime}=\xi^{\prime\prime}.$ Since $\tilde{\xi}%
^{\prime\prime}=\xi^{\prime\prime}$ changes sign at most once by (ii),
$\tilde{\xi}^{\prime}$ can vanish at most twice, and so $\tilde{\xi}$ has at
most two critical points; the last one must be a maximum since $\tilde{\xi
}(n)\rightarrow-\infty$ as $n\rightarrow\infty.$ The same then holds for
$\psi_{\theta_{0},\theta_{1}}=\exp\tilde{\xi}.$
\end{proof}
Figure 3 provides an example with two critical points, a minimum followed by a
maximum, which is thus the most that one may get in this normal setup.
\textbf{General normal prior: }When the prior on $\theta$ is a general
${{\mathcal{N}}}(\mu,\sigma_{\Theta}^{2})$ distribution, and the observations
$x_{i}$ given $\theta$ are iid ${{\mathcal{N}}}(\theta,\sigma_{X}^{2})$ (where
$\mu,\sigma_{\Theta}^{2},\sigma_{X}^{2}$ are fixed), the linear transformation
$x\rightarrow(x-\mu)/\sigma_{\Theta}$ reduces it to the above case with $\sigma
^{2}=\sigma_{X}^{2}/\sigma_{\Theta}^{2}$. Thus, $\sigma^{2}$ is now the ratio
of the variance of the observations to the variance of the prior. The above result
(iii), for instance, says that when the variance of the observations is
not too large relative to the variance of the prior (specifically, when
$\sigma_{X}^{2}\leq\sqrt{2}\sigma_{\Theta}^{2}),$ the sequence of expected
posteriors is log-concave, and thus unimodal, for $n\geq1$.
\subsection{Exponential observations with exponential prior\label{sec:expexp}}
In this section we consider exponential observations whose parameter is also
exponentially distributed; specifically, $x_{1},x_{2},\ldots$ are iid
${{\mathrm{Exp}}}(\theta)$ observations, and the prior distribution of
$\theta{{\mathbb{\ }}}$is ${{\mathrm{Exp}}}(1).$ Thus $p_{\theta}%
(x)=\theta\operatorname{exp}(-\theta x)$ and $\pi(\theta)=\operatorname{exp}%
(-\theta)$ for $x\geq0$ and $\theta>0$ (with $\Theta=(0,\infty)).$ Working
again with the sufficient statistic $u_{n}:=\sum_{i=1}^{n}x_{i},$ whose
distribution conditional on $\theta$ is the $\Gamma(n,\theta)$ distribution
(the $n$-fold convolution of ${{\mathrm{Exp}}}(\theta)\equiv\Gamma
(1,\theta)),$ we have $p_{\theta}(u_{n})=\theta^{n}u_{n}^{n-1}%
\operatorname{exp}(-\theta u_{n})/(n-1)!,$ and $p(u_{n})=\int_{0}^{\infty
}p_{\theta}(u_{n})\pi(\theta)~{{\mathrm{d}}}\theta=nu_{n}^{n-1}(u_{n}%
+1)^{-(n+1)},$ and thus\footnote{This can be obtained also from the known fact
that in this setup the posterior distribution on $\Theta$ is a gamma
distribution, specifically, $\Gamma(n+1,u_{n}+1),$ and so $q_{n}^{\theta_{0}%
}=(u_{n}+1)^{n+1}\theta_{0}^{n}\operatorname{exp}(-(u_{n}+1)\theta_{0}%
)/n!$\thinspace.}
\begin{equation}
\psi_{\theta,\theta}(n)=\pi(\theta)\int_{0}^{\infty}\frac{p_{\theta}%
(u_{n})^{2}}{p(u_{n})}\,{{\mathrm{d}}}u_{n}=\frac{e^{-\theta}\theta^{2n}%
}{(n-1)!n!}\int_{0}^{\infty}u^{n-1}(u+1)^{n+1}\operatorname{exp}(-2\theta
u)\,{{\mathrm{{{\mathrm{d}}}}}}u. \label{eq:exppsi}%
\end{equation}
Our result is
\begin{theorem}
\label{th:exp} Let $x_{1},x_{2},...$ be iid ${{\mathrm{Exp}}}(\theta)$
observations, and let the prior distribution of $\theta$ be the
${{\mathrm{Exp}}}(1)$ distribution on $\Theta=(0,\infty).$
(i) For every $\theta>0$ and $n\geq1$ we have
\[
\psi_{\theta,\theta}(n)=\frac{\theta^{n-1/2}}{2^{n+1/2}n!\sqrt{\uppi }%
}[(n+\theta)K_{n+1/2}(\theta)+\theta K_{n-1/2}(\theta)],
\]
where $K_{\nu}$ denotes the modified Bessel function of the second
kind,\footnote{See, e.g., \cite{Abram}, Chapter 9.6.} and by \eqref{eq:expp} for every
$\theta_{0},\theta_{1}>0$ with $(\theta_{0}+\theta_{1})/2=\theta$, we have
\[
\psi_{\theta_{0},\theta_{1}}(n)=\operatorname{exp}(\theta-\theta_{0}%
)\psi_{\theta,\theta}(n)\left( \frac{\theta_{0}\theta_{1}}{\theta^{2}%
}\right) ^{n}.
\]
(ii) For every $\theta_{0},\theta_{1}>0$ the sequence $\psi_{\theta_{0}%
,\theta_{1}}(n)$ is strictly log-concave for $n\geq1.$
\end{theorem}
We start with two preliminary results. Put
\[
k_{n}{\;:=\;}K_{n+1/2}(\theta)
\]
and
\[
I_{n,m}:=\int_{0}^{\infty}u^{n}(u+1)^{m}e^{-2\theta u}{{\mathrm{\,{{\mathrm{d}%
}}}}}u.
\]
\begin{lemma}
\label{l:I_n,n}For every $n\geq0$ we have
\[
I_{n,n}=\frac{e^{\theta}}{\sqrt{\uppi }}\frac{n!}{(2\theta)^{n+1/2}}k_{n}.
\]
\end{lemma}
\begin{proof}
The change of variable $v=2u+1$ and Equation (9.6.23) of \cite{Abram} give
\begin{align*}
I_{n,n} & =\int_{0}^{\infty}u^{n}(u+1)^{n}\operatorname{exp}(-2\theta
u){{\mathrm{\,{{\mathrm{d}}}}}}u=\frac{e^{\theta}}{2^{2n+1}}\int_{0}^{\infty
}((2u+1)^{2}-1)^{n}\operatorname{exp}(-\theta(2u+1))\,{{\mathrm{d}}}(2u+1)\\
& =\frac{e^{\theta}}{2^{2n+1}}\int_{1}^{\infty}(v^{2}-1)^{n}%
\operatorname{exp}(-\theta v)\,{{\mathrm{d}}}v=\frac{e^{\theta}}{2^{2n+1}%
}\frac{n!}{\sqrt{\uppi}}\left( \frac{2}{\theta}\right) ^{n+1/2}%
K_{n+1/2}(\theta).
\end{align*}
\end{proof}
\begin{lemma}
\label{l:I-n-1,n+1}For every $n\geq1$ we have
\[
I_{n-1,n+1}=\frac{n+\theta}{n}I_{n,n}+\frac{1}{2}I_{n-1,n-1}.
\]
\end{lemma}
\noindent\textit{\textbf{Proof.}}\qua
First, the identity $u^{n-1}(u+1)^{n}=u^{n}(u+1)^{n-1}+u^{n-1}(u+1)^{n-1}$
gives
\begin{equation}
I_{n-1,n}=I_{n,n-1}+I_{n-1,n-1}.\label{eq:I=I+I}%
\end{equation}
Second, integration by parts yields
\begin{align*}
2\theta I_{n,n} & =\int_{0}^{\infty}u^{n}(u+1)^{n}2\theta\operatorname{exp}%
(-2\theta u)\,{{\mathrm{d}}}u=\left[ u^{n}(u+1)^{n}\left(
-\operatorname{exp}(-2\theta u)\right) \right] _{0}^{\infty}\\
& -\int_{0}^{\infty}nu^{n-1}(u+1)^{n}\operatorname{exp}(-2\theta
u)\,{{\mathrm{d}}}u-\int_{0}^{\infty}nu^{n}(u+1)^{n-1}\operatorname{exp}%
(-2\theta u)\,{{\mathrm{d}}}u\\
& =[0-0]+nI_{n-1,n}+nI_{n,n-1}%
\end{align*}
(we used $n>0$ for the value of the integrand at $u=0).$ By (\ref{eq:I=I+I})
we get
\[
2\theta I_{n,n}=(n+n)I_{n,n-1}+nI_{n-1,n-1},
\]
and thus
\begin{equation}
I_{n,n-1}=\frac{\theta}{n}I_{n,n}-\frac{1}{2}I_{n-1,n-1}.\label{eq:n,n-1}%
\end{equation}
Finally, the identity $u^{n-1}(u+1)^{n+1}=u^{n}(u+1)^{n}+u^{n-1}(u+1)^{n-1}%
+$\linebreak$u^{n}(u+1)^{n-1}$ and (\ref{eq:n,n-1}) yield
\[
I_{n-1,n+1}=I_{n,n}+I_{n-1,n-1}+I_{n,n-1}=\left( 1+\frac{\theta}{n}\right)
I_{n,n}+\left( 1-\frac{1}{2}\right) I_{n-1,n-1}.
\qquad \qquad \qed \]
\begin{proof}
[Proof of Theorem \ref{th:exp}](i) By the previous two lemmas and
\eqref {eq:exppsi}, and setting $\psi(n):=\psi_{\theta,\theta}(n),$ we have
\[
\psi(n)=\frac{e^{-\theta}\theta^{2n}I_{n-1,n+1}}{(n-1)!n!}=\frac{1}%
{\sqrt{\uppi}}\frac{\theta^{2n}}{(n-1)!n!}\left( \frac{n+\theta}{n}\frac
{n!}{(2\theta)^{n+1/2}}k_{n}+\frac{1}{2}\frac{(n-1)!}{(2\theta)^{n-1/2}%
}k_{n-1}\right) ,
\]
which simplifies to the claimed formula.
(ii) Let $\phi_{n}:=(n+\theta)k_{n}+\theta k_{n-1}$ and $\rho_{n}%
:=k_{n-1}/k_{n}.$ Using the recursion
\begin{equation}
k_{n+1}=\frac{2n+1}{\theta}k_{n}+k_{n-1}\label{eq:recurse}%
\end{equation}
(by $K_{\nu+1}(u)=(2\nu/u)K_{\nu}(u)+K_{\nu-1}(u),$ which is (9.6.26) in
\cite{Abram}) we have
\begin{align*}
\phi_{n+1} & =(n+1+\theta)k_{n+1}+\theta k_{n}=(n+1+\theta)\left(
\frac{2n+1}{\theta}k_{n}+k_{n-1}\right) +\theta k_{n}\\
& =\left( \frac{2n^{2}+3n+1}{\theta}+2n+1+\theta\right) k_{n}%
+(n+1+\theta)k_{n-1}\\
& =\left[ \left( \frac{2n^{2}+3n+1}{\theta}+2n+1+\theta\right)
+(n+1+\theta)\rho_{n}\right] k_{n},\\
\phi_{n} & =(n+\theta)k_{n}+\theta k_{n-1}=\left[ n+\theta+\theta\rho
_{n}\right] k_{n},\\
\phi_{n-1} & =(n-1-\theta)k_{n-1}+\theta k_{n-2}=(n-1-\theta)k_{n-1}%
+\theta\left( k_{n}-\frac{2n-1}{\theta}k_{n-1}\right) \\
& =\theta k_{n}-(n-\theta)k_{n-1}=\left[ \theta-(n-\theta)\rho_{n}\right]
k_{n}.
\end{align*}
Therefore
\begin{equation}
R_{n}{\;:=\;}\frac{\psi(n-1)\psi(n+1)}{\psi(n)^{2}}=Q_{n}(\rho_{n}),\nonumber
\end{equation}
where
\begin{equation}
Q_{n}(\rho):=\frac{n\left[ \left( \frac{2n^{2}+3n+1}{\theta}+2n+1+\theta
\right) +(n+1+\theta)\rho\right] \left[ \theta-(n-\theta)\rho\right]
}{(n+1)(n+\theta+\theta\rho)^{2}}.\label{eq:Q}%
\end{equation}
The proof that
$Q_{n}(\rho_{n})<1\label{eq:R<1}$
for all $n\geq2$ and $\theta\ge 0$ is quite technical and is relegated to
Appendix \ref{sec-a:expexp}.
\end{proof}
\textbf{General exponential prior: }If the prior on $\theta$ is a general
${{\mathrm{Exp}}}(\lambda)$ distribution for some $\lambda>0$ not necessarily
equal to $1,$ then the linear transformation $x\rightarrow x/\lambda$ reduces
it to the case of Theorem \ref{th:exp}, and so the sequence of expected
posteriors is log-concave, and thus unimodal, for $n\geq1.$
\section{The omitted calculations of Appendix A.4}
At the end of Appendix A4 we write:
\begin{quote}
\textquotedblleft the coefficients of the powers of $m$ in $A$ and $E$ are
positive for all $\theta \geq 0$, as shown by direct calculations that we
omit.\textquotedblright
\end{quote}
\noindent We provide these calculations here.
\bigskip
\noindent Let $a_{i}$ denote the coefficient of $m^{i}$ in $A,$ and $e_{i}$
the coefficient of $m^{i}$ in $E.$
\bigskip
\noindent For $A$ we have:
\begin{itemize}
\item $a_{4}=8>0;$
\item $a_{3}=72>0;$
\item $a_{2}=4\,${$\theta $}$^{2}-16\,\theta +230>4${$\theta $}
^{2}-16\,\theta +16\geq 0;$
\item $a_{1}=8\,${$\theta $}$^{3}+4\,${$\theta $}$^{2}-56\,\theta
+302>4\theta ^{2}-56\theta +196\geq 0;$ and
\item $a_{0}>0,$ since its minimum over $\theta \geq 0$ is $\approx 108$
(attained at\footnote
The minimum here (and in the other cases below) was obtainded using Maple.}
\theta \approx 0.73).$\newpage
\end{itemize}
\noindent For $E$ we have:
\begin{itemize}
\item $e_{7}=128>0;$
\item $e_{6}=128\,${$\theta $}$^{2}-256\,\theta +1920>128\,${$\theta $}
^{2}-256\,\theta +128\geq 0;$
\item $e_{5}=256\,${$\theta $}$^{3}+1088\,${$\theta $}$^{2}-3200\,\theta
+12064>800\,${$\theta $}$^{2}-3200\,\theta +3200\geq 0;$
\item $e_{4}=208\,${$\theta $}$^{4}+2432\,${$\theta $}$^{3}+2816\,${$\theta
}$^{2}-16192\,\theta +41024>2000\,${$\theta $}$^{2}-17000\,\theta +40000>0;$
\item $e_{3}=64\,${$\theta $}$^{5}+1824\,${$\theta $}$^{4}+9344\,${$\theta $
$^{3}-416\,${$\theta $}$^{2}-42336\,\theta +81184>0$, since its minimum for
\theta \geq 0$ is $\approx 49317$ (attained at $\theta \approx 1.09);$
\item $e_{2}=480\,${$\theta $}$^{5}+6208\,${$\theta $}$^{4}+17664\,${$\theta
$}$^{3}-13408\,${$\theta $}$^{2}-60128\,\theta +92928>0,$ since its minimum
for $\theta \geq 0$ is $\approx 43609$ (attained at $\theta \approx 1.04);$
\item $e_{1}=1312\,${$\theta $}$^{5}+9200\,${$\theta $}$^{4}+16096\,${
\theta $}$^{3}-21120\,${$\theta $}$^{2}-43776\,\theta +56448>0,$ since its
minimum for $\theta \geq 0$ is $\approx 18075$ (attained at $\theta \approx
0.97);$
\item $e_{0}=16\,${$\theta $}$^{6}+1152\,${$\theta $}$^{5}+4896\,${$\theta $
$^{4}+5568\,${$\theta $}$^{3}-10368\,${$\theta $}$^{2}-12672\,\theta
+13824>0,$ since its minimum for $\theta \geq 0$ is $\approx 1981$ (attained
at $\theta \approx 0.90).$
\end{itemize}
\end{document}
|
\section{Introduction}
The communication distance of point-to-point quantum key distribution (QKD) is fundamentally limited by losses in optical channel to a hundreds of kilometers \cite{Pirandola2017}.
Current long-distance QKD systems use chain of connected trusted nodes that translate the generated key from one end of the chain to another \cite{Elliott_2002, Sasaki:11,Dynes2019,Chen2021}.
Alternatively, the development of the quantum internet pursuits the creation of entanglement-enriched quantum networks, such as quantum repeater, that could be used for QKD without a need of the trusted nodes \cite{Briegel98, Guha2015}. For a cohesive operation of an existing QKD lines together with the future quantum networks a mutual encoding of quantum information is necessary.
Conventional discrete variable quantum repeater provides photonic entanglement in the form of the Bell pairs in dual rail encoding (e.g. polarization) \cite{Sangouard11,Yu2020}.
The subcarrier wave (SCW) encoding is a promising for point-to-point \cite{PhysRevLett.82.1656,Gleim:16}, Plug\&Play \cite{Bannik:21}, continuous variables \cite{Melnik2018,SCW-CV} and twin-field QKD \cite{Chistiakov:19} due to its robustness to environmental effects acting on a fiber line, interferometer-free optical scheme and its capacity for spectral multiplexing \cite{mora2012simultaneous}.
In SCW-QKD the information is encoded in a value of the phase of the applied sinusoidal optical phase modulation, that leads to specific multi-mode state.
However, this multi-mode state complicates a creation entanglement distribution protocol for this encoding that limits potential long-distance SCW quantum communications.
One way to circumvent this issue would be a conversion of an information from SCW encoding into the one with available quantum repeater protocol.
In this letter, we propose and demonstrate proof of principle experiment on interfacing the SCW-QKD with dual rail encoding. First we describe the scheme of the interface and its simple theoretical model. Next we portray the experimental setup and present its results. Finally, we discuss the perspectives and further possible improvements of the interface.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Fig/SCW-interface-version-2.pdf
\caption{ Phase modulator (PM), spectral filter (SF), optical path matching delay line (DL), beam blocker (BB), non-polarizing beam splitter 50/50 (BS), polarization beam splitter (PBS), carrier frequency (0), The $\{-1,0,1\}$ indices in the circles indicate the presence of corresponding subcarrier frequencies with their phases on top.
}
\label{fig:scheme}
\end{figure}
\section{The principle scheme of the interface}
In SCW-QKD the information is encoded by applying a sinusoidal phase-modulation with modulation depth $\beta$, modulation frequency $\Omega$, and phase $\phi_{in}$ to a coherent state with complex amplitude $\alpha$ and carrier frequency $\omega_0$.
An optical state after phase modulator is described by wavefunction \cite{Miroshnichenko:17}:
\fla{
\ket{\Psi_{\text{SCW}}} = \bigotimes^{\infty}_{m=-\infty} \ket{ \alpha_0 J_{m}(\beta)e^{i m \phi_{in}} }_m, \label{eq::SCW-State}
}
where multiplication is carried over $m\in \mathbb{Z}$ indices of sidebands with frequencies $ \omega_0 + i m\Omega$, $J_{m}(\beta)$ is m-th order Bessel function of the first kind.
The value of the phase $\phi_{in}$ encodes the useful information.
For example, in typical SCW-QKD with BB-84-like protocol Alice sends to Bob logical bits in two non-orthogonal bases with single basis being formed by pairs of states (\ref{eq::SCW-State}) with $\phi_{in}$ being shifted by $\pi$ with respect to each other, such as ${0,\pi}$ and ${\pi/2,3\pi/2}$.
In turn, Bob measures the Alice's state in given basis by phase modulating it with the same $\beta$ and $\Omega$, but his own subcarrier phase and performing photon detection on sidebands ($m\neq0$) that effectively acts as a projection on a given state \cite{Miroshnichenko:17}.
For interfacing the state (\ref{eq::SCW-State}) with some dual encoding the phase $\phi_{in}$ has to be mapped into superposition of two modes.
The principle scheme of the interface is depicted in Fig. \ref{fig:scheme}.
The input state $\ket{\Psi_{\text{SCW}}}$ with horizontal polarization enters spectral filter that reflects the carrier component ($m=0$) with power coefficient $r$ and transmits it with power coefficient $1-r$.
Meanwhile the sidebands ($m\neq0$) are transmitted through the filter with power coefficient $1-\varrho$ and reflected with power coefficient $\varrho$, where $\varrho$ is called as filter's sideband suppression factor.
The reflected carrier is routed by an optical circulator to a phase modulator, that modulates the carrier with the same modulation depth $\beta$, modulation frequency $\Omega$, but a fixed phase $\phi_{\text{LO}}$.
After the phase modulator another spectral filter with the same $r$ and $\varrho$ separates the carrier frequency component from the generated sidebands.
The sidebands from the original state and the one generated locally are overlapped on 50-50 beam splitter.
The wavefunction for two spatial modes at the output of the beam splitter labeled as '$+$' and '$-$' is
\fla{
\ket{\Psi} = \ket{\psi_{+}} \otimes \ket{\psi_{-}},
}
with
\fla{
\ket{\psi_{\pm}} = \ket{\alpha_0\frac{J_0(\beta)((1-r) \pm J_0(\beta)r(1-r)) }{\sqrt{2} }}_0 \cdot \nonumber \\
\bigotimes^{\infty}_{m\neq0} \ket{\alpha_0 J_m(\beta) (1-\varrho) \frac{ e^{im\phi_{in}} \pm e^{im\phi_{\text{LO}}} r J_0(\beta) }{\sqrt{2}}}_m . \label{eq::dr}
}
For good enough filters ($r\approx 1$, $\varrho \approx 1$) and small modulation depth ($J_0(\beta)\approx 1$) the state \eqref{eq::dr} represents the dual rail encoding with two spatial modes for each individual sideband with $m\neq 0$, e.g. the states with phase difference of 0 or $\pi$ with respect to $\phi_{LO}$ will follow different channels of the beam splitter.
The subsequent conversion in polarization encoding is typical.
The both modes are combined on polarization beam splitter with one mode's polarization being rotated by $90^{o}$ with half-waveplate.
The resulted state for the given first order sidebands ($m=\pm 1$) up to a first order in photon number is:
\fla{
\ket{\Psi_{\pm 1}} \approx
\ket{0_H 0_V}_{\pm1} + \sqrt{2}\alpha J_{\pm1}(\beta) e^{\pm i(\phi_{in}-\phi_{\text{LO}}) /2 }\cdot \nonumber \\
\bigg(\cos \frac{\phi_{in}-\phi_{\text{LO}}}{2} \ket{1_H 0_V}_{\pm1} \pm \sin\frac{\phi_{in}-\phi_{\text{LO}}}{2} \ket{0_H 1_V}_{\pm1} \bigg).
\label{eq::pstate}
}
Once $\phi_{\text{LO}}$ is fixed to 0, the phase information $\phi_{in}$ is mapped on a probability amplitude of photon being in given polarization mode as in dual rail encoding.
\section{Experiment}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Fig/Exp-setup.pdf
\caption{
The scheme of the experimental setup that consists from laser diode (LD), optical isolator (OI), optical attenuator (ATN), 50-50 beam splitter (BS), polarization beam splitter (PBS),
phase modulators (PM), optical circulator (OC), spectral filter (SF), single photon counting module (SPC), quarter wave plate ($\lambda/4$), half wave plate ($\lambda/2$). \textcolor{red}{Red} and \textcolor{blue}{Blue} arrows indicate the carrier and sidebands direction and their polarization, respectively.
}
\label{fig:exp-setup}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Fig/HV-counts.pdf
\\
\includegraphics[width=\linewidth]{Fig/VisVSBeta.pdf}
\caption{ \textbf{a)} Dependence of the count rate for $\ket{H}$ (\textcolor{red}{red} dots) and $\ket{V}$ (\textcolor{blue}{blue} dots) states at different values of phase difference between two microwave generators. The data is obtained with modulation depth $\beta=0.15$. The solid lines represent theoretical fit.
\textbf{b)} Dependence of interference visibility on modulation depth. The solid curve is theoretical fit.
}
\label{fig:interference}
\end{figure}
\begin{figure*}[ht]
\centering
\includegraphics[width=0.48\linewidth]{Fig/H-DensityMatrix.pdf
\includegraphics[width=0.48\linewidth]{Fig/V-DensityMatrix.pdf
\\
\includegraphics[width=0.48\linewidth]{Fig/45-DensityMatrix.pdf}
\includegraphics[width=0.48\linewidth]{Fig/m45-DensityMatrix.pdf}
\caption{ Real and imaginary parts of reconstructed density matrices for polarization states at different values of the phase difference in $\{\ket{H},\ket{V}\}$ basis.}
\label{fig:tomography}
\end{figure*}
The experimental scheme is depicted in Fig. \ref{fig:exp-setup}.
The Fiber-Bragg-Grating stabilized laser diode (TeraXion PureSpectrum LM) protected by optical isolator together with an optical attenuator are used to generated continuous wave
at 1550 nm.
The laser is guided through an optical circulator into 4-port fiber 50/50 beam splitter (BS).
Two SCW states are generated using the same carrier in a Sagnac-type interferometer, where traveling clockwise and counter-clockwise waves are modulated by two separate $\text{LiNb}\text{O}_3$ phase modulators PM1 an PM2, respectively.
The phase modulators are driven by two phase-locked microwave generators with identical frequency of $\Omega=2\pi\cdot 4.8$ GHz, but different phases $\phi_{\text{in}}$ and $\phi_{\text{LO}}$, respectively.
The implemented design helps to circumvent a problem with a need of two identical spectral filters, that is necessary for principle scheme in Fig. \ref{fig:scheme}.
The generated optical sidebands are interfered on 50-50 BS and routed to ports 1 and 2 according to their phase difference.
The carrier component, meanwhile, is routed back to a incoming port 1 with isolation more than 30 dB due to constructive interference between clock and counter-clock wise waves in Sagnac loop.
The polarization of the sidebands travelling from the port 2 is rotated by 90 degree with fiber Faraday rotator.
The radiation from the port 1 is guide by the optical circulator to a fiber polarization beam splitter (PBS), where this radiation is combined with sidebands from port 2 having orthogonal polarization.
The polarization insensitive fiber Bragg grating spectral filter (TeraXion OF) with r=0.99 and $\rho=10^{-4}$ separates the prepared polarization state of the sidebands from the carrier.
The light is sent to free space via fiber output coupler for further analyze.
The polarization of the sidebands is characterized with conventional polarization analyzer consisting of quarter and half waveplates together with PBS, which both channels are coupled into single photon counting modules (ID Quantique ID210) with quantum efficiency $\epsilon=$10\% and dark count rate $\gamma\sim$100 Hz. The detectors are gated
at 100 MHz with gate duration of $\delta t=3.3$ ns. For this time window the amplitude of the carrier corresponds to value of $\alpha_0 \sim 10^{-3}$.
Multi-channel frequency counter is used to monitor the count rates from both single photon counting modules.
\section{Results and Discussion}
First, we scan $\phi_{\text{in}}-\phi_{\text{LO}}$ between two SCW states by manually changing phase of a single microwave generator. The normalized count rate of horizontally and vertically polarized photons at different value of the phase difference are presented in Fig. \ref{fig:interference}a, where each point is accumulated during $T=10$ s.
The interference with visibility of $\sim95\%$ clearly indicates the transfer of the phase information from SCW state into polarizational degree of freedom according to \eqref{eq::pstate}.
We measure visibility at different values of modulation depth $\beta$ with all sidebands being detected simultaneously.
The results are depicted in Fig \ref{fig:interference}b.
At small value ($\beta<0.1$) of modulation depth the number of detected photons drops below the dark count rate that leads to low visibility.
At large value of $\beta$ ($\beta>1$) the significant portion of energy is distributed at higher order sidebands ($m\ge2$) that interfere as $\sim \cos (\frac{m(\phi_{in}-\phi_{\text{LO}})}{2})$ and reduce overall visibility, when all sidebands are detected simultaneously by a single detector.
At the same time the visibility of $92 \pm 3\%$ is achievable within range $0.1 \le \beta\le0.7$.
It is worth noting, that the measured decrease in visibility at low $\beta$ is not fundamental, but rather implementation specific and can be circumvented, if detector with lower count rate and spectral filter with larger extinction ratio $r/(1-r)$ are used.
We model the number of clicks with given polarization as:
\fla{
n_{H(V)} \approx \left(\gamma + \frac{\epsilon}{\delta t}\bra{\psi_{+(-)}} n \ket{\psi_{+(-)}} \right) \cdot T,
\label{eq::clicks}
}
where $n = \sum_{m} \hat{a}^{\dagger}_m \hat{a}_m $ is number operator with $\hat{a}_m$ ($\hat{a}^{\dagger}_m$) being annihilation (creation) operator for m'th spectral mode, $\epsilon$ is a quantum efficiency of the detector, $\delta t$ is gate duration time, $T$ is sampling time, $\gamma$ is dark count rate.
The experimentally acquired data of normalized clicks and visibility is fitted using \eqref{eq::clicks}.
The polarization states on the opposite sidebands with m=1 and m=-1 have a $\pi$ phase shift with respect to their polarizational states, that results in a mixed polarization for $\phi_{in}-\phi_{\text{LO}} \in \{\pi/2,3\pi/2\}$ if both sidebands are measured simultaneously.
To characterize the conversion, we perform the polarization tomography for only a single sideband ($m=+1$) with $\beta=0.15$ by adjusting the laser's frequency with respect to spectral filter's reflection band such that only one sideband is transmitted. For each value of phase difference ($0,\pi,\pi/2,3\pi/2$) we measure polarization in different polarization basis.
The acquired data is used to reconstruct the polarizational density matrix by maximum likelihood method \cite{Rehacek2001} with reconstructed density matrices being presented in Fig. \ref{fig:tomography} with fidelity of more than 92\% with respect to the target state defined by \eqref{eq::pstate}. Fidelities for the states are summarized in Table \ref{table:fidelities}.
The small imaginary part of non-diagonal elements of the density matrix we attribute to a jitter in optical path difference between paths for $\ket{H}$ (port 1 to PBS) and $\ket{V}$ (port 2 to PBS) photons.
The jitter introduces extra complex phase between $\ket{H}$ and $\ket{V}$ that leads to unwanted ellipticity of the polarization.
We passively stabilized the optical scheme by placing it in thermally and acoustically isolating foam and reducing the fiber connection length. However, some sort of active stabilization may still be required for achieving fidelity $\ge 99$\%.
\begin{table}[htp]
\centering
\begin{tabular}{|c|c|c|}
\hline
$\phi_{in}-\phi_{LO}$ & Polarization state & Fidelity \\ \hline
$0$ & $\ket{H}$ & $96\pm3$ \% \\ \hline
$\pi$ & $\ket{V}$ & $96\pm3$ \% \\ \hline
$\pi/2$ & $(\ket{H}+\ket{V})/\sqrt{2}$ & $95\pm3$\% \\ \hline
$3\pi/2$ & $(\ket{H}-\ket{V})/\sqrt{2}$ & $95\pm3$\% \\ \hline
\end{tabular}
\caption{The correspondence between phase and polarization and the achieved fidelities.}
\label{table:fidelities}
\end{table}
The proposed interface may be implemented in integrated quantum optical fashion \cite{Sibson2017}.
This approach would minimize the phase fluctuations, device size and hence increase its stability. Moreover, this technology would allow to build two identical spectral filters for effective conversion of the external SCW state as in the proposed scheme.
In this case, this scheme can be used in SCW-QKD for simultaneous detection of two state in a single basis.
If single photon detectors are placed at output of each channel of 50-50 beam splitter, the states with phases $\phi_{\text{LO}}$ and $\phi_{\text{LO}}+\pi$ are distinguishable by a click of the corresponding detector.
The use of scheme in point-to-point SCW QKD may increase the key rate twice compare to the conventional approach \cite{Gleim:16} that detects only a single state in given basis at a time.
In summary, we presented the way for conversion of phase information from SCW into a dual rail encoding.
The experimentally realized in-fiber conversion from SCW into polarizational encoding shows fidelities above 92\%.
The main limitation for further increase of visibility is extinction ratio of the spectral filter and dark count rate of the single photon counting modules.
The proposed interface opens the way for connecting existing quantum networks with different physical layers and increase QKD distance.
\begin{backmatter}
\bmsection{Funding} Authors appreciate support within framework project \# 00075-02-2020-051/1 from 02.03.2020.
\bmsection{Acknowledgments} Authors profound thank S.A. Moiseev for fruitful discussions. We are grateful to M.M. Minnegaliev and M.A. Smirnov for lending us optomechanics and fiber-optics.
\bmsection{Disclosures} The authors declare no conflicts of interest.
\bmsection{Data Availability}
Data underlying the results presented in this paper may be obtained from the authors upon reasonable request.
\end{backmatter}
|
\section{INTRODUCTION}
\label{sec:intro}
A standard specification to avoid local turbulence is to keep all optical surface within $\pm$ 1.5$^\circ$ with respect to the ambient temperature. This specification can be challenging for active devices. This is especially true for Deformable Mirrors based on the voice-coil technology, whose actuators produce a significant amount of heat very close to the optical surface. This technology is currently used for instance for the Adaptive Optics of the CHARA \cite{2020SPIE11446E..22A}, the VLTI Auxiliary Telescope arrays \cite{2020SPIE11446E..06H,2019A&A...629A..41W}, several secondary mirrors, and is the baseline for the GRAVITY+ project\footnote{https://www.mpe.mpg.de/7480772/GRAVITYplus\_WhitePaper.pdf}.
This specification can also be challenging for optical devices preferentially operated at typical laboratory temperatures ($20^\circ$C), but which are deployed inside the Coudé train of telescopes where the ambient air temperature ranges from $0^\circ$C to $20^\circ$C. For a Deformable Mirror in which the stiffness is linked to the temperature, such a requirement imposes to provision larger margins on the critical trade-off between speed, stroke and heat. Yet, the NAOMI project for the VLTI-AT array ultimately had to implement a system to maintain the Deformable Mirror at a constant temperature of about $20^\circ$C, in order to achieve the desired stroke. This heating system deviates from the initial requirement of $\pm1.5^\circ$C for the surface temperature, without noticeable loss in performance. It questions the adequacy of this tight specification for the specific case of a Deformable Mirror which is ultimately part of an AO system correcting the turbulence.
\newpage
\section{INSTRUMENTAL SET-UP}
\label{sec:setup}
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=\linewidth]{Monture.png}
\end{tabular}
\end{center}
\caption[\label{fig:monture}]
{Sliced view of the gimbal without the Corrective Optics Mount. The red circle shows the position of the heater.}
\label{fig:monture}
\end{figure}
This section describes the hardware and the software of the experiment.
The experiment was executed in the actual test bench of the GRAVITY+ project, at the Observatoire de la Cote d'Azur \cite{2022_GRAVITYPLUS_BENCH}. The bench features several internal lights, and a long beam focalisation to create a pupil of about 100mm in the Deformable Mirror with a f-number of about 50. The Deformable Mirror is installed in a mount called the Corrective Optics (CO), where it is facing down. This CO is the original model that will be used inside the Coudé train of the Unit Telescopes (UTs) of the Very Large Telescope Interferometer (VLTI). The software part is a combination of both the software used for general testing of the CO (beyond this experiment) and dedicated software used/developed for this experiment.
\subsection{Hardware}
This mount can be decomposed into the mounting's structure, called the Corrective Optics Mount (COM) and the gimbal inside of it. The gimbal (see Fig\ref{fig:monture}) is the piece in which the Deformable Mirror will be inserted. It can be controlled to compensate for tip-tilt misalignment.
The final device of GRAVITY+ will be a 41x41 Deformable Mirror from ALPAO with a pitch 2.5mm for an active pupil of 100mm in diameter, and a clear aperture of 106mm. For the purpose of this experiment, we did not use a Deformable Mirror but a flat thick mirror instead (half grey cylinder on Fig.\ref{fig:monture}). Attached to this flat mirror is a mass placed in height so that it reproduces the geometry and mechanical constraints of the actual Deformable Mirror while leaving the mirror's backface accessible.
In the final configuration, the heat will be generated by the actuators of the Deformable Mirror, especially in bad seeing conditions. In the case of the flat mirror, we had to implement an external source of heat in order to be able to drive the temperature of the optical surface. We used a 40W heater\footnote{Reference : KHLVA-202/10 FLEXIBLE HEATER, POLYMIDE, 40W, 28VAC} glued to the backface of the flat mirror (this backface is thus facing up). Then, a Pt100 probe was placed on the same side of the mirror as the heater, at a distance comparable to the mirror's thickness so the monitored temperature corresponds to the optical surface's. A second Pt100 probe was used to check the ambient temperature. Overall, the whole set-up is to mimic the behaviour of a Deformable Mirror implemented in a realistic facing-down configuration (actually the exact one of GRAVITY+) and generating its own source of heat.
It should also be noted that the gimbal has a $13.26^\circ$ inclination with respect to the horizontal. We fixed the heater in the lowest part of the mirror, but we do not expect this to be important because the overall mirror should have a rather uniform temperature. The expected behaviour is that the heated surface would produce a flow of hot air in front of the field of view. This moving bubble would hence produce local turbulence. To observe this turbulence itself, we used a SID4\textsuperscript{\textregistered} wave front sensor \footnote{SID4 version 4.3 produced by PHASICS \copyright} that produce real-time intensity and phase map at a 7.5 Hz rate on a sensor of 3.6 x 4.8 mm$^2$ sampled with 120 x 140 pixels$^2$. This device was placed in the nominal optical path as designed for GRAVITY+. At this place, the pupil (106mm clear aperture) is slightly larger than the PHASICS wavefront sensor. We thus restrict the analysis to a reduced pupil of about 90mm in diameter.
\subsection{Software}
We monitored both the mirror's temperature and the power delivered to the heater thanks to a Python routine sent through the Beckhoff\textsuperscript{\textregistered} controller of the gimbal mount. We recovered phase maps thanks to the software of the SID4\textsuperscript{\textregistered} wave front sensor. A description of the maps we obtained is provided in section \ref{sec:measurements}.
\section{MEASUREMENTS}
\label{sec:measurements}
We performed a series of measurements covering the range of $+0^\circ$ to $+22^\circ$ above ambient temperature with a $2^\circ$ step between each measurement. Every temperature is tested twice. A first time during the increase in temperature of the mirror, and a second time during the decrease in temperature of the mirror, which is not driven, but only happens in a passive way. The warm up sequence took about 3h, and the cool down sequence about 4h. Altogether, this gives a set of 23 observations that are analysed in section \ref{sec:analysis}.
Each observation consists into the acquisition of a cube of 512 continuous phase images taken with the SID4\textsuperscript{\textregistered} at a 7.5 Hz rate. We obtain a 68.27 s film showing the evolution of local turbulence in front of the mirror. This is trade-off between the temporal resolution, resolution in frequency, and the size of the individual files.
For illustration, we extract a series of snapshots in Fig.\ref{fig:snaps} where we have subtracted the average image to each snapshot. These snapshots were taken at a fifth of the nominal rate (1.5 Hz) to make the speed and amplitude of the changes we observe visible. Time goes from the left to the right and from up to bottom. One can see the appearance of a bubble on the right of the field which is then transferred to the left. The process takes the 6 images to happen hence a timescale of $\sim 4$ seconds for the turbulence as a rough first estimate (see section \ref{sec:timeanalysis} for more precise analysis of the time-dependant behaviour of the turbulence). The overall wavefront residual in the images is about 20\,nm, which is smaller than the typical fitting error budget in modern high resolution adaptive optics. Our experiment should therefore be able to detect any additional problematic turbulence that may arise from the heating of the mirror.
The following analyses of the turbulence in section \ref{sec:analysis} below are entirely based on these measurements.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.33\linewidth, trim=52 0 59 0,clip]{Figure_1.png}
\includegraphics[width=0.33\linewidth, trim=52 0 59 0,clip]{Figure_2.png}
\includegraphics[width=0.33\linewidth, trim=52 0 59 0,clip]{Figure_3.png}
\end{tabular}
\end{center}
\end{figure}
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.33\linewidth, trim=52 0 59 0,clip]{Figure_4.png}
\includegraphics[width=0.33\linewidth, trim=52 0 59 0,clip]{Figure_5.png}
\includegraphics[width=0.33\linewidth, trim=52 0 59 0,clip]{Figure_6.png}
\end{tabular}
\end{center}
\caption[\label{fig:snaps}]
{Six snapshots taken at 1.5 Hz to illustrate the turbulence at ambient temperature (the entire cube is recorded at a frame rate of 7.5 Hz). Time goes from the left to the right and from up to bottom. Color scale is constant from one picture to the next and gives the path difference in nm.}
\label{fig:snaps}
\end{figure}
\section{RESULTS/ANALYSIS}
\label{sec:analysis}
\subsection{Time-dependant analysis}
\label{sec:timeanalysis}
\begin{figure}[ht]
\centering
\includegraphics[width=8cm]{MapTAU0.png}
\caption{Example of map of characteristic time $\tau_0$ (in seconds), here for a temperature difference $\Delta T=+16^\circ$ with respect to the ambiant temperature.}
\label{fig:tau0map}
\end{figure}
To understand the time-dependant behaviour of the turbulence, we performed an analysis on each observation (i.e. each series of 512 images for a given temperature).
This analysis consists into defining a characteristic timescale of variation $\tau_0$ for the turbulence in each pixel of the field. We chose to define it arbitrarily as in Equation \ref{eq:deftau}:
\begin{equation}
\label{eq:deftau}
\boxed{\frac{R_I(\tau_0)}{R_I(0)}=\frac{1}{2}}
\end{equation}
Where $R_I(\Delta)$ is the auto-correlation of the image matrix $I$ computed for a shift $\Delta$ as defined in Equation \ref{eq:defcor}.
\begin{equation}
\label{eq:defcor}
R_I(\Delta)=\int I(t).I^*(t-\Delta)~\text{d}t
\end{equation}
We obtain a map of $\tau_0$ over the whole field for a given temperature as shown in Fig.\ref{fig:tau0map}. That is 23 different maps for which we retain the mean value, minimum value and maximum value and display it in Fig.\ref{fig:taus}. It should be noted that the x axis corresponds to the number of the observation. Then the upper figure illustrates the variations of $\tau_0$ in a logarithmic scale while the lower figure shows the temperature associated to each observation.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.65\linewidth]{TAUS.png}
\end{tabular}
\end{center}
\caption[\label{fig:taus}]
{\textbf{Upper pannel :} Evolution of $\tau_0$. Red line is the mean value, surrounded by maximum and minimum. Dashed black line shows the acquisition rate (x axis represents the number of the measurement).\\
\textbf{Lower pannel :} Evolution of temperature for every measurement (x axis represents the number of the measurement).}
\label{fig:taus}
\end{figure}
One can see that the order of magnitude of $\tau_0$ ($\sim 1/10$ s) is well within range of any AO system. But more surprisingly, there is no correlation whatsoever between the timescale of variation of the turbulence and the temperature of the mirror. This result is surprising since one would expect the turbulence to be faster when the optical surface (and the air below it) is heated-up, hence the timescale should decrease when the temperature increases and vice-versa.
\subsection{Space-dependant analysis}
\label{sec:spaceanalysis}
Not only should the turbulence be slow enough so that our AO can follow it, but it should be spatially manageable too.
To make sure that this is the case, we projected our phase maps on the Zernike polynomial basis (limiting it to the 20 firsts polynomials in the Noll convention system). We obtain 22 observations of 512 snapshots for each temperature and for each polynomial. We averaged each time series, so that we get one average observation per temperature per polynomial. For further analysis, we will call "low order" polynomials those bellow the 5$^\text{th}$ Noll Zernike. "intermediate order" polynomials will be higher than the 5$^\text{th}$ and lower than the 10$^\text{th}$ Noll. Finally, "high order" modes will encompass 10$^\text{th}$ Noll and above.
In Fig.~\ref{fig:Zs} we show the evolution of the total power in all the Zernikes. Once again, there is no correlation with the temperature of the optical surface. However, one may expect the power to transfer from the low order to the high order modes when increasing the temperature and the other way around when decreasing it. This only seems to fit with the model of a bubble of air being heated-up, and forcing energy to be dissipated at smaller scales. To check this, we normalized the power in all the Zernikes to get figure \ref{fig:ZNs}. However, no variation whatsoever is observed in the relative power of the different orders. It seems that the heating of the mirror did not trigger more turbulence, nor did it influence the length scale of the pre-existing turbulence.
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.65\linewidth]{Z.png}
\end{tabular}
\end{center}
\caption[\label{fig:Zs}]
{\textbf{Upper pannel :} Evolution of the total power in all the observations projected in the Zernike basis. Blue is for the low order modes, red is for intermediate order modes, and grey is for the high order modes (x axis represents the number of the measurement). The power is expressed in nm$^2$. Typical wavefront residual when including all modes is about 15\,nm.\\
\textbf{Lower pannel :} Evolution of temperature for every measurement (x axis represents the number of the measurement).}
\label{fig:Zs}
\end{figure}
\begin{figure} [ht]
\begin{center}
\begin{tabular}{c}
\includegraphics[width=0.65\linewidth]{Z_N.png}
\end{tabular}
\end{center}
\caption[\label{fig:ZNs}]
{\textbf{Upper pannel :} Evolution of the normalized power in all the observations projected in the Zernike basis. Blue is for the low order modes, red is for intermediate order modes, and grey is for the high order modes (x axis represents the number of the measurement).\\
\textbf{Lower pannel :} Evolution of temperature for every measurement (x axis represents the number of the measurement).}
\label{fig:ZNs}
\end{figure}
\section{CONCLUSION}
\label{sec:conclusion}
Both our time-dependant analysis and our spatial analysis led to the conclusion that no influence of the temperature of the optical surface can be measured for our 100mm-diameter mirror, facing down, and mounted in a gimbal mount. We have explored temperature difference with respect to the ambient up to up $+20^\circ$.
To date, we explain this counter intuitive result with the hypothesis that the power injected in the mirror with the heater is mainly evacuated on the upper face of the mirror (see Fig.\ref{fig:monture}), and that the gimbal structure is sufficient to separate the upper flow from the flow in front of the optical surface. Further analysis replacing the flat mirror with an actual Deformable Mirror should confirm these results.
Our experiment validates \emph{a-posteriori} the choice of the NAOMI project to warm up the Deformable Mirror in order to maintain the temperature of the Deformable Mirror around $20^\circ$C (in order to ensure that the device delivers the required performances). We confirm by experiment in laboratory that the additional turbulence generated by the surface temperature being above ambient temperature (up to $+15^\circ$C on the coldest night at Paranal observatory) is negligible.
As a next step, we plan to perform a similar experiment but using the 41x41 Deformable Mirror of GRAVITY+. Instead of heating the mirror with an external heater, we will play typical sequence of correction for bad seeing which may warm up the optical surface and the housing of the Deformable Mirror. The exact value of this warm-up will depend on the actual heat extraction efficiency. The results presented in this work on a reference flat mirror will be an interesting reference point to compare with the the measurement obtained with the real Deformable Mirror.
In a nutshell, we performed an experiment which demonstrates that heating a facing-down mirror up to $+20^\circ$ above ambient temperature does not trigger significant turbulence. Hence the possibility to discuss the general statement that every optical surface should be kept within $\pm$1.5$^\circ$ around ambient temperature.
\acknowledgments
We address our greatest thanks to the Laboratoire Lagrange (Nice, France) and its members for their implication in this work.\\
This work is supported by the Action Spécifique Haute Résolution Angulaire (ASHRA) of CNRS/INSU co-funded by CNES.\\
This work is supported by the French National Research Agency in the framework of the “Investissements d’avenir” program (ANR-15-IDEX-02)\\
Finally, we thank the SPIE organisation for permitting us to present this work to the rest of the community, which allowed new fruitful development prospects.
|
\section{Entropy Decay for GNS-symmetric quantum Channels}
\subsection{GNS symmetric quantum channels}\label{sec:GNSsymmetric}
In this section, we discuss the entropy decay of quantum channels that are symmetric with respect to a state, a condition known as GNS detailed balance. The trace symmetric case is often referred to as infinite temperature, whereas the finite temperature cases at $\beta > 0$ usually satisfy the symmetric condition to the Gibbs state $\sigma_\beta\propto e^{-\beta H}$, where $\beta$ is an inverse temperature. Interesting examples from quantum mechanics and quantum Many-body system includes Bosonic Orstein-Unlenbeck (OU) semigroup\cite{huber2017geometric}, Fermic OU semigroup \cite{carlen2017gradient,carlen2014analog} and quantum Gibbs sampler (see e.g. \cite{kastoryano2016quantum,bardet2021entropy,bardet2021rapid,bardet2021modified,capel2020modified,temme2017thermalization,spohn1978irreversible,majewski1996quantum}). Mathematically, the difference between state symmetry and trace symmetry does not appear in the commutative setting, because every probability measure induces a trace. However, in the noncommutative setting, there are type III von Neumann algebras, which do not even admit a trace. In this situation, the state symmetry is more reasonable and also the only one can expect. We will show in this section that the state symmetric case can be obtained from trace symmetric results using a method call Haagerup reduction. This method is used to prove results for type {III} von Neumann algebras by reducing to the framework of tracial von Neumann algebras.
Let ${\mathcal M}$ be a von Neumann algebra and $\phi$ a normal faithful state.
Let us first state the symmetry condition with respect to a state.
\begin{defi}\label{defi:GNS}
We say a quantum Markov map $\Phi:{\mathcal M} \to {\mathcal M}$ is GNS-symmetric with respect to $\phi$ (in short, $\phi$-symmetric) if \[\phi(\Phi(x)y)=\phi(x\Phi(y)), \hspace{.1cm} \forall \hspace{.1cm} x,y\in {\mathcal M}\hspace{.1cm}. \]
We call the pre-adjoint $\Phi_*:{\mathcal M}_*\to {\mathcal M}_*$ a $\phi$-symmetric quantum channel
\end{defi}
This definition is a generalization of the detailed balance condition for classical Markov chains. In particular, $\phi=\phi\circ\Phi=\Phi_*(\phi)$ is always an invariant state of $\Phi$. For simplicity, we consider a semifinite
${\mathcal M}$ equipped with a normal faithful semi-finite trace $\tau$, but the discussion here applies to general von Neumann algebras with proper interpretation of notation. In the tracial setting, we can write $\phi(x)=\tau(d_\phi x)$ using the density operator $d_\phi$ of $\phi$.
Recall that the modular group $(\alpha_t^\phi)_{t\in {\mathbb R}}:{\mathcal M}\to {\mathcal M}$ is the one-parameter group of $*$-automorphism defined as
\[ \alpha_t^{\phi}(x) \pl = \pl d_{\phi}^{-it}xd_{\phi}^{it}\hspace{.1cm}, \hspace{.1cm} x\in {\mathcal M} \hspace{.1cm}, t\in \mathbb{R}\]
It is proven in \cite{wirth2022christensen} that every $\phi$-symmetric quantum Markov map commutes with the modular group
\[\alpha_t^{\phi}\circ\Phi\pl = \pl \Phi\circ\alpha_t^{\phi}\hspace{.1cm}, \hspace{.1cm} t\in {\mathbb R}\hspace{.1cm}.\]
Furthermore, $\Phi$ is also symmetric to the KMS inner product
\begin{align}\langle \Phi(x),y\rangle_{\phi} \pl = \pl \langle x,\Phi(y)\rangle_\phi, \hspace{.1cm} x,y \in {\mathcal M} \label{eq:KMS}\end{align}
where
\[\langle x,y\rangle_{\phi}:\pl = \pl \tau(d_\phi^{1/2}x^*d_\phi^{1/2} y)\hspace{.1cm}.\]
Let $\Phi_*:L_1({\mathcal M})\to L_1({\mathcal M})$ be the pre-adjoint quantum channel via trace duality. The KMS-symmetry implies that
\begin{equation}\label{eq:pre}
\Phi_*(d_\phi^{1/2}xd_\phi^{1/2})
\pl = \pl d_\phi^{1/2}\Phi(x)d_\phi^{1/2}\hspace{.1cm}, \hspace{.1cm} \forall x\in {\mathcal M}
\end{equation}
Let us introduce the contraction map $\Gamma_\phi: {\mathcal M}\to L_1({\mathcal M})\hspace{.1cm}, \hspace{.1cm} \Gamma_\phi(x)=d_\phi^{1/2}xd_\phi^{1/2}$. Then the above equation can be expressed as $\Phi_*\circ \Gamma_\phi=\Gamma_\phi\circ \Phi$.
For $1\le p\le \infty$, the weighted $L_p$-space $L_p({\mathcal M},\phi)$ is the completion of ${\mathcal M}$ under the norm
\[\norm{x}{p,\phi}=\norm{d_\phi^{1/2p} x d_\phi^{1/2p}}{p} = \norm{\Gamma_\phi^{1/p}(x)}{p}\hspace{.1cm},\]
where $\norm{y}{p}=\tau(|y|^p)^{1/p}$ is the tracial $p$-norm.
For $p=2$, $L_p({\mathcal M},\phi)$ is a Hilbert space with KMS-inner product $\norm{x}{2,\phi}^2=\langle x,x\rangle_\phi$. By equation \eqref{eq:pre}, $\Phi$ is also a contraction on $L_1({\mathcal M},\phi)$, and hence a contraction on $L_p({\mathcal M},\phi)$ for all $1\le p\le \infty$ by complex interpolation.
The lemma below is an analog of Proposition \ref{prop:cd} in the trace-symmetric case, which shows $\Phi$ is a ${\mathcal N}$-bimodule map for its multiplicative domain ${\mathcal N}$.
\begin{prop}\label{prop:cd2}Let $\Phi:{\mathcal M}\to{\mathcal M}$ be a $\phi$-symmetric quantum Markov semigroup for a normal faithful state $\phi$. Denote ${\mathcal N}$
as the multiplicative domain of $\Phi$. Then
\begin{enumerate}
\item[i)] ${\mathcal N}$ is invariant under $\alpha_t^\phi$. Hence there exists a $\phi$-symmetric normal condition expectation $E:{\mathcal M}\to {\mathcal N}$.
\item[ii)] $\Phi|_{\mathcal N}$ is an involutive $*$-automorphism satisfying
\begin{align}\label{eq:cd} \Phi^2\circ E=E\circ \Phi^2=E\hspace{.1cm}, \hspace{.1cm} E\circ \Phi=\Phi\circ E\hspace{.1cm} . \end{align}
Moreover, $\Phi^2$ is a ${\mathcal N}$-bimodule map satisfying $\Phi^2(axb)=a\Phi^2(x)b$.
\item[iii)] $\Phi$ is an isometry on $L_2({\mathcal N},\phi)$. If in addition, \[{\norm{\Phi(\iota_{\infty,2}^n-E):L_2({\mathcal M},\phi)\to L_2({\mathcal M},\phi)}{2}<1}\hspace{.1cm},\] then $E=\lim_{n}\Phi^{2n}$ as a map from $L_2({\mathcal M},\phi)$ to $L_2({\mathcal M},\phi)$ .
\end{enumerate}
\end{prop}
\begin{proof} It suffices to explain i). The rest follows similar as Proposition \ref{prop:cd} (see \cite[Lemma 2.5]{gao2022complete} in finite dimensional cases). Indeed, since $\Phi$ commutes with $\alpha_t^\phi$, for $x\in {\mathcal N}$
\begin{align*}\Phi\big(\alpha_t^\phi(x)y\big)=&\Phi\big(\alpha_t^\phi(x \alpha_{-t}^\phi(y) )\big)=\alpha_t^\phi\circ\Phi\big(x \alpha_{-t}^\phi(y) \big)\\ =&\alpha_t^\phi\Big(\Phi(x)\Phi(\alpha_{-t}^\phi(y)) \Big)=\alpha_t^\phi\circ\Phi(x) \alpha_t^\phi\circ \Phi\circ \alpha_{-t}^\phi(y)=\Phi(\alpha_t^\phi(x))\Phi(y)\hspace{.1cm}.
\end{align*}
The other side multiplicativity is similar and this implies $ \alpha_t^\phi(x)\in{\mathcal N}$. By Takesaki's theorem \cite{takesaki1972conditional},
there exists $\phi$-preserving condition expectation satisfying the defining property
\[\phi(xy)=\phi(xE(y)) \hspace{.1cm} \forall \hspace{.1cm} x\in {\mathcal N}, y\in {\mathcal M}\hspace{.1cm},\]
from which the $\phi$-symmetricness follows.
\end{proof}
\subsection{Haagerup's reduction}
A von Neumann algebra ${\mathcal M}$ is called type {III} if it does not admits a nontrivial semifinite trace.
We briefly review the basics of Haagerup's construction and refer to \cite{haagerup2010reduction} for more details. The key idea is to consider the additive subgroup $G=\bigcup_{n\in {\mathbb N}} 2^{-n}\mathbb{Z} \subset \mathbb{R}$.
Let ${\mathcal M}\subset B(H)$ be a von Neumann algebra and $\phi$ be a normal faithful state.
One can defined the crossed product by the action $\sigma^\phi:G\curvearrowright {\mathcal M}$
\[ \hat{{\mathcal M}}\pl = \pl {\mathcal M}\rtimes_{\sigma^{\phi}}G \hspace{.1cm} , \]
$\hat{{\mathcal M}}$ can be considered as the von Neumann subalgebra $ \hat{{\mathcal M}}=\{\pi({\mathcal M}),\lambda(G)\}'' \subset {\mathcal M}\overline\otimes B(\ell_2(G))$ generated by the embeddings
\begin{align}\label{eq:cross}
\pi:{\mathcal M}\to {\mathcal M}\rtimes_{\sigma^{\phi}}G\hspace{.1cm}, \hspace{.1cm} &\pi(a)\pl = \pl \sum_{g} \alpha_{g^{-1}}(a) \otimes \ketbra{g}\nonumber\\
\lambda:G\to {\mathcal M}\rtimes_{\sigma^{\phi}}G,\hspace{.1cm} &\lambda(g)(\ket{x}\otimes \ket{h})=\ket{x}\otimes \ket{gh}, \text{ and } \hspace{.1cm} \forall \hspace{.1cm} \ket{x}\in H\hspace{.1cm}, \ket{h}\in \ell_2(G) \hspace{.1cm}.
\end{align}
Basically, $\pi$ is the transference homomorphism ${\mathcal M}\to \ell_\infty(G,{\mathcal M})$, and $\lambda$ is the left regular representation on $\ell_2(G)$.
The set of finite sums
$\{\sum_g a_g \lambda(g)\hspace{.1cm} | \hspace{.1cm} a_g \in {\mathcal M} \}\subset \hat{{\mathcal M}}$ forms a dense $w^*$-subalgebra of $\hat{{\mathcal M}}$.
In the following, we identify ${\mathcal M}$ with $\pi({\mathcal M})$ (resp. $a$ with $\pi(a)$) and view ${\mathcal M}\subset \hat{{\mathcal M}}$ as a subalgebra.
The state $\phi$ admits a natural extension as a normal faithful state on $\hat{{\mathcal M}}$
\[ \hat{\phi}(\sum_g a_g \lambda(g)) \pl = \pl \phi(a_0) \hspace{.1cm},\]
Moreover, \[E_{\mathcal M}:\hat{{\mathcal M}}\to {\mathcal M}\hspace{.1cm}, \hspace{.1cm} (\sum_g a_g\lambda(g))=a_0\] is the canonical normal conditional expectation such that $\phi\circ E_{\mathcal M}\pl = \pl \hat{\phi}$.
The main object in Haagerup's construction is an increasing family of subalgebras
\[ {\mathcal M}_n \pl = \pl \hat{{\mathcal M}}_{\psi_n}:=\{x\in \hat{{\mathcal M}} \hspace{.1cm} |\hspace{.1cm} \alpha_t^{\psi_n}(x)=x \hspace{.1cm}, \hspace{.1cm}\forall\hspace{.1cm} t\in {\mathbb R}\}\hspace{.1cm},\]
given by the centralizer algebra $\hat{{\mathcal M}}_{\psi_n}$ for a suitable family of states $\psi_n$ so that $\bigcup_n {\mathcal M}_n$ is $w^*$-dense in $\hat{{\mathcal M}}$. The state $\psi_n$ is defined via a Radon-Nikodym density w.r.t to $\hat{\phi}$
\[ \psi_n(x) \pl = \pl \hat{\phi}(e^{-a_n}x) \hspace{.1cm} ,\hspace{.1cm}
a_n \pl = \pl -i2^n\text{Log} (\lambda(2^{-n})) \hspace{.1cm} .\]
Here $\text{Log}$ is the principal branch of the logarithmic function with $0\le \text{Log} z<2\pi$.
Each subalgebra ${\mathcal M}_n$ contains $\lambda(G)$ and there exists normal conditional expectation $E_{{\mathcal M}_n}:\hat{{\mathcal M}}\to {\mathcal M}_n$. Indeed, by the definition of $\psi_n$, the modular group $\alpha_t^{\psi_n}$ is $2^{-n}$ periodic. The explicit form (see \cite[Lemma 2.3]{haagerup2010reduction}) is given by
\[E_{{\mathcal M}_n}=2^{n}\int_0^{2^{-n}}\alpha_t^{\psi_n} dt\hspace{.1cm}.\]
The normalized state $\tau_n=\frac{\psi_n}{\psi_n(1)}$ is a normalized trace on ${\mathcal M}_n$.
The key properties of ${\mathcal M}_n$ are summarized in \cite[Theorem 2.1 \& Lemma 2.7]{haagerup2010reduction}
\begin{theorem}\label{thm:hr}
With above notations, ${\mathcal M}_n$ is an increasing family of von Neumann subalgebras satisfying the following properties
\begin{enumerate}
\item Each $({\mathcal M}_n,\tau_n)$ is a finite von Neumann algebra.
\item $\bigcup_{n\ge 1} {\mathcal M}_n$ is weak$^*$-dense in $\hat{{\mathcal M}}$.
\item There exists $\hat{\phi}$-preserving normal faithful conditional expectation $E_{{\mathcal M}_n}:\hat{{\mathcal M}}\to {\mathcal M}_n$ such that
\[ \hat{\phi}\circ E_{{\mathcal M}_n}=\hat{\phi}\hspace{.1cm}, \hspace{.1cm} \alpha_t^{\hat{\phi}}\circ E_{{\mathcal M}_n}= E_{{\mathcal M}_n}\circ \alpha_t^{\hat{\phi}}\hspace{.1cm}. \]
Moreover, $E_{{\mathcal M}_n}(x)\to x$ in $\sigma$-strong topology for any $x\in \hat{{\mathcal M}}$.
\end{enumerate}
\end{theorem}
We now look at the Haagerup reduction on the states. For a state $\rho\in S({\mathcal M})$, $\hat{\rho}=\rho\circ E_{\mathcal M}$ is the canonical extension on $\hat{{\mathcal M}}$. We denote $\rho_n:=\hat{\rho}|_{{\mathcal M}_n}\in {\mathcal M}_{n,*}$ as the restriction state of $\hat{\rho}$ on the subalgebra ${\mathcal M}_n \subset\hat{{\mathcal M}}$. Note that the predual ${\mathcal M}_{n,*}$ can be viewed as a subspace of $\hat{{\mathcal M}}_*$ via the embedding
\[ \iota_{n,*}: \hat{{\mathcal M}}_{n, *}\to \hat{{\mathcal M}}_*\hspace{.1cm}, \iota_{n,*}(\omega)=\omega\circ E_{{\mathcal M}_n}\hspace{.1cm}. \]
Via this identification, $\rho_n=\hat{\rho}|_{{\mathcal M}_n}\circ E_{{\mathcal M}_n}=\hat{\rho}\circ E_{{\mathcal M}_n}=E_{{\mathcal M}_n,*}(\hat{\rho})\in \hat{{\mathcal M}}_*$. Moreover, by the weak$^*$-density of the family ${\mathcal M}_n$, $\rho_n\to \hat{\rho}$ converges in the weak topology. One immediate consequence is the following approximation of relative entropy.
\begin{lemma}
\label{lemma:ultra}
Let $\rho$ and $\sigma$ be two normal states of ${\mathcal M}$. Then
\[ D(\rho||\sigma) \pl = \pl D(\hat{\rho}||\hat{\sigma})=\lim_{n\to \infty} D(\rho_n||\sigma_n) \hspace{.1cm} .\]
\end{lemma}
\begin{proof} Let $\iota:{\mathcal M}\subset \hat{{\mathcal M}}$ be the inclusion map. Because $\hat{\rho}=\rho\circ E_{\mathcal M}$ is an extension of $\rho$ , $\iota_*(\hat{\rho})=\hat{\rho}|_{\mathcal M}=\rho$ and similarly for $\sigma$. Both $\iota: {\mathcal M} \to \hat{{\mathcal M}}$ and $E_{\mathcal M}:\hat{{\mathcal M}}\to {\mathcal M}$ are quantum Markov maps (normal unital completely positive maps, dual of quantum channels). Then by the data processing inequality,
\[D(\rho||\sigma)=D(\iota_*(\hat{\rho})||\iota_*(\hat{\sigma}))\le D(\hat{\rho}||\hat{\sigma})=D(E_{{\mathcal M},*}(\rho)||E_{{\mathcal M},*}(\sigma))\le D(\rho||\sigma) \]
Thus $D(\rho||\sigma)=D(\hat{\rho}||\hat{\sigma})$. The limit follows from the lower semi-continuity of $D$ under weak-convergence (see e.g. \cite[Theorem 2.7]{hiai2021quantum}) and another use of the data processing inequality:
\begin{align*}
&D(\hat{\rho}||\hat{\sigma})\le\liminf_{n}D(\rho_n||\sigma_n)=\liminf_{n}D(E_{{\mathcal M}_n,*}(\rho_n)||E_{{\mathcal M}_n,*}(\sigma_{n,*}))\le D(\hat{\rho}||\hat{\sigma}) \hspace{.1cm}. \qedhere
\end{align*}
\end{proof}\vspace{0.5ex}
We shall also apply the Haagerup's reduction on GNS-symmetric maps. Let $\Phi:{\mathcal M}\to {\mathcal M}$ be a $\phi$-symmetric quantum Markov map. Its canonical extension map
\[\hat{\Phi}:\hat{{\mathcal M}}\to \hat{{\mathcal M}}\hspace{.1cm}, \hspace{.1cm} \hat{\Phi}(\sum_g a_g\lambda(g))
\pl = \pl \sum_g \Phi(a_g)\lambda(g) \hspace{.1cm} \]
is also a $\hat{\phi}$-symmetric quantum Markov map. Indeed, $\hat{\Phi}=\Phi\otimes \iota_{\infty,2}^n_{B(\ell_2(G))}|_{\hat{{\mathcal M}}}$ is the restriction of $\Phi\otimes \iota_{\infty,2}^n_{B(\ell_2(G))}$ on $\hat{{\mathcal M}}\subset {\mathcal M}\overline{\otimes} B(\ell_2(G)$. It is clear that $\hat{\Phi}$ has the multiplicative domain \[\hat{{\mathcal N}}:={\mathcal N}\rtimes_{\sigma^{\phi}} G\hspace{.1cm}.\]
where ${\mathcal N}$ be the multiplicative domain of $\Phi$. In particular, this crossed product is well-defined because $\alpha^\phi_t({\mathcal N})={\mathcal N}$. Moreover, the $\hat{\phi}$-conditional expectation $\hat{E}:\hat{{\mathcal M}}\to\hat{{\mathcal M}}$ is nothing but the canonical extension of $E:{\mathcal M}\to {\mathcal N}$.
Recall that we write $E_{\mathcal M}$ and $E_n$ as the normal conditional expectation from $\hat{{\mathcal M}}$ onto ${\mathcal M}$ and ${\mathcal M}_n$ respectively. The next lemma shows that the extension $\hat{\Phi}$ is well compatible with the approximation family ${\mathcal M}_n$.
\begin{lemma}\label{lemma:commute}
With the notations above,
\begin{enumerate}
\item[i)] $\hat{\Phi}$ commutes with $E_{\mathcal M}, \hat{E}$ and $E_{{\mathcal M}_n}$. In particular, $\hat{\Phi}({\mathcal M}_n)\subset {\mathcal M}_n$.
\item[ii)] The restriction $\Phi_n=\hat{\Phi}|_{{\mathcal M}_n}$ is a normal UCP map symmetric with respect to the tracial state $\tau_n$.
\item[iii)] Let ${\mathcal N}_n\subset {\mathcal M}_n$ be the multiplicative domain for $\Phi_n$. Then the restriction map $E_n:=\hat{E}|_{{\mathcal M}_n}:{\mathcal M}_n\to {\mathcal N}_n$ is the $\tau_n$-preserving conditional expectation.
\end{enumerate}
\end{lemma}
\begin{proof} The relation $\hat{\Phi}\circ E_{\mathcal M}=E_{\mathcal M}\circ \hat{\Phi}$ is clear from the definition of $\hat{\Phi}$, and $\hat{\Phi}\circ \hat{E}=\hat{E}\circ \hat{\Phi}$ follows from Lemma \ref{prop:cd2}.
Recall that $\psi_n(x)=\hat{\phi}(e^{-a_n}x)$ with density operator $e^{-a_n}\in \lambda(G)''$ and $\lambda(G)$ is in the centralizer of $\hat{\phi}$ \cite[Lemma 2.3]{haagerup2010reduction}. Then
\[ \alpha_t^{\psi_n} \pl = \pl u(t)^*\alpha_t^{\hat{\phi}}u(t)=\text{ad}_{u(t)}\alpha_t^{\hat{\phi}} \]
for the unitary $u(t)=e^{-ita_n}$. Note that $\hat{\Phi}$ commute with $\alpha_t^{\hat{\phi}}$ by $\hat{\phi}$-symmetry, and also commutes with $\text{ad}_{u(t)}$ because $u(t)\in \lambda(G)''$ is in $\hat{\Phi}$'s multiplicative domain. Thus $\hat{\Phi}$ commutes with $\alpha_t^{\psi_n}$ and hence the conditional expectation $E_{{\mathcal M}_n}=2^{-n}\int_0^{2^{-n}}\alpha_t^{\psi_n}$. This proves i).
For ii), we note that for $x,y\in {\mathcal M}_n$
\[\psi_n(x\Phi_n(y))=\hat{\phi}(e^{-a_n}x\hat{\Phi}(y))=\hat{\phi}(\hat{\Phi}(e^{-a_n}x)y)=\hat{\phi}(e^{-a_n}\hat{\Phi}(x)y)=\psi_n(\Phi_n(x)y)\hspace{.1cm},\]
where we used the fact that $\hat{\Phi}$ is $\hat{\phi}$-symmetric and $e^{-a_n}\in \lambda(G)''$ is in the fixed point subspace of $\hat{\Phi}$. Finally, iii) follows from applying i) and ii) to $\hat{E}$. \end{proof}\vspace{0.5ex}
\begin{rem}{\rm For the restriction state $\phi_n=\hat{\phi}_{{\mathcal M}_n}$ with density $d_{\phi_n}$ w.r.t to $\tau_n$, the above implies the KMS symmetry for each $n$
\begin{align*}
\langle\hat{\Phi}(x),y\rangle_{\phi_n}= \tau_n(d_{\phi_n}^{1/2}\hat{\Phi}(x^*) d_{\phi_n}^{1/2}y)
= \tau_n(d_{\phi_n}^{1/2}x^*d_{\phi_n}^{1/2} \hat{\Phi}(y))=\langle \hat{\Phi}(x),y\rangle_{\phi_n} \hspace{.1cm} .
\end{align*}
Hence the density $d_{\phi_n}^{1/2}$ is in the multiplicative domain of $\hat{\Phi}$.}
\end{rem}
To summarize the lemma above, we have the following commuting diagrams\\
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em]
{
{\color{white}\hat{l}}{\mathcal M} & \hat{{\mathcal M}} & {\mathcal M}_n{\color{white}\hat{l}} \\
{\color{white}\hat{l}}{\mathcal M} & \hat{{\mathcal M}} & {\mathcal M}_n{\color{white}\hat{l}} \\};
\path[-stealth]
(m-1-1) edge node [left] {$\Phi$} (m-2-1)
(m-1-2) edge node [above] {$E_{\mathcal M}$} (m-1-1)
edge node [right] {$\hat{\Phi}$} (m-2-2)
edge node [above] {$E_{{\mathcal M}_n}$} (m-1-3)
(m-1-3) edge node [right] {$\Phi_n$} (m-2-3)
(m-2-2) edge node [below] {$E_{\mathcal M}$} (m-2-1)
edge node [below] {$E_{{\mathcal M}_n}$} (m-2-3);
\end{tikzpicture}\hspace{1cm}\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em]
{
{\color{white}\hat{l}}{\mathcal M} & \hat{{\mathcal M}} & {\mathcal M}_n{\color{white}\hat{l}} \\
{\color{white}\hat{l}}{\mathcal N} & \hat{{\mathcal N}} & {\mathcal N}_n{\color{white}\hat{l}} \\};
\path[-stealth]
(m-1-1) edge node [left] {$E$} (m-2-1)
(m-1-2) edge node [above] {$E_{\mathcal M}$} (m-1-1)
edge node [right] {$\hat{E}$} (m-2-2)
edge node [above] {$E_{{\mathcal M}_n}$} (m-1-3)
(m-1-3) edge node [right] {$E_{n}$} (m-2-3)
(m-2-2) edge node [below] {$E_{\mathcal N}$} (m-2-1)
edge node [below] {$E_{{\mathcal N}_n}$} (m-2-3);
\end{tikzpicture} \\
Basically, $\Phi_n$ is a family of trace symmetric channel which approximates $\hat{\Phi}$ as a natural extension of $\Phi$. The same picture holds for the conditional expectation $E,\hat{E}$ and $E_n$
\subsection{Entropy Decay via CB return time}\label{sec:GNSentropydecay}
We are shall now discuss the entropy decay result for GNS-symmetric channels. The first step is to extend the entropy difference Lemma \ref{lemma:difference}. Define the state space that are bounded w.r.t $\phi$,
\[S_b({\mathcal M},\phi)=\{\rho \in S({\mathcal M})\hspace{.1cm} |\hspace{.1cm} c^{-1}\phi\le \rho\le c\phi\hspace{.1cm}, \hspace{.1cm} \text{for some } c>0 \}\hspace{.1cm}.\]
For all $\rho\in S_b({\mathcal M},\phi)$, $D(\rho||\phi) <\infty$ is finite. Such $S_b({\mathcal M},\phi)$ is a dense subset of $S({\mathcal M})$ because for any $\rho$ and $0<\varepsilon<1$,
$\rho_\varepsilon=(1-\varepsilon)\rho+\varepsilon \phi\in S_b({\mathcal M})$.
For $\rho\in S_b({\mathcal M})$, we define the entropy difference for a $\phi$-quantum channel $\Phi_*$ as
\[D_{\Phi_*}(\rho):=D(\rho||E_*(\rho))-D(\Phi_*(\rho)||\Phi_*\circ E_*(\rho))\hspace{.1cm}.\]
Recall the chain rule \cite[Theorem 2]{petz1991certain}
\[ D(\rho||\psi)=D(\rho||E_*(\rho))+D(\rho||E_*(\rho))\]
Thus for $\rho\in S_b({\mathcal M})$,
\[ D(\Phi_*(\rho)||\Phi_*\circ E_*(\rho))\le D(\rho||E_*(\rho))\le D(\rho||\psi)<\infty\hspace{.1cm}.\]
Then $D_{\Phi_*}(\rho)$ is well-defined and finite. Also by the chain rule, we have the alternative expression $D_{\Phi_*}(\rho)=D(\rho||\psi)-D(\Phi_*(\rho)||\psi)$. In the trace symmetric case, $D_{\Phi_*}(\rho)=D(\rho||1)-D(\Phi_*(\rho)||1)=H(\rho)-H(\Phi_*(\rho))$ as in Section \ref{sec:unital}.
\begin{lemma}\label{lemma:difference2}Let $\Phi_*$ be a $\phi$-symmetric quantum channel.
For any state $\rho,\omega\in S_b({\mathcal M},\phi)$,
\[ D(\rho||\Phi_*^2(\omega))\pl \le \pl D_{\Phi_*}(\rho)+D(\rho||\omega) \hspace{.1cm} .\]
\end{lemma}
\begin{proof}
Recall that we use $\rho_n=\hat{\rho}|_{{\mathcal M}_n}=E_{{\mathcal M}_n,*}(\hat{\rho})$ and $\omega_n=\hat{\omega}|_{{\mathcal M}_n}=E_{{\mathcal M}_n,*}(\hat{\omega})$,
as the states in the finite von Neumann algebra ${\mathcal M}_n\subset \hat{{\mathcal M}}$ obtained from the Haagerup reduction. By Lemma \ref{lemma:commute}, we know that $\Phi_n=\hat{\Phi}|_{{\mathcal M}_n}$ is a quantum Markov map symmetric with respect to the tracial state $\tau_n$. Thus by Lemma \ref{lemma:difference} in the tracial case,
\[ D(\rho_n||\Phi_{n}^2(\omega_n))
\pl \le \pl D_{\Phi_n}(\rho_n) +D(\rho_n||\omega_n) \hspace{.1cm} ,\]
where we identify $\Phi_{n}=\Phi_{n,*}$ by trace symmetry. Here,
\[D_{\Phi_n}(\rho_n)=D(\rho_n|| \hat{E}_{n}(\rho_n))-D(\Phi_n(\rho_n)|| \Phi_n\hat{E}_{n}(\rho_n)),\]
where $\hat{E}_{n}:{\mathcal M}_n\to {\mathcal N}_n$ is the $\tau_n$-preserving conditional expectation.
By the definition of $\Phi_n$ and $\rho_n$, recalling the hat "$\hat{}$" notation for states on $\hat{{\mathcal M}}$,
\begin{align*} &\Phi_{n}(\rho_n)=\hat{\rho}|_{{\mathcal M}_n}\circ \Phi_{n} =\hat{\rho}\circ \hat{\Phi}|_{{\mathcal M}_n}= \widehat{\Phi_*(\rho)}|_{{\mathcal M}_n}= \Phi_*(\rho)_n\hspace{.1cm} ,\\
&\Phi_{n}^2(\rho_n)=\Phi_{n}(\Phi_*(\rho)_n)=\Phi_*^2(\rho)_n \hspace{.1cm} .
\end{align*}
Apply this to $E_n= \hat{E}|_{{\mathcal M}_n}$, we also have
\begin{align*}
E_{n}(\rho_n)=E_*(\rho)_n\hspace{.1cm}, \hspace{.1cm}
\Phi_n\hat{E}_{n}(\rho_n)=\Phi_{n} (E_*(\rho)_n)=\Phi_*E_*(\rho)_n\hspace{.1cm}.
\end{align*}
Then by Lemma \ref{lemma:ultra}, we can approximate every entropic term
\begin{align*}\lim_{n}D(\rho_n||(\Phi_{n})^2(\omega_n))=&\lim_{n}D(\rho_n||\Phi_*^2(\omega)_n)=D(\rho||\Phi_*^2(\omega))\\
\lim_{n} D_{\Phi_n}(\rho_n)=&\lim_{n}D(\rho_n|| E_{n}(\rho_n))-D(\Phi_n(\rho_n)|| \Phi_nE_{n}(\rho_n))\\=&\lim_{n}D(\rho_n||E_*(\rho)_n))-D(\Phi_*(\rho)_n|| \Phi_*E_*(\rho)_n)=\lim_{n} D_{\Phi}(\rho)\\
\lim_{n}D(\rho_n||\omega_n)=&D(\rho||\omega) \qedhere
\end{align*}
\end{proof}\vspace{0.5ex}
The next lemma shows the CB-return time is also compatible with Haargerup reduction.
\begin{lemma}\label{lemma:transfer} Let $\Psi:{\mathcal M}\to {\mathcal M}$ be a $\phi$-symmetric quantum Markov map and $E$ be the conditional expectation on its multiplicative domain.
\begin{equation}\label{rett}
(1-\varepsilon) E \le_{cp} \Psi \le_{cp} (1+\varepsilon)E \hspace{.1cm} .
\end{equation}
Then for all $n\in \bN$,
\[ (1-\varepsilon)E_{n} \le_{cp} \Psi_n \le_{cp} (1+\varepsilon)E_{n} \hspace{.1cm} . \]
In particular, if $0.9 E \le_{cp} \Psi \le_{cp} 1.1E$ and $\Psi\circ E=E$, then for any $\rho\in S_b({\mathcal M},\phi)$
\[ \frac{1}{2}D(\rho||E_*(\rho))\le D(\rho||\Psi_*(\rho))\hspace{.1cm}.\]
\end{lemma}
\begin{proof} The CP order inequality follows from the fact that
both maps $E_n$ and $\Psi_n$ are the restriction of $E\otimes \iota_{\infty,2}^n$ and $\Psi\otimes \iota_{\infty,2}^n$ on the subalgebra ${\mathcal M}_n\subset \hat{{\mathcal M}}\subset {\mathcal M}\otimes B(\ell_2(G))$. Then the entropy inequality can be obtained by the tracial case Lemma \ref{lemma:approximate} and the approximation as in Lemma \ref{lemma:difference2}.
\end{proof}
Recall that we define
\[\alpha(\Phi)=\sup_{\rho} \frac{D(\Phi_*(\rho)||\Phi_*\circ E_*(\rho))}{D(\rho||E_*(\rho))}\hspace{.1cm},\]
as the entropy contraction coefficient and $\alpha_c(\Phi):=\sup_{{\mathcal Q}}\alpha(\iota_{\infty,2}^n_{\mathcal Q}\otimes\Phi)$ as the complete bounded coefficient. We obtain the identical estimate as in the tracical case.
\begin{theorem}\label{thm:GNSsymmetric}
Let $\Phi:{\mathcal M}\to {\mathcal M}$ be a $\phi$-symmetric quantum Markov map and $E$ be the $\phi$-preserving conditional expectation onto its multiplicative domain ${\mathcal N}$. Denote
\[ k_{cb}(\Phi) \pl = \pl \inf\{k\in {\mathbb N}^+ \hspace{.1cm} |\hspace{.1cm} 0.9 E\le_{cp}\Phi^{2k} \le_{cp} 1.1 E\hspace{.1cm} \} \]
Then
\[ \alpha(\Phi)\le \alpha_c(\Phi)\le \Big ( 1-\frac{1}{2k_{cb}(\Phi)} \Big ).\]
As a consequence, for any finite von Neumann algebra ${\mathcal Q}$, state $\rho\in S({\mathcal M}\overline{\otimes} {\mathcal Q})$ and $n\ge 1$,
\[ D(\Phi_*^n\otimes \iota_{\infty,2}^n(\rho)||\Phi_*^nE_*\otimes \iota_{\infty,2}^n(\rho)) \pl \le \pl \Big (1-\frac{1}{2k_{cb}(\Phi)}\Big )^nD(\rho||E_*\otimes \iota_{\infty,2}^n(\rho) \hspace{.1cm} .\]
\end{theorem}
\begin{proof}
For $\rho\in S_b({\mathcal M},\phi)$, the proof is same as the tracial case Theorem \ref{thm:unital} by using Lemma \ref{lemma:difference2} and Lemma \ref{lemma:transfer} above. The general case $\rho\in S({\mathcal M})$ can be approximated by $\rho_\varepsilon= (1-\varepsilon)\rho+ \varepsilon\phi$.
\end{proof}
\begin{rem}{\rm
The above theorem gives substantial improvement to \cite[Theorem 4.1]{gao2022complete} which is almost an existence proof of $\alpha_c(\Phi)<1$ given the spectral gap
\[\lambda(\Phi):=\norm{\Phi(\iota_{\infty,2}^n-E):L_2({\mathcal M},\phi)\to L_2({\mathcal M},\phi)}{}<1\]
and finite index $C_{cb}(E)<\infty$. Our results yield explicit bound depending on $\lambda$ and $C_{cb}(E)$ and also applies to infinite dimensional setting even with $C_{cb}(E)=\infty$.}\end{rem}
For a $\phi$-symmetric quantum Markov semigroup $T_t=e^{-tL}:{\mathcal M}\to {\mathcal M}$, we define the modified log-Sobolev (MLSI) constant $\alpha_1(L)$ as the largest $\alpha$ such that
\begin{align}D(T_{t,*}(\rho)||E_*(\rho)) \pl \le \pl e^{-2\alpha t}D(\rho||E_*\otimes \iota_{\infty,2}^n(\rho))\hspace{.1cm}, \hspace{.1cm} \forall \rho\in S({\mathcal M})\hspace{.1cm},\label{eq:expdecay}\end{align}
where $E$ is the $\phi$-preserving conditional expectation onto the fixed point subalgebra ${\mathcal N}$
The complete MLSI constant is then defined as $\alpha_c(L):=\sup_{\mathcal{{\mathcal Q}}}\alpha(\iota_{\infty,2}^n_{\mathcal Q}\otimes L)$.
\begin{theorem}\label{thm:GNSsemigroup} Denote
\[ t_{cb}(L) \pl = \pl \inf\{t>0 \hspace{.1cm}|\hspace{.1cm} 0.9 E\le_{cp}T_t \le_{cp} 1.1 E\} \]
Then
\[ \alpha_1(L)\le \alpha_c(L)\le \frac{1}{2t_{cb}(L)}.\]
Namely, for any finite von Neumann algebra ${\mathcal Q}$ and state $\rho\in S({\mathcal M}\overline{\otimes} {\mathcal Q})$, we have the exponential decay of relative entropy
\[ D(T_{t,*}\otimes \iota_{\infty,2}^n(\rho)||E_*\otimes \iota_{\infty,2}^n(\rho)) \pl \le \pl e^{-\frac{t}{t_{cb}(L)}}D(\rho||E_*\otimes \iota_{\infty,2}^n(\rho)) \hspace{.1cm},\hspace{.1cm} t\ge 0\hspace{.1cm}.\]
\end{theorem}
\begin{proof}This can be approximated using the tracial case Theorem \ref{thm:unital} as Lemma \ref{lemma:difference2} above.
\end{proof}
\begin{rem}{\rm In finite dimensions, the MLSI constant of a GNS-symmetric semigroup is defined as
\[ 2\alpha D(\rho||E_*(\rho))\le I_L(\rho):=\tau(L_*(\rho)(\ln \rho-\ln\phi) )\hspace{.1cm},\]
The right hand side $I_L(\rho)$ is the entropy production, and the equivalence to entropy decay rely on the de Bruijn identity
\begin{align}\label{eq:debruijn} I_L(\rho)=-\frac{d}{dt}D(T_*(\rho)||E_*(\rho))|_{t=0}\hspace{.1cm}, \end{align}
In infinite dimensions, the de Bruijn identity \eqref{eq:debruijn} is less justified even in $B(H)$ with $\dim(H)=+\infty$ (see discussions in \cite{huber2017geometric, konig2014entropy}). Here, our definition of MLSI \eqref{eq:expdecay} via exponential entropy decay avoids this issue, as the relative entropy is always well-defined even on Type III algebra via Araki's definition \eqref{eq:araki}. We note that this definition of MLSI also does not depend on any choice of reference state $\phi$.
}
\end{rem}
The next proposition shows that the GNS-symmetry is also independent of the choice of invariant state $\phi$.
\begin{lemma}\label{lemma:trace} Let $T_t:{\mathcal M}\to{\mathcal M}$ be a $\phi$-symmetric quantum Markov semigroup for a normal faithful state $\phi$. Denote $E:{\mathcal M}\to{\mathcal N}$ as the $\phi$-preserving conditional expectation onto the fixed point space. Suppose $\psi$ is an another normal faithful state invariant under $E$, i.e. $\psi\circ E=\psi$. Then $T_t:{\mathcal M}\to{\mathcal M}$ is also $\psi$-symmetric.
\end{lemma}
\begin{proof}
Without loss of generosity we assume $\psi\le C\phi$ for some $C>0$. We first view them as the state on the subalgebra ${\mathcal N}$ by restriction. By \cite[Theorem 3.17]{takesaki2}, there exists $h\in {\mathcal N}$ such that
\[ \psi( x )=\phi(h^*xh ) \hspace{.1cm}, \hspace{.1cm} \forall x\in {\mathcal N}\hspace{.1cm}. \]
This identity actually also holds for $y\in {\mathcal M}$. Indeed, because of $\phi\circ E=\phi$ and $\psi\circ E=\psi$,
\[ \psi( y )=\psi( E(y) )=\phi(h^*E(y)h )=\phi(E(h^*yh) )=\phi(h^*yh ) \hspace{.1cm}, \hspace{.1cm} \forall y\in {\mathcal M}\hspace{.1cm}. \]
Thus the GNS-symmetry with respect to $\psi$ follows that for $x,y\in{\mathcal M}$,
\[ \psi( xT_t(y) )=\phi(h^*xT_t(y)h )=\phi(h^*xT_t(yh)) =\phi(T_t(h^*x)yh)=\phi(h^*T_t(x)yh)=\psi( T_t(x)y ) \hspace{.1cm},\]
where we used the bimodule property of $T_t(axb)=aT_t(x)b$ for $a,b\in {\mathcal N}$. The general case can be obtained via $\psi_{\varepsilon}=(1-\varepsilon)\psi+\varepsilon \phi$.
\end{proof}\vspace{0.5ex}
We remark that under spectral gap or MLSI, one have $\lim_{t}T_t=E$ in $L_2$ or $L_1$ norm. Then the above $E$-invariant condition $\phi\circ E=\phi$ can be replaced by $\phi=T_{t}\circ \phi$ for any/some $t>0$.
\begin{rem}{\rm In the above Haagerup's reduction, both $\hat{\Phi}$ and $\Phi_n$ are always non-primitive even given $\Phi$ is primitive. From this point of view, our consideration for non-primitive cases is essential even for primitive $\Phi$. It also indicates that Haagerup's reduction does not work for LSI/hypercontractivity.}\end{rem}
As we have seen in the tracial case, a combination of heat kernel estimates and spectral gap allow us to bound CB return time. This remains true in the GNS-symmetric case. Here we briefly illustrate the idea, as a careful treatment in finite dimenisonal cases is already given in \cite{bardet2021entropy}.
Let $E:{\mathcal M}\to {\mathcal N}$ be a $\phi$-preserving conditional expectation. Our technical tool, following \cite{gao2020fisher}, is the use of the module basis $x_{\alpha},\alpha\in I$ such that
\[ E(x_{\alpha}^*x_{\beta}) \pl = \pl \delta_{\alpha,\beta} e_{\alpha} \]
holds for a family of projections $e_{\alpha}\in {\mathcal N}$. For each $a\in{\mathcal M}$, there is a unique decomposition $a=\sum_{\alpha} x_\alpha a_\alpha$ with $a_\alpha=E(x_\alpha^* a)\in L_2({\mathcal N},\phi)$. In full generality the elements $x_{\alpha}$ may no longer be in ${\mathcal M}$, but belong to the strong closure of ${\mathcal M}\subset B(L_2({\mathcal M},\phi))$ (see \cite{junge2002doob,junge2005noncommutative}). For an ${\mathcal N}$-bimodule map $T:{\mathcal M} \to{\mathcal M}$, the module Choi operator is defined as
\[\chi_T=\sum_{\alpha,\beta\in I}\ket{\alpha}\bra{\beta} \otimes T(x_\alpha^*x_\beta)\in B(\ell_2(I))\overline{\otimes}{\mathcal M} \]
$T$ and $\chi_T$ determine each other because for a family $a_i=\sum_{\alpha} x_\alpha a_{\alpha,i}$
\begin{align*}\iota_{\infty,2}^n \otimes T(\sum_{i,j}\ket{i}\bra{j}\otimes a_i^*a_j )=&\sum_{i,j}\ket{i}\bra{j}\otimes T(a_i^*a_j)\\
=&\sum_{i,j}\ket{i}\bra{j}\otimes T(a_i^*a_j)\\
=&\sum_{i,j}\ket{i}\bra{j}\otimes \sum_{\alpha,\beta}a_{\alpha,i}^*T(x_\alpha^*x_\beta) a_{\beta,j}\\
=& (\sum_{\alpha, i} \ket{i}\bra{\alpha} \otimes a_{\alpha,i}^*)\chi_T (\sum_{\beta, i} \ket{\beta}\bra{j}\otimes a_{\beta,i})\hspace{.1cm}.
\end{align*}
This also shows that $T$ is completely positive if and only if $\chi_T$ is positive.
For $1\le p\le \infty$, we define the $\phi$-weighted conditional $L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)$ space as the completion of ${\mathcal M}$ under the norm
\[ \norm{x}{L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)}=\sup\{\norm{axb}{p,\phi}\hspace{.1cm} |\hspace{.1cm} a,b\in {\mathcal N},\hspace{.1cm} \norm{aa^*}{p,\phi}=\norm{b^*b}{p,\phi}=1\hspace{.1cm}\}\hspace{.1cm}.\]
At $p=1$, we have for $x\ge 0$,
\[ \norm{x}{L_\infty^1({\mathcal N}\subset {\mathcal M}, \phi)}=\norm{E(x)}{\infty}\hspace{.1cm}.\]
For a bimodule map $T$,
\begin{align}\label{eq:modulechoi}\norm{\chi_T}{B(\ell_2(I))\overline{\otimes}{\mathcal M}}=\norm{T:L_\infty^1({\mathcal N}\subset {\mathcal M}, \phi)\to {\mathcal M}}{cb}\hspace{.1cm},\end{align}
Here, we only need the easy direction "$\le$", which stems from
\begin{align*} \norm{\sum_{\alpha,\beta}\ket{\alpha}\bra{\beta}\otimes x_\alpha^*x_\beta }{B(\ell_2(I))\underset{min}{\otimes} L_{\infty}^1({\mathcal N}\subset {\mathcal M})}=& \norm{\sum_{\alpha,\beta}\ket{\alpha}\bra{\beta}\otimes x_\alpha^*x_\beta }{L_{\infty}^1(B(\ell_2(I))\overline\otimes {\mathcal N}\subset B(\ell_2(I))\overline{\otimes}{\mathcal M})}
\\ =& \norm{\iota_{\infty,2}^n\otimes E(\sum_{\alpha,\beta}\ket{\alpha}\bra{\beta}\otimes x_\alpha^*x_\beta )}{\infty}
\\ =& \norm{\sum_{\alpha}\ket{\alpha}\bra{\alpha}\otimes E_\alpha}{\infty}\le 1\hspace{.1cm}.
\end{align*}
Denote $\chi_E=\sum_{\alpha,\beta}\ket{\alpha}\bra{\beta}\otimes x_\alpha^*x_\beta$ as the Choi operator of $E$,
\begin{align*} \norm{\chi_T}{B(\ell_2(I))\overline{\otimes}{\mathcal M}}=&\norm{\iota_{\infty,2}^n\otimes T( \chi_E)}{B(\ell_2(I))\overline{\otimes}{\mathcal M}}
\\\le &\norm{T: L_{\infty}^1(B(\ell_2(I))\overline{\otimes} {\mathcal N}\subset B(\ell_2(I))\overline{\otimes}{\mathcal M})\to B(\ell_2(I))\overline{\otimes}{\mathcal M})}{} \\&\cdot \norm{\chi_0}{ L_{\infty}^1(B(\ell_2(I))\overline{\otimes} {\mathcal N}\subset B(\ell_2(I))\overline{\otimes}{\mathcal M})}\\ \le& \norm{T:L_\infty^1({\mathcal N}\subset {\mathcal M}, \phi)\to {\mathcal M}}{cb}
\end{align*}
As a consequence, we have
\begin{align}\label{eq:equiv}\norm{T-E:L_\infty^1({\mathcal N}\subset {\mathcal M}, \phi)\to {\mathcal M}}{cb}\le \varepsilon \Longrightarrow (1-\varepsilon) E\le_{cp} T \le_{cp} (1+\varepsilon)E\hspace{.1cm}.\end{align}
It is clear that the right hand side only rely on positivity structure. Indeed, the $L_\infty^1({\mathcal N}\subset {\mathcal M}, \phi)$ norm is independent of the choice of the invariant state $\phi=\phi\circ E$.
\begin{lemma}\label{lemma:independent}
Let $\phi$ be faithful normal state and $E:{\mathcal M}\to {\mathcal N}$ be a $\phi$-preserving conditional expectation. Suppose $\psi=\psi\circ E$ is another normal faithful state preserved by $E$. Then for any $x\in {\mathcal M}$,
\[ \norm{x}{L_\infty^p({\mathcal N}\subset{\mathcal M}, \phi)}=\norm{x}{L_\infty^p({\mathcal N}\subset{\mathcal M}, \psi)}\]
The identity extends to all $x\in L_\infty^p({\mathcal N}\subset{\mathcal M}, \phi) $.
\end{lemma}
\begin{proof}
Note that if both $\phi$ and $\psi$ are trace-preserving, then $d_{\psi}^{-\frac{1}{2p}}d_\phi^{\frac{1}{2p}}$ is affiliated to ${\mathcal N}$. Indeed, as argued in Lemma \ref{lemma:trace}, if $\psi\le C\phi$, then $d_\psi= hd_\phi h^*$ for some $h\in {\mathcal N}$, and the general case follows from approximation $\psi\le \frac{1}{\varepsilon}((1-\varepsilon)\phi+\varepsilon \psi)$.
Then we have
\[\norm{aa^*}{\phi, p}=\norm{d_{\phi}^{\frac{1}{2p}}aa^*d_{\phi}^{\frac{1}{2p}}}{p}=\norm{d_{\psi}^{-\frac{1}{2p}}d_{\phi}^{\frac{1}{2p}}aa^*d_{\phi}^{\frac{1}{2p}}d_{\psi}^{-\frac{1}{2p}}}{\psi,2p} \]
Denote $a_1=d_{\psi}^{-\frac{1}{2p}}d_\phi^{\frac{1}{2p}}a$ and $b_1=bd_\phi^{\frac{1}{2p}}d_{\psi}^{-\frac{1}{2p}}$.
For $x\in {\mathcal M}$,
\begin{align*}
\norm{x}{L_\infty^p({\mathcal N}\subset{\mathcal M}, \phi)}=&\sup_{\norm{\hspace{.1cm} aa^*\hspace{.1cm}}{\phi,2p}=\norm{\hspace{.1cm} b^*b\hspace{.1cm}}{\phi,2p}=1}\norm{axb}{\phi,p}
=\sup_{\norm{\hspace{.1cm} aa^*\hspace{.1cm}}{\phi,p}=\norm{\hspace{.1cm} b_1b_1\hspace{.1cm} }{\phi,p}=1}\norm{d_{\psi}^{-\frac{1}{2p}}d_\phi^{\frac{1}{2p}}axbd_{\phi}^{\frac{1}{2p}}d_{\psi}^{-\frac{1}{2p}}}{\psi,p}
\\=& \sup_{\norm{\hspace{.1cm} a_1a_1^*\hspace{.1cm}}{2p}=\norm{\hspace{.1cm} b_1b_1^*\hspace{.1cm}}{2p}=1}\norm{a_1xb_1}{\psi, p}=\norm{x}{L_\infty^p({\mathcal N}\subset{\mathcal M},\psi)}
\end{align*}
where the supremum are for $a,b\in {\mathcal N}$.
\end{proof}
Based on above we have obtain
\begin{prop}\label{prop:finitetcb3} The Proposition \ref{prop:finitecb} and \ref{prop:finitetcb2} remains valid in the GNS-symmetric cases by replacing $L_2({\mathcal M},\tau)$ by $L_2({\mathcal M},\phi)$ and
$L_\infty^1({\mathcal N}\subset {\mathcal M})$ and $L_\infty^1({\mathcal N}\subset {\mathcal M},\phi)$.
\end{prop}
\begin{proof}The argument is identical to the tracial cases by using the property of $L_\infty^p({\mathcal N}\subset {\mathcal M},\phi)$ for general von Neumann algebra established \cite{junge2010mixed}. See also \cite[Section 5]{bardet2021entropy} for argument in finite dimensional GNS-symmetric cases.
\end{proof}
\begin{rem}{\rm \label{rem:phitr} For finite ${\mathcal M}$, one particular invariant state of $E$ used in \cite{bardet2022hypercontractivity,bardet2021entropy} is
\[\phi_{{\rm \text{tr}}}=E_*(1)\hspace{.1cm} .\]
This state is convenient because $\phi_{{\rm \text{tr}}}|_{\mathcal N}$ is a tracial state. Then by Lemma \ref{lemma:independent}, we have
\[ \norm{x}{L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)}=\norm{x}{L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi_{\rm \text{tr}})}=\sup\{\norm{axb}{p,\phi_{\rm \text{tr}}}\hspace{.1cm} | a,b\in {\mathcal N},\hspace{.1cm} \norm{a}{p,\phi_{\rm \text{tr}}}=\norm{b}{p,\phi_{\rm \text{tr}}}=1\hspace{.1cm}\}\]
where we used the fact $L_p({\mathcal N},\phi_{\rm \text{tr}})$ is a tracial $L_p$-space. We will use this point to simplify the discussion in Section \ref{sec:concentration}.
}
\end{rem}
\subsection{Applications to finite Quantum Markov chains}
\label{sec:symmetricQMS}
Let $T_t=e^{-Lt}:{\mathbb M}_d\to {\mathbb M}_d$ be a quantum Markov semigroup on matrix algebra ${\mathbb M}_d$. Its generator $L$ admits the following Lindbladian form (\cite{gorini1976completely,lindblad1976generators})
\[L(x)= i[h,x]+\sum_{j} \gamma_j(V_j^*[x,V_j]+[ V_j^*,x]V_j)\hspace{.1cm} , \]
where $h,V_j\in {\mathbb M}_d$ and $h=h^*$ is Hermitian. When $T_t$ is GNS-symmetric, Carlen and Maas proved the following simplified form
\[L(x)=\sum_{j}e^{-w_j/2}\Big(V_j^*[x,V_j]+[ V_j^*,x]V_j\Big)\hspace{.1cm} , \]
where $\{V_j\}=\{V_j\}^*$ is an orthonormal set with respect to trace inner product and the eigenvector of modular group
\[\alpha_t^{\phi}(V_j)=e^{-iw_j t}V_j\hspace{.1cm}.\]
Gao and Rouz\'e in \cite[Theorem 3.3]{gao2022complete} proved that in finite dimensions,
\begin{align}\label{eq:GRbound} 0<\frac{\lambda}{2C_{cb}(E)} \le \alpha_{c}\le \alpha_1 \le\lambda\hspace{.1cm} ,\end{align}
where $E$ is the conditional expectation onto the fixed point subalgebra.
Here, we obtain an exponential improvement on the dependence of the dimension constant $C_{cb}(E)$.
\begin{theorem}\label{thm:fdsemigroup}
Let $\phi$ be a faithful state.
For a $\phi$-symmetric quantum Makrov semigroup $T_t=e^{-tL}:{\mathbb M}_d\to {\mathbb M}_d$,
\begin{align}\frac{\lambda}{2\ln (10 C_{cb}(E))}\le \alpha_{c}(L)\le \alpha_1(L)\le \lambda(L) \hspace{.1cm}.\label{eq:fdQMS}\end{align}
\end{theorem}
\begin{proof}It follows from combining Theorem \ref{thm:GNSsemigroup} and Proposition \ref{prop:finitetcb3}.
\end{proof}
\begin{rem}{\rm
In the primitive case ${\mathcal N}=\mathbb{C}1$, the conditional expectation $E_\phi(x)=\phi(x)1$ has index
\[C(E_\phi)=\norm{\phi^{-1}}{\infty}\hspace{.1cm}, C_{cb}(E_\phi)\le \norm{\phi^{-1}}{\infty}^2\hspace{.1cm}.\]
The above theorem gives
\[ \alpha_1\ge \alpha_{c}\ge \frac{\lambda}{2\ln10+4
\ln\norm{\phi^{-1}}{\infty}}\hspace{.1cm}. \]
It can be compared to the bound
\begin{align} \frac{2(1-\frac{2}{\norm{\phi^{-1}}{\infty}})\lambda}{\ln(\norm{\phi^{-1}}{\infty}-1)}\le \alpha_2 \le \alpha_1 \le\lambda \label{eq:TK}\hspace{.1cm}.\end{align}
This bound was proved Diaconis and Saloff-Coste \cite{diaconis1996logarithmic} for symmetric classical Markov semigroups. In the quantum case, it is only partially obtained for unital semigroups \cite{kastoryano2013quantum} and $d=2$ \cite{beigi2020quantum}.
In both classical and quantum depolarizing semigroup $L(x)=x-\phi(x)1$, this bound is known to be tight for $\alpha_2$, which upper bounds $\alpha_1$. In this sense, our results gives a general $\mathcal{O}(\frac{\lambda}{\norm{\phi^{-1}}{\infty}})$ lower bound for $\alpha_1$, not only for the primitive case, but also for non-primitive cases and complete constant $\alpha_c$. We also note that for both classical and quantum depolarizing semigroup $L(x)=x-\phi(x)1$, this bound is known to be tight for $\alpha_2$, which upper bounds $\alpha_1$.
}
\end{rem}
\begin{rem}{\rm The Corollary \ref{eq:loglog} shows that the CMLSI constant $\alpha_{c}$ for classical Markov semigroup is lower bounded by LSI constant $\alpha_2$ up to a $O(\log\log \norm{\mu^{-1}}{\infty})$ term. This argument does not work for Quantum Markov semigroup $T_t:\mathbb{M}_d\to {\mathbb M}_d$ on matrix algebras, despite relation \eqref{eq:re}
remain valid for primitive quantum Markov semigroup. The difference is that for matrix algebra, the bounded return time
\[t_b:=\frac{1}{2}\inf\{t>0\hspace{.1cm} |\hspace{.1cm} \norm{T_t-E:L_1({\mathbb M}_d,\phi)\to L_\infty({\mathbb M}_d)}{}<1/e^2\}\]
and the CB return time of completely bounded norm
\begin{align*}
t_{cb}=\inf\{t>0 \hspace{.1cm} | \hspace{.1cm} \norm{T_t-E_\mu:L_1({\mathbb M}_d,\phi)\to L_\infty({\mathbb M}_d)}{cb}<1/10\hspace{.1cm}.\}
\end{align*}
are quite different.
In the classical setting, we used the fact \[\norm{T:L_1(\Omega)\to L_\infty(\Omega)}{}=\norm{T:L_1(\Omega)\to L_\infty(\Omega)}{cb}\hspace{.1cm}.\] So the $t_b$ and $t_{cb}$ are comparable by absolute constants.
In the noncommutative setting, we can only have \[\norm{T_t-E_\mu:L_1({\mathbb M}_d,\phi)\to L_\infty({\mathbb M}_d)}{cb}\le d\norm{T_t-E_\mu:L_1({\mathbb M}_d)\to L_\infty({\mathbb M}_d)}{}.\] In the trace symmetric case, $\norm{\mu^{-1}}{\infty}=d$ and $t_{cb}\le 3t_b +\ln d$,
\[ \alpha_{c}\ge \frac{1}{4t_{cb}}\ge \frac{1}{12t_b+ 4\ln d}\sim O(\frac{\alpha_2}{\ln d})\hspace{.1cm},\]
which is worse than the lower bound in Theorem \ref{thm:fdsemigroup} as $\alpha_2\le \lambda$.
}
\end{rem}
\subsection{An example of noncommutative Birth-Death Process}
We illustrate our estimate with a noncommutative Birth-Death process. This example is an extension of usual graph Laplacians to matrix algebras (see \cite{li2020graph,junge2019stability} for similar constructions).
To fix the notation, let $\mathcal{G}=(V, E)$ be an undirected graph with $n=|V|$ vertices and edge set $E$. For each edge $(r,s)$, we introduce the edge Lindbladian on ${\mathbb M}_n$,
\begin{align*} L_{rs}(x)=&e^{\beta_{rs}/2}L_{e_{rs}}(x) + e^{-\beta_{rs}}L_{e_{sr}}(x) \\
&= e^{\beta_{rs}/2}(e_{ss}x+xe_{ss}-2e_{sr}xe_{rs})
+e^{-\beta_{rs}/2}(e_{rr}x+xe_{rr}-2e_{sr}xe_{sr}) \hspace{.1cm} ,
\end{align*}
where $e_{rs}\in {\mathbb M}_n$ is the matrix unit with $1$ at the $(r,s)$ position. The total Lindbladian is a weighted sum over the edge set $E$,
\[ L \pl = \pl \sum_{(r,s)\in E} w(r,s) L_{rs}\]
where we assume $\beta_{rs}=-\beta_{sr}$ and $w(r,s)=w(s,r)>0$ for the GNS symmetry condition. Note that for $j\neq k$,
\begin{align*} &L(e_{jk})=2(\sum_{(r,k)\in E}w(r,k)e^{\beta_{rk}/2}
+ \sum_{(r,j)\in E} w(r,j)e^{\beta_{r,j}/2})e_{jk} \\
&L(e_{jj}) = 4 \sum_{(r,j)\in E} w(r,j)(e^{\beta_{r,j}/2} e_{jj}-e^{-\beta_{r,j}/2}e_{rr}) \hspace{.1cm} .
\end{align*}
Let us collect some relevant facts of such Lindbladian $L$ as noncommutative extension of graph Laplacian.
\begin{enumerate}
\item[i)] Denote $\ell_{\infty}(V)\subset {\mathbb M}_n$ as the diagonal subalgebra.
$L(\ell_{\infty}(V))\subset \ell_{\infty}(V)$, and $L|_{\ell_{\infty}(V)}$ is a weighted graph Laplacian;
\item[ii)] For $r\neq s$, the matrix unit $e_{rs}$ is an eigenvector of $L$
\[ L(e_{rs}) \pl = \pl \gamma_{rs} e_{rs} \]
where $\gamma_{rs}=2(\sum_{(r,j)\in E}w(r,j)e^{\beta_{rj}/2}+\sum_{(k,s)\in E}w(k,s)e^{-\beta_{ks}/2})$.
\item[iii)] $\ker(L)\subset \ell_{\infty}(V)$, and $\ker(L)=\mathbb{C}1$ if $\mathcal{G}=(V, E)$ is connected.
\item[iv)] Let $\mu=(\mu_k)\in \ell_{\infty}(V)$ be density operator in the diagonal subalgebra. Then $L$ is $\mu$-symmetric if $e^{\beta_{rs}}=\mu_s/\mu_r$ for any $s\neq r$.
\end{enumerate}
Assume $L=\sum_{(s,r)\in E}L_{sr}$ is a primitive graph Lindbladian satisfying $\mu$-symmetric condition for a diagonal density operator $\mu$. Denote $E_d$ as the projection onto diagonal subalgebra. We can decompose the semigroup $T_t=e^{-tL}$ on the diagonal part and off diagonal part.
\begin{align*}
T_t=T_tE_d +T_t(\iota_{\infty,2}^n-E_d):=T_t^{diag}+T_t^{off}\hspace{.1cm}.
\end{align*}
It is clear from i) and ii) that $T_tE_d$ is a classical graph random walk and $T_t(\iota_{\infty,2}^n-E_d)$ is a Schur multiplier on ${\mathbb M}_n$. Using this decomposition, we have the CB-return time of the semigroup
\[ t_{cb}(\varepsilon):=\inf\{t>0\hspace{.1cm} | \hspace{.1cm} \norm{T_t: L_1({\mathbb M}_n,\mu)\to {\mathbb M}_n}{}\}\]
satisfies
\[ t_{cb}(2\varepsilon)\le t^{diag}_{cb}(\varepsilon)+t^{off}_{cb}(\varepsilon)\]
where $t^{diag}_{cb}$ and $t^{off}_{cb}$ are the CB-return time for the diagonal part $T_tE_d$ and off diagonal part $T_t(\iota_{\infty,2}^n-E_d)$ respectively. For the diagonal part, this is a classical bounded mixing time, i.e. the smallest $t$ such that
\[ \norm{T_tE_d: L_1(V,\mu)\to L_\infty(V)}{}\le \varepsilon\]
For the off-diagonal term, we have by Effros-Ruan isomorphism that a Schur multiplier map
\begin{align*}\norm{T_t(\iota_{\infty,2}^n-E_d): L_1({\mathbb M}_n,\mu)\to {\mathbb M}_n}{cb}=&\norm{\sum_{r\neq s} \mu_{r}^{-1/2}e^{-\gamma_{rs}t}\mu_{s}^{-1/2}e_{rs}\otimes e_{rs} }{\infty}\\ =&\norm{\sum_{r\neq s} \mu_{r}^{-1/2}e^{-\gamma_{rs}t}\mu_{s}^{-1/2}e_{rs} }{\infty}\hspace{.1cm}.\end{align*}
Note that for each $t$,
\[A_t=\sum_{r\neq s} \mu_{r}^{-1/2}e^{-\gamma_{rs}t}\mu_{s}^{-1/2}e_{rs}\]
is a symmetric matrix with positive entry. A standard application of Schur's lemma for matrices with positive entries implies
\begin{align*}
\norm{A_t}{\infty}\le \sup_{r}\Big(\sum_{s}\mu_{r}^{-1/2}e^{-\gamma_{rs}t}\mu_{s}^{-1/2}\Big)\hspace{.1cm},
\end{align*}
which gives us an estimate for the off diagonal term.
Now we consider the birth-death process on a finite state space $V=\{1,\cdots,n\}$, which we denote as $L^{BD}_n$. The corresponding edge $E$ set consists of only successive vertices $E=\{(j,j+1)| 1\le j\le n-1\}$. The simplest case chooses the uniform weight $w(r,s)=1$ for $(r,s)\in E$ and allows only one Bohr frequency $e^{-\beta}=\frac{\mu_{j}}{\mu_{j+1}}$, and the resulting stationary measure is the well-studied thermal state \[\mu=Z_{\beta}^{-1}(e^{-\beta j})_{j=1}^n\]
where $Z_{\beta}=\sum_{j=1}^{n} e^{-\beta j}$ is the normalization constant. In this case, $\gamma_{rs}=8(\cosh \beta) t$, and the off diagonal CB norm can be estimated by
\begin{align*} \norm{A_t}{\infty}\le & \sup_{r}\Big(\sum_{s=1}^n e^{\beta r/2}e^{\beta s/2}\Big) Z_\beta e^{-8(\cosh \beta) t}
\\ \le & e^{\beta \frac{n-1}{2}}e^{\beta/2}e^{-\beta}\frac{1-e^{n\beta /2}}{1-e^{\beta /2}} \frac{ 1-e^{-n\beta }}{ 1-e^{-\beta}} e^{-8(\cosh \beta) t}
\end{align*}
Thus $t_{cb}^{off}(\varepsilon)\le C_1(\beta)n$ for some constant $C_1(\beta)$ depending on $\beta$. For the classic part, We refer to \cite{Miclo99} and \cite{Chen} for the fact that the spectral gap is of order $O(1)$, i.e.
\[ c(\beta) \le \lambda(L^{diag}_n)\le C(\beta) \]
for all $n\in \bN$. For the commutative system on the diagonal part this implies (see also \cite{diaconis1996logarithmic})
\[ t^{diag}_{cb}(\varepsilon) \pl \le \pl 2c(\beta)^{-1}(2+ |\log \mu_n|)\pl \le \pl C_2(\beta)n \hspace{.1cm} ,\]
(for $\varepsilon=e^{-2}$, but here the actual value of $\varepsilon$ does not matter the asymptotic).
On the other hand, Miclo \cite{Miclo99} proves that \[t^{diag}_{cb}\ge \alpha_1(L^{diag}_n)^{-1}\pl \ge \pl c(\beta)n\hspace{.1cm}.\] Combining the diagonal and off diagonal part, we know $t^{diag}_{cb}(L^{BD}_n)\sim \frac{1}{n}$. It turns out the same asymptotic holds for CMLSI constant $\alpha_c(L^{BD}_n)$, which indicates our $\alpha_{c}\ge \frac{1}{2t_{cb}}$ estimate is asymptotically tight.
\begin{theorem} \label{birthdeath} For $\beta > 0$, there exist constants $c(\beta), C(\beta)>0$ such that the CMLSI constant of noncommutative birth-death process $L_n^{BD}$ satisfies
\[ \frac{c(\beta)}{n} \pl \le \pl \alpha_c(L_n^{BD})\le \alpha_1(L^{BD}_n) \pl \le \pl \frac{C(\beta)}{n} \hspace{.1cm} .\]
The same $\Theta(\frac{1}{n})$ asymptotic holds for the CB return time $t_{cb}(L^{BD}_n)$.
\end{theorem}
\begin{proof} It suffices to show that
\[\alpha_c(L^{BD_n})\le \alpha_1(L^{BD}_n)\le c(\beta)/n \hspace{.1cm}.\]
For this, we consider the function in the commutative system on the diagonal \[f(k)=\frac{Z(\beta)}{n}e^{\beta k} \text{ and } \sum_{k=1}^nf(k)\mu(k)=\sum_{k=1}^{n} \frac{Z(\beta)}{n}e^{\beta k}\frac{1}{Z(\beta)}e^{-\beta k}=1\]
so that $f\mu:=\rho$ represents a probability density. The relative entropy term satisfies
\[ D(\rho||\mu)= D(f\mu||\mu)\pl = \pl \sum_k \frac{e^{-\beta k}}{Z(\beta)} f(k)(\beta k+\ln Z(\beta)-\ln n)\
\pl = \pl \ln Z(\beta)-\ln n+ \beta \frac{n+1}{2} \hspace{.1cm} .\]
Our density is $\rho\equiv(\frac{1}{Z(\beta)n})$ and the reference density if $\mu(k)=\frac{e^{-\beta k}}{Z(\beta)}$. Denote $a_k=|k\rangle\langle k+1|$. On the diagonal, we have
\begin{align*}
\frac{1}{2}L^*(f)&=
\sum_k e^{\beta/2}(a_ka_k^*f-a_k^*fa_k)+ e^{-\beta/2}(a_k^*a_kf-a_kfa_k^*) \\
&= \sum_k e^{\beta/2}(e_{k}f(k)-f(k)e_{k+1})
+ e^{-\beta/2}(f(k+1)e_{k+1}-f(k+1)e_k) \\
&= \frac{1}{Z(\beta)n} (e^{\beta/2}(e_0-e_{n})
+ e^{-\beta/2}(e_n-e_0)) \hspace{.1cm} .
\end{align*}
Note that
\[\ln \rho-\ln \mu \pl = \pl \ln f \pl = \pl \big(\beta k-\ln (Z(\beta)n)\big)_{k=1}^n \hspace{.1cm} .\]
Since there are only two terms, we have
\[ EP(\rho) \sim c(\beta) \hspace{.1cm} \]
for some constant $c(\beta)$ only depending on $\beta$. This holds for $n\pl \ge \pl n_0$ large enough.
\end{proof}\vspace{0.5ex}
\begin{rem}
When $\beta > 0$, $\sum_{k=1}^n e^{-\beta k} = O(1)$ as a geometric series. In the case that $\beta = 0$, Theorem \ref{birthdeath} no longer yields with $\alpha_c(L_n)\sim \frac{1}{n}$. For $\beta = 0$, the birth-death process reduces to a ``broken" version of the cyclic graph as in Example \ref{ex:cyclic2} with $\alpha_c(L_n) \sim 1/n^2$.
\end{rem}
\subsection{Noncommutative concentration inequality}\label{sec:concentration}
One important application of modified Log-Sobolev inequality is to derive concentration inequalities.
Such implication is discovered by Otto and Villani \cite{otto2000generalization} for classical Markov semigroups (see also the work of Erbar and Maas \cite{erbar2012ricci} in the discrete case), and recently extended to the noncommutative setting in \cite{rouze2019concentration,gao2020fisher,carlen2020non}.
In this section we will show that $\text{CMLSI}$ of a GNS symmetric semigroup also implies concentration inequalities. A key quantity in the discussion of concentration inequalities is the Lipschitz norm
\[ \|x\|_{\text{Lip}}^2 \pl = \pl \max\{ \norm{\Gamma_L(x,x)}{}\hspace{.1cm}, \hspace{.1cm} \norm{\Gamma_{L}(x^*,x^*)}{} \} \]
where the \emph{gradient form} (or Carr\'e du Champ operator) is
\[ \Gamma_L(x,y) \pl = \pl \frac{1}{2}\Big(L(x^*)y+x^*L(y)-L(x^*y)\Big)\hspace{.1cm}, \hspace{.1cm} \forall x,y\in \text{dom}(L)\hspace{.1cm}. \]
Note that $\|\cdot\|_{\text{Lip}}$ is a semi-norm (satisfying triangle inequality) because $\Gamma_L$ is completely positive bilinear form.
Our first lemma is to show that $\|x\|_{\text{Lip}}$ can be approximated by Haagerup reduction.
\begin{lemma}\label{down} Let $x\in {\mathcal M}$. Then for all $n\in \bN$
\[ \|E_{{\mathcal M}_n}(x)\|_{Lip}\pl \le \pl \|x\|_{Lip} \hspace{.1cm} .\]
\end{lemma}
\begin{proof} Recall the conditional expectation $E_{{\mathcal M}_n}:\hat{{\mathcal M}}\to {\mathcal M}_n$ is given by
\[ E_{{\mathcal M}_n}(x)=2^n\int_{0}^{2^{-n}} \alpha_t^{\psi_n}(x)d t \]
Note that $\alpha_t^{\psi_n}$ is an inner automorphism on ${\mathcal M}\rtimes_\alpha 2^{-n}\mathbb{Z}$. We note that for a modular group $\alpha_t$ such that $L\alpha_t=\alpha_tL$,
\[ \Gamma_L(\alpha_t(x),\alpha_t(y))
\pl = \pl \alpha_t(\Gamma_L(x,y)) \hspace{.1cm}, \]
which implies $\norm{x}{\text{Lip}}=\norm{\alpha_t(x)}{\text{Lip}}$.
Here both $\alpha_t^{\hat{\phi}}$ and $\alpha_t^{\psi_n}$ commutes with $\hat{L}=\iota_{\infty,2}^n_{\mathbb{T}}\otimes L$.
Then by triangle inequality,
\begin{align*}
\norm{E_{{\mathcal M}_n}(x)}{\text{Lip}} &= \Big \| 2^n\int_0^{2^{-n}} \alpha_{t}^{\psi_n}(x) dt \Big \|_{\text{Lip}}\le 2^n\int_0^{2^{-n}} \norm{x }{\text{Lip}}dt=\norm{x }{\text{Lip}}\hspace{.1cm}. \quad \qedhere
\end{align*}
\end{proof}\vspace{0.5ex}
\begin{lemma}\label{lemma:commuting}Let ${\mathcal M}_0,{\mathcal N}\subset {\mathcal M}$ be two subalgebras and $\phi$ be a normal faithful state $\phi$. Suppose $E_{0}:{\mathcal M}\to {\mathcal M}_0$ and $E:{\mathcal M}\to {\mathcal N}$ are $\phi$-preserving conditional expectations onto ${\mathcal M}_0$ and ${\mathcal N}$ respectively. Suppose $E\circ E_0=E_0\circ E$ satisfy the commuting square condition
\begin{center}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em]
{
{\mathcal M} & {\mathcal M}_0 \\
{\mathcal N} & {\mathcal N}_0 \\};
\path[-stealth]
(m-1-1) edge node [left] {$E$} (m-2-1)
(m-1-1) edge node [above] {$E_0$} (m-1-2)
(m-2-1) edge node [below] {$E_0$} (m-2-2)
(m-1-2) edge node [right] {$E$} (m-2-2)
;
\end{tikzpicture}\end{center}
Then for any $p\in [1,\infty]$ and any $x\in {\mathcal M}$,
\[ \norm{E_0(x)}{L_\infty^p({\mathcal N}_0\subset {\mathcal M}_0,\phi)}=\norm{E(x)}{L_\infty^p({\mathcal N}\subset {\mathcal M},\phi)}\le \norm{x}{L_\infty^p({\mathcal N}\subset {\mathcal M},\phi)}\]
In other words, $L_\infty^p({\mathcal N}_0\subset {\mathcal M}_0,\phi)\subset L_\infty^p({\mathcal N}\subset {\mathcal M},\phi)$ is a $1$-complemented subspace with projection $E_0$.
\end{lemma}
\begin{proof}We can assume $\phi=\phi_{\rm \text{tr}}$ in the Remark \ref{rem:phitr}. Using commuting square assumption, we know $E_0(a)\in {\mathcal N}_0$ for $a\in {\mathcal N}$.
By definition,
\begin{align*}
\norm{E_{0}(x)}{L_\infty^p({\mathcal N}_0\subset {\mathcal M}_0, \phi)}= &\sup_{a,b\in\hspace{.1cm} {\mathcal N}_0}\norm{aE_{0}(x)b}{\phi,p}
\le \sup_{a,b\in\hspace{.1cm} {\mathcal N}_0}\norm{E_0(axb)}{\phi,p}\\
\le &\sup_{a,b\in\hspace{.1cm} {\mathcal N}_0}\norm{axb}{\phi,p}\le \norm{x}{L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)}
\end{align*}
where the first supremum is for all $a,b$ in the corresponding subalgebra with $\norm{a}{\phi,p}=\norm{ b}{\phi,p}=1$. Now it suffices to show the other direction
$$\norm{x}{L_\infty^p({\mathcal N}_0\subset {\mathcal M}_0,\phi)}\ge \norm{x}{L_\infty^p({\mathcal N}_0\subset {\mathcal M}_0,\phi)}\hspace{.1cm},$$
for $x\in {\mathcal M}_0$. For that, we revoke that for $\frac{1}{p}+\frac{1}{q}$, $L_1^{p'}({\mathcal N}\subset {\mathcal M})\subset L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)^*$ is as a weak$^*$-dense subspace. Here for $x\in {\mathcal M}$,
\[ \norm{y}{L_1^{q}({\mathcal N}\subset {\mathcal M})}=\inf_{y=azb}\norm{a}{2p,\phi}\norm{y}{q,\phi}\norm{b}{2p,\phi}\]
where the infimum is over all factorization $a,b\in {\mathcal N}, z\in {\mathcal M}$. The duality pairing is given by the KMS inner product,
\[\langle x,y \rangle=\tau(x^*d_{\phi}^{1/2} y d_{\phi}^{1/2})=\langle x,y \rangle_{\phi} \]
Indeed, it was proved in \cite[Corollary 3.13]{junge2010mixed} that
\[E_0: L_1^{q}({\mathcal N}\subset {\mathcal M})\to L_1^{q}({\mathcal N}_0\subset {\mathcal M}_0)\]
is a contraction by the commuting square condition. Therefore, for $x\in {\mathcal M}_0$, by the $\phi$-symmetry of $E_0$,
\begin{align*}\norm{x}{L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)}=&\sup_{\norm{ \hspace{.1cm} y\hspace{.1cm} }{L_1^{q}({\mathcal N}\subset {\mathcal M})}=1}\langle x,y \rangle_\phi \\ =& \sup_{\norm{ \hspace{.1cm} y\hspace{.1cm} }{L_1^{q}({\mathcal N}\subset {\mathcal M})}=1}\langle x,E_0(y) \rangle_\phi
\\ \le& \sup_{\norm{ \hspace{.1cm} z\hspace{.1cm} }{L_1^{q}({\mathcal N}_0\subset {\mathcal M}_0)}=1}\langle x,z \rangle_\phi=\norm{x}{L_\infty^p({\mathcal N}_0\subset {\mathcal M}_0, \phi)} \hspace{.1cm}. \qedhere
\end{align*}
\end{proof}
\begin{lemma}\label{lemma:limit}For $x\in {\mathcal M}$, $\lim_{n}\|E_{{\mathcal M}_n}(x)\|_{L_\infty^p({\mathcal N}_n\subset {\mathcal M}_n, \psi_n)}= \|x\|_{L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)}$
\end{lemma}
\begin{proof} Recall the commuting square condition $E_{{\mathcal M}_n}\circ \hat{E}=\hat{E}\circ E_{{\mathcal M}_n}$. Then by Lemma \ref{lemma:independent} \& \ref{lemma:commuting},
\begin{align*}\norm{E_{{\mathcal M}_n}(x)}{L_\infty^p({\mathcal N}_n\subset {\mathcal M}_n, \psi_n)}=&\norm{E_{{\mathcal M}_n}(x)}{L_\infty^p({\mathcal N}_n\subset {\mathcal M}_n,\hat{\phi})}
=\norm{E_{{\mathcal M}_n}(x)}{L_\infty^p({\mathcal N}\subset {\mathcal M},\hat{\phi})}\le \norm{x}{L_\infty^p({\mathcal N}\subset {\mathcal M},\hat{\phi})}\hspace{.1cm}.
\end{align*}
The other direction follows from the weak$^*$-convergence $E_{{\mathcal M}_n}(x)
\to x$. Fix $\frac{1}{q}+\frac{1}{p}=1$.
For any $\varepsilon>0$, there exists $a_0,b_0\in \hat{{\mathcal N}}$ and $y_0\in \hat{{\mathcal M}}$ such that
\begin{align*}
&\norm{aa^*}{p,\hat{\phi}}=\norm{b^*b}{p,\hat{\phi}}=\norm{y}{\hat{\phi},q}=1\\
&\hat{\tau}(d_{\hat{\phi}}^{1/2}axb d_{\hat{\phi}}^{1/2}y)\ge \norm{x}{L_\infty^p(\hat{{\mathcal N}}\subset \hat{{\mathcal M}}, \hat{\phi})}-\varepsilon
\end{align*}
By the weak$^*$-density, we can choose $n_1,n_2,n_3$ and $n_4\ge \max\{n_1,n_2,n_3\}$ inductively such that
\begin{align*}
\tau(d_\phi^{1/2}E_{{\mathcal M}_{n_1}}(a)E_{{\mathcal M}_{n_4}}(x)E_{{\mathcal M}_{n_2}}(b)d_\phi^{1/2}E_{{\mathcal M}_{n_3}}(y))>\tau(d_\phi^{1/2}axbd_\phi^{1/2}y)-\varepsilon >\norm{x}{L_\infty^p(\hat{{\mathcal N}}\subset \hat{{\mathcal M}}, \hat{\phi})}-2\varepsilon\hspace{.1cm}. \end{align*}
Since $E_{{\mathcal M}_{n}}(\hat{{\mathcal N}})={\mathcal N}_n$ (see the commuting diagram after Lemma \ref{lemma:commute}), we have
\begin{align*}&\norm{E_{{\mathcal M}_{n_1}}(a)E_{{\mathcal M}_{n_1}}(a)^*}{\hat{\phi},p}\le \norm{E_{{\mathcal M}_{n_1}}(aa^*)}{\hat{\phi},p}\le \norm{aa^*}{\hat{\phi},p}= 1\\
&\norm{E_{{\mathcal M}_{n_3}}(b^*)E_{{\mathcal M}_{n_3}}(b)}{\hat{\phi},p}\le \norm{b^*b}{\hat{\phi},p}= 1\hspace{.1cm},
\\ & \norm{E_{{\mathcal M}_{n_4}}(y)}{\hat{\phi},q}\le \norm{y}{\hat{\phi},q}=1
\end{align*}
by the $\hat{\phi}$-symmetry of $E_{{\mathcal M}_n}$.
Then, for $n=n_4=\max\{n_1,n_2,n_3,n_4\}$,
\begin{align*}
\norm{x}{L_\infty^p(\hat{{\mathcal N}}_n\subset \hat{{\mathcal M}}_n, \hat{\phi})}&\ge \norm{E_{{\mathcal M}_{n}}(x)}{L_\infty^p(\hat{{\mathcal N}}_n\subset \hat{{\mathcal M}}_n, \hat{\phi})}\\ &\ge \tau(d_{\hat{\phi}}^{1/2}E_{{\mathcal M}_{n_1}}(a)xE_{{\mathcal M}_{n_2}}(b) d_{\hat{\phi}}^{1/2}E_{{\mathcal M}_{n_3}}(y))\\ &\ge \norm{x}{L_\infty^p(\hat{{\mathcal N}}\subset \hat{{\mathcal M}}, \hat{\phi})}-2\varepsilon
\end{align*}
which proves
\[\lim_{n}\|E_{{\mathcal M}_n}(x)\|_{L_\infty^p({\mathcal N}_n\subset {\mathcal M}_n, \psi_n)}= \|x\|_{L_\infty^p(\hat{{\mathcal N}}\subset \hat{{\mathcal M}}, \hat{\phi})}\hspace{.1cm}.\]
Finally, the assertion follows from
\[ \|x\|_{L_\infty^p(\hat{{\mathcal N}}\subset \hat{{\mathcal M}}, \hat{\phi})}=\|x\|_{L_\infty^p({\mathcal N}\subset {\mathcal M}, \phi)}\hspace{.1cm},\]
as a consequence of $E_0\circ \hat{E}=\hat{E}\circ E_0$ by Lemma \ref{lemma:commuting}.
\end{proof}
\begin{theorem} Let ${\mathcal M}$ be a finite von Neumann algebra and $T_t=e^{-tL}$ be a $\phi$-symmetric quantum Markov semigroup. Suppose $T_t$ satisfies \text{MLSI} with parameter $\alpha>0$. There exists an universal constant $c$ such that for $2\le p <\infty$
\[ \alpha\|x-E(x)\|_{L_p({\mathcal M},\phi)}\le \alpha\|x-E(x)\|_{L_{\infty}^p({\mathcal N}\subset {\mathcal M},\phi)}\pl \le \pl c\sqrt{p}\norm{x}{\text{Lip}} \hspace{.1cm} .\]
\end{theorem}
\begin{proof} We first show that if $T_t$ satisfying $\alpha$-MLSI, so does the approximation semigroup.
\[T_{n,t}=\hat{T}_t:{\mathcal M}_n\to {\mathcal M}_n\]
Indeed, as we see in the discussion above, ${\mathcal M}_n\subset {\mathcal M}\rtimes_{\alpha^{\phi}_t} 2^{-n}\mathbb{Z}\cong L_\infty(\mathbb{T}, {\mathcal M})$ and the extension $\hat{T}_t=T_t\otimes \iota_{\infty,2}^n_{\mathbb{T}}$ has $\alpha$-MLSI (because $L_\infty(\mathbb{T})$ is a commutative space). Note that since ${\mathcal M}_n\subset {\mathcal M}\rtimes_{\alpha^{\phi}_t} 2^{-n}\mathbb{Z}\subset {\mathcal M}\rtimes_{\alpha^{\phi}_t} G$, the restriction
$E_{{\mathcal M}_n}:{\mathcal M}\rtimes_{\alpha^{\phi}_t} 2^{-n}\mathbb{Z}\to {\mathcal M}_n$ is also a conditional expectation.
Then for any $\rho,\sigma\in S({\mathcal M}_n)$, we have
\[D(E_{{\mathcal M}_n,*}\rho|| E_{{\mathcal M}_n,*}\sigma)\le D(\rho|| \sigma)=D(\rho|_{{\mathcal M}_n}|| \sigma|_{{\mathcal M}_n})\le D(E_{{\mathcal M}_n,*}\rho|| E_{{\mathcal M}_n,*}\sigma)\hspace{.1cm}.\]
Now using the commutation relation $\hat{T}_t\circ E_{{\mathcal M}_n}=E_{{\mathcal M}_n}\circ \hat{T}_t$ and $E_{{\mathcal M}_n}\circ \hat{E}=E_n\circ E_{{\mathcal M}_n} $, we have
\begin{align*}D(T_{t,n,*}\rho|| E_{n,*}\rho)=&D(E_{{\mathcal M}_n,*}T_{t,n,*}\rho|| E_{{\mathcal M}_n,*}E_{n,*}\rho)
=D(T_{t,n,*}E_{{\mathcal M}_n,*}\rho|| \hat{E}_{*}E_{{\mathcal M}_n,*}\rho)\\ \le& e^{-2\alpha t} D(E_{{\mathcal M}_n,*}\rho|| \hat{E}_{*} E_{{\mathcal M}_n,*}\rho)\\=& e^{-2\alpha t} D(E_{{\mathcal M}_n,*}\rho|| E_{{\mathcal M}_n,*}E_{n,*}\rho)= e^{-2\alpha t} D(\rho|| E_{n,*}\rho) \hspace{.1cm}.
\end{align*}
Thus $T_{n,t} $ has $\alpha$-$\text{MLSI}$ on ${\mathcal M}_n$. Note that $T_{n,t}$ is both $\hat{\phi}$-symmetric for the extension state $\hat{\phi}$ and also $\psi_n$-symmetric for the trace $\psi_n$. Now, we may use the tracial version of the concentration inequality \cite[Theorem 6.10]{gao2020fisher} that for $x\in {\mathcal M}_n$
\[ \alpha\norm{x-E_n(x)}{L_\infty^p({\mathcal N}_n\subset {\mathcal M}_n)}\le C\sqrt{p}\norm{x}{\text{Lip}}\]
Now by the approximation of Lemma \ref{lemma:limit} and independence of $L_\infty^p$ on the reference state, for $x\in {\mathcal M}$,
\begin{align*}
\norm{{x-E(x)}}{L_\infty^p({\mathcal N}\subset {\mathcal M},\phi)}=&\lim_{n}\norm{E_{{\mathcal M}_n}(x-E(x))}{L_\infty^p({\mathcal N}_n\subset {\mathcal M}_n,\psi_n)}\\=&\lim_{n}\norm{E_{{\mathcal M}_n}(x)-E_{n}E_{{\mathcal M}_n}(x))}{L_\infty^p({\mathcal N}_n\subset {\mathcal M}_n,\psi_n)}
\\ \le& C\sqrt{p}\norm{E_{{\mathcal M}_n}(x)}{\text{Lip}}\le C\sqrt{p}\norm{x}{\text{Lip}}\hspace{.1cm}.
\end{align*}
The other inequality
\[ \norm{y}{L_\infty^p({\mathcal N}\subset {\mathcal M},\phi)}\ge \norm{y}{L_p({\mathcal M},\phi)}\]
is clear from definition of $L_\infty^p({\mathcal N}\subset {\mathcal M},\phi)$.
\end{proof}\vspace{0.5ex}
For Gaussian type concentration property, we need the following definition. \begin{defi}For an operator $O$ we say that
\[ \text{Prob}_{\phi}(|O|>t) \pl \le \pl \varepsilon \]
if there exist a projection $e$ such that
\[ \|eOe\|_{\infty} \le t \quad \mbox{and} \quad \phi(1-e)\le \varepsilon \hspace{.1cm} .\]
\end{defi}
The next lemma is a Cheybev inequality for $\phi$-weighted $L_p$ norm.
\begin{lemma}\label{lemma:cheb} Let $x\in L_p({\mathcal M},\phi)$ and $1<p<\infty$. Then
\[ \text{Prob}_\phi(|x|>t) \pl \le \pl 2 \Big ( \frac{t}{4} \Big )^{-p}\|x\|_{p,\phi} \hspace{.1cm} .\]
\end{lemma}
\begin{proof}We start with a positive element $x=y^2$ and assume $\|x\|_{p,\phi}=M$. Then we have
\[ M=\|x\|_{p,\phi} \pl = \pl \|d_{\phi}^{1/2p}xd_{\phi}^{1/2p}\|_p
\pl = \pl \|yd_{\phi}^{1/2p}\|_{2p}^2 \hspace{.1cm} .\]
Recall the asymmetric Kosaki $L_p$-space
\[ \norm{y}{L_{2p}^c({\mathcal M},\phi)}:=\|yd_{\phi}^{1/2p}\|_{2p} \hspace{.1cm}, \]
and the complex interpolation relation.
\[ L_{2p}^c({\mathcal M},\phi)=[{\mathcal M},L_2^c({\mathcal M},\phi)]_{1/p} \hspace{.1cm}. \]
By the relation between real and complex interpolation
\[ L_{2p}^c({\mathcal M},\phi)=[{\mathcal M},L_2^c({\mathcal M},\phi)]_{1/p}\subset [N,L_2^c({\mathcal M},\phi)]_{1/p,\infty} \hspace{.1cm} , \]
for every $s>0$ we have a decomposition $y \pl = \pl y_1+ y_2$ such that
\[ \|y_1\|_{\infty}+ s\norm{y_2}{L_2^c({\mathcal M},\phi)} \pl \le \pl s^{1/p}M^{1/2} \hspace{.1cm} .\]
The by Chebychev's inequality for the spectral projection $e=e_{[0,a]}(y_2^*y_2)$ of $y_2^*y_2$, we have
\[ a \phi(1-e)\pl \le \pl \phi(y_2^*y_2) \pl \le \pl s^{2/p-2}M \quad \mbox{and} \quad \|y_2e\|_{\infty}^2\pl = \pl \|ey_2^*y_2e\|_{\infty}\pl \le \pl a \hspace{.1cm} .\]
Choose $a=s^{2/p}$ and deduce that
\[ \|exe\|_{\infty}=\|ye\|_{\infty}^2 \pl \le \pl (\|y_1e\|_{\infty}+ \|y_2e\|_{\infty})^2\pl \le \pl (s^{1/p}+s^{1/p})^2
\pl \le \pl 4s^{2/p}M \hspace{.1cm} .\]
Then for $t=4s^{2/p}M$ and
\[ \phi(1-e) \pl \le \pl a^{-1} s^{2/p-2} \pl = \pl s^{-2} \pl = \pl (\frac{t}{4M})^{-p}=(\frac{t}{4})^{-p}M^p \hspace{.1cm}. \]
For an arbitrary $x$, we may write $x=x_1x_2$ such that
\[ \|d_\phi^{1/2p}x_1\|_{2p}\pl = \pl \|x_2d_\phi^{1/2p}\|_{2p} \pl = \pl \norm{x}{p,\phi}=1 \hspace{.1cm} .\]
Then for each $s>0$, we have decompotision
\[ x_1\pl = \pl x_{11}+x_{21} \hspace{.1cm} ,\hspace{.1cm} x_2=x_{21}+x_{22} \hspace{.1cm} .\]
with
\begin{align*} \|x_{11}\|_{\infty}+ s\norm{x_{21}}{L_2^c({\mathcal M},\phi)} \pl \le \pl s^{1/p}M^{1/2} \hspace{.1cm} ,
\|x_{12}\|_{\infty}+ s\norm{x_{21}}{L_2^c({\mathcal M},\phi)} \pl \le \pl s^{1/p}M^{1/2} \hspace{.1cm}
.\end{align*}
We then use the Chebychev inequality for
\[ \phi(x_{21}^*x_{21}+x_{22}^*x_{22}) \pl \le \pl 2s^{2/p-2}M \hspace{.1cm} .\]
Then, using the choice of $a=s^{2/p}$ and the spectral projection,
\begin{align*}
\|exe\|_{\infty}
&= \|e(x_1x_2)e\|_{\infty} \pl = \pl \|e(x_{11}+x_{12})(x_{21}+x_{22})e\|_{\infty} \\
&\pl \le \pl \|x_{11}x_{22}\|_{\infty} + \|ex_{12}x_{21}\|_{\infty}
\|x_{11}x_{22}e\|_{\infty} + \|ex_{12}x_{22}e\|_{\infty} \\
&\pl \le \pl 4s^{2/p}M \hspace{.1cm} .
\end{align*}
Thus for $t=4s^{2/p}M$,
\begin{align*}
\phi(1-e) &\le a^{-1} 2s^{2/p-2} \pl = \pl 2s^{-2}
\pl = \pl 2 (\frac{t}{4M})^{-p}= 2(\frac{t}{4})^{-p}M^p\hspace{.1cm} . \qedhere
\end{align*}
\end{proof}\vspace{0.5ex}
\begin{cor}\label{cor:concentration} Let $T_t=e^{-tL}$ be a $\phi$-symmetric quantum Markov semigroup. Suppose $T_t$ satisfies $\alpha$-MLSI. Then for any $x\in{\mathcal M}$ and $t>0$,
\[ \text{Prob}_{\phi}(|x-E_{fix}(x)|>t)
\pl \le \pl 2 \exp \Big ( -\frac{1}{e} \Big (\frac{\alpha t}{4c\norm{x}{\text{Lip}}} \Big )^2 \Big )
\]where $c$ is an universal constant.
\end{cor}
\begin{proof} By Lemma \ref{lemma:cheb}, we have
\[ \text{Prob}_{\phi}(|x-E(x)|>t)\le 2 (t/4)^{-p}\norm{x-E(x)}{L_p({\mathcal M},\phi)}^p\le 2 \big(\frac{4C\norm{x}{\text{Lip}}\sqrt{p}}{\alpha t}\big)^{p}\]
The standard optimization gives $p=\frac{1}{e}(\frac{\alpha t}{4C\norm{x}{\text{Lip}}})^2$, which implies
\begin{align*}
&\text{Prob}_{\phi}(|x-E(x)|>t)\le 2 \exp( -\frac{\alpha^2t^2}{16ec^2\norm{x}{\text{Lip}}^2})\qedhere
\end{align*}
\end{proof}\vspace{0.5ex}
We remark that our argument for the above concentration inequality applies to general von Neumann algebras. The details are left to readers.
\begin{rem}{\rm In the primitive case, the above results can be compared to \cite[Theorem 8]{rouze2019concentration}, which states that for self-adjoint $x=x*$
\begin{align*}
\phi(e_{\{|x-E(x)|>t\}})\le \exp\Big( -\frac{\alpha t^2}{8\norm{d_\phi^{-1/2}xd_\phi^{1/2}}{\tilde{\text{Lip}}}^2}\Big)
\end{align*}
with a different Lipschitz norm. Our results here use a more natural definition of Lipschitz norm and apply to non-primitive cases. Nevertheless, the projection we have for
\[ \text{Prob}_{\phi}(|x-E(x)|>t)\]
is not necessarily a spectral projection $e_{\{|x-E(x)|>t\}}$ and depends on the state $\phi.$}
\end{rem}
\begin{rem}{\rm
In the operator valued setting, let ${\mathcal Q}$ be any finite von Neumann algebra and $T_t\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$ be the amplification semigroup on ${\mathcal Q}\overline{\otimes} {\mathcal M}$. The conditional expectation for $T_t\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$ is $E\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$. Note that by Lemma \ref{lemma:independent},
$T_t\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$ is GNS-symmetric to the product state $ \phi\otimes \sigma$, for any state $\sigma\in S({\mathcal Q})$ and any invariant state $\phi\in E_*(S({\mathcal M}))$.
This means we obtain
\[ \text{Prob}_{\phi\otimes \sigma}( |x-E_{fix}(x)|>t) \le 2 e^{-\frac{\alpha^2t^2}{C\norm{x}{\text{Lip}}^2}} \]
for any product state $\phi\otimes \sigma$ of this specific form. The projection depends, of course, on $\phi$ and $\sigma$.
}\end{rem}
We illustrate our result with a special case as matrix concentration inequalities.
\begin{exam}[Matrix concentration inequality]{\rm Let $S_1,\cdots,S_n$ be an independent sequence of random $d\times d$-matrices $S_1,\cdots,S_n$ such that
\[ \norm{S_i-\mathbb{E}S_i}{\infty}\le M \hspace{.1cm} , \hspace{.1cm} a.e. \]
It was obtained in \cite[Corollary 6.1.2]{tropp2015introduction} by Tropp the matrix Bernstein inequality that for the sum $Z=\sum_{k=1}^nS_k$
\[ {\mathbb E}\norm{Z-\mathbb{E}Z}{\infty}\le \sqrt{2v(Z)\log(2d)}+\frac{1}{3}M\log(2d)\]
and the matrix Chernoff bound
\[ P(|Z-\mathbb{E}Z|>t) \le 2d\exp\big(-\frac{t^2}{v(Z)+\frac{t}{3}M}\big)\]
where
\[v(Z)=\max \{ \norm{{\mathbb E}((Z-\mathbb{E}Z)^*(Z-\mathbb{E}Z))}\hspace{.1cm}, \norm{{\mathbb E}((Z-\mathbb{E}Z)^*(Z-\mathbb{E}Z))}{}\}\hspace{.1cm}.\]
Now to apply our results, we recall the depolarizing semigroup with generator $L(f):=(I-E_\mu)(f)=f-\mu(f)1_\Omega$ on any probability space $(\Omega,\mu)$
has $\alpha_c\ge \frac{1}{2}$. For a random matrix $f:\Omega\to M_d$,
the Lipschitz norm is
\begin{align*} \norm{f}{\text{Lip}}^2= &\frac{1}{2}\max\{\norm{\hat{f}^*\hat{f}+E(\hat{f}^*\hat{f})}{\infty}\hspace{.1cm}, \hspace{.1cm} \norm{\hat{f}\hat{f}^*+E(\hat{f}\hat{f}^*)}{\infty}\}\\ \le &\frac{1}{2}(\norm{f}{\infty}^2+v(f)),\end{align*}
where $\hat{f}=f-E_\mu(f)$ is the mean zero part.
Then we consider $S_k:\Omega_k\to M_d$ as a random matrix on $(\Omega_k,\mu_k)$. Then on the product space $(\Omega,\mu)=(\Omega_1,\mu_1)\times\cdots \times(\Omega_n,\mu_n)$, we have for $Z=\sum_{k}S_k$
\begin{align*}
{\mathbb E}\norm{Z-{\mathbb E} Z}{\infty}\le \Big ( \frac{1}{d}{\mathbb E}\norm{Z-{\mathbb E} Z}{p}^p \Big )^{1/p}\le d^{1/p}\norm{Z-{\mathbb E} Z}{L_\infty(M_d, L_p(\Omega) )} \le 2cd^{1/p}\sqrt{p}\norm{Z}{\text{Lip}} \hspace{.1cm},
\end{align*}
where $\norm{\cdot}{p}$ is the $p$-norm for the normalized trace,
and
\[ \norm{Z}{\text{Lip}}^2\le \frac{1}{2}(v(Z)+M^2)\hspace{.1cm}.\]
Optimizing $p$ gives
\[ {\mathbb E}\norm{Z-{\mathbb E} Z}{\infty}\le 2ce^{-1/2} \sqrt{ (v(Z)+M^2)\log d}\]
And for the matrix Chernoff bound
\[ P(|Z-\mathbb{E}Z|>t)\le d\text{Prob}_{\mu\otimes \frac{{\rm \text{tr}}}{d}}(|Z-\mathbb{E}Z|>t)\le 2 \exp\Big( -\frac{t^2}{64ec^2(v(Z)+M^2)}\Big)\hspace{.1cm}.\]
}
\end{exam}
We will now prove a second concentration inequality under more restrictive assumptions. Our starting point is a finite von Neumann algebra $({\mathcal M},\tau)$ and a normal faithful weight $\phi$ whose density operator has finite spectrum
\[ d_\phi \pl = \pl \sum_{k=1}^m \lambda_k e_k \hspace{.1cm} , \hspace{.1cm} \alpha_t^{\phi}(x)\pl = \pl d_\phi^{-it}xd_\phi^{it} \hspace{.1cm} ,\]
where $\lambda_k >0$ and $\sum {e_k}=1$. By re-scaling $\phi$, we can always assume $\lambda_k\ge 1$. Let $T_t=e^{-Lt}:{\mathcal M}\to{\mathcal M}$ be a $\phi$-symmetric quantum Markov semigroup.
We now analyze Haagerup's construction for a fixed $n\in \bN$.
We work with the subgroup $2^{-n}\mathbb{Z}\subset {\mathbb R}$ of the modular automorphism group $\alpha_t, t\in \mathbb{R}$. By the duality $L(\mathbb{Z})\cong L_{\infty}(\mathbb{T})$, the crossed product can be represented as
\[ {\mathcal M}\rtimes_\alpha 2^{-n}\mathbb{Z} \cong L_{\infty}(\mathbb{T},{\mathcal M})\hspace{.1cm} \]
with the embedding
\begin{align*}
&\pi:{\mathcal M} \to L_{\infty}(\mathbb{T},{\mathcal M})\hspace{.1cm}, \hspace{.1cm} \pi(x)=1\otimes x\\
&\lambda:2^{-n}\mathbb{Z} \to L_{\infty}(\mathbb{T},{\mathcal M})\hspace{.1cm}, \hspace{.1cm} \lambda(g)=\exp(i 2^ng s)\otimes d_\phi^{i g}
\end{align*}
Under this identification we write
\[ w(1):=\lambda(2^{-n}) = \exp(is)\otimes d_\phi^{\frac{i}{2^n}} \hspace{.1cm},\hspace{.1cm} s\in [0,2\pi] \]
as the generator of the group $\lambda(2^{-n}\mathbb{Z})$. Then the canonical extension state is \[\hat{\phi}(x)=\int_0^{2\pi} \phi(x(s)) \frac{ds}{2\pi}\hspace{.1cm}, \hspace{.1cm} x\in L_{\infty}(\mathbb{T},{\mathcal M}) \]
Note that the representation $\pi$ and $\lambda$ above is different with the one we used in \eqref{eq:cross}. Let us also recall the rest of the construction under this identification
\begin{align*}
b_n(s) &\pl = \pl -i \text{Log}(w(1))
\pl = \pl \sum_{k=1}^m [s+2\frac{\log \lambda_k}{2^{n}}] e_k
\hspace{.1cm} ,\\
a_n(s) & \pl = \pl 2^n\sum_{k=1}^m [s+\frac{\log \lambda_k}{2^{n}}] e_k\hspace{.1cm} ,\\
\psi_n(x) & \pl = \pl \hat{\phi}(e^{-a_n(s)}x) \hspace{.1cm} .
\end{align*}
where $[s]$ is the number in $[0,2\pi)$ such that $s-[s]=2\pi l$ for some $l\in \mathbb{Z}$.
The key point here is that for $n$ large enough we may assume that $0\le\frac{\log \lambda_k}{2^{n}}< \pi $ and hence
\begin{align}\label{eq:interval} a_n(s) \pl = \pl \sum_k (2^ns + \ln \lambda_k) e_k \hspace{.1cm} \hspace{.1cm} \text{for} \hspace{.1cm} 0\le s\le \pi\hspace{.1cm} .\end{align}
\begin{lemma}\label{cancel} Suppose $2^n\pi> \max\{\log \lambda_j\}$. Then for $x\in {\mathcal M}$,
\begin{enumerate}
\item[i)] $E_{{\mathcal M}_n}(x)(s)=x$ for $s\in [0,\pi]$;
\item[ii)] $\|E_{{\mathcal M}_n}(x)\|_{\text{Lip}}= \|x\|_{\text{Lip}}$;
\end{enumerate}
\end{lemma}
\begin{proof} It suffices to consider the element $x=e_kxe_j$. Then the cocyle chain rule tells us that for $s\in[0,\pi]$, we have by \eqref{eq:interval}
\begin{align*}
\alpha_t^{\psi_n}(x)(s) & \pl = \pl e^{ita_n}(s)
\alpha_t^{\hat{\phi}}(e_kxe_j)e^{-ita_n}(s) \\
&\pl = \pl e^{ita_n}(s)
\lambda_k^{it}\lambda_j^{-it}(e_kxe_j)e^{-ita_n}(s)
\end{align*}
Because $e_k$ and $e_k$ are projectors, they leave only terms in the Taylor expansion for $e^{i t a_n}(s)$ that contain only powers of themselves. Hence we may rewrite the sums in the exponential to obtain
\begin{align*}
\alpha_t^{\psi_n}(x)(s) &\pl = \pl \lambda_k^{it}\lambda_j^{-it}e^{-it2^n[s+2^{-n}\ln \lambda_k]}e_kxe_j e^{it2^n[s+2^{-n}\ln \lambda_j]}
\\
&\pl = \pl \lambda_k^{it}\lambda_j^{-it} \lambda_k^{-it}\lambda_j^{it} e^{-its2^n}e^{its2^n} e_kxe_j \pl = \pl e_kxe_j \hspace{.1cm} .
\end{align*}
For the second assertion, we recall that the upper estimate holds thanks to Lemma \ref{down}. For the other direction follows, we have by i) that for $s\in [0,\pi]$
\[ \Gamma_{\hat{L}}(E_{{\mathcal M}_n}(x),E_{{\mathcal M}_n}(x))(s)=\Gamma_{L}(E_{{\mathcal M}_n}(x)(s),E_{{\mathcal M}_n}(x)(s))=\Gamma_L(x,x)\hspace{.1cm},\]
Therefore
\begin{align*}
\norm{\Gamma_{\hat{L}}(E_{{\mathcal M}_n}(x),E_{{\mathcal M}_n}(x))}{}=\sup_{s\in [0,2\pi]} \norm{\Gamma_{L}(E_{{\mathcal M}_n}(x)(s),E_{{\mathcal M}_n}(x)(s))}{}\ge\norm{\Gamma_L(x,x)}{},
\end{align*}
which implies $\norm{E_{{\mathcal M}_n}(x)}{\text{Lip}}\ge \norm{x }{\text{Lip}}$.\end{proof}
\begin{theorem} Let ${\mathcal M}$ be a finite von Neumann algebra with finite trace $\tau$ and let $\phi$ be a faithful density operator with finite spectrum. Suppose $L$ is a $\phi$-symmetric generator satisfying CMLSI with parameter $\alpha$. For $2\le p <\infty$, we have
\[ \alpha\|x-E(x) \|_{L_{\infty}^p({\mathcal N}\subset{\mathcal M},\tau)}\pl \le \pl c \norm{x}{\text{Lip}} \hspace{.1cm} .\]
Note that here $\|x-E(x) \|_{L_{\infty}^p({\mathcal N}\subset{\mathcal M},\tau)}$ is the $L_\infty^p$ norm depending on the trace.
\end{theorem}
\begin{proof} Let $n$ be as in Lemma \ref{cancel}. Now, we consider the subalgebra
\[ {\mathcal M}_n \pl = \pl \{x| \alpha_t^{\psi_n}(x)=x\} \subset L_{\infty}(\mathbb{T},{\mathcal M}) \hspace{.1cm} .\]
By definition of CMLSI, $\iota_{\infty,2}^n_{\mathbb{T}}\otimes L$ also satisfies $\alpha$-CMLSI on $L_{\infty}(\mathbb{T}, {\mathcal M})$. By restriction we have $\alpha$-CMLSI for each ${\mathcal M}_n$. On ${\mathcal M}_n$ we have two natural traces. Indeed, the Haagerup construction leads to $\tau_n(x)\pl = \pl \psi_n(x)$ and this is compatible with the restriction of $\hat{\phi}$ to ${\mathcal M}_n$. The other trace is $\mu\otimes \tau$, where $\mu$ can be probability measure on $\mathbb{T}$. By the independence of tracial state for CMLSI, we use the tracial version \cite[Theorem 6.10]{gao2020fisher} for $\mu\otimes \tau$ restricted on ${\mathcal M}_n$ that
\[ \alpha\|x\|_{L_{\infty}^p({\mathcal N}_n\subset {\mathcal M}_n,\mu\otimes \tau)}
\pl \le \pl c_0 \sqrt{p} \|x\|_{Lip} \hspace{.1cm} .\]
Applying this to $E_{{\mathcal M}_n}(x)|_{[0,\pi]}=1_{[0,\pi]}\otimes x$ yields the assertion, as $E_{{\mathcal M}_n}(E(x))=\hat{E}E_{{\mathcal M}_n}(x)$ is also of the form $1_{[0,\pi]}\otimes E(x)$. \end{proof}\vspace{0.5ex}
\section{Entropy Contraction of Symmetric Markov maps}\label{sec:unital}
\subsection{States, channels and entropies}
We briefly review the definitions of states, channels and entropies in the noncommutative setting.
Recall that a von Neumann algebra ${\mathcal M}$ is a unital $*$-subalgebra of $B(H)$ closed under weak$^*$-topology. A state $\phi:{\mathcal M}\to\mathbb{C}$ is a unital positive linear functional satisfying $\phi(1)=1$ and positive $\phi(x^*x)\ge 0$ for any $x\in {\mathcal M}$. We say $\phi$ is normal if $\phi$ is weak$^*$-continuous. Throughout the paper, we will only consider normal state and denote $S({\mathcal M})$ as the normal state space of ${\mathcal M}$. We write $s(\phi)$ as the support projection of a state $\phi$, which is the minimal projection $e$ such that $\phi(x)=\phi(exe)\hspace{.1cm}, \forall \hspace{.1cm} x\in {\mathcal M}$. A normal state $\phi$ is faithful if $s(\phi)=1$. For two normal states $\rho$ and $\sigma$, the relative entropy
is defined as
\begin{align}D(\rho||\sigma)=\begin{cases}
\langle \xi_\rho| \log\Delta(\rho/\sigma) |\xi_\rho \rangle\hspace{.1cm}, & \mbox{if } \text{supp}(\rho)\le \text{supp}(\sigma)\\
+\infty, & \mbox{otherwise}.
\end{cases}\hspace{.1cm},\label{eq:araki}\end{align}
where $\xi_\rho$ is a (or any) vector implementing the state $\rho$ and $\Delta(\rho/\sigma)$ is the relative modular operator.
The above form of definition \eqref{eq:araki} was introduced by Araki \cite{araki1976relative} for general von Neumann algebras.
In this section, we will focus on the case ${\mathcal M}$ is a finite von Neumann algebra. Namely, ${\mathcal M}$ is equipped with a normal faithful tracial state $\tau$ satisfying $\tau(x^*x)> 0$ if $x\neq 0$, and $\tau(x^*x)=\tau(xx^*)$ for all $x\in {\mathcal M}$. The tracial $L_p$-space $L_p({\mathcal M},\tau)$ is defined as the completion of ${\mathcal M}$ with respect to the $p$-norm $\norm{a}{p}=\tau(|a|^{p})^{1/p}$. We identify $L_\infty({\mathcal M})\cong {\mathcal M}$, and also $ L_1({\mathcal M})\cong {\mathcal M}_*$ via the trace duality
\[d_\phi\in L_1({\mathcal M})\longleftrightarrow \phi\in {\mathcal M}_*,\hspace{.1cm} \phi(x)=\tau(d_\phi x)\hspace{.1cm}.\]
We say $\rho \in L_1({\mathcal M})$ is a density operator if $\rho\ge 0$ and $\tau(\rho)=1$, which corresponds to a normal state. For example, the identity element $1$ is a trivial density operator corresponding to the trace $\tau$ itself. We will often identify normal states with their density operators if no ambiguity. Via this identification, relative entropy reduces to Umegaki's definition \cite{umegaki1962conditional},
\[D(\rho||\sigma)=\tau(\rho\log \rho-\rho\log \sigma )\]
whenever the trace is well-defined. For example, for $\rho$ and $\sigma$ in the bounded state space
\begin{align*}
&S_{b}({\mathcal M})=\{\rho\in S({\mathcal M})\hspace{.1cm} |\hspace{.1cm} \mu_1 {\bf 1}\le \rho \le \mu_2 {\bf 1} \text{ for some } \mu_1,\mu_2>0 \}.
\end{align*}
the Umegaki's formula is always well-defined and finite. For this reason, we will mostly work with bounded state from $S_{b}({\mathcal M})$ and derive general case $S({\mathcal M})$ by approximation.
When $\sigma={\bf 1}$, we recover the entropy functional
\[H(\rho)=D(\rho||{\bf 1})=\tau(\rho\log\rho)\hspace{.1cm}.\] Note that our definition of entropy functional $H(\rho)$ differs from the standard notion of von Neumann entropy by a negative sign.
We say a linear map $T:{\mathcal M}\to {\mathcal M}$ is a quantum Markov map if $T$ is normal, unital and completely positive. The pre-adjoint map $T_*:{\mathcal M}_*\to {\mathcal M}_*$ is called a quantum channel, which sends normal states to normal states. In the tracial setting, $T_*:L_1({\mathcal M})\to L_1({\mathcal M})$ given by \[\tau(T_*(\rho)y)=\tau(\rho T(y))\hspace{.1cm}, \hspace{.1cm} \forall\hspace{.1cm} y\in {\mathcal M}, \rho\in L_1({\mathcal M})\] is completely positive and trace preserving (in short, CPTP). One fundamental inequality about quantum channel is the data processing inequality (also called monotonicity of relative entropy)
\begin{align} D(\rho||\sigma)\ge D(T_*(\rho)||T_*(\sigma))\hspace{.1cm}, \hspace{.1cm} \forall \rho,\sigma\in S({\mathcal M})\hspace{.1cm}. \label{eq:DPI} \end{align}
The data processing inequality states that two quantum states cannot become more distinguishable under a quantum channel. The main result of this work is essentially an improved data processing inequality for quantum channels under certain symmetric condition.
\subsection{Improved data processing inequality for unital quantum channel}\label{sec:unital}
We start our discussion on entropy decay of unital quantum channels. A quantum channel $\Phi:L_1({\mathcal M})\to L_1({\mathcal M})$ is unital if $\Phi(1)=1$. Then $\Phi$ restricted on ${\mathcal M}$ is bounded and normal, thus can be viewed as the $L_1$-norm continuous extension of its restriction $\Phi:{\mathcal M}\to{\mathcal M}$.
By duality, its adjoint $\Phi^*:{\mathcal M}\to {\mathcal M}$
is a trace preserving quantum Markov map hence also extends to a unital quantum channel. For a state $\rho$ with $H(\rho)<\infty$, we define the entropy difference of $\Phi$ as follow,
\[D_\Phi(\rho):=H(\rho)-H(\Phi(\rho))\hspace{.1cm}.\]
Note that $D_\Phi(\rho)\ge 0$ because $\Phi(1)=1$ and by data processing inequality \eqref{eq:DPI},
\[H(\rho)=D(\rho||1)\geq D(\Phi(\rho)||\Phi(1))= H(\Phi(\rho))\hspace{.1cm}.\]
We start with a simple lemma that is key in our argument.
\begin{lemma}[Entropy Difference Lemma]\label{lemma:difference} Let $\Phi:L_1({\mathcal M})\to L_1({\mathcal M})$ be a unital quantum channel and $\Phi^*$ be its adjoint. Then for two bounded states $\rho,\omega\in S_B({\mathcal M})$,
\[ D(\rho\|\Phi^*\Phi(\omega))
\le D_\Phi(\rho) + D(\rho\|\omega) \le \tau((\iota_{\infty,2}^n-\Phi^*\Phi)(\rho)\ln \rho ) + D(\rho\|\omega) \hspace{.1cm} .\]
\end{lemma}
\begin{proof} By duality, $\Phi^*$ is also completely positive unital.
Then,
\begin{align*}
D(\rho\|\Phi^*\Phi(\omega))=&\tau(\rho\ln \rho -\rho \ln \Phi^*\Phi(\omega))
\\ =\hspace{.1cm}&\tau(\rho\ln \rho-\Phi(\rho)\log \Phi(\rho))+\tau( \Phi(\rho)\log \Phi(\rho)-\rho \ln \Phi^*\Phi(\omega))
\\ =\hspace{.1cm}&D_\Phi(\rho)+\tau\big( \Phi(\rho)\log \Phi(\rho)-\rho \ln \Phi^*\Phi(\omega)\big)
\\ \overset{(1)}{\le}\hspace{.1cm} &D_\Phi(\rho)+\tau\Big( \Phi(\rho)\log \Phi(\rho)-\rho \Phi^*
\big(\ln \Phi(\omega)\big)\Big)
\\ = \hspace{.1cm}&D_\Phi(\rho)+\tau\Big( \Phi(\rho)\log \Phi(\rho)-\Phi(\rho)\ln \Phi(\omega)\Big)
\\ =\hspace{.1cm} &D_\Phi(\rho)+D(\Phi(\rho)\|\Phi(\omega))
\\ \overset{(2)}{\le}\hspace{.1cm} &D_\Phi(\rho)+D(\rho\|\omega)\hspace{.1cm},
\end{align*}
where (2) follows from the monotonicity of relative entropy.
The inequality (1) uses the operator concavity \cite{choi74} of logarithm function $t\mapsto \ln t$ that for any positive operator $x\ge 0$
\[\Phi^*
(\ln x)\le \ln \Phi^*(x)\hspace{.1cm} .\]
This proves the first inequality in the assertion. For the second part, it suffices to note that
\begin{align*}
D_\Phi(\rho)=&\tau(\rho\log \rho-\Phi(\rho)\log \Phi(\rho))
\le \tau(\rho\log \rho-\Phi(\rho)\Phi(\log \rho))
=\tau(\rho\log \rho-\Phi^*\Phi(\rho)\log \rho),
\end{align*}
where we use the operator concavity $\Phi(\ln x)\le \ln \Phi(x)$ again.
\end{proof}
Our idea is to iterate the above lemma as follows,
\begin{align*}D(\rho|| (\Phi^* \Phi)^n(\rho))\le\hspace{.1cm} D_\Phi(\rho)+D(\rho|| (\Phi^* \Phi)^{n-1}(\rho))
\le\hspace{.1cm} n D_\Phi(\rho)+D(\rho|| \rho)=n D_\Phi(\rho)
\end{align*}
Then a relevant question is what would be the limit of $(\Phi^* \Phi)^n(\rho)$ as $n \to\infty$. This is related to the multiplicative domain of $\Phi$. Recall that the multiplicative domain of a unital completely positive map $\Phi$ is
\[{\mathcal N}_\Phi=\{ x\in{\mathcal M} \ |\ \Phi(y) \Phi(x)=\Phi(yx)\hspace{.1cm},\hspace{.1cm} \Phi(x) \Phi(y)= \Phi(xy),\forall y\in\mathcal{M}\}\hspace{.1cm}.\]
When $\Phi$ is normal, ${\mathcal N}_\Phi$ is a von Neumann subalgebra. A linear map $E:{\mathcal M}\to {\mathcal M}$ is called a conditional expectation if $E$ is a unital completely positive map and idempotent $E\circ E=E$. When ${\mathcal M}$ is a finite von Neumann algebra, for any subalgebra ${\mathcal N}\subset {\mathcal M}$, there always exists a (unique) trace preserving condition expectation $E$ onto ${\mathcal N}$ defined by
\begin{align}\tau(xy)=\tau(xE(y))\hspace{.1cm}, \hspace{.1cm} x\in {\mathcal N}, y\in {\mathcal M}\hspace{.1cm}. \label{eq:condexp}\end{align}
In particular, $E$ is also a unital quantum channel.
\begin{prop}\label{prop:cd}Let $\Phi:L_1({\mathcal M})\to L_1({\mathcal M})$ be a unital quantum channel and let $E:{\mathcal M}\to {\mathcal N}$ be the trace preserving conditional expectation onto the multiplicative domain ${\mathcal N}:={\mathcal N}(\Phi)$. Then
\begin{enumerate}
\item[i)] $\Phi:{\mathcal N}\to\Phi({\mathcal N})$ is a $*$-isomorphism with inverse $\Phi^*:\Phi({\mathcal N})\to {\mathcal N}$. Moreover, $\Phi({\mathcal N})$ is the multiplicative domain for $\Phi^*$, and
\begin{align}\label{eq:cd2} (\Phi^* \Phi)\circ E=E\circ (\Phi^* \Phi)=E\hspace{.1cm}, \hspace{.1cm} E_0\circ \Phi=\Phi\circ E\hspace{.1cm} , \end{align}
where $E_0:{\mathcal M}\to \Phi({\mathcal N})$ is the trace preserving conditional expectation onto $\Phi({\mathcal N})$.
\item[ii)] $\Phi$ is a isometry on $L_2({\mathcal N})$. If in addition $\norm{\Phi(\iota_{\infty,2}^n-E):L_2({\mathcal M})\to L_2({\mathcal M})}{2}<1$, then $E=\lim_{n}(\Phi^* \Phi)^n$ as a map from $L_2({\mathcal M})$ to $L_2({\mathcal M})$ .
\end{enumerate}
\end{prop}
\begin{proof} It is clear that $\Phi$ is a $*$-homomorphism.
For any $x,y\in L_2({\mathcal N})\subset L_2({\mathcal M})$,
\[\tau( y(\Phi^*\circ \Phi)(x))=\tau( \Phi(y) \Phi(x))=\tau(\Phi(xy))=\tau(xy)\]
Thus $\Phi^*\circ \Phi|_{{\mathcal N}}=\iota_{\infty,2}^n_{\mathcal N}$ is the identity map. This verifies $(\Phi^* \Phi)\circ E=E$. Since $E^*=E$, $E\circ (\Phi^* \Phi)=E$ follows from taking adjoint. Thus $\Phi:{\mathcal N}\to\Phi({\mathcal N})$ is a $*$-isomorphism with inverse $\Phi^*$. Denote ${\mathcal N}_0$ as the multiplicative domain for $\Phi^*$, we have $\Phi({\mathcal N})\subset {\mathcal N}_0$. Conversely, we also have $\Phi^*(N_0)\subset {\mathcal N}$ by
switching the role of $\Phi=(\Phi^*)_*$. Then $\Phi({\mathcal N})=N_0$ since $\Phi$ is bijective on ${\mathcal N}$.
For ii), we note that by \eqref{eq:cd2}
\begin{align*}&(\iota_{\infty,2}^n-E)\Phi^* \Phi(\iota_{\infty,2}^n-E)=(\iota_{\infty,2}^n-E)(\Phi^* \Phi-E)=\Phi^* \Phi-E,\\
&(\Phi^* \Phi-E)^n=\Phi^* \Phi^n-E.
\end{align*}
Therefore,
\begin{align*}&\norm{\Phi^* \Phi-E:L_2({\mathcal M})\to L_2({\mathcal M})}{}=\norm{\Phi(\iota_{\infty,2}^n-E)}{2}^2<1,
\\
&\norm{(\Phi^* \Phi)^n-E:L_2({\mathcal M})\to L_2({\mathcal M})}{}=\norm{(\Phi^* \Phi-E)^n:L_2({\mathcal M})\to L_2({\mathcal M})}{}=\norm{\Phi(\iota_{\infty,2}^n-E)}{2}^{2n}
\end{align*}
which goes to $0$ as $n\to \infty$.
\end{proof}
\begin{defi}\label{defi:lambda}Let $\Phi$ be a unital quantum channel and $E$ be the trace preserving conditional expectation onto the multiplicative domain of $\Phi$. We define the $L_2$-contraction coefficient on the mean zero part as
\begin{align}\lambda(\Phi):=\norm{\Phi(\iota_{\infty,2}^n-E):L_2({\mathcal M})\to L_2({\mathcal M})}{}.\label{eq:PI}\end{align}
\end{defi}
The condition $\lambda(\Phi)<1$ can be viewed as a Poincar\'e inequality for quantum Markov map $\Phi$, which implies $(\Phi^* \Phi)^n(\rho)\to E(\rho)$ in $L_2$.
In order to estimate entropic quantities, we need a stronger approximation in the complete positivity order.
Recall that for a density operator $\sigma\in S({\mathcal M})$, the Bogoliubov-Kubo-Mori (BKM) metric for $X\in{\mathcal M}$ is defined by
\begin{align*}\gamma_{\sigma}(X)=\int_{0}^\infty \tau(X^*(\sigma+s)^{-1}X(\sigma+s)^{-1})ds\hspace{.1cm}.\end{align*}
The BKM metric is a Riemannian metric on the state space $S({\mathcal M})$ that is monotone under any quantum channel $\Psi$,
\[ \gamma_{\Psi(\sigma)}(\Psi(X))\le \gamma_{\sigma}(X)\hspace{.1cm}, \forall X\in {\mathcal M}\hspace{.1cm}. \]
It connects to the relative entropy as follows
\begin{align} D(\rho||\sigma)=\int_0^1 (1-t) \gamma_{\rho_t}(\rho-\sigma)dt\hspace{.1cm}, \label{eq:st}\end{align}
where $\rho_t=t\rho+(1-t) \sigma$ for $t\in [0,1]$.
It is proved in \cite[Lemma 2.1 \& 2.2]{gao2022complete} that if $\rho\le c\sigma$,
\begin{align} &c\gamma_{\rho}(X)\le \gamma_{\sigma}(X)\hspace{.1cm}, \hspace{.1cm} \forall X\in {\mathcal M} \nonumber\\ &k(c) \gamma_{\sigma}(\rho-\sigma) \le D(\rho||\sigma) \le \gamma_{\sigma}(\rho-\sigma) \label{eq:keylemma}
\end{align}
where $k(c)=\frac{c\ln c-c+1}{(c-1)^2}$. For two completely positive maps $\Psi$ and $\Phi$, we write $\Phi\le_{cp} \Psi $ if $\Psi-\Phi$ is again completely positive.
\begin{lemma}\label{lemma:approximate}
Let $E$ be a conditional expectation (not necessarily trace preserving) and $\Psi$ be a quantum Markov map such that
\[ (1-\varepsilon)E\le_{cp} \Psi \le_{cp} (1+\varepsilon)E\hspace{.1cm}.\]
Assume that $E\circ \Psi=E$. Then for any $\rho\in S({\mathcal M})$,
\[D(\rho||\Psi_*(\rho))\ge \Big(\frac{1-\varepsilon}{1+\varepsilon}-\frac{\varepsilon}{(1-\varepsilon)k(2)}\Big) D(\rho||E_*(\rho))\]
In particular, for $\varepsilon=\frac{1}{10}$, \[D(\rho||\Psi_*(\rho))\ge \frac{1}{2} D(\rho||E_*(\rho))\]
\end{lemma}
\begin{proof}By assumption, $\Psi_*=(1-\varepsilon)E_*+\varepsilon\Psi_0$ for some quantum channel $\Psi_0\le 2E_*$.
We denote $\sigma=E_*(\rho), \tilde{\sigma}=\Phi_*(\rho)$ and $\omega=\Psi_0(\rho)$. Then $\tilde{\sigma}=(1-\varepsilon)\sigma+\varepsilon \omega$.
Note that for any bounded state $\sigma\in S_B({\mathcal M})$, $X\mapsto \sqrt{\gamma_\sigma(X)}$ is a Hilbert space norm. Then by the triangle inequality,
\begin{align*}\sqrt{\gamma(\rho-\tilde{\sigma})}=&\sqrt{\gamma(\rho-(1-\varepsilon)\sigma-\varepsilon \omega)}\\
=&\sqrt{\gamma((\rho-\sigma)+\varepsilon (\sigma-\omega))}
\\
\ge &\sqrt{\gamma(\rho-\sigma)}-\varepsilon\sqrt{\gamma(\sigma-\omega)}
\end{align*}
where $\gamma$ can be $\gamma_\phi$ for any bounded state $\phi \in S_B({\mathcal M})$. Then
\begin{align*}\gamma(\rho-\tilde{\sigma})
\ge &\gamma(\rho-\sigma)-2\varepsilon\sqrt{\gamma(\rho-\sigma)}\sqrt{\gamma(\sigma-\omega)}+\varepsilon^2\gamma(\sigma-\omega)\\
\ge &\gamma(\rho-\sigma)-2\varepsilon\sqrt{\gamma(\rho-\sigma)}\sqrt{\gamma(\sigma-\omega)}\\
\ge &\gamma(\rho-\sigma)-\varepsilon\gamma(\rho-\sigma)-\varepsilon\gamma(\sigma-\omega)\\
= &(1-\varepsilon)\gamma(\rho-\sigma)-\varepsilon\gamma(\sigma-\omega)
\end{align*}
Now take $\rho_t=t\rho+(1-t)\sigma$ and $\tilde{\rho}_t=t\rho+(1-t)\tilde{\sigma}$,
\begin{align*}
D(\rho||\tilde{\sigma})=&\int_0^1 (1-t)\gamma_{\tilde{\rho}_t}(\rho-\tilde{\sigma})dt \\
\ge & (1-\varepsilon)\int_0^1 (1-t) \gamma_{\tilde{\rho}_t}(\rho-\sigma)dt-\varepsilon\int_0^1(1-t) \gamma_{\tilde{\rho}_t}(\sigma-\omega)dt
\end{align*}
For the first term, because $\tilde{\rho}_t\le (1+\varepsilon)\rho_t$,
\begin{align*}
\int_0^1 (1-t) \gamma_{\tilde{\rho}_t}(\rho-\sigma)dt
\ge (1+\varepsilon)^{-1}\int_0^1 (1-t) \gamma_{\rho_t}(\rho-\sigma)dtds=(1+\varepsilon)^{-1}D(\rho||\sigma)
\end{align*}
For the second term, consider that $\tilde{\rho}_t\ge (1-\varepsilon)(1-t)\sigma$
\begin{align*}
\int_0^1 (1-t)\gamma_{\tilde{\rho}_t}(\sigma-\omega)dt
\le & \frac{1}{(1-\varepsilon)}\int_0^1 \int_0^s \gamma_{\sigma}(\omega-\sigma)dt\\ =&\frac{1}{(1-\varepsilon)}\gamma_{\sigma}(\omega-\sigma)\\ \overset{(1)}{\le}& \frac{1}{k(2)(1-\varepsilon)}D(\omega||\sigma)
\\ =& \frac{1}{(1-\varepsilon)k(2)}D(\Psi_*(\rho)||\sigma)\overset{(2)}{\le} \frac{1}{(1-\varepsilon)k(2)}D(\rho||\sigma)\hspace{.1cm}.
\end{align*}
Here, the inequality (1) above uses $\omega\le 2\sigma$ and \eqref{eq:keylemma}. The inequality (2) above follows from the monotonicity of relative entropy
and the fact $\Psi_*(\sigma)=\sigma$.
Combined the estimated above, we obtained
\[ D(\rho||\tilde{\sigma})\ge \frac{1-\varepsilon}{1+\varepsilon} D(\rho||\sigma)-\varepsilon k(2)^{-1}D(\rho||\sigma)=\Big(\frac{1-\varepsilon}{1+\varepsilon}-\frac{\varepsilon}{(1-\varepsilon)k(2)}\Big)D(\rho||\sigma) \hspace{.1cm},\]
where $k(2)=2\ln 2-1$. The above inequality is non-trivial for $\varepsilon$ such that
\[\frac{1-\varepsilon}{1+\varepsilon}-\frac{\varepsilon}{(1-\varepsilon)k(2)}>0\hspace{.1cm}. \]
Taking $\varepsilon=0.1$, the above expression is approximately $0.53$.
\end{proof}
\begin{rem}{\rm We remark that the above lemma does not need any symmetric condition, which will be used in Section \ref{sec:GNSsymmetric} for GNS-symmetric channels.
This lemma is similar to \cite[Corollary 2.15]{laracuente2022quasi-factorization} and is a variant of \cite[Theorem 5.3]{gao2022complete}, which proves
\[D(\rho||(\Phi_*)^2(\rho))\ge \frac{1}{1-\varepsilon^2 k(2)^{-1}} D(\rho||E_*(\rho))\]
for GNS-symmetric channel.
}
\end{rem}
Putting the above lemmas together, we obtain the main theorem in this section.
\begin{theorem}\label{thm:unital} Let $\Phi$ be a unital quantum channel and let $E:{\mathcal M}\to{\mathcal N}$ be the trace preserving condition expectation onto the multiplicative domain of $\Phi$. Define the CB return time
\begin{align}\label{eq:cbreturn} k_{cb}(\Phi):=\inf \{k\in \mathbb{N}^+\hspace{.1cm} |\hspace{.1cm} 0.9E \le_{cp} (\Phi^*\Phi)^{k}\le_{cp} 1.1E \} \end{align}
Then for $k=k_{cb}(\Phi)$ and any state $\rho\in S({\mathcal M})$,
\begin{align} D(\Phi(\rho)||\Phi\circ E(\rho))\le \Big (1-\frac{1}{2k} \Big ) D(\rho||E(\rho))\label{eq:SDPI}\end{align}
Furthermore, for any finite von Neumann algebra $\mathcal{Q}$ and state $\rho\in S(\mathcal{Q}\overline{\otimes}{\mathcal M}))$
\begin{align}\label{eq:CSDPI} D(\iota_{\infty,2}^n_{\mathcal Q}\otimes \Phi(\rho)||\iota_{\infty,2}^n_{\mathcal Q}\otimes(\Phi\circ E)(\rho))\le \Big (1-\frac{1}{2k} \Big ) D(\rho||\iota_{\infty,2}^n_{\mathcal Q}\otimes E(\rho))\hspace{.1cm}.\end{align}
\end{theorem}
\begin{proof} It suffices to consider bounded state $\rho \in S_B({\mathcal M})$. Note that by the condition expectation property
\eqref{eq:condexp}
\begin{align*}&D(\rho||E(\rho))=\tau(\rho\log \rho-\rho\log E(\rho))=\tau(\rho\log \rho)-\tau(E(\rho)\log E(\rho))=H(\rho)-H(E(\rho))\hspace{.1cm},\\
&D(\Phi(\rho)||\Phi\circ E(\rho))=D(\Phi(\rho)||E_0\circ \Phi(\rho))=H(\Phi(\rho))-H(E \circ \Phi(\rho)) =H(\Phi(\rho))-H(\Phi\circ E(\rho))\hspace{.1cm},
\end{align*}
where we used the property $\Phi\circ E=E \circ \Phi$ from Proposition \ref{prop:cd}. Moreover, $H(E(\rho))=H(\Phi\circ E(\rho))$ as $\Phi$ is a trace preserving $*$-isomorphism on ${\mathcal N}$. Thus we have
\[D_\Phi(\rho)=H(\rho)-H(\Phi(\rho))=D(\rho||E(\rho))-D(\Phi(\rho)||\Phi \circ E(\rho))\hspace{.1cm}.\]
Iterating the entropy difference Lemma \ref{lemma:difference}, we have
\begin{align*}D(\rho\|(\Phi^*\Phi)^{k}(\rho))
\pl = \pl & D_\Phi(\rho) + D(\rho\|(\Phi^*\Phi)^{k-1}(\rho))\\
\pl = \pl & k D_\Phi(\rho) + D(\rho\|\rho)\\
=\hspace{.1cm} & k (D(\rho||E(\rho))-D(\Phi(\rho)||\Phi \circ E(\rho))
\end{align*}
Now using Lemma \ref{lemma:approximate} above, for $k=k_{cb}(\Phi)$,
\[ D(\rho||E(\rho))\le 2 D(\rho\|(\Phi^*\Phi)^{n}\rho))\le 2k\big (D(\rho||E(\rho))-D(\Phi(\rho)||\Phi \circ E(\rho))\big)\hspace{.1cm}. \]
Rearranging the term gives the assertion. The general case $\rho\in S({\mathcal M})$ can be obtained via approximation using \cite[Lemma A.2]{brannan2022complete}. The same argument applies to $\iota_{\infty,2}^n_{\mathcal Q}\otimes \Phi$ as $k_{cb}(\Phi)=k_{cb}(\iota_{\infty,2}^n_{\mathcal Q}\otimes \Phi)$ by the definition of complete positivity.
\end{proof}\vspace{0.5ex}
The above theorem is an improved data processing inequality for the relative entropy $D(\rho||E(\rho))$ between a state $\rho$ and its conditional expectation $E(\rho)$. Here ${\mathcal N}$ is the ``decoherence free'' subalgebra, because
for any two states $\sigma_1,\sigma_2\in{\mathcal N}$
\[D(\sigma_1||\sigma_2)\ge D(\Phi(\sigma_1)||\Phi(\sigma_2))\ge D(\Phi^*\Phi(\sigma_1)||\Phi^*\Phi(\sigma_2))=D(\sigma_1||\sigma_2)\]
does not decay.
Outside the ``decoherence free'' subalgebra ${\mathcal N}$, the relative entropy from a state $\rho$ to its projection $E(\rho)$ on ${\mathcal N}$ is always strictly contractive under every use of the channel $\Phi$.
\begin{defi}\label{defi:alpha}Let $\Phi$ be a unital quantum channel and $E$ be the trace preserving conditional expectation onto the multiplicative domain of $\Phi$. We define the entropy contraction coefficient $\alpha$ as
\[\alpha(\Phi)=\sup_{\rho\notin E(\mathcal{M})} \frac{D(\Phi(\rho)||\Phi\circ E(\rho))}{D(\rho||E(\rho))}\hspace{.1cm}.\]
and $\alpha_c(\Phi):=\sup_{\mathcal{Q}}\alpha(\iota_{\infty,2}^n_{\mathcal Q}\otimes\Phi)$ as the complete bounded constant where the supremum is over all finite von Neumann algebra ${\mathcal Q}$.
\end{defi}
By definition, $\alpha_c(\Phi)$ controls not only the entropy decay of $\Phi$ but also $\iota_{\infty,2}^n_{\mathcal Q}\otimes\Phi$ with any environment system $\mathcal{Q}$. This property leads to the tensorization property of $\alpha_c$ that for two unital quantum channels,
\begin{align}\alpha_c(\Phi_1\otimes \Phi_2)=\max\{\alpha_c(\Phi_1),\alpha_c(\Phi_2)\}\label{eq:tensorizationchannel}\end{align}
For classical channels (Markov maps), the tensorization property \eqref{eq:tensorizationchannel} is also known to hold for the non complete constant $\alpha$. Nevertheless, in the noncommutative setting it was observed in \cite[Section 4.4]{brannan2022complete} that the tensorization of $\alpha$ fails for quantum channels, and in general, $\alpha(\Phi)$ can be strictly less than $\alpha_c(\Phi)$. From this perspective, Theorem \ref{thm:unital} can be reformulated as an upper bound of $\alpha_c(\Phi)$ using CB return time.
\begin{cor}\label{cor:CSDPI} For a unital quantum channel $\Phi$,
\[\alpha(\Phi)\le \alpha_c(\Phi)\le (1-\frac{1}{2k_{cb}(\Phi)})\hspace{.1cm}.\]
\end{cor}
\begin{rem}\rm{Complete positivity of $\Phi$ is not really needed in the proof of Proposition \ref{prop:cd}. Choi (\cite{choi74}) proved the operator concavity of logarithmic function $x\mapsto \ln x$ that
\begin{align*}
\Phi(\ln x)\leq \ln \Phi(x),\quad \forall x\geq 0
\end{align*}
for any unital positive map $\Phi$. Also,
the monotonicity of relative entropy
\begin{align*}
D(\rho\|\sigma)\geq D(\Phi(\rho)\|\Phi(\sigma))
\end{align*}
was generalized to any positive trace preserving map $\Phi$ in \cite{MR17}. Then we can relax the conditions in Lemma \ref{lemma:approximate} to $\Psi$ being normal unital positive and
\[ (1-\varepsilon)E\le\Psi \le (1+\varepsilon)E\hspace{.1cm}.\]
Therefore, we obtain \eqref{eq:SDPI} in Theorem \ref{thm:unital} for
any unital trace preserving positive $\Phi$ for the positivity return time
\begin{align}\label{eq:cbreturn} k(\Phi):=\inf \{k\in \mathbb{N}^+\hspace{.1cm} |\hspace{.1cm} 0.9E \le (\Phi^*\Phi)^{k}\le 1.1E \}. \end{align}
The complete positivity and the CB return time $k_{cb}$ are only required to obtain estimate of $\alpha_c(\Phi)$ in \eqref{eq:CSDPI}. The advantage using $k_{cb}$ instead of $k$ is connection to completely bounded norm discussed below.
}
\end{rem}
\subsection{CB-return time}
We now consider a common scenario that the CB return time of $\Phi$ is finite. Despite defined using CP order \eqref{eq:cbreturn}, the CB-return time can be characterized using completely bounded norm (which was the original motivation for the name).
Recall that $\Psi:{\mathcal M}\to{\mathcal M}$ is called a ${\mathcal N}$-bimodule map if
\[\Phi(axb)=a\Phi(x)b\hspace{.1cm}, \hspace{.1cm} \forall\hspace{.1cm} a,b\in {\mathcal N}, x\in {\mathcal M}\]
\begin{prop}\label{prop:equivalence}
Let ${\mathcal N}\subset{\mathcal M}$ be a subalgebra and $E:{\mathcal M}\to{\mathcal N}$ be the trace preserving conditional expectation. Let $\Psi:{\mathcal M}\to{\mathcal M}$ be a ${\mathcal N}$-bimodule $*$-preserving map. For any $\varepsilon>0$, the following two conditions are equivalent:
\begin{enumerate}\item[i)]$(1-\varepsilon) E\le_{cp} \Psi \le_{cp} (1+\varepsilon) E$;
\item[ii)] $\norm{\Psi-E:L_\infty^1({\mathcal N}\subset{\mathcal M})\to L_\infty({\mathcal M})}{cb}\le \varepsilon$.
\end{enumerate}
\end{prop}
The condition ii) above is on the completely bounded norm from the $L_\infty^1({\mathcal N}\subset{\mathcal M})$ space to ${\mathcal M}$. The $L_\infty^1({\mathcal N}\subset{\mathcal M})$ is called a conditional $L_\infty$ space, defined as the completion of ${\mathcal M}$ with respect to the norm
\[ \norm{x}{L_\infty^1({\mathcal N}\subset{\mathcal M})}=\sup_{a,b\in{\mathcal N} \hspace{.1cm} ,\hspace{.1cm} \|a\|_2=\|b\|_2=1}\norm{axb}{1}\hspace{.1cm}, \]
where the supremum takes over all $a,b\in L_2({\mathcal N})$ with $\|a\|_2=\|b\|_2=1$.
The operator space structure of $L_\infty^1({\mathcal N}\subset{\mathcal M})$ is given by the identification
\[{\mathbb M}_n(L_\infty^1({\mathcal N}\subset{\mathcal M}))=L_\infty^1({\mathbb M}_n({\mathcal N})\subset {\mathbb M}_n({\mathcal M}))\hspace{.1cm}.\]
(see \cite{junge2010mixed} and \cite[Appendix]{gao2020relative}). In particular, for trivial ${\mathcal N}=\mathbb{C}1$, $L_\infty^1({\mathcal N}\subset{\mathcal M})\cong L_1({\mathcal M})$, and
Proposition \ref{prop:equivalence} is self-evident, which we illustrate below.
\begin{exam}[Classical case]{\rm Let $(\Omega,\mu)$ be a probability space. Consider $P:L_\infty(\Omega)\to L_\infty(\Omega)$ be a linear map with kernel $P(f)(x)=\int_{\Omega}k(x,y)f(y)d\mu(y)$. It is clear that $P$ is $*$-preserving i.e. $P(\bar{f})=\overline{P(f)}$ if $k$ is real; $P$ is a positive map if and only if the kernel function $k\ge 0$. Recall that the (conditional) expectation map $E_\mu:L_\infty(\Omega)\to \mathbb{C}{\bf 1} $ is $E(f)=(\int_{}f\mu){\bf 1}$, where ${\bf 1}$ is the unit constant function. The kernel of $E_\mu$ is the constant function ${\bf 1}$ on w.r.t to the product measure $\mu\times \mu$ on $\Omega\times \Omega$. We have the following equivalence
\begin{align*} &(1-\varepsilon)E\le P \le (1+\varepsilon)E
\\ \Longleftrightarrow\hspace{.1cm} &\varepsilon E\le P-E \le \varepsilon E
\\ \Longleftrightarrow\hspace{.1cm} &\varepsilon {\bf 1}\le k-{\bf 1} \le \varepsilon {\bf 1}
\\ \Longleftrightarrow\hspace{.1cm} &\norm{k-{\bf 1}}{L_\infty(\Omega\otimes \Omega)}\le \varepsilon
\\ \Longleftrightarrow\hspace{.1cm} &\norm{P-E:L_1(\Omega)\to L_\infty(\Omega)}{}\le \varepsilon
\end{align*}
To see the equivalence in Proposition \ref{prop:equivalence}, it suffices to note that every positive (resp. bounded) map to $L_\infty(\Omega)$ is automatically completely positive (resp. completely bounded with same norm \cite{smith1983completely}).
}
\end{exam}
The above simple argument also applies to the noncommutative setting with ${\mathcal N}=\mathbb{C}1\subset{\mathcal M}$. The correspondence between the map $P$ and its kernel $k$ generalizes to the isomorphism between the map $T$ and its Choi operator $C_T\in {\mathcal M}^{op}\overline{\otimes}{\mathcal M}$
\[T(x)=\tau \otimes \iota_{\infty,2}^n (C_T(x\otimes 1))\hspace{.1cm},\hspace{.1cm} a\in L_1({\mathcal M})\cong ({\mathcal M}^{op})_*,\]
where ${\mathcal M}^{op}$ is the opposite algebra of ${\mathcal M}$. Here we requires completely bounded norm instead of simply bounded norm because of Effros-Ruan Theorem (see \cite{effros2000operator,blecher1991tensor}),
\[\norm{T:L_1({\mathcal M})\to L_\infty({\mathcal M})}{cb}=\norm{C_T}{{\mathcal M}^{op}\overline{\otimes}{\mathcal M}}\hspace{.1cm}.\]
For a ${\mathcal N}$-bimodule map $T$ of nontrivial ${\mathcal N}$, the above identity holds with more involved module Choi operator, which
we refer to \cite[Lemma 5.1]{bardet2021entropy} and \cite[Lemma 3.15]{gao2020fisher} for the complete proof of Proposition \ref{prop:equivalence}.
With Proposition \ref{prop:equivalence}, the CB-return time can be equivalently defined as
\begin{align}\label{eq:cbreturn2} k_{cb}(\Phi):=\inf \{k\in \mathbb{N}^+\hspace{.1cm} |\hspace{.1cm} \norm{(\Phi^*\Phi)^k-E:L_\infty^1({\mathcal N}\subset{\mathcal M})\to L_\infty({\mathcal M})}{cb}\le 0.1 \} \end{align}
Here we choose $\varepsilon=0.1$ instead of $\varepsilon=1/2$ in \cite{gao2020fisher}. This is of course motivated from the calculation in Lemma \ref{lemma:approximate}. The next proposition shows that the cb-return time is finite whenever $\Phi$ satisfies the Poincar\'e inequality and ultra-contractivity.
\begin{prop}\label{prop:finitecb}Let $\Phi:L_1({\mathcal M})\to L_1({\mathcal M})$ be a unital quantum channel and ${\mathcal N}$ be the multiplicative domain of $\Phi$. Denote $E:{\mathcal M}\to{\mathcal N}$ as the trace preserving conditional expectation. Suppose \begin{enumerate} \item[i)] $\Phi$ satisfies the Poincar\'e inequality
\[ \lambda(\Phi):=\norm{\Phi(\iota_{\infty,2}^n-E):L_2({\mathcal M})\to L_2({\mathcal M})}{}<1\hspace{.1cm}.\]
\item[ii)] There exists $k_0$ such that $\norm{(\Phi^*\Phi)^{2k_0}-E: L_\infty^1({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}\le C_0$.
\end{enumerate}
Then
\[ k_{cb}(\Phi)\pl \le \pl \frac{\ln (10 C_0)}{-2\ln \lambda}+2k_0 \hspace{.1cm} .\]
\end{prop}
\begin{proof}Denote $\Psi=\Phi^*\Phi$ Then
\[ (\iota_{\infty,2}^n-E)\Phi^*\Phi(\iota_{\infty,2}^n-E)=\Psi-E\hspace{.1cm}, \norm{\Psi-E:L_2({\mathcal M})\to L_2({\mathcal M})}{}=\lambda(\Phi)^2\hspace{.1cm}.\]
Moreover, by Proposition \ref{prop:cd} $\Psi$ is a ${\mathcal N}$-bimodule map, we have by \cite[Lemma 3.13]{gao2020fisher}
\begin{align*}\norm{\Psi^{2k_0}: L_\infty^1({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}=&\norm{\Psi^{k_0}: L_\infty^1({\mathcal M})\to L_\infty^2({\mathcal N}\subset {\mathcal M})}{cb}^2\\ =&\norm{\Psi^{k_0}: L_\infty^2({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}^2\hspace{.1cm}.\end{align*}
Then for $k\ge 2k_0$,
\begin{align*}
&\norm{\Psi^{k}-E: L_\infty^1({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}\\ \le & \norm{\Psi^{k_0}: L_\infty^1({\mathcal M})\to L_\infty^2({\mathcal N}\subset {\mathcal M})}{cb}\cdot \norm{\Psi^{k-2k_0}-E: L_\infty^2({\mathcal N}\subset {\mathcal M})\to L_\infty^2({\mathcal N}\subset {\mathcal M})}{cb}\\ &\cdot
\norm{\Psi^{k_0}: L_\infty^2({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}
\\ \le & \norm{\Psi^{2k_0}: L_\infty^1({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}\norm{(\Psi-E)^{k-2k_0}: L_2({\mathcal M})\to L_2({\mathcal M})}{cb}
\\ =&\lambda(\Phi)^{2(k-2k_0)}C_0\hspace{.1cm}.
\end{align*}
Here we use the property \cite[Lemma 3.12]{gao2020fisher} that for a ${\mathcal N}$-bimodule $T$,
\[ \norm{T: L_\infty^2({\mathcal N}\subset {\mathcal M})\to L_\infty^2({\mathcal N}\subset {\mathcal M})}{cb}= \norm{T: L_2({\mathcal N}\subset {\mathcal M})\to L_2({\mathcal N}\subset {\mathcal M})}{cb}\hspace{.1cm}.\]
Then assertion follows from taking $\lambda(\Phi)^{2(k-2k_0)}C_0\le \frac{1}{10}$.
\end{proof}
It was proved in \cite[Theorem 3.9]{gao2020relative} that
\[ \norm{\iota_{\infty,2}^n: L_\infty^1({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}=C_{cb}(E)\hspace{.1cm},\]
where $C_{cb}(E)$ is the complete version of von Neumann subalgebra index introduced by Popa and Pimsner \cite{pimsner1986entropy}
\[ C_{cb}(E):=\sup_{n}C(\iota_{\infty,2}^n_{{\mathbb M}_n}\otimes E)\hspace{.1cm}, \hspace{.1cm} C(E):=\{\mu>0 \hspace{.1cm} |\hspace{.1cm} \rho\le \mu E\rho\hspace{.1cm}, \hspace{.1cm} \forall \rho\in {\mathcal M}_{+}\}\]
For infinite dimensional ${\mathcal M}$, $C(E)$ and $C_{cb}(E)$ can be infinite, and the condition \[\norm{(\Phi^*\Phi)^{2n_0}: L_\infty^1({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}\le C_0\] can be viewed as an ultra-contractive type estimate since $L_\infty^1({\mathcal N}\subset {\mathcal M})$ is a generalization of $L_1({\mathcal M})$.
When ${\mathcal M}$ is of finite dimensions, both the index $C(E)$ and $C_{cb}(E)$ are finite and admits explicit formula in \cite[Theorem 6.1]{pimsner1986entropy}. Note that in finite dimensions, we always have $1-\lambda(\Phi)^2>0$ and $\lambda(\Phi)<1$, because $1-\lambda(\Phi)^2$ is spectral gap of the symmetric map $\Phi^*\Phi$.
\begin{theorem}\label{thm:finitecb}Let $\Phi:{\mathcal M}\to{\mathcal M}$ be a unital quantum channel on a finite von Neumann algebra ${\mathcal M}$ and $E$ is the trace preserving conditional expectation onto its multiplicative domain. Then
\[\alpha(\Phi)\le \alpha_c(\Phi)\le (1-\frac{-\ln \lambda(\Phi)}{\ln (10 C_{cb}(E))})\hspace{.1cm}.\]
\end{theorem}
\begin{proof}Taking $k_0=0$ in Proposition \ref{prop:finitecb}.
\end{proof}
\begin{rem}{\rm In the primitive case ${\mathcal N}=\mathbb{C}1$ and ${\mathcal M}={\mathbb M}_d$,
we have the trace map $E(x)={\rm \text{tr}}(x)\frac{1}{d}$ and the CB-index $C_{cb}(E)=d^2$.
The above estimate implies
\begin{align}\label{eq:CSDPId} \alpha(\Phi)\le \alpha_c(\Phi)\le (1-\frac{-\ln \lambda(\Phi)}{\ln (10 d^2)})\end{align}
On the other hand, by
\cite[Theorem 4.2]{muller2016entropy} and \cite[Corollary 27]{kastoryano2013quantum},
\begin{align}\label{eq:SDPId} \alpha(\Phi)\le 1-\frac{1}{2}\alpha_{2}(\iota_{\infty,2}^n-\Phi^*\Phi)\le 1-\frac{(1-\lambda(\Phi)^2)^2(1-\frac{2}{d})}{\ln(d-1)}\hspace{.1cm},\end{align}
where $\alpha_{2}(\iota_{\infty,2}^n-\Phi^*\Phi)$ is the log-Sobolev constant of $\iota_{\infty,2}^n-\Phi^*\Phi$ as a generator of quantum Markov semigroup, which will be defined in Section \ref{sec:symmetricCMLSI}. The two upper bounds in \eqref{eq:CSDPId} and \eqref{eq:SDPId} are comparable as $-\ln \lambda(\Phi)\ge 1-\lambda(\Phi)^2$, and both the dimension constant $\frac{(1-\frac{2}{d})}{\ln(d-1)}$ and $\frac{1}{\ln (10 d^2)}$ are asymptotically $\Theta(\frac{1}{\ln d})$. Nevertheless, our results \ref{thm:finitecb} \eqref{eq:CSDPId} is an improvement as it bounds the complete bounded constant $\alpha_c(\Phi)$ which has tensorization property in noncommutative setting.
}
\end{rem}
\begin{rem}{\rm Compared to \eqref{eq:SDPId}, one has
\[ 1-\alpha_1(\iota_{\infty,2}^n-\Phi^*\Phi)\le \alpha(\Phi),\]
where $\alpha_1$ is the entropy decay rate of the quantum Markov semigroup generated by $(\iota_{\infty,2}^n-\Phi^*\Phi)$. (See Section \ref{sec:CMLSI} the formal definition of $\alpha_1$.)
This follows from the inequality
\[D_\Phi(\rho)\le \tau\big((\iota_{\infty,2}^n-\Phi^*\Phi)(\rho)\ln\rho\big)\hspace{.1cm},\]
proved that in Lemma \ref{lemma:difference}. For classical Markov map, it was proved by Del Moral, Ledoux and Miclo \cite{del2003contraction} that there exists an universal constant $0<c<1$ such that
\begin{align}\label{eq:miclo} 1-\alpha_1(\iota_{\infty,2}^n-\Phi^*\Phi)\le \alpha(\Phi)\le 1-c\alpha_1(\iota_{\infty,2}^n-\Phi^*\Phi)\hspace{.1cm}.\end{align}
To the best of our knowledge, the upper bound in \eqref{eq:miclo} is open for quantum cases.
}
\end{rem}
\begin{exam}[Graph random walk]{\rm Let $G=(V,E)$ be a finite undirected graph with $|V|=d$ and the edge set $E\subset V\times V$. The random walk on $G$ is a finite Markov chain given by the stochastic matrix
\[K_G(u,v)=\begin{cases}
\frac{1}{d(u)}, & \mbox{if } (u,v)\in E \\
0, & \mbox{otherwise}.
\end{cases}\]
Here $d(u)$ is the degree of vertex $u\in V$. Then $K_G:l_\infty(V)\to l_\infty(V)$ is a Markov map.
If $G$ is not bipartite, $K_G$ admits a unique station distribution $\pi(u)=\frac{d(u)}{2m}$. It is clear that $K_G$ is symmetric to the measure $\pi$, also called time reversible. Hence $K_G$ is a primitive unital channel on $L_\infty(V,\pi)$ as $\pi(K_G(f))=\pi(f)$. The expectation map is $E_\pi(f)=\pi(f)1$ whose index is
\[ C_{cb}(E_\pi)=\norm{\pi^{-1}}{\infty}\]
Both $K_G$ and $E_\pi$ are symmetric operator on $L_2(V,\pi)$ and \[\lambda(G)=\norm{K_G-E_\pi:L_2(V,\pi)\to L_2(V,\pi)}{}<1\] if $K_G$ not bipartite (in the bipartite case $K_G$ has eigenvalue $-1$). Then our results implies
\[\alpha(K_G)\le \alpha_c(K_G)\le (1-\frac{1}{2k_{cb}(K_G)})\le (1-\frac{-\ln \lambda(K_G)}{\ln (10 \norm{\pi^{-1}}{\infty})})\hspace{.1cm}. \]
}
\end{exam}
\begin{exam}[Cyclic graphs]{\rm \label{ex:cyclic} Let us consider the cyclic graph $C_d=(V,E)$ with $d\ge 4$ where $V=\{1,\cdots,d\}$ and $E=\{(j,j+1)| j=1,\cdots, d \}$. Here the addition is understood in the sense of "mod d".
Then
\[ K_{C_d}(i,j)=\begin{cases}
\frac{1}{2}, & \mbox{if } |i-j|=1 \\
0, & \mbox{otherwise}.
\end{cases}\]
As $C_d$ is 2 regular, $K_{C_d}$ is symmetric to the uniform distribution $\pi(i)=1/d$. It is known that $K_{C_d}$ has spectrum \[\lambda_j=\cos(\frac{2\pi j}{d})\hspace{.1cm} ,\hspace{.1cm} j=0,\cdots, d-1\hspace{.1cm} .\]
The associated eigenvector is $e_j=\frac{1}{\sqrt{d}}(1,\omega^j,\omega^{2j},\cdots, \omega^{(d-1)j})$ where $\omega=\exp(\frac{2\pi i}{d})$.
When $d=2m+1$ is odd, $\pi$ is the unique stationary measure and $E_\pi$ is the projection onto the vector $e_0$. We have
\begin{align*} K_G^k-E_\pi= (K_G-E_\pi)^k=\sum_{j=1}^{2m}\lambda_j^k \ketbra{e_j}
\end{align*}
By triangle inequality, we have
\begin{align*} \norm{K_G^k-E_\pi:L_1(V,\pi)\to L_\infty(V,\pi)}{}&\le \sum_{j=1}^{2m}|\lambda_j|^k= 2\sum_{j=1}^{m}\cos(\frac{\pi j}{d})^k\\ &\le 2\frac{d}{\pi}\int_{0}^{\pi/2}\cos^k(x)dx
=2\frac{d}{\pi}W_k\le 2Cd\sqrt{\frac{1}{2k\pi }}
\end{align*}
where $C>0$ is some absolute constant by fact that the Wallis integrals $W_k=\int_{0}^{\pi/2}\cos^k(x)dx\sim \sqrt{\frac{\pi}{2k}}$. Thus
\[k_{cb}(K_{C_d})\le \frac{(10Cd)^2}{\pi}\sim \mathcal{O}(d^2) \]
And
\[ \alpha(K_{C_d})\ge \alpha_c(K_{C_d})\ge 1-\mathcal{O}(d^{-2})\]
By Miclo's result \eqref{eq:miclo}, this is asymptotically tight because the MLSI constant $\alpha_1(I-K_G^2)\sim \mathcal{O}(d^{-2})$ (see Example \ref{ex:cyclic2} below for detials).
The similar asymptotic estimate also holds for even circle $d=2m$.
}
\end{exam}
\subsection{Entropy convergence of symmetric quantum Markov map}
For quantum Markov maps on a finite von Neumann algebra ${\mathcal M}$, we can define trace symmetric condition.
This includes classical Markov maps satisfies the so-called detailed balance condition to a probability measure.
\begin{defi}{\rm
We say a quantum Markov map $\Phi$ is trace-symmetric (or simply, symmetric) if for all $x,y\in{\mathcal M}$, \[\tau(\Phi(x)y)=\tau(x\Phi(y))\hspace{.1cm}, \hspace{.1cm} \forall \hspace{.1cm} x,y\in{\mathcal M}\hspace{.1cm}. \] }
\end{defi}
A symmetric quantum Markov map $\Phi$ coincides with its pre-adjoint $\Phi=\Phi_*$ hence is always a unital quantum channel. Thus, all our results on unital quantum channels apply to symmetric quantum Markov maps. Because $\Phi=\Phi_*$, Proposition \ref{prop:cd} reduces to
\[\Phi\circ E=E\circ \Phi, \Phi^2\circ E= E\circ \Phi^2=E\hspace{.1cm}.\]
In particular, $\Phi$ is an involutive isomorphism on the multiplicative domain ${\mathcal N}$. Thus, we can iterate the entropy decay in Theorem \ref{thm:unital}.
\begin{align*} &D(\Phi^n(\rho)||\Phi^n\circ E(\rho))\le \alpha(\Phi)^n D(\rho||E(\rho))\\
&D(\Phi^{2n}(\rho)|| E(\rho))\le \alpha(\Phi)^{2n} D(\rho||E(\rho))
\end{align*}
This implies that \[\lim_{n\to \infty}\Phi^n-\Phi^n\circ E=0\hspace{.1cm}, \hspace{.1cm} \lim_{n\to \infty}\Phi^{2n}=E\] in terms of relative entropy. Recall that the diamond norm between quantum channels is defined as the CB norm from $L_1({\mathcal M})\to L_1({\mathcal M})$,
\begin{align*}\norm{\Phi-\Psi}{\diamond}:=&\norm{\Phi-\Psi:L_1({\mathcal M})\to L_1({\mathcal M})}{cb}\\ =&\sup_n\norm{\iota_{\infty,2}^n_{{\mathbb M}_n}\otimes \Phi-\iota_{\infty,2}^n_{{\mathbb M}_n}\otimes\Psi:L_1({\mathbb M}_n({\mathcal M}))\to L_1({\mathbb M}_n({\mathcal M}))}{}\end{align*}
In finite dimensions, it is proved in \cite{gao2020relative} that \[ \sup_{\rho\in S({\mathbb M}_n({\mathcal M}))}D(\rho||\iota_{\infty,2}^n_{{\mathbb M}_n}\otimes E(\rho))= C_{cb}(E),\]
where the supremum is over all states $\rho\in S({\mathbb M}_n({\mathcal M}))$ and $n\in {\mathbb N}^+$.
Then by the Pinsker inequality (see e.g. \cite{watrous2018theory})
\[D(\rho\|\sigma)\ge \frac{1}{2}\norm{\rho-\sigma}{1}^2\hspace{.1cm}, \]
we have
\begin{align}\label{eq:decaydiamond}\norm{\Phi^{2k}-E}{\diamond}\le \alpha_c(\Phi)^k\sqrt{2\sup_{\rho } D(\rho||\iota_{\infty,2}^n_\mathcal{R}\otimes E(\rho))}= \alpha_c(\Phi)^k\sqrt{2\ln C_{cb}(E)}\hspace{.1cm}.\end{align}
Define the diamond norm mixing time
\[ k_{\diamond}(\Phi)=\min \{ k \in \mathbb{N}^+|\norm{\iota_{\infty,2}^n_{{\mathbb M}_n}\otimes \Phi^{2k}(\rho)-\iota_{\infty,2}^n_{{\mathbb M}_n}\otimes E(\rho)}{1}\le \frac{1}{2}\hspace{.1cm}, \hspace{.1cm} \forall n\in {\mathbb N}^+, \rho\in S({\mathbb M}_n({\mathcal M}))\}\hspace{.1cm}.\]
We obtain the following corollary of Theorem \ref{thm:finitecb}.
\begin{cor}\label{cor:cbmix}
Let $\Phi$ be a symmetric quantum Markov map. Then
\[ k_{\diamond}(\Phi)\le \frac{\ln\ln C_{cb}(E)+2}{-2\ln( 1-\frac{-\ln \lambda(\Phi)}{\ln (10 C_{cb}(E))})}\hspace{.1cm}.\]
\end{cor}
\begin{exam}[Quantum Expander]{\rm A unital quantum channel $\Phi:{\mathbb M}_d\to {\mathbb M}_d$ is called $N$-regular if $\Phi(\rho)=\frac{1}{N}\sum_{i=1}^N U_i\rho U_i^*$ for some unitary $U_1,\cdots,U_N$. $\Phi$ is a $(N,\lambda)$-quantum expander if $\Phi$ is primitive, $N$-regular and $\lambda(\Phi)\le \lambda$. It was proved in \cite{ben2008quantum} that for every $\lambda<1$, there exists a family of $(N=O(\lambda^{-4}),\lambda)$ quantum expanders $\Phi_k:{\mathbb M}_{d_k}\to {\mathbb M}_{d_k}$ with $\lim_k d_k=+\infty$. Then
\[ \alpha_c(\Phi_k)\le 1-\frac{-\ln \lambda}{\ln (10 d^2)}\]
where we used the fact $C_{cb}(E_\tau)=d^2$.
Note that for a quantum expander $\Phi$, its symmetrization $\Psi=\frac{1}{2}(\Phi+\Phi^*)$ is always a $(2N,\lambda)$. Our Corollary \ref{cor:cbmix} implies the following upper bound for the $\diamond$-mixing time of a $(N,\lambda)$ symmetric quantum expander
\[ k_{\diamond}(\Psi)\le \frac{\ln \ln d+3}{-2\ln( 1-\frac{-\ln \lambda}{\ln (10 d^2)})}=O(\log d)\hspace{.1cm}, \hspace{.1cm}\text{ as } d\to \infty \hspace{.1cm}.\]
Such $O(\log d)$ mixing time bound is asymptotically optimal for classical expander graph. The above estimate extends it to quantum expander and completely bounded norm. We refer to \cite{ben2008quantum,hastings2007random,pisier2014quantum,hastings2009superadditivity} for the explicit constructions of quantum expander and its applications in quantum information theory.
}
\end{exam}
\begin{rem}\rm{The results in this and next section can be extended to $p$-relative entropy and $p$-Beckner inequality studied in \cite{li20}. This will be presented in a forthcoming publication.
}
\end{rem}
\section{Introduction}
Logarithmic Sobolev inequalities (LSI) were first introduced by Gross \cite{gross1}, who established them on Gaussian measures as a reformulation of Nelson's Hypercontractivity \cite{nelson1966quartic,nelson1973construction} in quantum field theory (see also \cite{bonami1968ensembles,bonami1970etude,rudin1960trigonometric,stam1959some} for earlier related results).
Since the seminal works of Gross \cite{gross1,gross2}, log-Sobolev inequalities have been widely studied on manifolds, graphs, as well as noncommutative spaces (see the surveys \cite{Ledoux,Gross14}). It appears that one natural framework of LSIs is given by the theory of Markov semigroups, i.e. a semigroup $(P_t)_{t\ge 0}$ of unital positivity preserving maps on some measure space. Recall that on a probability space $(\Omega, \mu)$, a Markov semigroup $(P_t)_{t\ge 0}: L_\infty(\Omega,\mu)\to L_\infty(\Omega,\mu)$ satisfies the \emph{logarithmic Sobolev inequality} ($\LSI$) with a parameter $\alpha>0$ if
\begin{align}\label{eq:introLSI}
\int f^2 \ln f^2 d\mu - \left( \int f^2d\mu \right) \ln \left( \int f^2d\mu \right) \leq \frac{2}{\alpha}\int (L f)f d\mu\hspace{.1cm} ,\hspace{.1cm} \forall \hspace{.1cm} f \in L_{2}(\Omega,\mu) \tag{$\LSI$}
\end{align}
where $L$ is the generator of the semigroup $P_t=e^{-tL}$. A variant of LSI, called \emph{modified log-Sobolev inequality} ($\text{MLSI}$), states that for all positive functions $g\ge 0$
\begin{align}\label{eq:MLSI}
\int g\ln g d\mu - \left( \int g d\mu\right) \ln \left( \int g d\mu \right) \leq \frac{1}{2\alpha} \int \left( L g \right)\ln g d\mu. \tag{$\text{MLSI}$}
\end{align}
These two forms of log-Sobolev inequality are closely related and inspire fruitful interactions between analysis, geometry, and probability. For instance, a classic theorem of Bakry and \'Emery \cite{BakryEmery1985} states that a positive Ricci curvature lower bound on a Riemannian manifold implies log-Sobolev inequality (LSI) for its heat semigroup. Later, Otto and Villani \cite{otto2000generalization} proved that the modified log-Sobolev inequality (MLSI) implies the transport cost inequalities, the latter was used by Talagrand \cite{Talagrand95} in deriving his concentration inequality. In the last decade, these connection between LSIs and other functional inequalities have also been extended to discrete spaces for finite Markov chain \cite{erbar2012ricci,maas2011gradient,mielke11}, as well as noncommutative spaces for quantum Makrov semigroups.
Quantum Markov semigroups are noncommutative generalizations of Markov semigroups, where the underlying probability spaces are replaced by operator algebras. Quantum Markov semigroups model Markovian approximations of open quantum systems, which undergo noise and interact with the environment. In recent years, log-Sobolev and related inequalities for quantum Markov semigroups are gaining increasing attentions for applications in quantum information theory and quantum many-body systems, such as quantum hypothesis testing \cite{beigi2020quantum}, quantum variational algorithms \cite{de2022limitations}, and quantum Gibbs samplers \cite{bardet2021rapid,capel2020modified}.
Starting from the work by Olkiewicz and Zegarlinski \cite{olkiewicz1999hypercontractivity} on equivalence of LSI and hypercontractivity on noncommutative $L_p$ spaces, much progress has been made on a systematic study of the connection of quantum LSI and MLSI to other noncommutative functional inequalities, such as the Poincar'e inequality (also known as spectral gap) \cite{KastoryanoTemme2013,bardet2017estimating}, entropic Ricci curvature \cite{datta1,carlen2017gradient,li2020graph,carlen2020non,wirth2021complete,wirth2021curvature}, and concentration inequalities \cite{rouze2019concentration,brannan2022complete}. Despite the progress mentioned above, one major challenge to useful application of LSIs is to obtain tight estimates of the Log-Sobolev constant $\alpha$. For example, Stroock and Zegarlinski in a series of works \cite{stroock1992equivalence,stroock1992logarithmic,stroock1992logarithmic1,stroock1995ergodic} proved that if the temperature is higher than certain critical value, the Glauber dynamics on a translation invariant classic spin system admit Log-Sobolev constant independent of lattice size, which implies the dynamics has rapid mixing property. Whereas the analogous results for quantum spin system are highly desired in theoretical study of quantum stimulation and quantum computation, the rapid mixing property remains widely open for quantum Gibbs sampler, excepts for some very recent progress on special cases \cite{capel2020modified,bardet2021entropy}.
The major difference and also the difficulty of LSIs in the quantum case is the tensorization property. For two classical Markov semigroups, if both satisfy LSI (resp. MLSI), so does their tensor product semigroup. Tensorization is a crucial property behind the Gross' dimension independent estimate of Gaussian measure \cite{gross1}, as well as the application of LSIs to concentration inequality \cite{OV}. Nevertheless, in the noncommutative settings, the tensorization property of MLSI fails in general, and the tensorization property of LSI is only known for limited $2$-dimensional examples \cite{king2014hypercontractivity,beigi2020quantum}. This problem was partially solved by the concept of \emph{complete modified log-Sobolev inequalities} introduced in \cite{gao2020fisher}: a semigroup $P_t$ satisfies complete Modified log-Sobolev inequalities (CMLSI) if the extended semigroup $P_t\otimes \iota_{\infty,2}^n_{{\mathbb M}_n}$ satisfies MLSI for $\alpha>0$ that is uniform for all matrix algebra ${\mathbb M}_n$.
This stronger definition naturally satisfies the tensorization property.
The purpose of this work is to derive tight CMLSI estimate for both classic and quantum Markov semigroups. Below we present our main results and illustrate relations to previous works.
\subsection{Entropy decay via CB return time}
We start with the relatively simple case of tracial von Neumann algebras. Let ${\mathcal M}$ be a finite von Neumann algebra equipped with a normal faithful tracial state $\tau$. A (trace) symmetric quantum Markov semigroup $T_t:{\mathcal M}\to {\mathcal M}$ is a continuous semigroup of completely positive unital map satisfying
\[\tau(T_t(x)y)=\tau(yT_t(x)),\quad \forall t\geq 0.\]
Under the symmetric assumption, the fixed point space
\[{\mathcal N}=\{x\in {\mathcal M} \hspace{.1cm} |\hspace{.1cm} T_t(x)=x,t\geq 0 \}\]
forms a subalgebra, which admits a trace preserving condition expectation $E:{\mathcal M}\to{\mathcal N}$. The main theorem of this work is the following CMLSI via a complete positivity mixing time. Recall that for two maps $\Phi$ and $\Psi$, we write $\Phi\le_{cp} \Psi$ if $\Psi-\Phi$ is completely positive.
\begin{theorem}\label{thm:main1}
Let $T_t=e^{-tL}:{\mathcal M}\to {\mathcal M}$ be a trace-symmetric quantum Markov semigroup with generator $L$ and $E:{\mathcal M}\to {\mathcal N}$ be the condition expectation onto its fixed point subalgebra ${\mathcal N}$. Define the completely bounded return time
\[t_{cb}:=\inf\{t>0 \hspace{.1cm} |\hspace{.1cm} 0.9E\le_{cp} T_t\le_{cp} 1.1E \}. \]
Then for any positive operator $\rho$, we have
\begin{align}\label{eq:symmetricMLSI}
2\alpha \hspace{.1cm} \tau(\rho \ln \rho-E(\rho)\ln E(\rho) )\le \tau(L (\rho)\ln \rho) \tag{MLSI} \hspace{.1cm},
\end{align}
where the modified log-Sobolev constant $\alpha$ satisfies $\alpha\ge \frac{1}{2t_{cb}}$. Equivalently,
\[ D(T_t(\rho)\|E(\rho))\le e^{-2\alpha t} D(\rho||E(\rho)) \hspace{.1cm}, \]
where $D(\rho||\sigma)=\tau(\rho \ln \rho-\rho\ln \sigma)$ denotes the quantum relative entropy. Moreover, the same inequality holds for the semigroup $T_t\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$ with any finite von Neumann algebra ${\mathcal Q}$.
\end{theorem}
Denote $\alpha_1$ as the optimal (largest) constant satisfying \eqref{eq:symmetricMLSI}, and $\alpha_c$ as the optimal CMLSI constant such that $T_t\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$ satisfying \eqref{eq:symmetricMLSI} for all ${\mathcal Q}$. The above theorem states that
\begin{align}\label{eq:inversetcb} \alpha_1\ge \alpha_c\ge \frac{1}{2t_{cb}}\hspace{.1cm}.\end{align}
The proof of Theorem \ref{thm:main1} uses entropic inequalities inspired from quantum information theory.
The quantum relative entropy is a fundamental information quantity that measures how well one can distinguish the state $\rho$ from $\sigma$ (see the survey for the ubiquitous role of relative entropy in quantum information).
The key property behind the wide applications of relative entropy is the data processing inequality that for any quantum Markov map $\Phi$ (normal, unital and completely positive)
\begin{align*} D(\Phi_*(\rho)||\Phi_*(\sigma))\le D(\rho||\sigma)\hspace{.1cm}, \end{align*}
where $\Phi_*$ is the pre-adjoint of $\Phi$ and $\Phi_*=\Phi$ if $\Phi$ is trace symmetric. From this perspective, MLSI is a refined data processing inequality that the relative entropy from a state $\rho$ to its equilibrium $E(\rho)$ decays exponentially fast. Indeed, the proof of Theorem \ref{thm:main1} can be reduced from the following entropy contraction inequality for symmetric quantum Markov maps.
\begin{theorem}\label{thm:main5} Let $\Phi$ be a symmetric quantum Markov map and let $E:{\mathcal M}\to{\mathcal N}$ be the condition expectation onto the multiplicative domain of $\Phi$. Define $k_{cb}\in \mathbb{N}$ as the smallest integer such that $0.9E \le_{cp} \Phi^{2k}\le_{cp} 1.1E$.
Then for $\alpha=1-\frac{1}{2k_{cb}}$ and any quantum state $\rho$,
\begin{align} D(\Phi(\rho)||\Phi\circ E(\rho))\le \alpha D(\rho||E(\rho))\label{eq:introSDPI}\end{align}
Furthermore, the same inequality holds for $\Phi\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$ for any finite von Neumann algebra $\mathcal{Q}$.
\end{theorem}
We denote by $\alpha(\Phi)$ the optimal (smallest) $\alpha$ such that \eqref{eq:introSDPI} is satisfied. In primitive cases (${\mathcal N}=\mathbb{C}1$), such contraction coefficients were considered in both classical \cite{del2003contraction} and quantum \cite{hirche2020contraction,muller2016entropy} setting. They are also closely related to the strong data processing inequality studied in \cite{polyanskiy2017strong,du2017strong,lesniewski1999monotone,hiai2016contraction}. In particular, the existence of some constant $\alpha<1$ in finite dimensions was obtained in \cite{gao2022complete}.
The above theorem is the first result giving a concrete estimate of $\alpha(\Phi)$ for non-primitive cases as well as the complete bounded version (as the assertion extends to $\Phi\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$). This is of course the analog of Theorem \ref{thm:main1} for a discrete time quantum Markov chain $(\Phi^n)_{n\ge 1}$, where \ref{thm:main1} follows easily from applying the inequality \eqref{eq:introSDPI} to a semigroup map $T_t$ and taking the continuous-time limit.
\subsection{Classical Markov semigroups}
The trace symmetric case is broad enough to cover many important scenarios including classical Markov chains. For a classical Markov semigroup $P_t:L_\infty(\Omega,\mu)\to L_\infty(\Omega,\mu)$, the notion CMLSI is simply a uniform MLSI for all its matrix value functions that for any positive matrix valued function $g:\Omega\to {\mathbb M}_n$ and $n\ge 1$ ,
\begin{align}\label{eq:classicCMLSI}
\mu\circ {\rm \text{tr}} (g\ln g - E_\mu(g) \ln E_\mu(g)) \leq \frac{1}{2\alpha} \mu \circ {\rm \text{tr}}( (L g) \ln g )
\end{align}
where $\mu(f)=\int fd\mu$ is the scalar valued mean, $E_\mu(g)=\int gd\mu\in {\mathbb M}_n$ is the matrix valued mean, and ${\rm \text{tr}}$ is the matrix trace. In this case, the definition $t_{cb}$ time reduces to $L_\infty$-mixing time
\[ t_{cb}=\{t >0| \norm{T_t-E_\mu:L_1(\Omega)\to L_\infty(\Omega)}{}\le 0.1\}\hspace{.1cm},\]
which is accessible by standard kernel estimate. Indeed, $t_{cb}$ is always finite if the semigroup $P_t$ satisfies both ultra contractivity and Poincar\'e inequality.
\begin{theorem}\label{thm:main9}
Let $P_t=e^{-Lt}:L_\infty(\Omega,\mu)\to L_\infty(\Omega,\mu)$ be a primitive Markov semigroup symmetric to the probability measure $\mu$. Suppose
\begin{enumerate}
\item[i)] $P_t$ satisfies $\lambda$-Poincar\'e inequality: for all $f$,
\begin{align} \lambda\mu(|f-E_\mu(f)|^2)\le \int f(Lf)d\mu \hspace{.1cm}.\label{eq:label}\end{align}
\item[ii)] There exists $t_0>0$ such that \begin{align}\label{eq:ultracon}\norm{P_{t_0}-E_\mu: L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}\le C_0\end{align}
\end{enumerate}Then
\begin{align}\alpha_1(L)\ge \alpha_{c}(L)\ge \frac{\lambda(L)}{2(\lambda(L)t_0+\ln C_0+\ln 10)}\hspace{.1cm}. \label{eq:CMLSIclassical}\end{align}
\end{theorem}
For LSI constant, a similar estimate was established by Diaconis and Saloff-Coste \cite{diaconis1996logarithmic} (translated to our normalization of $t_{cb}$) that
\begin{align*}\label{eq:inversetb}\alpha_1\ge\alpha_{2}\ge \frac{1}{t_{cb}}\hspace{.1cm}.\end{align*}
Here $\alpha_2$ is the optimal LSI constant in \eqref{eq:introLSI}.
We remark that this $O(t_{cb}^{-1})$ lower bound is asymptotically tight for both $\alpha_2$ and $\alpha_c$ for graph Laplacians on the odd circles. Moreover, when $\Omega$ is a finite state space, our result implies that $\alpha_c$ is equivalent to $\alpha_2$ up to a $\log\log$ term.
\begin{theorem}
\label{thm:main2}
Let $\Omega$ be a finite state space and $P_t=e^{-tL}:l_\infty(\Omega,\mu)\to l_\infty(\Omega,\mu)$ be a Markov semigroup symmetric to a probability density $\mu$. Then
\[ \alpha_c\ge \frac{4\alpha_2}{3(4+\log\log \norm{\mu^{-1}}{\infty})}\hspace{.1cm}.\]
\end{theorem}
Beyond finite state space, Theorem \ref{thm:main9} applies to sub-Laplacian generators of H\"ormander systems.
\begin{theorem}\label{thm:main3} Let $(M,g)$ be a compact connected Riemannian manifold without boundary and let $d\mu=\omega d\operatorname{vol}$ be a probability measure with a smooth density $\omega$. Suppose $H=\{X_i\}_{i=1}^k\subset TM$ is a family of vectors fields satisfies the \emph{H\"{o}rmander's condition} that at every point $x\in M$,
\[
T_xM=\operatorname{span}\{[X_{i_1},[X_{i_2},\cdots, [X_{i_{n-1}}, X_{i_n}]]] \hspace{.1cm} | \hspace{.1cm} 1\leqslant i_1,i_2\cdots i_n\leqslant k \}. \tag{\text{H\"{o}rmander condition}}
\]
Then the horizontal heat semigroup $P_t=e^{-\Delta_H}$ generated by the sub-Laplacian
\[
\Delta_H=\sum_{i} X_i^*X_i
\]
has CMLSI constant $\alpha_c(\Delta_H)>0$.
\end{theorem}The only sub-Laplacian example of CMLSI before this work is the canonical sub-Laplacian on $SU(2)$, which was obtained in \cite{gao2022complete2} using certain gradient estimate as a quasi-curvature condition. We remark that Theorem \ref{thm:main3} asserts CMLSI for all H\"ormander sub-Laplacians, avoiding any curvature condition.
For scalar valued functions, it is well-known that the MLSI and LSI constants coincide: $\alpha_1=\alpha_2$, as $P_t=e^{-\Delta_H}$ is a diffusive semigroup. The positivity of $\alpha_2$ was proved by \L ugiewicz and Zegarli\'nski \cite{lugiewicz2007coercive} using a similar hyper-contractive argument from \cite{diaconis1996logarithmic}.
The derivation of matrix valued MLSI through Theorem \ref{thm:main1} is quite different of the standard approaches for LSI. For example, both \cite{diaconis1996logarithmic} and \cite{lugiewicz2007coercive} rely on the Rothaus Lemma that
\begin{align} \mu(f^2\ln f^2)\le \mu(\hat{f}^2\ln \hat{f}^2)+\norm{\hat{f}}{2}^2\hspace{.1cm}, \label{eq:rothaus1}\end{align}
where $\hat{f}=f-{\mathbb E}_\mu(f)$ is the mean zero part for a real function $f$. A related alternative approach can uses uniform convexity of $L_p$-space plus hypercontractive estimates on mean zero part to conclude hypercontractivity. We remark that all these steps fail essentially for matrix valued functions (see Proposition \ref{prop:rothaus}).
\subsection{GNS-symmetric quantum Markov semigroup}
Beyond the trace symmetric semigroup, we extend the entropy decay result to quantum Markov semigroup symmetric to a state. The motivation is two fold: from physics perspective, most of physical model of quantum systems has equilibrium as a state but not necessarily a trace. For instance, the quantum Gibbs state at finite temperature is never a trace, and the tracial invariant state corresponds to the infinite temperature in the theorectial limit. On the mathematical side, a general von Neumann algebra may not admit a trace, called Type III von Neumann algebra. It turns out that entropy decay formulation of MLSI naturally extends to Type III setting without traces. We note that such difference does not appears for classic cases because every measure gives a trace.
\begin{theorem}\label{thm:main6}
Let ${\mathcal M}$ be a $\sigma$-finite von Neumann algebra equipped with a normal faithful state $\phi$. Suppose $T_t=e^{-tL}:{\mathcal M}\to {\mathcal M}$ is a quantum Markov semigroup satisfying the GNS-symmetric condition
\begin{align}\label{eq:introGNSsymmetry}
\phi(xT_t(y))=\phi(T_t(x)y)\hspace{.1cm}, \hspace{.1cm} \forall x, y\in {\mathcal M} , \ t\ge 0\hspace{.1cm}.
\end{align}
with respect to $\phi$. Denote $E:{\mathcal M}\to{\mathcal N}$ as the $\phi$-preserving conditional expectation onto the fixed point subalgebra ${\mathcal N}$.
Then for any state $\rho$,
\begin{align}D(T_{t,*}(\rho)||E_*(\rho)) \pl \le \pl e^{-2\alpha t}D(\rho||E_*(\rho))\hspace{.1cm}, \label{eq:introexpdecayintro}\end{align}
where the MLSI constant satisfies $\alpha\ge \frac{1}{2t_{cb}}$.
The same inequality holds for $T_t\otimes \iota_{\infty,2}^n_{{\mathcal Q}}$ for any $\sigma$-finite von Neumann algebra ${\mathcal Q}$.
\end{theorem}
The exponential decay of relative entropy \eqref{eq:introexpdecayintro} can be viewed as the definition of MLSI in Type III cases, whereas the right hand side of the tracial MLSI \eqref{eq:symmetricMLSI} is even not defined without trace. The key technical tool we used is the Haagerup's reduction \cite{haagerup2010reduction}, which is a method to prove results for type {III} von Neumann algebras by reducing to the cases of tracial von Neumann algebras. We show that the both the statement of entropy decay \eqref{eq:introexpdecayintro} and our tool $t_{cb}$ are well compatible with Haagerup's reduction.
In finite dimensions, we have the following estimate using the complete Pimsner-Popa index,
\[C(E)=\inf \{C>0| x\le CE(x)\hspace{.1cm}, \forall \hspace{.1cm} x\in {\mathcal M}_+\}\hspace{.1cm}, \hspace{.1cm} C_{cb}(E):=\sup_{{\mathcal Q}}C(E\otimes \iota_{\infty,2}^n_{{\mathcal Q}})\]
\begin{theorem}\label{thm:main7}
For GNS-symmetric quantum Makrov semigroups,
\begin{align}\frac{\lambda}{2\ln (10 C_{cb}(E))}\le \alpha_{c}\le \alpha_1\le \lambda \hspace{.1cm}.\label{eq:introfdQMS}\end{align}
where $\lambda$ is the spectral gap of the generator.
\end{theorem}
We note that the above $\mathcal{O}(\frac{\lambda}{\ln C_{cb}(E)})$ bound gives an exponential improvement on the dependence of $C_{cb}(E)$ to Gao and Rouz\'e's work \cite[Theorem 3.3]{gao2022complete}.
An application of noncommutative MLSI is the following concentration inequalities for a general quantum state. Recall that the Lipschitz semi-norm
\[ \|x\|_{\text{Lip}}^2 \pl = \pl \max\{ \norm{\Gamma_L(x,x)}{}\hspace{.1cm}, \hspace{.1cm} \norm{\Gamma_{L}(x^*,x^*)}\} \]
is defined through the \emph{gradient form} (or Carr\'e du Champ operator)
\[ \Gamma_L(x,y) \pl = \pl \frac{1}{2}\Big(L(x^*)y+x^*L(y)-L(x^*y)\Big)\hspace{.1cm}, \hspace{.1cm} \forall x,y\in \text{dom}(L)\hspace{.1cm}. \]
\begin{theorem}\label{thm:main8} Let ${\mathcal M}$ be a $\sigma$-finite von Neumann algebra and $T_t=e^{-tL}$ be a $\phi$-symmetric Quantum quantum Markov semigroup. Suppose $T_t$ satisfies MLSI with parameter $\alpha>0$. There exists an universal constant $c$ such that for $2\le p <\infty$
\[ \alpha\|x-E(x)\|_{L_p({\mathcal M},\phi)}\pl \le \pl c\sqrt{p}\norm{x}{\text{Lip}} \hspace{.1cm} .\]
Moreover, for any $t>0$, there exist a projection $e$ such that
\[ \|e(x-E(x))e\|_{\infty} \le t \quad \mbox{and} \quad \phi(1-e)\le 2 \exp( -\frac{\alpha^2t^2}{16ec^2\norm{x}{\text{Lip}}^2}) \hspace{.1cm} .\]
\end{theorem}
The rest of paper is organized as follows. Section 2 proves the entropy contraction Theorem \ref{thm:main5} for symmetric quantum Markov map. Based on that, we derive in Section 3 the main Theorem \ref{thm:main1} for trace symmetric semigroups and discuss its consequence Theorem \ref{thm:main2} \& \ref{thm:main3} in classical setting. Section 4 is devoted to the GNS-symmetric cases, in which Theorem \ref{thm:main6}, \ref{thm:main7} and \ref{thm:main8} are presented. We conclude with some discussion and open problems.
{\bf Notations.} We write calligraphic letters ${\mathcal M},{\mathcal N}$ for von Neumann algebras. We denote ${\mathbb M}_n$ as the algebra of $n\times n$ as complex matrices and ${\rm \text{tr}}$ as the standard matrix trace. The identity operator is denoted by $\mathbf{1}$, and the identity map between spaces is denoted as $\iota_{\infty,2}^n$.
{\bf Acknowlegdement.} The research of LG is partially supported by NSF grant DMS-2154903.
LG is also grateful to the support of AMS-Simons Travel Grants. Nicholas LaRacuente is supported as an IBM Postdoc at The University of Chicago. MJ was partially supported by NSF Grant DMS 1800872 and NSF RAISE-TAQS 1839177.
\section{Modified Log-Sobolev Inequality for Symmetric Markov semigroups}\label{sec:symmetricCMLSI}
\subsection{Complete modified log-Sobolev inequality}\label{sec:CMLSI}
In this section, we discuss the entropy decay for trace symmetric quantum Markov semigroups. This represents the continuous time analogue of the results in the previous section.
A quantum Markov semigroup $(T_t)_{t\ge 0}:{\mathcal M}\to{\mathcal M}$ is a family of maps satisfying
\begin{enumerate}
\item[i)] for each $t\ge 0$, $T_t$ is a quantum Markov map (i.e. normal, completely positive, and unital)
\item[ii)] $T_0=\iota_{\infty,2}^n_{\mathcal M}$ and $T_s\circ T_t=T_{s+t}$ for any $s,t \ge 0$.
\item[iii)] for $x\in {\mathcal M}$, $t\mapsto T_t(x)$ is weak$^*$-continuous.
\end{enumerate}
The generator of the semigroup is defined as
\[\hspace{.1cm} Lx=w^*\text{-}\lim_{t\to 0} \frac{1}{t}(x-T_t(x))\hspace{.1cm} \]
so that on the domain of $L$ the limit exists.
Again, we will consider ${\mathcal M}$ is a finite von Neumann algebra equipped with a normal faithful tracial state $\tau$. We say a quantum Markov semigroups $(T_t)$ is trace-symmetric (or simply, \emph{symmetric}) if each $T_t$ is trace-symmetric, i.e.
\[\tau(x^*T_t(y))=\tau(T_t(x)^*y)\hspace{.1cm} , \hspace{.1cm} \forall x,y\in {\mathcal M}, t\ge 0\]
In this case, $L$ is a positive, symmetric operator, densely defined on $L_2({\mathcal M})$. Its kernel is
the fixed-point subspace ${\mathcal N}:=\ker(L)=\{x\in {\mathcal M}\hspace{.1cm} | \hspace{.1cm} T_t(x)=x, \forall t\ge 0\}$. This coincides with the common multiplicative domain of all $T_t$, hence it is a von Neumann subalgebra. Moreover, each $T_t$ is an ${\mathcal N}$-bimodule map
\[T_t(axb)=aT_t(x)b\hspace{.1cm}, \hspace{.1cm} \forall\hspace{.1cm} a,b\in {\mathcal N} ,x\in {\mathcal M}\]
In particular, we have
\[T_t\circ E= E\circ T_t=E\hspace{.1cm}.\]
where $E:{\mathcal M}\to {\mathcal N}$ is the trace preserving conditional expectation onto the fixpoint algebra ${\mathcal N}$.
We say $(T_t)$ is \emph{ergodic} if ${\mathcal N}={\mathbb C} 1$ is trivial. Note that in the literature on many-body systems it is common to use \emph{primitive} instead of ergodic. In this case the semigroup admits a unique invariant (normal) state, namely the trace $\tau$, when acting on normalized densities in $L_1(\mathcal{M})$. Alternatively, the only invariant observable are multiples of the identity ${\bf 1}$.
In our setup we allow for not necessarily ergodic quantum Markov semigroups.
Recall that semigroup is equivalently determined by its
{\it Dirichlet form} \[{\mathcal E}:L_2({\mathcal M})\to [0,\infty]\hspace{.1cm} ,\hspace{.1cm} {\mathcal E}(x,x)=\tau(x^*L x)\hspace{.1cm}.\]
We write $\operatorname{dom} (L)$ for the domain of $L$ and $\operatorname{dom} ({\mathcal E})$ for the domain of ${\mathcal E}$. The Dirichlet subalgebra is defined as ${\mathcal A}_{\mathcal E}:=\operatorname{dom} ({\mathcal E})\cap {\mathcal M}$.
It was proved \cite{davies1992non} that ${\mathcal A}_{\mathcal E}$ is a dense $*$-subalgebra of ${\mathcal M}$ and a core of $L^{1/2}$.
We denote by
\[S({\mathcal A}_{\mathcal E})=S({\mathcal M})\cap {\mathcal A}_{\mathcal E} \hspace{.1cm}, \hspace{.1cm} S_B({\mathcal A}_{\mathcal E})=S_B({\mathcal M})\cap {\mathcal A}_{\mathcal E}\]
the set of bounded density operators from ${\mathcal A}$. Let us briefly review the definition of functional inequalities for quantum Markov semigroups.
\begin{defi}\label{defi:FI}
Let $T_t=e^{-Lt}:{\mathcal M}\to{\mathcal M}$ be a symmetric quantum Markov semigroup and $E:{\mathcal M}\to {\mathcal N}$ be the trace preserving conditional expectation onto its fixed point space. We say $T_t$ satisfies
\begin{enumerate}
\item[i)] the Poincar\'e inequality (PI) for $\lambda>0$ if
\begin{align}\label{eq:PI}
\lambda\norm{x-E(x)}{2}^2\le {\mathcal E}(x,x)\hspace{.1cm}, \hspace{.1cm} \forall x\in {\mathcal A}_{\mathcal E}\hspace{.1cm}.
\end{align}
\item[ii)] the log-Sobolev inequality (LSI) for $\alpha_2>0$ if
\begin{align}\label{eq:LSI}
\alpha_2 \tau\big(|x|^2\ln |x^2|-E(|x^2|)\ln E(|x^2|)\big)\le 2{\mathcal E}(x,x)\hspace{.1cm}, \hspace{.1cm} \forall x\in {\mathcal A}_{\mathcal E}\hspace{.1cm}.
\end{align}
\item[iii)]the modified Log-Sobolev inequality (MLSI) for $\alpha_1>0$ if
\begin{align}\label{eq:MLSI1}
2\alpha_1 D(\rho||E(\rho))\le {\mathcal E}(\rho,\ln\rho)\hspace{.1cm}, \hspace{.1cm} \forall \rho\in S_B({\mathcal A}_{\mathcal E})\hspace{.1cm},
\end{align}
\item[iv)]the complete modified Log-Sobolev inequality (CMLSI) for $\alpha_c>0$ if for any finite von Neumann algebra ${\mathcal Q}$, $\iota_{\infty,2}^n_{\mathcal Q}\otimes T_t$ satisfies $\alpha$-modified Log-Sobolev inequality
\end{enumerate}
We denote by $\lambda(L), \alpha_2(L),\alpha_{1}(L),$ and $\alpha_{c}(L)$ the optimal (largest possible) constants for PI, LSI, MLSI, and CMLSI respectively.
\end{defi}
The Poincar\'e inequality \eqref{eq:PI} is equivalent to the spectral gap of $L$ as a positive operator. LSI \eqref{eq:LSI} is equivalent to hypercontractivity \cite{olkiewicz1999hypercontractivity}
\begin{align} \norm{T_t:L_2({\mathcal M})\to L_p({\mathcal M})}{}\le 1\hspace{.1cm} \text{ if } \hspace{.1cm} p\le 1+e^{2\alpha t} \label{eq:HC}.\end{align}
MLSI \eqref{eq:MLSI1} is known to be equivalent to the exponential decay of relative entropy (\cite[Theorem 3.2]{bardet2017estimating} and \cite[Proposition A.3]{brannan2022complete}) that
\begin{align} D(T_t(\rho)||E(\rho))\le e^{-2\alpha t}D(\rho||E(\rho))\hspace{.1cm}, \hspace{.1cm} \forall \rho\in S({\mathcal M})\hspace{.1cm}.\label{eq:entropy}\end{align}
The equivalence of \eqref{eq:MLSI1} and \eqref{eq:entropy} is obtained by differentiating the relative entropy for $T_t$ at $0$. This leads to the entropy production
\[I_L(\rho):={\mathcal E}(\rho,\ln\rho)=-\frac{d}{dt}\vert_{t=0} D(T_t(\rho)||E(\rho))=\tau(L(\rho) \ln \rho) \hspace{.1cm},\]
where the last expression make sense for $\rho\in \operatorname{dom}(L)\cap S_B({\mathcal M})$. In the classical literature this term is also called Fisher information. The MLSI \eqref{eq:MLSI1} states that
\[ 2\alpha D(T_t(\rho)||E(\rho))\le -\frac{d}{dt}D(T_t(\rho)||E(\rho))|_{t=0}\hspace{.1cm},\]
which by Gronwall's lemma implies \eqref{eq:entropy}.
For primitive symmetric semigroups, the above constants satisfy the relation \cite{olkiewicz1999hypercontractivity,kastoryano2013quantum},
\[ \alpha_2 \le \alpha_1 \le\lambda\hspace{.1cm},\]
In particular, $\alpha$-LSI always implies $\alpha$-MLSI, but the converse is not true in general.
Indeed, LSI fails for non-primitive semigroup whenever ${\mathcal M}$ is noncommutative \cite{bardet2022hypercontractivity}. On the other a non-trivial bound for MLSI and also CMLSI has been proved for all finite dimensions symmetric quantum Markov semigroups \cite{gao2022complete}.
From this perspective, the complete MLSI is the more suitable completely bounded version of log-Sobolev inequalities. Note that $\iota_{\infty,2}^n_{\mathcal Q}\otimes T_t$ is never primitive. The advantage of CMLSI over MLSI and LSI is the tensorization property
\begin{align}\label{eq:tensorization} \alpha_{c}(L_1\otimes \iota_{\infty,2}^n+\iota_{\infty,2}^n\otimes L_2)=\min\{\alpha_{c}(L_1), \hspace{.1cm} \alpha_{c}(L_2)\}\hspace{.1cm},\end{align}
which fails for MLSI and LSI in the quantum cases \cite[Section 4.4]{brannan2022complete}.
The main results of this section is a lower bound of $\alpha_{c}$ via the CB-return time.
\begin{theorem}\label{thm:symmetricQMS} Let $T_t=e^{-Lt}:{\mathcal M}\to {\mathcal M}$ be a symmetric quantum Markov semigroup and $E:{\mathcal M} \to {\mathcal N}$ be the trace preserving conditional expectation onto the fixed point subalgebra ${\mathcal N}$. Define the CB return time as
\[ t_{cb}(L)=\inf \Big \{ t>0\hspace{.1cm} |\hspace{.1cm} \norm{T_t-E:L_\infty^1({\mathcal N}\subset{\mathcal M})\to L_\infty({\mathcal M})}{cb}\le \frac{1}{10} \Big \}\]
Then
\[ \frac{1}{2t_{cb}(L)}\pl \le \pl \alpha_{c}(L)\pl \le \pl\alpha_{1}(L)\hspace{.1cm}.\]
\end{theorem}
\begin{proof}
Take $t_m=t_{cb}(L)/2m$. Since $T_t$ is symmetric, $T_{t_m}^*T_{t_m}=T_{t_m}T_{t_m}=T_{2t_m}$. Hence $T_{t_m}$ has discrete return time $k_{cb}(T_{t_m})\le m$. By the Lemma \ref{lemma:difference}, for any $\rho\in S_B({\mathcal M})$
\begin{align*}D(T_{t_m}(\rho)||E(\rho))\le & (1-\frac{1}{2m}) D(\rho||E(\rho))
\end{align*}
Write $t_{cb}=t_{cb}(L)$. Now assume further $\rho\in \cup_{t>0} T_t({\mathcal M})\subset \operatorname{dom} (L)$.
We have
\begin{align*}I(\rho)&=\lim_{t\to 0}\frac{ D(\rho||E(\rho))-D(T_{t}(\rho)||E(\rho))}{t}\\
&=\lim_{m\to \infty}\frac{ D(\rho||E(\rho))-D(T_{\frac{t_{cb}}{2m}}(\rho)||E(\rho))}{\frac{t_{cb}}{2m}}\\
&\ge \lim_{m\to \infty}\frac{\frac{1}{2m}D(\rho||E(\rho))}{\frac{t_{cb}}{2m}}
=\frac{1}{t_{cb}} D(\rho||E(\rho))\hspace{.1cm} .
\end{align*}
The entropy decay for general $\rho\in S({\mathcal M})$ can be obtained by approximation as in the Appendix \cite[Appendix]{brannan2022complete}). This proves $\alpha_{1}(L)\ge \frac{1}{2t_{cb}(L)}$. The same argument applies to $\iota_{\infty,2}^n_{\mathcal Q}\otimes L$ as $t_{cb}(L)=t_{cb}(\iota_{\infty,2}^n_{\mathcal Q}\otimes L)$ by the definition of $t_{cb}$.
\end{proof}\vspace{0.5ex}
\begin{rem}{\rm
a) The tightness of the above estimate will be discussed in Example \ref{ex:cyclic2}. The above theorem shows that the CMLSI is lower bounded by inverse of CB return time up to a constant.\\
b) In \cite{brannan2022complete}, a similar estimate $\alpha_c\ge \Omega(\frac{1}{t_{cb}})$ was obtained for quantum Markov semigroups that admits non-negative entropy Ricci curvature lower bound.
The entropy Ricci curvature lower bound for quantum Markov semigroup was introduced by Carlen and Mass \cite{carlen2017gradient} using $\lambda$-displacement convexity of entropy functionals $H$ w.r.t to certain noncommutative Wasserstein distance. This was inspired from Lott and Villani \cite{lott2009ricci}, and Sturm's \cite{sturm2006geometry} work on metric measure spaces.\\
c) For heat semigroups on Riemmannian manifold, the entropy Ricci curvature lower bound follows from a lower bound of the Ricci curvature tensor as well as Bakry-Emery's curvature dimension. Nevertheless, unlike Bakry-Emery's curvature dimension, the entropy Ricci curvature lower bound for quantum Markov semigroup in general is hard to obtain. So far most examples rely on certain algebraic interwining relation $\nabla T_t= e^{-\lambda t}T_t \nabla$ between the semigroup $T_t$ and gradient operator $\nabla$ (see
\cite{carlen2017gradient,brannan2021complete,wirth2021complete}).\\ d) Our Theorem \ref{thm:symmetricQMS} here does not reply on any curvature condition. Furthermore, our approach uses a combination of information theoretic methods (entropic quantity, data processing inequality) and operator spaces technique (conditional $L_\infty^1$ space). This direct proof is even new in the classical setting.\\
e) It is worth point out that the definition of relative entropy as well as the exponential decay of relative entropy is independent of the choice of the trace. This also shows the naturality of our information-theoretic approach.
}\end{rem}
As an analog of Proposition \ref{prop:finitecb}, the CB-return time $t_{cb}(L)$ is finite if $T_t$ satisfies a Poincar\'e inequality (spectral gap) and $L_\infty^1\to L_\infty$ ultracontractivity for some $t$.
\begin{prop}\label{prop:finitetcb2}Let $T_t=e^{-L t}:{\mathcal M}\to {\mathcal M}$ be a symmetric quantum Markov semigroup and $E:{\mathcal M}\to {\mathcal N}$ be the trace preserving conditional expectation onto the fixed point subalgebra ${\mathcal N}$. Suppose
\begin{enumerate}
\item[i)] $T_t$ the satisfies $\lambda$-Poincar\'e inequality: for all $x$,
\[ \lambda\norm{x-E(x)}{2}^2\le {\mathcal E}(x,x)\hspace{.1cm}, \forall \hspace{.1cm} x\in {\mathcal M}\]
\item[ii)] There exists $t_0$ such that $\norm{T_{t_0}: L_\infty^1({\mathcal N}\subset {\mathcal M})\to L_\infty({\mathcal M})}{cb}\le C_0$
\end{enumerate}Then
\[ t_{cb}(L)\pl \le \pl \frac{\ln (10 C)}{\lambda(L)}+t_0 \hspace{.1cm} .\]
If in additional $C_{cb}(E)<\infty$, then
\[ t_{cb}(L)\pl \le \pl \frac{\ln (10 C_{cb}(E))}{\lambda(L)} \hspace{.1cm} .\]
\end{prop}
\begin{proof}
The argument is similar to the discrete time case Proposition \ref{prop:finitecb}.
\end{proof}
\subsection{Classical Markov Semigroups} In the remainder of this section, our focus lies on applications towards classical Markov map. We postpone the discussion of truly quantum semigroups to Section \ref{sec:GNSsymmetric}. Let us start with a probability space $(\Omega,\mu)$ and $T_t=e^{-Lt}:L_\infty(\Omega,\mu)\to L_\infty(\Omega,\mu)$ be a (classical) Markov semigroup symmetric to the probability measure $\mu$. For each $t$, $T_t$ is a unital positive map satisfying
\[\int gT_t(f) d\mu =\int T_t(g)f d\mu\hspace{.1cm}, \hspace{.1cm} \forall \hspace{.1cm} g,f\in L_\infty(\Omega,\mu)\hspace{.1cm}.\]
The corresponding Dirichlet form is given by
\[{\mathcal E} (f,g)=\int_{\Omega} \bar{f}L(g)d\mu\hspace{.1cm},\]
where $L$ is the generator. We denote the mean and expectation map with the following symbols:
\[\mu(f)=\int f d\mu\hspace{.1cm}, \hspace{.1cm} E_\mu(f)=(\int f d\mu){\bf 1},\]
where $1_\Omega$ is the unit constant function on $\Omega$. In the classical setting,
it is conventional to assume that $T_t$ is ergodic, i.e. $\mu$ is the unique invariant measure. Then the complete modified log-Sobolev inequality (CMLSI) is simply the modified log-Sobolev inequality for matrix valued functions: for all $n\in \mathbb{N}^+$ and positive matrix valued function $g:\Omega \to {\mathbb M}_n, g\ge 0$
\begin{align} \label{eq:classicalCMLSI}2\alpha \hspace{.1cm} \mu\circ {\rm \text{tr}} \big(g \log g - E_\mu(g) \log E_\mu(g) \big) \le \mu\circ {\rm \text{tr}}(L(g) \log g)\hspace{.1cm}.\end{align}
Here $g$ positive means that $g(\omega)$ is a positive semi-definite positive matrix $\mu$ almost everywhere. The left hand side of \eqref{eq:classicalCMLSI} is the classical-quantum relative entropy
$D(\rho||E_\mu(g))$, where $E_\mu(g)=\mu(g)1$ is interpreted as constant function with the matrix mean $\mu(g)=\int gd\mu\in {\mathbb M}_n$. The right hand side is the entropy production (or Fisher information)
\[I(g):=\mu\circ {\rm \text{tr}}(L(g) \log g)=-\frac{d}{dt} D(T_t(\rho)||E_\mu(g)) \vert_{t=0}\]
When $n=1$, \eqref{eq:classicalCMLSI} this reduces to the standard modified Log-Sobolev inequality (MLSI)
\begin{align} \label{eq:CMLSI}2\alpha \hspace{.1cm} \mu\big(g \log g - E_\mu(g) \log E_\mu(g) \big) \le {\mathcal E}(g, \log g)\hspace{.1cm}, \forall \hspace{.1cm} g\ge 0\end{align}
For classical Markov semigroups, the LSI, MLSI and CMLSI constant $\alpha_2,\alpha_1,\alpha_c$ all satisfy the tensorization property
\begin{align*}&\alpha(L_1\otimes \iota_{\infty,2}^n +\iota_{\infty,2}^n\otimes L_2)=\min\{ \alpha(L_1),\alpha(L_2)\}\hspace{.1cm}. \end{align*}
To apply Theorem \ref{thm:symmetricQMS}, we note that in the ergodic case $L_\infty^1(\mathbb{C}1\subset L_\infty(\Omega) )=L_1(\Omega,\mu)$. Recall that by Smith's lemma \cite{smith1983completely}, any bounded map $T:L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)$ is automatic completely bounded
\[\norm{T: L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}=\norm{T: L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{cb}\hspace{.1cm}. \]
Then the CB return time for the classical Markov semigroup reduces to the standard $L_\infty$ mixing time
\begin{align*} t_{cb}(L)=&\inf \{ t>0\hspace{.1cm} |\hspace{.1cm} \norm{T_t-E:L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}\le \frac{1}{10} \}\\
=&\inf \{ t>0\hspace{.1cm} |\hspace{.1cm} \norm{k_t-1}{ L_\infty(\Omega\times \Omega)}\le \frac{1}{10} \}\hspace{.1cm},
\end{align*}
where $k_t$ is the kernel of the semigroup $T_t$ written as an integral operator. Heat kernel estimates are accessible by standard tools in harmonic analysis.
In particular, we have the following corollary of Theorem \ref{thm:symmetricQMS} and Proposition \ref{prop:finitetcb2},
which shows that ultracontractivity plus spectral gap implies CMLSI.
\begin{cor}\label{cor:classical}
Let $T_t=e^{-Lt}:L_\infty(\Omega,\mu)\to L_\infty(\Omega,\mu)$ be a primitive Markov semigroup symmetric to the probability measure $\mu$. Suppose
\begin{enumerate}
\item[i)] $T_t$ satisfies $\lambda$-Poincar\'e inequality: for all $f$,
\begin{align} \lambda\mu(|f-E_\mu(f)|^2)\le {\mathcal E}(f,f)\hspace{.1cm}.\label{eq:label}\end{align}
\item[ii)] There exists $t_0>0$ such that \begin{align}\label{eq:ultracon}\norm{T_{t_0}-E_\mu: L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}\le C_0\end{align}
\end{enumerate}Then
\begin{align}\alpha_1(L)\ge \alpha_{c}(L)\ge \frac{\lambda(L)}{2(\lambda(L)t_0+\ln C_0+\ln 10)}\hspace{.1cm}. \label{eq:CMLSIclassical}\end{align}
\end{cor}
\begin{rem}{\rm
The above result can be compared to the bound of Diaconis and Saloff-Coste \cite[Theoem 3.10]{diaconis1996logarithmic}, which states
\begin{align} \label{eq:diaconis}\alpha_1\ge \alpha_{2}\ge \frac{\lambda}{\lambda t_0+\frac{1}{2}\ln (C_0)-1}\hspace{.1cm},\end{align}
In particular, $\alpha_1\ge \alpha_{2}\ge \frac{1}{t_b}$ for
the alternative $L_\infty$ mixing time
\begin{align*} t_b(L)=\inf \{ t>0\hspace{.1cm} |\hspace{.1cm} \norm{T_t-E_\mu:L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}\le (\frac{1}{e})^2\}
\end{align*}
Note that here $\norm{T_t-E_\mu:L_1\to L_\infty}{}$ is controlled by $\varepsilon=\frac{1}{e^2}$ not $\varepsilon=\frac{1}{10}$ in our definition of $t_{cb}$. Since $e^{-3}<0.1<e^{-2}$, we have $2t_b\le t_{cb}\le 3t_b$. Hence, in terms of lower bound for $\alpha_1$, \eqref{eq:diaconis} and \eqref{eq:classicalCMLSI} are equivalent up to an absolute constant. On the other hand, \eqref{eq:diaconis} lower bounds the LSI constant $\alpha_2$ and our estimate \eqref{eq:classicalCMLSI} bounds the complete MLSI constant $\alpha_{c}$. }
\end{rem}
\begin{rem} {\rm A
crucial step in the proof of \eqref{eq:diaconis} is the Rothaus lemma
\[H(f^2)\le H(|f-\mu(f)|^2)+\norm{f-\mu(f)}{2}^2,\]
where $H(f^2)=\mu(f^2\log f^2)-\mu(f^2)\log \mu(f^2)$ is the entropy functional.
The Rothaus lemma, closely connected to the notion of uniform convexity in Banach spaces, is a standard tool to improve a defective Logarithmic Sobolev inequality to a standard one (see e.g. \cite[Theoem 3.9]{diaconis1996logarithmic}). However, the Rothaus lemma fails for matrix-valued functions \cite[Section 7.5]{gao2020fisher}. Hence the Diaconis and Saloff-Coste argument brakes down for matrix-valued functions. Indeed, we will show in Section \ref{sec:failureLSI} that matrix-valued LSI inequality always fails for any classical Matrix semigroup.
}\end{rem}
\begin{exam}{\rm
Note that $\alpha_c(L)>0$ in turn implies $\lambda(L)\ge\alpha_1(L)\ge \alpha_c(L) >0$. Here is a very simple example with $\alpha_c(L)>0$ but the ultra-contractivity \eqref{eq:ultracon} is never satisfied for finite $t_0$. Take $L=I-E_\mu$. It generates the so called depolarizing semigroup
\[ T_t=e^{-t}\iota_{\infty,2}^n+(1-e^{-t})E_\mu\hspace{.1cm}, \hspace{.1cm} T_t(f)= e^{-t}f+(1-e^{-t})\mu(f){\bf 1}\hspace{.1cm},\]
Then for any $t<\infty$, \[\norm{T_{t}-E_\mu: L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}=\norm{e^{-t}\iota_{\infty,2}^n: L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}=e^{-t}C(E_\mu)\hspace{.1cm}.\]
Note however, that for infinite dimensional $L_{\infty}(\Omega,\mu)$ the constant $C(E_{\mu})=\infty$ is never finite.
}
\end{exam}
\subsection{Finite Markov Chain}\label{sec:finiteMC}
We now consider a finite set $\Omega=\{1,\cdots,d\}$ with a strictly positive probability measure $\mu$. Then we trivially find a finite index
\[C_{cb}(E_\mu)=C(E_\mu)=\inf\{C>0\hspace{.1cm} |\hspace{.1cm} f\le C\mu(f) \hspace{.1cm} \forall f\ge 0\}= \norm{\mu^{-1}}{\infty}\hspace{.1cm},\]
It was proved in \cite{diaconis1996logarithmic} that
\begin{align}&\frac{1}{t_b}\le \lambda \le \frac{2+\log\norm{\mu^{-1}}{\infty}}{2t_b}\hspace{.1cm} ,\\
&\frac{1}{t_b}\le \alpha_2 \le \frac{4+\log\log \norm{\mu^{-1}}{\infty}}{2t_b}\hspace{.1cm}. \label{eq:re}
\end{align}
Combined with our Theorem \ref{thm:symmetricQMS} and Proposition \ref{prop:finitetcb2}, we obtain:
\begin{cor}\label{eq:loglog}
For a finite Markov semigroup $T_t:L\infty(\Omega,\mu)\to L_\infty(\Omega,\mu)$ symmetric to $\mu$, the following estimates hold
\[ \min \Big \{ \frac{4\alpha_2}{3(4+\log\log \norm{\mu^{-1}}{\infty})}, \frac{\lambda}{2\log (10\norm{\mu^{-1}}{\infty})} \Big \} \le \alpha_c\le \alpha_1\le \lambda\]
\end{cor}
\begin{proof} Note that $2t_b\le t_{cb}\le 3t_b$. Then by Theorem \ref{sec:symmetricQMS} and \eqref{eq:re}
\[ \alpha_c\ge \frac{1}{2t_{cb}}\ge \frac{1}{6t_{b}} \ge \frac{4\alpha_2}{3(4+\log\log \norm{\mu^{-1}}{\infty})}\hspace{.1cm}.\]
The other lower bound $\alpha_c\ge \frac{\lambda}{2\log (10\norm{\mu^{-1}}{\infty})}$ follows from Corollary \ref{cor:classical} by choosing $t_0=0$.
\end{proof}
We illustrate our estimate with the following two examples.
\begin{exam}[Graph Laplacians]{\rm Let $G=(V,E)$ be a simple undirected graph with $|V|=d$. Let $w:E\to (0,\infty)$ be a positive weighted function on the edge set $E$. The (weighted) graph Laplacian is given by the matrix
\begin{align*}
L_G(u,v)=\begin{cases}
\sum_{e=(u,u')\in E}w_e, & \mbox{if } u=v \\
-w_e, & \mbox{if } (u,v)\in E \\
0, & \mbox{otherwise}.
\end{cases}
\end{align*}
$L_G$ generates the continuous time random walk $T_t=e^{-L_G t}$ as a Markov semigroup and is symmetric with to the uniform distribution $\pi$ on $V$. $T_t$ is ergodic if and only if $G=(V,E)$ is connected.
The expectation map $E_\pi(f)=\pi(f)1$ has index $C_{cb}(E_\pi)=d$. Then our estimates gives
\[ \frac{ \lambda(L_G)}{2(\ln d+\ln 10)}\le \alpha_c(L_G)\le \alpha(L_G)\le \lambda(L_G)\hspace{.1cm}.\]
This lower bound of $\alpha_c(L_G)$ has better dependence of the dimension $d$ than \cite[Lemma 5.2]{li2020graph}.
}
\end{exam}
\begin{exam}[Cyclic graphs]{\rm \label{ex:cyclic2} Let us revisit the cyclic graph $C_d$ with $d$ vertices. For the non-weighted case $w_e\equiv 1$, $L_{C_d}$ is a circulant matrix
\[L_{C_d}(i,j)=\begin{cases}
2, & \mbox{if } i=j \\
-1, & \mbox{if } |i-j|=1 \\
0, & \mbox{otherwise}.
\end{cases}\]
Thus, $L_{C_d}=2(I-K_{C_d})$ where $K_{C_d}$ is the random walk kernel in Example \ref{ex:cyclic}, $ L_{C_d}$ has spectrum $\lambda_j=2(1-\cos\frac{2\pi j}{d})$. As discussed in \cite[Example 3.6]{diaconis1996logarithmic}, \[\norm{T_t-E:L_1(V,\pi)\to L_\infty(V,\pi) }{}\le 2\exp(-\frac{4t}{d^2})(\sqrt{1+d^2/4t})\hspace{.1cm}\]
Choose $t_0=d^2$, we have
\[ \norm{T_t-E:L_1(V,\pi)\to L_\infty(V,\pi) }{}\le 2e^{-4}\sqrt{5/4}<\frac{1}{10}\hspace{.1cm}.\]
Thus by Theorem \ref{thm:symmetricQMS},
\[\frac{1}{2d^2}\le \alpha_c(L_{C_d})\le \alpha_1(L_{C_d})\le 2(1-\cos\frac{2\pi}{d})=\frac{8\pi^2}{d^2}+\mathcal{O}(\frac{1}{d^4})\hspace{.1cm}. \]
This shows that our inverse of $t_{cb}$ bound for $\alpha_c$ is asymptotically tight. Note that the LSI constant $\alpha_2(L_{C_d})$ is also $\mathcal{\Theta}(\frac{1}{d^2})$.
}
\end{exam}
We refer to \cite{diaconis1996logarithmic,bobkov2006modified} more examples on spectral gap $\lambda$, Log-Sobolev constants $\alpha_2,\alpha_1$, and $L_\infty$ mixing time $t_b$ of finite Markov chains.
\subsection{H\"ormander system}\label{sec:hormander}
We now switch to Markov semigroups on smooth manifolds generated by Laplacians and sub-Laplacians.
Let $(M,g)$ be a $d$-dimensional compact connected Riemannian manifold without boundary and let $d\mu=\omega d\operatorname{vol}$ be a probability measure with smooth density $\omega$ w.r.t to the volume form $d\operatorname{vol}$. A family of vectors fields $H=\{X_i\}_{i=1}^k\subset TM$ with $k\leqslant d$ is called a \emph{H\"{o}rmander's system} if at every point $x\in M$, the tangent space at $x$ can be spanned by the iterated Lie brackets of $X_i$s
\[
T_xM=\operatorname{span}\{[X_{i_1},[X_{i_2},\cdots, [X_{i_{n-1}}, X_{i_n}]]] \hspace{.1cm} | \hspace{.1cm} 1\leqslant i_1,i_2\cdots i_n\leqslant k \}. \tag{\text{H\"{o}rmander condition}}
\]
By compactness we can assume there is a global constant $l_H$ such that for every point $x\in M$, we need at most $l_H$th iterated Lie bracket in above expression (this is also called a strong H\"ormander system). Denote $\nabla=(X_1, \cdots, X_k)$ and by $X_i^*$ the adjoint of $X_i$ on $L^{2}(M,d\mu)$. Under the H\"{o}rmander condition, the sub-Laplacian
\[
\Delta_H=\nabla^*\nabla= \sum_{i} X_i^*X_i=\sum_{i}X_i^2+\div_{\mu}(X_i)X_i
\]
is a symmetric operator on $L^{2}(M,\mu)$ which generates an ergodic Markov semigroup $P_t=e^{-\Delta_H t}$, often called horizontal heat semigroup. Here $\operatorname{div}_{\mu}(X)$ is the divergence of $X$ w.r.t to $\mu$. When $H=\{X_i\}_{i=1}^d$ forms an orthonormal frame, $\Delta_H=\Delta$ recovers the Laplace-Beltrami operator and $P_t=e^{-\Delta t}$ is the standard heat semigroup on $M$.
Recall that the \emph{gradient form} (Carr\'e du Champ operator) is defined as
\[
\Gamma(f,g):=\frac{1}{2}(f\Delta_H(g)+\Delta_H(f)g-\Delta_H(fg))=
\sum_{i}\langle X_i f,X_ig\rangle\hspace{.1cm}. \]
It follows from the product rule that $\Gamma$ is diffusive, i.e. $\Gamma(fg,h)=f\Gamma(g,h)+g\Gamma(f,h)$. For diffusion semigroups, it is known \cite[Theorem 5.2.1]{bakry2014analysis} that the MLSI constant $\alpha_1$ and the LSI constant $\alpha_2$ coincides, i.e. $\alpha(L):=\alpha_1(L)=\alpha_2(L)$. It was proved in \cite[Theorem 3.1]{lugiewicz2007coercive} that for any H\"{o}rmander's system $H=\{X_i\}_{i=1}^k$ on a compact connected Riemannian manifold without boundary,
\[ \alpha(\Delta_H)>0 \hspace{.1cm}. \]
We obtain that $\alpha_c(\Delta_H)>0$ as a consequence of Corollary \ref{cor:classical}.
\begin{theorem}\label{thm:sublaplacian}Let $(M,g)$ be a $d$-dimensional compact and connected Riemannian manifold without boundary and let $H=\{X_i\}_{i=1}^k$ be a H\"ormander system. Then the horizantal heat semigroup $P_t=e^{-\Delta_H t}$ generated by the sub-Laplacian
\[\Delta_H=\sum_{i=1}^k X_i^*X_i\] satisfies $\alpha_c(\Delta_H)>0$.
\end{theorem}
\begin{proof}
Recall the following Sobolev-type inequality (see e.g. \cite[Lemma 2.1]{lugiewicz2007coercive})
\begin{align}\label{eq:sobolev}
\|f\|_{q} \pl \le \pl C \big( \langle \Delta_H f,f\rangle + \norm{f}{2}^2\big)^{1/2} \hspace{.1cm},
\end{align}
where $q=\frac{2dl_H}{dl_H-2}>2$ and $l_H$ is globoal Lie bracket length needed in the H\"ormander condition. By Varopoulos' Theorem (see \cite[Chapter 2]{varopoulos1991analysis}) on the dimension of semigroups, this implies the following ultra-contractivity estimate for $P_{t}=e^{-\Delta_Ht}$
\begin{align}\label{eq:ultra}
\norm{P_t:L_{1}(M,\mu)\to L_{\infty}(M,\mu)}{}\leqslant C^{\prime} t^{-m/2} \text{ for } 0 < t \le 1\hspace{.1cm},
\end{align}
where $m=dl_x$. Also, it was proved in \cite[Theorem 2.3]{lugiewicz2007coercive} that $\Delta_H$ always satisfies Poincar\'e inequality $\lambda(\Delta_H)>0$. Combining these with our Corollary \ref{cor:classical} yield the assertion.
\end{proof}
The Sobolev-type inequality \eqref{eq:sobolev} is also used in \cite{lugiewicz2007coercive} to prove that $\alpha(\Delta_H)>0$. As in the discrete case consider by Saloff-Coste \cite{diaconis1996logarithmic}, the proof of these authors relies on the Rothaus lemma, and hence does not apply to matrix-valued functions. Therefore $\alpha_c(\Delta_H)>0$ is out of scope by the classical hypercontractivity approach.
\begin{cor} Let $(M,g)$ be a compact manifold and let $H=\{X_1,...,X_m\}$ be a H\"ormander system. Let $a_j\in L_{\infty}(\mathcal{M},\Mz_d)$ be functions of Hermitian operators. Define $Y_j=X_j+a_j$ and
\[ \tilde{\Delta}_H(f) \pl = \pl \sum_j [Y_j,[Y_j(f)]] \]
be the twisted Laplacian. Then $\alpha_c(\tilde{\Delta}_H)>0$, where $\alpha_c(\tilde{L})$ depends on $d$ and the $a_j$'s.
\end{cor}
\begin{proof}We recall that the original semigroup $P_t\otimes id_{\Mz_d}$ generated by $\Delta_H=\sum_j X_j^*X_j$ is ergodic and satisfies the Sobolev inequality that
\[ \|f-E(f)\|_{L_2({\mathbb M}_d,L_q( \mu))} \le c_q {\rm \text{tr}}\circ {\mathcal E}(f, f)^{1/2} \hspace{.1cm} .\]
This implies
\begin{align*}
\|f\|_{L_q({\mathbb M}_d,L_q( \mu))}
&\le \|f-E(f)\|_{L_q({\mathbb M}_d,L_q( \mu)} \\
&\le d^{1/2-1/q} {\rm \text{tr}}\circ {\mathcal E}(f, f)^{1/2} \hspace{.1cm} .
\end{align*}
All these estimates hold in the completely bounded setting. Now, we note that the energy form satisfies
\begin{align*}
{\rm \text{tr}}\circ {\mathcal E}(f, f) &= \sum_j \langle [X_j,f],[X_j,f]\rangle \\
&\le
2\Big( \sum_j \langle [X_j+a_j,f],[X_j+a_j,f]\rangle + \sum_j \langle[a_j,f],[a_j,f]\rangle\Big) \\
&\le 2\langle \hat{L}(f),f\rangle+ 16 \sum_j \|a_j\|_{\infty}^2 \langle f,f\rangle \hspace{.1cm} .
\end{align*}
Note that
\[ [Y_j,f] \pl = \pl [X_j,f]+ [a_j,f]
\pl = \pl X_j(f)+[a_j,f] \]
is an inner derivation which sends smooth matrix-valued functions to smooth matrix-valued functions. Thus
\[ \tilde{\Delta}_H(f) \pl = \pl \sum_j [Y_j,[Y_j,f] \]
is a generator of Markov semigroups. Let $C=16\sum_j \|a_j\|_{\infty}^2$. Then $S_t=e^{-t(\hat{L}+Cid)}$ also satisfies a Sobolev inequality with constant $c'_q=2c_q(\mathcal{X})m^{1/2-1/q}$ and hence by the cb-version of Varopoulos' theorem (see \cite{zhao2018smoothing} and
\cite[Chapter 2]{varopoulos1991analysis}) we find that
\[ \|S_t:L_1(M,\mu)\to L_{\infty}(M)\|_{cb}
\pl \le \pl (\frac{C}{2K})^{K/2} t^{-K/2} \]
holds for $\frac{1}{2}-\frac{1}{q}=\frac{1}{K}$. We may use $t=1$ and find
\[ \|S_1-E_\mu:L_1(M,\mu)\to L_{\infty}(M)\| \pl \le \pl 1+ (\frac{C}{2K})^{K/2} \hspace{.1cm} .\]
Now Proposition \ref{prop:finitetcb2} applies and implies the assertion.\end{proof}\vspace{0.5ex}
In the cases of the heat semigroup $P_t=e^{-\Delta t}$ and the Laplace-Beltrami operator $\Delta$, CMLSI on compact manifolds was obtained in \cite{brannan2022complete} via a lower bound on the Ricci curvature of the underlying Riemannian manifold. Compared to the Laplacian in the Riemannian setting, the difficulty of sub-Laplacian in sub-Riemannian setting stems from the lack of a Ricci curvature bound. Informally, at points of degeneracy of $H=\{X_i\}_{i=1}^k$, the Ricci tensor is not well-defined and might be interpreted as being $-\infty$ in some directions. This issue was only partially overcome by using a weaker variant of Bakry-Emery Curvature dimension condition in the form
\begin{align}\Gamma(P_t f,P_t f)\le C(t) P_t \Gamma(f,f)\hspace{.1cm} . \label{eq:GE}\end{align}
Recall that when $C(t)=e^{-2\lambda t}$, this is exactly Bakry-Emery's $CD(\lambda,\infty)$ curvature-dimension inequality. In \cite{gao2022complete}, Gao and Gordina proved that for a sub-Laplacian
\begin{align} \alpha_c(\Delta_H)\ge \frac{1-\varepsilon}{2(\int_{0}^{t_\varepsilon} C(t)d t)}\label{eq:GG}\end{align}
where $\varepsilon\in [0,1)$ and $t_\varepsilon$ is the $L_\infty$-mixing time
\[ t_\varepsilon=\inf\{t>0\hspace{.1cm} | \hspace{.1cm} \norm{T_{t}-E_\mu: L_1(\Omega,\mu)\to L_\infty(\Omega,\mu)}{}\le \varepsilon\}\hspace{.1cm}.\]
As the finiteness of $t_\varepsilon$ is guaranteed by the H\"ormander condition, the estimate \eqref{eq:GG} shows that the gradient estimate \eqref{eq:GE} implies $\alpha_c(\Delta_H)$. Nevertheless, even this weaker gradient estimate in the sub-Laplacian setting is known for only a limited number of examples. Our Theorem \ref{thm:sublaplacian} further overcomes this issue and obtains CMLSI for all H\"ormander systems.
\begin{exam}{\rm \label{exam:su2}The special unitary group $\operatorname{SU}\left( 2 \right)$ is
\[
\operatorname{SU}\left( 2 \right)=\{ cI+xX+yY+zZ: c^2+x^2+y^2+z^2=1, x,y,x, c \in \mathbb{R} \}.
\]
where $X,Y,Z$ are the skew-Hermitian Pauli unitary
\[
X=\left[\begin{array}{cc} 0& 1\\ -1& 0
\end{array}\right],Y=\left[\begin{array}{cc} 0& i\\ i& 0
\end{array}\right], Z=\left[\begin{array}{cc} i& 0\\ 0& -i
\end{array}\right]\hspace{.1cm}.
\]
The Lie algebra is $\mathfrak{su}(2)=\operatorname{span}_{\mathbb{R}}\{X,Y,Z\}$ with Lie bracket rules as
\begin{align}[X,Y]=2Z\ , [Y,Z]=2X\ , [Z,X]=2Y\hspace{.1cm}. \label{eq:lie}\end{align}
The canonical sub-Riemannian structure is given by $H=\{X,Y\}$, a generating set of $\mathfrak{g}$ because $[X,Y]=2Z$. Thus the H\"ormander's condition is satisfied. The associated sub-Laplacian is
\begin{align}\label{eq:su2}
\Delta_H=-(X^2+Y^2).
\end{align}
The semigroup $P_t=e^{-\Delta_H t}$ on $\operatorname{SU}\left( 2 \right)$ has been studied as a prototype of horizontal heat semigroup. In particular, Baudoin and Bonnefont in \cite{baudoin2009subelliptic} proved that
\begin{align}\label{eq:baudoin}
\Gamma(P_t f,P_t f)\leqslant C e^{-4 t}P_t(\Gamma(f, f)),
\end{align}
for some constant $C>0$. Then one can apply \eqref{eq:GG} for $ t_0=\infty,\varepsilon=0$ to obtain
\[ \alpha_c(\Delta_H)\ge (2\int_0^\infty C e^{-4 t}dt )^{-1}=\frac{2}{C}\hspace{.1cm}.\]
}
\end{exam}
\begin{exam}{\rm Let $n\ge 3$. The special orthogonal group $\operatorname{SO}\left( n \right)$ is
\[
\operatorname{SO}\left( n \right)=\{ Q\in {\mathbb M}_n(\mathbb{R})\hspace{.1cm}|\hspace{.1cm} Q^tQ=1\hspace{.1cm}, \hspace{.1cm} \det(Q)=1\}.
\]
where $Q^\perp$ is the transpose of $Q$. Denote $e_{ij}$ as the matrix unit at $(i,j)$ position. The Lie algebra $\mathfrak{SO}(n)$ is the space of all traceless real matrices, spanned by
\[
R_{jk}=e_{jk}-e_{kj}\hspace{.1cm}, 1\le j<k\le n\hspace{.1cm}.
\] Let $V=\{1,\cdots, n\}$ be a vertex set and $E\subset V\times V$ as an edge set.
The set $H_E=\{R_{jk}\hspace{.1cm} |\hspace{.1cm} (j,k)\in E\}$ is a generating set of $\mathfrak{SO}(n)$ (hence H\"ormander system) if and only if $(V,E)$ is a connected graph. Our theorem \eqref{thm:sublaplacian} asserts that the associated sub-Laplacian
\[\Delta_E=\sum_{(j,k)\in E} R_{jk}^2\hspace{.1cm}, \]
satisfies $\alpha_c(\Delta_E)>0$ for all connected $(V,E)$. To the best of our knowledge, the gradient estimate \eqref{eq:GE} is unknown for any this type of generator, but the combination of heat kernel estimates and spectral gaps provides the CMLSI.
}
\end{exam}
\begin{exam}{\rm Let $n\ge 3$. The special unitary group $\operatorname{SU}\left( n \right)$ is
\[
\operatorname{SU}\left( n \right)=\{ u\in {\mathbb M}_n\hspace{.1cm}|\hspace{.1cm} u^*u=1\hspace{.1cm}, \hspace{.1cm} \det(u)=1\hspace{.1cm}.\}.
\]
The Lie algebra $\mathfrak{su}(n)$ is the space of all the skew-Hermitian matrices.
A natural basis $\mathfrak{su}(n)$ is given by $X_{j,k},Y_{j,k},Z_{j,k}, 1\le j< k\le n$ where
\[
X=e_{jk}-e_{kj}\hspace{.1cm} ,\hspace{.1cm} Y=i(e_{jk}+e_{kj})\hspace{.1cm} ,\hspace{.1cm} Z_k=i(e_{11}-e_{kk})
\]
which is $n^2-1$ dimensional. Similar to the previous example, the set \[H_E=\{X_{j,k},Y_{j,k} \hspace{.1cm} |\hspace{.1cm} (j,k)\in E\}\] is a H\"ormander system if and only if $(V,E)$ is a connected graph. The associated sub-Laplacian
\[\Delta_E=\sum_{(j,k)\in E} X_{j,k}^2+Y_{j,k}^2\hspace{.1cm}, \]
is a generalization of \eqref{eq:su2}. Again, our Theorem \eqref{thm:sublaplacian} implies that $\alpha_c(\Delta_E)>0$ for all connected $(V,E)$, but the gradient estimate \eqref{eq:GE} is not known for this type of generator.
}
\end{exam}
Let us discuss an immediate application of $\alpha_c(\Delta_H)>0$ towards symmetric Quantum Markov semigroup. Let $G$ be a compact Lie group and $H=\{X_1,\cdots,X_k\}$ be a generating set of its Lie algebra $\mathfrak{g}$. Then $\{X_1,\cdots,X_k\}$ satisfies H\"ormander condition and its sub-Laplacian
$\Delta_H=-\sum_{k}X_i^2$
generates a Markov semigroup $P_t=e^{-\Delta_H t}$ symmetric to the Haar measure. Let $u:G\to {\mathbb M}_n$ be a finite dimensional unitary representation and $d_u:\mathfrak{g}\to i({\mathbb M}_n)_{s.a.}$ be the corresponding Lie algebra morphism. $P_t=e^{-\Delta_H t}$ induces a quantum Markov semigroup $T_t=e^{-L_Xt}:{\mathbb M}_m\to{\mathbb M}_m$ with generator in the Lindbladian form \cite{lindblad1976generators}
\[L_H(\rho)=-\sum_{i=1}^k [d_u(X_i),[d_u(X_i),\rho]]\hspace{.1cm}.\]
The connection is given by the following commuting diagram
\begin{equation}\label{dia:transference}
\begin{tikzcd}
L^{\infty}(G, {\mathbb M}_m) \arrow{r}{P_t\otimes \iota_{\infty,2}^n_{{\mathbb M}_m}} & L^{\infty}(G, {\mathbb M}_m)
\\
{\mathbb M}_m \arrow{u}{\pi_u} \arrow{r}{T_{t}} & {\mathbb M}_m \arrow{u}{\pi_u},
\end{tikzcd}
\end{equation}
where the map $\pi_u$ is a $*$-endomorphism given by
\begin{align*}
\pi_u: {\mathbb M}_m\to L^{\infty}(G, {\mathbb M}_m)\hspace{.1cm},\hspace{.1cm} \pi_u(\rho)(g)=u(g)^*(\rho)u(g)\hspace{.1cm}.
\end{align*}
Such a $^*$-homomorphism $\pi_u$ is called a transference map, which embeds the matrix algebra ${\mathbb M}_m$ into the matrix-valued functions $L^{\infty}(G, {\mathbb M}_m)$ on the group $G$. By the diagram \eqref{dia:transference}, the quantum semigroup $T_t$ is exactly the restriction of the matrix-valued classical semigroup $P_t\otimes \iota_{\infty,2}^n_{{\mathbb M}_m}$ onto the image of $\pi({\mathbb M}_m)$. Such a transference relation holds for any (projective) unitary representation, which yields the following dimension-free estimate for both spectral gap and CMLSI constant (see \cite[Section 4]{gao2020fisher}),
\[\alpha_{c}(\Delta_X)\le \alpha_{c}(L_X)\hspace{.1cm} , \lambda(\Delta_X)\le \lambda(L_X)\hspace{.1cm}.\]
Combined with Theorem \ref{thm:sublaplacian}, we have the following dimension independent bound on matrix algebra.
\begin{cor}
Let $G$ be a compact Lie group and $H=\{X_1,\cdots,X_k\}$ be a generating set of its Lie algebra $\mathfrak{g}$. There exists a constant $\alpha=\alpha_{c}(\Delta_H)>0$ such that for all unitary representation $u$, the induced quantum Markov semigroup generated by
\[L_H(\rho)=-\sum_{i=1}^k [d_u(X_i),[d_u(X_i),\cdot]]\]
satisfies
\[ \alpha_{c}(L_X)\ge \alpha>0\hspace{.1cm}.\]
\end{cor}
\subsection{Failure of Matrix valued Log-Sobolev inequality}\label{sec:failureLSI}
As mentioned above, one standard approach to MLSI (or equivalently, exponential entropy decay) is to use the connection between Hypercontractivity and LSI, in which the Rothaus Lemma (see e.g. \cite{rothaus1985analytic,bakry1994hypercontractivite}) below plays an important role
\[H(|f|^2)\le H(|f-E_\mu(f)|^2)+\norm{f-E_\mu(f)}{2}^2\hspace{.1cm}.\]
In \cite{beigi2016hypercontractivity}, Beigi and King show that such an equivalence can also be extended to completely bounded setting. Nevertheless, Bardet and Rouz\'e shows \cite{bardet2022hypercontractivity} that neither Hypercontractivity nor LSI holds for non-primitive quantum Markov semigroups on matrix algebras ${\mathcal M}={\mathbb M}_m$. In this subsection, we show that Rothaus Lemma, Log-Sobolev inequality and Hypercontractivity all fail for matrix valued functions for any classical Markov semigroup. This is a strong indicator that the approach by Diaconis and Saloff-Coste's hypercontractive \cite{diaconis1996logarithmic} estimate (also used in \cite{lugiewicz2007coercive}) is destined to fail in proving lower bounds for the CMLSI constants.
We start with a lemma calculating the derivatives of the entropy functional $H(\rho)=\tau(\rho\log \rho)$. Recall the BKM metric of a operator $X\in {\mathcal M}$ at the base state $\rho$ is
\[\gamma_\rho(X)=\int_{0}^\infty\tau(X^*(\rho+s)^{-1}X (\rho+s)^{-1})\]
\begin{lemma}\label{eq:2nd}
Let $t\mapsto \rho_t \in S_B({\mathcal M}), t\in (a,b)$ be a smooth family of bounded density operator.
Denote
\begin{align*} F(t)=H(\rho_t)=\tau(\rho_t\log \rho_t)\hspace{.1cm}.
\end{align*}
Then
\begin{align*}F'(t)=\tau(\rho_t'(\log \rho_t+1))\hspace{.1cm},\hspace{.1cm} F''(t)=\tau(\rho_t''(\log \rho_t+1))+\gamma_{\rho_t}(\rho'_t)
\end{align*}
where $\rho_t'$ and $\rho_t''$ are the first and second order derivative of $\rho_t$.
\end{lemma}
\begin{proof}
The formula for $F'$ follows from Lemma \cite[Lemma 5.8]{wirth2018noncommutative}. For the second derivative, recall the noncommutative chain rule
\[\frac{d}{dt}(\log \rho_t)=\int_{0}^\infty(\rho_t+s)^{-1}\rho_t'(\rho_t+s)^{-1}ds\hspace{.1cm}.\]
By calculating the second derivative, we obtain the second assertion
\begin{align*}
F''(t)=&\tau(\rho_t''(\log \rho_t+1))+\int_{0}^\infty \tau(\rho_t'(\rho_t+s)\rho_t'(\rho_t+s))ds
\pl = \pl \tau(\rho_t''(\log \rho_t+1))+\gamma_{\rho_t}(\rho'_t) \qedhere
\end{align*}
\end{proof}
\begin{prop}\label{prop:rothaus}
Let $(\Omega,\mu)$ be a probability space and $T_t=e^{-tL}:L_\infty(\Omega,\mu)\to L_\infty(\Omega,\mu)$ be a primitive symmetric Markov semigroup. Denote $E_\mu(f)=(\int f d\mu)1_\Omega$ as the expectation map and $\hat{f}=f-E_\mu(f)$ as the mean zero part of $f$. Let $\alpha_R,\alpha_2,\alpha_h$ be the optimal (largest) constant such that the following inequalities holds for any $f\in L_\infty(\Omega,{\mathbb M}_2)\cap {\mathcal A}_{\mathcal E}$
\begin{align}&\alpha_R D\Big(|f|^2 \left| \right|E_\mu(|f|^2)\Big)\le D\Big(|\hat{f}|^2 \left| \right| E_\mu(|\hat{f}|^2)\Big)+\norm{\hat{f}}{2}^2\hspace{.1cm}, \tag{\text{Rothaus Lemma}} \\
&\alpha_2 D(f^2||E_\mu(f^2))\le 2{\mathcal E}(f,f) \tag{\text{LSI}}\\
&\norm{T_tf}{L_2({\mathbb M}_2,L_{p(t)}(\Omega))}\le \norm{f}{L_2({\mathbb M}_2,L_2(\Omega))} \text{ for } p(t)=1+e^{2\alpha_h t} \tag{\text{Hypercontractivity}}
\end{align}
Then $\alpha=\alpha_2=\alpha_h=0$.
\end{prop}
\begin{proof}
We write $\tau(f)=\int {\rm \text{tr}}(f)d\mu$ for the normalized trace on $L_\infty(\Omega,{\mathbb M}_2)$. We start with constant $\alpha_R$ in Rothaus Lemma. Without loss of generality, we may assume there is a measurable set $X\subset \Omega$ such that $\mu(X)=r$ with $0<r<1$. Let $\eta\in(0,1)$. Then $h_0=(1-r)1_X-r 1_{X^c}$ is a real mean zero function.
Consider the matrix valued function
$f_\varepsilon=f+\varepsilon h$ where
\[f=\left[\begin{array}{cc}1+\eta &0 \\ 0& 1-\eta
\end{array}\right]{\bf 1} \hspace{.1cm}, \hspace{.1cm} h=\left[\begin{array}{cc}0 & h_0 \\ h_0& 0
\end{array}\right]\hspace{.1cm},
\]
where $f$ is a constant matrix valued function.
Then $E_\mu f_\varepsilon=f, \hat{f}_\varepsilon=\varepsilon h$ and
\begin{align*}&f_\varepsilon^2=(f+\varepsilon h)^2=f^2+\varepsilon(fh+hf)+\varepsilon^2 h^2=f^2+2\varepsilon h+\varepsilon^2 h^2
\end{align*}
Then $E_\mu(|f_\varepsilon|^2)=f^2+\varepsilon^2 h^2$ and $E_\mu(|\hat{f}_\varepsilon|^2)=E_\mu( h^2)\varepsilon^2$. Using Lemma \eqref{eq:2nd}, the Taylor expansion of the left hand side of (LSI) is
\begin{align*}
D\Big(|f|^2 \left| \right|E_\mu(|f|^2)\Big)=&D(f^2+2\varepsilon h+\varepsilon^2 h^2\left| \right|f^2+\varepsilon^2 h^2)\\
=&H(f^2+2\varepsilon h+\varepsilon^2 h^2)-H(f^2+\varepsilon^2 h^2)\\
=&2\tau(h\log f)\varepsilon+\big(2\tau(h^2(\log f+1))+\gamma_{f}(2h)\big)\varepsilon^2+O(\varepsilon^3)\\&-\tau(2E(h^2)(\log f+1))\varepsilon^2+O(\varepsilon^3)\\
=&\gamma_{f}(2h)\varepsilon^2+O(\varepsilon^3) \hspace{.1cm} ,
\end{align*}
where we used the fact $\tau(h\log f)=0$ and $\tau(h^2\log f-E_\mu(h^2)\log f)=0$. For the right hand side of (Rothaus Lemma), we find
\begin{align*}
D(|\hat{f}_\varepsilon|^2||E_\mu(|\hat{f}_\varepsilon|^2))=D(h^2||E_\mu( h^2))\varepsilon^2\hspace{.1cm}, \hspace{.1cm}
\norm{\hat{f}_\varepsilon}{2}^2=\norm{h}{2}^2\varepsilon^2\hspace{.1cm}.
\end{align*}
While both $D(h^2||E_\mu( h^2))$ and $\norm{h}{2}^2$ are finite, we have
\begin{align*}\gamma_{f}(2h)=&\int_{0}^\infty\tau(h(f+s)^{-1}h (f+s)^{-1})ds\\
=&\int_{0}^\infty \int_\Omega {\rm \text{tr}}(\left[\begin{array}{cc}\frac{1}{(1-\eta+s)(1+\eta+s)}h^2 &0 \\ 0& \frac{1}{(1-\eta+s)(1+\eta+s)}h^2
\end{array}\right]1_\Omega )d\mu ds\\
=&\Big(\int_{0}^\infty \frac{1}{(1-\eta+s)(1+\eta+s)} ds\Big)\norm{h}{2}^2
\\ =& \frac{1}{2\eta}\ln\frac{1+\eta}{1-\eta}\norm{h}{2}^2
\end{align*}
Note that we can choose $\eta\to 1$ and $\frac{1}{2\eta}\ln(\frac{1+\eta}{1-\eta})\to +\infty$. This implies $\alpha_R=0$. The same example applies to LSI by choosing a mean zero function $h_0$ such that ${\mathcal E}(h_0,h_0)<\infty$. For the hypercontractivity, for $p\ge 2$ we recall the norms
\begin{align*} &\norm{f}{L_2({\mathbb M}_2, L_2(\Omega))}=\norm{f}{L_2({\mathbb M}_2)\otimes L_2(\Omega)}=(\int {\rm \text{tr}}(f^*f)d\mu)^{1/2}\hspace{.1cm}. \\
&\norm{f}{L_2({\mathbb M}_2, L_p(\Omega))}=\inf_{x,y\in ({\mathbb M}_2)_+\hspace{.1cm},\hspace{.1cm} \norm{\hspace{.1cm} x\hspace{.1cm} }{2r}=\norm{\hspace{.1cm} y\hspace{.1cm}}{2r}=1}\norm{x^{-1}fy^{-1}}{L_2({\mathbb M}_2, L_2(\Omega))}\end{align*}
where the infimum takes over all positive invertible $x,y\in {\mathbb M}_2$ with unit $2r$-norm for $\frac{1}{r}=\frac{1}{2}-\frac{1}{p}$. Since $T_t$ is a bimodule map for $\mathbb{C}1\otimes {\mathbb M}_2\subset L_\infty(\Omega,{\mathbb M}_2)$, we can equivalently consider the norm
\[\norm{T_t: L_2({\mathbb M}_2, L_2(\Omega))\to L_2({\mathbb M}_2, L_p(\Omega))}{}=\norm{T_t: L_2({\mathbb M}_2, L_2(\Omega))\to L_2^a({\mathbb M}_2, L_p(\Omega))}{}\hspace{.1cm},\]
where the asymmetric amalgamated $L_2^a({\mathbb M}_2, L_p(\Omega))$ space is equipped with norm
\begin{align*}\norm{f}{L_2^a({\mathbb M}_2, L_p(\Omega))}=\inf_{x\in ({\mathbb M}_2)_+\hspace{.1cm}, \norm{\hspace{.1cm} x\hspace{.1cm} }{r}=1}\norm{fa^{-1}}{L_2({\mathbb M}_2, L_2(\Omega))} \end{align*}
In particular,
\[ \norm{f}{L_2^a({\mathbb M}_2, L_p(\Omega))}^2=\norm{f^*f}{L_1({\mathbb M}_2, L_{\frac{p}{2}}(\Omega))}\]
and we have
\[D(|f|^2||E(|f|^2))=\lim_{q\to 1^+}\frac{\norm{|f|^2}{L_1({\mathbb M}_2, L_q(\Omega))} -\norm{|f|^2}{1}}{q-1 }\]
Now define $p(t)=2q(t)=1+e^{2\alpha_h t}$
\[G(t)=\norm{T_t f}{L_2^a({\mathbb M}_2, L_p(t)(\Omega))}^2=\norm{|T_tf|^2}{L_1({\mathbb M}_2, L_{q(t)}(\Omega))}\hspace{.1cm} \hspace{.1cm}, \]
By assumption $G(t)\le 1$, we have
\[G'(0)=-2{\mathcal E}(f,f)+\alpha_hD(|f|^2||E(|f|^2))\le 0,\]
which implies $\alpha_h\le \alpha_2=0$. Note however, that $\alpha_h\ge 0$ because $T_t$ is always contractive on $L_2({\mathbb M}_2, L_2(\Omega))$, and the proof is complete.
\end{proof}
\begin{rem}{\rm Similar to \cite[Corollary 5.2]{bardet2022hypercontractivity}, the above proposition implies that for $p\neq 2$, neither $L_2^a({\mathbb M}_2, L_p(\Omega))$ and $L_2({\mathbb M}_2, L_p(\Omega))$ are uniformly convex.}
\end{rem} |
\section{Introduction}\label{sec:introduction}}
\else
\section{Introduction}
\label{sec:introduction}
\fi
\IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file''
for IEEE Computer Society journal papers produced under \LaTeX\ using
IEEEtran.cls version 1.8b and later.
I wish you the best of success.
\hfill mds
\hfill August 26, 2015
\subsection{Subsection Heading Here}
Subsection text here.
\subsubsection{Subsubsection Heading Here}
Subsubsection text here.
\section{Conclusion}
The conclusion goes here.
\appendices
\section{Proof of the First Zonklar Equation}
Appendix one text goes here.
\section{}
Appendix two text goes here.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
The authors would like to thank...
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\section{Introduction}
This demo file is intended to serve as a ``starter file''
for IEEE Computer Society conference papers produced under \LaTeX\ using
IEEEtran.cls version 1.8b and later.
I wish you the best of success.
\hfill mds
\hfill August 26, 2015
\subsection{Subsection Heading Here}
Subsection text here.
\subsubsection{Subsubsection Heading Here}
Subsubsection text here.
\section{Conclusion}
The conclusion goes here.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
The authors would like to thank...
\section{Conclusion and Future Work}
In this work, we revealed that the aleatoric uncertainty can measure the mislabeling and tackle the positive-unlabeled issue. In this light, we proposed the AUR framework to estimate and leverage the aleatoric uncertainty.
To justify the rationality and effectiveness of AUR, we gave a theoretical proof
and conducted experiments on two real-world datasets with three different evaluation protocols.
The results demonstrated that our method can achieve better recommendation performance on tail items without sacrificing head items.
This work is an innovative attempt to introduce aleatoric uncertainty into recommendation.
In the future, we will extend our method to other training paradigms, such as BPR and BCE.
We are also interested in designing more powerful framework to estimate the uncertainty.
In addition, we will investigate whether the aleatoric uncertainty can mitigate other biases such as exposure bias~\cite{2016Modeling} and confounding bias~\cite{DCR}. Besides, we will further explore how to design a more effective uncertainty estimator.
\section{EXPERIMENTS}\label{sec:experiment}
In this section, we conduct experiments on two real-world datasets to answer the following questions:
\textbf{RQ1:}
How is the performance of AUR compared with expectation-based recommender models, the debiasing methods, and the probabilistic methods?
\textbf{RQ2:} How does each design of AUR affect its effectiveness \emph{w.r.t. } the recommendation performance?
\textbf{RQ3:} Do recommendations given by uncertainty exhibit properties different from those given by expectation?
\subsection{Experimental Settings}\label{ssec:exp_setting}
\textbf{Datasets. }
We conduct experiments on two benchmark datasets: Yelp2018\footnote{https://www.yelp.com/dataset} (Yelp) and Amazon-book\footnote{https://nijianmo.github.io/amazon/index.html} (Amazon), which are abbreviated as Yelp and Amazon, respectively.
We follow the setting of LightGCN~\cite{he2020lightgcn} to filter users and items, and divide each dataset into training and testing sets.
Considering that we need to evaluate the tail item recommendation performance, for both datasets, we define the tail item as the item with less interactions, and make sure that all tail items occupy $50\%$ of all interactions in the corresponding dataset. We treat the remaining items as the head items. Finally,
$86.3\%$ and $84\%$ of items are tail items in Yelp and Amazon, respectively.
\begin{table*}
\caption{Performance comparison of top-$K$ recommendation performance between expectation-based models and AUR on Yelp and Amazon, under the Tail Absolute, Tail Relative, and Overall evaluation protocols. The metric Recall and NDCG are denoted as R and N, respectively, and RI denotes the relative improvement of AUR over the corresponding expectation model.}
\vspace{-0.3cm}
\small
\resizebox{\linewidth}{35mm}{
\begin{tabular}{|c|c|cccc|cccc|cccc|}
\hline
\multirow{2}{*}{Datasets} & \multirow{2}{*}{Methods} & \multicolumn{4}{c|}{Tail Absolute} & \multicolumn{4}{c|}{Tail Relative} & \multicolumn{4}{c|}{Overall} \\ \cline{3-14}
& & R@20 & R@50 & N@20 & N@50 & R@20 & R@50 & N@20 & N@50 & R@20 & R@50 & N@20 & N@50 \\ \hline
\multirow{9}{*}{Yelp} & MF & 0.0046 & 0.0180 & 0.0025 & 0.0071 & 0.0522 & 0.1010 & 0.0343 & 0.0498 & 0.0639 & 0.1215 & 0.0531 & 0.0745 \\
& MF-AUR & 0.0184 & 0.0459 & 0.0112 & 0.0203 & 0.0650 & 0.1230 & 0.0430 & 0.0614 & 0.0689 & 0.1316 & 0.0563 & 0.0796 \\
& RI & 300.00\% & 155.00\% & 348.00\% & 185.91\% & 24.52\% & 21.78\% & 25.36\% & 23.29\% & 7.84\% & 8.31\% & 6.03\% & 6.85\% \\ \cline{2-14}
& LGCN & 0.0081 & 0.0296 & 0.0044 & 0.0116 & 0.0590 & 0.1126 & 0.0390 & 0.0562 & 0.0673 & 0.1280 & 0.0557 & 0.0782 \\
& LGCN-AUR & 0.0149 & 0.0426 & 0.0085 & 0.0177 & 0.0636 & 0.1211 & 0.0419 & 0.0602 & 0.0687 & 0.1307 & 0.0562 & 0.0793 \\
& RI & 83.95\% & 43.92\% & 93.18\% & 52.59\% & 7.80\% & 7.55\% & 7.44\% & 7.12\% & 2.08\% & 2.11\% & 0.90\% & 1.41\% \\ \cline{2-14}
& VAE & 0.0103 & 0.0316 & 0.0060 & 0.0133 & 0.0611 & 0.1157 & 0.0409 & 0.0584 & 0.0688 & 0.1302 & 0.0574 & 0.0801 \\
& VAE-AUR & 0.0196 & 0.0480 & 0.0123 & 0.0218 & 0.0648 & 0.1218 & 0.0430 & 0.0613 & 0.0698 & 0.1328 & 0.0575 & 0.0809 \\
& RI & 90.29\% & 51.90\% & 105.00\% & 63.91\% & 6.06\% & 5.27\% & 5.13\% & 4.97\% & 1.45\% & 2.00\% & 0.17\% & 1.00\% \\ \hline
\multirow{9}{*}{Amazon} & MF & 0.0036 & 0.0098 & 0.0022 & 0.0043 & 0.0374 & 0.0679 & 0.0251 & 0.0349 & 0.0412 & 0.0777 & 0.0325 & 0.0461 \\
& MF-AUR & 0.0235 & 0.0448 & 0.01524 & 0.0222 & 0.0637 & 0.1060 & 0.0439 & 0.0575 & 0.0550 & 0.1003 & 0.0434 & 0.0602 \\
& RI & 552.78\% & 354.14\% & 592.73\% & 416.28\% & 70.32\% & 56.11\% & 74.90\% & 64.76\% & 33.50\% & 29.09\% & 33.54\% & 30.59\% \\ \cline{2-14}
& LGCN & 0.0072 & 0.0190 & 0.0043 & 0.0083 & 0.0381 & 0.0708 & 0.0254 & 0.0359 & 0.0436 & 0.0820 & 0.0347 & 0.0489 \\
& LGCN-AUR & 0.0229 & 0.0468 & 0.0145 & 0.0224 & 0.0525 & 0.0918 & 0.0356 & 0.0483 & 0.0511 & 0.0936 & 0.0399 & 0.0557 \\
& RI & 218.06\% & 146.32\% & 237.21\% & 169.88\% & 37.80\% & 29.66\% & 40.16\% & 34.54\% & 17.20\% & 14.15\% & 14.99\% & 13.91\% \\ \cline{2-14}
& VAE & 0.0114 & 0.0293 & 0.0067 & 0.0126 & 0.0557 & 0.0955 & 0.0378 & 0.0508 & 0.0532 & 0.0982 & 0.0418 & 0.0586 \\
& VAE-AUR & 0.0285 & 0.0543 & 0.0187 & 0.0272 & 0.0607 & 0.1031 & 0.0416 & 0.0553 & 0.0581 & 0.1053 & 0.0458 & 0.0635 \\
& RI & 150.00\% & 85.32\% & 179.10\% & 155.87\% & 8.98\% & 7.96\% & 10.52\% & 8.86\% & 9.42\% & 7.23\% & 9.57\% & 8.36\% \\ \hline
\end{tabular}
}
\label{main result}
\end{table*}
\noindent \textbf{Evaluation Protocols. }
To measure the recommendation performance, we adopt two widely-used evaluation metrics: $Recall@K$ and $NDCG@K$ ($K= 20\,or\,50$). To evaluate both the overall recommendation and tail item recommendation performance, we define three evaluation protocols to rank items and compute the above metrics, as following:
\begin{itemize} [leftmargin=*]
\item \textbf{Overall} evaluation. In this protocol, we conduct \textbf{all-ranking}~\cite{PDA} -- all items that are not interacted by a user are the candidates and report the normal Recall and NDCG. Indeed, it measures the trade-off between recommendation of head items and tail items.
\item \textbf{Tail Absolute} evaluation. In this protocol, we focus on measuring the accuracy of recommending tail items when both head and tail items are candidates. We still conduct \textbf{all-ranking}, but just take the positive tail items in the testing set as ground-truth. We adjust the evaluation metrics,
\emph{e.g., } $Recall@K=\frac{1}{|\mathcal{U}|}\sum_{u}\frac{|\mathcal{T}_u \cap \mathcal{R}^{K}_u|}{|\mathcal{T}_u|}$,
where $T_{u} $ denotes the set of tail items that are interacted by the user $u$ in the testing set. $\mathcal{R}^{K}_{u}$ denotes the top-$K$ recommended items among all candidates in the all-ranking protocol, and $|\cdot |$ denotes the size of the corresponding set.
\item \textbf{Tail Relative} evaluation. It also evaluates the tail recommendation performance but focuses on the ranking of tail items where all tail items not interacted by a user are the candidates. Correspondingly, we adjust the metric
$Recall@K=\frac{1}{|\mathcal{U}|}\sum_{u}\frac{|\mathcal{T}_u \cap {\mathcal{R}^{'}}^{K}_u|}{|\mathcal{T}_u|}$,
where $ {\mathcal{R}^{'}}^{K}_u $ denotes the top-${K}$ recommended tail items.
\end{itemize}
\vspace{5pt}
\noindent \textbf{Compared Methods.} The proposed AUR generates recommendations with the consideration of uncertainties. To verify the effectiveness of the AUR framework, we implement AUR with three backbone models (MF, LGCN, and VAE), and compare AUR with corresponding backbone models that generate recommendations just based on expectations.
\begin{itemize}[leftmargin=*]
\item \textbf{MF}~\cite{2009Collaborative}. This refers to the Matrix Factorization method, which predicts user preference on an item as the inner product of user and item embedding.
\item \textbf{LGCN}~\cite{he2020lightgcn} refers to LightGCN, which takes the simplified graph convolution networks (GCN) to better capture the collaborative information in the user-item interaction graph. We fix the number of GCN layers to $3$ due to its good performance on both Yelp and Amazon~\cite{he2020lightgcn}.
\item \textbf{VAE} \cite{liang2018variational} refers to MultiVAE.
This is an item-based CF method built with the variational autoencoder (VAE).
We set the network architecture as: $1024 \to 512 \to 1024$ and take Equation~\eqref{loss_backbone}
as the optimization objective.
\item \textbf{MF-AUR/LGCN-AUR/VAE-AUR}, refer to the AUR with MF, LGCN, and VAE as the expectation estimator, respectively. We set the network size of $q_i$ and $p_i$ in the uncertainty estimator as 1024 to fairly compare with VAE.
\end{itemize}
To verify the ability of AUR to mitigate the popularity bias, we also compare MF-AUR with the following methods designed for addressing popularity bias applied to MF:
\begin{itemize}[leftmargin=*]
\item \textbf{IPS}~\cite{liang2016causal}. This is a famous de-biasing method --- inverse propensity scoring, which re-weights training samples to improve the tail recommendation.
\item \textbf{MACR} \cite{wei2021model}. This is a SOTA method to eliminate the popularity bias in recommendation. It takes the counterfactual inference to remove the influence of popularity on recommendations, promoting the recommendation of tail items. We use the code released by the authors and tune hyperparameters in the same range as the MACR paper.
\end{itemize}
Moreover, we compare MF-AUR with two well-known probabilistic recommender models:
\begin{itemize}[leftmargin=*]
\item \textbf{PMF} \cite{2007Probabilistic}, which assumes that latent features and labels obey different Gaussian distributions. The variance of each distribution is treated as a hyper-parameter.
For fair comparison, we implement PMF with the same mini-batch strategy as AUR and tune all hyper-parameters in the ranges taken by the PMF paper.
\item \textbf{ExpoMF} \cite{2016Modeling}.
ExpoMF incorporates user exposure to items into collaborative filtering. It could avoid suppressing potential positive but mislabeled samples by decreasing the weights of negative samples exposed with lower probabilities. Regarding the exposure model of ExpoMF, we follow the 'per-item' strategy to implement exposure covariates, \emph{i.e., } encoding exposure via item popularity.
\end{itemize}
\noindent \textbf{Hyper-parameter Settings. }
We optimize all models with Adam~\cite{2014Adam} and use the default learning rate $1e\text{-}4$ and default mini-batch size $32$. The $L_2$ regularization coefficient is searched in the range of $\{1e\text{-}4,1e\text{-}3,1e\text{-}2,1e\text{-}1\}$. For baselines with embedding layers (MF, LGCN, IPS, and MACR) and the expectation estimators of MF-AUR and LGCN-AUR, the embedding size is fixed to $128$. For MF, LGCN, VAE and all the expectation estimators of AUR, we take use of the same training strategy described in Section~\ref{expectation_training}, and the sampling rate $\mu$ of this strategy is searched in the range of $\{0.01,0.1,0.2,0.4\}$, and in most case the optimal value of $\mu$ is $0.1$. For AUR, $\gamma$ is searched in the range of $\{1e\text{-}3,1e\text{-}2\}$, and the best values are $1e\text{-}3$ and $1e\text{-}2$ for Yelp and Amazon, respectively; $\beta$ is fixed to $1e\text{-}2$, and $\alpha$ is searched in the range of $\{1,2,3,4,5\}$ and the default choice is $1$; $\lambda$ is searched from $\{0,0.2,0.4,0.6,0.8,1.0\}$.
\subsection{Performance Comparison (RQ1)}
In this subsection, we first compare the proposed AUR with different expectation models (\emph{i.e., } MF, LGCN and VAE), then compare AUR with SOTA debiasing methods (IPS and MACR) and probabilistic approaches (PMF and ExpoMF). Lastly, we discuss the fairness of the comparisons.
\noindent\textbf{Compared with Expectation Models.}\label{main_exp}
We compare three expectation models (MF, LGCN, and VAE) with the corresponding versions of AUR (MF-AUR, LGCN-AUR and VAE-AUR). Table~\ref{main result} shows their recommendation performance on the two datasets under the three evaluation protocols. From the table, we have the following observations:
\begin{itemize}[leftmargin=*]
\item Regarding the Overall evaluation, MF-AUR, LGCN-AUR, VAE-AUR distinctly outperform the corresponding expectation models --- MF, LGCN and VAE on Amazon.
On Yelp, MF-AUR still distinctly outperforms its corresponding expectation model MF; LGCN-AUR and VAE-AUR show comparable performance to LGCN and VAE. These results verify our theoretical analysis that taking uncertainty to generate recommendation scores will be at least on par with the expectation-based recommendation on overall performance. As the Amazon dataset is sparser than Yelp, which means a higher chance of mislabeling, the higher performance boost on Amazon than Yelp shows that uncertainty is more important when observed interactions are insufficient.
\item Regarding the Tail Absolute evaluation, all versions of the proposed AUR greatly outperform the corresponding expectation models. Specifically, compared to the corresponding expectation-based models, the averaged RI of AUR is $131.1\%$ for Yelp and $268.4\%$ for Amazon. This verifies that AUR has a higher opportunity to retrieve relevant tail items even if the head items are candidates, \emph{i.e., } the recommendations are not biased towards the head items. Note that the expectation-based model usually gives low scores for the tail samples who are mislabeled as negative. As such, the huge improvements imply that AUR can push the uncertainty estimator to mitigate the impact of mislabeling.
\item Regarding the Tail Relative evaluation, each version of AUR also outperforms its corresponding expectation model. The mean RI is $12.2\%$ for Yelp and $37.0\%$ for Amazon. Note that only tail items are taken as ranking candidates in this evaluation protocol.
Better Tail Relative performance implies that the AUR can better discern which items in the tail are more relevant to a user.
Together with the results of Tail Absolute, these results show that AUR is not blindly giving higher recommendation opportunities for tail items, but finding true relevant tail items, \emph{i.e., } potential positive tail items.
\item Generally, taking different types of backbone recommenders (MF-based, graph-based and VAE-based) as the expectation estimators, AUR can always show better tail performance and comparable even better overall performance on top-$K$ recommendation. This verifies that AUR is agnostic to the architecture of the expectation estimator. It also indicates that AUR can be grafted onto most existing models without architecture changes or even retraining. This will reduce the cost of deploying AUR for industry applications.
Meanwhile, getting better tail performances like AUR is also important for real-world applications. For example, when the item popularity drifts along time, promoting tail items that have the potential to be popular in the future could boost clicks~\cite{PDA}.
\item LGCN performs worse than VAE, which is different from the result in the original paper of LGCN~\cite{he2020lightgcn}. We think the reason is the difference of the training strategy. Note that the reported results of LGCN are better than the results reported in the LGCN paper, we believe the comparison is fair. Although LGCN outperforms MF, LGCN-AUR cannot outperform MF-AUR. This means although the AUR can improve different expectation model performance, the expectation estimator will also affect the performance of AUR, since the uncertainty estimation is based on results of the expectation estimator.
\end{itemize}
\begin{table}
\centering
\caption{The comparison between AUR, debiased models and probabilistic models regarding $Recall@50$ on Yelp and Amazon, under the Overall, Tail Absolute, and Tail Relative evaluation protocols.}
\vspace{-3mm}
\small
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{|c|c|ccc|}
\hline
Dataset & Methods & Overall & Tail Absolute & Tail Relative \\ \hline
& MF & 0.1215 & 0.0180 & 0.1010 \\
& IPS & 0.1103 & 0.0313 & 0.0916 \\
Yelp & MACR & 0.1193 & 0.0395 & 0.0950 \\
\multicolumn{1}{|l|}{} & ExpoMF & 0.1254 & 0.0315 & 0.1084 \\
\multicolumn{1}{|l|}{} & PMF & 0.1249 & 0.0226 & 0.1069 \\
& MF-AUR & \textbf{0.1316} & \textbf{0.0459} & \textbf{0.1230} \\ \hline
& MF & 0.0777 & 0.0098 & 0.0679 \\
& IPS & 0.0790 & 0.0298 & 0.0766 \\
Amazon & MACR & 0.0797 & 0.0183 & 0.0753 \\
\multicolumn{1}{|l|}{} & ExpoMF & 0.0815 & 0.0136 & 0.0718 \\
\multicolumn{1}{|l|}{} & PMF & 0.0781 & 0.0107 & 0.0682 \\
& MF-AUR & \textbf{0.1003} & \textbf{0.0448} & \textbf{0.1060} \\ \hline
\end{tabular}
}
\label{tab:sota-debiasing}
\vspace{-0.4cm}
\end{table}
\noindent\textbf{Compared with Debiased Models.}
We compare MF-AUR with IPS and MACR and present the results in Table~\ref{tab:sota-debiasing}. From the table, we can see that: 1) IPS and MACR can only achieve obvious improvements on Tail Absolute evaluation, compared to MF. While on the Overall and Tail Relative evaluations, they can even be worse than the basic MF. 2) MF-AUR can beat IPS and MACR on both the overall and tail evaluations, especially can achieve improvements in both Tail Absolute and Tail Relative evaluations.
These results show that IPS and MACR cannot recognize the relative relevance of tail items to a user more accurately than MF, \emph{i.e., } they indeed blindly promote the tail items, which is not an ideal solution for popularity bias~\cite{PDA}.
However, AUR can find truly relevant tail items from all tail candidates by identifying whether a tail item is indeed positive but mislabeled as negative.
\noindent\textbf{Compared with Probabilistic Models.}
\label{exp_prob}
Considering that AUR is also a probabilistic modeling framework, we compare the AUR framework with existing probabilistic frameworks ExpoMF and PMF. The results are also shown in Table~\ref{tab:sota-debiasing}. From the table, we can draw the following conclusions: 1) Both ExpoMF and PMF outperform the basic MF model on all three evaluation protocols.
2) Compared with debiased models, ExpoMF and PMF achieve better overall performance but lower tail performance,
while AUR brings improvements in both.
These results show that existing probabilistic approaches are better than non-probabilistic models. However, they fix the uncertainties (variance) as constants and take no account of uncertainty when generating recommendations.
They thus achieve much inferior performance than AUR, validating the superiority
of modeling the uncertainty in a learning manner and generating recommendation with the consideration of data uncertainty.
\begin{figure}[t]
\centering
\subfigbottomskip=1pt
\subfigure{
\includegraphics[width=0.8\columnwidth]{pic/tuli2.pdf}
}
\setcounter{subfigure}{0}
\subfigure[Overall performance]{
\includegraphics[width=0.44\columnwidth]{pic/overall_R20.pdf}
}
\subfigure[Tail absolute performance]{
\includegraphics[width=0.44\columnwidth]{pic/Tail_Absolute_R20.pdf}
}
\vspace{-2mm}
\caption{The influence of training strategy and embedding size on MF, MF-BPR, and MF-AUR on Amazon.}
\label{fair}
\end{figure}
\noindent\textbf{Comparison Fairness. }
Recall that we take a new training strategy in Equation (4), for MF, LGCN, and VAE, differing from typical training strategies, \emph{e.g., } optimizing BPR loss.
We question whether the training strategy decreases the effectiveness of these expectation-based models, leading to unfair comparison in Table~\ref{main result}. Thus we implement another MF with normal BPR loss, and denote it as MF-BPR. Besides, we also question whether the relative improvements of MF-AUR come from more model parameters, compared to the corresponding expectation-based model MF. Recall that the uncertainty estimator of AUR has an embedding size of $1024$, while the embedding size of MF is $128$. Thus we change the embedding size of MF, MF-BPR, expectation estimator (\emph{i.e., } MF) of MF-AUR to study its influence on model performance. The results are shown in Figure~\ref{fair}. From the table, we can find:
\begin{table}
\centering
\small
\caption{The performance of different variants of AUR \emph{w.r.t. } $Recall@50$.}
\vspace{-2mm}
\setlength{\tabcolsep}{2.5mm}{
\begin{tabular}{|c|cc|cc|}
\hline
\multirow{2}{*}{Methods} & \multicolumn{2}{c|}{Yelp} & \multicolumn{2}{c|}{Amazon} \\ \cline{2-5}
& Overall & Tail & Overall & Tail \\ \hline
MF-AUR & 0.1316 & 0.0459 & 0.1003 & 0.0448 \\
MF-AUR-M & 0.1280 & 0.0343 & 0.0900 & 0.0280 \\
MF-AUR-J & 0.1349 & 0.0406 & 0.1097 & 0.0650 \\ \hline
LGCN-AUR & 0.1296 & 0.0447 & 0.0936 & 0.0468 \\
LGCN-AUR-M & 0.1300 & 0.0422 & 0.0916 & 0.0353 \\
LGCN-AUR-J & 0.1341 & 0.0451 & 0.1088 & 0.0659 \\ \hline
VAE-AUR & 0.1328 & 0.0518 & 0.1053 & 0.0543 \\
VAE-AUR-M & 0.1340 & 0.0437 & 0.1029 & 0.0514 \\
VAE-AUR-J & 0.1248 & 0.0490 & 0.1074 & 0.0656 \\ \hline
UE & 0.1111 & 0.0355 & 0.0948 & 0.0288 \\ \hline
\end{tabular}}
\label{variant}
\end{table}
\noindent $\bullet$ Regarding the training strategy, compared to MF that takes the proposed training strategy, MF-BPR that utilizes the BPR loss shows poor performance in the overall evaluation especially when the embedding size increases, and has relatively better performance on Tail absolute evaluation. However, its superiority over MF on tail performance disappears when the embedding size is 1024. The tail and overall performances of MF-BPR are far lower than MF-AUR, which shows the advantages of the proposed training strategy.
\noindent $\bullet$ Regarding the embedding size, both MF and MF-AUR show increasing trends on overall and tail performance.
MF achieves comparable results on overall performance to MF-AUR when the embedding size is 1024, while its tail performance is still far lower than MF-AUR. These results show that modifying the embedding size of MF does not change the above conclusions in this subsection.
\subsection{Studies on Model Designs (RQ2)}
In this subsection, we investigate the influence of the key designs of AUR, including the training procedure, uncertainty estimator and the tail-controlling coefficient $\alpha$.
\noindent \textbf{The Influence of Training Procedure and Uncertainty Estimator. } To study the influence of different training procedures and the design of the estimator, and verify the importance of the uncertainty estimation, we study three variants of AUR: 1) \textbf{AUR-J} takes a joint learning way to synchronously learn the expectation estimator and uncertainty estimator, \emph{i.e., } optimizing the objective in Equation~\eqref{loss_overall} to update them together. 2) \textbf{AUR-M} changes the uncertainty estimator to MLP, which takes the same multi-hot vectors as VAE as the input. For the MLP, there are two hidden layers both with the size of $1024$, and the activation function is ReLU. 3) \textbf{UE}
estimates expectations using the same model and weighting strategy as uncertainty estimation, then generates recommendations based on the expectation.
We compare these models with AUR on the two datasets under the three evaluation protocols.
The results are shown in Table~\ref{variant}, where we omit the results on Tail Relative Evaluation that show similar trends to Tail Absolute Evaluation.
From the table, we find, 1) compared to AUR, AUR-J shows better tail and overall performance on Amazon, but shows worse tail performance on Yelp, which is even worse than the expectation model (\textit{c.f.}, Table~\ref{main result}). This suggests selecting different training procedures for different datasets. In other words, different training procedures are suitable for different scenarios. However, considering the stability of AUR and its graftability to different backbone recommenders, we suggest taking the proposed two-step training (\emph{i.e., } sequential training) procedure of AUR.
2) Although AUR-M takes a more complicated MLP as the uncertainty estimator, it performs worse than the original AUR in most cases. This demonstrates that our uncertainty estimator's architecture is effective.
3) UE shows much worse tail performance on Yelp and Amazon, which verifies the importance of estimating the uncertainty for tail item recommendation.
\vspace{+5pt}
\noindent \textbf{The Effect of Tail-controlling Coefficient $\alpha$. } Recall that the hyper-parameter $\alpha$ in Equation \eqref{weight}, \emph{i.e., } the tail-controlling coefficient, is to control the tail performance. To empirically study its influence on the model performance, we conduct an experiment that changes the $\alpha$ in the range of $\{1,2,3,4,5\}$ for MF-AUR, and draws the trends of overall and tail recommendation performance. The results \emph{w.r.t. } $Recall@50$ are shown in Figure~\ref{hyper}. We omit the result on other metrics that exhibit the same trend. We find the tail performance keeps increasing in both Yelp and Amazon with $\alpha$ increasing, which supports the discussion that larger $\alpha$ indicates higher attention on the tail items in Section~\ref{sec:algo}.
Regarding the overall performance, the result exhibits different trends in the two datasets, which may be attributed to the properties of the dataset. Note that the training and testing sets are chronologically split and the item popularity drifts largely in Amazon. This means the tail items of the training set may become head items in the testing set, thus dominating the overall performance, leading to both the increased overall and tail performances on Amazon.
Therefore, it may be preferable to set a large value of $\alpha$ for a dataset with huge popularity drifts.
\begin{figure}[t]
\vspace{-2mm}
\centering
\subfigbottomskip=1pt
\subfigure[$Recall@50$ on Yelp]{
\includegraphics[width=0.44\columnwidth]{pic/hyper_yelp.pdf}
}
\subfigure[$Recall@50$ on Amazon]{
\includegraphics[width=0.44\columnwidth]{pic/hyper_amazon.pdf}
}
\caption{The overall and tail performance of MF-AUR on Yelp and Amazon as changing the value of the
tail coefficient $\alpha$.}
\label{hyper}
\end{figure}
\vspace{-2mm}
\subsection{In-depth Analyses (RQ3)}
\noindent \textbf{Calibrating Recommendation.}
Although most candidate items are tail items, traditional models prefer to recommend many head items to users even though certain users usually consume tail items. Thus we further study whether AUR calibrates the recommendation list regarding the tail and head items. We divide users into two groups: 1) tail-focus users whose at least $60\%$ of interacted items are tail items; 2) head-focus users who are not the tail-focus users. Then we count the proportion of tail items in the recommendation lists generated by AUR (MF-AUR,LGCN-AUR, and VAE-AUR) and expectation models (MF, LGCN, and VAE), for the tail-focus users and head-focus users, respectively. The result is shown in Figure~\ref{tail proportion}. We find: 1) AUR recommends more tail items to tail-focus users, and the recommendation ratio of tail items is greater than $35\%$ on Yelp and $40\%$ on Amazon, which is closer to the splitting ratio ($60\%$) compared to expectation models. 2) For head-focus users, expectation models almost do not recommend tail items, while AUR can recommend more tail items.
The results indicate that the ratios of tail items recommended by AUR are more close to that computed on the historical interactions. These findings verify that AUR can calibrate the recommendation ratios of tail items and head items, \emph{i.e., } recommending more suitable ratios of tail items to different types of users, compared with expectation models.
\begin{figure}[t]
\centering
\subfigbottomskip=1pt
\subfigure{
\includegraphics[width=0.8\columnwidth]{pic/tuli.pdf}
}
\subfigure[Yelp]{
\includegraphics[width=0.44\columnwidth]{pic/tail_proportion_Yelp.pdf}
}
\subfigure[Amazon]{
\includegraphics[width=0.44\columnwidth]{pic/tail_proportion_Amazon.pdf}
}
\caption{Recommendation ratio of tail items in top 20 recommendations for tail-focus and head-focus users.}
\label{tail proportion}
\end{figure}
\begin{table}
\centering
\vspace{-2mm}
\caption{The averaged minimal length of recommendation list to cover all testing interactions over all users.}
\vspace{-3mm}
\resizebox{0.5\textwidth}{!}{
\begin{tabular}{|c|cccccc|}
\hline
Methods & MF & MF-AUR & LGCN & LGCN-AUR & VAE & VAE-AUR \\ \hline
Yelp & 18,823 & 15,546 & 16,738 & 10,894 & 21,041 & 12,387 \\ \hline
Amazon & 42,255 & 41,184 & 39,453 & 28,824 & 50,684 & 41,831 \\ \hline
\end{tabular}
}
\label{worst_case}
\end{table}
\noindent \textbf{Worst Case Analysis. }
We then investigate the effect of AUR under the worst case by counting the number of recommendations needed to recall all relevant items in the testing set, \emph{i.e., } studying the minimal length of the recommendation list to cover all interacted items in the testing set. We average the required length over users for AUR and expectation model, where smaller value indicates better performance.
We summarize the results in Table~\ref{worst_case}. Compared with the expectation model, the corresponding AUR requires shorter recommendation lists to cover all positive testing interactions, which indicates the better robustness of making recommendation with uncertainty. Moreover, it reflects that generating recommendations with uncertainty taken into account can result in more item consumption at a lower cost. Meanwhile, LGCN-AUR achieves the best result on both Yelp and Amazon, which suggests selecting LGCN as the backbone expectation estimator of AUR when all relevant items are needed to be recommended.
\begin{table}[t]
\centering
\caption{The comparison of theoretical and empirical time complexity between MF, LGCN, VAE, and AUR on Amazon. The actual time cost per epoch (s) is calculated by averaging 10 epochs, and the $95\%$ confidence interval is reported.}
\scriptsize
\resizebox{88mm}{8mm}{
\begin{tabular}{|c|cccc|}
\hline
Time Analysis & MF & LGCN & VAE & AUR \\ \hline
MACC (1e6) & 11.7 & 83.3 & 189.2 & 187.6 \\
Backbone (s) & $31.11_{\pm0.02}$ & $149.23_{\pm0.09}$ & $100.13_{\pm0.05}$ & - \\
Backbone+AUR (s) & $92.35_{\pm0.07}$ & $125.37_{\pm0.09}$ & $104.95_{\pm0.10}$ & - \\ \hline
\end{tabular}
}
\label{time cost table}
\end{table}
\noindent \textbf{Time Cost Analysis. }
In this work, our AUR framework introduces an additional uncertainty estimation module. It is necessary to analyze the time complexity of the backbone expectation estimator and our AUR framework\footnote{ Focusing on uncertainty estimator, we set $\lambda$ as $1$.}.
We theoretically calculate the Multiply-Accumulate Operations(MACC)~\cite{mcquillanfeasibility} of each backbone method and the uncertainty estimator of AUR. We also empirically record the averaged training time cost per epoch on the Amazon dataset, for all models run on the same GPU machine\footnote{CPU:Intel(R) Core(TM) i9-9900X;GPU:NVIDIA GeForce RTX 2080 Ti;RAM:128GB.}. The results are summarized in the Table~\ref{time cost table}.
From the table, we see that:
1) the actual time cost of AUR framework is at the same level as backbone models.
Considering that AUR can improve tail performance significantly, such additional cost is worthwhile.
2) Although the MACC of VAE is great than LGCN, VAE shows less actual time cost. This is because we only calculate the cost of the aggregation from neighbor nodes when computing the MACC of aggregation operation in LGCN. However, LGCN indeed implements the aggregation via matrix multiplication, bringing additional time cost.
3) The actual time cost of LGCN-AUR is even lower than LGCN. This is because the actual time cost contains not only the feed-forward time cost but also the back-propagation time cost. The gradient calculation and update of LGCN are complex and costly.
When training LGCN-AUR, the LGCN is fixed. Therefore, the actual time cost of LGCN-AUR could decrease.
\section{Introduction}\label{sec:introduction}}
\IEEEPARstart{R}{ecommender} systems play an irreplaceable role in various online platforms~\cite{news,e-commerce,video}, which aim to facilitate information seeking by providing personalized services.
A canonical paradigm is solving recommendation as a machine learning problem to model the interaction likelihood between user-item pairs for making recommendations.
A de facto standard is learning the recommender model from historical interactions, which however suffers from severe data sparsity issues~\cite{grvcar2005data}.
The ratio of missing data, \emph{i.e., } user-item pairs lacking the label of interaction, can reach $99\%$ in many practical cases such as e-commerce~\cite{e-commerce} and social media~\cite{video} due to the huge size of candidate item set which typically increases over time. Worse still, the historical interactions are unevenly distributed over items where long-tail items encounter more missing data, leading to notorious issues like popularity bias~\cite{PDA}.
Therefore, it is essential to properly account for the missing data in recommender training.
Existing work typically treats missing data as negative samples~\cite{pan2008one}, making the objective of learning recommender models becoming fitting different labels or expanding the distance between positive and negative samples.
The model then estimates the interaction likelihood of negative samples and recommends the top-ranked ones to each user.
However, such methods face a mislabeling issue (\textit{a.k.a.} positive-unlabeled issue~\cite{elkan2008learning}) since the potential positive samples are not distinguished from the true negative samples, \emph{i.e., } potential positive samples are wrongly labeled.
Consequently, the model makes inaccurate estimation of interaction likelihood, which may suppress
potential positive samples, especially the tail ones that have a relatively higher chance to be mislabeled.
A line of research attempts to tackle this mislabeling issue by sample weighting~\cite{2009Collaborative,2016Modeling,2019Unbiased},
which essentially reduces the weight of potential positive samples in missing data to alleviate their predicted values fitting to wrong labels.
Nevertheless, these methods will push the model to focus more on popular items, amplifying the popularity bias~\cite{2019Unbiased}.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\columnwidth]{pic/intuitive.pdf}
\vspace{-5pt}
\caption{An intuitive example about aleatoric uncertainty in recommendation. We assume $Y_{u_1, i_2} \sim N(0.3, 0.5)$ and $Y_{u_1, i_4} \sim N(0.35,0.1)$. As labels of $i_2$ and $i_4$ have similar expectations, we prefer to recommend $i_2$ that exhibits higher uncertainty.}
\label{intuitive}
\vspace{-10pt}
\end{figure}
To address the mislabeling issue, we resort to \textit{aleatoric uncertainty}~\cite{2017What} which represents the inherent randomness of the label, \emph{i.e., } a measure of mislabeling.
{Aleatoric uncertainty} has shown success in mitigating the impact of wrong labels in other machine learning problems, such as image segmentation~\cite{zheng2021rectifying} and classification~\cite{uncertainty_app}.
In this light, we consider incorporating aleatoric uncertainty into recommendation modeling, \emph{i.e., } modeling both the interaction likelihood and aleatoric uncertainty of each user-item pair.
The key lies in how to accurately estimate aleatoric uncertainty at the fine-grained level of user-item pair.
This is non-trivial due to the specific properties of recommendation data and recommender training that are distinct from the existing research about aleatoric uncertainty in other problems.
For instance, 1) there lacks a formal definition of aleatoric uncertainty in recommendation.
2) There are complex relations across different samples due to involving the same user or item.
3) The common recommendation losses dissatisfy the assumptions behind the theory of aleatoric uncertainty.
In this work, we propose a new \textit{Aleatoric Uncertainty-aware Recommendation} (AUR) framework.
In particular, following the representative work PMF~\cite{DBLP:journals/corr/GopalanHB13} and ExpoMF~\cite{2016Modeling}, we assume that the implicit label $Y_{ui}$ obeys a Gaussian Distribution $N(r_{ui},\sigma_{ui}^{2})$\footnote{Traditional
models only estimate the expectation $r_{ui}$, which can be seen as specifying a uniform
variance for all user-item pairs.}. Further, the variance $\sigma_{ui}^{2}$ is supposed as the corresponding aleatoric uncertainty for the labels.
As shown in Figure \ref{intuitive}, the potential positive sample $(u_1, i_2)$ with higher uncertainty exhibits a broader Gaussian Distribution than the true negative sample $(u_1, i_4)$.
Since we
label the missing data as 0
during training, the higher the variance, the more likely the item will be mislabeled as 0.
It is thus better to give items with higher uncertainty a greater chance of being recommended.
We then devise a simple yet effective uncertainty estimator that leverages the idea of collaborative filtering to estimate uncertainty from user and item embeddings, accounting for the relation of uncertainty across different samples. The estimator can be composed with common recommender models that estimates the mean $r_{ui}$ of the distribution.
To learn the parameters of the recommender model and uncertainty estimator, we derive a general training objective to maximize the posterior over both positive and negative samples.
We argue that aleatoric uncertainty is also an important criteria for making recommendations.
As items that will be interacted are mislabeled as negative during training, a higher uncertainty indicates a higher chance that the true label is positive (see Figure \ref{intuitive}).
We thus make recommendations with the combination of both the expectation and the aleatoric uncertainty.
We theoretically and empirically prove that our AUR framework can achieve similar overall performance with the backbone model at least and works particularly well on recommending the tail items.
Moreover, it can alleviate the popularity bias issue since long-tail items will obtain a relatively higher uncertainty.
This is because positive samples on long-tail items typically receive lower predictions of expectation (\emph{i.e., } larger gap to the value of label $1$), pushing the uncertainty estimator to increase its uncertainty prediction.
Consequently, the estimator learns a potential correlation between high uncertainty and low popularity.
Remarkably, the AUR framework is agnostic to the architecture of the recommender model for expectation estimation.
We demonstrate AUR on three representative backbone models: MF~\cite{koren2009matrix}, VAE~\cite{liang2018variational} and LightGCN~\cite{he2020lightgcn} and evaluate it on two real-world datasets, achieving better overall recommendation results and significant improvements on long-tail items.
Our main contributions are as follows:
\begin{enumerate}[leftmargin=*]
\item We present a new perspective of aleatoric uncertainty for recommendation modeling and propose the AUR framework to estimate the uncertainty.
\item We reveal the potential of aleatoric uncertainty in making recommendation and conduct theoretical analysis to demonstrate its superiority over the conventional expectation-based recommendation.
\item We instantiate AUR on three backbone models and conduct extensive experiments, validating the rationality and effectiveness of our method.
\end{enumerate}
\section{Proof of Theorem 1}\label{proof}
The training objective of uncertainty estimator, \emph{i.e., } Equation (\ref{loss}) can be written as:
\begin{equation}
\small
\begin{split}
&L(D)=\sum_u\sum_{i}(\frac{\Delta_{ui}^2}{e^{s_{ui}}}+\beta s_{ui}+ \gamma s_{ui}^2)
\end{split}
\end{equation}
here we assume $w_{ui}$ or $\alpha$ in equation (\ref{weight}) is 1, but it is trivial to obtain a similar inequality for $\alpha \neq 1$.
We obtain the normalized distribution of $p_u$ and $q_u$ as $p_{ui}= \frac{\Delta_{ui}^2}{K_{u}}$ and $q_{ui}=\frac{e^{s_{ui}}}{R_{u}}$,
where $K_{u}=\sum_{i}\Delta_{ui}^2, R_{u}=\sum_{i}e^{s_{ui}}$, and $p_{ui}$ and $q_{ui}$ denote the value of $p_u$ and $q_u$ on the $i$-$th$ item.
According to the common inequality: $lnx+1 \le x (x>0)$, we can derive the inequality:
\begin{equation}
\small
\frac{\Delta_{ui}^2}{e^{s_{ui}}}= \frac{\Delta_{ui}^2}{K_{u}}\frac{K_{u}}{e^{s_{ui}}}\ge \frac{\Delta_{ui}^2}{K_{u}}(ln\frac{\frac{\Delta_{ui}^2}{K_{u}}}{\frac{e_{s_{ui}}}{R_{u}}}+ln\frac{K_{u}^2}{R_{u}}+1-ln\Delta_{ui}^2)
\end{equation}
By substituting the above inequality into $L(D)$, we have:
\begin{equation}\small
\begin{split}
&L(\mathcal{D})\ge \sum_{u}\sum_{i}(p_{ui}(ln\frac{p_{ui}}{q_{ui}}-lnR_{u})+\beta s_{ui}+ \gamma s_{ui}^2)+c_1\\
&=\sum_{u}KL(p_{u}||q_{u})+\sum_{u}\sum_{i}(\beta s_{ui}+ \gamma s_{ui}^2 -lnR_{u})+c_1
\end{split}
\label{eq21}
\end{equation}
Assume that $s_{uk}=max_{i}s_{ui}$, we have $lnR_{u}\le ln|\mathcal{I}|e^{s_{uk}}=s_{uk}ln|\mathcal{I}|$ and $\mathcal{I}$ is the set of all items. According to the Extremum Theorem of a quadratic function, we can derive:
\begin{equation}
\begin{split}
&\beta s_{ui}+ \gamma s_{ui}^2 \ge \frac{-\beta ^2}{4\gamma} \ \ \ \ \ \ \ for \ i \neq k\\
&(\beta-ln|\mathcal{I}|) s_{uk}+\gamma s_{uk}^2 \ge \frac{-(\beta-ln|\mathcal{I}|) ^2}{4\gamma} \ for \ i = k
\end{split}
\end{equation}
Therefore, the second term in the RHS of the inequality (\ref{eq21}) is also lower bounded by a constant.
Overall, we can obtain the inequality:
\begin{equation}
\small
L(\mathcal{D}) \ge \sum_{u}KL(p_{u}||q_{u})+c
\end{equation}
This completes the proof of the theorem.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\section{METHOD}\label{sec:method}
\textbf{Problem Formulation.
Suppose we have an implicit
feedback dataset
$\mathcal{D}$ with $m$ users and $n$ items. The set of users and items are represented as $\mathcal{U}$ and $\mathcal{I}$.
For any $u \in \mathcal{U}$ and $i \in \mathcal{I}$, the label of pair $(u,i)$ is $Y_{ui}=1$, if $u$ has consumed $i$, and $Y_{ui}=0$ otherwise.
Let $\mathcal{R}=\{ r_{ui} \}_{m \times n}$ denotes the prediction given by a
backbone model (\emph{e.g., } MF) parameterized with $\theta$, \emph{i.e., } $r_{ui}=f_{\theta}(u,i)$.
We assume that $Y_{ui} \sim N(r_{ui},\sigma_{ui}^2)$ and learn the variance $\sigma_{ui}^2$ as our aleatoric uncertainty.
The backbone prediction and uncertainty estimation can be represented as $r_{ui}=f_{\theta}(u,i)$ and $\sigma_{ui}^2=g_{\phi}(u,i)$, where $g_{\phi}(\cdot, \cdot)$ denotes the uncertainty estimator parameterized with $\phi$.
\subsection{Aleatoric Uncertainty}\label{sec:method-overall
In practical recommender systems, there are many potential positive samples that are blindly labeled as 0. When training on such data, a certain amount of useful information would be discarded, resulting in an ineffectively trained model.
Therefore, it is an important task to quantify how likely a sample is mislabeled.
While traditional methods treat the recommendation problem as a homoscedastic regression task~\cite{2007Probabilistic}, in this work, we reformulate it as a heterogeneous regression problem where samples have different variances.
We assume that the implicit feedback label $Y_{ui}$ follows the Gaussian Distribution $N(r_{ui},\sigma_{ui}^2)$, where $r_{ui}$ denotes the prediction of any recommendation model which is trained from the dataset $\mathcal{D}$.
The variance $\sigma_{ui}^2$ represents the stability of the Gaussian Distribution, which is a variable learned from the implicit data.
A large variance $\sigma_{ui}^2$ usually indicates that the given label is unstable,
\emph{i.e., } a label with large variance is likely to be incorrect.
We thus use the variance of the label to quantify the likelihood of mislabeling, which is referred to as aleatoric uncertainty.
Similar to \cite{2017What}, we define the variance $\sigma_{ui}^2$ as our aleatoric uncertainty of the label $Y_{ui}$.
\subsubsection{Uncertainty Estimator}\label{unestimator}
We consider to model the uncertainty $\sigma_{ui}^2$ in a similar manner as the backbone model $r_{ui}=f_{\theta}(u,i)$.
To ensure the non-negativity of $\sigma_{ui}^2$, we naturally consider to denote it by following the exponential form:
\begin{equation}\small
\sigma_{ui}^2=g_{\phi}(u,i)=\frac{e^{h_{\phi}(u, i)}}{K}.
\label{exponential}
\end{equation}
$K$ controls
the scale of exponential function.
Notice that the uncertainty of the label $Y_{ui}$ is a combination of the uncertainty signals from the user and item.
We learn latent item representation $q_i$ and user representation $p_u$, similar to recommender models, and combine them as $h_{\phi}(u, i)$.
As to item $i$, we learn a projection $f_{item}:e_i \rightarrow q_i$ to
mine the latent factors affecting uncertainty,
where $e_i$ denotes the one-hot encoding of item $i$.
For user representation, intuitively, users with similar item interactions should have similar uncertainty.
We employ the combination of all item representations in the user's historical interaction as the user representation. Formally,
$p_u=\sigma(\frac{1}{\sqrt{|\mathcal{H}_u|}}\sum_{i\in \mathcal{H}_u}z_i),$
where $\mathcal{H}_u$ denotes the historical interaction set of user $u$, and $\sigma$
indicates an activation function. $z_i$ is the user history information encoded by another projection\footnote{This design is similar to \cite{koren2008factorization}, which validates the rationality of using two different projections.} $f_{user}:e_i \rightarrow z_i$.
Then $h_{\phi}(u, i)$ is simply defined as the inner product of item representation $q_i$ and user representation $p_u$, \emph{i.e., } $h_{\phi}(u,i)=<p_u,q_i>$. For briefness, we denote the value of $h_{\phi}(u, i)$ as {$s_{ui}$} in the rest of the paper.
\subsection{Training Procedure}\label{sec:training}
A default choice for estimating the parameters of backbone model and uncertainty estimator is minimizing the negative log-likelihood of training samples. Formally, the negative log-likelihood of $Y_{ui}$ is $-logP(Y_{ui}|r_{ui},\sigma_{ui}^2)=\frac{(r_{ui}-Y_{ui})^2}{2\sigma_{ui}^2}+\frac{1}{2}log \ \sigma_{ui}^2$.
Due to the huge amount of training instances, we usually consider the mini-batch training strategy.
We adopt the same user-batch strategy used in \cite{liang2018variational}. We first randomly sample a batch of users $\mathcal{U}_{batch}$ from $\mathcal{U}$, then the user batch loss $L(\mathcal{U}_{batch},\mathcal{D};\theta,\phi)$ can be formulated as:
\begin{equation}\small
\begin{split}
\sum_{u\in \mathcal{U}_{batch}, i \in \mathcal{I}}w_{ui}\left(
\frac{(r_{ui}-Y_{ui})^2}{
\frac{2}{K}e^{s_{u,i}}}+\frac{1}{2} \ s_{u,i}
\right)+\Omega,
\label{loss_overall}
\end{split}
\end{equation}
where $w_{ui}$ denotes the weight of different samples in the training set $\mathcal{D}$, and $\Omega$ denotes the regularization term.
Noticing that optimizing the two models simultaneously can lead to unstable issues and overfitting~\cite{yu2020sampler},
we propose to train them sequentially in an Expectation-Maximization like algorithm. In particular, we first train the backbone model as our expectation estimator without considering the uncertainty, and then train the uncertainty estimator while fixing the expectation model.
Sequential training also simulates the situation where our method is grafted onto an existing model, which is required by industry applications. We also show the results of the joint learning of expectation and uncertainty in the experiments (\emph{cf. } Table~\ref{variant}).
\subsubsection{Backbone Training}\label{expectation_training}
To alleviate the imbalance issue of positive and negative samples, we leverage a sampling strategy.
We assign the weight to each instance $w_{ui}$ as follows:
\begin{equation}\small
w_{ui}\sim
\begin{cases}
1 &\mbox{$Y_{ui}=1$,}\\
Bernoulli(\mu) &\mbox{$Y_{ui}=0$.}
\end{cases}
\label{weight}
\end{equation}
where $\mu$ denotes the sampling rate for negative samples.
The aleatoric uncertainty is not considered for the backbone model,
$s_{ui}$ is thus fixed for all users and items.
We set $s_{ui} = 0$ and use $l_2$-norm as regularization, \emph{i.e., } $\Omega=\lambda||\theta||^2$, simplifying the training objective for backbone model as:
\begin{equation}\small
\begin{split}
&L(\mathcal{U}_{batch},\mathcal{D};\theta)=\sum_{u\in \mathcal{U}_{batch}, i \in \mathcal{I}}w_{ui}\left(
\frac{(r_{ui}-Y_{ui})^2}{2}
\right)+\lambda ||\theta||^2.
\label{loss_backbone}
\end{split}
\end{equation}
The first term is essentially a mean square error, one of the common recommendation losses.
The form of the loss is decided by the prior assumption of data distribution. Other losses like BPR \cite{2009BPR} corresponds to distributions other than Gaussian Distribution.
We belief Gaussian Distribution is a reasonable assumption since Equation ~\eqref{loss_backbone} can lead to stronger backbone models than BPR (\emph{cf. } Figure~\ref{fair}).
\subsubsection{Uncertainty Estimator Training}
The uncertainty estimator is trained on top of the backbone model. In this stage, we also need to balance the positive and negative samples, although the learning objective is not the hard label of implicit data.
To speed up training, we replace the sampling process by directly assigning fixed weight to each instance.
The weight $w_{ui}$ there can be formulated as:
\begin{equation}\small
w_{ui}=
\begin{cases}
\alpha &\mbox{$Y_{ui}=1$,}\\
1 &\mbox{$Y_{ui}=0$.}
\end{cases}
\label{weight_1}
\end{equation}
where $\alpha$ is a key hyperparameter for training the uncertainty estimator. More specifically, $\alpha$ can be adjusted to control the strength of
emphasizing the tail items (\emph{cf. } Figure~\ref{hyper}).
To optimize the uncertainty estimator, we first derive the distribution of uncertainty $P(s_{ui}|r_{ui},Y_{ui})$. According to Bayes theorem, we can get:
\begin{equation}\small
P(s_{ui}|r_{ui},Y_{ui})=\frac{P(Y_{ui}|r_{ui},s_{ui})P(r_{ui}|s_{ui})P(s_{ui})}{P(r_{ui},Y_{ui})}.
\end{equation}
Since the expectation and variance of a Gaussian Distribution are independent statistics~\cite{casella2021statistical}, $s_{ui}$ and $r_{ui}$ are independent. We have $P(r_{ui}|s_{ui})=P(r_{ui})$. Note that when training the uncertainty estimator, $r_{ui}$ is fixed and treated as a constant. Then we have $P(s_{ui}|r_{ui},Y_{ui}) \propto P(Y_{ui}|r_{ui},s_{ui})P(s_{ui})$. We assume that $s_{ui}$ has a prior of zero-mean Gaussian Distribution: $s_{ui} \sim N(0,\tau^2)$.
Combined with Equation \eqref{exponential}, we obtain the posterior probability of $s_{ui}$ and minimize its negative logarithm to optimize the parameters of uncertainty estimator. Formally,
\begin{equation}\small
-logP(s_{ui}|r_{ui},Y_{ui})=\frac{(r_{ui}-Y_{ui})^2}{\frac{2}{K}e^{s_{ui}}}+\frac{1}{2}s_{ui}+\frac{s_{ui}^2}{\tau^2}+c,
\end{equation}
where $c$ denotes a constant. After simplifying the form, the final optimization objective is:
\begin{small}
\begin{equation}
\begin{split}
&L(\mathcal{U}_{batch},\mathcal{D};\phi)=\sum_{u \in \mathcal{U}_{batch},i \in \mathcal{I}}w_{ui}\left(
\frac{(r_{ui}-Y_{ui})^2}{e^{s_{ui}}}+\beta s_{ui}+ \gamma s_{ui}^2
\right),
\label{loss}
\end{split}
\end{equation}
\end{small}
where $\beta$, $\gamma$ are hyperparameters.
This training process can be understood as finding the best variances such that the data likelihood is maximized.
After obtaining the optimal values of the variance, our model makes recommendations based on both of the expectation and aleatoric uncertainty.
$Pred_{ui}$ denotes the final prediction of the pair $(u,i)$.
Given the pair $(u,i)$, the final prediction is $Pred_{ui}=(1-\lambda)r_{ui}+\lambda \sigma_{ui}$.
When $\lambda = 0 $, the prediction degenerates to the backbone model.
\begin{figure}[t]
\centering
\subfigure{
\includegraphics[width=0.44\columnwidth]{pic/sigma_r_Yelp.png}
}
\subfigure{
\includegraphics[width=0.44
\columnwidth]{pic/sigma_r_Amazon.png}
}
\vspace{-10pt}
\caption{Relationship between $r_{ui}^2$ and $\sigma_{ui}^2$.}
\vspace{-10pt}
\label{r_sigma}
\end{figure}
\subsection{Theoretical Analysis}\label{sec:algo}
In this subsection, we first show that the overall performance of the recommendation with aleatoric uncertainty is at least as good as the backbone models using expectation. The statement is further verified empirically with results from two real datasets (\emph{cf. } Table~\ref{main result}). We then explain the mechanism of improving the tail performance.
\noindent \textbf{$\bullet$ Overall Performance. }
We first state the property that connects the aleatoric uncertainty with the expectation.
\begin{theorem}
Denote the residual of the expectation as $\Delta_{ui}=r_{ui}-Y_{ui}$.
For user $u$, denote the distributions of $\Delta_{ui}^2$ and $e^{s_{ui}}/K$ over the items as $p_u$ and $q_u$ respectively.
If the weights $w_{ui}$ are the same for all user-item pairs, the following inequality holds:
\begin{equation}\small
L(\mathcal{D}) \ge \sum_{u}KL(p_{u}||q_{u}) + c,
\end{equation}
where $L(\mathcal{D})$ is the loss of uncertainty estimator in Equation \eqref{loss}; $KL(p_u||q_u)$ is the KL-Divergence of the two distributions; and $c$ is a constant.
\end{theorem}
The above theorem essentially states that the KL-Divergence between the distributions of expectation and uncertainty is bounded by the loss function that is being optimized. In other words, the lower the loss can be achieved, the closer the two distributions will become. Therefore, the expectation and the uncertainty are highly positively correlated, and thus the overall performance of uncertainty recommendation should be at least on par with the recommendation models using expectation. The proof is provided in \textit{Appendix} \ref{proof}.
To further validate the above statement, we report the empirical results on two real world datasets in Figure \ref{r_sigma}. Specifically, we randomly select one user from each dataset and show the visual relationship between $r_{ui}^2$ and $\sigma_{ui}^2$.
As shown in Figure \ref{r_sigma}, it is clear that expectation and uncertainty have strong linear positive correlation for both users, which confirms the theorem. The yellow line indicates the best fit line. Moreover, we calculate the mean pearson correlation coefficient on different datasets over the users, where it is 0.69 on Amazon-Book and 0.74 on Yelp (see dataset descriptions in Section \ref{ssec:exp_setting}).
\noindent \textbf{$\bullet$ Tail Performance. }
We also claim that using uncertainty to make recommendations improves the quality of tail items recommendations.
The reason is that when a positive sample is not able to be retrieved by the expectation $r_{ui}$, its variance $\sigma_{ui}^2$ would be learned to be a large value.
This positive sample is likely to be captured by our AUR model. Especially, when the hyperparameter $\alpha$ in Equation \eqref{weight} increases, the $\sigma_{ui}^2$ of tail items would also increase. Therefore, we name $\alpha$ as the \textbf{tail-controlling coefficient}.
We use an example to illustrate this statement in Table \ref{pop}. We select a user and four items from the Amazon-Book dataset to compare the learned expectations and uncertainties. The backbone model is MF. We count item popularity in the training set while the user feedback is retrieved from the testing set. From the table, we can see that $i_4$ has a much lower popularity compared to $i_1$, as well as its expectation.
Despite that item $i_1$ has been interacted many times, it is actually a true negative sample for the user.
Both the popularity and the expectation of $i_4$, a true positive sample, are the lowest among the four items, but it still has the highest uncertainty indicating that $i_4$ is a potential positive sample. We provide more analysis in the experiments.
\begin{table}
\small
\centering
\caption{Case study of selected user-item pairs. We show four items with their expectation, uncertainty, and popularity in Amazon to explain how AUR improves the tail performance. }
\vspace{-4mm}
\begin{tabular}{|c|cccc|}
\hline
Items & $i_1$ & $i_2$ & $i_3$ & $i_4$ \\ \hline
Expectation & 0.8038 & 0.7784 & 0.7654 & 0.4648 \\
Uncertainty & 0.3188 & 0.2952 & 0.3050 & 0.3277 \\
Popularity & 128 & 44 & 59 & 40 \\
User Feedback & 0 & 0 & 0 & 1 \\ \hline
\end{tabular}
\label{pop}
\vspace{-6mm}
\end{table}
\subsection{Discussion}\label{sec:discussion}
\noindent \textbf{$\bullet$ Relevance to Sample Weighting. }
We first discuss the connections and distinctions between our approach and previous methods.
In the previous discussion, \cite{2009Collaborative,reweight1} heuristically assigned weights $c_{ui}$ to each training sample. In our work, if we fix $s_{ui}=log(\frac{1}{c_{ui}})$ and $\alpha=1$ in Equation \eqref{loss} to optimize the backbone model, our approach is equivalent to \cite{2009Collaborative,reweight1}.
Intuitively, for the positive samples, high correlation between $(u,i)$ indicates low variance \cite{2009Collaborative}. For the negative samples, high correlation indicates that the item $i$ might be a potential positive sample for user $u$, resulting in a high variance \cite{reweight1}.
Therefore, assigning fixed weights to re-weight MSE loss in the training process is essentially to determine the uncertainty of each sample.
Previous methods specify/fix the variance to learn the expectation. In contrast, our approach first learns an expectation, and then estimates the variance based on the learned expectation. During the recommendation stage, previous methods recommend items using expectation while we make recommendations with variance to better capture the uncertainty in the data.
\noindent \textbf{$\bullet$ Relevance to Probabilistic Methods. }
Similar to AUR, some probabilistic models, \textit{e.g.}, PMF \cite{2007Probabilistic} and ExpoMF \cite{2016Modeling}, also assume the implicit labels obey Gaussian Distributions. Differently, they take the variances of Gaussian Distributions as hyper-parameters, and assume all $(u,i)$ pairs have the same variances.
AUR estimates the variances for each $(u,i)$ pair with the proposed learning strategy, obtaining more accurate uncertainties at a fine-grained user-item pair level.
Moreover, these probabilistic models generate recommendations with the predicted expectations, which may over-recommend popular items,as discussed in Section~\ref{exp_prob}. Fundamentally different to them, AUR makes recommendations based on the uncertainties, aiming at providing more accurate and more balanced recommendations.
\section{Problem Formulation}\label{sec:problem}
\section{Related Work}
\textbf{Uncertainty Quantification and Application.}
According to the source of uncertainty, there are two types of uncertainty: epistemic uncertainty and aleatoric Uncertainty~\cite{uncertainty_survey}. Aleatoric uncertainty is also known as data uncertainty, which originates from randomness in the data generation process~\cite{uncertainty_survey}. The data mislabeling can be detected and addressed with this uncertainty~\cite{2017What,wang2019aleatoric}. However, most methods are not designed for recommender systems. The epistemic uncertainty, which quantifies how uncertain a model is~\cite{uncertainty_app} , is less related to us. It is usually used to detect the out-of-distribution samples in risk-sensitive fields~\cite{al2019deep,leibig2017leveraging}, such as automatic control~\cite{al2019deep} and disease detection~\cite{leibig2017leveraging}. There is also some work~\cite{huang2022uncertainty} focus on the label noise problem by leveraging both epistemic uncertainty and aleatoric uncertainty.
It is worth mentioning that our AUR framework is inspired by~\cite{2017What}. We similarly assume that the labels obey Gaussian Distribution and learn the variance in a similar way. However, there are still some essential difference. \cite{2017What} trains the uncertainty to obtain a better expectation model, while we make recommendation based on the value of uncertainty. And they do not focus on
the improvement of tail prediction.
There are some efforts to model and utilize the uncertainty in recommendation~\cite{zeldes2017deep,wang2019bayesian,wang2020m2grl,liu2021personalised}. \cite{embedding_uncertainty,jiang2020convolutional} focus on the epistemic uncertainty. \cite{embedding_uncertainty} models the uncertainty of user/item representation to expand users' and items' interaction space, further alleviating the cold-start problem. \cite{jiang2020convolutional} employs Gaussian embedding, \emph{i.e., } endowing the randomness to the embedding, being adaptive to the uncertain preference exhibited by some users. Different from them, we aim at model the uncertainty of data instead of model. \cite{wang2020m2grl} also considers the the uncertainty from the data perspective. Differently, it considers the multi-task scenarios and takes the uncertainty to identify the importance of different task. \cite{xu2022ukd} introduces aleatoric uncertainty to mitigate the inherent noise in pseudo labels when distill the knowledge from the teacher model. While we consider a single task and directly take use uncertainties for recommendation. \cite{zeldes2017deep,wang2019bayesian} model both the epistemic uncertainty and aleatoric uncertainty in reinforcement learning-based recommendation. Their goal is to do better exploration and exploitation trade-off with both the uncertainty and model normal prediction. Different from them, we mainly focus on collaborative filtering and pay attention to the mislabeling issue instead of the exploration and exploitation trade-off.
\textbf{Popularity Bias in Recommendation.}
Tail item recommendation performance is very related to the popularity bias problem~\cite{ChenDQ0XCLY21,LeePL21}, which has gotten huge attention in the recommendation community. To deal with the popularity bias problem, existing methods can be divided into four different categories: 1) IPS-based methods~\cite{2019Unbiased,schnabel2016recommendations} deal with the problem by re-weighting training samples with propensities. However, the propensities are hard to set properly, causing high variance problems~\cite{PDA}. 2) Utilizing uniform data~\cite{liu2020general,zheng2021disentangling,ChenDQ0XCLY21} to guide the model to be unbiased. However, uniform data is usually small and expensive to gain. 3) The methods based on post-hoc re-ranking or model regularization~\cite{abdollahpouri2019managing,zhu2021popularity} usually need to heuristically design the re-ranking or regularization polices. 4) Causal inference-based methods~\cite{PDA,wei2021model} first take prior knowledge to create a causal graph to model the recommendation process, and then take intervention or counterfactual inference to eliminate the popularity bias. While the prior knowledge is not easy to get. Different from all these methods, we take the uncertainty to address the mislabeling issue to achieve the better long tail recommendation performance.
\textbf{Data mislabeling.}
To solve the mislabeling issue, \emph{i.e., } the positive-unlabeled issue, there are two main strategies: re-weighting and sampling. Reweghting-based methods try to solve the problem by identifying mislabeled samples and assigning lower/higher weights for them/others. \cite{steck2010training,pan2008one,2009Collaborative,gantner2012personalized} take heuristic strategies for reweighting, such as assigning higher weights for unlabeled but popular items~\cite{gantner2012personalized}. \cite{chen2020fast} proposes to learn these weights instead of heuristically designing.
To solve the mislabeling issue, sampling-based methods avoid selecting the potential positive but unlabeled samples. There are many sampling methods~\cite{2015Improving,2016Modeling,wang2020m2grl,hsieh2014pu,yu2020sampler}. \cite{2015Improving} proposes a neighborhood-aware sampling strategy -- avoiding sampling items interacted by neighbors. \cite{2016Modeling} gives higher sampling probabilities for unlabeled samples that have higher exposure probabilities. While the exposure probabilities are hard to be estimated well, \cite{wang2020m2grl} resorts to knowledge graph to design a sampling strategy. Besides, \cite{hsieh2014pu,yu2020sampler} directly estimates the probability of mislabeling with a Bayesian framework, and samples negatives based on this probability. Different to the two types of method, we learn the uncertainty of samples, and take the uncertainty as the mislabeling probability. Moreover, we directly take the uncertainty to form recommendation scores instead of utilizing them to re-weight or sample data.
|
\section{Introduction}\label{sec:intro}
The study of the interplay between the microscopic quantum world and the
macroscopic classical one is a fundamental research topic in contemporary
physics, although it dates back to the first days of quantum
mechanics~\cite{Zurek_2003, rossini2021coherent}. Typically, the interaction
with the environment is believed to destroy genuine quantum behaviors, although
consensus is emerging that this is not always the case.
Dissipation-based protocols have been devised to imprint nontrivial
correlations in quantum many-body systems
(see, e.g., Refs.~\cite{Syassen_2008, lin2013, diehl_2008, verstraete-2009,
Eisert_2010, Roncaglia_2010, diehl-2011, Bouchoule_2020, Rossini_2021,
seetharam2022correlation}).
A first crucial question is whether entanglement, which is the distinctive
feature of quantum mechanics, is robust against the presence of the
environment. Second, is it possible to enhance the entanglement content of a
quantum many-body state via an \emph{ad hoc} engineered environment?
Answering these questions is a daunting task, because there is no universal
approach (neither analytic nor numerical) to tackle generic \emph{open} quantum
many-body systems. With this state of affairs, one has to resort to approximate
treatments. Markovian master equations, such as the Lindblad master
equation~\cite{lindblad1976on, gorini1976completely, petruccione2002the},
provide some of the most successful tools to attack open quantum many-body
systems.
In this paper we focus on the paradigmatic one-dimensional setup illustrated in
Fig.~\ref{fig:sb}: a system of \emph{noninteracting} fermions is locally
coupled to ideal thermal baths.
To be specific, we focus on the tight-binding chain and on the Kitaev chain.
The fermions live on a lattice with $N$ sites, with either periodic boundary
conditions (PBC) or open boundary conditions (OBC).
The system is put in contact with two ideal fermionic reservoirs at
temperatures $T_L,T_R$, and with chemical potentials $\mu_L,\mu_R$. With OBC
the two baths are placed at the edges of the chain [Fig.~\ref{fig:sb} (a)],
whereas with PBC they are are at the maximum distance $N/2$
[Fig.~\ref{fig:sb} (b)].
The interaction between the chain and the reservoirs is treated within the
formalism of the Lindblad master equation~\cite{petruccione2002the}.
Specifically, we employ the \emph{nonlocal} description derived in
Ref.~\cite{dabbruzzo2021self}.
The Lindblad operators are obtained \emph{ab initio} from the microscopic
system-bath interaction, and are written in terms of the Bogoliubov modes of
the model without dissipation. As such, the Lindblad operators are non-local in
real space.
Interestingly, this allows to recover the Conformal Field Theory (CFT)
description of the chain in the low-temperature limit.
Furthermore, the nonlocal Lindblad approach allows to obtain a
thermodynamically consistent description of transport
properties~\cite{dabbruzzo2021self, dabbruzzo2021self2}.
We are interested in the quantum correlations emerging in the steady state of
finite chains of length $N$, in the limit $t \to \infty$. The ensemble
describing this state is written in terms of a convex combination of the
Fermi-Dirac distributions of the reservoirs~\cite{dabbruzzo2021self}.
Importantly, this ensemble is in general different from the finite-temperature
ensemble of the underlying fermionic chain. In particular, we will show that
ground-state criticality of the free chain Hamiltonian is associated with
nontrivial steady-state correlations in the dissipative model.
To monitor these correlations, we consider the quantum mutual information
$I(A_1:A_2)$ between two subregions $A_1$ and $A_2$ of the chain, defined
as~\cite{amico2008entanglement, calabrese2009entanglemententropy,
eisert2010colloquium, laflorencie2016quantum}
\begin{equation}\label{eq:mi-intro}
I(A_1:A_2) \coloneqq S_{A_1} + S_{A_2} - S_{A_1\cup A_2},
\end{equation}
where $S_A$ is the von Neumann entropy of the subregion $A$,
which is defined as
\begin{equation} \label{eq:vn}
S_A \coloneqq -\mathrm{Tr} \left( \rho_A \ln \rho_A \right),
\end{equation}
where $\rho_A$ is the reduced density matrix for subsystem $A$.
For pure states the von Neumann entropy $S_A$ of a subsystem quantifies its
entanglement with the rest of the system. Moreover, one has that
$S_A = S_{\overline{A}}$, with $\overline{A}$ being the complement of $A$
[see, for instance, Fig.~\ref{fig:sb} (a)], and $S_{A\cup\overline{A}}=0$.
However, in the presence of an environment, the global state is mixed, which
implies that neither the von Neumann entropy nor the mutual information are
proper measures of the entanglement shared between different regions.
Still, it has been shown recently that for out-of-equilibrium free-fermion and
free-boson models in the presence of quadratic global dissipation the mutual
information admits a hydrodynamic description in terms of a quasiparticle
picture~\cite{alba2021spreading,carollo2022dissipative,alba2022hydrodynamics,
alba2022logarithmic}.
Moreover, we numerically checked that our findings remain qualitatively valid
using the fermionic entanglement negativity, which instead is a proper measure
of entanglement~\cite{Vidal_2002}.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{cartoon.pdf}
\caption{
The setup used in this work. A one-dimensional chain of $N$
noninteracting fermions is put in contact with two thermal reservoirs
at temperatures $T_L$, $T_R$ and chemical potentials $\mu_L$, $\mu_R$.
(a) In the case of OBC, only the sites at the edges are in contact with
the baths, and we are interested in the steady-state mutual information
$I(A_1:A_2)$ between the two connected intervals $A_1$ and $A_2$ of
length $\ell$ and $N-\ell$, respectively.
(b) In the case of PBC, the sites in contact with the baths are at the
maximum distance $N/2$ and the intervals $A_1$ and $A_2$ are
respectively of length $\ell$ and $N/2-\ell$. Analytical calculations
are done in the thermodynamic limit $N,\ell \to \infty$.
}
\label{fig:sb}
\end{figure}
Both the full-system entropy, as well as the subsystems entropies, generally
exhibit volume-law scaling in the steady-state.
However, if the bulk of the system is tuned to a critical point, logarithmic
corrections appear. Specifically we show that, for a generic subsystem $A$
of length $\ell$, in the scaling limit $N,\ell\to\infty$ with arbitrary ratio
$\ell/N$, the entropy $S_A$ is given as
\begin{equation}\label{eq:ent-asy}
S_A = \alpha \ell+\frac{c(\Theta)}{3\nu}\ln\left[\frac{N}{\pi}\sin\left(
\frac{\pi\ell}{N}\right)\right]+{\mathcal O}(1).
\end{equation}
The prefactor $\alpha$ of the volume-law term depends on the full spectrum of
the model and on the properties of the bath.
A similar volume-law scaling has been found in free-fermion systems with
localized dissipative
impurities~\cite{alba2022noninteracting,alba2022unbounded}.
The prefactor $c(\Theta)$ is an effective ``central charge'', which contains
information only about the singularities in the single-particle spectrum of the
model. It is an even function of a parameter $\Theta \in [-1,1]$ which depends
on the properties of the baths, i.e., their temperatures and chemical
potentials.
The quantity $\nu \in \{1,2\}$ is the number of boundary points between
subsystem $A$ and the rest. For instance, for a chain with OBC
[see Fig.~\ref{fig:sb} (a)], $\nu=2$ if the subsystem is next to the chain
edge, whereas $\nu=1$ if it is in the bulk.
The argument inside the square brackets is the so-called
chord length~\cite{difrancesco1997conformal}.
In the limit $\ell/N\to 0$, the second term in~\eqref{eq:ent-asy} becomes
$c/(3\nu)\ln(\ell)$, which we prove analytically.
On the other hand, the result for finite ratio $\ell/N$ is a conjecture
inspired by the
zero-temperature CFT scaling~\cite{calabrese2009entanglemententropy}.
Similar logarithmic terms as in~\eqref{eq:ent-asy} have been found in a
tight-binding model, although for different nonequilibrium
settings~\cite{Eisler2014, Kormos2017,turkeshi2022entanglement,
turkeshi2022enhanced}.
Finally, the last term ${\mathcal O}(1)$ is a subleading constant, which can be
calculated, at least for the tight-binding chain, by using the Fisher-Hartwig
conjecture~\cite{fisher1969toeplitz,basor1991the,basor1994the,
forrester2004applications,deifts2011asymptotics,fagotti2011universal}.
In the limit $|\Theta|\to1$ one recovers the
zero-temperature result, i.e., $c(\Theta)$ becomes the central charge of the
CFT that describes ground-state properties of the model. Here one has $c=1$ and
$c=1/2$ for the tight-binding chain and the Kitaev chain, respectively. On the other hand, for $\Theta\to0$,
which corresponds to the high-temperature limit, $c(\Theta)$ vanishes. Remarkably,
for generic $\Theta$, the effective central charge of the tight-binding chain is twice that
of the Kitaev model. Moreover, we show that $c(\Theta)$ is always upper bounded by the
zero-temperature central charge of the models.
Upon substituting the asymptotic scaling~\eqref{eq:ent-asy} in~\eqref{eq:mi-intro}, one
obtains that the volume-law term cancels out, and the mutual information exhibits
logarithmic scaling.
The structure of the paper is as follows. In Sec.~\ref{sec:models} we introduce the
fermionic models we are interested in.
In particular, Sec.~\ref{sec:majo} contains the calculation of the Majorana correlation matrix,
which is important to determine entanglement properties, while the main formulas to determine
the von Neumann entropies in terms of correlation functions are reviewed in Sec.~\ref{sec:ent-free-f}.
In Sec.~\ref{sec:lindblad} we summarize the treatment of thermal environments within the
nonlocal Lindblad equation, which was derived in Ref.~\cite{dabbruzzo2021self}; the main result
is formula~\eqref{eq:statio}. In Sec.~\ref{sec:tb-ent} and Sec.~\ref{sec:kitaev-ent}
we derive Eq.~\eqref{eq:ent-asy} for the tight-binding chain and for the Kitaev chain, respectively.
We provide numerical benchmarks of our results in Sec.~\ref{sec:numerics}: in particular,
in Sec.~\ref{sec:vn} we overview the volume-law scaling of the von Neumann entropy, while in
section~\ref{sec:mi} we discuss the scaling of the mutual information.
Our conclusions are drawn in Sec.~\ref{sec:conclusions}.
\section{Models \& methods}
\label{sec:models}
In this section we describe the basic framework used in this
work. We first introduce the quadratic fermionic
Hamiltonians of interest, in Sec.~\ref{sec:free-fermions}.
Then we review the tight-binding chain (Sec.~\ref{sec:tb}) and the Kitaev chain (Sec.~\ref{sec:kitaev}).
In Sec.~\ref{sec:majo} we provide some general formulas for the
two-point correlation of Majorana operators in the two models. These
are essential to study entanglement properties.
In Sec.~\ref{sec:ent-free-f} we summarize the calculation of the von Neumann
entropy via correlation matrix techniques~\cite{jin2004quantum}.
In Sec.~\ref{sec:lindblad} we review the approach of Ref.~\cite{dabbruzzo2021self} to derive a
\emph{nonlocal} Lindblad description of fermionic chains in contact with localized thermal baths.
\subsection{Fermionic quadratic Hamiltonians}
\label{sec:free-fermions}
We focus on free-fermion chains~\cite{lieb1961two, pfeuty1970the}.
Let $\mathcal{S}$ be a quantum system on a lattice with $N$ sites
and $H_\mathcal{S}$ its second-quantized Hamiltonian, which
can be written in terms of fermionic raising and lowering operators $a_n^\dagger, a_n$,
with $n \in \{1,\ldots,N\}$. We assume $H_\mathcal{S}$ to be quadratic, i.e.,
\begin{equation}
\label{eq:ham-gen}
H_\mathcal{S} = \sum_{n,m = 1}^N \left[ Q_{nm} a_n^\dagger a_m +
\frac{P_{nm}}{2} \left( a_n^\dagger a_m^\dagger - a_n a_m \right) \right],
\end{equation}
with $Q,P$ being $N \times N$ real matrices satisfying $Q^T = Q$ and $P^T = -P$.
It is known that $\mathcal{H}_\mathcal{S}$ can be written in diagonal form as
\begin{equation}
\label{eq:ek}
H_\mathcal{S} = E_0 + \sum_k \omega_k b_k^\dagger b_k \, ,
\end{equation}
where $\omega_k \geq 0$ is the single-particle dispersion,
$E_0$ is an irrelevant constant, $b_k^\dagger, b_k$ are new fermionic operators
(the Bogoliubov modes), and the index $k$ denotes the quasimomentum.
The operators $b_k$ are written as linear superpositions of the original fermions.
Specifically, one has
\begin{equation}
\label{eq:bk-def}
b_k \coloneqq \sum_{n=1}^N \left( X_{kn} a_n + Y_{kn}a_n^\dagger \right),
\end{equation}
where $X$ and $Y$ are appropriately chosen $N \times N$ complex
matrices.
For the following it is useful to define the so-called
Lieb-Schultz-Mattis matrices~\cite{lieb1961two} $\phi$, $\psi$ as
\begin{equation}
\label{eq:phipsi-def}
\phi \coloneqq (X + Y)^\dagger, \qquad \psi \coloneqq (X-Y)^\dagger.
\end{equation}
In general, these are complex $N \times N$ matrices that encode
information about the Majorana correlation functions of the model
(see Sec.~\ref{sec:majo}).
\subsubsection{Tight-binding chain}
\label{sec:tb}
The tight-binding chain model is obtained from~\eqref{eq:ham-gen} by fixing
\begin{equation}
Q_{nm}=-J\delta_{n,m-1}-J\delta_{n,m+1}-h\delta_{nm},
\quad P_{nm}=0,
\end{equation}
where $h$ is an external magnetic field strength, and $J$ is the
hopping amplitude between nearest-neighbor sites.
Thus, the Hamiltonian reads as
\begin{equation}
\label{eq:ham-xx}
H_\mathcal{S} = -J\sum_{n=1}^{N} \left( a_n^\dagger a_{n+1} +
a_{n+1}^\dagger a_n \right) - h \sum_{n=1}^N a_n^\dagger a_n,
\end{equation}
where $a_{N+1}$ is determined by the boundary conditions: with OBC
we are choosing $a_{N+1} = 0$, whereas with PBC we have $a_{N+1} = a_1$.
The single-particle dispersion relation [cf.~Eq.~\eqref{eq:ek}] is given by~\footnote{For simplicity,
hereafter we set $J=1$ as the energy scale and work in units of $\hbar = k_B = 1$.}
\begin{equation}
\label{eq:xx-obc-disp}
\omega_k = \big| h + 2\cos(k) \big|,
\end{equation}
where ($n=1,\ldots,N$):
\begin{equation}\label{eq:k_bc}
k = \left\{
\begin{array}{lll}
n\pi/(N+1) & \mbox{(for OBC)}, \\
2\pi n/N & \mbox{(for PBC)}.
\end{array} \right.
\end{equation}
The functions $\phi_{nk}$ and $\psi_{nk}$ [cf.~Eqs.~\eqref{eq:phipsi-def}] are given by
\begin{subequations}
\label{eq:phipsi-xx}
\begin{align}
\label{eq:phi-xx}
& \phi_{nk} = \left\{ \begin{array}{ll} \sqrt{2/(N+1)} \, \sin(kn) & \mbox{(for OBC)}, \vspace*{1mm}\\
e^{ik n}/\sqrt{N} & \mbox{(for PBC)}, \end{array} \right.
\\
\label{eq:psi-xx}
& \psi_{nk} = \mathrm{sgn}\big(-h-2\cos(k) \big) \, \phi_{nk},
\end{align}
\end{subequations}
where $\mathrm{sgn}(x)$ is the sign function and
the corresponding quasimomentum index has to be chosen as in~\eqref{eq:k_bc}.
The ground state of the tight-binding model is annihilated by
all the Bogoliubov operators $b_k$ [cf.~Eq.~\eqref{eq:bk-def}],
and it exhibits criticality for $|h| \leq 2$. For $|h|<2$, its properties are described
by a CFT~\cite{difrancesco1997conformal} with central charge $c=1$.
The entanglement properties of free-fermion systems are encoded
in the fermionic two-point correlation functions~\cite{peschel2009reduced}.
Let us first discuss the open tight-binding chain.
In the limit $N\to\infty$, the ground-state fermionic correlation function
$C^{(\mathrm{obc})}_{nm} \coloneqq \langle a^\dagger_n a_m\rangle$
is given as~\cite{peschel2009reduced}
\begin{equation}
\label{eq:gs-corr}
C^{(\mathrm{obc})}_{nm} = \int_{-\pi}^\pi \frac{dk}{2\pi} \,
\Theta_{\mathrm{H}}(k_F-|k|) \left[ e^{ik(n-m)} - e^{ik(n+m)} \right],
\end{equation}
where $\Theta_{\mathrm{H}}(\cdot)$ is the Heaviside step function,
$n,m\in [1,\infty)$,
and $k_F$ is the Fermi momentum
\begin{equation}
\label{eq:kf}
k_F \coloneqq \arccos\left(-\frac{h}{2}\right).
\end{equation}
Performing the integral in~\eqref{eq:gs-corr}, one obtains
\begin{equation}
\label{eq:gs-corr-1}
C^{(\mathrm{obc})}_{nm}=\frac{\sin(k_F(n-m))}{\pi(n-m)}-
\frac{\sin(k_F(n+m))}{\pi(n+m)}.
\end{equation}
The result for the periodic infinite chain can be recovered from~\eqref{eq:gs-corr-1}
by taking the limit $n,m\to\infty$ with $n-m$ fixed, i.e., by
considering correlators in the bulk of the open chain.
Thus, only the first term in~\eqref{eq:gs-corr} survives and
one obtains
\begin{equation}
\label{eq:gs-corr-pbc}
C^{(\mathrm{pbc})}_{nm}=
\frac{\sin(k_F(n-m))}{\pi(n-m)}.
\end{equation}
It is useful to observe that the first term in~\eqref{eq:gs-corr}
depends only on the difference $n-m$, reflecting translation
invariance, and it defines a so-called Toeplitz matrix~\cite{deifts2011asymptotics},
with symbol $\Theta_H(k_F-|k|)$.
The second term in~\eqref{eq:gs-corr} depends only on $n+m$,
which defines a so-called Hankel matrix.
Thus, the full correlator exhibits a Toeplitz-plus-Hankel structure.
Crucially, for $|h|> 2$ the symbol of the correlator in~\eqref{eq:gs-corr}
is smooth as a function of $k$. On the other hand,
for $|h|<2$ it exhibits a jump discontinuity
at $\pm k_F$, which is the main signature of critical behavior.
This non analytic behavior gives rise to a logarithmic
violations of the area law in the ground-state entanglement
entropies~\cite{calabrese2009entanglemententropy,laflorencie2016quantum}.
In the following sections, by using the approach of Ref.~\cite{dabbruzzo2021self}
we will show that in the presence of thermal baths
locally coupled to the chain, the steady-state fermionic correlator
exhibits a similar structure as in Eq.~\eqref{eq:gs-corr}. In particular,
even though the symbol of the correlator is affected by the presence of the
bath, it is not smooth as a function of $k$. This gives rise to
logarithmic scaling of the steady-state mutual information.
\subsubsection{Kitaev chain}
\label{sec:kitaev}
Let us now consider the Kitaev chain~\cite{Kitaev_2001}. This is obtained
from~\eqref{eq:ham-gen} by choosing
\begin{subequations}
\begin{align}
&Q_{nm}=-J\delta_{n,m-1}-J\delta_{n,m+1}-h\delta_{n,m},\\
& P_{nm}=-\Delta\delta_{n,m-1}+\Delta\delta_{n,m+1},
\end{align}
\end{subequations}
with $J$ the hopping strength, $h$ a magnetic field,
and $\Delta$ the strength of the pairing term.
In the following, we will set $\Delta=J=1$.
Thus, the Hamiltonian of the Kitaev chain reads as
\begin{equation}
\label{eq:kit-ham}
H_\mathcal{S} = -\sum_{n=1}^N \left( a_n^\dagger a_{n+1}
+ a_n^\dagger a_{n+1}^\dagger + \text{h.c.} \right)
- h \sum_{n=1}^N a_n^\dagger a_n.
\end{equation}
For this model we exclusively employ PBC, choosing $a_{N+1} = a_1$, and we assume $h \neq 0$.
The Hamiltonian~\eqref{eq:kit-ham} can be rewritten as in~\eqref{eq:ek}
with single-particle dispersion
\begin{equation}
\label{eq:kit-disp}
\omega_k = \sqrt{h^2 + 4h\cos(k) + 4},
\end{equation}
where the index $k$ is chosen as in~\eqref{eq:k_bc} for PBC.
The functions $\phi_{jk}$ and $\psi_{jk}$ [cf.~Eqs.~\eqref{eq:phipsi-def}]
that encode the Fourier transform and the Bogoliubov transformation
needed to diagonalize~\eqref{eq:kit-ham} are given by
\begin{subequations}
\label{eq:phipsi-is}
\begin{align}
\label{eq:phi-is}
&\phi_{jk} = \frac{e^{ikj}}{\sqrt{N}}, \\
\label{eq:psi-is}
&\psi_{jk} = e^{i\xi(k)}\phi_{jk},
\end{align}
\end{subequations}
where we defined
\begin{equation}
\label{eq:xi-def}
e^{i\xi(k)} \coloneqq -\frac{h+2e^{ik}}{\omega_k}
\end{equation}
and the so-called Bogoliubov angle $\xi(k) \in \mathbb{R}$ is defined by
\begin{subequations}
\label{eq:xi}
\begin{eqnarray}
\label{eq:cosxi}
\cos\xi(k) & = & -\frac{h+2\cos k}{\sqrt{h^2 + 4h\cos k + 4}}, \\
\label{eq:sinxi}
\sin\xi(k) & = & \-\frac{2\sin k}{\sqrt{h^2 + 4h\cos k + 4}}.
\end{eqnarray}
\end{subequations}
It is clear that $\xi(k)$ is continuous as a function
of $k$, for $|h|\ne2$. For $h=2$, a jump discontinuity appears
at $k=\pm \pi$, while for $h=-2$ it emerges for $k=0$.
As for the tight-binding chain (see section~\ref{sec:tb}),
for $|h|=2$ long-wavelength properties of the
ground state of the Kitaev chain are described by a CFT with central
charge $c=1/2$. Consequently, the ground state exhibits logarithmic
violations of the area law for the von Neumann entropies.
Again, below we show that the singular structure of Eqs.~\eqref{eq:xi}
survives in the presence of localized baths,
giving rise to logarithmic scaling of the mutual information.
\subsubsection{Majorana correlation function}
\label{sec:majo}
To determine entanglement-related quantities,
it is convenient to introduce the Majorana operators~\cite{VidalLatorre_2003, VidalLatorre_2004}:
\begin{equation}
\label{eq:majo-def}
w_{2n-1} = \frac{1}{\sqrt{2}}\left( a_n^\dagger + a_n \right),
\quad
w_{2n} = \frac{i}{\sqrt{2}} \left( a_n^\dagger - a_n \right).
\end{equation}
It is straightforward to write these in terms of the Bogoliubov operators $b_k$
that diagonalize the models [cf.~\eqref{eq:ek}]:
\begin{equation}
\begin{bmatrix}
w_{2n-1} \\ w_{2n}
\end{bmatrix} = \frac{1}{\sqrt{2}} \sum_k \begin{bmatrix}
\phi_{nk} & \phi^*_{nk} \\ -i\psi_{nk} & i\psi^*_{nk}
\end{bmatrix} \begin{bmatrix}
b_k \\ b_k^\dagger
\end{bmatrix},
\end{equation}
where $\phi_{nk}$ and $\psi_{nk}$ are given in Eqs.~\eqref{eq:phipsi-def}.
For the tight-binding chain and the Kitaev chain
$\phi_{nk},\psi_{nk}$ are reported
in Eqs.~\eqref{eq:phipsi-xx} and~\eqref{eq:phipsi-is}, respectively.
One can show that the generic expectation value $\langle w_a w_b\rangle$ is
written as
\begin{equation}
\langle w_a w_b \rangle = \frac{\delta_{ab} + i\Gamma_{ab}}{2},
\end{equation}
where $\Gamma$ is a $2N \times 2N$ matrix of the form
\begin{equation}\label{eq:gamma-block-struct}
\Gamma = \begin{bmatrix}
\Pi_{11} & \Pi_{12} & \cdots & \Pi_{1N} \\
\Pi_{21} & \Pi_{22} & \cdots & \Pi_{2N} \\
\vdots & \vdots & \ddots & \vdots \\
\Pi_{N1} & \Pi_{N2} & \cdots & \Pi_{NN}
\end{bmatrix} ,
\end{equation}
with $\Pi_{nm}$ being a $2 \times 2$ block defined by
\begin{equation}\label{eq:gamma-def}
\Pi_{nm} \coloneqq \begin{bmatrix}
0 & \Re[\phi\,\theta\,\psi^\dagger]_{nm} \\
-\Re[\phi\,\theta\,\psi^\dagger]_{mn} & 0
\end{bmatrix}.
\end{equation}
In writing~\eqref{eq:gamma-def} we assumed the
matrix $K_{kq} \coloneqq \langle b_k^\dagger b_q \rangle$ to be
diagonal and the matrix $F_{kq} \coloneqq \langle b_k b_q \rangle$
to be zero, which will turn out to be true in our formalism
(see Sec.~\ref{sec:lindblad}).
Here $\theta_{kq}$ is the occupation of the Bogoliubov modes $b_k$ given by
\begin{equation}
\label{eq:theta-def}
\theta_{kq} = \delta_{kq} \big( 1-2\langle b^\dagger_k b_k\rangle \big).
\end{equation}
Notice that $\Gamma$ is a real skew-symmetric matrix
of even dimension. This means that it has pairs of
eigenvalues $\pm i \nu_r$ with $\nu_r \in \mathbb{R}$.
Let us now specialize the matrix $\Gamma$ to the case of
the tight-binding chain (see section~\ref{sec:tb}).
By using Eqs.~\eqref{eq:phipsi-xx} in~\eqref{eq:gamma-def} we obtain
\begin{equation}
\label{eq:G-def}
G_{nm} \coloneqq \Re [\phi \, \theta \, \psi^\dagger]_{nm}
= \delta_{nm} - 2 \, C_{nm},
\end{equation}
where $C_{nm}=\langle a^\dagger_n a_m\rangle$ is the fermion correlation
function.
This implies that
\begin{equation}
\label{eq:GC}
\Gamma = G \otimes \begin{bmatrix}
0 & 1 \\ -1 & 0
\end{bmatrix},
\end{equation}
from which we conclude that the eigenvalues of
$\Gamma$ are $\pm i \nu_r$ if and only if $\nu_r$ are the eigenvalues of $G$.
Let us now consider the Kitaev chain with PBC.
By using Eqs.~\eqref{eq:phipsi-is} we obtain
\begin{equation}
\label{eq:re-kit}
\Re[\phi \,\theta \,\psi^\dagger]_{nm} = \frac{1}{N} \Re \sum_k
\theta_{kk} e^{-i\xi(k)} e^{i k(n-m)},
\end{equation}
where $\xi(k)$ is defined in~\eqref{eq:xi-def}, and $\theta_{kk}$ is
given in~\eqref{eq:theta-def}.
Using the fact that $\theta_{kk}$ is an even function of $k$ and
$\xi(k)$ is an odd one, we can write
\begin{equation}
\label{eq:gamma-is}
\Pi_{nm} = \int_{-\pi}^\pi \frac{dk}{2\pi}
\begin{bmatrix}
0 & \theta_{kk} e^{-i\xi(k)} \\ -\theta_{kk}e^{i\xi(k)} & 0
\end{bmatrix}
e^{i k(n-m)},
\end{equation}
where we took the thermodynamic limit $N\to\infty$.
Eq.~\eqref{eq:gamma-is} holds for a generic thermodynamic
state, which is characterized by the functions $\theta_{kk}$.
Like for the tight-binding chain, the ground-state of the Kitaev chain is the vacuum state of
all the Bogoliubov operators $b_k$. In this case $\theta_{kk}=1$ [cf.~\eqref{eq:theta-def}],
and the ground-state is characterized by
\begin{equation}
\label{eq:gamma-gs}
\Pi^{(\mathrm{GS})}_{nm} = \int_{-\pi}^\pi \frac{dk}{2\pi}
\begin{bmatrix}
0 & e^{-i\xi(k)} \\ -e^{i\xi(k)} & 0
\end{bmatrix}
e^{i k(n-m)}.
\end{equation}
In the presence of external thermal baths, the Majorana correlator $\Gamma$ is determined
by~\eqref{eq:gamma-is}, with $\theta_{kk}$ encoding the properties of the baths.
\subsection{Entropy in free-fermion systems}
\label{sec:ent-free-f}
For free-fermion systems,
the von Neumann entropy~\eqref{eq:vn} of a subsystem $A$
of length $\ell$ (see Fig.~\ref{fig:sb}), and
the R\'enyi entropies in general~\cite{peschel2009reduced},
are obtained from the Majorana
correlation matrix $\Gamma_A$, which is obtained
from~\eqref{eq:gamma-block-struct} by restricting $n,m \in A$.
If $\pm i \nu_r$ are the eigenvalues of $\Gamma_A$, then
\begin{equation}
\label{eq:S-nu}
S_A = \sum_{r=1}^\ell e(1,\nu_r),
\end{equation}
where we defined the function $e(x,\nu)$ as
\begin{equation}
\label{eq:e-def}
e(x,\nu) \coloneqq -\frac{x-\nu}{2}\ln(\frac{x-\nu}{2}) -
\frac{x+\nu}{2}\ln(\frac{x+\nu}{2}).
\end{equation}
Notice that Eq.~\eqref{eq:S-nu} is well-defined because
one can show that $-1\le\nu_r\le1$.
It is convenient to rewrite the sum in Eq.~\eqref{eq:S-nu} as
an integral in the complex plane. To this purpose we define the determinant
\begin{equation}
\label{eq:D-def}
D_\ell(\lambda) \coloneqq
\det\left(\lambda\mathds{1}-i\Gamma_A\right).
\end{equation}
A straightforward application of Cauchy's theorem allows to
rewrite Eq.~\eqref{eq:S-nu} as
\begin{equation}
\label{eq:ent-C}
S_X = \lim_{\delta, \epsilon \rightarrow 0^+} \frac{1}{4\pi i}
\oint_\gamma d\lambda \, e(1+\epsilon,\lambda)
\frac{d \ln D_\ell(\lambda)}{d\lambda},
\end{equation}
where we used the fact that $e(1,\nu)=e(1,-\nu)$.
The contour $\gamma$ in the complex plane is shown in Fig.~\ref{fig:contour}.
Dashed blue lines in the figure are the branch cuts of $e(1+\epsilon,\lambda)$
at $(-\infty,-1-\epsilon]\cup [1+\epsilon,\infty)$.
The horizontals parts of the contour are shifted by $\delta$ from the real axis.
Finally, the function $d\ln(D_\ell)/d\lambda$
has simple poles in the interval $[-1,1]$ (green dots in the figure).
In the limit $\ell\to\infty$ the poles become dense, forming a new branch cut.
The strategy~\cite{jin2004quantum} to obtain the asymptotic scaling
of $S_A$ in the limit $\ell=|A|\to\infty$
is to first obtain $D_\ell$ in the limit $\ell\to\infty$,
then using it in Eq.~\eqref{eq:ent-C}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.75]{contour.pdf}
\caption{
Contour $\gamma$ in the complex plane for $\lambda$
used to compute the von Neumann
entropy of an interval [cf.~Eq.~\eqref{eq:ent-C}]. Dashed lines
at $(-\infty,-1-\epsilon]\cup [1+\epsilon,\infty)$ denote a branch cut.
Dots in the region $[-1,1]$ are the zeros of $D_\ell(\lambda)$ [cf.~Eq.~\eqref{eq:D-def}],
or, equivalently, the poles of $d D_\ell(\lambda)/d\lambda$.
Here we are interested in the limits $\epsilon\to 0$ and $\delta\to0$.
}
\label{fig:contour}
\end{center}
\end{figure}
\section{Global Lindblad master equation}
\label{sec:lindblad}
To make the paper self contained, we now recap the formalism used to treat self consistently
thermal baths in the Lindblad approximation, within quadratic models~\cite{dabbruzzo2021self}.
Let us consider the interaction between the fermionic chain
$\mathcal{S}$ and the environment $\mathcal{E}$ (see Fig.~\ref{fig:sb}).
The global system $\mathcal{U}={\mathcal S}\cup \mathcal{E}$
is described by the Hamiltonian
$H_\mathcal{U} = H_\mathcal{S} \otimes \mathds{1}_\mathcal{E} + \mathds{1}_\mathcal{S} \otimes
H_\mathcal{E} + H_I$, where $H_\mathcal{E}$ is the Hamiltonian of the
environment and $H_I$ models the interaction between system and
environment. We can always write $H_I$ in the form
\begin{equation}
\label{eq:se-ham}
H_I = \sum_\alpha O_\alpha \otimes R_\alpha,
\end{equation}
where $O_\alpha$ and $R_\alpha$ are Hermitian operators
acting on $\mathcal{S}$ and $\mathcal{E}$, respectively.
In the following we restrict ourselves to the situation in which
$O_\alpha$ act nontrivially only on a finite number of sites of the
chain (see Fig.~\ref{fig:sb}).
Given a state $\rho(t)$ of the entire system $\mathcal{S}\cup \mathcal{E}$,
we are interested in the evolution of the reduced density matrix
$\rho_\mathcal{S}(t) \coloneqq \Tr_\mathcal{E}[\rho(t)]$. In the Markovian regime,
the dynamics is described by a Lindblad master
equation of the form~\cite{petruccione2002the}
\begin{equation}
\label{eq:lin-eq}
\frac{d\rho_\mathcal{S}(t)}{dt} = -i[H, \rho_\mathcal{S}(t)]
+ \mathcal{D}(\rho_\mathcal{S}(t)),
\end{equation}
where both $H$ and $\mathcal{D}$ have to be determined. Let us assume
that the environment consists of a finite number of uncorrelated
fermionic infinite thermal baths, such that
\begin{equation}
H_\mathcal{E} = \sum_\alpha \int dp \, \varepsilon_{\alpha,p} \,
d^\dagger_{\alpha,p} d_{\alpha,p} ,
\end{equation}
where $\alpha$ is here an index that labels the bath, $d_{\alpha,p}$ the
fermionic operators of the bath, and $\varepsilon_{\alpha,p}$ the
bath dispersion. We also consider a generic linear coupling
[cf.~Eq.~\eqref{eq:se-ham}] between the system and the baths:
\begin{align}
\label{eq:o}
& O_\alpha = \sum_{j \in \mathcal{I}_\alpha} \left( a_j + a_j^\dagger \right),\\
\label{eq:r}
& R_\alpha = \int dp \, g_{\alpha,p} \left( d_{\alpha,p} +
d^\dagger_{\alpha,p} \right),
\end{align}
where $g_{\alpha,p}$ is the strength of the coupling and $\mathcal{I}_\alpha$ are
the sites of $\mathcal{S}$ that are coupled to the bath $\alpha$.
Here we focus on the situation in which each bath is coupled to a single
site of the system.
For instance, for the case in Fig.~\ref{fig:sb} (a), one has $\alpha=1,2$
with $\mathcal{I}_1=\{1\}$ and $\mathcal{I}_2=\{N\}$.
The dissipator $\mathcal{D}$ in Eq.~\eqref{eq:lin-eq} can be written as~\cite{dabbruzzo2021self}
\begin{multline}
\label{eq:D-self}
\mathcal{D}(\rho) = \sum_{\alpha,k} \gamma_{\alpha,k} \Big[
(1-f_\alpha(\omega_k)) (2b_k \rho b_k^\dagger - \{b_k^\dagger b_k, \rho\})\\
+ f_\alpha(\omega_k) (2b_k^\dagger \rho b_k - \{b_k b_k^\dagger, \rho\})
\Big] ,
\end{multline}
where we removed the subscript $\mathcal{S}$ in $\rho_{\mathcal{S}}$, and
where $f_\alpha$ is the Fermi-Dirac distribution associated
with bath $\alpha$,
\begin{equation}
\label{eq:fd}
f_\alpha(\omega)= \frac{1}{1 + e^{\beta_\alpha(\omega - \mu_\alpha)}},
\quad \beta_\alpha=\frac{1}{T_\alpha}.
\end{equation}
Here $T_\alpha,\mu_\alpha$ are the temperature of the bath and chemical
potential, respectively. In Eq.~\eqref{eq:D-self}, $b_k$ denote the Bogoliubov
operators that diagonalize the system [cf.~\eqref{eq:bk-def}], whereas $\omega_k$ are the
single-particle energies [cf.~\eqref{eq:ek}]. As is clear from Eq.~\eqref{eq:D-self},
dissipation is encoded via the action of the Bogoliubov modes.
The factors $f_\alpha(\omega_k)$ and $1-f_\alpha(\omega_k)$ reflect the effect of the baths.
Now, although the interaction Hamiltonian $H_I$ is local in space,
the dissipator $\mathcal{D}(\rho)$ is written in terms of Bogoliubov operators
$b_k$, which are nonlocal. In contrast, with common approaches the Lindblad
operators are chosen \emph{ad hoc} and are typically local.
Information about the locality of the baths is encoded in the functions
$\gamma_{\alpha,k}$, defined as
\begin{equation}
\label{eq:gamma-bath}
\gamma_{\alpha,k} \coloneqq J_\alpha(\omega_k) \,
\bigg|\sum_{j \in \mathcal{I}_\alpha}\phi_{jk} \bigg|^2,
\end{equation}
where $\phi_{jk}$ is given in Eq.~\eqref{eq:phipsi-def}
and $J_\alpha(\omega)$ is the spectral density of the bath,
\begin{equation}
\label{eq:J-alpha}
J_\alpha(\omega) \coloneqq \pi\int dp \,
|g_{\alpha,p}|^2 \, \delta(\omega-\varepsilon_{\alpha,p}).
\end{equation}
Besides the dissipative effect encoded in $\mathcal{D}(\rho)$, the presence
of the baths also renormalizes the unitary part of the Lindblad equation~\eqref{eq:lin-eq}.
Indeed, the effective Hamiltonian $H$ reads
\begin{equation}
H = \sum_k \widetilde{\omega}_k b_k^\dagger b_k,
\end{equation}
where the ``dressed'' single-particle dispersion $\widetilde\omega_k$ reads
\begin{equation}
\label{eq:tildeomega}
\widetilde{\omega}_k = \omega_k \Bigg( 1+ \frac{2}{\pi}\sum_\alpha
\Big|\sum_{j \in \mathcal{I}_\alpha}\phi_{jk} \Big|^2
\mathcal{P} \int_0^\infty \!\!\!d\epsilon
\frac{J_\alpha(\epsilon)}{\omega_k^2 - \epsilon^2} \Bigg),
\end{equation}
with $\mathcal{P}$ denoting Cauchy's principal value.
Crucially, the Lindblad equation~\eqref{eq:lin-eq} is derived by using a
\emph{full secular approximation}~\cite{dabbruzzo2021self,dabbruzzo2021self2},
which neglects rapidly oscillating terms $\propto \exp(i (\omega_k-\omega_{k'})t)$.
Moreover, we neglect degeneracy in the spectrum, assuming that $\omega_k\ne \omega_{k'}$ if $k\ne k'$.
Both these approximations are in general uncontrolled, and checking their validity would require
an \emph{ab initio} treatment of the baths.
The master equation~\eqref{eq:lin-eq}
is quadratic in the Bogoliubov operators $b_k, b_k^\dagger$.
This means that if the state of the system is
Gaussian at a certain initial time, it will remain Gaussian
at all times. Therefore, the state $\rho$ is completely
determined by the two-point functions of the Majorana fermions~\eqref{eq:majo-def}.
Equivalently, one can use the correlators
$K_{kq}$ and $F_{kq}$ defined as
\begin{subequations}
\begin{eqnarray}
K_{kq} & \coloneqq & \mathrm{Tr}(\rho \, b^\dagger_k b_q ), \\
F_{kq} & \coloneqq & \mathrm{Tr}(\rho \, b_k b_q).
\end{eqnarray}
\end{subequations}
A direct computation allows to obtain the evolution of $K_{kq}$
and $F_{kq}$ as~\cite{dabbruzzo2021self}
\begin{subequations}
\begin{multline}
\label{eq:Ckq}
\frac{dK_{kq}}{dt} = \Big[ i(\widetilde{\omega}_k - \widetilde{\omega}_q)\\
- \sum_\alpha (\gamma_{\alpha,k} + \gamma_{\alpha,q}) \Big] C_{kq}(t)
+ 2\delta_{kq} \sum_\alpha \gamma_{\alpha,k} f_\alpha(\omega_k),
\end{multline}
\begin{equation}
\label{eq:Fkq}
\frac{dF_{kq}}{dt} = \Big[ -i(\widetilde{\omega}_k + \widetilde{\omega}_q)
- \sum_\alpha (\gamma_{\alpha,k}+\gamma_{\alpha,q}) \Big] F_{kq}(t).
\end{equation}
\end{subequations}
Here $\widetilde\omega_k$ are the modified single-particle energies in Eq.~\eqref{eq:tildeomega},
the rates $\gamma_{\alpha,k}$ are defined in Eq.~\eqref{eq:gamma-bath}, and
$f_{\alpha}(\omega)$ is the Fermi-Dirac distribution of the bath [cf.~Eq.~\eqref{eq:fd}].
We anticipate that, for the tight-binding chain and for the Kitaev chain, due to the
simple structure of Eqs.~\eqref{eq:phipsi-xx} and~\eqref{eq:phipsi-is},
the dependence on $\gamma_{\alpha,k}$ drops out.
Assuming that $\gamma_{\alpha,k}$ are not all equal to zero (which is
obviously true if the system is actually coupled to the environment),
in the stationary limit $t \rightarrow \infty$ we obtain
\begin{equation}
\label{eq:statio}
K_{kq} = \delta_{kq}\frac{\sum_\alpha \gamma_{\alpha,k} f_\alpha(\omega_k)}
{\sum_\alpha \gamma_{\alpha,k}}, \quad F_{kq}=0.
\end{equation}
Thus the correlation function in momentum space becomes diagonal, and it is
a convex combination of the Fermi-Dirac distributions of the baths.
Equation~\eqref{eq:statio} is the main ingredient to extract steady-state
properties of the system (see Sec.~\ref{sec:tb-ent} and Sec.~\ref{sec:kitaev-ent}).
Note that, if the baths are identical ($f_\alpha$ does
not depend on $\alpha$), we obtain $K_{kq}=\delta_{kq}f(\omega_k)$.
Interestingly, even in this situation, the statistical ensemble that describes
the steady state is not the standard finite-temperature
ensemble of the underlying free-fermion model, due to the nonzero
chemical potential in Eq.~\eqref{eq:fd}. As we will show in the following,
this implies that the steady-state von Neumann entropy exhibits logarithmic additive
corrections to the expected volume-law scaling at finite temperature.
\section{Scaling of entropy in the tight-binding chain}
\label{sec:tb-ent}
In this section we derive the scaling equation~\eqref{eq:ent-asy}
of the steady-state von Neumann entropy
for a subinterval of the tight-binding chain [cf.~Eq.~\eqref{eq:ham-xx}].
A similar calculation was performed in~\cite{Eisler2014,Kormos2017},
but for a nonequilibrium setting which is different from ours.
First of all, using the definitions of $\phi_{nk}$ and $\psi_{nk}$ reported
in Eqs.~\eqref{eq:phipsi-xx}, the steady-state fermionic correlation matrix $G_{nm}$
of Eq.~\eqref{eq:G-def} for $N\to\infty$ reads
\begin{equation}
\label{eq:gnm}
G_{nm} = \int_{-\pi}^\pi \frac{dk}{2\pi} \,
\widetilde\theta_{kk} \left[ e^{ik(n-m)} - \zeta e^{ik(n+m)} \right],
\end{equation}
where $\zeta=0$ and $\zeta=1$ corresponds to PBC and OBC, respectively.
This equation defines a Toeplitz matrix for $\zeta=0$, whereas one
has a Toeplitz-plus-Hankel matrix~\cite{deifts2011asymptotics} for $\zeta=1$.
The so-called symbol of $G_{nm}$ is
\begin{equation}
\label{eq:tk}
\widetilde\theta_{kk} \coloneqq \theta_{kk}\,\mathrm{sgn} \big( -h-2\cos(k) \big),
\end{equation}
where $\theta_{kk}$ is given in Eq.~\eqref{eq:theta-def} and the sign function
is the same as in Eq.~\eqref{eq:psi-xx}. The function $\theta_{kk}$ encodes
the information about the steady state and is obtained from~\eqref{eq:statio}.
For now we consider generic $\theta_{kk}$, specializing to the situations
described in Fig.~\ref{fig:sb} (a) and (b), in Sec.~\ref{sec:xx-single}
and~\ref{sec:xx-two-baths} respectively.
The function $\widetilde\theta_{kk}$ is smooth, except for
the sign function which displays a singularity for $|h|<2$,
thus giving rise to logarithmic corrections to the von Neumann entropy.
To extract the scaling behavior of the mutual information~\eqref{eq:mi-intro}
one has to determine the asymptotic scaling of the von Neumann entropy
$S_A$ for a subsystem $A$ of length $\ell\to\infty$. To that purpose,
we first study the asymptotic behavior of
$D_\ell(\lambda)\equiv D_\ell[g_\lambda]$ in Eq.~\eqref{eq:D-def},
being the determinant of the Toeplitz (or Toeplitz-plus-Hankel) matrix with symbol
$g_\lambda(k) \coloneqq \lambda-\widetilde\theta_{kk}$ given by
\begin{equation}
\label{eq:tilde-simb}
g_\lambda(k)=\left\{
\begin{array}{cl}
\lambda-\theta_{kk} & \;\; k\in[-\pi,-k_F]\cup [k_F,\pi) ,\\
\lambda+\theta_{kk} & \;\; k\in[-k_F,k_F) .
\end{array}
\right.
\end{equation}
\subsubsection{Periodic boundary conditions}
Let us first consider the case of PBC [$\zeta=0$ in Eq.~\eqref{eq:gnm}].
From that equation we obtain
\begin{equation}
\label{eq:matg}
\lambda\delta_{nm}-G_{nm}=
\int_{-\pi}^\pi\frac{dk}{2\pi} e^{ik(n-m)} g_\lambda(k).
\end{equation}
We now apply a specialized version of the Fisher-Hartwig
conjecture~\cite{fisher1969toeplitz}~\footnote{The Fisher-Hartwig conjecture
was proven by Basor in Ref.~\cite{basor1991the,basor1994the}.}
in which the symbol $g_\lambda(k)$ is allowed to have only jump
discontinuities at a finite number of points $k_r$.
To apply it one has to rewrite $g_\lambda(k)$ in the form
\begin{equation}
\label{eq:symbol}
g_\lambda(k) = g_s(k) \prod_{r=1}^R e^{ib_r( k-k_r-\pi \, \mathrm{sgn}(k-k_r) )}.
\end{equation}
Here $g_s(k)$ is a smooth function of $k$, $R$ is the number of
discontinuities of the symbol,
and $b_r,k_r$ are real constants. The Fisher-Hartwig conjecture
states that, in the limit $\ell\to\infty$, one has
\begin{equation}
\label{eq:D-asy}
D_\ell[g_\lambda] \sim F[g_s]^\ell \Bigg(\prod_{j=1}^R\ell^{-b_j^2} \Bigg) E[g_\lambda],
\end{equation}
where we defined
\begin{equation}
\label{eq:F-capital}
F[g_s] \coloneqq \exp\left(\int_{-\pi}^\pi\frac{dk}{2\pi}\ln g_s(k)\right).
\end{equation}
From Eq.~\eqref{eq:ent-C} it is clear that the first factor in~\eqref{eq:D-asy}
gives a volume-law von Neumann
entropy, and it is not sensitive to the singularities in the symbol
$g_\lambda(k)$. The second factor is responsible for the logarithmic
scaling of the von Neumann entropy and contains information about
the singularities of $g_\lambda(k)$. The constant
$E$ is a known function of $g_\lambda(k)$. In the following, we are not considering
$E$ because we are interested only in the linear growth of the von Neumann entropy and
in the logarithmic correction. The Fisher-Hartwig conjecture
has been employed in the literature to determine the scaling behavior of the von Neumann
and R\'enyi entropies in the ground state of critical fermionic
chains~\cite{jin2004quantum,calabrese2010universal,fagotti2011universal}.
It is straightforward to check that in our case the symbol $g_\lambda(k)$
in~\eqref{eq:tilde-simb} can be written in the form~\eqref{eq:symbol}
with two discontinuities at $k_1=-k_F$ and $k_2=k_F$, i.e., with $R=2$, and
\begin{align}
\label{eq:b}
b_1=-b_2=\beta_\lambda+m,
\end{align}
where $m$ is an integer and
\begin{equation}
\label{eq:beta}
\beta_\lambda=\frac{1}{2\pi i}\ln\left(\frac{\lambda-
\Theta}{\lambda+\Theta}\right), \quad
\mathrm{with}\,
\Theta:=\left.\theta_{kk}\right|_{k=k_F}.
\end{equation}
Importantly, the parameter $\Theta$ for the tight-binding chain is the
steady-state density of Bogoliubov excitations at the Fermi level,
which contains information about the environment. We anticipate that $\Theta$
fully determines the logarithmic scaling of the von Neumann entropy.
The function $g_s(k)$ is given by
\begin{multline}
\label{eq:f}
g_s(k)=\left(\frac{\lambda+\Theta}{\lambda-\Theta}\right)^{\frac{k_F}{\pi}-1}
(\lambda+\theta_{kk})\Theta_H(k_F-|k|)\\+
\left(\frac{\lambda+\Theta}{\lambda-\Theta}\right)^{\frac{k_F}{\pi}}
(\lambda-\theta_{kk})\Theta_H(|k|-k_F).
\end{multline}
Notice that $g_s(k)$ is not smooth as a function of $k$ because it has a cusp-like
singularity at $k=\pm k_F$.
However, this is not expected to affect the logarithmic
subleading term in the von Neumann entropy, as we are going to verify numerically
in section~\ref{sec:numerics}. Now,
upon substituting~\eqref{eq:b}~\eqref{eq:beta}~\eqref{eq:f}
in~\eqref{eq:D-asy}, we obtain that the von Neumann entropy of an interval
$A$ of length $\ell$ embedded in an infinite chain, in the limit $\ell\to\infty$
is given by
\begin{equation}
\label{eq:ent-expansion}
S_A = \alpha\ell+\frac{c(\Theta)}{3}\ln(\ell)+{\mathcal O}(1).
\end{equation}
Here $c(\Theta)$, with $\Theta$ defined in~\eqref{eq:beta}, originates
from the term in the round brackets in~\eqref{eq:D-asy} and will
be determined below. The constant $\alpha$ in~\eqref{eq:ent-expansion}
is determined by the first term in~\eqref{eq:D-asy}. By using~\eqref{eq:ent-C}
one obtains
\begin{equation}
\label{eq:line-inte}
\alpha \! = \!
\lim_{\delta,\epsilon \rightarrow 0^+} \frac{1}{4\pi^2 i}
\oint_\gamma d\lambda\int_{-\pi}^\pi dk \frac{e(1+\epsilon,\lambda)}
{\lambda+\mathrm{sgn}(k_F-|k|)\theta_{kk}},
\end{equation}
where $e(x,\nu)$ is defined in~\eqref{eq:e-def}, $\theta_{kk}$ in~\eqref{eq:theta-def},
and $\gamma$ denotes the contour shown in Fig.~\ref{fig:contour}.
Performing such integral with the residue theorem, we get
\begin{eqnarray}
\label{eq:lin-coeff}
\alpha & = &\int_{-\pi}^\pi \frac{dk}{2\pi} e(1,\theta_{kk}), \\ \nonumber
& = & \! - \! \int_{-\pi}^\pi\frac{dk}{2\pi}\left[K_{kk}\ln(K_{kk})+(1-K_{kk})\ln(1-K_{kk})\right],
\end{eqnarray}
$K_{kk}$ being the steady-state correlation function given in Eq.~\eqref{eq:statio}.
Equation~\eqref{eq:lin-coeff} is the steady-state density (i.e., the von Neumann entropy per volume)
of the full system in the $N\to\infty$ limit. Specifically, we have that
\begin{equation}
\alpha=\lim_{\ell\to\infty}\frac{S_X}{\ell}=\lim_{N\to\infty}\frac{S_N}{N},
\end{equation}
where we denoted with $S_N$ the entropy of the full system. For pure quantum
states, one has either $K_{kk}=0$ or $K_{kk}=1$, which implies that
$S_N=0$ and $\alpha=0$, as it should be. This is not the case in the presence
of the environment, because the global system is not pure.
A similar behavior is typically observed in out-of-equilibrium
generic quadratic fermionic and bosonic
systems~\cite{alba2021spreading,carollo2022dissipative,alba2022hydrodynamics,alba2022logarithmic}.
Let us now derive the prefactor $c(\Theta)$ of the logarithmic
growth in Eq.~\eqref{eq:ent-expansion}. By using~\eqref{eq:ent-C},
the term in the brackets in~\eqref{eq:D-asy} yields
\begin{equation}
\label{eq:ctheta}
\frac{c(\Theta)}{3} = \lim_{\delta,\epsilon \rightarrow 0^+} \frac{1}{\pi i}
\oint_\gamma d\lambda \, e(1+\epsilon,\lambda) \frac{d(-\beta_\lambda^2)}{d\lambda},
\end{equation}
where $\beta_\lambda$ is defined in Eq.~\eqref{eq:beta} and,
again, $\gamma$ denotes the ``dogbone'' contour in Fig.~\ref{fig:contour}.
We can perform an integration by parts to obtain
\begin{equation}
\label{eq:dogbone}
\frac{c(\Theta)}{3}=
\lim_{\delta,\epsilon \rightarrow 0^+} \frac{1}{8\pi^3 i}
\oint_\gamma d\lambda \, \ln^2 \left( \frac{\lambda+\Theta}
{\lambda-\Theta} \right) \ln(\frac{1+\epsilon+\lambda}
{1+\epsilon-\lambda}).
\end{equation}
The contribution of the circles around $\pm 1$ in $\gamma$ (see Fig.~\ref{fig:contour})
vanish in the limit $\epsilon \to 0^+$. The integration
along the horizontal paths can be performed using the fact that, for $\delta\to0^+$, one has
\begin{equation}
\label{eq:id-delta}
\ln(\frac{x \pm i\delta + t}{x \pm i \delta - t})
\to \ln\left|\frac{t + x}{t - x}\right| \mp i\pi
\text{sgn}(t)\Theta_H(|t|-|x|).
\end{equation}
Finally, the integral~\eqref{eq:dogbone} gives
\begin{equation}
\label{eq:ctheta-int}
c(\Theta) = \frac{3}{2\pi^2} \int_{-\Theta}^{\Theta}
dx \log(\frac{\Theta + x}{\Theta - x})
\log(\frac{1+x}{1-x}),
\end{equation}
which can be performed analytically, yielding
\begin{multline}
\label{eq:ctheta-res}
c(\Theta)=\\
\frac{3}{\pi^2}\Big\{\mathrm{Li}_{2}\Big(\frac{2\Theta}{\Theta-1}\Big)
+\mathrm{Li}_{2}\Big(\frac{2\Theta}{\Theta+1}\Big)
+2\Theta\mathrm{Re}\Big[\mathrm{Li}_{2}\Big(\frac{1+\Theta}{1-\Theta}\Big)\Big]
\\-\frac{1}{3}\pi^2\Theta+\frac{\Theta}{2}\ln\Big(\frac{(1-\Theta)^3(1+\Theta)}
{16\Theta^4}\Big)\ln\Big(\frac{1-\Theta}{1+\Theta}\Big)\Big\}.
\end{multline}
Here $\mathrm{Li}_n(z)$ is the polylogarithm function~\cite{nist2022digital}.
In particular, $\mathrm{Li}_2(x)$ is
\begin{equation}
\mathrm{Li}_2(x) \coloneqq -\int_{0}^x dz\frac{\ln(1-z)}{z}.
\end{equation}
In Eq.~\eqref{eq:ctheta-res}
we restricted ourselves to $0\le \Theta\le 1$ because
$c(\Theta)$ is an even function of $\Theta$, as is clear from Eq.~\eqref{eq:ctheta-int}.
This result is somewhat reminiscent of the effective
central charge obtained in free fermion chains in the presence of
defects~\cite{eisler2010entanglement,eisler2012on,calabrese2011entanglement,calabrese2011the,calabrese2012entanglement}.
In the limit $\Theta\to 1$ one
recovers the case without baths. Thus, Eq.~\eqref{eq:ctheta-res} gives the
central charge $c=1$ of the model. For $\Theta\to1$, which corresponds
to a low-temperature limit, one has
\begin{multline}
\label{eq:c-low}
c(\Theta)=1+\frac{\Theta-1}{2\pi^2}\Big\{6+2\pi^2+3\ln^2(2)+6\ln(2)\\
+3\Big[\ln\Big(\frac{1-\Theta}{4}\Big)-2\Big]\ln(1-\Theta)\Big\}
+{\mathcal O}((\Theta-1)^2).
\end{multline}
Notice the logarithmic term $\ln(1-\Theta)$.
In the opposite limit $\Theta\to0$ one obtains that $c(\Theta)$ vanishes as
\begin{equation}
\label{eq:c-high}
c(\Theta)=\frac{1}{\pi^2}\Big(6\Theta^2+\frac{4}{3}\Theta^4\Big)+{\mathcal{O}(\Theta^6)}.
\end{equation}
The results derived so far hold for $|h|<2$. Away from this region, i.e, for
$|h|>2$, the von Neumann entropy is given by~\eqref{eq:ent-expansion} with
$c=0$ and $\alpha$ given by~\eqref{eq:lin-coeff}.
The behavior of $c(\Theta)$ as a function of $\Theta$ is illustrated in
Fig.~\ref{fig:theo}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.8]{charge.pdf}
\caption{Effective central charge $c(\Theta)$ for the tight-binding
chain coupled to localized thermal baths. Here $\Theta \in [-1,1]$
encodes the information about the baths. For a single bath
at the left edge of the chain, $\Theta$ is given by Eq.~\eqref{eq:theta-single}.
For two baths at the edges of the chain, $\Theta$ is given by Eq.~\eqref{eq:theta-is-3}.
In the limit $|\Theta|\to1$, one recovers the CFT result $c=1$.
For $\Theta=0$ one has $c\to0$, which corresponds to a high-temperature limit.}
\label{fig:theo}
\end{center}
\end{figure}
\subsubsection{Open boundary conditions}
Let us now discuss the case with OBC and consider a block
of $\ell$ sites starting at one edge of the chain [see Fig.~\ref{fig:sb} (a)].
Now, one has $\zeta=1$ in the fermionic correlator~\eqref{eq:gnm}, which has the
Toeplitz-plus-Hankel structure with the same symbol. A version of the Fisher-Hartwig
conjecture for Toeplitz-plus-Hankel matrices exists~\cite{deifts2011asymptotics}, and
the analysis can be carried out as for PBC.
However, the application of the Fisher-Hartwig conjecture is more cumbersome
and we do not report the details.
Following the approach of Ref.~\cite{fagotti2011universal} we obtain that
\begin{equation}
\label{eq:ent-expansion-open}
S_X = \alpha\ell +\frac{c(\Theta)}{6}\ln(\ell)+{\mathcal O}(1),
\end{equation}
where $c(\Theta)$ is the same as in Eq.~\eqref{eq:ent-expansion}.
\subsection{Single bath}
\label{sec:xx-single}
To illustrate our results, we first focus on the open tight-binding
chain with only one edge coupled to the environment
(see Fig.~\ref{fig:sb}) and $|h|<2$.
In the steady state, from Eq.~\eqref{eq:statio} we have
\begin{equation}
\label{eq:prev}
K_{kq} = \langle b^\dagger_k b_q\rangle=\delta_{kq}f(\omega_k),
\end{equation}
where $b_k$ are the Bogoliubov operators of Eq.~\eqref{eq:ek}, $\omega_k$ their
single-particle energies [cf.~Eq.~\eqref{eq:xx-obc-disp}], and $f(\omega_k)$
the Fermi-Dirac distribution describing the bath [cf.~Eq.~\eqref{eq:fd}].
Eq.~\eqref{eq:prev} implies that the function $\theta_{kk}$ [cf.~\eqref{eq:theta-def}]
is given by
\begin{equation}
\theta_{kk} = 1-2f(\omega_k).
\end{equation}
The prefactor of the logarithmic contribution in the von Neumann entropy
[cf.~\eqref{eq:ent-expansion}] depends on the value of $\theta_{kk}$ at the
Fermi point $k_F$ [cf.~\eqref{eq:beta}]. Thus, we have
\begin{equation}
\label{eq:theta-single}
\Theta \coloneqq \left.\theta_{kk}\right|_{k=k_F} = 1-2f(0)
= \tanh(-\frac{\mu_L}{2T_L}).
\end{equation}
Notice that $\Theta$ depends only on the ratio $\mu_L/T_L$.
The limit $\mu_L/T_L\to 0$ gives $\Theta\to0$ and a vanishing
$c(\Theta)$ (see Fig.~\ref{fig:theo}). In the limit $\mu_L/T_L\to 0$
the logarithmic correction to the entropy disappears.
Specifically, from~\eqref{eq:c-high} we
find that $c(\Theta)\sim (\mu_L/T_L)^2$. This regime corresponds to either
infinite temperature $T_L\to\infty$ or vanishing chemical potential $\mu_L\to0$.
On the other hand, for $|\mu_L/T_L|\to\infty$ we have $|\Theta|\to1$ and
$c\to1$. Moreover, in this limit $K_{kk}\to0$ or $K_{kk}\to1$, which implies
that $\alpha\to0$ [cf.~\eqref{eq:lin-coeff}]. Thus,
[see Eq.~\eqref{eq:ent-expansion-open}] we recover the ground-state scaling
of the von Neumann entropy.
\subsection{Two baths}
\label{sec:xx-two-baths}
Let us now discuss the case with two different thermal baths,
first considering the simpler case of PBC
with the baths placed at sites $n=1$ and $n=N/2$ [see Fig.~\ref{fig:sb} (b)].
Under the assumption that the spectral density $J(\omega)$
[cf.~\eqref{eq:J-alpha}] is the same for the two baths, we have for the couplings
\begin{equation}
\gamma_{L,k} = \gamma_{R,k} = \frac{J(\omega_k)}{N},
\end{equation}
since, from Eq.~\eqref{eq:phi-xx}, we have $|\phi_{1k}|^2 = |\phi_{Nk}|^2 = 1/N$.
Using Eq.~\eqref{eq:statio}, we thus obtain
\begin{align}
\label{eq:Ckq-1}
& K_{kq} = \delta_{kq}\frac{f_L(\omega_k) + f_R(\omega_k)}{2},\\
&
\label{eq:theta-is-2}
\theta_{kk}=\frac{1}{2}\left[\tanh\left(\frac{\omega_k-\mu_L}{2T_L}\right)
+\tanh\left(\frac{\omega_k-\mu_R}{2T_R}\right)
\right].
\end{align}
Notice that the correlator $K_{kq}$ does not depend on the
couplings $\gamma_{L/R,k}$. Moreover, as is clear from~\eqref{eq:Ckq-1},
the steady-state correlator is written in terms of the average between the
Fermi-Dirac distributions describing the baths.
From Eq.~\eqref{eq:theta-is-2} we obtain
\begin{equation}
\label{eq:theta-is-3}
\Theta =
\frac{1}{2}\left[\tanh\left(-\frac{\mu_L}{2T_L}\right)
+\tanh\left(-\frac{\mu_R}{2T_R}\right)
\right].
\end{equation}
Notice that $\Theta$ depends only the ratios $\mu_L/T_L$ and $\mu_R/T_R$,
which is structurally similar to what we obtained in the
single-bath scenario.
In the case of OBC and baths placed at the edges of the chain [see Fig.~\ref{fig:sb} (a)],
one has for the couplings
\begin{subequations}
\begin{align}
& \gamma_{L,k} = J(\omega_k) \frac{2 \sin^2(k)}{N+1},\\
& \gamma_{R,k} = J(\omega_k) \frac{2 \sin^2(kN)}{N+1}.
\end{align}
\end{subequations}
Making use of the quantization condition on $k$, one can show
with a straightforward calculation that
Eqs.~\eqref{eq:Ckq-1}, \eqref{eq:theta-is-2}, and~\eqref{eq:theta-is-3}
continue to remain valid.
\section{Scaling of entropy in the Kitaev chain}
\label{sec:kitaev-ent}
We now turn to the steady-state von Neumann entropy in the Kitaev chain with PBC.
The blocks of the Majorana correlation matrix $\Gamma$ [cf.~Eq.~\eqref{eq:gamma-is}]
read as
\begin{equation}
\label{eq:gamma-is-2}
\Pi_{nm} = \int_{-\pi}^\pi \frac{dk}{2\pi}
\begin{bmatrix}
0 & \theta_{kk} e^{-i\xi(k)} \\ -\theta_{kk}e^{i\xi(k)} & 0
\end{bmatrix}
e^{i k(n-m)},
\end{equation}
where $\xi(k)$ is defined in Eq.~\eqref{eq:xi-def} and $\theta_{kk}$ is the
same as in Eq.~\eqref{eq:theta-is-2}.
Notice that for $\mu_L=\mu_R=0$ and $T_L=T_R=T$ one recovers the correlator
for the finite-temperature Kitaev (and Ising)
chain~\cite{barouch1970statistical,barouch1971statistical,barouch1971statisticalmechanics}
at temperature $T$. For $T\to0$ one has that $\theta_{kk}\to1$, thus recovering
the zero-temperature correlator of the Kitaev chain.
Importantly, the presence of $\mu_L$ and $\mu_R$ in~\eqref{eq:theta-is-2} implies
that the statistical ensemble describing the steady state is not the
usual finite-temperature ensemble of the Kitaev chain.
For the following it is important to stress that the function $\xi(k)$ has
jump discontinuities for $|h|=2$, which determine the ground-state critical properties
of the system.
Now, Eq.~\eqref{eq:gamma-is-2} defines a Toeplitz matrix.
Given a subsystem $A$, we are interested in the matrix $\lambda\mathds{1}-i\Gamma_A$
[cf.~Eq.~\eqref{eq:D-def}], which is then also of the Toeplitz type. Let us define
its symbol $g_\lambda(k)$ (cf.~\eqref{eq:gamma-is-2}) as
\begin{equation}
\label{eq:glambda}
g_\lambda(k) \coloneqq
\begin{bmatrix}
\lambda & -i \theta_{kk} e^{-i\xi(k)} \\ i\theta_{kk}e^{i\xi(k)} & \lambda
\end{bmatrix}.
\end{equation}
In contrast with the tight-binding chain, for which it was a scalar,
now the symbol is a two-by-two matrix.
At zero temperature, the asymptotic behavior in the large-$\ell$
limit of the Toeplitz matrix~\eqref{eq:gamma-is-2} has been obtained
in Ref.~\cite{jin2004quantum}. Here we are only interested in the
logarithmic correction to the volume-law
scaling of the von Neumann entropy. Thus, we can use the techniques of
Refs.~\cite{ares2018entanglement,ares2019sublogarithmic}. The idea is that since the logarithmic
correction depends only on the singularities of the symbol, we are allowed to
modify the latter, provided that we do not change its singularity structure.
This eventually allows one to work with a scalar symbol.
To proceed, let us define the modified symbol $\widetilde g_\lambda(k)$ as
\begin{equation}
\label{eq:gtilde}
\widetilde{g}_\lambda(k) \coloneqq
\begin{bmatrix}
\lambda & -i\theta_{kk} e^{-i|\xi(k)|} \\
i\theta_{kk} e^{i|\xi(k)|} & \lambda
\end{bmatrix},
\end{equation}
which differs from $g_\lambda(\kappa)$ in~\eqref{eq:glambda} because of
the absolute value $|\xi(k)|$ in the phase factors.
Its inverse $\widetilde g^{-1}_\lambda(k)$ is
\begin{equation}
\label{eq:gtilde-i}
\widetilde{g}_\lambda^{-1}(k) =
\frac{1}{\lambda^2 - \theta_{kk}^2}
\begin{bmatrix}
\lambda & i\theta_{kk} e^{-i|\xi(k)|} \\
-i\theta_{kk} e^{i|\xi(k)|} & \lambda
\end{bmatrix}.
\end{equation}
Crucially, both $\widetilde g_\lambda$ and its inverse are smooth functions
of $k$. As a consequence, in the limit $\ell\to\infty$,
the corresponding Toeplitz determinants
$D_\ell[\widetilde g_\lambda]$ and $D_\ell[\widetilde g^{-1}_\lambda]$
do not contain logarithmic terms.
Their asymptotic behavior is determined
by the Szeg\"o-Widom theorem~\cite{widom1976asymptotic}. Given a generic
smooth symbol $z(k)$ of a Toeplitz matrix, the Szeg\"o-Widom theorem gives
\begin{equation}
\label{eq:widom}
\ln D_\ell[z] = \ell \int_{-\pi}^\pi \frac{dk}{2\pi} \ln\det(z(k))
+ \mathcal{O}(1).
\end{equation}
In our case we have $z(k)=\widetilde g_\lambda(k)$ or
$\widetilde g^{-1}_\lambda(k)$.
Moreover, we can use the so-called Basor localization
theorem~\cite{basor1979a}, which allows to write, in the limit $\ell\to\infty$,
\begin{equation}
\label{eq:Dlambda}
\ln D_\ell[{g}_\lambda] = \ln D_\ell[{g}_\lambda
\widetilde{g}_\lambda^{-1}] -
\ln D_\ell[\widetilde{g}_\lambda^{-1}] + \mathcal{O}(1).
\end{equation}
Here the first contribution contains logarithmic terms,
whereas the second one gives rise to volume-law terms as in~\eqref{eq:widom}.
To proceed, let us now notice that
\begin{equation}
\label{eq:ggtilde}
{g}_\lambda \widetilde{g}_\lambda^{-1} \!= \!
\frac{1}{\lambda^2 - \theta_{kk}^2}
\begin{bmatrix}
\lambda^2 - \theta_{kk}^2 e^{i(|\xi|-\xi)} &
i\lambda\theta_{kk} (e^{-i|\xi|} \!-\! e^{-i\xi}) \\
i\lambda\theta_{kk} (e^{i\xi} \!-\! e^{i|\xi|}) &
\lambda^2 - \theta_{kk}^2 e^{i(\xi - |\xi|)}
\end{bmatrix}\!.
\end{equation}
If $-\pi \leq k \leq 0$, then $\xi(k) \geq 0$ and $g_\lambda\widetilde g_\lambda^{-1}$
is the identity matrix. On the other hand, if $0 < k \leq \pi$, then $\xi(k) < 0$
and Eq.~\eqref{eq:ggtilde} becomes
\begin{equation}
\label{eq:ggtilde-1}
{g}_\lambda \widetilde{g}_\lambda^{-1} =
\frac{1}{\lambda^2 - \theta_{kk}^2}
\begin{bmatrix}
\lambda^2 - \theta_{kk}^2 e^{-2i\xi} &
-2\lambda\theta_{kk} \sin(\xi) \\
-2\lambda\theta_{kk} \sin(\xi) &
\lambda^2 - \theta_{kk}^2 e^{2i\xi}
\end{bmatrix}.
\end{equation}
Diagonalizing the matrix in~\eqref{eq:ggtilde-1}, one obtains that the eigenvalues $b_\pm$ are
\begin{equation}
\label{eq:is-eig}
b_\pm = \left[ \frac{\sqrt{\lambda^2 - \theta_{kk}^2 \cos^2(\xi)} \pm
|\theta_{kk} \sin(\xi)|}
{\sqrt{\lambda^2 - \theta_{kk}^2}} \right]^2 .
\end{equation}
One can easily verify that the corresponding eigenvectors are smooth
functions of $k$. Hence, a further application of Basor localization
theorem yields
\begin{equation}
\label{eq:semi-final}
\ln D_\ell[g_\lambda \widetilde{g}_\lambda^{-1}] =
\ln D_\ell[b_-] + \ln D_\ell[b_+] + \mathcal{O}(1).
\end{equation}
Here we also used that, according to the Szeg\"o-Widom theorem~\eqref{eq:widom},
the contribution in Eq.~\eqref{eq:semi-final} of the matrices that diagonalize~\eqref{eq:ggtilde-1}
would be a constant that we can neglect in the limit $\ell\to\infty$,
because it is subleading. Now in~\eqref{eq:semi-final},
$D_\ell[b_{\pm}]$ are determinants of Toeplitz matrices with scalar
symbols. Their asymptotic behavior for large $\ell$
can be determined by using the Fisher-Hartwig conjecture for
scalar symbols (as in section~\ref{sec:tb-ent}).
We obtain
\begin{multline}
\label{eq:bpm-asy}
\ln D_\ell[b_\pm] = \ell \int_{-\pi}^\pi \frac{dk}{2\pi} \, \ln b_\pm(k)
\\+ \ln^2 \left[ \frac{\sqrt{\lambda^2}
\pm |\Theta|}{\sqrt{\lambda^2 - \Theta^2}} \right] \frac{\ln(\ell)}{\pi^2} + \mathcal{O}(1),
\end{multline}
where
\begin{equation}
\label{eq:table}
\Theta = \left\{
\begin{array}{cl}
\left.\theta_{kk}\right|_{k=\pi} & \mathrm{for}\,\,h=2,\\ \\
\left.\theta_{kk}\right|_{k=0} & \mathrm{for}\,\,h=-2.
\end{array}\right.
\end{equation}
Since $\phi_{nk}$ for the Kitaev chain has the same form of
the corresponding quantity for the tight-binding chain,
one can see that in both cases
$\Theta$ is given by Eq.~\eqref{eq:theta-single} for the geometry with only
one edge of the chain coupled to the environment
and by Eq.~\eqref{eq:theta-is-3} with two baths.
Noticing that $b_- b_+ = 1$, and using Eqs.~\eqref{eq:widom}, \eqref{eq:Dlambda},
\eqref{eq:bpm-asy}, we get
\begin{eqnarray}
\ln D_\ell[g_\lambda] & = &
\ell \int_{-\pi}^\pi \frac{dk}{2\pi} \ln(\lambda^2 - \theta_{kk}^2) \nonumber\\
&&+ \ln^2\left[ \frac{\sqrt{\lambda^2} + |\Theta|}{\sqrt{\lambda^2 - \Theta^2}}
\right]\!\frac{2\ln(\ell)}{\pi^2} + \mathcal{O}(1). \quad
\label{eq:kitaev-fin}
\end{eqnarray}
We can now determine the scaling of the steady-state
von Neumann entropy. The von Neumann entropy
of a subsystem $A$ of size $\ell$ embedded in an infinite periodic
chain in the large $\ell$ limit is given by
\begin{equation}
\label{eq:kitaev-deco}
S_A = \alpha\ell+\frac{c(\Theta)}{3}\ln(\ell)+\mathcal{O}(1).
\end{equation}
Equation~\eqref{eq:kitaev-deco} has the same form
of Eq.~\eqref{eq:ent-expansion}, although $\alpha$ and $c(\Theta)$ are not
the same. The coefficient $\alpha$ of the volume-law term is easily
obtained by using~\eqref{eq:ent-C}:
\begin{equation}
\label{eq:alpha-kit}
\alpha =
\int_{-\pi}^\pi \frac{dk}{2\pi} e(1,\theta_{kk}),
\end{equation}
where $e(x,\lambda)$ is defined in Eq.~\eqref{eq:e-def}. Eq.~\eqref{eq:alpha-kit} is
obtained after replacing the first term of~\eqref{eq:kitaev-fin} in~\eqref{eq:ent-C}.
Equation~\eqref{eq:alpha-kit} has the same structure as for the tight-binding
chain [cf.~Eq.~\eqref{eq:lin-coeff}]. Specifically, $\alpha$
is the density of von Neumann entropy of the full system.
Let us now discuss the prefactor $c(\Theta)$ of the logarithmic term in Eq.~\eqref{eq:kitaev-deco}.
By substituting the second term of~\eqref{eq:kitaev-fin} in~\eqref{eq:ent-C}, we obtain
\begin{equation}
\frac{c(\Theta)}{3} =
\lim_{\delta,\epsilon \rightarrow 0^+}
\oint_\gamma \frac{d\lambda}{4\pi^3 i}
\ln(\frac{1+\epsilon+\lambda}{1+\epsilon-\lambda})
\ln^2 \! \left( \frac{\sqrt{\lambda^2}+|\Theta|}{\sqrt{\lambda^2 -
\Theta^2}} \right).
\end{equation}
As for the tight-binding chain, $\gamma$ is the same
dogbone contour of Fig.~\ref{fig:contour}.
After using~\eqref{eq:id-delta}, and proceeding as for the
tight-binding chain (cf.~section~\ref{sec:tb-ent}), we
obtain
\begin{equation}
\label{eq:kit-c}
c(\Theta) = \frac{3}{4\pi^2} \int_{-\Theta}^\Theta dx
\ln(\frac{1+x}{1-x}) \ln(\frac{\Theta + x}{\Theta - x}).
\end{equation}
Remarkably, Eq.~\eqref{eq:kit-c} is half of the result of Eq.~\eqref{eq:ctheta-int}
obtained for the tight-binding chain.
Clearly, in the zero-temperature limit one recovers the well-known
central charge $c=1/2$ of the critical Kitaev chain.
\section{Numerical results}
\label{sec:numerics}
We now provide numerical benchmarks
for the results of Sec.~\ref{sec:tb-ent} and Sec.~\ref{sec:kitaev-ent}.
In Sec.~\ref{sec:vn} we overview the behavior of the subsystem von Neumann entropy.
Our main results are contained in Sec.~\ref{sec:mi}, where we discuss the scaling behavior
of steady-state mutual information both for the tight-binding chain (Sec.~\ref{sec:tb-numerics})
and for the Kitaev chain (Sec.~\ref{sec:ki-numerics}). Our numerical
results confirm a logarithmic scaling for the mutual information,
being in perfect agreement with the predictions of the previous sections.
We should stress again that even though the mutual information is not a proper measure
of entanglement, we numerically checked that the logarithmic scaling
also occurs with the fermionic entanglement negativity~\cite{Vidal_2002}, thus
suggesting that the growth of the mutual information reflects
logarithmic entanglement growth.
\subsection{Volume-law scaling of von Neumann entropy}
\label{sec:vn}
In the presence of the external baths, the steady-state
von Neumann entropy exhibits a volume-law scaling as $\alpha\ell$,
with $\ell$ being the size of the subsystem. The prefactor $\alpha$ is given
by Eq.~\eqref{eq:lin-coeff} for the tight-binding chain and by Eq.~\eqref{eq:alpha-kit}
for the Kitaev chain. For both models, $\alpha$ is
the density of the von Neumann entropy of the full system. In the absence of baths,
the full system is in a pure state, and its von Neumann entropy is
zero ($\alpha=0$). The volume-law scaling is due to the fact that
the steady state is described by a finite-temperature-like
statistical ensemble. Here we focus on the tight-binding chain with OBC
and one thermal bath (see Fig.~\ref{fig:sb} (a)). Results for different boundary
conditions and for the Kitaev chain are qualitatively similar and will not be
discussed.
\begin{figure}[!b]
\begin{center}
\includegraphics[scale=0.7]{single_bath_entropy.pdf}
\caption{Volume-law scaling of the von Neumann
entropy in the open tight-binding chain with a single
bath on the left edge of the system (see Fig.~\ref{fig:sb}).
Subsystem $A$ is the half chain ($\ell=N/2$).
Here we choose $h=1$, $\mu_L=-1$, and $T_L=0.3,0.4,0.5$.
Straight lines denote the analytic predictions for the volume-law
scaling, in the limit $\ell\to\infty$ [cf.~Eq.~\eqref{eq:ent-expansion}].
The logarithmic correction
is not clearly visible at this scale.}
\label{fig:tb-volume}
\end{center}
\end{figure}
Figure~\ref{fig:tb-volume} reports the von Neumann entropy
$S_A$ as a function of the chain length $N$, where the subsystem $A$
is the half-chain with $\ell = N/2$ [Fig.~\ref{fig:sb}(a)].
Only the left edge is in contact with a thermal bath at $\mu_L=-1$ and temperature $T_L$.
Data in the figure correspond to different temperatures $T_L=T$.
A robust growth is visible at all temperatures, with a slope that decreases
as the temperature is lowered. This is expected, since at $T=0$ the scaling
of the von Neumann entropy is logarithmic with the interval size.
Continuous lines show the expected linear behavior as
$\alpha\ell$ in the limit $\ell\to\infty$, with a prefactor
$\alpha$ given by Eq.~\eqref{eq:lin-coeff}.
We observe a qualitative agreement with the theoretical predictions, at least
to the leading order in $\ell$. However, as also expected from Eq.~\eqref{eq:ent-expansion},
subleading logarithmic corrections are present. To reveal them it is convenient
to use the mutual information.
\subsection{Logarithmic scaling of mutual information}
\label{sec:mi}
We now discuss the scaling of the steady-state mutual information in the
presence of external baths.
The logarithmic prefactor is determined by the singular structure of the single-particle
energy dispersion, as discussed in Sec.~\ref{sec:tb-ent} and Sec.~\ref{sec:kitaev-ent}.
\subsubsection{Tight-binding chain}
\label{sec:tb-numerics}
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{single_bath_ell1.pdf}
\caption{Mutual information $I(A_1:A_2)$ between two intervals
in the open tight-binding chain with a thermal bath on the left edge.
Here we choose $\mu_L=-1$, $T_L=0.3$, and $h=1$. Different colors correspond
to different chain length $N$. The data are plotted versus $\ell/N$,
with $\ell$ being the size of $A_1$ (see Fig.~\ref{fig:sb} (a)). Notice
the symmetry under exchange of the two subsystems
$\ell \leftrightarrow N-\ell$.}
\label{fig:tb-single-no-scaling}
\end{center}
\end{figure}
For the tight-binding model of Eq.~\eqref{eq:ham-xx}, we consider the same setup
as in Sec.~\ref{sec:xx-single}, i.e, the open chain with a thermal bath on the left edge.
We fix $h=1$, $\mu_L=-1$, and $T_L=0.3$.
Our numerical data for the mutual information $I(A_1:A_2)$ between two complementary intervals
$A_1$ and $A_2=\overline{A}_1$ are plotted in Fig.~\ref{fig:tb-single-no-scaling}
versus $\ell/N$, with $\ell$ being the size of $A_1$. The three different
data sets correspond to different values of $N$. At each fixed $N$, $I(A_1:A_2)$
increases upon increasing $\ell$ up to $\ell\sim N/2$, after which it starts
decreasing. The behavior at intermediate $1\ll\ell\ll N$ is consistent with
a logarithmic increase, as predicted in Eq.~\eqref{eq:ent-expansion-open},
which should hold in the limit $N\to\infty$ and then $\ell\to\infty$ (with this order of
limits).
Looking now at the definition of the mutual information [cf.~Eq.~\eqref{eq:mi-intro}],
it is clear that, when constructing $I(A_1:A_2)$, the volume-law terms in
the entropies cancel out.
To derive the prefactor of the logarithmic scaling, we can use Eq.~\eqref{eq:ent-expansion-open}
for each term appearing in~\eqref{eq:mi-intro}.
Notice that no logarithmic contribution is expected from $S_{A_1\cup A_2}$, since the entropy of the full
system for large $N$ is exactly $\alpha N$, with $\alpha$ given by~\eqref{eq:lin-coeff}.
Let us also stress that, in principle, we are not allowed to use Eq.~\eqref{eq:ent-expansion-open}
for $S_{A_2}$ because the size $N-\ell$ of $A_2$ is comparable with $N$.
To proceed, we should then conjecture a generalization for an interval $A$ of generic size,
embedded in a finite-size chain. Following the standard strategy for critical systems
described by CFTs, we write~\cite{calabrese2009entanglement}
\begin{equation}
\label{eq:A1-1}
S_{A}=\alpha\ell+\frac{c(\Theta)}{6}\ln\left[\frac{N}{\pi}
\sin\left(\frac{\pi\ell}{N}\right)\right]+{\mathcal O}(1).
\end{equation}
Notice that the prefactor of the volume-law term is the same as before,
while in the logarithmic term of Eq.~\eqref{eq:ent-expansion-open}
we replaced
\begin{equation}
\label{eq:chord}
\ell \to X_\ell,\quad\mathrm{with}\,\,X_\ell \coloneqq \frac{N}{\pi}\sin\left(\frac{\pi\ell}{N}\right),
\end{equation}
where $X_\ell$ is the so-called chord length. Eq.~\eqref{eq:A1-1} holds in the thermodynamic
limit $\ell,N\to\infty$.
For systems with boundaries, as is the case here, the chord length differs from~\eqref{eq:chord}
by an overall factor $2$, which only affects the ${\mathcal O}(1)$ term, and can therefore
be neglected.
We conclude that
\begin{equation}
\label{eq:theory-xx}
I(A_1:A_2)= \frac{c(\Theta)}{3}\ln\left[\frac{N}{\pi}
\sin\left(\frac{\pi\ell}{N}\right)\right]+{\mathcal O}(1).
\end{equation}
The factor $1/3$ rather than $1/6$ is due to the fact that both subsystems $A_1$ and
$A_2$ contribute with a logarithmic term.
Importantly, Eq.~\eqref{eq:theory-xx} implies that for large $\ell,N$ the
data for the mutual information should collapse on the same curve, when plotted
as a function of $X_\ell$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{single_bath_ell2.pdf}
\caption{Mutual information $I(A_1:A_2)$ between two complementary
intervals [see Fig.~\ref{fig:sb} (a)] in the open tight-binding chain
with a thermal bath on the left edge. On the $x$-axis
$N\sin(\pi\ell/N)/\pi \equiv X_\ell$ is the chord length.
Data are for fixed $\mu_L = 1$, $h=1$, and for $T_L=0.3,0.5$,
while the various colors denote different $N$.
The lines are fits to $c(\Theta)\ln(X_\ell)/3 + b$,
$c(\Theta)$ being the effective central charge [cf.~Eq.~\eqref{eq:ctheta-res}]
and $b$ a fitting constant parameter.}
\label{fig:tb-single-scaling}
\end{center}
\end{figure}
The validity of Eq.~\eqref{eq:theory-xx} is investigated in Fig.~\ref{fig:tb-single-scaling}
for the tight-binding chain with one thermal bath on the left edge. We consider two
different temperatures $T_L=0.3,0.5$ at fixed $\mu_L=-1$. The mutual information $I(A_1:A_2)$ is
plotted versus $X_\ell$ (notice the logarithmic scale on the $x$-axis)
for several values of $N=200,600,1000$. For both temperatures, the data
exhibit collapse. The quality of the collapse improves upon increasing $N$, as expected.
Continuous lines are fits to Eq.~\eqref{eq:theory-xx}, where $c(\Theta)$ is kept fixed and
given by Eq.~\eqref{eq:ctheta-res}, while the additive ${\mathcal O}(1)$ term being the only fitting
parameter. For both temperatures. the agreement between the data and the fits
is very satisfactory.
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{single_bath_fit.pdf}
\caption{Scaling of the half-chain mutual information in
the open tight-binding chain with a thermal bath of the left edge
of the chain. Here we fix $\mu_L=-1$, $h=1$. Different
colors are for different values of $T_L$. Continuous lines
are fits to $a \ln(N)+b$, with $a,b$ fitting parameters.}
\label{fig:tb-fits}
\end{center}
\end{figure}
Eq.~\eqref{eq:theory-xx} also implies that, for large $N$, the mutual information between
the two halves of the chain scales logarithmically as $c(\Theta)\ln(N)/3$. This is shown
in Fig.~\ref{fig:tb-fits}, for the same setup as in Fig.~\ref{fig:tb-single-scaling}.
The various data sets correspond to different temperatures of the external bath.
The logarithmic increase is clearly visible (notice the semilog scale),
although oscillating corrections are present. The continuous lines are fits to
\begin{equation}
\label{eq:mi-fit}
I(A_1:A_2) = a\ln(N)+b,
\end{equation}
with $a,b$ fitting parameters.
Further checks of our results are provided in Fig.~\ref{fig:tb-check-single} where,
for each temperature, we numerically extract $c(\Theta)$ by fitting the mutual
information to Eq.~\eqref{eq:mi-fit}.
Symbols are the results of the fits, which are obtained as in Fig.~\ref{fig:tb-fits}
fitting the data with $N > 2^6$. At low temperature, one finds $c(\Theta)\to1$,
whereas $c(\Theta)$ vanishes in the high-temperature limit. The continuous line is the analytic
prediction in the limit $N\to\infty$, given by Eq.~\eqref{eq:ctheta-res}. The agreement
with the numerics is excellent.
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{single_bath_temp.pdf}
\caption{The effective central charge $c(\Theta)$ versus the
temperature $T_L$, as obtained from fits of numerical data
as those in Fig.~\ref{fig:tb-fits}.
Parameters are the same as in Fig.~\ref{fig:tb-fits}.}
\label{fig:tb-check-single}
\end{center}
\end{figure}
Finally we discuss a two-bath geometry, where the edges of the chain are connected
to two different thermal baths. Here we fix $h=1$, $\mu_L=0$, $T_L=T_R=1$,
and vary $\mu_R$. In Fig.~\ref{fig:mi-tb-check-double} we plot the numerically extracted
$c(\Theta)$ versus $\mu_R$. As for Fig.~\ref{fig:tb-check-single},
the continuous line denotes the theoretical result in the limit $N\to\infty$,
which is in perfect agreement with the numerics.
\begin{figure}
\begin{center}
\includegraphics[scale=0.7]{two_baths.pdf}
\caption{Logarithmic scaling of the mutual information $I(A_1:A_2)$ in the
tight-binding chain with OBC coupled to two
thermal baths at the edges. We plot the effective central charge
$c(\Theta)$ of the logarithmic growth versus the chemical potential
$\mu_R$ of the right bath. Here we fix $h=1$, $\mu_L=0$, and
$T_L=T_R=1$. Numerical results for $c(\Theta)$ are obtained by performing
a finite-size scaling analysis for the half-chain mutual information.}
\label{fig:mi-tb-check-double}
\end{center}
\end{figure}
\subsubsection{Kitaev chain}
\label{sec:ki-numerics}
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.7]{twobath_collapse_Kitaev.pdf}
\caption{Mutual information $I(A_1:A_2)$ between two intervals in the Kitaev
chain with PBC and two external baths (see Fig.~\ref{fig:sb}) versus the
chord length $X_{2\ell}=N\sin(2\pi\ell/N)/\pi$.
Here we choose $T_L=T_R=1$, $\mu_L=0$ and $\mu_R=2,4$. The varioua symbols correspond
to different chain sizes $N$. Continuous lines are fits to
$I(A_1:A_2)=c(\Theta)\ln(X_{2\ell})/6 + b$, with
$c(\Theta)$ the effective central charge, and $b$ a fitting constant.}
\label{fig:kitaev}
\end{center}
\end{figure}
We conclude by discussing the steady-state mutual information in the Kitaev chain.
Here we consider a PBC geometry, as depicted in Fig.~\ref{fig:sb} (b).
Two sites at mutual distance $N/2$ are put in contact with two external
baths at temperatures $T_{R/L}$ and with chemical potential $\mu_{R/L}$.
We choose $T_L=T_R=1$, $\mu_L=0$, $\mu_R=2,4$, and fix $h=2$. We consider the mutual information
$I(A_1:A_2)$ between two intervals of size $\ell$ and $N/2-\ell$ placed between the
baths (see Fig.~\ref{fig:sb}). First, we should observe that
in constructing the mutual information~\eqref{eq:mi-intro} all the entropies (i.e.,
$S_{A_1}$, $S_{A_2}$, and $S_{A_1\cup A_2}$) contain a subleading logarithmic term.
This happens because $S_{A_1\cup A_2}$ is not the full system. As for the tight-binding chain
the volume-law terms cancel out.
The final result is
\begin{equation}
\label{eq:theory-kit}
I(A_1:A_2)=\frac{c'(\Theta)}{3}\ln(X_{2\ell})
+{\mathcal O}(1),
\end{equation}
where $X_{2\ell}$ is the chord length in Eq.~\eqref{eq:chord} (notice the factor $2$), and
\begin{equation}
\label{eq:cp}
c'(\Theta)=\frac{c(\Theta)}{2}.
\end{equation}
Here $c(\Theta)$ is given in Eq.~\eqref{eq:ctheta-res} and $\Theta$ depends on the
parameters of baths (cf.~Eq.~\eqref{eq:theta-is-3}). The factor $1/2$ in Eq.~\eqref{eq:cp}
reflects the fact that the prefactor of the logarithmic growth for the Kitaev chain is half
that of the tight-binding chain, as derived in Sec.~\ref{sec:kitaev-ent}. To derive
Eq.~\eqref{eq:theory-kit} we used the fact that, for all the intervals $A_1$, $A_2$, and $A_1\cup A_2$,
\begin{equation}
\label{eq:Sw}
S_{W}\xrightarrow[]{\ell,N\to\infty}\frac{c'(\Theta)}{3}
\ln(X_\ell), \quad W=A_{1(2)},A_1\cup A_2,
\end{equation}
where $X_\ell$ is the chord length and $c'(\Theta)$ is given by Eq.~\eqref{eq:cp}.
After substituting~\eqref{eq:Sw} in the definition of the
mutual information~\eqref{eq:mi-intro}, we obtain~\eqref{eq:theory-kit}.
We point out that Eq.~\eqref{eq:theory-kit} holds only for the geometry
in Fig.~\ref{fig:sb} (b), although it could be easily generalized to
more general settings.
The validity of Eq.~\eqref{eq:theory-kit} is numerically verified in Fig.~\ref{fig:kitaev},
where we plot $I(A_1:A_2)$ versus $X_{2\ell}$. For both values of $\mu_R$,
the data exhibit collapse at large $\ell, N$. Continuous lines are fits to Eq.~\eqref{eq:theory-kit},
the only fitting parameter being the ${\mathcal O}(1)$ constant, whereas $c'(\Theta)$ is given
by Eq.~\eqref{eq:cp}. The agreement between the analytic prediction in the scaling limit
$\ell,N\to\infty$ and the numerics is nearly perfect already for relatively small chains with
$X_{2\ell}\sim 10$.
\section{Conclusions}
\label{sec:conclusions}
We investigated the quantum-information spreading in the tight-binding chain and the
Kitaev chain in the presence of external thermal baths coupled to individual sites of the
chains. To this purpose, we employed a self-consistent \emph{nonlocal} Lindblad master equation approach,
where the Lindblad operators modeling the baths are written in terms of the Bogoliubov modes
that diagonalize the isolated system, implying that they are, in principle, nonlocal
in real space~\cite{dabbruzzo2021self}.
The statistical ensemble describing the steady state is written in terms of a convex
combination of the Fermi-Dirac distributions of the baths.
We showed that the steady-state von Neumann entropy of a subsystem exhibits a volume-law
scaling with the subsystem size, reflecting that the system is not in a pure state. The mutual
information exhibits an area-law scaling for generic values of the system parameters. Interestingly,
we observe logarithmic violations of the area law in the presence of ground-state criticality.
This behavior reflects the singularity of the single-particle
energy dispersion of the models, which is present at all energies.
We analytically derived the prefactor of the logarithmic growth of the mutual information, which
depends on the system and bath parameters, such as the temperature and the chemical potential.
Let us now mention some promising directions for future work.
First of all, here we only analyzed the steady-state value of the mutual information:
it would be tempting to study the full-time dynamics, in order to
establish how the logarithmic scaling builds up during the evolution of the system.
A natural conjecture is that the same effective central charge governs a logarithmic increase in time, as
in Ref.~\cite{Kormos2017}.
Our analysis may be also extended to genuine quantum entanglement measures
for mixed states, such as the fermionic logarithmic negativity.
Finally, it would be important to check the validity of our results by
comparing them with \emph{ab initio} numerical simulations, or with results
obtained using different master equations. A crucial question to address
is whether the logarithmic scaling of the mutual information would survive
in interacting integrable systems or even in nonintegrable ones.
|
\section{Transverse momenta in different frames and kinematic corrections}
\label{apx:rotations}
Here we give the relations among the different transverse momenta, presented in Section~\ref{sec:2_h_prod}, in the hadron frame. Even if already discussed in the literature, it is worth presenting them once again, taking into account also mass effects.
More precisely, we work in a frame where the two final-state hadrons are exactly back-to-back along the $\pm \hat{\bm{z}}$ direction, the fragmenting quarks have a transverse momentum with respect to the direction of their parent hadrons and the photon has a transverse momentum as well. We label the momenta as follows:
\begin{itemize}
\item $k$ is the four-momentum of the quark fragmenting into the hadron $h_1$, with four-momentum $P_1$ and moving along $-\hat{\bm{z}}$;
\item $p$ is the four-momentum of the anti-quark fragmenting into the hadron $h_2$, with four-momentum $P_2$ and moving along $+\hat{\bm{z}}$;
%
\item $\bm{k}_T$ and $\bm{p}_T$ are respectively the transverse momenta of the quark and the anti-quark with respect to their associated hadron momenta, while $\bm{k}_{\perp}$ and $\bm{p}_{\perp}$ are respectively the transverse momenta of the two hadrons with respect to their parent quark directions of motion;
\item $q$ is the four-momentum of the virtual photon and $\bm{q}_T$ its transverse component.
\end{itemize}
By adopting standard light-cone coordinates ($a=(a^+,a^-, \bm{a}_T)$),
we define the quarks and photon momenta as follows~\cite{Boer:1997mf}:
\begin{align}
k &=\Bigg( \frac{k^2_T}{2 k^-},k^-,\bm{k}_T \Bigg) \quad \text{where} \quad k_T=\abs{\bm{k}_T}\\
%
p &= \Bigg( p^+,\frac{p^2_T}{2 p^+}, \bm{p}_T \Bigg) \quad \text{where} \quad p_T=\abs{\bm{p}_T}\\
%
q &= \Bigg( \frac{\widetilde{Q}}{\sqrt{2}},\frac{\widetilde{Q}}{\sqrt{2}},\bm{q}_T \Bigg) \,,
\end{align}
with $\widetilde{Q}^2 = Q^2 + q^2_T$, where $q_T=\abs{\bm{q}_T}$. Since $ k +p = q$, we have directly
\begin{align}
k^- = \frac{{Q}}{\sqrt{2}} +\mathcal{O}{(q^2_T/Q^2)}&; \qquad k^+ = \frac{k^2_T}{\sqrt{2}Q} +\mathcal{O}{(q^2_T/Q^2)} \\
%
k^0 = \frac{Q}{2} +\mathcal{O}{(k^2_T/Q^2)}&; \qquad k^3 = - \frac{Q}{2} +\mathcal{O}{(k^2_T/Q^2)}\\
%
p^+ = \frac{{Q}}{\sqrt{2}} +\mathcal{O}{(q^2_T/Q^2)}&; \qquad p^- = \frac{p^2_T}{\sqrt{2}Q} +\mathcal{O}{(q^2_T/Q^2)}\,,
\end{align}
with $\bm{k}_T + \bm{p}_T = \bm{q}_T$.
Concerning the hadron momenta, we have:
\begin{equation}
\begin{split}
P_1 &= \Bigg(\frac{M^2_1}{z_1 {Q}\sqrt{2}}, \frac{z_1 {Q}}{\sqrt{2}},\bm{0}_{T} \Bigg) \\
%
P_2 &= \Bigg( \frac{z_2 {Q}}{\sqrt{2}} ,\frac{M^2_2}{z_2 {Q}\sqrt{2}} ,\bm{0}_{T} \Bigg) \,, \\
\end{split}
\end{equation}
where we have used the light-cone momentum fractions:
\begin{equation}
z_1 = \frac{P_1^-}{k^-}; \qquad z_2 = \frac{P^+_2}{p^+}\,.
\end{equation}
\subsection{$k_T$ vs.~$k_{\perp}$ }
We now derive the relation between the transverse momenta $\bm{k}_T$ and $\bm{k}_{\perp}$.
Firstly, without loss of generality, we consider the hadron $h_1$ and the fragmenting quark momenta sitting on the $\widehat{xz}$ plane, namely
\begin{equation}
P_1=(P_1^0,0,0,P_1^3)\qquad k= (k^0, k_{T}, 0, k^3)\,,
\end{equation}
with the quark transverse momentum, $\bm{k}_T$, along the positive $\hat{\bm{x}}$-axis. Then, we move to a frame where the quark has zero transverse momentum component by employing the following rotation:
\begin{equation}
R_{y}(\xi, \lambda) = \begin{pmatrix}
1 & 0 & 0 & 0 \\
0 & \xi & 0 & -\lambda \\
0 & 0 & 1 & 0 \\
0 & \lambda & 0 & \xi \\
\end{pmatrix} \,,
\end{equation}
where $\xi = - k^3/k^0 \,$ and $\lambda =- k_T/k^0 $, with $\xi^2 + \lambda^2 =1$. The transformed quark four-momentum $k'$ is then given as:
\begin{equation}
k'=R_{y}(\xi, \lambda) k = (k^0,0,0,k'^3)\,.
\end{equation}
By applying the above rotation to the hadron four-momentum $P_1$, we get
\begin{equation}
P'_{1}= R_{y}(\xi, \lambda)\,P_{1} = (P_1^0, -\lambda{P}_1^3 ,0,{P'}_1^3)\,,
\end{equation}
with
\begin{align}
P_1^0 &= \frac{z_1 Q}{2} \Bigg( 1 + \frac{M^2_1}{ z^2_1 {Q}^2}\Bigg) \\
{P'}_1^3 &=- \xi \frac{z_1 Q}{2}\Bigg( 1 - \frac{M^2_1}{ z^2_1 {Q}^2}\Bigg)\,\\
%
-\lambda {P}_1^3 &=-k_T z_1 \Bigg( 1 - \frac{M^2_1}{ z^2_1 {Q}^2}\Bigg)\,.
\end{align}
The hadron transverse momentum with respect to the quark direction, usually called $\bm{k}_{\perp}$, is now along the negative $\hat{\bm{x}}$-axis.
Therefore, from the above relations, we can see that $\bm{k}_{T}$ and $\bm{k}_{\perp}$ are anti-parallel (neglecting $k_T^2/Q^2$ corrections) and, using Eq.~(\ref{eq:long_fract}), that $k_{\perp} = k_T\, z_{p_1}$ (this indeed can be derived also as an exact relation). This eventually leads to:
\begin{align}
\bm{k}_{\perp} &= -\bm{k}_T z_{p_1} \,,
\end{align}
that, for a massless hadron, reduces to the usual $\bm{k}_T = -\bm{k}_{\perp}/z_{1}$ relation. The same is valid for the second hadron, that is $\bm{p}_{\perp} = -\bm{p}_T z_{p_2}$.
\subsection{$P_{1T}$ vs.~$q_{T}$ }
In order to find the relation between $\bm{P}_{1T}$, the transverse momentum of $h_1$ with respect to $h_2$ in the hadron frame, and $\bm{q}_T$, the transverse momentum of the photon in the new frame, where the two hadrons are back-to-back, we can use more directly the tensor $g^{\mu \nu}_{\perp}$ \cite{Boer:1997mf}:
\begin{equation}
g^{\mu \nu}_{\perp} = g^{\mu \nu} -\hat{t}^{\mu} \hat{t}^{\nu} + \hat{z}^{\mu} \hat{z}^{\nu} \quad\quad {\rm with}\quad\quad
%
\hat{t}^{\mu} = \frac{q^{\mu}}{Q} \quad\quad
%
\hat{z}^{\mu} = 2 \frac{P^{\mu}_2}{z_{h_2}Q} - \hat{t}^{\mu} \,.
\end{equation}
By contracting the above tensor with the four-momentum of hadron $h_1$, we find
\begin{equation}
P^{\mu}_{1T} = g^{\mu \nu}_{\perp} P_{1 \nu} = P^{\mu}_1 + P^{\mu}_2 \bigg(2 \frac{z_1}{z_{h_2}} - \frac{z_{h_1}}{z_{h_2}} \bigg) - z_1 q^{\mu} \,,
\end{equation}
where we have neglected the mass of the second hadron, $M_{2} =0$. By explicit calculation one can then see that all components but the transverse one are zero, that is
\begin{equation}
\bm{P}_{1T} = - z_1 \bm{q}_T\,,
\end{equation}
where $z_1$ is the light-cone momentum fraction.
\section{Convolutions and Fourier transforms}
\label{apx2:conv_FT}
In order to exploit the CSS evolution equations, in Section~\ref{sec:2_h_prod} we showed how the convolutions, in $\bm{k}_T$-space, can be written in the conjugate $\bm{b}_T$-space through Fourier transforms of the TMD-FFs. We give here some details about this procedure. The first convolution we consider is $F_{UU}$, defined according to Eq.~(\ref{eq:conv_general}),
\begin{equation}
\begin{split}
\mathcal{F}[\omega D \bar{D}] = \sum_q e^2_q \int d^2\bm{k}_T d^2\bm{p}_T \, \delta^{(2)}(\bm{k}_T + \bm{p}_T - \bm{q}_T) \,\omega(\bm{k}_T, \bm{p}_T) D(z_1,\bm{k}_{\bot}) \bar{D}(z_2,\bm{p}_{\bot}) \,,
\end{split}
\end{equation}
as follows:
\begin{equation}
\begin{split}
F_{UU} =& \mathcal{F}[D_1 \bar{D}_1]\,.
%
\end{split}
\end{equation}
It is trivial to convert this convolution in $\bm{b}_T$-space by employing the Fourier transform definition of the TMD unpolarized FF:
\begin{equation}
\widetilde{D}_1(z,b_T) = \int d^2\bm{k}_T \, e^{i\bm{b}_T \cdot \bm{k}_T } D_1(z,{k}_{\perp})
= 2 \pi \int dk_T \, k_T J_0(b_T k_T) D_1(z,k_{\perp})\,,
\end{equation}
where we have used the integral representation of $J_0$, the Bessel function of the first kind
\begin{equation}
\int^{2\pi}_0 d\theta \, e^{i b_T k_T \cos\theta} = 2\pi J_0(b_T k_T) \,.
\label{bessel_0}
\end{equation}
With the above relations, the $F_{UU}$ convolution in $\bm{b}_T $-space can be written as:
\begin{equation}
\begin{split}
F_{UU} =& \mathcal{F}[D_1 \bar{D}_1]
= \sum_q e^2_q \int \frac{d^2 \bm{b}_T}{(2 \pi)^2} \, e^{-i\bm{b}_T \cdot \bm{q}_T} \widetilde{D}_1(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T) \\
%
=& \sum_q e^2_q \int \frac{d b_T}{(2 \pi)} \, b_T J_0(b_T\, q_T) \widetilde{D}_1(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T)
= \mathcal{B}_0 \Big[\widetilde{D}_1 \widetilde{\bar{D}}_1\Big] \,.
\end{split}
\end{equation}
The second convolution we consider is $F^{\sin(\phi_1 - \phi_{S_1})}_{TU}$, defined as follows:
\begin{equation}
\begin{split}
F^{\sin(\phi_1 - \phi_{S_1})}_{TU} =& \mathcal{F}\bigg[\frac{\hat{\bm{h}}\cdot \bm{k}_T}{M_{1}}D^{\perp}_{1T} \bar{D}_1\bigg]\,. \\
\end{split}
\end{equation}
To work out this convolution we need the Fourier transform of the unpolarized fragmentation function, presented above, and the one of the polarizing fragmentation function multiplied by $k^i_T$, the $i$-th component of the quark transverse momentum with respect to the hadron direction:
\begin{eqnarray}
\int d^2\bm{k}_T \, \frac{k^i_T}{M_{h}} e^{i\bm{b}_T \cdot \bm{k}_T} D^{\perp}_{1T}(z,k_{\perp})
&=& \frac{-i}{M_{h}} \pr{}{b^i_T} \int d^2\bm{k}_T \, e^{i\bm{b}_T \cdot \bm{k}_T} D^{\perp}_{1T}(z,k_{\perp}) \\
&=& -\frac{i}{M_{h}} \pr{}{b^i_T} \widetilde{D}^{\perp}_{1T}(z,b_T) \\
%
%
&=& - 2\frac{i}{M_{h}} b^i_T \pr{}{b^2_T} \widetilde{D}^{\perp}_{1T}(z,b_T) \\
%
&=& i b^i_T M_{h} \widetilde{D}^{\perp (1)}_{1T}(z,b_T)\,,
\end{eqnarray}
where we have used the definition of the Fourier transform of the polarizing FF
\begin{equation}
\widetilde{D}^{\perp}_{1T}(z,b_T) = \int d^2\bm{k}_T \, e^{i\bm{b}_T \cdot \bm{k}_T } D^{\perp}_{1T}(z,\bm{k}_{\perp}) \,.
\label{eq:pFF_bt}
\end{equation}
In the very last step we have also introduced
the first moment of the polarizing FF in $\bm{b}_T $-space:
\begin{equation}
\widetilde{D}^{\perp (1)}_{1T}(z,b_T) = \bigg(-\frac{2}{M^2_{h}}\pr{}{b^2_T} \widetilde{D}^{\perp}_{1T}(z,b_T) \bigg)\; .
\label{eq:apx_first_mom_bt}
\end{equation}
This term can be related to the first moment in $\bm{k}_T $-space, defined as:
\begin{equation}
{D}^{\perp (1)}_{1T}(z) = \int d^2 \bm{k}_{\perp} \, \bigg(\frac{\bm{k}^2_{\perp}}{2 z^2 M^2_{h}} \bigg) {D}^{\perp}_{1T}(z,{k}_{\perp})\,,
\label{eq:first_mom_apx}
\end{equation}
by using the relation:
\begin{equation}
\lim_{b_T \to 0} \widetilde{D}^{\perp (1)}_{1T}(z,b_T) =\frac{1}{z^2} {D}^{\perp (1)}_{1T}(z)\; .
\label{eq:first_mom_kt_bt}
\end{equation}
We have indeed that:
\begin{equation}
\begin{split}
\widetilde{D}^{\perp (1)}_{1T}(z,b_T) &= -\frac{2}{M^2_{h}}\pr{}{b^2_T} \widetilde{D}^{\perp}_{1T}(z,b_T) \\
%
&= -\frac{2 \pi}{b_T M^2_{h}}\pr{}{b_T}\int \frac{dk_{\perp} \, k_{\perp}}{z^2} J_0\bigg(\frac{b_T k_{\perp}}{z}\bigg) D^{\perp }_{1T}(z,k_{\perp }) \\
%
&=\frac{2 \pi}{M^2_{h} z^3} \int dk_{\perp} \frac{k^2_{\perp}}{b_T} J_1\bigg(\frac{b_T k_{\perp}}{z}\bigg)D^{\perp }_{1T}(z,k_{\perp })\,,
\end{split}
\end{equation}
where we have used
\begin{equation}
\pr{}{b_T} J_0(a b_T) = - a J_1(ab_T)\,.
\end{equation}
Then by employing the additional relation
\begin{equation}
\lim_{b_T\to0} \frac{J_1(a b_T)}{b_T} = \frac{a}{2}\,,
\end{equation}
we are able to proof Eq.~(\ref{eq:first_mom_kt_bt}):
\begin{equation}
\begin{split}
\lim_{b_T\to0} \widetilde{D}^{\perp (1)}_{1T}(z,b_T) = \frac{1}{z^2}\int d^2 \bm{k}_{\perp} \, \bigg(\frac{\bm{k}^2_{\perp}}{2 z^2 M^2_h} \bigg) {D}^{\perp}_{1T}(z,{k}_{\perp})\,.
\end{split}
\end{equation}
\noindent Finally, with the above results we can write the $F^{\sin(\phi_1 - \phi_{S_1})}_{TU}$ convolution in $\bm{b}_T $-space:
\begin{equation}
\begin{split}
& F^{\sin(\phi_1 - \phi_{S_1})}_{TU} = \mathcal{F}\bigg[\frac{\hat{h}\cdot \bm{k}_T}{M_{1}}D^{\perp}_{1T} \bar{D}_1\bigg] \\
%
=& \sum_q e^2_q \int d^2\bm{k}_T d^2\bm{p}_T \, \delta^{(2)}(\bm{k}_T + \bm{p}_T - \bm{q}_T) \,\frac{\hat{h}\cdot \bm{k}_T}{M_{1}} D^{\perp}_{1T}(z_1,\bm{k}_T) \bar{D}_1(z_2,\bm{p}_T) \\
%
=& M_{1} \sum_q e^2_q \int \frac{d^2 \bm{b}_T}{(2 \pi)^2} \, e^{-i\bm{b}_T \cdot \bm{q}_T} (i \hat{h}\cdot \bm{b}_T) \widetilde{D}^{\perp (1)}_{1T}(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T) \\
%
=& M_{1} \sum_q e^2_q \int \frac{d b_T}{2 \pi} \,b^2_T J_1(q_T\, b_T) \widetilde{D}^{\perp (1)}_{1T}(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T) \\
%
=& M_{1} \mathcal{B}_1 \Big[\widetilde{D}^{\perp (1)}_{1T} \widetilde{\bar{D}}_1\Big]\,,
\end{split}
\end{equation}
where we have used the integral definition of the $J_1$ Bessel function:
\begin{equation}
\int^{2\pi}_0 d\theta \, e^{i b_T k_T \cos\theta}\,\cos\theta = (2\pi i) J_1(b_T k_T) \,.
\label{bessel_1}
\end{equation}
\subsection{Gaussian model}
As a first example of parametrization for the transverse momentum dependence of the FFs we present a simple Gaussian model, deriving its Fourier transform and first moment. We use the following Gaussian model for a generic TMD fragmentation function:
\begin{equation}
{D}(z,k_{\perp}) = {D}(z,0) \,\frac{e^{-k^2_{\perp }/\langle k_\perp^2 \rangle_D }}{\pi \langle k_\perp^2 \rangle} .
\end{equation}
where $\langle k_\perp^2 \rangle$ is the Gaussian width of the unpolarized TMD-FF.
According to Eq.~(\ref{eq:first_mom_apx}), its first moment is:
\begin{equation}
D^{(1)}(z) = \int d^2 \bm{k}_{\perp} \, \bigg(\frac{\bm{k}^2_{\perp}}{2 z^2 M^2_h} \bigg) {D}(z,{k}_{\perp})
= {D}(z,0)\, \frac{\langle k_\perp^2 \rangle^2_D}{2 z^2 M^2_h \langle k_\perp^2 \rangle}
\end{equation}
and the $\bm{b}_T$-space fragmentation function is
\begin{eqnarray}
\widetilde{D}(z,b_T) & = & \int d^2\bm{k}_T \, e^{i\bm{b}_T \cdot \bm{k}_T } D(z,k_{\perp}) = 2 \pi \int dk_T \, k_T J_0(b_T k_T) D(z,k_{\perp})\nonumber \\
&=& 2 \pi \int \frac{dk_{\perp } \, k_{\perp}}{z^2} J_0\bigg(\frac{b_T k_{\perp}}{z}\bigg) D(z,k_{\perp})
= \frac{{D}(z,0) }{z^2} \frac{\langle k_\perp^2 \rangle_D}{\langle k_\perp^2 \rangle} e^{-b^2_T \langle k_\perp^2 \rangle_D/(4 z^2)}\,.\nonumber\\
\end{eqnarray}
We notice that, when $\langle k_\perp^2 \rangle_D = \langle k_\perp^2 \rangle$, as in the case of the unpolarized FF,
\begin{equation}
\lim_{b_T\to 0}\widetilde{D}_1(z,b_T) = \frac{1}{z^2}{D}_1(z,0) \,,
\end{equation}
where ${D}_1 $ coincides with $d_{j/h}$ in the OPE in Eq.~(\ref{OPE_unp}).
Employing the first-moment definition in $\bm{b}_T$-space, Eq.~(\ref{eq:apx_first_mom_bt}), we have:
\begin{equation}
\widetilde{D}^{ (1)}(z,b_T) = \bigg(-\frac{2}{M^2_{h}}\pr{}{b^2_T} \widetilde{D}(z,b_T) \bigg)
%
= {D}(z,0) \frac{1}{2 z^4 M^2_h} \frac{\langle k_\perp^2 \rangle^2_D}{\langle k_\perp^2 \rangle} \, e^{-b^2_T \langle k_\perp^2 \rangle_D/(4 z^2)} \,.
\end{equation}
Then, by using Eq.~(\ref{eq:first_mom_kt_bt}), we find:
\begin{equation}
\lim_{b_T\to 0}\widetilde{D}^{ (1)}(z,b_T) = \frac{1}{z^2} \bigg[ \frac{1}{2 M^2_h z^2} \frac{\langle k_\perp^2 \rangle^2_D}{\langle k_\perp^2 \rangle} {D}(z,0) \bigg]
= \frac{1}{z^2} D^{(1)}(z) \, .
\end{equation}
\subsection{Power-Law model}
The second model used to parametrize the transverse momentum dependence of FFs, of which we calculate here its Fourier transform and first moment, is the the Power-Law model. This is defined as follows:
\begin{equation}
D(z,k_{\perp}) = D(z,0) \frac{\Gamma(p)}{\pi \Gamma(p-1)}\frac{m^{2(p-1)}}{(k^2_{\perp} + m^2)^p} \,.
\label{eq:powerlaw_apx_p}
\end{equation}
Its Fourier transform is:
\begin{equation}
\widetilde{D}(z,b_T) = D(z,0)\frac{2^{2-p}}{\Gamma(p-1)}\,(b_T m/z)^{p-1}{K}_{p-1}(b_T m/z)\,.
\end{equation}
The integrated first moment and the first moment in $\bm{b}_T$ space are:
\begin{equation}
D^{(1)}(z) = D(z,0) \frac{1}{M^2_h z^2} \frac{m^2}{2(p-2)}
\end{equation}
\begin{equation}
\widetilde{D}^{(1)}(z,b_T) =D(z,0) \frac{2^{2-p}}{\Gamma(p-1)} \frac{ m^2}{M^2_h z^2} (b_T m/z)^{p-2}{K}_{p-2}(b_T m/z)\,.
\end{equation}
\section{Introduction}
\label{sec:introduction}
One of the outstanding challenges in QCD theory and phenomenology has been to describe and predict the observed large transverse single spin asymmetries (TSSAs) in deep inelastic scattering (DIS) processes using factorization theorems derived within perturbative QCD.
During the mid 70s, significant TSSAs in inclusive pion production in proton-proton collisions, at center of mass (cm) energy of few GeV, were observed at the Argonne Laboratory synchrotron~\cite{Dick:1975ty,Klem:1976ui,Dragoset:1978gg},
while during the same period at Fermilab, $\Lambda$-hyperons produced in unpolarized proton-nuclear collisions at $\sqrt{s}\approx$ 24 GeV and moderate transverse momentum, $P_T$ (below $\sim 1.5$~GeV), displayed large transverse polarization with respect to the production plane~\cite{Bunce:1976yb,Schachinger:1978qs,Heller:1978ty}. Large TSSAs (approx 30-40\% ) continued to be observed in the 90s at Fermilab~\cite{Adams:1991cs,Adams:1991ru,Adams:1991rx,Bravar:1996ki} in pseudo-scalar meson production with center of mass energies of $\sqrt{s}\approx 20$~GeV.
These results were confirmed by the STAR, PHENIX and BRAHMS Collaborations at the
Relativistic Heavy Ion Collider (RHIC), at cm energies up to $\sqrt{s}\approx 500$~GeV covering a wide range in Feynman $x_F=2 P_L/\sqrt{s}$ (where $P_L$ is the longitudinal momentum of the final hadron) and $P_T$~\cite{Adams:2003fx,Adams:2006uz,PHENIX:2003qdw,Adler:2005in,:2008mi,BRAHMS:2007tyt}.
More striking, is the transverse polarization data for $\Lambda$-hyperon production from {\em unpolarized} hadron collisions. Along with followup experiments from
Fermi-lab~\cite{Lundberg:1989hw,Ramberg:1994tk},
fixed target measurements of this reaction were reported by the NA48 Collaboration~\cite{Fanti:1998px} and the HERA-B Collaboration~\cite{Abt:2006da}. At CERN a $\Lambda$ polarization of approximately 35\% was also measured in $pp$ collisions at the ISR at cm energies of $\sqrt{s}=52$ and 63~GeV~\cite{Erhan:1979xm}.
From the theory side we know that large transverse polarization effects cannot be explained within the collinear QCD factorization at leading twist which, at large enough $P_T$, predicts negligible values~\cite{Kane:1978nd}. Presumably then hyperon polarization, observed in unpolarized collisions, necessarily has to originate from nonperturbative effects during the hadronization process, as they are produced from parity conserving strong interaction and in turn undergo self-analyzing weak decay. For this reason a study of the $\Lambda$ polarization enables us to obtain important information on this nonperturbative mechanism.
Within the context of QCD factorization theorems the hadronization of partons is described in terms of nonperturbative matrix elements of QCD operators which can be fitted to experimental data. This is a challenging endeavour on the basis of data taken from nucleon-nuclear scattering experiments, since these processes are mediated solely by the strong force where an analytical description is complicated due to competing effects that enter QCD factorization formulas for spin dependent $pp$ and/or $pA$ reactions.
A simplification emerges for processes involving electromagnetic interactions such as in semi-inclusive deep inelastic scattering (SIDIS) where polarized $\Lambda$'s can be produced in $ep\rightarrow e\Lambda + X$ or in quasi-real photoproduction. Such experimental studies were carried out by the HERMES collaboration~\cite{HERMES:2006lro, HERMES:2007fpi,HERMES:2014fmx} as well as in neutrino-nucleon scattering by the NOMAD Collaboration~\cite{NOMAD:2000wdf,NOMAD:2001iup}.
Probably the most direct process to gain access to (un)polarized $\Lambda$ fragmentation functions (FFs) is single and/or semi-inclusive hadron production electron-positron annihilation (SIA). Recently, the Belle Collaboration~\cite{Guan:2018ckx} has collected data for the transverse polarization of $\Lambda$-hyperons produced together with light mesons in an almost back-to-back configuration as well as for Lambda’s inclusively produced, where the $\Lambda$ transverse momentum is measured with respect to the thrust axis. Also, data on polarized $\Lambda$ fragmentation in this reaction has been provided by the OPAL Collaboration at LEP~\cite{OPAL:1997oem}. This measurement was performed on the $Z$-pole, that is at the cm energy equal to the mass of the $Z$ boson.
In this paper, we investigate associated production of transversely polarized $\Lambda$-hyperons in the process $e^+ e^-\rightarrow \Lambda^\uparrow (\pi /K) + X$
as well as its inclusive production in $e^+ e^-\rightarrow \Lambda^\uparrow (\rm{jet}) + X$.
We present a renewed analysis of Belle data by exploiting the CSS evolution equations and the recent theory developments on the factorization of single-inclusive hadron production in $e^+ e^-$ annihilation processes~\cite{Kang:2020yqw,Gamberg:2021iat}.
A critical issue in
the first phenomenological studies~\cite{DAlesio:2020wjq,Callos:2020qtu} of the $\Lambda$ polarizing FFs from the analysis of Belle data is that the
extraction of these TMD FFs was carried out at fixed scale ($Q=10.58\,\text{GeV}$). Thus these studies do not employ TMD evolution. Moreover, and more relevant, in Ref.~\cite{DAlesio:2020wjq} we used a simplified and phenomenological model to study the transversely polarized $\Lambda$ produced in a single-inclusive process\footnote{We notice that a first attempt, within the same simplified model, to describe the transverse $\Lambda$ polarization in unpolarized $pp$ collisions was carried out in Ref.~\cite{Anselmino:2000vs}.}, due to the lack of a generalized TMD factorization formalism.
Indeed, this approach does not enable one to make predictions at different energy scales. For that, one should employ TMD factorization theorems, and consequently proper
evolution equations for both the double and the single-inclusive hadron production cross sections.
In fact, unlike the cross section for double-hadron production in $e^+e^-$ annihilation processes, only recently new advancements in the TMD factorization of the cross section for single-inclusive production processes have appeared. Among them, are the works of Refs.~\cite{Makris:2020ltr,Kang:2020yqw}, where the factorization has been formulated within an effective theory context, and where a first phenomenological analysis of the Belle data based on the latter paper was carried out in~\cite{Gamberg:2021iat}. Moreover, a Collins-Soper-Sterman (CSS) formalism has been adopted in Refs.~\cite{Boglione:2020auc,Boglione:2020cwn,Boglione:2021wov}.
The purpose of this study is therefore to present a renewed analysis of Belle data by exploiting the TMD framework in its full glory, paying special attention to scale evolution effects and to the nonperturbative component of the polarizing FF. We will also touch a couple of fundamental issues, namely the $SU(2)$ isospin symmetry and the role of heavy quark contributions.
The paper is organised as follows: in Section~\ref{sec:2_h_prod} we present the main formulas and the cross section for the production of a transversely polarized spin-$1/2$ hadron, in association with a light hadron, in $e^+e^-$ annihilation processes, and their expressions in the impact parameter space. Then in Section~\ref{sec:CSS_RG_eqs} we show how these convolutions can be treated within TMD factorization, by employing the CSS evolution equations. In Section~\ref{sec:1-h_prod} we summarize some useful results already presented in Ref.~\cite{Kang:2020yqw}, giving expressions for the cross sections for single-inclusive hadron production and for the transverse polarization.
All these results will be exploited to re-analyze the Belle data in Section~\ref{sec:analysis}, where we show the outcomes of the fits for the double-hadron production data alone and the combined fit of both data set, discussing our main findings. Here we also consider the role of OPAL data, checking our predictions (based only on Belle data) against them or including them in a global fit. Lastly, in Section~\ref{sec:conclusions} we collect our concluding remarks.
\section{Double-hadron production}
\label{sec:2_h_prod}
We start considering the process $e^+e^-\to h_1(S_1)\, h_2(S_2) +X$ where $h_1$ and $h_2$ are two spin-1/2 hadrons, with spin polarization vectors $S_{1,2}$ and masses $M_{1,2}$, produced almost back-to-back in the center-of-mass frame of the incoming leptons. For more details we refer to Ref.~\cite{DAlesio:2021dcx}.
\subsection{Kinematics and cross section}
\label{subsec:kin}
We adopt the following kinematics set-up: we fix the $\hat{z}_L$ axis (in a \emph{laboratory} frame) along the momentum of the second hadron, $h_2$, with the first one, $h_1$, moving in the opposite hemisphere, with a small transverse momentum $\bm{P}_{1T}$ with respect to the second hadron direction. This is illustrated in Fig.~\ref{fig:kin2_2}, where $(\hat{\bm{x}}_L,\hat{\bm{y}}_L,\hat{\bm{z}}_L)$ are the unit vectors in the laboratory frame and $\bm{P}_{1}$ and $\bm{P}_{2}$ are the momenta of, respectively, the first and second hadron. %
We notice that the frame adopted here has the unit vectors,
\hat{y}_L, \hat{z}_L$, inverted with respect to those of the hadron frame used in~\cite{DAlesio:2021dcx}. This choice allows us to employ directly the convolutions adopted in Ref.~\cite{Boer:1997mf} (see also Ref.~\cite{Pitonyak:2013dsu}), with a direct connection with the convolutions in $\bm{b}_T$-space (see below).
We can define two planes: the \emph{Lepton Plane}, determined by the leptons and the hadron $h_2$ momenta, and the \emph{Production Plane}, determined by the momenta of the two observed hadrons, $h_{1,2}$, at an angle $\phi_1$ with respect to the \emph{Lepton Plane}.
\begin{figure}[!th]
\centering
\includegraphics[width=13cm]{Imm/had_frame22.png}
\caption{Kinematics for the process $e^+e^-\to h_1\, h_2 +X$ in the hadron-frame configuration.}
\label{fig:kin2_2}
\end{figure}
In this configuration, referred to as the hadron frame, one measures only the momenta of the two hadrons and the azimuthal distribution of the hadron $h_1$ around the hadron $h_2$ direction. No information on the original quark-antiquark direction is required.
From the theoretical point of view, it is however more convenient to adopt yet a different frame, where the two hadrons are back to back, along a new $\hat{\bm{z}}$ axis, and the hadron transverse unbalance is now carried out by the virtual photon. All details of this transformation are given in Appendix~\ref{apx:rotations}. In this frame the differential cross section can be expressed as~\cite{Boer:1997mf}
\begin{equation}
\frac{d\sigma^{e^+e^-\to h_1(S_1) h_2(S_2)\, X}}{2dy dz_1dz_2d^2\bm{q}_{T}}\,,
\label{eq:xsec2h}
\end{equation}
where $\bm{q}_{T}$ is the transverse momentum of the virtual photon. This is related to the transverse momentum of the first hadron as follows:
\begin{equation}
\bm{P}_{1T} = - z_1 \bm{q}_T\,.
\label{eq:qt_P1T}
\end{equation}
The two scaling variables in Eq.~(\ref{eq:xsec2h}) are the usual light-cone momentum fractions $z_{1,2}$ of the final-state hadrons, defined as
\begin{equation}
z_1 = \frac{P^-_{1}}{k^-}, \qquad z_2 = \frac{P^+_{2}}{p^+}\,,
\label{eq:light_con_def}
\end{equation}
where $k$ and $p$ are the four-momenta of the quark and the antiquark, fragmenting with a certain transverse momentum $\bm{k}_\perp$ and $\bm{p}_\perp$, respectively, into the hadron $h_1$ and $h_2$.
These scaling variables are in turn directly related to another set of variables, usually adopted in experimental analyses: the energy fraction
\begin{equation}
z_{h} = \frac{E_h}{E_q} =\frac{2E_h}{Q} \simeq z \Bigg( 1 + \frac{M^2_{h}}{ z^2 {Q}^2}\Bigg)\,,
\label{eq:en_fract}
\end{equation}
and the momentum fraction
\begin{equation}
\begin{split}
z_{p} =\frac{|\bm{P}_h|}{E_q} =\frac{2|\bm{P}_h|}{Q} \simeq z \Bigg( 1 - \frac{M^2_{h}}{ z^2 {Q}^2}\Bigg)\,,
\end{split}
\label{eq:long_fract}
\end{equation}
where $Q$ is the cm energy of the process, $Q^2 = q^2$, with $q$ being the four momentum of the virtual photon. The last equalities are obtained neglecting powers of $\bm{k}_\perp^2/Q^2$. Notice that the variable $z_h$, usually defined also as an invariant, $z_h= 2P_h\cdot q/ Q^2$, coincides with the energy fraction above in the hadron frame.
Finally, the fraction $y$ is defined as $y=P_2\cdot k_{e^+} /P_2\cdot q$ (with $k_{e^+}$ being the $e^+$ momentum), that in the hadron frame reduces to
\begin{equation}
y = \Big(1 - \sqrt{1-4M_2^2/z_{h_2}^2Q^2}\cos\theta\Big)/2 \simeq (1 - \cos\theta)/2\,,
\end{equation}
where $\theta$ is the angle between the hadron 2 momentum and the incoming lepton directions (see Fig.~\ref{fig:kin2_2}).
Notice that in all relevant variables we will keep kinematic corrections in $M^2/Q^2$, useful for the study of massive hadron production. On the other hand, as we will see below, the $y$ (or the $\theta$) dependence will not play any direct role in our analysis.
In general, the cross sections can be written in terms of convolutions of two generic TMD fragmentation functions \cite{DAlesio:2021dcx, Boer:1997mf}, defined as follows
\begin{equation}
\mathcal{F}[\omega D \bar{D}] = \sum_q e^2_q \int d^2\bm{k}_T d^2\bm{p}_T \, \delta^{(2)}(\bm{k}_T + \bm{p}_T - \bm{q}_T) \,\omega(\bm{k}_T, \bm{p}_T) D(z_1,\bm{k}_{\bot}) \bar{D}(z_2,\bm{p}_{\bot}) \,,
\label{eq:conv_general}
\end{equation}
where $\omega$ is a suitable weight-factor depending on the two transverse momenta, $D$ and $\bar{D}$ are the TMD-FFs of the first and second hadron and $\bm{k}_{T}$ and $\bm{p}_{T}$ are the transverse momenta of the quark/antiquark with respect to the hadron $h_1$ and $h_2$. Notice that one can easily relate these transverse momenta to $\bm{k}_{\bot}$ and $\bm{p}_{\bot}$, the transverse momenta of the hadrons with respect to their own parent quarks as follows (see Appendix \ref{apx:rotations})
\begin{equation}
\bm{k}_T = -\frac{\bm{k}_{\bot}}{z_{p_1}}; \qquad \bm{p}_T = -\frac{\bm{p}_{\bot}}{z_{p_2}}\,.
\end{equation}
\subsection{Transversely polarized hadron production}
\label{subsec:pol_definition}
We now consider specifically the associated production of a transversely polarized spin-$1/2$ hadron, $h_1$, with an unpolarized hadron, $h_2$.
If the polarization is measured only as a function of the hadron energy fractions, with the proper use of Eqs. (\ref{eq:en_fract}) and (\ref{eq:long_fract}), we can give it as the ratio of two $\bm{q}_T$-integrated convolutions
\begin{equation}
P^{h_1}_T(z_1,z_2) = - \sin(\phi_1 - \phi_{S_1}) \frac{ \int d^2\bm{q}_T \, F^{\sin(\phi_1 - \phi_{S_1})}_{TU} }{ \int d^2\bm{q}_T \, F_{UU}} \,,
\label{pol_qt}
\end{equation}
where $\phi_{S_1}$ is the azimuthal angle of the spin of the hadron $h_1$ and where we have simplified a common factor coming from the hard partonic subprocess (see below).
The two convolutions are defined as follows:
\begin{eqnarray}
F_{UU} & = & \mathcal{F}[D_1 \bar{D}_1] \\
F^{\sin(\phi_1 - \phi_{S_1})}_{TU} &=& \mathcal{F}\bigg[\frac{\hat{\bm{h}}\cdot \bm{k}_T}{M_{1}}D^{\perp}_{1T} \bar{D}_1\bigg]\,,
\label{eq:main_pol_conv}
\end{eqnarray}
where $D_1(z,{k}_{\bot})$ is the unpolarized TMD fragmentation function and $D^{\perp}_{1T}(z,{k}_{\bot})$ is the polarizing FF, with
$\hat{\bm{h}}= \bm{P}_{1T}/|\bm{P}_{1T}|$.
Notice that there is another common notation for the polarizing FF, related to the probability that an unpolarized quark fragments into a transversely polarized spin-1/2 hadron~\cite{Bacchetta:2004jz}:
\begin{equation}
\Delta^{N}\! D_{h^\uparrow/q}(z,{k}_\perp) = \frac{{k}_\perp}{zM_h} D_{1T}^\perp (z,{k}_\perp)\,.
\end{equation}
When the polarization is measured perpendicularly to the production plane, that is along the unit vector $\hat{\bm{n}}$ defined as:
\begin{equation}
\hat{\bm{n}} \equiv (\cos\phi_n, \sin\phi_n,0)= \frac{- \bm{P}_{2} \times \bm{P}_{1}}{|\bm{P}_{2} \times \bm{P}_{1}|} = -\sin\phi_1 \hat{\bm{x}}_L + \cos\phi_1 \hat{\bm{y}}_L\,,
\end{equation}
the factor entering Eq.~(\ref{pol_qt}) simplifies as~\cite{DAlesio:2021dcx}
\begin{equation}
- \sin(\phi_1 - \phi_{S_{1}}) = 1 \,.
\end{equation}
Generally, one uses the TMD fragmentation functions in the conjugate $\bm{b}_T$-space.
More precisely, the Fourier transform of the unpolarized FF is defined as:
\begin{equation}
\begin{split}
\widetilde{D}_1(z,b_T) = & \int d^2\bm{k}_T \, e^{i\bm{b}_T \cdot \bm{k}_T } D_1(z,{k}_{\perp})
= 2 \pi \int dk_T \, k_T J_0(b_T k_T) D_1(z,k_{\perp}) \,,
\end{split}
\end{equation}
\noindent where we have used Eq.~(\ref{bessel_0}), the integral definition of $J_0$, the Bessel function of the first kind of order zero. With the above relation, the $F_{UU}$ convolution in $\bm{b}_T $-space can be written as:
\begin{equation}
F_{UU} = \mathcal{F}[D_1 \bar{D}_1] = \mathcal{B}_0 \Big[\widetilde{D} \widetilde{\bar{D}}\Big]
%
= \sum_q e^2_q \int \frac{d b_T}{(2 \pi)} \, b_T J_0(b_T\, q_T) \widetilde{D}_1(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T)\,.
\end{equation}
Regarding the $\bm{b}_T $-space convolution of $F^{\sin(\phi_1 - \phi_{S})}_{TU}$, we first define the Fourier transform of the product of the polarizing fragmentation function with $k^i_T$, the $i$-th component of the quark transverse momentum with respect to the hadron direction, see Appendix~\ref{apx2:conv_FT}:
\begin{equation}
\int d^2\bm{k}_T \, \frac{k^i_T}{M_{1}} e^{i\bm{b}_T \cdot \bm{k}_T} D^{\perp}_{1T}(z,k_{\perp})
= i b^i_T M_{1} \widetilde{D}^{\perp (1)}_{1T}(z,b_T) \,,
\end{equation}
where $\widetilde{D}^{\perp (1)}_{1T}(z,b_T)$, the first moment of the polarizing fragmentation function in $\bm{b}_T $-space, is defined as
\begin{equation}
\widetilde{D}^{\perp (1)}_{1T}(z,b_T) = -\frac{2}{M^2_{1}}\pr{}{b^2_T} \widetilde{D}^{\perp}_{1T}(z,b_T) \; ,
\label{eq:bt_first_mom_1}
\end{equation}
and $ \widetilde{D}^{\perp}_{1T}(z,b_T)$ is the Fourier transform of the polarizing FF, Eq. (\ref{eq:pFF_bt}).
Notice that the first moment of the polarizing FF in $\bm{k}_T $-space, ${D}^{\perp (1)}_{1T}(z)$, defined as
\begin{equation}
{D}^{\perp (1)}_{1T}(z) = \int d^2 \bm{k}_{\perp} \, \bigg(\frac{\bm{k}^2_{\perp}}{2 z^2 M^2_h} \bigg) {D}^{\perp}_{1T}(z,{k}_{\perp}) \,,
\end{equation}
can be related to the corresponding one in $\bm{b}_T$-space, Eq.~(\ref{eq:bt_first_mom_1}), as follows:
\begin{equation}
\lim_{b_T \to 0} \widetilde{D}^{\perp (1)}_{1T}(z,b_T) =\frac{1}{z^2} {D}^{\perp (1)}_{1T}(z)\; .
\label{eq:firstmoment_z}
\end{equation}
Employing the above equations and using the integral definition of the Bessel function $J_1$, Eq.~(\ref{bessel_1}), we can find the expression of $F^{\sin(\phi_1 - \phi_{S_1})}_{TU}$ in $\bm{b}_T $-space:
\begin{eqnarray}
F^{\sin(\phi_1 - \phi_{S_1})}_{TU} & = &
\mathcal{F}\bigg[\frac{\hat{\bm{h}}\cdot \bm{k}_T}{M_{1}}D^{\perp}_{1T} \bar{D}_1\bigg] = M_{1} \mathcal{B}_1 \Big[\widetilde{D}^{\perp (1)}_{1T} \widetilde{\bar{D}}_1\Big] \nonumber\\
& = & M_{1} \sum_q e^2_q \int \frac{d b_T}{2 \pi} \,b^2_T J_1(b_T\,q_T) \widetilde{D}^{\perp (1)}_{1T}(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T)\,.
%
\end{eqnarray}
Finally, we can express the polarization of the final hadron, Eq.~(\ref{pol_qt}), along the $\hat{\bm{n}}$ direction as the ratio of the two convolutions in $\bm{b}_T $-space:
\begin{equation}
P^{h_1}_n(z_1,z_2) = \frac{ \int d^2\bm{q}_T \, F^{\sin(\phi_1 - \phi_{S_1})}_{TU} }{ \int d^2\bm{q}_T \, F_{UU}} = \frac{M_{1} \int dq_T\;q_T\;d\phi_1 \, \mathcal{B}_1 \Big[\widetilde{D}^{\perp (1)}_{1T} \widetilde{\bar{D}}_1\Big]}{ \int dq_T\;q_T\;d\phi_1 \, \mathcal{B}_0 \Big[\widetilde{D}_1 \widetilde{\bar{D}}_1\Big] }\,,
\label{pol_ratio}
\end{equation}
where
\begin{align}
\mathcal{B}_0 \Big[\widetilde{D}_1 \widetilde{\bar{D}}_1\Big] =& \sum_q e^2_q\int \frac{d b_T}{2 \pi} \, b_T J_0(b_T\, q_T) \widetilde{D}_1(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T)\label{eq:all_convs1} \\
%
\mathcal{B}_1 \Big[\widetilde{D}^{\perp (1)}_{1T} \widetilde{\bar{D}}_1\Big] =& \sum_q e^2_q \int \frac{d b_T}{2 \pi} \,b^2_T J_1( b_T\,q_T) \widetilde{D}^{\perp (1)}_{1T}(z_1,b_T) \widetilde{\bar{D}}_1(z_2,b_T)\label{eq:all_convs2}\,.
\end{align}
The last step is to integrate both convolutions on $\bm{q}_T $. %
The integration over the azimuthal angle, $\phi_1 $, is trivial, giving a factor of $2\pi$ that cancels out in the ratio. %
Moreover, since the only terms inside the convolutions depending on $q_T $ are the Bessel functions, we can separately integrate them, obtaining
\begin{equation}
\int^{q_{T_{\text{max}}}}_0 dq_T \, q_T J_0(b_T\, q_T) = \frac{q_{T_{\text{max}}}}{b_T}J_1(b_T\, q_{T_{\rm max}}) \\
\label{int_qtmax_1}
\end{equation}
\begin{equation}
%
%
\int^{q_{T_{\text{max}}}}_0 dq_T \, q_T J_1(b_T\, q_T) =\frac{\pi q_{T_{\text{max}}}}{2 b_T}
%
\{J_1(b_T\, q_{T_{\text{max}}})\bm{H}_0(b_T\, q_{T_{\rm max}}) - J_0(b_T\, q_{T_{\text{max}}})\bm{H}_1(b_T\, q_{T_{\text{max}}}) \}\,,
\label{int_qtmax_2}
\end{equation}
\noindent where ${\bm{H}}_{0,1}$ are the Struve functions of order zero and one respectively. Notice that in the above integration we have introduced a maximum value $q_{T_{\text{max}}}$,
that has to fulfil the condition $q_{T_{\text{max}}} \ll Q$, in order to respect the validity of the TMD factorization~\cite{Collins:2016hqq} . %
In the phenomenological analysis, Section~\ref{sec:analysis}, we will test different choices of the ratio $q_{T_{\text{max}}}/Q$.
\section{Double-hadron production: CSS formalism}
\label{sec:CSS_RG_eqs}
In this Section we elaborate on the convolutions, presented in Section~\ref{sec:2_h_prod}, with the proper treatment of the scale evolution within the Collins-Soper-Sterman (CSS) approach (see Refs.~\cite{collins_2011, Collins:2014jpa, Collins:2016hqq} for more details). According to the CSS formalism, the complete expressions of the two convolutions entering the transverse polarization observable, Eq.~(\ref{pol_ratio}),
are given by:
\begin{equation}
\begin{split}
\mathcal{B}_0 \Big[\widetilde{D}_1 \widetilde{\bar{D}}_1\Big] &= \sum_q e^2_q \,\mathcal{H}^{(e^+e^-)}(Q) \\
&\times\int \frac{d b_T}{(2 \pi)} \, b_T J_0(b_T\, q_T) \widetilde{D}_{1,q/h_1}(z_1,b_T;\zeta_1,\mu) \widetilde{\bar{D}}_{1,\bar{q}/h_2}(z_2,b_T;\zeta_2,\mu) \\
\end{split}
\label{B0_css}
\end{equation}
\begin{equation}
\begin{split}
\mathcal{B}_1 \Big[\widetilde{D}^{\perp (1)}_{1T} \widetilde{\bar{D}}_1\Big] =& \sum_q e^2_q \, \mathcal{H}^{(e^+e^-)}(Q)\\
\times&\int \frac{d b_T}{2 \pi} \,b^2_T J_1( b_T\,q_T) \widetilde{D}^{\perp (1)}_{1T, q/h_1}(z_1,b_T;\zeta_1,\mu) \widetilde{\bar{D}}_{1,\bar{q}/h_2}(z_2,b_T;\zeta_2,\mu)\,, \\
\end{split}
\label{B1_css}
\end{equation}
where $\mathcal{H}^{(e^+e^-)}(Q)$ is the hard scattering part (depending also on $y$), for the massless on-shell process $e^+e^-\to q \bar{q}$, at the center-of-mass energy $Q$.
With respect to the expressions given in the previous section, the two fragmentation functions now depend explicitly on two scale arguments: the renormalization scale $\mu$ and the $\zeta$ scale, that describes the effect of the recoil against the emission of soft gluons into an energy range determined approximately by $\mu$ and $\zeta$. The dependence on these two scales is regulated by the CSS and Renormalization Group (RG) equations.
\subsection{Evolution equations for TMD fragmentation functions}
The CSS evolution equation for the $\zeta$ dependence of the unpolarized TMD-FF has the following form:
\begin{equation}
\pr{\ln \widetilde{D}_1(z,b_T;\zeta,\mu)}{\ln{\sqrt{\zeta}}} = \widetilde{K}(b_T;\mu)\,,
\label{eq_rapidity}
\end{equation}
where $ \widetilde{K}$ is the CSS kernel~\cite{collins_2011}.
It is flavour and spin independent, but different for quarks and gluons. Its RG equation is
\begin{equation}
\prtot{\widetilde{K}(b_T;\mu)}{\ln{\mu}} = - \gamma_K(g(\mu))\,,
\label{CSS_kern}
\end{equation}
where the anomalous dimension $\gamma_K$ has no dependence on $b_T$, since the UV divergences only arise from virtual graphs \cite{collins_2011}. The corresponding RG equation for the fragmentation function is given by
\begin{equation}
\prtot{\ln \widetilde{D}_1 (z,b_T;\zeta,\mu)}{\ln{\mu}} = \gamma_D(g(\mu);\zeta/\mu^2)\,.
\label{eq_scalemu}
\end{equation}
Since the derivatives of the FF with respect to $\mu$ and $\zeta$ commute, we can finally obtain the energy dependence of $\gamma_D$:
\begin{equation}
\gamma_D(g(\mu);\zeta/\mu^2) = \gamma_D(g(\mu);1) - \frac{1}{2}\gamma_K(g(\mu)) \ln{\frac{\zeta}{\mu^2}}\,.
\end{equation}
In addition, the anomalous dimensions and the CSS kernel can be computed order by order perturbatively.
By solving Eq.~(\ref{eq_rapidity}), that gives us the evolution in $\zeta$, and by using it in Eq.~(\ref{eq_scalemu}), we get the scale evolution from $\mu_0$ to $\mu$ (with $\mu_0$ large enough to start already in the perturbative region). This eventually leads to
\begin{equation}
\begin{split}
\widetilde{D}_1(z,b_T;\zeta,\mu)
%
&=\widetilde{D}_1(z,b_T;\zeta_{0},\mu_0) \exp{\Bigg \{ \frac{1}{2}\widetilde{K}(b_T;\mu_0) \ln\frac{\zeta}{\zeta_{0}} \Bigg\}}\\
%
&\times \exp{\Bigg\{ \int^{\mu}_{\mu_0} \frac{d\mu'}{\mu'}\,\bigg[ \gamma_D(g(\mu');1) - \frac{1}{2}\gamma_K(g(\mu')) \ln{\frac{\zeta}{\mu'^2}} \bigg]\Bigg\}} \, .
\end{split}
\label{eq:unp_bt_space_1}
\end{equation}
The dependence of the FF on $\zeta$ involves the function $\widetilde{K}$, implying an energy dependence on the shape of the transverse momentum distribution. Moreover, the function $\widetilde{D}_1$, at its reference scales $\zeta_{0}$ and $\mu_0$, can be thought as the Fourier transform of an intrinsic transverse momentum distribution of the hadron with respect to its parent parton.
The full solution of the evolution equations in terms of the anomalous dimensions and the CSS kernel, and all the above results, can be directly extended to the $\widetilde{D}^{\perp (1)}_{1T}$ function~\cite{collins_2011}.
\subsection{Small-$b_T$ expansion}
The first term on the right hand side of Eq.~(\ref{eq:unp_bt_space_1}) is the TMD-FF at the reference energy scale: it is related to the short distance and small-$\bm{b}_T$ behaviour of $D_1$ and therefore computable in perturbation theory. For such reason, at small-$\bm{b}_T$, the unpolarized TMD-FF can be matched onto the corresponding integrated fragmentation function $d_{j/h}(z;\mu)$ via an Operator Product Expansion (OPE):
\begin{equation}
\widetilde{D}_{1,q/h}(z,b_T;\zeta_{0},\mu_0)
=\sum_j \int^1_{z} \frac{d\hat{z}}{\hat{z}^{3-2\epsilon}} \widetilde{C}_{j/q}(z/\hat{z},b_T;\zeta_{0},\mu_0,g(\mu_0)) \, d_{j/h}(\hat{z};\mu_0) + \mathcal{O}[(mb_T)^p]\,,
\label{OPE_unp}
\end{equation}
where the error term is suppressed by some power of the transverse position. The sum is over all parton types $j$, including gluons and antiquarks. When $b_T$ is small, the coefficient function $\widetilde{C}_{j/q}$ can be expanded in perturbation theory and calculated from Feynmann graphs with external on-shell partons of type $j$, with a double-counting subtraction in order to cancel all collinear contributions \cite{collins_2011}. The lowest-order coefficient is simply given as:
\begin{equation}
\widetilde{C}_{j/q}(z/\hat{z},b_T;\zeta,\mu,g(\mu)) = \delta_{jq}\,\delta(z/\hat{z} -1) + \mathcal{O}(g^2)\,.
\label{C0_coef}
\end{equation}
An OPE of the same kind applies also to other collinear fragmentation functions, e.g.~$G_{1L}$ and $H_{1T}$, but they generally have different coefficient functions beyond lowest-order. For the other, polarization-dependent, TMD fragmentation functions, like the Collins and the polarizing fragmentation functions, it is possible to generalize the OPE involving quantities that are associated to matrix elements of higher-twist operators~\cite{Boer:2010ya}. For the Sivers function, for instance, this would be the Qiu-Sterman function~\cite{Qiu:1991pp,Qiu:1991wg}.
\subsection{Matching the perturbative and nonperturbative $b_T$ regions}
TMD evolution follows from generalized renormalization properties of the operator definitions for TMD parton distribution and fragmentation functions. In order to combine the small-$\bm{b}_T$ dependence coming from perturbative calculations with the one from the nonperturbative part (that must be extracted from experimental data) it is necessary to introduce a matching procedure. To match the perturbative and nonperturbative contributions, one defines large and small $\bm{b}_T$ through a function of $\bm{b}_T$ that freezes above some $b_{\text{max}}$ and equals $\bm{b}_T$ for small values.
We adopt the following standard procedure~\cite{Collins:1984kg}. First, we introduce the parameter $b_{\text{max}}$, representing the maximum distance at which perturbation theory is to be trusted, usually taken within an interval of $[0.5-1.5]\,\text{GeV}^{-1}$. Then we define a function $\bm{b}_*(\bm{b}_T)$, that almost equals $b_T$ at small $b_T$ and saturates at $b_{\text{max}}$ at large $b_T$:
\begin{equation}
\bm{b}_* = \frac{\bm{b}_T}{\sqrt{1+b^2_T/b^2_{\text{max}}}}\,,
\label{b_star0}
\end{equation}
and re-define the CSS Kernel as:
\begin{equation}
\widetilde{K}(b_T;\mu) = \widetilde{K}(b_*;\mu) - g_K(b_T;b_{\text{max}})\,.
\label{K_star}
\end{equation}
In this way, $\widetilde{K}(b_*;\mu)$ is computed in a region where perturbation theory is appropriate and the correction term, $g_K$, is important only at large ${b}_T$. This last term, $g_K$, to be extracted from data fits, is a function of $b_T$ and can depend explicitly or not on the parameter $b_{\text{max}}$. Since it is the difference of $ \widetilde{K}$ calculated at two values of its position argument, it is RG invariant and has to vanish as $b_T \to 0$.
If we want to match the perturbative and nonperturbative part of the unpolarized FF $\widetilde{D}_1$, we can then use $\bm{b}_*$ as defined in Eq.~(\ref{b_star0}).
Generalizing Eq.~(\ref{K_star}), it is possible to define an intrinsically nonperturbative part of the FF with the following decomposition:
\begin{eqnarray}
&&\widetilde{D}_{1, q/h}(z,b_T;\zeta,\mu) = \widetilde{D}_{1, q/h}(z,b_*;\zeta,\mu) \Bigg[\frac{\widetilde{D}_{1, q/h}(z,b_T;\zeta,\mu)}{\widetilde{D}_{1, q/h}(z,b_*;\zeta,\mu)}\Bigg]\nonumber\\
&&= \widetilde{D}_{1, q/h}(z,b_*;\zeta,\mu)\exp\Bigg[ -g_{q/h}(z,b_T;b_{\text{max}}) - \,g_K(b_T;b_{\text{max}})\ln\frac{\sqrt{\zeta}}{\sqrt{\zeta_{0}}} \Bigg] \nonumber\\
&&= \widetilde{D}_{1, q/h}(z,b_*;\zeta,\mu)\exp\Bigg[ -g_{q/h}(z,b_T;b_{\text{max}}) - \,g_K(b_T;b_{\text{max}})\ln\frac{\sqrt{\zeta}\,z}{M_{h}} \Bigg] \, ,
\label{D_star}
\end{eqnarray}
where in the second line we have introduced the perturbatively calculable $\widetilde{D}(b_*)$ and in the last line we have used the reference value $\zeta_{0} = M^2_{h}/z^2$~\cite{collins_2011}.
By employing Eq.~(\ref{eq:unp_bt_space_1}), the anomalous dimensions, $\gamma_D$ and $\gamma_K$, cancel between numerator and denominator in the square brackets (first line) and only $g_K$ survives.
The remaining factor, $e^{-g_{q/h}}$, defined as~\cite{collins_2011, Collins:2014jpa, Collins:2016hqq}
\begin{equation}
\exp\big[-g_{q/h}(z,b_T;b_{\text{max}})\big] = \frac{\widetilde{D}_{1, q/h}(z,b_T;\zeta,\mu)}{\widetilde{D}_{1, q/h}(z,b_*;\zeta,\mu)} \exp\big[ g_K(b_T;b_{\text{max}})\ln\sqrt{\zeta}/\sqrt{\zeta_{0}}\big]\,,
\label{eq:interme_css_ev}
\end{equation}
can be interpreted as the nonperturbative part of the intrinsic transverse momentum distribution.
Some comments are in order: both $g_K$ and $g_{q/h}$ vanish approximately as $b^2_T$ at small $b_T$ \cite{collins_2011}, and become significant when $b_T$ approaches $b_{\text{max}}$ and beyond; they are independent of $\zeta$ and $\mu$, being invariant under the application of the CSS and RG equations, while they do depend on the choice of $b_{\text{max}}$. On the other hand, the full TMD fragmentation function and the function $\widetilde{K}$ are independent of $b_{\text{max}}$ and the use of the $b_*$ prescription.
The flavour and $z$ dependences of $g_K$ and $g_{q/h}$ follow from those of the corresponding parent functions, respectively $\widetilde{K}$ and the TMD fragmentation functions \cite{Collins:2014jpa}. Since $\widetilde{K}$ is independent of the quark's and hadron's type, polarization and fraction $z$, so is $g_K$. %
The same, of course, is not true in general for the TMD fragmentation functions and therefore for the factor $e^{-g_{q/h}}$. %
It is worth mentioning that this last term is usually written as $M_D( b_T,{z} ;b_{\text{max}})$ or $D_{NP}(b_T,{z};b_{\text{max}})$, a generic function of $b_T$: this is because it could assume also a non-exponential functional form, still preserving its properties, and the fact that at small $b_T$ it goes like 1 + ${\cal O}(b_T^2)$; it is referred to as the \emph {nonperturbative part} of the fragmentation function, and within a parton model, can be seen as the Fourier transform of the transverse momentum distribution.
Like $g_K$, the function $M_D$ can depend explicitly or not on the parameter $b_{\text{max}}$.
In a more general way, the phenomenological extraction of both nonperturbative functions is affected by the choice of the $b_{\text{max}}$ value.
To use the perturbative small-$b_T$ result from Eq.~(\ref{OPE_unp}), it is necessary to evolve the $\widetilde{D}$ term in Eq.~(\ref{D_star}), with the $b_*$ prescription, from a region where no large kinematic ratios appear in the coefficient function $\widetilde{C}$, whose logarithms could spoil the use of the perturbative approach~\cite{collins_2011}. The standard choice is to replace $\mu_0$ by:
\begin{equation}
\mu_{b} = \frac{C_1}{b_*(b_T)}\, ,
\end{equation}
where $C_1 = 2e^{-\gamma_E}$ (with $\gamma_E$ being the Euler-Mascheroni constant), and use for the reference value $\zeta_{0}$ the same value, that is $\zeta_0=\mu^2_{b} $. Then the TMD fragmentation function can be written as:
\begin{eqnarray}
\widetilde{D}_{1, q/h}(z,b_T;\zeta,\mu) &=&
\sum_j \int^1_{z} \frac{d\hat{z}}{\hat{z}^{3-2\epsilon}} \widetilde{C}_{j/q}(z/\hat{z},b_*;\mu^2_b,\mu_b,g(\mu_b)) \, d_{j/h}(\hat{z};\mu_b) \nonumber\\
&\times& M_D(b_T,z;b_{\text{max}}) \exp\Bigg\{ - \,g_K(b_T;b_{\text{max}})\ln\frac{\sqrt{\zeta}\,z}{M_{h}} \Bigg\} \label{D_full} \\
&\times &\exp{\Bigg\{ \frac{1}{2}\widetilde{K}(b_*;\mu_b) \ln\frac{\zeta}{\mu^2_b} + \int^{\mu}_{\mu_b} \frac{d\mu'}{\mu'}\,\bigg[ \gamma_D(g(\mu');1) - \frac{1}{2}\gamma_K(g(\mu')) \ln{\frac{\zeta}{\mu'^2}} \bigg]\Bigg\}}\nonumber \,.
\end{eqnarray}
Finally, in order to properly control the low-$b_T$ region (to ensure the matching at high $k_T$), we modify the $\bm{b}_T$ definition using~\cite{Collins:2016hqq}:
\begin{equation}
b_c(b_T) = \sqrt{b^2_T + b^2_{\text{min}}}\,,
\end{equation}
where $b_{\text{min}} = 2e^{-\gamma_E}/Q$, decreasing like $1/Q$ in contrast to $b_{\text{max}}$, which remains fixed. This definition reduces to $b_T$ when $b_T \gg 1/Q$ but it is of order $1/Q$ when $b_T$ is small, thereby providing an effective cutoff at small $b_T$. Consistently, $b_*$ has to be replaced by:
\begin{equation}
b_*(b_c(b_T)) =\sqrt{ \frac{b^2_T + b^2_{\text{min}}}{1+b^2_T/b^2_{\text{max}} +b^2_{\text{min}}/b^2_{\text{max}} } }\,,
\label{b_star}
\end{equation}
in order to ensure the requested behaviour simultaneously at small and large $b_T$. Indeed, we have:
\begin{equation}
b_*(b_c(b_T)) \to
\begin{cases}
b_{\text{min}} \quad b_T \ll b_{\text{min}} \\
b_T \qquad b_{\text{min}} \ll b_T \ll b_{\text{max}}\\
b_{\text{max}} \quad b_T \gg b_{\text{max}}
\end{cases}\,.
\label{b_star_gen}
\end{equation}
Lastly, we also redefine $\mu_b$, replacing it by:
\begin{equation}
\bar{\mu}_{b}(b_c(b_T)) = \frac{C_1}{ b_*(b_c(b_T))}\, ,
\end{equation}
implying a maximum cutoff on the renormalization scale equal to $\bar{\mu}_{b} \simeq C_1/b_{\text{min}}$. Recollecting all the above results we can now write the TMD fragmentation function, Eq.~(\ref{D_full}), employing the new definitions of ${b}_T$, as:
\begin{eqnarray}
&& \widetilde{D}_{1, q/h}(z,b_c(b_T);Q^2,Q) =
%
\sum_j \int^1_{z} \frac{d\hat{z}}{\hat{z}^{3-2\epsilon}} \widetilde{C}_{j/q}(z/\hat{z},b_*(b_c(b_T)); \bar{\mu}^2_b, \bar{\mu}_b,g( \bar{\mu}_b)) \nonumber\\
&& \mbox{}\times d_{j/h}(\hat{z}; \bar{\mu}_b)\, M_D(b_c(b_T),z;b_{\text{max}}) \exp\Bigg\{ - \,g_K(b_c(b_T);b_{\text{max}})\ln\frac{Q z}{M_{h}} \Bigg\} \nonumber\\
&&\mbox{} \times \exp{\Bigg\{ \frac{1}{2}\widetilde{K}(b_*;\bar{\mu}_{b}) \ln\frac{Q^2}{\bar{\mu}_{b}^2} + \int^{Q}_{\bar{\mu}_{b}} \frac{d\mu'}{\mu'}\,\bigg[ \gamma_D(g(\mu');1) - \frac{1}{2}\gamma_K(g(\mu')) \ln{\frac{Q^2}{\mu'^2}} \bigg]\Bigg\}}\,,
\label{D_full_bc_Q}
\end{eqnarray}
where we have adopted $\zeta = Q^2$ and $\mu = Q$.
\subsection{Convolutions}
Thanks to the evolution equations and the matching procedure, we can write the full form of the convolutions in Eqs.~(\ref{B0_css}) and (\ref{B1_css}). For the convolution $\mathcal{B}_0 $ we have:
\begin{eqnarray}
\mathcal{B}_0 \Big[\widetilde{D} \widetilde{\bar{D}}\Big] &=& \frac{\mathcal{H}^{(e^+e^-)}(Q)}{z^2_1 z^2_2} \sum_q e^2_q\int \frac{d b_T}{(2 \pi)} \, b_T J_0(b_T\, q_T) \, d_{q/h_1}(z_1; \bar{\mu}_b) \, d_{\bar{q}/h_2}(z_2; \bar{\mu}_b)\nonumber \\
&\times& M_{D_1}(b_c(b_T),\textcolor{black}{z_1}) \, M_{D_2}(b_c(b_T),\textcolor{black}{z_2})
\exp\Bigg\{\!\!-g_K(b_c(b_T);b_{\text{max}})\ln{\bigg(\frac{Q^2 z_1 z_2}{M_{1} M_{2}}\bigg)}\Bigg\} \nonumber\\
%
&\times& \exp{\Bigg\{ \widetilde{K}(b_*;\bar{\mu}_{b}) \ln\frac{Q^2}{\bar{\mu}_{b}^2} + \int^{Q}_{\bar{\mu}_{b}} \frac{d\mu'}{\mu'}\,\bigg[ 2\gamma_D(g(\mu');1) - \gamma_K(g(\mu')) \ln{\frac{Q^2}{\mu'^2}} \bigg]\Bigg\}}\,,
\label{B0_full}
\end{eqnarray}
where in the second line we have omitted the implicit $b_{\rm max}$ dependence in $M_D$ and used the lowest-order coefficient, Eq.~(\ref{C0_coef}), of the OPE expression for both the fragmentation functions.
Similarly, for the convolution $\mathcal{B}_1$ we have:
\begin{eqnarray}
\mathcal{B}_1 \Big[\widetilde{D}^{\perp (1)}_{1T} \widetilde{\bar{D}}_1\Big]
&=& \frac{\mathcal{H}^{(e^+e^-)}(Q)}{z^2_1 z^2_2}\sum_q e^2_q\int \frac{d b_T}{(2 \pi)} \, b^2_T J_1(b_T\, q_T) {D}^{\perp (1)}_{1T} (z_1;\bar{\mu}_b) \, d_{\bar{q}/h_2}(z_2; \bar{\mu}_b) \nonumber\\
&\times& M^{\perp}_{D_1}(b_c(b_T),\textcolor{black}{z_1}) \, M_{D_2}(b_c(b_T),\textcolor{black}{z_2})\exp\Bigg\{-g_K(b_c(b_T);b_{\text{max}})\ln{\bigg(\frac{Q^2 z_1 z_2}{M_{1} M_{2}}\bigg)}\Bigg\} \nonumber\\
%
&\times& \exp{\Bigg\{ \widetilde{K}(b_*;\bar{\mu}_{b}) \ln\frac{Q^2}{\bar{\mu}_{b}^2} + \int^{Q}_{\bar{\mu}_{b}} \frac{d\mu'}{\mu'}\,\bigg[ 2\gamma_D(g(\mu');1) - \gamma_K(g(\mu')) \ln{\frac{Q^2}{\mu'^2}} \bigg]\Bigg\}}\,,\nonumber\\
\label{B1_full}
\end{eqnarray}
where again we have used the lowest-order coefficient for the OPEs and $M^{\perp}_{D_1}$ as the nonperturbative part for the polarizing fragmentation function for the hadron $h_1$. %
The last lines of Eqs.~(\ref{B0_full}) and (\ref{B1_full}) are usually referred to as perturbative Sudakov factors and, as explained above, they can be computed analytically. The anomalous dimension of the fragmentation functions at order $\alpha_s(\mu)$
is~\cite{Aidala:2014hva}:
\begin{equation}
\gamma_D(\alpha_s(\mu);\zeta/\mu^2) = 4 C_F \bigg( \frac{3}{2} - \ln\frac{\zeta}{\mu^2} \bigg) \bigg(\frac{\alpha_s(\mu) }{4\pi} \bigg) + \mathcal{O}(\alpha^2_s(\mu) )\,,
\label{gamma_D}
\end{equation}
where $C_F = 4/3$. Meanwhile the anomalous dimension of the CSS Kernel $\widetilde{K}$ at one-loop order is:
\begin{equation}
\gamma_K(\alpha_s(\mu)) = 8 C_F\bigg(\frac{\alpha_s(\mu) }{4\pi} \bigg) + \mathcal{O}(\alpha^2_s(\mu) )\,,
\label{gamma_K}
\end{equation}
with $\widetilde{K} = 0$ at first order. For the running coupling \cite{Aidala:2014hva} we use the form:
\begin{equation}
\alpha_s(\mu) = \frac{A}{2 \ln(\mu/\Lambda_{\rm QCD})}\,,
\end{equation}
with
\begin{equation}
A = \frac{1}{\beta_0} = \frac{12 \pi}{33 - 2 n_f}\,,
\end{equation}
where $n_f$ is the number of active flavours.\footnote{Actually we consistently use $n_f=3$ with $\Lambda_{\rm QCD} =$ 0.2123~GeV~\cite{Aidala:2014hva}.} We can then get, by analytical integration, the following expression for the exponent in the perturbative Sudakov factor
\begin{eqnarray}
&& \widetilde{K}(b_*;\bar{\mu}_{b}) \ln\frac{Q^2}{\bar{\mu}_{b}^2} + \int^{Q}_{\bar{\mu}_{b}} \frac{d\mu'}{\mu'}\,\bigg[ 2\gamma_D(g(\mu');1) - \gamma_K(g(\mu')) \ln{\frac{Q^2}{\mu'^2}} \bigg] \\
&&= \frac{2A}{\pi}\Bigg[ \ln\bigg( \frac{\ln(Q/\Lambda_{\rm QCD})}{\ln(\bar{\mu}_{b}/\Lambda_{\rm QCD})} \bigg) - \frac{4}{3} \ln(Q/\Lambda_{\rm QCD}) \ln\bigg( \frac{\ln(Q/\Lambda_{\rm QCD})}{\ln(\bar{\mu}_{b}/\Lambda_{\rm QCD})} \bigg) + \frac{4}{3}\ln(Q/\bar{\mu}_{b})\Bigg] \nonumber\,.
\end{eqnarray}
\subsection{Nonperturbative parts}
We discuss now the nonperturbative contributions, entering the above convolution formulas. We will give some details on the corresponding parametrizations available in the literature, focusing on those we will use directly in our phenomenological analysis: namely $g_K$ and $M_D$. For the moment, we will leave apart the other fundamental quantity, $M_D^\perp$, one of the main focus of our study, that will be properly addressed when we discuss our fitting procedure.
Notice that, while for $g_K$ we will use functional forms depending explicitly on $b_{\rm max}$, for $M_D$ this dependence enters only implicitly (as discussed previously) and, for the sake of notation, we will drop it in the sequel.
\subsubsection{$g_K(b_T;b_{\rm max})$}
We start considering $g_k$, an intrinsically nonperturbative quantity, that cannot be computed from first principles. Nevertheless, as shown in Ref.~\cite{Collins:2014jpa}, it is possible to extract some of its properties from perturbative calculations. Indeed, the lowest-order formula for $\widetilde{K}$ gives:
\begin{equation}
g_K(b_T,b_{\text{max}}) = \frac{\alpha_s(C_1/b_*) C_F}{\pi} \ln(1+ b^2_T/b^2_{\text{max}})\,,
\label{ll_lgm}
\end{equation}
that behaves like $b^2_T$ at small $b_T$, but shows a slower logarithmic rise above $b_{\text{max}}$. A similar functional form has been adopted in Refs.~\cite{Aidala:2014hva, Sun:2014dqm} (AFGR/SIYY in the following) with the following expression
\begin{equation}
g_K(b_T;b_{\text{max}})= g_2 \ln\bigg(\frac{b_T}{b_*}\bigg) \,,
\label{}
\end{equation}
with an extracted value of $g_2 = 0.84$ .
For large-$b_T$ values, Eq.~(\ref{ll_lgm}) is not expected to be an accurate parametrization. Indeed, it is an extrapolation of a lowest-order perturbative calculation and it depends strongly on $b_{\text{max}}$ at large $b_T$. The $b^2_T$ behaviour at small $b_T$ can be found expanding Eq.~(\ref{ll_lgm}) in powers of $b_T$, obtaining:
\begin{equation}
g_K(b_T;b_{\text{max}})= \frac{C_F}{\pi} \frac{b^2_T}{b^2_{\text{max}}}\alpha_s(\mu_{b})\,,
\label{}
\end{equation}
with an explicit quadratic form of the $g_K$ function. This justifies the use of the following expression:
\begin{equation}
g_K(b_T;b_{\text{max}}) = \frac{g_2 b^2_T}{2}\,,
\end{equation}
employed and fitted to data in Ref.~\cite{Landry:2002ix} (BLNY in the following) and in Ref.~\cite{Konychev:2005iy} (KN) where they found, respectively, a value of $g_2 = 0.68\,\text{GeV}^2$, with $b_{\text{max}} = 0.5 \,\text{GeV}^{-1}$, and a value of $g_2 = 0.18\,\text{GeV}^2$, with $b_{\text{max}} = 1.5 \,\text{GeV}^{-1}$. Notice that such a large value of $b_{\rm max}$, as we will discuss in the next section, implies a too small renormalization scale, preventing its use in our calculation.
Since the real nonperturbative physics is at larger $b_T$ values, one wants to extract $g_K(b_T;b_{\text{max}})$ with a more general parametrization and be sure that the data used to extract it are sensitive to high values of $b_T$. Moreover, the complete TMD factorization formalism is $b_{\text{max}}$ independent and, in principle, optimized fits should not depend on its choice. %
For a more exhaustive comparison, we will also consider another set of nonperturbative functions, as those extracted from fits on SIDIS, Drell-Yan and $Z$-boson production and discussed in Ref.~\cite{Bacchetta:2017gcc}. Concerning $g_K$, this has the same functional form as BLNY, but with a different value for $g_2$, namely $g_2 = 0.13\,\text{GeV}^2$ (PV17). This implies a softer behaviour in $b_T$.
Moreover, this has been obtained with an \emph{ad hoc} $b_*$ prescription
\begin{equation}
b_* \equiv b_*(b_T;b_{\rm min},b_{\rm max}) = b_{\rm max}\bigg(\frac{1-e^{-b_T^4/b^4_{\rm max}}}{1-e^{-b_T^4/b^4_{\rm min}}} \bigg)^{1/4}
\end{equation}
and the corresponding parametrization of $M_D$ (see below).
Quite recently a new and refined global analysis, at N$^3$LL accuracy, has been performed by the Pavia group~\cite{Bacchetta:2022awv}. For consistency we will not adopt it in our study, where we are using the anomalous dimensions only at one-loop order.
Summarizing, in the phenomenological analysis, Section~\ref{sec:analysis}, we will employ the following functional forms of the nonperturbative function, $g_K$, adopting $b_{\rm max} = 0.6$~GeV$^{-1}$, see Fig.~\ref{fig:all_gk}:
\begin{equation}
\begin{split}
g_K(b_T;b_{\text{max}}) &= \frac{g_2 b^2_T}{2}\,; \quad g_2 = 0.68\,\text{GeV}^2 \qquad \qquad\;\, \text{BLNY} \\
%
g_K(b_T;b_{\text{max}}) &= \frac{C_F}{\pi} \frac{b^2_T}{b^2_{\text{max}}}\alpha_s(\mu_{b_*}) \qquad\qquad\qquad\qquad \text{Quadratic}\\
%
g_K(b_T;b_{\text{max}}) &= \frac{\alpha_s(C_1/b_*) C_F}{\pi} \ln(1+ b^2_T/b^2_{\text{max}}) \qquad \text{Logarithmic}\\
%
g_K(b_T;b_{\text{max}}) &= g_2 \ln\bigg(\frac{b_T}{b_*}\bigg) \quad g_2 = 0.84 \qquad\qquad\quad \text{AFGR/SIYY} \\
g_K(b_T;b_{\text{max}}) &= \frac{g_2 b^2_T}{2}\,; \quad g_2 = 0.13\,\text{GeV}^2 \qquad \qquad\;\,\, \text{PV17} \,. \\
\end{split}
\label{eq:all_gk}
\end{equation}
\begin{figure}[!bth]
\centerline{\includegraphics[trim = 50 30 30 0,width=9cm]{Imm/gk_mod_2_nolgm.pdf}}
\caption{Graphical representation of the different forms of the nonperturbative function $\exp(-g_K)$, listed in Eq.~(\ref{eq:all_gk}), with $b_{\rm max} = 0.6$~GeV$^{-1}$.}
\label{fig:all_gk}
\end{figure}
Thanks to their universality, all of them can be used to predict observables or be supportive in the extraction of other nonperturbative functions, in processes like $e^+e^-$ collisions, Semi-inclusive DIS and Drell-Yan processes.
\subsubsection{$M_D(b_T,z)$
\noindent
The other relevant nonperturbative function entering our convolutions is $M_D(b_T,z)$.
In Ref.~\cite{Collins:2014jpa} it has been shown that the arguments for the approximately quadratic behaviour of $g_K(b_T)$ at small $b_T$ are also valid for the function $g_{q/h}(b_T)$, and this corresponds, after an exponentiation, to a Gaussian model for the TMD-FF:
\begin{equation}
M_D(b_T,z
= \exp{\bigg(-\frac{a b^2_T}{2}\bigg)}\,.
\label{gauss_mod}
\end{equation}
This justifies the use of the following parameterization:
\begin{equation}
M_D(b_T,z)
= \exp{\bigg(-\frac{\langle p_\perp^2 \rangle b^2_T}{4 z^2}\bigg)}\,,
\label{eq:gaussian_np}
\end{equation}
corresponding in the conjugate $\bm{p}_{\perp}$-space\footnote{This space is equivalent to the $\bm{k}_{\perp}$-space used in previous sections.} to
\begin{equation}
\widetilde{M}_D(p_{\perp}) =\frac{e^{- p^2_{\perp}/\langle p_\perp^2 \rangle}}{\pi \langle p_\perp^2 \rangle}\,,
\label{eq:gaussian_pperp}
\end{equation}
where $\langle p_\perp^2 \rangle$ is the usual transverse momentum Gaussian width. Notice that, assuming it as a constant, in $p_\perp$-space there is no explicit $z$ dependence.
The commonly assumed quadratic behaviour of $g_K(b_T)$ and the Gaussian behaviour of the TMD fragmentation function can only be a valid approximation, at best, for moderate $b_T$.
Appropriate parametrizations for the nonperturbative large-$b_T$ behaviour of the TMD-FFs and of the CSS kernel need to be inferred from general principles of quantum field theory~\cite{Collins:2014jpa}, that suggest an exponentially decaying behaviour at large $b_T$. From several one-loop calculations of TMD quantities, a typical integral giving the proper $b_T$ dependence is of the form
\begin{equation}
\int d^2\bm{p}_T \frac{e^{i\bm{p}_T\cdot \bm{b}_T}}{p^2_T + m^2}\,.
\end{equation}
One possible functional parametrization that generalizes, in $\bm{b}_T$-space, the Fourier transform of the previous equation, and preserves the quadratic behaviour at small $b_T$, used in Refs.~\cite{Boglione:2017jlh,Boglione:2020auc,Boglione:2022nzq}, is the following:
\begin{equation}
M_D(b_T,z,p,m) = \frac{2^{2-p}}{\Gamma(p-1)}\,(b_T m/z)^{p-1}{K}_{p-1}(b_T m/z)\,,
\label{eq:pwrlw_mod_np}
\end{equation}
where $K_{p-1}$ is a Bessel function of the second kind, with the condition $p>1$. Its Fourier transform in $\bm{p}_{\perp}$-space is given by:
\begin{equation}
\widetilde{M}_D(p_{\perp}) = \frac{\Gamma(p)}{\pi \Gamma(p-1)}\frac{m^{2(p-1)}}{(p^2_{\perp} + m^2)^p}\,.
\label{eq:pwelw_md_pperp}
\end{equation}
We will refer to this as the Power-law model.
Notice that for a massive hadron, on the right-hand side of Eq.~(\ref{eq:gaussian_np}) one has to replace $z_p$ with $z$, in order to properly get Eq.~(\ref{eq:gaussian_pperp}). The same happens for going from Eq.~(\ref{eq:pwrlw_mod_np}) to Eq.~(\ref{eq:pwelw_md_pperp}).
\begin{figure}[!ht]
\centerline{\includegraphics[trim = 50 30 30 0,width=14cm]{Imm/unp_models_zs.pdf}}
\caption{Representation of nonperturbative hadronic models for $M_D(b_T,z)$ as a function of $b_T$ at different values of $z$: Gaussian (orange solid line), Power-Law (blue dot-dashed line) and PV17 (green dotted line) model.}
\label{fig:np_models}
\end{figure}
As mentioned above, the Pavia group~\cite{Bacchetta:2017gcc} provided a complete set of nonperturbative functions, for pions and kaons: for $M_D$, to be used together with its corresponding $g_K$ (last line of Eq.~(\ref{eq:all_gk})), they propose:
\begin{equation}
M_D(b_T,z) =\frac{g_3\, e^{-b^2_T\frac{g_3}{4z^2}} +\frac{\lambda_F}{z^2}g^2_4(1 -g_4\frac{b_T^2}{4z^2} )e^{-b^2_T\frac{g_4}{4z^2}} }{g_3 + \frac{\lambda_F}{z^2}g^2_4 }\,,
\label{PV17_md}
\end{equation}
where
\begin{equation}
g_{3,4} = N_{3,4}\frac{(z^{\beta}+\delta)(1-z)^{\gamma} }{(\hat{z}^{\beta}+\delta)(1-\hat{z})^{\gamma}} \quad %
\end{equation}
\begin{equation}
\begin{split}
\hat{z}&= 0.5; \quad N_3= 0.21 \,\text{GeV}^2; \quad N_4 = 0.13 \,\text{GeV}^2\\
\beta &= 1.65; \quad \delta = 2.28; \quad \gamma=0.14; \quad \lambda_F= 5.50 \,\text{GeV}^{-2}\,.
\end{split}
\end{equation}
The three parametrizations, discussed above, are shown in Fig.~\ref{fig:np_models}.
Finally, for the nonperturbative component of the polarizing FF, $M_D^\perp$, in the subsequent phenomenological analysis we will adopt the same functional forms as those in Eqs.~(\ref{eq:gaussian_np}) and (\ref{eq:pwrlw_mod_np}), extracting the corresponding parameters from the fit.
\section{Single-inclusive hadron production
\label{sec:1-h_prod}
As already mentioned, there is another interesting and related process relevant in this context: namely the single-inclusive production of (un)polarized spin-$1/2$ hadrons in $e^+e^-$ annihilation processes. In Ref.~\cite{DAlesio:2020wjq} a first attempt to consider this case, within a simplified phenomenological TMD scheme, was discussed.
As we will see, this case is more subtle and deserves a proper and dedicated treatment within the TMD factorization scheme.
We will present here only some relevant formulas, summarizing the kinematics and giving the expression of the cross section. We refer the reader to Refs.~\cite{Kang:2020yqw,Gamberg:2021iat}, where a complete TMD formulation of the process $e^+e^-\to h(S_{h})+X$ is discussed. It is worth noticing that the issues of proper factorization and universality for such a process have been formally addressed in a series of recent papers~\cite{Makris:2020ltr,Boglione:2020auc, Boglione:2020cwn, Boglione:2021wov}, where a detailed, and somehow complementary, approach can be found.
In this configuration, as shown in Fig.~\ref{fig:kin_1_h},
the hadron is produced with a transverse momentum $\bm{j}_{\perp}$ with respect to the thrust axis $\hat{\bm{T}}$, defined as the vector, $\hat{\bm{T}}$, which maximizes the thrust variable $T$
\begin{equation}
T = \frac{\sum_i |\bm{p}_i \cdot \hat{\bm{T}}|}{\sum_i |\bm{p}_i|}\,,
\end{equation}
\begin{figure}[!th]
\centering
\includegraphics[width=12cm]{Imm/thrust_frame4.png}
\caption{Kinematics for the process $e^+e^-\to h+X$ in the \emph{thrust-frame} configuration.}
\label{fig:kin_1_h}
\end{figure}
where $\bm{p}_i$ represent the three-momenta of the measured final-state particles. This is referred to as the \emph{thrust frame} configuration, where one plane is defined by the lepton direction and the thrust vector (Lepton plane) and the second plane by the thrust axis and the hadron momentum.
Moreover, the full phase space is divided into two hemispheres by the plane perpendicular to the thrust axis at the $e^+e^-$ interaction point.
Similarly to the previous case, the hadron has a certain energy fraction
\begin{equation}
z_{h} = \frac{2 P_{h} \cdot q}{Q^2}\,,
\end{equation}
where once again $q$ is the virtual photon four-momentum and $Q^2 = q^2 = s$.
The most important aspect is that, for this process, one hemisphere is fully inclusive, while
the single-inclusive measurement is carried out in the hemisphere that contains the thrust axis. Thus, only soft radiation which is emitted into this hemisphere will contribute to $\bm{j}_{\perp}$.
Indeed the factorized expression used in Ref.~\cite{Gamberg:2021iat} and given at next-to-leading logarithm accuracy (NLL) in \cite{Kang:2020yqw} introduces the hemisphere soft function $S_{\text{hemi}}$, that is different from the typical soft function $S$ usually defined to describe the almost back-to-back double hadron production in $e^+e^-$ collisions.
On the other hand, as demonstrated in \cite{Kang:2020yqw}, since $S_{\text{hemi}}$ at one-loop order accuracy equals $\sqrt{S}$, both the unpolarized and polarizing FFs, in the single-inclusive process, are the same FFs appearing in the double-hadron production process.
The cross section for the unpolarized hadron production, for $j_T\ll Q$, is then given by:
\begin{eqnarray}
\frac{d \sigma}{dz d^2 \bm{j}_{\perp}} &=& \frac{\sigma_0}{z^2} \sum_q e^2_q\int \frac{d b_T}{(2 \pi)} \, b_T J_0(b_T\, q_T)\,d_{q/h}(z; \bar{\mu}_b) \, U_{NG}(\bar{\mu}_b,Q) \nonumber\\
%
&\times & M_{D}(b_c(b_T),z) \, \exp\Bigg\{-g_K(b_c(b_T);b_{\text{max}})\ln{\bigg(\frac{Q z}{M_{h}}\bigg)}\Bigg\} \nonumber\\
%
&\times& \exp\Bigg\{ \Tilde{K}(b_*;\bar{\mu}_b)\ln{\frac{Q}{\bar{\mu}_b}} + \int^Q_{\bar{\mu}_b} \frac{d\mu'}{\mu'} \Bigg[ \gamma_D(g(\mu'),1) - \gamma_K(g(\mu'))\ln{\frac{Q}{{\mu'}}} \Bigg] \Bigg\}\,,
\label{eq:1-h_unp}
\end{eqnarray}
where $z$ is the hadron light-cone momentum fraction, related to $z_{h}$ as shown in Eq.~(\ref{eq:en_fract}), and
\begin{equation}
\sigma_0 = \frac{4 N_c \pi \alpha^2_{em}}{3 Q^2}\,.
\end{equation}
Here we find the same elements already discussed in Section~\ref{sec:CSS_RG_eqs}: $d_{q/h}$ is the integrated unpolarized FF, $M_D$ and $g_K$ are the nonperturbative model functions.
Similarly, the cross section for transversely polarized hadron production has the following form:
\begin{eqnarray}
\frac{d \Delta\sigma}{dz d^2\bm{j}_{\perp}}& =&\sin(\phi_{S_{h}} - \phi_j)\,\frac{\sigma_0}{z^2} \sum_q e^2_q\int \frac{d b_T}{(2 \pi)} \, b^2_T J_1(b_T\, q_T) {D}^{\perp (1)}_{1T}(z,\bar{\mu}_b) \, U_{NG}(\bar{\mu}_b,Q) \nonumber\\
%
&\times& M^{\perp}_{D}(b_c(b_T),z) \, \exp\Bigg\{-g_K(b_c(b_T);b_{\text{max}})\ln{\bigg(\frac{Q z}{M_{h}}\bigg)}\Bigg\} \nonumber\\
%
&\times &\exp\Bigg\{ \Tilde{K}(b_*;\bar{\mu}_b)\ln{\frac{Q}{\bar{\mu}_b}} + \int^Q_{\bar{\mu}_b} \frac{d\mu'}{\mu'} \Bigg[ \gamma_D(g(\mu'),1) - \gamma_K(g(\mu'))\ln{\frac{Q}{{\mu'}}} \Bigg] \Bigg\}\,,
\label{eq:1-h_pol}
\end{eqnarray}
where ${D}^{\perp (1)}_{1T}$, see Eq.~(\ref{eq:firstmoment_z}), is the small-${b}_T$ limit of the first moment of the polarizing fragmentation function.
Since soft radiation is restricted to only one hemisphere, the cross section is a non-global observable. The factorization formulas for this kind of observables have been derived within an effective field theory framework \cite{Becher:2015hka,Becher:2016mmh,Becher:2016omr,Becher:2017nof}, where a multi-Wilson-line structure \cite{Caron-Huot:2015bja,Nagy:2016pwq,Nagy:2017ggp} is the key ingredient to capture the non-linear QCD evolution effects from the so-called \emph{non-global logarithms}. %
For this reason in both cross sections, Eqs.~(\ref{eq:1-h_unp}) and (\ref{eq:1-h_pol}), we have introduced the function $ U_{NG}$ (see Ref. \cite{Kang:2020yqw}), which accounts for the effects of such non-global effects.
In the following we will use the parametrization given in Ref.~\cite{Dasgupta:2001sh}
\begin{equation}
U_{NG}(\bar{\mu}_b,Q) = \exp{\bigg[ -C_A C_F \frac{\pi^2}{3} u^2 \frac{1+ (au)^2}{1+(bu^c)} \bigg]}\,,
\end{equation}
with $a = 0.85 \, C_A$, $b = 0.86 \, C_A$, $c = 1.33$ and
\begin{equation}
u = \frac{1}{4 \pi \beta_0} \ln{\bigg[\frac{\alpha_s(\bar{\mu}_b)}{\alpha_s(Q)} \bigg]}\,,
\end{equation}
where $\beta_0 = (11 C_A - 4 T_F n_f)/12 \pi$, with $T_F = 1/2$, $C_A=3$, $C_F=4/3$ and $n_f$ is the number of the active flavours.
In addition, when the polarization is measured along the axis $\hat{\bm{n}}= \hat{\bm{T}} \times \hat{\bm{P}}_{h}$, the spin and transverse momentum azimuthal angles are such that $\sin(\phi_{S_{h}} - \phi_j) = 1$. Finally, the expression of the transverse polarization can be given as:
\begin{equation}
\mathcal{P}(z,j_{\perp}) = \frac{{d\Delta \sigma}/{dz d^2\bm{j}_{\perp}}}{{d \sigma}/{dz d^2\bm{j}_{\perp}}}\,,
\label{eq:1_h_polarization}
\end{equation}
that will be used to fit the Belle data. This will allows us to extract the first moment of the polarizing fragmentation function and its nonperturbative model function.
\section{Phenomenological analysis}
\label{sec:analysis}
We can now proceed with the analysis of Belle data \cite{Guan:2018ckx} for the transverse $\Lambda$ polarization measured in $e^+e^-$ collisions, employing the approach presented in the previous sections. Two data sets are available: one where the $\Lambda$ particle is produced almost back-to-back with respect to a light unpolarized hadron, that we will refer to as double-hadron production (2-h) data set, and one where the $\Lambda$ transverse momentum is measured with respect to the thrust axis, the single-inclusive production (1-h) data set.
We will start considering only the double-hadron production data set and present the corresponding results. In a second phase, we will include in the study also the single-inclusive hadron production case.
\subsection{Double-hadron production data fit}
In this section, by employing Eqs.~(\ref{pol_ratio}),~(\ref{int_qtmax_1}),~(\ref{int_qtmax_2}),~(\ref{B0_full}) and~(\ref{B1_full}), we present the analysis of the polarization of $\Lambda$/$\bar{\Lambda}$ hyperons produced with a light hadron, $\pi^{\pm}$ or $K^{\pm}$, measured at $\sqrt{s} = 10.58\,\text{GeV}$. The 128 data points are given as a function of $z_{\Lambda}$ and $z_{\pi/K}$, the energy fractions of the $\Lambda$/$\bar{\Lambda}$ and $\pi/K$ particles. For the current analysis we start imposing a cut on large values of the light-hadron energy fractions, $z_{\pi/K}<0.5$, keeping only 96 data points. We will come back to this point in the following.
Notice that here we consider the transverse polarization for inclusive $\Lambda$ particles, namely those directly produced from $q\bar{q}$ fragmentation and those indirectly produced from strong decays of heavier strange baryons.
The purpose of this analysis is to extract ${D}^{\perp (1)}_{1T,\, \Lambda/q}$ and $M^{\bot}_{D, \Lambda}$, the first moment and the nonperturbative component of the $\Lambda$ polarizing FF.
We will use the following expression to parametrize the $z$ dependence of ${D}^{\perp (1)}_{1T,\, \Lambda/q}$:
\begin{equation}
{D}^{\perp (1)}_{1T,\, \Lambda/q}(z;\mu_b)=\mathcal{N}^p_q(z)\,d_{q/\Lambda}(z;\mu_b)\,,
\label{eq:first_mom1}
\end{equation}
with, as adopted and motivated in Ref.~\cite{DAlesio:2020wjq}, $q = u,\, d,\, s$ and sea, and where
$\mathcal{N}^{\rm p}_q(z)$ (the apex refers to the polarizing FF) is parametrized as:
\begin{equation}
\mathcal{N}^{\rm p}_q(z) = N_q z^{a_q}(1-z)^{b_q}\frac{(a_q +b_q )^{(a_q +b_q )}}{a_q^{a_q}b_q^{b_q}}\,.
\label{eq:first_mom2}
\end{equation}
$d_{q/\Lambda}$ is the collinear unpolarized $\Lambda$ fragmentation function
for which we employ
the AKK08 set~\cite{Albino:2008fy}. This parametrization is given for $\Lambda + \bar{\Lambda}$ and adopts the longitudinal momentum fraction, $z_p$, as scaling variable. In order to separate the two contributions we assume
\begin{equation}
d_{q/\bar{\Lambda}}(z_p) = d_{\bar{q}/\Lambda}(z_p) = (1- z_p)\, d_{q/\Lambda}(z_p) \,.
\end{equation}
This is a common way to take into account the expected difference between the quark and antiquark FF with a suppressed sea at large $z_p$ with respect to the valence component. Other similar choices have a very little impact on the fit.
Concerning the nonperturbative function $M^{\bot}_{D, \Lambda}$ we employ two different functional forms. The first one is the Gaussian model, analogous to Eq.~(\ref{eq:gaussian_np}):
\begin{equation}
M^{\bot}_{D, \Lambda}(b_T,z) = \exp{\bigg(-\frac{\langle p_\perp^2 \rangle_\text{p} b^2_T}{4 z^2_p}\bigg)}\,,
\end{equation}
where
$\langle p_\perp^2 \rangle_{\text{p}} $ is the Gaussian width, a free parameter that we extract from the fit. The second one is the Power-Law model, Eq.~(\ref{eq:pwrlw_mod_np}):
\begin{equation}
M^{\bot}_{D, \Lambda}(b_T,z) = \frac{2^{2-p}}{\Gamma(p-1)}\,(b_T m/z_p)^{p-1}{K}_{p-1}(b_T m/z_p)\,,
\end{equation}
where we will extract the values of $p$ and $m$ (with the condition $p>1$).
Regarding the collinear FFs of the unpolarized light hadrons, $\pi$ and $K$, we adopt the DSS07 set~\cite{deFlorian:2007aj}, while for $M_D$ we assume a Gaussian model, compatible with previous extractions, with $\langle p_\perp^2 \rangle = 0.2 \, \text{GeV}^2$ \cite{Anselmino:2005nn}.
We will also consider the PV17 model with its proper nonperturbative functions.
For what concerns the $\Lambda$ unpolarized FFs, for $M_D$ we use either a Gaussian model, with the same width as for the light hadrons, or a Power-Law model, Eq.~(\ref{eq:pwrlw_mod_np}), with the parameters values $p=2$ and $m=1\,\text{GeV}$. These are represented in Fig.~\ref{fig:np_models}.
Notice that all the conversions among the different scaling variables $(z, z_h, z_p)$ involved, Eqs.~(\ref{eq:light_con_def}),~(\ref{eq:en_fract}) and~(\ref{eq:long_fract}), are properly taken into account.
For the $g_K$ function, we use the expressions presented in Section \ref{sec:CSS_RG_eqs}, see Eq. (\ref{eq:all_gk}) and Fig. \ref{fig:all_gk}.
Concerning the $\bm{b}_*$-prescription, we use:
\begin{equation}
b_{\text{min}} = 2e^{-\gamma_E}/Q \,; \qquad b_{\text{max}} = 0.6\,\text{GeV}^{-1}\,,
\nonumber
\end{equation}
with $Q = 10.58 \,\text{GeV}$. %
Bearing in mind that the larger is the value of $b_{\text{max}}$, the smaller is the value assumed by $\mu_b$, we chose the value of $b_{\text{max}}$ to be as large as possible, taking into account
at the same time that the AKK set is defined for scales $\ge$ 1 GeV. %
Lastly, for the integration in Eqs.~(\ref{int_qtmax_1}) and~(\ref{int_qtmax_2}), we use $q_{T_{\text{max}}} = 0.25\,Q $, exploring also the impact of different values of $q_{T_{\text{max}}}/Q$.
To perform the phenomenological analysis we use \emph{iMINUIT}~\cite{iminuit} as a minimizer for the $\chi^2$ function, and for the Fourier transforms we employ the \emph{Fast Bessel Transform} algorithm presented in \cite{DBLP:journals/cphysics/KangPST21}.
\subsection{Fit results}
Concerning the first moment of the polarizing FF, Eqs.~(\ref{eq:first_mom1}) and (\ref{eq:first_mom2}), we adopt the same parameter choice considered in Ref.~\cite{DAlesio:2020wjq}, that is:
\begin{equation}
N_u,\quad N_d,\quad N_s,\quad N_{\text{sea}},\quad a_s,\quad b_u,\quad b_{\text{sea}} \,,
\end{equation}
with all other $a$ and $b$ parameters set to zero. This indeed ensures, once again, a good quality of the fit, keeping the number of parameters under control.
Regarding the nonperturbative functions, we have explored various combinations of them, for a total of 36 fits, plus the PV17 set. %
We have also considered different initial values of the $p$ parameter
of the Power-Law model, noticing that this leads to different chi-square minimum values. This means that we have 8 or 9 free parameters depending on whether we use the Gaussian or the Power-Law model for the polarizing nonperturbative function.
%
The best results for the double-hadron production fit are reported in Tab.~\ref{tab:best_fit}, adopting different combinations of the $g_K$, $M_D$ and $M_D^\perp$ parametrizations.
\begin{table}[h!]
\centering
\begin{tabular}{|c c c c c|}
\hline
$M_D^\perp$ & $M_D$ & $g_K$ & $\chi^2_{\text{dof}}\,$(2-h) & $\chi^2_{\text{dof}}\,$(2-h + 1-h) \\ [.7ex]
\hline
Gaussian & Power-Law & Logarithmic & 1.192 & 2.9 \\ [.7ex]
Power-Law & Power-Law & Logarithmic & 1.21 & 2.43 \\ [.7ex]
\hline
Gaussian & Power-Law & PV17 & 1.198 & 3.159 \\ [.7ex]
\hline
\end{tabular}
\caption{Values of the $\chi^2_{\text{dof}}$ obtained fitting the double hadron production data set only (column ``2-h''), and those obtained for the combined fit (column ``2-h + 1-h''). }
\label{tab:best_fit}
\end{table}
As reported in Tab.~\ref{tab:parameters}, the parameters values extracted employing the Gaussian or Power-Law models are totally consistent and, similarly, the two polarizing nonperturbative models $M^{\perp}_D$ are compatible, as shown in Fig.~\ref{fig:two_pol_model}. As in the case of the previous extraction~\cite{DAlesio:2020wjq}, we find that only the first moment of the up quark is positive, confirming, moreover, that the contributions to the $\Lambda$ transverse polarization given by the up and the down quarks are opposite in sign. %
\begin{table}[h!]
\centering
\begin{tabular}{|ccc|c|}
\hline
Parameters & Gaussian & Power-Law & Gaussian (PV17) \\ [.7ex]
\hline
$N_u$ & $0.093^{-0.052}_{+0.092}$ & $0.100^{-0.054}_{+0.095}$ & $0.168^{-0.007}_{+0.008}$ \\ [.7ex]
$N_d$ & $-0.100^{-0.036}_{+0.035}$ & $-0.107^{-0.041}_{+0.036}$ & $-0.138^{-0.011}_{+0.012}$ \\ [.7ex]
$N_s$ & $-0.117^{-0.09}_{+0.059}$ & $-0.115^{-0.089}_{+0.057}$& $-0.161^{-0.03}_{+0.033}$ \\ [.7ex]
$N_{sea}$ & $-0.055^{-0.058}_{+0.033}$ & $-0.058^{-0.062}_{+0.034}$& $-0.104^{-0.008}_{+0.008}$ \\ [.7ex]
$a_s$ & $2.19^{-0.83}_{+1.07}$ & $2.12^{-1.0}_{+1.5}$ & $2.19^{-0.32}_{+0.28}$\\ [.7ex]
$b_u$ & $3.5^{-2.2}_{+2.8}$ & $3.5^{-1.9}_{+2.8}$ & $4.02^{-0.26}_{+0.28}$ \\ [.7ex]
$b_{sea}$ & $2.3^{-1.8}_{+2.5}$ & $2.3^{-1.9}_{+2.7}$ & $2.91^{-0.16}_{+0.18}$ \\ [.7ex]
$\langle p_\perp^2 \rangle_p$ & $0.066^{-0.031}_{+0.039}$ && $0.103^{-0.014}_{+0.015}$ \\ [.7ex]
$p$ & & $3.0^{-1.4}_{+2.5}$ &\\ [.7ex]
$m$ & & $0.35^{-0.22}_{+0.3}$& \\ [.7ex]
\hline
\end{tabular}
\caption{Best parameter values for the first moment of the polarizing FF and for the two nonperturbative functions employed to fit the double-hadron data set.}
\label{tab:parameters}
\end{table}
\begin{figure}[!h]
\centerline{\includegraphics[trim = 50 30 30 0,width=7cm]{Imm/pol_models_2.pdf}}
\caption{Nonperturbative functions extracted from the double-hadron data fit: Gaussian (orange solid line), Power-Law (blue dash-dotted line) and PV17 Gaussian (green dotted line) model.}
\label{fig:two_pol_model}
\end{figure}
In Fig.~\ref{fig:Lh_gauss} we show the estimates of the transverse $\Lambda$ polarization, produced in association with a light-hadron, compared against Belle data~\cite{Guan:2018ckx}, adopting the parameters extracted with the Gaussian model. The shaded areas, corresponding to a $2\sigma$ uncertainty, are computed according to the procedure explained in the Appendix of Ref.~\cite{Anselmino:2008sga}.
\begin{figure}[t]
{\includegraphics[trim = 0 50 0 140,clip,width=7.5cm]{Imm/Lb_pi_gauss.pdf}}
{\includegraphics[trim = 0 50 0 140,clip,width=7.5cm]{Imm/Lbbar_pi_gauss.pdf}}
\newline
{\includegraphics[trim = 0 50 0 140,clip,width=7.5cm]{Imm/Lb_k_gauss.pdf}}
{\includegraphics[trim = 0 50 0 140,clip,width=7.5cm]{Imm/Lbbar_k_gauss.pdf}}
\caption{Best-fit estimates (Gaussian model with the parameters of Tab.~\ref{tab:parameters}) of the transverse polarization for $\Lambda$, $\bar{\Lambda}$ in $e^+e^-\to \Lambda (\bar{\Lambda}) h +X $, for $\Lambda \pi^{\pm}$ (a), $\bar{\Lambda} \pi^{\pm}$ (b), $\Lambda K^{\pm}$ (c), $\bar{\Lambda} K^{\pm}$ (d), as a function of $z_h$ ($h= \pi, K$) for different $z_{\Lambda}$ bins. Data are from Belle \cite{Guan:2018ckx}. The statistical uncertainty bands, at $2\sigma$ level, are also shown. Data for $z_{\pi, K} > 0.5$ are not included in the fit. The results of the fit of Ref.~\cite{DAlesio:2020wjq} (dotted line) are also shown.}
\label{fig:Lh_gauss}
\end{figure}
As in the previous fit~\cite{DAlesio:2020wjq}, where the full TMD machinery was not employed, $\Lambda K^\pm$ data with $z_{K}>0.5$ cannot be described. This can due to different reasons, like the large uncertainty on the last $z_{\pi,K}$ bin and/or the uncertainty affecting the unpolarized FFs in this region.
In Tab.~\ref{tab:chi_range} we report the $\chi^2_{\text{dof}}$ ranges for different fits.
In this comparison we combine different nonperturbative functions to fit data. At fixed $g_K$ function and adopting four combinations of the polarizing and unpolarized FFs with Gaussian and Power-law model, the $\chi^2$ goes from a minimum to a maximum value as reported in Tab~\ref{tab:chi_range}.
It is worth noticing that the extractions are consistent and stable when we employ the same $g_K$.
Moreover, we see that the best fits are found when we make use of the \emph{Logarithmic} $g_K$ function (quite similar to the PV17 model), while the \emph{Quadratic} and \emph{AFGR} functional forms give similar results with worse $\chi^2_{\text{dof}}$. Finally, the worst $\chi^2_{\text{dof}}$s are obtained with the \emph{BLNY} functional form.
\begin{table}[ht!]
\centering
\begin{tabular}{|c c| }
\hline
$g_K$ & $\chi^2_{\text{dof}}$ range \\ [.7ex]
\hline
Logarithmic & 1.192 - 1.287 \\ [.7ex]
Quadratic & 1.4 - 1.472 \\[.7ex]
AFGR & 1.474 - 1.514 \\[.7ex]
BLNY & 1.67 - 1.783 \\[.7ex]
PV17 & 1.198 - 1.524 \\[.7ex]
\hline
\end{tabular}
\caption{Range of the $\chi^2_{\text{dof}}$ values for different nonperturbative $g_K$ functions.}
\label{tab:chi_range}
\end{table}
In Fig.~\ref{fig:qt_impact} we show the impact of choosing different values of $q_{T_{\text{max}}}/Q$ on the quality of the fit, obtained using the Power-Law and the Gaussian models. In general, the Gaussian model gives smaller $\chi^2_{\text{dof}}$ values. Both models reach their minimum $\chi^2_{\text{dof}}$ value around $q_{T_{\text{max}}}/Q = 0.22$, while the Gaussian model with the PV17 parametrization reaches its minimum $\chi^2_{\text{dof}}$ value at $q_{T_{\text{max}}}/Q \simeq 0.27$. %
The growth of the $\chi^2_{\text{dof}}$, as $q_{T_{\text{max}}}/Q$ increases, can be explained considering that we are gradually going out of the validity region of the TMD factorization. Meanwhile, the growth for lower values of $q_{T_{\text{max}}}/Q$ can be attributed to the fact that the smaller is $q_{T_{\text{max}}}/Q$, the larger is the distance between the nodes of the Bessel functions. Hence, the \emph{Fast Bessel Transform} algorithm~\cite{DBLP:journals/cphysics/KangPST21} is not anymore able to sample sufficiently well the integrand, whose Fourier transform is to be computed.
\begin{figure}[!t]
\centerline{\includegraphics[trim = 50 30 30 0,width=12cm]{Imm/qt_impact2.pdf}}
\caption{$\chi^2_{\text{dof}}$ values for the fits obtained adopting the Power-Law (blue short-dashed line), the Gaussian (red dotted line), both with the \emph{Logarithmic} $g_K$ function, and the PV17 model (green long-dashed line) with the PV17 $g_K$ functional form, as a function of $q_{T_{\text{max}}}/Q$.}
\label{fig:qt_impact}
\end{figure}
\subsection{Combined fit: double- and single-inclusive hadron production data}
We now discuss the role of single-inclusive polarization data in extracting the polarizing FF, in view of a combined fit of both data sets.
We start checking whether by adopting the results from the 2-h fit one is able to describe the single-inclusive hadron data. This data set is given as a function of $p_{\perp}$, the transverse momentum of the $\Lambda/\bar{\Lambda}$ particle with respect to the thrust axis, that coincides with $j_{\perp}$ in Eqs.~(\ref{eq:1-h_unp}) and (\ref{eq:1-h_pol}), for different bins of the energy fraction $z_{\Lambda}$~\cite{Guan:2018ckx}.
\begin{figure}[h!]
\begin{center}
\includegraphics[trim = 0 50 0 140,clip,width=12.5cm]{Imm/LB_thrust_0.25_chi_1.192_4.pdf}
\end{center}
\caption{Estimates of the transverse polarization for single-inclusive $\Lambda/\bar\Lambda$ production compared with Belle data. The results are obtained using Eqs.~(\ref{eq:1-h_unp}),~(\ref{eq:1-h_pol}) and~(\ref{eq:1_h_polarization}) and the parameter values of Tab.~\ref{tab:parameters}, for the Gaussian model (with the logarithmic functional form for $g_K$) and the 2-h fit.}
\label{fig:pred_1h}
\end{figure}
As shown in Fig.~\ref{fig:pred_1h} the Gaussian model from the 2-h fit (but the same happens also for the Power-Law model) cannot describe the pattern and the size of the polarization.
As a further step, we perform a combined fit including {\it both} the single-inclusive (1-h) and the 2-h hadron data sets. In such a case we obtain a large $\chi^2_{\text{dof}}$, as shown in the last column of Tab.~\ref{tab:best_fit}. The main outcome is that while the single-inclusive data can be described better than in the previous case,
the agreement for the associated light-hadron production case (in particular for pions) is spoiled. This is the main reason for the increasing of the $\chi^2$ (see also Tab.~\ref{tab:chi_square_point}, where we show the $\chi^2$ for data points).
Even exploring all other different combinations of nonperturbative models, as we have done in the double-hadron production section, we keep getting $ \chi^2_{\text{dof}}$ values ranging from $2.4$ to $5.4$.
Since the first moment of the polarizing FF is a collinear quantity, it is expected to be the same in both the double-hadron and the single-inclusive cross sections. Therefore, the fact that the two data sets cannot be fitted simultaneously could suggest that these processes cannot be described within the same factorization theorem and/or by the same nonperturbative function $M^{\perp}_D$ (see also Refs.~\cite{Boglione:2020cwn,Boglione:2020auc,Boglione:2021wov}). An attempt to explore this issue will be discussed in the following section.
\subsubsection{Different nonperturbative functions $M_D^\perp$ for the 2-h and 1-h cases }
In order to investigate the possible reasons why the combined fit is not satisfactory and why the parameters extracted in the double-hadron fit cannot describe the single-inclusive polarization data, we try to fit both data sets using the same parametrizations for the first moment of the polarizing FF and the same functional form for $M^{\perp}_D$, but with two different sets of parameters for the latter.
Notice that this has to be considered as an attempt to explore the compatibility of the two data sets with the use of a unique and universal nonperturbative function.
More precisely, for the Gaussian model we fit two different Gaussian widths, while in the case of the Power-Law model we fit two different $(p, m)$ parameter pairs, one for the 2-h data set and one for the 1-h data set. Concerning $g_K$ and the unpolarized nonperturbative functions we use the same functional forms as in Tab.~\ref{tab:best_fit}. As reported in Tab.~\ref{tab:double_mod_prm}, with this approach we can find much better $\chi^2_{\text{dof}}$'s with respect to those reported in the last column of Tab.~\ref{tab:best_fit}. Indeed, we obtain a $\chi^2_{\text{dof}}$ = 1.801 and 1.565 respectively for the Gaussian and the Power-Law models.
\begin{table}[!h]
\centering
\begin{tabular}{|c c c | c c c|}
\hline
\multicolumn{3}{|c|}{Gaussian} & \multicolumn{3}{c|}{Power-Law} \\ [.7ex]
\hline
\multicolumn{3}{|c|}{$\chi^2_{\text{dof}} = 1.801$} & \multicolumn{3}{c|}{$\chi^2_{\text{dof}} = 1.565$} \\ [1.ex]
\hline
& 2-h & 1-h & & 2-h & 1-h \\ [.7ex]
\hline
\multirow{2}{*}{$\langle p_\perp^2 \rangle_p$} & \multirow{2}{*}{$0.04^{-0.02}_{+0.03}$} & \multirow{2}{*}{$0.2^{-0.01}_{+0}$} & $p$ & $1.352^{-0.055}_{+0.068}$ & $1.623^{-0.011}_{+0.011}$ \\ [1.2ex]
& & & $m$ & $0.151^{-0.024}_{+0.026}$ & $0.48^{-0.005}_{+0.005}$ \\ [.9ex]
\hline
\end{tabular}
\caption{Values of the best fit parameters for the nonperturbative function, $M_D^\perp$, using two independent sets of parameters for the 2-h and 1-h data sets, for the Gaussian and Power-Law models.}
\label{tab:double_mod_prm}
\end{table}
\begin{figure}[!h]
\begin{center}
\hspace{0.cm}
\includegraphics[width=12cm]{Imm/L_jet.pdf}
\caption{Comparison of our fit estimates for the $\Lambda/\bar\Lambda$ single-inclusive polarization against Belle data, adopting the double parametrization for the Power-Law model.}
\label{fig:L1h_pwrlw}
\end{center}
\end{figure}
\begin{center}
\begin{figure}[!th]
\includegraphics[trim = 0 0 0 0,clip,width=9cm]{Imm/firsm_up.pdf}\hspace{-2.7cm}
\includegraphics[trim = 0 0 0 0,clip,width=9cm]{Imm/firsm_down.pdf}
\includegraphics[trim = 0 0 0 40,clip,width=9cm]{Imm/firsm_strange.pdf}\hspace{-2.5cm}
\includegraphics[trim = 0 0 0 40,clip,width=9cm]{Imm/firsm_sea.pdf}
\caption{First moments of the polarizing FFs, for the up (a), down (b), strange (c) and sea (d) quarks, as obtained from the combined fit (red dashed lines) and the 2-h fit (blue dot-dashed lines). The corresponding statistical uncertainty bands are also shown. }
\label{fig:first_mom}
\end{figure}
\end{center}
Focusing on the results obtained with the Power-Law model, that gives a better $\chi^2_{\text{dof}}$, in Fig.~\ref{fig:L1h_pwrlw} we can see how the estimates for the single-inclusive polarization describe the experimental data much better than those in Fig.~\ref{fig:pred_1h} (without spoiling the agreement with the 2-h data set, see below). The two different pairs of $(p, m)$ are given in Tab.~\ref{tab:double_mod_prm}.
In Fig.~\ref{fig:first_mom} we also show a comparison of the first moments of the polarizing FFs as extracted in the 2-h fit and in the combined fit, adopting the double model parametrization for $M_D^\perp$. As one can see the two extractions seem not to be compatible, at least within the uncertainty bands (the full theoretical uncertainty bands, very difficult to estimate, might be larger than the statistical ones). On the other hand, as already stressed, the combined fit requires further insight for what concerns the single-inclusive hadron production. We also notice that the qualitative behaviour and the size of the first moments are comparable with those extracted in Ref.~\cite{DAlesio:2020wjq} .
In Tab.~\ref{tab:chi_square_point} we try to summarize the main findings of our phenomenological analysis, by reporting the $\chi^2$ per data points for each case considered, and still focusing on the Power-Law model.
Starting with the 2-h fit (second column), we see that, besides a small tension in the $\Lambda K$ data set, the overall $\chi^2_{\rm point}$ is extremely good (as already discussed above).
Moving to the results for the combined fit (2-h+1-h), adopting a unique parametrization for $M_D^\perp$ (third column), we see that the description of the $\Lambda\pi$ data set is completely spoiled. Moreover, the $\chi^2_{\rm point}$ for the inclusive data is also extremely high and the overall $\chi^2_{\rm point}$ doubles its value with respect to the 2-h fit. Finally, the combined fit, but with two separate parametrizations for $M_D^\perp$ (last column), shows that for this scenario the 2-h data sets can be described at the same level of accuracy as in the 2-h fit. More important, the $\chi^2_{\rm point}$ for the inclusive data set reduces significantly, leading to an improvement in the overall description ($\chi^2_{\rm point}=1.43$)
\begin{table}[!h]
\centering
\begin{tabular}{| c || c c || c c|| c c|}
\hline
&\multicolumn{2}{|c||}{2-h} & \multicolumn{2}{c||}{2-h+1-h (one param.)} & \multicolumn{2}{c|}{2-h+1-h (two param.)} \\ [.7ex]
\hline
& Data &$\chi^2_{\rm point}$ & Data &$\chi^2_{\rm point}$ & Data &$\chi^2_{\rm point}$ \\ [1.ex]
$\Lambda \pi$ & 48 & 0.79 & 48 & 2.2 & 48 & 0.8 \\ [1.ex]
$\Lambda K$ & 48 & 1.4 & 48 & 1.7 & 48 & 1.4 \\ [1.ex]
$\Lambda $& & & 31 & 3.1 & 31 & 2.4 \\ [1.ex]
TOT. & 96 & 1.1 & 127 & 2.3 & 127 & 1.43 \\ [1.ex]
\hline
\end{tabular}
\caption{$\chi^2$ per data points for the 2-h, 2-h+1-h (single parametrization) and 2-h+1-h (double parametrization) fits, adopting the Power-Law model.}
\label{tab:chi_square_point}
\end{table}
Let us move to the results obtained for the soft nonperturbative functions, $M_D^\perp$.
Comparing the two Power-Law models as extracted from the 2-h and 1-h data fits in Fig.~\ref{fig:double_mod_comparison}, we can see that they have almost the same behaviour and size at small $b_T$, as expected since the two fragmentation functions should be the same in the collinear limit, but they differ significantly as $b_T$ increases: the model related to the 2-h data set, blue solid line, is wider than the model related to the 1-h data set, orange dash-dotted line.
This corresponds in $\bm{p}_{\perp}$-space to a similar behaviour at large $p_{\perp}$ values, Fig.~\ref{fig:double_mod_comp_pt}, and a sizeably different one at small $p_{\perp}$.
\begin{figure}[!th]
\begin{center}
\hspace{0.cm}
\includegraphics[trim = 20 40 0 100,clip,width=12cm]{Imm/double_model_pwrlw2.pdf}
\caption{Results within the Power-Law model fit for the two $M_D^\perp$ parametrizations in $b_T$-space: 2-h (blue solid line) and 1-h (orange dash-dotted line) case. }
\label{fig:double_mod_comparison}
\end{center}
\end{figure}
\begin{figure}[!th]
\centerline{\includegraphics[trim = 0 0 0 0,width=8cm]{Imm/power_lw_pt.pdf}}
\caption{Results within the Power-Law model fit for the two $M_D^\perp$ parametrizations in $p_{\perp}$-space: 2-h (blue solid line) and 1-h (orange dash-dotted line) case. The functional form is given in Eq.~(\ref{eq:powerlaw_apx_p}).}
\label{fig:double_mod_comp_pt}
\end{figure}
These findings suggest that the two models (same functional form but different parameters) could represent effective convolutions involving two distinct soft factors and that the effects of the recoil against the emission of soft gluons might be different in the two cross sections~\cite{Boglione:2021wov}.
Notice that the fragmentation functions of the single-inclusive hadron production, in the factorization scheme presented in Ref.~\cite{Kang:2020yqw}, coincide with those in the double-hadron production at one-loop order. Indeed, only in this case the hemisphere soft factor, $S_{\text{hemi}}$, corresponds to the soft factor, $\sqrt{S}$, convoluted with each one of the fragmentation functions in the double-hadron production cross sections.
Hence in future analyses, in order to have a more consistent and robust combined fit without using, as an ansatz, two different sets of parameters, one should try to calculate and employ the soft factor $S_{\text{hemi}}$ beyond the one-loop order. An alternative strategy would be to exploit the cross sections as formulated in Ref.~\cite{Makris:2020ltr} within an effective theory framework, or in Refs.~\cite{Boglione:2020cwn,Boglione:2021wov}, where the derivation is based on a CSS approach.
\subsection{Predictions}
We can now try to check explicitly the consistency of the whole approach as well as the role of the TMD evolution by considering another set of data, namely those from the OPAL Collaboration~\cite{OPAL:1997oem}, collected at much larger energy, $\sqrt s =~M_Z$. This data set refers to the single-inclusive $\Lambda$ production, integrated over almost the entire $z_h$ range [0.15-1]. Even if they have large errors, this is the only other available data set for this observable and it is therefore worth studying the impact on our findings.
In Fig.~\ref{fig:opal} we show a series of predictions obtained adopting the different models discussed in the previous sections: all of them are almost compatible with data, within the large error bars, with some differences that deserve several comments.
\begin{figure}[!bt]
\centerline{
\includegraphics[trim = 50 30 30 0,width=7cm]{Imm/opal_compare_gpm_2h_bands.pdf}
\includegraphics[trim = 50 30 30 0,width=7cm]{Imm/opal_compare_gpm_1h_bands2_fit.pdf}
}
\caption{Predictions for the transverse $\Lambda$ polarization in $e^+e^-\to \Lambda\, X$ as a function of $p_\perp$ at $\sqrt s = M_Z$, integrated over $z_h$ in the range $[0.15,1]$, against OPAL data~\cite{OPAL:1997oem}. Left panel, 2-h fit: Power-Law (blu dot-dashed line), Gaussian (red solid line) model and GPM results~\cite{DAlesio:2020wjq} (dashed green line) from double-hadron fits. Right panel, combined fits within the Power-Law parametrization: double-model fit (blue solid line), one-model fit (violet dotted line) and global analysis (Belle + OPAL data) within the double-model fit (orange dot-dashed line). We also show the results within the GPM approach~\cite{DAlesio:2020wjq} (green dashed line). For some cases uncertainty bands are shown.}
\label{fig:opal}
\end{figure}
Let us start from the 2-h fit extraction (left panel). Both the Power-law (blue dot-dashed line) and the Gaussian (red solid line) models are not able to describe the lowest $p_\perp$ data, while they work pretty well above 0.4~GeV. This behaviour shows the same features of the description of the single-inclusive Belle data adopting the 2-h parametrization, see Fig.~\ref{fig:pred_1h}. On the other hand, OPAL data are integrated over a single, much larger $z_h$ bin and this somehow reduces the discrepancies with the theoretical estimates.
For completeness we also show the predictions from the analysis performed in Ref.~\cite{DAlesio:2020wjq}, where a TMD scheme with a simple Gaussian parametrization at fixed scale was adopted (green dashed line, labelled GPM for simplicity). The main difference with respect to the results obtained in the present analysis is the strong suppression, starting already at $p_\perp \simeq 0.5$~GeV. We will come back to this point below.
In the right panel we present the corresponding predictions obtained from the combined fit. In such a case we focus on the Power-Law model that gives the best $\chi^2$ values.
Both extractions, the one based on the single-model parametrization (red dotted line) and the one coming from the double-parametrization fit, with the parameter set for $M_D^\perp$ from the single-inclusive hadron production (blue solid line), are able to describe the data in size and sign pretty well.
On the other hand, one has to recall, see Tab.~\ref{tab:chi_square_point}, that the single-model parametrization for the combined fit gives very large $\chi^2$ for the associated pion production data set.
Even if with some caution, we could observe how the flattening behaviour in $p_\perp$ of these predictions reproduce the \emph{puzzling} plateau of the transverse $\Lambda$ polarization observed in unpolarized hadron-hadron collisions as a function of the transverse momentum of the $\Lambda$ with respect to the direction of the incoming hadrons.
The corresponding GPM results (recall that for the single-inclusive case a simplified phenomenological TMD scheme was adopted), while qualitatively good, show once again their distinguishing Gaussian suppression at large $p_\perp$. The main difference with the corresponding GPM curve shown in the left panel is that the combined fit, even in the GPM approach, gives a much larger Gaussian width and the suppression is somehow shifted at larger $p_\perp$.
It is worth noticing that the overall common good agreement, within the large error bars, of the two approaches, CSS framework (this analysis) and GPM results from Ref.~\cite{DAlesio:2020wjq}, is due to the fact that both of them are controlled by the collinear DGLAP evolution. In the first case from the formal matching onto the collinear scale-dependent FFs, in the second one by construction. What makes the difference, and improves somehow the description in the present study, is the indirect scale dependence of the widths, through the CSS evolution~\cite{collins_2011}, not included in the simplified TMD scheme.
For the sake of completeness, we have also tried a sort of global fit, including the OPAL data set in our analysis. By adopting the double-model parametrization and the Power-Law model we get an overall $\chi^2_{\rm dof}$ =~1.52. We have to notice that this somehow very low value (at least for the combined fit) is affected by the fact that we have included further data points with large errors. The resulting estimate, shown in the right panel of Fig.~\ref{fig:opal} (orange dot-dashed line) almost coincides with the prediction obtained within the same scenario without including the OPAL data.
Obviously, the large error bars prevent to draw any definite conclusion. Nonetheless, the agreement is quite promising.
More precise data as well as more data at larger $p_\perp$ would be extremely useful to test the model predictions.
\subsection{$SU(2)$ symmetry and charm contribution}
We conclude our phenomenological analysis by discussing here some important (and somehow related)
issues, as we will see below. We stress that what follows has to be considered as a preliminary and exploratory study, to be addressed more carefully in future work.
The first aspect we would like to address is the $SU(2)$ isospin symmetry, advocated for instance in Ref.~\cite{Chen:2021hdn}. In our phenomenological analysis, presented in the previous sections, we have not imposed any constraint on the polarizing FFs for up and down quarks. As already found in the first extraction of the $\Lambda$ polarizing FFs~\cite{DAlesio:2020wjq,Callos:2020qtu}, the use of a unique FF for up and down quarks (within a three-flavour parametrization), even adopting the full TMD formalism, would lead to a very poor quality of the fit with a much larger $\chi^2_{\rm dof}$ (around 2, reaching up to 2.5 if no large-$z_h$ cuts are imposed). When the normalization of the up and down quark polarizing FFs are left free the $\chi^2$ minimization naturally leads to different sizes and, more importantly, to opposite signs.
Therefore, within a three-flavour fit and with present data $SU(2)$ symmetry seems to be ruled out. In this respect nothing changes adopting the proper TMD framework.
Another important issue is the relevance of the charm contribution, as explicitly discussed in the Belle experimental analysis~\cite{Guan:2018ckx}. As we will show below, this could play a role also concerning the isospin symmetry issue, and, quite interestingly, in the description of the large-$z_h$ $\Lambda K$ data. These indeed have not been included in the fit, since they would spoil its quality and, at variance with the large-$z_\pi$ $\Lambda\pi$ data, are very difficult to describe. We have also to notice that the Kaon FFs, especially at large $z$, are affected by large uncertainties and this has to be considered in conjunction with the large last experimental $z$ bin.
\begin{figure}[!ht]
\begin{center}
\hspace{0.cm}
\includegraphics[width=15cm]{Imm/Lb_k_p_partial_draft2.pdf}
\caption{Impact of $SU(2)$ symmetry, charm contribution and large $z_K$-cut in the description of $\Lambda K^+$ data: without $SU(2)$ sym., no charm and $z_K$ cut (blue solid line), without $SU(2)$ sym., with charm and $z_K$ cut (orange dot-dashed line), without $SU(2)$ sym., with charm and without $z_K$ cut (green dotted line) and imposing $SU(2)$ sym., with charm and without $z_K$ cut (red dashed line).}
\label{fig:charm2}
\end{center}
\end{figure}
At this stage we have tried to see what happens by including the charm contribution only in the unpolarized cross section, the denominator of the transverse polarization. While still obtaining a similar $\chi^2_{\rm dof}$ around 1.2 with no $SU(2)$ symmetry, now one can obtain a $\chi^2_{\rm dof}$ around 1.45 when imposing it. Quite interestingly also the description of the large $z_h$ bins for $\Lambda K$ data, even if not included in the fit, improves a lot, as one can see in Fig.~\ref{fig:charm2} for the $\Lambda K^+$ data set. Notice that the agreement for $\Lambda\pi$ data is preserved in all cases considered.
Concerning the polarizing FFs
we obtain the following results: the inclusion of the charm contribution leads to an increase, in size, of all polarizing FFs; this effect is mainly driven by the increase of the unpolarized cross section in the denominator of the transverse polarization. When we impose also $SU(2)$ symmetry, both the down- and up-quark polarizing FFs come out positive (it is worth recalling that without $SU(2)$ symmetry the down-quark polarizing FF was negative) and there is a further general increase of all polarizing FFs.
As a general conclusion we can say that by including also the charm contribution (at least in the unpolarized cross section) one can obtain similar good fits imposing or not $SU(2)$ symmetry and/or imposing or not the cut on the large-$z_h$ bin. On the other hand these choices could affect in a different way the corresponding predictions for the transverse $\Lambda$ polarization in other processes, like in SIDIS. We have indeed carried out a preliminary study confirming this hypothesis. A detailed analysis is in progress and will be presented in a future paper.
For its relevance, we also checked the impact of the above assumptions in the combined fit. Once again the one-model fit would give very large $\chi^2_{\rm dof}$. In the double-model fit, focusing on the Power-law parametrization, we obtain similar results as those discussed for the 2-h fit. In other words, imposing $SU(2)$ symmetry once again leads to a very large $\chi^2$.
On the other hand, if one includes the charm contribution, without imposing $SU(2)$ symmetry the $\chi^2$ does not change, while imposing it the $\chi^2$ increases up to 1.7. In both cases the agreement with the large $z_h$ bins for $\Lambda$-$K$ data, even if not included in the fit, improves significantly.
\section{Conclusions}
\label{sec:conclusions}
In this paper we have carried out a reanalysis of Belle data for the transverse $\Lambda$ polarization in $e^+e^-$ annihilation processes by employing proper TMD factorization theorems and QCD evolution equations within the CSS approach.
This is indeed an important observable in the context of hadron physics and the fragmentation process, in particular to reach deeper insights on the transverse polarization mechanism of $\Lambda$ hyperons. Moreover, it is well defined in terms of a TMD approach, allowing to extract the still poorly known polarizing fragmentation function.
One of our main findings is that this study confirms, in many aspects, the results previously obtained within a former, simplified, extraction of the $\Lambda$ polarizing FF, allowing, at the same time, to have a more robust framework for future studies at different energies and/or for different processes.
A common feature of the two analyses is that for the extracted polarizing FFs a clear separation in flavours can be achieved. Within a three flavour scenario, the best description is obtained with three different valence polarizing FFs (breaking the $SU(2)$ isospin symmetry), with their relative
sign and size determined quite accurately. The need of a sea contribution is also well supported.
Two data sets have been considered with their proper features, employing the corresponding formal description and paying special attention to the nonperturbative functions: the associated production case with a light unpolarized hadron and the inclusive case, where one has to reconstruct the thrust axis direction.
Focusing on the double-hadron production data, we have shown how they can be described extremely well through different combinations of nonperturbative functions. Moreover, the $M^{\perp}_D$ functions so extracted, within the Gaussian and the Power-Law model, are totally compatible. However, the other relevant nonperturbative function, $g_K$, plays a more distinctive role. Indeed, the smallest $\chi^2_{\text{dof}}$s are found by employing a specific functional form: the \emph{Logarithmic} one.
One main remark is that none of the models extracted from the double-hadron production data can describe either the size or the pattern of the transverse $\Lambda$ polarization data in the single-inclusive production.
Another striking feature is that the $\Lambda K^+$ and $\bar{\Lambda} K^-$ polarization data, with $z_{K}>0.5$, cannot be described even employing the CSS evolution equations.
As a further attempt we have performed a combined fit of the double and single-inclusive hadron production data sets, by extracting a single or two different sets of parameters for the $M^{\perp}_D$ function. In the first case, the results are still unsatisfactory, while the double-model fit, to be considered as a phenomenological attempt, allows for an overall quite good description.
These findings raise the issue that the two data sets could hardly analysed within the same factorization scheme. For the same reason, the discrepancy between the two models, extracted independently from the two data sets,
could highlight that there are different effects of recoil against emission of soft gluons, and distinct polarization mechanisms for $\Lambda$ hyperons between double and single-inclusive hadron production processes. This deserves a more detailed analysis.
For completeness, we have also considered another data set, the one from the OPAL Collaboration, for the inclusive production case, but at much larger energies. We have checked our predictions against the data as well tried to include them in a global fit. As for the Belle data, the predictions from the associated production data fit are not able to describe the OPAL data, while the combined fit works pretty well. Including the OPAL data set in the fit confirms this finding, with an even smaller $\chi^2$.
Finally, we have explored the role of $SU(2)$ symmetry and of the charm contribution (limiting it to the unpolarized cross section).
Even without carrying out a detailed phenomenological analysis, we can say that imposing $SU(2)$ symmetry alone spoils the description of the data. On the other hand, the inclusion of the charm contribution (with or without $SU(2)$ symmetry) allows still to get a good fit, improving also the description of the $\Lambda K$ data at large $z_K$. Also in this case further work seems necessary and is underway.
Future experimental analyses and further theoretical developments will certainly help in understanding and possibly unveiling the hadronization mechanism involved in the transverse $\Lambda$ polarization observed in the two processes considered.
This issue together with the universality properties of the polarizing FF and its TMD evolution, as well as the role of $SU(2)$ symmetry and heavy flavour contributions, could eventually be explored and clarified in other processes, like SIDIS (in particular at the Electron Ion Collider) and inclusive hadron production in $pp$ collisions.
|
\section{Introduction}
The studies on higher-dimensional black hole solutions to Einstein's equations have played important roles in the microscopic derivation of black hole entropy~\cite{Strominger:1996sh} and in the fundamental research on the scenario of large extra dimensions~\cite{Argyres:1998qn} through black hole production in an accelerator.
The recent developments in solution-generation techniques enable us to find various exact solutions of higher-dimensional black holes, our understanding of them is still not enough.
For example, according to the topology theorem for five-dimensional black holes~\cite{Galloway:2005mf,Cai:2001su,Hollands:2007aj,Hollands:2010qy}, the topology of the spatial cross section of the event horizon must be either sphere $S^3$, ring $S^1\times S^2$ or lens spaces $L(p,q)$, if the spacetime with commuting two rotational Killing vector fields and a timelike Killing vector field
is asymptotically flat.
As for the first two topologies, the exact solutions to vacuum Einstein's equations~\cite{Tangherlini:1963bw,Myers:1986un,Emparan:2001wn,Pomeransky:2006bd} have already been found.
For lens space topology, however, it has been difficult to find a regular vacuum solution since the resultant solutions have naked singularities.
\medskip
The inverse scattering method (ISM) is perhaps one of the most powerful tools to obtain exact solutions of Einstein equations with $(D-2)$ Killing isometries, where $D$ is a spacetime dimension.
In particular, combined with the concept of the rod structure~\cite{Harmark:2004rm}, this method has succeeded to derive five-dimensional vacuum black hole solutions.
The first example of the construction of black hole solutions by the ISM is the re-derivation of the five-dimensional Myers-Perry black hole solution~\cite{Pomeransky:2005sj}.
Next, the black ring with $S^2$-rotation was re-derived~\cite{Tomizawa:2005wv} by the ISM (this was first derived in Refs.~\cite{Mishima:2005id,Figueras:2005zp}),
but it turned out that the generation of the black ring with $S^1$ rotation has a certain problem on how to choose the seed, since an easy choice of the seed always results in the generation of a singular solution.
The suitable seed to derive the black ring with $S^1$-rotation was first considered in~\cite{Iguchi:2006rd,Tomizawa:2006vp}.
Subsequently, the more general black ring solution with both $S^1$ and $S^2$ rotations was constructed by Pomerasnky and Sen'kov~\cite{Pomeransky:2006bd}.
\medskip
Using the ISM, some authors attempted to construct asymptotically flat black lens solutions to the five-dimensional vacuum Einstein equations.
First, Evslin~\cite{Evslin:2008gx} attempted to construct a static black lens with the lens space topology of $L(n^2+1,1)$ but found that curvature singularities cannot be eliminated, whereas both conical and orbifold singularities can be removed.
Moreover, by using the ISM, Chen and Teo~\cite{Chen:2008fa} constructed black hole solutions with the horizon topology of $L(n,1)=S^3/{\mathbb Z_n}$ but these solutions must have either conical singularities or naked curvature singularities.
Thus, the major obstacle in constructing a black lens solution is always the existence of naked singularities.
However, the sudden breakthrough has come from supersymmetric solutions.
Kunduri and Lucietti~\cite{Kunduri:2014kja} succeeded in the derivation of the first regular supersymmetric solution of an asymptotically flat black lens with the horizon topology of $L(2,1)=S^3/{\mathbb Z}_2$.
This solution was further generalized to the more general supersymmetric black lens with the horizon topology $L(n,1)=S^2/{\mathbb Z}_n\ (n\ge 3)$~\cite{Tomizawa:2016kjh,Breunholder:2017ubu}.
Getting a useful clue from the work of Kundhuri and Lucietti, Ref~\cite{Tomizawa:2019acu} attempted to construct the vacuum solution of the black lens with $L(2,1)$ without singularities but the solution has unavoidable closed timelike curves (CTCs).
Thereafter, Ref.~\cite{Lucietti:2020phh} discuss nonexistence of vacuum black lenses.
\medskip
Supersymmetric black lenses carry the mass, electric charge (saturating BPS bound), two angular momenta and magnetic fluxes~\cite{Kunduri:2014kja,Tomizawa:2006vp}.
As is discussed in Ref.~\cite{Tomizawa:2006vp}, there exists no limit such that all the magnetic fluxes vanish.
Therefore, as for the supersymmetric solutions, the existence of the magnetic fluxes seem to play an essential role in supporting the horizon of the black lens.
In general, however, it is not clear whether such magnetic fluxes necessarily need to construct a black lens.
Recently, a different type of solutions within a class of generalized Weyl solutions, static black hole binaries and black rings in expanding bubbles of nothing, was studied in Ref.~\cite{Astorino:2022fge}, although so far equilibrium configuration of black holes in bubble had been studied in the context of Kaluza-Klein theory~\cite{Elvang:2002br,Tomizawa:2007mz,Iguchi:2007xs}.
As is well-known, an asymptotically flat, static black ring cannot be in equilibrium since the horizon collapse due to the self-gravitational force.
However, the black ring in~\cite{Astorino:2022fge} is allowed to be in static equilibrium by the balance between the expanding force of a bubble and the gravitational force, so it has no conical singularities.
This solution leads a simple, interesting question to us.
Is a non-rotating black lens in bubble of nothing allowed to be in equilibrium?
To study such a solution may enable us to know what (except for magnetic fluxes) is needed to obtain a regular black lens.
Thus, the end of this paper is to investigate whether expanding bubble of nothing admits the existence of a black lens in equilibrium.
In this paper, to derive such a solution, we apply the ISM to the five-dimensional vacuum Einstein equations with staticity and bi-axisymmetry, and construct a $1$-soliton solution by considering a static black ring inside of bubble in~\cite{Astorino:2022fge} as a seed solution.
Note that our procedure in the ISM is entirely the same as the work of Chen-Teo where the seed solution is chosen as a static black ring, namely, the only different point is the seed solution.
\medskip
We organize the remaining part of this paper as follows.
In the following section~\ref{sec:solution}, under the assumptions of staticity and bi-axisymmetry, we present a vacuum solution of a non-rotating black lens with the horizon topology $L(n,1)$ in bubble of nothing as a $1$-soliton solution in five dimensions by using the ISM.
In section~\ref{sec:boundary}, we impose the boundary conditions under such that the spacetime has
none of curvature, conical, and orbifold singularities on the axis and horizon.
In section~\ref{sec:ctcs}, we further impose no closed timelike curves (CTCs) in the domain of communication.
In section~\ref{sec:black lens}, we discuss whether the non-rotating black lens in bubble of nothing indeed exist.
In section~\ref{sec:limit}, we confirm the limit of our solution to the Chen-Teo static solution.
In the final section~\ref{sec:summary}, we devote ourselves to the summary and discussion on our results.
\section{ Static black ring in bubble of nothing as a seed solution}\label{sec:solution}
In general, the metric for a stationary and bi-axisymmetric spacetime can be written in the canonical coordinates as
\begin{eqnarray}
ds^2=g_{ab}dx^adx^b+f(d\rho^2+dz^2),\quad (a,b=t,\phi,\psi),
\end{eqnarray}
where $g_{ab}$ and $f$ depend on only $(\rho,z)$.
The following constraint condition must be satisfied
\begin{eqnarray}
{\rm det} (g_{ab})=-\rho^2. \label{eq:det}
\end{eqnarray}
According to the procedure of Chen and Teo~\cite{Chen:2008fa}, we construct the static black lens in bubble of nothing by the ISM, where the static black ring as the seed solution is replaced with the black ring in bubble of nothing~\cite{Astorino:2022fge}\ (see Fig.~\ref{fig:rod} on the rod structure.).
\begin{figure}[H]
\centering
\includegraphics[width=10cm]{rod1.pdf}
\caption{Rod structure of the black ring inside bubble of nothing. }
\label{fig:rod}
\end{figure}
\medskip
Therefore, let us start with the exact solution of the black ring in bubble of nothing to the five-dimensional vacuum Einstein equations, whose metric is given by
\begin{eqnarray}
G_0&=&{\rm diag\ }\left(-\rho^2 \frac{\mu_1}{\mu_0\mu_2},\frac{\mu_0\mu_2}{\mu_1\mu_3},\mu_3 \right),\\
f_0&=&C_f\frac{\mu_3W_{01}^2 W_{03} W_{12}^2 W_{23}}{W_{02}^2W_{13} W_{00} W_{11} W_{22} W_{33}},
\end{eqnarray}
where for $i,j=0,1,2,3$,
\begin{eqnarray}
\mu_i&:=&\sqrt{\rho^2+(z-z_i)^2}-(z-z_i), \\
\bar \mu_i&:=&-\frac{\rho^2}{\mu_i},\\
W_{ij} &:=&\rho^2+\mu_i\mu_j.
\end{eqnarray}
First, let us remove a trivial anti-soliton with the BZ vector $(0,0,1)$ at $z=z_3$:
\begin{eqnarray}
g_0&=&{\rm diag\ }\left(1,1,-\frac{\bar \mu_3^2}{\rho^2}\right) G_0={\rm diag\ }\left(1,1,-\frac{\rho^2}{\mu_3^2}\right) G_0,
\end{eqnarray}
In turn, let us add back a non-trivial anti-soliton with $(0,-a,1)$ to this, and then we can obtain a new $1$-soliton solution, a solution of {\it a static black lens in bubble of nothing},
\begin{eqnarray}
g_1&=&-\rho^2 \frac{\mu_1}{\mu_0\mu_2}dt^2
+\frac{\mu_0 \mu_2 ( \mu_1W_{03}^2W_{23}^2 + 4a^2z_3^2\mu_0\mu_2\mu_3^2 W_{13}^2 )}
{\mu_1\mu_3 ( \mu_1 W_{03}^2W_{23}^2 - 4a^2z_3^2\rho^2 \mu_0\mu_2 W_{13}^2 )}d\phi^2\notag\\
&&-2az_3\frac{ 2 \mu_0\mu_2 W_{03} W_{13} W_{23}W_{33} }
{ \mu_3 ( \mu_1 W_{03}^2W_{23}^2 - 4a^2z_3^2\rho^2 \mu_0\mu_2 W_{13}^2 ) } d\phi d\psi \notag\\
&&+\frac{ \mu_1 \mu_3^2 W_{03}^2 W_{23}^2 + 4a^2 z_3^2\rho^4 \mu_0\mu_2 W_{13}^2 }
{ \mu_3 ( \mu_1 W_{03}^2W_{23}^2 - 4a^2z_3^2\rho^2 \mu_0\mu_2 W_{13}^2 )} d\psi^2,
\end{eqnarray}
and
\begin{eqnarray}
f_1
&=& f_0 \frac{ \mu_1 W_{03}^2W_{23}^2 - 4a^2z_3^2\rho^2 \mu_0\mu_2 W_{13}^2 }
{\mu_1 W_{03}^2 W_{23}^2 }.
\end{eqnarray}
It is easy to confirm that in the limit of $a\to0$, this solution coincides with the static black ring in bubble of nothing~\cite{Astorino:2022fge}.
One should note that $t,\psi$ is dimensionless and $\phi$ has the dimension of length.
In the following section, after an appropriate coordinate transformation, we will impose the periodicity of $\phi, \psi$ so that conical singularities do not exist on symmetry of axises.
\section{Boundary conditions on the rods} \label{sec:boundary}
\begin{figure}[H]
\centering
\includegraphics[width=10cm]{rod2.pdf}
\caption{Rod structure of the black lens inside bubble of nothing. }
\label{fig:rod-lens}
\end{figure}
In order to impose the appropriate boundary conditions so that the solution has the rod structure in Fig.~\ref{fig:rod-lens} and neither conical singularities nor orbifold singularities,
let us introduce new parameters $b:=z_3a$ and new coordinates $(\phi',\psi')$ defined by
\begin{eqnarray}
\frac{\partial}{\partial \phi'}=\sqrt{\frac{z_{30}}{z_{30}-2b^2}} \left(\frac{\partial}{\partial \phi}+\frac{b}{z_{30}}\frac{\partial}{\partial \psi}\right),\quad
\frac{\partial}{\partial \psi'}=\sqrt{\frac{z_{30}}{z_{30}-2b^2}} \left(\frac{\partial}{\partial \psi}+2b\frac{\partial}{\partial \phi}\right),
\end{eqnarray}
and then the angular components are written as
\begin{eqnarray}
g_{\phi'\phi'}&=&\frac{z_{30}}{z_{30}-2b^2}\left(g_{\phi\phi}+\frac{b^2}{z_{30}^2} g_{\psi\psi}+\frac{2b}{z_{30}} g_{\phi\psi} \right), \\
g_{\psi'\psi'}&=&\frac{z_{30}}{z_{30}-2b^2}(g_{\psi\psi}+4b^2 g_{\phi\phi}+4b g_{\phi\psi}),\\
g_{\phi'\psi'}&=&\frac{z_{30}}{z_{30}-2b^2}\left[ \left( 1+\frac{2b^2}{z_{30}}\right) g_{\phi\psi}+2bg_{\phi\phi}+\frac{b}{z_{30}} g_{\psi\psi} \right],
\end{eqnarray}
where it should be noted that the constraint condition~(\ref{eq:det}) is preserved.
\medskip
Focusing on the two-dimensional space $\Sigma=\{(\rho,z)|\ \rho> 0,-\infty<z<\infty\}$, let us study the rod structure of the obtained solution.
The rod structure~\cite{Emparan:2001wk,Harmark:2004rm} enables us to understand stationary and axisymmetric solutions (more precisely, solutions with $(D-2)$ commuting Killing vectors) easily in a diagrammatic may.
The $z$-axis ($\rho=0$) of the metric, which corresponds to a fixed point set of a certain Killing isometry, is decomposed into five parts, $I_-=\{(\rho,z)\ | \ \rho=0,\ z<z_0\}$, $I_0=\{(\rho,z)\ | \ \rho=0,\ z_0<z<z_1\}$, $I_1=\{(\rho,z)\ | \ \rho=0,\ z_1<z<z_2\}$, $I_2=\{(\rho,z)\ | \ \rho=0,\ z_2<z<z_3\}$, $I_+=\{(\rho,z)\ | \ \rho=0,\ z_3<z\}$.
Thus, the boundary $\partial \Sigma$ of $\Sigma$ are composed of $I_\pm,I_i\ (i=0,\ldots,2)$ and the asymptotic region $I_\infty=\{(\rho,z)|\ \sqrt{\rho^2+z^2}\to \infty \ $ with $ z/\sqrt{\rho^2+z^2}$ finite$\}$.
\medskip
Now, we impose the conditions on each rod so that the solution has the same rod structure as Fig.~\ref{fig:rod-lens} and has no conical singularities.
\begin{itemize}
\item[(1)] $I_3= \{(\rho,z)\ | \ \rho=0,\ z>z_3\}$: \\
The Killing vector $v_3:=(0,0,1)=\partial/\partial\psi'$ vanishes.
The condition for the absence of conical singularities on $I_3$ is given by
\begin{eqnarray}
\lim_{\rho\to 0}\sqrt{\frac{\rho^2 f_1}{g_{\psi' \psi '} }}=\frac{\Delta \psi'}{2\pi}
\Longleftrightarrow
C_f\frac{z_{30}-2b^2 }{z_{30}}=\left(\frac{\Delta \psi'}{2\pi}\right)^2
\end{eqnarray}
for $z\in(z_3,\infty)$.
Hence, if we choose the periodicity of $\psi'$ as $\Delta \psi'=2\pi$,
the condition can be satisfied on $I_3$ by setting
\begin{eqnarray}
C_f=\frac{z_{30}}{z_{30}-2b^2 }.\label{eq:cf}
\end{eqnarray}
\item[(2)] $I_{0}= \{(\rho,z)\ | \ z_0<z<z_1,\rho=0\}$: \\
The Killing vector $v_{01}:=(0,1,0)=\partial/\partial\phi'$ vanishes.
The condition for the absence of conical singularities on $I_0$ is given by
\begin{eqnarray}
\lim_{\rho\to 0}\sqrt{\frac{\rho^2 f_1}{g_{\phi'\phi'}} }=\frac{\Delta \phi'}{2\pi}
\Longleftrightarrow
\frac{2C_fz_{10}^2(z_{30}-2b^2)}{z_{20}^2}=\left(\frac{\Delta \phi'}{2\pi}\right)^2. \label{eq:conical_01}
\end{eqnarray}
\item[(3)] $I_{2}= \{(\rho,z)\ | \ z_2<z<z_3,\rho=0\}$: \\
The Killing vector $v_{23}:=(0,1,\frac{bz_{21}}{z_{30}z_{32}-2b^2 z_{31}})=\partial/\partial\tilde \phi'$ vanishes.
The conical-free condition is given by
\begin{eqnarray}
\lim_{\rho\to 0}\sqrt{\frac{\rho^2 f_1}{g_{1ab} v_{23}^av_{23}^b} }=\frac{\Delta \tilde\phi'}{2\pi}
\Longleftrightarrow
\frac{2C_f(z_{30}z_{32}-2b^2z_{31})^2}{ z_{32} z_{31} (z_{30}-2b^2) }=\left(\frac{\Delta \tilde \phi'}{2\pi}\right)^2. \label{eq:conical_23}
\end{eqnarray}
\end{itemize}
Since $\phi'$ and $\tilde \phi'$ have the scale of length, it is useful to introduce angular coordinates $\varphi$ and $ \tilde \varphi$ with $2\pi$ periodicity by $\varphi:=L \phi'$ and $\tilde \varphi:=L \tilde \phi'$. Then, together with Eq.~(\ref{eq:cf}), the conditions (\ref{eq:conical_01}) and (\ref{eq:conical_23}) can be written as,
\begin{eqnarray}
(\ref{eq:conical_01}) &\Longleftrightarrow& \frac{2z_{10}^2z_{30}}{z_{20}^2}=L^2\left(\frac{\Delta \varphi}{2\pi}\right)^2 \\
(\ref{eq:conical_23}) &\Longleftrightarrow& \frac{2z_{30}(z_{30}z_{32}-2b^2z_{31})^2}{ z_{32} z_{31} (z_{30}-2b^2)^2 }=L^2\left(\frac{\Delta \tilde \varphi}{2\pi}\right)^2.
\end{eqnarray}
\medskip
Moreover, we impose the boundary condition on the parameters $(z_i,b)$ so that the spatial topology of the horizon is lens space $L(n;1)=S^3/{\mathbb Z}_n$ ($n\in {\mathbb N}$).
From the mathematical discussion in Ref.~\cite{Hollands:2007aj}, this condition is represented by
\begin{eqnarray}
{\rm det}( \bar v_{01},\bar v_{23})=n
\Longleftrightarrow
\frac{L b z_{21}}{z_{30}z_{32}-2b^2z_{31}}=n,
\end{eqnarray}
where $( \bar v_{01},\bar v_{23})=L( v_{01}, v_{23})$.
\medskip
(4) $I_{-}= \{(\rho,z)\ | \ z<z_0,\rho=0\}$:
The Killing vector $v_{-}:=(1,0,0)=\partial/\partial t$ vanishes.
This semi-infinite rod corresponds to an accelerating horizon.
\medskip
(5) $I_{1}= \{(\rho,z)\ | \ z_1<z<z_2,\rho=0\}$:
The Killing vector $v_{1}:=(1,0,0)=\partial/\partial t$ vanishes.
This finite rod corresponds to an event horizon.
\section{CTCs}\label{sec:ctcs}
We require absence of CTCs on $\Sigma \cup \partial \Sigma$.
The necessary and sufficient conditions to ensure that CTCs do not exist on $\Sigma \cup \partial \Sigma$ is such that $g_{\phi\phi}$ and $g_{\psi\psi}$ (or $g_{\phi'\phi'}$ and $g_{\psi'\psi'}$) become nonnegative in the region:
\medskip
The condition for the absence of CTCs is given by
\begin{align}
\mu_1 W_{03}^2 W_{23}^2-4 b^2 \rho^2 \mu_0 \mu_2 W_{13}^2 > 0,
\label{eq:noCTC}
\end{align}
which imposes an upper bound for $b^2$ at each point
\begin{align}
b^2 < U(\rho,z) :=\frac{\mu_1 W_{03}^2 W_{23}^2}{4 \rho^2 \mu_0 \mu_2 W_{13}^2}.
\end{align}
Therefore, if the minimum $U_{min}$ of $U(\rho,z)$ exists on $\Sigma \cup \partial \Sigma$ and $b^2< U_{min}$ holds, CTCs do not exist in the region.
To prove this, we show that the function $U(\rho,z)$ has a minimum at $(\rho,z)=(0,z_3)$ on the rod $I_\pm,I_i\ (i=0,\ldots,2)$.
\medskip
It is not difficult to show that the function $U(\rho,z)$ has a minimum not on $\Sigma$ but on $\partial \Sigma$ .
To see this, one should note that the norm of the gradient can be written as
\begin{align}
&( \partial_\rho U )^2 + (\partial_z U)^2= \frac{\mu_1^2 W_{03}^4 W_{23}^4}{4\rho^6 \mu_0^2 \mu_2^2 W_{00} W_{11} W_{22} W_{13}^4} \nonumber\\
& \qquad \times \left((\mu_0-\mu_1+\mu_2)^2 \rho^4 + 2 \mu_0\mu_2(2\mu_0 \mu_2-(\mu_0+\mu_2)\mu_1 + \mu_1^2)\rho^2+\mu_0^2\mu_1^2\mu_2^2 \right),
\end{align}
where the first line is always positive for $\rho>0$, and the second line is also positive since
\begin{align}
& (\mu_0-\mu_1+\mu_2)^2 \rho^4 + 2 \mu_0\mu_2(\mu_1^2-(\mu_0+\mu_2)\mu_1 + 2\mu_0 \mu_2)\rho^2+\mu_0^2\mu_1^2\mu_2^2\nonumber\\
& = \left(\left(\mu _0-\mu _1+\mu _2\right) \rho ^2+\frac{\mu _0 \mu _2 \left(\mu _1^2-(\mu_0+\mu_2)\mu_1+2 \mu_0\mu _2\right)}{\mu _0-\mu _1+\mu _2}\right)^2+\frac{4 \mu _0^3\mu_2^3 \left(\mu _0-\mu _1\right) \left(\mu _1-\mu _2\right) }{\left(\mu _0-\mu _1+\mu _2\right){}^2},
\end{align}
where the positivity of the last term can be shown from
\begin{align}
&(\mu_0-\mu_1)(\mu_1-\mu_2)
\nonumber\\
&=
\frac{z_{10} z_{21}(\mu_1 + \mu_0)(\mu_2 + \mu_1)}{(\sqrt{\rho^2+(z-z_1)^2}+\sqrt{\rho^2+(z-z_0)^2})(\sqrt{\rho^2+(z-z_2)^2}+\sqrt{\rho^2+(z-z_1)^2})}>0.
\end{align}
Therefore, the gradient of a smooth function $U(\rho,z)$ cannot be zero on $\Sigma$, which means that $U(\rho,z)$
must have a minimum not on $\Sigma$ but on $\partial\Sigma$.
Hence, in what follows, let us consider a minimum of $U(\rho,z)$ on $\partial\Sigma$, which corresponds to $ I_i\ (i=\pm,0,\ldots,2), I_\infty$.
\medskip
First, let us consider the minimum of $U(0,z)$ on the rod $\rho=0$, i.e., on $I_i\ (i=\pm,0,\ldots,2)$.
On $I_+$, we have
\begin{eqnarray}
U(0,z)=\frac{(z-z_0)(z-z_2)}{2(z-z_1)}, \quad
U_{,z}(0,z)=\frac{(z-z_1)^2+z_{10}z_{21}}{2(z-z_1)^2}>0,
\end{eqnarray}
and hence the monotonically increasing function $U(0,z)$ on $I_+$ has a minimum at $z=z_3$, which is written as
\begin{align}
U(0,z_3) = \frac{z_{30}z_{32}}{2 z_{31}}. \label{eq:localmin-1}
\end{align}
On the other hand, since on $I_2$
\begin{eqnarray}
U(0,z)=\frac{(z-z_1)z_{30}^2z_{32}^2}{2(z-z_0)(z-z_2)z_{31}^2}, \quad
U_{,z}(0,z)=-\frac{[(z-z_1)^2+z_{10}z_{21}]z_{30}^2z_{32}^2}{2(z-z_0)^2(z-z_2)^2z_{31}^2}<0,
\end{eqnarray}
the monotonically decreasing function $U(0,z)$ on $I_2$ has a minimum at $z=z_3$, and hence
\begin{eqnarray}
U(0,z)\ge U (0,z_3).
\end{eqnarray}
Moreover, observing on $I_0$,
\begin{eqnarray}
U(0,z)&=&\frac{(z-z_2)z_{30}^2}{2(z-z_0)(z-z_1)},\\
U_{,z}(0,z)&=&-\frac{[(z-z_2)^2-z_{20}z_{21}]z_{30}^2}{2(z-z_0)^2(z-z_1)^2}
\begin{cases}
<0 & ( z_0 <z< z_2-\sqrt{z_{20}z_{21}}) \\
>0 & (z_2-\sqrt{z_{20}z_{21}}<z<z_1),
\end{cases}
\end{eqnarray}
we find that the function $U(0,z)$ on $I_0$ has a local minimum at $z=z_*:=z_2-\sqrt{z_{20}z_{21}}$, and hence
\begin{eqnarray}
U(0,z)\ge U (0,z_*)= \frac{z_{30}^2}{2(z_{20}+z_{21}-2 \sqrt{z_{20}z_{21}})},\label{eq:localmin-2}
\end{eqnarray}
where we note that the ratio of these minima on $I_0,I_2,I_+$ is computed as
\begin{align}
\frac{U(0,z_*)}{U(0,z_3)} =
\left(1+\frac{z_{21}}{z_{32}}\right) \left(1+\frac{z_{32}}{z_{20}}\right) \left(1-\sqrt{\frac{z_{21}}{z_{20}}}\right)^{-2}> 1.
\end{align}
Furthermore, near $I_-$ and $I_1$, the function $U(\rho,z)$ behaves as, respectively,
\begin{eqnarray}
U(\rho,z)\simeq \frac{2(z_0-z)(z_2-z)(z_3-z)^2}{(z_1-z)\rho^2},\quad
U(\rho,z)\simeq \frac{2(z-z_1)(z_2-z)(z_3-z)^2z_{30}^2}{(z-z_0)z_{31}^2\rho^2},
\end{eqnarray}
which implies $U(0,z)=\infty$ on $I_-$ and $I_1$.
To summarize, we have shown that the minimum of $U(0,z)$ on $I_i\ (i=\pm,0,\ldots,2)$ is given by Eq.~(\ref{eq:localmin-1}).
\medskip
Next, let us consider the function $U(\rho,z)$ in the asymptotic region, namely, on $I_\infty$.
We can show that in the asymptotic region, $U(\rho,z)$ behaves as
\begin{align}
U(\rho,z) \simeq \frac{1}{1+\frac{z}{\sqrt{\rho^2+z^2}}}\sqrt{\rho^2+z^2}.
\end{align}
Hence, the function $U(\rho,z)$ diverges on $I_\infty$, so it cannot have a minimum on $I_\infty$.
Thus, we conclude that Eq.~(\ref{eq:localmin-1}) is also a minimum on $\Sigma\cup \partial \Sigma$ as well as on $\partial \Sigma$, and hence
the necessary and sufficient condition for the absence of CTCs is given by
\begin{align}
b^2 < \frac{z_{30}z_{32}}{2z_{31}}. \label{eq:noctc-0}
\end{align}
\section{On existence of solutions} \label{sec:black lens}
From the discussion in Sec~\ref{sec:boundary} and \ref{sec:ctcs}, we have shown that the absence of conical/orbifold singularities, the black lens condition require
\begin{align}
&\frac{2 z_{10}^2 z_{30}}{z_{20}^2} = L^2, \label{eq:regularity-con1}\\
& \frac{2 z_{30}( z_{32} z_{30}-2 b^2 z_{31})^2}{z_{32} z_{31}(z_{30}-2b^2)^2}=L^2, \label{eq:regularity-con2}\\
& \frac{L b z_{21}}{z_{30}z_{32}-2b^2 z_{31}} = n, \label{eq:regularity-con3}
\end{align}
Now, to confirm whether there is really a parameter range such that all these conditions can be satisfied, we reparameterize the rod interval $z_{i ,i-1}:=z_i-z_{i-1}\ (i=1,2,3)$ and the redefined BZ parameter $b$ as follows
\begin{align}
z_{10} = \ell ,\quad z_{21} := x \, \ell, \quad z_{32} = y \, \ell ,\quad L = \sqrt{\ell}\, \hat{L},\quad b = \sqrt{\ell} \, \hat{b},
\label{eq:regularity-param}
\end{align}
where $\ell$ fixes the size of the bubble on $I_0$,
$x$ and $y$ are the size of the horizon and the distance between the horizon and the center which is often called nut.
All the dimensionless parameters but $n$, $x,y,\ell,\hat L, \hat b$, are assumed to be positive.
The condition for avoiding CTCs~\label{noctc-0} is now given by
\begin{align}
b^2-\frac{y(1+x+y)}{2(x+y)}<0. \label{eq:noctc}
\end{align}
From Eq.~(\ref{eq:regularity-con3}), we have
\begin{align}
\hat{L} = \frac{n ( y^2-2 \tilde{b}^2 (x+y) + y (1+x))}{\hat{b}\, x} \label{eq:regularity-ell}.
\end{align}
Eliminating $\hat{L}$ from Eqs.~(\ref{eq:regularity-con1}) and (\ref{eq:regularity-con2}) in terms of Eq.(\ref{eq:regularity-ell}), we obtain
\begin{align}
&0= 2 \hat{b}^2 (x+y+1) \left\{ 2 n^2 (x+1)^2 y^2+2 n^2 (x+1)^2 x y+x^2\right \}\nonumber\\
&\qquad-4 \hat{b}^4 n^2 (x+1)^2
(x+y)^2-n^2 (x+1)^2 y^2 (x+y+1)^2, \label{eq:regularity-con2a}
\end{align}
and
\begin{align}
&0=\left\{ y (x+y+1)-2 \hat{b}^2 (x+y)\right\}^2\nonumber\\
& \times \left[ 4 \hat{b}^4 n^2 y (x+y)-2 \hat{b}^2 (x+y+1) \left\{ 2 n^2y( x+y)+x^2\right\}+n^2 y (x+y)
(x+y+1)^2 \right]. \label{eq:regularity-con2b}
\end{align}
First, we consider the conical singularity free condition on $I_2$ in Eq.~(\ref{eq:regularity-con2b}), which admits three branches for $\hat b^2$
\begin{eqnarray}
\hat b^2&=&\hat b_\pm^2:=\frac{(x+y+1) \left(x^2+2 n^2 (x+y)y \pm x \sqrt{x^2+ 4 n^2 x y+4 n^2 y^2}\right)}{4 n^2 (x+y)y}, \label{eq:b_pm}\\
\hat b^2&=&\hat b_0^2:=\frac{y(x+y+1)}{2(x+y)}.\label{eq:b_0}
\end{eqnarray}
From Eq.~(\ref{eq:b_pm}), we can show
\begin{eqnarray}
\hat b_\pm^2-\frac{y(1+x+y)}{2(x+y)}=\frac{x(x+y+1) \left(x+2n^2 y \pm \sqrt{x^2+ 4 n^2 x y+4 n^2 y^2}\right)}{4 n^2 (x+y)y}>0,
\end{eqnarray}
where we note
\begin{eqnarray}
(x+2n^2 y)^2-(x^2+ 4 n^2 x y+4 n^2 y^2)=4n^2(n^2-1)y^2>0.
\end{eqnarray}
Therefore, this shows that the nonexistence condition of CTCs, (\ref{eq:noctc}), cannot be satisfied for any $x>0$ and $y>0$.
On the other hand, substituting Eq.~(\ref{eq:b_0}) into Eq.~(\ref{eq:regularity-ell}), we can show
\begin{eqnarray}
\hat L=0,
\end{eqnarray}
which cannot satisfy $\hat L>0$.
Hence, the solution without the conical singularity on $I_2$ cannot avoid CTCs.
\medskip
Next, we consider the conical singularity free condition on $I_0$ in Eq.~(\ref{eq:regularity-con2a}), from which we can obtain two branches for $\hat b^2$,
\begin{eqnarray}
\hat b^2=\tilde b^2_\pm:=\frac{(x+y+1)\left\{x^2+2n^2(x+1)^2(x+y) y \pm x\sqrt{x^2+4n^2(x+1)^2(x+y)y} \right\}}{4n^2(x+1)^2(x+y)^2},
\end{eqnarray}
which leads to
\begin{eqnarray}
\tilde b_\pm^2-\frac{y(1+x+y)}{2(x+y)}=\frac{x(x+y+1)\left(x\pm \sqrt{x^2+4n^2(x+1)^2(x+y)y} \right)}{4n^2(x+1)^2(x+y)^2}.
\end{eqnarray}
From these, we find that only the branch $\tilde b_-^2$ can satisfy the nonexistence condition of CTCs~(\ref{eq:noctc}).
\medskip
In summary, if one imposes the absence of conical singularities on the whole symmetry of axis $I_0,I_2,I_+$, the presence of CTCs cannot be avoided around the center $(\rho,z)=(0,z_3)$.
However, if one imposes it only on $I_0$ and \blue{$I_+$}, one can obtain the solutions without CTCs, in which the horizon admits the lens space topology $L(n;1)$ for $n\ge 1$.
\section{Consistency with Chen-Teo static black lens}\label{sec:limit}
Here, we confirm that our solution coincides with the asymptotically flat, static black lens solution by Chen-Teo~\cite{Chen:2008fa} in a certain scaling limit, for which the following variables are used,
\begin{align}
z_0 = - \lambda^2 ,\quad z_1 = - \lambda c \kappa^2,\quad z_2 = \lambda c \kappa^2,\quad z_3 = \lambda \kappa^2,\quad L = \sqrt{2} \lambda \bar{L},\quad a = \frac{\bar{a}}{\sqrt{2}\kappa^2}.
\end{align}
With the rescaled coordinates,
\begin{align}
\rho \to \lambda \bar{\rho},\quad z \to \lambda \bar{z},
\end{align}
and the rescaled parameters,
\begin{align}
\bar{z}_0 := z_1/\lambda = - c\kappa^2,\quad \bar{z}_1 := z_2/\lambda = c\kappa^2, \quad\bar{z}_1 :=z_3/\lambda = \kappa^2
\end{align}
the limit $\lambda \to \infty$ pushes $z_0$ away to $-\infty$, and one can see that the rod structure in $(\bar{\rho},\bar{z})$ recovers that of the static black lens in Ref.~\cite{Chen:2008fa}.
In the limit $\lambda \to \infty$, Eq.~(\ref{eq:regularity-con1}) requires
\begin{align}
\bar{L} =1.
\end{align}
and then, Eqs.~(\ref{eq:regularity-con2}) and (\ref{eq:regularity-con3}) leads, respectively, to the nonexistence condition of the conical singularities on $\bar z\in (\bar z_2,\bar z_3)$ and the condition of horizon topology $L(n;1)$ in Ref.~\cite{Chen:2008fa},
\begin{align}
\frac{(1-c-\bar{a}^2(1+c))^2}{(1-\bar{a}^2)^2(1-c^2)}=1,\quad
\frac{2 \bar{a} c}{1-c-\bar{a}^2(1+c)}=n,
\end{align}
where $\bar{a}$ corresponds to $a$ in Ref.~\cite{Chen:2008fa}.
Moreover, the limit of the nonexistence condition for CTCs~(\ref{eq:noctc-0}) can be written as
\begin{align}
\bar{a}^2 < \frac{c-1}{c+1},
\end{align}
which corresponds to the parameter region, ``Region I", in Ref.~\cite{Chen:2008fa}.
Here, it should be noted that the solution in ``Region II" in \cite{Chen:2008fa}, which admits naked singularities and CTCs,
are excluded from our solution by the nonexistence condition (\ref{eq:noctc-0}) of CTCs.
\section{Summary and Discussions}\label{sec:summary}
In this paper, using the ISM for static and bi-axisymmetric Einstein equations, we have constructed the non-rotating black lens inside a bubble of nothing whose horizon is topologically lens space $L(n,1)=S^3/{\mathbb Z}_n$.
Our work is entirely the parallel to the work of Chen-Teo~ \cite{Chen:2008fa}, where the static black ring as a seed solution is replaced with a static black ring in bubble of nothing~\cite{Astorino:2022fge}.
Using this solution, we have investigated whether a static black lens can be in equilibrium by the force balance between the expansion and gravitational attraction.
If we require the absence of CTCs in the domain of outer communication,
the non-rotating black lens must have conical singularities between the horizon and the center.
It has been shown, however, that for black lens, the existence of expanding bubble does not exclude conical singularities and hence the two forces, the force of the bubble expansion and gravitational attraction, cannot be in static equilibrium unlike for the black ring.
\medskip
Refs.~\cite{Chen:2008fa,Tomizawa:2019acu} have argued whether a rotating black lens can be in equilibrium by the balance between
the gravitational force (attraction) and the centrifugal force (repulsive force), and concluded that it cannot be in equilibrium.
The generalization of this rotating black lens solution to one in expanding bubble may be an interesting issue,
since whether such a black lens without conical singularities exists depends on the balance among the gravitational force, the centrifugal force and the expansion of the bubble.
Such a rotating black lens solution is expected to be two or more than soliton solution.
This deserves our future work.
\acknowledgments
This work is supported by Toyota Technological Institute Fund for Research Promotion A.
RS was supported by JSPS KAKENHI Grant Number JP18K13541.
ST was supported by JSPS KAKENHI Grant Number 21K03560.
|
\section{Introduction}
Velocity map imaging (VMI) \cite{eppink1997velocity} is a widely used and extremely versatile technique for measuring a large range of ion or electron energies with high resolution\cite{ding2021composite,zhang2016aplanatic,kling2014thick,kregel2017multi}. The VMI technique utilizes a series of cylindrically symmetric electrodes to create an electrostatic field which serves as lens to map charged particles of a given transverse momentum (or energy) to a specific position on the surface of a detector. Over time a variety of lensing geometries have been developed to increase the relative energy resolution and range to > 1 keV \cite{ding2021composite,kling2014thick} and < 2\% \cite{kling2014thick,zhang2016aplanatic} respectively. One such development pushing previous limits is the thick-lens VMI by Kling et al. \cite{kling2014thick}. This design was shown to extend the VMI detectable energy range while maintaining high resolution by using 11 VMI electrodes rather than the standard 3. Additionally, new applications of mesh electrodes have become common \cite{kauczok2009three,ablikim2019coincidence,zhang2014method,ding2021composite}. These electrodes allow for equipotential surfaces that span the electrostatic lens and give fine control of the field while being transparent enough to limit particle loss.
Though these changes have enhanced VMI capabilities, the net result of VMI largely remained a 2D projection of the 3D momentum distribution. Despite this traditional limit of VMI, there has been a constant and continued interest in attaining the full 3D momentum distribution for use in studying various phenomena such as ionization \cite{pengel2017electron,maurer2012molecular} and photodissociation \cite{dinu2002application,chichinin2002three}. Other 3D methods such as cold-target recoil-ion-momentum spectroscopy (COLTRIMS) \cite{ullrich2003recoil,ullrich1997recoil,dorner2000cold} have proven effective to this end, but can be prohibitively expensive and difficult to implement relative to VMI.
3D VMI momentum measurements have been achieved either through reconstruction techniques or, more recently, direct measurement. Of the two methods, direct measurement is generally considered the ideal in most circumstances as reconstruction often comes with considerable drawbacks. The most common reconstruction technique utilizes the inverse Abel transform \cite{shepp1974fourier,strickland1991reconstruction} which requires cylindrical symmetry and is thus unsuitable for experiments using elliptically polarized light. Alternatively, tomographic imaging methods have been implemented \cite{smeenk2009momentum} which recover 3D information by rotating the momentum distribution (via the exciting laser field polarization) and reconstructing from multiple 2D projections. This has the obvious drawback of requiring multiple datasets to be taken to recover a single 3D distribution which in not practical in some situations, e.g., pump-probe experiments. Further, both reconstruction methods are liable to introduce noise and potentially artifacts into the reconstructed distribution.
Direct measurement of the 3D distribution can be performed by using the particle’s time-of-flight (TOF) which encodes its momentum along the axis of the electrostatic lens. DC slicing \cite{gebhardt2001slice} is a technique by which the VMI detector is gated on a range of TOFs by a fast voltage switch as to image only a ‘slice’ of the 3D momentum distribution. Scanning the gate then allows for the full 3D distribution to be measured. Unfortunately, not only does this method require multiple datasets to be taken to produce a single distribution, but it also effectively discards the majority of data by measuring only the small fraction of particles that fall within a given slice.
The most straightforward, though technically challenging, method for measuring the 3D distribution is by coincident measurement of each particle’s position and TOF. In this method the momentum components (p$_{x}$,p$_{y}$,p$_{z}$) are measured by (x,y,t) respectively, where x and y are the position of the particle on the surface of a detector and t is its TOF. This type of measurement has attracted significant interest in recent years despite the high temporal resolution (generally < 1 ns) needed to measure TOF, particularly with respect to electrons. Basnayake et al. performed 3D momentum measurements for electrons by correlating images with the output of a photomultiplier tube resulting in a temporal resolution of $\sim$30 ps \cite{basnayake2022three}. Cheng et al. accomplished similar measurements by coupling the signal from the phosphor anode of a microchannel plate (MCP) detector into an event-driven camera, TPX3CAM (TPX3) \cite{Amsterdam_instruments_2022}, which has two built-in time-to-digital converters (TDCs) \cite{cheng20223d}. The TPX3 TDCs (and overall detection) are limited to a TOF resolution of 260 ps, but with the tradeoff that the scheme is easily implementable on almost any VMI and requires no synchronization between devices. Further, Cheng et al. performed 3D momentum measurements for electrons and ions in coincidence using a single detector with a fast-switching voltage supply.
Although traditional (2D) VMI can measure electrons with transverse kinetic energy on the order of 1 keV, doing so requires increasing the gradient of the electrostatic field which reduces the TOF range of the particles; in turn this necessitates unrealizable temporal resolutions to achieve 3D measurement. This effectively means that measurements of this kind are limited to low energy while still requiring temporal resolution on the order of 100 ps. For reference, Cheng et al. run their VMI at voltages which limit it to the collection of electrons with kinetic energy less than $\sim$5 eV, but the electron TOF spread is still only $\sim$8 ns.
Here we expand on the work of both Cheng et al.\cite{cheng20223d} and Kling et al.\cite{kling2014thick} by using a TPX3 camera in conjunction with a coincident plano-convex thick-lens (PCTL) VMI and ion TOF spectrometer for measurement of 3D electron momentum distributions. By simply introducing a grounded mesh electrode at the far end of a thick-lens VMI, the electrostatic field forms a plano-convex lens which extends the detectable electron cutoff energy range while maintaining high spatial energy resolution. Additionally, the thick-lens serves to increase the electron TOF range, thus requiring less temporal resolution in electronics. The net result is the collection of electrons with transverse energies up to $\sim$7 eV while extending the electron TOF spread to $\sim$30 ns. Further, the coincident ion TOF measurement allows for significant noise reduction and is shown to be effective in extracting the unique 3D electron momentum distributions for multiple ion species measured simultaneously. These techniques for high resolution 3D momentum information have potential to lend themselves to more advanced measurements such as 3D holography , photoionization with intricate exciting fields (e.g. elliptical, two-color), and other systems for which electron momentum distributions are not cylindrically symmetric.
\begin{figure*}[ht!]
\includegraphics[width=\linewidth]{apparatus_v2}
\caption{\label{fig:apparatus}(a) is a depiction of the coincident PCTL-VMI and ion TOF spectrometer and its signal processing electronics. Electron-side electrodes are shown in blue and ion-side electrodes in orange. The repeller electrode is equipped with a mesh, as is an additional, grounded plate on the far end of the electron-side (grey). The lens assembly and detectors are encased in a mu-metal cylinder. Electrons are detected by an MCP stack with a phosphor screen anode and imaged by a TPX3 camera with two built in TDC inputs. Ions are detected by an MCP stack and stainless-steel anode. Both detectors are supplied high voltage through a decoupling circuit which allows the signal pulse from each detected particle to be separated and subsequently amplified. The amplified signals are each passed through a constant fraction discriminator (CFD) and a delay generator, the result of which is a positive TTL pulse. The output ion TTL pulses are subsequently combined with a reference laser trigger (also TTL) and the combined signal is recorded by TDC 1 of the TPX3 camera; the electron pulses are sent directly into TDC 2 of the TPX3 after the delay generator. (b) is a temporal depiction of the two TDC signals shown in a plot of voltage as a function of time. For each laser pulse, the laser trigger arrives first in TDC 1, followed by the electron signal in TDC 2, and then the ion signal which is also in TDC 1. (c) is a schematic of the VMI cross section in SIMION. In red are equipotential lines which show the plano-convex lens of the electron side. Electron trajectories for four different energies are shown in red, green, blue and yellow, and ion trajectories are shown in black.}
\end{figure*}
\section{Apparatus}
The general setup of the coincident PCTL-VMI and ion TOF spectrometer is depicted in Figure \ref{fig:apparatus}. Note the axes shown in Figure \ref{fig:apparatus} (c) are used throughout this manuscript, particularly, $\hat{z}$ is the axis of the electrostatic lens and $\hat{x}$ is the axis of laser propagation.
The electrostatic lens is formed by eleven stainless steel electrodes, five on the electron side (V1 through V5) and six on the ion side (repeller, iV1 through iV5). The position and potential of each electrode is detailed in Table \ref{tab:electrodes}. Both the electron and ion side electrodes are connected in a chain by 100 M$\Omega$ resistors, but the two sides are isolated from one another. In doing so, only four voltages need to be applied to the PCTL-VMI lens, one at either end of each electrode chain (V1, V5, repeller, and iV5), and the equal resistance resistors ensure a uniform voltage step between every two adjacent electrodes in each chain. The electron side geometry is based on a previous design \cite{kling2014thick}.
\begin{table}[b!]
\centering
\begin{tabularx}{\linewidth}{p{75pt}||p{75pt}|p{75pt}}
\hline
\textbf{Electrode} & \textbf{Position (mm)} & \textbf{Potential (V)} \\
\hline
\hline
Mesh & 106 & 0 \\
\hline
V5 & 94 & -37.3 \\
\hline
V4 & 89 & -49.7 \\
\hline
V3 & 84 & -62.1 \\
\hline
V2 & 79 & -74.6 \\
\hline
V1 & 74 & -87 \\
\hline
Repeller & 50 & -100 \\
\hline
iV1 & 40 & -100.2 \\
\hline
iV2 & 30 & -100.4 \\
\hline
iV3 & 20 & -100.6 \\
\hline
iV4 & 10 & -100.7\\
\hline
iV5 & 0 & -100.9\\
\end{tabularx}
\caption{\label{tab:electrodes}Geometry of the coincident PCTL-VMI and ion TOF spectrometer. Blue and orange shaded cells comprise the electron and ion side electrodes respectively. Electrodes for each side are connected by a series of equal resistance resistors to achieve a uniform step in potential across each two adjacent electrodes when voltages are applied at either end of the chain (V1 and V5, or repeller and iV5).}
\end{table}
In addition to the electrodes there is a grounded, copper mesh disk on the far end of the electron side. The mesh is 12 mm from the V5 electrode and 71 mm in diameter; it is commercially available from Precision eforming with 3.15 lines per mm and 80\% transmission efficiency. The grid is crucial to the formation of the plano-convex electrostatic lens on the electron side as it serves to be the ‘plano’ of the lens by providing a uniform voltage spanning the transverse profile of the lens. This is in contrast to typical annular electrodes which result in a convex field. The plano-convex field is clearly visible in Figure \ref{fig:apparatus} (c), a SIMION schematic of the VMI cross section in which the red lines indicate equipotential surfaces. Each of the five electrodes on the electron side are annular and, moving away from the interaction region, with each electrode the field lines become flatter up to the surface of the grid. The repeller is equipped with the same mesh except 23 mm in diameter; such a mesh on the repeller is not uncommon for a double sided VMI (or VMI-TOF). Note that when a mesh is referred to throughout the rest of this manuscript it is referring to the electron-side mesh, not the repeller mesh.
The electron detector is a stack of two 80 mm diameter microchannel plates (MCPs) in chevron configuration equipped with a P47 phosphor anode. The front face of the electron detector is grounded so, given that the mesh is also grounded, there is a 45 mm field free drift region between the exit of the plano-convex lens and the electron detector. The ion detector is a stack of two 25 mm MCPs (also in chevron configuration) equipped with a stainless-steel anode.
The lens electrodes, mesh, and detectors are encased in a mu-metal cylinder which shields the charged particles from any external magnetic field. The entire assembly is mounted in a high vacuum chamber with a baseline pressure of 2.5 x 10$^{-8}$ Torr.
External to the vacuum, the electron MCP and phosphor stack is imaged by a TPX3 camera which is equipped with two time-to-digital converters (TDC 1 and 2), both with a temporal resolution of 260 ps. Additionally, the TPX3 also natively records the time at which each pixel is activated (time of arrival, TOA) with a resolution of 1.6 ns. In our detection scheme TDC 1 is used to measure both the TOF for ions and a laser trigger, while TDC 2 is used for electron TOF. In this regard, the TPX3 camera represents a convenient platform for multiple laser-event synchronization. Pixel TOA is used to distinguish laser correlated electron events from noise by removing pixel events outside of a 1 $\mu$s window for each laser pulse. Further, because a pixel TOA is recorded for every pixel activation, a clustering algorithm is implemented to identify multiple pixel events as belonging to the same electron by evaluating their proximity in space and time. Clustering serves to effectively increase the spatial resolution of the camera \cite{bromberger2022shot}. Here we make use of an open-source algorithm, DBSCAN \cite{ester1996density}. Because we are grouping pixel events into clusters, we will associate each cluster with a cluster TOA that is the mean TOA, weighted by the duration of pixel activation (time-over-threshold, TOT), for the pixel events comprising that cluster. In practice, TOA is recorded by the TPX3 as an absolute time, however, throughout the rest of this manuscript we will refer to TOA as it is processed and referenced to a specific laser pulse. In other words, TOA is now the time difference between a laser timing reference signal and the recorded (absolute) TOA. This effectively makes the cluster TOA a measure of the electron TOF, but the 1.6 ns resolution is insufficient to resolve structure in p$_{z}$, so its use is limited to noise reduction.
A diagram of the electronic schemes for electron and ion detection is shown in Figure \ref{fig:apparatus} (a). Both the electron and ion detector anodes are supplied voltage through simple decoupling circuits which allow for single particles to be detected electronically. Signal pulses from the decoupling circuits are amplified and subsequently processed by constant fraction discriminators (CFDs, Ortec models 9327 and 583) with temporal jitter < 20 ps and < 75 ps for the electrons and ions respectively. Because the TPX3 TDCs call for a positive TTL pulse, the output of each CFD triggers a delay generator (SRS models DG535 for ions and DG645 for electrons) to produce such a pulse. We use two delay generators to avoid dead-time issues in the processing of the signals. All together the temporal resolution of the electronics configurations for both electrons and ions is limited by the resolution of the TDCs themselves at 260 ps.
In an event-driven experiment it is necessary to label each event; we achieve this by recording each laser pulse non-ambiguously using the TDC 1 channel of the TPX3. For this purpose, the ion TTL signal is first combined with a laser trigger TTL through a DC pass splitter/combiner (Mini-Circuits, model ZX10R-2-183-S+). This combined ion and laser trigger signal is critical as the current TPX3 model has only two TDCs, but to attain the 3D momentum distribution we require both the ion and electron signals, as well as the laser trigger to serve as reference for the electron and ion TOF. The laser trigger TTL pulse arrives well before the ion signal ($\sim$260 $\mu$s) meaning there is no concern of overlap, and has a significantly longer pulse width relative to the ion TTL (1 $\mu$s versus 50 ns) so that the two are easily distinguishable. The time sequence of the three signals as recorded by the two TDCs is shown in Figure \ref{fig:apparatus} (b).
\section{Results}
\subsection{Simulations}
To optimize the operating voltages and determine the energy range and resolution capabilities of the PCTL-VMI geometry shown in Figure \ref{fig:apparatus} (c), electron trajectory simulations were performed using SIMION \cite{dahl1990simion} (version 8.1). A key feature of the PCTL-VMI apparatus presented here is the capability to directly measure the 3D momentum distribution of electrons which requires the range and resolution of the PCTL-VMI to be assessed in two parts. First, the momentum of the electron transverse to the axis of the electrostatic lens (xy-plane) is related to the position of the electron on the MCP/phosphor detector; this is the standard operation mode for a 2D VMI. The corresponding (spatial) energy resolution, $\Delta$E/E, is proportional to the range of radii (distance from center of the detector, R) for electrons of a given energy. Specifically, because a VMI exhibits a linear relationship between the electron momentum and R, we can approximate $\Delta$E/E $\simeq$ 2$\Delta$R/R as is standard\cite{kling2014thick}. Here we take $\Delta$R to be the full width at half maximum (FWHM) of the electron radial distribution.
\begin{figure}[b!]
\centering
\includegraphics[width=80mm]{SIMION}
\caption{\label{fig:SIMION} Predicted energy resolution by SIMION simulations for spatially (a) and temporally (b) resolved electron energies. Note that negative energies in (b) correspond to electrons with velocity in -$\hat{z}$. Simulations were run with the thick-lens VMI for both plano-convex (PCTL-VMI) and conventional convex (CTL-VMI) electrostatic lens configurations, for a fixed -100 V applied to the repeller. The source volumes was taken as a 1 mm radius, 2 mm length cylinder and 5000 particles were flown for each electron energy.}
\end{figure}
Additionally, the component of momentum along the axis of the detector, p${_z}$, is measured by the electron TOF. Assessing this (temporal) $\Delta$E/E is less straightforward as there is not a linear relationship between the TOF and momentum. Further, for electrons emitted with momentum $\pm$$p_{z}$, there is a lack of symmetry up vs down in the electrostatic field. Thus, to evaluate the temporal energy resolution we require an additional simulation to serve as a reference and reliably relate a given electron TOF with its initial momentum. This reference simulation is performed with electrons originating from a single, ‘ideal’, point with energies from 0 eV to 7 eV in steps of 0.001 eV and initial velocity in both $\pm$$\hat{z}$. The reference simulation is presented later in the text as a comparison to experimental results and can be seen in Figure \ref{fig:calibration} (d). Using this simulation to correlate electron TOF to momentum $\Delta$E/E is now evaluated in a manner similar to before: $\Delta$E is the energy range corresponding to the FWHM of the TOF distribution for a given kinetic energy.
The simulated spatial and temporal energy resolutions of the PCTL-VMI are presented in Figure \ref{fig:SIMION} (a) and (b) respectively. The electron source volume was taken to be a cylinder of 1 mm radius and 2 mm length with its axis along $\hat{x}$. Electrons have initial momentum corresponding to kinetic energies from 0 eV to 7.5 eV in steps of 0.25 eV; this range is particularly relevant as it corresponds to low (up to fifth) order ATI from 800 nm light. The initial electron momenta are along $\hat{y}$ to determine the spatial energy resolution (Figure \ref{fig:SIMION} (a)) or $\pm$$\hat{z}$ for the temporal energy resolution (Figure \ref{fig:SIMION} (b)). Note that in Figure \ref{fig:SIMION} (b) positive and negative kinetic energies correspond to initial momenta in the respective $\pm$$\hat{z}$ direction.
Each plot of $\Delta$E/E for the PCTL-VMI shown in Figure \ref{fig:SIMION} is accompanied by simulation results for the same VMI geometry except with a conventional (convex) thick-lens (labelled CTL-VMI). This conversion from plano-convex to convex is achieved simply by removing the mesh and reoptimizing the electrode voltages while keeping the repeller voltage fixed at -100 V. Note that for both configurations the electrode voltages are optimized with the restriction that they are connected by a chain of equal resistance resistors, that is, there is a uniform voltage drop between adjacent electrodes. Thus, for a fixed repeller voltage, the optimization of the electron-side lens is a two-parameter problem consisting of the voltages applied on the ends of the electrode stack (electrodes V1 and V5). With optimized electrode voltages, it is shown in Figure \ref{fig:SIMION} (a) that both the PCTL-VMI and CTL-VMI configurations have spatial energy resolution better than or comparable to the resolution of most modern VMI designs over their respective energy ranges \cite{kling2014thick,zhang2014method,zhang2016aplanatic,ding2021composite}.
Although electrons were simulated with energies up to 7.5 eV, Figure \ref{fig:SIMION} (a) shows that, for initial velocities in $\hat{y}$ and the repeller at -100 V, the PCTL-VMI collects electrons up to a cutoff of 6.75 eV as compared to just 4.25 eV for the CTL-VMI. Further, while the CTL-VMI has about twice the resolution of the PCTL-VMI for electron energies between 0.5 eV and 2 eV, the PCTL-VMI configuration retains a high energy resolution ($\Delta$E/E < 1\%) over a range more than twice that of its convex counterpart. The energy resolution curve of the CTL-VMI is tunable by the voltage on V1 such that the $\Delta$E/E minimum can be shifted between 1.5 eV and 3.5 eV, but the cutoff energy cannot be extended without significantly compromising the resolution. In fact, tuning the voltage for better resolution at the higher end of this energy range further limits the cutoff energy by about 0.75 eV ($\sim$18\%). The PCTL-VMI configuration exhibits a similar dependence on the V1 electrode but because it already has a high energy resolution over much of its collection range, optimizing the $\Delta$E/E curve for specific energies comes at the cost of resolution over the rest of the range. In general, the PCTL-VMI is shown more valuable for reliable spatial energy resolution over a large range of electron energies, whereas the CTL-VMI may be preferred in instances where extremely high resolution is required over a narrow range of energies. With respect to 3D momentum measurements, measurement of p$_{x}$ and p$_{y}$ is already extremely limited by the low voltages required to maintain a resolvable TOF spread. In this case the PCTL-VMI design offers a means to extend the detectable energy range without increasing voltages, thus retaining a broad TOF range.
\begin{figure*}[b!]
\includegraphics[width=\linewidth]{raw_v3}
\caption{\label{fig:raw} Ionization of Xe gas with linearly polarized light oriented in $\hat{y}$ for $\sim$600k laser pulses. (a) is a histogram of cluster TOA and ion TOF (iTOF); also overlain is a portion of the TOF spectrum scaled by a factor of 100 to visualize other ion species which are present. (b) is a clustered VMI image with no gating used, i.e., every pixel activation included, while (c) is a clustered VMI image with gating on both cluster TOA (1 $\mu$s gate, centered on the pixel TOA peak in (a)) and ion TOF (1.75 $\mu$s, vertical black lines in (a)). (c) also shows the projection of this image onto the polarization axis, $\hat{y}$.}
\end{figure*}
Figure \ref{fig:SIMION} (b) shows the temporal energy resolution for the two configurations. Here the PCTL and CTL are almost indistinguishable as the electron TOF is most dependent on the potential difference between the repeller and V1 rather than the lensing itself. The temporal energy resolution is shown to be less than its spatial resolution, but it compares well to the resolution of some current, conventional VMI designs (e.g. this design by Garcia et al. \cite{garcia2013delicious}). There are two other points to be made regarding Figure \ref{fig:SIMION} (b); first, although the simulated electrons only have velocity along the axis of the lens ($\hat{z}$), it is shown that the CTL-VMI cannot collect electrons with velocities corresponding to more than $\sim$5.8 eV in -$\hat{z}$ (E < -5.8 eV in Figure \ref{fig:SIMION} (b)). This is because, given enough velocity along -$\hat{z}$, an electron can escape the region between the repeller and V1 into the (largely) field free ion side of the apparatus and will not have a trajectory that ends at the electron detector. To this end, the PCTL does have a slight advantage over the CTL because its optimized V1 voltage is slightly less negative than in the CTL configuration. On the other hand, although trajectories for electrons with initial velocity in +$\hat{z}$ will always reach the electron detector and are shown to have reasonable energy resolution, this does not mean that in practice we can resolve energies in this regime. The issue is that although $\Delta$E/E appears reasonable, the differences in the electron TOF become so small for similar energies that the TPX3 TDC does not have the temporal resolution to distinguish them. We have found (experimentally) for our setup we are able to resolve energetic structures for electrons up to $\sim$3.5 eV with velocities along +$\hat{z}$ and $\sim$7.5 eV for velocities along -$\hat{z}$ (see Figure \ref{fig:calibration} (b) and (d)).
\subsection{Experiment}
\begin{figure*}[b!]
\centering
\includegraphics[width=150mm]{calibration_v2}
\caption{\label{fig:calibration} Spatial and temporal energy calibration of the PCTL-VMI. (a) is a normalized histogram of the radius (distance from center of mass) for cluster events corresponding to Xe singly ionized by linearly polarized laser pulses oriented in $\hat{y}$ (image shown in Figure \ref{fig:raw} (c)). It is fit by the sum of multiple Gaussian functions to find the position of ATI peaks. The ATI peaks are plotted against their energy and fit to a quadratic then compared to SIMION results in (c). Note that though five ATI peaks are fit in (a), the fifth is close to the edge of the detector and thus omitted from the calibration in (c). Similarly, (b) is a histogram of electron TOF (eTOF) for Xe singly ionized by linearly polarized laser pulses but now oriented in $\hat{z}$. This is also fit to a sum of Gaussian functions and subsequently the ATI peaks are fit to a fifth order polynomial and compared to SIMION results in (d). Note that the x-axis in (b) and (d) has been shifted such that a TOF of zero corresponds to zero kinetic energy (denoted by a dashed vertical line in (b) and (d)). Also note that negative energies correspond to initial velocities in -$\hat{z}$.}
\end{figure*}
To demonstrate the 3D measurement capabilities of PCTL-VMI, above-threshold ionization (ATI) was observed in Xenon (Xe). For these experiments Xe gas was injected into the VMI chamber to a pressure of 2.2 x 10$^{-7}$ Torr. The voltages applied to the electrostatic lens were measured to be within $\pm$0.2\% of the target voltages shown in Table \ref{tab:electrodes}. Ionization is achieved with a Titanium:Sapphire system producing 30 fs pulses at a repetition rate of 1 kHz. Pulses were spectrally filtered to a 10 nm width centered at 800 nm using an interferometric band pass filter yielding a temporal FWHM (in intensity) of 80 fs with a pulse energy of 12.6 $\mu$J. A wire grid polarizer and $\lambda$/2 plate (both Thorlabs products) are used to clean and orient the laser polarization before being focused into the VMI chamber by a f = +300 mm focusing lens. The intensity in the interaction region is estimated to be on the order of 10$^{13}$ W/cm$^{2}$. The net result of these conditions is that $\sim$15\% of laser pulses have a Xe ionization event.
In Figure \ref{fig:raw} we compare VMI images of Xe$^{+}$ before and after processing. Figure \ref{fig:raw} (a) and (b) present the ion TOF spectrum and clustered VMI image for linearly polarized light oriented along $\hat{y}$. Also included in Figure \ref{fig:raw} (a) is a histogram of cluster TOA. Figure \ref{fig:raw} (b) is clustered, but otherwise there is no data processing, meaning that it includes every pixel activation during data collection. Figure \ref{fig:raw} (c) is a processed image where cluster events are gated by two means. First, we remove clusters with a TOA outside a 1 $\mu$s window (centered about the cluster TOA peak in Figure \ref{fig:raw} (a)) as they are not correlated with electrons emitted by a laser pulse. In doing so $\sim$40\% of clusters are removed, which includes any faulty or dead pixels. Second, we gate on the ion species of interest. Specifically, we only take clusters that occur in coincidence with an ion measured within a particular TOF range. Here we gate on a 1.75 $\mu$s window in the ion TOF (vertical black lines in Figure \ref{fig:raw} (a)) corresponding to Xe$^{+}$.
Much in the same way that the energy resolution for the PCTL-VMI was simulated spatially and temporally, we now require both spatial and temporal energy calibration to access the 3D momentum distribution; this is shown in Figure \ref{fig:calibration}. Spatial calibration is done in a standard manner: ATI peaks for a VMI image (with laser polarization in the xy-plane) are fit under the condition that each is separated by a photon energy and that clusters at R = 0 correspond to having zero kinetic energy. The ATI peaks and corresponding calibration fit are shown to be in good agreement compared to SIMION simulations (Figure \ref{fig:calibration} (a) and (c)). Temporal energy calibration requires a separate dataset to be taken with the laser polarization oriented along $\hat{z}$ so that the ATI are now clearly discernible in the electron TOF spectrum (shown in Figure \ref{fig:calibration} (b)). Notice that in this ATI spectrum there is an additional peak corresponding to zero kinetic energy along $\hat{z}$ present in the electron TOF spectrum. This feature can also be seen in Figure \ref{fig:raw} (c) as the projection of the image on the polarization axis which is now effectively being measured as a TOF distribution. By calibrating the spatial ATI first we have the absolute energy for each remaining (E $\neq$ 0) peak in the electron TOF spectrum. The ATI energies are plotted as a function of TOF and fit to a fifth order polynomial to serve as a conversion from TOF to energy. Figure \ref{fig:calibration} (d) shows the measured ATI TOFs and conversion fit to be in good agreement with SIMION simulations.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{threeD_v3}
\caption{\label{fig:threeD} 3D momentum distribution results for singly ionized Xe by linearly polarized light. (a) is a 3D histogram of the electron momentum distribution with the laser polarization along $\hat{y}$ and plotted in atomic momentum units (amu). (b) and (c) are slices from this 3D histogram in the xy and yz plane respectively. (d) and (e) present slices in the yz and xy planes (respectively), but for the laser polarization aligned with the $\hat{z}$-axis. Red lines denote axes for which the momentum is measured by the electron TOF (p$_{z}$).}
\end{figure}
It is important to note that with the PCTL-VMI design the electron TOF spectrum is far broader than with previously reported 3D VMIs. For example, Cheng et al. observe electrons of $\pm$2.12 eV over a range of $\sim$5 ns \cite{cheng20223d}, as compared to $\sim$10 ns for the PCTL-VMI as shown in Figure \ref{fig:calibration}. This means that, for the same timing resolution, the PCTL-VMI design presented here will have greater overall TOF energy resolution. Further, even in the low voltage configuration required for a broad electron TOF spectrum, the PCTL-VMI also retains a higher spatial photoelectron cutoff energy than that shown by Cheng et al.
Figure \ref{fig:threeD} (a) presents the calibrated 3D histogram of the electron momentum distribution, for laser polarization along $\hat{y}$, in coincidence with Xe$^{+}$. This is the same dataset presented in Figure \ref{fig:raw} and Figure \ref{fig:calibration} (a) and (c). Figure \ref{fig:threeD} also presents slices of the 3D histogram for p$_z$ = 0 (b) and p$_x$ = 0 (c) which are then compared to slices from the same experimental conditions, but now with the laser polarized along $\hat{z}$ (Figure \ref{fig:threeD} (d) and (e) anti-respectively). In doing so we have simply rotated the momentum distribution by 90$^{o}$, and, as expected, the p$_z$ = 0 slice for $\hat{y}$ polarized light (b) is in good agreement with the p$_x$ = 0 slice for $\hat{z}$ polarized light (d). The same can be said of Figure \ref{fig:threeD} (c) and (e). Note that this figure shows the VMI lens does exhibit some aberrations, particularly in (c) and (d). From (b) we can see that the momentum distribution is not perfectly centered on the detector in $\hat{y}$, indicating that the laser focus is not properly positioned along this axis. That being the case, we suspect electrons with sufficiently high momentum in -$\hat{z}$ can have trajectories which approach the V1 electrode too closely, altering their course. This results in a measurable effect on the measured y coordinate of the cluster, but is negligible in TOF: an apt description of what is seen in (c) and (d).
\begin{figure*}[t!]
\includegraphics[width=\linewidth]{coin_v3}
\caption{\label{fig:coin} Ion coincidence of the 3D electron momentum distribution. (a) is a histogram of ion events in mass to charge ratio (m/q) for a mix of ambient air and a small amount of Xe gas for 240M laser shots. A portion of the histogram is scaled by a factor of 50 and overlain for easy visualization of the less abundant species. The full 3D histogram was gated for electrons measured in coincidence with H$_{2}$O$^{+}$, N$_{2}^{+}$, and Xe$^{+}$ (vertical black lines in (a)) and a histogram of the total (3D) momentum for each is shown in (b). Slices of each 3D histogram at p$_{x}$ = 0 are shown for each ion in (c), (d), and (e). Additionally, slices of N$_{2}^{+}$ for p$_{x}$ = $\pm$0.2 amu, and p$_{x}$ = $\pm$0.4 amu are shown in (f).}
\end{figure*}
From Figure \ref{fig:threeD} we can draw two conclusions. First and foremost, by comparing results for laser polarization along $\hat{y}$ and $\hat{z}$ we can see that the 3D momentum distribution can be accurately retrieved in a single measurement. Particularly, in this we have verified reliable measurement of p$_{z}$ by measurement of TOF. Second, as predicted by the SIMION simulations, the energy (or momentum) resolution along the TOF axis is notably less than that of the spatially resolved axes. For example, splitting of the first ATI ring can be seen in both Figures \ref{fig:raw} (c) and \ref{fig:threeD} (b) which is not observed in \ref{fig:threeD} (d). This can also be seen by comparing the width of each ATI ring in Figure \ref{fig:threeD} (c) and (e). Despite this, having access to the 3D histogram provides information that was previously unavailable from the standard (projected) VMI image, e.g. the structures perpendicular to the polarization axis ($\hat{x}$) in the 2D image Figure \ref{fig:raw} (c) are shown to be rings around $\hat{y}$ in Figure \ref{fig:threeD} (a). Further, each ATI is a spherical shell with preferential emission along the polarization axis.
Finally, to demonstrate the efficacy of the ion coincidence scheme, a dataset was taken for a mix of ambient air and Xe gas using light polarized in $\hat{z}$. To ionize molecular species in air the pulse energy was raised to 48.8 $\mu$J (as compared to 12.6 $\mu$J for all other results presented). A relatively large dataset was taken, $\sim$240M laser shots over 3 days, to ensure good statistics for each ion species. Figure \ref{fig:coin} (a) shows the ion TOF spectrum for this run; of particular interest are the three dominating species: H$_{2}$O$^{+}$, N$_{2}^{+}$, and Xe$^{+}$. Histograms of the total (3D) momentum were produced for electrons in coincidence with each of these ion species and are shown in Figure \ref{fig:coin} (b). It is seen clearly here that there is a shifting of ATI peaks, due in part to differing ionization potentials (I$_{p}$). This can also be attributed to laser intensity effects, that is, ions with lower I$_{p}$ will be ionized from a larger volume and thus from a larger range of field intensities which results in a change in ATI structure. Further, Figure \ref{fig:coin} (c) through (e) present slices of the 3D histogram for each ion at p$_{x}$ = 0 and (f) presents additional slices for N$_{2}^{+}$ at p$_{x}$ = $\pm$0.2 amu, and p$_{x}$ = $\pm$0.4 amu. This figure shows that not only are the ATI peaks shifted, but there is an entirely different angular momentum structure for each species. The cumulative result is three unique ATI spectra, measured within a single dataset, as expected for three ionic species with vastly different electronic structure.
\section{Conclusion}
We have demonstrated the efficacy of a plano-convex electrostatic field in a thick-lens VMI design to enhance its 3D electron momentum resolution capabilities. SIMION simulations show that, with the simple addition of a mesh electrode to form the plano-convex field, the PCTL-VMI not only retains its high spatial energy resolution but also extends the detectable electron cutoff energy for a given repeller voltage. This is of particular importance for 3D measurements which rely on the electron TOF as they necessitate low repeller voltages to maintain a large enough TOF spread to resolve energetic features. Further, the thick-lens serves to spread the electron TOF distribution, thus requiring less temporal resolution of the electronic detection scheme to attain similar overall energy resolution. We have confirmed these simulations experimentally by demonstrating the PCTL-VMI capability to collect electrons up to $\sim$7 eV with a TOF spread of $\sim$30 ns, both being improvements over previous work by factors of $\sim$1.4 and $\sim$3.75 respectively. Additionally, the PCTL-VMI is equipped with a coincident ion TOF spectrometer which allows for effective gating of electrons from different ionic species in a gas mixture and recovers the unique 3D electron momentum distribution for each. These techniques have the potential to lend themselves to more advanced measurements, particularly involving systems where the electron momentum distributions possess non-trivial symmetries.
\section{Acknowledgments}
The authors would like to recognize the Air Force Office of Scientific Research (AFOSR) for financial support of this work under grant no. FA9550-21-1-0387. T.S. was partially funded by the U.S. Department of Energy (DOE) under grant no. DE-SC0019098. C.C., G.M., and T.W. gratefully acknowledge support from the Department of Energy, Basic Energy Sciences Division, under Award No. DEF-G02-08ER15983.
\section{Data Availability Statement}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
\section{Conflict of Interest}
The authors have no conflicts to disclose.
\bibliographystyle{unsrt}
\section{Introduction}
\lipsum[2]
\lipsum[3]
\section{Headings: first level}
\label{sec:headings}
\lipsum[4] See Section \ref{sec:headings}.
\subsection{Headings: second level}
\lipsum[5]
\begin{equation}
\xi _{ij}(t)=P(x_{t}=i,x_{t+1}=j|y,v,w;\theta)= {\frac {\alpha _{i}(t)a^{w_t}_{ij}\beta _{j}(t+1)b^{v_{t+1}}_{j}(y_{t+1})}{\sum _{i=1}^{N} \sum _{j=1}^{N} \alpha _{i}(t)a^{w_t}_{ij}\beta _{j}(t+1)b^{v_{t+1}}_{j}(y_{t+1})}}
\end{equation}
\subsubsection{Headings: third level}
\lipsum[6]
\paragraph{Paragraph}
\lipsum[7]
\section{Examples of citations, figures, tables, references}
\label{sec:others}
\lipsum[8] \cite{kour2014real,kour2014fast} and see \cite{hadash2018estimate}.
The documentation for \verb+natbib+ may be found at
\begin{center}
\url{http://mirrors.ctan.org/macros/latex/contrib/natbib/natnotes.pdf}
\end{center}
Of note is the command \verb+\citet+, which produces citations
appropriate for use in inline text. For example,
\begin{verbatim}
\citet{hasselmo} investigated\dots
\end{verbatim}
produces
\begin{quote}
Hasselmo, et al.\ (1995) investigated\dots
\end{quote}
\begin{center}
\url{https://www.ctan.org/pkg/booktabs}
\end{center}
\subsection{Figures}
\lipsum[10]
See Figure \ref{fig:fig1}. Here is how you add footnotes. \footnote{Sample of the first footnote.}
\lipsum[11]
\begin{figure}
\centering
\fbox{\rule[-.5cm]{4cm}{4cm} \rule[-.5cm]{4cm}{0cm}}
\caption{Sample figure caption.}
\label{fig:fig1}
\end{figure}
\subsection{Tables}
\lipsum[12]
See awesome Table~\ref{tab:table}.
\begin{table}
\caption{Sample table title}
\centering
\begin{tabular}{lll}
\toprule
\multicolumn{2}{c}{Part} \\
\cmidrule(r){1-2}
Name & Description & Size ($\mu$m) \\
\midrule
Dendrite & Input terminal & $\sim$100 \\
Axon & Output terminal & $\sim$10 \\
Soma & Cell body & up to $10^6$ \\
\bottomrule
\end{tabular}
\label{tab:table}
\end{table}
\subsection{Lists}
\begin{itemize}
\item Lorem ipsum dolor sit amet
\item consectetur adipiscing elit.
\item Aliquam dignissim blandit est, in dictum tortor gravida eget. In ac rutrum magna.
\end{itemize}
\section{Conclusion}
Your conclusion here
\section*{Acknowledgments}
This was was supported in part by......
\bibliographystyle{unsrt}
|
\section{#1}\setcounter{equation}{0}\setcounter{theorem}{0}}
\newcommand{\hfill\nonumber}{\hfill\nonumber}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{definition-theorem}[theorem]{Definition-Theorem}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\newtheorem{remark}[theorem]{Remark}
\newtheorem{conjecture}[theorem]{Conjecture}
\newtheorem{notation}[theorem]{Notation}
\newcommand \bth[1] { \begin{theorem}\label{t#1} }
\newcommand \ble[1] { \begin{lemma}\label{l#1} }
\newcommand \bdeth[1] { \begin{definition-theorem}\label{dt#1} }
\newcommand \bpr[1] { \begin{proposition}\label{p#1} }
\newcommand \bco[1] { \begin{corollary}\label{c#1} }
\newcommand \bde[1] { \begin{definition}\label{d#1}\rm }
\newcommand \bex[1] { \begin{example}\label{e#1}\rm }
\newcommand \bre[1] { \begin{remark}\label{r#1}\rm }
\newcommand \bcj[1] { \begin{conjecture}\label{j#1}\rm }
\newcommand \bqu[1] { \medskip\noindent{\it{Question #1}} }
\renewcommand {\eth} { \end{theorem} }
\newcommand {\ele} { \end{lemma} }
\newcommand {\edeth}{ \end{definition-theorem} }
\newcommand {\epr} { \end{proposition} }
\newcommand {\eco} { \end{corollary} }
\newcommand {\ede} { \end{definition} }
\newcommand {\eex} { \end{example} }
\newcommand {\ere} { \end{remark} }
\newcommand {\ecj} { \end{conjecture} }
\newcommand {\equ} {\smallskip}
\newcommand {\enota} { \end{notation} }
\newcommand \thref[1]{Theorem \ref{t#1}}
\newcommand \leref[1]{Lemma \ref{l#1}}
\newcommand \prref[1]{Proposition \ref{p#1}}
\newcommand \coref[1]{Corollary \ref{c#1}}
\newcommand \cjref[1]{Conjecture \ref{j#1}}
\newcommand \deref[1]{Definition \ref{d#1}}
\newcommand \exref[1]{Example \ref{e#1}}
\newcommand \reref[1]{Remark \ref{r#1}}
\newcommand \lb[1]{\label{#1}}
\def \mathfrak{d} {{\partial}}
\def \mathrm{d} {\mathrm{d}}
\def \Rset {{\mathbb R}}
\def {\mathbb C} {{\mathbb C}}
\def {\mathbb K} {{\mathbb K}}
\def {\mathbb Z} {{\mathbb Z}}
\def {\mathbb N} {{\mathbb N}}
\def {\mathbb Q} {{\mathbb Q}}
\def {\mathbb P} {{\mathbb P}}
\def {\mathbb T} {{\mathbb T}}
\def {\Cset \Pset^1} {{{\mathbb C} {\mathbb P}^1}}
\def \Aa {{\mathcal A}}
\def \Ab {{\mathcal A}_q}
\def {\mathcal A}_q^{1/2} {{\mathcal A}_q^{1/2}}
\def {\mathcal A}_\varepsilon {{\mathcal A}_\varepsilon}
\def {\mathcal A}_\varepsilon^{1/2} {{\mathcal A}_\varepsilon^{1/2}}
\def \A {{\boldsymbol{\mathsf A}}}
\def \C {{\boldsymbol{\mathsf C}}}
\def \U {{\boldsymbol{\mathsf U}}}
\def {\widetilde{B}} {{\widetilde{B}}}
\newcommand \ex {{\bf{ex}}}
\newcommand \inv {{\bf{inv}}}
\newcommand \ninv {{\bf{ninv}}}
\newcommand \nc {{\bf{nc}}}
\def {\boldsymbol{\cdot}} {{\boldsymbol{\cdot}}}
\def {\mathrm{im}} {{\mathrm{im}}}
\def \x {{\widetilde{\mathbf{x}}}}
\newcommand{\genfrac{[}{]}{0pt}{}}{\genfrac{[}{]}{0pt}{}}
\def \AA {{\mathcal{A}}}
\def {\mathcal{B}} {{\mathcal{B}}}
\def{\mathcal{D}}{{\mathcal{D}}}
\def {\mathcal{I}} {{\mathcal{I}}}
\def {\mathcal{Q}} {{\mathcal{Q}}}
\def {\mathcal{P}} {{\mathcal{P}}}
\def {\mathcal{J}} {{\mathcal{J}}}
\def {\mathcal{V}} {{\mathcal{V}}}
\def {\mathcal{U}} {{\mathcal{U}}}
\def {\mathcal{R}} {{\mathcal{R}}}
\def {\mathcal{S}} {{\mathcal{S}}}
\def {\mathcal{L}} {{\mathcal{L}}}
\def {\mathscr{L}} {{\mathscr{L}}}
\def \TT {{\mathcal{T}}}
\def {\mathcal{Z}} {{\mathcal{Z}}}
\def {\mathcal{O}} {{\mathcal{O}}}
\def {\mathscr{O}} {{\mathscr{O}}}
\def {\mathcal{E}} {{\mathcal{E}}}
\def {\mathcal{F}} {{\mathcal{F}}}
\def {\mathcal{X}} {{\mathcal{X}}}
\def {\mathcal{Y}} {{\mathcal{Y}}}
\def {\mathcal{Z}} {{\mathcal{Z}}}
\def {\textbf{p}} {{\textbf{p}}}
\def {\textbf{r}} {{\textbf{r}}}
\def {\textbf{r}} {{\textbf{r}}}
\def {\textbf{r}} {{\textbf{r}}}
\def \De {\Delta}
\def \delta {\delta}
\def \alpha {\alpha}
\def \beta {\beta}
\def \varpi {\varpi}
\def \lambda {\lambda}
\def \Lambda {\Lambda}
\def\kappa{\kappa}
\def\Kappa{\Kappa}
\def \omega {\omega}
\def \Omega {\Omega}
\def \gamma {\gamma}
\def \delta {\delta}
\def \Gamma {\Gamma}
\def \Sigma {\Sigma}
\def \sigma {\sigma}
\def \Psi {\Psi}
\def \varphi {\varphi}
\def \varepsilon {\varepsilon}
\def \oplus {\oplus}
\def \mapsto {\mapsto}
\def \ra {\rightarrow}
\def \longrightarrow {\longrightarrow}
\def \tilde {\tilde}
\def \hookrightarrow {\hookrightarrow}
\def \leftharpoonup {\leftharpoonup}
\def \rightharpoonup {\rightharpoonup}
\def \subset {\subset}
\def \supset {\supset}
\def \ci {\circ}
\def \bullet {\bullet}
\def \rangle {\rangle}
\def \langle {\langle}
\def \partial {\partial}
\def \otimes {\otimes}
\def \overline {\overline}
\def \widetilde {\widetilde}
\def \widehat {\widehat}
\def \hspace{.2in} {\hspace{.2in}}
\newcommand\pr[2]{\langle{#1},{#2}\rangle}
\def {\mathrm{Id}} { {\mathrm{Id}} }
\def {\mathrm{id}} { {\mathrm{id}} }
\def {\mathrm{st}} { {\mathrm{st}} }
\def {\mathrm{Stab}} { {\mathrm{Stab}} }
\def \coeff { {\mathrm{coeff}} }
\def {\mathrm{sign}} { {\mathrm{sign}} }
\def {\mathrm{Ad}} { {\mathrm{Ad}} }
\def {\mathrm{const}} { {\mathrm{const}} }
\def {\mathrm{rank}} { {\mathrm{rank}} }
\def {\mathrm{nrf}} { {\mathrm{nrf}} }
\def {\mathrm{Lie \,}} { {\mathrm{Lie \,}} }
\def {\mathrm{reg}} { {\mathrm{reg}} }
\def {\mathrm{sing}} { {\mathrm{sing}}}
\def \g {\mathfrak{g}}
\def \mathfrak{gl} {\mathfrak{gl}}
\def \mathfrak{sl} {\mathfrak{sl}}
\def \mathfrak{h} {\mathfrak{h}}
\def \mathfrak{d} {\mathfrak{d}}
\def \mathfrak{f} {\mathfrak{f}}
\def \mathfrak{n} {\mathfrak{n}}
\def \mathfrak{m} {\mathfrak{m}}
\def \mathfrak{b} {\mathfrak{b}}
\def \mathfrak{p} {\mathfrak{p}}
\def \mathfrak{r} {\mathfrak{r}}
\def \mathfrak{z} {\mathfrak{z}}
\def \mathfrak{l} {\mathfrak{l}}
\def \mathfrak{l} {\mathfrak{l}}
\def \mathfrak{sl} {\mathfrak{sl}}
\def \mathrm{k} {\mathrm{k}}
\def M_{m,n} {M_{m,n}}
\def (X^+_i) {(X^+_i)}
\def (X^+) {(X^+)}
\def (X^-) {(X^-)}
\def (X^-_i) {(X^-_i)}
\DeclareMathOperator \Cl {{\mathrm{Cl} }}
\DeclareMathOperator \Span { {\mathrm{Span}} }
\DeclareMathOperator \Aut { {\mathrm{Aut}} }
\DeclareMathOperator \Max { {\mathrm{Max}} }
\DeclareMathOperator \charr { {\mathrm{char}} }
\DeclareMathOperator \GSV {{\mathrm{GSV}}}
\DeclareMathOperator \PSpec { P \mbox{-} {\mathrm{Spec}}}
\DeclareMathOperator \PPrim { P \mbox{-} {\mathrm{Prim}}}
\DeclareMathOperator \mSpec { m \mbox{-} {\mathrm{Spec}}}
\DeclareMathOperator \ord { {\mathrm{ord}} }
\DeclareMathOperator \ad { {\mathrm{ad}} }
\DeclareMathOperator \diag { {\mathrm{diag}} }
\DeclareMathOperator \Ker { {\mathrm{Ker}} }
\DeclareMathOperator \tr { {\mathrm{tr}} }
\DeclareMathOperator \rk { {\mathrm{rk}} }
\DeclareMathOperator \qtr { {\mathrm{qtr}} }
\DeclareMathOperator \End { {\mathrm{End}} }
\DeclareMathOperator \Gr { {\mathrm{Gr}} }
\DeclareMathOperator \Hom { {\mathrm{Hom}} }
\DeclareMathOperator \Dom { {\mathrm{Dom}} }
\DeclareMathOperator \Ann { {\mathrm{Ann}} }
\DeclareMathOperator \Ind { {\mathrm{Ind}} }
\DeclareMathOperator \MaxSpec { {\mathrm{MaxSpec}} }
\DeclareMathOperator \Irr { {\mathrm{Irr}} }
\DeclareMathOperator \Res { {\mathrm{Res}} }
\DeclareMathOperator \gr { {\mathrm{gr}} }
\DeclareMathOperator \Ext { {\mathrm{Ext}} }
\DeclareMathOperator \GKdim {{\mathrm{GK \, dim}}}
\DeclareMathOperator \Symp { {\mathrm{Symp}} }
\DeclareMathOperator \Dix { {\mathrm{Dix}} }
\DeclareMathOperator \Prim { {\mathrm{Prim}} }
\DeclareMathOperator \Supp { {\mathrm{Supp}} }
\DeclareMathOperator \Der { {\mathrm{Der}}}
\renewcommand \Im { {\mathrm{Im}} }
\renewcommand \Re { {\mathrm{Re}} }
\renewcommand \max { {\mathrm{max}} }
\newcommand \Spec { {\mathrm{Spec}} }
\newcommand \Fract {{\mathrm{Fract}}}
\begin{document}
\title
[Poisson geometry and Azumaya loci of cluster algebras]
{Poisson geometry and Azumaya loci \\ of cluster algebras}
\author[G. Muller]{Greg Muller}
\thanks{B. N. has been supported supported by an AMS-Simons travel grant and M.Y. by NSF grants DMS-1901830 and DMS-2131243.}
\address{Department of Mathematics, University of Oklahoma, Norman, OK 73019}
\email{<EMAIL>}
\author[B. Nguyen]{Bach Nguyen}
\address{
Department of Mathematics, Xavier University of Louisiana, New Orleans, LA 70125}
\email{<EMAIL>}
\author[K. Trampel]{Kurt Trampel}
\address{
Department of Mathematics, University of Notre Dame, Notre Dame, IN, 46556}
\email{<EMAIL>}
\author[M. Yakimov]{Milen Yakimov}
\address{
Department of Mathematics, Northeastern University, Boston, MA 02115}
\email{<EMAIL>}
\keywords{Cluster algebras, Gekhtman--Shapiro--Vainshtein Poisson brackets, torus orbits of symplectic leaves,
root of unity quantum cluster algebras, fully Azumaya loci, Poisson orders}
\subjclass[2010]{Primary: 13F60, Secondary: 53D17, 16G30, 17B37}
\begin{abstract} There are two main types of objects in the theory of cluster algebras: the upper cluster algebras $\U$
with their Gekhtman--Shapiro--Vainshtein Poisson brackets and their root of unity quantizations $\U_\varepsilon$.
On the Poisson side, we prove that (without any assumptions) the spectrum of every finitely generated upper cluster algebra $\U$ with its GSV Poisson structure
always has a Zariski open orbit of symplectic leaves and give an explicit description of it.
On the quantum side, we describe the fully Azumaya loci of the quantizations $\U_\varepsilon$ under the assumption that
$\A_\varepsilon = \U_\varepsilon$ and $\U_\varepsilon$ is a finitely generated algebra. All results allow frozen variables to be either inverted or not.
\end{abstract}
\maketitle
\sectionnew{Introduction}
\lb{Intro}
\subsection{Setting}
\label{1.1}
Cluster Algebras were defined by Fomin and Zelevinsky \cite{FZ1} in 2001, and since then have played a prominent role in
many areas of mathematics and mathematical physics. There are two main classes of algebras in this theory. The algebras in these classes are defined
from an integer matrix ${\widetilde{B}}$ (called {\em{exchange matrix}}) of size $N \times \ex$ where $N$ is a positive integer and $\ex \subset [1,N]$
is a set of mutable (exchangeable) variables. Furthermore, one partitions
\[
[1,N] \backslash \ex = \inv \sqcup \ninv.
\]
The two subsets $\inv$ and $\ninv$ will index the inverted and non-inverted frozen variables. It is important to allow this degree of flexibility because
many key examples in Lie theory require that not all frozen variables are inverted.
\begin{enumerate}
\item The {\em{upper cluster algebra}} $U({\widetilde{B}},\inv)$ is defined as the intersection
\[
\U({\widetilde{B}},\inv)= \bigcap_{(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})} {\mathbb C}[(x'_k)^{\pm 1}, x'_i ; k \in \ex \sqcup \inv, i \in \ninv],
\]
where the intersection ranges over all seeds $(\x':= (x'_1, \ldots, x'_N), {\widetilde{B}})$ in the mutation class of the initial seed
$(\x=(x_1, \ldots, x_N), {\widetilde{B}})$. Gekhtman, Shapiro and Vainshtein proved in \cite{GSV} that,
if the exchange matrix ${\widetilde{B}}$ is compatible with a skew-symmetric integer matrix $\Lambda$ of size $N \times N$ (cf. \eqref{compatibility}),
then $\U({\widetilde{B}},\inv)$ admits a canonical {\em{Poisson structure}}. The cluster
variables in each seed are in log-canonical form, meaning that
\[
\{ x'_k, x'_i \} = \lambda'_{ki} x'_k x'_i, \quad \forall k,i \in [1,N]
\]
for some skew-symmetric integer matrix $(\lambda'_{ki})$ depending on the seed.
\item Let $\varepsilon^{1/2} \in {\mathbb C}$ be a primitive $\ell$-th root of unity for a positive integer $\ell$.
The root of unity upper quantum cluster algebra $\U_\varepsilon(M_\varepsilon,{\widetilde{B}},\inv)$ is a non-commutative algebra,
defined in a similar way to $\U({\widetilde{B}},\inv)$, by intersecting
mixed quantum tori/quantum affine spaces
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv):= \bigcap_{(M'_\varepsilon,{\widetilde{B}}')\sim (M_\varepsilon,{\widetilde{B}})} \TT_\varepsilon(M'_\varepsilon)_\geq,
\]
where $M'_\varepsilon$ are root of unity toric frames, see Sect. \ref{3.1}. Each such algebra has a
canonical central subalgebra $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ obtained by intersecting the mixed Laurent polynomial
rings in the $\ell$-th powers of the generators of $M'_\varepsilon$. Under mild assumptions on $\ell$,
\begin{equation}
\label{u-CUep-isom}
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \cong \U({\widetilde{B}}, \inv),
\end{equation}
see Sect. \ref{3.1}.
\end{enumerate}
There is a canonical action of the torus
\[
T({\widetilde{B}}) := ( {\mathbb C}^\times)^{\dim \Ker {\widetilde{B}}}
\]
on both algebras $\U({\widetilde{B}}, \inv)$ and $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ by algebra automorphisms. In the first case, $T({\widetilde{B}})$ acts by
Poisson automorphisms. In the second case, it preserves the central
subalgebra $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ and the isomorphism \eqref{u-CUep-isom} intertwines the two actions.
\subsection{Results on the Poisson geometric side}
\label{1.2}
The affine Poisson variety corresponding to the GSV Poisson structure
\[
Y({\widetilde{B}}):= \MaxSpec \U({\widetilde{B}},\inv)
\]
is of much interest in Lie theory and integrable systems (we suppress the dependance of $Y({\widetilde{B}})$ on $\inv$ for brevity).
However, little is known about its global geometry. One only knows that the cluster tori inside $Y({\widetilde{B}})$ are regular Poisson \cite{GSV}.
This is a local result because the symplectic leaves inside each cluster torus never entirely belong to the cluster torus. So,
general Hamiltonian flows on $Y({\widetilde{B}})$ leave each cluster torus after some time.
In Lie theory and combinatorics, one knows \cite{GY0} that each Schubert cell $X_w^\circ$ in the full flag variety $G/B_+$
of a complex simple Lie group $G$, equipped with the standard Poisson structure, has a dense torus orbit of symplectic leaves which equals the complement of the
Richardson divisor \cite{KLS} of $X_w^\circ$. The latter equals the union of the closures in $X_w^\circ$ of the
open Richardson varieties $R_{w,s_i}$ for $i$ ranging over the support of the Weyl group element, see Sect. \ref{7.1}.
Remarkably, such a fact holds for the spectrum of every upper cluster algebra $\U({\widetilde{B}}, \inv)$ with the GSV Poisson structure,
without any assumptions on $\U({\widetilde{B}}, \inv)$ except for finite generation which is needed to be even able to talk about symplectic leaves.
This is the topic of our first main result on the the global description of the $T({\widetilde{B}})$-orbit of the symplectic leaves ${\mathcal{S}}$ of $(Y({\widetilde{B}}), \pi)$ of maximal dimension,
\[
T({\widetilde{B}}) \cdot {\mathcal{S}}.
\]
\noindent
{\bf{Theorem A.}}
{\em{Assume that $\U({\widetilde{B}}, \inv)$ is a finitely generated upper cluster algebra for which the exchange matrix ${\widetilde{B}}$ admits a compatible skew-symmetric integer matrix.
Then the affine Poisson variety $(Y({\widetilde{B}}), \pi)$ has a Zariski open $T({\widetilde{B}})$-orbit of symplectic leaves, which equals
\[
Y({\widetilde{B}})^ {\mathrm{reg}} \backslash \bigcup_{i \in \ninv} {\mathcal{V}} (\x_i),
\]
where $Y({\widetilde{B}})^ {\mathrm{reg}} $ is the nonsingular part of $Y({\widetilde{B}})$.
}}
\medskip
The theorem is proved by using sections of the anticanonical bundle of $Y({\widetilde{B}})^ {\mathrm{reg}} $ coming from the GSV Poisson structure $\pi$ and the action of
$T({\widetilde{B}})$, coupled with normality of $Y({\widetilde{B}})$.
At this point it might be tempting to conjecture that a much stronger result than Theorem A holds, namely that there are
finitely many $T({\widetilde{B}})$-orbits of symplectic leaves of $Y({\widetilde{B}})$ (or $Y({\widetilde{B}})^ {\mathrm{reg}} $). However, that is not correct.
For large classes of Belavin--Drinfeld Poisson structures \cite[Sect. 3.2]{CP} on ${\mathrm{GL}}_n({\mathbb C})$,
Gekhtman, Shapiro and Vainshtein proved \cite{GSV1,GSV2} that the coordinate ring ${\mathbb C}[{\mathrm{GL}}_n]$ admits
upper cluster algebra structures with compatible Poisson structures given by the ones in the list.
Their symplectic leaves were classified in \cite{Y}, where it was proved that they are classified
by Weyl group datum and twisted conjugacy classes of reductive groups. Because of the last bit of data, in general,
those Poisson structures have infinitely many torus orbits of symplectic leaves.
\subsection{Results on the quantum side}
\label{1.3}
It was proved in \cite[Theorem B]{HLY} that, if the ${\mathbb C}$-algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is finitely generated,
then it is a finitely generated module over $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ and $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \cong \U({\widetilde{B}}, \inv)$ is a finitely
generated commutative ${\mathbb C}$-algebra. The root of unity upper quantum cluster algebras $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ form
a vast family of algebras that includes as special cases many important classes of quantum algebras at roots of unity
arising in Lie theory and topology. A fundamental open problem for them is to classify their irreducible representations.
By \cite[Theorem III.1.6]{BrG} all of their representations are finite dimensional of dimension less than or equal to the PI degree
of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$. In this generality, the problem of classifying the irreducible representations of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is extremely difficult.
The first step towards its resolution
is to classify the representations of maximal dimension. For this, Brown and Gordon defined \cite[Sect. 1.3]{BG1} the
{\em{fully Azumaya locus}} of a finitely generated prime algebra $R$ with respect to a central subalgebra $Z$ such that
$R$ is a finitely generated $Z$-module. This locus is a Zariski open subset of $\MaxSpec Z$ consisting of those maximal ideals $\mathfrak{m}$ of $Z$
such that all irreducible representations of $R$ annihilated by $\mathfrak{m}$ have maximal dimension (equal to the PI degree of $R$),
see Sect. \ref{2.1} for details.
Our second main result gives an explicit description of the fully Azumaya locus of the root of unity upper
quantum cluster algebras $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ with respect to their central subalgebras $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \cong \U({\widetilde{B}}, \inv)$:
\medskip
\noindent
{\bf{Theorem B.}} {\em{Assume that $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a finitely generated strict root of unity upper quantum cluster algebra such that
the order of $\varepsilon^{1/2}$ is odd and coprime to the diagonal entries of the skew-symmetrizing matrix $D$
for the principal part of the exchange matrix ${\widetilde{B}}$ and
\begin{equation}
\label{A=U}
\A_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv) = \U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv).
\end{equation}
Then the fully Azumaya locus $\AA$ of
$\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ with respect to the central subalgebra
\[
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \cong \U({\widetilde{B}}, \inv)
\]
satisfies
\[
Y({\widetilde{B}})^ {\mathrm{reg}} \backslash \bigcup_{i \in \ninv} {\mathcal{V}} (\x_i) \subseteq \AA \subseteq
Y({\widetilde{B}}) \backslash \bigcup_{i \in \nc} {\mathcal{V}} (\x_i),
\]
where $\nc$ denotes the set of those non-inverted frozen variables $M(e_i)$ that are not in the center on $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$.
}}
\medskip
The assumption that $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a finitely ${\mathbb C}$-algebra is needed to be even able to define the fully Azumaya locus.
The strictness assumption on $\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv)$ means that ${\widetilde{B}}$ admits a skew-symmetric integer matrix which is compatible
with ${\widetilde{B}}$ over ${\mathbb Z}$ and not just over ${\mathbb Z}/\ell$, which is what is needed to
define root of unity quantum cluster algebras in general. This assumption ensures that there is an associated upper quantum cluster algebra
$\U_q(M_q, {\widetilde{B}},\inv)$ in the sense of Berenstein and Zelevinsky \cite{BerZe}. The assumption \eqref{A=U} is needed to ensure that
$\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv)$ is a specialization of $\U_q(M_q, {\widetilde{B}},\inv)$. This in turn is used to
construct a Poisson order structure on the pair $(\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv), \C\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv))$ in the sense of
Brown and Gordon \cite{BG1} to be able to link Poisson geometry to representation theory. A property of the form $\A=\U$
was established in numerous situations on the classical and quantum levels \cite{CGGLSS,GY1,GY1,GY2,GY3,M1,M2,SW}.
Undoubtably, these methods will be extended in the future to show that the assumption \eqref{A=U} is satisfied in broad generality.
In \cite{BG1a} the fully Azumaya loci of all quantum function algebras of complex simple Lie groups $G$ at roots of unity were determined.
In \cite{GJS} a second proof of the Bonahon--Wang unicity conjecture \cite{BW} was given, stating that the Azumaya loci of the skein algebras of oriented surfaces
at roots of unity contain the smooth parts of the spectra of their centers (this was first proved in \cite{FKBL}).
Both \cite{BG1a,GJS} relied on Poisson orders.
The result of \cite{BG1a} is that the fully Azumaya locus inside $G$ is the open double Bruhat cell (a torus orbit of symplectic leaves), while the proofs of \cite{GJS} relied on
an argument that in that situation, the smooth part in question is a single symplectic leaf. Our Theorems A and B prove that such phenomena hold in much wider
generality in the world of cluster algebras, rather than quantum groups or skein algebras.
The zero loci of frozen variables and singular part of $Y({\widetilde{B}})$ were exactly the points that were thrown out
in \cite{BG1a} and \cite{GJS}.
\medskip
\noindent
{\bf{1.4 Organization of the paper and notation.}} The paper is organized as follows. Section 2 provides background on Azumya loci,
Poisson orders, cluster algebras, compatible Poisson structures and quantum cluster algebras. Section 3 contains background on
root of unity quantum cluster algebras and results on torus actions on them and the related Poisson cluster algebras.
Section 4 proves Theorem A. Section 5 construct Poisson orders on root of unity quantum cluster algebras.
Section 6 proves Theorem B. Section 7 discusses the motivation for the main theorem from the stand point of
Richardson varieties, the special case of the theorems on acyclic cluster algebras, and a Kronecker type
example.
The following notation will be used throughout the paper. The standard basis of ${\mathbb Z}^N$ will be denoted by
\[
e_1, \ldots, e_N.
\]
The dot product on ${\mathbb Z}^N$ will be denoted by $\mu \cdot \nu$. The transpose of a matrix $B$ will be denoted by $B^\top$.
For a complex affine algebraic variety $Y$, its singular part will be denoted by $Y^ {\mathrm{sing}}$ and its smooth part by $Y^ {\mathrm{reg}} $. Given a regular function
$f \in {\mathbb C}[Y]$, we will denote by ${\mathcal{V}}(f)$ the zero locus of $f$.
\sectionnew{Preliminaries on Poisson orders and cluster algebras}
\lb{prelim}
In this section we gather background material on Poisson orders and cluster algebras that
will be used in the paper.
\subsection{Poisson orders}
\label{2.1}
We follow Brown and Gordon \cite{BG2}.
Let $R$ be a ${\mathbb C}$-algebra and $Z$ be a central subalgebra of $R$.
\bde{poi order} \cite{BG2}
The pair $(R,Z)$ is called a {\em{Poisson order}} if $R$ is a finitely generated $Z$-module satisfying the following conditions:
\begin{enumerate}
\item[(i)] $Z$ is has the structure of a Poisson algebra with bracket $\{\cdot,\cdot\}$;
\item[(ii)] There exists a ${\mathbb C}$-linear map $\partial: Z\longrightarrow \Der_{\mathbb C}(R)$ such that $\partial_z|_Z=\{z,-\}$ for all $z\in Z$.
\end{enumerate}
\ede
By assumption (ii), the Poisson structure is uniquely determined from the linear map $\partial$.
Because of this we will denote Poisson orders as triples $(R,Z, \partial)$.
{\em{Restriction of Poisson orders:}} For a Poisson order $(R, Z, \partial)$, if $C$ is a Poisson subalgebra of $Z$ with respect to
the underlying Poisson structure, then $(R,C, \partial|_C)$ is also a Poisson order.
{\em{Poisson orders from specialization:}} The following is a well known fact for obtaining Poisson order structures from specialization,
see e.g. \cite[Sect. 2.2]{BG2}:
\ble{Pord-spec}
Assume that $R$ and $S$ are ${\mathbb C}$-algebras and $\eta : S \to R$ is a surjective ${\mathbb C}$-algebra homomorphism with kernel $(h)=hS$
for a regular central element $h \in S$. Choose a ${\mathbb C}$-linear map
\[
\iota : {\mathcal{Z}}(R) \to S
\]
such that $\eta \circ \iota = {\mathrm{id}} _{{\mathcal{Z}}(R)}$. If $R$ is a finitely generated ${\mathcal{Z}}(R)$-module, then
the pair $(R, {\mathcal{Z}}(R))$ admits a Poisson order structure with $\partial_z : {\mathcal{Z}}(R) \to \Der_{\mathbb C}(R)$
given by
\begin{equation}
\label{partial-spec}
\partial_z (r) := \eta \Big( \frac{\iota(z) \widetilde{r} - \widetilde{r} \iota(z)}{h} \Big), \quad \forall z \in {\mathcal{Z}}(R), r \in R,
\end{equation}
where $\widetilde{r}$ is any preimage of $r$ inder $\eta$. Its underlying Poisson structure is given by
\[
\{z_1, z_2 \} := \eta \Big( \frac{\iota(z_1) \iota(z_2) - \iota(z_2) \iota(z_1)}{h} \Big), \quad \forall z_1, z_2 \in {\mathcal{Z}}(R).
\]
This Poisson structure is independent on the choice of ${\mathbb C}$-linear section $\iota : {\mathcal{Z}}(R) \to S$.
\ele
In \eqref{partial-spec}, $\iota(z) \widetilde{r} - \widetilde{r} \iota(z) \in (h)$ because $z \in {\mathcal{Z}}(R)$ and $\eta$ is an
algebra homomorphism. The right hand side of \eqref{partial-spec} is independent on the choice of preimage $\widetilde{r}$
by a similar argument.
Recall that an affine Poisson variety is an affine variety $X$ whose coordinate ring is equipped with a Poisson algebra structure.
Its singular part $X^ {\mathrm{sing}}$ is automatically an affine Poisson variety as well \cite[Corollary 2.4]{P}. The symplectic leaves of
a complex affine Poisson variety $X$ are defined recursively as the symplectic leaves of the smooth complex manifold $X^ {\mathrm{reg}} $
together with the symplectic leaves of the lower dimensional Poisson variety $X^ {\mathrm{sing}}$.
\bth{P-ord-symp-leaves} {\em{(Brown--Gordon)}} \cite{BG2} Assume that $(R,Z)$ is a complex Poisson order and $\mathfrak{m}, \mathfrak{m}' \in \MaxSpec Z$ lie in the
same symplectic leaf. Then we have the isomorphism of finite dimensional complex algebras
\[
R/\mathfrak{m} R \cong R/ \mathfrak{m}' R.
\]
\eth
The theorem proved in \cite{BG2} has a stronger conclusion using Poisson cores, but we will not need that fact in this paper.
For the rest of this subsection we assume that
(*) {\em{$R$ is a finitely generated prime ${\mathbb C}$-algebra, which is a finitely generated $Z$-module over a central subalgebra $Z$.}}
By the Artin--Tate Lemma (see e.g. \cite[Sect. I.13.4]{BrG}), $Z$ is a finitely generated ${\mathbb C}$-algebra.
The primeness assumption on $R$ implies that ${\mathcal{Z}}(R)$, and thus also $Z$, are integral domains. Hence, $\MaxSpec {\mathcal{Z}}(R)$ and $\MaxSpec Z$
are irreducible affine varieties.
Recall that $\mathfrak{m} \in \MaxSpec {\mathcal{Z}}(R)$ is in the {\em{Azumaya locus of $R$}} if $R_\mathfrak{m}$ is an Azumaya algebra over
$Z_\mathfrak{m}$. This is equivalent to saying that $R$ has an irreducible module, annihilated by $\mathfrak{m}$, of maximal dimension
among the irreducible $R$-modules (which equals the PI degree of $R$);
such a representation is automatically unique (see \cite[Theorem III.1.6]{BrG}).
We have the canonical map
\begin{equation}
\label{ZRtoZ}
\MaxSpec {\mathcal{Z}}(R) \to \MaxSpec Z,
\end{equation}
induced by the inclusion $Z \subseteq {\mathcal{Z}}(R)$.
\bde{fullyAzumaya}
A point $\mathfrak{m} \in \MaxSpec Z$ is said to be in the {\em{fully Azumaya locus of $R$ with respect to $Z$}} if all of its preimages are
in the Azumaya locus of $R$, see \cite[Sect. 1.3]{BG1}. In other words, one requires that all irreducible modules of $R/ \mathfrak{m} R$ have
dimensions equal to the PI degree of $R$.
\ede
The map \eqref{ZRtoZ} is closed by \cite[Lemma III.1.5]{BrG}.
This and the fact that the Azumaya locus of $R$ is open, and hence dense in $\MaxSpec {\mathcal{Z}}(R)$, imply
\ble{FAzumaya-dense} In the above setting, the fully Azumya locus of $R$ with respect to $Z$ is an open and hence dense subset of $\MaxSpec Z$.
\ele
This fact is stated in \cite[Proposition III.4.10]{BrG} under a Hopf algebra assumption on $R$ and $Z$, but this assumption is not used in its proof.
Finally, we have the following corollary of \thref{P-ord-symp-leaves}:
\bco{P-ort-FAzumaya}
Assume that $R$ is a finitely generated ${\mathbb C}$-algebra and
$(R,Z)$ is a complex Poisson order. Then the fully Azumaya locus of $R$ with respect to $Z$ is a union of symplectic leaves of $\MaxSpec Z$.
\eco
\begin{proof} If $\mathfrak{m}, \mathfrak{m}' \in \MaxSpec Z$ lie in the same symplectic leaf of $\MaxSpec Z$,
then $R/\mathfrak{m} R \cong R/ \mathfrak{m}' R$ by \thref{P-ord-symp-leaves}. In particular, if all irreducible representations of the
algebra $R/ \mathfrak{m} R$ have dimensions equal to the PI degree of $R$, then the same holds for the algebra
$R/ \mathfrak{m}' R$.
\end{proof}
\subsection{Cluster algebras of geometric type}
\label{2.2}
We follow Berenstein, Fomin and Zelevinsky \cite{FZ1,BFZ05} with the exception that all algebras are defined over ${\mathbb C}$ instead of ${\mathbb Z}$.
Let $N$ be a positive integer, $\ex\subseteq [1,N]$ and ${\mathcal{F}}$ be a purely transcendental field extension of ${\mathbb C}$
of transcendence degree $N$. We say that a pair $(\x, {\widetilde{B}})$ is a {\em{seed}} if
\begin{enumerate}
\item[(i)] $\x = \{x_1, \dots, x_N\}$ is a transcendence basis of ${\mathcal{F}}$ over ${\mathbb C}$;
\item[(ii)] ${\widetilde{B}}\in M_{N\times \ex}({\mathbb Z})$ and its $\ex\times \ex$ submatrix $B$, called the principal part of ${\widetilde{B}}$,
is skew-symmetrizable for some matrix $D=\diag(d_j,~j\in\ex)$ with diagonal entries $d_j \in {\mathbb Z}_+$.
\end{enumerate}
The elements $x_i \in {\mathcal{F}}$ are called {\em{cluster variables}}. A matrix ${\widetilde{B}}$ satisfying the condition in (ii) is called an {\em{exchange matrix}}.
For each $k\in \ex$, the mutation of ${\widetilde{B}}$ in the direction of $k$ is the matrix $\mu_k({\widetilde{B}})$, where
\[\mu_k({\widetilde{B}}) = (b_{ij}'):=
\begin{cases}
-b_{ij} & \text{if } i = k \text{ or } j=k \\
b_{ij} + \frac{|b_{ik}|b_{kj} + b_{ik}|b_{kj}|}{2} & \text{otherwise.}
\end{cases}
\]
For a choice of sign, $s = \pm$, define the matrices $E_s \in M_N({\mathbb Z})$ and $F_s \in M_\ex({\mathbb Z})$ to be
\[
E_s := (e_{ij}) =
\begin{cases}
\delta_{ij} & \text{if } j \neq k \\
-1 & \text{if } i = j = k \\
\max(0,-sb_{ik}) & \text{if } i \neq j = k,
\end{cases} \; \;
F_s := (f_{ij})=
\begin{cases}
\delta_{ij} & \text{ if } i \neq k \\
-1 & \text{ if } i = j = k \\
\max(0,sb_{kj}) & \text{ if } j \neq i = k.
\end{cases}
\]
Then we also have $\mu_k({\widetilde{B}})= E_s {\widetilde{B}} F_s$ for both $s = \pm$.
The mutation of a seed $(\x, {\widetilde{B}})$ in the direction of $k \in \ex$
is defined to be $\mu_k (\x, {\widetilde{B}}) := (\x', \mu_k({\widetilde{B}}))$ where
\begin{equation}
\label{classic-mut}
\x' := \{x_k'\} \cup \x \backslash \{x_k\}
\quad
\mbox{and}
\quad
x_k x_k' := \prod_{b_{ik}>0}x_i^{b_{ik}} + \prod_{b_{ik}<0}x_i^{-b_{ik}}.
\end{equation}
The pair $\mu_k (\x, {\widetilde{B}})$ is also a seed, the principal part of $\mu_k({\widetilde{B}})$ equals $\mu_k(B)$, and $\mu_k(B)$ is skew-symmetrizable
with respect to the same matrix $D$ that skew-symmetrizes $B$.
Mutation is an involution. Two seeds are {\em{mutation equivalent}}, $(\x', {\widetilde{B}}')\sim (\x'', {\widetilde{B}}'')$, if one can be obtained from the other by a finite sequence of mutations.
Any seed which is mutation-equivalent to $(\x, {\widetilde{B}})$ contains $x_i$ for $i \in [1,N]\backslash \ex $ and we call these, {\em{frozen variables}}.
Fix a decomposition of the set of frozen variables into a disjoint union of two sets:
\[
[1,N]\backslash \ex= \inv \sqcup \ninv.
\]
The set $\inv$ will index the set of those frozen variables that will be {\em{inverted}}. The set $\ninv$ will index the {\em{non-inverted frozen variables}}.
The {\em{cluster algebra}} $\A({\widetilde{B}}, \inv)$ is defined to be the ${\mathbb C}$-subalgebra of ${\mathcal{F}}$ generated by
the cluster variables in all the seeds $(\x',{\widetilde{B}}')\sim( \x, {\widetilde{B}})$ together with $\{ x_i^{-1} \mid i \in \inv\}$.
For a seed $(\x',{\widetilde{B}}')\sim (\x,{\widetilde{B}})$ denote the Laurent polynomial ring
\[
{\mathcal{L}}(\x') = {\mathbb C}[(x'_k)^{\pm 1}; 1 \leq k \leq N]
\]
and its mixed polynomial/Laurent polynomial subring
\begin{equation}
\label{Lx'}
{\mathcal{L}}(\x')_\geq = {\mathbb C}[(x'_k)^{\pm 1}, x'_i; k \in \ex \sqcup \inv, i \in \ninv].
\end{equation}
The {\em{upper cluster algebra}} $\U({\widetilde{B}}, \inv)$ is the intersection
\begin{equation}
\label{upper-c-a}
\U({\widetilde{B}},\inv)= \bigcap_{(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})} {\mathcal{L}}(\x')_\geq.
\end{equation}
By the {\em{Laurent phenomenon}} \cite{FZlp}, we have $\A({\widetilde{B}}, \inv) \subseteq ~\U({\widetilde{B}}, \inv)$.
We will need the algebra \eqref{upper-c-a} in the special case when all frozen variables are inverted:
\[
\U({\widetilde{B}}):= \U({\widetilde{B}}, [1,N]\backslash \ex) = \bigcap_{(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})} {\mathcal{L}}(\x').
\]
It is easy to verify that the latter is obtained as a localization:
\[
\U({\widetilde{B}}) = \U({\widetilde{B}}, \inv)[x_i^{-1}; i \in \ninv].
\]
\subsection{Quantum cluster algebras}
\label{2.3}
We follow Berenstein and Zelevinsky \cite{BerZe} with the exception that we allow for an arbitrary subset of frozen quantum
cluster variables not to be inverted and we work over the algebra
\[
{\mathcal A}_q^{1/2}:= {\mathbb C}[q^{\pm 1/2}]
\]
instead of ${\mathbb Z}[q^{\pm 1/2}]$.
By abuse of notation, we will identify a skew-symmetric bilinear form $\Lambda: {\mathbb Z}^N \times {\mathbb Z}^N \to {\mathbb Z}$
with the skew-symmetric integer matrix with entries $\Lambda(e_i, e_j)$. Recall that $e_1, \ldots, e_N$ denotes the standard basis of ${\mathbb Z}^N$.
The \emph{based quantum torus} $\TT_{q}(\Lambda)$ associated with $\Lambda$ is the ${\mathcal A}_q^{1/2}$-algebra with the ${\mathcal A}_q^{1/2}$-basis
$\{ \hspace{1pt} x^f \hspace{1pt} | \hspace{1pt} f \in {\mathbb Z}^N \hspace{1pt} \}$ and multiplication given by $x^f x^g = q^{\Lambda(f,g)/2} x^{f+g},$
where $f,g \in {\mathbb Z}^N$.
A \emph{toric frame} $M_q$ for a division ${\mathbb C}(q^{1/2})$-algebra ${\mathcal{F}}_q$
is a map $M_q : {\mathbb Z}^N \to {\mathcal{F}}_q$ for which there exists an ${\mathcal A}_q^{1/2}$-algebra embedding $\phi: \TT_q(\Lambda) \hookrightarrow {\mathcal{F}}_q$
for some skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$, such that $\phi(x^f)=M_q(f)$ for all $f\in {\mathbb Z}^N$ and ${\mathcal{F}}_q \simeq \Fract\left( \phi( \TT_q(\Lambda) ) \right)$.
Denote by $\Lambda_{M_q}$ the skew-symmetric matrix (bilinear form) of a toric frame $M_q$, defined by
\[
M_q(f) M_q(g) = q^{\Lambda(f,g)/2} M_q(f+g), \quad \forall f, g \in {\mathbb Z}^N.
\]
For a torus frame $M_q$, the image of $\phi$ in ${\mathcal{F}}_q$ will be denoted by $\TT_q(M_q)$; its basis is $\{ \hspace{1pt} M_q(f) \mid f \in {\mathbb Z}^N \hspace{1pt} \}$.
We have the isomorphism of quantum tori $\TT_q(M_q) \simeq \TT_q(\Lambda_{M_q})$.
Let ${\widetilde{B}}\in M_{N\times \ex}({\mathbb Z})$ be an exchange matrix and $\Lambda= (\lambda_{ij}) \in M_N({\mathbb Z})$ be a skew-symmetric matrix.
The pair $(\Lambda, {\widetilde{B}})$ is called {\em{compatible}} if
\[
\sum_{k=1}^N b_{kj}\lambda_{ki} = \delta_{ij} d_j, \quad \forall k \in [1,N], j \in \ex
\]
for a collection of
positive integers $(d_j, j \in \ex)$. In terms of the diagonal matrix $D := \diag(d_j, j \in \ex)$, this condition is written as
\begin{equation}
\label{compatibility}
{\widetilde{B}}^\top \Lambda = [D \; 0],
\end{equation}
where where 0 denotes the zero matrix of size $\ex \times ([1, N] \backslash \ex)$.
If $(\Lambda, {\widetilde{B}})$ is a compatible pair, then ${\widetilde{B}}$ has full rank and its principal part $B$ is skew-symmetrized by $D$.
For each $k\in \ex$, the mutation of a compatible pair $(\Lambda, {\widetilde{B}})$ in the direction of $k$ is
\begin{equation}
\label{mut-com-pairs}
\mu_k(\Lambda, {\widetilde{B}}) := (\Lambda',{\widetilde{B}}'),
\end{equation}
where ${\widetilde{B}}' = E_s {\widetilde{B}} F_s$ as in Sect. \ref{2.2} and $\Lambda':= E_s^\top \Lambda E_s$, $s = \pm$.
The pair $(\Lambda',{\widetilde{B}}')$ is compatible with respect to the same diagonal matrix $D$
and is independent on the choice of sign $s$.
A {\em{quantum seed}} is a pair $(M_q, {\widetilde{B}})$, consisting of a toric frame $M_q$ of ${\mathcal{F}}_q$ and an exchange matrix ${\widetilde{B}}$
such that $(\Lambda_{M_q}, {\widetilde{B}})$ is compatible.
We call $M_q(e_j)$, $j\in [1,N]$, \emph{cluster variables} of the seed, among which the {\em{frozen}} ones are those indexed by $[1,N]\backslash \ex$.
Denote $b^k$ be the $k$-th column of ${\widetilde{B}}$. The mutation of a quantum seed $(M_q, {\widetilde{B}})$ in the direction of $k \in \ex$ is
\[
\mu_k(M_q, {\widetilde{B}}) = (\mu_k(M_q), \mu_k({\widetilde{B}}) ) := (\rho^{M_q}_{b^k, s} M_q E_s , E_s {\widetilde{B}} F_s )
\]
for any choice of sign $s$,
where $\rho_{b^k, s} = \rho^{M_q}_{b^k, s}$ is the unique automorphism of ${\mathcal{F}}_q$ such that
\[
\rho_{b^k, s}(M_q(e_j)) =
\begin{cases}
M_q(e_k) + M_q(e_k + s b^k) & \text{if } j=k \\
M_q(e_j) & \text{if } j \neq k.
\end{cases}
\]
The skew-symmetric matrix associated to the toric frame $\mu_k(M_q)$ is $\Lambda_{\mu_k(M_q)} = \mu_k(\Lambda_{M_q})$.
The mutation process is involutive.
The {\em{quantum cluster algebra}} $\A_q(M_q, {\widetilde{B}}, \inv)$ is the ${\mathcal A}_q^{1/2}$-subalgebra of ${\mathcal{F}}_q$ generated by all cluster variables of quantum seeds $(M_q',{\widetilde{B}}')\sim (M_q, {\widetilde{B}})$
and by the inverted frozen variables $\{ M_q(e_j)^{-1} \mid j \in \inv \}$.
The corresponding \emph{upper quantum cluster algebra} is given by the intersection
\[
\U_q(M_q, {\widetilde{B}}, \inv):= \bigcap_{(M_q',{\widetilde{B}}')\sim (M_q,{\widetilde{B}})} \TT_q(M_q')_\geq
\]
of the \emph{mixed quantum tori}
\[
\TT_q(M'_q)_{\geq}:={\mathcal A}_q^{1/2}\langle \hspace{1pt} M_q'(e_k)^{\pm 1}, M_q'(e_i)
\mid k \in \ex \sqcup \inv,
i \in \ninv \rangle
\subset \TT_q(M'_q).
\]
Here, the term mixed refers to the fact that these algebras are mixtures of quantum tori and quantum affine spaces.
We have the quantum Laurent phenomenon:
\begin{equation}
\label{qLaurent}
\A_q(M_q, {\widetilde{B}}, \inv) \subseteq \U_q(M_q, {\widetilde{B}}, \inv),
\end{equation}
proved in \cite{BerZe} for $\inv = [1,N] \backslash \ex$ and \cite{GY1} in general.
The {\em{exchange graphs}} of the upper cluster algebra $\U({\widetilde{B}}, \inv)$ and the upper quantum cluster algebra $\U_q(M_q, {\widetilde{B}}, \inv)$
are the labelled graphs with vertices corresponding to the seeds that are mutation-equivalent to $(\widetilde{\mathbf{x}}, {\widetilde{B}})$ and $(M_q, {\widetilde{B}})$, respectively,
and edges given by seed mutation, labelled by the corresponding mutation number. (The exchange graph does not depend on the set $\inv$ of inverted frozen variables.)
\bth{isom-e-graphs} {\em{(Berenstein--Zelevinsky)}} \cite[Theorem 6.1]{BerZe}
There exists a unique isomorphism of labelled graphs, between the exchange graphs of $\U({\widetilde{B}}, \inv)$ and $\U_q(M_q, {\widetilde{B}}, \inv)$
that sends the vertex corresponding to the seed $(\widetilde{\mathbf{x}}, {\widetilde{B}})$ to that of the seed $(M_q, {\widetilde{B}})$.
\eth
\subsection{Poisson structures on cluster algebras}
\label{2.4}
We follow Gekhtman, Shapiro and Vainshtein \cite{GSV}. Consider an upper cluster algebra $\U({\widetilde{B}}, \inv)$ for which there exists a skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$
such that $(\Lambda, {\widetilde{B}})$ is a compatible pair. This is equivalent to saying that there exists a quantum cluster algebra with a seed with exchange matrix ${\widetilde{B}}$.
For such a seed $(M_q, {\widetilde{B}})$, the skew-symmetric matrix of the toric frame $M_q$ equals $\Lambda$.
In this setting, for every seed $(\x',{\widetilde{B}}') \sim (\x, {\widetilde{B}})$, there exists a skew symmetric matrix $\Lambda_{(\x',{\widetilde{B}}')}$ satisfying the following two conditions:
\begin{enumerate}
\item[(i)] the pair $\big( \Lambda_{(\x', {\widetilde{B}}')}, {\widetilde{B}}' \big)$ is compatible and
\item[(ii)] for all $k \in \ex$,
\[
\mu_k \big( \Lambda_{(\x', {\widetilde{B}}')}, {\widetilde{B}}' \big) = \big( \Lambda_{(\mu_k (\x') , \mu_k({\widetilde{B}}'))}, \mu_k({\widetilde{B}}')\big)
\]
where the left hand side uses mutation of compatible pairs, see \eqref{mut-com-pairs}.
\end{enumerate}
For a seed $(\x',{\widetilde{B}}') \sim (\x, {\widetilde{B}})$, we choose a sequence of mutations such that
\[
\mu_{k_1} \ldots \mu_{k_j} (\x, {\widetilde{B}}) = (\x', {\widetilde{B}}').
\]
Applying the sequence of mutations to the compatible pair $(\Lambda, {\widetilde{B}})$ gives that
\[
\mu_{k_1} \ldots \mu_{k_j} (\Lambda, {\widetilde{B}}) = (\Lambda', {\widetilde{B}}')
\]
for a skew-symmetric matrix $\Lambda' \in M_N({\mathbb Z})$. The exchange graph isomorphism from \thref{isom-e-graphs} implies that the matrix $\Lambda'$ is independent
on the choice of mutation sequence $\mu_{k_1} \ldots \mu_{k_j}$. We set $\Lambda_{(\x', {\widetilde{B}}')} := \Lambda'$.
It is clear that conditions (i)-(ii) are satisfied.
For each seed $(\x', {\widetilde{B}}')\sim (\x, {\widetilde{B}})$, the mixed polynomial/Laurent polynomial ring ${\mathcal{L}}(\x')$ has a Poisson algebra structure such that
\begin{equation}
\label{Poisson-L}
\{ x'_k, x'_i \} = \Lambda_{(\x', {\widetilde{B}}')}(e_k,e_i) x'_k x'_i, \quad \forall k,i \in [1,N].
\end{equation}
All algebras ${\mathcal{L}}(\x')$ have a common field of fractions ${\mathcal{F}}$, so all Poisson brackets \eqref{Poisson-L} automatically extend to Poisson field structures on ${\mathcal{F}}$.
Those extensions coincide and, hence, the intersection \eqref{upper-c-a}, $\U({\widetilde{B}}, \inv)$, inherits a Poisson algebra structure,
called the Gekhtman--Shapiro--Vainshtein (GSV) Poisson algebra structure of $\U({\widetilde{B}}, \inv)$.
\sectionnew{Root of unity quantum cluster algebras and torus automorphisms}
In this section we gather background material on root of unity quantum cluster algebras from \cite{NTY2,HLY} and prove auxiliary properties
that will be used in the next sections. In particular, we describe certain torus actions on these algebras, which will play a key role in the paper.
\subsection{Root of unity quantum cluster algebras}
\label{3.1}
Let $\ell$ be a positive integer and $\varepsilon^{1/2}\in {\mathbb C}$ be a primitive $\ell$-th root of unity. Denote
\[
{\mathbb Z}/\ell:= {\mathbb Z}/{\ell{\mathbb Z}}.
\]
In this paper we will work over ${\mathbb C}$, while the construction in \cite{NTY2} was carried out over ${\mathbb Z}[\varepsilon^{1/2}]$.
We start with a skew-symmetric bilinear form $\Omega : {\mathbb Z}^N \times {\mathbb Z}^N \to {\mathbb Z}/\ell$
and identify it with the skew-symmetric matrix $( \Omega(e_i, e_j))_{i,j=1}^N \in M_N({\mathbb Z}/\ell)$.
The associated {\em root of unity based quantum torus} is
\begin{equation}
\label{Tep}
\TT_\varepsilon(\Omega) := \Span_{{\mathbb C}} \{ x^f \mid f \in {\mathbb Z}^N \}, \; \;
\mbox{where} \; \; x^f x^g = \varepsilon^{\Omega(f,g)/2} x^{f+g}, \; \; \forall f,g \in {\mathbb Z}^N.
\end{equation}
A {\em root of unity toric frame} $M_\varepsilon$ of a division algebra ${\mathcal{F}}_\varepsilon$ over ${\mathbb C}$ is a map
\[
M_\varepsilon : {\mathbb Z}^N \to {\mathcal{F}}_\varepsilon
\]
for which there exists an ${\mathbb C}$-algebra embedding $\phi: \TT_\varepsilon(\Omega) \hookrightarrow {\mathcal{F}}_\varepsilon$ for a skew-symmetric matrix $\Omega\in M_N({\mathbb Z}/\ell)$
with the properties that $\phi(x^f)=M_\varepsilon(f)$ for all $f\in {\mathbb Z}^N$ and ${\mathcal{F}}_\varepsilon \simeq \Fract \left( \TT_\varepsilon(\Omega) \right)$.
Denote by $\Omega_{M_\varepsilon}$ the skew-symmetric matrix (bilinear form) with values in ${\mathbb Z}/\ell$ of the root of unity toric frame $M_\varepsilon$,
given by
\[
M_\varepsilon(f) M_\varepsilon(g) = \varepsilon^{\Omega(f,g)/2} M_\varepsilon(f+g), \quad \forall f, g \in {\mathbb Z}^N.
\]
The mixed quantum tori for the root of unity setting are given by
\begin{equation}
\label{Tep-geq}
\TT_\varepsilon(M_\varepsilon)_{\geq} :={\mathbb C} \langle M_\varepsilon(e_k)^{\pm 1}, \ M_\varepsilon(e_i) \mid k \in \ex \sqcup \inv, i \in \ninv \rangle \subset \TT_\varepsilon( M_\varepsilon).
\end{equation}
A pair $(\Omega, {\widetilde{B}})$, consisting of a skew-symmetric matrix $\Omega \in M_N({\mathbb Z}/\ell)$ and an exchange matrix ${\widetilde{B}} \in M_{N \times \ex}({\mathbb Z})$,
is said to be {\em{$\ell$-compatible}} if there exists a diagonal matrix $D := \diag(d_j, j \in \ex)$ with $d_j \in {\mathbb Z}_+$
such that the principal part $B$ of ${\widetilde{B}}$ is skew-symmetrized by $D$ and
\[
\overline{{\widetilde{B}}}^\top \hspace{-0.15cm} \Omega =
\begin{bmatrix}
\, \overline{D} \; 0
\end{bmatrix} \hspace{-0.1cm}.
\]
Here and below, for an integer matrix $C$, $\overline{C}$ denotes its reduction modulo $\ell$.
The \emph{mutation} in direction $k \in \ex$ of an $\ell$-compatible pair is defined to be $\mu_k ( \Omega, {\widetilde{B}}) := ( \overline{E}_s^\top \Omega \overline{E}_s, E_s {\widetilde{B}} F_s)$ for $s= \pm$,
and as in the quantum case; it is independent of the choice of sign $s$. Further, the pair
$\mu_k(\Omega, {\widetilde{B}})$ is also $\ell$-compatible with respect to the same diagonal matrix $D$.
A pair $(M_\varepsilon, {\widetilde{B}})$ is called a {\em root of unity quantum seed} if $(\Omega_{M_\varepsilon},{\widetilde{B}})$ is an $\ell$-compatible pair. The mutation in direction $k\in\ex$ of a root of unity quantum seed is similar to that in the quantum seed case,
$\mu_k(M_\varepsilon, {\widetilde{B}}) := (\rho^{ M_\varepsilon}_{b^k, s} M_\varepsilon E_s, E_s {\widetilde{B}} F_s)$, where $\rho_{b^k,s}=\rho^{M_\varepsilon}_{b^k,s}$ is the unique automorphism of ${\mathcal{F}}_\varepsilon$ given by
\[ \rho_{b^k, s}^{M_\varepsilon}(M_\varepsilon(e_j)) =
\begin{cases}
M_\varepsilon(e_k) + M_\varepsilon(e_k + sb^k) & \text{if } j=k \\
M_\varepsilon(e_j) & \text{if } j \neq k.
\end{cases}\]
The skew-symmetric matrix associated to $\mu_k(M_\varepsilon)$ is $\Omega_{\mu_k(M_\varepsilon)} = \mu_k(\Lambda_{M_\varepsilon})$.
Moreover, mutation of root of unity quantum seeds does not depend on the choice of sign and is an involution.
The {\em root of unity quantum cluster algebra} $\A_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is defined to be the ${\mathbb C}$-subalgebra of ${\mathcal{F}}_\varepsilon$
generated by all cluster variables of the seeds $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$ and by
$M_\varepsilon(e_i)^{-1}$ for $i \in \inv$.
The corresponding \emph{root of unity upper quantum cluster algebra}
is defined as the intersection
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) := \bigcap_{ (M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})}\TT_\varepsilon(M_\varepsilon')_{\geq}.
\]
We have $\A_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)\subseteq \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ by \cite[Theorem 3.10]{NTY2}.
For every root of unity toric frame $M'_\varepsilon$ of ${\mathcal{F}}_\varepsilon$ and $1 \leq i \leq N$,
\[
M'_\varepsilon(e_i)^\ell \in {\mathcal{Z}}({\mathcal{F}}_\varepsilon).
\]
Consider the central subalgebra
\begin{equation}
\label{Lep}
{\mathcal{L}}_\varepsilon(\Omega') = {\mathbb C} [ (x^{e_k})^{\pm \ell}; 1 \leq k \leq N] \subset {\mathcal{Z}}(\TT_\varepsilon(\Omega'))
\end{equation}
and the mixed polynomial/Laurent polynomial central subalgebras
\begin{align}
\label{Lep-geq0}
{\mathcal{L}}_\varepsilon( \Omega')_{\geq} &:= {\mathbb C} [ (x^{e_k})^{\pm \ell}, \ (x^{e_i})^\ell ;
k \in \ex \sqcup \inv, i \in \ninv ] \subset {\mathcal{Z}}( \TT_\varepsilon(\Omega')_\geq), \\
\label{Lep-geq}
{\mathcal{L}}_\varepsilon( M'_\varepsilon)_{\geq} &:= {\mathbb C} [ M'_\varepsilon(e_k)^{\pm \ell}, \ M'_\varepsilon(e_i) ;
k \in \ex \sqcup \inv, i \in \ninv ] \subset {\mathcal{Z}}( \TT_\varepsilon(M'_\varepsilon)_\geq).
\end{align}
Denote the central subalgebra
\begin{equation}
\label{CU}
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) := \bigcap_{ (M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})} {\mathcal{L}}_\varepsilon(M_\varepsilon')_{\geq}
\subset {\mathcal{Z}} (\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)).
\end{equation}
The {\em{exchange graph}} of the root of unity upper quantum cluster algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is
the labelled graph with vertices corresponding to the root of unity quantum seeds that are mutation-equivalent to $(M_\varepsilon, {\widetilde{B}})$
and edges given by seed mutations, labelled by the corresponding mutation number. (The exchange graph is independent of the set $\inv$.)
\bth{cent-isom} Assume that $\ell$ is an odd positive integer that is coprime to the diagonal entries of the skew-symmetrizing matrix $D$
for the principal part of the exchange matrix ${\widetilde{B}}$. Then the following hold:
\begin{enumerate}
\item[(i)] \cite[Theorem 4.8]{NTY2}
The exists a unique isomorphism of labelled graphs, between the exchange graphs of $\U({\widetilde{B}}, \inv)$ and $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
that sends the vertex corresponding to the seed $(\widetilde{\mathbf{x}}, {\widetilde{B}})$ to that of the seed $(M_\varepsilon, {\widetilde{B}})$.
\item[(ii)] \cite[Proposition 4.4]{NTY2} For all seeds $(M_\varepsilon', {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$ and $k \in \ex$,
\[
M'_\varepsilon(e_k)^\ell \left( \mu_k M'_\varepsilon (e_k) \right)^{\ell} = \prod_{b'_{ik}>0} (M'_\varepsilon(e_i)^\ell)^{b'_{ik}} + \prod_{b'_{ik}<0} (M'_\varepsilon(e_i)^\ell)^{ - b'_{ik}},
\]
which is precisely the mutation formula for the seeds of $\U({\widetilde{B}}, \inv)$.
\item[(iii)] \cite[Proposition 3.9]{HLY} We have an isomorphism of ${\mathbb C}$-algebras
\begin{equation}
\label{U-CU-isom}
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \simeq \U({\widetilde{B}}, \inv),
\end{equation}
which is uniquely determined by sending $M'_\varepsilon(e_k)^\ell \mapsto x'_k$, where the seed $(M'_\varepsilon, {\widetilde{B}}')$ corresponds
to the seed $(\x', {\widetilde{B}}')$ under the graph isomorphism in part (i).
\item[(iv)] \cite[Theorem B]{HLY} $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a Cayley--Hamilton algebra of degree $\ell^N$ over its central subalgebra $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
in the sense of Procesi \cite[Definition 4.2]{DP}.
$\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is finitely generated ${\mathbb C}$-algebra if and only if $\U(M_\varepsilon, {\widetilde{B}}, \inv)$ is a finitely generated
$\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$-module and $\U({\widetilde{B}}, \inv)$ is a finitely generated ${\mathbb C}$-algebra.
\end{enumerate}
\eth
We will need the root of unity upper quantum cluster algebra in the special case when all frozen variables are inverted:
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}):= \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, [1,N]\backslash \ex) = \bigcap_{(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})} \TT_\varepsilon(M_\varepsilon').
\]
It is easy to verify that it is a localization of the ones for other choices of $\inv$:
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}) = \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)[M(e_i)^{-1}; i \in \ninv].
\]
The above treatment can be viewed as defining quantum cluster $\mathcal{A}$-varieties at roots of unity.
Quantum cluster $\mathcal{X}$-varieties at roots of unity were defined and studied by Fock and Goncharov in \cite{FG}, who obtain analogous algebraic results
to \thref{cent-isom}(ii)-(iii) under the stronger assumption that the order of the root of unity is coprime to the entries of the exchange matrices of all seeds of the algebra.
However, they only consider varieties up to birational isomorphism and therefore do not consider such phenomena as singularities.
\subsection{Strict root of unity quantum cluster algebras}
\label{3.2}
\bde{strict-seed} \cite[Sect. 5]{NTY2} We say that a root of unity quantum seed $(M_\varepsilon, {\widetilde{B}})$ is {\em{strict}} if there exists a skew-symmetric integer matrix
$\Lambda \in M_N ({\mathbb Z})$ such that
\begin{enumerate}
\item[(i)] $\Omega_{M_\varepsilon} = \overline{\Lambda}$ and
\item[(ii)] $(\Lambda, {\widetilde{B}})$ is a compatible pair,
see Sect. \ref{2.3}.
\end{enumerate}
The corresponding root of unity upper quantum cluster algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ will be also called strict.
\ede
\bpr{struct-mut} Assume that $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a strict root of unity upper quantum cluster algebra such that $\ell$
is an odd positive integer that is coprime to the diagonal entries of the skew-symmetrizing matrix $D$ for the principal part of the exchange matrix ${\widetilde{B}}$.
Then for every seed $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$ there exists a unique skew-symmetric integer matrix
$\Lambda_{(M'_\varepsilon, {\widetilde{B}}')} \in M_N ({\mathbb Z})$, such that $\Lambda_{(M'_\varepsilon, {\widetilde{B}}')} = \Lambda'$, the matrix from \deref{strict-seed}, and
\begin{enumerate}
\item[(i)] the pair $\big( \Lambda_{(M'_\varepsilon, {\widetilde{B}}')}, {\widetilde{B}}' \big)$ is compatible and $\overline{\Lambda_{(M'_\varepsilon, {\widetilde{B}}')}} = \Omega_{M'_\varepsilon}$;
\item[(ii)] for all $k \in \ex$,
\[
\mu_k \big( \Lambda_{(M'_\varepsilon, {\widetilde{B}}')}, {\widetilde{B}}' \big) = \big( \Lambda_{(\mu_k (M'_\varepsilon, {\widetilde{B}}') )}, \mu_k({\widetilde{B}}') \big),
\]
where the left hand side uses mutation of compatible pairs, defined in \eqref{mut-com-pairs}.
\end{enumerate}
\epr
For the proof of the proposition we will need the surjective ${\mathbb C}$-algebra homomorphism
\begin{equation}
\label{kappa}
\kappa_\varepsilon : \TT_q(M_q)_\geq \to \TT_q(M_\varepsilon)_\geq,
\end{equation}
given by
\[
q^{1/2} \mapsto \varepsilon^{1/2}, M_q(e_k)^{\pm 1} \mapsto M_\varepsilon(e_k)^{\pm 1},
M_q(e_i) \mapsto M_\varepsilon(e_i), \forall k \in \ex \sqcup \inv, i \in \ninv.
\]
Its kernel is
\begin{equation}
\label{ker-kappa}
\ker \kappa_\varepsilon = (q^{1/2} - \varepsilon^{1/2}) \TT_q(M_q)_\geq,
\end{equation}
see \cite[Lemma 5.5]{NTY2}.
\medskip
\noindent
{\em{Proof of \prref{struct-mut}.}}
For a given seed $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$ consider a sequence of mutations $\mu_{k_1} \ldots \mu_{k_j}$ such that
\begin{equation}
\label{mut-sequence}
\mu_{k_1} \ldots \mu_{k_j} (M_\varepsilon, {\widetilde{B}}) = (M'_\varepsilon, {\widetilde{B}}').
\end{equation}
Applying this sequence of mutations to the compatible pair $(\Lambda, {\widetilde{B}})$ gives that
\begin{equation}
\label{comp-pair}
\mu_{k_1} \ldots \mu_{k_j} (\Lambda, {\widetilde{B}}) = (\Lambda', {\widetilde{B}}')
\end{equation}
for a skew-symmetric matrix $\Lambda' \in M_N({\mathbb Z})$. The two exchange graph isomorphisms from Theorems
\ref{tisom-e-graphs} and \ref{tcent-isom}(i) imply that the matrix $\Lambda'$ is independent on the choice of mutation sequence $\mu_{k_1} \ldots \mu_{k_j}$
satisfying \eqref{mut-sequence}. Define
\[
\Lambda_{(M'_\varepsilon, {\widetilde{B}}')} := \Lambda'.
\]
Since the mutations of a compatible pair are compatible pairs (see Sect. \ref{2.3}), $\big( \Lambda_{(M'_\varepsilon, {\widetilde{B}}')}, {\widetilde{B}}' \big)$
is a compatible pair.
By \cite[Theorem 5.7]{NTY2}, $\kappa_\varepsilon$ restricts to a surjective ${\mathbb C}$-algebra homomorphism $\A_q(M_q, {\widetilde{B}}, \inv) \to \A_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
that sends cluster variables to cluster variables and commutes with mutation. The toric frame
$M_q$ of ${\mathcal{F}}_q$ is such that $\Lambda_{M_q} = \Lambda$. The existence of such a homomorphism implies that
\[
M'_\varepsilon(e_i) M'_\varepsilon(e_k) = \varepsilon^{\Lambda'(e_i, e_k)/2} M'_\varepsilon(e_i+e_k), \quad \forall i,k \in [1,N].
\]
Hence,
\[
\overline{\Lambda_{(M'_\varepsilon, {\widetilde{B}}')}} = \Omega_{M'_\varepsilon}.
\]
Finally, property (ii) in the statement of the proposition follows at once from the definition \eqref{comp-pair} of $\Lambda_{(M'_\varepsilon, {\widetilde{B}}')}$.
\qed
\subsection{Torus action on $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$}
\label{3.3}
Denote by $\Ker({\widetilde{B}}^\top)$ the null space of ${\widetilde{B}}^\top$ in ${\mathbb Z}^N$.
Let $\nu=(\nu_1, \dots, \nu_N) \in \Ker({\widetilde{B}}^\top)$. For each $k\in \ex$,
we define the vector $\mu_k(\nu)=(\nu_1, \dots,\nu_{k-1}, \nu_k', \nu_{k+1}, \dots, \nu_N)$, where
\[\nu_k' := \nu\cdot[b^k]_+ - \nu_k.\]
Recall that $b^k$ denotes the $k$-th column of ${\widetilde{B}}$ and $f \cdot g$ the dot product on ${\mathbb Z}^N$.
By \cite[Lemma 2.3]{GSV},
\begin{equation}
\label{ker-nu}
\nu\in \Ker({\widetilde{B}}^\top) \quad \Rightarrow \quad
\mu_k(\nu)\in \Ker(\mu_k({\widetilde{B}})^\top), \; \; \forall k \in \ex.
\end{equation}
For $\nu \in \Ker({\widetilde{B}}^\top)$, we have the ${\mathbb C}^\times$-action on the root of unity mixed quantum torus $\TT_\varepsilon(M_\varepsilon, {\widetilde{B}})_\geq$, given by
\begin{equation}
\label{phi-act}
\varphi_\nu(t) \cdot M_\varepsilon(f):= t^{\nu \cdot f} M_\varepsilon(f), \; \; \forall f \in {\mathbb Z}^N.
\end{equation}
It induces ${\mathbb C}^\times$-action on the skew field of fractions ${\mathcal{F}}_\varepsilon$ of $\TT_\varepsilon(M_\varepsilon, {\widetilde{B}})$.
Analogously to the proof of \cite[Lemma 2.3]{GSV} one verifies that
\begin{equation}
\label{phi-nu-mut}
\varphi_\nu(t)\cdot \mu_k(M_\varepsilon)(f) := t^{\mu_k(\nu) \cdot f} \mu_k(M_\varepsilon)(f), \; \; \forall f \in {\mathbb Z}^N.
\end{equation}
By recursively applying \eqref{ker-nu}, we obtain that this action preserves the root of unity mixed quantum tori $\TT_\varepsilon(M'_\varepsilon, {\widetilde{B}}')_\geq$
associated to all seeds $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$, and thus, the action preserves $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$.
This proves the first part of the following lemma:
\ble{Caction}
\hfill
\begin{enumerate}
\item[(i)]
For every $\nu \in \Ker({\widetilde{B}}^\top)$, there is a ${\mathbb C}^\times$-action $\varphi_\nu$ on the root of unity upper quantum
cluster algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$, given by \eqref{phi-act}. For every seed $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$ and $k \in [1,N]$,
\begin{equation}
\label{phi-nu-M'}
\varphi_\nu(t) \cdot M'_\varepsilon(e_k) = t^a M'_\varepsilon(e_k)
\end{equation}
for some $a \in {\mathbb Z}$ depending on the seed and $k$.
\item[(ii)] The action $\varphi_\nu$ preserves $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$. Under the isomorphism \eqref{U-CU-isom},
it corresponds to the ${\mathbb C}^\times$-action $\psi_\nu$ on $\U({\widetilde{B}}, \inv)$ uniquely determined from
\[
\psi_\nu(t) \cdot x_k := t^{\ell (\nu \cdot e_k) } x_k, \; \; \forall k \in [1,N].
\]
For every seed $(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})$ and $k \in [1,N]$,
\[
\psi_\nu(t) \cdot x'_k := t^{\ell a } x'_k
\]
for some $a \in {\mathbb Z}$.
\item[(iii)] For every $\nu, \eta \in \Ker({\widetilde{B}}^\top)$, the ${\mathbb C}$-actions $\varphi_\nu$ and $\varphi_\eta$ commute.
\end{enumerate}
\ele
The second and third parts of the lemma follow from \eqref{phi-nu-M'} and the concrete form of the isomorphism \eqref{U-CU-isom}
from \thref{cent-isom}(iii). Up to the $\ell$-th power the ${\mathbb C}$-action $\psi_\nu$ on $\U({\widetilde{B}}, \inv)$ is the one constructed in
\cite[Lemma 2.3]{GSV}.
\bde{nB} For an exchange matrix ${\widetilde{B}}$, define the nullity
\[
n({\widetilde{B}}) := \dim \Ker({\widetilde{B}}^\top)
\]
and the complex torus
\[
T({\widetilde{B}}) := ({\mathbb C}^\times)^{n({\widetilde{B}})}.
\]
\ede
$\Ker({\widetilde{B}}^\top)$ is a free abelian group of rank $n({\widetilde{B}})$. Fix a basis
\begin{equation}
\label{basis-KerB}
\{\nu^1, \dots, \nu^{n({\widetilde{B}})} \}
\end{equation}
of it. Parts (i) and (iii) \leref{Caction} imply that we have a
$T({\widetilde{B}})$-action on $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$, given by
\[
\varphi(t_1, \dots, t_{n({\widetilde{B}})}) \cdot y := \varphi_{\nu^1}(t_1)\cdots \varphi_{\nu^{n({\widetilde{B}})}}(t_{n({\widetilde{B}})}) \cdot y,
\]
for $t_i \in {\mathbb C}$ and $y \in \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$.
For every seed $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$ and $k \in [1,N]$, there exists a character $\theta : T({\widetilde{B}}) \to {\mathbb C}^\times$ such that
\begin{equation}
\label{phi-M'}
\varphi(t) \cdot M'_\varepsilon(e_k) = \theta(t) M'_\varepsilon(e_k), \; \; \forall t \in T({\widetilde{B}}).
\end{equation}
For different choices of a basis $\{\nu^1, \dots, \nu^m\}$ of $\Ker({\widetilde{B}}^\top)$, the actions $\varphi$ differ from each other by an automorphism of $T({\widetilde{B}})$.
Part (ii) of \leref{Caction} implies that the action $\varphi$ preserves $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$. Under the isomorphism \eqref{U-CU-isom},
it corresponds to the $T({\widetilde{B}})$-action $\psi$ on $\U({\widetilde{B}}, \inv)$ given by
\[
\psi(t_1, \dots, t_{n({\widetilde{B}})}) \cdot y := \psi_{\nu^1}(t_1)\cdots \psi_{\nu^{n({\widetilde{B}})}}(t_{n({\widetilde{B}})}) \cdot y,
\]
for $t_i \in {\mathbb C}$ and $y \in \U({\widetilde{B}}, \inv)$.
For every seed $(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})$ and $k \in [1,N]$, there exists a character $\theta : T({\widetilde{B}}) \to {\mathbb C}^\times$ such that
\begin{equation}
\label{rescale-x}
\psi(t) \cdot x'_k = \theta(t^\ell) x'_k, \; \; \forall t \in T({\widetilde{B}}).
\end{equation}
By the results in \cite{GSV}, the GSV Poisson algebra structure on $\U({\widetilde{B}}, \inv)$ is invariant under the
action $\psi$ of $T({\widetilde{B}})$. The action $\psi$ is defined for every upper cluster algebra $\U({\widetilde{B}}, \inv)$.
Finally, we note that the $T({\widetilde{B}})$-action $\psi$ on $\U({\widetilde{B}}, \inv)$ is defined for any upper cluster algebra, without the need of
existence of root of unity quantization $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
\sectionnew{Poisson side: the open torus orbit of symplectic leaves}
\lb{Torbits}
In this section we prove our first main result proving that the GSV Poisson structure on every finitely generated
upper cluster algebra always has a Zariski open torus orbit of symplectic leaves and explicitly describe this set.
\subsection{Statement of main theorem}
\label{5.1}
Throughout this section, we will assume that $\U({\widetilde{B}}, \inv)$ is a finitely generated, upper cluster algebra for which there exists a skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$
such that $(\Lambda, {\widetilde{B}})$ is a compatible pair. The GSV Poisson algebra structure $\{.,.\}$ on $\U({\widetilde{B}}, \inv)$ (see Sect. \ref{2.4})
gives rise to a Poisson structure $\pi$ on the affine variety
\[
Y({\widetilde{B}}):= \MaxSpec \U({\widetilde{B}}, \inv).
\]
This variety is in general singular. It is normal because each mixed polynomial/Laurent polynomial ring ${\mathcal{L}}(\x')_\geq$ is integrally closed,
and thus the upper cluster algebra $\U({\widetilde{B}}, \inv)$, given by the intersection \eqref{upper-c-a}, is integrally closed too.
Denote the product of non-inverted frozen variables
\begin{equation}
\label{x-prod}
x := \prod_{i \in \ninv} x_i \in {\mathbb C}[Y({\widetilde{B}})].
\end{equation}
Denote
\[
Y({\widetilde{B}})^\circ = Y({\widetilde{B}}) \backslash {\mathcal{V}}(x) \cong \MaxSpec \U({\widetilde{B}}).
\]
For a seed $(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})$, denote
\begin{align*}
S_{\x'} &:=\MaxSpec {\mathcal{L}}(\x')_{\geq} \cong ({\mathbb C}^\times)^{ | \ex | + | \inv |} \times {\mathbb C}^{| \ninv |}, \\
S_{\x'}^\circ &:=\MaxSpec {\mathcal{L}}(\x') \cong ({\mathbb C}^\times)^N.
\end{align*}
Clearly,
\[
S_{\x'}^\circ \cong S_{\x'} \backslash {\mathcal{V}}(x).
\]
The algebras ${\mathcal{L}}(\x')_{\geq}$ and ${\mathcal{L}}(\x')$ are localizations of $\U({\widetilde{B}}, \inv)$ by the multiplicative subsets generated by $x_i$ for $i \in \ex$
and $i \in \ex \sqcup \ninv$, respectively. So, $S_{\x'}$ and $S_{\x'}^\circ$ are Zariski open, and thus, dense subsets of $Y({\widetilde{B}})$.
Recall from Sect. \ref{3.3} that the Poisson structure $\pi$ is invariant under the action $\psi$ of $T({\widetilde{B}})$. Thus it makes sense to consider the
$T({\widetilde{B}})$-orbits of symplectic leaves of $(\U({\widetilde{B}}, \inv), \pi)$, which are regular Poisson submanifolds.
\bth{open-T-orb-sl} Assume that $\U({\widetilde{B}}, \inv)$ is a finitely generated upper cluster algebra for which there exists a skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$
such that $(\Lambda, {\widetilde{B}})$ is a compatible pair. Then the affine Poisson variety $(Y({\widetilde{B}}), \pi)$ has a Zariski open $T({\widetilde{B}})$-orbit of symplectic leaves, which equals
\[
Y({\widetilde{B}})^ {\mathrm{reg}} \backslash {\mathcal{V}}(x) = (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} .
\]
\eth
\subsection{One-sided containment}
\label{5.2}
\bde{P-prime} \cite[Sect. 4]{GY3}, \cite[Sect. 2.2]{NTY1} Let $R$ be a Poisson algebra which is an integral domain considered as a commutative algebra.
An element $p \in R$ is called Poisson prime, if the principal ideal $(p)$ is a Poisson ideal and a prime ideal of $R$, considered
as a commutative algebra.
Equivalently, $p \in R$ is a prime element and
\[
p \mid \{ p, r \}, \quad \forall r \in R.
\]
\ede
\bpr{p-primes} Assume that $\U({\widetilde{B}}, \inv)$ is an upper cluster algebra, such that ${\widetilde{B}}$ admits a compatible skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$.
Then the following hold:
\begin{enumerate}
\item[(i)] Every prime element of ${\mathbb C}[x_i; i \in \ninv]$ is a prime element of $U({\widetilde{B}}, \inv)$.
\item[(ii)] Every non-inverted frozen variable $x_i$, {\em{(}}i.e. $i \in \ninv${\em{)}}
is a Poisson prime of $\U({\widetilde{B}}, \inv)$ with respect to the Poisson structure associated to $\Lambda$.
\end{enumerate}
\epr
\begin{proof} (i) Let $p \in {\mathbb C}[x_i; i \in \ninv]$ be a prime element. Analogously to the proof of \cite[Proposition 3.5]{GLS-spec}, one shows that
\begin{equation}
\label{inters}
{\mathcal{L}}(\x')_{\geq} \cap ( p {\mathcal{L}}(\mu_k \x')_{\geq}) = ( p {\mathcal{L}}(\x')_{\geq}) \cap {\mathcal{L}}(\mu_k \x')_{\geq}
\end{equation}
for all seed $(\x', {\widetilde{B}}') \sim (\x,{\widetilde{B}})$ and $k \in \ex$. Assume that $p | ab$ for some $a, b \in U({\widetilde{B}}, \inv)$. Then $a, b \in {\mathcal{L}}(\x)_\geq$,
and since $p$ is a prime element of ${\mathcal{L}}(\x)_\geq$, either $p \mid a$ or $p \mid b$. Say $p \mid a$.
Then $a \in p {\mathcal{L}}(\x)_{\geq}$, and by \eqref{inters},
\[
a \in p {\mathcal{L}}(\mu_k \x)_{\geq}, \quad \forall k \in \ex.
\]
Iterating this argument gives that $a \in p {\mathcal{L}}(\x')_{\geq}$ for all $(\x', {\widetilde{B}}') \sim (\x,{\widetilde{B}})$. Hence,
\[
a \in \bigcap_{(\x', {\widetilde{B}}') \sim (\x,{\widetilde{B}})} p {\mathcal{L}}(\x')_{\geq} = p \, U({\widetilde{B}}, \inv).
\]
This implies the statement of (i) because $p$ is not a unit of $U({\widetilde{B}}, \inv)$.
(ii) Fix $b \in U({\widetilde{B}}, \inv)$. It follows from \eqref{Poisson-L} that
\[
\{x_i, b \} /x_i \in {\mathcal{L}}(\x')
\]
for all $(\x', {\widetilde{B}}') \sim (\x,{\widetilde{B}})$. Hence,
\[
\{x_i, b \} /x_i \in U({\widetilde{B}}, \inv),
\]
which completes the proof of part (ii).
\end{proof}
\bpr{containment} In the setting of \thref{open-T-orb-sl}, $Y({\widetilde{B}})^ {\mathrm{reg}} \backslash {\mathcal{V}}(x) = (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} $ is a union of $T({\widetilde{B}})$-orbits of symplectic leaves.
\epr
\begin{proof} Let $i \in \ninv$. By \eqref{rescale-x}, $x_i$ is rescaled by the $T({\widetilde{B}})$-action. So ${\mathcal{V}}(x_i)$ is $T({\widetilde{B}})$-stable.
It follows from \cite[Proposition 2.3]{GY3} that ${\mathcal{V}}(x_i)$ is a union of symplectic
leaves. Therefore, ${\mathcal{V}}(x_i)$ is a union of $T({\widetilde{B}})$-orbits of symplectic leaves.
Since $T({\widetilde{B}})$ acts on $U({\widetilde{B}}, \inv)$ by algebra automorphisms, $Y({\widetilde{B}})^ {\mathrm{sing}}$ is $T({\widetilde{B}})$-stable. By \cite[Corollary 2.4]{P}, it is also
a union of symplectic leaves. Thus, $Y({\widetilde{B}})^ {\mathrm{sing}}$ is a union of $T({\widetilde{B}})$-orbits of symplectic leaves.
Hence, $Y({\widetilde{B}})^ {\mathrm{sing}} \cup {\mathcal{V}}(x)$ is a union of $T({\widetilde{B}})$-orbits of symplectic leaves, and the same applies to its complement,
\[
Y({\widetilde{B}})^ {\mathrm{reg}} \backslash {\mathcal{V}}(x) = Y({\widetilde{B}}) \backslash (Y({\widetilde{B}})^ {\mathrm{sing}} \cup {\mathcal{V}}(x)).
\qedhere
\]
\end{proof}
\subsection{A section of the anticanonical bundle of $Y({\widetilde{B}})^ {\mathrm{reg}} $}
\label{5.3}
For $\nu \in \Ker( B^\top)$, denote by
\[
w_\nu \; \; \mbox{the vector field on $Y({\widetilde{B}})$},
\]
which represents the infinitesimal action of the ${\mathbb C}^\times$-action $\psi_\nu$ on $Y({\widetilde{B}})$,
cf. \leref{Caction}(ii).
The rank of the skew-symmetric matrix $\Lambda = (\lambda_{ik}) \in M_N({\mathbb Z})$ equals
\[
\rk (\Lambda) = 2 r
\]
for some $r \in {\mathbb Z}_+$. It follows from \eqref{compatibility} that
\begin{align}
\label{nLa-ineq}
2r + n({\widetilde{B}}) &= \dim \Im (\Lambda) + \dim \Ker ({\widetilde{B}}^\top) \geq \dim \Im ({\widetilde{B}}^\top \Lambda) + \dim \Ker ({\widetilde{B}}^\top)
\\
&= \dim \Im ({\widetilde{B}}^\top) + \dim \Ker ({\widetilde{B}}^\top) =N.
\hfill\nonumber
\end{align}
The Poisson structure $\pi$ on $Y({\widetilde{B}})$ has rank $r$ (meaning $\wedge^r \pi \neq 0$ and $\wedge^{r+1} \pi =0$).
This follows from the fact that
its rank equals the rank of its restriction to the Zariski open subset $S_\x$ on $Y({\widetilde{B}})$, on which the Poisson structure is
explicitly given by
\[
\pi|_{S_\x} = \sum_{1 \leq i, k \leq N} \lambda_{ik} \frac{\partial}{\partial x_i} \wedge \frac{\partial}{\partial x_k} \cdot
\]
Recall from \eqref{basis-KerB} that the action $\psi$ of the torus $T({\widetilde{B}})$ on $Y({\widetilde{B}})$ is defined from a basis $\{\nu^1, \ldots, \nu^{n({\widetilde{B}})}\}$
of $\Ker( B^\top)$.
Denote by $\Delta$ the set of $(N-2r)$-element subsets of $[1,n({\widetilde{B}})]$, recall \eqref{nLa-ineq}.
For $\theta := \{i_1, \ldots, i_{N- 2r}\} \in \Delta$,
denote the section of the anticanonical bundle of $Y({\widetilde{B}})$,
\[
\chi_\theta := w_{v^{i_1}} \wedge \ldots \wedge w_{\nu^{i_{N-2r}}} \wedge (\wedge^r \pi) \in \Gamma(Y({\widetilde{B}}), K^*_{Y({\widetilde{B}})}).
\]
\bth{sections-antican} Assume that $\U({\widetilde{B}}, \inv)$ is a finitely generated upper cluster algebra, such that ${\widetilde{B}}$ admits a compatible skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$.
\begin{enumerate}
\item[(i)] There exist a unique up to rescaling non-zero global section $\chi$ of the anticanonical bundle $K_{Y({\widetilde{B}})}^*$ such that for every $(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})$,
\[
\chi|_{S_{\x'}} = c_{(\x', {\widetilde{B}}')} x'_1\ldots x'_N \frac{\partial}{\partial x'_1} \wedge \cdots \wedge \frac{\partial}{\partial x'_N}
\]
for some $c_{(\x', {\widetilde{B}}')} \in {\mathbb C}^\times$.
\item[(ii)] For each $\theta \in \Delta$, either $\chi_\theta=0$ or $\chi_\theta$ is a non-zero scalar multiple of $\chi$.
The latter is the case for at least one $\theta \in \Delta$.
\end{enumerate}
\eth
\begin{proof} First we show that for all $\theta \in \Delta$ and $(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})$,
\begin{equation}
\label{chi-theta}
\chi_\theta|_{S_{\x'}} = c_{\theta, (\x', {\widetilde{B}}')} x'_1\ldots x'_N \frac{\partial}{\partial x'_1} \wedge \cdots \wedge \frac{\partial}{\partial x'_N}
\end{equation}
for some $c_{\theta, (\x', {\widetilde{B}}')} \in {\mathbb C}$. \leref{Caction}(ii) implies that
\[
\psi_\nu(t) \cdot x'_k := t^{\ell a_k } x'_k, \; \; \forall k \in [1,N]
\]
for some $a_k \in {\mathbb Z}$. Therefore,
\[
w_\nu|_{S_{\x'}^\circ} =
\sum_{k=1}^N \ell a_k x'_k \frac{\partial}{\partial x'_k} \cdot
\]
This equality and the fact that
\[
\pi|_{S_{\x'}} = \sum_{1 \leq i, k \leq N} \Lambda_{(\x', {\widetilde{B}}')}(e_i,e_k) x'_i x'_k \frac{\partial}{\partial x'_i} \wedge \frac{\partial}{\partial x'_k}
\]
imply \eqref{chi-theta}. Since $S_{\x'}$ is a Zariski open subset of the irreducible affine variety $Y({\widetilde{B}})$, the following are equivalent
for $\theta \in \Delta$:
\begin{enumerate}
\item $\chi_\theta \neq 0$,
\item $c_{\theta, (\x, {\widetilde{B}})} \neq 0$;
\item $c_{\theta, (\x', {\widetilde{B}}')} \neq 0$ for all $(\x', {\widetilde{B}}') \sim (\x, {\widetilde{B}})$.
\end{enumerate}
The proof of \cite[Lemma 2.4]{GSV} gives that $T({\widetilde{B}})$ acts transitively on the symplectic leaves of $(S_{\x'}^\circ, \pi|_{V})$. Furthermore,
for $y \in Y({\widetilde{B}})^ {\mathrm{reg}} $ the tangent space at $y$ of the orbit $T({\widetilde{B}}) \cdot y$ is
\begin{equation}
\label{span-T}
T_y \big( T({\widetilde{B}}) \cdot y \big) = \Span_{\mathbb C} \{ w_{\nu^1, y}, \ldots, w_{\nu^{N- 2r}, y} \}.
\end{equation}
Therefore, there exists $\theta_0 \in \Delta$ such that
\[
c_{\theta_0, (\x, {\widetilde{B}})} \neq 0,
\]
which, in view of the above equivalence, implies that
\[
\chi := \chi_{\theta_0}
\]
satisfies the conditions in the proposition. Furthermore, for all $\theta \in \Delta$, $\chi_\theta \chi_{\theta_0}^{-1}$
is a meromorphic function on $Y({\widetilde{B}})$, which by \eqref{chi-theta} is constant on every Zariski open subset $S_{\x'}$, and thus,
it should be constant on $Y({\widetilde{B}})$. The same argument proves the uniqueness of a section $\chi$ with stated properties in part (i).
\end{proof}
\subsection{Normality and codimension two results}
\label{5.4}
Recall that an upper cluster algebra $\U({\widetilde{B}})$ is called totally coprime \cite[Sect. 1.2]{BFZ05} if every two columns of all of its extended exchange matrices are linearly independent.
If the exchange matrix ${\widetilde{B}}$ has full rank, then this condition is satisfied. In turn, this condition is
satisfied whenever ${\widetilde{B}}$ has a compatible skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$, \cite[Proposition 1.8]{BFZ05}.
\bpr{codim2} If $\U({\widetilde{B}})$ is a finitely generated, totally coprime upper cluster algebra, then
\[
(Y({\widetilde{B}})^\circ)^ {\mathrm{sing}} \quad \mbox{and} \quad
Y({\widetilde{B}})^\circ \backslash \big( S_{\x} \cup (\cup_{k \in \ex} S_{\mu_k(\x)}) \big)
\]
are closed subvarieties of $Y({\widetilde{B}})^\circ= \MaxSpec \U({\widetilde{B}})$, each irreducible component of which lies in codimension $\geq 2$.
\epr
\begin{proof}
Since $\U({\widetilde{B}})$ is Noetherian and normal, Serre's Criterion (e.g.~\cite[Theorem 11.5.i]{Eis95}) implies that each irreducible component of the singular locus $(Y({\widetilde{B}})^\circ)^ {\mathrm{sing}}$ has codimension $\geq2$ (this is the \emph{R1 condition}).
Because $\U({\widetilde{B}})$ is totally coprime, \cite[Corollary 1.7]{BFZ05} implies that the coordinate ring of the open subvariety $ S_{\x} \cup (\cup_{k \in \ex} S_{\mu_k(\x)}) \subseteq Y({\widetilde{B}})^\circ$ (called the \emph{upper bound cluster algebra} in \textit{loc.~cit.}) is equal to the coordinate ring of $Y({\widetilde{B}})^\circ$ (which is the upper cluster algebra $\U({\widetilde{B}})$). By the Algebraic Hartogs Lemma, the complement of $ S_{\x} \cup (\cup_{k \in \ex} S_{\mu_k(\x)})$ in $Y({\widetilde{B}})^\circ$ must have codimension $\geq2$; see \cite[Lemma~4.3.1]{MM15} for the general statement and \cite[Lemma~4.4.2]{MM15} for the statement in this setting.
\end{proof}
\subsection{Proof of \thref{open-T-orb-sl}}
\label{5.5}
We prove a stronger result than that of \thref{open-T-orb-sl}, namely that $Y({\widetilde{B}})^ {\mathrm{reg}} \backslash {\mathcal{V}}(x)$
is a single torus orbit of symplectic leaves of $(Y({\widetilde{B}}), \pi)$ for a subtorus of $T({\widetilde{B}})$ of rank equal to
\[
\dim Y({\widetilde{B}}) - 2 \rk \pi = \dim Y({\widetilde{B}}) - \dim {\mathscr{L}},
\]
where ${\mathscr{L}}$ is a symplectic leaf of $Y({\widetilde{B}})$ of maximal dimension.
\bth{small-T-orbit} Assume that in the setting of \thref{sections-antican},
$\theta = \{ i_1, \ldots, i_{N -2r} \}$ is one of the elements of $\Delta$ such that
$\chi_\theta \neq 0$. Let $T({\widetilde{B}})_0 \cong ({\mathbb C}^\times)^{N-2r}$ be the subtorus
of $T({\widetilde{B}}) \cong ({\mathbb C}^\times)^{n({\widetilde{B}})}$,
corresponding to coordinates $i_1, \ldots, i_{N-2r}$. Then the following hold:
\begin{enumerate}
\item[(i)]
The restriction $\chi_\theta|_{(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} }$ is a nowhere vanishing section of the
anticanonical bundle $K^*_{(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} }$.
\item[(ii)] The Zariski open subset
\[
Y({\widetilde{B}})^ {\mathrm{reg}} \backslash {\mathcal{V}}(x) = (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}}
\]
of $Y({\widetilde{B}})$ is a single $T({\widetilde{B}})_0$-orbit of symplectic leaves of $(Y({\widetilde{B}}), \pi)$.
\end{enumerate}
\eth
\begin{proof} (i) The section $\chi_\theta|_{(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} }$ does not vanish on
\[
(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} \backslash \big( S_{\x} \cup (\cup_{k \in \ex} S_{\mu_k(\x)}) \big)
\]
by \thref{sections-antican}. It can not vanish anywhere on $(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} $ because otherwise its zero locus
will be a codimension 1 subvariety of ${(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} }$, which is disjoint from the union
$S_{\x} \cup (\cup_{k \in \ex} S_{\mu_k(\x)})$. This would contradict \prref{codim2}.
(ii) Let ${\mathscr{L}}_0$ be a symplectic leaf of $(S_\x, \pi|_{S_\x})$. There exists a symplectic leaf ${\mathscr{L}}$ of $(Y({\widetilde{B}}), \pi)$
such that ${\mathscr{L}}_0$ is a connected component of the intersection ${\mathscr{L}} \cap S_\x$. The proof of \cite[Lemma 2.4]{GSV}
gives that
\[
T({\widetilde{B}})_0 \cdot {\mathscr{L}}_0 = S_\x.
\]
This proves the first inclusion in
\begin{equation}
\label{2containments}
S_\x \subseteq T({\widetilde{B}})_0 \cdot {\mathscr{L}} \subseteq (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} .
\end{equation}
The second inclusion follows from \prref{containment}. Consider a point
\[
y \in (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}}
\]
and let ${\mathscr{L}}_y$ be a neighborhood (in the $C^\infty$ topology) of the symplectic leaf
of $(Y({\widetilde{B}}), \pi)$ through $y$. Since
\[
w_{\nu_{i_1}, y} \wedge \ldots \wedge w_{\nu_{i_{N-2r}}, y} \wedge (\wedge^{2r} \pi_y) = \chi_{\theta, y} \neq 0,
\]
there exists a neighborhood (in the $C^\infty$ topology) ${\mathscr{O}}_1$ of the identity element of the torus
$T({\widetilde{B}})_0$ such that ${\mathscr{O}}_1 \cdot {\mathscr{L}}_y$ contains a neighborhood (in the $C^\infty$ topology) of $y$ in $(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} $.
Since $S_\x$ is a Zariski open subset of the irreducible variety $(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} $,
\[
T({\widetilde{B}})_0 \cdot {\mathscr{L}}_0 \cap {\mathscr{O}}_1 \cdot {\mathscr{L}}_y =
S_\x \cap {\mathscr{O}}_1 \cdot {\mathscr{L}}_y \neq \varnothing.
\]
Hence, there exist $t \in {\mathscr{O}}_1$ and $t' \in T({\widetilde{B}})_0$ such that
\[
\psi(t) \cdot {\mathscr{L}}_y \subseteq \psi(t') \cdot {\mathscr{L}}.
\]
Note that the right hand side is a symplectic leaf of $(Y({\widetilde{B}}), \pi)$, while the left hand side is a
$C^\infty$ open subset of a symplectic leaf of $(Y({\widetilde{B}}), \pi)$.
Therefore,
\[
y \in {\mathscr{L}}_y \subseteq \psi( t^{-1} t') {\mathscr{L}},
\]
which shows that the second containment in \eqref{2containments} is an equality.
\end{proof}
\sectionnew{Poisson order structures}
\lb{qtorus}
This section contains a construction of Poisson order structures on strict root of unity upper quantum cluster algebras that will be
used to link Poisson geometry and representation theory.
\subsection{Poisson orders on mixed root of unity quantum tori}
\label{4.1}
Denote
\[
{\mathbb Z}^N_\geq := \{ (m_1, \ldots m_N) \in {\mathbb Z}^N \mid m_i \geq 0, \forall i \in \ninv \}.
\]
\bpr{Pord-tori} Assume that $\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv)$ is a strict root of unity upper quantum cluster algebra and $\ell$ is an odd positive integer that is coprime to the diagonal entries
of the skew-symmetrizing matrix $D$ for the principal part of the exchange matrix ${\widetilde{B}}$.
Then, for every seed $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$, the pair
\[
(\TT_\varepsilon(M_\varepsilon)_\geq, {\mathcal{L}}_\varepsilon(M_\varepsilon)_\geq)
\]
has a unique Poisson order structure $\partial' : {\mathcal{L}}_\varepsilon(M_\varepsilon)_\geq \to \Der_{\mathbb C}(\TT_\varepsilon(M_\varepsilon)_\geq)$ such that
\begin{equation}
\label{partial'}
\partial'_{M'_\varepsilon(\ell f)} (M_\varepsilon(g)) = \frac{1}{\ell} \Lambda_{(M'_\varepsilon, {\widetilde{B}}')}(f,g) M_\varepsilon(\ell f +g).
\end{equation}
The underlying Poisson structure is given by
\begin{equation}
\label{Poisson-L2}
\{ y'_k, y'_i \}_{\partial'} = \Lambda_{(M'_\varepsilon, {\widetilde{B}}')}(e_k,e_i) y'_k y'_i, \quad \forall k,i \in [1,N],
\end{equation}
where $y'_k := M'_\varepsilon(e_k)^\ell$ for $1 \leq k \leq N$.
\epr
Recall from \prref{struct-mut} the construction of the skew-symmetric matrices $\Lambda_{(M'_\varepsilon, {\widetilde{B}}')} \in M_N({\mathbb Z})$,
identified with the corresponing bilinear forms
\[
\Lambda_{(M'_\varepsilon, {\widetilde{B}}')} : {\mathbb Z}^N \times {\mathbb Z}^N \to {\mathbb Z}.
\]
In the proposition we suppress the dependence of the ${\mathbb C}$-linear map $\partial'$ on the choice of seed $(M'_\varepsilon, {\widetilde{B}}')$
for simplicity of the notation. Note that, in view of \eqref{Tep-geq} and \eqref{Lep-geq},
\[
\TT_\varepsilon(M'_\varepsilon)_\geq = \Span_{\mathbb C} \{ M'_\varepsilon(f) \mid f \in {\mathbb Z}^N_\geq\}, \quad
{\mathcal{L}}_\varepsilon(M'_\varepsilon)_\geq = \Span_{\mathbb C} \{ M'_\varepsilon(\ell f) \mid f \in {\mathbb Z}^N_\geq\}.
\]
\begin{proof} Set $\Lambda':= \Lambda_{(M'_\varepsilon, {\widetilde{B}}')}$. Let $(M'_q, {\widetilde{B}}') \sim (M_q, {\widetilde{B}})$ be the seed
corresponding to $(M'_\varepsilon, {\widetilde{B}}')$ under the exchange graph isomorphisms of Theorems \ref{tisom-e-graphs} and \ref{tcent-isom}(i).
Consider the surjective ${\mathbb C}$-algebra homomorphism from \eqref{kappa} for the seed $(M'_\varepsilon, {\widetilde{B}}')$:
\begin{equation}
\label{kappa-ep'}
\kappa'_\varepsilon : \TT_q(M'_q)_\geq \to \TT_\varepsilon(M'_\varepsilon)_\geq
\end{equation}
and its ${\mathbb C}$-linear section
\begin{equation}
\label{iota}
\iota'_\varepsilon : \TT_\varepsilon(M'_\varepsilon)_\geq \to \TT_q(M'_q)_\geq, \quad M'_\varepsilon(f) \mapsto M'_q(f), \; \; \forall f \in {\mathbb Z}^N_\geq.
\end{equation}
By \leref{Pord-spec} and eq. \eqref{ker-kappa}, we obtain
a Poisson order structure ${\mathcal{D}}'$ on the pair $(\TT_\varepsilon(M'_\varepsilon), {\mathcal{Z}}(\TT_\varepsilon(M'_\varepsilon)))$. It is given by
\begin{align*}
{\mathcal{D}}'_{M'_\varepsilon(\ell f)} (M_\varepsilon(g)) &= \kappa'_\varepsilon \Bigg( \frac{q^{\ell \Lambda'(f,g)/2}- q^{-\ell \Lambda'(f,g)/2}}{q^{1/2} - \varepsilon^{1/2}} M'_\varepsilon(\ell f +g) \Bigg),
\\&= c M'_\varepsilon(\ell f +g), \quad \forall f,g \in {\mathbb Z}^N_\geq,
\end{align*}
where
\[
c = \frac{q^{\ell \Lambda'(f,g)/2}- q^{-\ell \Lambda'(f,g)/2}}{q^{1/2} - \varepsilon^{1/2}} \Bigg{|}_{q^{1/2} = \varepsilon^{1/2}} = 2 \ell \Lambda'(f,g) \varepsilon^{\ell \Lambda'(f,g)/2} \varepsilon^{-1/2}
= 2 \ell \varepsilon^{-1/2} \Lambda'(f,g).
\]
The underlying Poisson structure satisfies
\[
\{ y'_k, y'_i \}_{{\mathcal{D}}'} =
{\mathcal{D}}'_{M'_\varepsilon(\ell e_k)} \big( M'_\varepsilon(\ell e_i) \big) =
2 \ell^2 \varepsilon^{-1/2} \Lambda'(e_k,e_i) y'_k y'_i, \quad k,i \in [1,N].
\]
In particular, ${\mathcal{L}}_\varepsilon(M'_\varepsilon)_\geq$ is a Poisson subalgebra of ${\mathcal{Z}}({\mathcal{L}}_\varepsilon(M'_\varepsilon)_\geq)$ with respect to the bracket
$\{-, -\}_{{\mathcal{D}}'}$. Therefore,
\[
\partial' := \frac{\varepsilon^{1/2}}{2 \ell^2} {\mathcal{D}}'|_{{\mathcal{L}}_\varepsilon(M'_\varepsilon)_\geq}
\]
is a Poisson order structure on the pair $(\TT_\varepsilon(M'_\varepsilon)_\geq, {\mathcal{L}}_\varepsilon(M'_\varepsilon)_\geq)$, which is given by \eqref{partial'}.
Eq. \eqref{Poisson-L2} follows at once from \eqref{partial'}. In light of the ${\mathbb C}$-linearity of $\partial'$, there
is a unique Poisson order structure satisfying \eqref{partial'}.
\end{proof}
\subsection{Poisson order structures on root of unity upper quantum cluster algebras}
\label{4.2}
\bth{Uep-Pord} Assume that $\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv)$ is a strict root of unity upper quantum cluster algebra and $\ell$ is an odd positive integer that is coprime to the diagonal entries
of the skew-symmetrizing matrix $D$ for the principal part of the exchange matrix ${\widetilde{B}}$ such that
\begin{equation}
\label{assume}
\A_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv) = \U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv).
\end{equation}
Then the pair $(\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv), \C\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv))$ admits a Poisson order structure such that under the isomorphism
\[
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv) \cong \U({\widetilde{B}},\inv)
\]
from \eqref{U-CU-isom} the underlying Poisson structure on $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv)$ corresponds to the GSV Poisson structure on $\U({\widetilde{B}},\inv)$.
\eth
\begin{proof} Recall the surjective ${\mathbb C}$-algebra homomorphism
\[
\kappa_\varepsilon : \TT_q(M_q)_\geq \to \TT_\varepsilon(M_\varepsilon)_\geq
\]
from \eqref{kappa} and consider its restriction to $\U_q(M_\varepsilon, {\widetilde{B}},\inv)$. For all seeds $(M'_q, {\widetilde{B}}') \sim (M_q, {\widetilde{B}})$ and indices $k \in \ex$, we have
\begin{multline}
\label{eq-int}
(q^{1/2} - \varepsilon^{1/2}) \TT_q(M'_q)_\geq \cap \TT_q(M''_q)_\geq = \TT_q(M'_q)_\geq \cap (q^{1/2} - \varepsilon^{1/2}) \TT_q(M''_q)_\geq
\\
= (q^{1/2} - \varepsilon^{1/2}) \big( \TT_q(M'_q)_\geq \cap \TT_q(M''_q)_\geq \big).
\end{multline}
The first equality follows from \cite[Proposition 5.10]{NTY2} and the second is an immediate consequence of the first one.
Since $\Ker \kappa_\varepsilon = (q^{1/2} - \varepsilon^{1/2}) \TT_q(M_q)_\geq$ by eq. \eqref{ker-kappa},
\begin{equation}
\label{ker-kep-Upe}
\Ker \kappa_\varepsilon|_{\U_q(M_\varepsilon, {\widetilde{B}},\inv)} = (q^{1/2} - \varepsilon^{1/2}) \U_q(M_\varepsilon, {\widetilde{B}},\inv).
\end{equation}
The exchange graph isomorphism from \thref{cent-isom}(i) implies that
\begin{equation}
\label{subseteq}
\kappa_\varepsilon \big( \U_q(M_\varepsilon, {\widetilde{B}},\inv) \big) \subseteq \U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv).
\end{equation}
By \cite[Theorem 5.7]{NTY2}
\[
\kappa_\varepsilon (\A_q(M_q, {\widetilde{B}}, \inv)) =\A_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv).
\]
Using the quantum Laurent phenomenon \eqref{qLaurent} and the assumption \eqref{assume}, we obtain
\[
\kappa_\varepsilon (\U_q(M_q, {\widetilde{B}}, \inv)) \supseteq \kappa_\varepsilon (\A_q(M_q, {\widetilde{B}}, \inv)) =
\A_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) = \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv).
\]
This, combined with \eqref{subseteq}, gives
\[
\kappa_\varepsilon (\U_q(M_q, {\widetilde{B}}, \inv)) = \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv).
\]
In view of \eqref{ker-kep-Upe} and \leref{Pord-spec}, the restriction of $\kappa_\varepsilon$ to $\U_q(M_q, {\widetilde{B}}, \inv)$
gives rise to a Poisson order structure ${\mathcal{D}}_\U$ on
\[
( \U_\varepsilon(M_q, {\widetilde{B}}, \inv), {\mathcal{Z}}(\U_\varepsilon(M_q, {\widetilde{B}}, \inv)) ).
\]
Denote
\[
\partial_\U : = \frac{\varepsilon^{1/2}}{2 \ell^2} {\mathcal{D}}_\U.
\]
Consider a ${\mathbb C}$-linear section
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \to \U_q(M_q, {\widetilde{B}}, \inv)
\]
of the restriction of $\kappa_\varepsilon$ to $\U_q(M_q, {\widetilde{B}}, \inv)$. For an arbitrary seed $(M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})$,
extend it to a section of the homomorphism $\kappa'_\varepsilon : \TT_q(M'_q)_\geq \to \TT_\varepsilon(M'_\varepsilon)_\geq$ from \eqref{kappa-ep'}.
\prref{Pord-tori} and the independence of the underlying Poisson structure of a Poisson order obtained by specialization (\leref{Pord-spec})
imply that
\begin{equation}
\label{P-bra-U}
\{ y'_k, y'_i \}_{\partial_\U} =
\{ y'_k, y'_i \}_{\partial'} = \Lambda_{(M'_\varepsilon, {\widetilde{B}}')}(e_k,e_i) y'_k y'_i, \quad \forall k,i \in [1,N],
\end{equation}
for $y'_i := M'_\varepsilon(e_i)^\ell$. Therefore,
\[
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) := \bigcap_{ (M'_\varepsilon, {\widetilde{B}}') \sim (M_\varepsilon, {\widetilde{B}})} {\mathcal{L}}_\varepsilon(M_\varepsilon')_{\geq}
\]
is closed under the Poisson bracket $\{., .\}_{\partial_\U}$ and $\partial_U$ restricts to a Poisson order structure
\[
\partial'_\U : \C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \to \Der_{\mathbb C} \big( \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \big).
\]
Eqs. \eqref{Poisson-L} and \eqref{P-bra-U} imply that
the underlying Poisson structure on $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv)$ corresponds to the GSV Poisson structure on $\U({\widetilde{B}},\inv)$.
\end{proof}
\sectionnew{Root of unity quantum side: the fully Azumaya loci}
\lb{Azumaya}
In this section we prove our second main result describing the fully Azumaya loci of all finitely generated
strict upper cluster algebras $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ with the property $\U_\varepsilon = \A_\varepsilon$.
\subsection{Statement of the main result}
\label{6.1}
Consider a root of unity upper quantum cluster algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ and its central subalgebra
\[
\C \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \subset {\mathcal{Z}}(\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)),
\]
given by \eqref{CU}. If, the order of $\varepsilon$ is odd and coprime to the diagonal entries of the skew-symmetrizing matrix $D$
for the principal part of the exchange matrix ${\widetilde{B}}$, then by \thref{cent-isom}(iii),
\[
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \cong \U({\widetilde{B}}, \inv).
\]
If, in addition, $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a finitely generated ${\mathbb C}$-algebra, then by \thref{cent-isom}(iii),
it is a module finite domain over $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ and $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
is a finitely generated normal commutative ${\mathbb C}$-algebra. So, we can consider the fully Azumaya locus $\AA$
of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ with respect to $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$, recall \deref{fullyAzumaya}.
We will identify
\[
\MaxSpec \C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \cong Y({\widetilde{B}}) = \MaxSpec \U({\widetilde{B}}, \inv)
\]
and think of $\AA$ as of a Zariski open subset of $Y({\widetilde{B}})$, recall \leref{FAzumaya-dense}.
Recall \eqref{x-prod}. Denote the set of {\em{non-central frozen variables among the non-inverted ones}}:
\begin{equation}
\label{nc-def}
\nc := \{ i \in \ninv \mid M(e_i) \notin {\mathcal{Z}}(\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)) \}
\subset \ninv
\end{equation}
and set
\begin{equation}
\label{x-prod-nc}
x_\nc := \prod_{i \in \nc} x_i.
\end{equation}
Our main result provides an explicit characterization of the fully Azumaya locus $\AA$.
\bth{Azumaya}
Assume that $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a finitely generated strict root of unity upper quantum cluster algebra such that
the order of $\varepsilon$ is odd and coprime to the diagonal entries of the skew-symmetrizing matrix $D$
for the principal part of the exchange matrix ${\widetilde{B}}$ and
\[
\A_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv) = \U_\varepsilon(M_\varepsilon, {\widetilde{B}},\inv).
\]
Then the fully Azumaya locus $\AA$ of
$\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ with respect to the central subalgebra
\[
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) \cong \U({\widetilde{B}}, \inv)
\]
satisfies
\[
(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} \subseteq \AA \subseteq Y({\widetilde{B}}) \backslash {\mathcal{V}}(x_{\nc}).
\]
\eth
All important root of unity quantum cluster algebras that we are aware of do not have central frozen variables. In those situations
\[
\nc = \ninv \quad \mbox{and} \quad
x_\nc = x,
\]
so the upper bound for the fully Azumaya locus in \thref{Azumaya} becomes
\[
Y({\widetilde{B}}) \backslash {\mathcal{V}}(x_{\nc}) = Y({\widetilde{B}}) \backslash {\mathcal{V}}(x) = Y({\widetilde{B}})^\circ.
\]
\bco{nc1} If $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a strict root of unity upper quantum cluster algebra as in
\thref{Azumaya} such that
\[
M(e_i) \notin {\mathcal{Z}}(\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)), \quad \forall i \in \ninv,
\]
then
\[
(Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} \subseteq \AA \subseteq Y({\widetilde{B}})^\circ.
\]
\eco
However, in general, one cannot replace the upper bound in \thref{Azumaya}
with $Y({\widetilde{B}})^\circ$ as shown in \prref{nc2}.
\subsection{Proof of the first inclusion in \thref{Azumaya}}
\label{6.2}
The first inclusion in \thref{Azumaya} follows from the following:
\bpr{Azumaya2} In the setting of \thref{Azumaya}, for all
\[
\mathfrak{m} \in (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} \subset Y({\widetilde{B}}) \cong \MaxSpec \C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv),
\]
the algebra
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)
\]
is isomorphic to
\[
\TT_\varepsilon(\overline{\Lambda})/\mathfrak{n} \TT_\varepsilon(\overline{\Lambda}),
\]
where $\mathfrak{n}$ is any maximal ideal of ${\mathcal{L}}_\varepsilon(\overline{\Lambda})$ {\em{(}}recall \eqref{Tep} and \eqref{Lep}{\em{)}}
and $\Lambda \in M_N ({\mathbb Z})$ is the skew-symmetric integer matrix compatible with ${\widetilde{B}}$ as in \deref{strict-seed}. All irreducible representations
of the last algebra have dimension
\begin{equation}
\label{index}
\sqrt{[{\mathbb Z}^N:\Ker(\overline{\Lambda})]},
\end{equation}
which equals the PI degree of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$.
\epr
Here, $[G:H]$ denotes the index of a subgroups $H$ of a group $G$.
The matrix $\overline{\Lambda} \in M_N({\mathbb Z}/\ell)$ is identified with the corresponding bilinear
form ${\mathbb Z}^N \times {\mathbb Z}^N \to {\mathbb Z}/\ell$, and $\Ker (\overline{\Lambda})$ denotes the kernel of this
form. The index in \eqref{index} is finite because $\Ker (\overline{\Lambda}) \supseteq (\ell {\mathbb Z})^N$.
\begin{proof} The algebras
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)
\]
are isomorphic to each other for all
\[
\mathfrak{m} \in (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} .
\]
This follows by combining Theorems \ref{tUep-Pord} and \ref{topen-T-orb-sl}, \coref{P-ort-FAzumaya}, and the fact that $\varphi$ is an action of $T({\widetilde{B}})$ on
$\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ by algebra automorphisms that preserves the central subalgebra $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$, see \leref{Caction}.
So, we can restrict ourselves to the special case when
\begin{equation}
\label{special-m}
\mathfrak{m} \in \MaxSpec {\mathcal{L}}_\varepsilon(M_\varepsilon) \cong \MaxSpec {\mathcal{L}}(\x) \subset (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} .
\end{equation}
Denote by $E$ the multiplicative subset of $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$, generated by $M_\varepsilon(e_i)^{\ell}$ for $i \in [1,N] \backslash (\ex \sqcup \inv)$.
We have
\begin{equation}
\label{UTT-isom0}
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [E^{-1}] \cong \TT_\varepsilon(M_\varepsilon) \cong \TT_\varepsilon(\overline{\Lambda})
\end{equation}
and, on the level of their centers,
\begin{equation}
\label{UTT-isom}
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [E^{-1}] \cong {\mathcal{L}}_\varepsilon(M_\varepsilon) \cong {\mathcal{L}}_\varepsilon(\overline{\Lambda}).
\end{equation}
For the rest of the proof assume \eqref{special-m}. This implies that $\mathfrak{m}$ is a maximal ideal of $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
that is disjoint from $E$. This property and eq. \eqref{UTT-isom0} give
\begin{align*}
&\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)
\\
\cong &\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [E^{-1}] / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [E^{-1}]
\cong \TT_\varepsilon(\overline{\Lambda})/ \mathfrak{n} \TT_\varepsilon(\overline{\Lambda}).
\end{align*}
Here $\mathfrak{n}$ is the maximal ideal of ${\mathcal{L}}_\varepsilon(\overline{\Lambda})$ corresponding to the maximal ideal
$\mathfrak{m} [E^{-1}]$ of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [E^{-1}]$ under the isomorphism
\eqref{UTT-isom}.
Since $\TT_\varepsilon(\overline{\Lambda})$ is an Azumaya algebra of PI degree $\sqrt{[{\mathbb Z}^N:\Ker(\overline{\Lambda})]}$ by \cite[Proposition 6.1(3-4)]{HLY}, all irreducible representations
of $\TT_\varepsilon(\overline{\Lambda})$ have dimension equal to the same integer. The latter equals the PI degree of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ by \cite[Proposition 6.4]{HLY}.
\end{proof}
\subsection{Proof of the second inclusion in \thref{Azumaya}}
\label{6.3}
Consider a strict root of unity quantum cluster algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ with skew-symmetric integer matrix $\Lambda \in M_N ({\mathbb Z})$ as in \deref{strict-seed}.
For $j \in [1,N]\backslash (\ex \sqcup \inv)$, denote by $\Lambda_j \in M_{N-1}({\mathbb Z})$ the submatrix of $\Lambda$, obtained by removing the $j$-th row and column. Set
\[
{\mathbb Z}^{N-1}_j := \bigoplus_{i \neq j} {\mathbb Z} e_i \subset {\mathbb Z}^N.
\]
The bilinear form ${\mathbb Z}^{N-1}_j \times {\mathbb Z}^{N-1}_j \to {\mathbb Z}/\ell$ associated to $\overline{\Lambda}_j$ is the restriction to ${\mathbb Z}^{N-1}_j$ of the bilinear form
${\mathbb Z}^N \times {\mathbb Z}^N \to {\mathbb Z}/\ell$ associated to $\overline{\Lambda}$.
Let $\TT_\varepsilon(\overline{\Lambda}_j)$ be the root of unity quantum subtorus of $\TT_\varepsilon(\overline{\Lambda})$ spanned by $x^f$ for $f \in {\mathbb Z}^{N-1}_j$, cf. \eqref{Tep}.
Denote
\begin{equation}
\label{J-def}
J:= \ninv \backslash \{ j \}.
\end{equation}
It is easy to see that we have a second description of $\TT_\varepsilon(\overline{\Lambda}_j)$:
\begin{equation}
\label{Tquot}
\big( \TT_\varepsilon(\overline{\Lambda})_\geq [(x^{e_i})^{-1} ; i \in J] \big) / (x^{e_j}) \cong \TT_\varepsilon(\overline{\Lambda}_j).
\end{equation}
Recall that an element $a$ of an algebra $R$ is called normal if $Ra = aR$.
Here and below, for such an element $a \in R$, we denote the principal ideal
\[
(a) := Ra = aR.
\]
\bpr{Azumaya3} Assume the setting of \thref{Azumaya}.
Then the open subset
\[
\big( {\mathcal{V}}(x_j) \cap S_{\x} \big) \backslash {\mathcal{V}}(x /x_j)
\]
of ${\mathcal{V}}(x_j)$ is non-empty and for all
\begin{equation}
\label{m-assume}
\mathfrak{m} \in \big( {\mathcal{V}}(x_j) \cap S_{\x} \big) \backslash {\mathcal{V}}(x /x_j) \subset Y({\widetilde{B}}) \cong \MaxSpec \C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv),
\end{equation}
the algebra
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)
\]
is isomorphic to
\[
\TT_\varepsilon(\overline{\Lambda}_j)/\mathfrak{n} \TT_\varepsilon(\overline{\Lambda}_j),
\]
where $\mathfrak{n}$ is a maximal ideal of ${\mathcal{L}}_\varepsilon(\overline{\Lambda}_j)$
and $\Lambda \in M_N ({\mathbb Z})$ is the skew-symmetric integer matrix as in \deref{strict-seed}. All irreducible representations
of the last algebra have dimension
\[
\sqrt{[{\mathbb Z}^{N-1}_j:\Ker(\overline{\Lambda}_j)]}.
\]
\epr
\begin{proof} Recall \eqref{J-def}.
We have
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [M_\varepsilon(e_i)^{-\ell} ; i \in J] \cong \TT_\varepsilon(M_\varepsilon)_{\geq} [M_\varepsilon(e_i)^{-\ell} ; i \in J] \cong \TT_\varepsilon(\overline{\Lambda})_\geq [(x^{e_i})^{-\ell} ; i \in J].
\]
Combining this with \eqref{Tquot} gives
\begin{equation}
\label{UTj-isom}
\Big( \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [M_\varepsilon(e_i)^{-\ell} ; i \in J] \Big)/ (M(e_j)) \cong \TT_\varepsilon(\overline{\Lambda}_j).
\end{equation}
Analogously, on the level of centers one shows that
\begin{equation}
\label{isom-for-ideals}
\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [M_\varepsilon(e_i)^{-\ell} ; i \in J] \cong {\mathcal{L}}_\varepsilon(\overline{\Lambda})_\geq [(x^{e_i})^{-\ell} ; i \in J].
\end{equation}
The assumption \eqref{m-assume} implies that the maximal ideal $\mathfrak{m}$ of $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
is disjoint from the multiplicative set $[M_\varepsilon(e_i)^{\ell} ; i \in J]$.
Denote by $\mathfrak{n}$ the maximal ideal of ${\mathcal{L}}_\varepsilon(\overline{\Lambda})_\geq [x_i^{-\ell} ; i \in J]$ that corresponds to the maximal ideal
$\mathfrak{m} [M_\varepsilon(e_i)^{-\ell} ; i \in J]$ of $\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [M_\varepsilon(e_i)^{-\ell} ; i \in J]$ under the isomorphism
\eqref{isom-for-ideals}. We have,
\begin{align*}
R:=&\big(\TT_\varepsilon(\overline{\Lambda})_\geq [x_i^{-\ell} ; i \in J]) / \mathfrak{n} \big(\TT_\varepsilon(\overline{\Lambda})_\geq [x_i^{-\ell} ; i \in J] \big),
\\
\cong&\big( \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [M_\varepsilon(e_i)^{-\ell} ; i \in J] \big) / \mathfrak{m} \big( \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) [M_\varepsilon(e_i)^{-\ell} ; i \in J] \big)
\\
\cong&\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv).
\end{align*}
It follows from \eqref{m-assume} that
\[
(x^{e_j})^\ell \in \mathfrak{n},
\]
and thus, $(x^{e_j})^\ell =0$ as an element of the algebra $R$. Therefore, all irreducible representations of $R$ are annihilated by $x^{e_j}$,
and so, those representations are in bijection with the irreducible representations of $R/(x^{e_j})$. Now we invoke \eqref{UTj-isom}, to obtain
\[
R/(x^{e_j}) \cong \big( \TT_\varepsilon(\overline{\Lambda})_\geq [(x^{e_i})^{-\ell} ; i \in J] \big)
/ (x^{e_j}) \mathfrak{n}
\cong \TT_\varepsilon(\overline{\Lambda}_j)
\]
for an ideal $\mathfrak{n}'$ of $\TT_\varepsilon(\overline{\Lambda}_j)$. By \cite[Proposition 6.1(3-4)]{HLY}, all irreducible representations
of $\TT_\varepsilon(\overline{\Lambda}_j)$ have dimension equal to
\[
\sqrt{[{\mathbb Z}^{N-1}_j:\Ker(\overline{\Lambda}_j)]},
\]
and hence, the same holds for the algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$.
\end{proof}
\bpr{Azumaya4} Assume the setting of \thref{Azumaya}. Let $j \in \nc$. Let
\[
\mathfrak{m} \in \big( {\mathcal{V}}(x_j) \cap S_{\x} \big) \backslash {\mathcal{V}}(x /x_j) \subset Y({\widetilde{B}}).
\]
The all irreducible representations of the algebra
\[
\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) / \mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)
\]
have dimension strictly less than the PI degree of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$,
\[
\sqrt{[{\mathbb Z}^N:\Ker(\overline{\Lambda})]}.
\]
\epr
\begin{proof}
Obviously, $\Ker (\overline{\Lambda}) \cap {\mathbb Z}_j^{N-1} \subseteq \Ker ( \overline{\Lambda}_j)$. Therefore,
\[
[{\mathbb Z}^N:\Ker(\overline{\Lambda})] \geq [{\mathbb Z}^{N-1}_j: (\Ker (\overline{\Lambda}) \cap {\mathbb Z}_j^{N-1}) ] \geq
[{\mathbb Z}^{N-1}_j:\Ker(\overline{\Lambda}_j)].
\]
The first inequality is in fact strict. If it is an equality, then $e_j \in \Ker(\overline{\Lambda})$, which implies that
\[
x^{e_j} \in {\mathcal{Z}}(\TT_\varepsilon(\overline{\Lambda})),
\]
which implies that
\[
M(e_j) \in {\mathcal{Z}}(\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv))
\]
because $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ is a subalgebra of the skew field of fractions of $\TT_\varepsilon(M_\varepsilon)$.
This contradicts the definition \eqref{nc-def} of the set $\nc$. Hence,
\[
\sqrt{[{\mathbb Z}^N:\Ker(\overline{\Lambda}_j)]} <
\sqrt{[{\mathbb Z}^N:\Ker(\overline{\Lambda})]}.
\]
and the proposition now follows from \prref{Azumaya3}.
\end{proof}
With this proposition we complete the proof of \thref{Azumaya}:
\medskip
\noindent
{\em{Proof of the second inclusion in \thref{Azumaya}.}} Let $j \in \nc$. Propositions \ref{pAzumaya3} and
\ref{pAzumaya4} imply that
\[
\big( {\mathcal{V}}(x_j) \cap S_{\x} \big) \backslash {\mathcal{V}}(x /x_j) \subseteq Y({\widetilde{B}}) \backslash \AA
\]
and that the first set is a non-empty Zariski open subset of ${\mathcal{V}}(\x_j)$.
Since ${\mathcal{V}}(x_j)$ is irreducible by \prref{p-primes},
\[
\overline{\big( {\mathcal{V}}(x_j) \cap S_{\x} \big) \backslash {\mathcal{V}}(x /x_j)} = {\mathcal{V}}(x_j).
\]
Furthermore, \leref{FAzumaya-dense} implies that $Y({\widetilde{B}}) \backslash \AA$
is a Zariski closed subset of $Y({\widetilde{B}})$. Therefore,
\[
{\mathcal{V}}(x_j) \subseteq Y({\widetilde{B}}) \backslash \AA, \quad \forall j \in \nc.
\]
hence, ${\mathcal{V}}(x_\nc) \subseteq Y({\widetilde{B}}) \backslash \AA$, and thus
$\AA \subseteq Y({\widetilde{B}}) \backslash {\mathcal{V}}(x_{\nc})$.
\qed
\medskip
Last we prove that one cannot replace the upper bound in \thref{Azumaya}
with $Y({\widetilde{B}})^\circ$.
\bpr{nc2} Assume the setting of \thref{Azumaya}. Let
\[
j \in \ninv \backslash \nc.
\]
All points
\[
\mathfrak{m} \in \Big( {\mathcal{V}}(x_j) \cap \Big( \bigcup_{(\x', {\widetilde{B}}') \sim (\x,{\widetilde{B}})} S_{\x'} \Big) \Big) \backslash {\mathcal{V}}(x/x_j)
\]
belong to the fully Azumaya locus of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$ with respect to
$\C\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$.
\epr
\begin{proof}
It is sufficient to prove the statement for
\[
\mathfrak{m} \in ({\mathcal{V}}(x_j) \cap S_{\x} ) \backslash {\mathcal{V}}(x/x_j).
\]
Applying \prref{Azumaya3}, we obtain that all irreducible representations
of the algebra $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv) /\mathfrak{m} \U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv)$
have dimension
\[
\sqrt{[{\mathbb Z}^{N-1}_j:\Ker(\overline{\Lambda}_j)]},
\]
Since $j \in \ninv \backslash \nc$, we have $M(e_j) \in {\mathcal{Z}}(U_\varepsilon(M_\varepsilon, {\widetilde{B}}, \inv))$, and thus,
\[
e_j \in \Ker (\overline{\Lambda}).
\]
Therefore,
\[
[{\mathbb Z}^N:\Ker(\overline{\Lambda})] = [{\mathbb Z}^{N-1}_j: (\Ker (\overline{\Lambda}) \cap {\mathbb Z}_j^{N-1}) ] =
[{\mathbb Z}^{N-1}_j:\Ker(\overline{\Lambda}_j)].
\]
Hence,
\[
\sqrt{[{\mathbb Z}^N:\Ker(\overline{\Lambda}_j)]} <
\sqrt{[{\mathbb Z}^N:\Ker(\overline{\Lambda})]}.
\]
and the proposition now follows from \prref{Azumaya3}.
\end{proof}
\sectionnew{Special cases and examples}
\label{spec-cases}
\subsection{The Richardson divisor of a Schubert cell}
\label{7.1}
The explicit Zariski open torus orbit of symplectic leaves from \thref{open-T-orb-sl} is a far reaching generalization of the
complement of the Richardson divisor of a Schubert cell for a complex simple Lie group $G$.
Let $G$ be a complex simple Lie group with a pair of opposite Borel subgroups $B_\pm$. Denote by
$T := B_+ \cap B_-$ the corresponding maximal torus of $G$ and by $U_\pm$ the unipotent radicals of $B_\pm$.
Let $W$ be the Weyl group of $G$, identified with $N(T)/T$, where $N(T)$ is the normalizer of $T$ in $G$.
Denote by $s_1, \ldots, s_r$ the set of simple reflections of $W$ and by $l : W \to {\mathbb Z}_{\geq 0}$
the length function on $W$.
The full flag variety has a canonical Poisson structure $\pi$ which is the descent of the standard Poison--Lie
structure on $G$, \cite[Sect. 1.2-1.3]{CP}. This Poisson structure is invariant under $T$.
The {\em{Schubert cell}} of the full flag variety $G/B_+$ corresponding to $w \in W$ is
\[
X_w^\circ := B_+ w B_+/B_+ \subset G/B_+.
\]
The {\em{open Richardson variety}} \cite{KLS} in $G/B_+$ corresponding to the pair $(w,u) \in W \times W$ is
\[
R_{w,u} := \big( B_+ w B_+ \cap B_- u B_+ \big) /B_+ \subset G/B_+.
\]
It is non-empty precisely when $u \leq w$ in the Bruhat order. We have the partition:
\[
X_w^\circ = \bigsqcup_{u \in W, u \leq W} R_{w,u}.
\]
Denote the support of $w \in W$,
\[
\Supp w := \{ i \in [1,r] \mid s_i \leq w\}.
\]
This is precisely the set of indices $i \in [1,r]$ such that $s_i$ appears in one, and thus, in any reduced decomposition of $w$.
The {\em{Richardson divisor}} of $X_w^\circ$ is
\[
RD_w : = \bigcup_{i \in \Supp(w)} \Cl_{X_w^\circ}(R_{w,s_i}) = \bigsqcup_{u \in W, u \leq w, u \neq 1} R_{w,u},
\]
where $\Cl_Y(Z)$ stands for the Zariski closure of $Z$ in $Y$.
\bth{T-orbits} (Goodearl--Yakimov) \cite[Theorem 0.4]{GY0} The $T$-orbits of symplectic leaves of the Schubert cell $(X_w^\circ, \pi_{X_w^\circ})$ are
the open Richardson varieties $R_{w,u}$ for $u \in W$, $u \leq w$. There is a Zariski open $T$-orbit of leaves, which is the complement
of the Richardson divisor of the Schubert cell $X_w^\circ$:
\[
R_{w,1} = X_w^\circ \backslash RD_w.
\]
\eth
\bex{Richardson} The complement of the Richardson divisor $RD_w$ of the Schubert cell $X_w^\circ$ is a very special case of the Zariski open torus orbit
of symplectic leaves of $Y({\widetilde{B}}) = \MaxSpec U({\widetilde{B}}, \inv)$ from \thref{open-T-orb-sl}, as we show next.
(This can be shown in the more general case of symmetrizabe Kac--Moody groups, but the setting requires more technical details.)
To each $w \in W$, one associates an exchange matrix ${\widetilde{B}}_w$ of size
\[
l(w) \times (l(w) - |\Supp(w)|)
\]
by the first display before Theorem 10.1 of \cite{GY1}. It is known \cite{GLS1,GY3,SW} that for the corresponding cluster algebras
without inverted frozen variables (i.e., $\inv = \varnothing$):
\[
{\mathbb C}[X_w^\circ] = \A({\widetilde{B}}_w, \varnothing) = \U({\widetilde{B}}_w, \varnothing).
\]
Thus,
\begin{equation}
\label{YX}
Y({\widetilde{B}}_w) \cong X_w^\circ,
\end{equation}
and in particular,
\[
Y({\widetilde{B}}_w)^ {\mathrm{sing}} = \varnothing, \quad Y({\widetilde{B}}_w)^ {\mathrm{reg}} = X_w^ {\mathrm{sing}}.
\]
The frozen variables of the cluster algebra $\A({\widetilde{B}}_w, \varnothing)$ are the generalized minors (cf. \cite[Eq. (2.5)]{BFZ05}),
\[
\Delta_{\varpi, w \varpi}, \quad i \in \Supp(w),
\]
considered as functions on $X_w^\circ$ via the identification $U_+ \cap w U_- w^{-1} \cong B_+ w B_+/B_+$,
given by $g \mapsto g. wB_+/B_+$. There is a canonical isomorphism $T({\widetilde{B}}_w) \cong T$ and the corresponding
actions on \eqref{YX} coincide.
The exchange matrix ${\widetilde{B}}_w$ admits a compatible
skew-symmetric matrix $\Lambda_w \in M_{l(w)}({\mathbb Z})$ and the cluster algebra $\A({\widetilde{B}}_w, \varnothing)$,
admits a quantization, isomorphic to the corresponding integral quantum
unipotent cell \cite{GY2}. By specialization, the GSV Poisson structure
on $Y({\widetilde{B}}_w) = \MaxSpec \U({\widetilde{B}}_w, \varnothing) \cong X_w^\circ$ associated to $\Lambda_w$
concides with $\pi$. We have
\[
{\mathcal{V}}(\Delta_{\varpi, w \varpi}) = R_{w,s_i}, \quad \forall i \in \Supp(w).
\]
Thus,
\[
X_w^\circ \backslash RD_w = Y({\widetilde{B}})^\circ = (Y({\widetilde{B}})^\circ)^ {\mathrm{reg}} ,
\]
which shows how the Zariski open $T$-orbit of symplectic leaves in \thref{T-orbits} is a special case of that in
\thref{open-T-orb-sl} for a small class of cluster algebras.
\eex
\subsection{All frozen variables inverted}
\label{7.2}
There is an important special case of Theorems \ref{topen-T-orb-sl} and \ref{tAzumaya} when all frozen variables are inverted.
In those situations our results take on a particularly strong form. On the Poisson side we have:
\bco{open-T-orb-sl} If $\U({\widetilde{B}})$ is a finitely generated upper cluster algebra with all frozen variables inverted for which there exists a skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$
such that $(\Lambda, {\widetilde{B}})$ is a compatible pair, then the non-singular part of $\MaxSpec U({\widetilde{B}})$ is a single $T({\widetilde{B}})$-orbit of symplectic leaves of
the affine Poisson variety $(\MaxSpec U({\widetilde{B}}), \pi)$.
\eco
On the Azumaya loci side we have:
\bco{Azumaya-inv}
If $\U_\varepsilon(M_\varepsilon, {\widetilde{B}})$ is a finitely generated strict root of unity upper quantum cluster algebra
with all frozen variables inverted such that
the order of $\varepsilon$ is odd and coprime to the diagonal entries of the skew-symmetrizing matrix $D$
for the principal part of the exchange matrix ${\widetilde{B}}$ and $\U_\varepsilon(M_\varepsilon, {\widetilde{B}}) = \A_\varepsilon(M_\varepsilon, {\widetilde{B}})$,
then fully Azumaya locus of $\U_\varepsilon(M_\varepsilon, {\widetilde{B}})$
over $\C\U_\varepsilon(M_\varepsilon,{\widetilde{B}})$ contains the nonsingular part of $\MaxSpec \U({\widetilde{B}})$.
\eco
\subsection{Acyclic cluster algebras}
\label{7.3}
Another important special case of Theorems \ref{topen-T-orb-sl} and \ref{tAzumaya} is the case of acyclic cluster algebras
when $\U_\varepsilon(M_\varepsilon,{\widetilde{B}})=\A_\varepsilon(M_\varepsilon,{\widetilde{B}})$ and
an explicit presentation of $\U({\widetilde{B}})$ and $\U_\varepsilon(M_\varepsilon, {\widetilde{B}})$ can be given, which in particular implies that those algebras
are finitely generated.
Recall that the sign pattern of an exchange matrix ${\widetilde{B}}$ is encoded in the graph $\Gamma({\widetilde{B}})$ with vertex set $\ex$ and directed edges $(i,j)$
for the vertices $i,j$ with $b_{ij} >0$. We say that ${\widetilde{B}}$ (and the corresponding cluster algebras of various kinds) are acyclic
if $\Gamma({\widetilde{B}})$ has no oriented cycles, cf. \cite[Definition 1.14]{BFZ05}.
\bth{acycl1} (Berenstein-Fomin-Zelevinsky) \cite[Theorems 1.18 and 1.20]{BFZ05} If ${\widetilde{B}}$ is an acyclic exchange matrix
and all variables are exchangeable {\em{(}}$\ex = [1,N]${\em{)}}, then $\A({\widetilde{B}})=\U({\widetilde{B}})$ and this algebra is isomorphic to the ${\mathbb C}$-algebra
with generators $x_1, x'_1, \ldots, x_n, x'_n$ and relations
\begin{equation}
\label{x'x-rel}
x'_k x_k = \prod_{i, b_{ik}>0} x_i^{b_{ik}} + \prod_{i, b_{ik}<0} x_i^{-b_{ik}}, \quad \forall k \in [1,N].
\end{equation}
\eth
Analogously to this result and \cite[Theorem 7.5]{BerZe}, one proves the following:
\bpr{acycl2} If $\A_\varepsilon(M_\varepsilon, {\widetilde{B}})$ is an acyclic root of unity quantum cluster algebra for which
all variables are exchangeable {\em{(}}$\ex = [1,N]${\em{)}}, then $\A_\varepsilon(M_\varepsilon,{\widetilde{B}})=\U_\varepsilon(M_\varepsilon,{\widetilde{B}})$ and this algebra
is isomorphic to the ${\mathbb C}$-algebra with generators $y_1, y'_1, \ldots, y_n, y'_n$ and relations
\begin{equation}
\label{y'y-rel}
y'_k y_k = \varepsilon^{\mu_{ki}/2} \prod_{i, b_{ik}>0} y_i^{b_{ik}} + \varepsilon^{\nu_{ki}/2} \prod_{i, b_{ik}<0} y_i^{-b_{ik}}, \quad \forall k \in [1,N],
\end{equation}
where
\begin{align*}
\mu_{ki} &= \sum_{i<j, \, b_{ik}>0, b_{jk} >0} b_{ik} b_{jk} \phi_{ij} - \sum_{i, b_{ik}>0 } b_{ik} \phi_{ik},
\\
\nu_{ki} &= \sum_{i<j, \, b_{ik}<0, b_{jk} <0} b_{ik} b_{jk} \phi_{ij} + \sum_{i, b_{ik}<0 } b_{ik} \phi_{ik}
\end{align*}
and $(\phi_{ij}) \in M_N({\mathbb Z}/\ell)$ is the matrix of the root of unity toric frame $M_\varepsilon$.
\epr
An immediate consequence of \thref{open-T-orb-sl} is the following:
\bco{Poisson-acycl} If an acyclic exchange matrix ${\widetilde{B}}$ has a compatible skew-symmetric matrix $\Lambda \in M_N({\mathbb Z})$
and all variables are exchangeable {\em{(}}$\ex = [1,N]${\em{)}}, then the maximum spectrum $Y({\widetilde{B}})$
of the algebra with generators $x_1, x'_1, \ldots, x_n, x'_n$ and relations \eqref{x'x-rel} is an affine
Poisson manifold whose non-singular part $Y({\widetilde{B}})^ {\mathrm{reg}} $ is a single $T({\widetilde{B}})$-orbit of symplectic leaves.
\eco
An immediate consequence of \thref{Azumaya} is the following:
\bco{Azumaya-inv2}
Assume that ${\widetilde{B}}$ is an acyclic exchange matrix, $(\lambda_{ij}) \in M_N({\mathbb Z})$ is a compatible
skew-symmetric matrix and $\varepsilon$ is an odd root of unity whose order is coprime to the diagonal entries
of the skew-symmetrizing matrix $D$ for the principal part of ${\widetilde{B}}$. Consider the ${\mathbb C}$-algebra
$\U_\varepsilon$ with generators $y_1, y'_1, \ldots, y_n, y'_n$ and relations \eqref{y'y-rel} with $\phi_{ij}:= \overline{\lambda}_{ij}$.
Then the following hold:
\begin{itemize}
\item[(i)] The unital subalgebra $\C\U_\varepsilon$ of $\U_\varepsilon$ generated by $y_1^\ell, (y'_1)^\ell, \ldots, y_n^\ell, (y'_n)^\ell$ is central and isomorphic
to the algebra with generators $x_1, x'_1, \ldots, x_n, x'_n$ and relations \eqref{x'x-rel}
via $y_k^\ell \mapsto x_k$, $(y'_k)^\ell \mapsto x_k$. $\U_\varepsilon$ is a finitely generated module over $\C\U_\varepsilon$.
\item[(ii)] The fully Azumaya locus of $\U_\varepsilon$ with respect to $\C\U_\varepsilon$ contains
the non-singular part of $\MaxSpec(\C\U_\varepsilon)$.
\end{itemize}
\eco
\subsection{An acyclic example}
As an example of an acyclic cluster algebra, consider the following compatible pair.
\[ {\widetilde{B}} = B=
\begin{bmatrix}
0 & -2 \\
2 & 0
\end{bmatrix}
\hspace{1cm}
\Lambda =
\begin{bmatrix}
0 & 1 \\
-1 & 0
\end{bmatrix}
\]
As in the previous subsection, every cluster variable is mutable, and so, $\ex = [1,N] = [1,2]$ and $\inv=\varnothing$.
Because $\Ker({\widetilde{B}}^\top)=0$, the torus $T({\widetilde{B}})$ is trivial,
and so \thref{open-T-orb-sl} predicts that $Y({\widetilde{B}})^{ {\mathrm{reg}} }$ is a single symplectic leaf.
The quiver of $B$ is the Kronecker quiver, which is acyclic. By \thref{acycl1},
\[
\A({\widetilde{B}}) = \U({\widetilde{B}})
= {\mathbb C}[x_1,x_2,x_1',x_2']/( x_1x_1'-x_2^2-1,x_2x_2' - x_1^2-1).
\]
There is an even nicer presentation in terms of the following element
\[
z := x_1'x_2' - x_1x_2 \in \U({\widetilde{B}}).
\]
This element satisfies several notable identities.
\[
x_1x_2 z = x_1^2+x_2^2+1,
\hspace{1cm}
x_1 z = x_2+x_2',
\hspace{1cm}
x_2 z = x_1+x_1'.
\]
The latter two equations imply that $x_1,x_2$, and $z$ generate $\U({\widetilde{B}})$, and the first equation implies that the relations
among these generators are generated by a single element; therefore,
\[
\A({\widetilde{B}}) = \U({\widetilde{B}}) = {\mathbb C}[x_1,x_2,z]/( x_1x_2 z - (x_1^2 + x_2^2 + 1) ).
\]
Since
\[
f(x_1,x_2,z) := x_1x_2z - (x_1^2 + x_2^2 + 1)
\]
is an irreducible polynomial, the maximum spectrum of $Y(\widetilde{B})$ is isomorphic to the irreducible algebraic variety
${\mathcal{V}}(x_1x_2 z-(x_1^2+x_2^2+1)) \subset {\mathbb C}^3$ defined by $f(x_1,x_2,z)=0$.
The GSV Poisson structure on $Y(\widetilde{B})$ extends to the Poisson structure on $\mathbb{C}^3$ with potential $f(x_1, x_2, z)$.
The latter is defined on the coordinates by
\[
\{x_1,x_2\} := f_z= x_1x_2
,\;\;\;
\{x_1,z\} := - f_{x_2} = 2x_2-x_1 z
,\;\;\;
\{x_2, z\} := f_{x_1} = x_2 z - 2x_1.
\]
At each point in $Y({\widetilde{B}})$, the latter two functions cannot simultaneously be zero, since
\[
x_1\{x_1,z\} -x_2\{x_2,z\} =
x_1(x_2 z-2x_1) + x_2(x_1 z-2x_1)
=2x_1x_2 z -2x_1^2 -2x_2^2
\]
equals $2$ everywhere on $Y(\widetilde{B})$. This has two important consequences.
\begin{itemize}
\item The differential of the polynomial $f(x_1,x_2, z)$ is
\[ df = \{x_2, z\} dx_1 +\{z,x_1\}dx_2 + \{x_1,x_2\} d z. \]
Since $df$ does not vanish on ${\mathcal{V}}(f)$, it is a smooth variety, and so $Y({\widetilde{B}})^{ {\mathrm{reg}} }=Y({\widetilde{B}})$.
\item The Hamiltonian vector field of $z$ is
\[
H_{z}
= \langle \{z,x_1\}, \{z,x_2\},0\rangle.
\]
Since $H_z$ does not vanish on $Y({\widetilde{B}})$, the rank of the Poisson bracket cannot be 0 anywhere on $Y({\widetilde{B}})$. Since $Y({\widetilde{B}})$ is smooth and 2-dimensional, the rank must be $2$ everywhere.
\end{itemize}
Since $Y({\widetilde{B}})$ is connected, it is a single symplectic leaf as predicted.
An interesting feature of this example is that the cluster tori do not cover $Y({\widetilde{B}})=Y({\widetilde{B}})^{ {\mathrm{reg}} }$.
By \cite{SZ04}, the cluster variables may be indexed by the integers so that the clusters are pairs of adjacent variables
$(x_i,x_{i+1})$. These cluster variables may be defined recursively by the mutation identity
\[
x_{n-1}x_{n+1} = x_n^2+1.
\]
If $x_{n}=0$ at a point, then the mutation identities force $x_{n-1}x_{n+1}=1$ and
$x_{n-1}^2+1=0$, and so $x_{n-1}=\pm i$ and $x_{n+1}=\mp i$. It follows that there are four points in
$Y({\widetilde{B}})$ which are not in any cluster torus, on which the sequence of cluster variables takes a periodic sequence of values
\[...0,i,0,-i,0,-i,0,i,0,...\]
Under the embedding $(x_1,x_2,z):Y({\widetilde{B}}) \hookrightarrow {\mathbb C}^3$, these points are sent to
\[ (0,\pm i,0), (\pm i,0,0) \in {\mathbb C}^3. \]
The embedding $Y({\widetilde{B}})\hookrightarrow{\mathbb C}^3$ and these four points can be visualized as follows.
Fixing a value of $z$ is equivalent to intersecting the image of $Y({\widetilde{B}})$ with a plane, and the result is the curve
\[{\mathcal{V}}(x_1^2- z x_1x_2 +x_2^2+1)\subset {\mathbb C}^2.\
For every value of $z$, this curve is a conic that passes through the four points $(0,\pm i)$ and $(\pm i,0)$; Figure \ref{fig: pencilofconics} depicts these curves for five values of $z$. One may show that every conic through the points $(0,\pm i)$ and $(\pm i,0)$ appears as $z$ varies except one: the singular conic ${\mathcal{V}}(x_1x_2)$.\footnote{There are two other singular conics through the four points, which correspond to $z=2$ and $z=-2$.}
\begin{figure}[h!t]
\begin{centering}
\begin{tikzpicture}
\begin{scope}[xshift=-6cm,scale=.45]
\draw[->] (-3,0) to (3.25,0);
\draw[->] (0,-3) to (0,3.25);
\draw (-2,-.1) to (-2,.1);
\draw (-1,-.1) to (-1,.1);
\draw (1,-.1) to (1,.1);
\draw (2,-.1) to (2,.1);
\draw (-.1,-2) to (.1,-2);
\draw (-.1,-1) to (.1,-1);
\draw (-.1,1) to (.1,1);
\draw (-.1,2) to (.1,2);
\node at (0,-4) {$z=-3$};
\clip (-3,-3) rectangle (3,3);
\draw[thick,blue,variable=\t,domain=-3:2] plot ({2*cosh(\t)/sqrt(5)},{(-3*cosh(\t)+sqrt(5)*sinh(\t))/sqrt(5)});
\draw[thick,blue,variable=\t,domain=-3:2] plot ({-2*cosh(\t)/sqrt(5)},{(3*cosh(\t)+sqrt(5)*sinh(\t))/sqrt(5)});
\draw[thick,red,variable=\t,domain=-3:2] plot ({-2*sinh(\t)/sqrt(5)},{(3*sinh(\t)+sqrt(5)*cosh(\t))/sqrt(5)});
\draw[thick,red,variable=\t,domain=-3:2] plot ({2*sinh(\t)/sqrt(5)},{-(3*sinh(\t)+sqrt(5)*cosh(\t))/sqrt(5)});
\end{scope}
\begin{scope}[xshift=-3cm,scale=.45]
\draw[->] (-3,0) to (3.25,0);
\draw[->] (0,-3) to (0,3.25);
\draw (-2,-.1) to (-2,.1);
\draw (-1,-.1) to (-1,.1);
\draw (1,-.1) to (1,.1);
\draw (2,-.1) to (2,.1);
\draw (-.1,-2) to (.1,-2);
\draw (-.1,-1) to (.1,-1);
\draw (-.1,1) to (.1,1);
\draw (-.1,2) to (.1,2);
\node at (0,-4) {$z=-1.5$};
\path[use as bounding box] (-3,-3) rectangle (3,3);
\draw[thick,red,variable=\t,domain=-300:200,samples=100] plot ({2*cos(\t)/sqrt(4-1.5^2)},{-(1.5*cos(\t)+sqrt(4-1.5^2)*sin(\t))/(sqrt(4-1.5^2))});
\end{scope}
\begin{scope}[xshift=0cm,scale=.45]
\draw[->] (-3,0) to (3.25,0);
\draw[->] (0,-3) to (0,3.25);
\draw (-2,-.1) to (-2,.1);
\draw (-1,-.1) to (-1,.1);
\draw (1,-.1) to (1,.1);
\draw (2,-.1) to (2,.1);
\draw (-.1,-2) to (.1,-2);
\draw (-.1,-1) to (.1,-1);
\draw (-.1,1) to (.1,1);
\draw (-.1,2) to (.1,2);
\node at (0,-4) {$z=0$};
\clip (-3,-3) rectangle (3,3);
\draw[thick,red] (0,0) circle (1);
\end{scope}
\begin{scope}[xshift=3cm,scale=.45]
\draw[->] (-3,0) to (3.25,0);
\draw[->] (0,-3) to (0,3.25);
\draw (-2,-.1) to (-2,.1);
\draw (-1,-.1) to (-1,.1);
\draw (1,-.1) to (1,.1);
\draw (2,-.1) to (2,.1);
\draw (-.1,-2) to (.1,-2);
\draw (-.1,-1) to (.1,-1);
\draw (-.1,1) to (.1,1);
\draw (-.1,2) to (.1,2);
\node at (0,-4) {$z=1.5$};
\path[use as bounding box] (-3,-3) rectangle (3,3);
\draw[thick,red,variable=\t,domain=-300:200,samples=100] plot ({2*cos(\t)/sqrt(4-1.5^2)},{(1.5*cos(\t)+sqrt(4-1.5^2)*sin(\t))/(sqrt(4-1.5^2))});
\end{scope}
\begin{scope}[xshift=6cm,scale=.45]
\draw[->] (-3,0) to (3.25,0);
\draw[->] (0,-3) to (0,3.25);
\draw (-2,-.1) to (-2,.1);
\draw (-1,-.1) to (-1,.1);
\draw (1,-.1) to (1,.1);
\draw (2,-.1) to (2,.1);
\draw (-.1,-2) to (.1,-2);
\draw (-.1,-1) to (.1,-1);
\draw (-.1,1) to (.1,1);
\draw (-.1,2) to (.1,2);
\node at (0,-4) {$z=3$};
\clip (-3,-3) rectangle (3,3);
\draw[thick,blue,variable=\t,domain=-3:2] plot ({2*cosh(\t)/sqrt(5)},{(3*cosh(\t)+sqrt(5)*sinh(\t))/sqrt(5)});
\draw[thick,blue,variable=\t,domain=-3:2] plot ({-2*cosh(\t)/sqrt(5)},{(-3*cosh(\t)+sqrt(5)*sinh(\t))/sqrt(5)});
\draw[thick,red,variable=\t,domain=-3:2] plot ({2*sinh(\t)/sqrt(5)},{(3*sinh(\t)+sqrt(5)*cosh(\t))/sqrt(5)});
\draw[thick,red,variable=\t,domain=-3:2] plot ({-2*sinh(\t)/sqrt(5)},{-(3*sinh(\t)+sqrt(5)*cosh(\t))/sqrt(5)});
\end{scope}
\end{tikzpicture}
\end{centering}
\caption{Five curves in the family ${\mathcal{V}}(x_1^2- z x_1x_2 +x_2^2+1)$ (real part in blue and imaginary part in red)}
\label{fig: pencilofconics}
\end{figure}
For historical reasons, this family of curves is called the \textbf{pencil of conics} through the four \textbf{base points} $(0,\pm i)$ and $(\pm i,0)$.
The cluster $Y({\widetilde{B}})$ is then identified with the total space of the pencil of conics through $(0,\pm i)$ and $(\pm i,0)$ minus one of the three singular fibers.
Under this identification, the four points not in any cluster torus correspond to the base points in the $z=0$ fiber, which is the complex circle of radius -1.
|
\section{Introduction}
Let $G$ be a discrete group acting by isometries on a proper $\cat(0)$ space $X$. Let $\mu $ be a Borel probability measure on $G$, which we always assume admissible, meaning that the support of $\mu$ generates $G$ as a semigroup. Consider the sequence $\omega = (\omega_i)_i$, where the $\omega_i's $ are chosen independently according to the measure $\mu$. The random walk $(Z_n(\omega))_n$ on $G$ generated by $\mu $ is then defined by $Z_n(\omega) = \omega_1 \dots \omega_n$. Taking $o \in X $, we want to study the asymptotic behaviour of the random variables $(Z_n(\omega)o )_n$. To be more precise, we want to study limit laws of the random walk in a natural compactification of $X$. Even though these questions may be hard to solve for general metric spaces, the theory is very rich when $X$ possesses nice linear or hyperbolic-like properties. In the fundamental paper of V. Kaimanovich \cite{kaimanovitch00}, the convergence of $(Z_n o)_o$ to a point of the visual boundary is proven for groups acting geometrically on proper hyperbolic spaces and several other classes of actions. More recently this result has been extended by J. Maher and G. Tiozzo in \cite{maher_tiozzo18} for groups acting by isometries on non proper hyperbolic spaces. A major difficulty in the proof of the latter result was that in the non proper setting, the completion of a hyperbolic space by its Gromov boundary might be non compact. The results of Maher and Tiozzo will be fundamental in the sequel because we will deal with hyperbolic spaces without properness assumption. In non-specified $\cat(0)$ space, Karlsson and Margulis proved in \cite[Theorem 2.1]{karlsson_margulis} a first general result of convergence of the random walk, under the assumption that the escape rate $\lambda = \liminf \frac{d(Z_no, o)}{n}$ is positive.
In \cite{LeBars22} we proved that if $G$ acts on a $\cat(0)$ space $X$ with rank one isometries, then the random walk $(Z_n(\omega))_n$ almost surely converges to a point of the boundary of the visual compactification $\partial_\infty X$. A rank one element is an axial isometry whose axes do not bound any flat half plane. We give more details on this notion in Section \ref{background}, but a rank one element must be thought of as a contracting isometry with features that typically arise in hyperbolic settings. In this context, we also prove that the speed at which the random walk converges (the drift) is almost surely positive: there exists $\lambda >0$ such that almost surely, $\lim_n \frac{d(Z_n o, o )}{n} = \lambda$. We review these results in Section \ref{results hyp}. The present paper can be thought of as a continuation of \cite{LeBars22}, and the goal of this work is study further limit laws of the random walk $(Z_n o)_n$, and more specifically central limit theorems for the random variables $(d(Z_n o, o ))_n$.
\newline
The study of central limit theorem in noncommutative settings has a long history. In the case of a random product of matrices, a classical result of Furstenberg \cite{furstenberg63} is the following. Take $(M_n)$ a sequence of matrices in $GL_n(\mathbb{R})$, independent and identically distributed according to a probability measure $\mu$ whose support generates a noncompact subgroup of $GL_n(\mathbb{R})$ that does not preserve any proper linear subspace of $\mathbb{R}^n $. Assume that $\mu$ has finite first moment. Then there exists $\lambda >0$ such that for all $v \in \mathbb{R}^n - \{0\}$,
\begin{equation*}
\frac{1}{n} \log \| M_n \dots M_1 v\| \longrightarrow_n \lambda.
\end{equation*}
This result can be thought of as an analogue of a law of large numbers on the random walk $(M_n \dots M_1 v)_n$. In this context, central limit theorems and other limit laws were proven by Furstenberg-Kesten \cite{furstenberg_kesten60}, Le Page \cite{lepage82} and Guivarc'h-Raugi \cite{guivarch_raugi85}. These state that there exists $\sigma_\mu >0 $ such that for every $v \in \mathbb{R}^n$,
\begin{equation*}
\frac{\log \| M_n \dots M_1 v\| - n \lambda}{\sqrt{n}} \underset{n}{\longrightarrow} \mathcal{N}(0, \sigma_\mu^2),
\end{equation*}
where $\mathcal{N}(0, \sigma^2)$ is a centred Gaussian law on $\mathbb{R}$. We recall that the convergence in law means that for any bounded continuous function $F : \mathbb{R} \rightarrow \mathbb{R}$, one has
\begin{equation*}
\lim_{n \rightarrow \infty} \int_G F(\frac{\log \| M_n \dots M_1 v\| - n \lambda}{\sqrt{n}})d\mu^{\ast n}(g) = \int_\mathbb{R} F(t) \frac{\exp(-t^2 / 2 \sigma^2)}{\sqrt{2\pi \sigma^2}}dt.
\end{equation*}
Those kinds of results were also obtained in negative curvature settings, for example in Gromov-hyperbolic groups \cite{bjorklund09}. However, the results stated thus far were obtained under rather strong moment conditions. Typically, $\mu$ was assumed to have a finite exponential moment, meaning that there exists $\alpha >0 $ such that $\int_G \exp(\alpha d(o, g o)) d\mu(g)<\infty$.
More recently, Benoist and Quint have developed a new approach to this question and have proven central limit theorems in the linear context \cite{benoist_quint16CLTlineargroups} and for hyperbolic groups \cite{benoist_quint16}. They could weaken the moment condition and only assume that the measure $\mu $ has finite second moment $\int_G (\log\|g v\|)^2 d\mu(g) < \infty$. Namely, if $\mu$ is such a measure on a group $G$ acting non elementarily on a proper hyperbolic space $Y$ with basepoint $o$, then there exists $\lambda >0$ such that the random variables $\frac{d(Z_n(\omega)o, o) - n \lambda}{\sqrt{n}} $ converge in law to a non-degenerate Gaussian distribution \cite[Theorem 1.1]{benoist_quint16}.
\newline
Using this approach, C. Horbez proved central limit theorems for mapping class groups of closed connected orientable hyperbolic surfaces and on $\Out(F_N)$ \cite{horbez18}. More recently, T. Fern\'os, J. Lécureux and F. Mathéus proved that if $G$ is a group acting non-elementarily on a finite-dimensional $\cat(0)$ cube complex, then we also have a central limit theorem for the random variables $(d(Z_n(\omega) o,o))_n$ \cite{fernos_lecureux_matheus21}. In both cases, the authors only assume a second moment condition.
The main result of this paper is to prove a similar result in the context of a group acting on a general $\cat(0)$ space with a rank one isometry. We say that the group action $G \curvearrowright X$ is non-elementary if there are no fixed points in $\overline{X}$ nor a fixed pair of points in $\partial_\infty{X}$.
\begin{thm}\label{thm clt cat0}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Let $\lambda$ be the (positive) drift of the random walk. Then the random variables $\frac{1}{\sqrt{n}}(d(Z_n o, o) - n \lambda) $ converge in law to a Gaussian distribution $N_\mu$.
\end{thm}
Our strategy also relies heavily on the approach developed by Benoist and Quint. To summarize, they give a condition under which a given random cocycle converges in law to a Gaussian distribution. However, in order to prove that we can apply this result, one needs to obtain good estimates on this cocycle. In the aforementioned papers \cite{horbez18} and \cite{fernos_lecureux_matheus21}, the authors use hyperbolic models, on which they could derive quantitative estimates. To summarize, one studies the behaviour of the random walk on a hyperbolic model, on which we know many properties thanks to Maher and Tiozzo \cite{maher_tiozzo18}. Then, one must lift this information back on the original space $X$ using contracting properties.
\begin{itemize}
\item For $\Mod(S)$, the hyperbolic model is the curve complex $C(S)$, and the lifting to $\mathcal{T}(S)$ is done in \cite[Section 3.4]{horbez18}.
\item For a $\cat(0)$ cube complex, the hyperbolic model is the contact graph $\mathcal{C}X$, and the lifting is implemented in \cite[section 5]{fernos_lecureux_matheus21}.
\end{itemize}
In \cite{petyt_spriano_zalloum22}, H. Petyt, D. Spriano and A. Zallum introduced analogues of curve graphs and cubical hyperplanes for the class of $\cat(0)$ spaces. Using a generalized notion of hyperplane, they build a family of hyperbolic metrics $(d_L)_L$ on $X$ which conserve many of the geometric features of the original $\cat(0)$ space. These spaces capture hyperbolic behaviours in $X$ and behave very well under the isometric action of a group. Moreover, a rank one isometry of $X$ acts on some hyperbolic model as a loxodromic isometry. The strategy we use is then to chose a good hyperbolic model $X_L = (X, d_L)$, and then to make use of the limit laws proven by Maher and Tiozzo in \cite{maher_tiozzo18}.
Another interesting question in the study of $(Z_n(\omega))_n$ is the proportion of steps that are "hyperbolic". In the context of random walks on hyperbolic spaces, Maher and Tiozzo show that the probability that a random walk of size $n$ is a loxodromic isometry goes to $1$ as $n$ goes to infinity \cite[Theorem 1.4]{maher_tiozzo18}. For a non-elementary action on an irreducible $\cat(0)$ cube complex, Fern\'os, Lécureux and Mathéus show that the proportion of steps $Z_n$ that are contracting goes to $1$ as $n$ goes to infinity. They use this result to show that if a group $G$ acts non elementarily and essentially on a (possibly reducible) finite-dimensional $\cat(0)$ cube complex, then there exist regular elements, extending a result of Caprace and Sageev \cite{caprace_sageev11}. In our context, we also prove that "most" of the steps in the random walk are rank one. This result is not involved in the proof of Theorem \ref{thm clt cat0}, but is of independent interest.
\begin{thm}[Rank one elements in the random walk]
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$, and assume that $G $ contains a rank one element. Then
\begin{eqnarray}
\mathbb{P}(\omega \, : \, Z_n (\omega) \text{ is a contracting isometry }) \underset{n\rightarrow \infty}{\rightarrow} 1 \nonumber.
\end{eqnarray}
\end{thm}
While we were working on this project, Inhyeok Choi released a paper in which he states central limit theorems along with other limit laws in $\cat(0)$ spaces, Teichmüller spaces and outer spaces \cite{choi22}. One of the main assumptions is still the presence of a pair of independent contracting isometries in the group, but the methods and the proofs are different. Indeed, the author uses a pivotal technique introduced by Gouëzel in \cite{gouezel21}, while our paper uses hyperbolic models that depend on specific features of $\cat(0)$ spaces. We think that this approach is natural from a geometric point of view, and we hope that the interplay between $\cat(0)$ spaces and their underlying hyperbolic models will be useful in the study of still open questions about limit laws.
While this is not stated in this paper, our approach can be of use in order to determine if the boundary $\partial_\infty X$ endowed with the hitting measure is actually the Poisson boundary of $(G, \mu)$. Moreover, it seems natural to use these hyperbolic spaces to prove that almost surely, limit points of the random walk belong to the sublinear Morse boundary constructed by Qing and Rafi in \cite{qing_rafi22}.
\newline
In Section \ref{background}, we review basic definitions about random walks, rank one isometries and explain our setting. In Section \ref{hyperbolic models}, we
explicit the construction and properties of the hyperbolic models $(X, d_L)$, and give various geometric lemmas that will be useful afterwards. Section \ref{results hyp} is dedicated to presenting the works of Maher and Tiozzo in \cite{maher_tiozzo18}, and the first results in proper $\cat(0) $ spaces that were found in \cite{LeBars22}. We explain the strategy developed by Benoist and Quint in Section \ref{strategy}, and give the proof of our main Theorem in Section \ref{Section proof}.
\begin{ackn}
This work was done during the trimester "Groups acting on fractals, hyperbolicity and self-similarity", which was held at IHP from April 11th to July 8th, 2022. The author acknowledges support of the Institut Henri Poincaré (UAR 839 CNRS-Sorbonne Université), and LabEx CARMIN (ANR-10-LABX-59-01). We also thank Jean Lécureux for his help and commentaries on this paper. Last, we are very grateful to H. Petyt, D. Spriano and A. Zalloum for allowing us to discuss the details of their construction and for their friendly remarks.
\end{ackn}
\section{Background}\label{background}
\subsection{Random walks and $\cat(0)$ spaces}
Let $G$ be a discrete countable group and $\mu \in \text{Prob}(G)$ a probability measure on $G$. Throughout the article we will assume that $\mu $ is admissible, i.e. $\supp(\mu)$ generates $G$ as a semigroup. Let $(\Omega, \mathbb{P}) $ be the probability space $(G^{\mathbb{N}}, \delta_e \times \mu^{\mathbb{N^\ast}})$. The application
\begin{equation*}
(n, \omega) \in \mathbb{N} \times \Omega \mapsto Z_n(\omega) = \omega_1 \omega_2 \dots \omega_n,
\end{equation*}
where $\omega$ is chosen according to the law $\mathbb{P}$, defines the random walk on $G$ generated by the measure $\mu$.
Let now $(X,d)$ be a proper $\cat(0)$ metric space, on which $G$ acts by isometries. If the reader wants a detailed introduction to $\cat (0) $ spaces, the main references that we will use are \cite{bridson_haefliger99} and \cite{ballman95}. We recall that the boundary $ \partial_\infty X$ of a $\cat (0) $ space $X$ is the set of equivalent classes of rays $\sigma : [0, \infty) \rightarrow X$, where two rays $\sigma_1, \sigma_2$ are equivalent if they are asymptotic, i.e. if $d(\sigma_1(t), \sigma_2 (t))$ is bounded uniformly in $t$.
Given two points on the boundary $\xi$ and $\eta$, if there exists a geodesic line $\sigma : \mathbb{R} \rightarrow X$ such that the geodesic ray $\sigma_{[0, \infty)}$ is in the class of $\xi$ and the geodesic ray $t \in [0, \infty) \mapsto \sigma(-t)$ is in the class of $\eta$, we will say that the points $\xi $ and $\eta$ are joined by a geodesic line. The reader should be aware that in general, such a geodesic need not exist between any two points of the boundary, as can be seen in $\mathbb{R}^2$. A point $\xi$ of the boundary that is called a \textit{visibility point} if, for all $\eta \in \partial_\infty X - \{\xi\}$, there exists a geodesic from $\xi$ to $\eta$. We will see in the next section a criterion to prove that a given boundary point is a visibility point.
An important feature in $\cat(0)$ spaces is the existence of closest-point projections on complete convex subsets. More precisely, given a complete convex subset $C$ in a $\cat(0)$ space, there exists a map $\pi_C : X \rightarrow C$ such that $\pi_C(x)$ minimizes the distance $d(x,C)$:
\begin{prop}[{\cite[Lemma 2.4]{bridson_haefliger99}}]\label{projection prop}
The projection $\pi_C$ onto a convex complete subset in a $\cat(0)$ space satisfies the following properties:
\begin{itemize}
\item $\forall x \in X$, $\pi_C(x)$ is uniquely defined and $d(x, \pi_C(x))= d(x, C) = \inf_{c \in C} d(x, c) $;
\item if $x'$ belongs to the geodesic segment $[x, \pi(x)]$, then $\pi_C(x')= \pi_C(x)$;
\item $\pi $ is a retraction of $X$ onto $C$ that does not increase the distances: for all $x, y \in X$, we have $d(\pi_C(x), \pi_C(y )) \leq d(x,y)$.
\end{itemize}
\end{prop}
It is immediate to see that the above properties can be applied to geodesic segments, which are convex and complete with the induced metric. When $\gamma : [a, b] \rightarrow X$ is a geodesic segment, we will write $\pi_\gamma$ for the projection onto the image $[\gamma(a), \gamma(b)]\subseteq X$.
When $X$ is a proper space, the space $\overline{X} = X \cup \partial X $ is a compactification of $X$, that is, $\overline{X}$ is compact and $X$ is an open and dense subset of $\overline{X}$. We recall that the action of $G$ on $X$ extends to an action on $\partial_\infty X$ by homeomorphisms.
Another equivalent construction of the boundary can be done using horofunctions. If $x_n \rightarrow \xi \in \partial_\infty X$ and $x \in X$, we denote by $b_\xi^{x} : X \mapsto \mathbb{R}$ the horofunction given by
\begin{equation*}
b_\xi^{x}(z) = \lim_n d(x_n, z) - d(x_n, x).
\end{equation*}
It is a standard result in $\cat(0)$ geometry (see for example \cite[Proposition II.2.5]{ballman95}) that this limit exists and that given any basepoint $x$, a horofunction characterizes the boundary point $\xi $. When the context is clear. we will often omit the basepoint and just write $ b_\xi $.
\subsection{Rank one elements}\label{rank one section}
Let $g \in G$. We say that $g$ is a \textit{semisimple} isometry if its displacement function $ x \in X \mapsto \tau_g(x) = d(x , gx)$ has a minimum in $X$. If this minimum is non-zero, it is a standard result (see for example \cite[Proposition II.3.3]{ballman95}) that the set on which this minimum is obtained is of the form $C \times \mathbb{R}$, where $C$ is a closed convex subset of X. On the set $\{c\}\times \mathbb{R}$ for $c \in C$, $g$ acts as a translation, which is why $ g$ is called \textit{axial} and the subset $\{c\}\times \mathbb{R}$ is called an \textit{axis} of $g$. A \textit{flat half-plane} in $X$ is defined as a euclidean half plane isometrically embedded in $X$.
\begin{Def}
We say that a geodesic in $X$ is \textit{rank one} if it does not bound a flat half-plane. If $g$ is an axial isometry of $X$, we say that $g$ is rank one if no axis of $g$ bounds a flat half-plane.
\end{Def}
If $G$ acts on $X$ by isometries and possesses a rank one element $g \in G$ for this action, we may say that $G$ is rank one. However, the theory of $\cat(0)$ groups is not as clear as for Gromov hyperbolic groups. For example, there is no good (i.e. invariant under quasi isometry) notion of boundary of a $\cat(0)$ group, as shown by Croke and Kleiner in \cite{croke_kleiner00}. To summarize, it is better to keep in mind that "rank one" is always attached to a given action $G \curvearrowright X$ on a $\cat(0)$ space.
\newline
More information on rank one isometries and geodesics can be found in \cite[Section III. 3]{ballman95}, and more recently in \cite{caprace_fujiwara10} and in \cite{bestvina_fujiwara09}.
\begin{Def}
We say that the action $G \curvearrowright X$ of a rank one group $G$ on a $\cat (0) $ space $X$ is \textit{non-elementary} if $G$ neither fixes a point in $\partial_\infty X$ nor stabilizes a geodesic line in $X$.
\end{Def}
To justify this definition, we use a result from Caprace and Fujiwara in \cite{caprace_fujiwara10}. What follows comes from the aforementioned paper.
\begin{Def}
Let $g_1, \, g_2 \in G$ be axial isometries of $G$, and fix $x_0 \in X$. The elements $g_1, g_2 \in G$ are called independent if the map
\begin{equation}
\mathbb{Z} \times \mathbb{Z} \rightarrow [0, \infty) : (m,n) \mapsto d(g_1^m x_0, g_2^nx_0)
\end{equation}
is proper.
\end{Def}
\begin{rem}
In particular, the fixed points of two independent axial elements form four distinct points of the visual boundary.
\end{rem}
Let us end this section by stating two results about rank one isometries. The first one was proven by P-E.~Caprace and K.~Fujiwara in \cite{caprace_fujiwara10}.
\begin{prop}[{\cite[Proposition 3.4]{caprace_fujiwara10}}]\label{non elem caprace fuj}
Let $X$ be a proper $\cat (0) $ space and let $G < \iso (X)$. Assume that $G$ contains a rank one element. Then exactly one of the following assertions holds:
\begin{enumerate}
\item \label{alt elem} $G$ either fixes a point in $\partial_\infty X$ or stabilizes a geodesic line. In both cases, it possesses a subgroup of index at most 2 of infinite Abelianization. Furthermore, if $X$ has a cocompact isometry group, then $\overline{G} < \iso (X)$ is amenable.
\item \label{alt non elem} G contains two independent rank one elements. In particular, $\overline{G}$ contains a discrete non-Abelian free subgroup.
\end{enumerate}
\end{prop}
As a consequence, the action $G \curvearrowright X$ of a rank one group $G$ on a $\cat (0) $ space $X$ is non-elementary if and only if alternative \ref{alt non elem} of the previous Proposition holds.
\newline
Rank one isometries are especially interesting because they induce natural contracting properties on the space. These properties mimic how loxodromic isometries behave in the hyperbolic setting.
\begin{Def}
A geodesic $\sigma$ in a $\cat(0)$ space is said to be \textit{$C$-contracting} with $C >0$ if for every metric ball $B$ disjoint from $\sigma$, the projection $\pi_\sigma (B)$ of the ball $B$ onto $\sigma$ has diameter at most $C$. An axial isometry is contracting if there exists $C>0$ such that one of its axes is $C$-contracting.
\end{Def}
It is clear that a contracting isometry is rank one. It turns out that the converse is true if $X$ is a proper $\cat(0)$ space, as was shown by M. Bestvina and K. Fujiwara in \cite{bestvina_fujiwara09}. This result will allow us to use the hyperbolic models described in Section \ref{hyperbolic models}.
\begin{thm}[{\cite[Theorem 5.4]{bestvina_fujiwara09}}]\label{bestv fuj rank one contracting}
Let $X$ be a proper $\cat (0)$ space, $g : X \rightarrow X$ be an axial isometry and $\sigma$ be an axis of $g$. Then there exists $B$ such that $\sigma$ is $B$-contracting if and only if $\sigma$ does not bound a half-flat. In other words, $g$ is contracting if and only if $g$ is a rank one isometry.
\end{thm}
\subsection{Gromov products}
Let $(X,d) $ be a metric space. One defines the Gromov product of $x, y \in X$ with respect to $o \in X$ as
\begin{eqnarray}
(x |y )_o = \frac{1}{2}(d(x, o) + d(y, o) - d(x,y)). \nonumber
\end{eqnarray}
The quantity $(x |y )_o$ must be thought of as representing the distance between $o$ and the geodesic between $x$ and $y$. This notion is particularly interesting because it does not require $X$ to be actually geodesic, and in fact we often deal with only quasigeodesic spaces. Also, we can use Gromov products to characterize hyperbolic spaces. We recall that a metric space $(X,d) $ is hyperbolic if there is $\delta >0$ such that for all $x, y, z \in X$,
\begin{eqnarray}
(x|z)_o \geq \min((x|y)_o, (y|z)_o) - \delta \nonumber.
\end{eqnarray}
If the reader wants a detailed introduction to hyperbolic spaces, a standard reference is \cite{bridson_haefliger99}.
\newline
If $(X,d)$ is a proper $\cat(0) $ space, the Gromov product can be extended to the visual boundary $\partial_\infty X$ of $X$ by the following formulas: for $x, y \in \partial_\infty X$, $o, m \in X$,
\begin{eqnarray}
(m|x)_o & := &\frac{1}{2}(d(o, m) - b^o_x(m)); \nonumber \\
(x |y)_o & := &- \frac{1}{2}\, \underset{q \in X}{\inf}\, (b_x (q) + b_{y}(q)) \nonumber.
\end{eqnarray}
If there is a geodesic line $\gamma$ between $x$ and $x'$, then by triangular inequality this infimum is actually a minimum, and is obtained at any point on $\gamma$. As a consequence, if $(x_n) $ and $(y_n)$ are sequences converging to $x,y \in \partial_\infty X $ respectfully, then $(x |y)_o = \lim_{n, m}(x_n| y_m)_o$.
\section{Hyperbolic models for proper $\cat(0)$ spaces}\label{hyperbolic models}
The goal of this section is to briefly present some ideas of \cite{petyt_spriano_zalloum22}, in which the authors build a way of attaching a family of hyperbolic metric spaces $X_L = (X, d_L)_L$ to a proper $\cat(0)$ space. What is interesting about these spaces is that they convey much of the geometry of the original space, especially at infinity, and they behave very well under isometric actions. More specifically, rank one isometries will act on some well-chosen spaces as loxodromic isometries. This construction can be understood as the analogue (and generalization) of the curve graphs that exist in the context of $\cat(0)$ cube complexes, see \cite{hagen14} and \cite{genevois19}.
\begin{Def}
Let $X$ be a $\cat(0) $ space, and let $\gamma: I \rightarrow X$ be a geodesic. Let $\pi_\gamma$ be the projection onto the geodesic $\gamma$ characterized by Proposition \ref{projection prop}. Let $t \in I$ be such that $[t - \frac{1}{2} , t + \frac{1}{2}]$ belongs to $I$. Then the \textit{curtain} dual to $\gamma$ at $t$ is
\begin{eqnarray}
h = h_{\gamma, t} = \pi^{-1}_\gamma (\gamma ([t - \frac{1}{2} , t + \frac{1}{2}])). \nonumber
\end{eqnarray}
The \textit{pole} of $h_{\gamma, t}$ is $ \gamma ([t - \frac{1}{2} , t + \frac{1}{2}]) $. Borrowing from the vocabulary of hyperplanes, we will call $h^{-} = \pi^{-1}_\gamma (\gamma ((-\infty, t- \frac{1}{2})\cap I))$ and $h^{+} = \pi^{-1}_\gamma (\gamma ((t +\frac{1}{2}, + \infty )\cap I))$ the \textit{halfspaces} determined by $h$. Note that $\{h^{-} , h , h^{+}\}$ is a partition of $X$. If $A \subseteq h^{-}$ and $B \subseteq h^{+}$ are subsets of $X$, we say that $h$ \textit{separates} $A$ from $B$.
\end{Def}
We will often denote a curtain by the letter $h$, even though one must keep in mind that $h = h_{\gamma, t}$ is characterized by a given geodesic $\gamma : I \rightarrow X $ and a point $t \in I$ (which defines a unique pole $P \subseteq \gamma$). Sometimes, we may also write $h= h_{\gamma, P}$ to emphasize on the pole $P$.
\begin{rem}\label{remark curtains thick}
By Proposition \ref{projection prop}, it is immediate that a curtains are closed subsets of $X$, and that they are thick: if $h$ is a curtain, then $d(h^{-}, h^{+})= 1$.
\end{rem}
Curtains can fail to be convex: if $x, y \in h^{-}$, it may happen that there exists $z \in [x, y ]\cap h^{+}$. Nonetheless, we have a weaker notion of convexity that the authors call star convexity:
\begin{prop}[{\cite[Lemma 2.6]{papasoglu_swenson09}}]\label{star convexity}
Let $h$ be a curtain dual to $\gamma$ and $P\subseteq \gamma$ be its pole. For every $x \in h$, then $[x, \pi_P (x)] \subseteq h$.
\end{prop}
\begin{Def}
A family of curtains $\{h_i\}$ is said to be a chain if $h_i $ separates $h_{i-1}$ from $h_{i+1}$ for every $i $. Chains can be used in order to define a metric on $X$ by the following: for $x \neq y \in X$,
\begin{eqnarray}
d_\infty (x, y) = 1 + \max \{\, |c| \, : \, c \text{ is a chain separating }x \text{ from } y \}. \nonumber
\end{eqnarray}
\end{Def}
One can check that this definition gives a metric. Let $h$ be a curtain, we have seen that $d(h^{-}, h^{+}) = 1$, hence for any $x,y \in X$, $d_\infty (x, y ) \leq \lceil d(x,y ) \rceil$. Conversely, it turns out that $d$ and $d_\infty$ may differ by at most 1.
\begin{lem}[{\cite[Lemma 2.10]{petyt_spriano_zalloum22}}]
Let $x, y \in X$. Then there is a chain of curtains $c$ dual to $[x,y]$ that realizes $d_\infty (x, y) = 1 + |c| $.
\end{lem}
We are now ready to refine the notion of separation in order to capture only some of the hyperbolic features of the space.
\begin{Def}[$L$-separation]
Let $L \in \mathbb{N}^\ast$, we say that disjoint curtains are $L$\textit{-separated} if every chain meeting both has cardinality at most $L$. A chain of pairwise $L$-separated curtains is called an $L$\textit{-chain}.
\end{Def}
The following geometric Lemma is a key ingredient for the proof that of Theorem \ref{theorem hyperbolic models}, and will be used several times in the sequel. It means that $L$-separation induces good Morse properties. The picture one has to keep in mind is given by Figure \ref{bottleneck figure}.
\begin{lem}[{\cite[Lemma 2.14]{petyt_spriano_zalloum22}}]\label{bottleneck}
Suppose that $A$, $B$ are two sets which are separated by an L-chain $\{h_1, h_2, h_3\}$ all of whose elements are dual to a geodesic $\gamma= [x_1, y_1]$ with $x_1 \in A$ and $y_1 \in B$. Then for any $x_2 \in A$, $y_2 \in B$, if $p \in h_2 \cap [x_2, y_2]$, then $d(p, \pi_\gamma(p))\leq 2L + 1$.
\end{lem}
\begin{figure}
\centering
\begin{center}
\begin{tikzpicture}[scale=1.2]
\draw (0,0) -- (4,0) ;
\draw (2.8,2) -- (2.8,-1) ;
\draw (1.2,2) -- (1.2,-1) ;
\draw (2,2) -- (2,-1) ;
\draw (2,0.5) node[right]{$\leq 2L+1$} ;
\draw (-1, 2.5) to[bend right = 80] (5, 3);
\draw (0,0) node[below left]{$x_1$} ;
\draw (2, -1) node[below]{$h_2$} ;
\draw (1.2, -1) node[below]{$h_1$} ;
\draw (2.8, -1) node[below]{$h_3$} ;
\draw (-1, 2.5) node[above left]{$x_2$} ;
\draw (4,0) node[below right]{$y_1$} ;
\draw (5,3) node[above]{$y_2$};
\filldraw[black] (2,0) circle(1pt);
\filldraw[black] (2,1) circle(1pt);
\end{tikzpicture}
\end{center}
\caption{Illustration of Lemma \ref{bottleneck}.}\label{bottleneck figure}
\end{figure}
The next Lemma states that if there is a $L$-chain separating two points $x $ and $y$, we can find a (smaller) $L$-chain of curtains separating those, which is dual to the geodesic $[x,y]$ and whose size can be controlled. It will prove useful later on, especially when we want to use Lemma \ref{bottleneck}.
\begin{lem}[{\cite[Lemma 2.22]{petyt_spriano_zalloum22}}]\label{dual chain}
Let $L, n \in \mathbb{N}$, and let $\{h_1, \dots, h_{(4L +10)n}\}$ be an $L$-chain separating $A$, $B \subseteq X$. Take $x \in A$, $y \in B$. Then $A$ and $B$ are separated by an $L$-chain of size $\geq n+1 $ dual to $[x,y]$.
\end{lem}
We are now ready to define a family of metrics using $L$-separation.
\begin{Def}
Given distinct points $x\neq y \in X$, we define
\begin{eqnarray}
d_L(x,y) = 1 + \max\{|c| \, : \, c \text{ is an } L\text{-chain separating }x \text{ from } y\}. \nonumber
\end{eqnarray}
\end{Def}
It turns out that for every $L$, $d_L$ gives a metric on $X$ \cite[Lemma 2.17]{petyt_spriano_zalloum22}. We will denote by $X_L = (X, d_L)$ the resulting metric space. With this definition in hand, Petyt, Spriano and Zalloum prove the that the metric spaces $(X, d_L)$ are hyperbolic.
\begin{thm}[{\cite[Theorem 3.1]{petyt_spriano_zalloum22}}] \label{theorem hyperbolic models}
For any $\cat(0)$ hyperbolic space $X$ and any integer $L$, the space $(X, d_L)$ is a quasi-geodesic hyperbolic space with hyperbolicity constants depending only on $L$. Moreover, $\iso(X)$ acts by isometries on $(X, d_L)$.
\end{thm}
We will then call $(X, d_L)$ a hyperbolic model for the $\cat(0)$ space $X$. Another useful fact about these spaces is that they behave well under isometries with "hyperbolic-like" properties.
\begin{thm}[{\cite[Theorem 4.9]{petyt_spriano_zalloum22}}] \label{rank one loxodromic}
Let $g$ be a semisimple isometry of $X$. The following are equivalent:
\begin{enumerate}
\item $g$ is a contracting isometry of the $\cat(0)$ space $X$;
\item there exists $L \in \mathbb{N}$ such that $g$ acts loxodromically on $X_L$.
\end{enumerate}
\end{thm}
Another piece of information brought by this construction is the relation between the Gromov boundaries $\partial X_L$ of the hyperbolic models $X_L = (X, d_L)$ and the visual boundary of the original $\cat(0)$ space $(X, d)$.
\begin{Def}
We say that a geodesic ray $\gamma : [0, \infty) \rightarrow X$ \textit{crosses} a curtain $h $ if there exists $t_0 \in [0, \infty)$ such that $h$ separates $\gamma(0) $ from $\gamma ([t_0, \infty))$. Alternatively, we may say that $h$ separates $\gamma(0)$ from $\gamma(\infty)$. Similarly, we say that a geodesic line $\gamma: \mathbb{R} \rightarrow X$ crosses a hyperplane $h$ if there exists $t_1, t_2 \in \mathbb{R}$ such that $h$ separates $\gamma ((-\infty, t_1])$ from $\gamma ([t_2, \infty))$. We say that $\gamma $ crosses a chain $c = \{h_i\}$ if it crosses each individual curtain $h_i$.
\end{Def}
As a consequence of Lemma \ref{bottleneck} and Lemma \ref{dual chain}, if two geodesic rays with the same starting point cross an infinite $L$-chain $c$, then they are asymptotic, and hence equal.
\begin{rem}\label{infinite dual chain}
Since curtains are not convex, it is not obvious that any geodesic ray $\gamma$ meeting a given curtain $h$ must cross it ($\gamma$ could meet $h$ infinitely often). However, by \cite[Corollary 3.2]{petyt_spriano_zalloum22} if $\gamma$ is a geodesic ray that meets every element of an infinite $L$-chain $c = \{h_i\}_{i \in \mathbb{N}}$, then $\gamma$ must cross $c$: for every $i$, there exists $t_i \in [0, \infty)$ such that $h_i$ separates $\gamma(0) $ from $\gamma ([t_0, \infty))$.
\end{rem}
Given $o \in X$, we define $\mathcal{B}_L $ as the subspace of $\partial_\infty X$ consisting of all geodesic rays $\gamma : [0, \infty) \rightarrow X$ starting from $o$ and such that there exists an infinite $L$-chain crossed by $\gamma$. In the case of the contact graph associated to a $\cat(0)$ cube complex $X$, we had the existence of an $\iso(X)$-equivariant embedding of the boundary of the contact graph into the Roller boundary $\partial_{\mathcal{R}}X$. The following result is the analogue in the context of $\cat(0)$ spaces.
\begin{thm}[{\cite[Theorem 7.1]{petyt_spriano_zalloum22}}] \label{equivariant embedding of boundaries}
Let $X$ be a proper $\cat(0)$ space. Then, for every $L \in \mathbb{N}^\ast$, the identity map $\iota : X \longrightarrow X_L$ induces an $\iso(X)$-equivariant homeomorphism $\partial_L : \mathcal{B}_L \longrightarrow \partial X_L$.
\end{thm}
\begin{Def}
We say that the action by isometries of a group $G$ on a hyperbolic space $Y$ (not assumed to be proper) is \textit{non-elementary} if there are two loxodromic isometries with disjoint fixed points on the Gromov boundary. A probability measure $\mu$ on $G$ is said to be \textit{non-elementary} if its support generates a group acting non-elementarily on $Y$.
\end{Def}
In order to use the results concerning random walks in hyperbolic spaces, we must show that the action of a group $G$ on a proper $\cat(0)$ space with rank one isometries induces a non-elementary action on some hyperbolic model $(X, d_L)$.
\begin{prop}\label{prop non elem}
Let $G$ be a group acting non-elementarily by isometries on a proper $\cat(0)$ space $(X,d)$, and assume that $G$ possesses a rank one element for this action. Then there exists $L \in \mathbb{N}$ such that $G$ acts on the hyperbolic space $(X, d_L)$ non-elementarily by isometries.
\end{prop}
\begin{proof}
The action $G \curvearrowright (X,d)$ is non elementary and contains a rank one element, hence by Theorem \ref{non elem caprace fuj} there exist two independent rank one isometries $ g,h$ in $G$. By Theorem \ref{bestv fuj rank one contracting}, those rank one isometries are $B$-contracting for some $B$. Now, applying Theorem \ref{rank one loxodromic}, there exists $L \in \mathbb{N}$ such that $g$ and $h$ act on $(X, d_L)$ as loxodromic isometries. As $g $ and $H$ are independent, their fixed points form four distinct points of the visual boundary $\partial_\infty X$. Now seen in $X_L = (X, d_L)$, their fixed points sets must also form four distinct points of $\partial X_L$ because of the homeomorphism $\partial_L : \mathcal{B}_L \longrightarrow \partial X_L$. This means that the action $G \curvearrowright X_L$ is non-elementary.
\end{proof}
\section{Random walks and hyperbolicity}\label{results hyp}
The results of Section \ref{hyperbolic models} allow us to read some information about the random walk in the hyperbolic models $X_L = (X, d_L)$, and then translate this information back to the original $\cat(0)$ space. As the theory of random walks on hyperbolic spaces is well-studied, one may hope that this process is fruitful.
\subsection{Random walks on hyperbolic spaces}
In this section, we summarize what is known concerning random walks in hyperbolic spaces. Most of the work for the non-proper case was done by Maher and Tiozzo in \cite{maher_tiozzo18}. The first result is the convergence of the random walk to the Gromov boundary.
\begin{thm}[{\cite[Theorem 1.1]{maher_tiozzo18}}] \label{mt conv}
Let $G$ be a countable group of isometries of a separable hyperbolic space $Y$. Let $\mu$ be a non-elementary probability distribution on $G$, and $o \in Y$ a basepoint. Then the random walk $(Z_n (\omega)o )_n $ induced by $\mu$ converges to a point $z^{+} (\omega)\in \partial_\infty X$, and the resulting hitting measure is the unique $\mu$-stationary measure on $\partial_\infty X$.
\end{thm}
Now assume that the measure $\mu $ has finite first moment $\int d(go, o) d\mu(g) < \infty$. Let us define the drift (or escape rate) of the random walk.
\begin{Def}
The \textit{drift} of the random walk $(Z_n o)_n$ on a hyperbolic space $(Y, d)$ is defined as
\begin{eqnarray}
\lambda: = \inf_n \int_\Omega d(Z_n(\omega) o, o)d\mathbb{P}(\omega) = \inf_n \int_G d(g o, o)d\mu^{\ast n}(g). \nonumber
\end{eqnarray}
\end{Def}
By a classical application of Kingmann subadditive Theorem, if $\mu $ has finite first moment the drift can also be defined as
\begin{eqnarray}
\lambda = \lim_n \frac{1}{n}d(Z_n (\omega)o , o),
\end{eqnarray}
and the above limit is essentially a constant (up to measure $0$, does not depend on $\omega$).
Maher and Tiozzo prove that in their context, the drift is almost surely positive. This can be seen as a law of large numbers.
\begin{thm}[{\cite[Theorem 1.2 and 1.3]{maher_tiozzo18}}]\label{mt drift}
Let $G$ be a countable group of isometries of a separable hyperbolic space $(Y,d_Y)$. Let $\mu$ be a non-elementary probability distribution on $G$, and $o \in Y$ a basepoint. Assume that $\mu$ has finite first moment. Then the drift $\lambda:= \lim_n \frac{1}{n}d(Z_n o, o )$ is well-defined and almost surely positive.
Moreover, for every $\lambda' < \lambda$, there exists $\kappa >0$ such that
\begin{eqnarray}
\mathbb{P}\big( \omega \in \Omega \, : \, d_Y(Z_n(\omega) o, o) \leq \lambda'n \big)< e^{-\kappa n }.
\end{eqnarray}
\end{thm}
Another piece of information that can be given about the random walk is the proportion of hyperbolic isometries in the random variables $(Z_n)_n$. Recall that the translation length of an isometry in a hyperbolic space is defined as $|g| := \lim_n \frac{1}{n} d(g^n o , o ) $, which does not depend on the basepoint $o$.
\begin{thm}[{\cite[Theorem 1.4]{maher_tiozzo18}}]\label{mt 4}
Let $G$ be a countable group of isometries of a separable hyperbolic space $Y$. Let $\mu$ be a non-elementary probability distribution on $G$, and $o \in Y$ is a basepoint. Then the translation length $|Z_n(\omega)| $ grows almost surely at least linearly in $ n $: there exists $K>0 $ such that
\begin{eqnarray}
\mathbb{P}(\omega \, : \, |Z_n(\omega)| \leq Kn) \underset{n \rightarrow \infty}{\longrightarrow} 0 \nonumber.
\end{eqnarray}
\end{thm}
The above result thus implies that the probability that $Z_n(\omega)$ is not a loxodromic isometry goes to zero as $n$ goes to infinity.
\subsection{First results for random walks in $\cat(0)$ spaces}
In $\cat(0) $ spaces, many of the previous theorems hold if we assume that $G$ induces "hyperbolic-like" properties. Namely, if $X$ is a proper $\cat(0)$ space, we will assume that $G$ contains rank one isometries of $X$. The first result deals with stationary measures on $\overline{X}$. Recall that a measure $\nu \in \text{Prob}(\overline{X})$ is called stationary if $\mu \ast \nu = \nu$.
\begin{thm}[{\cite[Theorem 1.1]{LeBars22}}]\label{measure thm cat0}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$, and assume that $G $ contains a rank one element. Then there exists a unique $\mu$-stationary measure $\nu \in \text{Prob} (\overline{X})$.
\end{thm}
The convergence of the random walk to the boundary can then be established in this setting. It is the analogue of Theorem \ref{mt conv}.
\begin{thm}[{\cite[Theorem 1.2]{LeBars22}}]\label{convergence thm}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$, and assume that $G $ contains a rank one element. Then for every $x \in X$, and for $\mathbb{P}$-almost every $\omega \in \Omega$, the random walk $(Z_n (\omega) x)_n $ converges almost surely to a boundary point $z^{+}(\omega) \in\partial_\infty X$. Moreover, $z^{+}(\omega)$ is distributed according to the stationary measure $\nu$.
\end{thm}
Interestingly, we can prove that the limit points are almost surely rank one, meaning that for almost any pair of limit points $\xi, \eta \in \partial_\infty X$, there exists a rank one geodesic in $X$ joining $\xi$ to $\eta$ (\cite[Corollary 1.3]{LeBars22}). This feature suggests the use of hyperbolic models. First, we establish a result concerning the proportion of rank one elements in the random walk.
\begin{thm}\label{thm prop rank one elements}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$, and assume that $G $ contains a rank one element. Then
\begin{eqnarray}
\mathbb{P}(\omega \, : \, Z_n (\omega) \text{ is a contracting isometry }) \underset{n\rightarrow \infty}{\rightarrow} 1 \nonumber.
\end{eqnarray}
\end{thm}
\begin{proof}
Because of Proposition \ref{prop non elem}, we can then apply the results of Maher and Tiozzo. In particular, by Theorem \ref{mt conv}, the random walk ${(Z_n o)_n}$ in $(X_L, d_L)$ converges almost surely to a point of the Gromov boundary $\partial X_L$, and by Theorem \ref{mt 4}, the translation length $|Z_n(\omega)|_L$ of $(Z_n(\omega))_n$ grows almost surely at least linearly in $n$. In particular, the probability that $Z_n(\omega)$ is a loxodromic element of $X_L$ goes to 1 as $ n $ goes to $\infty$. But thanks to Theorem \ref{rank one loxodromic}, an isometry $g$ of the $\cat(0)$ space $X$ is contracting if and only if there is an $L $ such that $g$ acts as a loxodromic isometry on $X_L$. The previous now implies that the probability that $Z_n(\omega)$ is a contracting isometry of $X$ goes to 1 as $n $ goes to $ \infty$.
\end{proof}
The analogue of Theorem \ref{mt drift} also holds in the context of $\cat(0)$ spaces with rank one isometries.
\begin{thm}[{\cite[Theorem 1.4]{LeBars22}\label{drift thm}}]
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite first moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Then the drift $\lambda$ is almost surely positive:
\begin{equation}
\lim_{n \rightarrow \infty} \frac{1}{n} d(Z_n o, o) = \lambda >0. \nonumber
\end{equation}
\end{thm}
Actually H. Izeki worked on the drift-free case in \cite{izeki22}. The author proves a strengthening of Theorem \ref{drift thm}, in that it is valid even for finite dimensional, non proper $\cat(0)$ spaces, and without the assumption that there are rank one isometries. The counterpart is that one needs to assume that $\mu$ has finite second moment. Namely, Izeki proves that in this context, either the drift $\lambda$ is strictly positive, or there is a $G$-invariant flat subspace in $X$ \cite[Theorem A]{izeki22}. However, for our purpose, we will only need Theorem \ref{drift thm}.
\newline
In the proof of Theorem \ref{drift thm}, we actually show that the displacement $d(Z_n(\omega) x, x)$ is almost surely well approximated by the Busemann functions $b_\xi (Z_n(\omega)x)$. This result will be used later when we give geometric estimates for the action.
\begin{prop}[{\cite[Proposition 5.2]{LeBars22}}]\label{approx displacement horo}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite first moment, and assume that $G $ contains a rank one element. Let $x \in X$ be a basepoint. Then for $\nu$-almost every $\xi \in \partial X$, and $\mathbb{P}$-almost every $\omega \in \Omega$, there exists $C > 0 $ such that for all $n \geq 0$ we have
\begin{equation}
|b_\xi (Z_n(\omega)x) - d(Z_n(\omega) x, x)| < C.
\end{equation}
\end{prop}
\section{Central Limit Theorems and general strategy} \label{strategy}
In order to prove our main result, we use a strategy that is largely inspired by the works of Benoist and Quint on linear spaces and hyperbolic spaces, see \cite{benoist_quint16} and \cite{benoist_quint16CLTlineargroups}. They developed a method for proving central limit theorems for cocycles, relying on results due to Brown in the case of martingales \cite{brown71}.
\subsection{Centerable cocycle}
Let $G$ be a discrete group, $Z$ a compact $G$-space and $c$ a cocycle $c : G \times Z \rightarrow \mathbb{R}$, meaning that $c(g_1g_2, x) = c(g_1, g_2 x ) + c(g_2, x)$, and assume that $c$ is continuous. Let $\mu$ be a probability measure on $G$.
\begin{Def}
Let $c$ be a continuous cocycle $c : G \times Z \rightarrow \mathbb{R}$. We say that $c$ has \textit{constant drift} $c_\mu$ if $c_\mu = \int_G c(g, x) d\mu(g)$ does not depend on $x \in Z$. We say that $c$ is \textit{centerable} if there exists a bounded measurable map $\psi : Z \rightarrow \mathbb{R} $ and a cocycle $c_{0} : G \times Z \rightarrow \mathbb{R}$ with constant drift $c_{0, \mu} = \int_G c_0(g, x) d\mu(g)$ such that
\begin{eqnarray}
c(g, x) = c_0(g,x) + \psi(x) - \psi(gx).
\end{eqnarray}
We say that $c$ and $c_0$ are cohomologous. In this case, the \textit{average} of $c$ is defined to be $c_{0, \mu}$.
\end{Def}
\begin{rem}\label{rem average cocycle}
Let $\nu \in \text{Prob}(Z)$ be a $\mu$-stationary measure, and let $c : G \times Z \rightarrow \mathbb{R}$ be a centerable continuous cocycle: for $g, \in G, x \in Z$, $c(g, x) = c_0(g,x) + \psi(x) - \psi(gx)$ with $c_0$ having constant drift and $\psi$ bounded measurable. Then
\begin{eqnarray}
& & \int_{G\times Z} c_0(g, x) d\mu(g) d\nu (x) + \int_Z \psi(x) d\nu(x) - \int_{G \times Z} \psi(gx) d\mu(g) d\nu(x) \nonumber \\
& = & \int_{G} c_0(g, x) d\mu(g) + \int_Z \psi(x) d\nu(x) - \int_{G \times Z} \psi(gx) d\mu(g) d\nu(x) \nonumber \\
& = & \int_{G} c_0(g, x) d\mu(g) + \int_Z \psi(x) d\nu(x) - \int_{ Z} \psi(x) d\nu(x) \text{ by stationarity} \nonumber \\
& = & c_{0, \mu} \text{ because $c_0$ has constant drift}. \nonumber
\end{eqnarray}
Hence the average of $c$ is given by $ c_{0, \mu} =\int c(g, x) d\mu(g) d\nu(x) $, which explains the terminology and shows that it does not depend on the choices of $c_0$ and $\psi$.
\end{rem}
The reason why we study limit laws on cocycles is the following result. This version is borrowed from Benoist and Quint, who improved previous results from Brown about central limit theorems for martingales \cite{brown71}.
\begin{thm}[{\cite[Theorem 3.4]{benoist_quint16CLTlineargroups}}]\label{criterium clt benoist quint}
Let $G$ be a locally compact group acting by homeomorphisms on a compact metrizable space $Z$. Let $c : G \times Z \rightarrow \mathbb{R}$ be a continuous cocycle such that $\int_G \sup_{x \in Z} | c(g,x)|^2 d\mu (g) < \infty $. Let $\mu $ be a Borel probability measure on $G$. Assume that $c$ is centerable with average $\lambda_c$ and that there exists a unique $\mu$-stationary probability measure $\nu $ on $Z$.
Then the random variables $\frac{1}{\sqrt{n}}(c(Z_n, x) - n \lambda_c) $ converge in law to a Gaussian law $N_\mu$. In other words,
for any bounded continuous function $F$ on $\mathbb{R}$, one has
\begin{eqnarray}
\int_G F \big( \frac{c(g,x) - n \lambda_c}{\sqrt{n}}\big) d(\mu^{\ast n})(g) \longrightarrow \int_{\mathbb{R}} F(t) dN_\mu (t) \nonumber.
\end{eqnarray}
\end{thm}
\subsection{Busemann cocycle and strategy}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite first moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Theorems \ref{convergence thm} and \ref{drift thm} ensure that the random walk $(Z_n(\omega)o)_n$ converges to a point of the boundary and that the drift $\lambda = \lim_n \frac{1}{n} d(Z_n(\omega) o , o)$ is well-defined and almost surely positive.
We denote by $\check{\mu}$ the probability measure on $G$ defined by $\check{\mu}(g) = \mu(g^{-1})$. Let $(\check{Z}_n)_n$ be the right random walk associated to $\check{\mu}$. Since $\mu $ is admissible and has finite first moment, so does $\check{\mu}$. We can then apply Theorems \ref{measure thm cat0}, \ref{convergence thm} and \ref{drift thm} to $\check{\mu}$. We will denote by $\check{\nu} $ the unique $\check{\mu}$-stationary measure on $\overline{X}$, and by $\check{\lambda} $ the positive drift of the random walk $(\check{Z}_n o)_n$.
\begin{rem}
One can check that
\begin{eqnarray}
\check{\lambda} &=& \inf_n \frac{1}{n} \int d(go, o) d\check{\mu}^{\ast n}(g) \nonumber \\
& = & \inf_n \frac{1}{n} \int d(o, g^{-1} o) d\check{\mu}^{\ast n}(g) \nonumber \\
& = & \inf_n \frac{1}{n} \int d(o, g o) d\mu^{\ast n}(g) \nonumber,
\end{eqnarray}
hence $\lambda = \check{\lambda}$.
\end{rem}
In our context, the continuous cocycle that we consider is the Busemann cocycle on the visual compactification of the $\cat(0)$ space $X$: for $x \in \overline{X}, \ g \in G$ and $o \in X$ a basepoint,
\begin{equation*}
\beta(g,x) = b_x (g^{-1} o).
\end{equation*}
It is straightforward to show that $\beta$ is continuous. Observe that for all $g_1, g_2 \in G, \, x \in Y$, horofunctions satisfy a cocycle relation:
\begin{eqnarray}
b_\xi (g_1g_2 o ) & = & \lim_{x_n \rightarrow \xi}d(g_1g_2 , x_n) - d(x_n, x) \nonumber \\
& = & \lim_{x_n \rightarrow \xi}d(g_2 , g_1^{-1}x_n) - d(g_1 o, x_n ) + d(g_1 o, x_n ) - d(x_n, o) \nonumber \\
& = & \lim_{x_n \rightarrow \xi}d(g_2x , g_1^{-1}x_n) - d(o, g_1^{-1}x_n ) + d(g_1 x, x_n ) - d(x_n, o) \nonumber \\
& = & b_{g_1^{-1}\xi} (g_2 o) + b_\xi (g_1 o). \label{cocycle horof}
\end{eqnarray}
By \eqref{cocycle horof}, $\beta$ satisfies the cocycle relation $\beta(g_1g_2, x) = \beta(g_1, g_2 x ) + \beta(g_2, x)$. Thanks to Proposition \ref{approx displacement horo}, for every $o \in X$, for $\nu$-almost every $x \in \partial X$, and $\mathbb{P}$-almost every $\omega \in \Omega$, there exists $C > 0 $ such that for all $n \geq 0$ we have
\begin{equation}\label{approx busemann displacement}
|\beta (Z_n(\omega)^{-1}, x) - d(Z_n(\omega) o,o)| < C.
\end{equation}
Equation \eqref{approx busemann displacement} shows that the cocycle $\beta(Z_n(\omega),x)$ "behaves" like $d(Z_n(\omega) o,o)$. Thus it makes sense to try and apply Theorem \ref{criterium clt benoist quint} to the Busemann cocycle $\beta(g,x)$.
\newline
Henceforth, we will assume that $\mu$ is an admissible probability measure on $G$ with finite second moment $\int_{G} d(go, o)^2 d\mu(g) < \infty$.
The following proposition summarizes some properties of the Busemann cocycle. It shows that obtaining a central limit theorem on $\beta$ will imply our main result.
\begin{prop}\label{Busemann cocyle prop}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Let $\lambda$ be the (positive) drift of the random walk, and $\beta : G \times \overline{X} \rightarrow \mathbb{R}$ be the Busemann cocycle $\beta(g, x) = b_x(g^{-1} o)$. Then
\begin{enumerate}
\item $\int_{G} \sup_{x \in \overline{X}}|\beta (g, x)|^2 d\mu(g) < \infty$ and $\int_{G} \sup_{x \in \overline{X}}|\beta (g, x)|^2 d\check{\mu}(g) < \infty$;\label{moment}
\item For $\nu$-almost every $\xi \in \partial_\infty X$, $\lambda = \lim_n \frac{1}{n} \beta (Z_n (\omega), \xi)$ $\mathbb{P}$-almost surely; \label{drift cocycle}
\item $\mathbb{P}$-almost surely, $\lambda = \int_{G \times \overline{X}} \beta(g, x) d\mu(g)d\nu(x) = \int_{G \times \overline{X}} \beta(g, x) d\check{\mu}(g)d\check{\nu}(x)$. \label{average Busemann}
\end{enumerate}
\end{prop}
\begin{proof}
As a consequence of Proposition \ref{approx displacement horo}, equation \eqref{approx busemann displacement} gives that for $\nu$-almost every $x \in \partial X$, and $\mathbb{P}$-almost every $\omega \in \Omega$, there exists $C > 0 $ such that for all $n \geq 0$ we have
\begin{eqnarray}
|\beta (Z_n(\omega)^{-1}, x) - d(Z_n(\omega) o,o)| < C. \nonumber
\end{eqnarray}
Because the action is isometric and $\mu$ has finite second moment $\int_{G} d(g o, o)^2 d\mu(g)~<~\infty$, we obtain
\begin{eqnarray}
\int_{G} \sup_{x \in \overline{X}}|\beta (g, x)|^2 d\mu(g)<\infty. \nonumber
\end{eqnarray}
With the same argument:
\begin{eqnarray}
\int_{G} \sup_{x \in \overline{X}}|\beta (g, x)|^2 d\check{\mu}(g) < \infty. \nonumber
\end{eqnarray}
Now thanks to Theorem \ref{drift thm}, the drift of the random walk $\lambda = \lim_n \frac{1}{n} d(Z_n(\omega) o , o)$ is almost surely positive. Together with equation \ref{approx busemann displacement}, this gives that for $\nu$-almost every $\xi \in \overline{X}$,
\begin{eqnarray}
\lambda = \lim_n \frac{1}{n} \beta (Z_n (\omega), \xi) \text{ $\mathbb{P}$-almost surely}.\nonumber
\end{eqnarray}
The ideas in the proof of \ref{average Busemann} are classical. We give the details for the convenience of the reader.
Let $T : (\Omega \times \overline{X}, \mathbb{P} \times \check{\nu} ) \rightarrow (\Omega \times \overline{X}, \mathbb{P} \times \check{\nu} )$ be defined by $T(\omega, \xi ) \mapsto (S\omega, \omega_0^{-1} \xi)$, with $S((\omega_i)_{i \in \mathbb{N}}) = (\omega_{i+1})_{i \in \mathbb{N}}$ the usual shift on $\Omega$. By \cite[Proposition 5.4]{LeBars22}, $T$ preserves the measure $\mathbb{P}\times \check{\nu}$ and is an ergodic transformation. Define $H : \Omega \times \overline{X} \rightarrow \mathbb{R}$ by
\begin{equation*}
H(\omega, \xi) = h_\xi (\omega_0 o) = \beta(\omega_0^{-1}, \xi).
\end{equation*}
By \ref{moment}, it is clear that $\int |H(\omega, \xi)| d\mathbb{P}(\omega)d\check{\nu}(\xi) < \infty$.
By cocycle relation \eqref{cocycle horof} one gets that
\begin{eqnarray}
h_\xi (Z_n o) = \sum_{k=1}^{n} h_{Z_k^{-1}\xi} (\omega_k o ) = \sum_{k=1}^{n} H(T^k (\omega, \xi)) \label{transient cocycle}.
\end{eqnarray}
Then $\beta(Z_n(\omega)^{-1}, \xi) = \sum_{k=1}^{n} H(T^k (\omega, \xi))$, and by \ref{drift cocycle},
\begin{eqnarray}
\lambda = \lim_n \frac{1}{n} \sum_{k=1}^{n} H(T^k (\omega, \xi)).
\end{eqnarray}
Now, by Birkhoff ergodic theorem, one obtains that almost surely,
\begin{eqnarray}
\lambda &=& \int_{\Omega \times \overline{X}} H(\omega, \xi) d\mathbb{P}(\omega)d\check{\nu}(x). \nonumber \\
& = & \int_{\Omega \times \overline{X}} h_\xi(\omega_0 o) d\mathbb{P}(\omega)d\check{\nu}(x) \nonumber \\
& = & \int_{G \times \overline{X}} \beta(g^{-1}, \xi) d\mu(g)d\check{\nu}(x) \nonumber \\
& =& \int_{G \times \overline{X}} \beta(g, \xi) d\check{\mu}(g)d\check{\nu}(x)
\end{eqnarray}
The previous computations can be done similarly for $\mu $ and $\nu $, hence we also have that
\begin{eqnarray}
\lambda = \int_{G \times \overline{X}} \beta(g, x) d\mu(g)d\nu(x). \nonumber
\end{eqnarray}
\end{proof}
In order to apply Theorem \ref{criterium clt benoist quint} on the Busemann cocycle $\beta$, it remains to show that $\beta$ is centerable. If this is the case, by \ref{average Busemann} and Remark \ref{rem average cocycle}, its average must be the positive drift $\lambda$. In other words, we need to show that there exists a bounded measurable function $\psi : \overline X \rightarrow \mathbb{R}$ such that the cohomological equation
\begin{eqnarray}
\beta(g, x) = \beta_0(g,x) + \psi(x) - \psi(gx). \label{cohom equation}
\end{eqnarray}
is verified. Then, proving the Central Limit Theorem in our context amounts to finding such a $\psi $ that is well defined and bounded. This will be done by using a hyperbolic model that can give nice estimates on the random walk.
\section{Proof of the Central Limit Theorem}\label{Section proof}
\subsection{Geometric estimates}
In this section, we prove our main Theorem, following the strategy explained in Section \ref{strategy}. First, we will provide geometric estimates on the random walk that will be used later on. This is where we use the specific contraction properties provided by the curtains and the hyperbolic models exposed in Section \ref{hyperbolic models}. The goal is ultimately to prove that the candidate $\psi$ for the cohomological equation is bounded.
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Recall that $B_L$ is defined to be the subspace of $\partial_\infty X$ consisting of all geodesic rays $\gamma : [0, \infty) \rightarrow X$ starting from $o$ and such that there exists an infinite $L$-chain crossed by $\gamma$. By Theorem \ref{equivariant embedding of boundaries}, there exists an $\iso(X)$-equivariant embedding $\mathcal{I} : \partial X_L \rightarrow \partial_\infty X$, whose image lies in $\mathcal{B}_L$.
\begin{prop}\label{geometric lemma}
Let $(g_n)$ be a sequence of isometries of $G$, and let $o \in X$, $x, y \in \partial_\infty X$. Assume that there exists $\lambda, \epsilon, A >0 $ such that:
\begin{enumerate}[label= (\roman*)]
\item $ \{g_n o \}_n $ converges in $(\overline{X_L}, d_L)$ to a point of the boundary $z_L \in \partial X_L$, whose image in $\partial_\infty X $ by the embedding $\mathcal{I}$ is not $y$; \label{assum cv}
\item $d_L(g_n o, o) \geq An$; \label{assum drift hyp}
\item $|b_x (g_n^{-1} o ) - n \lambda | \leq \varepsilon n$; \label{assum gnx}
\item $|b_y (g_n o ) - n \lambda | \leq \varepsilon n$; \label{assum y}
\item $|d(g_n o , o) - n \lambda | \leq \varepsilon n$. \label{assum drift cat}
\end{enumerate}
Then, one obtains:
\begin{enumerate}
\item $(g_nx |g_n o )_o \geq (\lambda - \varepsilon) n$; \label{geom estimate 1}
\item $(y | g_n o )_o \leq \varepsilon n$.\label{geom estimate 2}
\end{enumerate}
If moreover $A \geq 2(4L+10) \varepsilon$, then we have:
\begin{enumerate}[resume]
\item $(y| g_n x )_o \leq \varepsilon n + (2L+ 1)$. \label{geom estimate 3}
\end{enumerate}
\end{prop}
The proof of points \ref{geom estimate 1} and \ref{geom estimate 2} is straightforward, so we begin by these.
\begin{proof}[Proof of estimates \ref{geom estimate 1} and \ref{geom estimate 2}]
A simple computation gives that
\begin{eqnarray}
(g_n x | g_n o)_o = \frac{1}{2}(b_x (g^{-1}_n o) + d(g_no, o)) \nonumber
\end{eqnarray}
Then using assumptions \ref{assum gnx} and \ref{assum drift cat} gives immediately that $(g_nx | g_no ) \geq (\lambda - \epsilon ) n $, which proves \ref{geom estimate 1}.
Now, by definition,
\begin{eqnarray}
(y |g_n o ) = \frac{1}{2}(d(g_no, o) - b_y (g_no))\nonumber
\end{eqnarray}
Then by assumptions \ref{assum y} and \ref{assum drift cat}, we obtain \ref{geom estimate 2}.
\end{proof}
The proof of point \ref{geom estimate 3} is the hard part. We prove it in two steps. First, we show that under the assumptions, for $n$ large enough, there exist at least three $L$-separated curtains dual to $[o, g_no]$ separating $\{g_n o, g_nx\}$ on the one side and $\{o, y \}$ on the other, see Figure \ref{figure separating hyperplanes}. Then we show that the presence of these hyperplanes implies the result.
\begin{figure}
\centering
\begin{center}
\begin{tikzpicture}[scale=1.2]
\draw (0,0) -- (4,0) ;
\draw (2.8,2) -- (2.8,-1) ;
\draw (1.2,2) -- (1.2,-1) ;
\draw (2,2) -- (2,-1) ;
\draw (0,0) node[below left]{$o$} ;
\draw (4,0) [-stealth] to[bend left = 20] (5, 3);
\draw (0,0) [-stealth] to[bend right = 20] (-1, 2.5);
\draw (-1, 2.5) to[bend right = 55] (5,3);
\draw (-1, 2.5) node[above left]{$y$} ;
\draw (4,0) node[below right]{$g_no$} ;
\draw (5,3) node[above]{$g_n x$};
\end{tikzpicture}
\end{center}
\caption{A "hyperbolic-like" 4 points inequality in Proposition \ref{geometric lemma}.}\label{figure separating hyperplanes}
\end{figure}
By assumption \ref{assum drift hyp}, there exists an $L$-chain of size $S(n) $ that separates $o$ and $g_no$. By Proposition \ref{dual chain}, there exists an $L$-chain dual to $[o, g_no ]$ of size greater than or equal to $\lfloor\frac{S(n)}{4L + 1}\rfloor$ that separates $o$ and $g_no$. Denote by $c_n = \{h^n_i\}_{i=1}^{S'(n)}$ a maximal $L$-chain dual to $[o, g_no]$, separating $o $ and $g_n o $, and orient the half-spaces so that $o \in h_i ^{-}$ for all $i$. When the context is clear, we might omit the dependence in $n$ for ease of notations, and just write $\{h_i\}_{i=1}^{S'(n)}$ for a maximal $L$-chain dual to $[o, g_n o ]$. Recall that $S(n) \geq An$, hence $c_n$ must be of length $S'(n) \geq A'n$, where $A' = \frac{A}{4 L+1}$.
\begin{lem}\label{estimate y}
Under the assumptions of Proposition \ref{geometric lemma}, there exists a constant $C$ such that for all $n \in \mathbb{N}$, the number of $L$-separated hyperplanes in $c_n$ that do not separate $\{o, y\}$ and $\{g_n o\} $ is less than $C$.
\end{lem}
\begin{proof}[Proof of Lemma \ref{estimate y}]
By assumption, $ \{g_n o \}_n $ converges in $(\overline{X_L}, d_L)$ to a point of the boundary $z_L \in \partial X_L$. By Theorem \ref{equivariant embedding of boundaries}, there exists an $\iso(X)$-equivariant embedding $\mathcal{I} : \partial X_L \rightarrow \partial_\infty X$ that extends the canonical inclusion $X_L \rightarrow X$, and whose image lies in $\mathcal{B}_L$. Denote by $z := \mathcal{I}(z_L)$ the image in $\partial_\infty X$ of the limit point $z_L$ by this embedding.
Denote by $\beta : [0, \infty) \rightarrow X$ a geodesic ray joining $o$ to $z$. Since $z \in \mathcal{B}_L$, there exists an infinite $L$-chain $c = \{k_i\}_{i \in \mathbb{N}}$ that separate $o$ from $z$. Note that because of Lemma \ref{dual chain} and Remark \ref{infinite dual chain}, we can assume that $c$ is a chain of curtains which is dual to the geodesic ray $\beta$. Since $\{g_n o \}_n $ converges in $(\overline{X_L}, d_L)$ to $z_L \in \partial X_L$, and $z$ is the image of $z_L$ by the equivariant embedding $\mathcal{I}$, it implies that $\{g_no\}_n $ converges to $z$ in $X$. The fact that $z \in \mathcal{B}_L$ implies that for all $i \in \mathbb{N}$, there exists $n_0 \in \mathbb{N}$ such that for all $n\geq n_0$, $k_i$ separates $o $ from $g_n o $. Now, we denote by $\gamma : [0, \infty) \rightarrow X$ the geodesic ray that represents $y \in \partial_\infty X $. See figure \ref{figure estimate y}.
\begin{figure}
\centering
\begin{center}
\begin{tikzpicture}[scale=1]
\draw (0,0) [-stealth]-- (7,0) ;
\draw (1.2,4) -- (1.2,-1.5) ;
\draw (2.5,4) -- (2.5,-1.5) ;
\draw (3.5,4) -- (3.5,-1.5) ;
\draw (4.5,4) -- (4.5,-1.5) ;
\draw (0,0) node[below left]{$o$} ;
\draw (2, 3) node[left]{$\gamma$} ;
\draw (7,0) node[below right]{$z$} ;
\draw (2, 4.5) node[above left]{$y$} ;
\draw (2.5, 1) node[above right]{$k_p$} ;
\draw (3.5, 1) node[above right]{$k_{p+1}$} ;
\draw (3.5, 0) node[above right]{$r$} ;
\draw (4.5, 1) node[above right]{$k_{p+2}$} ;
\draw (5.5,-1) node[below right]{$g_no$} ;
\draw (6,0) node[above right]{$\beta$} ;
\filldraw[black] (0,0) circle (1.5pt) ;
\draw (3.3,-0.6) node[below left]{$r'$} ;
\filldraw[black] (3.5,0) circle (1.5pt) ;
\filldraw[black] (3.3,-0.6) circle (1.5pt) ;
\filldraw[black] (5.5,-1) circle (1.5pt) ;
\draw (0, 0 ) [-stealth] to[bend right] (2, 4.5);
\draw (0, 0 ) -- (5.5,-1);
\end{tikzpicture}
\end{center}
\caption{Illustration of Lemma \ref{estimate y}.}\label{figure estimate y}
\end{figure}
Due to Remark \ref{infinite dual chain}, meeting $c$ infinitely often is equivalent to crossing it, then since $y \neq z$, there exists $p \in \mathbb{N} $ such that $\gamma \subseteq k_p^{-}$. Now consider $n_0$ such that for $n\geq n_0$, $g_n o \in k_{p+2}^{+}$. Fix $n \geq n_0$. Recall that $c_n$ is a maximal $L$-chain dual to $[o, g_n o]$ separating $o$ and $g_n o $.
Denote by $r \in \beta$ a point in the pole of $k_{p+1}$, and denote by $r'= r'(n)$ the projection of $r$ onto the geodesic $[o, g_n o ]$. Then by Lemma \ref{bottleneck},
\begin{equation*}
d(o, r'(n)) \leq d(o , r) + 2L +1 \nonumber.
\end{equation*}
Due to the thickness of the curtains (Remark \ref{remark curtains thick}), the number of curtains in $c_n$ that separate $o $ and $r' (n) $ is $\leq d(o , r) + 2L +2$. We emphasize that this number does not depend on $n\geq n_0$, because for all $n \geq n_0$, $g_n \in k_{p+2}^{+}$ and the previous equation holds.
Recall that $\gamma \subseteq k_{p}^{-}$, so in particular $\gamma \subseteq k_{p+1}^{-}$. Then by star convexity of the curtains (Lemma \ref{star convexity}), every curtain in $c_n$ whose pole belongs to $[r'(n), g_n o]$ separates $\{o,y\}$ from $g_n o $. Then by the previous argument, the number of curtains that do not separate $\{o, y\}$ from $\{g_no\}$ is less than $d(o , r'(n))$. In particular, the number of curtains that do not separate $\{o, y\}$ from $\{g_no\}$ is less than $d(o, r) + 2L+2 $. Since this quantity does not depend on $n$, we have proven the Lemma.
\end{proof}
Now, for a fixed $n$, let us give an estimate for the number of curtains in $ c_n = \{h^n_1 , \dots , h^n_{S'(n)}\} $ that separate $o $ and $g_n x $. When a given $n$ is fixed, we omit the dependence in $n$ and just write $c_n = \{h_1 , \dots , h_{S'(n)}\}$ to ease the notations. Let $\gamma_n : [0, \infty) \rightarrow X$ be the geodesic ray joining $o $ and $g_n x$. Let us take $k_0= k_0(n)$ (depending on $n$) large enough so that for all $k \geq k_0$,
\begin{equation}
| (g_n o | g_n x )_o - (g_no | \gamma_n(k)_o) | \leq 1. \nonumber
\end{equation}
\begin{lem}\label{estimate gnx}
Under the assumptions of Proposition \ref{geometric lemma}, the number of $L$-separated hyperplanes in $c_n$ that separate $\{o\}$ and $\{g_n o, \gamma_n(k_0)\} $ is unbounded in $n$. More precisely, for all $M \in \mathbb{N}$, there exists $n_0$ such that for all $n \geq n_0$, the number of $L$-separated hyperplanes in $c_n$ that separate $\{o\}$ and $\{g_n o, \gamma_n(k)\} $ is greater than $M$ for all $k \geq k_0$.
\end{lem}
\begin{proof}[Proof of Lemma \ref{estimate gnx}]
Let $k \geq k_0$. Suppose that the number of curtains in $ c_n = \{h_1 , \dots , h_{S'(n)}\} $ separating $o $ and $\gamma_n(k) $ is less than or equal to $p \in [0, S'(n)-4]$. Then $\{h_{p+2} , \dots , h_{S'(n)}\} $ is an $L$-chain separating $\{o, \gamma_n(k)\} $ and $\{g_n o\} $. We then denote by $r(n)$ a point on $h_{p+3} \cap [\gamma_n(k), g_no]$ and by $r'(n)$ the projection of $r(n)$ onto $[o, g_n o]$, see Figure \ref{figure estimate gnx}. By hypothesis on $k$,
\begin{eqnarray}
2((g_nx | g_n o)_o - 1) &\leq& 2(\gamma_n(k)| g_no)_o \nonumber \\
&=& d(\gamma_n(k) , o ) + d(g_n o , o ) - d(g_n o, \gamma_n (k) ). \nonumber
\end{eqnarray}
Now by the bottleneck Lemma \ref{bottleneck} and the triangular inequality,
\begin{eqnarray}
2(\gamma_n(k)| g_no)_o & = & d(\gamma_n(k) , o ) + d(g_n o , o ) - (d(g_n o, r(n) ) + d(r(n) , \gamma_n(k))) \nonumber \\
&\leq & d(\gamma_n(k) , o ) + d(g_n o , o ) - (d(g_n o, r' (n)) - (2L + 1) + d(r (n), \gamma_n(k))) \nonumber \\
&\leq & d(r (n) , o ) + d(g_n o , o ) - d(g_n o, r'(n) ) + 2L + 1 \nonumber \\
&\leq & d(r' (n) , o ) + 2L + 1 + d(g_n o , o ) - d(g_n o, r'(n) ) + 2L + 1 \nonumber \\
& \leq & 2 d(r' (n), o ) + 2(2L + 1). \nonumber
\end{eqnarray}
Because the pole of a curtain is of diameter 1, $d(o , r'(n)) \leq d(g_n o, o) - (S'(n) - (p +1))$. However, by assumptions \ref{assum drift hyp} and \ref{assum drift cat} of Lemma \ref{geometric lemma}, one gets that $d(g_no , o ) \leq (\lambda + \varepsilon) n $ and $S(n) \geq An $. Recall that by Lemma \ref{dual chain}, this means that $S'(n) \geq A'n$, where $A' = \frac{A}{4L +1}$. Combining this with the previous result yields
\begin{eqnarray}
& & (g_nx | g_n o)_o - 1 \leq d(o, r'(n)) + 2L+1 \nonumber \\
& \Rightarrow& (\lambda - \varepsilon) n - 1 \leq (\lambda + \varepsilon )n - (A'n - (p+1)) + 2L+1 \text{ by Lemma \ref{geometric lemma}, \ref{geom estimate 1}}\nonumber \\
& \Rightarrow & 0 \leq (2\varepsilon - A' ) n + 2L + p+2. \nonumber
\end{eqnarray}
If $A' > 2 \varepsilon$, there exists $n_0 $ large enough such that for all $n \geq n_0$, the above inequality gives a contradiction. As a consequence, if $A' > 2 \varepsilon$, or equivalently if $ A > 2(4L +1) \varepsilon$, there exists $n_0$ such that for all $n \geq n_0$, the number of curtains in $c_n$ separating ${o } $ and $\{\gamma_n (k), g_no\}$ is greater than $p$.
\end{proof}
\begin{figure}
\centering
\begin{center}
\begin{tikzpicture}[scale=2]
\draw (0,0) -- (4,0) ;
\draw (3.6,2) -- (3.6,-0.5) ;
\draw (2.5,2) -- (2.5,-0.5) ;
\draw (3,2) -- (3,-0.5) ;
\draw (1.5,2) -- (1.5,-0.5) ;
\draw (0,0) node[below left]{$o$} ;
\draw (2.2, 2.5) to[bend right = 40] (4,0);
\draw (0,0) [-stealth]to[bend right = 20] (2.3,2.8);
\draw (2.2, 2.5) node[below left]{$\gamma_n(k)$} ;
\draw (2.3, 2.8) node[above right]{$g_n x$} ;
\draw (3,0.5) node[above right]{$r(n)$} ;
\filldraw[black] (2.2, 2.5) circle (1 pt) ;
\filldraw[black] (3, 0.5) circle (1 pt) ;
\filldraw[black] (0,0) circle (1 pt) ;
\filldraw[black] (2.85,0) circle (1 pt) ;
\filldraw[black] (4,0) circle (1 pt) ;
\draw (3,-0.5) node[below ]{$h_{p+3}$} ;
\draw (1.5,-0.5) node[below ]{$h_p$} ;
\draw (2.85,0) node[above]{$r'(n)$} ;
\draw (2.5,-0.5) node[below ]{$h_{p+2}$} ;
\draw (4,0) node[below right]{$g_no$} ;
\end{tikzpicture}
\end{center}
\caption{Illustration of Lemma \ref{estimate gnx}.}\label{figure estimate gnx}
\end{figure}
We can now conclude the proof of Lemma \ref{geometric lemma}.
\begin{proof}[Proof of estimate \ref{geom estimate 3}]
Recall that we denote by $\gamma : [0, \infty) \rightarrow X$ the geodesic ray that represents $y \in \partial_\infty X $ such that $\gamma(0) = o$ and by $\gamma_n : [0, \infty) \rightarrow X$ the geodesic ray joining $o $ and $g_n x$. Combining Lemma \ref{estimate y} and Lemma \ref{estimate gnx}, we get that if $A > 2(4L+10)\varepsilon$, there exists $n_0$, $k_0$ such that for all $n \geq n_0$ and all $k\geq k_0$, $c_n$ contains at least 3 pairwise $L$-separated curtains that separate $\{o, \gamma(k)\} $ on the one side and $\{g_no, \gamma_n(k)\}$ on the other. Call these hyperplanes $\{h_1, h_2, h_3\}$ and arrange the order so that $h_i \subseteq h_{i+1}^{-}$. Denote by $m_k (n) \in h_2$ some point on the geodesic segment joining $\gamma(k) $ to $\gamma_n(k)$, and $m'_k(n)$ belonging to the geodesic segment $[o, g_no]$ such that $d(m_k(n), m'_k(n)) \leq 2 L + 1$, see Figure \ref{figure proof geom estimate}
\begin{figure}
\centering
\begin{center}
\begin{tikzpicture}[scale=1]
\draw (0,0) -- (8,0) ;
\draw (2.5,4) -- (2.5,-0.5) ;
\draw (4,4) -- (4,-0.5) ;
\draw (5.5, 4) -- (5.5,-0.5) ;
\draw (0,0) node[below left]{$o$} ;
\draw (-0.3, 2) node[left]{$\gamma$} ;
\draw (0, 0) to[bend right = 20](-1, 4) ;
\draw (0, 0) to[bend right = 10] (9,3) ;
\draw (-1, 4) to[bend right = 30] (9,3);
\draw (8,0) to[bend left = 50] (9,3);
\draw (-1, 4) node[above left]{$\gamma(k)$} ;
\draw (2.5, -0.5) node[below]{$h_1$} ;
\draw (4, -0.5) node[below]{$h_2$} ;
\draw (5.5, -0.5) node[below]{$h_3$} ;
\draw (8,0) node[below right]{$g_no$} ;
\draw (9,3) node[above right]{$\gamma_n(k)$};
\draw (4,2) node[above right]{$m_k(n)$};
\draw (4,0) node[above right]{$m'_k(n)$};
\draw (7,2) node[below]{$\gamma_n$};
\filldraw[black] (-1, 4) circle (1 pt) ;
\filldraw[black] (4,0) circle (1 pt) ;
\filldraw[black] (4, 2.0) circle (1 pt) ;
\filldraw[black] (9,3) circle (1 pt) ;
\filldraw[black] (8,0) circle (1 pt) ;
\filldraw[black] (0,0) circle (1 pt) ;
\end{tikzpicture}
\end{center}
\caption{Proof of Proposition \ref{geometric lemma}}\label{figure proof geom estimate}
\end{figure}
Then we have
\begin{eqnarray}
2(\gamma(k)| \gamma_n(k))_o & = & d(\gamma(k) , o ) + d(o, \gamma_n(k)) - d(\gamma(k), \gamma_n(k)) \nonumber\\
& \leq & d(\gamma(k) , o ) + d(o, m'_k(n)) + d(m'_k(n), m_k(n) ) \nonumber \\
& & + \, d(m_k (n), \gamma_n(k)) - d(\gamma(k), \gamma_n(k)) \text{ by the triangular inequality }\nonumber\\
& \leq & d(\gamma(k) , o ) + d(o, m'_k(n)) - d(\gamma(k), m_k(n)) + 2L + 1 \nonumber
\end{eqnarray}
by Lemma \ref{bottleneck}. Since $m'_k$ is on $[o , g_no]$, $d(o , m'_k)= d(o, g_no ) - d(g_no , m'_k(n))$.
We then have:
\begin{eqnarray}
2(\gamma(k)| \gamma_n(k))_o & \leq & d(\gamma(k) , o ) + d(o, g_no ) - d(g_no , m'k(n)) - d(\gamma(k), m_k(n)) + 2L + 1 \nonumber \\
& = & d(\gamma(k) , o ) + d(o, g_no ) - (d(g_no , m'_k(n)) + d(\gamma(k), m_k(n))) + 2L + 1. \nonumber
\end{eqnarray}
Now observe that
\begin{eqnarray}
d(\gamma(k), g_no) &\leq& d(g_no , m'_k(n)) + d(\gamma(k), m_k(n)) + d(m_k, m'_k) \nonumber \\
& \leq & d(g_no , m'_k(n)) + d(\gamma(k), m_k(n)) + 2L + 1 \text{ by Lemma \ref{bottleneck}}, \nonumber
\end{eqnarray}
hence $d(\gamma(k), g_n o ) - (2 L + 1) \leq d(g_no , m'_k(n)) + d(\gamma(k), m_k(n))$. Then
\begin{eqnarray}
2(\gamma(k)| \gamma_n(k))_o & \leq & d(\gamma(k) , o ) + d(o, g_no ) - (d(\gamma(k), g_n o ) - (2 L + 1)) + 2L + 1 \nonumber \\
& = & d(\gamma(k) , o ) + d(o, g_no ) - d(\gamma(k), g_n o ) + 2(2 L + 1) \nonumber \\
& = & 2(\gamma(k)| g_n o )_o + 2(2 L + 1) \nonumber.
\end{eqnarray}
As $k \rightarrow \infty$, one obtains that $(g_nx| y)_o \leq (g_no |y)_o + (2 L + 1)$, and the result follows from \ref{estimate 2}.
\end{proof}
\subsection{Proof of the Central Limit Theorem}
In this section, we prove the main result of the paper. Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Let $\lambda$ be the (positive) drift of the random walk provided by Theorem \ref{drift thm}.
We assume the action on $X$ to be non elementary and rank one, hence due to Proposition \ref{prop non elem}, there exists a number $L \geq 0$ such that $G$ acts by isometries on $X_L = (X, d_L) $ non elementarily. Then one can consider the random walk $(Z_n(\omega) o)_n$ as a random walk on $(X, d_L) $, which we will write $(Z_n \tilde{o})_n $ when the context is not clear. The model $(X, d_L) $ is hyperbolic, so we can apply the results of Maher and Tiozzo \cite{maher_tiozzo18} summarized in Section \ref{results hyp}. In particular, due to Theorem \ref{mt conv}, the random walk $(Z_n \tilde{o})_n$ in $X_L$ converges to a point of the Gromov boundary $\partial X_L$ of $(X, d_L)$.
Moreover, since we assume $\mu $ to have finite first moment (for the action on the $\cat(0)$ space $X$), and since $d(x, y ) \geq d_L(x, y )$ for all $x, y \in X$, the measure $\mu $ is also of finite first moment for the action on the hyperbolic model $(X, d_L)$. In particular, the drift $\tilde{\lambda}$ of the random walk $(Z_n \tilde{o})_n$ is almost surely positive. In other words, we have that $\mathbb{P}$-almost surely,
\begin{equation}
\lim_{n \rightarrow \infty} \frac{1}{n} d(Z_n(\omega) \tilde{o}, \tilde{o}) = \tilde{\lambda} >0. \nonumber
\end{equation}
Due to Theorem \ref{measure thm cat0}, there exists a unique $\mu$-stationary probability measure $\nu$ on $\overline{X}$. If we define $\check{\mu} \in \text{Prob}(G)$ by $\check{\mu} (g) = \mu(g^{-1})$, $\check{\mu} $ is still admissible and of finite second moment. We denote by $\check{\nu}$ the unique $\check{\mu}$-stationary measure on $\overline{X}$.
We recall that the Busemann cocycle $\beta : G \times \overline{X} \rightarrow \mathbb{R}$ is defined by:
\begin{equation*}
\beta(g,x) = b_x (g^{-1} o).
\end{equation*}
Our goal is to apply Theorem \ref{criterium clt benoist quint} to the Busemann cocycle $\beta$. The results of Section \ref{strategy} show that proving a central limit theorem for the random walk $(Z_n(\omega)o)_n$ amounts to proving that $\beta$ is centerable. As in the works of \cite{benoist_quint16}, \cite{horbez18} and \cite{fernos_lecureux_matheus21}, the natural candidate to solving the cohomological equation \eqref{cohom equation} is the function:
\begin{equation*}
\psi ( x ) = -2 \int_{\overline{X}} (x | y)_o d\check{\nu}(y).
\end{equation*}
\begin{prop}\label{prop centerable cocycle}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Then the Borel map
$\psi(x) = \int_{\overline{X}} (x |y)_o \, d\check{\nu}(y)$ is well-defined and essentially bounded. Thus, the cocycle $\beta(g, x) = h_x (g^{-1} o )$ is centerable.
\end{prop}
In order to show that $\psi$ is well-defined and bounded, we need the following statement, which resembles \cite[Proposition 4.2]{benoist_quint16}.
\begin{prop}\label{estimates}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint for the random walk $(Z_n(\omega)o)_n$. Let $\lambda$ be the (positive) drift of the random walk, and $\nu$ a $\mu$-stationary measure on $\overline{X}$. Assume that there exists $a >0 $ and $(C_n)_n \in l^1 (\mathbb{N})$ such that for almost every $x, y \in \overline{X}$, we have, for every $n$:
\begin{enumerate}
\item $\mathbb{P}((Z_n o | Z_n x )_o \leq an) \leq C_n $; \label{estimate 1}
\item $\mathbb{P}((Z_n o | y )_o \geq an) \leq C_n $; \label{estimate 2}
\item $\mathbb{P}((Z_n x | y )_o \geq an) \leq C_n $. \label{estimate 3}
\end{enumerate}
Then one has:
\begin{eqnarray}
\sup_{x \in \overline{X}} \int_{\overline{X}} (x |y)_o d\nu(y) < \infty \nonumber.
\end{eqnarray}
\end{prop}
\begin{proof}
Suppose that there exist $a>0$, $(C_n)_n \in l^1 (\mathbb{N})$ such that for almost every $x, y \in \overline{X}$, we have estimates \ref{estimate 1} to \ref{estimate 3}. We get:
\begin{eqnarray}
\nu(\{x \in X | (x|y) \geq an \}) & = & \int_{\overline{X}} \mu^{\ast n }(\{g\in G \, | \, (gx |y )_o \geq an \})d\nu(x) \text{ by $\mu$-stationarity} \nonumber \\
& \leq & \int_{\overline{X}} C_n d\nu(x) = C_n \text{ by estimate \ref{estimate 3}}.\nonumber
\end{eqnarray}
Then, define $A_{n, y } := \{ x \in \overline{X} \, | \, (x|y )_o \geq an\}$, so that by splitting along the subsets $A_{n-1, y } - A_{n, y }$, one gets
\begin{eqnarray}
\int_{\overline{X}} (x |y)_o d\nu(x) & \leq & \sum_{n \geq 1} a n (\nu(A_{n-1, y }) - \nu(A_{n, y })) \nonumber \\
& \leq & \sum_{n \geq 1} an (C_{n-1} - C_n) \nonumber \\
& = & a + \sum_{n \geq 1} aC_n (n+1 - n) < \infty. \nonumber
\end{eqnarray}
\end{proof}
We want to show that estimates from Proposition \ref{estimates} hold. As we will see, estimates \ref{estimate 1} are quite straightforward to check using the positivity of the drift. Most of the work concerns estimate \ref{estimate 3}.
Combining Proposition \ref{Busemann cocyle prop} with Theorem \ref{drift thm} and \cite[Proposition 3.2]{benoist_quint16CLTlineargroups}, one obtains the following:
\begin{prop}\label{proba control}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Let $\lambda$ be the (positive) drift of the random walk. Then, for every $\varepsilon >0 $, there exists $(C_n)_n \in l^1 (\mathbb{N})$ such that for any $x \in \overline{X}$,
\begin{eqnarray}
& & \mathbb{P}(|\beta(Z_n, x) - n \lambda | \geq \varepsilon n) \leq C_n;\label{prob estimate 1}\\
& & \mathbb{P}(|\beta(Z^{-1}_n, x) - n \lambda | \geq \varepsilon n) \leq C_n ; \label{prob estimate 2} \\
& & \mathbb{P}(|d(Z_n o, o) - n \lambda | \geq \varepsilon n) \leq C_n. \label{prob estimate 3}
\end{eqnarray}
\end{prop}
\begin{proof}
Recall that by Proposition \ref{Busemann cocyle prop}, $\beta $ is a continuous cocycle such that
\begin{eqnarray}
& & \int_G \sup_{x \in \overline{X}} |\beta(g,x)|^2 d\mu(g) < \infty \text{ and }
\int_G \sup_{x \in \overline{X}} |\beta(g,x)|^2 d\check{\mu}(g) < \infty. \nonumber
\end{eqnarray}
Moreover,
\begin{eqnarray}
\lambda = \int_{G \times \overline{X}} \beta(g, x) d\mu(g)d\nu(x)= \int_{G \times \overline{X}} \beta(g, x) d\check{\mu}(g)d\check{\nu}(x). \nonumber
\end{eqnarray}
We can then apply \cite[Proposition 3.2]{benoist_quint16CLTlineargroups}: for every $\varepsilon >0 $, there exists a sequence $(C_n) \in l^1 (\mathbb{N})$ such that for every $x \in \overline{X}$,
\begin{eqnarray}
\mathbb{P}\big(\omega \in \Omega \, : \, \big|\frac{\beta(Z_n(\omega), x)}{n} - \lambda\big| \geq \epsilon\big) \leq C_n \nonumber.
\end{eqnarray}
The same goes for $\check{\mu} $ and $\check{\nu}$, which gives estimates \eqref{prob estimate 1} and \eqref{prob estimate 2}.
Estimate \eqref{prob estimate 3} is then a straightforward consequence of Proposition \ref{approx displacement horo}.
\end{proof}
The following Lemma will also be important in the proof of Proposition \ref{prop centerable cocycle}.
\begin{lem}\label{Lemma drift hyp}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Let $\lambda$ be the (positive) drift of the random walk.
Then there exists $L >0$, $\lambda_L > 0$ such that almost surely, $\liminf_n \frac{d_L (Z_n o, o)}{n} = \lambda_L$. Moreover, there exists $A > 0$ and $(C_n) \in l^1 (\mathbb{N}) $ such that
\begin{eqnarray}
\mathbb{P}\big( d_L (Z_n o, o) < An\big) \leq C_n \nonumber.
\end{eqnarray}
\end{lem}
\begin{proof}
The action $G \curvearrowright (X,d)$ is non elementary and contains a rank one element, hence by Proposition \ref{prop non elem}, there exists $L$ such that the action $G \curvearrowright (X, d_L)$ is non-elementary as the loxodromic isometries $g$ and $h$ are independent. We can then apply Theorem \ref{mt drift}, which gives the Lemma.
\end{proof}
Let us now complete the proof of Proposition \ref{prop centerable cocycle}.
\begin{proof}[Proof of Proposition \ref{prop centerable cocycle}]
By assumptions, we can apply Theorem \ref{mt conv}: there exists $L >0 $ such that $(Z_n(\omega) o )_n$ converges in $(X_L, d_L)$ to a point $z_L$ of the boundary. By Theorem \ref{measure thm cat0}, there is a unique $\mu$-stationary measure $ \nu$ on $\partial_\infty X$, and this measure is non-atomic.
Fix $A$ as in Lemma \ref{Lemma drift hyp}, and $(C_n)_n \in l^1 (\mathbb{N})$ such that
\begin{eqnarray}
\mathbb{P}\big( d_L (Z_n o, o) < An\big) < C_n \nonumber.
\end{eqnarray}
Now take $0 < \varepsilon < \min(\frac{A}{2(4L+10)}, \lambda/2)$. Due to Proposition \ref{proba control}, there exists a sequence $C'_n \in l^1( \mathbb{N})$ such that
\begin{eqnarray}
& & \mathbb{P}(|\beta(Z_n, x) - n \lambda | \geq \varepsilon n) \leq C'_n\nonumber\\
& & \mathbb{P}(|\beta(Z^{-1}_n, x) - n \lambda | \geq \varepsilon n) \leq C'_n \nonumber \\
& & \mathbb{P}(|d(Z_n o, o) - n \lambda | \geq \varepsilon n) \leq C'_n. \nonumber
\end{eqnarray}
We can assume that $C_n = C'_n $ for all $n$. Then for $\nu$-almost every $x, y \in \partial_\infty X$, we have the quantitative assumptions in Proposition \ref{geometric lemma}: with
\begin{enumerate}[label= (\roman*)]
\item $ \{Z_no \}_n $ converges in $(\overline{X_L}, d_L)$ to a point of the boundary $z_L \in \partial X_L$, whose image in $\partial_\infty X $ by the embedding $\mathcal{I}$ is not $y$;
\item $\mathbb{P}\big( d_L(Z_n o, o) \geq An \big) \geq 1 - C_n$;
\item $\mathbb{P}\big(|b_x (Z_n^{-1} o ) - n \lambda | \leq \varepsilon n\big) \geq 1-C_n$;
\item $\mathbb{P}\big(|b_y (Z_n o ) - n \lambda | \leq \varepsilon n\big) \geq 1- C_n$;
\item $\mathbb{P}\big(|d(gZ_n o , o) - n \lambda | \leq \varepsilon n\big) \geq 1- C_n$.
\end{enumerate}
As a consequence, one obtains that for $\nu$-almost every $x, y \in \partial_\infty X$, the probability that these estimates are not satisfied is bounded above by $4 C_n$. Now choosing $a \in (\varepsilon, \lambda - \varepsilon)$, we get that for $n$ large enough,
\begin{enumerate}
\item $\mathbb{P}\big((g_nx |g_n o )_o \geq a n\big) \geq 1 - 4C_n$;
\item $\mathbb{P}\big((y | g_n o )_o \leq a n\big)\geq 1 - 4C_n$;
\item $\mathbb{P}\big((y| g_n x )_o \leq a n\big) \geq 1 - 4 C_n$.
\end{enumerate}
Since the sequence $(4C_n)_n$ is still summable, we can apply Proposition \ref{estimates}. Since $\psi $ is measurable by Fubini, the proof is complete.
\end{proof}
We can now state our main theorem.
\begin{thm}
Let $G$ be a discrete group and $G \curvearrowright X$ a non-elementary action by isometries on a proper $\cat (0)$ space $X$. Let $\mu \in \text{Prob}(G) $ be an admissible probability measure on $G$ with finite second moment, and assume that $G $ contains a rank one element. Let $o \in X$ be a basepoint of the random walk. Let $\lambda$ be the (positive) drift of the random walk. Then the random variables $(\frac{1}{\sqrt{n}}(d(Z_n o, o) - n \lambda))_n $ converge in law to a Gaussian distribution $N_\mu$.
\end{thm}
\begin{proof}
By Proposition \ref{prop centerable cocycle}, the cocycle $\beta$ is centerable, with average $\lambda$. Since the measure $\nu$ is the unique $\mu$-stationary measure on $\overline{X}$, we can then apply Theorem \ref{criterium clt benoist quint}: the random variables $(\frac{1}{\sqrt{n}}\beta(Z_n(\omega) - n \lambda))_n$ converge to a Gaussian law $N_\mu $. But thanks to Proposition \ref{approx displacement horo}, this is equivalent to the convergence of the random variables $(\frac{1}{\sqrt{n}}d(Z_n(\omega)o, o) - n \lambda)_n$ to a Gaussian law. This proves the theorem.
\end{proof}
\bibliographystyle{alpha}
|
\section{Introduction}
Models of multidimensional data -- panel data with more than two dimensions -- are fast becoming popular in econometric analysis as large data sets with a multidimensional structure become available.
For example, in gravity models of trade that are repeated over time one may be interested in studying trade patterns between an importer, $i$, an exporter, $j$ that is repeated every quarter or year, $t$.
One may also be interested in studying demand elasticities through consumption data that may vary by product, $i$, store, $j$ with repeated observation over week or month, $t$.\footnote{A non-exhaustive list of related examples can be found in the introduction of \cite{matyas2017econometrics} in trade, housing and prices, migration, country productivity and consumer price setting. }
In these examples it is clear that there may exist unobserved characteristics in each dimension that can determine variation across both the dependent and independent variables that needs to be controlled for to avoid issues with endogeneity.
For example, this could be shifts in taste preferences, that are unobserved by the econometrician, that may effect sales of particular products in certain stores differently over time.
Thus far most analysis has addressed unobserved heterogeneity in the higher-dimensional setting by using a combination of additive scalar fixed-effects.
These additive scalar fixed-effects approaches, however, can only accommodate variation in unobserved heterogeneity over a subset of dimensions with any one of the scalar fixed-effects terms.
For example, in the three-dimensional model this type of fixed-effect approach can only control for variation over $ij$, $it$ and $jt$, but not over all $ijt$.
In the face of more complicated relationships that admit multiplicative variation across dimensions, these additive effects are unsatisfactory to control for unobserved heterogeneity.
This paper develops tools to control for unobserved heterogeneity in the form of interactive fixed-effects in models of multidimensional panel data.
The main body of the paper focuses on the linear and additively separable model.
More generic applications of these tools are discussed in the introduction but are not formally studied.
To fix ideas consider linear parameter estimation in the following interactive fixed-effects model with three dimensions,
\begin{align}\label{eqn:IntroModel}
Y_{ijt} = X_{ijt}^\prime\beta + \sum_{\ell = 1}^L \varphi^{(1)}_{i\ell} \varphi^{(2)}_{j\ell} \varphi^{(3)}_{t\ell} +\varepsilon_{ijt},
\end{align}
where all terms in $\sum_{\ell = 1}^L \varphi^{(1)}_{i\ell} \varphi^{(2)}_{j\ell} \varphi^{(3)}_{t\ell}$ are unobserved and $L$, which is defined later, can be interpreted as a rank parameter.
Reducing the problem to three dimensions is without loss of generality for the methods considered herein.
Additive fixed-effects are omitted for brevity but can be removed with a simple within transformation.
Let $X_{ijt}$ be arbitrarily correlated with the unobserved interactive fixed-effects term, $\sum_{\ell = 1}^L \varphi^{(1)}_{i\ell} \varphi^{(2)}_{j\ell} \varphi^{(3)}_{t\ell} $, but uncorrelated with the noise term, $\varepsilon_{ijt}$.
The challenge to estimating $\beta$ is isolating variation in $X_{ijt}$ that is not correlated with the interactive fixed-effects term.
This paper develops the multidimensional group fixed-effects and kernel weighted transformations to project out this unobserved heterogeneity and also shows settings were standard factor methods work well.
The group fixed-effects method used below is similar to \cite{bonhomme2021discretizing} and the within-cluster transformation in \cite{freeman2022linear}.
The model for interactive fixed-effects has precedent in the standard two-dimensional panel data setting. For instance take the model considered in \cite{Bai2009} and similar to \cite{Pesaran2006},
\begin{align}\label{eqn:bai}
Y_{it} = X_{it}^\prime\beta + \sum_{\ell = 1}^L\lambda_{i\ell} f_{t\ell} + e_{it}.
\end{align}
In that setting, \cite{Bai2009} show that the interactive term $\sum_{\ell = 1}^L\lambda_{i\ell} f_{t\ell}$ also sufficiently captures variation in additive individual and time effects without the need to specify these separately, so these are again naturally omitted.
For multidimensional applications
it may be preferable to simply transform the problem in \eqref{eqn:IntroModel} to a two dimensional problem and estimate \eqref{eqn:bai} directly using the transformed data.
However, and as will be explained in further detail in Section~\ref{sect:estimation}, problems persist when $L$ is large and only a subset of the unobserved heterogeneity parameters are low-dimensional.
For consistent estimation of $\beta$, transforming the multidimensional array to a matrix then estimating \eqref{eqn:bai} requires either: (a) all fixed-effects are low-dimensional, or; (b) that a subset of the fixed-effects are low-dimensional and the analyst knows which ones are.
The requirement that the analyst has this knowledge can be highly restrictive.
Alternatively, the within-cluster and kernel weighted transformations analysed in this paper requires only that a subset of the fixed-effect parameters are low-dimensional, though the analyst does not need to know which of the fixed-effect parameters make up this subset.
Under sufficient regularity conditions, the methods considered in this paper may also control for variation from arbitrary functions of the fixed-effects.
Similar to that considered in \cite{zeleneev2020identification} and \cite{freeman2022linear}, the functional representation of model \eqref{eqn:IntroModel} could be,
\begin{align*}
Y_{ijt} = X_{ijt}^\prime \beta + h( \varphi^{(1)}_{i}, \varphi^{(2)}_{j}, \varphi^{(3)}_{t}) + \varepsilon_{ijt},
\end{align*}
for vector-valued $\varphi^{(1)}_{i}$, $\varphi^{(2)}_{j}$ and $\varphi^{(3)}_{t}$. The set of fixed-effects to be transformed by the function $h(\cdot,\cdot,\cdot)$ could also extend to fixed-effects over multiple indices, e.g. $\alpha_{ij}$ from above.
It should be noted that the setting considered in \cite{zeleneev2020identification} requires that the transformation is non-smooth, and it is not trivial to see that a ``within-type'' transformation will sufficiently project this type of heterogeneity.
With sufficient smoothness conditions on the function transforming the fixed-effects, existing literature could be generalised to show consistency using the proposed within-cluster transformation in the multidimensional case.
Models with discrete explanatory variables
\citep{chernozhukov2013average, hoderlein2012nonparametric,evdokimov2010identification,fernandez2021low},
provide another interesting application of these group fixed-effects estimators. Take the following regression line for discrete valued $X_{ijt}$,
\begin{align*}
Y_{ijt} = h\left(X_{ijt}, \varphi^{(1)}_{i}, \varphi^{(2)}_{j}, \varphi^{(3)}_{t}, \varepsilon_{ijt}\right).
\end{align*}
Then, under sufficient smoothness conditions on the function $h$, the unobserved heterogeneity may also be projected out with a group fixed-effect estimator.
Tensor completion techniques also have useful generalisations in this setting, for example,
\citet{tomioka2010estimation,li2019statistical, xu2020tensor},
for some examples of methods that consider sparse multidimensional arrays.
The sparse multidimensional array problem has similar complexities to the low-rank tensor approximation problem in that they do not extend from the matrix problems in a straightforward way, hence require non-trivial extensions.
It is also important to consider unobserved heterogeneity in applications that admit discrete dependent variables. For example, for binary response variable with known $F(\cdot)$,
\begin{align}\label{eqn:nonlinearDependent}
Y_{ijt} = F\big(X_{ijt}^\prime\beta + \sum_{\ell = 1}^L \varphi^{(1)}_{i\ell} \varphi^{(2)}_{j\ell} \varphi^{(3)}_{t\ell}\big).
\end{align}
Estimation of the unobserved heterogeneity term may then be performed with a similar iterative scheme as that proposed in \cite{chen2021nonlinear} or the sufficient statistic approach in Chapter 6 of \cite{matyas2017econometrics}.
The incidental parameter problem in this setting can be alleviated using methods in this paper by allowing cluster sizes to grow with data size coupled with taking grouped fixed-effects along fewer dimensions, for example in \cite{bonhomme2021discretizing} and also Appendix~\ref{sect:reduceParameters}.
\cite{menzel2017bootstrap} consider a special case of multidimensional data for bootstrapping methods where the data is $D$-adic.
That is, each dimension of the data refers to the same set of observations, like a network graph where each index refers to an individual in the network.
An example of the multidimensional version of this could be a binary indicator of a three step path, $Y_{ijk} = G_{ij}G_{jk}$, detailing if there exists a path from $i$ to $k$.
In any case, the type of multidimensional data considered in that work is a distinct special case of the type of data structures considered in this paper.
The technical component of this paper is highly related to the numerical analysis literature on low-rank approximations of multidimensional arrays.
As pointed out in \cite{de2008tensor}, the optimisation problem of finding low-rank approximations in the tensor setting is not well-posed, hence most results in this literature rely on numerical evidence.
See \cite{kolda2009tensor} for a summary of the multidimensional array decomposition problem and \citet{vannieuwenhoven2012new, rabanser2017introduction} for examples of numerical results.
As such, it is necessary to innovate on this tensor low-rank problem to find appropriate analytical results.
To this end, this paper utilises well-posed components of the numerical analysis literature for use in nuisance parameter applications.
These applications have the advantage that they do not require the multidimensional array of fixed-effects to be reconstructed, hence do not attempt to directly solve the low-rank tensor problem.
It is worth a note that \cite{elden2011perturbation}, along with related papers, suggest a reformulation of the low multilinear rank problem that may have promising applications in econometrics, but this is left for future research.
The paper is organised as follows.
Section~\ref{sect:motivation} introduces the model and motivates the within-cluster transformation,
Section~\ref{sect:ExistLit} discusses the existing literature on matrix methods and how these extend to the multidimensional setting,
Section~\ref{sect:tensordecomp} discusses estimation with group fixed-effects and kernel methods, which is the main methodological contribution of the paper,
Section~\ref{sect:asymResults} discusses the asymptotic performance of the estimators,
Section~\ref{sect:sims} displays the simulation results and
Section~\ref{sect:conclusion} concludes.
\section{Motivation}\label{sect:motivation}
The model in full dimensional generality is,
\footnote{
For example, in index notation this model can be written as,
\begin{align*}
Y_{i_1, i_2, \dots, i_d} = \sum_{k = 1}^K X_{i_1, i_2, \dots, i_d; k}\beta_k
+ \mathcal{A}_{i_1, i_2, \dots, i_d}
+ \mathcal{B}_{i_1, i_2, \dots, i_d}
+ \varepsilon_{i_1, i_2, \dots, i_d}
\end{align*}
with $\mathcal{A}_{i_1, i_2, \dots, i_d} = \sum_{\ell = 1}^L \varphi^{(1)}_{i_1\ell} \dots \varphi^{(d)}_{i_d\ell}$ and the additive fixed-effects term could be for example,
$\mathcal{B}_{i_1, i_2, \dots, i_d} = \gamma^{(1)}_{i_2, \dots, i_{d}}
+ \gamma^{(2)}_{i_1,i_3, \dots, i_{d}}
+ \gamma^{(3)}_{i_1,i_2, i_4, \dots, i_{d}}
+ \dots
+ \gamma^{(d)}_{i_1, \dots, i_{d-1}}.
$
}
\begin{align}\label{eqn:fullModel}
\boldsymbol{Y} = \sum_{k = 1}^K \boldsymbol{X}_k \beta_k
+ \boldsymbol{\mathcal{A}}
+ \boldsymbol{\mathcal{B}}
+ \boldsymbol{\varepsilon},
\end{align}
where $\boldsymbol{Y}, \boldsymbol{X}_k, \boldsymbol{\varepsilon} \in\mathbb{R}^{N_1\times N_2\times\dots \times N_d}$.
$\boldsymbol{\mathcal{A}} = \sum_{\ell = 1}^L \varphi^{(1)}_\ell \circ \dots\circ \varphi^{(d)}_\ell$ where $\varphi^{(n)}_\ell\in\mathbb{R}^{N_n}$ for each $n = 1, \dots, d$ and ``$\circ$'' is the outer product.
$L$ is naturally restricted to have upper bound $\min_n \{\prod_{n^\prime\neq n}N_{n^\prime}\}$, see \cite{kruskal1989rank}.
$\boldsymbol{\mathcal{B}}$ is the sum of additive fixed-effects terms that can each vary over $d-1$ indices.
For example, this could be, $\mathcal{B}_{i_1, i_2, \dots, i_d} = \gamma^{(1)}_{i_2, \dots, i_{d}}
+ \gamma^{(2)}_{i_1,i_3, \dots, i_{d}}
+ \gamma^{(3)}_{i_1,i_2, i_4, \dots, i_{d}}
+ \dots
+ \gamma^{(d)}_{i_1, \dots, i_{d-1}}.$
However, each of these scalar terms need not vary over $d-1$ dimensions and may vary over $d-2$ or $d-3$ dimensions, and so forth.
$\boldsymbol{\varepsilon}$ is a noise term uncorrelated with all $\boldsymbol{X}_k$ and all unobserved fixed-effects terms.
Take $i_n\in\{1,\dots,N_n\}$ for all $n\in\{1,\dots,d\}$ as the dimension specific index.
The regressors $\boldsymbol{X}_k$ may be arbitrarily correlated with $\boldsymbol{\mathcal{A}}$ and $\boldsymbol{\mathcal{B}}$.
$\boldsymbol{\mathcal{B}}$ cannot be ignored without loss of generality for arbitrary estimators hence is included for completeness. The main estimation procedure considered below is robust to $\boldsymbol{\mathcal{B}}$ being included or excluded, so it may be ignored in that context.
Model \eqref{eqn:fullModel} can be seen as a natural extension of the \cite{Bai2009} model to three (or more) dimensions with the $\boldsymbol{\mathcal{A}}$ term interpreted as a ``higher-dimensional'' factor stucture.
Similar to this strain of the literature, all terms in $\boldsymbol{\mathcal{A}}$ and $\boldsymbol{\mathcal{B}}$ are considered fixed nuisance parameters.
There are potentially many extensions to the factor model setting in \cite{Bai2009} to the higher dimension case.
This paper starts with what seems the most natural extension.
\subsection{Within-cluster transformation}\label{sect:withinClTrans}
Ignore the additive fixed-effects and consider the model,
\begin{align*}
{\mathbf Y} = \sum_{k=1}^K \beta_k \mathbf{X}_k + \boldsymbol{\mathcal{A}} + \pmb{\varepsilon}.
\end{align*}
Allowing for arbitrary correlation between $\mathbf{X}_k$ and $\boldsymbol{\mathcal{A}}$ for some or all $k$ introduces an endogeneity problem unless estimation methods can account for this.
For example, $\mathbf{X}_k$ may be generated as,
\begin{align*}
\mathbf{X}_k = h_k(\boldsymbol{\mathcal{A}}) + \pmb{\nu}_k.
\end{align*}
For the below motivation it is useful to reconsider \eqref{eqn:fullModel} with just three-dimensions again.
A popular method to project fixed-effects is the within transformation on both $\mathbf{Y}$ and $\mathbf{X}$ as follows,
\begin{align}\label{withinEstimator}
\widetilde{{Y}}_{ijt} = {Y}_{ijt} - \Bar{{Y}}_{\cdot jt} - \Bar{{Y}}_{i\cdot t} - \Bar{{Y}}_{ij\cdot} + \Bar{{Y}}_{\cdot\cdot t} + \Bar{{Y}}_{\cdot j\cdot}+ \Bar{{Y}}_{i\cdot \cdot} - \Bar{{Y}}_{\cdot \cdot \cdot}
\end{align}
where the variables with bars simply denote the average taken over the ``dotted'' index for the entire sample.
That is,
$\Bar{{Y}}_{\cdot jt} := \frac{1}{N_1}\sum_{i = 1}^{N_1} Y_{ijt}$,
$\Bar{{Y}}_{\cdot\cdot t} := \frac{1}{N_1N_2}\sum_{i = 1}^{N_1}\sum_{j = 1}^{N_2} Y_{ijt}$, etc.
Note, this is not a unique representation of the within transformation, and can be achieved with many different transformations.
In the presence of interactive fixed-effects this transformation leaves the following fixed-effects residual,
\begin{align*}
\widetilde{\mathcal{A}}_{ijt} =\sum_{\ell= 1}^L (\varphi^{(1)}_{i\ell} - \Bar{\varphi}^{(1)}_{\ell}) (\varphi^{(2)}_{j\ell} - \Bar{\varphi}^{(2)}_{\ell}) (\varphi^{(3)}_{t\ell} - \Bar{\varphi}^{(3)}_{\ell}),
\end{align*}
where
$\Bar{\varphi}^{(1)}_{\ell} := \frac{1}{N_1}\sum_{i = 1}^{N_1} \varphi^{(1)}_{i\ell}$,
$\Bar{\varphi}^{(2)}_{\ell} := \frac{1}{N_2}\sum_{j = 1}^{N_2} \varphi^{(2)}_{j\ell}$, and
$\Bar{\varphi}^{(3)}_{\ell} := \frac{1}{N_3}\sum_{t = 1}^{N_3} \varphi^{(3)}_{t\ell}$.
Sufficient projection of the unobserved heterogeneity term, $\widetilde{\mathcal{A}}_{ijt}$, then relies on either $\varphi^{(1)}_{i\ell} \rightarrow \Bar{\varphi}^{(1)}_{\ell}$, $\varphi^{(2)}_{j\ell} \rightarrow \Bar{\varphi}^{(2)}_{\ell}$ or $\varphi^{(3)}_{t\ell} \rightarrow \Bar{\varphi}^{(3)}_{\ell}$ for each $\ell$, which is highly restrictive and actually implies the simple additive fixed-effect model is the true model.
Consider as an alternate the within-cluster transformation.
Take for now as given some grouping of units within each dimension.
That is, there exist some allocations of all $i$'s into groups, all $j$'s into groups and all $t$'s into groups.
How these groups are constructed is formally defined later.
With a slight abuse of notation in how each $\Bar{\mathcal{A}}$ is defined, the within-cluster transformation is,
\begin{align}\label{withinGroupEstimator}
\begin{split}
\widetilde{\mathcal{A}}_{ijt} &= \mathcal{A}_{ijt} - \Bar{\mathcal{A}}_{i^*jt} - \Bar{\mathcal{A}}_{ij^* t} - \Bar{\mathcal{A}}_{ijt^*} + \Bar{\mathcal{A}}_{i^*j^* t} + \Bar{\mathcal{A}}_{i^* jt^*}+ \Bar{\mathcal{A}}_{ij^* t^*} - \Bar{\mathcal{A}}_{i^*j^*t^*}
\end{split}
\end{align}
where the bar variables combined with the star indices denote means taken within that index's group.
For example, $\Bar{\mathcal{A}}_{i^*jt}$ is the mean value of all $i^*$'s assigned to $i$'s group, $\Bar{\mathcal{A}}_{i^*j t^*}$ is the mean across both $i^*$ in $i$'s group and $t^*$ in $t$'s group, and so on.
Taking group assignments as a given, this projection is equivalent to solving the following least squares problem
\begin{align}\label{eqn:fullWithClusterObjective}
Q(\beta,\mathbf{g}) = \min_{\alpha,\gamma,\delta}\sum_{i,j,t}\left( Y_{ijt} - X_{ijt}^\prime\beta - \alpha_{g_1(i)jt}-\gamma_{ig_2(j)t}-\delta_{ijg_3(t)}\right)^2
\end{align}
where $\alpha \in \mathbb{R}^{{\rm ncol}(g_{1})\times N_2\times N_3}$, $\gamma \in \mathbb{R}^{N_1\times {\rm ncol}(g_{2})\times N_3}$ and $\delta \in \mathbb{R}^{N_1 \times N_2\times {\rm ncol}(g_{3})}$ where ${\rm ncol(\cdot)}$ returns the number of columns of a matrix.
$\mathbf{g}$ is a list of group assignment for each dimension.
For example, $g_1(i)$ maps to the group identity of individual $i$.
This is why $\alpha$ is restricted to vary across only ${\rm ncol}(g_{1})$ different values in the first dimension, which is less than $N_1$.
Notice $\alpha$, $\gamma$ and $\delta$ can be defined as combinations of the within-cluster projection as follows,
\begin{align*}
\alpha_{g_1(i)jt} &:= \Bar{\mathcal{A}}_{i^*jt} - \Bar{\mathcal{A}}_{i^*j^* t} + \Bar{\mathcal{A}}_{i^*j^*t^*}\\
\gamma_{ig_2(j)t} &:= \Bar{\mathcal{A}}_{ij^* t} - \Bar{\mathcal{A}}_{i^* j^*t} \\
\delta_{ijg_3(t)} &:= \Bar{\mathcal{A}}_{ijt^*} - \Bar{\mathcal{A}}_{i^*j t^*},
\end{align*}
though this representation is not unique.
This projection also differences out any additive fixed effects and leaves the following interactive fixed-effects residual,
\begin{align}\label{eqn:withinClusterApproxError}
\widetilde{\mathcal{A}}_{ijt} =\sum_{\ell= 1}^L (\varphi^{(1)}_{i\ell} - \Bar{\varphi}^{(1)}_{i^*\ell}) (\varphi^{(2)}_{j\ell} - \Bar{\varphi}^{(2)}_{j^*\ell}) (\varphi^{(3)}_{t\ell} - \Bar{\varphi}^{(3)}_{t^*\ell}).
\end{align}
where $\Bar{\varphi}^{(1)}_{i^*\ell}$ is the group mean of $\varphi^{(1)}_{i^*\ell}$ for the $i^*$'s in $i$'s group, and so on for the other terms.
Hence, sufficient projection of the interactive fixed-effects terms relies on the weaker condition that parameters converge to their group means, namely, $\varphi^{(1)}_{i\ell} \rightarrow \Bar{\varphi}^{(1)}_{i^*\ell}$, $\varphi^{(2)}_{j\ell} \rightarrow \Bar{\varphi}^{(2)}_{j^*\ell}$ or $\varphi^{(3)}_{t\ell} \rightarrow \Bar{\varphi}^{(3)}_{t^*\ell}$ for each $\ell$. This condition is clearly weaker than requiring the parameters to converge to the population means.
In fact, the group mean differencing could be seen as a weighted mean difference across the population, with equal weight given to observations within the cluster and zero weight to observations outside of each cluster.
This fact will be utilised to also develop a more generic kernel type weighted difference estimator, synonymous to a Nadaraya-Watson type estimator.
So far it has been shown that with the relatively innocuous shift from the within transformation to the within-cluster transformation, projection of the additive heterogeneity term has been preserved and less restrictive conditions are required to also eliminate the interactive term.
Choice of clusters for this transformation is key to suffice this less restrictive condition.
Given a set of proxies to cluster on, clustering or matching methods can be used to find these groups, for example \cite{bonhomme2021discretizing}.
Developing a set of proxies and choice of cluster approach is discussed in Section~\ref{sect:tensordecomp}.
A sufficient condition for consistency can be concisely summed up by Assumption~\ref{ass:cluster}, which considers the model in full generality as in \eqref{eqn:fullModel}.
In Assumption~\ref{ass:cluster}, the set $\mathcal{M}$ is defined to be referenced at a later stage in the statement of the consistency result but can be ignored at this stage.
Importantly this set must be non-empty.
Reference to the cluster mechanism $\mathcal{C}$ is also made in the assumption as these may change and impact how reasonable each bound in Assumption~\ref{ass:cluster} is.
Generically, these cluster mechanisms are simply any way to group units together.
Again, at this stage reference to $\mathcal{C}$ can be ignored, it is just useful for later reference.
\begin{assumption}[Clustering]~
\label{ass:cluster}
Let $j_n(i_n)$ be any unit in the same cluster as $i_n$ from using cluster mechanism $\mathcal{C}$. Then,
\begin{enumerate}[(i).]
\item For all $n$ as $N_n\rightarrow\infty$,
\begin{align*}
\frac{1}{N_n}\sum_{i_n = 1}^{N_n}\left\|{\varphi}^{(n)}_{i_n} - {\varphi}^{(n)}_{j_n(i_n)}\right\|^2 \lesssim O_p(1)
\end{align*}
and,
\item For a non-empty subset $\mathcal{M}\subset\{1,\dots,d\}$ take for any $n^*\in\mathcal{M}$ a sequence $\xi_{N_{n^*}}\rightarrow 0$ as $N_{n^*}\rightarrow\infty$. Then,
\begin{align*}
\frac{1}{N_{n^*}}\sum_{i_n = 1}^{N_{n^*}}\left\|{\varphi}^{({n^*})}_{i_{n^*}} - {\varphi}^{({n^*})}_{j_{n^*}(i_{n^*})}\right\|^2
=
O_p(\xi_{N_{n^*}})
\end{align*}
\end{enumerate}
\end{assumption}
Assumption~\ref{ass:cluster}.(i) restricts the chosen clustering procedure to allocate stable clusters.
This implies that as $\{N_1,\dots,N_d\}\rightarrow\infty$ the cluster allocations do not become increasingly disparate in the underlying parameter space.
Assumption~\ref{ass:cluster}.(ii) states that for at least one dimension the clustering procedure finds matches with asymptotically negligible difference in the underlying parameter space.
These assumptions restrict the cluster mechanism to uncover closeness in the true parameter space, which implies a restriction on how the cluster proxies are sampled and on how these proxies are used to cluster.
Since proxies are not always estimated, and actually sometimes group assignments may be given extraneously, it is useful to state Assumption~\ref{ass:cluster} in generic terms that ignore these clustering mechanics.
With some additional regularity conditions on the set of regressors, this high level assumption is sufficient to guarantee consistency using the within-cluster transformation in conjunction with pooled OLS.
Details of the estimation procedure are given in Section~\ref{sect:tensordecomp} and of the regularity conditions and theoretical bounds in Section~\ref{sect:asymResults}.
Some data generating processes and clustering mechanisms that suffice these high level conditions are discussed in Section \ref{sect:asymResults}.
These are not exhaustive, however, and may suggest avenues for future research.
\section{Existing Literature}\label{sect:ExistLit}
This section details some existing literature on factor models and multidimensional low-rank problems.
This section is also utilised to make some useful definitions that are vital to the remainder of the paper.
The section starts with a brief heuristic explanation of how the problem may be transformed into a simple two-dimensional panel data model and estimated using methods in that literature.
The multidimensional array framework is then discussed along with the literature related to low-rank approximation in those settings.
\subsection{Matrix low-rank approximation estimator}\label{sect:estimation}
This section provides a heuristic description of some matrix methods that can be applied directly to the multidimensional model.
Specific regularity conditions are left to the results in Section~\ref{sect:asymResults} for brevity.
As such, this subsection serves to highlight settings where matrix methods are sufficient and where they are not, and to justify the use of the more elaborate multidimensional methods.
Note that \cite{kapetanios2021estimation} employ a similar approach for three-dimensional arrays in conjunction with the \cite{Pesaran2006} common correlated effects estimator.
Consider, first, recasting the multidimensional array problem into a two dimensional panel problem by reorganising the indices, e.g. with the first dimension varying over the row space and the remaining indices jointly varying over the column space. This procedure is known as flattening, and the formal definition is reserved until Section~\ref{sect:tensordecomp} for reference.
Again, assume a three dimensional model for simplicity.
This amounts to the regression line,
\begin{align*}
Y_{(1)} = X_{(1)}^\prime \beta + \varphi^{(1)} \Gamma^\prime +\varepsilon_{(1)}
\end{align*}
where $Y_{(1)}, X_{(1)} ,\varepsilon_{(1)} \in \mathbb{R}^{N_1\times N_2N_3}$, $\varphi^{(1)}$ is an $N_1\times L_{1}$ matrix and $\Gamma$ is an $N_2N_3\times L_{1}$ matrix that accounts for variation in $\varphi^{(2)}$ and $\varphi^{(3)}$.
The term $L_{1}$ is indexed by the dimension $1$ because it may vary non-trivially according to the flattened dimension.
It should then be apparent that this is exactly the model described in \eqref{eqn:bai}, that is, the standard linear model with factor structure unobserved heterogeneity as studied in \cite{Bai2009}, where the second dimension is simply a combination of $j$ and $t$.
Hence, a standard factor model that estimates at least $L_1$ factors should result in consistent estimation of the slope coefficients, see \cite{MoonWeidner2017}.
However, this relies on an important structural feature of the unobserved heterogeneity term.
When flattened in the chosen dimension -- the first dimension in the above example -- the rank of the matrix after flattening must be low relative to data size.
This implies that to successfully project out the variation in the fixed-effect term either the matrix of fixed-effects from any flattening is low-rank, or, at least one flattening leads to a low-rank matrix of fixed-effects and the analyst knows which flattening this is.
To use the above example again, this means the analyst knows that $\varphi^{(1)} \Gamma^\prime$ is low-rank, hence flattening in the first dimension is the correct way to recast the model to a panel data model, and so forth for the other flattenings.
Whilst requiring low-rankness in at least one dimension may be an acceptable restriction, having knowledge of which dimension this low-rankness resides in is potentially more restrictive.
To understand the problem, consider the following two examples, one where the flattening is not low-rank and one where it is.
First, assume $\varphi^{(1)}$ varies in a high-dimensional parameter space, e.g. with $N_1 < N_2N_3$, $\varphi^{(1)} \in \mathbb{R}^{N_1\times N_1}$ and $\Gamma\in\mathbb{R}^{N_2N_3\times N_1}$ with each column mutually orthogonal for both these matrices.
Then the product of these matrices is full-rank and any factor projection approach will not fully control for this term.
On the contrary, consider $\varphi^{(1)} \in \mathbb{R}^{N_1\times N_1}$ where all columns are linearly dependent.
Then the matrix $\varphi^{(1)} \Gamma^\prime$ is rank-1 regardless of $L$ and of how $\varphi^{(2)}$ and $\varphi^{(3)}$ vary, thus can be projected with a factor model estimated with 1 factor.
Hence it is important which dimension the analyst chooses to flatten over.
Well established diagnostics in \cite{BaiNg2002}, \cite{ahn2013eigenvalue} and \cite{hallin2007determining} can be used to determine the number of factors.
These diagnostics can be repeated across different flattenings, which may be informative of the dimension to use for flattening.
Note these procedures require an initial guess of $\beta$ and relies on this guess not eradicating the factor structure in the residual; see the beginning of Section~\ref{sect:covariateCluster} for a concrete example of this.
It should also be noted that these diagnostics are not without restrictions and can lead to spurious conclusions on the optimal number of factors.
For example, the eigenvalue ratio test in \cite{ahn2013eigenvalue} can undershoot the number of factors when singular values decay quickly for the leading few factors.
This does not interfere with the asymptotic result in that paper but can have implications in small sample estimation.
Indeed, however, these diagnostics can be helpful in both the matrix recasting of the problem and the group fixed-effects estimation in the sequel.
\subsection{Multidimensional arrays}\label{sect:HOSVD}
Discussed here are some important features of multidimensional arrays and their use in this paper.
This includes a discussion on how to uncover cluster proxies from multidimensional array data.
It also includes a discussion on why standard matrix methods do not extend well to the multidimensional setting.
This section is also utilised to define important objects referenced in later sections.
First, consider how to cluster in the within-cluster transformation.
In most clustering algorithms, for example $K$-means or $K$-nearest neighbour, there is some notion of a distance metric between units considered for each cluster.
To arrive at a distance there must be some space to measure that distance over.
For example, using some vector $u_i$ and the Euclidean norm of differences, $\|u_i - u_j\|$, to measure the distance between units.
Algorithms to arrive at these groupings are well established when the distance metric and variable to take distance over are given.
However, in this setting there is no clear variable through which to take distance over.
Motivated here are methods to extract proxies that serve to measure distance across units in a way that isolates variation in each dimension of the unobserved heterogeneity term.
It is important to find proxies that isolate variation in each dimension since clustering is to be performed one index at a time.
Discussed here are decompositions of multidimensional arrays that can perform this, \cite{kolda2009tensor} contains a nice summary of some candidate decompositions.
The method discussed here uses the higher order singular value decomposition (HOSVD), and focuses on components of this decomposition that have well formulated theoretical properties.
The HOSVD is traditionally used in pursuit of a low-rank tensor decomposition by either direct truncation of left singular vectors or by some iteration approach similar to this, see for example the higher order orthogonal iteration scheme (HOOI).
The problem of direct truncation, however, is not well-posed because the solution to the low-rank tensor problem may not be unique and reformulating the original tensor after the aforementioned truncation is not guaranteed to be lower tensor rank.
See \cite{de2008tensor} for an extensive explanation of the ill-posedness issues.
Hence this method cannot be used in the pursuit of analytic consistency results.
Problems also arise in this setting where the reformulated tensors can be arbitrarily well approximated by a tensor of lower tensor rank, which is a result of the border rank issue of the tensor rank decomposition.
Whilst the HOOI scheme is implemented in simulations in Section~\ref{sect:sims} and may be well motivated from a numerical standpoint, this type of tensor decomposition is not discussed further in this paper.
Reconsider the three dimensional model with heterogeneity of the following form
\begin{align}\label{eqn:hosvdMotivation}
\mathcal{A}_{ijt} = \sum_{\ell = 1}^L \varphi^{(1)}_{i\ell} \varphi^{(2)}_{j\ell} \varphi^{(3)}_{t\ell}.
\end{align}
It would be convenient to use some Eckart–Young–Mirsky like theorem to arrive at a low-rank approximation for the multidimensional array $\boldsymbol{\mathcal{A}}$.
This would provide an appropriate method to project out low-rank variation in the unobserved heterogeneity in the original regression model.
Unfortunately such a theorem does not exist in the multidimensional setting with more than two dimensions.
In fact, if multidimensional array rank is defined as the least number of outer products required to exactly decompose a multidimensional array, the low-rank problem is not well-posed, see again \cite{de2008tensor}.
This motivates the use of the group fixed-effects as an alternative solution.
Take the standard singular value decomposition for some $N_1\times N_2$ matrix $A$,
\begin{align*}
A = U\Sigma V^\prime.
\end{align*}
This decomposition is well-known in the literature and can be used, for example, to estimate fixed-effects parameters in the interactive fixed-effects model.
The elements of any fixed column of $U$ explains variation across a row space in the original matrix $A$, likewise the elements of any fixed column of $V$ explain variation over a column space of $A$.
As is convention, the columns of $U$ and $V$ are arranged in descending order according to their joint explanatory value of the original matrix.
Thus, if $A$ is a low-rank object then the leading few columns of $U$, respectively $V$, are useful proxies to measure closeness in the row space, respectively the column space of $A$.
Extending this concept to the multidimensional setting turns out to be remarkably simple.
First to define some objects.
Take the factor-$n$ flattening, abbreviated to just $n$-flattening, of a $d$-order multidimensional array
$\boldsymbol{\mathcal{A}}\in \mathbb{R}^{N_1\times\dots\times N_d}$ to be the matrix
$\mathcal{A}_{(n)}\in\mathbb{R}^{N_n\times N_{n+1}\dots N_d N_1\dots N_{n-1}}$ that is simply a rearrangement of the original array.
This matrix arranges the data so the $n$\textsuperscript{th} dimension varies along the row space and data from the remaining $d-1$ dimensions vary along the column space.
How these remaining $d-1$ dimensions are arranged is not important - what is important is isolating variation in each dimension one at a time along the row space.
With this definition of a flattening the multilinear rank can be defined as the vector $\mathbf{r} = \{r_1,\dots,r_d\}$ where each $r_n$ is
\begin{align*}
r_n := rank\big(\mathcal{A}_{(n)}\big).
\end{align*}
Reconsider the singular value decomposition for matrices, applied to each of the $n$-flattenings of $\boldsymbol{\mathcal{A}}\in \mathbb{R}^{N_1\times,\dots,\times N_d}$ as
\begin{align}\label{eqn:nFlatSVD}
\mathcal{A}_{(n)} = U^{(n)}\Sigma_n V^{(n)\,^\prime} .
\end{align}
The rank of $\mathcal{A}_{(n)}$, defined above as $r_n$, is simply the number of non-zero elements of the diagonal matrix $\Sigma_n$.
By the same logic as in the matrix case, variation over the rows of each $U^{(n)}$ explains variation over the $n$\textsuperscript{th} dimension of the multidimensional array $\boldsymbol{\mathcal{A}}$.
Thus, if $\mathcal{A}_{(n)}$ is low rank, the leading few columns of $U^{(n)}$ provide good proxies for closeness in the row space of $\mathcal{A}_{(n)}$, which is the vector space of the $n$\textsuperscript{th} dimension.
This shows that a simple rearrangement of the data provides readily available techniques to measure closeness in each dimension separately.
Consider for any of the dimensions $n$ the corresponding matrix of left singular vectors from above, $U^{(n)}$, estimated with noise $\varepsilon_{ijt}$.
That is, each $U_n$ are calculated from the object $\boldsymbol{\mathcal{V}} = \boldsymbol{\mathcal{A}} + \boldsymbol{\varepsilon}$.
Under reasonable regularity conditions on the noise term $\boldsymbol{\varepsilon}$, the left singular vectors from this decomposition comprise of a signal of the underlying parameter and noise from $\boldsymbol{\varepsilon}$.
For example, in the three dimensional case,
define the $L_1$-vector $\widehat{U}^{(1)}_{i}$ as the $i$-{th} row of the left singular matrix of $\boldsymbol{\mathcal{A}} + \boldsymbol{\varepsilon}$ flattened in the first dimension.
Then the vector $\widehat{U}^{(1)}_{i}$ may comprise of,
\begin{align*}
\widehat{U}^{(1)}_{i} = {\varphi_i^{(1)}} + O_p\left(\frac{1}{\sqrt{\min\{N_1,N_2N_3\}}}\right).
\end{align*}
Likewise, for any dimension $n$ define $\widehat{U}^{(n)}$ as the matrix of singular vectors from $\boldsymbol{\mathcal{A}} + \boldsymbol{\varepsilon}$ flattened in the $n$-{th} dimension, where $\boldsymbol{\mathcal{A}}$ is the unobserved fixed-effects component of interest and $\boldsymbol{\varepsilon}$ is the usual idiosyncratic noise term.
Then, \cite{BaiNg2002} detail conditions required for the following ``up-to-rotation'' consistency result, which has been amended to this paper's setting;
\begin{lemma}[Theorem 1 from \cite{BaiNg2002}]\label{lemma:cons_proxy}
For any fixed integer $k\geq 1$, there exists an $(r_n\times k)$ matrix $H^k_n$ with ${\rm rank}(H^k_n) = \min\{k,r_n\}$ and $C_n = \min\big\{\sqrt{N_n}, \prod_{n^\prime\neq n}\sqrt{N_{n^\prime}}\big\}$ such that for each $n$ under some regularity conditions
\begin{align*}
C_n^2\norm{\widehat{U}^{(n)}_{i_n} - H^{k\,\prime}_n\varphi^{(n)}_{i_n}}^2 = O_p(1).
\end{align*}
\end{lemma}
This establishes a consistency result for estimating cluster proxies and suggests these left singular vectors are viable options to cluster in each dimension.
It also makes concrete the limitation implied by the value of $C_n$ for each index - that short indices have poorly estimated proxies.
However, given that the error term displayed in \eqref{eqn:withinClusterApproxError} is multiplicative across dimension, the error from this poor approximation should become negligible as long as enough other dimension proxies are well estimated.
Also, the presence of the rotation matrices, $H^k_n$, in Lemma~\ref{lemma:cons_proxy} can be ignored since these do not change relative distances of each unit under standard distance metrics used to cluster.
\section{Group fixed-effects estimator}\label{sect:tensordecomp}
Discussed below is the main estimation procedure for the group fixed-effects estimator.
In the main description one type of clustering mechanism is used, however, this can be extended to other methods to group units.
These other methods, including a kernel function approach, are discussed later in the section.
\subsection{Main estimator}\label{sect:estClusterResid}
Let $\Theta_\mathcal{C}$ be the space of group fixed-effects parameters associated to cluster mechanism $\mathcal{C}$, which is stated generically here.
That is, each $\boldsymbol{\theta} \in \Theta_\mathcal{C}$ is a $d$-list of $\bigtimes_{n=1}^d N_n$ tensors.
For each $n$ in $\{1,\dots,d\}$, the tensor $\theta_n$ varies freely over dimensions $\{1,\dots, n-1\}$ and $\{n+1,\dots, d\}$ but is fixed within each cluster along dimension $n$.
The objective function for the group fixed-effect estimation of $\beta$ under cluster mechanism $\mathcal{C}$ is
\begin{align}\label{eqn:GFEobj}
Q(\beta,\mathcal{C}) = \min_{\theta\in\Theta_\mathcal{C}}\norm{ \mathbf{Y} - \sum_{k=1}^K\mathbf{X}_k\beta_k - \sum_{n=1}^d\theta_n }^2_F
\end{align}
and $\widehat{\beta}_{GFE,\mathcal{C}}:= \operatorname*{argmin}_{\beta\in \mathbb{R}^K}Q(\beta,\mathcal{C})$.
It should be clear that the parameter space $\Theta_\mathcal{C}$ is indexed by clustering mechanism $\mathcal{C}$ because this mechanism defines how the parameters may vary.
That is, this is the estimated parameter space under a specific group fixed-effects estimator, which may only be an approximation of the true parameter space.
Consider taking cluster proxies from the estimated error term $\mathbf{W} = \mathbf{Y} - \mathbf{X}^\prime\widetilde{\beta}$.
Define $\widetilde{\beta}$ as the interim estimator used to obtain $\mathbf{W}$, and notice that this forms the basis of an iterative procedure, between forming clusters and estimating slope coefficients.
This is illustrated in the following two-step procedure.
For the below let $\widehat{L}_n$ be a hyperparameter that defines the number of singular vectors to use in the clustering stage.
\begin{enumerate}
\item For given $\widetilde{\beta}$, take the left singular matrices from each $n$-flattening of $\mathbf{W} = \mathbf{Y} - \mathbf{X}^\prime\widetilde{\beta}$ to obtain $\{\widehat{U}_1,\widehat{U}_2,\widehat{U}_3\}$.
\item
Cluster on the leading $\widehat{L}_n$ columns of $\widehat{U}_n$ to generate cluster assignments in the $n$\textsuperscript{th} dimension.
Use these cluster assignments in the within-cluster transformation on $\mathbf{Y}$ and $\mathbf{X}$ then perform pooled OLS to obtain $\widehat{\beta}$.
\item Iterate steps 1 and 2 until convergence in the slope coefficients
\end{enumerate}
This procedure may also be used as a debias estimator for a given initial estimate of $\widetilde{\beta}$ by ignoring step 3. Iteration here may not be stable given that step 1 and 2 do not optimise the same objective function, hence for theoretical purposes it may be convenient to only consider this as a debias procedure. In practice, iterating between step 1 and 2 after some initial grid search to initialise $\beta$ may be optimal.
Of course, other clustering or transformations may be used in place of the residual clustering and within-cluster transformation.
In the below, two alternatives are provided.
The first maintains the within-cluster transformation but considers a different set of proxies.
The second approach considers a kernel weighted transformation procedure that uses a generic set of proxies.
At this stage and in the below estimator refinements the analyst may be concerned with the number of parameters required to conduct these transformation.
Appendix~\ref{sect:reduceParameters} discusses a number of ways to reduce the size of the parameter space, including only projecting fixed-effects over a subset of dimensions and letting group sizes increase to reduce the number of groups.
\subsection{Clusters on covariates}\label{sect:covariateCluster}
Whether used as an iterative scheme or an update, the above method has some identification issues.
As an illustration take the data generating process for model \eqref{eqn:IntroModel} with just one covariate,
\begin{align*}
X_{ijt} = -\mathcal{A}_{ijt} + \mu_{ijt},
\end{align*}
where $\mu_{ijt}$ is a white noise term. Consider an initial guess of $\widetilde{\beta} = 0$ when the true value is $\beta^0 = 1$.
This leaves the residual term from Step 1 to base cluster assignment on as, $\mathbf{W} = \mathbf{Y} - \mathbf{X}\widetilde{\beta} = \mathbf{Y}$, which reduces to $\mathbf{W} = \boldsymbol{\mu} + \boldsymbol{\epsilon}$.
Thus, clustering is based solely on noise and can be reasonably described as random.
The associated within-cluster transformation will not project variation in the $\mathcal{A}$ terms that appear in both $\mathbf{Y}$ and $\mathbf{X}$ such that the OLS step in stage 2 produces
\begin{align*}
\widehat{\beta} \approx \frac{Var(\mu_{ijt})}{Var(\mu_{ijt}) + Var(\mathcal{A}_{ijt})} + o_p(1).
\end{align*}
For $\frac{Var(\mu_{ijt})}{Var(\mathcal{A}_{ijt})} \rightarrow 0$, $\widehat{\beta}\rightarrow 0$ and the algorithm does not update the initial guess of $\widetilde{\beta} = 0$.
This problem also arises in the matrix methods in Section~\ref{sect:estimation} and is a more fundamental issue with this algorithmic approach.
This example clearly displays some identification issues with the above method.
Worth noting is that this may be alleviated with a grid search approach, though this can be computationally infeasible even for a moderate number of covariates since the grid grows exponentially in the number of covariates.
To avoid this, proposed below is a method to extract cluster allocations from only variation in the set of covariates.
As discussed below, this clustering may also be conducted on control variables extraneous to the regression line.
The two-step procedure works as follows.
\begin{enumerate}
\item Take the left singular matrices from each $n$-flattening of ${X}$ to obtain $\{\widehat{U}_1,\widehat{U}_2,\widehat{U}_3\}$.
\item
Cluster on the leading $\widehat{L}_n$ columns of $\widehat{U}_n$ to generate cluster assignments in the $n$\textsuperscript{th} dimension.
Use these cluster assignments in the within-cluster transformation on $\mathbf{Y}$ and $\mathbf{X}$ then perform pooled OLS to obtain $\widehat{\beta}$.
\end{enumerate}
An advantage of using covariate clustering is that it can make use of control variables that are a good signal of cluster but are not included in the regression line.
For example, a control variable $Z_i$ that is constant across $j$ and $t$ may be a good candidate to cluster along the $i$ dimensions but will be projected out with the within-cluster transformation, so cannot be used directly in the pooled OLS estimation of $\beta$ stage.
This refinement also makes optimisation over $\beta$ a convex problem, and no iteration is required because clustering is not a function of $\beta$ estimates like in the approach in Section~\ref{sect:estClusterResid}.
\subsection{Kernel Methods}\label{sect:kernel}
The distance metrics used for the group fixed-effects methods may also be used in kernel functions to difference fixed-effects with a Nadaraya-Watson style estimator.
The estimator is then simply the within-cluster transformation with weighted means across the whole population rather than the mean within each cluster.
To see this take the usual distance metric as
$d(u_{i_n}, u_{i_n^*}) := \norm{u_{i_n} - u_{i_n^*}}$
and apply a kernel function to take weighted difference over the fixed-effects parameter space.
Take the proxy parameters used in the distance metric calculations generically as $\widehat{\varphi}_{i_n}^{(n)}$.
Then for each $n\in\{1,\dots,d\}$ and $i_n\in\{1,\dots,N_n\}$, define the kernel weight
\begin{align*}
w(i_n,j_n) := \frac{k\left(\norm{\widehat{\varphi}_{i_n}^{(n)} - \widehat{\varphi}_{j_n}^{(n)}}\right)}{\sum_{i^\prime_n}k\left(\norm{\widehat{\varphi}_{i_n}^{(n)} - \widehat{\varphi}_{i^\prime_n}^{(n)}}\right)}.
\end{align*}
These weights can then be used in the standard within-transformation \eqref{withinEstimator}.
Take, for example, the term $Y_{i^*_1 i^*_2 i_3}$ from \eqref{withinGroupEstimator}, where the more generic index notation is used here, as,
\begin{align*}
Y_{i^*_1 i^*_2 i_3} = \sum_{j_1 = 1}^{N_1} \sum_{j_2 = 1}^{N_2} w(i_1,j_1) w(i_2,j_2) Y_{j_1j_2i_3}.
\end{align*}
This transformation then creates an analogous estimation procedure to the cluster methods mentioned above.
Proxies used in the kernel functions to create the weights can be the same singular vectors used for forming clusters and a similar analysis for the estimator performance follows.
Of course, as with the group fixed-effects case, the proxies used in these kernel functions need not be the singular vectors and may be other observed control variables or other unobserved proxies.
After this weighted difference, pooled OLS is used on the transformed variables to arrive at the estimator for $\beta$.
Specifically,
let $\check{\mathbf{A}}$ denote a tensor after the kernel weighted differencing transformation described above is applied.
Then $\widehat{\beta}_{KER}$ is the pooled OLS estimator of $\check{\mathbf{Y}}$ regressed on $\check{\mathbf{X}}$.
\section{Asymptotic results}\label{sect:asymResults}
This section discusses the main asymptotic results for the low-rank approximation matrix method and group fixed-effects method.
The result for the group fixed-effects method is stated generically and applies to different approaches that comply with Assumption~\ref{ass:cluster}.
Results for the kernel methods are then stated.
In the following, let $\beta^0$ denote the true value of the slope coefficient.
\subsection{Matrix methods}\label{sect:baiCons}
The following restrictions are required for consistency when using the \cite{Bai2009} estimator after first flattening over a chosen dimension.
The tensor rank parameter, $L$, may without loss be restricted to the upper bounded by $L \leq \min_n\{\prod_{n^\prime\neq n } N_{n^\prime}\}$.
This is a result of elementary bounds on the tensor rank of an arbitrary tensor.
If the multilinear rank terms depend on data dimensions then the tensor rank may be indexed as $L_N$, where $N$ generically refers to the collection of dimensions $N_n$ for $n\in\{1,\dots,d\}$.
The multilinear rank of $\boldsymbol{\mathcal{A}}$ is restricted such that it is low-rank along at least one of the flattenings.
Let $\beta^0$ denote the true parameter value for the slope coefficients.
In the following assumptions let $\widehat{L}_n$ be the estimated number of factors for the $(n)$-flattening of the regression line when applying the least square methods in \cite{Bai2009}.
Also, let $\mathcal{L}\subset \{1,\dots,d\}$, and this this subset be arbitrary for now.
\begin{assumption}[Bounded norms of covariates and exogenous error]~
\label{ass:norms}
\begin{enumerate}[(i).]
\item $\big\|{{X}_k}\big\|_F = O_p\left(\prod_{n=1}^d \sqrt{N_n}\right)$ for each $k$
\item $\norm{{\varepsilon}_{(n)}} = O_p\left(\max\{\sqrt{N_{n^*}},\prod_{m\neq n^* }\sqrt{N_m}\}\right)$ for each $n^*\in\mathcal{L}$
\end{enumerate}
\end{assumption}
\begin{assumption}[Weak exogeneity]~
\label{ass:Exog}
${\rm vec}(X_k)^\prime{\rm vec}(\varepsilon) = O_p\left(\prod_{n=1}^d \sqrt{N_n}\right)$ for each $k$
\end{assumption}
\begin{assumption}[Low multilinear rank]~
\label{ass:multiLrank}
Let $(r_1,\dots,r_d)$ be the multilinear rank of $\boldsymbol{\mathcal{A}}$.
Then for some positive integer, $c$, $r_{n^*} < c $ for all $n^*\in\mathcal{L}$.
\end{assumption}
\begin{assumption}[Non-singularity]~
\label{ass:NonSing}
Let $\sigma_s(A)$ be the $s$\textsuperscript{th} singular value for a matrix $A$.
For each dimension $n^*\in\mathcal{L}$ that satisfies Assumption~\ref{ass:multiLrank}, there exists a $K\times 1$ unit vector $\delta_{n^*}$ such that
\begin{align*}
\sum_{s = 2\widehat{L}^{(n^*)} + 1}^{\min\{N_{n^*}, \prod_{m \neq n^*}N_m\}} \sigma_s\left( \frac{(\delta_{n^*}\cdot X_{(n^*)}) (\delta_{n^*}\cdot X_{(n^*)})^\prime }{\prod_{n}\sqrt{N_n}} \right) > b > 0 \quad \quad wpa1.
\end{align*}
\end{assumption}
Assumptions~\ref{ass:norms}, \ref{ass:Exog} and \ref{ass:NonSing} are standard regularity assumptions already well established in the literature, e.g. see \cite{MoonWeidner2015}.
Assumption~\ref{ass:norms}.(i) ensures that the covariates have bounded norms, for example having bounded second moments.
Assumption~\ref{ass:norms}.(ii) allows for some weak correlation across dimensions, see \cite{MoonWeidner2015}, or is otherwise implied if the noise terms are independently distributed with bounded fourth moments, see \cite{latala2005some}.
Assumption~\ref{ass:Exog} is implied if $X_{i_1,i_2,\dots,i_d;k}\varepsilon_{i_1,i_2,\dots,i_d}$ are zero mean, bounded second moment and only admits weak correlation across dimensions for each $k = 1,\dots, K$.
Assumption~\ref{ass:NonSing} simply states that, after factor projection, the set of covariates still collectively admit full-rank variation.
Assumption~\ref{ass:multiLrank} is new and asserts that there exists at least one flattening of the interactive term, $\boldsymbol{\mathcal{A}}$, that is low-dimensional or simply low-rank.
Given that the true value for $L$ is left mostly unrestricted at this stage, this requires that at least one of the unobserved terms $\varphi^{(n)}$ is low dimensional.
Note that not all dimensions must satisfy Assumption~\ref{ass:multiLrank} for the below result.
If the correct dimension is chosen then variation from the interactive term can be sufficiently projected out using the factor model approach.
This makes up the statement of the following Proposition.
\begin{proposition}
\label{prop:bai}
Let $\widehat{\beta}_{(n)}$ be the estimator from \cite{Bai2009} after first flattening along dimension $n$.
If Assumptions~\ref{ass:norms}-\ref{ass:NonSing} hold, the subset $\mathcal{L}$ is non-empty, and the estimated number of factors $\widehat{L}_{n^*}\geq r_{n^*}$, then, for each $n^*\in\mathcal{L}$ satisfying Assumption~\ref{ass:multiLrank},
\begin{align}
\norm{\widehat{\beta}_{(n^*)} - \beta^0} = O_p\left(\frac{1}{\sqrt{\min\{N_{n^*}, \prod_{n\neq n^*}N_n\}}}\right).
\end{align}
\end{proposition}
Proposition~\ref{prop:bai} follows directly from \cite{MoonWeidner2015} since the flattening procedure reduces the problem to the standard linear factor model.
Notice that this result only applies to estimates in the dimension(s) that satisfy the low-rank assumption in Assumption~\ref{ass:multiLrank}.
That is, implicit in Proposition~\ref{prop:bai} is that the analyst has chosen the correct dimension to flatten over when reformulating the problem as a two-dimensional panel.
Assumption~\ref{ass:multiLrank} can be relaxed to $r_{n^*} = o\left(\min\{N_{n^*}, \prod_{n\neq n^*}N_n\}\right)$ as long as the estimated number of factors is allowed to increase with data size at a faster rate than this.
The constraint $\widehat{L}_{n^*}\geq r_{n^*}$ can also be changed to $\widehat{L}_{n^*}\geq c$, however, this is more conservative than required for the statement of the result.
The estimation procedure from Proposition~\ref{prop:bai} can also be augmented to flatten over multiple indices.
For instance, the analyst may flatten such that both the rows and columns in the matrix contain multiple indices from the original array.
Of course, this augmentation makes Assumption~\ref{ass:multiLrank} harder to satisfy as it requires multiple parameters to vary in low-dimensional space.
To see this take the tensor $\boldsymbol{\mathcal{A}}$ flattened over the first two indices as $\mathcal{A}_{(1,2)}\in \mathbb{R}^{N_1N_2\times\prod_{n\notin\{1,2\}}N_n}$.
If the parameters $\varphi^{(n)}$ for $n = 3,\dots,d$ are high-dimensional, Assumption~\ref{ass:multiLrank} is only satisfied when both $\varphi^{(1)}$ and $\varphi^{(2)}$ and their product space is low-dimensional.
Clearly this is more restrictive than requiring only one of the parameter spaces to be low-dimensional.
However, flattening along multiple dimensions can improve the convergence rate in Proposition~\ref{prop:bai} to $O_p\left(\frac{1}{\sqrt{\min\{N_{1}N_{2}, \prod_{n\notin\{1,2\}}N_n\}}}\right)$, so there are benefits if this more restrictive assumption can be made.
As stated already, Proposition~\ref{prop:bai} takes for granted the correct choice in dimension to flatten across. Under Assumption~\ref{ass:norms}.(ii) the singular values of the flattened normalised noise term dissipates as follows;
\begin{align*}
\frac{1}{\sqrt{\prod_nN_n}}\norm{{\varepsilon}_{(n)}} = O_p\left(\frac{1}{\sqrt{\min\{N_n,\prod_{m\neq n }N_m\}}}\right).
\end{align*}
Since $\mathcal{A}$ is a collection of {fixed}-effects, the normalised singular values of its flattenings are $O_p\left(1\right)$, that is, the singular values are not asymptotically negligible like those of the noise term.\footnote{To see this consider the standard two dimension model and take the Frobenius norm any arbitrary component of the interactive fixed-effects term, $\lambda_rf_r^\prime$, normalised by ${1}/{\sqrt{NT}}$
\begin{align*}
\frac{1}{\sqrt{NT}}\norm{\lambda_rf_r^\prime}_F
&= \sqrt{\frac{1}{NT}\sum_i \sum_t (\lambda_{ir}f_{tr})^2}
= \sqrt{\frac{1}{N}\sum_i \lambda_{ir}^2} \sqrt{\frac{1}{T}\sum_t f_{tr}^2} = O(1).
\end{align*}
The last equality comes from $\lambda_{ir}$ and $f_{tr}$
having bounded second moments.
}
This ensures that, after flattening $\mathcal{A}$, each of the singular values eventually dominate those of the noise term.
These conditions make up similar restrictions imposed in \cite{ahn2013eigenvalue} that allow for the use of the eigenvalue ratio test (ER) to diagnose the number of factors.
Hence, in large samples, the analyst maybe able to use this test or similar to not only decide how many factors to use but also decide which dimension is likely to be low-dimensional.
Consider the factor model applied to a flattening that may not be low-rank.
For a concrete example of when this can occur see the data generating process in the simulations in Section~\ref{sect:sims}, where $\varphi^{(n)}$ are designed to be low-dimensional for some $n$, and high-dimensional otherwise.
Along the dimensions of $\mathcal{A}$ that do not conform to the low-rank assumption in Assumption~\ref{ass:multiLrank}, the tail singular values may become difficult to discern from the singular values of the noise term in small samples.
This means variation from those tail factors are less likely to be projected out from the factor model unless many factors are used in this projection.
If $r_n$ for $n\notin \mathcal{L}$ is allowed to increase adversely, for example at exactly the upper bound, then factor projection may never sufficiently project all relevant factors.
Also, as the number of estimated factors increases, Assumption~\ref{ass:NonSing} becomes harder to satisfy since variation in the set of covariates is also projected out.
This demonstrates the importance of choosing the correct dimension to flatten over, which is supported by the simulation results in Section~\ref{sect:sims}.
\subsection{Group fixed-effects methods}\label{sect:GFE}
Below some regularity conditions additional to Assumption~\ref{ass:cluster} are stated that are required for consistency of the group fixed-effects estimator.
A brief discussion of some alternative restrictions precedes the statement of the result.
Below are refinements to the regularity conditions contained within the Assumptions listed in Section~\ref{sect:baiCons} that account for the within-cluster transformation.
\begin{assumption}[Regularity conditions]~
\label{ass:regCond}
Let $\widetilde{T}_{i_1,\dots ,i_d}$ be the entries of tensor $\mathbf{T}$ with group fixed-effects projected out according to the within-cluster transformation with cluster mechanism $\mathcal{C}$. Then,
\begin{enumerate}[(i).]
\item $\left(\frac{1}{\prod_n N_n}\sum_{i_1}\dots\sum_{i_d}\widetilde{X}_{i_1,\dots ,i_d}\widetilde{X}_{i_1,\dots, i_d}^\prime\right) = \Omega_{X,N}$ exists and is positive definite along the sequence
$N_1,\dots, N_d \rightarrow \infty$.
Specifically, let $\Omega_{X,N} = O_p(1)$.
\item $\frac{1}{\prod_n N_n}\sum_{i_1}\dots\sum_{i_d}\widetilde{X}_{i_1,\dots ,i_d}\varepsilon_{i_1,\dots, i_d}= \Omega_{\varepsilon,N}$ exists as $N_1,\dots, N_d \rightarrow \infty$. \\
Specifically, $\Omega_{\varepsilon,N} = O_p\left(\frac{1}{\sqrt{\prod_nN_n}}\right)$.
\end{enumerate}
\end{assumption}
Assumption~\ref{ass:regCond}.(i) is very similar to Assumption~\ref{ass:NonSing} except that here full rank is required after the within-cluster projection rather than the factor projection.
Assumption~\ref{ass:regCond}.(ii) is an exogeneity condition that requires weak exogeneity in the covariates after the within-cluster transformation, which can be viewed as similar to Assumption~\ref{ass:Exog}.
This is stricter than Assumption~\ref{ass:Exog} because the noise term $\varepsilon$ can foreseeably impact cluster allocation via its impact on the estimation of cluster proxy terms.
This limitation is alleviated with, for instance, the use of cluster proxies extraneous to the regression line or perhaps through some sample splitting methods such as that proposed in \cite{freeman2022linear}.
The method proposed in Section~\ref{sect:covariateCluster}, clustering on covariates only, should also alleviate this issue if the non-transformed covariates are weakly exogenous.
\begin{proposition}[Upper bound on group fixed-effects estimator]
\label{lemma:consGFE}
Let Assumptions~\ref{ass:cluster} and \ref{ass:regCond} hold for cluster mechanism $\mathcal{C}$. Let $\mathcal{M}$ be the set defined in Assumption~\ref{ass:cluster}.(ii).
Then, for tensor rank $L_N$ that may depend on data size,
\begin{align*}
\big\|\widehat{\beta}_{GFE,\mathcal{C}} - \beta^0\big\|
=
\sqrt{L_N}O_p\left(
\prod_{n^*\in\mathcal{M}} \sqrt{\xi_{N_{n^*}}} \right)
+
O_p\left(\prod_{n = 1}^d\frac{1}{\sqrt{N_{n}}} \right)
.
\end{align*}
\end{proposition}
Before discussing this result an alternative restriction on the clustering mechanism is proposed.
If clustering is performed on proxies then Assumption~\ref{ass:cluster} can be stated in terms of the proxies, which forms the statement of Remark~\ref{remark:cluster}.
This requires that the proxies are estimated consistently and form an injective mapping to the true parameters.
An example of this are the conditions imposed in \cite{freeman2022linear}, stated in similar terms here:
\begin{remark}[Clustering]\label{remark:cluster}
The statement of Assumption~\ref{ass:cluster} can be reformulated in terms of the cluster proxies as follows.
Let $\widehat{\varphi}^{(n)}_{i_n}:= \widehat{\varphi}^{(n)}\big(\varphi^{(n)}_{i_n}\big)\in\mathbb{R}^{\widehat{L}_n}$ be the proxy for individual $i_n$ used to cluster along dimension $n$.
Then,
\begin{enumerate}[(i).]
\item For all $n$ as $N_n\rightarrow\infty$,
\begin{align*}
\frac{1}{N_n}\sum_{i_n}^{N_n}\left\|\widehat{\varphi}^{(n)}_{i_n} - \widehat{\varphi}^{(n)}_{j_n(i_n)}\right\|^2 \lesssim O_p(1)
\end{align*}
\item For a non-empty subset $\mathcal{M}\subset\{1,\dots,d\}$ take for any $n^*\in\mathcal{M}$ a sequence $\xi_{N_{n^*}}\rightarrow 0$ as $N_{n^*}\rightarrow\infty$. Then,
\begin{align*}
\frac{1}{N_{n^*}}\sum_{i_{n^*}}^{N_{n^*}}\left\|\widehat{\varphi}^{({n^*})}_{i_{n^*}} - \widehat{\varphi}^{({n^*})}_{j_{n^*}(i_{n^*})}\right\|^2
=
O_p(\xi_{N_{n^*}})
\end{align*}
\item Let $\varphi^{(n)}_{i_n}\in\Phi_n$ be the $L_n$-column vector of fixed effects, where $\Phi_n$ are convex sets for each $n$.
For each $a,b\in\Phi_n$ there exists a scalar $c_n>0$ such that
$\norm{a - b}\leq c_n \cdot \big\|\widehat{\varphi}^{(n)}(a)- \widehat{\varphi}^{(n)}(b)\big\|$
\end{enumerate}
If these alternate restrictions hold along with Assumption~\ref{ass:regCond}, then the bound in Proposition~\ref{lemma:consGFE} holds for the GFE estimator.
\end{remark}
Restrictions (i) and (ii) in Remark~\ref{remark:cluster} are exactly Assumption~\ref{ass:cluster}.(i) and (ii) but with cluster proxies in place of the true parameter values.
These are high level restrictions on the clustering mechanism that requires the mechanism to find closeness in the proxy space.
Restriction (iii) in Remark~\ref{remark:cluster} is an injectivity assumption on the proxy functions that demands closeness in the underlying parameter space given closeness in the proxy space.
This requires that the proxies do actually provide a mapping to the true parameter space, that is, that they are reasonable proxies.
An example of proxies that do this are the singular vectors from Section~\ref{sect:HOSVD} that fit the requirements of Lemma~\ref{lemma:cons_proxy}.
To see this expand the term $\|a - b\|$ and use the triangle inequality to see,
$\|a - b\|
\leq
\|a - \widehat{\varphi}^{(n)}(a)\| + \|\widehat{\varphi}^{(n)}(b) - b\| + \|\widehat{\varphi}^{(n)}(a) - \widehat{\varphi}^{(n)}(b)\|$,
where the first two terms are bound at the rate $O_p(C_n^{-1})$.
Note the rotation matrices are ignored for brevity and $C_n$ is the convergence rate from Lemma~\ref{lemma:cons_proxy}. Hence, asymptotically, Remark \ref{remark:cluster}.(iii) can be achieved with $c_n = 1$.
This display also makes clear the bottle-neck when clustering in high-dimensional objects. The distance of the proxies, $\|\widehat{\varphi}^{(n)}(a) - \widehat{\varphi}^{(n)}(b)\|$, is difficult to bound using clustering methods when the dimension of the proxies are larger than two, see \cite{graf2002rates} and further discussion in \cite{bonhomme2021discretizing}.
This implies that a low-dimensional set of proxies must bound the true parameter values for clustering methods to work well in this setting.
Hence, whilst the relationship in restriction (iii) of Remark~\ref{remark:cluster} may be satisfied for an arbitrarily high-dimensional set of proxies, for a reasonable family of cluster mechanisms to bound these proxies as per restrictions (i) and (ii) of this remark, restriction (iii) must also hold for a low-dimensional set of proxies.
This can be highly restrictive with, for example, fixed-effects $\varphi^{(n)}_{i_n}$ that are high-dimensional.
Of note in Assumption~\ref{ass:cluster} and Remark~\ref{remark:cluster} is the subset $\mathcal{M}$ of dimensions that admit the well-behaved cluster property.
As per the above, an implicit requirement on the latent parameters from this subset of dimensions is some form of low-dimensionality in the vectors $\varphi_{i_{n^*}}^{(n^*)}$ for $n^*\in\mathcal{M}$.
A sufficient condition for parameters in these dimension to be low-dimensional is low multilinear rank for each $n^*\in \mathcal{M}$ for a given sample.
Hence, it is expected that the set $\mathcal{M}$ should roughly coincide with the set $\mathcal{L}$, or at least be a subset of $\mathcal{L}$, from Assumption~\ref{ass:multiLrank}.
How the sequences $\xi_{N_{n^*}}$ converges to zero and how $L_N$ is bounded are important for the convergence result in Proposition~\ref{lemma:consGFE}.
First note that for fixed $L_N$ the first term in the result simplifies to
$O_p\left(
\prod_{n^*\in\mathcal{M}} \sqrt{\xi_{N_{n^*}}} \right)$.
Also note, if the conditions for Lemma~\ref{lemma:cons_proxy} hold and clustering is based on singular vector estimates that adhere to Remark~\ref{remark:cluster},
then it is possible to achieve $\xi_{N_{n^*}} = O_p\left(\frac{1}{\min\{N_{n^*} , \prod_{n\neq n^*}N_n\} } \right)$.
If each $N_n$ grow at the same rate
then the consistency result is,
\begin{align*}
\big\|\widehat{\beta}_{GFE,\mathcal{C}} - \beta^0\big\|
=
\sqrt{L_N}O_p\left( N_{n}^{-|\mathcal{M}|/2} \right).
\end{align*}
In the worst case scenario $\sqrt{L_N}$ is upper bound by $\sqrt{L_N}\lesssim N_n^{(d-1)/2}$, which is taken from $L_N \leq \min_n\prod_{n^\prime \neq n} N_{n^\prime}$.
The convergence result is then $O_p\left( N_{n}^{(d - 1 -|\mathcal{M}|)/2} \right)$, which is of course conservative but shows that if $|\mathcal{M}|= d$, then consistency is guaranteed albeit at the slow rate of $N_n^{1/2}$.
This means that all dimensions must have good cluster assignments, which is obviously not an ideal worst case but shows the limitations of this method when $L_N$ is unrestricted.
For the special case of $d = 3$ it can be shown that $L_N \leq \min_n \prod_{n^\prime\neq n}r_{n^\prime}$.
From the discussion above, it is expected that $n\in\mathcal{M}$ is sufficient for $n\in\mathcal{L}$, that is, $r_n$ is small for the set of dimensions $n \in \mathcal{M}$.
This tightens the bound in Proposition~\ref{lemma:consGFE} to $O_p\left( N_{n}^{\max\{ -|\mathcal{M}|/2, 1 - |\mathcal{M}|\}} \right)$, such that only $|\mathcal{M}|\geq 1$ is required for consistency.
The analogous tensor rank bound is so far not known for the case with $d\geq 4$.
\subsection{Kernel Methods}
Described here are the restrictions required to show consistency using the kernel methods in Section~\ref{sect:kernel}.
\begin{assumption}[Kernels]\label{ass:Kernels}
Take the class of kernel functions used as $\mathcal{K}$.
For $k \in \mathcal{K}$, $k$ is bounded.
Then for $h>0$ there exists an $\alpha>0$ such that
$
k(a/h)a \lesssim O(h^\alpha)$.
\end{assumption}
Assumption~\ref{ass:Kernels} refers to a bandwidth parameter, $h$, and restricts the kernels to penalise distance at a rate equal to or faster than $O(h^\alpha/a)$.
For consistency using the kernel methods, the sequence $h\rightarrow0$ is considered, such that an upper bound on $\alpha$ is the critical object of interest in Assumption~\ref{ass:Kernels} for the kernel function under consideration.
As an example of a class of kernel functions that satisfies Assumption~\ref{ass:Kernels}, the exponential class of the form considered in Remark~\ref{remark:Kernels} may be utilised.
\begin{remark}\label{remark:Kernels}
For $c_1, c_2>0$, let $k^\prime(a) \propto c_1exp(-c_2a^2)$ for all $a\geq 0$ and $k^\prime \in \mathcal{K}^\prime$.
Then $\operatorname*{argmax}_a k^\prime(a/h)a = h/\sqrt{2c_2}$, and,
\begin{align*}
\max_a k^\prime(a/h)a \propto \frac{c_1}{\sqrt{2c_2}}e^{-1/2}h = O(h)
\end{align*}
Thus, Assumption~\ref{ass:Kernels} is satisfied for the exponential class of kernel functions $\mathcal{K}^\prime$ with $\alpha = 1$.
Further, for $h\rightarrow 0 $, it suffices that $\alpha\in(0,1]$.
\end{remark}
Assumption~\ref{ass:Kernels} is stated more generically than Remark~\ref{remark:Kernels} as there is a larger class of bounded kernel functions that satisfy the sufficient restriction for result below.
\begin{assumption}[Sampling]\label{ass:sampling}
Let $\widehat{\varphi}_{i_n}^{(n)}\in\widehat{\Phi}_n$ be the proxy space for the fixed-effects and let $\mathcal{K}$ be the family of kernel functions used.
Let
$K_{i_n}(h_n)
:=
\max_{j_n}k\left( \frac{1}{h_n} \left\| \widehat{\varphi}_{i_n}^{(n)} - \widehat{\varphi}_{j_n}^{(n)} \right\|\right)$.
For $k\in \mathcal{K}$ and $0<\varepsilon_{i_n}<K_{i_n}(h_n)$ define
$$
M_n\left(\widehat{\varphi}^{(n)}_{i_n}, \varepsilon_{i_n} \right)
:=
\sum_{j= 1}^{N_n}
\mathbbm{1}\left(
k\left( \frac{1}{h_n} \left\| \widehat{\varphi}_{i_n}^{(n)} - \widehat{\varphi}_{j_n}^{(n)} \right\| \right) > \varepsilon_{i_n}
\right).$$
Then for any $\varepsilon_{i_n}\in(0, K_{i_n}(h_n))$,
\begin{align}\label{eqn:sampling}
\operatorname*{plim}_{N_{n}\rightarrow\infty} \frac{
M_n\left(\widehat{\varphi}^{(n)}_{i_n}, \varepsilon_{i_n} \right)
}{N_{n}}
\geq c_{i_{n}} \in(0,1].
\end{align}
for all ${i_{n}}\in 1,\dots,N_{n}$.
\end{assumption}
The upper bound on $\varepsilon_{i_n}$, $K_{i_n}$, is expected to be $k(0)$ for most classes of kernels.
That is, the kernel function evaluated at $\widehat{\varphi}_{i_n}^{(n)} = \widehat{\varphi}_{j_n}^{(n)}$ should maximise the value of the kernel function.
For example, $k(0) = c_1$ in the exponential class of kernel functions in Remark~\ref{remark:Kernels}.
Assumption~\ref{ass:sampling} may, for example, be satisfied for a reasonable class of kernel functions when the parameter space is generated such that the neighbourhood around each realisation grows proportionally with the sample size. Remark~\ref{remark:sampling} formalises this.
\begin{remark}[Sampling]\label{remark:sampling}
Let $\widehat{\varphi}_{i_n}^{(n)}\in\widehat{\Phi}_n$ and redefine $M_n^\varepsilon\left(\widehat{\varphi}^{(n)}_{i_n}\right)$ as
$$
M^\varepsilon_n\left(\widehat{\varphi}^{(n)}_{i_n}\right):=
\sum_{j_n= 1}^{N_n}
\mathbbm{1}\left(
\widehat{\varphi}_{j_n}^{(n)} \in B_\varepsilon\left(
\widehat{\varphi}_{i_n}^{(n)}\right) \right),$$
where $B_\varepsilon\left(x\right)$ is the $\varepsilon$-neighbourhood around $x$.
Then Assumption~\ref{ass:sampling} is satisfied for the dimension $n$
for any $\varepsilon>0$ if,
$$
\operatorname*{plim}_{N_{n}\rightarrow\infty} \frac{M_n^\varepsilon\left(\widehat{\varphi}^{({n^*})}_{i_{n}}\right) }{N_{n}} \geq c_{i_{n}} \in(0,1].
$$
for all $\widehat{\varphi}_{i_{n}}^{({n})}\in\widehat{\Phi}_{n}$.
\end{remark}
Assumption~\ref{ass:sampling} and Remark~\ref{remark:sampling} are restrictions on the data generating process of the fixed-effect proxy parameter space.
These are similar to Assumption 5.5 in \cite{altonji2005cross} except related to the fixed-effect parameter space.
Note that for these to be satisfied, the probability over the support of the fixed-effect space must be strictly positive.
Defined in Remark~\ref{remark:sampling} is essentially the building blocks of the probability space for the fixed-effects proxy parameter space.
Whilst this paper focuses on fixed-effects, that is, effects that are taken as given and not modelled as random variables, it is still useful to understand that these parameters are sampled from some space.
This space is what the restrictions in Assumption~\ref{ass:sampling} and Remark~\ref{remark:sampling} pertain to.
\begin{proposition}[Upper bound on kernel estimator]\label{lemma:consKernel}
Let the class of kernel functions used to perform weighted means difference and the proxy space used in these kernel functions satisfy Assumption~\ref{ass:Kernels} and \ref{ass:sampling}.
Extend Assumption~\ref{ass:regCond} to the set of covariates after kernel weighted means are differenced.
Let
$\left\| \varphi_{i_{n^*}}^{({n^*})} - \widehat{\varphi}_{i_{n^*}}^{({n^*})} \right\|^2 = O_p(C_{n^*}^{-2})$ for ${n^*}\in\mathcal{M}$
and
$\left\| \varphi_{i_{n^*}}^{({n^*})} - \widehat{\varphi}_{i_{n^*}}^{({n^*})} \right\|^2 = O_p(1)$
for $n^\prime \notin \mathcal{M}$,
where $\mathcal{M}$ is a non-empty subset of dimensions.
Let $h_n$ be the bandwidth parameter from Assumption~\ref{ass:sampling}.
Then,
\begin{align*}
\norm{\widehat{\beta}_{KER} - \beta^0}
=
\sqrt{L_N}O_p\left(\prod_{n^*\in\mathcal{M}}
\sqrt{O_p\left({C_{n^*}^{-2}} \right)
+
O_p\left(C_{n^*}^{-1}h_{n^*}^\alpha\right)
+
O_p\left(h_{n^*}^{2\alpha}\right)}
\right)
+
O_p\left(\prod_{n = 1}^d\frac{1}{\sqrt{N_{n}}} \right)
.
\end{align*}
For $h_n^\alpha \lesssim O(C_n^{-1})$ this reduces to
\begin{align*}
\norm{\widehat{\beta}_{KER} - \beta^0}
=
\sqrt{L_N}O_p\left(\prod_{n^*\in\mathcal{M}}
{O_p\left({C_{n^*}^{-1}} \right)}
\right)
+
O_p\left(\prod_{n = 1}^d\frac{1}{\sqrt{N_{n}}} \right)
.
\end{align*}
\end{proposition}
Proposition~\ref{lemma:consKernel} shows that the convergence rate for the kernel estimator is bounded by the convergence rate of the proxy estimates.
That is, as long as the bandwidth parameter approaches zero sufficiently fast, the kernel estimator converges at a rate no worse than the convergence of the proxies when proxies are estimated at or slower than $\sqrt{N}$-convergence.
This is expected and also a good result that the kernel method does not hinder the convergence rate from these proxies.
These kernel methods do, however, suffer the same curse of dimensionality as the group fixed-effects estimators since Assumption~\ref{ass:sampling} becomes increasingly difficult to satisfy as the dimension of the fixed-effects increases.
\section{Simulation}\label{sect:sims}
Table~\ref{tbl:sim} shows simulation results for the following DGP,
\begin{align*}
\begin{split}
Y_{ijt} &= X_{ijt}\beta + \mathcal{A}_{ijt} + \mathcal{B}_{ijt} + \varepsilon_{ijt}\\
X_{ijt} &= \mathcal{A}_{ijt} + \mathcal{B}_{ijt} + \nu_{ijt}
\end{split}
\end{align*}
with $\mathcal{A}_{ijt} = \sum_{\ell = 1}^{N_1} \varphi^{(1)}_{i\ell} \varphi^{(2)}_{j\ell} \varphi^{(3)}_{t\ell} $, $\mathcal{B}_{ijt} = \alpha_{ij} + \gamma_{it} + \delta_{jt}$. Also,
\begin{align*}
\varepsilon_{ijt}, \nu_{ijt}, \alpha_{ij}, \gamma_{it}\textrm{ and }\delta_{jt} &\stackrel{ i.i.d.}{\sim} N(0,1) \textrm{ and for each } \ell, \,\, \varphi^{(1)}_{i\ell}, \varphi^{(3)}_{t\ell} \stackrel{ i.i.d.}{\sim} \,\,N(0,1). \\
\varphi^{(2)}_{j1} &\stackrel{ i.i.d.}{\sim} N(0,1) \textrm{ with } \varphi^{(2)}_{j1}=\varphi^{(2)}_{j2}=\dots=\varphi^{(2)}_{jN_1}
\end{align*}
$\mathcal{A}_{ijt}$ and $\mathcal{B}_{ijt}$ are normalised to have unit variance.
$\boldsymbol{\mathcal{A}}$ is specified such that it is rank 1 when flattened in the second dimension and rank $N_1$ when flattened in either dimension one or three.
That is, the multilinear rank is $\mathbf{r} = (N_1,1,N_1)$.
This comes directly from the data generating process for each $\varphi^{(n)}$, where the matrix $\varphi^{(2)}$ is designed to be rank-1 and the matrices $\varphi^{(1)}$ and $\varphi^{(3)}$ are designed to be rank-$N_1$.
In Table~\ref{tbl:sim}, the estimators OLS and Fixed-effects are simply the pooled OLS estimator and the pooled OLS estimator after additive fixed-effects are projected out, respectively.
As expected both of these two have poor bias. The four GFE estimators perform well with reasonably low bias and standard deviation.
GFE (K-means) is GFE estimator with clustering based on the K-means algorithm, with proxies taken from the residual.
GFE (K-means on X) is the same estimator with proxies taken from the scalar covariate of interest.
GFE (1-NN) and GFE (1-NN on X) are likewise the same estimators but using the one nearest neighbours clustering.
The factor model is used after first flattening along each dimension as Factor(dim = $n$), where $n$ is the dimension used for flattening.
In each case, 2 factors are projected.
The results show the theoretical result succinctly, where the bias is close to zero when the correct dimension is flattened over (the second dimension in this case) and very poor bias when the incorrect dimension is used (the first and third dimensions).
Lastly, the kernel differencing estimator is estimated with kernel function bandwidths 0.5, 1 and 1.5; which all have comparable bias but substantially better standard deviation for bandwidth equal 1 and 1.5.
This analysis is repeated for the four dimensional case in Table~\ref{tbl:sim4way}, where the second and third dimensions admit low-dimensional unobserved interactive fixed-effects parameters.
For computational reasons, the GFE nearest neighbour estimators are omitted.
The simulations suggest similar results as the three dimensional case, where the factor models perform well when flattened in the low-dimensional dimensions (second and third) and poorly in the high-dimensional dimensions (first and fourth).
\begin{table}[H]
\centering
\begin{tabular}{|l|ccc|}
\hline
3-D & Mean bias & St. dev. & MSE \\
\hline
OLS & 0.6668 & 0.0033 & 4.45e-01 \\
Fixed-effects & 0.4997 & 0.0114 & 2.50e-01 \\
GFE (K-means) & 0.0118 & 0.0096 & 2.32e-04 \\
GFE (K-means on X) & 0.0129 & 0.0112 & 2.91e-04 \\
GFE (1-NN) & 0.0112 & 0.0153 & 3.61e-04 \\
GFE (1-NN on X) & 0.0111 & 0.0154 & 3.61e-04 \\
Kernel (h = 0.5) & 0.0030 & 0.0090 & 8.94e-05 \\
Kernel (h = 1.0) & 0.0031 & 0.0068 & 5.58e-05 \\
Kernel (h = 1.5) & 0.0037 & 0.0062 & 5.24e-05 \\
Factor (dim = 1) & 0.4319 & 0.0135 & 1.87e-01 \\
Factor (dim = 2) & 0.0030 & 0.0050 & 3.40e-05 \\
Factor (dim = 3) & 0.4319 & 0.0135 & 1.87e-01 \\
\hline
\end{tabular}
\caption{3D model ($N_1 = N_2 = N_3 = 36$), with 10,000 Monte Carlo rounds.
All results are in relation to $\beta$ estimation.
}\label{tbl:sim}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{|l|ccc|}
\hline
4-D & Mean bias & St. dev. & MSE \\
\hline
OLS & 0.6670 & 0.0018 & 4.45e-01 \\
Fixed-effects & 0.4981 & 0.0282 & 2.49e-01 \\
GFE (K-means) & 0.0012 & 0.0049 & 2.58e-05 \\
GFE (K-means on X) & 0.0013 & 0.0051 & 2.82e-05 \\
Kernel (h = 0.5) & 5.23e-05 & 0.0114 & 1.00e-04 \\
Kernel (h = 1.0) & 4.42e-05 & 0.0057 & 3.26e-05 \\
Kernel (h = 1.5) & -1.32e-05 & 0.0045 & 2.03e-05 \\
Factor (dim = 1) & 0.3733 & 0.0311 & 1.40e-01 \\
Factor (dim = 2) & 0.0030 & 0.0030 & 1.82e-05 \\
Factor (dim = 3) & 0.0030 & 0.0030 & 1.82e-05 \\
Factor (dim = 4) & 0.3734 & 0.0311 & 1.40e-01 \\
\hline
\end{tabular}
\caption{
4D model ($N_1 = N_2 = N_3 = N_4 = 20$), with 10,000 Monte Carlo rounds.
}\label{tbl:sim4way}
\end{table}
A two-dimensional simulation exercise is also performed to compare the grouped fixed-effects approach to the factor model approach in a setting where theoretical results for the factor model are well known.
Table~\ref{tbl:simtwoway} shows the results of this two-way setting where the data generating process is a factor model with two factors.
The GFE estimators have less bias than the factor model even when the factor model overestimates the number of factors.
To see this, compare the factor estimates with 2, 4 and 6 factors projected out with the GFE estimator.
For increase in variance of order $\approx 4$, the GFE estimator reduces bias by an order $\approx 10$.
This is a surprising improvement in estimates for a setting that is purpose designed for the factor model.
Where this comparison falls down is for models with a larger number of factors because generally clustering does not perform well when the latent parameter space has dimension greater than 2.
\begin{table}[H]
\centering
\begin{tabular}{|l|ccc|}
\hline
& Mean bias & St. dev. & MSE \\
\hline
OLS & 0.6672 & 0.0033 & 4.45e-01 \\
Fixed-effects & 0.5000 & 0.0043 & 2.50-e01 \\
GFE & 0.0002 & 0.0090 & 8.05e-05 \\
Kernel (h = 0.5) & 0.0003 & 0.0055 & 3.06e-05 \\
Kernel (h = 1.0) & 0.0003 & 0.0054 & 2.87e-05 \\
Kernel (h = 1.5) & 0.0004 & 0.0053 & 2.81e-05 \\
Factor (R = 2) & 0.0024 & 0.0047 & 2.76e-05 \\
Factor (R = 4) & 0.0031 & 0.0048 & 3.25e-05 \\
Factor (R = 6) & 0.0024 & 0.0049 & 3.03e-05 \\
\hline
\end{tabular}
\caption{2D model ($N_1 = N_2 = 216$), with 10,000 Monte Carlo rounds.
}\label{tbl:simtwoway}
\end{table}
\section{Empirical application - demand estimation for beer}
The methods proposed in this paper are applied to estimated the demand elasticity for beer.
Price and quantity for beer sales is taken from the Dominick's supermarket dataset for the years 1991-1995 and is related to supermarkets across the Chicago area.
Price and quantity vary over three dimensions in this example -- product ($i$), store ($j$) and month ($t$).
Fixed-effects that interact across all three dimensions can control for taste shocks to beer consumption that differ over both product and store.
Take for instance a large sporting event (temporary $t$ shock) that changes preferences differently across locations ($j$) and across certain subsets of sponsored beer ($i$).
For example,
in the stadiums for the many NBA finals playoffs the Chicago Bulls played in the early 1990's,
Miller Lite beer advertisements could be seen alongside advertisements for the substitute product Canadian Club whisky.
This suggests these events attracted large marketing campaign spends for these and other beer substitute brands that most likely also included price offers at local supermarkets.
Whilst the impact of these advertisements and price offers on the demand for or price of beer is not clear and,
further, that it is reasonably safe to assume the econometrician does not observe the plethora of marketing campaigns around these events,
the analyst would most likely still want to control for aggregate shocks like these.
For this reason it is important to use methods that robustly control for unobserved fixed-effects, such as unobserved marketing campaigns, that may impact both quantity demanded and prices in unforeseen ways.
Models for demand estimation ideally account for endogenous variation in prices and quantity.
The classic instrumental variable approach is to find a variable that varies exogenously to the production process but can reasonably describe price fluctuations.
A popular instrument in the estimation of beer demand is the commodity price for barley, one of the product's main ingredients, see e.g. \citet{saleh2014simple,tremblay1995advertising, richards2021dynamic}.
Since the price of barley is arguably not driven by the demand for it by any one supplier of beer, it can be a useful variable to instrument for price shifts.
In the following, it is taken as given that the price of barley is exogenous with respect to the noise term, $\varepsilon$.
For validity the instrument is also required to be strong, in the sense that it is strongly correlated with price.
In this dataset correlation between the price of barley, which varies over only $t$, and price of beer depends on how beer price is aggregated.
If beer price is first integrated over $i$ and $j$, such that it only varies over $t$, then it is highly correlated with the price of barley, at 0.61.
However, if beer price is not aggregated at all it is only correlated at 0.05.
This suggests there are important product and store level price drivers for beer that are not accounted for by fluctuations in the price of barley.
This implies that price fluctuations in barley alone may not be viable to fully capture beer prices when considering variation over all three dimensions.
For exogeneity, the price of barley must be independent of common unobserved shocks to both price and demand, which translates to being independent of $\varphi_{t,\ell}^{(3)}$ and any scalar fixed-effects that vary over $t$ in the interactive fixed-effects model.
More details are deferred to Appendix~\ref{sect:FEinEVframework}.
An alternative method is to follow the external variable approach of \cite{altonji2005cross}, that requires exogeneity of beer price conditional on an external variable.
If the price of barley is treated as the external variable, the beer demand example requires conditional exogeneity of beer prices across time periods of similar barley prices.
For example, that the price of beer is exogenous when compared to other beer prices during time periods of high barley price, and so on.
Note, the external variable approach provides much more flexibility than the traditional IV approach in the multidimensional setting because it may viably be used in conjunction with the group and kernel fixed-effects approaches even when the external variable varies over a subset of dimensions.
See Appendix~\ref{sect:FEinEVframework} for an explanation of how external variables are used in this setting and how the assumption works in conjunction with the fixed-effects estimators.
Table~\ref{tbl:demandappLogit} refers to estimates from the standard logit demand model,
\begin{align*}
\log(quantity_{ijt}) - \log(quantity_{1jt}) = price_{ijt}\beta + \mathcal{A}_{ijt} + \varepsilon_{ijt}
\end{align*}
where $\mathcal{A}_{ijt}$ is the usual interactive fixed-effects and no control variables are included since the set of available controls are rank-deficient and automatically projected out with standard scalar fixed-effects and from differencing out the outside option.
The outside option is encoded as product number 1 and is the aggregate consumption of products with small quantities consumed.
This serves the purpose of creating an outside option to do the necessary logit demand transformation as well as to avoid issues related to an unbalanced panel for the many niche products with sparse consumption amounts.
Own price elasticity is calculated as $\eta_{ijt} = price_{ijt}\beta (1 - quantity_{ijt}/\sum_{ijt}quantity_{ijt})$ and the mean elasticity is taken as the mean of this measure for each estimator.
The pooled instrumental variable and external variable estimates estimate relatively large elasticities.
However, all of the fixed-effects approaches estimate statistically similar slope coefficients and elasticities at the mean.
This implies that whilst some fixed-effects may exist in the true model for demand, they are unlikely complex enough to require the high-dimensional projections from the GFE or kernel methods.
This of course also takes for granted that the IIA logit model is the true model for demand.
\begin{table}[h]
\centering
\begin{tabular}{|l|c|c|}
\hline
Estimator & Coefficient (bootstrap st. dev.) & Elasticity at mean \\
\hline
Pooled OLS & -0.60 (0.04) & -3.26 \\
Pooled {IV} & -0.72 (0.04) & -3.91 \\
Pooled {EV} & -0.71 (0.05) & -3.86 \\
Additive FE & -0.32 (0.05) & -1.74 \\
Factor (dim = 1) & -0.29 (0.04) & -1.58 \\
Factor (dim = 2) & -0.32 (0.05) & -1.74 \\
Factor (dim = 3) & -0.37 (0.05) & -2.01 \\
GFE & -0.32 (0.05) & -1.74 \\
GFE ({EV})& -0.31 (0.05) & -1.68 \\
Kernel (Gaussian) & -0.30 (0.05) & -1.63 \\
Kernel ({EV}, Gaussian) & -0.34 (0.08) & -1.85 \\
\hline
\end{tabular}
\caption{
Logit demand estimates (73 products, 41 stores, 57 months).\\
{\footnotesize
Standard deviations were bootstrapped by resampling along each dimension separately.
In the first dimension, product 1 is fixed across bootstrap samples as the outside option and the remaining products are resampled with replacement.
}
}\label{tbl:demandappLogit}
\end{table}
The second column from Table~\ref{tbl:demandappLogLog} refers to the estimates for demand elasticities for the following regression model,
\begin{align}\label{eqn:LogLogNoCross}
\log(quantity_{ijt}) = \log(price_{ijt})\beta + \mathcal{A}_{ijt} + \varepsilon_{ijt}
\end{align}
where $\mathcal{A}_{ijt}$ is the usual interactive fixed-effects term from the prequel.
This amounts to estimating the standard log-log model for demand with fixed-effects.
That is,
\begin{align*}
quantity_{ijt} = price_{ijt}^\beta \exp({\mathcal{A}_{ijt}} + \varepsilon_{ijt}).
\end{align*}
Again, no controls are included here since they are low-dimensional and subsumed by the fixed-effects term.
This model specification estimates reasonably similar elasticities as the logit case across each of the different fixed-effects estimators
but relatively large differences in estimates for pooled OLS, IV and external variable estimators.
The similar elasticities for the different fixed-effects estimators within Table~\ref{tbl:demandappLogLog} again suggests that whilst some form of fixed-effects should be included, they may not need be as complex as implied by the GFE and kernel methods.
The third column from Table ~\ref{tbl:demandappLogLog} reports estimates of the same log-log model controlling for the average log price of other products,
\begin{align}\label{eqn:LogLogCross}
\log(quantity_{ijt})
=
\log(price_{ijt})\beta +
\delta\sum_{i^\prime \neq i} \log(price_{i^\prime jt}) +
\mathcal{A}_{ijt} + \varepsilon_{ijt}.
\end{align}
This model assumes homogeneous cross-elasticity over all other beer products.
That is, it refers to the demand model,
\begin{align*}
quantity_{ijt} = price_{ijt}^\beta \prod_{i^\prime \neq i}price_{i^\prime jt}^{\delta_{i i^\prime }} \exp({\mathcal{A}_{ijt}} + \varepsilon_{ijt}),
\end{align*}
where $\delta_{i i^\prime } = \delta$ for all $i$ and $i^\prime$.
Whilst this may oversimplify the system of cross-elasticities in the market for beer, it does significantly change the estimates for $\beta$ in the log-log model.
This suggests that cross-elasticities should probably be controlled for since $\beta$ estimates do seem sensitive to their inclusion.
Note that most estimators returned a negative value for $\delta$, which opposes the theory that other brands of beer, on aggregate, are substitutes.
However, since prices are aggregated in such a crude way, the cross-elasticity estimates should not be taken too seriously.
If interested in the cross-elasticities, then some care should be taken to segment or group products in such a way that actual substitution is being identified here, not just aggregate market forces.
For this model, all fixed-effects estimates are within statistical noise of each other, this time with the external variable approach being closely aligned.
IV is estimated with very high variation in both log-log models, which may be due to barley being a weak instrument.
\begin{table}[h!]
\centering
\begin{tabular}{|l|c|c|}
\hline
Estimator & $\widehat{\beta}$ (St. dev.) from \eqref{eqn:LogLogNoCross} & $\widehat{\beta}$ (St. dev.) from \eqref{eqn:LogLogCross} \\
\hline
Pooled OLS & 0.06 (0.31) & 0.09 (0.30) \\
Pooled {IV} & -4.76 (2.76) & -3.47 (2.30)\\
Pooled {EV} & -3.05 (0.48) & -2.74 (0.27) \\
Additive FE & -1.78 (0.34) & -2.80 (0.28) \\
Factor (dim = 1) & -1.61 (0.30) & -2.65 (0.25) \\
Factor (dim = 2) & -1.78 (0.33) & -2.78 (0.27) \\
Factor (dim = 3) & -2.09 (0.32) & -2.89 (0.27) \\
GFE & -1.85 (0.33) & -2.86 (0.30) \\
GFE ({EV})& -1.84 (0.36) & -2.70 (0.30) \\
Kernel (Gaussian) & -1.72 (0.33) & -2.58 (0.29) \\
Kernel ({EV}, Gaussian) & -1.97 (0.47) & -2.81 (0.43) \\
\hline
\end{tabular}
\caption{
Log-log demand elasticities (73 products, 41 stores, 57 months).\\
{\footnotesize
Standard deviations were bootstrapped by resampling along each dimension separately.
In the first dimension, product 1 is fixed across bootstrap samples.
}
}\label{tbl:demandappLogLog}
\end{table}
\section{Conclusion}\label{sect:conclusion}
This paper shows methods to generalise the interactive fixed-effect to the multidimensional case with more than two dimensions.
Theoretical results show that standard matrix methods can be applied to this setting but require additional knowledge of the data generating process.
The multiplicative interactive error from the group fixed-effects and kernel methods show a potential improvement on the asymptotic rate of convergence and suggest a more robust approach to projecting fixed-effects.
Simulations corroborate these theoretical results and show the relative advantage of using a standard factor model when the structure of the interactive term is known.
They also show the robustness of the group fixed-effects estimator to not having this same knowledge.
Inference in these models is still an open question for further research.
The model is applied to a simple demand model for beer consumption.
The application demonstrated how the GFE and kernel methods integrate well with the external variable approach in \cite{altonji2005cross} as apposed to instrumental variable approaches that do not allow for fixed-effect estimation when the instrument is rank-deficient.
The application shows that whilst some fixed-effects should likely be included in the model for beer demand, they are unlikely to be overly complicated to justify the GFE or kernel methods.
This is a useful analysis, as it provides a robustness check for the specification of fixed-effects in model specifications.
\setlength{\bibsep}{2pt}
\bibliographystyle{chicago3}
|
\section{Introduction}
\begin{table*}[t]
\centering
\caption{\label{table: summary of existing defenses} A summary of the existing defenses and our work.}
\resizebox{0.92\textwidth}{!}{
\begin{threeparttable}
\begin{tabular}{ccccccccccc}
\toprule
\multirow{2}{*}{\textbf{Type}} & \multirow{2}{*}{\textbf{Approaches}} & \multicolumn{3}{c}{\textbf{Detection Target}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\textbf{Black-box}\\~\textbf{Access}\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\textbf{No Need of}\\\textbf{Clean Data}\end{tabular}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\textbf{All-to-all}\\\textbf{Attack}\end{tabular}} & \multicolumn{3}{c}{\textbf{Trigger Assumption}} \\
\cmidrule{3-5}\cmidrule{9-11}
& & \multicolumn{1}{c}{\textbf{Input}} & \textbf{Model} & \textbf{Trigger} & & & & \textbf{Universal} & \textbf{Partial} & \textbf{Dynamic} \\
\midrule
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}I:\ Input \\Masking\end{tabular}} & STRIP~\cite{gao2019strip} & \pie{360} & \pie{90} & \pie{90} & \pie{360} & \pie{90} & \pie{90} & \pie{360} & \pie{90} & \pie{90}\\
& Februus~\cite{doan2020februus} & \pie{360} & \pie{90} & \pie{360} & \pie{90} & \pie{90} & \pie{360} & \pie{360} & \pie{90} & \pie{90} \\
& SentiNet~\cite{chou2020sentinet} & \pie{360} & \pie{90} & \pie{360} & \pie{90} & \pie{90} & \pie{360} & \pie{360} & \pie{90} & \pie{90}\\
\midrule
\multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}II:\ Model \\Inspection\end{tabular}} & Neural Cleanse~\cite{wang2019neural} & \pie{90} & \pie{360} & \pie{360} & \pie{90} & \pie{90} & \pie{90} & \pie{360} & \pie{90} & \pie{90} \\
& ABS~\cite{liu2019abs} & \pie{90} & \pie{360} & \pie{360} & \pie{90} & \pie{90} & \pie{90} & \pie{360} & \pie{90} & \pie{90} \\
& MNTD~\cite{xu2021detecting} & \pie{90} & \pie{360} & \pie{90} & \pie{360} & \pie{90} & \pie{360} & \pie{360} & \pie{90} & \pie{90} \\
\midrule
\multirow{5}{*}{\begin{tabular}[c]{@{}c@{}}III:\ Feature\\
Representation\end{tabular}} & Activation-Clustering~\cite{chen2019detecting} & \pie{90} & \pie{360} & \pie{90} & \pie{90} & \pie{360} & \pie{360} & \pie{360} & \pie{90} & \pie{90} \\
& Spectral-Signature~\cite{tran2018spectral} & \pie{90} & \pie{360} & \pie{90} & \pie{90} & \pie{360} & \pie{360} & \pie{360} & \pie{90} & \pie{90} \\
& SPECTRE~\cite{hayase2021spectre} & \pie{90} & \pie{360} & \pie{90} & \pie{90} & \pie{90} & \pie{360} & \pie{360} & \pie{90} & \pie{90} \\
& SCAn~\cite{tang2021demon} & \pie{360} & \pie{360} & \pie{90} & \pie{90} & \pie{90} & \pie{360} & \pie{360} & \pie{360} & \pie{90} \\
& \textbf{Beatrix (our work)} & \pie{360} & \pie{360} & \pie{90} & \pie{90} & \pie{90} & \pie{360} & \pie{360} & \pie{360} & \pie{360} \\
\bottomrule
\end{tabular}
\begin{tablenotes}
\item[]\footnotesize\pie{90}: the item is not supported by the defense; \pie{360}: the item is supported by the defense.
\end{tablenotes}
\end{threeparttable}
}
\vspace{-3mm}
\end{table*}
With an explosive growth of Machine Learning (ML) and Artificial Intelligence (AI), Deep Neural Networks (DNNs) are widely adopted in many significant real-world and security-critical scenarios, including facial recognition~\cite{sharif2016accessorize}, self-driving navigation~\cite{sitawarin2018darts}, and medical diagnosis~\cite{rajkomar2018scalable}. Despite these surprising advances, it has been known that DNNs suffer from severe security issues, such as privacy leakage~\cite{shokri2017membership}, adversarial attacks~\cite{szegedy2013intriguing, goodfellow2014explaining} and backdoor attacks (a.k.a. Trojan attacks)~\cite{gu2019badnets, chen2017targeted}. In particular, the backdoor attack is a technique of embedding a hidden malicious functionality into the DNN, which is activated only when a certain trigger appears.
This hidden functionality is usually the misclassification of an input sample to the attacker's desired target class, given the presence of a predefined trigger.
For example, a stop sign corrupted by a few pieces of tape will be recognized as a speed limit sign by the navigation system in self-driving cars, which may lead to fatal consequences~\cite{eykholt2018robust}.
BadNets~\cite{gu2019badnets} is one of the first works to study the threat of neural backdoors. After that, many variants of backdoor attacks have been proposed~\cite{chen2017targeted, liu2018trojaning, yao2019latent, jia2022badencoder}. Despite varying in mechanisms and scenarios, all these existing backdoor attacks are premised on adopting a universal (or sample-agnostic) trigger, \emph{i.e.,}\xspace different poisoned samples carry the same trigger. This uniform backdoor trigger becomes the Achilles' heel of the backdoor attacks. Based on the fact that the trigger is fixed and universal, existing defensive techniques~\cite{wang2019neural, liu2019abs, gao2019strip, chou2020sentinet, chen2019detecting} can easily reconstruct or detect the trigger according to the same behaviors among different poisoned samples.
For example, Neural Cleanse~\cite{wang2019neural} utilizes an optimization scheme to synthesize potential trigger patterns that can convert all benign images of other classes to a specific class. The synthesized trigger pattern with abnormally small norm is considered as the attack pattern used by the adversary. Additionally, a defender can also perform run-time detection on each input sample. To examine a malicious input, STRIP~\cite{gao2019strip} superimposes an input image to a set of randomly selected images and measures the entropy of the prediction outputs. If the predictions of the blending images are consistent (\emph{i.e.,}\xspace low entropy of prediction outputs), this input is regarded as a malicious one. In addition, SentiNet~\cite{chou2020sentinet} exploits the explanation technique (\emph{e.g.,}\xspace Grad-CAM~\cite{selvaraju2017grad}) to locate a potential trigger region by finding a highly salient contiguous region of a given input.
Witness to the success of the existing defenses, one might think that the threat of backdoor attacks is mitigated or neutralized.
Unfortunately, the crucial weakness of such static and sample-agnostic trigger became known to adversaries and they started exploring more advanced approaches in their attacks. In the new attack paradigms, backdoor triggers (referred to as \textit{dynamic}~\cite{nguyen2020input} or \textit{sample-specific}~\cite{li2021invisible} triggers) vary from sample to sample.
The success of existing defensive techniques~\cite{wang2019neural, gao2019strip, chou2020sentinet} mostly relies on the assumption that the triggers are sample-agnostic.
However, the sample-specific backdoor attacks break the fundamental assumption of the existing defensive techniques, as the dynamic backdoor introduces diverse information into the trigger pattern, which makes it harder for the defender to model the trigger.
As shown in Table~\ref{table: summary of existing defenses}, current backdoor defensive techniques mainly focus on universal backdoor attacks, leaving dynamic backdoor attacks as an unaddressed crucial threat to DNNs (see more discussions in Section~\ref{sec: Limitations against Dynamic Backdoors}).
Although a poisoned sample is misclassified to a target label, its intermediate representation has been shown to be different from those of the normal samples in the target class~\cite{tran2018spectral, chen2019detecting, tang2021demon}.
This observation provides an important indicator to distinguish the malicious samples from the normal ones.
However, when you zoom in with order information, out-of-distribution (OOD) samples present more details than trojaned ones since OOD detection requires much higher-order in screening out OOD samples~\cite{liang2017enhancing, lee2018simple, zisselman2020deep, sastry2020detecting}. This observation renders OOD detection methods to be less appealing to trojan detection as the OOD detection signal is too strong.
To further demystify the reasons, OOD detection requires sufficient and uncontaminated data as \textit{a priori} knowledge, which is not practical in backdoor detection. Moreover, although the Gram matrix based OOD detector achieves successful performance~\cite{sastry2020detecting}, our experimental results demonstrate that it lacks robustness (using fragile deviation metrics) and efficiency (computing an over-powerful Gramian, \emph{e.g.,}\xspace 10-order Gramian as used in the work~\cite{sastry2020detecting}, see details in Section~\ref{sec: Effectiveness Against Dynamic Backdoor}) in detecting trojaned inputs.
\noindent \textbf{Our work.}
In this paper, we show that Gramian information of dynamically trojaned data points is highly distinct from that of the benign ones.
Therefore, if we carefully design the order information (\emph{e.g.,}\xspace less than 10-order) and detection metrics, a Gram matrix could be an effective tool for backdoor detection.
Our method, \textit{\textbf{Beatrix}} (\underline{b}ackdoor d\underline{e}tection via gr\underline{a}m ma\underline{trix}), captures not only the feature correlations but also the order information of the intermediate representations to reveal subtle changes in the activation pattern caused by backdoor triggers.
Beatrix learns robust class-conditional statistics from the activation patterns of legitimate samples to effectively and efficiently harness the Gramian information in trojan detection.
In the presence of a backdoor attack, Beatrix can capture the anomalies in the activation patterns since the difference in the feature representations of poisoned samples and legitimate samples is highlighted by our detection metrics.
\noindent \textbf{Contributions.} Our main contributions are summarized as follows.
\begin{itemize}[leftmargin=*]
\item We present a comprehensive analysis and insights of main-stream defenses to unveil their limitations against dynamic backdoor attacks.
\item We develop and implement Beatrix, a novel approach to defend against backdoor attacks. Beatrix utilizes a statistically robust deviation measurement with Gramian information to capture the anomalies in the activation patterns induced by poisoned samples. Beatrix also leverages Regularized Maximum Mean Discrepancy to further improve the performance in identifying infected classes.
\item We demonstrate the effectiveness and robustness of our proposed method through theoretical justifications and extensive comparisons with state-of-the-art defensive techniques. We show that Beatrix can effectively detect sample-specific backdoor attacks and significantly outperform the existing defenses.
\end{itemize}
\section{Background}
\label{sec: Background}
In this section, we begin by briefly introducing the concept of Gram Matrix and the advances in backdoor attacks. We then discuss the limitations of existing defenses.
\subsection{Gram Matrix in DNNs}
Gramian information is widely used in areas such as Gaussian process regression~\cite{rasmussen2003gaussian} and style transfer learning~\cite{gatys2016image}. It computes the inner products of a set of $m$-dimensional vectors. The vectors, for instance, can be random variables in Gaussian process or vectorized internal activation patterns in style transfer.
Formally, we suppose $A:=\{a_k|a_k\in\mathbb{R}^{m}\}_{k=1}^{n}$ is a set of $m$-dimensional random variables, and then the gramian information between $a_i, a_j\in A$ is defined as $G_{ij} = \sum_{k}^{m} a_{ik} \cdot a_{jk}$.
Thus, the Gram matrix $G$ is an $n \times n$ symmetric matrix containing the gramian information between each pair of random variables in $A$.
Since the off-diagonal entries of $G$ represent the pairwise correlation between $a_{i}$ and $a_{j}$, Gram matrix can be used as an covariance matrix in Gaussian process regression \cite{rasmussen2003gaussian}. On the other hand, due to its effectiveness in feature learning, the Gram matrix shows remarkable performance in capturing stylistic attributes (\textit{e.g.}, textures and patterns) in neural activations~\cite{li2017demystifying}.
The high-order form of the Gram matrix has also been leveraged to improve OOD detectability~\cite{sastry2020detecting}.
The entries of $p$-th order of the Gram matrix is defined as $G_{ij}^{p} = \left(a_{i}^{p}{a_{j}^{p}}^{T}\right)^{1/p}$, where $p$ is the exponent.
\subsection{Backdoor Attacks}
Backdoor attacks are a technique of injecting some hidden malicious functionality into ML systems~\cite{pang2022trojanzoo, gao2020backdoor, liu2020survey}. The injected backdoor is activated only when a certain trigger appears in the input. This hidden functionality usually results in misclassifying the input sample into a target class predefined by the attacker.
\noindent \textbf{Universal (sample-agnostic) backdoor.~}
Although various backdoor attacks~\cite{gu2019badnets, chen2017targeted, yao2019latent, jia2022badencoder} have been proposed, the majority of them have a static trigger setting, meaning that there is only one universal trigger and any clean sample with that trigger will be misclassified to the target label~\cite{li2020rethinking}.
Particularly, in the most common backdoor attack (\textit{i.e}., BadNets~\cite{gu2019badnets}), an adversary can form a backdoor trigger $t=(m,p)$, where $m$ and $p$ denote the blending mask and the trigger pattern, respectively.
During the training of a DNN, a clean training sample pair such as $(x,y)$ is randomly replaced by the poisoned pair $(x_{bd}, y_{bd})$ with a certain probability using the trigger embedding function $\mathcal{B}$, which is defined as
\begin{align}
x_{bd} &= \mathcal{B}(x,t) \\
&= x\cdot(1-m) + p\cdot m
\end{align}
\noindent \textbf{Partial (source-specific) backdoor.~}
In the partial attack, only samples in a specific source class can activate the backdoor and be misclassified into the target class set by the trigger~\cite{wang2019neural, tang2021demon}. As for samples in other classes, the trigger will not activate the backdoor. It is worth noting that all the trojaned samples still share the same uniform trigger in the source-specific backdoor attack.
\noindent \textbf{Dynamic (sample-specific) backdoor.~}
Compared to the universal and partial backdoor attacks, dynamic backdoor attacks~\cite{salem2020dynamic, nguyen2020input, li2021invisible} make triggers that vary from sample to sample and this complicates the detection of such backdoors.
To the best of our knowledge, there exist three dynamic backdoor attacks~\cite{salem2020dynamic, nguyen2020input, li2021invisible}~\footnote{We include~\cite{li2021invisible} here because it is a sample-specific backdoor attack which is very similar to the input-aware dynamic attack of \cite{nguyen2020input}, though the authors do not use the term \textit{dynamic} in their paper~\cite{li2021invisible} explicitly.}. All of them utilize a trigger generating network to launch dynamic backdoor attacks.
In this work, in addition to universal backdoors, we try defending against the state-of-the-art dynamic ones, and more specifically, invisible sample-specific ~\cite{li2021invisible} and input-aware dynamic backdoor~\cite{nguyen2020input} attacks. (We do not consider~\cite{salem2020dynamic} since its triggers are not sample-specific and the code is not released.)
Both of them consider the uniqueness and exclusiveness of triggers, \textit{i.e.}, each sample has a unique trigger which is non-reusable for any other sample. Herein, we will brief their attack paradigms, and readers can find more details from the original papers~\cite{nguyen2020input, li2021invisible}.
Compared to the fixed and universal backdoor triggers, a \textit{dynamic} or \textit{sample-specific} trigger is a function of the corresponding input sample $x$. Suppose $g$ is a trigger generator. It can be defined as a function mapping an input from the sample space $\mathcal{X}$ to a trigger in the trigger space $\mathcal{T}$:
\begin{align}
g: & x\in\mathcal{X} \rightarrow t\in\mathcal{T}.
\end{align}
Additionally, the dynamic triggers should be non-reusable and unique. Henceforth,
\begin{equation}
\mathop{\arg\max}_{y^{*}} f_{y^{*}}(\mathcal{B}(x,g(\hat{x}))) = y_{bd}\cdot\mathds{1}(x = \hat{x}) + y\cdot\mathds{1}(x \neq \hat{x}),
\end{equation}
where $y_{bd}$ is an backdoor target and $y$ is the ground truth label of $x$. In other words, a clean sample $x$ with the trigger generated based on another sample will not activate the hidden backdoor in the model $f$.
Both dynamic backdoor attacks and adversarial attacks aim to make models misbehave and share many similarities. Still, they have certain differences.
Given a classifier $f$, an adversary injects dynamic backdoor by jointly training a trigger generation function $g$ with $f$ on a clean distribution $p_{data}$ and a poisoning distribution $p_{bd}$. The overall training objective is
\begin{equation}
\max_{g, f}\Pr_{(x,y_{bd})\sim p_{bd}}[f(x+g(x))=y_{bd}] + \max_{f}\Pr_{(x,y)\sim p_{data}}[f(x)=y].
\end{equation}
The training objective may recall a similar objective in targeted evasive (adversarial) attacks in which an adversary $\hat{g}: \max_{\hat{g}}\Pr[f(x+\hat{g}(x))=y_t]$. However, it is easily found that the dynamic backdoor attacker has the capability of training $g$ on the training dataset of $f$ while the adversarial attacker optimizes $\hat{g}$ either in a per-sample manner or over an external dataset whose distribution is similar to that of the training dataset. As a ramification, $f_{y_{bd}}(x+g(x)) \gg f_{y^*:y^*\neq y_{bd}}(x+g(x))$, which means $f$ tends to be overfitted to the data distribution modified by $g(x)$. Such attack was found hard to be detected by adversarial detection methods due to the statistics of $f$ has been changed and the high confidence in the misclassification of $x+g(x)$~\cite{athalye2018obfuscated}.
In addition, further modifying the triggers during the inference stage harms the stealthiness of the dynamic backdoor attacks and requires more capability from the attacker's side, which breaks the threat model of the dynamic backdoor. Such modification generates a separate adversarial attack rather than being a part of the dynamic backdoor attack, which falls out of the scope of this paper.
\subsection{Existing Defenses}
\label{sec: Limitations against Dynamic Backdoors}
\begin{table}[t]
\caption{\label{table:SoL}Limitations against Dynamic Backdoors}
\centering
\resizebox{1.\linewidth}{!}{%
\begin{tabular}{ll}
\toprule
\multicolumn{1}{c}{\textbf{Defense}} & \multicolumn{1}{c}{\textbf{Limitation}} \\
\midrule
Type-I & \begin{tabular}[c]{@{}l@{}}Perturbation-resistant assumption of triggers\end{tabular}\\
Type-II & \begin{tabular}[c]{@{}l@{}}Can only reconstruct sample-agnostic triggers\end{tabular} \\
Type-III & \begin{tabular}[c]{@{}l@{}}Strong assumption on the distribution of feature representations\end{tabular}\\
\bottomrule
\end{tabular}
}
\end{table}
As summarized in Table~\ref{table:SoL}, Type-\uppercase\expandafter{\romannumeral1} defenses assume that the backdoor trigger is resistant to perturbations. Thus, the trigger regions or trigger-carrying images can cause the same misclassification when overlaid on other clean images~\cite{gao2019strip, chou2020sentinet, doan2020februus}. However, this assumption is violated in dynamic backdoors where the trigger is only activated for a specific sample. Moreover, Type-\uppercase\expandafter{\romannumeral2} defenses try to reconstruct a universal trigger that can convert any clean sample to the same target class~\cite{wang2019neural, liu2019abs}. Unfortunately, this reconstructed trigger is only valid for sample-agnostic backdoor. In contrast, the dynamic backdoor triggers vary from sample to sample, rendering the reconstructed universal trigger to be totally different from the actual dynamic triggers. In addition, Type-\uppercase\expandafter{\romannumeral3} defenses attempt to distinguish the difference between the representations of clean samples and those of trojaned samples. However, they either use trivial clustering techniques~\cite{chen2019detecting, tran2018spectral} or model the representations with Gaussian distributions~\cite{tang2021demon, hayase2021spectre}, which cannot resist dynamic backdoor attacks. The detailed results can be found in our experimental analysis in Section~\ref{sec: comparison}.
\noindent \textbf{A case study of SCAn.}
Although SCAn~\cite{tang2021demon} reveals the drawbacks of current defenses relying on the sample-agnostic backdoor assumption, \textit{it only considers the partial backdoor but leaves the sample-specific attack as an open problem}. We argue that there are three limitations of SCAn, which indicates SCAn cannot be extended to defending against the sample-specific backdoor.
Firstly, SCAn assumes that the representations of normal and trojaned samples can be distinguished by the first moment (mean) discrepancy (\textit{Two-component decomposition assumption} in SCAn). However, in the dynamic backdoor attack, the first moment information becomes less discriminative.
Secondly, SCAn models the feature distribution by a Gaussian distribution under the Linear Discriminant Analysis (LDA)~\cite{mika1999fisher} assumption, \textit{i.e.}, different mean values but same covariance for the distributions of clean and trojaned feature representations (\textit{Universal Variance assumption} in SCAn).
However, our normality test shown in Figure~\ref{fig:normality_test} resonates with the observation of previous work~\cite{zisselman2020deep} that the feature space of DNNs does not necessarily conform with a Gaussian distribution.
Thirdly, as demonstrated by our theoretical analysis in Appendix~\ref{appendix: Theoretical Analysis of SCAn}, the essence of SCAn is to compute the weighted Mahalanobis distance between the representations of clean samples and those of trojaned samples, indicating that the effectiveness of SCAn is also dependent on the weighted value (the ratio of trojaned samples).
According to their estimate~\cite{tang2021demon}, SCAn needs to discern roughly 50 trojaned images before it can reliably detect further attacks. This is a severe drawback for security-critical applications. For example, the adversary may have bypassed an authentication system dozens of times before being caught.
\section{Overview and Motivation}
\label{sec: Overview}
To overcome the limitations of existing defenses, we propose a novel backdoor detection approach that covers the sample-specific backdoor attack. We first introduce the threat model in our work and then present our key observations and ideas.
Finally, we analyze the superiority of Gram matrix.
\subsection{Threat Model}
We consider the standard threat model which is consistent with that of the most recent backdoor attack and defense studies~\cite{gu2019badnets, wang2019neural, tang2021demon}.
\noindent \textbf{Adversary.~}
Similar to most backdoor poisoning settings, the goal of the adversary is to deliberately inject one or more backdoors into the target model.
The compromised model performs well on clean samples, whereas it misclassifies attack samples (trigger-carrying samples) to the predefined target label. We assume that the adversary can access the training set of the model, and is capable of poisoning the training data without major constraints, but has no direct access to the model. This scenario allows us to study the attack under worst-case conditions from the defender's point of view.
\noindent \textbf{Defender.~} The goal of the defender is to perform input-level detection to determine whether an input will trigger a malicious behavior in a untrusted model in an online setting such as the Machine-Learning-as-a-Service scenario. Furthermore, the defender aims to tell the infected classes of a backdoored model based on the instances it classifies.
We assume that the defender has white-box access to the target model, including the feature representation in the intermediate layers. Additionally, the defender needs a small set of clean data to help it with the detection, which was also a requirement in the previous works~\cite{chou2020sentinet,gao2019strip,tang2021demon, xu2021detecting}.
\subsection{Intuition and Key Idea}
\label{sec: Intuition and Key Idea}
\begin{figure}[t]
\centering
\includegraphics[width=0.3\textwidth]{image/Shapiro_Wilk_test_for_multivariate_normality.pdf}
\caption{\label{fig:normality_test} Normality Test by Shapiro-Wilk test. We can find that about 60\% features do \textbf{NOT} follow a normal distribution under a 95\% confidence score. This demonstrates that the Gaussian distribution assumption in \cite{tang2021demon,hayase2021spectre} is untenable in more advanced attacks such as the dynamic backdoor attack.}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.88\textwidth]{image/overview1.pdf}
\caption{\label{fig: workflow} An overview of Beatrix.}
\vspace{-7mm}
\end{figure*}
\noindent \textbf{Intuition.~} A key observation is that the clean samples of a certain class and the trojaned (trigger-carrying) samples targeted at that class are disjoint in the pixel space.
Consequently, even though a trojaned sample is misclassified into the target label, its intermediate representation is somehow different from those of normal samples of the target class.
The anomaly triggered by the trojaned samples can be characterized by inconsistencies between the intermediate feature representations and their predicted labels. This observation provides a basis for investigating the problem of characterizing trojaned samples from the perspective of OOD detection.
A predictive uncertainty study~\cite{amodei2016concrete} reveals that DNNs perform well on the samples drawn from the distribution seen in the training phase, but tend to behave unexpectedly when they encounter OOD samples that lie far from the training distribution.
Analogously, trojaned samples can be thought of as OOD samples drawn from a distinct distribution in contrast to the distribution of clean samples. Therefore, we believe there is a link between the detection of trojaned samples and the OOD detection~\cite{liang2017enhancing, lee2018simple, zisselman2020deep, sastry2020detecting}.
However, we notice there exist differences between OOD detection and backdoor detection.
First, the feature representations of OOD samples can be effectively modeled by Gaussian distributions. This assumption shows superior performance on OOD detection tasks~\cite{lee2018simple, zisselman2020deep}. On the contrary, Gaussian distributions are less capable of modeling the features of trojaned samples, due to the complexity and diversity of backdoor triggers. This dilemma is further emphasized by dynamic backdoors. As our experimental results show in Section~\ref{sec: comparison}, backdoor detection methods that use Gaussian distribution, \textit{e.g.}, SCAn~\cite{tang2021demon}, achieve suboptimal performance in the identification of dynamic backdoors. A normality test in Figure~\ref{fig:normality_test} also exposes the problem.
Second, less attention has been paid to adversarial robustness in OOD detection methods. \textit{A priori} knowledge of in-distribution/clean samples is a key piece of information for many OOD detectors~\cite{lee2018simple, zisselman2020deep, sastry2020detecting} and backdoor detection methods~\cite{gao2019strip, wang2019neural}. The OOD detection task assumes that a set of clean samples used for training the detector can be well curated. However, this assumption is challenged in the scenario of backdoor detection since poisoned samples may carry invisible triggers~\cite{li2021invisible, shafahi2018poison} which can hardly be filtered even by manual inspection.
The lack of adversarial robustness restricts the deployment of OOD detectors in detecting trojaned inputs (see~\ref{sec: Effectiveness Against Dynamic Backdoor} in more detail).
Finally, OOD detectors require sufficient in-distribution data, or even OOD data~\cite{liang2017enhancing, lee2018simple, zisselman2020deep}. This requirement ensures that the parameters of the detection models can be effectively estimated. However, there are usually limited clean data available to the backdoor defender. Thus, the requirements for statistical robustness under limited data are different in OOD and backdoor detection methods.
\noindent \textbf{The key idea.~}
From our observation, backdoor defense can be viewed as the problem of detecting OOD. We argue that the problem of finding a robust detector for neural backdoors can be connected to feature modeling methods used in areas such as Gaussian process regression, style transfer, and OOD detection~\cite{rasmussen2003gaussian, gatys2016image, sastry2020detecting}. In this paper,
we find that although the trojaned and clean samples are deeply fused in the original feature space, they are distinguishable in the Gramian feature space, indicating Gram matrix is an effective tool for feature modeling. The Gram matrix derives from the inner products of feature maps across different channels.
Thus, Gram matrices not only consider features in each individual channel but also incorporate the feature correlations across channels~\cite{sastry2020detecting,li2017demystifying}.
We note that the previous representation-based backdoor detection methods either use trivial clustering techniques~\cite{chen2019detecting, tran2018spectral} or Gaussian models~\cite{tang2021demon, hayase2021spectre} in the detection.
These methods ignore the high-order information and only consider the first moment (mean) discrepancy between clean and trojaned samples. However, the simplification reduces the discriminative power of the methods against more complex attacks such as dynamic backdoors.
To tackle the problem, we turn to high-order statistics of the feature representations since they are more sensitive to the changes in the feature space, due to its higher-power format. Our method employs not only first-order moments but also high-order moments for feature modeling.
We utilize the Gram matrix and its appropriately high-order forms to capture not only the feature correlations but also appropriately high-order information to detect trojaned samples.
In addition, considering the adversarial robustness and statistical robustness in backdoor detection, we do not use the multivariate Gaussian to model the trojaned samples in the deviation measurement as Gaussian models only perform well with sufficient data~\cite{psutka2015sample}. Instead, we utilize Median Absolute Deviation (MAD), a more robust estimation of statistical dispersion, to measure the deviation of trojaned samples.
Finally, when the training dataset of a given class is contaminated, the set of examples can be viewed as a mixture of two subgroups~\cite{chen2019detecting, tran2018spectral, tang2021demon, hayase2021spectre}. In contrast to previous works~\cite{tang2021demon, hayase2021spectre} assuming that the two subgroups follow Gaussian distributions with two different means but the same covariance, we employ Regularized Maximum Mean Discrepancy (RMMD)~\cite{danafar2013testing} to enhance the adversarial robustness of our method. RMMD is a Kernel-based two-sample testing technique which does not have any assumption on the distributions. RMMD performs a hypothesis test on whether the feature representations in a given class are drawn from a mixture group (\textit{i.e.}, contaminated class) or a single group (\textit{i.e.}, uncontaminated class).
\subsection{Theoretical Analysis of Gram Matrix}\label{appendix: Theoretical Analysis of Gram Matrix}
Let $x^p\sim P_{c,p}$ and $y^p\sim P_{t,p}$ be $p$-th order representation vectors sampled from the clean distribution $P_{c,p}$ and the trojaned distribution $P_{t,p}$, respectively. Beatrix models the feature representations without any assumption on $P_{c,p}$ and $P_{t,p}$. Instead, Beatrix relies on Gram matrix to extract discriminative information of $P_{c,p}$ and $P_{t,p}$ from their statistical moments~\cite{bishop2006pattern}.
Let $u^p_1$ and $S^p_1$ be the mean vector and the covariance matrix of $x^p$, respectively. The second raw moment of $x^p$ over $P_{c,p}$ is:
\begin{align}
E({x^p}{x^p}^T) &= E({x^p})E({x^p})^T + E[({x^p}-u^p_1)({x^p}-u^p_1)^T] \\\nonumber
&= u^{p}_{1}{u^p_{1}}^T + S^p_1.
\end{align}
The Gramian feature of ${x^p}$ is defined as $G_{x^p} = {x^p}{x^p}^T$. Then, $E(G_{x^p}) = E({x^p}{x^p}^T) = u^p_{1}{u^p_{1}}^T + S^p_1$.
Similarly, let $u^p_2$ and $S^p_2$ be the mean vector and the covariance of $y^p$, we have $E(G_{y^p}) = u^p_{2}{u^p_{2}}^T + S^p_2$
over $y^p\sim P_t$.
Therefore, the expected discriminative information captured by the Gram matrices over different exponents can be represented from the prospective of statistical moments:
\begin{align}
\label{equ: moment discrimination captured by Gram matrix}
M(P_c,P_t) &= E_{p\in\mathbb{Z^+}} [E(G_{x^p}) - E(G_{y^p})]\\\nonumber
&= E_{p\in\mathbb{Z^+}} [u^{p}_{1}{u^p_{1}}^T - u^{p}_{2}{u^p_{2}}^T + S^p_1 - S^p_2],
\end{align}
where $P_c$ is a collection of distributions of clean representations with elements of different powers and $P_t$ is that of trojaned representations.
Equation~\ref{equ: moment discrimination captured by Gram matrix}
shows that Gramian features not only capture the first moment discrepancy (\textit{i.e.}, $u^p_{1}{u^p_{1}}^T - u^p_{2}{u^p_{2}}^T$) like SCAn (when $p=1$, see Equation~\ref{equ: scan likelihood ratio}), but also the second moment discrepancy (\textit{i.e.}, $S^p_1 - S^p_2$).
Compared to previous methods such as SCAn, a trojaned $y$ can still be distinguished from clean representations when the clean and trojaned distributions have the same mean. Moreover, by considering various $p$ values, the second moment discrepancy can capture more information about high-order features and better models $M(P_c,P_t)$.
\section{Design of Beatrix}
\label{sec: Design}
In this section, we provide the details of our approach to detecting backdoor attacks. The framework of Beatrix is illustrated in Figure~\ref{fig: workflow}.
\subsection{Feature Modeling via Gram Matrices}\label{subsec:gramian}
Formally, let $h$ be the sub-model up to the $l$-th layer of a DNN model. Then, the feature representation of the input sample $x$ at the $l$-th layer of the DNN model is defined as $h(x) = v \in \mathbb{R}^{n \times m}$, where $n$ is the number of channels at layer $l$ and $m$ is the height times the width of each feature map. The features correlations between channels can be expressed by
\begin{align}
G = vv^{T},
\end{align}
where $G \in \mathbb{R}^{n\times n}$ denotes the Gram matrix of the feature maps in an inner product space.
In order to capture more prominent activations in feature maps, we also use the high-order Gram matrix
\begin{align}
G^{p} = \left(v^{p}{v^{p}}^{T}\right)^{1/p},
\end{align}
where $v^{p}$ denotes the $p$-th power of the feature representation~$v$, and $G^{p}$ is the $p$-th order Gram matrix of $v$.
The off-diagonal entries of $G^{p}$ represent the pairwise correlation between feature maps at the $l$-th layer while the entries at the diagonal only relate to a single feature map.
Since the matrix $G^{p}$ is symmetric, we only need the upper (or lower) triangular part of it. In particular, the vectorized triangular matrix which contains the entries on and above (or below) the main diagonal, can form a $\frac{1}{2}n(n+1)$-dimensional vector like $\Vec{G^{p}}$.
We can compute $\Vec{G^{p}}$ for each order $p \in \{1,...,P\}$, where $P$ is a hyperparameter representing the bound of the order.
By concatenating all the output vectors $\Vec{G^{p}}$, we can derive a new representation vector $s = [\Vec{G^{1}},\ \Vec{G^{2}}, \ ..., \ \Vec{G^{P}}] \in \mathbb{R}^{\frac{1}{2}n(n+1)P}$ for the input sample $x$.
Let $\mathcal{X}_t$ denote a set of clean samples in class $t$. $\mathcal{X}_t$ has a feature presentation set $\mathcal{V}_t = \left\{ v_i:= h(x_i),\ x_{i} \in \mathcal{X}_t \right\}$ and a concatenated Gramian feature set $\mathcal{S}_t = \left\{s_i,\ i \in \{1,2,...,|\mathcal{X}_t|\} \right\}$. The task of backdoor detection can be formulated as an outlier detection problem: given the feature $\hat{s}$ of an input sample $\hat{x}$ and its predicted label $\hat{y_t}$ by the target model $f$, we try to determine whether $\hat{s}$ is an outlier, with respect to $\mathcal{S}_t$ based on the statistical properties of $\mathcal{S}_t$.
\subsection{Deviation Measurement}\label{subsec:dm}
A natural choice of computing the deviation of a point $\hat{s}$ is to build a multivariate Gaussian model of $\mathcal{S}_t$. However, the problem is non-trivial because of \textit{i)} the large dimensionality of the feature vector $s$ and \textit{ii)} the limited number of clean samples for estimating Gaussian parameters (especially, the covariance matrix).
Therefore, building a multivariate Gaussian model for high dimensional variables is not statistically robust when there are limited data samples available.
Additionally, since the feature modeling with Gram matrices has already considered the feature correlations, we can just simplify the problem and model each element in $\mathcal{S}_{t}$ independently. Thus, high dimensional estimation can be simplified into one dimensional estimation for each independent element.
In the simplified case, one may still consider using the Gaussian model but its univariate version to estimate mean and standard deviation of the features. However, recall that there is limited clean data available for the defender, the estimation results (\textit{i.e.}, mean and standard deviation) are easier to be affected by the outliers as Gaussian models only perform well with sufficient data~\cite{psutka2015sample}. More importantly, the individual elements of $s$ may not follow a Gaussian distribution strictly.
Instead of using a univariate Gaussian model, we propose to utilize Median Absolute Deviation (MAD) which is known to be more resilient to outliers in a dataset than the standard deviation $\hat{\sigma}$~\cite{leys2013detecting}. The absolute deviations between all data points and their medians are gained before MAD is employed as the median of these absolute deviations.
Given the set of concatenated Gramian features $\mathcal{S}_t$ of all clean samples $\mathcal{X}_t$ in class $t$, we can compute the median and the MAD with respect to each concatenated Gramian vector:
\begin{align}
\tilde{s}_j =& \ median(\{s_{ij},\ \forall i\in\{1,2,...,|\mathcal{X}_t|\}\}), \\
MAD_{j} =&\! \ median(\{|s_{ij}\!-\!\tilde{s}_j|,\!\ \forall i\!\in\!\{1,2,...,|\mathcal{X}_t|\}\}).
\end{align}
Then the deviation of the observed $j$-th value $\hat{s}_j$ in the candidate feature point $\hat{s}$ is defined as:
\begin{align}
\delta_j &= \delta(\hat{s}_j) \\
&=\left\{
\begin{aligned}
0\quad\quad\quad,&\quad if \quad min \leq \hat{s}_j \leq max, \\
\frac{min-\hat{s}_j}{min},&\quad if \quad \hat{s}_j \leq min, \\
\frac{\hat{s}_j-max}{max},&\quad if \quad max \leq \hat{s}_j,
\end{aligned}
\right.
\end{align}
where $min=\tilde{s}_j - k\cdot MAD$, $max=\tilde{s}_j + k\cdot MAD$, and $k$ is a predefined scale factor that is set to 10 in our case.
Then the deviation of the candidate feature point $\hat{s}$ is the sum of the deviation values over all entries in $\hat{s}$:
\begin{align}
\delta = \frac{2}{n(n+1)P}\sum_{j=1}^{\frac{1}{2}n(n+1)P} \delta_j.
\end{align}
\vspace{-1mm}
\noindent \textbf{Threshold determination.~}
The detection boundary of Beatrix is estimated by benign inputs. Due to the limited number of clean data available to the defender, we employ bootstrapping to compute the deviations of benign inputs~\cite{efron1994introduction}. Specifically, we randomly draw $\frac{1}{T}$ samples from the clean dataset as testing samples. The remaining samples are used as training samples to estimate the min/max values. The procedure is repeated for $T$ iterations to obtain the deviations of benign samples.
The detection boundary can be determined by the defender when choosing different percentiles like STRIP~\cite{gao2019strip}.
For example, the defender can choose 95\% as the detection boundary. This means that 95\% of the deviations of benign samples are less than this detection boundary.
\subsection{Identifying Infected Labels}
The performance of Beatrix can be further improved through a local refinement of the detection results to reduce false positives. Since the detection threshold in Section~\ref{subsec:dm} is predefined, there could be false positive in the detection results. In an offline setting, Beatrix accumulates the historical detection results for a second statistical analysis to ablate false trojaned targets given by the threshold thereof.
In the presence of a backdoor attack, the feature representations of samples in the infected class can be considered as a mixture of two subgroups~\cite{chen2019detecting, tran2018spectral, tang2021demon, hayase2021spectre}. However, previous works~\cite{tang2021demon, hayase2021spectre} assume that these two subgroups follow Gaussian distributions with two different means but the same covariance. Therefore, they perform a hypothesis testing to determine whether these two distributions are significantly different. However, as we discussed in Section~\ref{sec: Limitations against Dynamic Backdoors}, the Gaussian assumption is not tenable in more complex scenarios, such as dynamic backdoor attacks.
Therefore, we resort to the Kernel-based two-sample testing which addresses whether two sets of samples are identically distributed without assumption on their distributions~\cite{gretton2012kernel, danafar2013testing, chwialkowski2015fast, varoquaux2019comparing}.
A popular test statistic for this problem is the Maximum Mean Discrepancy (MMD)~\cite{gretton2012kernel}, which is defined based on a positive definite kernel function $k$~\cite{scholkopf2002learning}. Kernel methods provide the embedding of a distribution in a reproducing kernel Hilbert space (RKHS). For MMD with a linear kernel, $k(x, x^{\prime}) = \langle x, x^{\prime}\rangle$, it measures the distribution distance under their first moment discrepancy~\cite{muandet2016kernel}. In practice, a common option is to use the Gaussian kernel $k(x, x^{\prime}) = exp(-\beta\Vert x - x^{\prime} \Vert ^{2}_{2})$, which contains infinite order of moments by looking at its Taylor series~\cite{li2015generative}.
In this work, we use an extension of MMD metric, termed Regularized MMD (RMMD), which incorporates two penalty terms to achieve better performance when the two sample sets are small and imbalanced~\cite{danafar2013testing}:
\vspace{-1mm}
\begin{small}
\begin{align}
RMMD(P,Q) \!&=\! MMD(P, Q)^{2} \!-\! \lambda_{P} \Vert\mu_{P}\Vert_{\mathcal{H}}^{2} \!-\! \lambda_{Q} \Vert\mu_{Q}\Vert_{\mathcal{H}}^{2} \\
& =
\! \Vert\mu_{P} - \mu_{Q}\Vert_{\mathcal{H}}^{2} \!-\! \lambda_{P} \Vert\mu_{P}\Vert_{\mathcal{H}}^{2} \!-\! \lambda_{Q} \Vert\mu_{Q}\Vert_{\mathcal{H}}^{2},
\end{align}
\end{small}
where $P$ and $Q$ denote the representation distributions represented by the samples in the two subgroups obtained from the deviation measurement.
According to the work~\cite{danafar2013testing}, this test statistic follows an asymptotic normal distribution based on theorems in~\cite{hoeffiding1948class, serfling2009approximation}. Similar to Neural Cleanse~\cite{wang2019neural} and SCAn~\cite{tang2021demon}, we can leverage the Median Absolute Deviation to identify the infected label(s) with abnormally large values of RMMD statistic instead of directly computing the p-value of this test.
Specifically, we denote $R_{t}$ as the RMMD statistic of class $t$, and then anomaly index $R_{t}^{*}$ is defined as:
\vspace{-1mm}
\begin{align}
R_{t}^{*} &= |R_{t} - \tilde{R}|/(MAD(\tilde{R})*\eta), \\
\text{where} \quad \tilde{R} &= median(\{R_{t}: t \in \mathcal{L}\}),\\
MAD(\tilde{R}) &= median(\{R_{t} - \tilde{R}|: t \in \mathcal{L}\}).
\end{align}
As $R_{t}$ follows an asymptotic normal distribution, we apply a constant factor $\eta = 1.4826$ to the anomaly index. We identify any label with an anomaly index $R_{t}^{*} \geq e^{2}$ as an infected label with the confidence probability $\geq (1 - 10^{-9})$~\cite{tang2021demon}.
\begin{table*}[t]
\centering
\caption{\label{tabel: information of benigh models} Detailed information about dataset, model architecture and clean accuracy.}
\resizebox{0.98\textwidth}{!}{
\begin{tabular}{ccccccc}
\toprule
\textbf{Dataset} & \# \textbf{of Classes} & \# \textbf{of Training Images} & \# \textbf{of Testing Images} & \textbf{Input size} & \textbf{Model Architecture} & \textbf{Top-1 accuracy} \\ \midrule
CIFAR10 & 10 & 50000 & 10000 & 32 $\times$ 32 $\times$ 3 & PreActResNet18 & 94.5\% \\
GTSRB & 43 & 39209 & 12630 & 32 $\times$ 32 $\times$ 3 & PreActResNet18 & 99.1\% \\
VGGFace & 100 & 38644 & 9661 & 224 $\times$ 224 $\times$ 3 & VGG16 & 90.1\% \\
ImageNet & 100 & 50000 & 10000 & 224 $\times$ 224 $\times$ 3 & ResNet101 & 83.8\% \\
\bottomrule
\end{tabular}}
\vspace{-3mm}
\end{table*}
\section{Evaluation}
\label{sec: Evaluation}
In this section, we first evaluate the effectiveness of our proposed method against the dynamic backdoor and then compare Beatrix with state-of-the-art defensive techniques. Finally, we also demonstrate the robustness of Beatrix against other attacks. The datasets and model structures used in our experiments are summarized in Table~\ref{tabel: information of benigh models}. We provide detailed introduction of the experiment setup in Appendix~\ref{sec: Experiment Setup}.
\subsection{Effectiveness Against Dynamic Backdoor}
\label{sec: Effectiveness Against Dynamic Backdoor}
\noindent \textbf{Attack configuration.~} We implement the input-aware dynamic backdoor attack~\cite{nguyen2020input} using the code released by the authors~\cite{codeinputaware}. Figure \ref{fig: examples of dynamic} illustrates several examples of poisoned samples.
We conduct the common single-target attack, \textit{i.e.}, the target label is the same for all trojaned samples. We set both the backdoor probability (trojaned samples with their paired triggers) and the cross-trigger probability (trojaned samples with inconsistent triggers) as 0.1.
For all the four datasets, the backdoor attack success rates (ASR) are almost 100\% while still achieving a comparable performance on clean data as the benign models do, as shown in Table~\ref{tabel: accuracy of infected models}. It is worth noting that the cross-trigger accuracy (the accuracy of classifying images containing dynamic triggers deliberately generated for other images)
is over 80\% on all the four datasets, and this shows the nonreusability and uniqueness of the triggers on mismatched clean images. The evaluation results on the invisible sample-specific backdoor attack~\cite{li2021invisible} are shown in Section~\ref{sec: Robustness Against Other Attacks}.
\begin{figure}[t]
\centering
\subfigure[]{
\begin{minipage}[t]{0.065\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/GTSRB_dynmaic_clean.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.065\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/GTSRB_dynmaic_poison.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.065\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/ImageNet_dynmaic_clean.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.065\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/ImageNet_dynmaic_poison.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.065\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/VGGFace_dynmaic_clean.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.065\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/VGGFace_dynmaic_poison.png}
\end{minipage}}
\caption{\label{fig: examples of dynamic} Examples of poisoned samples under the input-aware dynamic backdoor. (a), (c) and (e) are clean images, (b), (d) and (f) are poisoned images.}
\end{figure}
\begin{table}[t]
\centering
\caption{\label{tabel: accuracy of infected models} Attack success rate, cross-trigger accuracy and classification accuracy of infected models.}
\resizebox{\linewidth}{!}{
\begin{tabular}{ccccc}
\toprule
\multirow{2}{*}{\textbf{Dataset}}& \multicolumn{3}{c}{\textbf{Infected Model}} & \textbf{Benign Model} \\
\cmidrule(r){2-4}\cmidrule(r){5-5}
& \begin{tabular}[c]{@{}c@{}}\textbf{Attack Success}\\ \textbf{Rate}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\textbf{Cross-trigger}\\ \textbf{Accuracy}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\textbf{Clean}\\ \textbf{Accuracy}\end{tabular} & \begin{tabular}[c]{@{}c@{}}\textbf{Clean}\\ \textbf{Accuracy}\end{tabular} \\
\midrule
CIFAR10 & 99.4\% & 88.6\% & 93.9\% & 94.5\% \\
GTSRB & 99.7\% & 96.1\% & 99.2\% & 99.1\% \\
VGGFace & 98.5\% & 82.5\% & 89.8\% & 90.1\% \\
ImageNet & 99.5\% & 81.3\% & 83.5\% & 83.8\% \\
\bottomrule
\end{tabular}}
\end{table}
\begin{figure*}[t]
\centering
\begin{minipage}[t]{0.24\linewidth}
\centering
\includegraphics[width=0.98\linewidth]{image/detection_results_anomaly.png}
\caption{\label{fig: detection results anomaly} The logarithmic anomaly index of infected labels on the four datasets.}
\end{minipage}
\quad
\begin{minipage}[t]{0.72\linewidth}
\centering
\includegraphics[scale=0.59]{image/detection_results.pdf}
\caption{\label{fig: detection results} Deviation distribution of benign and trojaned samples. The trojaned sample shows a much larger deviation than benign samples. The color boundary in the background indicates the decision threshold (same for the figures in the following sections).}
\end{minipage}
\vspace{-6mm}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[scale=0.43]{image/clean_data_for_deviation.pdf}
\caption{\label{fig: clean_data_for_deviation} The logarithmic anomaly index of infected labels when using different number of clean data.
}
\end{minipage}
\quad
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[scale=0.43]{image/clean_data_for_deviation_under_contamination.pdf}
\caption{\label{fig: clean_data_for_deviation_under_contamination} The logarithmic anomaly index of infected labels when clean data is contaminated.}
\end{minipage}
\quad
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[scale=0.43]{image/order_for_deviation.pdf}
\caption{\label{fig: order_for_deviation} False positive rate of benign images when incorporating different bound on the order of Gram matrix.
}
\end{minipage}
\vspace{-7mm}
\end{figure*}
\noindent \textbf{Effectiveness on various datasets.~} From each dataset, we randomly select 30 images per class as a clean dataset for the defender. The clean dataset accounts for no more than 6\% of the whole dataset.
The bound of the order of Gram matrix is set as 9. As is shown later, the detection effectiveness is stable when $P \geq 4$.
The experimental results on the four datasets show that our defensive technique is very effective in detecting the dynamic backdoor attack. Figure~\ref{fig: detection results anomaly} illustrates the logarithmic anomaly index values $ln(R^{*})$ of infected and uninfected labels. All infected labels have much larger anomaly index values, compared to uninfected labels. This demonstrates that our defensive technique can effectively detect target classes in infected models on various datasets and model architectures. Moreover, Figure~\ref{fig: detection results} illustrates that the deviations of the trojaned samples are larger than those of the benign ones. This demonstrates that our method can also effectively distinguish benign from polluted samples.
\noindent \textbf{Clean data for deviation measurement.~}
To achieve a high accuracy in discriminating the mixed representations of benign and poisoned samples, a small set of clean samples is required for estimating the threshold in advance~\cite{chou2020sentinet,gao2019strip,tang2021demon, xu2021detecting}. The above experiments show that our method can effectively detect the dynamic attack with 30 clean images per class. Our study further demonstrates that our method can perform effectively with less clean data and even with the contaminated data. As shown in Figure~\ref{fig: clean_data_for_deviation}, we can find that even with only 8 clean images, Beatrix can still accurately identify the infected class. Moreover, we also test the robustness of Beatrix when the clean data is moderately contaminated. Figure~\ref{fig: clean_data_for_deviation_under_contamination} shows that Beatrix is still effective when no more than 16\% (or 5 images) of the clean images per class are contaminated with poisoned ones. We also compare Beatrix with the OOD detection method~\cite{sastry2020detecting} under this data contaminating scenario. Figure~\ref{fig: ood vs beatrix} shows that Beatrix is much more robust than the OOD detector when the clean data is contaminated by trojaned samples. As a result, the OOD detection method~\cite{sastry2020detecting} cannot be directly applied to the backdoor detection.
\noindent \textbf{The order of Gram matrix.~}
In the above experiments, we consider Gram matrices from the first to the ninth order. However, incorporating high-order information induces much more computational overhead. Particularly, this overhead is of vital importance for the online scenario. Thus, it is crucial to choose an appropriate
set of orders
to achieve a better trade-off between detection effectiveness and computational complexity. Specifically, we construct the detector with different values of $P$ and evaluate them in an online setting. Figure~\ref{fig: order_for_deviation} illustrates that the testing time for an input sample increases from 8.9$\times 10^{-6}$s to 78.5$\times 10^{-6}$s when the order bound increases from 1 to 9. Additionally, we find that the detection capability (false positive rate) of our method is stabilized when $P \geq 4$.
We note that the OOD detector \cite{sastry2020detecting} needs much higher-order Gram matrices (\textit{i.e.}, $P=10$) to discriminate between in-distribution and out-of-distribution datasets, especially when the two datasets are similar-looking.
This experimental result shows that, considering the computational efficiency, it is sufficient to employ up to the third or the fourth order information to capture discriminative characteristics of benign and malicious inputs. Therefore, for the remaining experiments, we set $P$ as~4 (see efficiency comparison in Section~\ref{sec: comparison}).
\noindent \textbf{Defending against all-to-all attack.~} Here, we consider another type of adversary who launches all-to-all attacks~\cite{nguyen2020input}. Specifically, for a $c$-way classifier, the trojaned samples originally in the $i$-th class are misclassified into the (($i$+1) mod $c$)-th class. Since samples from all classes are infected by the all-to-all attack, the anomaly index values $R_{t}^{*}$ are no longer effective. However, as shown in Figure~\ref{fig: all2all}, most of the RMMD statistics $R_{t}$ of different infected labels in the all-to-all infected models are much larger than those of uninfected labels in the single-target attack. This demonstrates that Beatrix can still effectively defend against all-to-all attack relying on the RMMD statistics. In addition, we note that the all-to-all attack is the worst case of multi-target attacks that all labels are infected. However, the more labels that are infected, the less stealthy and lower performance the backdoor attack shows, especially for large datasets. As discussed in the recent research~\cite{tang2021demon}, when half of the labels are infected on the 1000-class ImageNet model, its clean accuracy and attack success rate drops 5\% and 41\%, respectively. Additionally, we consider the all-to-all attack with dynamic backdoors, which is much stronger than the all-to-all attack with universal backdoors~\cite{xu2021detecting}.
\begin{figure*}[t]
\centering
\begin{minipage}[t]{0.26\linewidth}
\centering
\includegraphics[width=0.95\linewidth]{image/ood_vs_beatrix.pdf}
\caption{\label{fig: ood vs beatrix} Robustness comparison between Beatrix and the OOD detection method \cite{sastry2020detecting}.}
\end{minipage}
\quad
\begin{minipage}[t]{0.68\linewidth}
\centering
\includegraphics[width=1\textwidth]{image/all2all.pdf}
\caption{\label{fig: all2all} RMMD statistics $R_{t}$ of different labels in all-to-all infected models (red) and in single-target (label 2 is target) infected models (blue). Although the anomaly index $R_{t}^{*}$ may not be effective for the all-to-all attack, their RMMD statistics $R_{t}$ are much larger than those of uninfected labels in the single-target attack.}
\end{minipage}
\vspace{-8mm}
\end{figure*}
\subsection{Comparison}
\label{sec: comparison}
In this subsection, we compare our method with several state-of-the-art methods under two scenarios that are used in~\cite{tang2021demon}. In the offline protection setting, a dataset containing benign and malicious samples are processed at once, and the defender is supposed to determine whether a data class is benign or infected. On the other hand, in the online setting, samples are processed one by one, and the defender is supposed to determine whether a sample is legitimate or malicious.
In the offline setting, we launch the dynamic backdoor attack~\cite{nguyen2020input} to generate 43 infected models with respect to the 43 classes in GTSRB and 10 infected models for CIFAR10. Additionally, we launch the conventional backdoor attack~\cite{gu2019badnets}
to generate 172 (43$\times$4) infected models for GTSRB and 40 (10$\times$4) infected models for CIFAR10 with four different static triggers (Figure \ref{fig: examples of universal triggers}) which are also used in~\cite{chou2020sentinet, gao2019strip, wang2019neural, tang2021demon}. For the online setting, we randomly select 4,000 samples from each dataset as testing samples, half of which carry dynamic (or static) triggers.
\begin{table}[t]
\centering
\caption{\label{tabel: comparison_on_dynamic_backdoor} Defense performance against dynamic backdoors on GTSRB and CIFAR10.}
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{cccccccc}
\toprule
\multirow{2}{*}{\textbf{Scenario}} & \multirow{2}{*}{\textbf{Method}} & \multicolumn{3}{c}{\textbf{GTSRB}} & \multicolumn{3}{c}{\textbf{CIFAR10}} \\
\cmidrule(r){3-5}\cmidrule(r){6-8}
& & \textbf{REC}(\%) & \textbf{PRE}(\%) & \textbf{F1} (\%) & \textbf{REC}(\%) & \textbf{PRE}(\%) & \textbf{F1}(\%) \\
\midrule
\multirow{5}{*}{Offline}
& NC & 18.6 & 9.3 & 12.4 & 40.0 & 25.0 & 30.8 \\
& ABS & 27.0 & 41.5 & 32.7 & 37.0 & 49.3 & 42.3 \\
& MNTD & 55.8 & 91.8 & 69.4 & 57.4 & 63.8 & 60.4 \\
& AC & 74.4 & 6.0 & 11.1 & 80.0 & 28.26 & 42.1 \\
& SCAn & 44.2 & 31.7 & 36.9 & 40.0 & 57.1 & 47.1 \\
& Beatrix & \textbf{95.3} & \textbf{87.2} & \textbf{91.1} & \textbf{100.0} & \textbf{83.3} & \textbf{90.9} \\
\midrule
\multirow{4}{*}{Online}
& STRIP & 23.0 & 41.9 & 29.7 & 20.9 & 43.9 & 28.3 \\
& SentiNet & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 \\
& SCAn & 83.3 & 68.7 & 75.3 & 28.6 & 39.1 & 33.0 \\
& Beatrix & \textbf{99.8} & \textbf{99.8} & \textbf{99.8} & \textbf{99.0} & \textbf{95.4} & \textbf{97.2} \\
\bottomrule
\end{tabular}}
\end{table}
\subsubsection{Offline defense}
In the offline setting, we consider four competitive methods, namely Neural Cleanse (NC)~\cite{wang2019neural}, ABS~\cite{liu2019abs}, MNTD~\cite{xu2021detecting}, Activation Clustering (AC)~\cite{chen2019detecting}, and SCAn~\cite{tang2021demon}. We re-implemented AC according to the paper~\cite{chen2019detecting} since the source code is not publicly available. We also re-implemented SCAn using TensorFlow Probability (TFP) based on the original MATLAB version~\cite{codescan}. Moreover, we used the PyTorch version of NC~\cite{codeinputaware}, ABS~\cite{codeabs} and MNTD~\cite{codemntd}. The comparison results are presented in Tables~\ref{tabel: comparison_on_dynamic_backdoor} and~\ref{tabel: comparison_on_conventional_backdoor}.
The results indicate that our proposed method largely outperforms all four benchmark methods against dynamic backdoors while slightly outperforming them against universal backdoors.
NC leverages an optimization-based reverse engineering approach to find a trigger pattern that causes any benign input from other classes to be misclassified into a target label. However, in dynamic backdoor attacks, triggers are unique and non-reusable instead of being static and universal. As a result, the reverse-engineered triggers obtained by NC are visually and functionally different from the actual dynamic triggers.
As for the 43 trojaned models infected by the dynamic backdoor on GTSRB, there are $43\times1=43$ poisoned classes (positives) and $43\times42=1806$ benign classes (negatives). The experimental results show that NC is not effective against dynamic backdoor attacks, achieving 18.6\% (8/43) recall, and 9.3\% (8/86) precision and 12.4\% F1-score in GSTRB.
AC utilizes a two-class clustering method to separate the benign and malicious samples based on their feature vectors (activations). Specifically, AC performs dimensionality reduction using Independent Component Analysis (ICA), and then clusters them using 2-means. A high silhouette score~\cite{rousseeuw1987silhouettes} of the clustering results indicates the class is infected because the two clusters obtained by 2-means do fit the data well. However, under the dynamic backdoor attack, the feature presentations become less distinguishable so that AC cannot effectively separate the activations of benign and trigger-carrying samples. Therefore, AC achieves 92.5\% F1-score against universal backdoor attacks whereas it yields only 11.1\% F1-score against dynamic backdoor attacks on GTSRB.
\begin{table}[t]
\centering
\caption{\label{tabel: comparison_on_conventional_backdoor} Defense performance against universal backdoors on GTSRB and CIFAR10. }
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{cccccccc}
\toprule
\multirow{2}{*}{\textbf{Scenario}} & \multirow{2}{*}{\textbf{Method}} & \multicolumn{3}{c}{\textbf{GTSRB}} & \multicolumn{3}{c}{\textbf{CIFAR10}} \\
\cmidrule(r){3-5}\cmidrule(r){6-8}
& & \textbf{REC}(\%) & \textbf{PRE}(\%) & \textbf{F1}(\%) & \textbf{REC}(\%) & \textbf{PRE}(\%) & \textbf{F1}(\%) \\
\midrule
\multirow{5}{*}{Offline}
& NC & 97.1 & 61.6 & 75.4 & 92.5 & 51.4 & 66.7 \\
& ABS & 95.3 & 81.2 & 87.7 & 90.0 & 48.6 & 63.2 \\
& MNTD & 90.8 & 81.5 & 85.8 & 77.2 & 77.4 & 77.3 \\
& AC & 96.5 & 88.8 & 92.5 & 87.5 & 70.6 & 79.1 \\
& SCAn & 95.9 & \textbf{96.5} & 96.2 & 92.5 & 90.2 & 91.4 \\
& Beatrix & \textbf{97.7} & 96.0 & \textbf{96.8} & \textbf{95.0} & \textbf{90.5} & \textbf{92.7} \\
\midrule
\multirow{4}{*}{Online}
& STRIP & 86.7 & 97.9 & 91.9 & 87.8 & 96.1 & 91.7 \\
& SentiNet & 91.5 & 96.2 & 93.8 & 90.3 & 96.9 & 93.5 \\
& SCAn & 88.0 & 95.1 & 91.4 & 91.0 & 96.2 & 93.5 \\
& Beatrix & \textbf{99.8} & \textbf{96.4} & \textbf{98.1} & \textbf{97.2} & \textbf{97.1} & \textbf{97.2} \\
\bottomrule
\end{tabular}}
\end{table}
SCAn models the representation distribution by a Gaussian distribution so that it uses a set of clean samples to estimate the covariance matrix. Then SCAn leverages Linear Discriminant Analysis (LDA) to separate the feature representations into two subgroups. A high statistic value from the likelihood-ratio test indicates the class is infected. As we discussed in Section~\ref{sec: Limitations against Dynamic Backdoors}, the mean discrepancy is ineffective and the universal covariance assumption is not true in the dynamic backdoor attack. As a result, SCAn is ineffective against the dynamic backdoor attack because of its intrinsic limitations.
Its F1-score against universal attacks on GSTRB is 96.2\%, but it drops to 36.9\% when it encounters dynamic attacks.
Figure~\ref{fig: scan vs beatrix} provides a more in-depth analysis from the representation space perspective on CIFAR10. It shows that SCAn cannot raise the alarm when the feature representations of clean and poisoned samples are deeply fused in the dynamic backdoor. In contrast, Beatrix remains effective against the dynamic backdoor since Gram matrix is capable of capturing the subtle differences between clean and poisoned representations, as shown in Figure~\ref{fig: scan vs beatrix}(c).
ABS can only determine whether a model is infected or not. Therefore, we trained 100 clean models and 100 infected models with random initialization for the evaluations in universal and dynamic attacks. ABS assumes that each target label is associated with only one trigger and the trigger subverts all benign samples to the target label. This assumption is broken by the dynamic backdoor attack, in which each trojaned sample has its own unique trigger.
Additionally, based upon the universal trigger assumption, ABS assumes there is only one compromised neuron activated at a time by the trigger. Put differently, the changes in the activation of the intermediate layer only depend on this single neuron when encountering a trigger-carrying sample. However, due to the uniqueness of dynamic triggers, the abnormal changes in the activation pattern are dispersed across multiple neurons.
Therefore, their reverse engineered trigger cannot reflect the malfunction in the model infected by dynamic backdoor attacks, and consequently, it yields only 32.7\% and 42.3\% F1-score on GSTRB and CIFAR10 against dynamic backdoor attacks.
MNTD, similar to ABS, can only flag a model as either trojaned or benign. Therefore, we evaluate MNTD with 100 clean/infected models used in the evaluation of ABS. MNTD uses jumbo learning to generate thousands of shadow models and then train a meta-classifier to learn the output differences of trojaned between clean models. MNTD fails to detect infected models with dynamic backdoor as the representations are already deep fused in the middle activation layers (see Figure~\ref{fig: scan vs beatrix}(b)). Therefore, MNTD achieves 85.8\% F1-score against universal backdoor attacks on GTSRB whereas it drops to 69.4\% when it encounters dynamic attacks.
\begin{figure*}[t]
\centering
\subfigure[Universal backdoor]{
\begin{minipage}[t]{0.22\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{image/cifar10_badnet_target_0_balance=0_PCA.pdf}
\end{minipage}}
\subfigure[Dynamic backdoor]{
\begin{minipage}[t]{0.22\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{image/cifar10_dynamic_target_0_balance=0_PCA.pdf}
\end{minipage}}
\subfigure[Gramian of dynamic backdoor]{
\begin{minipage}[t]{0.22\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{image/cifar10_dynamic_target_0_balance=0_order=4_PCA.pdf}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.22\textwidth}
\centering
\includegraphics[width=0.88\textwidth]{image/scan_vs_beatrix.pdf}
\end{minipage}}
\caption{\label{fig: scan vs beatrix} A case study of SCAn and Beatrix on CIFAR10 with 0-th label being the trojan target. (a) and (b) illustrate the target class’ representations projected onto their first two principle components under the universal and dynamic backdoor attack, respectively. (c) shows the projection of the Gramian features $s$ of representations in (b). (d) shows the logarithmic anomaly index returned by SCAn and Beatrix.}
\vspace{-8mm}
\end{figure*}
\subsubsection{Online defense}
In the online setting, we consider three existing defenses, STRIP~\cite{gao2019strip}, SentiNet~\cite{chou2020sentinet}, and SCAn~\cite{tang2021demon}. We re-implemented SentiNet according to the paper, and used the Pytorch version of STRIP~\cite{codeinputaware}. SCAn is configured for online defense following the paper. The experimental results are shown in Tables~\ref{tabel: comparison_on_dynamic_backdoor} and~\ref{tabel: comparison_on_conventional_backdoor}.
STRIP works by superimposing a set of randomly selected clean images to an input image and measuring the entropy of the prediction outputs. In the conventional backdoor attack, a trojaned image with a static trigger is resistant to this perturbation, leading to a much lower entropy of the outputs compared to that of a benign input. The effectiveness of STRIP relies on the dominant impact of the trigger~\cite{tang2021demon}. However, this assumption no longer holds in the dynamic backdoor attack where images with mismatching triggers will deactivate the backdoors~\cite{nguyen2020input, li2021invisible}, and subsequently the F1-score of STRIP drops from 91.9\% and 91.7\% against universal backdoors to only 29.7\% and 28.3\% against dynamic backdoors.
\begin{table}[t]
\centering
\caption{\label{table: efficiency comparison} Efficiency comparison.}
\resizebox{0.4\textwidth}{!}{
\begin{tabular}{ccccc}
\toprule
\textbf{Method} & STRIP & SentiNet & SCAn & Beatrix \\
\midrule
\textbf{Time} (s) & 0.04 & 0.11 & 13.58 & 35.14$\times10^{-6}$ \\
\bottomrule
\end{tabular}}
\end{table}
SentiNet utilizes the model interpretability technique~\cite{selvaraju2017grad} to locate highly salient contiguous region (\textit{i.e.}, a potential trigger-region) of a given input. The extracted salient region is then overlaid on a set of clean images whose classification results are used to distinguish normal images from malicious images, since the trigger region is much more likely than normal region to subvert the clean images to the target label. What has been assumed here is that the attack is \textit{localized} (\textit{i.e.}, the trigger-region is constrained to a small contiguous part) and \textit{universal} (\textit{i.e.}, the attack is sample-agnostic). Both assumptions are broken by the dynamic backdoor attack where triggers are sample-specific and are distributed over different and disjointed portions of an image. Thus, the recall, precision and F1-score of SentiNet are 0\%, indicating that it is no longer effective against this advanced attack.
SCAn builds a composition model (covariance estimation) as well as an untangling model (mean estimation for each class) on a set of clean samples in an offline manner. Therefore, for each input sample, SCAn needs to update the untangling model for the image class.
The incoming sample is tagged as malicious if it results in a class anomaly index larger than the threshold ($e^{2}$) and falls into the smaller subgroup.
As we discussed above, SCAn models the feature representations under the LDA assumption (see Section~\ref{sec: Limitations against Dynamic Backdoors}), which is violated in the dynamic backdoor attack. Consequently, the error in the composition model leads to the ineffectiveness of the untangling model in distinguishing representations of benign inputs from those of malicious inputs.
As a result, the F1-score of SCAn drops from 93.5\% (in universal attacks) to 33.5\% (in dynamic attacks).
Additionally, due to \textit{its dependency on the accumulation of adversarial inputs}~\cite{tang2021demon}, SCAn shows a suboptimal performance against universal attacks in the online setting while it achieves similar performance in the offline setting compared to Beatrix.
\noindent \textbf{Efficiency comparison.~} As the overhead is important for the online scenario, we also compare the efficiency of Beatrix with those of other online defenses on the GTSRB dataset.
The experiment is conducted on one NVIDIA GeForce RTX 3090 GPU. The running time is averaged over 1,000 testing samples. As shown in Table~\ref{table: efficiency comparison}, our approach is much faster than the three baseline methods. SCAn consumes the longest time (13.58s) compared to other three methods since it needs to update the untangling model when there is an incoming sample. SentiNet pastes the potential trigger region of the input image on clean and noise images while STRIP directly superimposes each input on clean images. Thus, SentiNet (0.11s) is slightly slower than STRIP (0.04s).
The main computation overhead of Beatrix is building the deviation measurement model with clean samples, but this measurement model can be obtained through an offline training. Therefore, Beatrix only needs a forward pass to get the intermediate presentation for each sample and computes Gramian features of different orders, which can be obtained simultaneously when the user (defender) trains or tests her/his model with a dataset.
Therefore, Beatrix takes only 35.14 $\times10^{-6}$s (for $P=4$). Furthermore, as demonstrated in Section~\ref{sec: Effectiveness Against Dynamic Backdoor}, the defender can choose a different number of orders to trade off efficiency and effectiveness.
\begin{figure}[t]
\centering
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/n01855672_24054.JPEG}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/n01855672_24054_residual.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/n01855672_24054_hidden.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/n01882714_14807.JPEG}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/n01882714_14807_residual.png}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/n01882714_14807_hidden.png}
\end{minipage}}
\caption{\label{fig: examples of issba} Examples of poisoned samples under ISSBA. (a) and (d) are clean images, (b) and (e) are sample-specific triggers, and (c) and (f) are poisoned images.}
\vspace{-6mm}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{image/issb.pdf}
\caption{(a) The logarithmic anomaly index of infected and uninfected labels under ISSBA. (b) Deviation distribution of benign and trojaned samples in the infected class under ISSBA.}
\label{fig: defend_against_issb}
\end{figure}
\subsection{Robustness Against Other Attacks}
\label{sec: Robustness Against Other Attacks}
\begin{figure}[t]
\centering
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/00001_00029_00029_aug_180+2008_000104-background.jpg}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/00001_00029_00029_aug_180+2008_000104-reflection.jpg}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/00001_00029_00029_aug_180+2008_000104-input.jpg}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/00014_00007_00024+2008_000126-background.jpg}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/00014_00007_00024+2008_000126-reflection.jpg}
\end{minipage}}
\subfigure[]{
\begin{minipage}[t]{0.068\textwidth}
\centering
\includegraphics[width=0.9\textwidth]{image/00014_00007_00024+2008_000126-input.jpg}
\end{minipage}}
\caption{\label{fig: examples of refool} Examples of poisoned samples under \textit{Refool}. (a) and (d) are clean images, (b) and (e) are reflection patterns, and (c) and (f) are poisoned images.}
\vspace{-6mm}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.95\linewidth]{image/refool.pdf}
\caption{(a) The logarithmic anomaly index of infected and uninfected labels under \textit{Refool}. (b) Deviation distribution of benign and trojaned samples in the infected class under \textit{Refool}.}
\label{fig: defend_against_refool}
\end{figure}
\noindent \textbf{Invisible sample-specific backdoor attack.~}
Motivated by the advances in DNN-based image steganography~\cite{tancik2020stegastamp, baluja2017hiding}, Li \textit{et al.} proposed an invisible sample-specific backdoor attack (ISSBA), where triggers are generated by a pre-trained encoder network~\cite{li2021invisible}. The generated triggers are invisible additive noise containing the information of a representative string of the target label. The attacker can flexibly design the string as the name and the index of the target class or even a random character. The encoder network embeds the string into a clean image to obtain a poisoned image. Thus, the poisoned image generator (encoder) is conditioned on the input images, indicating the backdoor triggers vary from input to input. A DNN classifier trained on the poisoned images will misclassify images trojaned by the same encoder into the classes indicated by the embedded strings. Some examples of the poisoned ImageNet samples and their corresponding triggers are shown in Figure~\ref{fig: examples of issba}.
We evaluate Beatrix on this attack using the infected model and the dataset shared by the authors~\cite{li2021invisible}.
Since they only release their implementation on the ImageNet dataset, we use the released infected model trained on a ImageNet subset which contains 200 classes with the 0-th label (goldfish) being the trojan target.
As shown in Figure~\ref{fig: defend_against_issb}, the anomaly index of the infected label (label~0) is much larger than the uninfected labels (labels 1-199), indicating that Beatrix can effectively defend against this attack. More detailed comparison with other detection methods can be found in Appendix~\ref{sec: Defense Performance Against ISSBA}.
\noindent \textbf{Reflection backdoor attack.~} Liu \textit{et al.} proposed \textit{reflection backdoor} (\textit{Refool}) using a special backdoor pattern based on a nature phenomenon --- reflection~\cite{liu2020reflection}. Reflection occurs wherever there are glasses or smooth surfaces. \textit{Refool} generates poisoned images by adding reflections to clean images based on mathematical models of physical reflection scenarios. Different from conventional backdoor attacks that rely on a fixed trigger pattern, \textit{Refool} can utilize various reflections as the trigger pattern, making it stealthier than other attacks.
Additionally, \textit{Refool} blends the clean image with a triggering reflection pattern, so that the trigger is complex and spans all over the image.
Some examples of clean images and their poisoned counterparts with reflection triggers are illustrated in Figure~\ref{fig: examples of refool}.
We evaluate Beatrix against this attack using the code and datasets shared by the authors~\cite{liu2020reflection}. The released datasets include three traffic sign datasets: GTSRB, BelgiumTSC~\cite{timofte2014multi} and CTSRD~\cite{CTSRD}.
In this experiment, we use the GTSRB dataset and randomly choose the reflection images from PascalVOC~\cite{everingham2011pascal} following the original implementation.
The target class is the speed limit sign of 30 km/h. Our detection results are demonstrated in Figure~\ref{fig: defend_against_refool}. It shows that the anomaly index of the infected label is larger than the threshold, and those of the uninfected labels are all below the threshold. For online detection, Beatrix can effectively distinguish clean images from trigger-carrying ones with 99.99\% TPR @ 5\% FPR and 98.50\% TPR @ 1\% FPR, as shown in Figure~\ref{fig: defend_against_refool}(b).
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{image/detection_results_deviation_against_badencoder.pdf}
\caption{\label{fig: defend_against_badencoder} Deviation distribution of benign and trojaned samples in the infected class of (a) Imagnet encoder and (b) CLIP encoder under BadEncoder attack.}
\end{figure}
\noindent \textbf{BadEncoder attack.~} Another recently introduced backdoor attack is BadEncoder~\cite{jia2022badencoder}, which has been proposed for self-supervised learning pipelines. Self-supervised learning aims to pre-train an image encoder using a large amount of unlabeled data. Thus, the pre-trained encoder can be used as a feature extractor to build downstream classifiers in many different tasks. BadEncoder aims to compromise the self-supervised learning pipeline by injecting backdoors into a pre-trained image encoder such that the downstream classifier built upon this trojaned encoder will inherit the backdoor behavior.
To craft a trojaned image encoder, BadEncoder fine-tunes a clean image encoder with two additional loss terms named effectiveness loss and utility loss. The effectiveness loss measures the similarity between feature vectors of reference inputs (\textit{i.e., clean inputs in the target class}) and those of trigger-carrying inputs produced by the trojaned encoder.
To maintain utility as well as stealthiness, BadEncoder applies the utility loss to encourage the trojaned encoder and the clean encoder produces similar outputs given the same clean inputs. In this way, a downstream classifier built upon the trojaned encoder will behave normally on clean inputs but misclassify trigger-carrying inputs into the target class since their feature representations are similar to those of clean target inputs.
We put Beatrix into test against BadEncoder using the infected models and datasets published along with the paper that introduced the attack itself~\cite{jia2022badencoder}. In the experiments, we consider two real-world image encoders: ImageNet encoder originally released by Google~\cite{chen2020simple} and CLIP encoder by OpenAI~\cite{radford2021learning}. The target downstream dataset is GTSRB, and the target class is the 12th-label (\textit{i.e.}, the priority road sign). Instead of building a downstream classifier, we directly evaluate Beatrix on distinguishing clean and poison inputs based on their feature vectors produced by the backdoor encoder.
As shown in Figure~\ref{fig: defend_against_badencoder}, Beatrix can effectively defend against BadEncoder. Specifically, Beatrix achieves 99.8\% TPR @ 1\% FPR on the infected ImageNet encoder, and 100\% TPR @ 1\% FPR on the infected CLIP encoder.
\subsection{Beyond the Image Domain}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{image/detection_results_deviation_against_nlp.pdf}
\caption{\label{fig: defend_against_nlp} Deviation distribution of benign and trojaned samples in the infected class under (a) Homograph Backdoor Attack and (b) Dynamic Sentence Backdoor Attack.}
\vspace{-6mm}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.95\linewidth]{image/audio.pdf}
\caption{(a) The logarithmic anomaly index of infected and uninfected labels of a speech recognition backdoor model. (b) Deviation distribution of benign and trojaned samples in the infected class under the speech recognition backdoor attack.}
\label{fig: defend_against_audio}
\end{figure}
Akin to previous works, we mainly focus on the
image classification tasks. There are backdoor attacks in other domains, such as natural language processing (NLP) \cite{li2021hidden, azizi2021t, chen2021badnl}, acoustics signal processing \cite{zhai2021backdoor} and malware detection \cite{li2021backdoor, severi2021explanation}. Here, we extend our approach to mitigate the threats posed by backdoor attacks in speech recognition and text classification domains.
For the NLP task, we evaluate Beatrix on the homograph backdoor attack and dynamic sentence backdoor attack proposed by Li \textit{et al.} \cite{li2021hidden}.
The homograph backdoor attack inserts triggers by replacing several characters of the clean sequences with their homograph equivalent. Given an original sentence, the dynamic sentence backdoor attack uses pre-trained language models to generate a suffix sentence to act as the trigger.
We use the code and dataset shared by the authors to train poisoned BERT model on the toxic comment classification dataset \cite{toxic2020}. To balance the number of positive (\textit{i.e.}, toxic) and negative (\textit{i.e.}, non-toxic) samples, it draws 16225 negative samples from the negative texts so the final dataset consists of 32450 samples. The dataset is then split to give 29205 (90\% of the dataset) in the training set and 3245 (10\%) in the test set.
Since this is a binary classification task, we directly evaluate the online defense performance of Beatrix. As shown in Figure \ref{fig: defend_against_nlp}, Beatrix achieves 89.8\% TPR @ 5\% FPR on homograph attack and 75.2\% TPR @ 5\% FPR on dynamic sentence attack.
For the speech task, we use the speech recognition
backdoor implementation provided in the work \cite{xu2021detecting} on the SpeechCommand dataset \cite{warden2018speech}. The original dataset consists of 65,000 one-second audio files of 35 classes. Following the previous work \cite{xu2021detecting}, we use the files of 10 classes (\textit{i.e.}, “yes”, “no”, “up”, “down”, “left”, “right”, “on”, “off”, “stop”, “go”), which gives 30,769 training samples
and 4,074 testing samples. It first extracts the mel-spectrogram of each file and then trains an LSTM model over the mel-spectrograms. The backdoor trigger is a consecutive noise signal whose length is 0.05 seconds. Our detection results shown in Figure~\ref{fig: defend_against_audio} demonstrate that Beatrix can detect the infected label effectively and distinguish clean signal from trigger-carrying ones with 77.5\% TPR @ 5\% FPR.
\subsection{Adaptive Attack}
We study an adaptive adversary who targets at the deviation measurement of Beatrix. We consider a strong white-box adversary who controls the training process of the victim model. The objective of the adversary is to force the activation patterns of poisoned images to resemble those of clean images, so that Beatrix cannot separate the benign and malicious inputs. To achieve this goal, we design an adaptive loss to minimize the distance between poisoned and clean images of a target class, in the representation space based on multiple high-order Gram matrices. In the experiment, we set the order upper bound $P$ as 4 in the adaptive loss. Therefore, the loss function of the adaptive attack can be denoted as follows:
\begin{align*}
L &= L_{o} + \lambda L_{a}, \\
L_{a} = \mathbb{E}_{x\in X_{/y_t}, x_t \in X_{y_t}}&\left[\sum^{P}_{p=1}\Vert G^{p}(\mathcal{B}(x,g(x))) - G^{p}(x_{t})\Vert^{2}\right],
\end{align*}
where $L_{o}$ is the original dynamic backdoor loss for the victim model $f$ ~\cite{nguyen2020input}, and $L_{a}$ denotes the adaptive loss. $G^{p}$ denotes the $p$-th order Gram matrix of the internal activation from an input image. $X_{y_t}$ and $X_{/y_t}$ denotes target and non-target training data, respectively. $\lambda$ is a hyperparameter balancing the model performance and the adaptive strength.
We perform this attack on the GTSRB and CIFAR10 datasets. We evaluate the online performance of Beatrix to determine whether an input image is benign or malicious during the inference phase. In the experiment, we randomly select 500 clean images from the target class and 500 poisoned images from other classes.
The results are shown in Table~\ref{tabel: adaptive attack}. We can find that the true positive rate of Beatrix slightly decreases when the hyper-parameter $\lambda$ increases from 0.05 to 0.5. In particular, when $\lambda$ increases to 1, Beatrix yields only 14.0\% TPR @ 1\% FPR on CIFAR10 dataset, indicating that Beatrix is no longer effective. However, the model performance (\textit{i.e.}, clean accuracy on benign inputs and attack success rate on malicious ones) also decreases significantly.
\begin{table}[t]
\centering
\caption{\label{tabel: adaptive attack} Adaptive attack.}
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{ccccccc}
\toprule
& $\lambda$& 0.05 & 0.1 & 0.5 & 1 & 5 \\
\midrule
\multirow{5}{*}{GTSRB}
&CA & 98.5\% & 98.1\% & 93.8\% & 79.7\% & 4.7\% \\
&ASR & 97.5\% & 96.8\% & 95.2\% & 89.1\% & - \\
&5\% FPR & 99.8\% & 98.6\% & 97.4\% & 92.4\% & - \\
&1\% FPR & 98.8\% & 96.6\% & 93.6\% & 81.0\% & - \\
\midrule
\multirow{5}{*}{CIFAR10}
&CA & 92.7\% & 91.8\% & 89.8\% & 69.5\% & 10.0\% \\
&ASR & 98.1\% & 96.2\% & 94.6\% & 85.1\% & - \\
&5\% FPR & 97.6\% & 94.3\% & 83.3\% & 40.5\% & - \\
&1\% FPR & 87.2\% & 76.6\% & 51.7\% & 14.0\% & - \\
\bottomrule
\end{tabular}}
\end{table}
\section{Related Work}
\label{sec: related work}
\noindent \textbf{Backdoor attacks.}
With the increasing use of pre-trained DNNs in security-sensitive domains, backdoor attacks have recently been recognized as a new threat. Current attacks inject backdoors to DNNs by either poisoning the training dataset or directly manipulating the model parameters.
BadNets was first introduced as a poisoning-based backdoor attack in which the adversary has full control over the model and training data~\cite{gu2019badnets}. Subsequently, a series of attempts were made to launch attacks with fewer poisoned samples and limited access to training data~\cite{chen2017targeted, liu2018trojaning}. On the other end, to make the poisoning data stealthier, there are attacks that inject poisoned samples without altering their labels~\cite{shafahi2018poison,turner2019label}.
Recently, backdoors activated by semantic triggers with diverse patterns were studied to further conceal the attacks~\cite{barni2019new,liu2020reflection,li2020invisible,saha2020hidden}. The threat of neural backdoors in emerging machine learning paradigms has also been investigated~\cite{yao2019latent, shafahi2018poison,xie2019dba,jia2022badencoder}. Interestingly, model replacement has been identified as an attack surface in federated learning~\cite{bagdasaryan2020backdoor}.
Although various backdoor attacks are proposed, most of them employ static triggers
which can be easily detected by existing defensive techniques. On the contrary, recently proposed dynamic backdoor attacks craft input-aware triggers, which hardens the detection of such backdoors~\cite{salem2020dynamic, nguyen2020input, li2021invisible}.
\noindent \textbf{Defenses against backdoor attacks.}
Current defensive techniques can be broadly categorized into two branches. The first line of works conducts a model analysis to identify backdoors. Otherwise, defenders can perform run-time detection against triggered backdoors.
A common wisdom in model analysis is to discover anomalies in the learned representations~\cite{tran2018spectral, chen2019detecting}. Moreover, other methods are reverse-engineering possible triggers and unlearning inserted backdoors~\cite{wang2019neural, liu2019abs, liu2021ex}. However, these methods are reported as ineffective against large triggers and source-label-specific attacks~\cite{liu2019abs, tang2021demon}. There is also a method that adopts a meta classifier trained on a set of clean and trojaned models to identify compromised models~\cite{xu2021detecting}.
A defender can also implement run-time detection methods when the DNN has been deployed. STRIP was proposed to detect backdoors controlled by dominant triggers~\cite{gao2019strip}. SentiNet adopts Grad-Cam to detect potential backdoor locations and malicious inputs~\cite{selvaraju2017grad, chou2020sentinet}. A similar idea is also explored and extended in Februus~\cite{doan2020februus}. However, the aforementioned methods are vulnerable to adaptive attacks such as the source-label-specific backdoor. Lately, SCAn proposed a statistical method to defend against this attack~\cite{tang2021demon}; however, it is less desirable against attacks with dynamic triggers.
\section{Conclusion}
\label{sec: Conclusion}
In this paper, we demonstrated that the existing defensive techniques heavily rely on the premise of the universal backdoor trigger, in which poisoned samples share the same trigger and thus show the same abnormal behavior. Once this prerequisite is violated, they can no longer effectively detect the advanced backdoor attacks like dynamic backdoors, where the adversary injects sample-specific triggers to each input. Based on the observation that Gramian information of dynamically trojaned data points is highly distinct from that of the benign ones, we developed Beatrix to capture not only the feature correlations but also the appropriately high-order information of the representations of benign and malicious samples, and utilized the Kernel-based two-sample testing to identify the infect labels. Experimental results show the effectiveness and robustness of our proposed approach. Beatrix can successfully defend against backdoor attacks for not only the conventional ones but also the advanced attacks, such as dynamic backdoors which can defeat the aforementioned defensive techniques.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:00000000000}
The study of non-Hermitian systems~\cite{PhysRevLett.109.230405,arXiv:2201.05367} has paved the way for recent developments across the subjects of quantum sensing~\cite{PhysRevLett.125.240506,PhysRevLett.127.026404}, $\mathcal{PT}$-symmetry and exceptional points~\cite{Zyablovsky_2016,PhysRevLett.119.190401,PhysRevLett.127.186602,Yuto_Ashida2020,Bergholtz2021,PhysRevResearch.2.033022,PhysRevResearch.3.013015,PhysRevA.106.023309}, linear response theory~\cite{PhysRevLett.128.016802}, quantum many-body systems~\cite{Korff_2007,Korff_2008,Castro_Alvaredo_2009,NatCommun_8_15791_2017,Xiao2019,PhysRevB.101.121109,PhysRevLett.123.090603,PhysRevLett.125.260601,ProgTheorExpPhys_12_2020}, skin effect~\cite{PhysRevB.102.205118,arXiv:2201.10318,PhysRevResearch.4.033041}, bulk-edge correspondence~\cite{PhysRevLett.121.086803,PhysRevLett.125.226402}, phase transitions~\cite{PhysRevB.103.155417,PhysRevA.105.022219}, and the quantum boomerang effect in localized systems~\cite{arXiv:2206.02922,PhysRevX.12.011035,PhysRevA.99.023629,PhysRevB.106.L060301}, to cite a few.
Recent theoretical achievements discussed the effects of postselection on the dynamics of open quantum systems, thus reconciling the approaches of effective non-Hermitian Hamiltonians and Liouvillian superoperators~\cite{PhysRevA.100.062131,PhysRevA.101.062112}. In this setting, several works have addressed the role of non-Hermitian features on legitimate quantum mechanical signatures, e.g., quantum coherence and entanglement~\cite{PhysRevB.104.155141,PhysRevA.104.052405,PhysRevLett.126.170503,PhysRevB.105.L121115}. For example, it has been shown that quantum coherence can be characterized under the framework of multiple quantum coherences~\cite{PhysRevB.105.075103,PhysRevA.99.052354,PhysRevA.102.012429}. In turn, the dynamics of entanglement in non-Hermitian systems has been widely probed with entropic measures, but their evaluation generally involves the full spectral decomposition of the state driven by the non-Hermitian Hamiltonian~\cite{JStatMech_033102_2016,QIP_10_234_2017,arXiv:2105.09793,PhysRevB.105.075126}. This can be a challenging computational task, especially for interacting quantum many-body systems.
To overcome this issue, the onset growth of entanglement at earlier times of the dynamics can be addressed through the so-called linear entropy, a useful information-theoretic quantifier that is related to the second-order R\'{e}nyi entropy and quantum purity~\cite{PhysRevX.7.031011,NJP_20_053058_2018}. Remarkably, those quantities have been experimentally probed in optical lattices~\cite{20159_arxiv_1509.01160,PhysRevLett.120.050406,Science_353_6301_794}, and trapped ion setups~\cite{PhysRevA.98.052334,Brydges260}. In the Hermitian case, it is known that the short-time perturbative expansion of the linear entropy implies a universal timescale for the entanglement dynamics of inte\-rac\-ting bipartite systems~\cite{PhysRevD.97.066008,PhysRevA.97.022317,PhysRevA.71.042321}. It is worth mentioning that this timescale is inversely proportional to the fluctuations of the coupling between subsystems~\cite{NEMES1986367,RevModPhys.75.715}. Importantly, this result also assigns a timescale for the decoherence mechanism within the subsystems of such composite quantum systems~\cite{PhysRevLett.77.207,PhysRevLett.70.1187,JPhysChemLett_17_8_2017}. We also mention the study of the growth of entanglement through a perturbative expansion of the entanglement negativity~\cite{PhysRevA.99.012322,arXiv:2209.03976}, and also quantum fidelities~\cite{PhysRevA.56.4466,GORIN200633}. To the best of our knowledge, despite these remarkable achievements in the Hermitian setting, deriving an analogous timescale for non-Hermitian systems remains a gap to be filled.
Here we address timescales for the growth of the li\-near entropy for finite-dimensional quantum systems described by effective non-Hermitian Hamiltonians. The physical system is initialized in a quantum state which can be chosen as either a pure or mixed one, possibly an entangled state or even an uncorrelated one. We investigate the short-time perturbative expansion of the linear entropy for a given input state driven by a general non-Hermitian Hamiltonian. In this setting, up to the se\-cond order in time, the onset growth of mixedness of the evolved state is governed by two competing timescales that are intrinsically related to the anti-Hermitian part of the non-Hermitian Hamiltonian. In particular, we specialize these results in the case of a driven non-Hermitian two-level system and discuss the mixedness of a single-qubit state.
Next, focusing on the reduced dynamics of bipartite systems described by non-Hermitian Hamiltonians, we derive the short-time perturbative expansion of the linear entropy for a given evolved marginal state of the composite system. In the Hermitian limit, these results recover the perturbative expansion for coherence loss in Hermitian systems~\cite{PhysRevLett.77.207}. In particular, for initial pure and uncorrelated states, we find the lowest order entanglement timescale for quantum systems described by Hermitian Hamiltonians~\cite{PhysRevD.97.066008,PhysRevA.97.022317}. To illustrate these findings, we consider a paradigmatic many-body non-Hermitian Hamiltonian, and present analytical calculations and numerical simulations to support our theo\-re\-ti\-cal predictions. We verify that, unlike the Hermitian case, the non-Hermitian Hamiltonian is responsible for an enhancement in the short-time dynamics of the linear entropy for the multiparticle states that have been considered.
The paper is organized as follows. In Sec.~\ref{sec:00000000001}, we briefly review useful properties regarding the linear entropy. In Sec.~\ref{sec:00000000002} we investigate the short-time perturbative expansion of the linear entropy for finite-dimensional quantum systems whose dynamics are driven by a non-Hermitian Hamiltonian. In Sec.~\ref{sec:00000000002A}, we illustrate our findings by means of the two-level system. In Sec.~\ref{sec:00000000003} we derive a universal entanglement timescale for bipartite quantum systems evolving under the action of a given non-Hermitian Hamiltonian. In Sec.~\ref{sec:00000000003A}, we specialize these results to the case of two initially uncorrelated subsystems. In addition, Sec.~\ref{sec:00000000003B} addresses the case of a many-body system with non-Hermitian Hamiltonian describing the transverse-field XY model perturbed by a fully connected Ising Hamiltonian with imaginary exchange coupling. Finally, in Sec.~\ref{sec:00000000004} we summarize our conclusions.
\section{Linear entropy}
\label{sec:00000000001}
In this section, we review the main properties of li\-near entropy, i.e., a versatile information-theoretic measure that quantifies the degree of mixedness of a given state~\cite{Schlosshauer_book}. Linear entropy has been used to witness multipartite entanglement~\cite{PhysRevA.68.022318,JOptB_6_S730_2004,PhysRevA.75.032301}. Let us consider a quantum system with finite-dimensional Hilbert space $\mathcal{H}$, with $d = \dim\mathcal{H}$. The space of quantum states $\mathcal{S} \subset \mathcal{H}$ is a convex set of Hermitian, positive semidefinite, trace-one, $d\times d$ matrices, i.e., $\mathcal{S} = \{\rho \in \mathcal{H} \mid {\rho^{\dagger}} = \rho,~\rho\geq 0,~\text{Tr}(\rho) = 1\}$. The normalized linear entropy of the quantum state $\rho$ is defined as~\cite{PhysRevA.70.052309}
\begin{equation}
\label{eq:00000000001}
{S_L}(\rho) := \frac{d}{d - 1}[1 - f(\rho)] ~,
\end{equation}
where $f(\rho) = \text{Tr}({\rho^2})$ stands for the quantum purity. The latter quantity is bounded as $1/d \leq f(\rho) \leq 1$, which implies that the linear entropy ranges as $0 \leq {S_L}(\rho) \leq 1$ for all $\rho \in \mathcal{S}$. In addition, given the spectral decomposition $\rho = {\sum_j}\, {p_j}|j\rangle \langle{j}|$ in terms of the basis of states ${\{|j\rangle\}_{j = 1,\ldots, d}}$, with $0 \leq {p_j} \leq 1$ and ${\sum_j}\, {p_j} = 1$, one readily concludes that ${S_L}(\rho) = [d/(d - 1)][1 - {\sum_j}\, {p_j^2}]$.
Linear entropy remains invariant under unitary transformations over the input state, i.e., ${S_L}(V\rho{V^{\dagger}}) = {S_L}(\rho)$, with $V{V^{\dagger}} = {V^{\dagger}}V = \mathbb{I}$ and for all $\rho \in \mathcal{S}$. It is related to the second-order R\'{e}nyi entropy ${S_2}(\rho)$~\cite{Renyi1961,10.1063.1.4838856}, also known as collision entropy~\cite{PhysRevA.85.012108}, and thus becomes ${S_L}(\rho) = [d/(d - 1)][1 - e^{-{S_2}(\rho)}]$. Furthermore, Eq.~\eqref{eq:00000000001} is also written as ${S_L}(\rho) = [d/(d - 1)] {{\text{H}}_{2}}(\rho)$, with ${{\text{H}}_{2}}(\rho)$ being the second-order Tsallis entropy~\cite{Tsallis1988}. Interestingly, for a given state $\rho \in \mathcal{S}$ and observable $\Lambda \in \mathcal{H}$, it has been shown that the linear entropy satisfies the lower bound ${S_L}(\rho) \geq [2d/(d - 1)][{\langle{\Lambda^2}\rangle_{\rho}} - {\langle{\Lambda}\rangle^2_{\rho}} - (1/4)F(\rho,\Lambda)]/{[{\lambda_{\text{max}}}(\Lambda) - {\lambda_{\text{min}}}(\Lambda)]^2}$, with $\langle \bullet \rangle_{\rho} = \text{Tr}(\bullet \rho)$, where $F(\rho, \Lambda)$ is the quantum Fisher information, while ${\lambda_{\text{max}}}(\Lambda)$ and ${\lambda_{\text{min}}}(\Lambda)$ are the largest and smallest eigenvalues of $\Lambda$, respectively~\cite{arXiv:1701.07461}.
Recently, linear entropy has also been discussed for systems described by non-Hermitian Hamiltonians~\cite{JStatMech_033102_2016,Entropy_18120451,QIP_10_234_2017}. In this regard, at least two approaches can be highlighted. The first one consists of equipping the Hilbert space with a generalized inner product structure, called metric operator~\cite{Bender2007,PhysScr_82_038110}. In the second case, one introduces a renormalized density operator whose dynamics is governed by modified Heisenberg equations of motion~\cite{PhysRevLett.109.230405,PhysRevA.42.1467,doi:10.1142,EPJD_253_69_2015}. Throughout this paper, we will follow this last perspective.
\section{Mixedness timescale for non-Hermitian systems}
\label{sec:00000000002}
We consider a quantum system with finite-dimensional Hilbert space $\mathcal{H}$, with $d = \text{dim}\mathcal{H}$, which is initialized in the state ${\rho_0} \in \mathcal{S}$ [see Sec.~\ref{sec:00000000001}]. This input state undergoes a nonunitary evolution governed by the time-independent non-Hermitian Hamiltonian $H = {H_1} + i{H_2}$, where ${H_1} = ({1}/{2})(H + {H^{\dagger}})$, and ${H_2} = ({1}/{2i})(H - {H^{\dagger}})$. In turn, the noncommuting observables $H_1$ and $H_2$ stand for the Hermitian and anti-Hermitian parts of the Hamiltonian $H$, res\-pec\-tively. In the remainder of the paper, we set $\hbar = 1$. The dynamics of the normalized time-dependent density matrix ${\widetilde{\rho}_t} = {\rho_t}/{\text{Tr}({\rho_t})}$ fulfills the equation of motion~\cite{PhysRevA.42.1467,doi:10.1142,EPJD_253_69_2015}
\begin{equation}
\label{eq:00000000002}
\frac{d{\widetilde{\rho}_t}}{dt} = - {i} \, [{H_1}, {\widetilde{\rho}_t}] + \{ {H_2} , {\widetilde{\rho}_t}\} - {2} \, \text{Tr}({\widetilde{\rho}_t}{H_2}){\widetilde{\rho}_t} ~,
\end{equation}
which in turn represents a completely positive and trace-preserving operation. We point out that the dynamical map in Eq.~\eqref{eq:00000000002} drives a nonunitary evolution under which the state ${\widetilde{\rho}_t}$ exhibits a time-dependent mixedness. In the Hermitian setting, the mixedness stands as a conserved quantity for any quantum state undergoing a unitary evolution generated by a Hermitian Hamiltonian.
Here, we choose the normalized linear entropy ${S_L}({\widetilde{\rho}_t}) = [d/(d - 1)][1 - f({\widetilde{\rho}_t})]$ as a useful quantum information-theoretic quantifier to probe the mixedness of state ${\widetilde{\rho}_t}$, with $f({\widetilde{\rho}_t}) = \text{Tr}({\widetilde{\rho}_t^{\, 2}})$. As discussed in Sec.~\ref{sec:00000000001}, for the case of unitary evolutions generated by Hermitian Hamiltonians, both the purity and the linear entropy remain invariant for all $t > 0$. However, for the nonunitary dynamics dictated by non-Hermitian Hamiltonians, the linear entropy becomes a time-dependent quantity, and its evaluation requires the full spectral decomposition of the evolved state. This task has a high computational cost for many-body quantum systems.
In this section, we are interested in the short-time perturbative expansion of ${S_L}({\widetilde{\rho}_t})$ to understand the initial growth of the mixedness of the evolved state ${\widetilde{\rho}_t}$ [see Eq.~\eqref{eq:00000000002}]. The Taylor expansion of the linear entropy up to second order in $t$ around $t = 0$ yields
\begin{equation}
\label{eq:00000000003}
{S_L}({\widetilde{\rho}_t}) \approx {S_L}({\rho_0}) - \frac{d}{d - 1}\left(\frac{t}{T_1} + \frac{t^2}{T_2^2}\right) + O({t^3}) ~,
\end{equation}
where we define
\begin{equation}
\label{eq:00000000004}
{T_1^{-1}} := {4} \, {\text{cov}_{\rho_0}}({\rho_0},{H_2}) ~,
\end{equation}
and
\begin{align}
\label{eq:00000000005}
&{T_2^{-2}} := - {4} f({\rho_0})\, {\text{var}_{\rho_0}}({H_2}) - {8} \, {\langle{H_2}\rangle_{\rho_0}} \, {\text{cov}_{\rho_0}}({\rho_0},{H_2}) \nonumber\\
&+ {8} \, {\text{cov}_{\rho_0}}({H_2},{\rho_0}{H_2}) - {2i} \, {\text{cov}_{\rho_0}}({\rho_0},[{H_2},{H_1}]) ~,
\end{align}
with $\langle \bullet \rangle_{\rho_0} = \text{Tr}({\rho_0} \, \bullet)$ being the expectation value at time $t = 0$, while ${\text{cov}_A}(B,C) = ({1}/{2})\text{Tr}(A\{B,C\}) - \text{Tr}(AB)\text{Tr}(AC)$ defines the covariance functional. In particular, for $B = C$, note that the covariance reduces to the variance ${\text{var}_A}(B) \equiv {\text{cov}_A}(B,B) = \text{Tr}(A{B^2}) - {\text{Tr}(AB)^2}$.
Equations~\eqref{eq:00000000003},~\eqref{eq:00000000004} and~\eqref{eq:00000000005} are the first main result of the paper. The coefficients $|1/{T_1}|$ and $|1/{T_2}|$ provide timescales for the linear entropy at earlier times of the dynamics, thus predicting the initial growth of the mixedness of the evolved state of the quantum system. Importantly, they can be evaluated once the input state $\rho_0$ and the Hamiltonian $H = {H_1} + i {H_2}$ of the system have been specified. Note that $1/T_1$ and $1/T_2$ depend on the fluctuations of the obser\-va\-ble $H_2$ that are captured by its covariance respective to the input state. In particular, choosing $H_2$ a zero-valued operator, Eqs.~\eqref{eq:00000000004} and~\eqref{eq:00000000005} vanish and one gets that ${S_L}({\widetilde{\rho}_t}) \approx {S_L}({\rho_0})$. In fact, this is expected since the linear entropy remains invariant for any quantum state undergoing a unitary evolution ge\-ne\-ra\-ted by a Hermitian operator. In addition, for any initial pure state with ${\rho_0^2} = {\rho_0}$, and $f({\rho_0}) = \text{Tr}({\rho_0^2}) = 1$, one can verify that Eqs.~\eqref{eq:00000000004} and~\eqref{eq:00000000005} imply that $1/{T_1} = 0$ and $1/{T_2^2} = 0$, regardless of the operators $H_1$ and $H_2$. Finally, we note that the last term on the right-hand side of Eq.~\eqref{eq:00000000005} vanishes for $[{H_1},{H_2}] = 0$, i.e., for two commuting operators $H_1$ and $H_2$.
The proof of Eq.~\eqref{eq:00000000003} is as follows. We shall begin with the linear entropy of the evolved state ${\widetilde{\rho}_t}$, and evaluate its Taylor expansion up to second order in $t$, around $t = 0$, which yields
\begin{align}
\label{eq:00000000006}
{S_L}({\widetilde{\rho}_t}) &\approx {S_L}({\rho_0}) - \frac{d}{d - 1} \, t \, {[ {f^{(1)}}({\widetilde{\rho}_t}) ]_{t = 0}} \nonumber\\
& - \frac{d}{d - 1} \, \frac{t^2}{2} {[ {f^{(2)}}({\widetilde{\rho}_t}) ]_{t = 0}} + O({t^3}) ~,
\end{align}
where the $n$-th order derivative of the quantum purity becomes
\begin{equation}
\label{eq:00000000007}
{f^{(n)}}({\widetilde{\rho}_t}) = {\sum_{k = 0}^n} \, \frac{n!}{{(n - k)!} \, {k!}} \, \text{Tr}\left( \frac{{d^k}{\widetilde{\rho}_t}}{d{t^k}} \,\frac{{d^{n - k}} {\widetilde{\rho}_t}}{d{t^{n - k}}} \right) ~.
\end{equation}
Therefore, starting from Eqs.~\eqref{eq:00000000002} and~\eqref{eq:00000000007}, both the first-order and second-order derivatives of the quantum purity at the vicinity of $t = 0$ yield
\begin{align}
\label{eq:00000000008}
{[ {f^{(1)}}({\widetilde{\rho}_t}) ]_{t = 0}} &= {4} \left( \text{Tr}({\rho_0^2}{H_2}) - \text{Tr}({\rho_0^2}) \text{Tr}({\rho_0}{H_2}) \right) \nonumber\\
&= {4} \, {\text{cov}_{\rho_0}}({\rho_0},{H_2}) ~,
\end{align}
and
\begin{align}
\label{eq:00000000009}
&(1/2){[ {f^{(2)}}({\widetilde{\rho}_t}) ]_{t = 0}} = - {4} f({\rho_0})\, {\text{var}_{\rho_0}}({H_2}) \nonumber\\
&- {8} \, {\langle{H_2}\rangle_{\rho_0}} \, {\text{cov}_{\rho_0}}({\rho_0},{H_2}) + {8} \, {\text{cov}_{\rho_0}}({H_2},{\rho_0}{H_2}) \nonumber\\
&- {2i} \, {\text{cov}_{\rho_0}}({\rho_0},[{H_2},{H_1}]) ~,
\end{align}
where we recognize the aforementioned covariance and variance functionals. Finally, bringing together the results in Eqs.~\eqref{eq:00000000006},~\eqref{eq:00000000008} and~\eqref{eq:00000000009}, one readily concludes the main result in Eq.~\eqref{eq:00000000003}. We point out that the first-order derivative ${[ {f^{(1)}}({\widetilde{\rho}_t}) ]_{t = 0}}$ was previously investigated in the context of gain-loss systems~\cite{PhysRevLett.109.230405}, and in the quantum-classical description of non-Hermitian systems~\cite{PhysRevLett.109.230405,Entropy_18120451}.
\begin{figure*}[!th]
\begin{center}
\includegraphics[scale=0.9]{FIG_01.pdf}
\caption{(Color online) Density plot of the dimensionless quantities $1/\gamma{T_1}$ [see Eq.~\eqref{eq:00000000010}] and $1/{\gamma^2}{T_2^2}$ [see Eq.~\eqref{eq:00000000011}], as a function of the mixing parameter $r\in [0,1]$ and the azimuthal angle $\theta \in [0,\pi]$, respective to the perturbative expansion of the linear entropy for the driven dissipative two-level system around [see Eq.~\eqref{eq:00000000003}]. In panels (b)--(d), we set the polar angle $\phi = \pi/4$, and consider the ratio $\Omega/\gamma = 0.1$ [panel (b)], $\Omega/\gamma = 1$ [panel (c)], and $\Omega/\gamma = 0.1$ [panel (d)].}
\label{fig:FIG01}
\end{center}
\end{figure*}
\subsection{Example: Dissipative two-level system}
\label{sec:00000000002A}
\begin{figure*}[!th]
\begin{center}
\includegraphics[scale=1.0]{FIG_02.pdf}
\caption{(Color online) Plot of the linear entropy ${S_L}({\widetilde{\rho}_t})$, as a function of the dimensionless parameter $\gamma t$, for the driven two-level system described by the Hamiltonian $H = \Delta{|1\rangle}{\langle{1}|} + (\Omega/2)(|{0}\rangle\langle{1}| + |{1}\rangle\langle{0}|)$. Here we choose the ratio $\Delta/\gamma = 0.5$, and also set $\Omega/\gamma = 0.1$ [panels (a),~(d)], $\Omega/\gamma = 1$ [panels (b),~(e)], and $\Omega/\gamma = 10$ [panels (c),~(f)]. The system is initialized in the single-qubit state $\rho_0 = (1/2)(\mathbb{I} + \vec{r}\cdot\vec{\sigma})$, with $\{r, \theta, \phi\} = \{ 1/4, \pi/4, \pi/4 \}$ [panels (a),~(b),~(c)], and $\{r, \theta, \phi\} = \{ 1/4, 3\pi/4, \pi/4 \}$ [panels (d),~(e),~(f)].}
\label{fig:FIG02}
\end{center}
\end{figure*}
To illustrate our findings, we consider a driven two-level system described by the Hamiltonian $H_1 = \Delta{|1\rangle}{\langle{1}|} + (\Omega/2)(|{0}\rangle\langle{1}| + |{1}\rangle\langle{0}|)$, where the two vectors $|0\rangle$ and $|{1}\rangle$ stand for ground and excited states, respectively, with $\Delta$ the energy detuning, and $\Omega$ being their coupling. The system interacts with a zero-temperature thermal reservoir, so that it decays from the excited state $|1\rangle$ to the ground state $|0\rangle$ emitting a photon at a rate $\gamma$. The dynamics is governed by the Markovian master equation $d{\rho_t}/dt = -i({H_{\text{eff}}}{\rho_t} - {\rho_t}{H^{\dagger}_{\text{eff}}}) + \gamma{L}{\rho_t}{L^{\dagger}}$, where ${H_{\text{eff}}} = H - i(\gamma/2){L^{\dagger}}{L}$ is the effective non-Hermitian Hamiltonian, while $L = |{0}\rangle\langle{1}|$ is the jump operator~\cite{Breuer_Petruccione_book}.
In the semiclassical regime, i.e., assuming that the effect of quantum jumps is negligible in the time interval under consideration, an effective description of the master equation can be obtained in terms of the co\-he\-rent nonunitary dissipation of the system, the latter related to the non-Hermitian Hamiltonian ${H_{\text{eff}}}$~\cite{PhysRevA.100.062131}. In this case, by discarding the quantum jump term $\gamma{L}{\rho_t}{L^{\dagger}}$, the dynamics of the system is dictated by the equation $d{\rho_t}/dt \approx -i({H_{\text{eff}}}{\rho_t} - {\rho_t}{H^{\dagger}_{\text{eff}}})$, which no longer des\-cri\-bes a completely positive and trace-preserving evolution.
To overcome this issue, one introduces the norma\-li\-zed time-dependent density matrix ${\widetilde{\rho}_t} = {\rho_t}/{\text{Tr}({\rho_t})}$, which in turn fulfills Eq.~\eqref{eq:00000000002}, with ${H_1} = \Delta {|1\rangle}{\langle{1}|} + (\Omega/2)(|{0}\rangle\langle{1}| + |{1}\rangle\langle{0}|)$ and $H_2 = - (\gamma/2){|1\rangle}{\langle{1}|}$. The system is initialized in a single-qubit state $\rho_0 = (1/2)(\mathbb{I} + \vec{r}\cdot\vec{\sigma})$, where $\vec{r} = \{ r\sin\theta\cos\phi, r\sin\theta\sin\phi , r\cos\theta \}$ is the Bloch vector, with $r \in [0,1]$, $\theta \in [0,\pi]$ and $\phi \in [0,2\pi [$, while $\vec{\sigma} = \{{\sigma_x},{\sigma_y},{\sigma_z}\}$ is the vector of Pauli matrices, and $\mathbb{I}$ is the $2\times 2$ identity matrix. We will not show the analytical expressions for the exact linear entropy ${S_L}({\widetilde{\rho}_t})$ of the evolved state as they are cumbersome [see Eq.~\eqref{eq:00000000001}]. However, it is straightforward to obtain the short-time series expansion of ${S_L}({\widetilde{\rho}_t})$ applying Eq.~\eqref{eq:00000000003}, with the linear entropy of the input state as ${S_L}({\rho_0}) = 1 - {r^2}$. Using Eqs.~\eqref{eq:00000000004} and~\eqref{eq:00000000005}, we obtain the following dimensionless coefficients
\begin{equation}
\label{eq:00000000010}
\frac{1}{\gamma \, {T_1}} = (1 - {r^2}) \, r \cos\theta ~,
\end{equation}
and
\begin{equation}
\label{eq:00000000011}
\frac{1}{{\gamma^2} {T_2^2}} = \frac{1}{4}(1 - {r^2})\left(1 - 3{r^2}{\cos^2}\theta + 2 \frac{\Omega}{\gamma} \, r \sin\theta\sin\phi \right) ~,
\end{equation}
respectively. We see that $1/\gamma{T_1}$ is a function of $r$ and $\theta$, while $1/{\gamma^2}{T_2^2}$ depends on the parameters $r$, $\theta$, $\phi$, and $\Omega/\gamma$. We notice that ${S_L}({\rho_0})$, $1/{T_1}$ and $1/{T_2}$ approach zero for any initial single-qubit pure state with $r = 1$, thus implying that the linear entropy ${S_L}({\widetilde{\rho}_t})$ is a vanishing quantity in this case.
In Fig.~\ref{fig:FIG01} we show the plots of the dimensionless quantities $1/\gamma{T_1}$ and $1/{\gamma^2}{T_2^2}$, as a function of the mixing parameter $r$ and the azimuthal angle $\theta$. In Fig.~\ref{fig:FIG01}(a), we see that $1/\gamma{T_1} > 0$ for $\theta \in [0,\pi/2)$, while $1/{T_1} < 0$ for $\theta \in (\pi/2,\pi]$, regardless of the mixing parameter $0 \leq r \leq 1$. In addition, it follows that $1/{\gamma}{T_1} = 0$ for any chosen initial state with $\theta = \pi/2$ [see Eq.~\eqref{eq:00000000010}]. Next, Figs.~\ref{fig:FIG01}(b)--(d) show the plots of $1/{\gamma^2}{T_2^2}$ in Eq.~\eqref{eq:00000000011}, where we consider the cases $\Omega/\gamma = 0.1$ [see Fig.~\ref{fig:FIG01}(b)], $\Omega/\gamma = 1$ [see Fig.~\ref{fig:FIG01}(c)], and $\Omega/\gamma = 10$ [see Fig.~\ref{fig:FIG01}(d)], also fixing the polar angle $\phi = \pi/4$. On the one hand, for input states with either $\theta = 0$ or $\theta = \pi$, that are all incohe\-rent states respective to the computational basis $\{ |0\rangle, |1\rangle \}$, Eq.~\eqref{eq:00000000011} reduces to ${1}/{{\gamma^2} {T_2^2}} = (1/4)(1 - {r^2})(1 - 3{r^2})$, which is positive for $0 \leq r \leq 1/\sqrt{3}$ [see Figs.~\ref{fig:FIG01}(b)--(d)]. On the other hand, for initial states lying in the equatorial $xy$-plane with $\theta = \pi/2$, one gets that ${1}/{{\gamma^2} {T_2^2}} = (1/4)(1 - {r^2})[ 1 + 2 r (\Omega/\gamma)\sin\phi ]$, which is po\-si\-tive for $0 \leq \phi \leq \pi$ and $0 \leq r \leq 1$. We emphasize that the timescales related to the growth of mixedness can be obtained from the absolute values $|1/\gamma{T_1}|$ and $|1/{\gamma}{T_2}|$.
Figure~\ref{fig:FIG02} shows the plots of the linear entropy ${S_L}({\widetilde{\rho}_t})$, as a function of the dimensionless parameter $\gamma t$, for the aforementioned driven two-level system. The blue solid line refer to the exact linear entropy ${S_L}({\widetilde{\rho}_t})$ [see Eq.~\eqref{eq:00000000001}], while the red dashed line depicts the short time expansion of this quantity in Eq.~\eqref{eq:00000000003}. We set input states with $\{r, \theta, \phi\} = \{ 1/4, \pi/4, \pi/4 \}$ [see Figs.~\ref{fig:FIG02}(a)--\ref{fig:FIG02}(c)], and $\{r, \theta, \phi\} = \{ 1/4, 3\pi/4, \pi/4 \}$ [see Figs.~\ref{fig:FIG02}(d)--\ref{fig:FIG02}(f)]. In addition, for a fixed ratio $\Delta/\gamma = 0.5$, we consider the cases $\Omega/\gamma = 0.1$ [see Figs.~\ref{fig:FIG02}(a) and~\ref{fig:FIG02}(d)], $\Omega/\gamma = 1$ [see Figs.~\ref{fig:FIG02}(b) and~\ref{fig:FIG02}(e)], and $\Omega/\gamma = 10$ [see Figs.~\ref{fig:FIG02}(c) and~\ref{fig:FIG02}(f)]. In this setting, Eqs.~\eqref{eq:00000000010} and~\eqref{eq:00000000011} show that $1/{T_1}$ and $1/{T_2^2}$ stand as positive quantities, respectively, regardless of the ratio $\Omega/\gamma$ and the probe states that have been considered. This implies that the linear entropy in Eq.~\eqref{eq:00000000003} is a concave function.
Overall, Fig.~\ref{fig:FIG02} shows that the short-time approximation of ${S_L}({\widetilde{\rho}_t})$ correctly reproduces its growth at earlier times of the dynamics. Indeed, we find good quantitative agreement for $0 \leq \gamma t \lesssim 0.1$. However, for $\gamma t \gtrsim 0.1$, we have that the latter result is loose and fails to capture the changes in the eigenvalues of the state ${\widetilde{\rho}_t}$ driven by the non-Hermitian Hamiltonian. We emphasize that one should look to higher orders in its Taylor expansion to accurately predict the mixedness degree of the evolved state for later times.
\section{Entanglement timescale for non-Hermitian systems}
\label{sec:00000000003}
In this section, we provide an entanglement timescale for quantum systems whose dynamics can be effectively described by a non-Hermitian Hamiltonian. In detail, using the linear entropy as a useful measure of entanglement, we investigate its short-time expansion up to the second order in $t$ for certain time-dependent marginal states of the composite system.
We consider a bipartite quantum system with a finite-dimensional Hilbert space ${\mathcal{H}_A} \otimes {\mathcal{H}_B}$ split into the subsystems ${\mathcal{H}_A}$ and ${\mathcal{H}_B}$, with ${d_{A,B}} = \text{dim} \, {\mathcal{H}_{A,B}}$. This com\-po\-si\-te system is initialized in the quantum state ${\rho^{AB}_0}$, which in turn can be chosen either a pure or mixed state, entangled or uncorrelated one, from which the mixed marginal states ${\rho^{A,B}_0} = {\text{Tr}_{B,A}}({\rho^{AB}_0})$ can be obtained. The state ${\rho^{AB}_0}$ undergoes a nonunitary evolution ge\-ne\-ra\-ted by the time-independent non-Hermitian Hamiltonian $H = {H_1} + i{H_2}$, with ${H_1} = ({1}/{2})(H + {H^{\dagger}})$ and ${H_2} = - ({i}/{2})(H - {H^{\dagger}})$ being noncommuting obser\-va\-bles acting over ${\mathcal{H}_A} \otimes {\mathcal{H}_B}$. It is noteworthy that the operators $H_1$ and $H_2$ play the role of the Hermitian and anti-Hermitian parts of $H$, respectively. In this setting, it can be proved that the effective dynamics of subsystem $\mathcal{H}_{A,B}$ is go\-ver\-ned by the equation of motion~\cite{PhysRevA.42.1467,doi:10.1142,EPJD_253_69_2015}
\begin{align}
\label{eq:00000000012}
\frac{d}{dt}{\widetilde{\rho}^{A,B}_t} &= - i \, {\text{Tr}_{B,A}}([{H_1}, {\widetilde{\rho}^{AB}_t}]) + {\text{Tr}_{B,A}}(\{ {H_2} , {\widetilde{\rho}^{AB}_t}(t)\}) \nonumber\\
& - 2 \,{\text{Tr}_{AB}}({\widetilde{\rho}^{AB}_t}{H_2})\,{\widetilde{\rho}^{A,B}_t} ~,
\end{align}
where ${\widetilde{\rho}^{A,B}_t} := {\text{Tr}_{B,A}}({\widetilde{\rho}^{AB}_t})$ stand for the time-dependent reduced density matrices, while ${\widetilde{\rho}^{AB}_t} := {{\rho^{AB}_t}}/{{\text{Tr}_{AB}}({\rho^{AB}_t})}$ is the nor\-ma\-li\-zed state of the whole system.
Without loss of generality, hereafter we will address the dynamics of the marginal state ${\widetilde{\rho}^A_t}$, and investigate the short-time behavior of its linear entropy
\begin{equation}
\label{eq:00000000013}
{S_L}({\widetilde{\rho}^A_t}) = \frac{d_A}{({d_A} - 1)}(1 - f({\widetilde{\rho}^A_t})) ~,
\end{equation}
where $f({\widetilde{\rho}^A_t}) = {\text{Tr}_A}[{({\widetilde{\rho}^A_t})^2}]$ is the purity of the aforementioned reduced density matrix. In this case, by per\-for\-ming a Taylor expansion of ${S_L}({\widetilde{\rho}^A_t})$ up to se\-cond order in $t$, around $t = 0$, one gets
\begin{align}
\label{eq:00000000014}
{S_L}({\widetilde{\rho}^A_t}) &\approx {S_L}({\rho_0^A}) - \frac{d_A}{({d_A} - 1)} \left(\frac{1}{T_{1,h}} + \frac{1}{T_{1,nh}}\right)t \nonumber\\
&- \frac{d_A}{({d_A} - 1)}\left(\frac{1}{T_{2,h}^2} + \frac{1}{T_{2,nh}^2}\right){t^2} + O({t^3}) ~,
\end{align}
with ${T_{1,h}}$ and ${T_{1, nh}}$ being coefficients related to the first-order derivative of the linear entropy around $t = 0$, and defined as
\begin{equation}
\label{eq:00000000015}
{T_{1,h}^{-1}} := {2i} \, {\langle {\text{Tr}_{B}}( [ {\rho^{AB}_0} , {H_1} ]) \rangle_{A}} ~,
\end{equation}
and
\begin{equation}
\label{eq:00000000016}
{T_{1,nh}^{-1}} := 2\, {\langle{\text{Tr}_B}(\{ {\rho^{AB}_0} , {H_2} \})\rangle_A} - 4 \, f({\rho_0^A}) {\langle{H_2}\rangle_{AB}} ~,
\end{equation}
while ${T_{2,h}}$ and ${T_{2, nh}}$ arise from the second-order derivative of the linear entropy at the vicinity of $t = 0$ as follows
\begin{align}
\label{eq:00000000017}
{T_{2,h}^{- 2}} &:= - {\langle {\text{Tr}_{B}}([[{\rho^{AB}_0} , {H_1}] , {H_1}]) \rangle_A} \nonumber\\
& - {\text{Tr}_{A}}\left({({\text{Tr}_{B}}([{\rho^{AB}_0} , {H_1}]) )^2}\right) ~,
\end{align}
and
\begin{align}
\label{eq:00000000018}
&{T_{2, nh}^{- 2}} :=\nonumber\\
&{\langle {\text{Tr}_{B}}(\{ \{ {\rho^{AB}_0} , {H_2} \} , {H_2} \}) \rangle_A} + {\text{Tr}_{A}}\left({({\text{Tr}_{B}}(\{{\rho^{AB}_0} , {H_2} \} ) )^2}\right) \nonumber\\
&+ i {\langle {\text{Tr}_{B}}(\{ [{\rho^{AB}_0} , {H_1} ] , {H_2} \}) \rangle_A} + i {\langle {\text{Tr}_{B}}( [ \{ {\rho^{AB}_0} , {H_2} \} , {H_1} ]) \rangle_A} \nonumber\\
&- 8 \, {\langle {H_2} \rangle_{AB}} \left( {\langle {\text{Tr}_{B}}( \{ {\rho^{AB}_0} , {H_2} \} ) \rangle_A} + i \, {\langle {\text{Tr}_{B}}([ {\rho^{AB}_0} , {H_1} ]) \rangle_A} \right) \nonumber\\
&+ 2 f({\rho_0^A}) \left( i\, {\langle [{H_2},{H_1}] \rangle_{AB}} - 2 \left({\langle{H_2^2}\rangle_{AB}} - 3\, {\langle{H_2}\rangle_{AB}^2} \right) \right) \nonumber\\
&- 2i \, {\text{Tr}_A}\left( {\text{Tr}_{B}}([ {H_1} , {\rho^{AB}_0} ]) \, {\text{Tr}_{B}}(\{ {\rho^{AB}_0} , {H_2} \}) \right) ~.
\end{align}
Here $\langle\bullet\rangle_{\mu} := {\text{Tr}_{\mu}}(\bullet \, {\rho^{\mu}_0})$ defines the expectation value at time $t = 0$, with $\mu = \{A, B, AB\}$.
We point out that Eq.~\eqref{eq:00000000014} [see also Eqs.~\eqref{eq:00000000015}--\eqref{eq:00000000018}] is the second main result of the paper. Overall, we see that the coefficients $|{1/{T_{1,h}}} + {1/{T_{1,nh}}}|$ and ${|{1/{T_{2,h}^2}} + {1/{T_{2,nh}^2}}|}$ represent first-order and second-order timescales in the initial growth of the entanglement dynamics signaled by linear entropy. On the one hand, both the coefficients $T_{1,h}$ and $T_{2,h}$ depend on the initial state of the bipartite system and the Hermitian part ${H_1}$ of the non-Hermitian Hamiltonian. Importantly, Eqs.~\eqref{eq:00000000015} and~\eqref{eq:00000000017} recover the two lowest-order terms in the short-time expansion in the so-called idempotency deffect for composite systems described by Hermitian Hamiltonians, and constitute a timescale for the entanglement dynamics of subsystems~\cite{PhysRevLett.77.207,PhysRevD.97.066008,PhysRevA.97.022317}. On the other hand, the coefficients $1/{T_{1,nh}}$ and $1/{T_{2 , nh}}$ depend on the anti-Hermitian part ${H_2}$ of the effective non-Hermitian Hamiltonian. In particular, note that the result in Eqs.~\eqref{eq:00000000016} and~\eqref{eq:00000000018} approach zero in the Hermitian limit $H^{\dagger} = H = H_1$, i.e., when one sets $H_2$ as a zero-valued observable, regardless of the observable $H_1$. This means that $1/{T_{1,nh}}$ and $1/{T_{2,nh}}$ assign first-order and second-order nontrivial corrections to the entanglement timescales that are induced by the effective non-Hermitian Hamiltonian.
The proof of Eq.~\eqref{eq:00000000014} is as follows. We begin evaluating Taylor's series of ${S_L}({\widetilde{\rho}^A_t})$ up to the second order in $t$, around $t = 0$, which yields
\begin{align}
\label{eq:00000000019}
{S_L}({\widetilde{\rho}^A_t}) &\approx {S_L}({\rho_0^A}) - \frac{d_A}{{d_A} - 1} \, t \, {[ {f^{(1)}}({\widetilde{\rho}^A_t}) ]_{t = 0}} \nonumber\\
& - \frac{d_A}{{d_A} - 1} \, \frac{t^2}{2}{[ {f^{(2)}}({\widetilde{\rho}^A_t}) ]_{t = 0}} + O({t^3}) ~.
\end{align}
The first-order and second-order derivatives of the purity of the reduced state around $t = 0$ read as
\begin{align}
\label{eq:00000000020}
&{[ {f^{(1)}}({\widetilde{\rho}^A_t}) ]_{t = 0}} = - {2i} \, {\langle {\text{Tr}_{B}}( [ {H_1} , {\rho^{AB}_0} ]) \rangle_{A}} \nonumber\\
&+ 2\, {\langle{\text{Tr}_B}(\{ {\rho^{AB}_0} , {H_2} \})\rangle_A} - 4 \, {{f_A}(0)} {\langle{H_2}\rangle_{AB}} ~,
\end{align}
and
\begin{align}
\label{eq:00000000021}
&(1/2){[ {f^{(2)}}({\widetilde{\rho}^A_t}) ]_{t = 0}} =\nonumber\\
&- {\langle {\text{Tr}_{B}}([[{\rho^{AB}_0} , {H_1}] , {H_1}]) \rangle_A} - {\text{Tr}_{A}}\left({ ({\text{Tr}_{B}}([{\rho^{AB}_0} , {H_1}]) )^2}\right) \nonumber\\
& + {\langle {\text{Tr}_{B}}(\{ \{ {\rho^{AB}_0} , {H_2} \} , {H_2} \}) \rangle_A} + {\text{Tr}_{A}}\left({ ({\text{Tr}_{B}}(\{{\rho^{AB}_0} , {H_2} \} ) )^2}\right) \nonumber\\
& + i {\langle {\text{Tr}_{B}}(\{ [{\rho^{AB}_0} , {H_1} ] , {H_2} \}) \rangle_A} + i {\langle {\text{Tr}_{B}}( [ \{ {\rho^{AB}_0} , {H_2} \} , {H_1} ]) \rangle_A} \nonumber\\
& - 8 \, {\langle {H_2} \rangle_{AB}} \left( {\langle {\text{Tr}_{B}}( \{ {\rho^{AB}_0} , {H_2} \} ) \rangle_A} + i \, {\langle {\text{Tr}_{B}}([ {\rho^{AB}_0} , {H_1} ]) \rangle_A} \right) \nonumber\\
&+ 2 f({\rho_0^A}) \left( i\, {\langle [{H_2},{H_1}] \rangle_{AB}} - 2 \left({\langle{H_2^2}\rangle_{AB}} - 3\, {\langle{H_2}\rangle_{AB}^2} \right) \right) \nonumber\\
& - 2i \, {\text{Tr}_A}\left({\text{Tr}_{B}}([ {H_1} , {\rho^{AB}_0} ]) \, {\text{Tr}_{B}}(\{ {H_2} , {\rho^{AB}_0} \}) \right) ~.
\end{align}
Finally, by combining Eqs.~\eqref{eq:00000000019},~\eqref{eq:00000000020} and~\eqref{eq:00000000021}, one readi\-ly recovers the short-time expansion of the linear entropy in Eq.~\eqref{eq:00000000014}.
To gain insights into understanding the results in Eqs.~\eqref{eq:00000000014}--\eqref{eq:00000000018}, in the following, we investigate two cases of interest in view of the nonunitary dynamics of non-Hermitian Hamiltonians. The first case des\-cribes bipartite quantum systems with initial uncorrelated states. The second one addresses a multiparticle system whose non-Hermitian Hamiltonian corresponds to the transverse-field XY model with next-nearest neighbor couplings and a perturbing term given by an all-to-all Ising Hamiltonian with an imaginary exchange coupling.
\subsection{Separable initial pure states}
\label{sec:00000000003A}
Here we specialize the result in Eq.~\eqref{eq:00000000014} to the particular case of uncorrelated initial pure state ${\rho^{AB}_0} = {\rho^A_0} \otimes {\rho^B_0}$, with ${\rho^A_0}$ and ${\rho^B_0}$ normalized pure marginal states, i.e., ${\text{Tr}_{\mu}}[{({\rho^{\mu}_0})^2}] = {\text{Tr}_{\mu}}({\rho^{\mu}_0}) = 1$ for all $\mu = \{A,B,AB\}$. We consider the non-Hermitian Hamiltonian $H = {H_1} + i{H_2}$ with ${H_1} = {\sum_n}\,{A_n}\otimes{B_n}$ and ${H_2} = {\sum_n}\,{C_n}\otimes{D_n}$, where ${A_n}, {C_n} \in {\mathcal{H}_A}$ and ${B_n}, {D_n} \in {\mathcal{H}_B}$ represent non\-commu\-ting local obser\-va\-bles. In this setting, one can prove that both the coefficients $1/{T_{1,h}} = 0$ [see Eq.~\eqref{eq:00000000015}] and $1/{T_{1,nh}} = 0$ [see Eq.~\eqref{eq:00000000016}] identically vanish, and the linear entropy in Eq.~\eqref{eq:00000000014} becomes
\begin{equation}
\label{eq:00000000022}
{S_L}({\widetilde{\rho}^A_t}) \approx - \frac{d_A}{({d_A} - 1)}\left(\frac{1}{T_{2,h}^2} + \frac{1}{T_{2,nh}^2} \right){t^2} + O({t^3}) ~,
\end{equation}
with the following nonzero coefficients
\begin{align}
\label{eq:00000000023}
{T_{2,h}^{-2}} &= - 2\, {\sum_{k,l}}\left({\langle{A_k}{A_l}\rangle_A} - {\langle{A_k}\rangle_A}{\langle{A_l}\rangle_A}\right)\nonumber\\
&\times\left({\langle{B_k}{B_l}\rangle_B} - {\langle{B_k}\rangle_B}{\langle{B_l}\rangle_B}\right) ~,
\end{align}
and
\begin{align}
\label{eq:00000000024}
{T_{2,nh}^{-2}} &= - 2\, {\sum_{k,l}}\left({\langle{C_k}{C_l}\rangle_A} - {\langle{C_k}\rangle_A}{\langle{C_l}\rangle_A}\right)\nonumber\\
&\times\left({\langle{D_k}{D_l}\rangle_B} - {\langle{D_k}\rangle_B}{\langle{D_l}\rangle_B}\right) \nonumber\\
& + 4 \, {\sum_{k,l}}\, \text{Im}[ \left({\langle{A_k}{C_l}\rangle_A} - {\langle{A_k}\rangle_A}{\langle{C_l}\rangle_A}\right)\nonumber\\
&\times\left({\langle{B_k}{D_l}\rangle_B} - {\langle{B_k}\rangle_B}{\langle{D_l}\rangle_B}\right)] ~.
\end{align}
Overall, Eq.~\eqref{eq:00000000022} implies that the linear entropy varies quadratically at earlier times of the dynamics. We see that $1/{T_{2,h}}$ is proportional to the so-called correlated quantum uncertainty of observables ${A_n} \in {\mathcal{H}_A}$ and ${B_n} \in {\mathcal{H}_B}$, thus being entirely determined by the expectation values of these operators with respect to the initial marginal states $\rho^{A,B}_0$. It is noteworthy that $|1/{T_{2,h}}|$ assigns a universal timescale for two initially pure subsystems to become entangled by means of the coupling with a Hermitian Hamiltonian $H_1$~\cite{PhysRevLett.77.207,PhysRevD.97.066008,PhysRevA.97.022317}.
In turn, the coefficient $1/{T_{2,nh}}$ depends on the correlated quantum uncertainty of observables ${C_n} \in {\mathcal{H}_A}$ and ${D_n} \in {\mathcal{H}_B}$, and the imaginary part of cross-correlations of the set of local observables. We see that $1/{T_{2,nh}}$ re\-pre\-sents a true signature of the non-Hermitian features of $H$ in the entanglement dynamics. Thus, in addition to the coefficient $1/{T_{2,h}}$, the effective non-Hermitian Hamiltonian induces the factor $1/{T_{2,nh}}$ on the entanglement timescale for initially separable states. In particular, one verifies that $1/{T_{2,nh}} = 0$ vanishes in the Hermitian limit $H = H^{\dagger} = H_1$, i.e., when choosing zero valued obser\-va\-bles $C_n$ and $D_n$.
\subsection{$1$D quantum many-body systems}
\label{sec:00000000003B}
%
\begin{figure*}[!th]
\begin{center}
\includegraphics[scale=1.0]{FIG_03.pdf}
\caption{(Color online) Plot of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$ for the $k$-particle reduced density matrix ${\widetilde{\rho}_t^A}$, as a function of the dimensionless parameter $Jt$. The nonunitary evolution of subsystem $A$ is governed by Eq.~\eqref{eq:00000000012}, with $H_1$ being the transverse field XY Hamiltonian in Eq.~\eqref{eq:00000000025}, and $H_2$ as the all-to-all Ising model in Eq.~\eqref{eq:00000000026}. The system $A + B$ is initialized in the GHZ mixed state ${\rho^{AB}_0} = ((1 - p)/d) \mathbb{I} + p\,|{\text{GHZ}_N}\rangle\langle{\text{GHZ}_N}|$, where $|{\text{GHZ}_N}\rangle = ({1}/{\sqrt{2}})(\,{|0\rangle^{\otimes N}} + {|1\rangle^{\otimes N}} )$. Here we set $N = 8$, $\gamma = 0.75$, ${J_z}/J = 0.5$, and the mixing parameter $p = 0.5$. The blue solid lines correspond to the exact expression of linear entropy in Eq.~\eqref{eq:00000000013}, and the red dashed lines indicates its the short-time perturbative expansion in Eq.~\eqref{eq:00000000014} [see also Eqs.~\eqref{eq:00000000029},~\eqref{eq:00000000030}, and~\eqref{eq:00000000031}].}
\label{fig:FIG03}
\end{center}
\end{figure*}
We set the non-Hermitian Hamiltonian $H = {H_1} + i{H_2}$, where ${H_1}$ describes the transverse-field XY model with open boundary conditions as~\cite{LIEB1961407,PhysRevA.2.1075,PhysRevA.3.786,PhysRevA.3.2137,PhysRevA.4.2331}
\begin{equation}
\label{eq:00000000025}
{H_1} = - J {\sum_{j = 1}^{N - 1}}\left({\gamma_+}{\sigma_j^x}{\sigma_{j + 1}^x} + {\gamma_-}{\sigma_j^y}{\sigma_{j + 1}^y}\right) - h {\sum_{j = 1}^N}\,{\sigma_j^z} ~,
\end{equation}
where $J$ is the coupling constant, $h$ represents the external magnetic field along the $z$-axis, and $\gamma_{\pm} = (1 \pm \gamma)/2$, with $\gamma$ being the anisotropy parameter. For $\gamma = 0$ this Hamiltonian reduces to the isotropic XX model, while for $\gamma = \pm 1$ we recover the Ising model. Furthermore, this model exhibits phase transitions at the isotropic line $\gamma = 0$ ($|h| \leq 1$), and at the critical magnetic field $|h| = 1$. In turn, $H_2$ denotes the many-body fully connected quantum Ising model given by
\begin{equation}
\label{eq:00000000026}
{H_2} = \frac{J_z}{N}\,{\sum_{j < l}}\,{\sigma_j^z}{\sigma_l^z} ~,
\end{equation}
where $J_z$ is the coupling strength, $N$ is the number of spins, and $\{ \sigma^{x,y,z}_s\}_{s = 1,\ldots,N}$ are the Pauli matrices.
\begin{figure*}[!th]
\begin{center}
\includegraphics[scale=1.0]{FIG_04.pdf}
\caption{(Color online) Plot of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$ in the short-time approximation [see Eq.~\eqref{eq:00000000014}], for the $k$-particle reduced density matrix ${\widetilde{\rho}_t^A}$, as a function of the dimensionless parameter $Jt$. The nonunitary evolution of subsystem $A$ is governed by Eq.~\eqref{eq:00000000012}, with $H_1$ being the transverse field XY Hamiltonian in Eq.~\eqref{eq:00000000025}, and $H_2$ as the all-to-all Ising model in Eq.~\eqref{eq:00000000026}. The system $A + B$ is initialized in the GHZ mixed state ${\rho^{AB}_0} = ((1 - p)/d) \mathbb{I} + p\,|{\text{GHZ}_N}\rangle\langle{\text{GHZ}_N}|$, where $|{\text{GHZ}_N}\rangle = ({1}/{\sqrt{2}})(\,{|0\rangle^{\otimes N}} + {|1\rangle^{\otimes N}} )$. Here we set $N = 8$, $\gamma = 0.75$, and the mixing parameter $p = 0.5$. The blue solid lines correspond to the case ${J_z}/J = 0.5$, and the red dashed lines depict the case ${J_z}/J = 0$.}
\label{fig:FIG04}
\end{center}
\end{figure*}
We consider a bipartition into first sequential $k$ sites ($1,\ldots,k$) as the subsystem $A$, and its complement of sequential $N - k$ sites ($k + 1, \ldots, N$) as subsytem $B$. The system $A + B$ is initialized in the mixed state
\begin{equation}
\label{eq:00000000027}
{\rho^{AB}_0} = \left(\frac{1 - p}{d}\right)\mathbb{I} + p\,|{\text{GHZ}_N}\rangle\langle{\text{GHZ}_N}| ~,
\end{equation}
with $d = 2^N$, $0 \leq p \leq 1$, and $|{\text{GHZ}_N}\rangle$ is the GHZ state of $N$ particles defined as
\begin{equation}
\label{eq:00000000028}
|{\text{GHZ}_N}\rangle = \frac{1}{\sqrt{2}}\left(\,{|0\rangle^{\otimes N}} + {|1\rangle^{\otimes N}}\right) ~,
\end{equation}
and its purity is written as $f({\rho^{AB}_0}) = (1/{2^N})(1 + ({2^N} - 1)\,{p^2})$. Furthermore, one can evaluate the averaged va\-lues ${\langle{H_2}\rangle_{AB}} = ({J_z}/2){(N - 1)p}$ and ${\langle{H_2^2}\rangle_{AB}} = {J_z^2}[(N - 1)/(4N)]( 2 + (N - 2)(N + 1)p)$ of the observable $H_2$ respective to the probe state of system $A + B$. The many-body state ${\rho^{AB}_0}$ undergoes a nonunitary evolution gene\-rated by the non-Hermitian Hamiltonian $H = {H_1} + i{H_2}$, and the subsystem $A$ is described by the reduced $k$-particle state ${\widetilde{\rho}^A_t} = {\text{Tr}_{N - k}}({\widetilde{\rho}^{AB}_t})$ whose dynamics is go\-ver\-ned by Eq.~\eqref{eq:00000000012}. The linear entropy ${S_L}({\widetilde{\rho}^A_t})$ of this marginal state is given in Eq.~\eqref{eq:00000000013}, which in turn reduces to ${S_L}({\rho_0^A}) = [{d_A}/({d_A} - 1)][1 - (1/{2^k})(1 + ({2^{k - 1}} - 1)\, {p^2})]$ at time $t = 0$. The short-time expansion of ${S_L}({\widetilde{\rho}^A_t})$ is given in Eq.~\eqref{eq:00000000014}. In this setting, it is possible to verify that the coefficient $1/{T_{1,h}} = 0$ vanishes [see Eq.~\eqref{eq:00000000015}], while Eq.~\eqref{eq:00000000016} implies the following nonzero contribution
\begin{align}
\label{eq:00000000029}
&{T^{-1}_{1,nh}} =\nonumber\\
& - \frac{{J_z} \, p (1 - p)}{{2^{k - 1}} N}\left( k(k - 1) + N(N - 1)({2^{k - 1}} - 1)p \right) ~.
\end{align}
Noteworthy, Eq.~\eqref{eq:00000000029} shows that $1/{T_{1,nh}}$ exhibits a polynomial dependence on the mixing parameter $p$, thus being a negative quantity for all $0 < p < 1$, and $k \in \{1,\ldots,N\}$. In particular, it follows that $1/{T_{1,nh}} = 0$ for the initial pure state $|{\text{GHZ}_N}\rangle\langle{\text{GHZ}_N}|$ ($p = 1$) and also for the maximally mixed state $\mathbb{I}/d$ ($p = 0$). We find that $1/{T_{1,nh}}$ is proportional to the coupling strength $J_z$, and identically vanishes in the Hermitian limit (${J_z} = 0$).
Next, by using Eq.~\eqref{eq:00000000017}, we obtain
\begin{equation}
\label{eq:00000000030}
{T^{-2}_{2,h}} = ({\delta_{{d_B},2}} - 1){\gamma^2}{J^2}{p^2} ~,
\end{equation}
which depends on the coupling $J$, anisotropy parameter $\gamma$, and va\-ni\-shes whenever $\mathcal{H}_B$ is a two-dimensional subspace, i.e., one gets $1/{T_{2,h}} = 0$ for the case ${d_B} = 2$. Finally, by applying Eq.~\eqref{eq:00000000018} and performing lengthy calculations, one obtains the result
\begin{align}
\label{eq:00000000031}
&{T^{-2}_{2,nh}} = - \frac{{J_z^2}}{{2^{k - 1}} \, {N^2}} \left\{ 3{N^2}{(N - 1)^2}(1 - {2^{k - 1}}) \, {p^4} \right. \nonumber\\
&\left. + N(N - 1)[({2^{k - 1}} - 1)(5N(N - 1) - 2) - 4k(k - 1)] \, {p^3} \right. \nonumber\\
&\left. + [ k({k^2}(k - 6) + k + 4) + 2{N}{k}(k - 1)(3N - 1) \right.\nonumber\\
&\left. - 2{N}({2^{k - 1}} - 1)({N^3} - 2{N^2} + 1)] \, {p^2} - 2k(k - 1) \right.\nonumber\\
&\left. - k(k - 1)[k(k - 5) + 2((N - 1)(N + 2) - 1)] \, {p} \right\} ~.
\end{align}
We find that $1/{T^{2}_{2,nh}}$ behaves polynomially with the mi\-xing parameter $p$. In particular, it follows that $1/{T^{2}_{2,nh}} = 0$ for $p = 1$, while for $p = 0$ one obtains that $1/{T^{2}_{2,nh}} = {2^{2 - k}} {J_z^2}k(k - 1)/{N^2}$. Hence, for $N \gg k$, the latter case implies $1/{T^{2}_{2,nh}} \sim {J_z^2}/{N^2}$ for the initial maximally mixed state ($p = 0$), i.e, it scales with the inverse square of the number of particles.
\begin{figure}[!th]
\begin{center}
\includegraphics[scale=0.9]{FIG_05.pdf}
\caption{(Color online) Plot of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$ [see Eq.~\eqref{eq:00000000013}] in the short-time approximation [see Eq.~\eqref{eq:00000000014}], for the $k$-particle reduced density matrix ${\widetilde{\rho}_t^A}$, as a function of the dimensionless parameter $Jt$. The nonunitary evolution of subsystem $A$ is governed by Eq.~\eqref{eq:00000000012}, with $H_1$ being the transverse field XY Hamiltonian in Eq.~\eqref{eq:00000000025}, and $H_2$ as the all-to-all Ising model in Eq.~\eqref{eq:00000000026}. The system $A + B$ is initialized in the GHZ mixed state ${\rho^{AB}_0} = ((1 - p)/d) \mathbb{I} + p\,|{\text{GHZ}_N}\rangle\langle{\text{GHZ}_N}|$, where $|{\text{GHZ}_N}\rangle = ({1}/{\sqrt{2}})(\,{|0\rangle^{\otimes N}} + {|1\rangle^{\otimes N}} )$. Here we consider the system size $N = 8$, with $k = 5$, and $\gamma = 0.75$. The blue solid lines correspond to the case ${J_z}/J = 0.5$, while the red dashed lines depict the case ${J_z}/J = 0$.}
\label{fig:FIG05}
\end{center}
\end{figure}
In the following, we will numerically address the short-time dynamics of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$ in Eq.~\eqref{eq:00000000014}. The system $A + B$ is initialized at the GHZ mixed state in Eq.~\eqref{eq:00000000027}, with $H_1$ being the transverse field XY Hamiltonian in Eq.~\eqref{eq:00000000025}, and $H_2$ stand for the all-to-all Ising model in Eq.~\eqref{eq:00000000026}. In this case, bearing in mind that $1/{T_{1,h}} = 0$, we also apply the results in Eqs.~\eqref{eq:00000000029},~\eqref{eq:00000000030},~\eqref{eq:00000000031}. Unless otherwise stated, we set the system size $N = 8$, the anisotropy parameter $\gamma = 0.75$, and the ratio ${J_z}/J = 0.5$.
In Fig.~\ref{fig:FIG03} we plot of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$, as a function of the dimensionless parameter $Jt$, and set the mixing parameter $p = 0.5$. The solid blue line refers to the exact linear entropy ${S_L}({\widetilde{\rho}^A_t})$ [see Eq.~\eqref{eq:00000000013}], while the red dashed line depicts the short-time expansion of this quantity in Eq.~\eqref{eq:00000000014}. We consider the subsystem $A$ with the number of sites $k \in \{2,\ldots,7\}$, and respective dimensions ${d_A} = \{ {2^k} \}_{k = 2,\ldots,7}$. Figures~\ref{fig:FIG03}(a)-\ref{fig:FIG03}(f) show that the short-time expansion of ${S_L}({\widetilde{\rho}^A_t})$ reproduces its growth at early times. In each panel, we find an accurate quantitative agreement for $0 \leq Jt \lesssim 0.1$. Nevertheless, for $Jt \gtrsim 0.1$, we have that the result in Eq.~\eqref{eq:00000000013} becomes loose and it is bounded from above by the exact linear entropy in Eq.~\eqref{eq:00000000014}, failing to predict the entanglement dynamics within the subsystem $A$.
Next, Fig.~\ref{fig:FIG04} shows the short-time dynamics of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$ in Eq.~\eqref{eq:00000000014}, for the initial state $\rho_0^{AB}$ with $p = 0.5$ [see Eq.~\eqref{eq:00000000027}]. The blue solid lines correspond to the case ${J_z}/J = 0.5$ (non-Hermitian Hamiltonian), while the red dashed lines depict the case ${J_z}/J = 0$ (Hermitian Hamiltonian). Overall, Figs.~\ref{fig:FIG04}(a)--\ref{fig:FIG04}(e) show that the linear entropy is a concave function for ${J_z}/J \neq 0$, while it is a convex function for ${J_z}/J = 0$. In turn, Fig.~\ref{fig:FIG04}(f) shows that ${S_L}({\widetilde{\rho}^A_t})$ stand as a concave function for ${J_z}/J \neq 0$, while for ${J_z}/J = 0$ it saturates to a fixed value for all $Jt > 0$. To see this, we first note that $1/{T_{1,nh}} = 0$ and $1/{T_{2,nh}} = 0$ for $J = 0$ [see Eqs.~\eqref{eq:00000000029} and~\eqref{eq:00000000031}, respectively], while one readily obtains that $1/{T_{2,h}} = 0$ as the subsystem $B$ has dimension $d_B = 2$ [see Eq.~\eqref{eq:00000000030}]. Hence, bearing in mind that $1/{T_{1,h}} = 0$, it follows that the linear entropy ${S_L}({\widetilde{\rho}^A_t}) \approx {S_L}({\rho^A_0})$ is time-independent in the short-time approximation, for ${J_z}/J = 0$.
In Fig.~\ref{fig:FIG05} we display the short-time dynamics of the li\-near entropy ${S_L}({\widetilde{\rho}^A_t})$ in Eq.~\eqref{eq:00000000014}, as a function of $Jt$, and consider the subsystem $A$ with $k = 5$ sites. We emphasize that each of the blue solid lines corresponds to the case ${J_z}/J = 0.5$ (non-Hermitian Hamiltonian), and the red ones represent the case ${J_z}/J = 0$ (Hermitian Hamiltonian). We set the mixing parameters $p = 0.25$ [see Fig.~\ref{fig:FIG05}(a)], $p = 0.5$ [see Fig.~\ref{fig:FIG05}(b)], $p = 0.75$ [see Fig.~\ref{fig:FIG05}(c)], and $p = 1$ [see Fig.~\ref{fig:FIG05}(d)]. In Figs.~\ref{fig:FIG05}(a) and~\ref{fig:FIG05}(b), one finds that the li\-near entropy is concave whenever ${J_z}/J \neq 0$, while it turns into a convex function in the Hermitian limit with ${J_z}/J = 0$. In Fig.~\ref{fig:FIG05}(c), the linear entropy turn to be a convex function, which is due to the fact that $1/{T_{2,h}} + 1/{T_{2,nh}} < 0$ for $p = 0.75$. Figure~\ref{fig:FIG05}(d) show that, for $p = 1$, the two linear entropies coincide regardless of the generator $H$. Indeed, we have seen from Eqs.~\eqref{eq:00000000029} and~\eqref{eq:00000000031} that $1/{T_{1,nh}} = 0$ and $1/{T_{2,nh}} = 0$ for initial pure states, respectively. In this case, given that $1/{T_{1,h}} = 0$ and ${d_B} = 8$, the onset growth of the linear entropy satisfies ${S_L}({\widetilde{\rho}^A_t}) \approx (16/31)(1 + 2 {\gamma^2}{J^2}{t^2})+ O({t^3})$ [see Eqs.~\eqref{eq:00000000014} and~\eqref{eq:00000000030}].
As a final remark, Figs.~\ref{fig:FIG04} and~\ref{fig:FIG05} show that the non-Hermitian Hamiltonian (${J_z}/J \neq 0$) enhances the short-time dynamics of ${S_L}({\widetilde{\rho}^A_t})$, which bounds from above the respective linear entropy for the Hermitian Hamiltonian (${J_z}/J = 0$). This result suggests that non-Hermitian Hamiltonians can play the role of a resource in the short-time dynamics of subsystem correlations, with potential use in some quantum metrological tasks as shown recently in Refs.~\cite{PhysRevLett.125.240506,PhysRevLett.127.026404}. Lastly, Figs.~\ref{fig:FIG04} and~\ref{fig:FIG05} show a crossover behavior between both the non-Hermitian (${J_z}/J \neq 0$) and Hermitian (${J_z}/J = 0$) cases, but it should be noted that it occurs in a time window that extrapolates the validity of the short-time approximation. Indeed, Fig.~\ref{fig:FIG03} shows that our results find good agreement with the numerical simulation of ${S_L}({\widetilde{\rho}^A_t})$ in Eq.~\eqref{eq:00000000013} for $0 \leq Jt \lesssim 0.1$.
\section{Discussion and Conclusions}
\label{sec:00000000004}
In this paper, we discuss the timescales related to the onset growth of linear entropy for finite-dimensional quantum systems described by effective non-Hermitian Hamiltonians. We investigate the short-time perturbative expansion of the linear entropy for a given input state driven by a general non-Hermitian Hamiltonian. We emphasize that our approach takes in account initial quantum states that can be either pure or mixed, possibly entangled or even uncorrelated states.
We address the degree of mixedness of a quantum state that undergoes the nonunitary dynamics generated by an effective non-Hermitian Hamiltonian $H = {H_1} + i{H_2}$ [see Sec.~\ref{sec:00000000002}]. In this setting, Eq.~\eqref{eq:00000000003} stands for the short-time expansion of the linear entropy up to second order in time $t$, around $t = 0$, which in turn depends on the coefficients $1/{T_1}$ and $1/{T_2}$ in Eqs.~\eqref{eq:00000000004} and~\eqref{eq:00000000005}, respectively. Both quantities can be evaluated once the input state and the Hamiltonian $H$ have been specified. We emphasize that Eqs.~\eqref{eq:00000000004} and~\eqref{eq:00000000005} provide two competing timescales in the initial growth of the mixedness of the evolved state at earlier times of the dynamics. In particular, both coefficients vanish whenever the system is initialized in a pure state, regardless of the non-Hermitian part of the Hamiltonian. Moreover, in the Hermitian limit, we have that $1/{T_1} = 0$ and $1/{T_2} = 0$ independently of the initial state of the system. We note that, since the linear entropy defines a conserved quantity for Hermitian quantum systems, it can be proved that any of the coefficients in its perturbative expansion must vanish in this limiting case [see Eq.~\eqref{eq:00000000003}].
We specialize these results to the case of a dissipative non-Hermitian two-level system initialized in a mixed single-qubit state [see Sec.~\ref{sec:00000000002A}]. We find analytical expressions for the coefficients $1/T_1$ [see Eq.~\eqref{eq:00000000010}] and $1/T_2$ [see Eq.~\eqref{eq:00000000011}] in terms of the Bloch sphere parameters. In this case, we compare the exact linear entropy ${S_L}({\rho_t})$ with its aforementioned short-time expansion around $t = 0$. We find good quantitative agreement between these two quantities at earlier times of the dynamics. Of course, for later times one should include higher orders in the Taylor expansion to obtain tighter results for the mixedness of the evolved state.
Next, we investigate the reduced dynamics of composite systems described by non-Hermitian Hamiltonians [see Sec.~\ref{sec:00000000003}]. We derived the short-time perturbative expansion of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$ for a given time-dependent marginal state of a bipartite system [see Eq.~\eqref{eq:00000000014}]. We found that, up to the second order in time $t$, the growth of the linear entropy is governed by the coefficients $1/{T_{1,h}}$ and $1/{T_{1,nh}}$ in Eqs.~\eqref{eq:00000000015} and~\eqref{eq:00000000016}, respectively, and also $1/{T_{2,h}}$ and $1/{T_{2,nh}}$ in Eqs.~\eqref{eq:00000000017} and~\eqref{eq:00000000018}, respectively. On the one hand, one gets that $1/{T_{1,h}}$ and $1/{T_{2,h}}$ depend on $H_1$ and the input state of the system. On the other hand, we have that $1/{T_{1,nh}}$ and $1/{T_{2,nh}}$ depend on $H_2$, thus being intrinsically related to the non-Hermitian features of the Hamiltonian. In the Hermitian limit, i.e., when one sets $H_2$ being a zero-valued operator, we find $1/{T_{1,nh}} = 0$ and $1/{T_{2,nh}} = 0$ for any bipartite system. In this case, Eqs.~\eqref{eq:00000000015} and~\eqref{eq:00000000017} remain nonzero and recover the perturbative expansion of the idempotency defect measuring the coherence losses for composite systems described by Hermitian Hamiltonians~\cite{PhysRevLett.77.207}.
In particular, specifying an initial pure and uncorrelated state, we find the vanishing coefficients $1/{T_{1,h}} = 0$ and $1/{T_{1,nh}} = 0$, and the lowest order of the short-time perturbative expansion of the linear entropy ${S_L}({\widetilde{\rho}^A_t})$ depends on $1/{T_{2,h}^2}$ and $1/{T_{2,nh}^2}$ that are given in Eqs.~\eqref{eq:00000000023} and~\eqref{eq:00000000024}, respectively [see Sec.~\ref{sec:00000000003A}]. In the Hermitian limit, the latter result identically vanishes, and $|1/{T_{2,h}}|$ signals the entanglement timescale for quantum systems described by Hermitian Hamiltonians~\cite{PhysRevD.97.066008}. It is noteworthy that this result is also related to the timescale that governs the growth of entanglement for R\'{e}nyi entropies~\cite{PhysRevA.97.022317}.
To illustrate these findings, we investigated the linear entropy of the $k$-particle evolved marginal state for a quantum many-body system described by the transverse-field XY model coupled to the imaginary fully connected Ising Hamiltonian [see Sec.~\ref{sec:00000000003B}]. We found analytical expressions for $1/{T_{1,nh}}$ and $1/{T_{2,nh}^2}$, which in turn scale linearly with the coupling strength of the all-to-all Ising Hamiltonian [see Eqs.~\eqref{eq:00000000029} and~\eqref{eq:00000000031}, respectively]. In addition, it follows that $1/{T_{1,h}}$ vanishes, while $1/{T_{2,h}^2}$ depends on the anisotropy para\-me\-ter of the XY model [see Eq.~\eqref{eq:00000000030}]. We compared the short-time expansion of the linear entropy with its exact numerical simulation [see Fig.~\ref{fig:FIG03}], and discussed its dynamical behavior in both cases of non-Hermitian and Hermitian Hamiltonians [see Figs.~\ref{fig:FIG04} and~\ref{fig:FIG05}]. We find that non-Hermiticity enhances the short-time dynamics of the linear entropy, pro\-vi\-ding an upper bound for the respective linear entropy for Hermitian Hamiltonian.
Our findings provide insightful qualitative and quantitative information about the initial growth of linear entropy at early times.
Importantly, the results require low computational cost and their evaluation involves minimal ingredients as the initial state and the non-Hermitian Hamiltonian that governs the nonunitary dynamics. This might be of interest for higher dimensional systems, where evaluating the linear entropy would require the full spectral decomposition of the evolved system. We point out that one could generalize the present discussion in terms of $\alpha$-R\'{e}nyi entropies~\cite{PhysRevA.97.022317}. Furthermore, one can investigate the interplay of the aforementioned timescales and the quantum speed limit for nonunitary evolutions generated by non-Hermitian Hamiltonians~\cite{PhysRevA.106.012403,PhysRevA.104.052223}. We hope to address these questions in further investigations. The results in this paper could find applications in the subjects of non-Hermitian quantum sensing~\cite{Lau2018,PhysRevA.103.042418}, quantum thermodynamics of non-Hermitian systems~\cite{deffner_scirep_23408_2016}, non-Hermitian long-range interacting quantum systems~\cite{arXiv:2109.01063}, and $\mathcal{PT}$-symmetric quantum field theory~\cite{Bender_2020}.
\begin{acknowledgments}
This work was supported by the Brazilian ministries MEC and MCTIC, and the Brazilian funding agencies CNPq, and Coordena\c{c}\~{a}o de Aperfei\c{c}oamento de Pessoal de N\'{i}vel Superior--Brasil (CAPES) (Finance Code 001). D. P. P. acknowledges Funda\c{c}\~{a}o de Amparo \`{a} Pesquisa e ao Desenvolvimento Cient\'{i}fico e Tecnol\'{o}gico do Maranh\~{a}o (FAPEMA). T. M. acknowledges the hospitality of ITAMP-Harvard where part of this work was done. T. M. acknowledges support from CAPES. This work was supported by the Serrapilheira Institute (Grant No. Serra-1812-27802).
\end{acknowledgments}
|
\section{Introduction}
Let $C$ be a compact complex manifold. We say that two holomorphic embeddings $f\colon C \hookrightarrow M$ and $\tilde f\colon C \hookrightarrow\tilde M$ are {\it holomorphic equivalent} if there is a biholomorphic mapping $F$ from a neighborhood of
$f(C)$ in $M$ into
a neighborhood of
$\tilde f(C)$ in $\tilde M$ such that $F f=\tilde f$. To classify such neighborhoods $M$, we identity $C$ with $f(C)$ via $f$. We also fix the normal bundle $N_C$ of $C$ in $M$ and $T_CM$, the restriction of $TM$ on $C$. To determine the holomorphic equivalence of two neighborhoods, Grauert~~\cite{MR0137127} introduced
the \emph{formal principle} that
asserts two neighborhoods of $C$ are holomorphic equivalent if they are formally equivalent. The formal principle holds for neighborhoods when $N_C$ has nice geometry properties. When $N_C$ is \emph{weakly negative}, Grauert~~\cite{MR0137127} proved that the formal principle holds when $C$ has codimension one. Hironaka and Rossi~\cite{MR0171784} extended Grauert's result to arbitrary positive codimension as well as to the reduced complex spaces $C$ that are exceptional in the sense of Grauert. Griffiths~\cite{MR206980} showed that the formal principle holds when $N_C$ is \emph{sufficiently positive} and $\dim C\geq3$. This result was improved by Hirschowitz~\cite{MR621013} including $\dim C\geq2$ with \emph{weakly positive} $N_C$,
and then by Commichau-Grauert~~\cite{MR627752} for $1$-\emph{positive} normal bundle $N_C$ for $\dim C\geq1$.
Despite all these positive results on the formal principle, it remains unknown until now if the equivalence classes of neighborhoods of $C$ with a fixed $T_CM$ form a \emph{finite} dimensional space
when the formal principle holds.
The main result of this paper provides a structure for the neighborhoods of $C$ as follows.
\begin{thm}\label{m-finite} Let $C$ be a compact complex manifold.
Assume that $N_C$ is either weakly negative or $2$-positive.
There is an injective mapping from the set of holomorphic equivalence classes of neighborhoods of $C$ into the finite-dimensional space
$$
\mathcal H^1(T_CM):=\bigoplus_{\ell\geq2} H^1(C,T_CM\otimes S^\ell N_C^*),
$$
where $S^\ell N_C^*$ is the $\ell$-th symmetric power of the dual bundle $N_C^*$ of $N_C$.
\end{thm}
Note that $\dim\mathcal H^1(T_CM)<\infty$ follows from the assumption on $N_C$.
Indeed, by definitions in~\cite[Def.~1, p.~342]{MR0137127} and ~\cite[Def.~1, p.~108]{MR627752}, $N_C$ is \emph{weakly negative} (resp. $q$-positive) if the zero section of $N_C$ admits a tubular neighborhood
$Tub(C)$
in $N_C$ whose boundary is strictly pseudoconvex (resp. of at least
$q$ negative Levi eigenvalues).
It is known that $\dim \mathcal H^1(T_CM)$ is finite, when the zero section of $N_C$ admits a relatively compact neighborhood $W$ in $N_C$ that is either weakly negative or $2$-positive.
The formal principle does not rule out the existence of an
infinite dimensional moduli space, i.e. equivalence classes.
In fact, given a (compact and smooth) Riemann surface $C$ and a \emph{positive} line bundle $N_C$, Morrow-Rossi~~\cite[p.~323]{MR627765} constructed a
complete set of the equivalence classes of holomorphic transverse foliations of such $C$, under the smaller group that preserves the foliations, showing that the moduli space for the equivalent classes is infinite-dimensional. See \rp{realization-MR} for details.
For other closely related results, we mention that when $C$ is the Riemann sphere and $\operatorname{rank} N_C=1$, Hurtubise-Kamran~~\cite{MR1152943} showed that the space of neighborhoods with $1$-positive $N_C$ is infinite-dimensional, and Mishustin~~\cite{MR1250979} constructed a normal form with infinite-dimensional invariants.
See Ilyashenko~~\cite{MR704627} when $C$ is elliptic curve and $N_C$ is a positive line bundle
and recent results of
Falla Luza and Loray~~\cite{MR3488114} on the Riemann sphere.
On the other hand, when $C$ is an elliptic curve embedded in a complex surface with $N_C$ topologically trivial, Arnol'd~~\cite{MR0431285} showed that the formal principle
holds under an extra condition, namely a suitable {\it Diophantine condition} on $T_CM$ when it splits as the direct sum $TC\oplus N_C$. Doing so, Arnol'd~ showed for the first time that the formal principle \emph{fails} when $N_C$ is topologically trivial for an elliptic curve $C$ and a certain Diophantine condition is violated (see ~~\cite[sect. 5.4]{GS3} for construction of counter-examples). Arnold's theorem was extended by Ilyashenko and Pjartli~~\cite{MR549623} to the case when $C$ is the product of finitely many elliptic curves and $N_C$ is the direct sum of line bundles, and by the authors~~\cite{GS4} for a complex torus with a Hermitian flat $N_C$ satisfying a Diophantine condition. Of course, the study of neighborhood of embedded compact complex manifolds has a long history. The reader is referred to~~\cite{GS3} and references therein on neighborhoods of compact manifolds. See also recent work of Hwang~\cite{hwang-annals}, Koike~~\cite{koike-fourier}, and Loray-Thom-Touzet~~\cite{loray-moscou}.
The paper is organized as follows.
In Section 2, we construct a formal normal form for holomorphic neighborhoods of $C$ that is realized as a subset of
$
\mathcal H^1(T_CM)$.
Our injectivity assertion in \rt{m-finite} remains true if $N_C$ is merely $1$-positive
and $H^0(C,T_CM\otimes S^\ell N_C^*)=0$ for all $\ell>1$.
In Section 3, we find a formal normal form for \emph{tangential} foliations of neighborhoods of $C$ that contain $C$ as a leaf.
In Section 4, we find a formal normal form for transverse foliations of neighborhoods of $C$. In Section 5, we apply the theorems in~\cite{MR0137127, MR627752} to show \rt{m-finite} and analogous classification for transverse foliations. In Section 6, we use a theorem of Camacho-Movasati-Sad~~\cite{MR1967036} to show that when the genus of the compact Riemann surface $C$ is bigger than one, there are neighborhoods of $C$ that are not linearizable. Therefore, the equivalence classes contain at least two elements.
{\bf Acknowledgments.} Part of work was carried out when X.~G. was supported by CNRS and UCA for a visiting position at UCA.
\setcounter{thm}{0}\setcounter{equation}{0}
\section{A formal normal form to classify neighborhoods}
We have mentioned Grauert's formal principle asserting two neighborhoods are holomorphic equivalent if they are formally equivalent.
To further motivate our results, let us describe the following problems about the classifications of neighborhoods. We recall the \emph{Kuranishi problem} mentioned by Morrow-Rossi~~\cite{MR627765} for the study of neighborhoods of $C$, which is to
construct a {\it moduli space} or a parametrization that classify neighborhoods of $C$ completely and understand the structures of the moduli space.
This paper provides partial answers to this problem. We will also study the analogous problems for the foliations of neighborhoods of $C$.
In this section, we will construct a formal normal form for formal equivalence classes of holomorphic neighborhoods. Certain features of the normal form will be described in details as they will be useful in the convergent proof.
To study classifications, we will use transition functions
for various vector bundles in coordinate charts. Let $C$ be a compact complex manifold embedded in a complex manifold $M$. %
We cover a neighborhood of $C$ in $M$ by open sets $V_j$ and choose coordinate charts
$(z_j,w_j)$ on $V_j$ %
for $M$
such that
$$
U_j:=C\cap V_j=\{w_j=0\}.
$$
To have Leray coverings, we may assume for instance that $U_j,V_j$ are biholomorphic to polydiscs.
Let ${\mathcal U}=\{U_i\}$ be a finite open covering of $C$ with coordinate
charts $z_i=\varphi_i(p)=(z_i^1,\dots, z_i^n)$
defined on $U_i$.
Let
\begin{equation}\label{transitionC}
z_k=
\varphi_{kj}
(z_j)=
\varphi_k\varphi_j^{-1}(z_j)
\nonumber
\end{equation}
be the transition function of $C$ on $U_{kj}:=U_k\cap U_j$. Thus the neighborhood $M$ has transition functions on $V_{kj}:=V_j\cap V_k$
\eq{transitionN-}\nonumber
\Phi_{kj}\colon
\begin{array}{rcl}
z_k &= & \Phi^h_{kj}(z_j,w_j):=
\varphi_{kj}(z_j)+l_{kj}(z_j)w_j+\phi^h_{kj}(z_j,w_j),\vspace{.75ex}
\\
w_k &= &\Phi^v_{kj}(z_j,w_j):=
t_{kj}(z_j)w_j+\phi^v_{kj}(z_j,w_j).
\end{array}
\end{equation}
Here, $\phi^h_{kj}$ (resp. $\phi^v_{kj}$) are holomorphic functions that vanish to order $\geq 2$ along $w_j=0$:
\eq{vord2}\nonumber
\phi_{kj}^h(z_j,w_j)=O(|w_j|^2), \quad \phi_{kj}^v(z_j,w_j)=O(|w_j|^2).
\end{equation}
For abbreviation, we also call $\Phi:=\{\Phi_{kj}\}$ a neighborhood of $C$
and $O(2)$ denotes such a $\phi_{kj}$.
Note that the transition functions of $N_C$ are
\eq{Nkjh}
N_{kj}(h_j,v_j)=(\varphi_{kj}(h_j),t_{kj}(h_j)v_j)
\end{equation}
and the transition functions of $T_{C}M:=(TM)|_C$ are
\begin{equation}\label{transitionTCM}
g_{kj}:=\left(\begin{matrix}s_{kj}& \ell_{kj}\\0&
t_{kj}\end{matrix}\right)
(h_j)\quad \text{on } U_j\cap U_k
\end{equation}
for some $n\times d$ matrix $l_{kj}$,
while $s_{kj}$ is the Jacobian matrix of $\varphi_{kj}$. When $l_{kj}=0$, $T_C M$ splits as $TC\oplus N_C$ (see
\cite[Prop.~2.9]{MR627765}). With $g_{kj}$, $T_CM$ has a basis $\{ \tilde e_k^\mu;\mu=1,\dots,n+ d\}$. Here
\eq{ekhv}
e_k^h=\{\tilde e_k^\tau\colon \tau=1,\dots,n\},
\quad e_k^v=\{\tilde e_k^{n+\nu}\colon\nu=1,\dots, d\}
\end{equation}
with $e_k^v$ being the basis of $N_C$ on $U_k$. To study the foliations, we will choose $e_k^v$ to be a flat basis when $N_C$ is \emph{flat}, i.e. its transition functions are locally constant.
Throughout the paper, we fix
\eq{Nkjad}\nonumber
N_{kj}^*(z_j,w_j)=(\varphi_{kj}(z_j)+\ell_{kj}(z_j)w_j,t_{kj}(z_j)w_j).
\end{equation}
We emphasize that $\{N^*_{kj}\}$ does not necessarily define a neighborhood of $C$ as it may not be a cocycle. Thus, we introduce the following.
\begin{defn}\label{admN} Fix $T_CM$.
Fix a holomorphic neighborhood $N^1=\{N^1_{kj}\}$ of $C$ such that $N^1_{kj}=N^*_{kj}+O(2)$.
If $T_CM$ splits, we always take $N_{kj}^*=N_{kj}$ defined by \re{Nkjh}.
Let $\mathcal M$ be the set of germs of holomorphic neighborhoods $\Phi$ of $C$ in complex
manifolds $\tilde M$ such that $T_CM=T_C\tilde M$ and $\Phi=N^1+O(2)$.
\end{defn}
Throughout the paper, we assume that $\mathcal M$ is non-empty. Therefore the existence of $N^1$ is ensured.
For the holomorphic (resp. formal) equivalence of two neighborhoods of $C$, we restrict to biholomorphic (resp. formal) mappings $F$ from
$\Phi$ to $\tilde\Phi$ fixing $C$ pointwise. Then $F$ can be expressed as
\begin{equation}\label{change}
F_k: \begin{array}{rcl}
z_k &= & F^h_{k}(h_k,v_k):=h_k+f^h_{k}(h_k,v_k),\vspace{.75ex
\\
w_k &= & F^v_{k}(h_k,v_k):=v_k+f^v_{k}(h_k,v_k)
\end{array}
\end{equation}
with $f_k(h_k,0)=0$ such that $F_k\Phi_{kj}
F_j^{-1}=\tilde \Phi_{kj}$ or $F\Phi F^{-1}=\tilde\Phi$ for short. By formal equivalence, we mean that $f_{k}(h_k,v_k)$ is a formal power series in the variable $v_k$ with holomorphic coefficients in the variables $h_k\in U_k$, for some finite cover of $C$ by open sets $\{U_k\}$.
\begin{defn}
We say that $F=\{F_j\}$ defined by \rea{change} is \emph{tangent to the identity} along $C$ and write $F=Id+f$ with $f=O(2)$, where the latter means that $f^h_{k}(h_k,v_k)$ and $f^v_{k}(h_k,v_k)$ are holomorphic functions vanishing to order $\geq 2$ at $v_k=0$.
Let ${\mathcal M}/{\sim}$ (resp. ${\mathcal M}/{\overset{f}{\sim}}$) the equivalence of \emph{holomorphic} neighborhoods of $C$ under all such biholomorphisms (resp. formal) $F$.
\end{defn}
Next, we identity sections in symmetric powers of $N_C^*$ with coordinate changes. We will associate
\eq{fkmh}
[ f_k]^m(h_k,v_k):=\left\{\left(\sum_{|Q|=m}f^h_{k, Q}(h_k)v_k^Q, \sum_{|Q|=m}f^v_{k, Q}(h_k)v_k^Q\right)\right\}\end{equation}
with a $0$-th cochain $[\tilde f]^m\in C^0(\mathcal U, T_CM\otimes S^mN_C^*)$ defined by
\eq{tfkm}
[\tilde f_k]^m(p):=\sum_{\mu=1}^{n+d}\sum_{|Q|=m}f_{k,Q}^\mu(h_k(p),v_k(p)) \tilde e_k^\mu(p)\otimes (w_k^*(p))^Q.
\end{equation}
Here and in what follow, $Q\in \mathbb{N}^d$. We will also associate
\eq{def-tilde-f}
[ \phi_{kj} ]^m(z_{j},w_{j}):=\sum_{|Q|=m} \phi_{kj;Q} (z_{j}) w_{j}^Q
\end{equation}
with the $1$-cochain $\{[\tilde\phi_{kj}]^m\}\in C^1(\{U_j\},\mathcal O(T_CM\otimes S^m(N_C^*)))$
defined by
\begin{gather}\label{def-tilde-f+}
[\tilde\phi_{kj }]^m(p)= \sum_{\mu=1}^{d+d}\sum_{|Q|=m} \phi_{{kj}; Q}^\mu (h_k(p),v_k(p))\tilde e^\mu_{k }(p)\otimes (w_{j}^*(p))^Q.
\end{gather}
Both associations are ${\mathbb C}$-linear, one-to-one and onto. By abuse of notation, we drop tildes in $\tilde\phi_{kj},\tilde f_k$ and we interchange \re{fkmh} and \re{tfkm} (resp. \re{def-tilde-f} and \re{def-tilde-f+}) for computation as we wish.
We also need to identity the coboundary operator $\delta$ for cochain \re{def-tilde-f+} with a coboundary operator for \re{def-tilde-f}.
By Lemma 2.7 in~~\cite{GS3}, applied to $E=T_CM$, we have
$\widetilde{\delta[f]^m}=[\tilde\phi]^m$ being equivalent to $\delta[f]^m=[\phi]^m$.
Writing in column vector $[f_{ij}]^m:=([f_{ij}^1]^m, \dots, [f_{ij}^{n+d}]^m)^t$ and recalling \re{transitionTCM}, this reads
\begin{align}\label{dfphi}
(\delta [f]^m)_{ij}(z_j,v_j)&:=
g_{ij}[f_j]^m-[f_i^m]\circ N_{ij}
=[\phi_{ij}]^m(z_j,v_j).
\end{align}
To construct a formal normal form for ${\mathcal M}$, recall that $\dim H^q(C,V)<\infty$ for $q>0$ and any holomorphic vector bundle $V$ on $C$; see ~\cite[Thm.~3.20 and Cor., p.~161]{MR2109686}.
\begin{defn}\label{choose-base}Let $C$ be a compact complex manifold and fix a holomorphic vector bundle $N_C$ on $C$. Fix a basis
$[e^m]=([e_1^m], \dots, [e_{k_ m }^m])$ for
$H^1(C, T_CM\otimes S^ m N_C^*)$ that is not zero. We also fix a representative $e_i^m\in Z^1(\mathcal U, T_CM\otimes S^ m N_C^*)$ for $[e_i^m]$.
Define $c^m\cdot e^m:=\sum c^m_ie^m_i$ for $c^m_i\in{\mathbb C}$ and $(c^2,\dots c^m)\cdot(e^2,\dots, e^m)=\sum_{j=2}^m c^j\cdot e^j$. For convenience, if $H^1(C, T_CM\otimes S^ m N_C^*)=0$, set $c^m=0$ and $c^m\cdot e^m=0$.
\end{defn}
\le{the-lem-} Fix $m>1$. Let $\Phi,\tilde \Phi\in \mathcal M$ and let $F=I+f$ be a formal mapping satisfying $f=O(m)$.
\begin{list}{$\hspace{-1em}(\alph{pp})$}{\usecounter{pp}}
\item If $\tilde\Phi-\Phi=O(m)$, then $[\tilde\Phi]^m-[\Phi]^m\in Z^1(\mathcal U,T_CM\otimes S^m N_C^*)$; in particular,
\eq{identify}\nonumber
[\Phi-N^1]^2\in Z^1(\mathcal U,T_CM\otimes S^2 N_C^*),\quad \forall \Phi\in\mathcal M.
\end{equation}
\item If $
F\Phi F^{-1}=\tilde\Phi+O(m)$ holds if and only if
$ \delta [f]^m=
[\Phi]^m
-[\tilde\Phi]^m.
$
\item If $\tilde \Phi=\Phi+O(m)$, there exist a unique $c^m\cdot e^m$ and some $\tilde F=I+O(m)$ such that $\tilde F\tilde\Phi \tilde F^{-1}=\Phi+c^m\cdot e^m+O(m+1)$.
\end{list}
\end{lemma}
\begin{proof}$(a)$ Recall that $N_{kj}^*(h_j,v_j)=(\varphi_{kj}(h_j)+\ell_{kj}(h_j)v_j,t_{kj}(h_j)v_j)$ and
\eq{oldphi}\nonumber
\Phi_{kj}(h_j,v_j)=N^*_{kj}(h_j,v_j)+\sum_{\ell=2}^m[\phi_{kj}]^\ell(h_j,v_j)+O(m+1).
\end{equation}
The Jacobian of $N^*_{kj}$ at $(h_j,v_j)$ applied to $(\tilde h_j,\tilde v_j)$ is given by
$$
DN_{kj}^*(h_j,v_j)(\tilde h_j,\tilde v_j)=(s_{kj}(h_j)\tilde h_j+\ell_{kj}(h_j)\tilde v_j
+\partial_{h_j}\ell_{kj}(z_j)\tilde h_jv_j,t_{kj}(h_j)\tilde v_j+\partial_{h_j}t_{kj}(h_j)\tilde h_jv_j).
$$
If functions $(\tilde h_j,\tilde v_j)=O(|v_i|)$, we simplify it as
$$
DN_{kj}^*(h_j,v_j)(\tilde h_j,\tilde v_j)=DN_{kj}^*(h_j,0)(\tilde h_j,\tilde v_j)+O(|v_i|^2).
$$
Let us use transition functions $g_{kj}$ of $T_CM$ to write
$$
DN_{kj}^*(h_j,0)(\tilde h_j,\tilde v_j)=g_{kj}(h_j)(\tilde h_j,\tilde v_j)=(s_{kj}(h_j)\tilde h_j+\ell_{kj}(h_j)\tilde v_j,t_{kj}(h_j)\tilde v_j).
$$
Then
\begin{align}\nonumber
\Phi_{kj}\Phi_{ji}(h_i,v_i) - N_{kj}^*N_{ji}^*(h_i,v_i)&=[\phi_{kj}]^{\leq m}(\varphi_{ji}(h_i)+\ell_{ji}(h_i)v_i,t_{ji}(h_i)v_i)
\\
&\quad+DN_{kj}^*(N_{ji}^*(h_i,v_i))[\phi_{ji}]^{\leq m}+O(m+1)\nonumber \\
&=[\phi_{kj}]^m (N_{ji}(h_i,v_i))
+g_{kj}(\varphi_{ji}(h_i))[\phi_{ji}]^m(h_i,v_i)\nonumber \\
&\quad + R_{kji}(h_i,v_i,[\phi_{\bullet}]^2,\dots, [\phi_{\bullet}]^{m-1})+O(m+1),\nonumber\\
&=\Phi_{ki}(h_i,v_i)-N_{kj}^*N_{ji}^*(h_i,v_i)
\nonumber
\end{align}
where function $R_{kji}$
is independent of $[\phi_{\bullet}]^\ell$ for $\ell\geq m$. Applying the same computation to $\tilde \Phi=\Phi+O(m)$ and subtracting them, we obtain, writing $\psi_{kj}:=\phi_{kj}-\tilde\phi_{kj}$,
$$
[\psi_{kj}]^m (N_{ji}(h_i,v_i))
+g_{kj}(\varphi_{ji}(h_i))[\psi_{ji}]^m(h_i,v_i)-[\psi_{ki}]^m(h_i,v_i)=0.
$$
According to \cite[Lemma 2.7]{GS3}, this is equivalent to saying that
$$
\{[\phi_{kj}-\tilde\phi_{kj}]^m\}_{kj}\in Z^1(\mathcal U,T_CM\otimes S^m N_C^*).
$$
$(b)$ Note that when $f=O(m)$ and $\tilde\Phi=F\Phi F^{-1}$, we have $\tilde\Phi=\Phi+O(m)$. Then $[\tilde\Phi]^m=[F\Phi F^{-1}]^m$ is equivalent to $\delta [f]^m=[\phi]^m-[\tilde \phi]^m$ according to \cite[Lemma 2.16 (2.34)]{GS3}.
$(c)$ By (a), we have $\delta( [\tilde\Phi]^m-[\Phi]^m)=0$. By the definition of the basis $e^\ell$, we can find $[f]^m$ such that $[\Phi]^m+ c^m\cdot e^m-[\tilde\Phi]^m=\delta[f]^m$ for some $[f]^m=\{[f_k]^m\}$. Set $F_k=I+[f_k]^m$. We get $F^{-1}\tilde\Phi F=\Phi+c^m\cdot e^m+O(m+1)$ by (b).
\end{proof}
To construct normal forms, we must refine the above
degree-by-degree normalization.
Define the approximate automorphism groups
$$
{\operatorname{Aut}}_ m (\Phi)=\{F=I+O(2)\colon F^{-1} \Phi F= \Phi+O( m +1)\}, \quad m=2,3,\dots.
$$
Here it is important that we allow $F$ admits lower order terms.
Since $F(\Phi+O( m +1))=F(\Phi)+O( m +1)$, then $F\in\operatorname{Aut}_ m (\Phi)$ if and only if
$$
F\Phi-\Phi F=O({ m +1}).
$$
The latter implies that $
\Phi F^{-1}=F^{-1}\Phi +O({ m +1}).
$
Thus, ${\operatorname{Aut}}_ m (\Phi)$ is indeed a group. The group structure implies that the conjugacy by elements in $\operatorname{Aut}_ m (\Phi)$ induces an equivalence relation on
$$
\mathcal M_m(\Phi):=\{\tilde\Phi\in \mathcal M\colon \tilde\Phi=\Phi+c^ m\cdot e^ m +O( m +1) \}.
$$
Namely, if $\Psi,\tilde\Psi\in \mathcal M_m(\Phi)$, define the equivalence relation $\sim_{\operatorname{Aut}_m(\Phi)}$ such that
$$
\Psi\sim_{\operatorname{Aut}_m(\Phi)}\tilde\Psi
$$
if and only if there is an $F\in \operatorname{Aut}_m(\Phi)$ such that $F\Phi F^{-1}=\tilde\Phi+O(m+1)$.
Let $\mathcal M_m(\Phi)/\operatorname{Aut}_m(\Phi)$ be the set of equivalent classes. For each equivalence class, \emph{fix} a representative.
Define $\hat{\mathcal C}_m(\Phi)$ be the set of elements $c^me^m$ such that $\Phi(c^m):=\Phi+c^m\cdot e^m+O(m+1)$ are among the (chosen) representatives (It will be clear from the context that $\Phi(c^m)$ do not stand for the evaluation of $\Phi$ at $c^m$).
Thus, we can express the set $\mathcal N_m(\Phi)$ of representatives as
$$
\{\Phi(c^m):=\Phi+c^m\cdot e^m+O(m+1)\colon c^m\cdot e^m\in \hat{\mathcal C}_m(\Phi)\}.
$$
To ensure stability, $\Phi$ is always a representative, i.e.
\eq{}\nonumber
0\cdot e^m\in \hat{\mathcal C}_m(\Phi), \quad \Phi(0)=\Phi.
\end{equation}
Using this equivalence relation, we define convergent \emph{partial normal forms} $N^m$ and formal normal forms $N^\infty$ as follows.
\begin{defn}\label{the-def} Fix transition functions $N^1=\{N^1_{kj}\}$ for all neighborhoods
of $C$ with a given $T_CM$ as in \rd{admN}.
\begin{list}[align=left]{$(\roman{ppp})$}{\usecounter{ppp}}
\item Define $\mathcal N_1=\{N^1\}$, which has one element.
Let $\mathcal N_2(N^1)$ to be the set of representatives $N^2(c^2)=N^{1}+c^2\cdot e^2+O(3)$ for
elements in $\mathcal M_2(N^1)/\operatorname{Aut}_2(N^1)$, which are determined by
\eq{}\nonumber
c^2\cdot e^2\in \hat {\mathcal C}_2(N^1).
\end{equation}
Inductively, let $\mathcal N_m$ be the set of representatives
\eq{}\nonumber
N^m(c^2,\dots, e^m):=N^{m-1}(c^2,\dots, c^{m-1})(c^m)\end{equation}
for elements in $\mathcal M_m(N^{m-1}(c^2,\dots, c^{m-1}))/\operatorname{Aut}_m(N^{m-1}(c^2,\dots, c^{m-1}))$, which are determined by
\eq{}\nonumber
c^m\cdot e^m\in \hat C_m(N^{m-1}(c^2,\dots, c^{m-1})).
\end{equation}
\item Define $N^\infty(c^2,c^3,\dots)$ to be the formal transition functions such that
$$
N^\infty(c^2,c^3,\dots)=N^m(c^2,c^3,\dots, c^m)+O(m+1), \quad \forall m.
$$
Let $\mathcal N_\infty$ be the set of all such formal transition functions.
\end{list}
\end{defn}
By definition, $N^{m-1}(c^2,\dots, c^{m-1})(0)=N^{m-1}(c^2,\dots, c^{m-1})$. Thus, we have
\begin{prop}
Assume that $H^1(C,\tcm{m})=0$ for all $m>m_0$ and $m_0$ is finite. Then all $N^\infty=N^{m_0}$ define convergent neighborhoods and $\mathcal N_\infty$ is finite-dimensional.
\end{prop}
Recall that
\eq{defH=}\nonumber
\mathcal H^q(T_CM):=\bigoplus_{\ell\geq2} H^q(C,T_CM\otimes S^\ell N_C^*).
\end{equation}
We now prove a formal version of \rt{m-finite}.
\begin{thm}
\label{injection} There exists a mapping $\frak C^f$ from $\mathcal M/{\, \overset{f}{\sim}}$ into $\mathcal H^1(T_CM)$. Furthermore, the mapping $\frak C^f$ is injective and there are formal mappings transforming $\Phi\in\mathcal M$ into $N^\infty(\Phi)\in\mathcal N_\infty$, provided
\begin{gather}\label{for-F1}
\dim \mathcal H^1(T_CM)<\infty, \quad \text{or} \\
\label{for-F0}\mathcal H^0(T_CM)=0.
\end{gather}
\end{thm}
\begin{proof}Recall that we fix a representant
basis $e^\ell$ for (a basis of)
$H^1(C,T_CM\otimes S^\ell N_C^*)$. Let $\Phi=N^1+\phi$ as in
\rd{admN}
define a neighborhood. We have $\delta[\phi]^2=0$. Applying \rl{the-lem-} there is a unique constant vector $c_0^2(\Phi)$ such that for some $ f^2=:\{f_j^2\}\in C^{ 0}(\mathcal U,\tcm{2})$,
$$
[\Phi
-N^1]^2=c_0^2(\Phi)\cdot e^2+\delta f^2.
$$
Set $F^2_j=I+f_j^2$ and $F_2=\{F^2_j\}$. Then $F_2\Phi F_2^{-1}=N^1+c_0^2(\Phi)\cdot e^2+O(3)$. Recall the definition
\eq{}\nonumber
\mathcal M_2(N^1):=\{\widetilde\Phi\in\mathcal M\colon\widetilde\Phi_{kj}=N^1_{kj}+c^2\cdot e^2+O(3)\}.
\end{equation}
In other words, we have achieved $F_2\Phi F_2^{-1}\in \mathcal M_2(N^1)$. We now apply the refinement. We find a unique element $c^2\cdot e^2\in \hat{\mathcal C}_2(N^1)$ and $\tilde F_2\in \operatorname{Aut}_2(N^1)$ such that
$$
\Phi_2:=\tilde F_2F_2\Phi F_2^{-1}\tilde F_2^{-1}=N^2(c^2)+O(3).
$$
We denote this $c^2$ by $c^2(\Phi)$ and we will show that $c^2(\Phi)$ is a formal invariant.
We repeat the above two-step normalization.
Inductively, using \rl{the-lem-} we find $F_\ell$ for $\ell>2$ such that
$$F_\ell\Phi_{\ell-1} F_\ell^{-1}\in\mathcal M_\ell(N^{\ell-1}(c^2,\dots, c^{\ell-1})).
$$
We then find $\tilde F_\ell\in \operatorname{Aut}_\ell(N^{\ell-1}(c^2,\dots, c^{\ell-1}))$ such that
$$
\Phi_\ell:=\tilde F_\ell F_\ell\Phi_{\ell-1}F_\ell^{-1}\tilde F_\ell^{-1}=N^\ell(c^2,\dots, c^\ell), \quad c^\ell\in \hat{\mathcal C}_\ell(N^{\ell-1}(c^2,\dots, c^{\ell-1})).
$$
Again, we denote $c^\ell$ by $c^\ell(\Phi)$. Let $\ell\to\infty$. Since
$$N^{\ell+1}(c^2,\dots, c^{\ell+1})=N^{\ell}(c^2,\dots, c^{\ell})+O(\ell+1)$$
it is clear that the sequence defines formal transition functions $N^\infty(c^2(\Phi), c^2(\Phi),\dots)$, denoted by $N_\infty(\Phi)$.
Next we want to show that $c^2(\Phi),\dots, c^\ell(\Phi)$ are uniquely determined by the equivalence class of $\Phi$ under formal changes of coordinates that are tangent to the identity.
Suppose that $\tilde\Phi=G\Phi G^{-1}$ with $G=I+O(2)$. We have $c^2(\Phi)=c^2(\tilde\Phi)$ immediately. Suppose that $c^\ell(\Phi)=c^\ell(\tilde\Phi)=c^\ell$ for $\ell<m$. Then we can find $F,\tilde F$ so that $F\Phi F^{-1}=N^{m-1}(c')+c^me^m+O(m+1)$ and $\tilde F\tilde\Phi \tilde F^{-1}=N^{m-1}(c')+\tilde c^me^m+O(m+1)$ with $c'=(c^2,\dots, c^{m-1})$. Then $K:=\tilde FG F^{-1}$ satisfies
$$N^{m-1}(c')+\tilde c^me^m+O(m+1)=K(N^{m-1}(c')+c^me^m+O(m+1))K^{-1}\in\mathcal M.
$$
Thus $K\in\operatorname{Aut}_{m}(N^{m-1}(c'))$. Now, $\tilde c^m,c^m$ are in $\mathcal C_{m}(N^{m-1}(c'))$ and $N^{m-1}(c')+\tilde c^me^m+O(m+1), N^{m-1}(c')+\tilde c^me^m+O(m+1)$ are equivalent module $O(m+1)$ by $K\in\operatorname{Aut}_{m}(N^{m-1}(c'))$. The equivalence class under the conjugacy of $\operatorname{Aut}_{m}(N^{m-1}(c'))$ can only represented by a unique element in $\hat{\mathcal C}_{m}(N^{m-1}(c'))$. We obtain $c^m(\Phi)=\tilde c^m(\tilde\Phi)$. This shows that $N_\infty(\Phi)=N^\infty(c^2(\Phi),\dots,)$ is well-defined on $\mathcal M/{\, \overset{f}{\sim}}$. Define $N_\infty(\Phi)$ to be the image under $\frak C^f$ for the equivalence class $\Phi \mod{\!\!\overset{f}{\sim}}$ of $\Phi$ in $\mathcal M/{\, \overset{f}{\sim}}$.
We know that $F_\ell=I+O(\ell)$. Since a subsequence of $\tilde F_\ell$ may contain terms of a fixed order, we need to verify that the sequence $\hat F_\ell:=\tilde F_\ell F_\ell\cdots \tilde F_2F_2$ still defines a formal change of coordinates as $\ell\to\infty$. When $ H^1(T_CM\otimes S^\ell N_C^*)=0$ for all $\ell>\ell_0$, we have $N^\infty(c^2,\dots)=N^{\ell_0}(c^2,\dots, c^{\ell_0})$ and $\tilde F_\ell=I$ and $F_\ell=I+O(\ell)$ for $\ell>\ell_0$. When $\mathcal H^0(T_CM)=0$, we can show inductively that $\tilde F_\ell=I$ for all $\ell\geq2$. This shows that the sequence of coordinate changes define a formal transformation $F$ and $F\Phi F^{-1}=N^\infty(\Phi)$.
Finally, we show that if $c^\ell(\Phi)=c^\ell(\tilde\Phi)$ for all $\ell$, then $\tilde\Phi$ and $\Phi$ are equivalent. Indeed, we have $F\Phi F^{-1}=N^\infty(c^2(\Phi),\dots)=\tilde F\tilde\Phi\tilde F^{-1}$. This shows that $\tilde\Phi=\tilde F^{-1}F\Phi F^{-1}\tilde F$.
\end{proof}
\begin{rem}We remark that there exist second-order obstructions to realize an element in $H^1(C,T_CM\otimes S^2N_C^*)$ by a neighborhood of $C$. See Griffiths~\cite{MR206980} and Morrow-Rossi~~\cite[Prop. 3.4]{MR627765}.
\end{rem}
Our classification is achieved under a group of biholomorphisms $F$ that is smaller that the whole group of biholomorphisms, by restricting to $F$ to be tangent to the identity along $C$. Therefore, the equivalence classes under non-restricted biholomorphisms might be a smaller set; in fact by a simple dilation we can further reduce the set of equivalence classes to be a \emph{compact} but possibly a non-Hausdorff set as the case of $\dim M=2$ in~~\cite[p. 323]{MR627765}.
For instance, let us consider the case of $N_C$ is a line bundle. Let $t_{kj}$ be the transition functions of $N_C$. An isomorphism of $N_C$ is given by $g_k^{-1}t_{kj}g_j=t_{kj}$. Thus $g_j$ define a global holomorphic function on $C$ without zero. Since $C$ is compact, then the function must constant. Now it is easy to see that the transition functions for the neighborhood $\Phi_{kj}(h_j,v_j)$ is transformed into $(\Phi_{kj}^h(h_j,cv_j), c^{-1}\Phi_{kj}^v(cv_j))$.
We will give examples in Section 6 for \re{for-F1} and \re{for-F0}.
\setcounter{thm}{0}\setcounter{equation}{0}
\section{A formal normal form for tangential foliations}
In this section, we study \emph{tangential} foliations of neighborhoods of $C$.
A neighborhood admits a tangential foliation, denoted by $M_\tau$, if there are $d$ holomorphic functions $\tilde v_k^1, \dots, \tilde v^d_k$ such that $d\tilde v^1_k\wedge\cdots\wedge d\tilde v^d_k\neq0$ while $\tilde v_k=c_k$ and $\tilde v_j=c_j$ define the same foliation on $V_k\cap V_j$ and $v_k=0$ on $C$. The set of tangential foliations of $C$ will be denoted by $\mathcal M_\tau$. A biholomorphism $F$ sends $M_\tau$ into $\tilde M_\tau$, if it sends leaves of $M_\tau$ into leaves of $\tilde M_\tau$. The set of such transformations $F$ that are tangent to the identity is denoted by $\mathcal T(M_\tau,\tilde M_\tau)$, which depends on $M_\tau,\tilde M_\tau$.
Clearly, the equivalences of foliations of neighborhoods of $C$ implies the equivalence of the neighborhoods. Therefore, the classification of tangential foliations is a refinement to that of neighborhoods. This should be reflected in the construction of our normal forms for tangential foliations.
\begin{defn} Let $M_\tau$ be a tangential foliation. We call $(h_j,v_j)$ are \emph{tangential coordinates} if $v_j=cst$ define the foliation and consequently the transition functions satisfy $\Phi_{kj}^v(h_j,v_j)=\Phi_{kj}^v(v_j)$. Such transition functions, denoted by $\Phi_\tau$, are called \emph{tangential transition functions}.
\end{defn}
\le{h-trans}
\begin{list}{$\hspace{-1em}(\alph{pp})$}{\usecounter{pp}}
\item A tangential foliation $M_\tau$ admits tangential coordinates $ (h_j,v_j)$. That $(h_k,v_k)$ are tangential coordinates for $M_\tau$ if and only if the transverse components $\Phi^v_{kj}$ of their transition functions $\Phi_{kj}$ depend only on $v_j$.
\item Suppose $F=I+O(2)$. Then $F$ sends a tangential foliation $\Phi_{kj}$ into another tangential foliation $\tilde\Phi_{kj}$ if and only if $F\Phi F^{-1}=\tilde\Phi$ and $F^v_j(h_j,v_j)=F^v_j(v_j)$ depends only on $v_j$, in which case
\eq{}\nonumber
\tilde\Phi_{kj}^v=F_k^v\Phi^v_{kj}( F_j^v)^{-1}, \quad \tilde\Phi^h_{kj}=
F_k^h\Phi_{kj}F_j^{-1}.
\end{equation}
\end{list}
\end{lemma}
\begin{proof} (a) Suppose that $M_\tau$ is defined by $\tilde v_j=cst$ such that $v_j=0$ and $d\tilde v_j^1\wedge\cdots\wedge d\tilde v_j^d\neq0$ on $C$. Without loss of generality, we may assume that $\tilde v_j^\ell(h_j,v_j)=v_j+O(2)$. Define $\hat v_k=\hat\Phi^v_{kj}(\hat v_j)$. Consequently, for $F_j(h_j,v_i)=(h_j,\hat v_j)$, we get $F_kF_j^{-1}(h_j,\hat v_j)=(\hat\Phi_{kj}^h(h_j,\hat v_j),
\hat\Phi_{kj}^v(\hat v_j))$.
(b) Suppose that $(h_j,v_j), (\tilde h_j,\tilde v_j)$ are tangential coordinates and $\Phi_{kj},\tilde\Phi_{kj}$ are the corresponding transition functions for foliations $M_\tau,\tilde M_\tau$. Suppose that $F$ sends $(\mathcal T,C)$ into $(\widetilde{\mathcal T},C)$. Write $F_j=(F_j^h,F_j^v)$. Then $F^v_j(h_j,v_j)$ must be constant if $v_j$ is constant, i.e. $\tilde v_j=F_j^v(h_j,v_j)=F^v_j(v_j)$. Combining it with $\tilde v_k=\tilde\Phi^v_{kj}(\tilde v_j)$ and $v_k=\Phi_{kj}^v(v_j)$, we obtain $\tilde \Phi_{kj}^v=F_k^v\Phi^v_{kj}(F_j^v)^{-1}$. Of course, we still have $\tilde\Phi^h_{kj}=F_k^h\Phi_{kj}F_j^{-1}$.
\end{proof}
\begin{defn}
The set of transformations $F=I+O(2)$ satisfying $F^v(h_j,v_j)=F_j^v(v_j)$ is denoted by $\mathcal T_\tau$. The set of transformations satisfying additional condition $F=I+O(m)$ is denoted by $\mathcal T^\tau_m$.
\end{defn}
It is known, see for instance~~\cite{GS3}, that if a tangential foliation exists, then $N_C$ is flat. We fix a flat basis $\hat e_k$ for $N_C$ over $U_k$. Recall that when we chose the basis $(e_k^h,e_k^v)$ in \re{ekhv} for $T_CM$ on $U_k\subset C$, we have decided $e_k^v$ to be a flat basis of $N_C$ when $N_C$ is flat. Let $w_k^v$ be the dual of $e_k^v$.
Denote by
$
\mathcal C_{cst}^\ell(\omega)$
the sections of $T_CM\otimes S^\ell N_C^*$ of the form
$$
\sum_{i=1}^n\sum_{|Q|=\ell} a_{i Q}(h_k)e_{k,i}^h\otimes (w^v_{k})^Q+\sum_{\alpha=1}^d\sum_{|Q|=\ell} b_{\alpha Q}e_{k,\alpha}^v\otimes (w_{k}^v)^Q
$$
where $a_{iQ}$ are holomorphic functions and $b_{\alpha Q}$ are locally constant functions on $\omega$ if $\omega$ is a subset of $U_k$.
Then $\delta$ defined by \re{dfphi} is still a coboundary operator from $C_{cst}^q$ to $B_{cst}^{q+1}$. Denote by $\check H^q(\mathcal U,\mathcal C_{cst}^\ell)$ the \v{C}ech cohomology groups.
Define
\eq{}
\check{\mathcal H}^q_\tau(T_CM):=\bigoplus_{\ell>1}\check{ H}^q(\mathcal U,\mathcal C_{cst}^\ell).
\end{equation}
We now start to adapt the normal form of neighborhoods of $C$ for the tangential foliations of neighborhoods of $C$.
\begin{defn}\label{admN-3.4} Fix $T_CM$ with flat $N_C$.
Fix a tangential foliation $N_\tau^1=\{N^1_{\tau,kj}\}$ of $C$ such that $N^1_{\tau, kj}=N^*_{kj}+O(2)$.
If $T_CM$ additionally splits, we take $N_\tau^1=N$ defined by \re{Nkjh}. \end{defn}
Throughout the paper, we assume that $\mathcal M_\tau$ is non-empty. Each element in $\mathcal M_\tau$ is given by tangential transitions functions. Therefore the existence of $N_\tau^1$ is ensured regardless if $T_CM$ splits or not.
\begin{defn
Let $\mathcal M_\tau$ be the set of holomorphic tangential foliations containing $C$ as a leaf. Let $\mathcal M_\tau/{\,\overset{f}{ \sim}}$ the set of equivalence classes under formal tangential-foliation mappings in $\mathcal T_\tau$.
\end{defn}
We now state a formal classification for tangential foliations.
\begin{thm}
\label{injection-h} Let $N_C$ be flat.
Assume that $\check{ H}^1_\tau(C,\mathcal C_{cst}^\ell)$ are finite dimensional for all $\ell>1$.
There is a mapping $\frak C^f_\tau$ from $\mathcal M_\tau/{\,\overset{f}{ \sim}}$ into
$\check{\mathcal H}_\tau^1(T_CM)$ such that if
\begin{gather}\label{for-F1-tau}\nonumber
\dim \mathcal H_\tau^1(T_CM)<\infty, \quad \text{or} \\
\label{for-F0-tau}\mathcal H^0_\tau(T_CM)=0,\nonumber
\end{gather}
then $\frak C^f_\tau$ is injective and there are formal mappings in $\mathcal T_\tau$ transforming $\Phi\in\mathcal M_\tau$ into $N_\tau^\infty(\Phi)$.
In particular, if $\mathcal H^1_\tau(T_CM)=0$, all tangential foliations are formally equivalent.
\end{thm}
We fix a basis $e_\tau^\ell$ for $\check{ H}^1(\mathcal U,\mathcal C_{cst}^\ell)$.
The proof is almost identical to the proof for the general case. We will only give an outline below.
\le{the-lem-TAU} Fix $m>1$. Consider transformations $F=I+f\in \mathcal T_\tau$ with $f=O(m)$. Suppose that $\Phi_\tau,\tilde\Phi_\tau\in\mathcal M_\tau$ and
$$
\tilde\Phi_\tau=\Phi_\tau+O(m),\quad
\Phi_\tau=N^1_\tau+\phi, \quad \tilde\Phi_\tau=N^1_\tau+\tilde\phi.
$$
Then we have the following.
\begin{list}{$\hspace{-1em}(\alph{pp})$}{\usecounter{pp}}\item $[\tilde\Phi]^m-[\Phi]^m\in Z^1(\mathcal U,\mathcal C_{cst}^m)$; in particular,
$$
[\Phi-N_\tau^1]^2\in Z_{cst}^1(\mathcal U,\mathcal C_{cst}^2),\quad \forall \Phi\in\mathcal M_\tau.
$$
\item
$F\in\mathcal T_m^{\tau}(\Phi_\tau,\tilde\Phi_\tau)$ holds if and only if
$ \delta [f]^m=[\tilde\Phi_\tau]^m-[\Phi_\tau]^m$.
\item There exist a unique $c_\tau^m\cdot e_\tau^m\in \check{ H}^1(\mathcal U,\mathcal C_{cst}^m)$ and some $f=O(m)$ such that $ \hat\Phi:=F^{-1}\tilde\Phi_\tau F=\Phi_\tau+c_\tau^m\cdot e_\tau^m+O(m+1)$, i.e.
$$
\hat\Phi=\Phi_\tau+c_\tau^m\cdot e_\tau^m+O(m+1), \quad \hat\Phi\in\mathcal M_\tau.
$$
\end{list}
\end{lemma}
For $\Phi\in\mathcal M_\tau$, define
\begin{gather}{}\nonumber
\mathcal M_m^\tau(\Phi):=\{\tilde\Phi\in\mathcal M_\tau\colon \tilde\Phi=\Phi+c_\tau^2\cdot e_\tau^2+O(m+1)\},
\\
\operatorname{Aut}_m^\tau(\Phi)= \{F\in\mathcal T_\tau\colon F^{-1} \Phi F= \Phi+O( m +1)\}.\nonumber
\end{gather}
Then $\operatorname{Aut}^\tau_m(\Phi)$ is a group inducing an equivalence relation on $\mathcal M_m^\tau(\Phi)$. Fix an element in each equivalence class. Thus the set $\mathcal M_m^\tau(\Phi)/\operatorname{Aut}^\tau_m(\Phi)$ of equivalent classes can be written as
$$
\Phi(c^m_\tau)=\Phi+c_\tau^m e^m+O(m+1)\in\mathcal M_\tau, \quad c^\tau e^\tau\in \mathcal H^m_\tau(\Phi)\subset \check H^1_{cst}(\mathcal U,\mathcal C_{cst}^m).
$$
For stability, we always choose $\Phi(0)=\Phi$.
\begin{defn}\label{the-def-TAU}Let $\Phi\in\mathcal M_\tau$. \begin{list}[align=left]{$(\roman{ppp})$}{\usecounter{ppp}} \item By \rl{the-lem-TAU}, we find $F_2=I+O(2)\in \mathcal T_\tau$ such that $F_2\Phi F_2^{-1}=N_\tau^1+\tilde c_\tau^2e_\tau^2+O(3)$, i.e.
$$
F_2\Phi F_2^{-1}\in\mathcal M^\tau_2(N_\tau^1).
$$
Take $\tilde F_2\in\operatorname{Aut}^\tau_2(N_\tau^1)$ such that
\eq{}\nonumber
\Phi_2:=\tilde F_2F_2\Phi F_2^{-1}\tilde F_2^{-1}=N_\tau^1(c_\tau^2)+O(3), \quad c_\tau^2 e^2_\tau\in\mathcal H^2_\tau(N_\tau^{1} ).
\end{equation}
\item
Let $ m >2$. Find $F_m=I+O(m)\in\mathcal T_\tau$ such that
$$
F_m\Phi_{m-1}F_m^{-1}\in \mathcal M^\tau_m(N^{m-1}_\tau(c^2_\tau,\dots, c^{m-1}_\tau)).
$$
Choose $\tilde F_m\in\operatorname{Aut}_m^\tau((N^{m-1}_\tau(c_2^\tau,\dots, c^{m-1}_\tau))$ such that
\begin{gather}{}\nonumber
\tilde F_mF_m\Phi_{m-1}F_m^{-1}\tilde F_m^{-1}=N^{m-1}_{\tau}(c^2_\tau,\dots, c^{m-1}_\tau)(c^m_\tau)+O(m+1), \\ c^m_\tau e^m_\tau \in \mathcal H^m_\tau(N_\tau^{m-1}(c^2_\tau,\dots, c^{m-1}_\tau))\nonumber
\end{gather}
\item Set $N^m_\tau(c^2_\tau,\dots, c^m_\tau)=N^{m-1}_{\tau}(c^2_\tau,\dots, c^{m-1}_\tau)(c^m_\tau)$. The formal normal form of $\Phi$ is $N_\tau^\infty$ with
\eq{}
N_\tau^\infty=N^m_\tau(c^2_\tau,\dots, c^m_\tau)+O(m+1), \quad m=1,2,\dots.\nonumber
\end{equation}
Define $(c^2_\tau e^2_\tau,\dots)\in \mathcal H^1_{\tau}(T_CM)$ to be $\frak C^f_\tau(\Phi)$ for the equivalence class of $\Phi$ under $\mathcal T_\tau$.
\end{list}\end{defn}
We leave the rest of details for the proof of \rt{injection-h} to the reader.
We refer to \cite{GS3} for a different approach to the existence of holomorphic tangential foliation when the formal obstructions are absent in a stronger sense that the normal component of $\Phi$ is formally linearizable.
Under a small divisor condition on cohomology groups depending only on a (flat) unitary $N_C$, a convergence for linearizing the normal components is achieved in~\cite{GS3}.
\setcounter{thm}{0}\setcounter{equation}{0}
\section{A formal normal form of transverse foliations}
In this section, we will use normal forms in Section 2 to describe neighborhoods that admit transverse foliations.
By a holomorphic \emph{transverse} foliation of $(M,C)$, we mean that on a neighborhood of $C$ in $M$ there a smooth holomorphic foliation with all leaves intersect $C$ transversely.
First-order obstructions to transverse foliations for $(M,C)$ in higher dimensions or higher codimension were considered in~\cite{MR206980,MR627765}; however they did not settle the existence of transverse foliations when formal obstructions vanish except for the case of $\dim M=2$ mentioned earlier.
In this section, we will obtain an existence result on transverse foliations and the classification on them under suitable conditions on $T_CM$.
A neighborhood admits a \emph{vertical} or \emph{transverse} foliation, if there are $n$ holomorphic functions $\tilde h_k^1, \dots, \tilde h^n_k$ such that
\eq{dhdt}\nonumber
d\tilde h^1_k\wedge\cdots\wedge d\tilde h^n_k\wedge dt_k^1\wedge\cdots\wedge dt_k^d\neq0
\end{equation}
while $\tilde h_k=c_k$ and $\tilde h_j=c_j$ define the same foliation on $V_k\cap V_j$ on $C$. By an abuse of notation, we still denote the vertical foliation on a neighborhood $(M,C)$ by $(\mathcal V,C)$, or simply by $\Phi$.
In the previous section, we have seen that for a neighborhood to admit tangential foliation, $N_C$ must be flat. A transverse foliation does not impose conditions on $N_C$ as any $N_C$ is already foliation by its fibers. It however impose a useful condition that $T_CM$ must split~~\cite{MR627765}. The formal obstructions for transverse foliations were obtained in~\cite{MR206980, MR627765}. In this section, we obtain a normal form for transverse foliations.
\begin{defn}Let $(\mathcal V,C)$ be a transverse foliation. We say that $(h_j,v_j)$ are foliated coordinates for $(\mathcal V,C)$ if $h_j=cst$ define the foliation and consequently the transition functions satisfy $\Phi_{kj}^h(h_j,v_j)=\Phi_{kj}^h(h_j)$, in which case we call $\{\Phi_{kj}\}$ a transverse foliation for abbreviation.
\end{defn}
\begin{defn}
Two transverse foliations $(\mathcal V,C)$ in $(M,C)$, $(\tilde{\mathcal V},C)$ in $(\tilde M,C)$ are equivalent by a biholomorphic mapping $F$ if $F=I+O(2)$ sends the neighborhood $(M,C)$ into $(\tilde M,C)$ and it sends each leave of $({\mathcal V},C)$ into a leaf of $(\tilde{\mathcal V},C)$.
\end{defn}
\le{H-trans}
\begin{list}{$\hspace{-1em}(\alph{pp})$}{\usecounter{pp}}\item Let $(\mathcal V,C)$ be a transverse foliation. Then the foliation admits foliated coordinates $(h_j,v_j)$. That $(\tilde h_k,\tilde v_k)$ are foliated coordinates for $(\mathcal V,C)$ if and only if their transition functions $\Phi_{kj}$ satisfy $\Phi_{kj}^h=\varphi_{kj}(h_j)$.
\item There is a biholomorphic mapping $F=I+O(2)$ sends a transverse foliation $\Phi$ in foliated coordinates into another transverse foliation $\tilde\Phi$ in the foliated coordinates if and only if $F^{-1}\Phi F=\tilde\Phi$ and $F^h_j(h_j,v_j)=
h_j$, in which case
\eq{}\nonumber
\tilde\Phi_{kj}^h=\Phi^h_{kj}=\varphi_{kj}(h_j), \quad \tilde\Phi^v_{kj}=F_k^v\Phi_{kj}F_j^{-1}.
\end{equation}
The set of all transformations $F=I+O(2)$ with $F_j^h(h_j,v_j)=h_j$ are denoted by $\mathcal T_\nu$.
\end{list}
\end{lemma}
\begin{proof} The proof is almost verbatim to \rl{h-trans}, which is leave to the reader.
\end{proof}
Note that the transformations $F_j=I+f_j$ with $f_j=O(2)$ that preserves transverse foliations are rather restrictive as $f^h_j=0$. This is quite different from the study of tangential foliations in the previous section whereas transformations that preserve the tangential foliations can be higher-order perturbations in both tangential and normal components. The advantage is that the formal classification of transverse foliations is almost identical to the formal classification of the neighborhoods.
Define
$$\mathcal H_\nu^q(T_CM):=\bigoplus_{\ell>1}H^q(C,N_C\otimes S^\ell N_C^*).$$
Using \rl{H-trans}, we can obtain the following formal normal form.
\begin{thm}
\label{injection-nu} Let $\mathcal M_\nu$ be the set of holomorphic transverse foliations of $C$.
There is a mapping $\frak C^f_\nu$ from $\mathcal M_\nu/{\,\overset{f}{ \sim}}$ into $\mathcal H_\nu^1(T_CM)$ such that if
\begin{gather}\label{for-F1-nu}\nonumber
\dim \mathcal H^1_\nu(T_CM))<\infty, \quad \text{or} \\
\label{for-F0-nu}\mathcal H^0_\nu(T_CM)=0,\nonumber
\end{gather}
then $\frak C^f_\nu$ is injective and there are formal mappings in $\mathcal T_\nu$ transforming $\Phi\in\mathcal M_\nu$ into $N^\nu_\infty(\Phi)$.
In particular, if $\mathcal H^1_\nu(T_CM)=0$, all transverse foliations are formally equivalent.
\end{thm}
\begin{proof}Given a transverse foliation, we know that $T_CM$ splits.
Let $\Phi\in\mathcal M_\nu$. Define
$$
\operatorname{Aut}_m^\nu(\Phi)=\{F\in \mathcal T_\nu\colon F\Phi F^{-1}=\Phi+O(m+1)\}.
$$
By conjugacy the group yields an equivalence relation on
$$
\mathcal M_m(\Phi):=\{\Phi+c^m_\nu e^m_\nu+O(m+1)\}\cap \mathcal M_\nu.
$$
Select representatives for the equivalence classes and denote the set of corresponding elements $c^m_\nu e^m_\nu$ by
$
\mathcal H^\nu_m(\Phi).
$
Thus the set of equivalence classes is given by
$$
\mathcal N_m(\Phi)=\{\Phi(c^m_\nu)=\Phi+c^m_\nu e^m_\nu+O(m+1)\colon c^m_\nu e^m_\nu\in \mathcal H_m^\nu(\Phi)\}\subset\mathcal M_\nu.
$$
By \rl{H-trans}, we find $F_2=I+O(2)\in\mathcal T_\nu$
such that $F_2\Phi F_2^{-1}=N^1_\nu+c_\nu^2e_\nu^2+O(3)$
with $N^1_\nu=N^1$. Thus
$$
F_2\Phi F_2^{-1}\in\mathcal M^\nu_2(N_\nu^1).
$$
Take $\tilde F_2\in\operatorname{Aut}^\nu_2(N^1)$ such that
\eq{}\nonumber
\Phi_2:=\tilde F_2F_2\Phi F_2^{-1}\tilde F_2^{-1}=N_\nu^1(c_\nu^2)+O(3), \quad c_\nu^2 e^2_\nu\in\mathcal H^2_\nu(N_\nu^{1} ).
\end{equation}
Set $N^2_\nu(c^2_\nu e^2_\nu):=N^1_\nu(c^2_\nu e^2_\nu)$.
Let $ m >2$. Find $F_m=I+O(m)\in \mathcal T_\nu(\mathcal M_\nu)$ such that
$$
F_m\Phi_{m-1}F_m^{-1}\in \mathcal M^\nu_m(N^{m-1}_\nu(c_2^\nu,\dots, c^{m-1}_\nu)).
$$
Choose $\tilde F_m\in\operatorname{Aut}_m^\nu((N^{m-1}_\nu(c_2^\nu,\dots, c^{m-1}_\nu))$ such that
\begin{gather}{}\nonumber
\tilde F_mF_m\Phi_{m-1}F_m^{-1}\tilde F_m^{-1}=N^{m-1}_{\nu}(c^2_\nu,\dots, c^{m-1}_\nu)(c^m_\nu)+O(m+1), \\ c^m_\nu e^m_\nu \in \mathcal H^m_\nu(N^{m-1}(c^2_\nu,\dots, c^{m-1}_\nu)).\nonumber
\end{gather}
Set $N^m_\nu(c^2_\nu,\dots, c^m_\nu)=N^{m-1}_{\nu}(c^2_\nu,\dots, c^{m-1}_\nu)(c^m_\nu)$. The formal normal form of $\Phi$ is $N_\nu^\infty$ with
\eq{}\nonumber
N_\nu^\infty=N^m_\nu(c^2_\nu,\dots, c^m_\nu)+O(m+1), \quad m=1,2,\dots.
\end{equation}
Define $\frak C^f_\nu(\Phi)$ to be $c^2_\nu e^2_\nu,\dots)\in \mathcal H^1_{\nu}(T_CM)$ for equivalence class of $\Phi$ under $\mathcal T_\nu$.
We have defined the normal forms $N_\nu^\infty(c^2_\nu,\dots)$ for $\Phi\in \mathcal M_\nu$. The rest is similar to the proof of \rt{injection}. We leave the details to the reader.
\end{proof}
For the rest of the section, we deal with the existence of transverse foliations using the normal form for the neighborhoods in Section 2. Let us first improve \rd{the-def} as follows.
\begin{defn}\label{the-def+}In \rd{the-def}, we select $N^\ell(c^2,\dots, c^\ell)$ satisfying
$$(N^\ell_{kj})^h(c^2,\dots, c^\ell)=\varphi_{kj}(h_j),
$$ if the set of neighborhoods $\Phi$ satisfies
$$
\Phi=FN^{\ell-1}(c^2,\dots, c^{\ell-1})F^{-1}+O(\ell)
$$
for some $F\in\operatorname{Aut}_{\ell-1}(N^{\ell-1}(c^2,\dots, c^{\ell-1}))$ has a neighborhood that admits a holomorphic transverse foliation.
\end{defn}
\begin{prop}\label{existence-formal}Let $\ell_0>1$ be an integer. Assume that
\begin{gather}{}\nonumber
H^1(C,T_CM\otimes S^\ell N_C^*)=0,\quad \forall\ell>\ell_0.
\end{gather}
Then $\Phi$ admits a formal transverse foliation if and only if there is a formal mapping $F\in\mathcal T$ such that
$F\Phi F^{-1}=N^{\ell_0}(c^2,\dots, c^{\ell_0})+O(\ell_0+1)$ with $(N^{\ell_0}(c^2,\dots, c^{\ell_0}))_{kj}^h=\varphi_{kj}(h_j)$.
\end{prop}
\begin{prop}Assume that $H^0(C, TC\otimes S^\ell N_C^*)=0$ for all $\ell>1$. Then two transverse foliations of neighborhoods of $C$ are equivalent if (and only if) the neighborhoods are equivalent.
\end{prop}
\begin{proof}
Suppose that $F=I+O(2)$ sends a transverse foliation $\Phi$ of a neighborhood $M$ into a transverse foliation $\tilde\Phi_{kj}$ of a neighborhood $\tilde M$. We have $\tilde\Phi_{kj}F_j=F_k\Phi_{kj}$.
Suppose that $F_j^h=I+[f^h_j]^m+O(m+1)$. We want to show that $[f^h_j]^m=0$. We know that $\Phi_{kj}^h(h_j,v_j)=\phi_{kj}(h_j)$ and $\tilde\Phi^h_{kj}(\tilde h_j,\tilde v_j)=\varphi_{kj}(\tilde h_j)$. We have
$$
\varphi_{kj}(I+[f_j^h]^m+O(m+1))=\varphi_{kj}(h_j)+[f_k^h(\Phi_{kj})]^m+O(m+1).
$$
Collecting terms of order $m$ in $v_j$, we obtain $\delta^h\{[f^h]^m\}=0$. Thus $[f^h]^m$ is a global section of $TC\otimes S^m N_C^*$. We conclude $[f^h]^m=0$.
\end{proof}
When $C$ is a compact Riemann surface with genus $g$, $H^0(C, TC\otimes S^\ell N_C^*)=0$
if $\deg N_C>\max\{0,g-1\}$ and $\ell>1$; see Section 6.
\setcounter{thm}{0}\setcounter{equation}{0}
\section{convergence of two classifications and criteria for transverse foliations}
In this section, we establish the convergence results for normalizations of full neighborhoods and transverse foliations by applying convergent results of Grauert~~\cite{MR0137127}, Griffiths~\cite{MR206980}, and Commichau-Grauert~~\cite{MR627752}. We also obtain a criteria for the existence of transverse foliations using our normal forms.
The first-order obstructions for the existence of transverse foliations were studied in~\cite{MR206980}, Morrow-Rossi~~\cite{MR627765}, where the convergence of transverse foliations are not addressed with the exception that $\dim M=2$ in~~\cite{MR627765} for which the classification of the foliations was addressed.
According to Grauert~\cite[Def. 1, p.~342]{MR0137127}, we say that $E$ is weakly negative if its zero section has a relatively compact strictly pseudoconvex neighborhood. We say that $E$ is weakly positive if $E^*$ is weakly negative~\cite[Def.~2, p.~342]{MR0137127}.
Grauert~\cite[Satz 1, p.~341]{MR0137127} proved that if the zero section of a vector bundle $E$ is exceptional, then $E$ is weakly negative.
According to Grauert~\cite[Def. 3, p.~339]{MR0137127}, a connected compact complex manifold $C$ of positive dimension in $M$ is called \emph{exceptional}, if there is a complex manifold $M'$ and a proper surjective holomorphic map $\phi\colon M\to M'$ such that $\phi(C)$ is a point $p$, $\phi\colon X\setminus C\to M'\setminus\{p\}$ is biholomorphic. In~\cite[Thm. 5,
p.~340]{MR0137127}, Grauert proved that $C$ is exceptional, if $C$ has a basis of strongly pseudoconvex neighborhoods.
It was proved by Grauert for codimension one compact complex manifold $C$ and by Hironaka and Rossi for higher codimension that the formal principle holds for exceptional sets.
Following Commichau-Grauert~~\cite{MR627752}, we say that a vector bundle $V$ on $C$ is \emph{$1$-positive}, if there is a tube neighborhood $W$ of the zero section $C$ of $V$ such that the Levi-form of $V$ has at least $1$-negative eigenvalue and $W\cap V_x$ is star-shaped for each $x\in C$ and $V_x$ intersects $\partial W$ transversely. We say that $V$ is \emph{$q$-positive}, if the $\partial W$ has $q$ negative Levi eigenvalues.
We recall a lemma mentioned by Griffiths~\cite{MR206980}.
\begin{lemma}[\!\!\cite{MR206980}, Lemma 2.1]\label{griffiths}
Let $\pi\colon N_C\to C$ be the normal bundle of $C$. Let $W$ be a neighborhood of $C$ in $N_C$. Let $F$ be any holomorphic vector bundle on $C$ and let $\pi^*F|_W$ be the pull-back bundle on $W$. Then
\eq{finite-dim}
\sum_{\ell\geq0}\dim H^q(C,F\otimes S^\ell N_C^*)\leq\dim H^q(W,\pi^*F|_W).
\end{equation}
\end{lemma}
\begin{proof} We provide a proof using our notation for $q=1$ only.
Let $\mathcal U$ be a covering of $C$ by open sets $U_j$. We may assume that there are open sets $V_j$ in $W$ so that $U_j=V_j\cap C$. Using additional open subset of $W$, we assume that $\{V_j\}$ is a covering $\mathcal V$ for $W$. Furthermore, we may assume that $U_j$ and $ V_k$ are Stein, $E$ and $ F$ holomorphically trivial on $U_j$.
The vector bundles $F$ and $N_C$ play different roles. For $F$, we simply pull it back as $\pi^*F$ as vector bundle on $N_C$. This allows turn sections of $F\otimes S^\ell N_C^*$ into $\pi^*F$ valued homogeneous polynomials on the manifold $N_C$. To be specific, let us identify a section of $F\otimes S^\ell N_C^*$ by a ${\mathbb C}$-linear injection
$$
\iota \colon C^q(\mathcal U, F\otimes S^\ell N_C^*)\to C^q(\mathcal V, \mathcal O(\pi^*F|_W)).
$$
Recall that $\pi^{-1}(U_j)$ has coordinates $(h_j,v_j)$. Let $\{\tilde e_j^{\mu}\}$ (resp. $\{e_j^{\mu}\}$) be a basis of $F_{|U_j}$ (resp. $N_C$). Let $ \{f_{kj}\}\in C^1(\mathcal U, F\otimes S^mN_C^*)$. Then by \re{def-tilde-f+},
\eq{tfkm-M}\nonumber
f_{kj}(p)=\sum_{\mu=1}^{n+d}\sum_{|Q|=m}f_{kj,Q}^\mu(h_k(p),v_k(p)) \tilde e_k^\mu(p)\otimes (w_j^*(p))^Q.
\end{equation}
Let $\tilde p\in \pi^{-1}(p)$ and $p\in U_j\cap U_k$. In coordinates, we have $\tilde p=(p,e_j(p)\cdot v_j(p))$. Define
$$
(\iota f)_{kj}(\tilde p)=\sum_{\mu=1}^{n+d}\sum_{|Q|=m}f_{kj,Q}^\mu(h_k(p), v_k(p)) v_j^Q(p) \tilde e_k^\mu(p).
$$
Now $\{\pi^{-1}(U_j)\}$ is an open covering $\hat{\mathcal V}$ of $F$ as a complex manifold and $\iota f\in C^{1}(\hat{\mathcal V}, \mathcal O(F))$. Let $\tilde{\mathcal V}$ be the covering $W$ defined by $\{V_i\cap\pi^{-1}U_j\}$. Note that $V_i\cap\pi^{-1}U_j$ are still Stein. Analogously, if $ \{u_j\}\in C^0(\mathcal U, F\otimes S^mN_C^*)$, then
$$
u_{k}(p)=\sum_{\mu=1}^{n+d}\sum_{|Q|=m}u_{k,Q}^\mu(h_k(p),v_k(p)) \tilde e_k^\mu(p)\otimes (w_k^*(p))^Q.
$$
Define
$
(\iota u)_{k}(\tilde p):=\sum_{\mu=1}^{n+d}\sum_{|Q|=m}u_{k,Q}^\mu(h_k(p),v_k(p)) v_k^Q(p) \tilde e_k^\mu(p).
$
In the other way around, given such en expression $U_k(q)$, we associate a 0-cochain $ \tilde U=\{\tilde U_j\}\in C^0(\mathcal U, F\otimes S^mN_C^*)$.
Then $\delta \iota =\iota\delta$.
We want to show that $\iota$ induces an injection
$$
\iota\colon \check H^{ 1}(\mathcal U,F\otimes S^\ell N_C^*)\to \check H^{ 1}(\tilde{\mathcal V}, \mathcal O(\pi^*F|_W)).
$$
Suppose that $f\in Z^{ 1}(\mathcal U,F\otimes S^\ell N_C^*)$ and $\iota f=\delta u$ with $u\in C^{0}(\tilde{\mathcal V}, \mathcal O(\pi^*F|_W))$. If $V_j\cap C$ is non-empty, we have $V_j\cap C=U_j$ and we can expand
$
u=\sum u_{m},
$
where $u_m$ is a homogeneous polynomial in $v_j$ of degree $m$.
Then $f=\delta\tilde u_\ell$. This gives us \re{finite-dim}.
\end{proof}
It is known from the Andreotti-Grauert
theory~~\cite{MR150342} that if $\partial W$ has $(q+1)$ negative Levi eigenvalue or $(n+d-q)$ positive Levi eigenvalues at each boundary point, then
$$
\dim H^q(W,\mathcal F)<\infty
$$
for any coherent analytic sheaf $\mathcal F$ on $W$.
Our convergent classifications are based on the following theorems, which can be regarded as strong forms of the formal principle.
\begin{thm}[\!\!\cite{MR0137127}, Satz 7, p.~363]
\label{conv1} Let $N_C$ be negative. Then there exists an integer $\ell_0$ such that
$H^1(C,T_CM\otimes S^\ell N_C^*)=0$ for $\ell>\ell_0$. Furthermore, if $\hat F_j$ are holomorphic mappings such that $\tilde\Phi_{kj}=\hat F_k^{-1}\Phi \hat F_j+O(\ell_0+1)$, then there are holomorphic mappings $F_j$ such that $F_j=\hat F_j+O(\ell+1)$ and $F_k^{-1}\Phi_{kj}F_j=\tilde\Phi_{kj}$.
\end{thm}
\begin{rem}\label{HR-remark}
Let $C$ be a compact Riemann surface and let $N_C$ be a line bundle with $\deg N_C<0$,
By Riemann-Roch (see \re{expansion}-\re{expansion+} below), we have
$$\lim_{\ell\to\infty}\ell^{-1}\dim H^0(C,T_CM\otimes S^\ell N_C^*)=-\deg N_C.$$ Using this one can prove that there are divergent formal mappings that preserve the germs of neighborhoods of the zero section of $N_C$; see \rp{MR-issue} below.
\end{rem}
On the other hand, with $1$-positivity, we have the following {\em strong } formal principle.
\begin{thm}[\!\!\cite{MR627752}, Satz 4, p.~119] \label{conv2}Let $N_C$ be $1$-positive. Let $ F_j$ be formal biholomorphic mappings such that $\tilde\Phi_{kj}=F_k^{-1}\Phi_{kj}F_j$. Then $F_j$ must converge.
\end{thm}
Commichau-Grauert~~\cite{MR627752} proved the theorem for codimension $1$ case first. Their proof for the higher codimensions follows from a blow-up ~~\cite[p.~115 and p.~126]{MR627752} along $C$. Indeed, one can verify easily as follows. By blowing up a neighborhood $M$ along $C$, we obtain $\tilde M$ and $\tilde C$ such that $\tilde C$ has codimension $1$ while $\partial\tilde M$ is biholomorphic to $\partial M$. If $F\colon (M,C)\to(M',C)$ is a formal mapping that is tangent to the identity along $C$, then the blow-up induces a formal mapping $\tilde F$ from $(\tilde M,\tilde C)$ to $(\widetilde {M'},\tilde C)$ (that may not be tangent to the identity along $\tilde C$). However, the theorem for codimension $1$ case, which is proved for any formal mapping that preserves $C$, implies that $\tilde F$ is convergent. Consequently, $F$ is also convergent.
We now prove the convergence classification for the neighborhoods.
\begin{cor}\label{full-fd} Let $N_C$ be weakly negative or $2$-positive. Then the set of holomorphic equivalence classes of neighborhoods of $C$ is identified with a subset of finite dimensional space $\mathcal H(T_CM)<\infty$. Furthermore, a neighborhood $\Phi$ admits a transverse foliation if and only if its normal formal $N^\infty(\Phi)$ satisfies
$(N_{kj}^\infty(\Phi))^h=\varphi_{kj}$.
\end{cor}
\begin{proof}Let us summarize the proof of \rt{injection}.
Let $\Phi$ be a holomorphic (convergent) neighborhood of $C$ such that $\Phi=N^1+O(2)$, where $N^1$ is a convergent neighborhood. Then we find a (convergent) biholomorphism $F_2$ such that $F_2^{-1}\Phi F_2=N^1+c^2(\Phi)e^2+O(3)$ and choose a convergent neighborhood $N(c^2(\Phi))$ such that $F_2^{-1}\Phi F_2=N(c^2(\Phi))+O(3)$. Inductively, we find biholomorphism $F_2,\dots, F_m$ and convergent neighborhoods $N(c^2(\Phi)e^2, \dots, c^m(\Phi))$ such that
$$
F_m^{-1}\cdots F_2^{-1}\Phi F_2\cdots F_m=N(c^2(\Phi), \dots, c^m(\Phi))+O(m+1).
$$
Choose $m$ so that $H^1(C,T_CM\otimes S^\ell N_C^*)=0$ for $\ell>m$.
Hence they are holomorphically equivalent by \rt{conv1}.
\end{proof}
We obtain the classification for transverse foliations.
\begin{cor}
Let $C$ be a compact complex manifold.
Suppose that $N_C$ is $1$ positive.
\begin{list}[align=left]{$(\roman{ppp})$}{\usecounter{ppp}}\item
If two transverse foliations are equivalent by a formal biholomorphic mapping $F$ preserving the foliations, then $F$ is also convergent.
\item Assume further that
$
\dim H^1(C,N_C\otimes S^\ell N_C^*)<\infty
$
for some $\ell_0$.
There is an injective mapping from the set ${\mathcal M_\nu}/{\,\sim}$ of holomorphic equivalence classes of the transverse foliations into the finite-dimensional space
$\mathcal H^1_\nu(T_CM).$
\end{list}
\end{cor}
\begin{proof} Assertion $(i)$ is a consequence of \rt{conv2}. Assertion $(ii)$ follows from $(i)$ and the formal classification by \rt{injection-h}.
\end{proof}
When $C$ is a compact Riemann surface and $N_C$ is a positive line bundle, Morrow-Rossi, using the theorem of Commichau-Grauert, the equivalence classes of transverse foliations under foliation-preserving transformations are actually infinite dimensional.
We conclude this section by considering the classification of neighborhoods of $C$ that are not necessary tangent to the identity. We will however use coordinate changes that fix $C$ pointwise and preserve $N_C$. We consider the case of $N_C$ is a line bundle. Let $t_{kj}$ be the transition functions of $N_C$. An isomorphism of $N_C$ is given by $g_k^{-1}t_{kj}g_j=t_{kj}$. Thus $g_j$ define a global holomorphic function on $C$ without zero. Since $C$ is compact, then the function must constant. Now it is easy to see that the transition functions for the neighborhood $\Phi_{kj}(h_j,v_j)$ is transformed into $(\Phi_{kj}^h(h_j,cv_j), c^{-1}\Phi_{kj}^v(cv_j))$. Note this non-homogenous dilation is used by Morrow-Rossi~~\cite{MR627765} to get a complete set of moduli spaces when $\dim C$ and $\operatorname{codim} C$ are $1$, and $N_C$ is negative.
\setcounter{thm}{0}\setcounter{equation}{0}
\section{ Riemann surfaces in complex surfaces}
We illustrate Grauert's result by showing that the obstructions exists for a neighborhood with negative normal line bundle being not holomorphically equivalent to a neighborhood of its zero section; compare \cite[p.~211]{MR947141} on Grauert's work.
We also prove the assertion in \rrem{HR-remark}.
This explains why we could not apply Grauert's formal principle for weakly negative $N_C$ to the classification of transverse foliations, and it seems to the authors that an application of a formal principle to the transverse foliations
needs a statement stronger than \rt{conv1}. An interested reader should consult
Morrow-Rossi~~~\cite[p.~323, line~3, and Thm. 6.3]{MR627765} for negative $N_C$.
Let us first recall some facts on the dimensions of $0$-th and first cohomology groups of line bundles.
Note that a line bundle on a compact Riemann surface is positive if and only if its degree is positive.
Let $C$ be a compact Riemann surface with genus $g$. When $L\to C$ is a line bundle with degree $\nu_L<0$, $L$ has no global section. Recall the duality
$$
H^q(C,\Omega^p(E))=H^{1-q}(C,\Omega^{1-p}(E^*))^*, \quad q=0,1.
$$
We have
$\deg TC=2-2g$ and $\deg K_C=2g-2$, where $K_C$ is the canonical line bundle of $C$.
The Riemann-Roch theorem says that if $h^0(L)$ is the dimension of $H^0(C,\mathcal O(L))$ then
\eq{RR}\nonumber
h^0(L)-h^0(K_C\otimes L^{-1})=\deg L+1-g.
\end{equation}
Thus
\eq{crit}\nonumber
h^0(L)\geq \deg L-g+1.
\end{equation}
This provides the following useful estimates for positive $N_C$:
\begin{gather}\label{expansion}
\dim H^1(C, NC\otimes S^\ell N_C^*)\geq g-1+(\ell-1)\deg N_C,\\
\dim H^1(C, TC\otimes S^\ell N_C^*)\geq 3g-3+\ell\deg N_C.\label{expansion+}
\end{gather}
\subsection{Normal forms on neighborhoods}
\subsubsection{Negative $N_C$} In this case, the formal principle and \re{for-F1} hold.
We have
\eq{}\nonumber
H^1(C, (T_C\oplus N_C)\otimes S^\ell N_C^*)=(H^0(C, K_C\otimes(K_C\otimes S^\ell N_C+N_C^*\otimes S^\ell N_C))^*.
\end{equation}
Now $d_\ell^h:=\deg K_C\otimes(K_C\otimes S^\ell N_C)=4(g-1)+\ell\deg N_C$ and $ d_\ell^v:=\deg K_C\otimes(N_C^*\otimes S^\ell N_C)=2(g-1)+( \ell-1)\deg N_C$. Thus
\begin{gather}{}\nonumber\dim H^1(C, TC\otimes S^2N_C^*)\geq 3g-3+2\deg N_C,\\
\dim H^1(C, N_C\otimes S^2N_C^*)\geq d_2^v-g+1=g-1+2\deg N_C.\nonumber
\end{gather}
By Morrow-Rossi~~\cite[Thm.~6.3]{MR627765}, each element in $H^1(C,N_C\otimes S^2 N_C^*)$ can be realized by transverse foliations.
See Camacho-Movasati-Sad~~\cite{MR1967036} for a different approach.
This shows that an element corresponding to a non-zero element in $H^1(C,N_C\otimes S^2 N_C^*)$ is not equivalent to a neighborhood of the zero section of $N_C$.
\subsubsection{Positive $N_C$ and negative $T_CM\otimes S^2 N_C^*$} This occurs if and only if $\deg N_C> \max\{0,g-1\}$, in which case \re{for-F0} holds. Then the formal principle holds. Also, both
$$
\oplus_{\ell>1} H^1(TC\otimes S^\ell N_C^*),\quad \oplus_{\ell>1} H^1(C,N_C\otimes S^\ell N_C^*).
$$ are infinite dimensional.
\begin{prop}\label{realization-MR} Let $C$ be a compact Riemann surface with genus $g$. Suppose $N_C$ is a line bundle. For any finite $r$, the following hold:
\begin{gather}\label{finite-sum-case}
\oplus_{m=2}^rH^1(C,T_CM\otimes S^m N_C^*)\subset\frak C(\mathcal M/\sim)\subset\oplus_{m=2}^\infty H^1(C,T_CM\otimes S^m N_C^*)\end{gather}
if $ \deg N_C>\max\{0,g-1\}$; if $\deg N_C>0$ and $T_CM$ splits then
\begin{gather}{}
\oplus_{m=2}^rH^1(C,N_C\otimes S^m N_C^*)\subset\frak C_\nu(\mathcal M_\nu/\sim)\subset\oplus_{m=2}^\infty H^1(C,N_C\otimes S^m N_C^*).
\label{finite-sum-case-nu}
\end{gather}
\end{prop}
\begin{proof}
Our proof is based on a construction in ~\cite{MR627765} to realize $\oplus_{\ell>1}H^1(C, N_C\otimes S^\ell N_C^*)$ for transverse foliations. Thus \re{finite-sum-case-nu} is essentially in ~\cite{MR627765}. As indicated in ~\cite{MR627765}, the construction applies to
$$\oplus_{\ell>1}H^1(C, T_CM\otimes S^\ell N_C^*)$$
as well, which we show below. We will need
$$
H^0(C,T_CM\otimes S^\ell N_C^*)=0, \quad\forall \ell>0
$$
which holds for $\deg N_C>\max\{0,g-1\}$. Thus $\operatorname{Aut}_m(N^m(c^2,\dots, c^m))$, defined as
$$\{F=Id+O(2)\colon FN^m(c^2,\dots, c^m)=N^m(c^2,\dots, c^m) F+O(m+1)\}$$
consists of mappings of the form $F=Id+O(m+1)$.
Take a holomorphic disk $U_0$ in $C$ and let $U_0'$ be a smaller disc in $U_0$. Then $U_0,U_1:=C\setminus U_0'$ form a Leray covering of $C$ as both $U_0,U_1$ are Stein. $U_0\cap U_1$ is an annulus biholomorphic to
$\{z\in{\mathbb C}\colon r<|z|<1/r\}$. Since $C$ is covered by two sets, then
\eq{Z=C}\nonumber
Z^1(\mathcal U,L)=
C^1(U_0\cap U_1, L).
\end{equation}
Since a line bundle on an open Riemann surface is holomorphically trivial~\cite[p.~52]{MR0207977}, then $N_C$ is completely determined by
$$
(\varphi_{10}(z_0),t_{10}(z_0)v_0)
$$
where $t_{10}(z)$ is non-vanishing holomorphic function on $U_0\cap U_1$ and $\varphi_{10}$ is holomorphic and injective on $U_0\cap U_1$. Also $\deg N_C$ is the winding number of $t_{10}$ on the unit circle. A neighborhood of $C$ with $N_C|_{U_1}$ being trivial is precisely given by
$$
\Phi_{10}(z_0,v_0):=(\varphi_{10}(z_0)+l_{10}(z_0)v_0,t_{10}(z_0)v_0)+
\sum_{\ell>1}\Phi_{10,\ell}(z_0)v_0^\ell
$$
by patching
$
(U_0\sqcup U_1)\times\Delta_\epsilon/\!\sim
$
with $(z_1,v_1)\sim\Phi_{10}(z_0,v_0)$. Here $\Phi_{10,\ell}$ are otherwise any holomorphic functions on ${U_{01}}\times \Delta_\epsilon$ subject to the condition that \eq{constr}
\sum_{\ell>1}\sup_{K}|\Phi_{10,\ell}|<C_K<\infty
\end{equation}
for any compact subset $K$ of $U_{01}\times\Delta_\epsilon$. By an abuse of notation, the $(z_1,U_1)$ stands for a finite collection of coordinate charts $(z_j,U_j)$, where $U_1=\cup \tilde U_j$ and $z_j$ maps $\tilde U_j$ onto the unit disc.
When $l_{10}=0$, $T_CM$ splits. When $\Phi_{10,\ell}^h=0$, the neighborhood admits transverse foliation. When $t_{10}, l_{10}$ and $\Phi_{10,\ell}^v$ are constant, the neighborhood admits tangential foliations.
The degree of $N_C$ is the winding number of $t_{10}$ on the unit circle.
To realize each element in \re{finite-sum-case}, we recall briefly the construction in \rt{injection}. Since $C$ is covered by two sets, we can take $$N^1=(\varphi_{10}(z_0)+l_{10}(z_0)v_0,t_{10}(z_0)v_0).
$$
We now take the advantage that $\Phi_{01,\ell}$ need only to satisfy \re{constr} and are otherwise arbitrary. Fix a finite basis $e^2$ for $H^1(\mathcal U,T_CM\times S^2N_C^*)$. Let $e^2_j$ be represented by $C^1(\mathcal U, T_CM\times S^2N_C^*)$, still denoted by $e^2_j$. We also use the identification via \re{def-tilde-f} and \re{def-tilde-f+}. Thus each element $c^2e^2\in H^1(\mathcal U,T_CM\times S^2N_C^*)$ is associated with $c^2e^2$ in the form
$$
N^2(c^2)=(\varphi_{10}(z_0)+l_{10}(z_0)v_0,t_{10}(z_0)v_0)+ c^2\hat N_2^2(z_0)v_0^2.
$$
Since $C$ is covered by two open sets, $N^2(c^2)$ is indeed a cocycle for transition functions of a neighborhood of $C$.
For finitely many $c^2,\dots, c^r$, $\sum_{\ell=2}^r c^\ell e^\ell$ is associated to $N^r(c^2,\dots, c^r)$ satisfying the convergence constrain \re{constr}.
\end{proof}
\subsection{Verifying \rrema{HR-remark}}
\begin{prop}\label{MR-issue}Assume that $T_CM$ splits. Assume that $H^1(C,T_CM\otimes S^\ell N_C^*)=0$ for $\ell\geq \kappa$ for a finite $\kappa$.
\begin{list}{$\hspace{-1em}(\alph{pp})$}{\usecounter{pp}}\item Suppose $H^0(C,T_CM\otimes S^\ell N_C^*)\neq0$ for a sequence $\ell=\ell_j\to\infty$. Then there are divergent formal mappings that preserve the germs of neighborhoods of the zero section of $N_C$.
\item
Suppose $H^0(C,N_C\otimes S^\ell N_C^*)\neq0$ for some $\ell>1$. Then there is a possibly divergent mapping $F$ preserving the germ of the zero section of $N_C$ while $F$ does not preserve the transverse foliation of $N_C$.
\end{list}
\end{prop}
\begin{proof}
If $T_CM$ has locally constant transition functions, the proof is straightforward as each global section $[\phi]^m$ of $T_CM\otimes S^mN_C^*$ gives us an automorphism of $T_CM$ of the form $F=I+f$ where $f$ is homogeneous of degree $m$. Then $F=I+\sum_{\ell>1}f_m$ diverges for suitable choices of coefficients of the global sections of $T_CM\otimes N_
c^\ell$ for $\ell>1$.
When $T_CM$ is not flat, we verify the assertion by a \emph{reversing} linearization procedure. Take $j$ such that $\ell_{j}\geq \kappa$.
Using a global section $f^*_{\ell_{\ell_j}}$ of $T_CM\otimes S^{\ell_j} N_C^*$, we find $F_{\ell_0}=I+f_{\ell_j}$ where $f_{k, \ell_j}$ is homogeneous of degree $\ell_j$ in $v_k$ variables. Then
\eq{m-comm}
F_{\ell_j}N^1 F_{\ell_j}^{-1}=N^1+O(\ell_j+1)
\end{equation}
By condition on $H^1$ and \re{m-comm}, we find a formal mapping $\tilde F_{\ell+1}=I+O(\ell_j+1)$ such that $\tilde F_{\ell_j+1}F_{\ell_j}N^1 F_{\ell_j}^{-1}\tilde F_{\ell_j+1}^{-1}=N^1$. We repeat this and find
$
F_{\ell_i}, \tilde F_{\ell_i+1}
$
for $i>j$
such that $\tilde F_{\ell_{i}+1}F_{\ell_i}$ and $N^1$ commute.
Then
$$
F_\infty=\lim_{i\to\infty} \tilde F_{\ell_i+1}F_{\ell_i}\circ\cdots\circ \tilde F_{\ell_j+1}F_{\ell_j}
$$
commutes with $N^1$. We have
$$
F_\infty=F_{\ell_i}\cdots \tilde F_{\ell_j+1}F_{\ell_j}+O(\ell_i+2),\quad
F_\infty=\tilde F_{\ell_i+1}F_{\ell_i}\cdots \tilde F_{\ell_j+1}F_{\ell_j}+O(\ell_i+2).
$$
In other words,
$$
[F_\infty]_{\ell_i}=f_{\ell_i}+R(f_{\ell_j},\dots, f_{\ell_{i-1}})
$$
When the coefficients of $\tilde f_{\ell_i}$ grows sufficiently fast as $j\to\infty$, we get a divergent $F_\infty$. This proves the first assertion.
The second assertion can be proved by a similar argument.
\end{proof}
As an example, the condition in \rp{MR-issue} (a) is satisfied when $C$ is the Riemann sphere and $N_C$ is a negative line bundle.
\newcommand{\doi}[1]{\href{http://dx.doi.org/#1}{#1}}
\newcommand{\arxiv}[1]{\href{https://arxiv.org/pdf/#1}{arXiv:#1}}
\def\MR#1{\relax\ifhmode\unskip\spacefactor3000 \space\fi%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{MR#1}}
|
\section*{Abstract}
{\bf
The generalized Rosenzweig--Porter model arguably constitutes the simplest random matrix ensemble displaying a non-ergodic delocalized phase, which we characterize here by using replica methods. We first derive analytical expressions for the average spectral density in the limit in which the size $N$ of the matrix is large but finite. We then focus on the number of eigenvalues in a finite interval and compute its cumulant generating function as well as the level compressibility, i.e., the ratio of the first two cumulants: these are useful tools to describe the local level statistics. In particular, the level compressibility is shown to be described by a universal scaling function, which we compute explicitly, when the system is probed over scales of the order of the Thouless energy.
We confirm our results with numerical tests.
}
\vspace{10pt}
\noindent\rule{\textwidth}{1pt}
\tableofcontents\thispagestyle{fancy}
\noindent\rule{\textwidth}{1pt}
\vspace{10pt}
\section{Introduction}
Quantum non-interacting particles in a disordered potential undergo the
Anderson localization transition as the disorder is increased~\cite{anderson1958absence}. In one and two dimensions an infinitesimal amount of disorder is sufficient to localize all eigenstates of the Hamiltonian, while in dimension larger than two a critical value of the disorder separates a metallic phase, where the eigenstates are similar to plane
waves and spread over the whole volume uniformly, from an insulating phase, where
the eigenstates are instead exponentially localized around specific points
in space and thereby occupy a finite $\cor{O}(1)$ portion of the total volume. It
is well established that exactly at the Anderson localization critical
point the wave-functions are \emph{multifractal}~\cite{wegner1980inverse,rodriguez2011multifractal}. This means that they are
neither fully delocalized (as in the metallic regime), nor fully localized
(as in the insulating phase), since their support set grows with the system
size but remains a vanishing fraction of the total volume.
In the last decade, the Hilbert space localization properties of quantum
disordered many-body systems have attracted much interest. In this context,
the emergence of this kind of multifractal states has been discussed as a
key and robust feature of their
phase diagram and has been invoked to explain some of their unconventional
properties beyond the single-particle limit. In the many-body setting,
multifractal eigenstates that do not cover the whole accessible Hilbert
space generically lead to the violation of the eigenstate thermalization
hypothesis (ETH)~\cite{srednicki1994chaos,rigol2008thermalization}, and are therefore often called partially delocalized but
non-ergodic (in contrast with the ergodic fully delocalized eigenstates
which satisfy ETH).
In the context of Many Body Localization (MBL), recent studies indicate that the many-body
eigenstates are in fact multifractal in the whole insulating phase~\cite{mace2019multifractal,de2021rare,gornyi2017spectral,tarzia2020many}. Furthermore, the seminal work by Basko, Aleiner and Altshuler~\cite{basko2006metal} predicted the existence of a novel unconventional ``bad
metal'’ regime in between the fully ergodic metallic phase at low disorder
and the insulating one at strong disorder. Following the pioneering ideas
of~\cite{altshuler1997quasiparticle}, the unusual properties of the bad metal regime have been
put in relation with the possible multifractal nature of the many-body
eigenstates. Recent investigations of the out-of-equilibrium phase diagram of
the quantum random energy model~\cite{Faoro_2019,baldwin2018quantum,biroli2021out,parolini2020multifractal,kechedzhi2018efficient} and Josephson junction arrays~\cite{pino2017multifractal,pino2016nonergodic} seem
to support these ideas. The existence of partially extended but non-ergodic
wave-functions is also believed to have relevant practical and conceptual
implications in the efficient population transfer in the context of quantum
computing~\cite{parolini2020multifractal,kechedzhi2018efficient,Smelyanskiy_2020}. Moreover, recent studies of the Sachdev-Ye-Kitaev model in
high-energy physics and quantum gravity have reported evidence for the
emergence of a non-ergodic extended phase when the model is perturbed by a
single-body term~\cite{Atland_2019,monteiro2021minimal}.
Matrix models have been an invaluable tool to describe and help
understanding complex physical systems, in particular those with quenched
randomness. The physical mechanism at the origin of the above-mentioned
partially
extended but non-ergodic eigenstates is one such problem, and specific
matrix models have recently been used as proxies which capture the peculiar
spectral properties associated with them. In this respect, the
Rosenzweig-Porter (RP) random matrix ensemble~\cite{RP_1960}, originally introduced to
reproduce the spectral properties of complex atomic spectra, provides an
archetypal illustration of a system in which a non-ergodic extended phase
featuring fractal eigenstates (along with other unconventional spectral
properties that will be extensively discussed below) appears in an
intermediate region of the phase diagram between a fully ergodic
delocalized phase and a fully Anderson localized phase. For this reason the
RP model has been the focus of a strong resurgence of attention over the
last few years~\cite{Kravtsov_2015,vonSoosten_2019,Facoetti_2016,Truong_2016,Bogomolny_2018,DeTomasi_2019,amini2017spread,pino2019ergodic,berkovits2020super}. Although one cannot expect that simple random matrix models could capture all the properties of interacting quantum systems, they
provide natural and powerful tools to understand the deep physical
mechanisms behind some of their features, which are often elusive to
analytical treatments in more realistic settings.
\comments{
In the last few years, the appearance of delocalized but non-ergodic eigenstates has emerged as a key feature of disordered many-body quantum systems. These extended but non-ergodic wave-functions are neither fully localized nor fully delocalized. \ao{A customary tool that characterize the degree of delocalization of an eigenstate $\vert \psi \rangle$ is provided by the so-called inverse participation ratio (IPR), defined as $I_2 = \sum_i \vert \langle i \vert \psi \rangle \vert^4$, which is essentially a measure of the inverse volume ``occupied'' by $\vert \psi \rangle$. In fact, the amplitudes of fully Anderson localized eigenstates are significantly different from $0$ only on a finite region of space and $I_2 \sim \cor{O}(1)$. Fully delocalized wave-functions are similar to plane waves, they spread over the whole volume with coefficients $\vert \langle i \vert \psi \rangle \vert \sim O(\sqrt{N})$ to ensure normalization, and $I_2 \propto 1/N$. Partially delocalized but non-ergodic eigenstates, instead, are \emph{(multi)fractal}, in the sense that they occupy a subextensive portion of the whole accessible Hilbert space. The wave-functions' coefficients are significantly different from $0$ on a support set that scales as $N^{D}$ with $0<D<1$, i.e., it grows with $N$, but remains much smaller than the total volume, yielding $I_2 \propto N^{-D}$.}
\dav{I hope this sounds constructive as I mean it, but by reading this paragraph I had the impression that the same message could be conveyed without introducing the IPR (it may even require an extra effort to understand it...). Can I suggest to keep the structure as it is, but simply remove the references to $I_2$?}
\ao{On the one hand, it is well established that exactly at the Anderson localization critical point the single-particle eigenstates are multifractal~\cite{}. On the other hand, many recent investigation have advocated the existence of non-ergodic wave-functions to explain the exotic properties of several interacting quantum disordered systems: } In the context of the Many-Body Localization (MBL) transition~\cite{} recent studies suggest that the many-body eigenstates are in fact multifractal in the whole insulating phase~\cite{}. Furthermore, the so-called ``bad metal'' regime, separating the thermal phase from the insulating one, has also been associated to the non-ergodic properties of the many-body wave-functions~\cite{}. Recent investigations of the nonequilibrium phase diagram of the quantum random energy model~\cite{} and Josephson junction arrays~\cite{} seem to support these ideas~\cite{}, which are believed to have relevant implications in the efficient population transfer in the context of quantum computing~\cite{}. Recent investigations of the Sachdev-Ye-Kitaev model in high-energy physics and quantum gravity have reported evidence for the emergence of a non-ergodic extended phase when the model is perturbed by a single-body term~\cite{}. \gs{[I still find this first paragraph a bit hard to read for a non-specialist!]}.\dav{I agree that it's not really RMT-community friendly.}
\lfccomm{me too, this is what I meant by "I have trouble with the 1st paragraph" in the email}
In this respect, the Rosenzweig-Porter (RP) random matrix ensemble~\cite{RP_1960}, originally introduced to reproduce the spectral properties of complex atomic spectra, provides an archetypal illustration of a system in which a non-ergodic extended phase featuring fractal eigenstates (along other unconventional spectral properties that will be extensively discussed below) appears in an intermediate region of the phase diagram between a fully ergodic delocalized phase and a fully Anderson localized phase. For this reason the RP model has been the focus of a strong resurgence of attention over the last few years~\cite{Kravtsov_2015}. Although one cannot expect that simple random matrix models could capture all properties of interacting quantum systems, they provide natural and powerful tools to understand the deep physical mechanisms behind some of their features, which are often elusive to analytical treatments in more realistic settings.
}
The Hamiltonian of the RP model $\cor{H}=A+c(N)B$ can be written as the sum of an $N \times N$ diagonal matrix $A$, whose entries $a_i$'s are independent and identically distributed (i.i.d.) random variables drawn from a Gaussian distribution $p_a(a_i)$, and another $N \times N$ random matrix $B$ belonging to the Gaussian orthogonal (or unitary) ensemble (GOE or GUE, respectively). If the variances of the matrix elements $b_{ij}$ are chosen of $\order{1}$, then the width of the spectrum of $B$ is of $\order{\sqrt{N}}$: thus the matrix $A$ (whose spectral width is of $\order{1}$) can produce significant deviations from the GOE/GUE behavior only if $c(N)$ decays sufficiently fast for large $N$.
The properties of this model have been extensively studied by using different techniques, such as a mapping to the Dyson Brownian motion \cite{Pandey_1995}, supersymmetry \cite{Guhr_1996,Guhr_1997}, resolvent methods \cite{Brezin_1996,Kunz_1998}, and first order perturbation theory \cite{Atland_1997}.
Due to the strong surge of attention towards non-ergodic extended states in quantum many-body disordered systems,
a generalized version of the RP model in which the distribution $p_a(a_i)$ is not necessarily Gaussian has then been introduced in Ref.~\cite{Kravtsov_2015}, and thoroughly investigated by using the techniques recalled above \cite{vonSoosten_2019,Facoetti_2016,Truong_2016,Bogomolny_2018,DeTomasi_2019,amini2017spread,pino2019ergodic,berkovits2020super} -- we will refer to this as the GRP model. In addition, new connections and applications have been pointed out in disordered elastic systems \cite{Krajenbrink_2021}, many body localization \cite{Faoro_2019,baldwin2018quantum,biroli2021out,tarzia2020many}, quantum gravity \cite{Atland_2019,monteiro2021minimal}, quantum information \cite{parolini2020multifractal,Smelyanskiy_2020,kechedzhi2018efficient}, or models of theoretical ecology \cite{Mergny_2021}.
In this work we revisit the generalized RP model by analyzing some properties of the energy levels and their correlations which have not been investigated in the literature yet. In particular, we perform a thorough study of the finite-$N$ corrections to the average spectral density and compute the level compressibility in the intermediate phase, thereby providing a deeper understanding of the properties of the non-ergodic extended regime. Our analysis uses the replica method largely exploited in the analysis of spin glass models~\cite{Mezard_1987}, but not only -- for example, this tool has been recently applied to study the properties of the ground states in a deformed GOE ensemble~\cite{Harukuni_2022}. In our study, and for simplicity, we will restrict to the case in which the entries of $\cor{H}$ are real numbers, so that $\cor{H}$ belongs to the deformed GOE ensemble (the GUE case can be analyzed with only minor changes of the calculations that we develop).
In the rest of this introductory section we will present the RP model and some of its salient properties (\cref{subsec:RPmodel}),
and we will outline our study and main results
(\cref{subsec:outline}).
\begin{figure}
\centering
\littlescheme
\caption{Sketch of the different phases of the RP model, depending on the value of the parameter $\gamma$ in \cref{eq:hamiltonian}. Considering the average spectral density $\rho(\lambda)$, a transition is first observed at $\gamma=1$ which separates a fully delocalized phase where $\rho(\lambda)=\rho_\T{GOE}(\lambda)$ from a non-ergodic delocalized phase where $\rho(\lambda)=p_a(\lambda)$ in the thermodynamic limit (see the main text). Focusing instead on the local level statistics, another transition is found at $\gamma=2$ from the delocalized (non-ergodic) phase characterized by the Wigner-Dyson statistics, to an Anderson-localized phase characterized by the Poisson statistics. The shaded region $1<\gamma<2$ indicates the non-ergodic extended phase studied in this work.}
\label{fig:scheme}
\end{figure}
\subsection{The generalized Rosenzweig-Porter model}
\label{subsec:RPmodel}
We consider the Hamiltonian represented by the $N \times N$ matrix
\begin{equation}
\cor{H} = A + \frac{\nu}{N^{\gamma/2}} B \, ,
\label{eq:hamiltonian}
\end{equation}
where the matrix $B$ belongs to the GOE ensemble: its elements are Gaussian random variables with zero mean and unit variance (i.e., $\langle b_{ij}^2 \rangle = 1$ for $i \neq j$ and $\langle b_{ii}^2 \rangle = 2$). With this choice, the spectrum of $B$ in the limit $N\to \infty$ converges to a Wigner semicircle supported within $\lambda \in [-\sqrt{2N},\sqrt{2N} ]$, where we denote hereafter with $\lambda$ the eigenvalues of $\cor{H}$. The parameter $\nu$ is of $\order{1}$ and does not scale with $N$. The deformation matrix $A$ is instead diagonal, with independent entries $a_i$'s identically distributed according to a generic distribution $p_a (a)$, hence the name Generalized Rosenzweig-Porter model (GRP).
Following the analogy with disordered quantum many-body systems, each matrix index can be thought of as a site of the reference Hilbert space, which is connected to every other site with the transition rates distributed according to the Gaussian law. Different phenomenologies are expected depending on the value of the parameter $\gamma$, which renders one of the two matrices $A$ or $B$ subleading with respect to the other in the limit of large $N$. As summarized in~\cref{fig:scheme}, the model features three distinct phases (and two transition points between them): a fully delocalized phase for $\gamma<1$, a fully Anderson localized phase for $\gamma>2$, and an intermediate delocalized but non-ergodic phase for $1<\gamma<2$.
The ergodic transition at $\gamma=1$ manifests itself as a transition for
the average spectral density $\rho(\lambda)$, which reproduces the Wigner semicircle law for $\gamma <1$ in the $N\to \infty$ limit, while it reduces to $\rho(\lambda) = p_a(\lambda)$ if $\gamma>1$
and the same $N\to\infty$ limit is taken. At $\gamma=1$,
$\rho(\lambda)$ interpolates between $p_a(\lambda)$ and the Wigner semicircle as the value of the parameter $\nu$ is increased. This transition becomes sharp (i.e., it occurs at a particular value of $\nu=\nu_c$) provided that $p_a(a)$ has a compact support and vanishes sufficiently fast at its upper edge \cite{Claeys_2018,Krajenbrink_2021}.
The value $\gamma=2$ corresponds instead to a genuine Anderson localization transition.
Indeed, the region $\gamma <1$ is characterized by the Wigner-Dyson statistics, meaning that the eigenvectors of $\cor{H}$ are uniformly delocalized on $N$ sites in the large $N$ limit, and the average level spacing follows the Wigner surmise \cite{Livan_2018}, signaling level repulsion. Conversely, in the region $\gamma > 2$ the eigenvectors are completely localized over $\cor{O}(1)$ sites and the mean level spacing exhibits Poisson statistics.
The intermediate region with $1 < \gamma < 2$ is particularly interesting, because the average spectral density tends to $p_a(\lambda)$, but the \textit{local} level statistics remains of the Wigner-Dyson type. Here the eigenvectors are known to be delocalized over a large number of sites $N^{D_\gamma}$, which represent, however, a vanishing fraction of the total number of sites $N$ in the thermodynamic limit, their fractal dimension being $D_\gamma = 2-\gamma <1$ \cite{Kravtsov_2015}.
The simplest and most intuitive way to understand the spectral properties in the intermediate region is provided by the Fermi golden rule: in the limit in which the off-diagonal matrix $B$ is absent and all the eigenvectors are trivially localized on a single site, one has $\vert \psi_i \rangle =\vert i \rangle$, with corresponding eigenenergies $\lambda_i = a_i$. When the GOE perturbation $\nu N^{-\gamma/2} B$ is turned on, the transition probability per unit time from a state $i$ to another state $j$ can be evaluated perturbatively as
\[
\Gamma_{i \to j} = \frac{2 \pi \rho}{\hbar} \frac{4 \eta}{N} |b_{ij}|^2 \, ,
\]
where $\rho=\rho(\lambda)$ and, for future convenience, we have introduced the combination
\begin{equation}
\eta \equiv N^{1-\gamma} \nu^2/4 \, .
\label{eq:eta}
\end{equation}
Hence, the average escape rate per unit time for a ``particle'' created in site $i$ at time $t=0$ reads
\begin{equation} \label{eq:FGR}
\Gamma = \sum_{j \neq i} \langle \Gamma_{i \to j} \rangle = \frac{2 \pi \rho}{\hbar} 4 \eta \, .
\end{equation}
The quantity $\hbar \Gamma$ can thus be interpreted as the bandwidth $\Delta E$ that can be reached in a time of $\cor{O}(1)$ from a given site $i$. This implies that the eigenvectors within this energy window are hybridized by the GOE perturbation. For $1<\gamma<2$ such energy band grows with the system size as $\Delta E \propto N^{1 - \gamma}$ and is much larger than the mean level spacing
\begin{equation}
\delta_N \simeq [N \rho]^{-1} \, ,
\label{eq:mean_level_spacing}
\end{equation}
entailing that the system is not Anderson localized; still, $\Delta E$ remains much smaller than the total bandwidth, which is of $\cor{O}(1)$, signifying that the system is non-ergodic. The Anderson localization transition thus occurs when $\Delta E$ becomes smaller than the mean level spacing, i.e., for $\gamma \ge 2$: this implies that the average escape time from site $i$ grows with $N$, and thus the eigenfunctions remain localized on $\cor{O}(1)$ sites. Conversely, the transition to the fully ergodic phase takes place when $\Delta E$ becomes of the order of the total bandwidth, i.e., for $\gamma \le 1$: this implies that, starting from site $i$, the wave-packet can reach any other site in a time of $\cor{O}(1)$.
In the intermediate phase, $1 < \gamma < 2$, the support set of the eigenvectors (i.e., the number of sites which are hybridized by the perturbation) is simply given by the spreading of the energy interval divided by the average gap between adjacent energy levels, and thus scales as $\Delta E / \delta_N \sim N^{D_\gamma}$, with $D_\gamma = 2 - \gamma$.
The partially extended but non-ergodic eigenstates are therefore linear combinations of a bunch of $N^{D_\gamma}$ localized states associated to nearby energy levels, i.e.,
\[
\vert \psi_i \rangle \approx
\!\! \sum_{\substack{i^\prime \; s.t.\\
|a_i - a_i^\prime| \le \Delta E}} \!\!
c_{i^\prime} \, \vert i^\prime \rangle \, ,
\]
with coefficients $c_{i^\prime}$ of order $N^{-D_\gamma/2}$ to ensure normalization. These eigenstates give
rise to the so-called \emph{mini-bands} in the local spectrum~\cite{Kravtsov_2015}. The width of the mini-bands sets the energy scale $E_T \sim \Delta E \sim N^{D_\gamma - 1} = N^{1 - \gamma}$, often called the Thouless energy~\cite{altshuler1986repulsion,cuevas2007two}, within which GOE-like spectral correlations (and in particular level repulsion) have been established. All the moments of the wave-functions' coefficients (the so-called generalized \textit{inverse participation ratios}, IPR) behave as $I_q = \sum_i \vert \langle i \vert \psi \rangle \vert^{2 q} \propto N^{D_\gamma (1-q)}$: this implies that all the fractal dimensions $D_q$ are degenerate and equal to $D_\gamma$ for all $q$, i.e., that the intermediate phase of the GRP model is fractal but not multifractal.
As discussed above, the emergence of such non-ergodic extended phase is particularly relevant in many physical contexts. Its existence was first suggested in \ccite{Kravtsov_2015} and then rigorously proven in \ccite{vonSoosten_2019}. In recent years several generalizations of the RP model have been put forward and analysed~\cite{kravtsov2020localization,khaymovich2020fragile,monthus2017multifractality,biroli2021levy,buijsman2022circular,khaymovich2021dynamical}, and many other random matrix ensembles have been shown to have an intermediate partially delocalized but non ergodic phase with similar spectral properties~\cite{sarkar2021mobility,roy2018multifractality,wang2016phase,nosov2019correlation,duthie2022anomalous,kutlin2021emergent,motamarri2021localization,tang2022non,tarzia2022fully}. Yet, the GRP setting is still a very useful playground to analyze the properties of fractal states in
a controlled framework.
\subsection{Outline of this work and summary of the main results}
\label{subsec:outline}
As explained above, the GRP model has been intensively investigated over the past few years with a great variety of analytical and numerical techniques.
In this paper we tackle this model by applying yet another approach, namely the replica formalism~\cite{Mezard_1987}, which allows us to obtain new results on the average spectral density, and the statistics of the number of energy levels within a finite interval.
In \cref{par:spectral_replica_approach} we start by analyzing the {average density of states}
$\rho(\lambda)$.
When the size $N$ of the matrix is large, we find the leading order estimate
\begin{equation}
\rho(\lambda) = -\frac{1}{\pi \eta} \lim_{\varepsilon \to 0^+} \Re C(\lambda_\varepsilon) +\order{1/N} \, ,
\label{eq:rho_N_intro}
\end{equation}
where $\lambda_\varepsilon=\lambda-i\varepsilon$, the parameter
$\eta$ depends on $N$ and was introduced in \cref{eq:eta}, and the function $C(\lambda)$ is implicitly defined by the self-consistency equation
\begin{equation}
C(\lambda) = i \eta \, \cor{G}_a\left[\lambda+2iC(\lambda)\right] \, .
\label{eq:C(l)_intro}
\end{equation}
Here $\cor{G}_a$ is the resolvent associated to the distribution $p_a(a)$ of the entries of $A$ (c.f. \cref{eq:resolvent}). In general, it is quite difficult to solve explicitly \cref{eq:C(l)_intro} for $C(\lambda)$ for any arbitrary distribution $p_a(a)$. However, we show in \cref{par:exactly_solvable} that it can be solved explicitly for two special cases: (i) when $p_a(a)$ is a Wigner semicircle -- which is expected, since it is stable under free-convolution \cite{voiculescu1992free} -- and (ii) when it is a Cauchy distribution, which is more surprising.
As we show in \cref{par:resolvent}, by taking the limit $N\to \infty$ with $\eta$ kept finite, \cref{eq:rho_N_intro,eq:C(l)_intro} reduce to the free addition formula \cite{voiculescu1992free}, sometimes called the ``Zee formula'' in the physics literature~\cite{Zee_1996}. However, as we show in \cref{par:approximate_solutions}, these equations contain more information and allow, in the case where $\eta=\eta(N) \ll 1$ defined in \cref{eq:eta} with $1<\gamma<2$, to obtain the leading $1/N$ corrections to the limiting density $\rho(\lambda) = p_a(\lambda)$ in a controlled way, even when the resolvent $\cor{G}_a(z)$ does not admit a closed-form analytic expression. These corrections turn out to be quite difficult to compute using the standard free addition formula, which in principle holds only in the limit $N \to \infty$.
Next, in \cref{par:level_compress} we analyze the behavior of the \textit{level compressibility} $\chi(E)$, which is a simple indicator of the degree of level repulsion and is defined as follows\footnote{It is known in statistics under the name of ``Fano factor''~\cite{fano1947ionization}, and it has been studied recently in physics in the context of extremes and record statistics of time series~\cite{majumdar2019exactly,majumdar2021universal}.}. We first introduce the empirical density of the (real) eigenvalues $\lambda_i$ of $\cal{H}$, defined as
\begin{equation} \label{eq:def_density}
\rho_N(\lambda) = \frac{1}{N} \sum_{i=1}^N \delta(\lambda - \lambda_i) \, .
\end{equation}
Note that $\rho_N(\lambda)$ is normalised to unity. Let
\begin{equation}
I_N[\alpha,\beta] \equiv N \int_\alpha^\beta \dd{\lambda} \rho_N(\lambda)
\label{eq:levels_number}
\end{equation}
denote the number of eigenvalues lying in the interval
$ [\alpha,\beta]\subseteq \mathbb{R}$, which is a random variable. Then
\begin{equation}
\chi(E) \equiv \frac{\kappa_2(E)}{\kappa_1(E)} = \frac{\expval{I_N^2[-E,E]}-\expval{I_N[-E,E]}^2}{\expval{I_N[-E,E]}} \, ,
\label{eq:level_compress}
\end{equation}
where $\kappa_1$ and $\kappa_2$ are the first two cumulants of $I_N$. For Poisson statistics, one has $\kappa_2(E) \simeq \kappa_1(E)$ for small $E$, and then $\chi(E)\simeq 1$ (see \cref{app:iid}). On the contrary, for a rigid spectrum like that of the GOE matrix $B$ in \cref{eq:hamiltonian}, the mean number of eigenvalues behaves as $\expval{I_N[-E,E]} \propto \tilde E$, with $\tilde E \equiv N \rho_N(0)\, E$ and where $[N \rho_N(0)]^{-1}$ is the mean level spacing close to $E=0$, while $\expval{I_N^2[-E,E]}_c \propto \ln \tilde E$ for large $\tilde E$. Hence in the GOE case one finds $\chi(E) \to 0$ for $\tilde E \to \infty$, i.e., for $E \gg [N \rho_N(0)]^{-1}$ (but still much smaller than $E\sim\order{1}$).
In \cref{par:symmetric_interval} we
provide the cumulant generating function of the variable $I_N[\alpha,\beta]$ at leading order for large $N$.
For a symmetric interval $[-E,E]$ and a symmetric distribution $p_a(a)$, the result reads
\begin{equation}
\cor{F}_{[-E,E]}(s) \equiv \frac{1}{N} \ln \expval{e^{-s I_N[-E,E]}} = -m s + \ln \expval{e^{-s f(a) }}_a + \order{\eta/N} \, ,
\label{eq:intro_F}
\end{equation}
where
\begin{equation}
m= - \frac{2\eta}{\pi} \Im \left[ \cor{G}_a\left(-i \Delta^{-1} \right) \right]^2 \, , \;\;\;\;\;
f(a) \equiv \frac{1}{\pi} \arctan \left( \frac{ \sin 2\theta}{a^2r^2 + \cos 2\theta} \right) \in [0,1] \, ,
\end{equation}
and $\Delta(E) = r(E) e^{i\theta(E)}$ has to be determined by solving the self-consistency equation
\begin{equation}
\Delta^{-1}= \varepsilon-iE -2i\eta \cor{G}_a\left(-i \Delta^{-1} \right) \, .
\end{equation}
before sending $\varepsilon \to 0$.
Again, we show that a closed-form solution can be found in some particular cases. Our results are supported by the comparison with the numerical diagonalization of large random matrices.
The ratio of the first two cumulants of $I_N[-E,E]$ gives the level compressibility $\chi(E)$ introduced in \cref{eq:level_compress}. In agreement with the non-ergodic picture presented above for the region $1<\gamma<2$, and having identified $E_T \propto \eta \propto N^{1-\gamma}$ as the Thouless energy of the system, we explicitly verify in \cref{par:scaling} that $\chi(E)\sim 0$ for $E\ll E_T$ -- but still much larger than the mean level spacing $\delta_N \propto N^{-1}$ -- corresponding to level repulsion, while $\chi(E)$ follows the Poisson statistics for $E\gg E_T$. Finally, in the scaling limit $E = 2\pi p_a(0) \eta \cdot y $ with $N^{-1} \ll \eta \ll 1$, we show that $\chi(E)$ takes the universal form
\begin{equation}
\chi(E) \simeq \chi_T \left( y= \frac{E}{2\pi p_a(0) \eta} \right) \,, \quad\quad \chi_T(y) = \frac{1}{\pi y}\left[2y \arctan(y) -\ln (1+y^2) \right] \, ,
\label{eq:intro_scaling_comp}
\end{equation}
where the scaling function $\chi_T(y)$ is
independent of the specific choice of $p_a(a)$. This function is plotted in \cref{fig:scaling_chi}: it behaves as $\chi_T(y) \simeq y/\pi$ for small $y$, while it tends to $1$ for large $y$.
Note that the crossover energy scale $2 \pi p_a(0) \eta$ coincides (apart from a factor of $4$) with the width of the mini-bands identified in Eq.~\eqref{eq:FGR} using the Fermi golden rule, which allows us to put the intuitive arguments given above on the structure of spectral correlations on a much firmer basis.
In \cref{par:low_energy} we finally inspect, using extensive numerical diagonalization of large random matrices, the low-energy region where $E$ is chosen on the scale of the Thouless energy.
The scaling form of $\chi(E)$ presented in \cref{eq:intro_scaling_comp} is thus shown to represent a universal crossover between the classical GOE result $\chi(E \sim N^{-1} )\simeq \chi_\T{GOE}(E)$ for energies of the order of the mean level spacing (and much smaller than the Thouless energy, see \cref{chi_GOE_final}), and the model-dependent prediction of \cref{eq:level_compress,eq:intro_F}, valid for energies of the order of the total spectral band-width, i.e., $E\sim \order{1}$.
\section{Average spectral density}
\label{par:spectral_replica_approach}
Let us begin by considering the density of states of the matrix $\cor{H}$ in \cref{eq:def_density}.
Its mean value can be obtained by means of the Edwards-Jones (E-J) formula \cite{Edwards_1976,Livan_2018}, which we briefly recall here. One starts from the Plemelj-Sokhotski formula: if $f(x)$ is a complex-valued function which is continuous on the real axis, and given $\alpha <0<\beta$, then
\begin{equation}
\lim_{\varepsilon \to 0^+} \int_{\alpha}^\beta \dd{x} \frac{f(x)}{x\pm i \varepsilon} = \mp i \pi f(0) + \pv{\int_{\alpha}^\beta \dd{x} \, \frac{f(x)}{x}} \, ,
\label{eq:plemelj}
\end{equation}
where $\pv$ indicates the Cauchy principal value of the integral. From \cref{eq:def_density} we then have
\begin{equation}
\expval{\rho_N(\lambda)} = \frac{1}{\pi N} \lim_{\varepsilon \to 0^+} \Im \expval{ \sum_{i=1}^N \frac{1}{\lambda-\lambda_i-i \varepsilon} } = \frac{1}{\pi N} \lim_{\varepsilon \to 0^+} \Im \pdv{\lambda} \expval{ \sum_{i=1}^N \ln(\lambda_i -\lambda +i\varepsilon) } \, ,
\end{equation}
where the average is taken over the distribution of the entries of $\cor{H}$, and where in the last step we indicated by $\ln(z)$ the principal branch of the complex logarithm. Using the properties of Gaussian integrals, we finally obtain the E-J formula~\cite{Edwards_1976}
\begin{align}
\rho(\lambda) &\equiv \expval{\rho_N(\lambda)} = -\frac{2}{\pi N} \lim_{\varepsilon \to 0^+} \Im \dv{\lambda} \expval{\ln \cor{Z}(\lambda_\varepsilon) } \, ,
\label{eq:edwards_jones} \\
\cor{Z}(\lambda) &\equiv \det ( \cor{H}- \lambda \mathbb{1} )^{-1/2} = (2\pi i)^{-N/2}\int_{\mathbb{R}^N} \dd{\vb{r}} e^{-\frac{i}{2} \vb{r}^T (\lambda \mathbb{1} -\cor{H} ) \vb{r} } \, , \label{eq:partition_function}
\end{align}
where $\lambda_\varepsilon \equiv \lambda -i \varepsilon$ with $\varepsilon>0$. Note that the negative imaginary part of $\lambda_\varepsilon$ ensures the convergence of the integral in \cref{eq:partition_function}.
The expectation value of the logarithm can then be handled by using the replica trick \cite{Mezard_1987}
\begin{equation}
\expval{\ln \cor{Z}(\lambda) } = \lim_{n\to 0} \frac{1}{n} \ln \expval{ \cor{Z}^n(\lambda) } \, .
\label{eq:replica_trick}
\end{equation}
Computing the last expectation value by standard methods (see \cref{app:replica}) leads to
\begin{equation}
\expval{ \cor{Z}^n(\lambda) } \propto \int \cor{D}\mu \, \cor{D}\hat \mu \, e^{N \cor{S}_n[\mu , \hat \mu; \lambda] } \, ,
\label{eq:average_Zn}
\end{equation}
where the proportionality holds up to an irrelevant numerical constant, and where we introduced the action
\begin{align}
\cor{S}_n[\mu , \hat \mu; \lambda] \equiv& - i \int \dd{\vec{y}} \mu(\vec{y}) \hat \mu(\vec{y}) - \frac{\eta}{2} \int \dd{\vec{y}} \dd{\vec{w}} \mu(\vec{y}) \mu(\vec{w}) \left( \vec{y} \cdot \vec{w} \right)^2 \nonumber\\
&+ \ln \int \dd{\vec{y}} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) |\vec{y}|^2+i \hat \mu(\vec{y}) ] \, .
\label{eq:action}
\end{align}
The parameter $\eta$ is defined in \cref{eq:eta},
while $\vec{y},\vec{w}$ are $n$-dimensional vectors (one component for each of the replicas).
The strategy to obtain the finite-$N$ averaged $\rho(\lambda)$ is the following:
\begin{enumerate}
\item For large $N$, we look for a saddle-point estimate of the path integral in \cref{eq:average_Zn} in the form
\begin{equation}
\expval{ \cor{Z}^n(\lambda) } \propto e^{N \cor{S}_n[\mu^* , \hat \mu^*; \lambda] +\order{1} } \, ,
\end{equation}
where the proportionality holds up to a $\lambda$-independent (even though possibly $N$-dependent) prefactor.
\item Using \cref{eq:edwards_jones}, we recover the spectral density via
\begin{align}
&\rho(\lambda) \simeq -\frac{2}{\pi} \lim_{\varepsilon \to 0^+} \Im \lim_{n\to 0} \frac{1}{n} \dv{\lambda} \cor{S}_n[\mu^* , \hat \mu^*; \lambda_\varepsilon] \label{eq:strategy_level} \\
& \qquad\, = \frac{1}{\pi} \lim_{\varepsilon \to 0^+} \Im{ \lim_{n\to 0} \frac{i}{n} \frac{ \int \dd{\vec{y}} |\vec{y}|^2 \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda_\varepsilon -a) |\vec{y}|^2 +i \hat \mu^*(\vec{y}) ]}{ \int \dd{\vec{y}} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda_\varepsilon -a) |\vec{y}|^2 +i \hat \mu^*(\vec{y}) ]}} \, . \nonumber
\end{align}
Indeed, only the third term in the action of \cref{eq:action} will contribute, because the dependence on $\lambda$ in the first two terms is only implicit,
\begin{equation}
\dv{\lambda} \cor{S}_n[\mu , \hat \mu; \lambda_\varepsilon] = \partial_\lambda \cor{S}_n + \int \dd{\vec{y}} \left[ \fdv{ \cor{S}_n}{\mu(\vec{y})} \dv{\mu(\vec{y})}{\lambda} +\fdv{ \cor{S}_n}{\hat\mu(\vec{y})} \dv{\hat\mu(\vec{y})}{\lambda} \right] \, ,
\end{equation}
and the term under the integral vanishes at the saddle-point (where the action is stationary by construction).
In turn this implies that, to compute $\rho(\lambda)$ from Eq. \eqref{eq:strategy_level}, we do not need to determine $\mu^*(\vec{y})$, but only $\hat \mu^*(\vec{y})$. Finally, by the $\simeq$ symbol in \cref{eq:strategy_level} we mean that the corrections are \textit{at most}
$\order{1/N}$.
\end{enumerate}
\subsection{Saddle-point equations and rotationally-invariant Ansatz}
\label{par:saddle_point_density}
The leading contribution to \cref{eq:average_Zn} for large $N$ can be found by minimizing the action in \cref{eq:action}. Omitting (to ease the notations) the superscript $*$ from $\mu^*(\vec{y})$ and $\hat \mu^*(\vec{y})$, and understanding the dependencies on $\lambda$ as computed in correspondence of $\lambda_\varepsilon$, the saddle-point equations read
\begin{align}
0 &\equiv \fdv{\cor{S}_n}{\mu(\vec{x})} = - i \hat \mu(\vec{x}) - \eta \int \dd{\vec{w}} \mu(\vec{w}) \left( \vec{x} \cdot \vec{w} \right)^2 \, , \label{eq:saddle_mu} \\
0 &\equiv \fdv{\cor{S}_n}{\hat \mu(\vec{x})} = - i \mu(\vec{x})+i \frac{ \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) |\vec{x}|^2 +i \hat \mu(\vec{x}) ] }{\int \dd{\vec{y}} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) |\vec{y}|^2 +i \hat \mu(\vec{y}) ]} \, . \label{eq:saddle_muhat}
\end{align}
Substituting the expression for $\mu(\vec{x})$ obtained from the second equation \eqref{eq:saddle_muhat} in the first one \eqref{eq:saddle_mu}, one obtains a closed equation for $\hat \mu(\vec{x})$ which reads
\begin{equation}
\hat \mu(\vec{x}) = i\eta \frac{ \int \dd{\vec{y}} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) |\vec{y}|^2 +i \hat \mu(\vec{y}) ] \left( \vec{x} \cdot \vec{y} \right)^2 }{\int \dd{\vec{y}} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) |\vec{y}|^2 +i \hat \mu(\vec{y}) ]} \, .
\label{eq:saddle_tot}
\end{equation}
To make progress, we plug in the Ansatz $\hat \mu(\vec{x})=\hat \mu(x)$, with $x\equiv |\vec{x}|$,
which is rotationally symmetric in the space of replicas (i.e., it is invariant under $O(n)$ transformations). Note that requiring invariance under $O(n)$ is a stronger request than the mere replica-symmetry (RS): indeed, the exchange between any pair of components of $\vec{x}$ can be obtained by means of a $O(n)$ transformation. Stepping to spherical coordinates and using the identity
\begin{equation}
\int \dd{\Omega_n} \left( \vec{x} \cdot \vec{y} \right)^2 = \frac{(xy)^2}{n} \int \dd{\Omega_n} \, ,
\end{equation}
where $\dd{\Omega_n}$ is the differential of the $n$-dimensional solid angle, we find
\begin{equation}
\hat \mu(x) = i\frac{\eta }{n} \frac{ \int_0^\infty \dd{y} y^{n-1} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) y^2 +i \hat \mu(y) ] \left( xy \right)^2 }{\int_0^\infty \dd{y} y^{n-1} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) y^2 +i \hat \mu(y) ] } \, .
\label{eq:replica_symmetrization}
\end{equation}
Let us now introduce the auxiliary function
\begin{equation}
G(y;\lambda) \equiv \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda -a) y^2 +i \hat \mu(y) ] = \exp[-\frac{i}{2} \lambda y^2 +i \hat \mu(y) ] \psi_a(-y^2/2) \, ,
\label{eq:G(y)}
\end{equation}
where in the second equality we recognized the characteristic function of $p_a(a)$ (i.e., its Fourier transform), namely
\begin{equation}
\psi_a(k) = \int \dd{a} p_a (a) e^{-ika} \, .
\label{eq:characteristic}
\end{equation}
Equation \eqref{eq:replica_symmetrization} can readily be expressed in terms of $G(y;\lambda)$. We then integrate by parts in the denominator of \cref{eq:replica_symmetrization}, finding that boundary terms disappear at least as long as $\varepsilon >0$ in $G(y;\lambda_\varepsilon)$ -- we will check \textit{a posteriori} that the presence of $\hat \mu(y)$ does not spoil the convergence. We thus get
\begin{equation}
\hat \mu(x) = -i\eta x^2 \frac{ \int_0^\infty \dd{y} y^{n+1} G(y;\lambda) }{\int_0^\infty \dd{y} y^{n} G'(y;\lambda) } \, ,
\end{equation}
where $G'(y;\lambda) = \partial_y G(y;\lambda)$. Under this form, we can now take the limit $n\to 0$, which yields
\begin{equation}
i \hat \mu(x) = x^2 \eta \frac{ \int_0^\infty \dd{y} y \, G(y;\lambda) }{\int_0^\infty \dd{y} G'(y;\lambda) } \equiv x^2 C(\lambda) \, .
\label{eq:C(l)}
\end{equation}
The function $C(\lambda)$ must now be determined self-consistently via
\begin{equation}
C(\lambda) =\eta \frac{ \int_0^\infty \dd{y} y \, G(y;\lambda) }{\int_0^\infty \dd{y} G'(y;\lambda) } \, ,
\label{eq:C(l)_defining}
\end{equation}
where $G(y;\lambda)$ can be read off from \cref{eq:G(y)} upon setting $i\hat \mu(y)=y^2 C(\lambda)$.
\subsection{General result}
We now repeat the same steps at the level of the saddle-point action in \cref{eq:strategy_level}: we plug in the rotationally symmetric Ansatz, we integrate by parts in the denominator and we finally take the limit $n\to 0$. This gives
\begin{equation}
\rho(\lambda) = -\frac{1}{\pi} \lim_{\varepsilon \to 0^+} \Re \frac{ \int_0^\infty \dd{y}y G(y;\lambda_\varepsilon)}{ \int_0^\infty \dd{y} G'(y;\lambda_\varepsilon)} +\order{1/N} = -\frac{1}{\pi \eta} \lim_{\varepsilon \to 0^+} \Re C(\lambda_\varepsilon) +\order{1/N}
\, .
\label{eq:rho_general}
\end{equation}
We then go back to \cref{eq:C(l)_defining}, which contains an exact differential in its denominator. If $\Re C(\lambda_\varepsilon) \leq 0$, then $G(y\to \infty;\lambda_\varepsilon)=0$ and $G(0;\lambda_\varepsilon)=1$, and we obtain a self-consistency equation for $C(\lambda_\varepsilon)$,
\begin{equation}
C(\lambda)= -\eta \int_0^\infty \dd{z} \psi_a(-z) \, e^{-i\lambda z +2z C(\lambda)} \, ,
\label{eq:C(l)_self_consistent}
\end{equation}
where we recall that $\psi_a(x)$ is the characteristic function associated with $p_a(a)$ (see \cref{eq:characteristic}). This determines $C(\lambda)$ implicitly. Equations~\eqref{eq:rho_general} and \eqref{eq:C(l)_self_consistent} represent the main result of this section.
\subsubsection{Limiting cases}
Here we briefly check the consistency of the result we just obtained in the limiting cases in which only one of the two matrices $A$ or $B$ in \cref{eq:hamiltonian} are retained.
First, when $\nu=\eta=0$ we find $\cor{H}=A$, and the level statistics can only be determined by $p_a (a)$. Indeed, from \cref{eq:saddle_mu} we read $\hat \mu(\vec{x})=0$, and even without invoking a particular Ansatz we get from \cref{eq:strategy_level}, upon sending $N \to \infty$,
\begin{equation}
\rho(\lambda)
= -\frac{1}{\pi} \lim_{\varepsilon \to 0^+} \Re \frac{ \int_0^\infty \dd{y}y G(y;\lambda_\varepsilon)}{ \int_0^\infty \dd{y} G'(y;\lambda_\varepsilon)} \, ,
\end{equation}
where $G(y;\lambda)$ is given by \cref{eq:G(y)} with $\hat \mu=0$. For any $\varepsilon >0$ the function $G(y;\lambda_\varepsilon)$ is well behaved at $y\to+\infty$, so that
\begin{equation}
\rho(\lambda)
= \frac{1}{\pi} \lim_{\varepsilon \to 0^+} \Re \int_0^\infty \dd{y}y G(y;\lambda_\varepsilon) = \frac{1}{\pi} \lim_{\varepsilon \to 0^+} \Re \int_0^\infty \dd{z} \psi_a(-z) e^{-i\lambda_\varepsilon z} = p_a (\lambda) \, ,
\end{equation}
where we changed variable to $z=y^2/2$ in the first step, and in the last one we used $\psi_a^*(-z)=\psi_a(z)$.
Conversely, the GOE case is formally recovered by setting $\gamma=1$ and $a_i=0$ identically.\footnote{The way we constructed the action in \cref{eq:action} does not allow us to explore the region $\gamma<1$ where $A$ becomes subleading for large $N$ (although this could of course be achieved with minor modifications of the calculation presented in \cref{app:replica}).} With this choice, the limiting spectrum of ${\mathcal H}$ is supported within $\lambda \in [-\sqrt{2}\nu,\sqrt{2}\nu ]$, and we set for simplicity $\nu=1$ (corresponding to $\eta=1/4$). Using \cref{eq:C(l)_defining} (without integrating by parts in the denominator, but really computing $G'(y;\lambda)$) one easily finds
\begin{equation}
C(\lambda) = \frac14 \left( i\lambda \pm \sqrt{2-\lambda^2} \right) \, .
\end{equation}
We choose the branch with the minus sign, as we assumed above that $\Re C(\lambda_\varepsilon)\leq 0$. Using \cref{eq:rho_general} then immediately renders, in the limit $N\to \infty$, the semicircle law
\begin{equation}
\rho(\lambda) \to \rho_\T{GOE}(\lambda) = \frac{\sqrt{2-\lambda^2}}{\pi}\, \Theta(\sqrt{2}-|\lambda|) \, ,
\label{eq:rho_GOE}
\end{equation}
where $\Theta(x)$ is the Heaviside step function, with $\Theta(x) = 1$ if $x>0$ and $\Theta(x) = 0$ if $x \leq 0$.
\subsection{Resolvent formulation and connection with the Zee formula}
\label{par:resolvent}
One can observe that, under the same convergence hypotheses as above, the self-consistency equation \eqref{eq:C(l)_self_consistent} which determines $C(\lambda)$ can be rewritten as
\begin{equation}
C(\lambda) = i \eta \, \cor{G}_a\left[\lambda+2iC(\lambda)\right] \, ,
\label{eq:self_consistence_resolvent}
\end{equation}
where
\begin{equation}
\cor{G}_a(z) = \int \dd{a} \frac{p_a(a)}{z-a}
\label{eq:resolvent}
\end{equation}
is the resolvent (or Cauchy-Stieltjes transform) of the distribution $p_a(a)$. The resolvent can be inverted to give back
\begin{equation}
p_a(x) = \frac{1}{\pi} \lim_{\varepsilon \to 0^+} \Im \cor{G}_a (x-i\varepsilon) \, ,
\label{eq:inversion_resolvent}
\end{equation}
which can be easily proved by using the Plemelj-Sokhotski formula in \cref{eq:plemelj}.
By comparing \cref{eq:inversion_resolvent} with \cref{eq:rho_general}, one immediately realizes that our function $C(\lambda)$ is nothing but
\begin{equation}
C(\lambda) = i \eta \, \cor{G}(\lambda) + \order{1/N} \, ,
\label{eq:connection_zee}
\end{equation}
where we denoted by $\cor{G}(\lambda)$ the resolvent of the spectral density $\rho(\lambda)$ of our model. Choosing $\gamma=1$ (so that $\eta$ in \cref{eq:eta} becomes $N$-independent) and taking the limit $N\to \infty$, the correction in \cref{eq:connection_zee} vanishes, and \cref{eq:self_consistence_resolvent} takes the form
\begin{equation}
\cor{G}(\lambda_\varepsilon) = \cor{G}_a (\lambda_\varepsilon -2\eta \cor{G}(\lambda_\varepsilon)) \, .
\label{eq:resolvent_ledoussal}
\end{equation}
This is analogous to Eq.~(148) in Ref. \cite{Krajenbrink_2021}, which was derived in the case in which the matrix $B$ belongs to the GUE (and not the GOE) ensemble, and it is consistent with the results of \ccite{Bouchaud_Les_Houches}. Moreover, in \cref{app:zee} we show how \cref{eq:resolvent_ledoussal} can be recovered by direct application of the Zee formula for the addition of two random matrices derived in \ccite{Zee_1996}.
One may legitimately wonder, at this point, whether our calculation is solely another way of obtaining Zee's result \cite{Zee_1996} for the particular case in which one of the two matrices being summed is a GOE matrix.
In \cref{par:exactly_solvable} we will analyze a few cases in which the self-consistency equation \eqref{eq:C(l)_self_consistent} (or \eqref{eq:self_consistence_resolvent}) can be solved exactly, so that from \cref{eq:rho_general} one can find an expression for $\rho(\lambda)$ which is correct, up to $\order{1/N}$, for any value of $\eta$. These cases are in fact the very same which could be cracked by applying the Zee formula to the deformed GOE matrix.
The advantage of our framework is that, whenever $\eta$ is $N$-dependent and decays slower than $1/N$ (as it happens, for instance, in \cref{eq:eta} for $1<\gamma<2$), then we are still able to keep track of the finite-$N$ corrections. Moreover, in \cref{par:approximate_solutions} we will provide approximate solutions which can be used whenever the resolvent $\cor{G}_a$ is not available in closed form, so that the Zee route is not viable.
\subsection{Exactly solvable cases}
\label{par:exactly_solvable}
In some particular cases, the self-consistency equation \eqref{eq:C(l)_self_consistent} admits an analytic solution, and we can access the limiting distribution $\rho(\lambda)$ for any value of $\eta$ (i.e., not necessarily small). This happens whenever the following conditions are met:
\begin{enumerate}[(i)]
\item{the resolvent $\cor{G}_a(z)$ associated with $p_a(a)$ is known analytically, and}
\item{the self-consistency equation for $C(\lambda)$ resulting from \cref{eq:C(l)_self_consistent} or \cref{eq:self_consistence_resolvent} is not transcendental, so that we can solve for $C(\lambda)$.}
\end{enumerate}
Below we present two such examples, which will also prove useful in our discussion of the level compressibility presented later in \cref{par:comp_exactly_solvable}.
\begin{figure}[ht]
\centering
\subfloat[]{
\includegraphics[width=0.47\textwidth]{cauchy_numeric.pdf}
\label{fig:cauchy_numeric}
}
\subfloat[]{
\includegraphics[width=0.47\textwidth]{cauchy_numeric_large_eta.pdf}
\label{fig:cauchy_numeric_large_eta}
}
\caption{Distribution of the eigenvalues $\rho(\lambda)$ in the case in which $p_a(a)$ is the Cauchy distribution (see \cref{par:cauchy}). We show a numerical check of \cref{eq:density_cauchy} in the regime of (a) small $\eta$ (here $\eta=0.12$) and (b) large $\eta$ (here $\eta=11.7$). The histogram was built using $\mu=0$, $\omega=1$, $\gamma=1.1$ and $N=2000$, with $\nu=1$ in (a) and $\nu=10$ in (b).}
\label{fig:cauchy}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{cauchy.pdf}
\caption{Limiting distribution of the rescaled eigenvalues $\kappa=\lambda/\sqrt{4\eta}$, with $\rho(\lambda)$ given in \cref{eq:density_cauchy} for the Cauchy case (see \cref{par:cauchy}). In this plot we used $\mu=0$ and $\omega=0.5$.}
\label{fig:cauchy_large_eta}
\end{figure}
\subsubsection{Cauchy distributed $a_i$}
\label{par:cauchy}
Let us choose $p_a(a)$ to be a Lorentzian of width $\omega$ and centered at $\mu$,
\begin{equation}
p_a(a) = \frac{1}{\pi \omega}\left[\frac{\omega^2}{(a-\mu)^2+\omega^2} \right] \, .
\label{eq:lorentzian}
\end{equation}
Its characteristic function is an exponential, $\psi_a(z) = \exp(-i\mu z -\omega |z|)$, and then by using \cref{eq:C(l)_self_consistent} we can compute $C(\lambda)$ in closed form:
\begin{equation}
C(\lambda) = \frac14 \left\lbrace \omega +i(\lambda-\mu) \pm \sqrt{\left[\omega +i(\lambda-\mu)\right]^2+8\eta } \right\rbrace \, .
\label{eq:C(l)_cauchy}
\end{equation}
We choose the branch with the minus sign for which $\Re C(\lambda)\leq \omega/2$. Using \cref{eq:rho_general} then yields
\begin{equation}
\rho(\lambda) = -\frac{\omega -\left\lbrace 4\omega^2 (\lambda-\mu)^2 + \left[ 8\eta +\omega^2 -(\lambda-\mu)^2 \right]^2 \right\rbrace^{1/4}\cos(\theta_\lambda /2)}{4\pi \eta} +\order{1/N} \, ,
\label{eq:density_cauchy}
\end{equation}
with
\begin{equation}
\theta_\lambda \equiv \arg \left\lbrace \left[\omega +i(\lambda-\mu)\right]^2+8\eta \right \rbrace \, .
\end{equation}
In \cref{fig:cauchy_numeric} we plot \cref{eq:density_cauchy} against numerical results in the small-$\eta$ region: we find a good agreement with the theoretical prediction, as well as visible departures from the Cauchy distribution, especially in the bulk.
Another interesting limit is the one of large $\eta$, i.e., $\gamma=1$ and $\nu$ large. It has been shown in \ccite{Krajenbrink_2021} that, whenever $p_a(a)$ is rapidly decaying close to the edge of its finite support, then the spectral density $\rho(\lambda)$ interpolates between $p_a(a)$ and $\rho_\T{GOE}(\lambda)$ as the value of $\nu$ is increased. Note, however, that in the present case $p_a(a)$ decays algebraically and its support is not compact, so the outcome is less clear. The correct way of taking this limit is to rescale the eigenvalues as $\kappa \equiv\lambda/\sqrt{4\eta}$ and look for the distribution $\rho_\kappa(\kappa) = \sqrt{4\eta}\rho(\sqrt{4\eta} \kappa)$. From \cref{eq:C(l)_cauchy} we see that, if $\eta\gg \omega,\mu$, then
\begin{equation}
C(\sqrt{4\eta}\kappa) = \frac{\sqrt{\eta}}{2}\left[i \kappa- \sqrt{2-\kappa^2} \right] +\order{\eta^0,1/N} \, ,
\end{equation}
and for large $N$ we get from \cref{eq:rho_general} that $\rho_\kappa(\kappa) \to \rho_\T{GOE}(\kappa)$ (see \cref{eq:rho_GOE}). For large but finite $\eta$ and $N\to \infty$, on the other hand, the bulk distribution of $\kappa$ looks like a semicircle (as in the GOE ensemble), but with fat tails whose width grows with $\omega$ (see \cref{fig:cauchy_large_eta}). Moreover, the whole distribution shifts rigidly by changing its center $\mu/\sqrt{4\eta}$.
Numerical results in the large-$\eta$ region are again nicely reproduced by \cref{eq:density_cauchy}, as shown in \cref{fig:cauchy_numeric_large_eta}.
Note that one can equivalently get to \cref{eq:C(l)_cauchy} by first computing the resolvent associated with the Lorentzian distribution in \cref{eq:lorentzian}, i.e.,
\begin{equation}
\cor{G}_\T{Cauchy}(\lambda) = \frac{1}{\lambda - \mu \pm i\omega} \, ,
\label{eq:cauchy_resolvent}
\end{equation}
where the $\pm$ branches correspond to $\Im{\lambda}>0$ or $\Im{\lambda}<0$, respectively. This can be obtained by explicitly performing the complex integral in \cref{eq:resolvent}, which only entails simple poles for a Cauchy distribution. One can then easily solve the self-consistency equation \eqref{eq:self_consistence_resolvent}, which turns out to be quadratic, and thus recover \cref{eq:C(l)_cauchy}.
\subsubsection{Wigner distributed $a_i$}
\label{par:wigner}
Another simple case is the one in which $p_a(a)$ itself is chosen as the ($\mu$-centered) Wigner distribution
\begin{equation}
p_a(a) \equiv \frac{2}{\pi \sigma^2} \sqrt{\sigma^2-(a-\mu)^2} \, ,
\label{eq:wigner}
\end{equation}
whose corresponding resolvent is
\begin{equation}
\cor{G}_a(z) = \frac{2}{\sigma^2} \left[z-\mu-\sqrt{(z-\mu)^2-\sigma^2} \right] \, .
\label{eq:wigner_resolvent}
\end{equation}
The self-consistency equation \eqref{eq:self_consistence_resolvent} is again quadratic and it yields
\begin{equation}
C(\lambda) = 2\eta \frac{i(\lambda-\mu) \pm \sqrt{\sigma^2+8\eta-(\lambda-\mu)^2}}{\sigma^2+8\eta} \, ,
\end{equation}
so that by choosing the branch with the minus sign and using \cref{eq:rho_general} we find
\begin{equation}
\rho(\lambda) = \frac{2\sqrt{\sigma^2+8\eta-(\lambda-\mu)^2}}{\pi(\sigma^2+8\eta)} \Theta\left(\sigma^2+8\eta-(\lambda-\mu)^2 \right) + \order{1/N}\, .
\label{eq:wigner_deformed}
\end{equation}
As expected, this is still a Wigner distribution centered in $\lambda=\mu$, but its width gets corrected as $\sigma^2 \to \sigma^2+8\eta$.
\subsection{Approximate solutions}
\label{par:approximate_solutions}
For most choices of the diagonal disorder distribution $p_a(a)$, \cref{eq:C(l)_self_consistent} cannot be solved exactly. However, since $\eta$ is a small parameter for large $N$ and $\gamma>1$ (which is the case we are mostly interested in), it may be sufficient to note that $C(\lambda) \sim \order{\eta} \sim \order{N^{1-\gamma}}$: this can be taken as a starting point for constructing self-consistent approximations for large $N$.
By expanding recursively the exponential in \cref{eq:C(l)_self_consistent} and using that $C(\lambda) \sim \order{\eta}$, one gets
\begin{align}
C(\lambda) = &
-\eta \int_0^\infty \dd{z} \psi_a(-z) e^{-i\lambda z}
\nonumber\\
& + 2\eta^2 \int_0^\infty \dd{z} \int_0^\infty \dd{s} z \psi_a(-z) \psi_a(-s) e^{-i\lambda (z+s)} + \order{\eta^3} \, .
\label{eq:C(l)_largeN}
\end{align}
Note that
\begin{equation}
\Re C(\lambda_\varepsilon) = -\eta \, p_a(\lambda_\varepsilon) +\order{\eta^2} \xrightarrow[\varepsilon \to 0]{N\gg 1} -\eta\, p_a(\lambda) \leq 0 \, ,
\end{equation}
in such a way that $G(y)$ is indeed well-behaved at $y\to +\infty$, as we had assumed in \cref{par:saddle_point_density}.
Plugging \cref{eq:C(l)_largeN} into \cref{eq:rho_general} now gives
\begin{align}
\rho(\lambda) = p_a (\lambda) - \frac{2\eta}{\pi} \Re \int_0^\infty \dd{z} \int_0^\infty \dd{s} z \psi_a(-z) \psi_a(-s) e^{-i\lambda (z+s)} + \order{\eta^2,N^{-1}} \, .
\label{eq:rho_first_order}
\end{align}
Of course this approximation becomes meaningless if $\gamma \geq 2$, where the higher order corrections we neglected when we took the saddle-point approximation mingle with our correction in $\eta$.
Alternatively, one can expand the exponential in \cref{eq:C(l)_self_consistent} as
\begin{equation}
C(\lambda) \simeq -\eta \int_0^\infty \dd{z} \psi_a(-z) e^{-i\lambda z} - 2\eta C(\lambda) \int_0^\infty \dd{z} z \psi_a(-z) e^{-i\lambda} \, ,
\label{eq:resummed}
\end{equation}
corresponding to a partial resummation of the perturbative series in $\eta$: this is the so-called self-consistent Hartree-Fock approximation \cite{parisi_statistical_1988}. From \cref{eq:resummed} we then obtain
\begin{equation}
C(\lambda) \simeq \frac{-\eta \int_0^\infty \dd{z} \psi_a(-z) e^{-i\lambda z}}{1+2\eta \Sigma(\lambda)} \, ,
\label{eq:C(l)_hartree_fock}
\end{equation}
with the \textit{self-energy}
\begin{equation}
\Sigma(\lambda) = \int_0^\infty \dd{z} z \psi_a(-z) e^{-i\lambda z} \, .
\end{equation}
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{gaussian_HF.pdf}
\caption{Distribution of the eigenvalues $\rho(\lambda)$ in the case in which $p_a(a)$ is Gaussian with unit variance, and $B$ is a GOE matrix as in \cref{eq:hamiltonian}. Numerical results are compared to the first $\order{\eta}$ correction given in \cref{eq:gaussian_correction}, and to the Hartree-Fock approximation given in \cref{eq:C(l)_hartree_fock}. The histogram was built using $N=2000$, $\nu=1$ and $\gamma=1.1$, corresponding to $\eta=0.12$.}
\label{fig:gaussian_HF}
\end{figure}
\noindent\textbf{An explicit example: the Gaussian case.} We finally test both approximations in the case where $p_a(a)$ is chosen to be Gaussian (with zero mean and variance $\sigma^2$), as in the original Rosenzweig-Porter model \cite{RP_1960}. By using $\psi_a(z)=\exp(-z^2 \sigma^2/2)$ in \cref{eq:rho_first_order}, one finds
\begin{equation}
\rho(\lambda) = p_a (\lambda) + \sqrt{\frac{2}{\pi}}\frac{\eta}{ \sigma^3} e^{-\lambda^2/(2\sigma^2)}\left[ \frac{2\sqrt{2}\lambda}{\sigma}F\left(\frac{\lambda}{\sqrt{2}\sigma}\right)-1 \right]+ \order{\eta^2,N^{-1}} \, ,
\label{eq:gaussian_correction}
\end{equation}
where $F(z)$ is the Dawson function defined as
$
F(z) \equiv e^{-z^2} \int_0^z e^{y^2} \dd{y} \, .
$
The corrected spectral density is plotted in \cref{fig:gaussian_HF}, and comparison with the eigenvalue spectrum computed numerically shows a fair agreement.
Using instead the Hartree-Fock approximation for $C(\lambda)$ given in \cref{eq:C(l)_hartree_fock} results in a better agreement with the numerical data in the central region of the spectrum.
\section{Number of eigenvalues in an interval and level compressibility}
\label{par:level_compress}
In this Section we consider the number of eigenvalues $I_N[\alpha,\beta]$ in a finite interval $[\alpha,\beta]$, as given by \cref{eq:levels_number}, and compute its cumulant generating function. This, in turn, can be used to access the level compressibility $\chi(E)$ defined in \cref{eq:level_compress}.
As we explained in the Introduction, $\chi(E)$ represents a simple measure of the rigidity of the spectrum, which in turn allows us to distinguish between the phases of the model.
This program can be achieved by following the replica-based procedure introduced and exploited in \ccite{Metz_2016,Metz_2017}, which we briefly outline here. Starting from the definition of the spectral density in \cref{eq:def_density}, we first rewrite \cref{eq:levels_number} as
\begin{equation}
I_N[\alpha,\beta] = \sum_{i=1}^N \left[ \Theta(\beta-\lambda_i) - \Theta(\alpha-\lambda_i) \right] \, .
\end{equation}
Now we recall that the Heaviside function can be represented in terms of the discontinuity of the complex logarithm,
\begin{equation}
\Theta(-x) = \frac{1}{2\pi i}\lim_{\varepsilon \to 0^+} \left[ \ln(x+i\varepsilon) - \ln(x-i\varepsilon) \right] \, ,
\end{equation}
so that we interpret
\begin{equation}
\sum_{i=1}^N \Theta(\alpha-\lambda_i) = \frac{1}{2\pi i}\lim_{\varepsilon \to 0^+} \left[ \ln \det ( \cor{H}- \alpha_\varepsilon \mathbb{1} ) - \ln\det ( \cor{H}- \alpha_\varepsilon^* \mathbb{1} ) \right] \, ,
\end{equation}
where we called as before $\alpha_\varepsilon \equiv \alpha-i\varepsilon$ with $\varepsilon>0$. This allows us to express $I_N[\alpha,\beta]$ in terms of the partition function given in \cref{eq:partition_function},
\begin{equation}
I_N[\alpha,\beta] = -\frac{1}{\pi i}\lim_{\varepsilon \to 0^+} \ln \left[\frac{\cor{Z}(\beta_\varepsilon)\cor{Z}(\alpha_\varepsilon^*) }{\cor{Z}(\beta_\varepsilon^*)\cor{Z}(\alpha_\varepsilon)} \right] \, .
\label{eq:I_N_complex_representation}
\end{equation}
In order to compute the moments of $I_N$, we first address its cumulant generating function
\begin{align}
\cor{F}_{[\alpha,\beta]}(s) \equiv
\frac{1}{N} \ln \expval{e^{-s I_N[\alpha,\beta]}}
=
\frac{1}{N} \ln
\langle
e^{\frac{s}{\pi i} \lim_{\varepsilon \to 0^+}
\ln \{
{\mathcal Z}(\beta_\epsilon)
{\mathcal Z}(\alpha_\epsilon^*)
[{\mathcal Z}(\beta^*_\epsilon) {\mathcal Z}(\alpha_\epsilon)]^{-1}
\}
}
\rangle
\,.
\end{align}
Assuming now that one can move the limit $\varepsilon \to 0^+$ at the front of this expression, we obtain
\begin{align}
\cor{F}_{[\alpha,\beta]}(s)
=
\lim_{\varepsilon \to 0^+} \frac{1}{N} \ln Q_{[\alpha,\beta]}(s) \, ,
\label{eq:cgf}
\end{align}
where we introduced
\begin{equation}
Q_{[\alpha,\beta]}(s) \equiv \expval{\left[ \cor{Z}(\beta_\varepsilon^*)\cor{Z}(\alpha_\varepsilon)\right]^{is/\pi} \left[\cor{Z}(\beta_\varepsilon) \cor{Z}(\alpha_\varepsilon^*)\right]^{-is/\pi} } \, .
\end{equation}
The latter can be accessed by first evaluating
\begin{equation}
Q_{[\alpha,\beta]}(n_\pm) \equiv \expval{\left[ \cor{Z}(\beta_\varepsilon^*)\cor{Z}(\alpha_\varepsilon)\right]^{n_+} \left[\cor{Z}(\beta_\varepsilon) \cor{Z}(\alpha_\varepsilon^*)\right]^{n_-} }
\label{eq:Q_ab}
\end{equation}
within the replica formalism with $n_\pm$ integer, and then performing its analytic continuation to
\begin{equation}
Q_{[\alpha,\beta]}(s) = \lim_{n_\pm \to \pm is/\pi} Q_{[\alpha,\beta]}(n_\pm) \, .
\end{equation}
To obtain the level compressibility in \cref{eq:level_compress}, we finally compute the cumulants
\begin{equation}
\frac{\kappa_j[\alpha,\beta]}{N} = (-1)^{j} \eval{\partial_s^j \cor{F}_{[\alpha,\beta]}(s)}_{s=0} \, ,
\label{eq:cumulants}
\end{equation}
and we evaluate them at $\alpha=-E$, $\beta=E$.
We note that this calculation involves the product of partition functions and not a logarithm of them, and thus there is no question of quenched vs. annealed averages here.
\subsection{Replica action and saddle-point equations}
The details of this derivation are reported in \cref{app:replica_level_compress}. In analogy with \cref{par:spectral_replica_approach}, here we express
\begin{align}
Q_{[\alpha,\beta]}(n_\pm) \propto & \int \cor{D}(i\varphi) \, \exp{N \cor{S}_{n_\pm}[\varphi; \hat \Lambda] } \, , \label{eq:Q_n_varphi} \\
\cor{S}_{n_\pm}[\varphi ;\hat \Lambda] \equiv& \frac{1}{2\eta} \int \dd{\vec{\tau}} \dd{\vec{\tau}'} \varphi(\vec{\tau}) M(\vec{\tau},\vec{\tau}') \varphi(\vec{\tau}')
\nonumber\\
&+ \ln \int \dd{\vec{\tau}} \exp[-\frac{i}{2}\vec{\tau}\, \hat \Lambda \,\vec{\tau} -\int \dd{\vec{\tau}'} M(\vec{\tau},\vec{\tau}') \varphi(\vec{\tau}') ] \psi_a\left(-\frac{1}{2} \vec{\tau} \, \hat L \, \vec{\tau}\right) \, ,
\label{eq:action_comp_varphi}
\end{align}
where the replica vectors $\vec{\tau}\in \mathbb{R}^n$ live in dimension $n=2(n_+ + n_-)$. Each of the four replica sets corresponds to one sector in the block matrices
\begin{equation}
\hat{\Lambda} \equiv \mqty(\dmat{\alpha_\varepsilon \mathbb{1}_{n_+}, \bar{\beta}_\varepsilon \mathbb{1}_{n_+}, \beta_\varepsilon \mathbb{1}_{n_-},\bar{\alpha}_\varepsilon \mathbb{1}_{n_-}}) \, ,
\;\;\;\qquad \hat{L} \equiv \mqty(\dmat{\mathbb{1}_{n_+}, -\mathbb{1}_{n_+}, \mathbb{1}_{n_-},- \mathbb{1}_{n_-}}) \, ,
\label{eq:lambda_L_def}
\end{equation}
where we called $\bar{\alpha}_\varepsilon=-\alpha^*_\varepsilon$, and similarly for $\bar\beta_\epsilon$. Finally, in \cref{eq:action_comp_varphi} we introduced the function
\begin{equation}
M(\vec{\tau},\vec{\tau}') \equiv \left( \vec{\tau} \, \hat L \, \vec{\tau}' \right)^2\, ,
\label{eq:M_def}
\end{equation}
while we recall that $\psi_a(z)$ was given in \eqref{eq:characteristic}. Next, we need to compute the integration over $\cor{D}(i\varphi)$ in \cref{eq:Q_n_varphi} in the limit of large $N$.
From \cref{eq:action_comp_varphi}, the saddle-point equation follows simply as
\begin{equation}
\varphi_0(\vec{\tau}) = \frac{\eta}{Z_\varphi} \exp[-\frac{i}{2}\vec{\tau}\, \hat \Lambda \,\vec{\tau} -\int \dd{\vec{\tau}'} M(\vec{\tau},\vec{\tau}') \varphi_0(\vec{\tau}') ] \psi_a\left(-\frac{1}{2} \vec{\tau} \, \hat L \, \vec{\tau}\right) \, ,
\label{eq:saddle_point_comp}
\end{equation}
with
\begin{equation}
Z_\varphi \equiv \int \dd{\vec{\tau}} \exp[-\frac{i}{2}\vec{\tau}\, \hat \Lambda \,\vec{\tau} -\int \dd{\vec{\tau}'} M(\vec{\tau},\vec{\tau}') \varphi_0(\vec{\tau}') ] \psi_a\left(-\frac{1}{2} \vec{\tau} \, \hat L \, \vec{\tau}\right) \, .
\end{equation}
In order to better quantify the finite-size corrections and to make contact with the calculation performed in \ccite{Metz_2017} for the pure GOE case, in \cref{app:gaussian_fluctuations} we also compute the Gaussian fluctuations around the saddle-point $\varphi_0(\vec{\tau})$, leading to
\begin{equation}
Q_{[\alpha,\beta]}(n_\pm) = \exp{N\cor{S}_{n_\pm}[\varphi_0 ; \hat \Lambda]+\frac12 \sum_{k=1}^\infty \frac{(-1)^k}{k} \Tr T^k}+\order{1/N^2} \, ,
\label{eq:Q_n_gaussian}
\end{equation}
with
\begin{equation}
T (\vec{\tau}_1,\vec{\tau}_2) \equiv \varphi_0(\vec{\tau}_1) \left[ M (\vec{\tau}_1,\vec{\tau}_2) - \frac{1}{\eta} \int \dd{\vec{\tau}'} M (\vec{\tau}_2,\vec{\tau}') \varphi_0(\vec{\tau}') \right] \, .
\label{eq:T_def}
\end{equation}
Note that $\varphi_0(\va{\tau}) \sim \order{\eta}$ (see \cref{eq:saddle_point_comp}), so that the function $T (\vec{\tau}_1,\vec{\tau}_2)$ itself is also of $\order{\eta}$.
\subsection{Rotationally-invariant Ansatz}
In the pure GOE case (which is recovered, for instance, by setting $\psi_a(z)=1$ identically in the expressions above), the saddle-point equation \eqref{eq:saddle_point_comp} suggests to look for a replica-symmetric solution in the form of a Gaussian, i.e.,
\begin{equation}
\varphi_0(\vec{\tau}) = \cor{N} \exp(-\frac{1}{2} \vec{\tau} \, \hat C^{-1} \, \vec{\tau}) \, ,
\label{eq:ansatz}
\end{equation}
where $\hat C$ is another block-diagonal matrix of the form
\begin{equation}
\hat C \equiv \mqty(\dmat{\Delta_\alpha \mathbb{1}_{n_+}, \bar{\Delta}_\beta \mathbb{1}_{n_+}, \Delta_\beta \mathbb{1}_{n_-},\bar{\Delta}_\alpha \mathbb{1}_{n_-}}) \, .
\label{eq:hat_C_def}
\end{equation}
The four parameters $\Delta_\alpha, \bar\Delta_\alpha, \Delta_\beta$ and $\bar\Delta_\beta$ and are to be fixed, together with the prefactor $\cor{N}$ in \cref{eq:ansatz}, by substituting the Ansatz in \cref{eq:ansatz} back into the saddle-point equation \eqref{eq:saddle_point_comp}.
This strategy has proven effective in \ccite{Metz_2017} and, for completeness, in \cref{app:GOE} we sketch the entire calculation for the GOE case. However, in the presence of a nontrivial $\psi_a(z)$ in \cref{eq:saddle_point_comp}, there is no reason why the Ansatz in \cref{eq:ansatz} should work. We are thus naturally led to extend it to the form
\begin{equation}
\varphi_0(\vec{\tau}) = \cor{N} \exp(-\frac{1}{2} \vec{\tau} \, \hat C^{-1} \, \vec{\tau}) \psi_a\left(-\frac{1}{2} \vec{\tau} \, \hat L \, \vec{\tau}\right) \, .
\label{eq:ansatz_new}
\end{equation}
This can be plugged into \cref{eq:saddle_point_comp} to first obtain $\cor{N} = \eta / Z_\varphi $, where
\begin{equation}
Z_\varphi = \int \dd{\vec{\tau}} \varphi_0(\vec{\tau}) \, .
\label{eq:Z_varphi}
\end{equation}
By using Gaussian integration one can then show that
\begin{equation}
\int \dd{\vec{\tau}'} M(\vec{\tau},\vec{\tau}') \varphi_0(\vec{\tau}') = \eta \vec{\tau} \, \hat K \, \vec{\tau} \, ,
\label{eq:property}
\end{equation}
where $\hat K$ is a diagonal matrix given by
\begin{equation}
\hat K = -i \hat L \, \cor{G}_a \left( (i\hat L \hat C)^{-1} \right)\, ,
\label{eq:K_def}
\end{equation}
$\cor{G}_a$ is the resolvent in \cref{eq:resolvent}, and $\hat{L}$ is given in \cref{eq:lambda_L_def}.
The remaining free parameters in \cref{eq:ansatz} can be determined by solving the set of four self-consistency equations which follow from \cref{eq:saddle_point_comp}, i.e.,
\begin{equation}
\hat C^{-1} \,= 2\eta \, \hat K +i \, \hat \Lambda \, .
\label{eq:four_set_p(a)}
\end{equation}
Finally, both the action and its Gaussian fluctuation matrix $T$ (see \cref{eq:T_def}) can be computed
by using the saddle-point solution in \cref{eq:ansatz_new}. First, notice that the action in \cref{eq:action_comp_varphi} takes the form
\begin{align}
\cor{S}_{n_\pm}[\varphi_0 ; \hat \Lambda] &= \frac12 \int \dd{\vec{\tau}} \left( \vec{\tau} \, \hat K \, \vec{\tau} \right) \varphi_0(\vec{\tau})+ \ln Z_\varphi \nonumber\\
&= \frac{\eta}{2} \sum_i K_{ii}^2 + \ln \int_{-\infty}^\infty \dd{a} p_a(a) \exp{ -\frac12 \Tr \ln[ (i\hat L \hat C)^{-1} - a ] } + \T{const.} \, .
\end{align}
The constant term vanishes upon taking the analytic continuation $n_\pm \to \pm is/\pi$, yielding
\begin{align}
\cor{S}_{\pm \frac{is}{\pi}}[\varphi_0 ; \hat \Lambda]=& \frac{is \eta}{2\pi} \left( k_\alpha^{2} + \bar{k}_\beta^{2} - k_\beta^{2} - \bar{k}_\alpha^{2} \right) \nonumber\\
&+ \ln \int_{-\infty}^\infty \dd{a} p_a(a) \exp{ -\frac{is}{2\pi} \ln \left[ \frac{(\bar{\Delta}_\beta^{-1}+ia)(\Delta_\alpha^{-1}-ia)}{(\bar{\Delta}_\alpha^{-1}+ia)(\Delta_\beta^{-1}-ia)} \right] }\, ,
\label{eq:action_prolungata}
\end{align}
where we introduced for brevity $\hat{K} \equiv \T{diag}(k_\alpha \mathbb{1}_{n_+}, \bar{k}_\beta \mathbb{1}_{n_+}, k_\beta \mathbb{1}_{n_-},\bar{k}_\alpha \mathbb{1}_{n_-})$. Note that this action coincides at leading order with the cumulant generating function in \cref{eq:cgf}, i.e.,
\begin{align}
\cor{F}_{[\alpha,\beta]}(s) = \lim_{\varepsilon \to 0^+} \cor{S}_{\pm \frac{is}{\pi}}[\varphi_0 ; \hat \Lambda] +\order{\eta/N} \, ,
\label{eq:cgf_general_solution}
\end{align}
where we used \cref{eq:Q_n_gaussian} (the estimate of the large-$N$ correction will soon be justified). In the next Section we will make this result more explicit in the case of an interval which is symmetric around the origin.
The Gaussian fluctuations around the saddle-point are studied in \cref{app:gaussian_fluctuations}. A closed-form result is not available in this case (in contrast to the GOE case, see \cref{app:GOE}), because the calculation involves increasingly complex generalizations of the resolvent $\cor{G}_a(z)$ which encode higher order correlations (see \cref{eq:gaussian_fluc_first_term}). However, we can show that the Gaussian fluctuations bring to the leading order term in \cref{eq:cgf_general_solution} a correction of $\order{\eta/N}=\order{N^{-\gamma}}$, which is strongly suppressed for large $N$ in the region $\gamma>1$ which we focus on here.
\subsection{General result in the case of a symmetric interval}
\label{par:symmetric_interval}
We consider now the simpler case in which $\alpha=-E$ and $\beta=E$, and we take a symmetric distribution $p_a(a)$. From \cref{eq:K_def,eq:four_set_p(a)} one can deduce that the entries of the matrix $\hat{C}$ are related by the following symmetries:
\begin{equation}
\Delta_\alpha \equiv \Delta \equiv r e^{i\theta}\,, \;\;\;\;\; \bar{\Delta}_\alpha=\Delta_\alpha^* = \Delta_\beta \,, \;\;\;\;\; \bar{\Delta}_\beta = \Delta_\beta^* \, .
\label{eq:symmetries}
\end{equation}
The same holds for the entries of $\hat{K}$ (see \cref{eq:K_def}), hence we will simply call $k_\alpha \equiv k$. The problem is then reduced to computing one single unknown, namely $\Delta$: from \cref{eq:K_def,eq:four_set_p(a)}, this amounts to solving the self-consistency equations
\begin{equation}
\begin{dcases}
\Delta^{-1}= \varepsilon-iE +2\eta k \, , \\
k = -i\cor{G}_a \left(\frac{1}{i\Delta}\right) \, .
\label{eq:self_consistency_p(a)_symm}
\end{dcases}
\end{equation}
The action in \cref{eq:action_prolungata} then takes the form
\begin{align}
\cor{S}[\varphi_0 ; \hat \Lambda]= -\frac{2\eta s}{\pi} \Im{k^2} + \ln \int_{-\infty}^\infty \dd{a} p_a(a) \exp[ -\frac{s}{\pi} \arctan \left( \frac{ \sin 2\theta}{a^2r^2 + \cos 2\theta} \right) ]\, ,
\label{eq:action_comp_symmetric_interval}
\end{align}
where the branch of the $\arctan$ is chosen so that it returns an angle in $[0,\pi]$. From \cref{eq:cgf} we can then read out the leading order contribution to the rate function, namely
\begin{align}
\cor{F}_{[-E,E]}(s) = \cor{S}[\varphi_0 ; \hat \Lambda] + \order{\eta/N} = -m s + \ln \expval{e^{-s f(a) }}_a + \order{\eta/N} \, .
\label{eq:cgf_GRP}
\end{align}
Here we used the notation $\expval{\bullet}_a$ to indicate an average over $p_a(a)$, and we introduced for brevity
\begin{align}
m&\equiv \frac{2\eta}{\pi} \Im{k^{2}} = - \frac{2\eta}{\pi} \Im \left[ \cor{G}_a \left(\frac{1}{i\Delta}\right) \right]^2 \, , \label{eq:m_def} \\
f(a) &\equiv \frac{1}{\pi} \arctan \left( \frac{ \sin 2\theta}{a^2r^2 + \cos 2\theta} \right) \in [0,1] \, ,
\label{eq:f_def}
\end{align}
where in the first line we used \cref{eq:self_consistency_p(a)_symm}.
The cumulant generating function in \cref{eq:cgf_GRP}, together with the self-consistency equations \eqref{eq:self_consistency_p(a)_symm}, represent our second main result. As we stressed above, when $\eta$ is given by \cref{eq:eta}, the correction to \cref{eq:cgf_GRP} is of $\order{\eta/N}=\order{N^{-\gamma}}$, which is strongly suppressed for large $N$ in the region $\gamma>1$.
Expanding \cref{eq:cgf_GRP} in powers of $s$ we get
\begin{align}
\cor{F}_{[-E,E]}(s) \simeq -s \left[ m+ \expval{f(a)}_a \right] + \frac{s^2}{2} \left[ \expval{f^2(a)}_a-\expval{f(a)}_a^2 \right]+ \order{s^3,\eta/N}\, ,
\end{align}
and by comparison with \cref{eq:cumulants} we can identify the first two cumulants
\begin{align}
\frac{\kappa_1}{N} &= m +\expval{f(a)}_a + \order{\eta/N}\, , \label{eq:first_cumulant} \\
\frac{\kappa_2}{N} &= \expval{f^2(a)}_a - \expval{f(a)}_a^2+ \order{\eta/N} \, .
\end{align}
From \cref{eq:level_compress}, we finally obtain the level compressibility
\begin{equation}
\chi(E) = \frac{\expval{f^2(a)}_a - \expval{f(a)}_a^2}{ m +\expval{f(a)}_a} + \order{\eta/N}\, .
\label{eq:comp_general_prediction}
\end{equation}
We remark that not only the level compressibility, but actually all the moments of $I_N[-E,E]$ can be simply computed starting from Eq. \eqref{eq:cgf_GRP}: they read (at leading order for large $N$)
\begin{eqnarray}
\langle (I_N[-E,E])^m\rangle \simeq N \langle [f(a)]^m \rangle_a \, , \quad m \geq 2 \;.
\end{eqnarray}
\subsubsection{Limit of a pure diagonal matrix with random i.i.d. entries}
It is instructive, at this point, to consider the limit $\eta\to 0$. In this case the GOE part of \cref{eq:hamiltonian} is neglected, and the spectral properties are completely determined by the matrix $A$, whose entries are independent and identically distributed according to $p_a(a)$. The self consistency equation \eqref{eq:four_set_p(a)} then reduces to
\begin{equation}
\hat C \,= (i \, \hat \Lambda)^{-1} \, ,
\label{eq:SCE_iid}
\end{equation}
whence
\begin{equation}
\Delta =
\eval{\Delta_\alpha}_{\alpha=-E} = \frac{iE+\varepsilon}{E^2+\varepsilon^2} \equiv re^{i\theta} \, ,
\end{equation}
and there is no need to determine the entries of $\hat K$ since it does not enter the expression of the saddle point action \eqref{eq:action_comp_symmetric_interval} for $\eta = 0$. From \cref{eq:action_comp_symmetric_interval} we obtain, for $\varepsilon \to 0^+$,
\begin{align}
\cor{S}[\varphi_0 ; \hat \Lambda] &= \ln \int_{-\infty}^\infty \dd{a} p_a(a) \exp{ -\frac{s}{\pi} \arctan \left[ \frac{ 0^+}{(a/E)^2 -1} \right] } \nonumber\\
&= \ln \left[ 1+(e^{-s}-1) \int_{-E}^E \dd{a} p_a(a) \right] \, ,
\end{align}
where we used that the branch of $\arctan (z)\in [0,\pi]$ has a discontinuity in $z=0$, i.e., it jumps from $\pi$ to $0$ as $z$ becomes positive.
From \cref{eq:cgf_GRP,eq:levels_number}, we can then read out the cumulant generating function
\begin{equation}
\cor{F}_{[-E,E]}(s) = \ln \left[ 1+(e^{-s}-1) \frac{\expval*{I_N[-E,E]}_a}{N} \right] \, .
\label{eq:cgf_iid}
\end{equation}
In this way we recover the standard textbook result for the cumulant generating function in the case of i.i.d. random variables, which we sketch for completeness in \cref{app:iid}. In particular, the level compressibility in \cref{eq:level_compress} reads in this case
\begin{equation}
\chi(E) = 1 - \frac{\expval{I_N[-E,E]}_a}{N} \, ,
\label{eq:comp_Poisson}
\end{equation}
so that in general $\chi(E)\sim 1$ for small $E$, and $\chi(E)\to 0$ for large $E$.
Finally, we note that \cref{eq:cgf_iid} is exact, because we have stressed above that the Gaussian fluctuations around the saddle point are at least of $\order{\eta}$ and so they vanish in the limit $\eta \to 0$.
\subsubsection{Limit of a pure GOE matrix}
\label{par:comp_pureGOE}
The opposite limit in which $A$ is neglected in \cref{eq:hamiltonian} is obtained by setting $p_a(a)=\delta(a)$, whose corresponding resolvent is $\cor{G}_a(z) = 1/z$ (indeed, summing zeros to the matrix $B$ in \cref{eq:hamiltonian} does not change its spectrum). Equation \eqref{eq:K_def} then implies $\hat K = \hat C$, where we used $\hat L^2 = \mathbb{1}$ (see \cref{eq:lambda_L_def}). The self-consistency equations \eqref{eq:four_set_p(a)} are then seen to coincide with \cref{eq:four_set_GOE}, corresponding to the GOE case studied in \ccite{Metz_2017} and here revisited in \cref{app:GOE}. From \cref{eq:action_comp_symmetric_interval} we obtain, in terms of $r$ and $\theta$ introduced in \cref{eq:symmetries},
\begin{align}
\cor{S}[\varphi_0 ; \hat \Lambda]= -\frac{2\eta r^2 s}{\pi} \sin 2\theta - \frac{2s\theta}{\pi} \, ,
\label{eq:action_GOE}
\end{align}
which is linear in $s$: we deduce that the cumulants higher than the average $\expval{I_N[-E,E]}$ are subleading for large $N$, and they are only accessible by explicitly computing the Gaussian fluctuations around the saddle-point (see \cref{app:gaussian_fluctuations}).
Still, the leading order term in the first cumulant (see \cref{eq:k1_GOE}) is correctly reproduced by \cref{eq:action_GOE}.
\subsection{Exactly solvable cases}
\label{par:comp_exactly_solvable}
The cumulant generating function we found in \cref{eq:cgf_GRP} is formal, in that $\Delta = re^{i\theta}$ must first be determined by solving the self-consistency equation \eqref{eq:four_set_p(a)}. In the following, we will present two cases in which the result can be expressed in closed form, namely those in which $p_a(a)$ is the Cauchy or the Wigner distribution, respectively. While the former presents fat tails and is thus slowly decaying, the latter has a compact support and a well-defined edge.
\subsubsection{Cauchy distributed $a_i$}
Let us start from the case in which $p_a(a)$ is the Cauchy distribution, see \cref{eq:lorentzian}.
We set $\mu=0$, so that the problem remains symmetric around the origin. Using the expression of $\cor{G}_a(z)$ in \cref{eq:cauchy_resolvent}, $\hat K$ in \cref{eq:K_def} reduces to
\begin{equation}
\hat K = \frac{\hat C}{1+\omega \hat C} \, .
\label{eq:K_cauchy}
\end{equation}
Solving the self-consistency equation \eqref{eq:four_set_p(a)} then yields
\begin{equation}
\Delta=\eval{\Delta_\alpha}_{\alpha=-E} = 2 \left[ \varepsilon -iE -\omega + \sqrt{8\eta+\left(\varepsilon -iE +\omega \right)^2 } \right]^{-1} \equiv re^{i \theta} \, ,
\end{equation}
where we chose the positive branch of the square root so that $\Re \Delta_\alpha >0$. After this choice, the constant $\varepsilon$ can be safely sent to zero. The resulting cumulant generating function is given in \cref{eq:cgf_GRP}, where the averages are taken over the Cauchy distribution in \cref{eq:lorentzian}, and the first linear contribution reads
\begin{equation}
m= \frac{4\eta r^2 \sin \theta \left( r \omega +\cos\theta \right)}{\pi \left[ 1+(r \omega)^2+2 r \omega \cos\theta \right]^2}\, .
\end{equation}
\begin{figure}[ht]
\centering
\subfloat[]{
\includegraphics[width=0.47\textwidth]{cauchy_n=1000_ntot=5000.pdf}
\label{fig:cauchy_comp}
}
\subfloat[]{
\includegraphics[width=0.47\textwidth]{wigner_n=1000_ntot=5000.pdf}
\label{fig:wigner_comp}
}
\caption{Numerical check of the level compressibility $\chi(E)$ predicted in \cref{eq:comp_general_prediction} (solid blue line), for the case in which the elements of the diagonal matrix $A$ belong to the (a) Cauchy, or (b) Wigner distribution. Numerical data (symbols) are obtained from the numerical diagonalization of $N_\T{tot}=1000$ random matrices of size $N=5000$. The (dashed yellow) line denoted by \textit{Poisson} shows the level compressibility as it would be in the absence of level repulsion -- see the main text.
In both plots we used the parameters $\mu=0$, $\gamma=1.1$. In (a) we chose $\omega=1$ and $\nu=1$, corresponding to $\eta=0.125$, while in (b) we set $\sigma=1$ and $\nu=0.2$, yielding $\eta=0.005$.}
\label{fig:comp_numerics}
\end{figure}
The resulting level compressibility $\chi(E)$ (see \cref{eq:comp_general_prediction}) is plotted in \cref{fig:cauchy_comp}, and it is compared to numerical results showing excellent agreement. We show in the same plot the level compressibility under the hypothesis that the level statistics is of the Poisson type, i.e., that the energy levels do not repel each other. This is given by \cref{eq:comp_Poisson} upon interpreting the average $\expval{\bullet}$ as taken over the average eigenvalue density $\rho(\lambda)$ for the case in which the matrix $A$ has Cauchy-distributed entries: this has been found previously in \cref{eq:density_cauchy}. Even for very small values of $\eta$, the behavior at low energies $E$ of the compressibility $\chi (E)$ is qualitatively very different: in the Poisson case we have $\chi (E)\sim 1$, while in the GRP model it is $\chi (E) \sim 0$. The latter $\chi(E)$ increases up to a maximum, whose position $E_\T{max}(\eta)$ grows monotonically (and sublinearly) with $\eta$.
\subsubsection{Wigner distributed $a_i$}
Let us now consider the case in which $p_a(a)$ is the Wigner distribution, see \cref{eq:wigner}.
Again we set $\mu=0$, so that the problem remains symmetric around the origin. Using the expression of $\cor{G}_a(z)$ in \cref{eq:wigner_resolvent}, the quantity $\hat K$ in \cref{eq:K_def} becomes
\begin{equation}
\hat K = - \frac{2}{\sigma^2 \hat C}\left[ 1-\sqrt{1+(\sigma \hat C)^2} \right] \, .
\label{eq:K_wigner}
\end{equation}
Solving the self-consistency equation \eqref{eq:four_set_p(a)} yields, after choosing the positive branch of the square root so that $\Re \Delta_\alpha >0$ and letting $\varepsilon\to 0$,
\begin{equation}
\Delta = \eval{\Delta_\alpha}_{\alpha=-E} = \frac{iE(\sigma^2+4\eta) + 4\eta \sqrt{\sigma^2+8\eta-E^2}}{16\eta^2+\sigma^2E^2} \equiv re^{i \theta} \, .
\end{equation}
The resulting cumulant generating function is again given in \cref{eq:cgf_GRP}, where the averages are taken over the Wigner distribution in \cref{eq:wigner}. The analytical expression of the quantity $m$ in \cref{eq:m_def} is cumbersome, but it follows readily from \cref{eq:K_wigner}.
In \cref{fig:wigner_comp} we plot the level compressibility $\chi (E)$ as we did for the Cauchy case, finding a qualitatively similar behavior.
\subsection{Scaling limit and Thouless energy}
\label{par:scaling}
In this Section we focus on the limit in which $E=x \eta^\delta$ and $\eta \ll 1$, while $x\sim \order{1}$. We can envision a different behavior depending on whether the exponent $\delta>1$, $\delta<1$ or $\delta=1$. The latter case turns out to be particularly interesting: we will show that the level compressibility computed in $\chi(E=x\eta^\delta)$ assumes for $\delta=1$ a \textit{universal} scaling form, which is independent of the particular choice of the distribution $p_a(a)$ of the entries of the diagonal matrix $A$.
To this end, let us go back to the self-consistency equations \eqref{eq:self_consistency_p(a)_symm}, which we can rewrite for $\delta=1$ and $\eta \ll 1$ as
\begin{align}
\Delta^{-1} &= \varepsilon-ix \eta -2i \eta \cor{G}_a\left(-i \Delta^{-1} \right) \nonumber\\
&\simeq \varepsilon-ix \eta -2i \eta \cor{G}_a\left(-i \varepsilon + \order{\eta} \right) \, ,
\label{eq:SCE_scaling}
\end{align}
where again $\Delta=r e^{i\theta}$.
By taking the complex conjugate of \cref{eq:SCE_scaling} and by summing and subtracting the two equations, we obtain the two conditions
\begin{align}
r^{-1} \cos\theta &= \varepsilon + 2 \eta \Im \cor{G}_a\left(-i \varepsilon + \order{\eta} \right) \xrightarrow[\varepsilon\to 0^+]{\eta \ll 1} 2 \pi \eta p_a(0) \, , \\
r^{-1} \sin\theta &= x \eta + 2 \eta \Re \cor{G}_a\left(-i \varepsilon + \order{\eta} \right) \xrightarrow[\varepsilon\to 0^+]{\eta \ll 1} x \eta \, ,
\end{align}
where we used the Plemelj-Sokhotski formula recalled in \cref{eq:plemelj}, and the fact that for a symmetric distribution $p_a(a)$ one has
\begin{equation}
\pv{\int \dd{a} \frac{p_a(a)}{a}} = 0 \, .
\end{equation}
One then easily obtains, at leading order for small $\eta$ (and with $x=E/\eta$),
\begin{align}
\tan\theta &\simeq \frac{x}{2\pi p_a(0)} \equiv y \, , \\
r^{-1} &\simeq 2\pi \eta p_a(0) \sqrt{1+y^2} \, . \label{eq:rho_scaling}
\end{align}
We can now estimate the level compressibility in this limit. First, notice that
\begin{equation}
\cor{G}_a(-i\Delta^{-1}) = i\pi p_a(0) +\order{\eta} \, ,
\end{equation}
so that from \cref{eq:m_def} we read
\begin{equation}
m = - \frac{2\eta}{\pi} \Im \left[ \cor{G}_a(-i\Delta^{-1}) \right]^2 = \order{\eta^2} \, .
\end{equation}
From \cref{eq:first_cumulant}, the first cumulant $\kappa_1$ thus reduces to
\begin{align}
\frac{\kappa_1}{N} &\simeq \expval{f(a)}_a = \int \dd{a} p_a(a) f(a) \simeq \frac{p_a(0)}{r} \int_{-\infty}^\infty \dd{u} f(u/r) \nonumber\\
&= \frac{p_a(0)}{r} \frac{2y}{\sqrt{1+y^2}} = 2 p_a(0) x \eta \, ,
\end{align}
where in the first line we changed variable to $u= r a$ and we used the fact that $r^{-1} \sim \order{\eta}$ (see \cref{eq:rho_scaling}), while in the second line we explicitly computed the integral
\begin{equation}
\frac{1}{\pi} \int_0^\infty \dd{u} \arctan \left( \frac{\sin 2\theta}{u^2+\cos 2\theta} \right) = \frac{\tan \theta}{\sqrt{1+\tan^2 \theta}} =\sin{\theta} \, , \;\;\;\;\;\;\; \theta \in \left[0,\frac{\pi}{2}\right] \, ,
\end{equation}
and we inserted the expression for $r$ found in \cref{eq:rho_scaling} (note that $f(u/r)$ is actually $r$-independent -- see \cref{eq:f_def}). One could alternatively compute $\kappa_1$ by taking the average of \cref{eq:levels_number}: this leads to the same result upon expanding for small $E$ and $\eta$, since
\begin{equation}
\rho(\lambda) = p_a(\lambda) + \order{\eta} \, .
\end{equation}
The same steps can be repeated for the second (and possibly any other) cumulant $\kappa_2$, yielding
\begin{equation}
\frac{\kappa_2}{N} \simeq \expval{f^2(a)}_a \simeq \frac{p_a(0)}{r} \int_{-\infty}^\infty \dd{u} f^2(u/r) \, .
\label{eq:k2_approx}
\end{equation}
From \cref{eq:comp_general_prediction} we thus obtain the leading order estimate for the level compressibility when $\eta \ll 1$, which takes the universal scaling form
\begin{equation}
\chi \left( y= \frac{E}{2\pi p_a(0) \eta} \right) \simeq \chi_T(y) \equiv \frac{\sqrt{1+y^2}}{\pi^2 y} \int_0^\infty \dd{u} \left\lbrace \arctan \left[\frac{2y}{u^2(1+y^2)+1-y^2 }\right] \right\rbrace^2 ,
\end{equation}
where we stress that we have chosen the branch $\arctan(z) \in [0,\pi]$. Upon integrating by parts and performing some algebra \cite{stack,table}, the integral over $u$ can be computed explicitly to give
\begin{equation}
\chi_T(y) = \frac{1}{\pi y}\left[2y \arctan(y) -\ln (1+y^2) \right] \, .
\label{eq:comp_universal}
\end{equation}
The function $\chi_T(y)$ grows monotonically from $0$ to $1$ as we increase $y$, and it is plotted in \cref{fig:scaling} together with the level compressibility for the two cases explicitly solved above, i.e., Cauchy and Wigner. We find a good agreement at low energies $y$, while we observe a departure at large energies: here the scaling prediction keeps growing, while the actual compressibility must hit a maximum and start decreasing -- see \cref{fig:comp_numerics}. The same trend can be observed in \cref{fig:collapse_cauchy}, where we evaluate the level compressibility for different values of $\eta$, and show that they collapse on a common master curve when they are plotted as a function of $y \propto E/\eta$.
\begin{figure}[ht]
\centering
\subfloat[]{
\includegraphics[width=0.47\textwidth]{scaling.pdf}
\label{fig:scaling}
}
\subfloat[]{
\includegraphics[width=0.47\textwidth]{collapse_cauchy.pdf}
\label{fig:collapse_cauchy}
}
\caption{Scaling form of the level compressibility, in the limit in which $E\propto y \eta$ and $\eta \ll 1$. In (a) we compare the universal prediction in \cref{eq:comp_universal} (solid red line) with the two exactly solvable cases studied in \cref{par:comp_exactly_solvable} (symbols), showing a good agreement at low energies $y =E/[2\pi p_a(0) \eta]$ (i.e., the condition under which \cref{eq:comp_universal} was derived). We chose $\eta=5\cdot 10^{-4}$ and $\omega=0.5$, $\sigma=1$, so that $p_a(0)$ assumes the same value for the two distributions (see inset). In (b) we exemplify in the Cauchy case how the curves corresponding to different values of $\eta$ collapse onto the same master curve, when plotted as a function of $y\propto E/\eta$ for $\eta\ll 1$.}
\label{fig:scaling_chi}
\end{figure}
The other two cases ($\delta>1$ or $\delta<1$) can be easily addressed by the same token. When $\delta>1$, by studying the self-consistency equations as in \cref{eq:SCE_scaling} we obtain at leading order
\begin{align}
\tan\theta &\simeq y \eta^{\delta-1} \, , \\
r^{-1} &\simeq 2\pi \eta p_a(0) \, .
\end{align}
It can be readily seen that $\kappa_1/N \simeq 2p_a(0)x\eta^\delta$ and $\kappa_2/N \sim \order{\eta^{2\delta-1}}$, so that in this limit
\begin{equation}
\chi(E=x\eta^\delta) \sim \order{\eta^{\delta-1}} \, , \;\;\;\;\;\;\; \delta >1 \, , \; \eta \ll 1 \, .
\label{eq:comp_lowE}
\end{equation}
This resembles the behavior of $\chi(E)$ in the case of a pure GOE matrix, see \cref{par:comp_pureGOE,app:GOE}.
Conversely, for $\delta<1$ the self-consistency equations \eqref{eq:self_consistency_p(a)_symm} reduce to
\begin{equation}
\Delta^{-1} \simeq \varepsilon -ix\eta^\delta \, ,
\end{equation}
and, by comparison with \cref{eq:SCE_iid}, we identify this limit as that in which the eigenvalues behave as i.i.d. random variables. In particular,
\begin{equation}
\chi(E=x\eta^\delta) \simeq 1- \frac{\expval{I_N[-E,E]}_a}{N} \, , \;\;\;\;\;\;\; \delta <1 \, , \; \eta \ll 1 \, .
\label{eq:comp_largeE}
\end{equation}
The above analysis suggests to interpret the quantity $E_T \sim 2\pi p_a(0) \eta $ as the \textit{Thouless energy} of the system. Indeed, consider again the limit in which $\eta\ll 1$, and let $E\propto \eta^\delta$. For $E\ll E_T$ (i.e., $\delta>1$), the eigenvalues organize in multiplets (or mini-bands \cite{Kravtsov_2015}) and they repel each other as in the GOE ensemble -- as a result, the level compressibility is zero at leading order (see \cref{eq:comp_lowE}). For $E\gg E_T$ ($\delta<1$), on the other hand, the various multiplets no longer interact, and we recover the Poisson statistics -- see \cref{eq:comp_largeE}. Finally, the case $\delta=1$ marks a crossover in which the level compressibility $\chi(E/E_T)$ assumes the universal scaling form given in \cref{eq:comp_universal}.
Indeed, the asymptotics of the function $\chi_T(y)$ in \cref{eq:comp_universal} can be checked to give
\begin{equation}
\chi_T(y) \simeq
\begin{dcases}
\frac{y}{\pi} \, , & y \ll 1 \, , \\
1- \frac{2(1+\ln y)}{\pi y} \, , & y \gg 1 \, ,
\end{dcases}
\end{equation}
showing that $\chi_T(y)$ interpolates between Wigner-Dyson statistics at low energy, and Poisson statistics at higher energy.
\begin{figure}[ht]
\centering
\subfloat[]{
\includegraphics[width=0.47\textwidth]{numerics_large_scale_single.pdf}
\label{fig:numerics_universal}
}
\subfloat[]{
\includegraphics[width=0.47\textwidth]{numerics_various_N.pdf}
\label{fig:numerics_N}
}
\caption{Behavior of the level compressibility $\chi(E)$ at low energies. The symbols correspond to numerical results, and we indicated with vertical lines the mean level spacing $\delta_N$ (see \cref{eq:mean_level_spacing}) and the Thouless energy $E_T \sim N^{1-\gamma}$. In panel (a), $p_a(a)$ is chosen Gaussian with unit variance or uniform, and $N=2^{16}$. The region $E \lesssim \delta_N$ is described by the GOE prediction in \cref{chi_GOE_final}, while the crossover region $E\sim E_T$ is described by the universal scaling form in \cref{eq:comp_universal}. In panel (b), $p_a(a)$ is Gaussian, and we show the approach to the universal curve $\chi_T(E)$ for increasing values of the matrix size $N$. We used $\gamma=1.5$ throughout, and the simulations with $N=2^{14},2^{15},$ or $2^{16}$ are averaged over $N_\T{tot}=256,64,$ or $32$ samples, respectively.}
\label{fig:numerics_scaling}
\end{figure}
\subsection{Behavior at low energy}
\label{par:low_energy}
In this final Section we use extensive numerical exact diagonalization of large random matrices in order to inspect the low-energy behavior of the level compressibility $\chi(E)$. Indeed, our prediction of \cref{par:scaling} is expected to break down for energies of the order of the mean level spacing $\delta_N$ of the finite-sized matrix $\cor{H}$, which is given by $\delta_N \simeq [N p_a(0)]^{-1}$, Eq.~\eqref{eq:mean_level_spacing}.
Equivalently, we expect that the replica calculation provides the correct result in the $N \to \infty$ limit, while for a matrix of size $N$ and for sufficiently small $E$ we should eventually recover the exact GOE result $\chi_\T{GOE}(E)$ \cite{Dyson_1962_I,Dyson_63_IV,Mehta_2004_book,Forrester_2010_book,Mirlin_2000,marino2014phase,Marino_2016, Tikhonov_2019}, whose derivation is reported in \cref{app_chi_GOE} -- see \cref{chi_GOE_final}.
The overall picture is thus the one we present in \cref{fig:numerics_scaling}, and which we support by numerical results. The region $E \lesssim \delta_N$ is described by the GOE prediction in \cref{chi_GOE_final}. For $1< \gamma < 2$ the Thouless energy $E_T \sim N^{1-\gamma}$ is such that $\delta_N \ll E_T \ll 1$, so that the crossover region with $E \sim E_T$ is described by the universal function $\chi_T(E)$ given in \cref{eq:comp_universal}. For larger $E \gtrsim \order{1}$, the level compressibility becomes model dependent and it is described by \cref{eq:comp_general_prediction} (see also \cref{fig:comp_numerics}).
The datapoints\footnote{Although the scaling function in \cref{eq:comp_universal} has been derived under the assumption that the spectral density is symmetric, the numerical results are obtained by averaging the cumulants of the number of eigenvalues within many energy windows across the whole bandwidth, for which the symmetry with respect to the center of the window is lost. However, on the scale of the Thouless energy $E \simeq E_T \propto \eta$, the corrections due to the fact that $p_a(-E) \neq p_a(E)$ are of $\order{\eta}$, and therefore yield a contribution which is of the same order as the finite-size corrections, and which can be neglected for sufficiently large $N$.} presented in panel (a) of \cref{fig:numerics_scaling} correspond to the choices of $p_a(a)$ Gaussian or uniform, which supports our claim of universality of $\chi_T(E)$ in the region $E\sim E_T$. Note, in fact, that the data follow the predicted curves (up to finite-size effects) with no adjustable parameters (i.e., no fitting was needed). The datapoints are eventually observed to deviate from the scaling prediction, as they reach a maximum and start decaying to zero as in \cref{fig:comp_numerics}. By increasing $N$, however, this maximum is observed to shift to larger values of $E$, and the plateau around $\chi(E)\sim 1$ broadens accordingly.
\section{Conclusions}
In this work we have used the replica method to study the average spectral density (\cref{par:spectral_replica_approach}) and the local level statistics (\cref{par:level_compress}) of a deformed GOE random matrix ensemble known as the generalized RP model. We have focused in particular on its non-ergodic extended phase \cite{Kravtsov_2015}, which is conveniently characterized in terms of the level compressibility $\chi(E)$ (see \cref{eq:level_compress}). In particular, we have shown that $\chi(E)$ assumes a universal form independent of the character $p_a(a)$ of the deformation matrix $A$ (see \cref{eq:hamiltonian}), provided that the system is probed over energy scales of the order of the Thouless energy $E_T$ (see \cref{par:scaling}).
It is natural to conjecture that this universal regime should persist in structurally similar random matrix ensembles. For instance, one could numerically inspect the case in which the GOE matrix is replaced by a Wigner matrix (i.e., any real symmetric matrix with i.i.d. random entries taken from a probability distribution with finite variance), whose limiting average spectral density is still given by the semi-circle law \cite{arous2011wigner}. Similarly, it would be interesting to study the effect of the diagonal deformation matrix $A$ on a Wishart matrix (see, e.g., \cite{potters2020first}), or else on a sparse (rather than dense) matrix $B$, such as those describing the Erdös-Rényi random graph \cite{Semerjian_2002,Rogers_2008,Metz_2014},
which can still be treated analytically (at least to some extent);
if the average connectivity is chosen finite, the spectral density of Erdös-Rényi is no longer a semi-circle, but the local statistics is still of the Wigner-Dyson type. Finally, it is worth mentioning that the spectral properties of the intermediate phase of the generalized RP model are particularly simple: differently from realistic interacting quantum systems, the mini-bands are compact and the eigenvectors are fractal but not \emph{multi}fractal (meaning that all the moments of the wave-functions' amplitudes are described by the same fractal exponent $D_\gamma = 2 - \gamma$). In order to overcome some of these issues, several extensions of the RP model have been proposed in the last few years, featuring either log-normal~\cite{kravtsov2020localization,khaymovich2020fragile} or power-law~\cite{biroli2021levy} distributed off-diagonal matrix elements. In these models the mini-band structure turns out to be non-compact but rather multifractal. It would be therefore illuminating to study the behavior of the level compressibility at small energy and check whether or not the universal form discussed here is robust with respect to these modifications.
\section*{Acknowledgements}
We thank F. Metz, M. Potters, and P. Vivo for illuminating insights. DV would like to thank F. Ares and L. Capizzi for interesting discussions, and the LPTHE for the kind hospitality during a substantial part of the preparation of this work.
\paragraph{Funding information}
DV acknowledges partial financial support from Erasmus+ 2020-1-IT02-KA103-078180.
LFC acknowledges partial financial support from FCM2b ANR-19-CE30-0014. GS acknowledges partial financial support from ANR Grant No. ANR-17-CE30-0027-01 RaMaTraF.
\begin{appendix}
\numberwithin{equation}{section}
\section{Number of i.i.d. variables in an interval}
\label{app:iid}
In this Appendix we revise the standard textbook result for the statistics of the number of eigenvalues in a finite interval, when such eigenvalues behave as independent and identically distributed random variables. This will help clarifying the behavior of the level compressibility $\chi(E)$ in the case of Poisson level statistics.
Given $N$ random variables $a_i$ distributed according to $p_a(a_i)$ (e.g., the eigenvalues of the matrix $A$ in \cref{eq:hamiltonian}), the number of variables contained in the interval $[\alpha,\beta]$ can be written as
\begin{equation}
I_N[\alpha,\beta] = \sum_{i=1}^N \theta_i \, ,
\end{equation}
where we introduced the indicator function
\begin{equation}
\theta_i \equiv \mathbb{1}_{[\alpha,\beta]}(a_i) \, , \;\;\;\;\;\qquad \mathbb{1}_{[\alpha,\beta]}(x) =
\begin{cases}
1 & x\in [\alpha,\beta] \, , \\
0 & x \not\in [\alpha,\beta] \, .
\end{cases}
\end{equation}
Its cumulant generating function can be constructed by noting that
\begin{equation}
e^{-s I_N[\alpha,\beta]} = \prod_{i=1}^N e^{-s \theta_i} = \prod_{i=1}^N \left[1+(e^{-s}-1) \theta_i \right] \, ,
\end{equation}
and then
\begin{equation}
\ln \expval{e^{-s I_N[\alpha,\beta]}} = \sum_{i=1}^N \ln \left[1+(e^{-s}-1) \int_\alpha^\beta \dd{a} p_a(a) \right] = N \ln \left[1+(e^{-s}-1) \frac{\expval{I_N[\alpha,\beta]}}{N} \right] \, .
\label{eq:cgf_iid_textbook}
\end{equation}
Note that this coincides with the limiting case in \cref{eq:cgf_iid} of our general result.
By expanding in powers of $s$ the two sides of \cref{eq:cgf_iid_textbook} and comparing with \cref{eq:cgf,eq:cumulants}, we can in particular extract the first two cumulants
\begin{equation}
\kappa_1 = \expval{I_N[\alpha,\beta]} \, , \;\;\;\;\;\qquad \kappa_2 = \expval{I_N[\alpha,\beta]} \left(1 - \frac{\expval{I_N[\alpha,\beta]}}{N}\right) \, ,
\end{equation}
and from \cref{eq:level_compress} we obtain the level compressibility
\begin{equation}
\chi(E) = 1 - \frac{\expval{I_N[-E,E]}}{N} \, .
\end{equation}
We thus generically expect $\chi(E)\sim 1$ for small $E$, and $\chi(E)\to 0$ for large $E$.
\section{Details of the replica calculation for the spectral density}
\label{app:replica}
In this Appendix we fill in the missing steps which lead from \cref{eq:replica_trick} to \cref{eq:average_Zn} in \cref{par:spectral_replica_approach}. A replica-based calculation for the pure GOE ensemble can be found in \ccite{Livan_2018}, from which we partially adopt the notation. We start by expressing the average of the replicated partition function as
\begin{equation}
\expval{ \cor{Z}^n(\lambda) } \propto \expval{ \int_{\mathbb{R}^{Nn}} \left( \prod_{\alpha=1}^n \dd{\vb{r}_\alpha}\right) \exp[-\frac{i}{2} \sum_{i,j=1}^N \sum_{\alpha=1}^n r_{i\alpha} (\lambda_\varepsilon \delta_{ij} -h_{ij} )r_{j\alpha} ]}_{A,B},
\label{eq:representation_Zn}
\end{equation}
where we indicated by $h_{ij}\equiv a_i \delta_{ij}+J b_{ij}$ the elements of the random matrix ${\mathcal H}$ given in \cref{eq:hamiltonian}, and $J= J(N) \equiv \nu N^{-\gamma/2}$. The average symbol means
\begin{equation}
\expval{\bullet}_{A,B} \equiv \int_{\mathbb{R}^N} \left(\prod_{i\leq j}^N \dd{b_{ij}}\right) p_B(\{b_{ij}\}) \left( \prod_{i=1}^N \int \dd{a_i} p_a(a_i) \right) \left( \bullet \right) \, ,
\label{eq:averages_def}
\end{equation}
where the probability distribution of the elements $b_{ij}$ of the GOE matrix $B$ reads
\begin{equation}
p_B(\{b_{ij}\}) = \prod_{i=1}^N \frac{e^{-b_{ii}^2/2}}{\sqrt{2\pi}} \prod_{i\leq j} \frac{e^{-b_{ij}^2}}{\sqrt{\pi}} \, .
\end{equation}
Here and henceforth, Latin indices run up to $N$ in real space, while Greek indices run up to $n$ in replica space.
Computing the Gaussian integrals over $b_{ij}$ gives (up to a numerical constant)
\begin{align}
\expval{ \cor{Z}^n(\lambda) } \propto & \int_{\mathbb{R}^{Nn}} \left( \prod_{\alpha=1}^n \dd{\vb{r}_\alpha}\right) \Bigg\langle \exp[-\frac{i}{2} \sum_{i=1}^N \sum_{\alpha=1}^n (\lambda_\varepsilon -a_i )r_{i\alpha}^2 ] \Bigg\rangle_{A} \nonumber \\
& \times \exp{-\frac{J^2}{4} \left[\frac12 \sum_{i=1}^N \left( \sum_{\alpha=1}^n r_{i\alpha}^2 \right)^2 + \sum_{i\leq j} \left( \sum_{\alpha=1}^n r_{i\alpha} r_{j\alpha} \right)^2 \right] } \, ,
\label{eq:after_gaussian_integrals}
\end{align}
where $\expval{\bullet}_{A}$ indicates the reduced averaged over the entries of $A$ -- see \cref{eq:averages_def}. The interacting term in the second line can be usually decoupled by means of the Hubbard-Stratonovich transformation \cite{Mezard_1987}, which is however ineffective in our case, for a generic choice of $p_a(a)$.
We introduce instead the normalized density
\begin{equation}
\mu(\vec{y}) \equiv \frac{1}{N} \sum_{i=1}^N \prod_{\alpha=1}^n \delta (y_\alpha-r_{i\alpha}) \, ,
\end{equation}
where $\vec{y}\in \mathbb{R}^n$ has components $y_a\in \mathbb{R}$, and we insert into \cref{eq:after_gaussian_integrals} the functional integral representation of the identity
\begin{equation}
1 = N^{\T{dim}(\mu)}\int \cor{D}\mu \, \cor{D}\hat \mu \, \exp{-i \int \dd{\vec{y}} \hat \mu(\vec{y}) \left[ N\mu(\vec{y}) - \sum_{i=1}^N \prod_{\alpha=1}^n \delta (y_\alpha-r_{i\alpha}) \right] } \, .
\label{eq:identity}
\end{equation}
Here $\T{dim}(\mu)$ is the dimension of the field $\mu$, which renders the prefactor on the right hand side formally infinite -- this will be of no consequence in the following calculation, since this prefactor is $\lambda$-independent. Equation~\eqref{eq:identity} is useful because it allows us to rewrite
\begin{align}
&\frac12 \sum_{i=1}^N \left( \sum_{\alpha=1}^n r_{i\alpha}^2 \right)^2 + \sum_{i\leq j} \left( \sum_{\alpha=1}^n r_{i\alpha} r_{j\alpha} \right)^2 = \frac12 \sum_{i,j=1}^N \left( \sum_{\alpha=1}^n r_{i\alpha} r_{j\alpha} \right)^2 \\
&= \frac{N^2}{2}\int \dd{\vec{y}} \dd{\vec{w}} \mu(\vec{y}) \mu(\vec{w}) \left( \sum_{\alpha=1}^n y_\alpha w_\alpha \right)^2 = \frac{N^2}{2}\int \dd{\vec{y}} \dd{\vec{w}} \mu(\vec{y}) \mu(\vec{w}) \left( \vec{y} \cdot \vec{w} \right)^2 \nonumber \, ,
\end{align}
so that inserting the identity in \cref{eq:identity} into \cref{eq:after_gaussian_integrals} leads to
\begin{align}
&\expval{ \cor{Z}^n(\lambda) } \propto
\int \cor{D}\mu \, \cor{D}\hat \mu \, \exp{-i N \int \dd{\vec{y}} \hat \mu(\vec{y}) \mu(\vec{y}) -\frac{(JN)^2}{8}\int \dd{\vec{y}} \dd{\vec{w}} \mu(\vec{y}) \mu(\vec{w}) \left( \vec{y} \cdot \vec{w} \right)^2 }\nonumber \\
&\times \int_{\mathbb{R}^{Nn}} \left( \prod_{\alpha=1}^n \dd{\vb{r}_\alpha}\right) \expval{ \exp[-\frac{i}{2} \sum_{i=1}^N \sum_{\alpha=1}^n (\lambda_\varepsilon -a_i )r_{i\alpha}^2 +i\sum_{i=1}^N \int \dd{\vec{w}} \hat \mu(\vec{w}) \prod_{\alpha=1}^n \delta (w_\alpha-r_{i\alpha}) ]}_A . \label{eq:huge}
\end{align}
Staring at \cref{eq:huge} for long enough, one realizes that the second line contains $N$ copies of the same integral,
\begin{align}
&\int_{\mathbb{R}^{Nn}} \left( \prod_{\alpha=1}^n \dd{\vb{r}_\alpha}\right) \expval{ \exp[-\frac{i}{2} \sum_{i=1}^N \sum_{\alpha=1}^n (\lambda_\varepsilon -a_i )r_{i\alpha}^2 +i\sum_{i=1}^N \int \dd{\vec{w}} \hat \mu(\vec{w}) \prod_{\alpha=1}^n \delta (w_\alpha-r_{i\alpha}) ]}_A \nonumber\\
&=\left\lbrace \int_{\mathbb{R}^n} \dd{\vec{y}} \int \dd{a} p_a (a) \exp[-\frac{i}{2} \sum_{\alpha=1}^n (\lambda_\varepsilon -a )y_\alpha^2 +i \int \dd{\vec{w}} \hat \mu(\vec{w}) \prod_{\alpha=1}^n \delta (w_\alpha-y_{\alpha}) ] \right\rbrace^N \nonumber\\
&= \left\lbrace \int_{\mathbb{R}^n} \dd{\vec{y}} \int \dd{a} p_a (a) \exp[-\frac{i}{2} (\lambda_\varepsilon -a )|\vec{y}|^2 +i \hat \mu(\vec{y}) ] \right\rbrace^N \, .
\end{align}
Note that it was crucial to assume independent entries $a_i$, so that their distribution in \cref{eq:averages_def} is factorized.
Plugging this expression back into \cref{eq:huge} allows us to rewrite $\expval{ \cor{Z}^n(\lambda) }$ as reported in \cref{eq:average_Zn} of the main text.
\section{Connection with the Zee formula}
\label{app:zee}
In this Appendix we show why \cref{eq:resolvent_ledoussal} is hiddenly the Zee formula. In Ref. \cite{Zee_1996}, the recipe for computing the spectrum $\rho_{1+2}(\lambda)$ of the sum of two random matrices $M_1+M_2$ is given as follows:
\begin{enumerate}[(i)]
\item Compute the resolvents (or Green's functions) associated to $\rho_1(\lambda)$ and $\rho_2(\lambda)$, i.e., $\cor{G}_1(z)$ and $\cor{G}_2(z)$.
\item Compute their functional inverses $B_1(z)$ and $B_2(z)$, or Blue's functions, defined by
\begin{equation}
B(\cor{G}(z))=z \, .
\end{equation}
\item Apply the sum rule
\begin{equation}
B_{1+2}(z) = B_1(z) + B_2(z) -1/z \, .
\end{equation}
\item Invert the result back (see \cref{eq:inversion_resolvent}) to find
\begin{equation}
B_{1+2}(z) \to \cor{G}_{1+2}(z) \to \rho_{1+2}(\lambda) \, .
\end{equation}
\end{enumerate}
Another interesting object is however the $R$-function, which is simply defined as
\begin{equation}
R(z) \equiv B(z) - 1/z \, ,
\end{equation}
and which is easily seen to satisfy the free-sum rule \cite{Livan_2018,Voiculescu_1991,voiculescu1992free}
\begin{equation}
R_{1+2}(z) = R_1(z) + R_2(z) \, .
\end{equation}
It follows that
\begin{equation}
B_1(x) = B_{1+2}(x) - R_2(x) \, ,
\end{equation}
which we can choose to apply in particular on $x=\cor{G}_{1+2}(z)$, yielding by construction
\begin{equation}
B_1(\cor{G}_{1+2}(z)) = z - R_2(\cor{G}_{1+2}(z)) \, .
\end{equation}
Applying $\cor{G}_1$ on both sides finally yields
\begin{equation}
\cor{G}_{1+2}(z) = \cor{G}_1(z-R_2(\cor{G}_{1+2}(z))) \, .
\end{equation}
The analogy with \cref{eq:resolvent_ledoussal} is readily established once we recall that, if $M_2$ is a GOE matrix, then its $R$-function is simply $R_2(z) = z$ \cite{Biane_97}.
\section{Details of the replica calculation of the level compressibility}
In this Appendix we provide the technical steps for the derivation of the cumulant generating function and the level compressibility presented in \cref{par:level_compress}. A similar calculation for the pure GOE/GUE ensemble can be found in \ccite{Metz_2017}, while the derivation in the case of the Erdös-Rényi graph and the Anderson model on a random regular graph was reported in \ccite{Metz_2016}.
\subsection{Functional representation}
\label{app:replica_level_compress}
The target of this Section is to express $Q_{[\alpha,\beta]}(n_\pm)$ given in \cref{eq:Q_ab} within the replica formalism, as we did in \cref{app:replica}. The first step is to choose a suitable representation for the partition function $\cor{Z}(z)$ which appears in \cref{eq:Q_ab}: indeed, the one we introduced in \cref{eq:partition_function} is only appropriate if $\Im{z}<0$, being the integral not convergent otherwise. If on the contrary $\Im{z}>0$, then one should choose instead
\begin{equation}
\cor{Z}_+(z) \equiv \left(\frac{i}{2\pi}\right)^{N/2}\int_{\mathbb{R}^N} \dd{\vb{r}} e^{\frac{i}{2} \vb{r}^T (z \mathbb{1} -\cor{H} ) \vb{r} } \, .
\label{eq:partition_function_+}
\end{equation}
Since the various prefactors in front of the integral in \cref{eq:partition_function} will cancel out in \cref{eq:Q_ab} after we take the analytic continuation to $n_\pm \to \pm i s/\pi$, we will not need to keep track of them in the following.
In analogy with the representation in \cref{eq:representation_Zn}, we can still write
\begin{equation}
Q_{[\alpha,\beta]}(n_\pm) \propto \expval{ \int_{\mathbb{R}^{Nn}} \left( \prod_{\sigma=1}^n \dd{\vb{r}_\sigma}\right) \exp[-\frac{i}{2} \sum_{i,j=1}^N \sum_{\sigma=1}^n r_{i\sigma} (\Lambda_{\sigma \sigma} \delta_{ij} -L_{\sigma \sigma}h_{ij} )r_{j\sigma} ]}_{A,B},
\end{equation}
but now we interpret
\begin{equation}
n=2(n_+ + n_-) ,
\label{eq:n}
\end{equation}
because each of the four partition functions in \cref{eq:Q_ab} requires its own set of replicas (here labelled by the Greek index $\sigma$, to avoid confusion with the left boundary $\alpha$ of the interval). We have also replaced the eigenvalue $\lambda_\varepsilon$ by the block matrix $\hat \Lambda$, which is defined in \cref{eq:lambda_L_def} together with the block matrix $\hat{L}$. Notice that the elements $\bar{\alpha}_\varepsilon=-\alpha_\varepsilon^*$ of the matrix $\hat \Lambda$ follow from the representation in \cref{eq:partition_function_+}.
The very same steps which in \cref{app:replica} led us to \cref{eq:average_Zn} of the main text now give
\begin{equation}
Q_{[\alpha,\beta]}(n_\pm) \propto \int \cor{D}\mu \, \cor{D}\hat \mu \, \exp{N \cor{S}_{n_\pm}[\mu , \hat \mu;\hat \Lambda] } \, ,
\label{eq:average_Qn}
\end{equation}
with the action
\begin{align}
\cor{S}_{n_\pm}[\mu , \hat \mu;\hat \Lambda] \equiv& - i \int \dd{\vec{\tau}} \mu(\vec{\tau}) \hat \mu(\vec{\tau}) - \frac{\eta}{2} \int \dd{\vec{\tau}} \dd{\vec{\tau}'} \mu(\vec{\tau}) \mu(\vec{\tau}') \left( \vec{\tau} \, \hat L \, \vec{\tau}' \right)^2 \label{eq:action_comp_mu_muhat}\\
&+ \ln \int \dd{\vec{\tau}} \exp[-\frac{i}{2}\vec{\tau}\, \hat \Lambda \,\vec{\tau} +i \hat \mu(\vec{\tau}) ] \int \dd{a} p_a (a) \exp(\frac{i}{2} a \vec{\tau} \, \hat L \, \vec{\tau}) \, . \nonumber
\end{align}
This generalizes the action in \cref{eq:action}, and the vector $\vec{\tau}\in \mathbb{R}^n$ plays the same role as the vector $\vec{y}$ but in an extended replica space, with $n$ given in \cref{eq:n}.
By noting that the action in \cref{eq:action_comp_mu_muhat} is quadratic in $\mu$, we can evaluate the Gaussian functional integral in $\cor{D}\mu$ to obtain
\begin{align}
Q_{[\alpha,\beta]}(n_\pm) \propto & \int \cor{D}\hat \mu \, \exp{N \cor{S}_{n_\pm}[\hat \mu;\hat \Lambda] } \, , \label{eq:Q_n_muhat} \\
\cor{S}_{n_\pm}[\hat \mu;\hat \Lambda] \equiv& - \frac{1}{2\eta} \int \dd{\vec{\tau}} \dd{\vec{\tau}'} \hat\mu(\vec{\tau}) M^{-1}(\vec{\tau},\vec{\tau}') \hat\mu(\vec{\tau}') \label{eq:action_comp_muhat}\\
&+ \ln \int \dd{\vec{\tau}} \exp[-\frac{i}{2}\vec{\tau}\, \hat \Lambda \,\vec{\tau} +i \hat \mu(\vec{\tau}) ] \psi_a\left(-\frac{1}{2} \vec{\tau} \, \hat L \, \vec{\tau}\right) \, , \nonumber
\end{align}
where $\psi_a(z)$ was given in \cref{eq:characteristic}, and we introduced the function $M(\vec{\tau},\vec{\tau}')$ as in \cref{eq:M_def}. We omitted from \cref{eq:Q_n_muhat} a $\hat \Lambda-$independent prefactor coming from the Gaussian integration, which will in general depend on $n$. However, one can check that all the prefactors cancel out smoothly by including the Jacobian of the variable transformations (see later), and after taking the functional integral over the Gaussian fluctuations in \cref{app:gaussian_fluctuations}. We will thus avoid reporting these prefactors, so as to lighten the notation.
The saddle-point equation follows simply from \cref{eq:action_comp_muhat} as
\begin{equation}
\hat\mu(\vec{\tau}) = i \eta \frac{\int \dd{\vec{\tau}'} M(\vec{\tau},\vec{\tau}') \exp[-\frac{i}{2}\vec{\tau}'\, \hat \Lambda \,\vec{\tau}' +i \hat \mu(\vec{\tau}') ] \psi_a\left(-\frac{1}{2} \vec{\tau}' \, \hat L \, \vec{\tau}'\right) }{ \int \dd{\vec{\tau}'} \exp[-\frac{i}{2}\vec{\tau}'\, \hat \Lambda \,\vec{\tau}' +i \hat \mu(\vec{\tau}') ] \psi_a\left(-\frac{1}{2} \vec{\tau}' \, \hat L \, \vec{\tau}'\right) } \, ,
\label{eq:saddle_comp_muhat}
\end{equation}
which is analogous to \cref{eq:saddle_tot}. In the following, we will look for an explicit rotationally-invariant solution: to this end, it is useful to introduce the new variable $\varphi(\vec{\tau})$ defined via
\begin{equation}
\hat\mu(\vec{\tau}) = i \int \dd{\vec{\tau}'} M(\vec{\tau},\vec{\tau}')\varphi(\vec{\tau}') \, .
\end{equation}
Changing variables from $\hat \mu$ to $\varphi$ in \cref{eq:Q_n_muhat} leads to the expression reported in \cref{eq:action_comp_varphi}.
\subsection{Gaussian fluctuations around the saddle-point}
\label{app:gaussian_fluctuations}
In order to go beyond the saddle-point approximation, we introduce the fluctuation $\phi (\vec{\tau})$ around the saddle-point $\varphi_0 (\vec{\tau})$ in the form
\begin{equation}
\varphi(\vec{\tau}) = \varphi_0(\vec{\tau}) + \phi (\vec{\tau}) \, .
\end{equation}
Calling for brevity $\cor{S}_{n_\pm}[\varphi; \hat \Lambda]\equiv \cor{S}[\varphi]$, we then have up to $\order{N^{-2}}$
\begin{equation}
Q_{[\alpha,\beta]}(n_\pm) \propto e^{N\cor{S}[\varphi_0]} \int \cor{D}(i\phi) \exp{\frac{N}{2} \int \dd{\vec{\tau}_1} \dd{\vec{\tau}_2} \phi(\vec{\tau}_1) \eval{\frac{\delta^2 \cor{S}[\varphi]}{\delta \varphi(\vec{\tau}_1) \delta \varphi(\vec{\tau}_2) }}_{\varphi=\varphi_0} \phi(\vec{\tau}_2) } \, ,
\label{eq:gaussian_fluc}
\end{equation}
and one can check that we can express
\begin{equation}
\eval{\frac{\delta^2 \cor{S}[\varphi]}{\delta \varphi(\vec{\tau}_1) \delta \varphi(\vec{\tau}_2) }}_{\varphi=\varphi_0} = \frac{1}{\eta} M(\vec{\tau}_1,\vec{\tau}_2) \left[ \mathbb{1}(\vec{\tau}_1,\vec{\tau}_2) + T (\vec{\tau}_1,\vec{\tau}_2) \right]
\label{eq:fluctuation_matrix}
\end{equation}
in terms of the functions $M$ and $T$ given in \cref{eq:M_def,eq:T_def}, respectively.
Computing the Gaussian integral in \cref{eq:gaussian_fluc} we thus find
\begin{equation}
Q_{[\alpha,\beta]}(n_\pm) = \exp{N\cor{S}[\varphi_0]-\frac12 \ln \det (\mathbb{1}+T)}+\order{1/N^2} \, .
\end{equation}
Expanding the logarithm in series as
\begin{equation}
\ln(1+x) = -\sum_{k=1}^\infty \frac{(-x)^k}{k}
\end{equation}
we finally get \cref{eq:Q_n_gaussian}, where the trace and matrix operations are intended over the replica vectors as
\begin{equation}
\Tr T = \int \dd{\vec{\tau}} \, T(\vec{\tau},-\vec{\tau}) \, , \;\;\;\;\;\;\;\;\;\;
T^2(\vec{\tau}_1,\vec{\tau}_2) = \int \dd{\vec{\tau}} \,T(\vec{\tau}_1,\vec{\tau})T(\vec{\tau},\vec{\tau}_2) \, .
\label{eq:trace}
\end{equation}
We may try to specialize \cref{eq:Q_n_gaussian} to the rotationally invariant Ansatz in \cref{eq:ansatz_new}. The fluctuation matrix in \cref{eq:fluctuation_matrix} becomes
\begin{equation}
T (\vec{\tau}_1,\vec{\tau}_2) = \varphi_0(\vec{\tau}_1) \left[ \left( \vec{\tau}_1 \, \hat L \, \vec{\tau}_2 \right)^2 - \left( \vec{\tau}_2 \, \hat K \, \vec{\tau}_2 \right) \right] \, ,
\end{equation}
and by using Wick's theorem together with \cref{eq:property} we obtain, for instance,
\begin{equation}
\Tr T = \int \dd{\vec{\tau}} T (\vec{\tau},\vec{\tau}) = \eta \left[ \sum_i \left( 2K_{ii}^{(2)}-K_{ii}^2 \right) + \sum_{ij}L_{ii}L_{jj} K_{ij}^{(2)}\right] \, ,
\end{equation}
where we introduced the matrix
\begin{equation}
K_{ij}^{(2)} \equiv \int \dd{a} p_a(a) \left( \frac{\hat C}{1-ia\hat L \hat C} \right)_{ii} \left( \frac{\hat C}{1-ia\hat L \hat C} \right)_{jj} \, .
\label{eq:gaussian_fluc_first_term}
\end{equation}
We recognize in the last expression a generalization of the resolvent (see \cref{eq:resolvent}) which encodes higher order correlations. The next terms $\Tr T^k$ with $k>1$ in the series of \cref{eq:Q_n_gaussian} will involve some matrices $K_{ij}^{(k+1)}$ with increasingly higher order correlations, which are nontrivial to compute in general. However, it is straightforward to show that $\Tr T^k = \order{\eta^k}$, so that when $\eta$ is small the series in \cref{eq:Q_n_gaussian} is dominated by its first few terms. To the best of our efforts, it has not been possible to resum the whole series in \cref{eq:Q_n_gaussian}, as it happens instead in the pure GOE case -- see \cref{app:GOE} and \ccite{Metz_2017}.
Specializing to the Cauchy distribution in \cref{eq:lorentzian} with $\mu=0$, one finds for instance that $K_{ii}$ is given in \cref{eq:K_cauchy} in terms of the matrix elements of $\hat C$, while using complex integration in \cref{eq:gaussian_fluc_first_term} yields
\begin{equation}
K_{ij}^{(2)} =
\begin{cases}
\dfrac{C_{ii}C_{jj}}{(\omega C_{ii}-1)(\omega C_{jj}-1)} \, , & \qquad\quad L_{ii}\cdot L_{jj} > 0 \, , \\
\dfrac{C_{ii}C_{jj}}{(\omega C_{ii}-1)(\omega C_{jj}-1)} \left[ 1- \dfrac{2\omega C_{ii}C_{jj}}{ C_{ii}+ C_{jj}} \right] \, , & \qquad\quad L_{ii}\cdot L_{jj} < 0 \, .
\end{cases}
\end{equation}
\section{Level compressibility in the pure GOE case}
\label{app:GOE}
In this Appendix we recover the results of \ccite{Metz_2017} concerning the level compressibility for a GOE matrix. This allows us to inspect the similarities and the differences with respect to the GRP case analyzed in this manuscript. In the final section \ref{app_chi_GOE}, we will repeat the derivation of $\chi_\T{GOE}(E)$ using more standard techniques in order to address the low-energy region $E\ll \delta_N$ (see \cref{eq:mean_level_spacing} and \cref{par:low_energy}).
The pure GOE case can be formally obtained from \cref{eq:hamiltonian} by letting the distribution $p_a(a)$ of the diagonal elements of the matrix $A$ tend to a delta function, so that $\psi_a(z) \to 1$.
The calculation then becomes analogous to that reported in \ccite{Metz_2017}, whose main steps we detail here for completeness. By replacing the replico-symmetric Ansatz of \cref{eq:ansatz} into the saddle point equation \eqref{eq:saddle_point_comp} one first obtains $\cor{N} = \eta / Z_\varphi $, where $Z_\varphi$ is given in \cref{eq:Z_varphi}. By using Gaussian integration one can then show that
\begin{equation}
\int \dd{\vec{\tau}'} M(\vec{\tau},\vec{\tau}') \varphi_0(\vec{\tau}') = \eta \vec{\tau} \, \hat C \, \vec{\tau} \, ,
\label{eq:property_GOE}
\end{equation}
and thus the remaining free parameters in \cref{eq:ansatz} can be determined by solving the set of four self-consistency equations which follow from \cref{eq:saddle_point_comp} as
\begin{equation}
\hat C^{-1} = 2\eta \, \hat C +i \, \hat \Lambda.
\label{eq:four_set_GOE}
\end{equation}
Note that this can be recovered from \cref{eq:K_def,eq:four_set_p(a)} by using the fact that the resolvent corresponding to $p_a(a)=\delta(a)$ is $\cor{G}_a(z) = 1/z$.
\subsection{Action and fluctuations around the saddle-point}
Both the action and its Gaussian fluctuation matrix $T$ can now be computed in correspondence of the saddle-point solution in \cref{eq:ansatz}. By using the definition of the action given in \cref{eq:action_comp_varphi} together with the saddle-point equation \eqref{eq:saddle_point_comp} and the property in \cref{eq:property_GOE}, one can easily deduce
\begin{align}
\cor{S}_{n_\pm}[\varphi_0 ; \hat \Lambda]= &\; \frac{\eta}{2}\left[ n_+ \left( \Delta_\alpha^{2} + \bar{\Delta}_\beta^{2}\right) +n_- \left( \Delta_\beta^{2} + \bar{\Delta}_\alpha^{2} \right) \right] + \frac12 (n_+ + n_-) \ln(2\pi) \nonumber\\
& + \frac12 n_+ \ln(\Delta_\alpha \bar{\Delta}_\beta) + \frac12 n_- \ln(\Delta_\beta \bar{\Delta}_\alpha) \, .
\label{eq:S_saddlepoint_first}
\end{align}
The computation of $\Tr T^k$ in \cref{eq:Q_n_gaussian} requires more work. First, we rewrite in correspondence of the Ansatz in \cref{eq:ansatz}
\begin{equation}
T (\vec{\tau}_1,\vec{\tau}_2) = \varphi_0(\vec{\tau}_1) \left[ \left( \vec{\tau}_1 \, \hat L \, \vec{\tau}_2 \right)^2 - \left( \vec{\tau}_2 \, \hat C \, \vec{\tau}_2 \right) \right] \, ,
\end{equation}
where we have used the definition of the function $M(\vec{\tau}_1,\vec{\tau}_2)$ in \cref{eq:M_def} and the property in \cref{eq:property_GOE}. The first few powers of $T$ can then be computed by applying Wick's theorem: by introducing the notation
\begin{equation}
\expval{\bullet}_1 \equiv \frac{1}{\eta} \int \dd{\vec{\tau}_1} (\bullet) \varphi_0(\vec{\tau}_1) \, ,
\end{equation}
it is sufficient to note that
\begin{equation}
\expval{\tau_{1i}\tau_{1j}}_1 = C_{ij} \, ,
\end{equation}
so that more complicated averages can be handled as
\begin{equation}
\expval{\tau_{1i}\tau_{1j}\tau_{1k}\tau_{1l}}_1 = C_{ij} C_{kl} + C_{ik} C_{jl}+C_{il} C_{jk} \, .
\end{equation}
Upon noting that $\hat L^2 = \mathbb{1}_n$, one can then prove by induction the relation
\begin{equation}
T^{k+1} (\vec{\tau}_1,\vec{\tau}_2) = (2\eta)^k \varphi_0(\vec{\tau}_1) \left[ \left( \vec{\tau}_1 \, \hat L^{k+1} \, \hat C^k \, \vec{\tau}_2 \right)^2 - \left( \vec{\tau}_2 \, \hat C^{2k+1} \, \vec{\tau}_2 \right) \right] \, .
\end{equation}
Using \cref{eq:trace} now yields
\begin{equation}
\Tr T^k = 2^{k-1}\eta^k \left[ \left( \sum_i L_{ii}^k C_{ii}^k \right)^2 + \sum_i C_{ii}^{2k} \right] \, ,
\end{equation}
and inserting the definition of the matrices $\hat L$ and $\hat C$ given in \cref{eq:lambda_L_def,eq:hat_C_def} gives
\begin{align}
\Tr T^k = &
\;
2^{k-1}\eta^k \Big\lbrace \left[ n_+\left( \Delta_\alpha^{k} + (-\bar{\Delta}_\beta)^{k} \right) + n_- \left( \Delta_\beta^{k} + (-\bar{\Delta}_\alpha)^{k}\right) \right]^2 \nonumber\\
& + n_+ \left( \Delta_\alpha^{2k} + \bar{\Delta}_\beta^{2k} \right) + n_- \left( \Delta_\beta^{2k} + \bar{\Delta}_\alpha^{2k} \right) \Big\rbrace \, .
\label{eq:Tk_first}
\end{align}
Taking the limit $n_\pm \to \pm is/\pi $ in \cref{eq:S_saddlepoint_first,eq:Tk_first} then results in
\begin{align}
\cor{S}_{\pm \frac{is}{\pi}}[\varphi_0 ; \hat \Lambda]=& \frac{is}{2\pi}\Bigg[ \eta \left( \Delta_\alpha^{2} + \bar{\Delta}_\beta^{2} - \Delta_\beta^{2} - \bar{\Delta}_\alpha^{2} \right) + \ln(\frac{\Delta_\alpha \bar{\Delta}_\beta}{\Delta_\beta \bar{\Delta}_\alpha}) \Bigg]\, , \label{eq:S_saddlepoint} \\
\eval{\Tr T^k}_{\varphi=\varphi_0} \!\!\!\!\!\!\! = & \;
2^{k-1}\eta^k \Big\lbrace \frac{is}{\pi} \left( \Delta_\alpha^{2k} + \bar{\Delta}_\beta^{2k} - \Delta_\beta^{2k} - \bar{\Delta}_\alpha^{2k} \right) - \frac{s^2}{\pi^2} \left[ \Delta_\alpha^{k} + (-\bar{\Delta}_\beta)^{k} - \Delta_\beta^{k} - (-\bar{\Delta}_\alpha)^{k} \right]^2 \Big\rbrace \, .
\label{eq:Tk}
\end{align}
Comparing with the definitions of the cumulant generating function and the cumulants in \cref{eq:cgf,eq:cumulants}, respectively, we can finally identify
\begin{align}
\frac{\kappa_1}{N}=& -\frac{i}{2\pi}\Bigg[ \eta \left( \Delta_\alpha^{2} + \bar{\Delta}_\beta^{2} - \Delta_\beta^{2} - \bar{\Delta}_\alpha^{2} \right) + \ln(\frac{\Delta_\alpha \bar{\Delta}_\beta}{\Delta_\beta \bar{\Delta}_\alpha}) \Bigg] \nonumber\\
&- \frac{i}{4\pi N } \sum_{k=1}^\infty \frac{(-2\eta)^k}{k} \left( \Delta_\alpha^{2k} + \bar{\Delta}_\beta^{2k} - \Delta_\beta^{2k} - \bar{\Delta}_\alpha^{2k} \right) \, , \label{eq:k1} \\
\frac{\kappa_2}{N} =& - \frac{i}{2\pi^2 N } \sum_{k=1}^\infty \frac{(-2\eta)^k}{k}
\left[ \Delta_\alpha^{k} + (-\bar{\Delta}_\beta)^{k} - \Delta_\beta^{k} - (-\bar{\Delta}_\alpha)^{k} \right]^2 \, . \label{eq:k2}
\end{align}
In the next Section we will specialize these results to the case in which the interval $[\alpha,\beta]$ is chosen symmetric.
\subsection{Case of a symmetric interval}
We consider here the case in which $\alpha=-E$ and $\beta=E$. With this choice, solving \cref{eq:four_set_GOE} gives\footnote{The angle $\theta$ should be compared with $\theta_L$ in Ref. \cite{Metz_2017}.}
\begin{equation}
\eval{\Delta_\alpha}_{a=-E} = \frac{1}{4\eta} \left[ -(\varepsilon-iE) \pm \sqrt{8\eta+\left(\varepsilon -iE\right)^2 } \right] \equiv r e^{i \theta} \, ,
\end{equation}
where we choose the positive branch of the square root so that $\Re \Delta_\alpha \geq 0$ for any positive $\eta$ (recall that $\Delta_\alpha$ represents the variance of a Gaussian distribution, see \cref{eq:ansatz}). Similarly, from \cref{eq:four_set_GOE} one finds for the entries of $\hat C$ the same symmetries as in \cref{eq:symmetries}. The first two cumulants in \cref{eq:k1,eq:k2} are then found to yield
\begin{align}
\frac{\kappa_1}{N} &= \frac{x}{\pi} \sin 2\theta +\frac{2\theta}{\pi} +\frac{i}{2\pi N} \ln( \frac{1+x e^{2i\theta}}{1+x e^{-2i\theta}} )\, , \label{eq:k1_GOE} \\
\frac{\kappa_2}{N} &= \frac{1}{\pi^2 N} \ln \left[ 1+ \left( \frac{2x \sin2 \theta}{1-x^2} \right)^2 \right]\, , \label{eq:k2_GOE}
\end{align}
where we called $x\equiv 2\eta r^2 $. As a first check, one can easily verify that both $\kappa_1$, $\kappa_2\to 0$ in the limit of a vanishing interval $E \to 0$.
By choosing $2\eta = 1$, we obtain in the $N\to \infty$ limit an eigenvalue spectrum distributed within the interval $[-2,2]$. Sending $\varepsilon \to 0^+$ as prescribed by \cref{eq:cgf}, one can check that $r \to 1$ (hence $x\to 1$), while
\begin{equation}
\theta \xrightarrow[\varepsilon \to 0^+]{} \theta_0 \equiv \arctan( \frac{E}{\sqrt{4-E^2}} ) \in \left[-\frac{\pi}{2},\frac{\pi}{2}\right] \, .
\end{equation}
This concludes the calculation of $\kappa_1$ (see \cref{eq:k1_GOE}), which includes both the leading order term and its $\order{N^0}$ correction (note that the latter is actually real-valued). However, the second cumulant $\kappa_2$ is seen to diverge in the limit $\varepsilon\to 0^+$; the problem is addressed in \ccite{Metz_2017} by introducing a $N$-dependent regularization of the infinite sum which appears in \cref{eq:k2}.
Nonetheless, we have shown that such infinite sum (and hence $\kappa_2$ itself in \cref{eq:k2_GOE}) does \textit{not} diverge for any finite value of $\varepsilon$. This hints at the well-known fact that the limit $\varepsilon \to 0^+$ and that for $N\to \infty$ in \cref{eq:cgf} are not interchangeable. It is then useful to expand for small $\varepsilon$
\begin{equation}
x = r^2\simeq 1-\frac{2\varepsilon}{\sqrt{4-E^2}} \, ,
\end{equation}
so as to rewrite
\begin{equation}
\kappa_2 \simeq \frac{2}{\pi^2} \ln \left[ \frac{\sqrt{4-E^2} \sin 2 \theta}{2\varepsilon} \right] \, .
\end{equation}
In order to recover the leading order result $\kappa_2 \sim \ln N$ found in previous literature \cite{Dyson_62,Cavagna_2000,Majumdar_2009,Majumdar_2011,Castillo_2014}, one has to assume some type of functional relation $\varepsilon = \varepsilon(N)$, so that the limit $\varepsilon \to 0^+$ is taken by controlling the product $\varepsilon N$ \cite{Cavagna_2000}. This goes however beyond the scope of the present paper.
\subsection{Derivation of the scaling function $\chi_{\rm GOE}(y)$}
\label{app_chi_GOE}
The prediction for $\chi(E)$ which we have pursued in the previous Section using the replica method is in any case expected to fail if $E$ lies below the mean level spacing, i.e., $E\ll \delta_N$ (as discussed in \cref{par:low_energy}).
In this Appendix, we thus derive in this limit an explicit expression for the level compressibility for a pure $N \times N$ GOE matrix (such as the matrix $B$ in Eq. \eqref{eq:hamiltonian}) using an alternative and more standard technique. We denote by $\rho_B(\lambda)$ (defined as in Eq. \eqref{eq:def_density}) the average density of eigenvalues normalized to unity. In the large $N$ limit, $\rho_B(\lambda)$ has a finite support over $[-\sqrt{2N}, \sqrt{2N}]$ and it is given by the standard Wigner semi-circle. We focus here on the number of eigenvalues $I_N[-E,E]$ in a symmetric interval $[-E,E]$ in the bulk of the spectrum, i.e., we choose $E$ of the order of the mean inter-particle spacing $[N \rho_B(0)]^{-1} = O(N^{-1/2})$.
The mean number $\langle I_N[-E,E]\rangle$ is easily obtained as
\begin{eqnarray} \label{av_app_GOE}
\langle I_N[-E,E]\rangle = N \int_{-E}^E \rho_B(\lambda) \, \dd \lambda \approx 2 N \rho_B(0) E = 2 \tilde E \, , \quad\qquad \tilde E = N \rho_B(0) E \;.
\end{eqnarray}
The variance of $I_N[-E,E]$ has been studied since the pioneering works of Dyson and Mehta \cite{Dyson_63_IV}. However, an explicit expression for it, valid for any value of $E$ in the bulk, seems hard to find in the literature. This observable was recently revisited in the context of full counting statistics of interacting fermions in Ref. \cite{smith2021full}, which provides a useful starting point, namely [see Eqs.~(31)-(37) therein]
\begin{align}
\langle (I_N[-E,E])^2\rangle - \langle I_N[-E,E]\rangle^2 &= - N \rho_B(0)^2 \int_{-E}^E \dd x \left(\int_{-\infty}^{-E} \dd y + \int_{E}^{+ \infty} \dd y \right) C(x,y) \, , \nonumber \\
C(x,y) &= - N \rho_B(0)^2 Y_{21}(N \rho_B(0)|x-y|) \;,
\label{var_app_GOE}
\end{align}
where the ``cluster'' function $Y_{21}(r)$ is given by \cite{pandey1979statistical}
\begin{eqnarray} \label{cluster}
Y_{21}(r) = \left(\frac{\sin(\pi r)}{\pi r} \right)^2 - \left(\frac{{\rm Si}(\pi r)}{\pi} - \frac{1}{2} \right)\left(\frac{\pi r \cos(\pi r) - \sin{(\pi r)}}{\pi r^2} \right) \, ,
\end{eqnarray}
with ${\rm Si}(z) = \int_0^z \sin(t)/t\, \dd t$ denoting the sine-integral function. By inserting this explicit expression \eqref{cluster} in \cref{var_app_GOE} and performing explicitly the integrals over $x$ and $y$, one obtains for the level compressibility \begin{eqnarray} \label{com_app_GOE}
\chi(E) = \frac{\langle (I_N[-E,E])^2\rangle - \langle I_N[-E,E]\rangle^2}{\langle I_N[-E,E]\rangle} \sim \chi_{\rm GOE}(y = N \rho_B(0)\,E) \, ,
\end{eqnarray}
where the function $\chi_{\rm GOE}(y)$ is given by
\begin{align}
\label{chi_GOE_final}
\chi_{\rm GOE}(y) = & \frac{1}{2 \pi^2 y}
\big\lbrace [\text{Si}(2 \pi y)]^2 -2\, \text{Ci}(4 \pi y) -\pi \, \text{Si}(2 \pi y)\\
&+2 \left[-4
\pi y \, \text{Si}(4 \pi y)+2 \pi ^2 y+\log (4 \pi y)-\cos (4 \pi y)+\gamma_E
+1\right]\big\rbrace \, , \nonumber
\end{align}
where ${\rm Ci}(z) = - \int_z^\infty \cos(t)/t\, \dd t$ is the cosine integral function and $\gamma_E = 0.577216\ldots$ is the Euler-gamma constant. Its asymptotic behaviors are given by
\begin{eqnarray} \label{chi_GOE_asympt}
\chi_{\rm GOE}(y) =
\begin{dcases}
&1 - 2y + O(y^2) \, , \quad y \to 0 \\
&\frac{\ln y}{\pi^2 y} + O(1/y) \, , \quad y \to \infty \;.
\end{dcases}
\end{eqnarray}
\end{appendix}
|
\section{Introduction}
The notion of 2-cocycle twist was introduced by Doi and
Takeuchi \cite{Doi93,DT94} as a dual version of the Drinfeld twist. This notion is given by a deformation of the algebraic structure of a Hopf algebra $H$ by a 2-cocycle $\sigma$ on $H$. It is well-known that $H$ and its 2-cocycle twist $H^\sigma$ are Morita--Takeuchi equivalent, that is, their corresponding comodule categories are tensor equivalent. Under this equivalence of tensor categories, the induced 2-cocycle twist of a comodule algebra over $H$ and its properties were studied earlier by various authors (see e.g., \cite{CKWZ, CS1, CS2,Montgomery}). In this paper, we take a categorical point of view to consider
superpotentials as morphisms in the comodule category over $H$, and to define their twistings as the corresponding morphisms in the equivalent comodule
category over $H^\sigma$.
Another approach of twisting a graded algebra $A$ was observed by Artin, Tate, and Van den Bergh in \cite{ATV1991} and Zhang in \cite{Zhang1996} as a deformation of the original graded product of $A$ by a graded automorphism of $A$. Two graded algebras have equivalent graded module categories if and only if they are twists of each other \cite[Theorem 3.5]{Zhang1996}. The twist by a graded automorphism later became known as a Zhang twist.
In \cite[Theorems E and F]{HNUVVW21}, the authors of this paper showed that Zhang twists of quadratic algebras yield Morita--Takeuchi equivalent universal quantum groups. Let $A$ be a graded algebra and $\underline{\rm aut}^r(A)$ be Manin's universal quantum group associated to $A$, the Hopf algebra which universally right coacts on $A$ preserving the grading (see \Cref{def:ManinU} for details). Here, we compare a cocycle twist of Manin's universal quantum group with twisting the underlying algebra $A$ as a comodule algebra of $\underline{\rm{aut}}^r(A)$.
\begin{introdethm}[\Cref{thm:Manintwist}]
Let $A$ be any $\mathbb Z$-graded locally finite algebra and $\underline{\rm aut}^r(A)$ be its associated right universal quantum group. For any left $2$-cocycle $\sigma$ on $\underline{\rm aut}^r(A)$, we have
\[\underline{\rm aut}^r(A_{\sigma^{-1}})~\cong~ \underline{\rm aut}^r(A)^\sigma\]
as Hopf algebras. Here, $A_{\sigma^{-1}}$ is the 2-cocycle twist of $A$ by $\sigma^{-1}$ and $\underline{\rm aut}^r(A)^\sigma$ is the 2-cocycle twist of $\underline{\rm aut}^r(A)$ by $\sigma$.
As a consequence, $\underline{\rm aut}^r(A_{\sigma^{-1}})$ and $\underline{\rm aut}^r(A)$ are Morita--Takeuchi equivalent.
\end{introdethm}
When a graded algebra $A$ is viewed as a comodule algebra over $\underline{\rm aut}^r(A)$, we now show that a Zhang twist of $A$ can be
realized as a 2-cocycle twist, thus, proving a comodule algebra analogue of our previous result \cite[Theorem 2.2.6]{HNUVVW21}.
\begin{introdethm}[\Cref{lem:2cocycleGrade}, \Cref{lemma:2coZhang}]\label{introthm:A}
Let $m\geq 2$ be an integer, and $A$ be a connected graded algebra generated in degree one with $m$-homogeneous relations. We view $A$ as a graded comodule algebra over $\underline{\rm aut}^r(A)$. For any graded automorphism $\phi$ of $A$, there is a left 2-cocycle $\sigma$ on $\underline{\rm aut}^r(A)$ (defined in \eqref{eq:2cocycleA}), such that
\[A^{\phi}~\cong~A_{\sigma^{-1}}\]
as graded comodule algebras over $\underline{\rm aut}^r(A)^\sigma$. Here $A^\phi$ denotes the (right) Zhang-twist of $A$ by $\phi$.
\end{introdethm}
Next, we examine twists of superpotential algebras and their related universal quantum groups. Preregular forms, which are dual to superpotentials, were used by Dubois-Violette in \cite{Dubois-Violette2005, DVM} to construct finitely generated connected graded Gorenstein algebras with relations given by its higher derivations. In \cite{Mori-Smith2016}, Mori and Smith investigated twisting of superpotentials in relation to Zhang twists of their associated superpotential algebras. We extend their notion of twisting to twisting by an arbitrary 2-cocycle of a Hopf algebra $H$ that coacts on the superpotential algebra while preserving its grading.
\begin{introdethm}[\Cref{prop:2.2.5}]
Let $1\leq N\leq m$ be integers and $H$ be a Hopf algebra. Let $V$ be a finite-dimensional right $H$-comodule, $j:W \hookrightarrow V^{\otimes m}$ be an $H$-subcomodule and $\sigma$ be a left 2-cocycle on $H$. Then there is an isomorphism \[A(W_{\sigma},N)~\cong~A(W,N)_{\sigma^{-1}}\] of $H^{\sigma}$-comodule algebras, where $A(W,N)$ is the derivation-quotient algebra given in \Cref{defn:derivation}, and $W_\sigma$ is from \Cref{def:twistform}.
\end{introdethm}
We also investigate the Artin--Schelter (AS) regularity of a comodule algebra that is connected graded under 2-cocycle twists. AS-regular algebras, introduced in \cite{ArtinSchelter1987}, are a foundational object of study in noncommutative projective algebraic geometry \cite{ArtinTateVanDenBergh1990, ATV1991, ArtinZhang1994}. We use the fact that a connected graded algebra $A$ is AS-regular if and only if its Ext-algebra $\Ext^\ast_A(\Bbbk,\Bbbk)$ is Frobenius. This result was first proved for any Koszul algebra by Smith \cite{Smith1994} and was later proved in its full generality by Lu et.~ al.~ \cite{LPWZ}.
\begin{introdethm}[\Cref{lem:twistkoszul}, \Cref{thm:AS}]
Let $H$ be a Hopf algebra and $\sigma$ be a 2-cocycle on $H$. For any connected graded comodule algebra $A$ over $H$, $A$ is $N$-Koszul if and only if $A_{\sigma^{-1}}$ is. In this case, $A$ is AS-regular of dimension $d$ if and only if $A_{\sigma^{-1}}$ is AS-regular of the same dimension.
\end{introdethm}
Previous results along the same lines have been obtained by Chan et.~ al.~ \cite{CKWZ} for $H$-module algebras when $H$ is semisimple in the context of noncommutative invariant theory. Recently, Chirvasitu and Smith \cite{CS2} proved a similar statement for any $H$-comodule algebra when $H$ is finite-dimensional using torsors, where they constructed ``exotic" elliptic algebras by taking 2-cocycle twists of the $4$-dimensional Sklyanin algebras coming from the Klein four-group. The key ingredient in our argument here is the fact that Frobenius algebras can be defined in any rigid monoidal category in terms of nondegenerate pairings. It allows us to prove that Frobenius algebras are invariant under the 2-cocycle twists (see \Cref{lem:twistF}). The only technical assumption here is that the connected graded comodule algebra $A$ is $N$-Koszul. In this case, we can show that their Ext-algebras are 2-cocycle twists of each other if the original comodule algebras are as well. Though $N$-Koszul AS-regular algebras contribute to a large portion of the family of AS-regular algebras, we conjecture the same result to hold without the $N$-Koszul assumption.
In \cite[Theorem 7.2.3]{vdb2017}, Raedschelders and Van den Bergh proved that Manin’s universal quantum groups associated to Koszul AS-regular algebras of the same dimension are always Morita--Takeuchi equivalent. Viewing any two Koszul AS-regular algebras of the same dimension and same Hilbert series as comodule algebras over their respective universal quantum groups, we show that they are always 2-cocycle twists of each other.
\begin{introdethm}[\Cref{Cor:TwistAS}]
Let $A,B$ be two Koszul AS-regular algebras of the same dimension and same Hilbert series. Then there is a left 2-cocycle $\sigma$ on $\underline{\rm aut}^r(A)$ such that $B\cong A_{\sigma^{-1}}$ as graded comodule algebras over $\underline{\rm aut}^r(B)$.
\end{introdethm}
The paper is organized as follows. In \Cref{sec:Manin twists}, we investigate how Manin's universal quantum group associated to a graded algebra $A$ behaves under 2-cocycle twists. When $A$ is a connected graded algebra generated in degree one with $m$-homogeneous relations, $m \geq 2$, we relate a Zhang twist of $A$ by any graded automorphism to a 2-cocycle twist of $A$ by a 2-cocycle on $\underline{\rm aut}^r(A)$. In \Cref{sec:superpotential}, we study 2-cocycle twists of superpotential algebras by introducing 2-cocycle twists of superpotentials. Finally, in \Cref{sec:properties}, we observe some ring-theoretic properties invariant under 2-cocycle twists, such as $N$-Koszulness, Frobenius property, and AS-regularity.
\subsection*{Acknowledgements}
Some results in this paper were formulated at the Structured Quartet Research Ensembles (SQuaREs) program in March 2022. The authors thank the American Institute of Mathematics for their hospitality and support of the SQuaREs program. Nguyen was partially supported by the Naval Academy Research Council and NSF grant DMS-2201146. Vashaw was partially supported by NSF Postdoctoral Fellowship DMS-2103272. Wang was partially supported by Simons collaboration grant \#688403 and AFOSR grant FA9550-22-1-0272.
\subsection*{Conventions}
Throughout the paper, let $\Bbbk$ be a base field with $\otimes$ taken over $\Bbbk$ unless stated otherwise. All categories are $\Bbbk$-linear and all algebras are associative over $\Bbbk$. We use the Sweedler notation for the coproduct in a coalgebra $B$: for any $h \in B$, $\Delta(h) = \sum h_1 \otimes h_2$. When a bialgebra $B$ right (resp.~ left) coacts on an algebra $A$, we denote the \emph{right} coaction $\rho: A \to A \otimes B$ by $a \mapsto \sum a_0 \otimes a_1 $ (resp.~ the \emph{left} coaction $\rho: A \to B \otimes A$ by $a \mapsto \sum a_{-1} \otimes a_0$). The category of all (resp.~ finite-dimensional) right $B$-comodules is denoted by ${\rm comod}(B)$ (resp.~${\rm comod}_{\rm fd}(B))$.
\section{Twisting of Manin's universal quantum groups}
\label{sec:Manin twists}
In this section, we review the generalization of Manin's universal quantum group $\underline{\rm aut}(A)$ of a quadratic algebra, as described in \cite{Manin2018}, to any locally finite $\mathbb Z$-graded algebra $A$. In \Cref{thm:Manintwist}, we investigate their twists by 2-cocycles.
\begin{defn}
\label{def:ManinU}
Let $A$ be any $\mathbb Z$-graded algebra.
\begin{enumerate}
\item We say a Hopf algebra $H$ \emph{left coacts on $A$ preserving the grading of $A$} via $\rho: A\to H\otimes A$ if each homogeneous component of $A$ is a left $H$-comodule via $\rho$ and $\rho$ is an algebra map. In this case, we say $A$ is a right graded comodule algebra over $H$.
\item The \emph{left universal quantum group $\underline{\rm aut}^l(A)$ associated to $A$} is the Hopf algebra that left coacts on $A$ preserving the grading of $A$ via $\rho: A\to \underline{\rm aut}^l(A)\otimes A$ satisfying the following universal property: If $H$ is any Hopf algebra that left coacts on $A$ preserving the grading of $A$ via $\tau: A\to H\otimes A$, then there is a unique Hopf algebra map $f: \underline{\rm aut}^l (A)\to H$ such that the following diagram
\begin{align}\label{def:aut}
\xymatrix{
A\ar[r]^-{\rho}\ar[dr]_-{\tau} & \underline{\rm aut}^l(A)\otimes A\ar[d]^-{f\otimes \id} \\
& H\otimes A
}
\end{align}
commutes. Similarly, we can define $\underline{\rm aut}^r(A)$ by using the universal right coaction on $A$ preserving the grading of $A$.
\end{enumerate}
\end{defn}
By \cite[Example 4.8(1)-(2)]{AGV}, we know $\underline{\rm aut}^l(A)$ always exists if $A$ is locally finite, namely, $\dim_\Bbbk A_i<\infty$ for all $i\in \mathbb Z$. Next we recall the 2-cocycle twist introduced by Doi and Takeuchi \cite{Doi93,DT94}.
\begin{defn}[{\cite[Definition 7.7.1]{Radford2012}}]
\label{defn:cocycle}
Let $B$ be a bialgebra over $\Bbbk$.
A \emph{left 2-cocycle} on $B$ is a convolution invertible bilinear map $\sigma: B \times B \rightarrow \Bbbk$ so that
\[
\sum \sigma(x_1, y_1)\, \sigma (x_2 y_2, z) = \sum \sigma(y_1,z_1)\, \sigma(x,y_2z_2),
\]
for all $x,y,z \in B$. Similarly, \emph{a right 2-cocycle} is a convolution invertible bilinear map $\sigma: B \times B \rightarrow \Bbbk$ satisfying
\[
\sum \sigma(x_2, y_2)\, \sigma (x_1 y_1, z) = \sum \sigma(y_2,z_2)\, \sigma(x,y_1z_1).
\]
For any 2-cocycle $\sigma$, we denote its convolution inverse by $\sigma^{-1}$.
\end{defn}
Throughout, for simplicity, we assume all 2-cocycles are normal, that is, $$\sigma(x,1) = \sigma(1,x) = \varepsilon(x),$$ for all $x \in B$. Note that if $\sigma$ is a left 2-cocycle for a Hopf algebra $H$, then $\sigma^{-1}$ is a right 2-cocycle for $H$ \cite[Lemma 7.7.2]{Radford2012}.
Given a left 2-cocycle $\sigma: B \times B \to \Bbbk$ for a bialgebra $B$, let $B^\sigma = B$ as a coalgebra, endowed with the original unit and deformed product
\[
x*_\sigma y:=\sum \sigma(x_1,y_1)\,x_2y_2\,\sigma^{-1}(x_3,y_3),
\]
for any $x,y\in B$. In fact, for a Hopf algebra $H$ with antipode $S$, $H^\sigma$ is again a Hopf algebra with deformed antipode $S^\sigma$ given in \cite[Theorem 1.6]{Doi93} and \cite[Proposition 7.7.3]{Radford2012}. We call $H^\sigma$ the \emph{2-cocycle twist of $H$} by $\sigma$. It is well-known that two Hopf algebras are 2-cocycle twists of each other if and only if there exists a bicleft object between them (see e.g., \cite{Sch1996}).
Moreover, suppose $A$ is a right $H$-comodule algebra via coaction $\rho: A \to A \otimes H$ with $a \mapsto \sum a_0 \otimes a_1$ for any $a \in A$. For any left 2-cocycle $\sigma$ on $H$, one can define a \emph{twisted algebra} $A_{\sigma^{-1}}$ of $A$ by $\sigma^{-1}$ as follows. As a vector space $A_{\sigma^{-1}}=A$ and the product of $A_{\sigma^{-1}}$ is defined to be
\[
a\cdot_\sigma b=\sum \sigma^{-1}(a_1,b_1)\,a_0b_0, \qquad \text{for all } a,b\in A.
\]
Note that $A_{\sigma^{-1}}$ becomes a right comodule algebra over $H^\sigma$. In fact, under the tensor equivalence
\[F: {\rm comod}(H) \xrightarrow{\sim} {\rm comod}(H^\sigma) \]
the $H$-comodule algebra $A$ is taken to the $H^\sigma$-comodule algebra \, $A_{\sigma^{-1}}$ (see \Cref{sec:superpotential} for details).
Similarly, when $A$ is a left $H$-comodule algebra, we can define the algebra $\!_\sigma A$. In this case, \,$\!_\sigma A$ is a left comodule algebra over $H^\sigma$.
On the other hand, if $A$ is a graded algebra and $\phi: A \to A$ is a graded automorphism, then the \emph{right Zhang twist of $A$}, denoted $A^{\phi}$, has the same graded vector space as $A$ with deformed multiplication \[a \cdot_{\phi} b=a\phi^{|a|}(b),\]
for any homogeneous elements $a,b\in A$, where $|a|$ denotes the degree of $a$. In \cite{HNUVVW21}, the following twisting conditions were given so that a Zhang twist of a $\mathbb Z$-graded bialgebra (resp. Hopf algebra) is again a bialgebra (resp. Hopf algebra). By \cite[Theorem 1.3.3]{HNUVVW21} a certain Zhang twist of a Hopf algebra $H$ satisfying these twisting conditions can be realized as a 2-cocycle twist $H^\sigma$, where the 2-cocycle $\sigma$ was explicitly described.
\begin{defn}
\label{defn:twisting conditions}
A bialgebra $(B,m,u,\Delta,\varepsilon)$ satisfies the \emph{twisting conditions} if
\begin{enumerate}
\item[(\textbf{T1})] as an algebra $B=\bigoplus_{n\in \mathbb{Z}} B_n$ is $\mathbb{Z}$-graded, and
\item[(\textbf{T2})] the comultiplication satisfies $\Delta(B_n)\subseteq B_n\otimes B_n$ for all $n\in \mathbb Z$.
\end{enumerate}
\end{defn}
As pointed out in \cite[Corollary 1.1.10]{HNUVVW21}, if $H=\bigoplus_{n\in \mathbb Z} H_n$ is a Hopf algebra that satisfies the twisting conditions, then $S(H_n)\subseteq H_{-n}$ for all $n\in \mathbb Z$.
An equivalent description of twisting conditions on $B$ was given by Bichon et.~ al.~ in \cite[Lemma 1.3]{Bichon-Neshveyev-Yamashita2016} including the existence of a cocentral homomorphism from $B$ to the group algebra $\Bbbk[\mathbb Z]$. Recall that a homomorphism of coalgebras $f: C\to D$ is said to be \emph{cocentral} if $\sum f(c_1)\otimes c_2=\sum f(c_2)\otimes c_1 \in D\otimes C$, for all $c \in C$.
\begin{lemma}
\label{TCautA}
Let $A=\bigoplus_{n\in \mathbb Z} A_n$ be a $\mathbb Z$-graded algebra that is locally finite, and $\underline{\rm aut}^r(A)$ be its associated right universal quantum group via the universal coaction $\rho: A\to A\otimes \underline{\rm aut}^r(A)$. Then $\underline{\rm aut}^r(A)=\bigoplus_{n\in \mathbb Z} \left(\underline{\rm aut}^r(A)\right)_n$ has a natural $\mathbb Z$-grading such that $\rho: A_n\to A_n\otimes\left(\underline{\rm aut}^r(A)\right)_n$, for each $n \in \mathbb Z$. Moreover, $\underline{\rm aut}^r(A)$ satisfies the twisting conditions. Similar results hold for $\underline{\rm aut}^l(A)$.
\end{lemma}
\begin{proof}
Recall that $\underline{\rm end}^r(A)$ is the universal bialgebra with right coaction $\rho$ on $A$ preserving the $\mathbb Z$-grading of $A$ which has the universal property satisfying a diagram similar to \eqref{def:aut}.
By an analogue of \cite[Proposition 2.1.3(1)]{HNUVVW21},
$\underline{\rm aut}^r(A)$ is the Hopf envelope of $\underline{\rm end}^r(A)$. Hence, by \cite[Lemma 1.1.9(1)]{HNUVVW21}, it suffices to prove the above statement for $\underline{\rm end}^r(A)$.
Now by \cite[Lemma 1.3]{Bichon-Neshveyev-Yamashita2016}, it is enough to show that there is a cocentral bialgebra map $f: \underline{\rm end}^r(A)\to \Bbbk[\mathbb Z]$. Note that since $A$ is $\mathbb Z$-graded, it follows that $A$ is a right $\Bbbk[\mathbb Z]$-comodule algebra via $a \mapsto a\otimes g^n$ for any $a\in A_n$ and $n \in \mathbb Z$, where $\mathbb Z=\langle g\rangle$. By the universal coaction $\rho: A\to A\otimes \underline{\rm end}^r(A)$, where $\rho(a)=\sum a_{0}\otimes a_{-1}$, we get a unique bialgebra map $f: \underline{\rm end}^r(A)\to \Bbbk[\mathbb Z]$ such that
\begin{align}\label{lem:eq211}
\sum a_0\otimes f(a_1)=a\otimes g^n, \quad \text{ for all } a\in A_n.
\end{align}
It remains to show that $f: \underline{\rm end}^r(A)\to \Bbbk[\mathbb Z]$ is cocentral. Denote by $V_n$ the cosupport of the left comodule $A_n$ in $\underline{\rm end}^r(A)$, that is, the smallest subspace $V_n$ of $\underline{\rm end}^r(A)$ such that $\rho(A_n)\subseteq A_n\otimes V_n$. It is straightforward to check that $V_n$ is a subcoalgebra of $\underline{\rm end}^r(A)$. Since $\underline{\rm end}^r(A)$ coacts on $A$ inner-faithfully, $\underline{\rm end}^r(A)$ is generated by $\{V_n\}_{n\in \mathbb Z}$ as an algebra. Therefore, it is enough to show that
\begin{align*}
\sum f(v_1)\otimes v_2=\sum f(v_2)\otimes v_1, \quad \text{ for all } v\in V_n.
\end{align*}
Using the fact that the $V_n$-comodule structure on $A_n$ satisfies \eqref{lem:eq211}, we have
$\sum v_1\otimes f(v_2)=v\otimes g^n$ and so $f(v)=g^n\varepsilon(v)$.
Hence
\begin{align*}
\sum f(v_1)\otimes v_2=\sum f(v_2)\otimes v_1=g^n\otimes v.
\end{align*}
It follows that $f$ is cocentral. A similar argument holds for $\underline{\rm end}^l(A)$ and $\underline{\rm aut}^l(A)$.
\end{proof}
The following result shows how Manin's universal quantum groups behave under 2-cocycle twists.
\begin{thm}
\label{thm:Manintwist}
Let $A$ be any $\mathbb Z$-graded locally finite algebra and $\underline{\rm aut}^r(A)$ be its associated right universal quantum group. For any left $2$-cocycle $\sigma$ on $\underline{\rm aut}^r(A)$, we have
\[\underline{\rm aut}^r(A_{\sigma^{-1}})~\cong~ \underline{\rm aut}^r(A)^\sigma\]
as Hopf algebras. As a consequence, $\underline{\rm aut}^r(A_{\sigma^{-1}})$ and $\underline{\rm aut}^r(A)$ are Morita--Takeuchi equivalent.
\end{thm}
In order to prove the above theorem, we first state some results, without proof, for 2-cocycle twists. Parts (3) and (4) appear in the dual context of Drinfeld twists in \cite[Lemma 2.7]{FKMW-Semisimple-reflection-Hopf-algebras-of-dim-16}.
\begin{lemma}
\label{rem:2cocycle}
Let $H$ and $K$ be two arbitrary Hopf algebras with a Hopf algebra map $\alpha: H\to K$. Let $\sigma$ be a left 2-cocycle on $K$ and $B$ be a right $K$-comodule algebra via $\nu: B\to B\otimes K$.
\begin{enumerate}
\item We have a left 2-cocycle on $H$, which is denoted by
\[\alpha^*\sigma:=\left( H\times H\xrightarrow{\alpha\times \alpha}K\times K\xrightarrow{\sigma} \Bbbk\right)
\]
with convolution inverse $\alpha^*\sigma^{-1}$.
\item The Hopf algebra map $\alpha: H\to K$ induces a Hopf algebra map $\alpha^\sigma:H^{\alpha^*\sigma} \to K^\sigma$ for the 2-cocycle twists of the corresponding Hopf algebras, where $\alpha=\alpha^{\sigma}$ as linear maps.
\item $\sigma^{-1}$ is a left 2-cocycle on $K^\sigma$ with $(K^\sigma)^{\sigma^{-1}}=K$. Similarly, $(\alpha^{\sigma})^*{\sigma^{-1}}=\alpha^*\sigma^{-1}$ is a left 2-cocycle on $H^{\alpha^*\sigma}$ with $(H^{\alpha^*\sigma})^{\alpha^*\sigma^{-1}}=H$.
In particular, we have $(\alpha^{\sigma})^{\sigma^{-1}}=\alpha: H\to K$.
\item $B_{\sigma^{-1}}$ is a right $K^\sigma$-comodule algebra via $\nu^{\sigma}: B_{\sigma^{-1}}\to B_{\sigma^{-1}}\otimes K^{\sigma}$, where $\nu^{\sigma}=\nu$ as linear maps. In particular, $(B_{\sigma^{-1}})_\sigma=B$ and $(\nu^{\sigma})^{\sigma^{-1}}=\nu$.
\item Suppose $B$ is also a right $H$-comodule algebra via $\tau: B\to B\otimes H$ such that the following
diagram \[
\xymatrix{
&B\ar[dl]_-{\tau}\ar[dr]^-{\nu} &\\
B\otimes H\ar[rr]^-{\id \otimes \alpha}&& B\otimes K
}
\]
commutes. Then $B_{\sigma^{-1}}$ is also a right $H^{\alpha^*\sigma}$-comodule algebra via $\tau^{\sigma}: B_{\sigma^{-1}}\to B_{\sigma^{-1}}\otimes H^{\alpha^*\sigma}$ satisfying $\nu^{\sigma}=(\id\otimes \alpha^{\sigma})\, \circ \, \tau^{\sigma}$, where $\tau^{\sigma}=\tau$ as linear maps.
\end{enumerate}
\end{lemma}
\begin{proof}[Proof of \Cref{thm:Manintwist}]
For a $\mathbb Z$-graded locally finite algebra $A$, we consider Manin's universal quantum groups $\underline{\rm aut}^r(A)$ and $\underline{\rm aut}^r(A_{\sigma^{-1}})$, for any left 2-cocycle $\sigma$ on $\underline{\rm aut}^r(A)$, together with universal Hopf coactions
\begin{align*}
\rho_A: A\to A\otimes \underline{\rm aut}^r(A) \quad \text{and} \quad \rho_{A_{\sigma^{-1}}}: A_{\sigma^{-1}}\to A_{\sigma^{-1}} \otimes \underline{\rm aut}^r(A_{\sigma^{-1}}),
\end{align*}
respectively. By \Cref{rem:2cocycle}(4), $A_{\sigma^{-1}}$ is also a right $\underline{\rm aut}^r(A)^\sigma$-comodule algebra via $\rho_A^{\sigma}: A_{\sigma^{-1}}\to A_{\sigma^{-1}}\otimes \underline{\rm aut}^r(A)^\sigma$. Thus the universal property of $\rho_{A_{\sigma^{-1}}}$ implies that there is a unique Hopf algebra map \begin{align}\label{thm211f}f:\underline{\rm aut}^r(A_{\sigma^{-1}})\to \underline{\rm aut}^r(A)^\sigma\end{align} such that the following diagram
\begin{equation}\label{dia:1}
\xymatrix{
&A_{\sigma^{-1}}\ar[dl]_-{\rho_{A_{\sigma^{-1}}}}\ar[dr]^-{\rho_A^\sigma} &\\
A_{\sigma^{-1}}\otimes \underline{\rm aut}^r(A_{\sigma^{-1}})\ar[rr]^-{\id \otimes f}&& A_{\sigma^{-1}}\otimes \underline{\rm aut}^r(A)^\sigma
}
\end{equation}
commutes. By \Cref{rem:2cocycle}(3), $\sigma^{-1}$ is a left 2-cocycle on $\underline{\rm aut}^r(A)^\sigma$ and so by \Cref{rem:2cocycle}(2)-(3), there is a Hopf algebra map
\begin{equation}
\label{eq:map1}
f^{\sigma^{-1}}:\underline{\rm aut}^r(A_{\sigma^{-1}})^{f^*\sigma^{-1}} \to \left(\underline{\rm aut}^r(A)^{\sigma}\right)^{\sigma^{-1}}=\underline{\rm aut}^r(A).
\end{equation}
Moreover, by \Cref{rem:2cocycle}(4) and \eqref{dia:1}, we have \[\rho^{\sigma^{-1}}_{A_{\sigma^{-1}}}: A=(A_{\sigma^{-1}})_{\sigma}\to A\otimes \underline{\rm aut}^r(A_{\sigma^{-1}})^{f^*\sigma^{-1}} \, \text{ and }\]
\[ \rho_A: A=(A_{\sigma^{-1}})_{\sigma}\to A\otimes (\underline{\rm aut}^r(A)^{\sigma})^{\sigma^{-1}}.\]
Now \Cref{rem:2cocycle}(5) yields the following commutative diagram:
\[
\xymatrix{
&A\ar[dl]_-{\rho^{\sigma^{-1}}_{A_{\sigma^{-1}}}}\ar[dr]^-{\rho_A} &\\
A\otimes \underline{\rm aut}^r(A_{\sigma^{-1}})^{f^*\sigma^{-1}}\ar[rr]^-{\id \otimes f^{\sigma^{-1}}}&& A\otimes \underline{\rm aut}^r(A).
}
\]
On the other hand, the universal property of $\rho_{A}$ implies that there is a unique Hopf algebra map $g: \underline{\rm aut}^r(A) \to \underline{\rm aut}^r(A_{\sigma^{-1}})^{f^*\sigma^{-1}}$ making the diagram
\[
\xymatrix{
&&A\ar[d]^-{\rho^{\sigma^{-1}}_{A_{\sigma^{-1}}}}\ar[dll]_-{\rho_A}\ar[drr]^-{\rho_A} &&\\
A\otimes \underline{\rm aut}^r(A)\ar[rr]_-{\id \otimes g} && A\otimes \underline{\rm aut}^r(A_{\sigma^{-1}})^{f^*\sigma^{-1}}\ar[rr]_-{\id \otimes f^{\sigma^{-1}}}&& A\otimes \underline{\rm aut}^r(A)
}
\]
commute, which implies that $f^{\sigma^{-1}}\circ g= \id$ on $\underline{\rm aut}^r(A)$.
Similarly by \eqref{eq:map1}, we have $f^*\sigma=(f^{\sigma^{-1}})^*\sigma$ is a left 2-cocycle on $\underline{\rm aut}^r(A_{\sigma^{-1}})^{f^*\sigma^{-1}}$. Note that $g^*( (f^{\sigma^{-1}})^*\sigma)=(f^{\sigma^{-1}}\circ g)^*\sigma=\sigma$ and so we have a Hopf algebra map
\begin{align}\label{thm211g}g^{(f^{\sigma^{-1}})^*\sigma}: \underline{\rm aut}^r(A)^{\sigma} \to \underline{\rm aut}^r(A_{\sigma^{-1}})\end{align}
making the diagram
\[
\xymatrix{
&&A_{\sigma^{-1}}\ar[dll]_-{\rho_{A_{\sigma^{-1}}}}\ar[drr]^-{\rho_{A_{\sigma^{-1}}}}\ar[d]^-{\rho_A^\sigma} & &\\
A_{\sigma^{-1}}\otimes \underline{\rm aut}^r(A_{\sigma^{-1}})\ar[rr]_-{\id \otimes f}&& A_{\sigma^{-1}}\otimes \underline{\rm aut}^r(A)^\sigma\ar[rr]_-{\id \otimes g^{(f^{\sigma^{-1}})^*\sigma}}&& A_{\sigma^{-1}}\otimes \underline{\rm aut}^r(A_{\sigma^{-1}})
}
\]
commute, which yields that $g^{\left(f^{\sigma^{-1}}\right)^*\sigma}\circ f= \id$ on $\underline{\rm aut}^r(A_{\sigma^{-1}})$.
Finally, it is immediate that $f$ and $g^{(f^{\sigma^{-1}})^*\sigma}$ (see \eqref{thm211f} and \eqref{thm211g}) are inverse to each other since $g^{(f^{\sigma^{-1}})^*\sigma}=g$ as linear maps by \Cref{rem:2cocycle}(5) and the fact that $f^{\sigma^{-1}}\circ g$ and $g^{(f^{\sigma^{-1}})^*\sigma}\circ f$ are both identities on $\underline{\rm aut}^r(A)$ and $\underline{\rm aut}^r(A_{\sigma^{-1}})$, respectively. \end{proof}
Let $m\ge 2$ be an integer. In the following, we extend the explicit construction in \cite[Chapter 6]{Manin2018} and \cite[\S 2.1]{CWZ2014} to determine the coaction of a bialgebra on a connected graded algebra with $m$-homogeneous relations. Let $A$ be a connected graded algebra generated by $x_1,\ldots,x_n$ in degree one, subject to the $m$-homogeneous relations $\{r_w=0\mid 1\leq w \leq t\}$, where
\[
r_w:=\sum_{1\leq i_1,\ldots,i_m\leq n} c_w^{i_1\cdots i_m} x_{i_1}\cdots x_{i_m},
\]
with $c_w^{i_1\cdots i_m}\in \Bbbk$. We assume that $\{r_1,\ldots,r_t\}$ are linearly independent. We write $A=\Bbbk\langle A_1\rangle/(R)$, where $A_1={\rm span}_\Bbbk(x_1,\ldots,x_n)$ and $R=\bigoplus_{i=1}^t\Bbbk r_i\subseteq (A_1)^{\otimes m}$. The \emph{$m$-Koszul dual} of $A$, denoted by $A^!$, is an algebra $\Bbbk\langle A_1^*\rangle/(R^\perp)$, where $R^\perp$ is the subspace
\begin{equation}\label{eq:Rperp}
R^\perp:=\left\{a\in (A_1^*)^{\otimes m}\,|\, \langle a,R\rangle =0\right\}=\left\{a\in (A_1^*)^{\otimes m}\,|\, \langle a,r_w\rangle=0 \text{ for all $w$}\right\}.
\end{equation}
By abuse of notation, let $\{x_1,\ldots,x_n\}$ be a basis for $A_1$ and let $\{x^1,\ldots,x^n\}$ be the dual basis for $(A^!)_1= (A_1)^*$. Also choose a basis for $R^\perp$, say $r^u$ and write
\[
r^u=\sum_{1\leq i_1,\ldots,i_m\leq n} c^u_{i_1\cdots i_m} x^{i_1}\cdots x^{i_m}
\]
for all $1\leq u\leq n^m-t$. According to \eqref{eq:Rperp}, we have
\begin{equation}\label{eq:mrelation}
\sum_{1\leq i_1,\ldots,i_m\leq n} c^u_{i_1\cdots i_m}c_w^{i_1\cdots i_m}=0
\end{equation}
for all possible $u,w$.
The following lemma extends the well-known result in quadratic case \cite[Lemma 2.2]{CWZ2014} to any connected graded algebra with $m$-homogeneous relations.
\begin{lemma}\label{lemma:bialgebraM}
Let $A$ be the quotient algebra of the free algebra $A'=\Bbbk\langle x_1,\ldots,x_n\rangle$ by $m$-homogeneous relations $R=\bigoplus_{w=1}^t\Bbbk\, r_w$ as above. Let $B$ be a bialgebra that right coacts on $A'$ with $\rho(x_i)=\sum_{j=1}^n x_j\otimes b_{ji}$ for $b_{ij}\in B$.
Then the following are equivalent.
\begin{itemize}
\item[(a)] The coaction $\rho: A'\to A'\otimes B$ satisfies $\rho(R)\subseteq R\otimes K$, that is
\begin{equation}\label{eq:coactionR}
\sum_{\substack{1\leq i_1,\ldots,j_m\leq n \\ 1\leq j_1,\ldots,j_m\leq n}} c_w^{i_1\cdots i_m}c^u_{j_1\cdots j_m}b_{j_1i_1}\cdots b_{j_mi_m}=0
\end{equation}
for all possible $u,w$.
\item[(b)] The map $\rho$ naturally induces a coaction of $B$ on $A$ such that $A$ is a right $B$-comodule algebra.
\end{itemize}
\end{lemma}
\begin{proof}
It is immediate that every right coaction of $B$ on $A$ is induced by a right $B$-coaction on $A'$ that factors through $R$ and vice versa. So it suffices to show that in part (a) the condition $\rho(R)\subseteq R\otimes B$ holds if and only if \eqref{eq:coactionR} holds for all possible $u,w$. Since $R\subseteq (A_1)^{\otimes m}$ with $A_1={\rm span}_\Bbbk(x_1,\ldots,x_n)$, we have $\rho(R)\subseteq (A_1)^{\otimes m} \otimes B$. By duality,
\[
R=\left\{r\in (A_1)^{\otimes m} \mid \langle R^\perp,r\rangle=0\right\}.
\]
Therefore $\rho(R)\subseteq R\otimes B$ if and only if $\langle R^\perp \otimes \id_B,\rho(R)\rangle=0$, if and only if
\begin{align*}
&\langle r^u\otimes \id_B,\ \rho(r_w)\rangle\\
&=\left\langle \sum_{1\leq j_1,\ldots, j_m\leq n} c_{j_1\cdots j_m}^u x^{j_1}\cdots x^{j_m}\otimes \id_B,\ \rho\left(\sum_{1\leq i_1,\ldots,i_m\leq n} c_w^{i_1\cdots i_m} x_{i_1\cdots i_m}\right)\right \rangle\\
&=\left\langle \sum_{1\leq j_1,\ldots, j_m\leq n} c_{j_1\cdots j_m}^u x^{j_1}\cdots x^{j_m}\otimes \id_B, \sum_{\substack{1\leq k_1,\ldots,k_m\leq n\\ 1\leq i_1,\ldots,i_m\leq n}} c_w^{i_1\cdots i_m} x_{k_1\cdots k_m}\otimes b_{k_1i_1}\cdots b_{k_mi_m}\right \rangle\\
&= \sum_{\substack{1\leq i_1,\ldots, i_m\leq n\\ 1\leq j_1,\ldots,j_m\leq n\\1\leq k_1,\ldots,k_m\leq n}} c_w^{k_1\cdots k_m} c_{j_1\cdots j_m}^u \left\langle x^{j_1}\cdots x^{j_m},\ x_{k_1\cdots k_m}\right\rangle \otimes b_{k_1i_1}\cdots b_{k_mi_m}\\
&= \sum_{\substack{1\leq i_1,\ldots,j_m\leq n \\ 1\leq j_1,\ldots,j_m\leq n}} c_w^{i_1\cdots i_m}c^u_{j_1\cdots j_m}b_{j_1i_1}\cdots b_{j_mi_m}\\
&= 0
\end{align*}
for all possible $u,w$.
\end{proof}
For two connected graded algebras $A=\Bbbk\langle A_1\rangle/(R(A))$ and $B=\Bbbk\langle B_1\rangle/(R(B))$ with $m$-homogeneous relations $R(A)\subseteq (A_1)^{\otimes m}$ and $R(B)\subseteq (B_1)^{\otimes m}$ respectively, we extend Manin's bullet product \cite[\S 4.2]{Manin2018} to $A$ and $B$ such that
\[
A\bullet B~:=~\frac{\Bbbk \langle A_1\otimes B_1\rangle}{\left(\tau(R(A)\otimes R(B))\right)},
\]
where $\tau: (A_1)^{\otimes m} \otimes (B_1)^{\otimes m}\to (A_1\otimes B_1)^{\otimes m}$ is the permutation considered by shuffling $m$-tensor factors of $B_1$ into $m$-tensor factors of $A_1$. When $B=A^!=\Bbbk\langle (A_1)^*\rangle/(R(A)^\perp)$ is the $m$-Koszul dual algebra of $A$, we see that $A\bullet A^!$ is a connected graded bialgebra with matrix comultiplication defined below on the generators of $A_1\otimes (A_1)^*$. As before, choose a basis $\{x_1,\ldots,x_n\}$ for $A_1$ and let $\{x^1,\ldots,x^n\}$ be the dual basis for $(A^!)_1= (A_1)^*$. Write $z_j^k=x_j\otimes x^k \in A_1\otimes (A_1)^*$ as the generators for $A\bullet A^!$. Then the comultiplication on $A\bullet A^!$ is given by
\[
\Delta(z_j^k)=\sum_{1\leq i\leq n} z_i^k\otimes z_j^i, \quad \text{for any } 1 \leq j,k \leq n.
\]
The following result is a straightforward generalization of the quadratic case \cite[Proposition 2.1.2, Proposition 2.1.3, Lemma 2.2.1]{HNUVVW21}.
\begin{lemma}\label{lem:ManinM}
Let $A=\Bbbk\langle A_1\rangle/(R(A))$ and $B=\Bbbk\langle B_1\rangle/(R(B))$ be two connected graded algebras finitely generated in degree one with $m$-homogeneous relations. Then we have:
\begin{itemize}
\item[(1)] $\underline{\rm end}^r(A)\cong A\bullet A^!$ and $\underline{\rm end}^l(A)\cong A^!\bullet A$;
\item[(2)] $\underline{\rm aut}^r(A)$ is the Hopf envelope of $\underline{\rm end}^r(A)$ and $\underline{\rm aut}^l(A)$ is the Hopf envelope of $\underline{\rm end}^l(A)$;
\item[(3)] $\underline{\rm aut}^r(A)\cong \underline{\rm aut}^l(A^!)$;
\item[(4)] For any algebra map $f: A\to A\otimes B$ satisfying $f(A_1)\subseteq A_1\otimes B_1$, there exists a unique graded algebra map $g: \underline{\rm end}^r(A)\to B$ such that $f=(\id \otimes g)\circ \rho_A$ where $\rho_A: A\to A\otimes \underline{\rm end}^r(A)$ is the universal bialgebra coaction of $\underline{\rm end}^r(A)$ on $A$;
\item[(5)] For any algebra map $f: A\to B\otimes A$ satisfying $f(A_1)\subseteq B_1\otimes A_1$, there exists a unique graded algebra map $g: \underline{\rm end}^l(A)\to B$ such that $f=(g\otimes \id)\circ \rho_{A^!}$ where $\rho_{A^!}: A^!\to \underline{\rm end}^l(A)\otimes A^!$ is the universal bialgebra coaction of $\underline{\rm end}^l(A)$ on $A$.
\end{itemize}
\end{lemma}
\begin{proof}
(1) We use the above notation to denote $A=\Bbbk\langle A_1\rangle/(R)$, where $R=\bigoplus_{w=1}^t\Bbbk\, r_w\in (A_1)^{\otimes m}$. Its $m$-Koszul dual algebra is $A^!=\Bbbk\langle A_1^*\rangle/(R^\perp)$ with $R^\perp =\bigoplus_{u=1}^{n^m-t}\Bbbk\, r^u\in (A_1^*)^{\otimes m}$ satisfying $\langle R^\perp, R\rangle=0$. Hence by definition, we have
\[A\bullet A^!=\frac{\Bbbk \langle z^k_j\rangle}{\left(\sum_{\substack{1\leq i_1,\ldots,i_m\leq n\\ 1\leq j_1,\ldots,j_m\leq n}} c_w^{i_1\cdots i_m}c^u_{j_1\cdots j_m} z_{i_1}^{j_1}\cdots z_{i_m}^{j_m}\right)}.\]
By \Cref{lemma:bialgebraM}, we have the bialgebra $A\bullet A^!$ right coact on $A$ via $\rho: A\to A\otimes (A\bullet A^!)$ induced by
\[
\rho(x_i)=\sum_{1\leq j\leq n} x_j\otimes z^j_i.
\]
Suppose $D$ is another bialgebra that right coacts on $A$ via $\rho_D(x_i)=\sum x_j\otimes d_{ji}$ for some $d_{ji}\in D$. It is clear that $\Delta(d_{ji})=\sum d_{jk}\otimes d_{ki}$ and $\varepsilon(d_{ji})=\delta_{ji}$. Since the $d_{ji}$'s satisfy the relations \eqref{eq:coactionR} by \Cref{lemma:bialgebraM}, we have a well-defined bialgebra map $f: A\bullet A^!\to D$ induced by $f(z^k_j)=d_{kj}$. Moreover, it is clear that $f$ is the unique bialgebra map such that $(\id_D\otimes f)\circ \rho=\rho_D$. Therefore, we have $\underline{\rm end}^r(A)\cong A\bullet A^!$. Similarly, we can prove $\underline{\rm end}^l(A^!)\cong A^!\bullet A$.
(2) is clear from the universal properties of Hopf envelope and Manin's universal quantum groups. (3) follows from (1) and (2) since $(A^!)^! \bullet A^! \cong A \bullet A^!$ as bialgebras. Finally, (4) and (5) can be proved in the similar fashion as (1) using \Cref{lemma:bialgebraM}.
\end{proof}
For any graded automorphism $\phi$ of $A$, we denote by $\phi^!$ the induced graded automorphism on the dual algebra $A^!$.
\begin{lemma}\label{lem:2cocycleGrade}
Let $A$ be a connected graded algebra finitely generated in degree one with $m$-homogeneous relations. For any graded automorphism $\phi$ of $A$, there is a unique pair of graded automorphisms
\begin{align}
\label{eq:twistpairA}
\left(\underline{\rm aut}^r(\phi),\ \underline{\rm aut}^l((\phi^{-1})^!))\right)
\end{align}
of $\underline{\rm aut}^r(A)$ making the following diagrams
\begin{align}\label{lem:caut}
\xymatrix{
A\ar[r]^-{\rho}\ar[d]_-{\phi} & A\otimes \underline{\rm aut}^r(A)\ar[d]^-{\id\otimes\underline{\rm aut}^r(\phi) }\\
A\ar[r]^-{\rho} & A\otimes \underline{\rm aut}^r(A)
}
\quad\quad
\xymatrix{
A^!\ar[r]^-{\rho^!}\ar[d]_-{(\phi^{-1})^!} & \underline{\rm aut}^r(A)\otimes A^!\ar[d]^-{\underline{\rm aut}^l((\phi^{-1})^!)) \otimes \id}\\
A^!\ar[r]^-{\rho^!} & \underline{\rm aut}^r(A)\otimes A^!
}
\end{align}
commute. In particular, the automorphisms in \eqref{eq:twistpairA} are a twisting pair as in \cite[Definition D]{HNUVVW21}. Moreover, we have a left 2-cocycle $\sigma$ on $\underline{\rm aut}^r(A)$ with convolution inverse $\sigma^{-1}$ via
\begin{align}
\label{eq:2cocycleA}
\sigma(x,y)=\varepsilon(x)\, \varepsilon((\underline{\rm aut}^l(\phi^{-1})^{!})^{|x|}(y))\,\, {\rm and}\,\, \sigma^{-1}(x,y)=\varepsilon(x)\, \varepsilon((\underline{\rm aut}^r(\phi)^{|x|}(y))
\end{align}
for any homogeneous elements $x,y\in \underline{\rm aut}^r(A)$.
\end{lemma}
\begin{proof}
By \Cref{lem:ManinM}, we have a unique pair of graded automorphisms
\begin{align}\label{eq:twistpairB}
\left(\underline{\rm end}^r(\phi),\ \underline{\rm end}^l((\phi^{-1})^!))\right)
\end{align}
of $\underline{\rm end}^r(A)$ making the diagrams analogous to \eqref{lem:caut} for $\underline{\rm end}^r(A)$ commute. By an identical proof of \cite[Lemma 2.2.3]{HNUVVW21}, we know it is a twisting pair of $\underline{\rm end}^r(A)$ in terms of \cite[Definition D]{HNUVVW21}. Since $\underline{\rm aut}^r(A)$ is the Hopf envelope of $\underline{\rm end}^r(A)$, we apply \cite[Proposition 1.1.12]{HNUVVW21} to conclude that the twisting pair \eqref{eq:twistpairB} extends uniquely to that of $\underline{\rm aut}^r(A)$ via the natural bialgebra map $\underline{\rm end}^r(A)\to \underline{\rm aut}^r(A)$. We denote the unique extended twisting pair of $\underline{\rm aut}^r(A)$ by \eqref{eq:twistpairA}. Then the diagrams \eqref{lem:caut} commute since the universal coactions of $\underline{\rm aut}^r(A)$ on $A$ and $A^!$ factors through $\underline{\rm end}^r(A)$, respectively. Finally, the fact that $\sigma$ is a well-defined left 2-cocycle on $\underline{\rm aut}^r(A)$ with convolution inverse $\sigma^{-1}$ follows from \cite[Proposition 1.3.2]{HNUVVW21}.
\end{proof}
In the following result, we show that a right Zhang twist of $A$ by $\phi$ is isomorphic to the twisted algebra $A_{\sigma^{-1}}$ by a 2-cocycle $\sigma$ as defined in \eqref{eq:2cocycleA}.
\begin{thm}
\label{lemma:2coZhang}
Let $A$ be a connected $\mathbb Z$-graded algebra finitely generated in degree one with $m$-homogeneous relations and $\phi$ be any graded automorphism of $A$. If $\sigma$ is a left 2-cocycle on $\underline{\rm aut}^r(A)$ defined by $\phi$ as in \eqref{eq:2cocycleA}, then \[A^{\phi}\cong A_{\sigma^{-1}}\]
as graded comodule algebras over $\underline{\rm aut}^r(A)^\sigma$.
\end{thm}
\begin{proof}
Since $A$ is a comodule algebra over $\underline{\rm aut}^r(A)$, $A_{\sigma^{-1}}$ is a comodule algebra over $\underline{\rm aut}^r(A)^\sigma$. One can see that $A^\phi=A_{\sigma^{-1}}=A$ as graded vector spaces. It suffices to show that $A^\phi\cong A_{\sigma^{-1}}$ as graded algebras. For any two homogeneous elements $a,b \in A$, we have
\begin{align*}
a \cdot_{\sigma} b&= \sum a_0b_0\, \sigma^{-1}(a_1, b_1) \\
\overset{\eqref{eq:2cocycleA}}&{=} \sum a_0b_0\,\varepsilon(a_1)\varepsilon(\underline{\rm aut}^r(\phi)^{|a_1|}(b_1))\\
&=\sum a\, b_0\,\varepsilon(\underline{\rm aut}^r(\phi)^{|a|}(b_1))\\
&=a\,(\id \otimes \varepsilon)\circ\left ((\id \otimes \underline{\rm aut}^r(\phi)^{|a|})\circ \rho\right)(b)\\
\overset{\eqref{lem:caut}}&{=}a\,(\id \otimes \varepsilon)\circ\left(\rho \circ \phi^{|a|}\right)(b)\\
&=a\,\phi^{|a|}(b) \\
&=a \cdot_\phi b,
\end{align*}
where the second to last equality follows from the counit axiom of the comodule structure $\rho$. Therefore, the $\Bbbk$-linear identity map $\id$ preserves the product structure and so $A^{\phi}\cong A_{\sigma^{-1}}$ as desired.
\end{proof}
By \Cref{lemma:2coZhang}, \Cref{thm:Manintwist} generalizes the results in \cite[Theorem 2.2.5 and Theorem 2.2.6]{HNUVVW21} for the special case when the 2-cocycle is derived from a graded automorphism of the underlying graded algebra.
\section{Twisting of generalized superpotential algebras}
\label{sec:superpotential}
Superpotential algebras and their associated universal quantum groups have attracted significant study in recent years \cite{DVL1990,Dubois-Violette2005, DVM,BDV13,WaltonWang2016, Chirvasitu-Walton-Wang2019}. They include all Gorenstein Koszul algebras of finite global dimension as a subset \cite[Theorem 11]{DVM}. In this section, we consider a generalization of superpotential algebras, termed derivation quotient algebras, and determine their behavior under 2-cocycle twists arising from coactions of arbitrary Hopf algebras.
Throughout, let $H$ be a Hopf algebra over $\Bbbk$. Suppose $V$ is a finite-dimensional right $H$-comodule. With a fixed basis $\{v_1,\ldots,v_n\}$ of $V$, we denote the right $H$-comodule structure on $V$ as $\rho(v_i)=\sum_{j=1}^n v_j\otimes h_{ji}$ for some elements $h_{ji}\in H$ and any $1\leq i\leq n$. Then the vector space dual $V^*={\rm Hom}_\Bbbk(V,\Bbbk)$ is again a right $H$-comodule via $\rho^*(v^i)=\sum_{j=1}^n v^j\otimes S(h_{ij})$, where $\{v^1,\ldots,v^n\}$ is the dual basis of $\{v_1,\ldots,v_n\}$ in $V^*$. We can view $V^*$ as the left dual of $V$ in the category ${\rm comod}(H)$ of right $H$-comodules, where the canonical evaluation map ${\rm ev}: V^*\otimes V\to \Bbbk$ and coevaluation map ${\rm coev}: \Bbbk \to V\otimes V^*$ are both $H$-comodule maps. Denote by $\Bbbk$ the trivial $H$-comodule. Recall that there are $H$-comodule isomorphisms $\Bbbk\otimes W\cong W \cong \Bbbk\otimes W$ for any right $H$-comodule $W$.
\begin{defn}
\label{defn:derivation}
Let $N$ and $m$ be integers with $2\leq N\leq m$. Let $V$ be a finite-dimensional right comodule over a Hopf algebra $H$ with a subcomodule $j: W \hookrightarrow V^{\otimes m}$. For any integer $1\leq i\leq m$, we define
\begin{itemize}
\item[(1)] subcomodules $\partial^iW$ of $V^{\otimes (m-i)}$ such that
\begin{align*} { \partial^1W = {\rm Im}\left(V^*\otimes W\xrightarrow{\id \otimes j} V^*\otimes (V^{\otimes m})\xrightarrow{\cong } (V^*\otimes V)\otimes V^{\otimes (m-1)} \xrightarrow{{\rm ev}\otimes \id} V^{\otimes (m-1)}\right) } \end{align*}
and $\partial^{i+1}W=\partial(\partial ^iW) $, inductively, and
\item[(2)] the $(m-N)$-th order derivation-quotient algebra $$A(W,N):=TV/(\partial^{m-N}W),$$ where $TV$ is the tensor algebra of $V$.
\end{itemize}
\end{defn}
Note that when $H=\Bbbk$, \Cref{defn:derivation} yields the same derivation-quotient algebras discussed in \cite[\S 1.5.1]{Mori-Smith2016}. In particular, when $H=\Bbbk$ and $W$ is one-dimensional, $A(W,N) $ is the superpotential algebra discussed in \cite{Dubois-Violette2005}.
Next, we describe how $A(W,N)$ behaves under twisting by a $2$-cocycle $\sigma$ on $H$. When $\sigma$ is derived from a graded automorphism of $A(W,N)$, we recover results of \cite{Mori-Smith2016} in terms of Zhang twists.
\begin{lemma}
Let $1\leq N\leq m$ be integers. Let $V$ be a finite-dimensional right $H$-comodule with a subcomodule $W\subseteq V^{\otimes m}$. Then $A(W, N)$ is a right $H$-comodule algebra induced by the right $H$-comodule structure on $V$.
\end{lemma}
\begin{proof}
By \Cref{defn:derivation} and the discussion above, we know $\partial^{m-N}(W)\subseteq V^{\otimes N}$ is a $H$-subcomodule. It is then straightforward to check that $A(W, N)=TV/(\partial^{m-N}W)$, as a quotient algebra of $TV$, is again a right $H$-comodule algebra.
\end{proof}
Now suppose $\sigma: H \times H\to \Bbbk$ is a left 2-cocycle on $H$. For any two right $H$-comodules $U$ and $V$, we define the following invertible linear map
\begin{align}\label{def: MTF}
\xi_{U,V}: U\otimes V&\to U\otimes V\\
u\otimes v&\mapsto \sum \sigma(u_1,v_1)\,u_0\otimes v_0, \notag
\end{align}
for any $u\in U$ and $v\in V$, with inverse $\xi_{U,V}^{-1}: u\otimes v\mapsto \sum \sigma^{-1}(u_1,v_1)u_0\otimes v_0$. Note that there is a monoidal equivalence
\begin{align}
\label{eq:equiv2}
F : \mathrm{comod}(H) \overset{\sim}{\rightarrow} \mathrm{comod}(H^\sigma)
\end{align}
between right comodule categories over $H$ and $H^\sigma$, respectively, where $F$ sends an $H$-comodule algebra $A$ to the $H^\sigma$-comodule algebra $A_{\sigma^{-1}}$. We write $\otimes$ and $\otimes_\sigma$ for the tensor products in the corresponding right comodule categories. In detail, $F$ is given by the identity functor on objects, since $H\cong H^\sigma$ as coalgebras, together with the isomorphism of $H^\sigma$-comodules
\begin{align*}
\xi_{U,V}: F(U \otimes V )\xrightarrow{\sim} F(U) \otimes_\sigma F(V).
\end{align*}
In particular, we have the following commutative diagram
\begin{align}\label{eq:cdsigma}
\xymatrix{
F(U\otimes V\otimes W) \ar[rr]^-{\xi_{U\otimes V,W}}\ar[d]_-{\xi_{U, V\otimes W}} && F(U \otimes V) \otimes_\sigma F(W) \ar[d]^-{\xi_{U,V}\otimes \id}\\
F(U) \otimes_\sigma F(V \otimes W) \ar[rr]^-{\id \otimes \xi_{V,W}} && F(U) \otimes_\sigma F(V) \otimes_\sigma F(W)
}
\end{align}
for any right $H$-comodules $U,V,W$.
\begin{defn}\label{def:twistform}
Retain the above notation.
\begin{itemize}
\item[(1)] For any $v\in V^{\otimes m}$, we define a new element
\begin{align*}
v_{\sigma}:=\left((\id_{V^{\otimes (m-2)}}\otimes \xi_{V,V})\circ\cdots\circ(\id_V\otimes \xi_{V, V^{\otimes(m-2)}})\circ \xi_{V,V^{\otimes (m-1)}}\right)(v)
\end{align*}
in $V^{\otimes m}$; and
\item[(2)] for any subspace $W\subseteq V^{\otimes m}$, we define a new subspace
\begin{align*}
W_{\sigma}:&=\left((\id_{V^{\otimes (m-2)}}\otimes \xi_{V,V})\circ\cdots\circ(\id_V\otimes \xi_{V, V^{\otimes(m-2)}})\circ \xi_{V,V^{\otimes (m-1)}}\right)(W)\notag\\
&=\{w_{\sigma}\,|\, w\in W\}
\end{align*}
in $V^{\otimes m}$.
\end{itemize}
\end{defn}
As a consequence of the commutative diagram \eqref{eq:cdsigma}, there exists a unique $H^{\sigma}$-comodule isomorphism $F(V^{\otimes m})\to F(V)^{\otimes_\sigma m}$ given by all possible compositions of identity maps, $\xi_{V^{\otimes i},V^{\otimes j}}$, $\id_{V^{\otimes(m-i-j)}}$ and parentheses, which we denote
\begin{align}\label{eq:coherence}
\xi_{V^{\otimes m}}:F(V^{\otimes m})\to F(V)^{\otimes_\sigma m}.
\end{align}
By abuse of notation, we view $\xi_{V^{\otimes m}}$ as an invertible endomorphism on the vector space $V^{\otimes m}$. Thus we can write $v_{\sigma}=\xi_{V^{\otimes m}}(v)$ and $W_{\sigma}=\xi_{V^{\otimes m}}(W)$ as in \Cref{def:twistform}.
In view of \Cref{lem:2cocycleGrade}, the following result extends \cite[Lemma 5.1]{Mori-Smith2016} from Zhang twists to 2-cocycle twists.
\begin{lemma}\label{lem:twistR}
Let $V$ be a finite-dimensional right $H$-comodule, $R$ be a subcomodule of $V^{\otimes m}$ for some $m\ge 2$, and $A=TV/(R)$. Then for any left $2$-cocycle $\sigma$ on $H$, we have $A_{\sigma^{-1}}\cong TV/(R_{\sigma})$ as graded right $H^\sigma$-comodule algebras.
\end{lemma}
\begin{proof}
We apply the monoidal equivalence functor $F: {\rm comod}(H)\xrightarrow{\sim} {\rm comod}(H^\sigma)$ described in \eqref{eq:equiv2} to the $H$-comodule embedding $j:R \hookrightarrow V^{\otimes m}$. Thus we have
\[
R_{\sigma}={\rm Im}\left(F(R)\xrightarrow{F(j)} F(V^{\otimes m})\xrightarrow{\xi_{V^{\otimes m}}} F(V)^{\otimes_\sigma m}\right).
\]
Therefore $R_{\sigma}\subseteq F(V)^{\otimes_\sigma m}$ is a $H^{\sigma}$-subcomodule and hence $TV/(R_{\sigma})$ is a graded right $H^\sigma$-comodule algebra.
We have a graded $H^\sigma$-comodule algebra map $f: TV\to \left(TV/(R)\right)_{\sigma^{-1}}$ satisfying $f|_V=\id$, since $TV$ is free. Note that $f|_{V^{\otimes \ell}}=\xi^{-1}_{V^{\otimes \ell}}$. Hence
\[
f(R_\sigma)=\xi^{-1}_{V^{\otimes \ell}}(R_\sigma)=\xi^{-1}_{V^{\otimes \ell}}\circ \xi_{V^{\otimes \ell}}(R)=R,
\]
which is $0$ in $\left(TV/(R)\right)_{\sigma^{-1}}$, and so $f$ factors through $TV/(R_\sigma)$, which we still denote as $f: TV/(R_\sigma)\to \left(TV/(R)\right)_{\sigma^{-1}}$.
Similarly, we have a graded $H$-comodule algebra map $g: TV/R\to \left(TV/(R_\sigma)\right)_\sigma$ satisfying $g|_V=\id$. It induces a graded $H^\sigma$-comodule algebra map
\[
g_{\sigma^{-1}}: (TV/R)_{\sigma^{-1}}\xrightarrow{g} \left((TV/(R_\sigma))_\sigma\right)_{\sigma^{-1}}=TV/(R_\sigma)
\]
still satisfying $g_{\sigma^{-1}}|_V=\id$. Now it is routine to check that $f$ and $g_{\sigma^{-1}}$ are inverse to each other.
\end{proof}
Our next result is a 2-cocycle version of \cite[Proposition 5.2]{Mori-Smith2016}.
\begin{proposition}\label{prop:2.2.5}
Let $1\leq N\leq m$ be integers. Let $V$ be a finite-dimensional right $H$-comodule, $j:W \hookrightarrow V^{\otimes m}$ be an $H$-subcomodule and $\sigma$ be a left 2-cocycle on $H$. Then
\begin{itemize}
\item[(1)] $\partial^i(W_{\sigma})=(\partial^i W)_{\sigma}$;
\item[(2)] if $A(W,N)=TV/(R)$, where $R\subseteq V^{\otimes N}$ is some $H$-subcomodule, then $A(W_{\sigma},N)=TV/(R_{\sigma})$;
\item[(3)] $A(W_{\sigma},N)\cong A(W,N)_{\sigma^{-1}}$.
\end{itemize}
\end{proposition}
\begin{proof}
(1) One checks that
\begin{align*}
\partial^i(W_{\sigma}) &= {\rm Im}\left((F(V)^*)^{\otimes_\sigma i}\otimes_{\sigma} F(W)\xrightarrow{\id \otimes_{\sigma} F(j)}(F(V)^*)^{\otimes_\sigma i}\otimes_{\sigma} F(V^{\otimes m})\right. \\
&\quad \xrightarrow{\id \otimes_{\sigma} \xi_{V^{\otimes m}}}(F(V)^*)^{\otimes_\sigma i}\otimes_{\sigma} F(V)^{\otimes_\sigma m}\\
&\quad \xrightarrow{\cong }((F(V)^*)^{\otimes_\sigma i}\otimes_{\sigma} F(V)^{\otimes_\sigma i})\otimes_{\sigma} F(V)^{\otimes_{\sigma} (m-i)}\left.\xrightarrow{{\rm ev}\otimes \id} F(V)^{\otimes_\sigma (m-i)}\right)
\end{align*}
and
\begin{align*}
(\partial^i W)_{\sigma} &= {\rm Im}\left(F((V^*)^{\otimes i}\otimes W)\xrightarrow{F(\id \otimes j)}F((V^*)^{\otimes i}\otimes V^{\otimes m})\right. \\
&\quad \xrightarrow{F(\cong)}F(((V^*)^{\otimes i}\otimes V^{\otimes i})\otimes V^{\otimes (m-i)})\xrightarrow{F({\rm ev}\otimes \id)}F(V^{\otimes (m-i)})\\
&\quad \left.\xrightarrow{\xi_{V^{\otimes (m-i)}}}F(V)^{\otimes_\sigma (m-i)}\right).
\end{align*}
Note that there is a natural isomorphism $\iota:(F(V)^*)^{\otimes_\sigma i} \cong F(V^*)^{\otimes_\sigma i}\cong F((V^*)^{\otimes i})$. Then the two above images are equal by the following commutative diagram:
\[
\xymatrix{
(F(V)^*)^{\otimes_\sigma i} \otimes_\sigma F(W) \ar[r]^-{\iota\otimes\id} \ar[d]_-{\id \otimes_{\sigma}\, F(j)} &F\left( (V^*)^{\otimes i}\right) \otimes_\sigma F(W)\ar[d]^-{\xi^{-1}_{(V^*)^{\otimes i},W}}\\
(F(V)^*)^{\otimes_\sigma i} \otimes_\sigma F(V^{\otimes m }) \ar[d]_-{\id \otimes_\sigma\, \xi_{V^{\otimes m}}} & F\left( (V^*)^{\otimes i} \otimes W \right) \ar[d]^-{F(\id \otimes j)} \\
F(V^*)^{\otimes_\sigma i} \otimes_\sigma F(V)^{\otimes_\sigma m }\ar[d]_-{\cong } & F\left( (V^*)^{\otimes i} \otimes V^{\otimes m} \right) \ar[d]^-{F(\cong)}\\
\left((F(V)^*)^{\otimes_\sigma i} \otimes_\sigma F(V)^{\otimes_\sigma i }\right)\otimes_\sigma F(V)^{\otimes_\sigma (m-i)}\ar[d]_-{{\rm ev} \otimes_{\sigma}\, \id} & F\left(((V^*)^{\otimes i} \otimes V^{\otimes i}) \otimes V^{\otimes (m-i)}\right)\ar[d]^-{F({\rm ev}\otimes \id)} \\
F(V)^{\otimes_\sigma (m-i)} & F(V^{\otimes (m-i)}). \ar[l]^-{\xi_{V^{\otimes (m-i)}}}
}
\]
(2) If $A(W,N)=TV/(R)$, then $R=\partial^{m-N}W$ by definition. Thus by (1), we have
\[
A(W_\sigma,N)=TV/(\partial^{m-N}W_\sigma)=TV/((\partial^{m-N}W)_\sigma)=TV/(R_\sigma).
\]
(3) Denote by $A(W,N)=TV/(R)$, where $R=\partial^{m-N}W$. It follows from (2) and \Cref{lem:twistR} that
\[
A(W_\sigma,N)=TV/(R_\sigma)\cong(TV/(R))_{\sigma^{-1}}=A(W,N)_{\sigma^{-1}}. \qedhere
\]
\end{proof}
Recall $V={\rm Span}_\Bbbk(v_1,\ldots,v_n)$ is a right $H$-comodule via $\rho: v_i\mapsto \sum_{j=1}^nv_j\otimes h_{ji}$ for some $h_{ji}\in H$. In the following result, we will view $V^*={\rm Span}_\Bbbk(v^1,\ldots,v^n)$ as a left $H$-comodule via
\begin{align}\label{RLcoaction}
\rho^*: v^i \mapsto \sum_{j=1}^n h_{ij}\otimes v^j.
\end{align}
Suppose $\sigma$ is a left 2-cocycle on $H$. For two right $H$-comodules $U$ and $V$, we define the invertible linear map
\begin{align}\label{def:adjointF}
\zeta_{U^*,V^*}: U^*\otimes V^*&\to U^*\otimes V^*&\\
f\otimes g&\mapsto \sum \sigma(f_{-1},g_{-1})\, f_0g_0 \notag
\end{align}
for any $f\in U^*$ and $g\in V^*$ with inverse $\zeta^{-1}_{U^*,V^*}: f\otimes g\mapsto \sum \sigma^{-1}(f_{-1},g_{-1})f_0g_0$, where $U^*$ and $V^*$ are viewed as left $H$-comodules as above. We denote by
\begin{align}\label{eq:coherenceE}
\zeta_{(V^*)^{\otimes m}}: (V^*)^{\otimes m}\to (V^*)^{\otimes m}
\end{align}
the unique invertible linear map given by all possible compositions of $\id_{{V^*}^{\otimes(m-i-j)}}$, $\zeta_{(V^*)^{\otimes i},(V^*)^{\otimes j}}$, and parentheses.
Let $m\ge 2$ be any integer, and $R\subseteq V^{\otimes m}$ be a right $H$-subcomodule. So the connected graded algebra $A=TV/(R)$ is a right $H$-comodule algebra with the induced right $H$-coaction on $V$. We still denote by $R^\perp=\{f\in (V^*)^{\otimes m}\,|\, f(r)=0\text{ for all $r\in R$}\}$ and the dual algebra of $A$ by $A^!=TV^*/(R^\perp)$. By a higher degree version of \cite[Proposition 2.5]{CWZ2014} for $A=TV/(R)$, we know $A^!$ is a left $H$-comodule algebra with the left $H$-coaction induced by \eqref{RLcoaction}. In particular, $R^\perp\subseteq (V^*)^{\otimes m}$ is a left $H$-subcomodule.
\begin{proposition}\label{lem:dualtwist}
Retain the above notation. Then we have
\[
\left(A_{\sigma^{-1}}\right)^!\cong \ \!_\sigma(A^!).
\]
as left $H^\sigma$-comodule algebras.
\end{proposition}
\begin{proof}
In view of \Cref{lem:twistR}, we have
\[
\left(A_{\sigma^{-1}}\right)^!=\left((TV/(R))_{\sigma^{-1}}\right)^!\cong \left(TV/(R_\sigma)\right)^!=TV^*/((R_\sigma)^\perp).
\]
By the canonical pairing \[\langle-,-\rangle:
(V^*)^{\otimes m} \times V^{\otimes m}\to \Bbbk,\]
one checks that
\[
\zeta_{(V^*)^{\otimes m}}=\left(\xi_{V^{\otimes m}}\right)^*.
\]
Note that $(R_\sigma)^\perp$ is defined as $\langle (R_\sigma)^\perp,R_\sigma\rangle=0$, which implies that \begin{align*}
0=&\,\langle (R_\sigma)^\perp,\,R_\sigma\rangle\\
=&\,\langle (R_\sigma)^\perp,\, \xi_{V^{\otimes m}}(R)\rangle \quad \quad ({\rm by}\,\, \eqref{eq:coherence})\\
=&\,
\langle (\xi_{V^{\otimes m}})^*(R_\sigma)^\perp,\, R\rangle\\
=&\,\langle \zeta_{(V^*)^{\otimes m}}(R_\sigma)^\perp,\, R\rangle.
\end{align*}
Thus, $(R_\sigma)^\perp=(\zeta_{(V^*)^{\otimes m}})^{-1}\,(R^\perp) = \zeta^{-1}_{(V^*)^{\otimes m}}\,(R^\perp)$. Then we get
\[
\left(A_{\sigma^{-1}}\right)^! \cong TV^*/((R_\sigma)^\perp) = TV^*/(\zeta^{-1}_{(V^*)^{\otimes m}}\,(R^\perp)) \cong \ \!_\sigma(TV^*/(R^\perp))= \ \!_\sigma(A^!),
\] where the last isomorphism follows by a left-sided version of \Cref{lem:twistR}.
\end{proof}
\section{Invariant properties under 2-cocycle twists}
\label{sec:properties}
In this section, we discuss 2-cocycle twists of AS-regular algebras that are comodule algebras over a Hopf algebra $H$. When $H$ is semisimple, it was proved in \cite[Lemma 4.12]{CKWZ} that AS-regularity is preserved under Drinfeld twists coming from $H$-actions. When $H$ is finite-dimensional, the same result was proved in \cite[Theorem 4.8]{CS2} for any 2-cocycle twist for $H$-coactions. Here, we mainly focus on the $H$-coactions when $H$ is infinite dimensional.
Let $A$ be a right $H$-comodule algebra together with the right $H$-coaction $\rho: A\to A\otimes H$. A vector space $M$ is called a \emph{relative $(A,H)$-module} if $M$ is a left $A$-module and a right $H$-comodule such that
\[
\sum (ax)_0\otimes (ax)_1=\sum a_0x_0\otimes a_1x_1,
\]
for all $a\in A$ and $x\in M$. We denote by $\!_A\mathcal M^H$ the category of relative Hopf modules with morphisms given by $A$-linear and $H$-colinear maps.
Let $m\ge 2$ be an integer and $V$ be a right $H$-comodule with subcomodule $R\subseteq V^{\otimes m}$. Suppose $A=TV/(R)$ is a connected graded algebra with $\deg(V)=1$. We set inductively $R_0=\Bbbk$, $R_1=V$, and for $k\ge m$,
\[
R_k~=~\bigcap_{i+j+m=k} V^{\otimes i}\otimes R \otimes V^{\otimes j}\subseteq V^{\otimes k}.
\]
We recall the \emph{generalized Koszul complex} of $\Bbbk$ over $A$ as follows (c.~f.~ \cite[\S 2.5]{Mori-Smith2016})
\begin{align}\label{lem:GKC}
\cdots \to A\otimes R_{\delta(k)}\to A\otimes R_{\delta(k-1)}\to\cdots \to A\otimes R_{\delta(1)}\to A\otimes R_{\delta(0)}\to \Bbbk\to 0,
\end{align}
where we define
\[
\delta(k):=\begin{cases} \frac{k}{2}\,m &k\equiv 0 \pmod 2 \\[0.1in] \frac{(k-1)}{2}\, m+1 & k\equiv 1 \pmod 2, \end{cases}
\]
and the differentials $d_k$ are induced by the multiplication $\mu$ in $A$ such that
\begin{equation}\label{GKD}
\xymatrix{
A\otimes R_{\delta(k)}\ar[r]^-{d_k}\ar@{^{(}->}[d] & A\otimes R_{\delta(k-1)}\ar@{^{(}->}[d] \\
A\otimes V^{\otimes \delta(k)}\ar[d]^-{} & A\otimes V^{\otimes \delta(k-1)} \\
\left(A\otimes V^{\otimes \left(\delta(k)-\delta(k-1)\right)}\right)\otimes V^{\otimes \delta(k-1)} \ar@{^{(}->}[r] & \left(A\otimes A\right)\otimes V^{\otimes \delta(k-1)} \ar[u]_-{\mu\otimes \id}
}
\end{equation}
commutes. We denote $\Gr(A)$ the category of graded modules over $A$ with degree-preserving $A$-module maps. We say $A$ is \emph{$N$-Koszul} if the trivial graded $A$-module $\Bbbk$ has a free resolution
\begin{align}\label{eq:freegraded}
\cdots\to P_k\to P_{k-1}\to\cdots \to P_1\to P_0\to \Bbbk\to 0
\end{align}
in $\Gr(A)$, where each term $P_k$ is generated in degree $\delta(k)$. If $A$ is $N$-Koszul, then the generalized Koszul complex \eqref{lem:GKC} is exact and the $\Ext$-algebra \[E:=\bigoplus_{i\ge 0} \Ext_A^i(\Bbbk,\Bbbk)\] is isomorphic to the subalgebra $\displaystyle B=\bigoplus_{k\ge 0} (A^!)_{\delta(k)}$ of $\displaystyle A^!=\bigoplus_{k\ge 0} (A^!)_{k}$.
We know $A$ is a right $H$-comodule algebra induced by the $H$-coaction on $V$. We use $\!_{\Gr(A)}\mathcal M^H$ to denote the subcategory of $\!_A\mathcal M^H$ consisting of all graded relative Hopf modules with degree-preserving $A$-linear and $H$-colinear maps. Let $\sigma$ be a left 2-cocycle on $H$. One can check that the monoidal equivalence $F: {\rm comod}(H)\xrightarrow{\sim} {\rm comod}(H^\sigma)$ induces an equivalence between the two categories of graded relative Hopf modules $\!_{\Gr(A)}{\mathcal M}^H$ and $\!_{\Gr(A_{\sigma^{-1}})}{\mathcal M}^{H^\sigma}$, respectively. For any $M\in \!_{\Gr(A)}{\mathcal M}^H$, we denote by $M_{\sigma^{-1}}$ its image in $\!_{\Gr(A_{\sigma^{-1}})}{\mathcal M}^{H^\sigma}$ under $F$, where $M_{\sigma^{-1}}=M$ as graded $H^{\sigma}$-comodules and the graded $A_{\sigma^{-1}}$-module structure is given by
\begin{align}\label{eq:twistmod}
a\cdot_\sigma x =\sum \sigma^{-1}(a_1,x_1)a_0x_0,
\end{align}
for any $a\in A_{\sigma^{-1}}$ and $x \in M_{\sigma^{-1}}$.
When $H$ is finite-dimensional, the following result was observed in \cite[Corollary 4.5]{CS2}, where the Koszul property is preserved under any 2-cocycle twist coming from finite Hopf coactions. We now prove it for an arbitrary Hopf algebra $H$.
\begin{lemma} \label{lem:twistkoszul}
Let $A=TV/(R)$ be a connected graded algebra with homogeneous relation $R$ in degree $m\ge 2$. Let $H$ be a Hopf algebra that right coacts on $A$ preserving the grading of $A$, and let $\sigma$ be a 2-cocycle on $H$. Then $A$ is $N$-Koszul if and only if $A_{\sigma^{-1}}$ is $N$-Koszul.
\end{lemma}
\begin{proof}
We call a graded relative $(A,H)$-module $M$ \emph{free} if $M\cong A\otimes V$ in $\!_{\Gr(A)}{\mathcal M}^H$ for some graded right $H$-comodule $V$ with the $A$-module structure coming from the left hand tensorand. One checks that \eqref{lem:GKC} consists of free graded relative $(A,H)$-modules since each $R_{\delta(k)}$ is a graded $H$-comodule concentrated in degree $\delta(k)$ and all the differentials are morphisms in $\!_{\Gr(A)}{\mathcal M}^H$ by \eqref{GKD}.
Next, we know the monoidal equivalence $F: {\rm comod}(H)\xrightarrow{\sim} {\rm comod}(H^\sigma)$ induces an equivalence between $\!_{\Gr(A)}{\mathcal M}^H$ and $\!_{\Gr(A_{\sigma^{-1}})}{\mathcal M}^{H^\sigma}$. Moreover, one checks that this equivalence sends free graded relative $(A,H)$-modules to free graded relative $(A_{\sigma^{-1}},H^\sigma)$-modules since $(A\otimes V)_{\sigma^{-1}}\cong A_{\sigma^{-1}}\otimes_\sigma V$.
Now suppose $A$ is $N$-Koszul. Applying $F$ to \eqref{lem:GKC}, we get a resolution of $\Bbbk$ in $\!_{A_{\sigma^{-1}}}{\mathcal M}^{H^\sigma}$ consisting of free graded relative $(A_{\sigma^{-1}},H^\sigma)$-modules, and hence a projective resolution of the trivial $A_{\sigma^{-1}}$-module $\Bbbk$. Moreover, it is clear that $(A\otimes R_{\delta(k)})_{\sigma^{-1}}\cong A_{\sigma^{-1}}\otimes_\sigma R_{\delta(k)}$, where $R_{\delta(k)}$ is treated as a graded $H^\sigma$-comodule concentrated in degree $\delta(k)$. By definition, we have $A_{\sigma^{-1}}$ is $N$-Koszul.
We replace $A$ with $A_{\sigma^{-1}}$ and $H$ with $H^{\sigma}$. Since $\sigma^{-1}$ is a left 2-cocycle on $H^\sigma$ and $(A_{\sigma^{-1}})_\sigma =A$, the other direction follows by the same argument.
\end{proof}
Recall a finite-dimensional algebra $A$ is said to be \emph{Frobenius} if there exists an associative nondegenerate bilinear form $f: A \times A \to \Bbbk$ such that $f(ab,c)=f(a,bc)$, for any $a,b,c \in A$ (see e.g., \cite[\S2]{KJ} and \cite[\S8]{CMZ} for equivalent definitions).
\begin{lemma}
\label{lem:twistF}
Let $A$ be a finite-dimensional right $H$-comodule algebra. Suppose $A$ is Frobenius with an associative nondegenerate bilinear form induced from an $H$-colinear map $f: A \otimes A \to D$ for some one-dimensional $H$-comodule $D$. Then for any left 2-cocycle $\sigma$ on $H$, the twisted algebra $A_{\sigma^{-1}}$ is again Frobenius.
\end{lemma}
\begin{proof}
We apply the monoidal equivalence $F:{\rm comod}_{\rm fd}(H) \to {\rm comod}_{\rm fd}(H^\sigma)$ in \eqref{eq:equiv2} to obtain the $H^{\sigma}$-comodule $F(A)=A_{\sigma^{-1}}$ together with multiplication
\[
m_\sigma: F(A)\otimes_\sigma F(A)\xrightarrow{\xi^{-1}_{A,A}} F(A\otimes A)\xrightarrow{F(m)} F(A),
\]
where $m$ is the multiplication of $A$. We define a bilinear form on $A_{\sigma^{-1}}$ as
$$
f_{\sigma}:=\left(F(A)\otimes_\sigma F(A)\xrightarrow{\xi^{-1}_{A,A}} F(A\otimes A)\xrightarrow{F(f)} F(D)\right),
$$
recalling that $F(D)$ is a one-dimensional $H^\sigma$-comodule.
One can check that $f_{\sigma}$ is associative due to the following commutative diagram:
\[
\xymatrix{
F(A)\otimes_\sigma F(A)\otimes_\sigma F(A)\ar[rrr]^-{{\rm id}\otimes m_\sigma}\ar[ddd]_-{m_\sigma\otimes {\rm id}} &&& F(A)\otimes_\sigma F(A)\ar[ddd]^{f_\sigma}\\
&F(A\otimes A\otimes A)\ar[ul]^-{\xi_{A^{\otimes 3}}}\ar[r]^-{F({\rm id}\otimes m)}\ar[d]_-{F(m\otimes {\rm id})}&F(A\otimes A)\ar[d]^-{F(f)}\ar[ur]_-{\xi_{A^{\otimes 2}}}&\\
&F(A\otimes A)\ar[dl]_-{\xi_{A^{\otimes 2}}}\ar[r]^-{F(f)}&F(D)\ar@{=}[dr]&\\
F(A)\otimes_\sigma F(A)\ar[rrr]^-{f_\sigma}&&& F(D).
}
\]
Note that in the sense of \cite[\S 1.5.1]{TV}, since $f: A\otimes A\to D$ is a nondegenerate bilinear map, it induces a nondegenerate pairing between $D^*\otimes A$ and $A$ via the following $H$-comodule maps
\[
f':=\left((D^*\otimes A)\otimes A\xrightarrow{\cong} D^*\otimes (A\otimes A)\xrightarrow{\id \otimes f} D^*\otimes D\xrightarrow{\cong} \Bbbk\right)
\]
with inverse
\[
g:=\left(\Bbbk\xrightarrow{{\rm coev}} A\otimes A^*\xrightarrow{\id \otimes \eta^{-1}} A\otimes (D^*\otimes A)\right),
\]
where
\begin{align*}
\eta:&=\left(D^*\otimes A\xrightarrow{\id \otimes {\rm coev}} (D^*\otimes A)\otimes (A\otimes A^*)\xrightarrow{\cong} (D^*\otimes (A\otimes A))\otimes A^*\right.\\
&\left.\xrightarrow{(\id \otimes f)\otimes \id} (D^*\otimes D)\otimes A^* \xrightarrow{\cong} \Bbbk\otimes A^*=A^*\right)
\end{align*}
is an isomorphism of $H$-comodules. Hence by \cite[Lemma 1.5]{TV}, we have
\[
f_{\sigma}':=\left(F(D)^*\otimes_{\sigma} F(A)\otimes_{\sigma} F(A) \xrightarrow{\id \otimes_{\sigma} \, f_\sigma} F(D)^*\otimes_{\sigma} F(D) \xrightarrow{\cong} \Bbbk\right)
\]is again a nondegenerate paring with inverse
\begin{align*}
g_\sigma:=&\left(\Bbbk=F(\Bbbk)\xrightarrow{F(g)} F(A\otimes (D^*\otimes A))\xrightarrow{\xi_{A,D^*\otimes A}} F(A)\otimes_\sigma F(D^*\otimes A)\right.\\
&\left.\xrightarrow{\id \otimes \, \xi_{D^*,A}} F(A)\otimes_\sigma (F(D^*)\otimes_\sigma F(A)) \xrightarrow{\cong} F(A)\otimes_\sigma(F(D)^* \otimes_\sigma F(A))\right).
\end{align*}
Here we make identifications $F(D^*)=F(D)^*$ as one-dimensional $H^\sigma$-comodules. In particular, the following identities hold
\begin{align*}
(f_{\sigma}'\otimes \id)\circ (\id\otimes g_\sigma)=\id_{F(D)^*\otimes F(A)}\quad{\text{and}}\quad (\id\otimes f_{\sigma}')\circ (g_\sigma \otimes \id)=\id_{F(A)}.
\end{align*}
As finite-dimensional vector spaces, we can identify $F(D^*)=F(D)=\Bbbk$ and $F(D)^*\otimes F(A)=A_{\sigma^{-1}}=A$, where $f=f'$ and $f_\sigma'=f_\sigma$ as linear maps. In this case, we write $g_\sigma(1)=\sum v_i\otimes v^i$ with $v_i\in F(A)=A_{\sigma{-1}}$ and $v^i\in F(D)^*\otimes F(A)=A_{\sigma^{-1}}$. Then the above identities imply that
\[
\sum f_{\sigma}(a,v_i)\,v^i~=~a~=~\sum v_i\,f_{\sigma}(v^i,a),
\]
for any $a\in A_{\sigma^{-1}}$. Suppose there is some $a\in A_{\sigma^{-1}}$ such that $f_{\sigma}(a,-)=0$, whence $a=0$. So $f_{\sigma}: A_{\sigma^{-1}}\times A_{\sigma^{-1}}\to \Bbbk$ is nondegenerate in its usual sense. As a consequence, $f_{\sigma}$ is an associative nondegenerate bilinear form on $A_{\sigma^{-1}}$, and hence $A_{\sigma^{-1}}$ is Frobenius.
\end{proof}
Recall that an algebra is called \emph{AS-regular} if it has finite global dimension, finite GK-dimension, and satisfies the Gorenstein property (c.~ f.~ \cite[\S 2]{Ro2016}). At last, we prove that AS-regularity is invariant under 2-cocycle twist of any Hopf algebra. The dual version of Drinfeld twist was first proved in \cite[Lemma 4.12]{CKWZ} for semisimple Hopf actions and later for any finite-dimensional Hopf actions \cite[Theorem 4.8]{CS2} using torsors.
\begin{thm}\label{thm:AS}
Let $A$ be an $N$-Koszul, AS-regular algebra of dimension $d$. Let $H$ be any Hopf algebra that right coacts on $A$ preserving the grading of $A$. Then for any left 2-cocycle $\sigma$ on $H$, the twisted algebra $A_{\sigma^{-1}}$ is still $N$-Koszul, AS-regular of the same dimension.
\end{thm}
\begin{proof}
Suppose $A$ is $N$-Koszul, AS-regular of dimension $d$. By \Cref{lem:twistkoszul}, we know $A_{\sigma^{-1}}$ is also $N$-Koszul. By a result of \cite{LPWZ}, a connected graded algebra $B$ is AS-regular if and only if its $\Ext$-algebra $E(B):=\bigoplus_{i\ge 0}\Ext_B^i(\Bbbk,\Bbbk)$ is Frobenius.
Denote $B:=A_{\sigma^{-1}}$. Therefore, it suffices to show that $E(B)=\bigoplus_{i\ge 0}(B^!)_{\delta(i)}$ is Frobenius whenever $E(A)=\bigoplus_{i\ge 0}(A^!)_{\delta(i)}$ is Frobenius.
Since $A$ is AS-regular, $E(A)$ is Frobenius. In particular, we have
\[
E(A)=\Bbbk\oplus (A^!)_{\delta(1)}\oplus\cdots \oplus (A^!)_{\delta(d)}
\]
is a connected graded Frobenius algebra. This means that $\dim_\Bbbk(A^!)_{\delta(d)}=1$ and the multiplication of $A^!$ induces a nondegenerate pairing \begin{align*}\label{thm:Frodual}
(A^!)_{\delta(i)}\times (A^!)_{\delta(d-i)}\to (A^!)_{\delta(d)}
\end{align*}
that is left $H$-colinear, for any $0\leq i\leq d$. This extends to an associative nondegenerate bilinear form on $E(A)$ that is $H$-colinear. \Cref{lem:dualtwist} shows that $(A_{\sigma^{-1}})^!\cong \!_\sigma (A^!)$ using the left $H$-coaction on $A^!$, which implies that $E(B)= \!_\sigma E(A)$. Thus, by the left-sided version of \Cref{lem:twistF}, $E(B)$ is Frobenius as desired.
\end{proof}
In \cite[Theorem 7.2.3]{vdb2017}, Raedschelders and Van den Bergh proved the Morita--Takeuchi equivalence between Manin’s universal quantum groups associated to Koszul AS-regular algebras of the same dimension (in fact, their theorem holds more generally without any assumption of finite GK-dimension on the algebra). Using their theorem, we are now able to deduce the following.
\begin{Cor}\label{Cor:TwistAS}
Suppose $\Bbbk$ is algebraically closed and that $A$ is a Koszul, AS-regular algebra of dimension $d$. Let $B$ any connected graded algebra generated in degree one. Then $B$ is a Koszul, AS-regular algebra of the same dimension with the same Hilbert series
of $A$ if and only if $B$ is a 2-cocycle twist of $A$.
\end{Cor}
\begin{proof}
One direction is clear from \Cref{thm:AS}. For the other direction, we follow the discussion in \cite[\S 15]{Manin2018}. Consider the rigid monoidal category $\mathcal U$ constructed in \cite[\S 15.5.1]{Manin2018}, whose objects are given by the free monoid
\[
\Lambda=\langle r_1,\ldots,r_{d-1},r_d^{\pm 1}\rangle
\]
and morphisms are generated by
\[
\phi_{a,b}: r_{a+b}\to r_a\otimes r_b\quad \text{and}\quad \theta_{a,b}: r_a\otimes r_d^{-1}\otimes r_b\to r_{a+b-d},
\]
satisfying certain relations. Let $A=TV/(R)$ be a Koszul, AS-regular algebra of dimension $d$. Set $R_1=V$ and $R_k=\bigcap_{i+j+2=k} V^{\otimes i}\otimes R\otimes V^{\otimes j}$ for $k\ge 2$, which are all finite-dimensional right comodules over Manin's universal quantum group $\underline{\rm aut}^r(A)$. In particular, $R_d$ is a one-dimensional invertible comodule. As discussed in \cite[\S 15.5.2]{Manin2018}, we have a monoidal functor $M_A: \mathcal U\to {\rm Vect_{fd}} (\Bbbk)$ such that $M(r_i)=R_i$ and $M(\phi_{a,b}): R_{a+b}\to R_a\otimes R_a$ and $M(\theta): R_a\otimes R_d^{-1}\otimes R_b\to R_{a+b-d}$, which are natural comodule morphisms over $\underline{\rm aut}^r(A)$. According to \cite[\S 15.5.2]{Manin2018}, we have ${\rm coend}(M_A)\cong \underline{\rm aut}^r(A)$. Moreover, one sees that
\[A^!=\bigoplus_{0\leq i\leq d} M_A(r_i)^*,\]
where the multiplication map is induced by
\begin{align*}
M_A(r_a)^*\otimes M_A(r_b)^*\xrightarrow{\cong} \left(M_A(r_a)\otimes M_A(r_b)\right)^*\xrightarrow{\cong} \left(M_A(r_a\otimes r_b)\right)^*\xrightarrow{M_A(\phi_{a,b})^*}\left(M_A(r_{a+b})\right)^*.
\end{align*}
In particular, $A^!$ is a right comodule algebra over $\underline{\rm aut}^r(A)$.
Suppose $B$ is another Koszul, AS-regular algebra of dimension $d$. By \cite[Theorem 15.45]{Manin2018}, we have ${\rm cohom}(M_A,M_B)\neq 0$ and cotensoring with it, denoted by $F: {\rm comod}\left(\underline{\rm aut}^r(A)\right)\xrightarrow{\sim}{\rm comod}\left(\underline{\rm aut}^r(B)\right)$, induces a Morita--Takeuchi equivalence between $\underline{\rm aut}^r(A)$ and $\underline{\rm aut}^r(B)$. Now assume that $A$ and $B$ have the same Hilbert series. Hence
\[\dim_\Bbbk M_A(r_i)=\dim_\Bbbk (A^!)_i=\dim_\Bbbk (B^!)_i=\dim_\Bbbk M_B(r_i)\]
for all $1\leq i\leq d$. This implies that the monoidal equivalence $F$ preserves vector space dimensions. Hence $\underline{\rm aut}^r(B)$ is a 2-cocycle twist of $\underline{\rm aut}^r(A)$ (see \cite[Theorem 1.17]{Bichon2014}). In particular,
\[
F(A^!)=F\left(\bigoplus_{0\leq i\leq d} M_A(r_i)^*\right)=\bigoplus_{0\leq i\leq d} F\left(M_A(r_i)\right)^*=\bigoplus_{0\leq i\leq d} M_B(r_i))^*=B^!
\]
as the monoidal equivalence $F$ maps the comodule algebra $A^!$ over $\underline{\rm aut}^r(A)$ to the comodule algebra $B^!$ over $\underline{\rm aut}^r(B)$. As a consequence, $B^!$ is a 2-cocycle twist of $A^!$. Then \Cref{lem:dualtwist} implies that $B=(B^!)^!$ is a 2-cocycle twist of $A=(A^!)^!$.
\end{proof}
|
\section{Introduction}
If $A$ is an abelian group, the \textit{$A$-fibered Burnside ring} of a finite group $G$, denoted by $B^A(G)$, is the Grothendieck ring of the category of finite $A$-fibered $G$-sets.
This ring was first introduced by Dress in \cite{Dress}. The case of $A=k^{\times}$ for a field $k$ is particularly interesting, since $B^{k^\times}(G)$ is naturally isomorphic to the ring $D^k(G)$ of monomial $k$-representations of $G$.
Recalling that $B^A(G) \cong B(G)$, the Burnside ring of $G$, if and only if $A$ has trivial $|G|$-torsion, we will say that two non-isomorphic finite groups $G$ and $H$ provide a \textit{non-trivial counterexample to the isomorphism problem of the $A$-fibered Burnside ring} if $B^A(G)$ and $B^A(H)$ are isomorphic as rings and there is a non-trivial element $a$ in $A$ such that $a^{|G|}=1$. Examples of non-isomorphic finite groups $G$ and $H$ with isomorphic Burnside rings have already been given for instance by Th\'evenaz in \cite{Thev}.
The notion of a \textit{species isomorphism} for fibered Burnside rings was introduced by the second author in \cite{Gar1} for a ring isomorphism preserving the standard bases given by conjugacy classes of monomial pairs, which is analogous to an isomorphism of mark tables in the case of Burnside rings.
In Section 2 of this note we present a sufficient condition on finite groups $G$ and $H$ for the existence of a species isomorphism between their fibered Burnside rings. In Section 3, we use this result to prove that Thévenaz' counterexamples to the isomorphism problem for Burnside rings (see \cite{Thev}) of order $p^2q$ for primes $p$ and $q$ such that $q|(p-1)$, provide also non-trivial counterexamples for the $A$-fibered Burnside ring when the fiber $A$ has trivial $p$-torsion and elements of order $q$.
\subsection*{Notation}
Throughout this note, the letters $G$ and $H$ stand for finite groups.
We write $\mathcal{S}_G$ for the set of subgroups of $G$ and $[\mathcal{S}_G]\subseteq \mathcal{S}_G$ for a set of representatives of the conjugacy classes. For an element $g\in G$, we denote the resulting conjugation map $c_g$ also by $\lexp{g}{-}\colon G\to G,\ x\mapsto gxg^{-1}$. For subgroups $K$ and $L$ of $G$ we write $K=_G L$ and $K\le_G L$ if there exists $g\in G$ with $K=\lexp{g}{L}$ and $K\le \lexp{g}{L}$, respectively.
If $G$ acts on a set $X$, we write $G\backslash X$ for the set of its orbits and we write $x=_Gy$ if two elements $x$ and $y$ of $X$ are in the same orbit.
\section{A criterion for species isomorphisms}
We first recall some basic definitions and results on fibered Burnside rings. We refer the reader to \cite{BY}, \cite{Dress} and \cite{Gar1} for further details.
The $A$-fibered Burnside ring can be presented as follows: let $\mathcal{M}^A_G$ be the set consisting of all pairs $(K,\phi)$, where $K\leq G$ and $\phi\colon K\longrightarrow A$ is a group homomorphism, also called \textit{subcharacters} or \textit{monomial pairs}. Note that $G$ acts on $\mathcal{M}^A_G$ by conjugation. We write $[K,\phi]_G$ for the orbit of $(K,\phi)$. Then $B^A(G)$ is the free abelian group over the basis $G\backslash \mathcal{M}^A_G$, with multiplication given by
$$[K,\phi]_G\cdot [L,\psi]_G=\sum_{KsL\in K\backslash G/ L}\bigl[K\cap {^sL},\phi|_{K\cap {^sL}}{^s\psi}|_{K\cap {^sL}}\bigr]_G$$
and with identity element $[G,1]_G$.
\smallskip
The group $G$ acts by conjugation on the ring $\prod_{K\leq G}\mathbb{Z} \mathrm{Hom}(K,A)$, and the \textit{ghost ring} of $B^A(G)$ is the subring $\widetilde{B^A}(G)=\left(\prod_{K\leq G}\mathbb{Z} \mathrm{Hom}(K,A)\right)^G$.
Here, we consider $\mathrm{Hom}(K,A)$ as an abelian group via pointwise multiplication and $\mathbb{Z}\mathrm{Hom}(K,A)$ as the associated group ring.
The map
$$\Phi^A_G\colon B^A(G)\longrightarrow \widetilde{B^A}(G)\,,\quad [L,\psi]_G\mapsto
\left(\sum_{\phi\in\mathrm{Hom}(K,A)}\gamma^G_{(K,\phi),(L.\psi)}\phi\right)_{K\leq G}\,,$$
is an injective ring homomorphism known as the \textit{mark morphism},
where
$$\gamma^G_{(K,\phi),(L,\psi)}=|\{sL\in G/L\;|\; (K,\phi)\leq {^s(L,\psi)} \}|\,,$$
for $(K,\phi),(L,\psi)\in \mathcal{M}^A_G$. Several properties of these numbers are listed in \cite[Section 1]{Bolt} and in \cite[Lemma 2.2]{Gar1}.
\smallskip
The natural projection map $\pi_{[\mathcal{S}_G]}:\prod_{K\leq G}\mathbb{Z} \mathrm{Hom}(K,A)\longrightarrow \prod_{K\in [\mathcal{S}_G]}\mathbb{Z} \mathrm{Hom}(K,A)$ is a ring homomorphism and injective when restricted to $\widetilde{B^A}(G)$; this is easily seen by considering the $\mathbb{Z}$-basis of $\widetilde{B^A}(G)$ given by the elements $\widetilde{\phi}$ for $[K,\phi]_G\in G\backslash\mathcal{M}^A_G$ defined in \cite[Eq. 6]{Gar1}. For $L\le G$, the $L$-entry of $\widetilde{\phi}$ is equal to the sum of the distinct $N_G(\lexp{g}{K})$-conjugates of $\lexp{g}{\phi}$, if there exists $g\in G$ with $L=\lexp{g}{K}$ and $0$ otherwise.
Moreover, one has
$$\widetilde{B^A}(G)\cong \pi_{[\mathcal{S}_G]}\left(\widetilde{B^A}(G)\right)=\prod_{K\in [\mathcal{S}_G]}\left(\mathbb{Z} \mathrm{Hom}(K,A)\right)^{N_G(K)}.$$
We set $\overline{B^A}(G)=\prod_{K\in [\mathcal{S}_G]}\left(\mathbb{Z} \mathrm{Hom}(K,A)\right)^{N_G(K)}$, $\overline{\phi}=\pi_{[\mathcal{S}_G]}\left(\widetilde{\phi}\right)$ and $\overline{\Phi}^A_G=\pi_{[\mathcal{S}_G]}\Phi^A_G$ for simplicity.
Thus, if $K\in[\mathcal{S}_G]$ and $\phi\in\mathrm{Hom}(K,A)$ then the $K$-entry of $\overline{\phi}\in\overline{B^A}(G)$ is the sum of the distinct $N_G(K)$-conjugates of $\phi$ and all the other entries of $\overline{\phi}$ are equal to $0$.
\smallskip
If $H$ is another finite group, a ring isomorphism $\Theta:B^A(G)\longrightarrow B^A(H)$ is called a \textit{species isomorphism} if $\Theta([K,\phi]_G)=[R,\rho]_H\in H\backslash \mathcal{M}^A_H$ for every $[K,\phi]_G\in G\backslash \mathcal{M}^A_G$ \cite[Def. 3.1]{Gar1}. We will make use of the following theorem which is part of the statement of Theorem~3.14 in \cite{Gar1}.
\begin{theorem}\thlabel{Previous Species Thm}
Let $A$ be an abelian group and let $G$ and $H$ be finite groups. There exists a species isomorphism from $B^A(G)$ to $B^A(H)$ if and only if there exist bijections $\theta_{\mathcal{S}}\colon \mathcal{S}_G\longrightarrow \mathcal{S}_H$ and $\theta_K\colon \mathrm{Hom}(K,A)\longrightarrow\mathrm{Hom}(\theta_{\mathcal{S}}(K),A)$, for $K\le G$, satisfying the following two conditions:
\smallskip
{\rm (a)} $\gamma_{(K,\phi),(L,\psi)}^G = \gamma_{(\theta_{\mathcal{S}}(K),\theta_K(\phi)), (\theta_{\mathcal{S}}(L),\theta_L(\psi))}^H$ for all $(K,\phi),(L,\psi)\in\mathcal{M}^A_G$.
\smallskip
{\rm (b)} The group homomorphism $\widetilde{\Theta}\colon \widetilde{B^A}(G) \longrightarrow \widetilde{B^A}(H)$ determined by $\widetilde{\phi}\mapsto \widetilde{\theta_K(\phi)}$, for $(K,\phi)\in\mathcal{M}^A_G$, is a ring isomorphism.
\end{theorem}
\begin{rem}\label{old remark}
In order to clarify the statement in Theorem~\ref{Previous Species Thm}, we will show that the map $\widetilde{\Theta}$ in (b) is well-defined, provided that the condition in (a) holds. That is, if $(K,\phi), (L,\psi) \in \mathcal{M}^A_G$ are $G$-conjugate, then $\widetilde{\theta_{L}(\psi)} = \widetilde{\theta_K(\phi)}\in \widetilde{B^A}(H)$. For this it suffices to show that $(\theta_{\mathcal{S}}(K),\theta_K(\phi))$ and $(\theta_{\mathcal{S}}(L), \theta_{L}(\psi))$ are $H$-conjugate. By part 2 of \cite[Lemma 2.2]{Gar1}, $(K,\phi)=_G(L,\psi)$ if and only if $0\neq \gamma^G_{(K,\phi),(L,\psi)}$ and $0\neq \gamma^G_{(L,\psi),(K,\phi)}$, then the condition in (a) implies that $(K,\psi)=_G(L,\psi)$ if and only if $(\theta_{\mathcal{S}}(K),\theta_K(\phi))=_H(\theta_{\mathcal{S}}(L), \theta_{L}(\psi))$. Moreover, if $K$ and $L$ are conjugate subgroups of $G$, then $(\theta_{\mathcal{S}}(K),\theta_K(1))=_H(\theta_{\mathcal{S}}(L),\theta_L(1))$, which implies that $\theta_{\mathcal{S}}(K)=_H\theta_{\mathcal{S}}(L)$. Conversely, if $\theta_{\mathcal{S}}(K)=_H\theta_{\mathcal{S}}(L)$, then $(K,\theta_K^{-1}(1))=_G(L,\theta_L^{-1}(1))$, implying $K=_GL$. Thus, $K=_GL$ if and only if $\theta_{\mathcal{S}}(K)=_H\theta_{\mathcal{S}}(L)$.
\end{rem}
Next we prove a slight modification of the above Theorem.
\begin{theorem} \thlabel{SpeciesThm}
Let $A$ be an abelian group and let $G$ and $H$ be finite groups. Then there exists a species isomorphism from $B^A(G)$ to $B^A(H)$ if and only if for any given $[\mathcal{S}_G]$ and $[\mathcal{S}_H]$, there are bijections $\theta_{[\mathcal{S}]}\colon [\mathcal{S}_G]\longrightarrow [\mathcal{S}_H]$ and $\theta_K\colon\mathrm{Hom}(K,A)\longrightarrow \mathrm{Hom}(\theta_{[\mathcal{S}]}(K),A)$ for $K\in [\mathcal{S}_G]$, satisfying the following two conditions:
\smallskip
{\rm (a)} $\gamma^H_{\left(\theta_{[\mathcal{S}]}(K),\theta_{K}(\phi)\right),\left(\theta_{[\mathcal{S}]}(L),\theta_L(\psi)\right)}=\gamma^G_{\left(K,\phi\right),\left(L,\psi\right)}$ for all $K,L\in [\mathcal{S}_G]$, $\phi\in \mathrm{Hom}(K,A)$ and $\psi\in \mathrm{Hom}(L,A)$.
\smallskip
{\rm (b)} The map $\overline{\Theta}\colon \overline{B^A}(G) \longrightarrow \overline{B^A}(H)$, $\overline{\phi}\mapsto \overline{\theta_K(\phi)}$, for $\phi\in \mathrm{Hom}(K,A)$ and $K\in[\mathcal{S}_G]$, is a ring isomorphism.
\end{theorem}
\begin{rem}\label{new remark}
With the same arguments as in Remark~\ref{old remark} one can show that in the situation of Theorem~\ref{SpeciesThm} the condition in (a) implies that the map $\overline{\Theta}$ in (b) is well-defined. More precisely, (a) implies that for any $K\in[\mathcal{S}_G]$, $\phi,\psi\in\mathrm{Hom}(K,A)$, after setting $K':=\theta_{[\mathcal{S}]}(K)$, $\phi':=\theta_K(\phi)$, $\psi':=\theta_K(\psi)$, one has $(K,\phi)=_G(K,\psi)$ if and only if $(K',\phi')=_H(K',\psi')$.
\end{rem}
\medskip\noindent
{\it Proof of Theorem~\ref{SpeciesThm}.}\quad
First suppose that there exists a species isomorphism from $B^A(G)$ to $B^A(H)$. Then there exist bijections $\theta_{\mathcal{S}}$ and $\theta_K$ for $K\le G$ as in Theorem~\ref{Previous Species Thm}, satisfying the conditions (a) and (b) in Theorem~\ref{Previous Species Thm}. Let $[\mathcal{S}_G]$ and $[\mathcal{S}_H]$ be given. Then for every $K\in[\mathcal{S}_G]$ there exists $h_K\in H$ such that $\lexp{h_K}{\theta_{\mathcal{S}}(K)}\in[\mathcal{S}_H]$. We replace the bijection $\theta_{\mathcal{S}}$ by the map $L\mapsto \lexp{h_K}\theta_{\mathcal{S}}(L)$ whenever $L=_GK$. This map is again a bijection, since $\theta_{\mathcal{S}}$ maps the $G$-conjugacy class of $K$ bijectively onto the $H$-conjugacy class of $\theta_{\mathcal{S}}(K)$, by Remark~\ref{old remark}. Moreover, for any $g\in G$, we replace $\theta_{\lexp{g}{K}}$ by $c_{h_K}\circ\theta_{\lexp{g}{K}} \colon \mathrm{Hom}(\lexp{g}{K},A)\longrightarrow \mathrm{Hom}(\lexp{h_K}{\theta_{\mathcal{S}}(^gK)},A)$. Then these new bijections satisfy again conditions (a) and (b) in Theorem~\ref{Previous Species Thm}. In fact, the map $\widetilde{\Theta}$ has stayed the same. Now we can restrict $\theta_{\mathcal{S}}\colon \mathcal{S}_G\to\mathcal{S}_H$ to the bijection $\theta_{[\mathcal{S}]}\colon [\mathcal{S}_G]\to[\mathcal{S}_H]$, and for every $K\in[\mathcal{S}_G]$ we choose the given bijection $\theta_K$. Then condition (a) in Theorem~\ref{SpeciesThm} is immediately satisfied and condition (b) holds, since with $\widetilde{\Theta}$ also $\overline{\Theta}=\pi_{[\mathcal{S}_H]}\widetilde{\Theta}\pi_{[\mathcal{S}_G]}^{-1}$ is a ring isomorphism.
Conversely, for each $[K,\phi]_G\in G\backslash\mathcal{M}^A_G$ we can assume $K\in [\mathcal{S}_G]$, and mapping $[K,\phi]_G$ to $[\theta_{[\mathcal{S}]}(K),\theta_K(\phi)]_H$ gives a bijection from $G\backslash \mathcal{M}^A_G$ onto $H\backslash \mathcal{M}^A_H$, thus it extends to an isomorphism of abelian groups $\Theta:B^A(G)\longrightarrow B^A(H)$. Then the diagram
$$\xymatrix{
B^A(G)\ar[rr]^-{\Theta}\ar[d]_{\overline{\Phi}^A_G} &&B^A(H)\ar[d]^{\overline{\Phi}^A_H}\\
\overline{B^A}(G) \ar[rr]_{\overline{\Theta}} &&\overline{B^A}(H)
}$$
commutes, and since the bottom and the vertical arrows are injective ring homomorphisms, $\Theta$ is a ring isomorphism.
\qed
\bigskip
As remarked in \cite[Cor. 3.12]{Gar1}, some of the $\theta_K$ are necessarily group isomorphisms, but we ignore whether this has to be the case for all these maps. However, when the $\theta_K$ are isomorphisms, we can drop Condition~(b) in Theorem~\ref{SpeciesThm}.
\begin{prop}\thlabel{SpeciesCriterion}
Let $A$ be an abelian group and let $G$ and $H$ be finite groups. Assume that there is a bijection $\theta_{[\mathcal{S}]}:[\mathcal{S}_G]\longrightarrow [\mathcal{S}_H]$ and, for each $K\in[\mathcal{S}_G]$, a group isomorphism $\theta_K\colon\mathrm{Hom}(K,A)\longrightarrow \mathrm{Hom}(\theta_{\mathcal{S}}(K),A)$ such that
$$\gamma^G_{(K,\phi),(L,\psi)}=\gamma^H_{(\theta_{[\mathcal{S}]}(K),\theta_{K}(\phi)),(\theta_{[\mathcal{S}]}(L),\theta_{L}(\psi))}$$
for all $K,L\in [\mathcal{S}_G]$, $\phi\in \mathrm{Hom}(K,A)$, $\psi\in \mathrm{Hom}(L,A)$. Then the assignment $[K,\phi]_G\mapsto [\theta_{[\mathcal{S}]}(K),\theta_K(\phi)]_H$ for $K\in [\mathcal{S}_G]$ and $\phi\in\mathrm{Hom}(K,A)$ extends to a species isomorphism $\Theta\colon B^A(G) \longrightarrow B^A(H)$.
\end{prop}
\begin{proof}
For $K\in [\mathcal{S}_G]$ and $\phi\in \mathrm{Hom}(K,A)$, the assingment $\overline{\phi}\mapsto \overline{\theta_K(\phi)}$ extends to an isomorphism $\overline{\Theta}:\overline{B^A}(G)\longrightarrow \overline{B^A}(H)$ of abelian groups. Then the diagram
$$\xymatrix{
\overline{B^A}(G)\ar@{^{(}->}[d]\ar[rr]^{\overline{\Theta}} &&\overline{B^A}(H)\ar@{^{(}->}[d]\\
\prod_{K\in[\mathcal{S}_G]}\mathbb{Z} \mathrm{Hom}(K,A)\ar[rr]_{(\theta_{K})} &&\prod_{L\in[\mathcal{S}_H]}\mathbb{Z} \mathrm{Hom}(L,A)\\
}$$
where $\theta_K\colon\mathbb{Z} \mathrm{Hom}(K,A)\longrightarrow \mathbb{Z} \mathrm{Hom}(\theta_{\mathcal{S}}(K),A)$ is the $\mathbb{Z}$-linear extension of $\theta_K$ and the vertical arrows are the inclusions, commutes by Remark~\ref{new remark}. Note that the bottom map is a ring isomorphism, since each $\theta_K$ was a group isomorphism. Therefore, as the bottom and the vertical arrows are injective, also $\overline{\Theta}$ is a ring isomorphism. By \thref{SpeciesThm} and its proof, the maps $\theta_{[\mathcal{S}]}$ and $\theta_K$ determine a species isomorphism.
\end{proof}
\section{Nontrivial counterexamples}
We recall the construction of Thévenaz' counterexamples in \cite{Thev}. Let $p$ and $q\geq 3$ be prime numbers such that $q|(p-1)$, and take elements $a\neq b$ of order $q$ in $(\mathbb{Z}/p\mathbb{Z})^{\times}$. Let $P_a=\mathbb{Z}/p\mathbb{Z}=\langle x\rangle$, $P_b= \mathbb{Z}/p\mathbb{Z}=\langle y\rangle$ and $Q=C_q=\langle z\rangle$, let $Q$ act on $P_a\oplus P_b$ by $\lexp{z}{x} = ax$ and $\lexp{z}{y}=by$, and consider the resulting semidirect product $G(a,b)=(P_a\oplus P_b)\rtimes Q$. A complete set of representatives of the conjugacy classes of subgroups of $G(a,b)$ is $\{1\}$, $P_a$, $P_b$, $P(j) = \langle x+jy\rangle$ for $j\in [(\mathbb{Z}/p\mathbb{Z})^{\times}/\langle a\rangle]$, $P_a\oplus P_b$, $Q$, $P_a\rtimes Q$, $P_b\rtimes Q$ and $G(a,b)$. Taken in this order, the table of marks of $G(a,b)$ is independent of the choice of $\{a,b\}$. For fixed $p$ and $q$, there are precisely $\frac{q-1}{2}$ isomorphism classes of these groups; in fact, if also $c\neq d$ are elements of order $q$ in $(\mathbb{Z}/p\mathbb{Z})^{\times}$ then $G(a,b)\cong G(c,d)$ if and only if there exists $n\in\{1.\ldots,q-1\}$ such that $\{c,d\}=\{a^n,b^n\}$ (see \cite{Thev}). Note that there are infinitely many choices for such $p$ and $q$: taking any prime $q\geq 3$, then by Dirichlet's theorem there are infinitely many primes $p$ in the arithmetic progression $1+q,1+2q,1+3q,\ldots$.
\begin{theorem} \thlabel{ThevenazGroups}
Let $p$ and $q\ge3$ be a primes with $q$ dividing $p-1$ and let $a\neq b$ and $c\neq d$ be elements of order $q$ in $(\mathbb{Z}/p\mathbb{Z})^{\times}$. If $A$ has trivial $p$-torsion, then $B^A(G(a,b))$ and $B^A(G(c,d))$ are isomorphic rings.
\end{theorem}
\begin{proof}
Since we already know that these groups have isomorphic Burnside rings, we can assume that $A$ has elements of order $q$. For simplicity, set $G:=G(a,b)$ and $H:=G(c,d)$ and take $[\mathcal{S}_G]$ and $[\mathcal{S}_H]$ as in the first paragraph of this section. We let $\theta_{[\mathcal{S}]}:[\mathcal{S}_G]\longrightarrow [\mathcal{S}_H]$ be the obvious bijection inducing an isomorphism of the tables of marks and set $K':=\theta_{[\mathcal{S}]}(K)$ for $K\in [\mathcal{S}_G]$. Next we define group isomorphisms $\theta_K\colon\mathrm{Hom}(K,A)\longrightarrow \mathrm{Hom}(K',A)$ for $K\in[\mathcal{S}_G]$: if $K$ is a $p$-group so is $K'$, and $\mathrm{Hom}(K,A)$ and $\mathrm{Hom}(K',A)$ are trivial, hence there is only one choice for $\theta_K$; if $K$ is not a $p$-subgroup, then a homomorphism $\phi\colon K\longrightarrow A$ is determined by the value $\phi(z)$ which is either an element of order $q$ or $1$, and we define $\phi':=\theta_K(\phi)\colon K'\longrightarrow A$ by requiring $\phi'(z)=\phi(z)$.
\smallskip
We now compare $\gamma^G_{(K,\phi),(L,\psi)}$ and $\gamma^H_{(K',\phi'),(L',\psi')}$ for $K,L\in [\mathcal{S}_G]$, $\phi\in \mathrm{Hom}(K,A)$ and $\psi\in \mathrm{Hom}(L,A)$. First, since $\theta_{[\mathcal{S}]}$ preserves the marks, we have
$$\gamma^G_{(K,1),(L,1)}=|(G/L)^K|=|(H/L')^{K'}|=\gamma^H_{(K',1),(L',1)},$$
for all $K$ and $L$ in $[\mathcal{S}_G]$. Note that the subgroups $P_a$, $P_b$ and $P_a\oplus P_b$ are normal, while $N_G(P(j))=P_a\oplus P_b$, and $P_a\rtimes Q$, $P_b\rtimes Q$ and $Q$ are self-normalizing. Moreover, it is straightforward to verify that if $K\not\leq L$ then $K\not\leq_{G} L$, and since $\theta_{[\mathcal{S}]}$ preserves containments, then $\gamma^G_{(K,\phi),(L,\psi)}=\gamma^H_{(K',\phi'),(L',\psi')}=0$ whenever $K\not\leq L$.
\smallskip
Therefore, we are left with the case when $K\leq L$, $L$ is not a $p$-subgroup and $\psi\neq 1$, and we distinguish the following cases for $K$:
\smallskip
(i) If $K\in\{\{1\}, P_a, P_b, P_a\oplus P_b\}$ then $K\unlhd G$ and $\phi=1$. Therefore $K\leq {\lexp{g}{L}}$ and $\lexp{g}{\psi}|_K=1$ for all $g\in G$ and hence
$$\gamma^G_{(K,1),(L,\psi)}=[G:L]=[H:L']=\gamma^H_{(K',1),(L',\psi')}\,.$$
\smallskip
(ii) If $K=P(j)$ we may assume that $L=G$. In this case $\gamma_{(K,1),(G,\psi)}^G= 1 = \gamma_{(K',1),(H,\psi')}^H$.
\smallskip
(iii) If $K$ is not a $p$-subgroup, then since $L$ is self-normalizing and both $K$ and $L$ contain $Q$, we have that if $g\notin L$ then $z\not\in \lexp{g}{L}$, hence $K\not\leq \lexp{g}{L}$. We can conclude that
$$\gamma^{G}_{(K,\phi),(L,\psi)}=\begin{cases}
1 &\text{if}\;\phi(z)=\psi(z),\\
0 &\text{otherwise},
\end{cases}$$
and by the way we have defined $\theta_K$, we have that $\gamma^{H}_{(K',\phi'),(L',\psi')}=\gamma^{G}_{(K,\phi),(L,\psi)}$.
\smallskip
We conclude that $\theta_{[\mathcal{S}]}$ and the isomorphisms $\theta_K$ for $K\in [\mathcal{S}_G]$ satisfy the condition of \thref{SpeciesCriterion}, hence they determine a species isomorphism.
\end{proof}
\begin{rem}
The above theorem shows that Thévenaz' counterexamples to the isomorphism problem for the Burnside ring are also non-trivial counterexamples for the $A$-fibered Burnside ring if $A$ is any abelian group with trivial $p$-torsion and having elements of order $q$. In particular, we have a negative answer to the isomorphism problem of the $C_q$-fibered Burnside ring for any prime $q\geq 5$, and for the ring of monomial representations over any field $k$ of characteristic $p$, since in this case $D^k(G(a,b))\cong B^{C_q}(G(a,b))$. However, the existence of non-isomorphic finite groups $G$ and $H$ with $D^{\mathbb{C}}(G)\cong D^{\mathbb{C}}(H)$ remains open.
\end{rem}
|
\section{Acknowledgements}
We thank Rodrigo Corti\~nas, Brice Ravon and Igor Dotsenko for experimental support, Fernando Gago for fruitful discussions. We acknowledge support from the European Union's Horizon 2020 under grant agreement No 786919 (TRENSCRYBE) and from the Agence Nationale de la Recherche under project 167754 (SNOCAR). H.W. acknowledges a China Scholarship Council No 201806190206.
\section{Supplementary Information}
We describe the preparation of the initial states, the experimental details, the model and CST calculations and the fitting procedures.
\subsection{Preparation of initial states and experimental details}
The circular state preparation starts with a stepwise excitation process using three resonant photons with wavelengths 780~nm ($5S_{1/2}\rightarrow 5P_{3/2}$ transition), 776~nm ($5P_{3/2}\rightarrow 5D_{5/2}$) and 1257~nm ($5D_{5/2}\rightarrow58f$). The three lasers illuminate the MOT cloud simultaneously for 1~$\mu$s at the start of each experimental sequence. During the laser excitation, a small electric field (5~V/m) aligned with the $Oz$ quantization axis, lifts the degeneracy between the $58f$ sublevels with different $|m|$. The frequency and the polarization of the 1257~nm beam are chosen to only excite the $58f$ level with $m=2$.
After the laser excitation, the field is raised to 195~V/m. The atoms are brought into the $58C$ circular state by an rapid adiabatic passage induced by a $\sigma^+$-polarized rf pulse at 178~MHz, applied while the field is ramped down from 195~V/m to 95~V/m in 6~$\mu$s~\cite{SignolesCoherentTransferLowAngularMomentum2017}. At the end of the `circularization' process, the field is set at 160~V/m.
The purity of the prepared state is not perfect. The circularization is thus followed by a combination of level-selective MW transitions towards the target state, rf transitions within Stark manifolds and partial ionization to increase the purity:
\begin{enumerate}
\item[(i)] For the circular states $nC$ with $n=57-54$, the initial $58C$ level is transferred to the target state by one or two MW $\pi$-pulses (duration $\approx 1 \ \mu$s) selectively addressing the circular-to-circular transitions. The atoms left in Rydberg manifolds with $n' > n$ are then removed by a partial field-ionization ramp that does not affect the target states (circular states remain circular up to their ionization limit).
\item[(ii)] For $60C$, a two-photon $\pi$-MW pulse first transfers $58C$ to $60C$. The spurious non-circular high-$m$ levels left in the 58 manifold are then driven by a 1~$\mu$s rf pulse towards low-$m$ states in the 58 manifold. This rf pulse is resonant with the transitions between Stark levels within the 58 manifold but not resonant with the ones in the 60 manifold. It thus does not perturb $60C$. Afterwards, the low-$m$ atoms of 58 manifold, which have a lower ionization threshold than $60C$, are removed by partial ionization.
\item[(iii)] For $58C$, in order to have a higher purity than that provided directly by the circularization processs, we prepare $60C$ as explained in (ii) and transfer the atoms back to $58C$ with a two-photon MW $\pi$-pulse. A partial ionization ramp clears all remaining atoms with $n'>58$.
\item[(iv)] Finally, the MW source limitations prevent us from preparing $59C$ from $60C$. We thus prepare the $59C$ state by a single $\pi$-pulse from the purified $58C$ prepared as in (iii).
\end{enumerate}
The MOT lasers are always present during the measurement sequence. They do not contribute to the measured lifetime, because circular states have no optical transitions. Photoionization of circular states in these weak laser powers is expected to be negligible~\cite{NguyenQuantumSimulationCircular2018}. The gradient of the MOT magnetic field is estimated to be on the order of 10~Gauss/cm. The induced Zeeman effect is always much lower that the Stark effect.
Dipole-Dipole Rydberg-Rydberg interactions might be an important perturbation, by inducing mixing with non-circular Rydberg states. To rule out such an effect, we reduced the number of atoms initially excited in a Rydberg state by a factor of 5. We did not observe any change in the measured lifetime within statistical errors.
The compact design of our setup requires the Rb dispenser to be placed very close to the inhibition capacitor. Its operating temperature (700 K) probably makes the effective BBR temperature higher than 300 K. This effect might contribute to the discrepancy between the calculated and observed transition rates.
\subsection{Model and CST calculations}
The modification of the transition rates within the inhibition capacitor is entirely described by the altered classical electromagnetic field-mode density between the plates~\cite{HarocheCavityQuantumElectrodynamics1992, HindsCavityQuantumElectrodynamics1990}. For a numerical estimation of this classical effect, we compute the power emitted by a classical dipole inside a detailed model of the capacitor, using the CST Studio software suite. The 3D model of the capacitor structure is shown in Fig. 5 (a). The yellow base is the capacitor plate. It is electrically isolated from the 4 rf electrodes above it (also in yellow). All these elements, gold-plated in the set-up, are assumed to have the conductivity of gold at 300 K. The transparent top plate is lead glass material. We mimic the effect of ITO coating with a 2D ohmic sheet covering the whole bottom of this plate with a direct current (dc) sheet resistance of $8 \ \Omega/\text{square}$, matching the value directly measured on the ITO coating by a 4-terminal dc method (the ITO impedance at MW frequency is not expected to be drastically different from its dc value~\cite{Meinert_Indium_2020}).
We compute the power radiated by a dipole antenna with a length on the order of the size of a Rydberg atom and a $\sigma$ or $\pi$-polarization placed at the position of the atoms [marked by a red cone in figure 5(a)]. For each frequency, we compare the total power radiated to the power radiated in free space. The ratio of these quantities directly gives the factor modifying the transition rates at that frequency. The results of this calculation over the frequency range of interest is shown as solid lines in Fig. 5(b). The black and red curves correspond respectively to the $\sigma$ and $\pi$ polarizations. For reference the rate modification factors for an ideal, infinite plane-parallel capacitor with the same 4.1~mm spacing is also represented (dashed lines). The theoretical rates shown in Fig. 3 are then computed using the calculated free-space rates of the relevant transitions at a 300~K background temperature.
\begin{figure}
\includegraphics[width=1\linewidth]{figure5.pdf}
\caption{ (a) 3D model of the experimental capacitor used for microwave simulations with the CST Studio softwares suite. The capacitor (4.1~mm spacing) is made up of the gold square electrode and of the resistive coating on the glass plate. Note that the surrounding glass cell is not included in the computation. (b) Ratio of the powers radiated by a classical dipole in the capacitor and in free space. Solid lines correspond to the model capacitor, dashed lines to an infinite ideal plane-parallel capacitor with 4.1~mm plates spacing. Black (red) corresponds to $\sigma$ ($\pi$) radiation. Dotted vertical lines mark circular-to-circular transition frequencies involved in experiment.}
\end{figure}
\subsection{Fitting procedures}
To estimate the transition rates between Rydberg levels, we use a \emph{single} fit of a rate equation model over 16 independent datasets [Figure 2(b) shows only 4 out of these 16 datasets]. They correspond to 16 different experiments that differ in either the chosen initial state (among the seven possible ones, $54C$--$60C$) or the delay range (0 to 150, 0 to 300 or 0 to 900~$\mu$s).
Sequences with smaller delay range are required for a better description of the population transfers at short times, particularly for the non-inhibited transitions.
Figure 6 shows the rates taken into account in the model. For strongly inhibited states ($n \ge 57$), we use two independent fit parameters for the $n\rightarrow n+1$ and the $n+1 \rightarrow n$ rates (red and blue arrows). For levels $n=57$ to $53$, there is little inhibition. As explained in the main text, we thus use the same rate for upward and downward transfers (double headed purple arrows). We have numerically checked this assumption in additional fits. Note that already for $n=57$, the fit result for the two independent rates $58\rightarrow 57$ and $57\rightarrow 58$ are very close ($710\pm39$~Hz and $715\pm20$~Hz respectively).
Finally, we use two independent parameters for the rates $60\leftrightarrow$ ``high-$n$" and $50\leftrightarrow$ ``low-$n$". Transition back from the ``high-$n$" or ``low-$n$" windows do not represent true inter-level transitions but transitions from detection slots gathering many levels. As a result, they are not represented on Fig. 3.
In addition to the rates, we also use as fit parameters, for each dataset, the initial populations of all relevant levels at zero delay. The fit is finally performed using the Levenberg-Marquardt algorithm of the SciPy library.
\begin{figure}
\includegraphics[width=1\linewidth]{figure6.pdf}
\caption{Level scheme for the rate equation model used for the fit. Each arrow represents an independent parameter of the fit.}
\end{figure}
In order to estimate the statistical error bars on the rates (shown in Fig. 3), we split each of the 16 datasets in 5 independent sets. We perform separate fits on the resulting 5 subsets. We thus obtain 5 independent sets of rates. The standard errors of these sets provide the error bars in figure 3.
The lifetimes presented in Fig. 4 are calculated from the rates in figure 3. The error bars on these lifetimes are obtained from error propagation of the statistical rates error bars.
Note that we have performed other fits, with rate equation models based on different transitions schemes, taking into account for instance explicitly sub-sets of the elliptical states in each manifold or direct transitions between $nC$ and the $n+2$ manifolds. The estimated lifetimes always agreed with those of Fig. 4 within the statistical error bars.
|
\section{Introduction}
Our motion through the Universe generates a dipole in both the temperature anisotropies of the Cosmic Microwave Background (CMB) \citep{2020A&A...641A...6P} and in the angular distribution of electromagnetic sources \citep{2017MNRAS.471.1045C,2018JCAP...04..031B,2021ApJ...908L..51S,2021A&A...653A...9S,Secrest:2022uvx}. If the cosmological principle is valid, these two measures should have consistent values. However, it is a longstanding problem that number counts of radio sources and of quasars at low and intermediate redshifts exhibit a dipole that is well aligned with that of the CMB but with an amplitude which is 2-5 times larger than expected, leading to a tension reaching up to $\sim 5 \sigma$. In \citet{Dalang:2021ruy} it is argued that this tension might be alleviated once one takes into account the redshift evolution of the population of sources, and the value that the evolution rate should have in order to remove the tension is found.
Gravitational-wave (GW) sources observed at cosmological distances can shed light on the cosmic dipole problem. Since the first historical detection of~\citet{LIGOScientific:2016aoc} during the first scientific observation run (O1), the rate of observed GW events has drastically increased, to reach roughly one detection per week in the O3 run. In total, about 90 binary black hole (BBH) coalescences, as well as two binary neutron star (BNS) and two neutron star - black hole mergers have been detected so far~\citep{2021arXiv211103606T}. Thanks to this abundance of new data, which will further accumulate in the forthcoming future \citep{2018LRR....21....3A}, it is possible to probe cosmology \citep{2021ApJ...909..218A,https://doi.org/10.48550/arxiv.2111.03604,2022PhRvD.105f4030M,2022arXiv220200025L} and astrophysical rates of compact binary coalescences \citep{2019ApJ...882L..24A,2021ApJ...913L...7A,2021arXiv211103634T} using GW observations.
It is thus not surprising that recently, it has been proposed to use GW sources as a new and independent probe to measure the cosmic dipole. A first approach proposed to measure the cosmic dipole is by studying anisotropies in the GW stochastic background \citep{Cusin:2022cbb, DallArmi:2022wnq, LISACosmologyWorkingGroup:2022kbp}. However, stochastic GW backgrounds have not been detected yet.
Another possibility, is to study the sky distribution of the transient GW sources currently detected \citep{2018PhRvD..97j3005C,2022arXiv220407472K,2022arXiv220705792E}. Although GW sources are the central paradigm of all these works, the methods employed significantly differ. In \citet{2018PhRvD..97j3005C}, the authors perform a forecast on the accuracy with which the cosmic dipole would be detected by fitting for the (modified) luminosity distance distribution of GW events. In \citet{2022arXiv220407472K}, the authors try instead to constrain the dipole anisotropy using the BBHs mass distribution from current GW events \citep{2021arXiv211103606T} finding that the mean mass is higher in the direction of the CMB dipole and invoking the need of further study to understand the origin of their findings. In contrast to \citet{2021MNRAS.501..970S}, \citet{2022arXiv220407472K}, \citet{2022arXiv220705792E} find no-evidence of anisotropies, using 63 GW sources distributed over the sky, and implementing a hierachical Bayesian analysis that takes into account selection biases.
In our work, we focus on the detection and estimation of the cosmic dipole with Einstein Telescope (ET) \citep{2010CQGra..27s4002P} and Cosmic Explorer (CE) \citep{Reitze:2019iox}.
These detectors, along with LISA \citep{2017arXiv170200786A}, belong to the so-called next generation (XG) of GW detectors and represent the aim of the community to substantially scale-up the experience of the LIGO-Virgo era. In this paper, we show that the number counts of BBH sources will likely offer an optimal tool to detect the cosmic dipole. We build an estimator, that has the advantage of being independent of unknown characteristics of the GW sources (or their evolution) and that will therefore allow us to infer our motion with respect to the cosmological frame in a robust way.
This paper is structured as follows. In Sec.~\ref{sec:2}, we derive the dipole modulation in the GW sources number counts, due to the observer velocity with respect to the cosmological frame.
We show in this section that GW number counting of BBHs offers a clean tool to evaluate anisotropies.
In Sec.~\ref{sec:3}, we simulate GW detections using a detector network composed by ET and two CEs (ET+2CE). We discuss detection capabilities and the measurement process using a frequentist and Bayesian approach. In Sec.~\ref{sec:4}, we discuss possible limiting factors of our approach in light of the possibility of constraining the cosmic dipole with GW sources. Finally, in Sec.~\ref{sec:5} we draw our conclusions.
\section{Modeling the impact of the observer velocity on GW number counts}
\label{sec:2}
We write down explicitly the expression for the cosmic dipole in the GW source distribution. In Sec.~\ref{sec:2.1}, we derive the theoretical modeling for the dipole induced by the observer velocity in the number counts of GW signals emitted by compact binary coalescences. In Sec.~\ref{sec:2.3}, we define a statistical estimator for the cosmic dipole.
\subsection{Theoretical framework}
\label{sec:2.1}
We define the number of GW sources $N_{\rm det}$ detectable per unit solid angle $d\Omega$ and distance bin $dr$, in direction ${\bf{n}}$ at comoving distance $r$ and with signal-to-noise ratio (SNR) $\rho$ larger than a given threshold $\rho_*$ as
\begin{equation}\label{int}
\frac{d N_{\rm det}}{d\Omega dr}(r, {\bf{n}}, \rho>\rho_*)\equiv \int_{\rho_*}^{\infty} d\rho \frac{d N}{d\Omega dr}(r, {\bf{n}}, \rho)\,.
\end{equation}
This number depends on the direction of observation ${\bf{n}}$ due to three effects. First, the GW sources are not perfectly homogeneously distributed: they live in galaxies which follow the large-scale structure of the Universe. Second, the propagation of GWs is affected by inhomogeneities along the trajectory, that change their apparent distribution over the sky, see e.g.~\cite{Cusin:2017fwz} in the context of stochastic background. And finally, the motion of the observer with respect to the source rest frame generates a further anisotropy in the observed distribution. In this work, we are interested in the latter effect, which gives a dipolar modulation in the distribution of sources. The other two effects will also have a dipolar contribution, but, as has been shown for quasars~\citep{2021ApJ...908L..51S}, this clustering dipole is expected to be negligible compared to the kinematic dipole.
The cosmic dipole averaged over all distances $r$ is obtained from the difference between Eq.~\eqref{int} and its angular average, integrated over $r$. Namely,
\begin{align}
&\Delta({\bf{n}})\equiv \mathcal{D} [{\bf{n}}\cdot \hat{{\bf{v}}}_o]\nonumber\\
&=\dfrac{\int dr \left[\frac{dN_{\rm det}}{d\Omega dr}(r, {\bf{n}}, \rho>\rho_*)-\frac{d\bar{N}_{\rm det}}{d\Omega dr}(r, \rho>\rho_*)\right]}{\int dr \frac{d\bar{N}_{\rm det}}{d\Omega dr}(r, \rho>\rho_*)}\,,
\label{dipole0}
\end{align}
where ${\bf{v}}_o$ denotes the observer velocity and $\hat{{\bf{v}}}_o\equiv {\bf{v}}_o/|{\bf{v}}_o|$ is its direction. The angular average over the sky, denoted with a bar, is given by
\begin{equation}
\frac{d\bar{N}_{\rm det}}{d\Omega dr}(r, \rho>\rho_*) \equiv \frac{1}{4\pi} \int d\Omega \frac{dN_{\rm det}}{d\Omega dr}(r, {\bf{n}}, \rho>\rho_*)\,.
\end{equation}
Eq.~\eqref{int} depends on the observer velocity through two effects. First, the observed solid angle is affected by aberration:
\begin{equation}
d\Omega=d\bar{\Omega}\left(1-2 {\bf{n}}\cdot \frac{{\bf{v}}_o}{c}\right)\,.
\label{eq:domegabar}
\end{equation}
And second, the SNR threshold at the observer, $\rho_*$, (which is a fixed number) corresponds to different emitted GW power for sources situated in different directions. To account for this, we factorise the SNR into an intrinsic part, $I$, which is related to the intrinsic properties of the source and the detector, and a propagation part, $P(r,{\bf{n}})$, which accounts for the impact of GWs propagation on the SNR
\begin{align}
\rho^2(r,{\bf{n}})=I\times P(r,{\bf{n}}) \, .
\end{align}
A source sitting at position $(r,{\bf{n}})$ with an SNR above threshold, $\rho(r,{\bf{n}})>\rho_*$, must have intrinsic properties $I>I_*(r,{\bf{n}})$ where
\begin{align}
I_*(r,{\bf{n}})\equiv \frac{\rho_*^2}{P(r,{\bf{n}})}\, . \label{eq:Istar}
\end{align}
The fixed SNR threshold $\rho_*$ corresponds therefore to different intrinsic properties $I_*(r,{\bf{n}})$ in different directions. Using that the observer velocity is small compared to the Hubble flow, we can then Taylor expand Eq.~\eqref{int} around the homogeneous background. We obtain
\begin{align}
\frac{d N_{\rm det}}{d\Omega dr}(r, {\bf{n}}, \rho>\rho_*)
&=\frac{d N_{\rm det}}{d\Omega dr}(r, {\bf{n}}, I>I_*(r,{\bf{n}}))\nonumber\\
&\simeq\frac{d N_{\rm det}}{d\bar{\Omega }dr}(r, I>\bar{I}_*(r))\left(1-\frac{\delta \Omega}{\delta \bar{\Omega}}\right)\nonumber\\
&+\frac{\partial }{\partial I_*}\left(\frac{d N_{\rm det}}{d\Omega dr}(r, I>I_*)\right)_{I_*=\bar{I}_*}\delta I_*(r,{\bf{n}})\,. \label{eq:Npert}
\end{align}
To compute $\delta I_*(r,{\bf{n}})$ we need a model for the SNR. For a binary system of compact objects, the SNR at the zero Post-Newtonian (0PN) order is given by~\citep{Finn:1992xs}
\begin{equation}\label{Finn1}
\rho^2(r,{\bf{n}}, m, \mathcal{M})= \frac{5}{96\pi^{4/3}} \frac{\Theta^2}{D_L^2(r,{\bf{n}})} (G \mathcal{M}_z)^{5/3} \mathcal{F}(f^z_{\rm ISCO}(m))\,,
\end{equation}
where $\mathcal{M}_z=\mathcal{M}(1+z)$ is the \textit{redshifted} chirp mass of the system, $m=m_1+m_2$ is the total mass, $\Theta^2$ is a geometrical factor that depends on the inclination of the binary and on the antenna pattern of the detector, and $D_L$ denotes the luminosity distance. The $\mathcal{F}$ quantifies the sensitivity of the GW detector, namely
\begin{equation}\label{f731}
\mathcal{F}(f^z_{\rm ISCO}(m))\equiv \int_0^{2f^z_{\text{ISCO}}} df\left[f^{7/3} S_n(f)\right]^{-1}\,,
\end{equation}
where $S_n(f)$ is the detector Power Spectral Density (PSD).
In Eq.~\eqref{f731} the upper integration bound is given by the redshifted frequency corresponding to the innermost stable circular orbit ($\text{ISCO}$) of the system, i.e. the frequency at which we consider the inspiraling phase of the system to end in our approximation. It is defined as \citep{Maggiore:1900zz}
\begin{equation}\label{ISCO}
f_{\text{ISCO}}\equiv\frac{1}{6\sqrt{6}(2\pi)}\frac{c^3}{Gm}\simeq 2.2. \text{kHz}\left(\frac{M_{\odot}}{m}\right)\,,
\end{equation}
and $f^z_{\text{ISCO}}=f_{\text{ISCO}}/(1+z)$.
Note that at this frequency the 0PN approximation in Eq.~\eqref{Finn1} used for the SNR estimate becomes inaccurate.
We now define the intrinsic part of the SNR as
\begin{align}
I(\mathcal{M},m)\equiv \frac{5}{96\pi^{4/3}} \Theta^2 (G \mathcal{M})^{5/3} \mathcal{F}(f_{\rm ISCO}(m))\, ,
\end{align}
and the propagation part is given by
\begin{align}
P(r,{\bf{n}})\equiv \frac{(1+z(r,{\bf{n}}))^{5/3}}{D_L^2(r,{\bf{n}})}\frac{\mathcal{F}(f^z_{\rm ISCO}(m,r,{\bf{n}}))}{\mathcal{F}(f_{\rm ISCO}(m))}\, . \label{eq:P}
\end{align}
The propagation factor, $P$, is affected by the observer peculiar velocity through the redshift and through the luminosity distance. Note that the redshift perturbation also enter via the upper bound of the integral $f^z_{\rm ISCO}$, which reflects the fact that the observer velocity shifts the observed frequency of the ISCO. Inserting~\eqref{eq:P} into~\eqref{eq:Istar} and expanding at linear order in the velocity we obtain
\begin{align}
\delta I_*(r,{\bf{n}})=\bar{I}_*(r)&\Bigg[2\frac{\delta D_L}{\bar{D}_L}-\frac{5}{3}\frac{\delta z}{1+\bar{z}}\\
&+\frac{2f^{\bar{z}}_{\rm ISCO}}{\mathcal{F}(f^{\bar{z}}_{\rm ISCO})}
\frac{\left(2f^{\bar{z}}_{\rm ISCO}\right)^{-7/3}}{
S_n\left(2f^{\bar{z}}_{\rm ISCO}\right)}\frac{\delta z}{1+\bar{z}}\Bigg] \, .\nonumber
\end{align}
Using that the redshift pertubation and luminosity distance perturbations are given by
\begin{equation}\label{eq:pert}
\frac{\delta z}{1+\bar{z}}=-{\bf{n}}\cdot \frac{{\bf{v}}_o}{c}\,,\quad \mbox{and}\quad
\frac{\delta D_L}{\bar{D}_L}=-{\bf{n}}\cdot \frac{{\bf{v}}_o}{c}\,,
\end{equation}
we find
\begin{align}
\delta I_*(r,{\bf{n}})=-\bar{I}_*(r)&\Bigg[\frac{1}{3}+\frac{2f^{\bar{z}}_{\rm ISCO}}{\mathcal{F}(f^{\bar{z}}_{\rm ISCO})}\frac{\left(2f^{\bar{z}}_{\rm ISCO}\right)^{-7/3}}
{S_n\left(2f^{\bar{z}}_{\rm ISCO}\right)}\Bigg]{\bf{n}}\cdot \frac{{\bf{v}}_o}{c} \, . \label{eq:Ipert}
\end{align}
Finally, we need to compute the variation of the cumulative number of events above threshold:
\begin{equation}
\frac{\partial }{\partial I_*}\left(\frac{d N_{\rm det}(r,I>I_*)}{d\Omega dr}\right)_{I_*=\bar{I}_*}\,.
\end{equation}
Without loss of generality, we assume that GW events are distributed in a window of $I$, i.e.\ $I \in [I_{\text{min}}, I_{\text{max}}]$. The intrinsic part of the SNR, $I$, is indeed directly related to the chirp mass of the system, and to the total mass (through $f_{\rm ISCO}$). Both these quantities have a given distribution with finite width. Therefore only a range of values of $I$ are physical. We can write
\begin{align}
\label{eq:defs}
\frac{\partial }{\partial I_*}\left(\frac{d N_{\rm det}(r,I>I_*)}{d\Omega dr}\right)_{I_*=\bar{I}_*}=&-\frac{s(I_*)}{I_*}\frac{d N_{\rm det}(r,I>I_*)}{d\Omega dr}\,,
\end{align}
where the parameter $s$ is defined through
\begin{align}
s(I_*)=s(\rho_*,r)\equiv -\frac{\partial\ln \left(\frac{d N_{\rm det}(r,I>I_*)}{d\Omega dr}\right)}{\partial\ln I_*}\, . \end{align}
This parameter directly depends on the population of sources. It is non-zero only if $I_{\rm min}<I_*<I_{\rm max}$, or similarly if $\rho_{\rm min}(r)<\rho_*<\rho_{\rm max}(r)$, i.e. when we are dealing with a population of sources with a significant fraction of events across threshold:
\begin{align}
s(\rho_*,r)\propto \Theta(\rho_*-\rho_{\text{min}})\Theta(\rho_{\text{max}}-\rho_*)\, .
\end{align}
Inserting Eqs.~\eqref{eq:defs},~\eqref{eq:Ipert} and~\eqref{eq:domegabar} into~\eqref{eq:Npert}, and subtracting the angular average, we find for the dipole
\begin{align}
\mathcal{D}\left[{\bf n}\cdot{\bf v}_0\right]&={\bf{n}}\cdot \frac{{\bf{v}}_o}{c}\int_0^{\infty} dr f(r)\left[2+s(r,\rho_*)\left(\frac{1}{3}+\mathcal{A}(r)\right)\right] \nonumber\\
&\equiv\alpha\,{\bf{n}}\cdot \frac{{\bf{v}}_o}{c} \,,
\label{eq:dipole}
\end{align}
where we implicitly defined the parameter $\alpha$ that will be used later on in the analysis. The function $f(r)$ denotes the radial distribution of sources
\begin{equation}\label{fr}
f(r)\equiv\frac{\frac{dN}{drd\Omega}}{\int_0^{\infty} dr \frac{dN}{d\Omega dr}}\,,
\end{equation}
and $\mathcal{A}$ is given by the second term in Eq.~\eqref{eq:Ipert}, averaged over all sources at distance $r$:
\begin{align}
\mathcal{A}(r) \equiv & \int dm_{1} dm_{2}\, p(m_{1}, m_{2}) \frac{1}{\mathcal{F}\left(\frac{f_{\text{ISCO}}}{(1+z)}\right)} \nonumber \\
& \times\left(\frac{2f_{\text{ISCO}}}{(1+z)}\right)^{-7/3} S_n\left(\frac{2f_{\text{ISCO}}}{(1+z)}\right)^{-1}\frac{2f_{\text{ISCO}}}{(1+z)}\,.
\end{align}
Here $p(m_{1}, m_{2})$ is the probability density function (PDF) of the source-frame masses and $\mathcal{A}$ depends on $r$ through the redshift $z=z(r)$. We have checked that $\mathcal{A}$ is factor of $\mathcal{O}(1)$.
We observe that in Eq.~\eqref{eq:dipole}, the term proportional to $s$ is relevant only if a significant part of the population is across threshold.
This term is due to the third line of Eq.~\eqref{eq:Npert}.
For a population of sources where $s\simeq 0$, i.e.\ such that nearly all sources are detected, then $\alpha(r) \simeq 2$, i.e. it is a fixed constant, independent on distance.
In this case, the dipole can be used as a direct estimator of the observer velocity. More precisely, for a perfectly isotropic distribution of sources, i.e. with PDF $P(\Omega)=1/4\pi$, we can build the following observable (choosing $v_0$ aligned along the azimutal axis)
\begin{equation}
v_{{\bf n}'}\equiv \frac32\int {\rm d}\Omega\ P(\Omega)\ \left({\bf n} \cdot {\bf n}'\right)\ \mathcal{D}\left[{\bf n}\cdot {\bf v}_0\right]=\frac{\alpha v_0}{2c} \cos\theta'\simeq \frac{v_0}{c} \cos\theta'\,,
\label{eq:observable}
\end{equation}
where ${\bf n}'\equiv\left(\sin\theta'\ \cos\phi'\,,\sin\theta'\ \sin\phi'\,,\cos\theta'\right)$ is a vector pointing towards a generic fixed direction.
This observable is maximized when evaluated along the a priori unknown dipole direction and is exactly equal to the observer velocity, $v_0/c$, at the maximum.
\subsection{A statistical estimator for the cosmic dipole}
\label{sec:2.3}
Let us now build an estimator for the observable $v_{{\bf n}'}$. We divide the sky in $N_{\rm sky}$ pixels of same solid angle, and associate a vector ${\bf{n}}_i$ pointing to the center of each pixel. The estimator is given by
\begin{equation}
\label{eq:estimator}
\hat{v}_{{\bf n}'}=\frac3{2 N_{\rm tot}}\sum_{i=1}^{N_{\rm sky}} N_{\rm det}^{i}\cdot({\bf n}_i\cdot {\bf n}')\,,
\end{equation}
where $N_{\rm tot}$ is the total number of events and $N_{\rm det}^{i}$ is the number of sources that falls in the sky pixel $i$. This number can be written as
\begin{align}
\label{eq:Ni}
N_{\rm det}^{i}=\bar{N}_{\rm det}\left(1+\alpha\,{\bf{n}}_i\cdot\frac{{\bf{v}}_o}{c} \right) +\Delta N^i\,,
\end{align}
where $\bar{N}_{\rm det}=N_{\rm tot}/N_{\rm sky}$ is the mean number of events per pixel, and
$\Delta N^i$ accounts for the fact that
the actual distribution of detected GW signals will not be exactly isotropic (even in the absence of a dipole), because of the stochastic nature of the coalescence distribution and of the detection process. Inserting~\eqref{eq:Ni} in~\eqref{eq:estimator}, the expected value of the estimator becomes
\begin{align}
\langle\hat{v}_{{\bf n}'}\rangle&= \frac3{2 N_{\rm tot}}\alpha \frac{v_o}{c}\sum_{i=1}^{N_{\rm sky}}({\bf n}_i\cdot \hat{{\bf{v}}}_o)({\bf n}_i\cdot {\bf n}')=\frac{\alpha v_o}{2c} \cos(\theta')=v_{\bf n'}\, ,\label{eq:exp}
\end{align}
because the isotropic part of the actual sky distribution averages to zero due to the $({\bf n}_i\cdot {\bf n}')$ factor, and the expectation value of the stochastic noise is, by definition, zero. Hence what survives is exactly the dipole component discussed in the previous section, which reduces to the observer velocity when $\alpha=2$, i.e.\ when threshold effects are negligible.
In order to assess the statistical detectability of the dipole, one must determine the variance of the estimator, due to the stochastic fluctuations of events, $\Delta N^i$, around the monopole.
Since the pixels have associated the same solid angle, the number of detections contained in every pixel is drawn from the same distribution, that is a Poissonian with $\mu=\bar{N}_{\rm det}$. One can then determine the variance of $\hat{v}_{{\bf n}'}$ just by summing in quadrature the variances within each pixel:
\begin{equation}
\label{eq:variance}
\langle\Delta \hat{v}_{{\bf n}'}^2\rangle=\frac{9}{4N_{\rm tot}^2}\sum_{i=1}^{N_{\rm sky}}\bar{N}_{\rm det}\cos^2\theta_i \cos^2\theta'
\simeq\frac{3}{4N_{\rm tot}}\cos^2\theta'\, ,
\end{equation}
as the sum in $\cos^2\theta_i$ quickly converges to $N_{\rm sky}/3$, already for $N_{\rm sky}< 10$. We thus have the intuitive result that the precision attainable in the dipole measurement scales down with the square root of the total number of detections, as expected in an essentially poissonian process.
From Eq.~\eqref{eq:variance}, one expects that in order to detect a velocity dipole $v_0/c$ of order $10^{-3}$,
one needs a very large number of GW detections, $N_{\rm tot}\simeq 10^6$, which can be reached only with XG detectors.
Another way to obtain the same result is to artificially generate a large number of random sky distributions, for instance by randomly reshuffling the positions of the detected GW events, and compute the variance of the estimator on such a collection of sky distributions. As will be shown in the next section, the result of this estimate is in very good agreement with the simple poissonian computation outlined above.
\section{Forecasts with XG GW detectors}
\label{sec:3}
In this section, we forecast the detectability of the cosmic dipole with GW events. In Sec.~\ref{sec:sim} we start by discussing the detection prospects for compact binary mergers with XG detectors. In Sec.~\ref{sec:cdet} we study the significance of a possible cosmic dipole detection using the estimator in Eq.~\eqref{eq:estimator}. Finally, in Sec.~\ref{sec:cbayes} we show an example of how to estimate the cosmic dipole using Bayesian statistic.
\subsection{Simulating compact binaries mergers with XG detectors}
\label{sec:sim}
We consider a network of XG detectors composed by ET \citep{2010CQGra..27s4002P} and two CE \citep{PhysRevD.91.082001,Reitze:2019iox}. For ET we assume the same PSD used in \citet{2022arXiv220702771I}, while for CE we use the PSD provided by the CE consortium\footnote{\url{https://cosmicexplorer.org/sensitivity.html}}. In this exploratory approach, we detect a GW event using a SNR threshold calculated with Eq.~\eqref{Finn1}\footnote{This is a reasonable approximation for BNS events, but not so much for BBHs; however in the latter case threshold effects are less important (see end of this section) and this crude approximation should not affect our result.}. As we detect the time-evolution of the antenna patterns of the GW detectors, we set a lower frequency cut-off in Eq.~\eqref{f731} at $5$ Hz.
Given the sensitivity curves of the network and the two values of the source-frame masses, Eq.~\eqref{Finn1} can be used to calculate the maximum redshift at which we will be able to observe binaries with a given SNR as threshold.
\begin{figure}
\centering
\includegraphics{horizons.pdf}
\caption{Detection horizons for a ET+2CE detectors network. The redshift horizons (vertical axis) are calculated as a function of total source-frame mass (horizontal axis) for an equal mass binary ``edge-on'' with respect to the observer (worst-case scenario). The different colors mark the horizons for 3 SNR thresholds of 6, 9 and 12. Detection horizons are calculated using a flat $\Lambda$CDM cosmology with $H_0=67.7 \mathrm{km \, s^{-1} \, Mpc^{-1}}$ and $\Omega_m=0.3097$. Note that we only consider binaries with \textsc{ISCO} frequency above $5$~Hz for our simulation.}
\label{fig:horizons}
\end{figure}
Fig.~\ref{fig:horizons} reports the maximum redshift, as a function of total source frame mass, up to which we will be able to detect compact binary coalescences with a SNR $>6,9$ and $12$. We note that the horizons reported in Fig.~\ref{fig:horizons} are calculated for edge-on binaries ($\cos \iota=0$) and therefore they represent the worst-case scenario at which we will be able to detect compact binaries. Fig.~\ref{fig:horizons} indicates that if we consider a SNR for detection of $9$, we will be able to detect all the BNSs merging below redshift $\sim 0.8$ and all the BBHs merging below redshift $\sim 2$.
We want to determine if these detection horizons can introduce a significant selection bias when detecting populations of BNSs and BBHs. If a strong selection bias is present, then the theoretical prediction for the dipole depends on the parameter $s(\rho_*,r)$ in Eq.~\eqref{eq:defs}, that we need to model. If no strong selection bias is induced, i.e.\ if $s(\rho_*,r)\simeq 0$, then $\alpha\simeq 2$ and our estimator provides a model-independent measurement of the observer velocity $v_o$.
\begin{figure}
\centering
\includegraphics{rates.pdf}
\caption{\textit{Top panel}: Total number (vertical axis) of GWs from BNS and BBH mergers arriving on Earth within a given redshift shell (horizontal axis). The shaded areas mark the 90\% credible intervals associated to the population model, while the solid lines their median values. \textit{Bottom panel}: Total fraction of GWs that will be detectable (with various SNR thresholds) within a given redshift shell.
The figures are generated using the same population model for BNS and BBHs as in \citet{2022arXiv220702771I}. The solid lines indicate the fiducial model for the merger rates, while the shaded area the contours identified by the 90\% credible interval uncertainties on the rate models.}
\label{fig:rates}
\end{figure}
We simulate BNSs and BBHs following the same population models for BNSs and BBHs reported in \citet{2022arXiv220702771I} and supported by current observations \citep{2021ApJ...913L...7A}. More details about the BNSs and BBHs mass distributions are given in App.~\ref{app:pop}. We use the same merger rate model as a function of redshift for both BNSs and BBHs, given by, see~\citet{2020ApJ...896L..32C}
\begin{equation}
R(z)=R_0 [1+(1+z_p)^{-\gamma-k}] \frac{(1+z)^\gamma}{1+\left(\frac{1+z}{1+z_p}\right)^{\gamma+k}}\,,
\end{equation}
where $R_0$ is the merger rate today, and $\gamma$ and $k$ are two parameters encoding the redshift evolution of the merger rate. Since $k$ and $\gamma$ are not constrained by current observations, we assume fiducial values of $k=3$ and $z_p=2$, which are consistent with the Star Formation Rate \citep{2014ARA&A..52..415M}. For $\gamma$ and $R_0$ we take values consistent with the 90\% credible intervals values found by \citet{2021arXiv211103634T}. For BBHs we take $R_{0,\rm BBH}=17^{+10}_{-6.7} {\rm Gpc^{-3} yr^{-1} }$ and $\gamma= 2.7^{+1.7}_{-1.8}$, while for BNSs we take $R_{0,\rm BNS}=106^{+190}_{-93} {\rm Gpc^{-3} yr^{-1} }$ and the same $\gamma$ as for BBHs (since it was not possible to constrain this parameter from current observations). We take as fiducial rate models the ones corresponding to the median values of the parameters.
The number of GWs, emitted within a redshift shell with $z'<z$, crossing the Earth per year is then given by
\begin{equation}
\frac{dN}{dt_d}= \int_0^{z} R(z')\frac{1}{1+z'} \frac{dV_c}{dz'} dz'\,,
\end{equation}
where $dV_c/dz$ is the differential of the comoving volume and $t_d$ is the detector time in years.
In the top panel of Fig.~\ref{fig:rates} we report the total number of GWs from BNS and BBH mergers arriving in 1 year of observing time. With the assumed model for the merger rate of BNSs and BBHs we find that the total number of BNS and BBH mergers saturates at redshift $\sim 2$. This is expected from the assumed merger rate model that has a peak around redshift $2$. If we consider all the observable Universe, we find that we might expect to have between $(1-50) \cdot 10^4$ GWs from BBHs arriving on Earth per year and $10^4-10^7$ GWs from BNSs per year. However, given the detection horizons in Fig.~\ref{fig:horizons} not all the GWs will be detectable.
The bottom panel of Fig.~\ref{fig:rates} shows the ``\textit{detectable fraction}'' of BNS and BBH mergers within a certain redshift. On one hand, the plot shows that we will be basically able to detect all the BBH mergers in the Universe. This means that all events are above threshold, leading to $s(\rho_*,r)\simeq 0$ at all distances. For BNSs, on the other hand, we see that above $z\simeq 1$, a sizeable fraction of events will not be detectable. As a consequence $s(\rho_*,r)$ may be large, leading to a non-negligible contribution of threshold effects to the parameter $\alpha$. Therefore, even though we expect significantly less detections from BBHs than from BNSs, the former will offer a cleaner measure of the observer velocity, which is independent of the characteristics of the population.
\begin{figure}
\centering
\includegraphics[scale=0.35]{example_map_CMB.pdf}
\includegraphics[scale=0.35]{example_map_AGN.pdf}
\caption{Sky distribution of the estimator $\hat{v}_{\mathbf{n}}$ for the CMB fiducial value (top map) and the AGN fiducial value (bottom map). The maps are centered around the injected direction of the observer velocity. The figure is generated by dividing the sky in equal size pixels of 53 deg$^2$.}
\label{fig:maps}
\end{figure}
\subsection{Detection and Estimation of the cosmic dipole}
\label{sec:cdet}
In this section we study the detectability of the GW dipole, considering two fiducial values for its amplitude. The first one assumes that the GW dipole is of purely kinematic nature, with a velocity consistent with the one inferred from CMB observations \citep{2020A&A...641A...6P}, i.e.\ $v^{\rm CMB}_{\hat{{\bf{v}}}_o}=v_{o,{\rm CMB}}/c=1.2 \cdot 10^{-3}$. Since observations from radio sources and quasars find a dipole which is 2-5 times larger than expected~\citep{2017MNRAS.471.1045C,2018JCAP...04..031B,2021ApJ...908L..51S,2021A&A...653A...9S,Secrest:2022uvx}, we also consider a second fiducial value for the GW dipole, that would be consistent with these observations. Indeed, if the large dipole found in these studies is due to a breaking of the Copernician principle~\citep{Secrest:2022uvx}, i.e.\ to an instrinsic large anisotropy in the distribution of structures, then this anisotropy should be present also in the distribution of GW sources, that follow the large-scale structure of the Universe. It is thus interesting to assess the detectability of such a large dipole with GW sources. For this we take the extreme case of a dipole which would be 5 times larger than the one expected from CMB velocity, and with a direction aligned with it, i.e. $v^{\rm AGN}_{\hat{{\bf{v}}}_o}=6.0 \cdot 10^{-3}$.
For each fiducial value of the dipole, we simulate $10^4,10^5,10^6$ and $10^7$ BBHs detections $N_{\rm tot}$, using the ET+2CE network with a SNR thresold of $9$ for detection and the population of BBHs described in the previous section. The BBH detections are simulated as follow.
Each simulated BBH mass and redshift are drawn from the distributions described in the previous section and in App.~\ref{app:pop}. The original distribution of BBHs is isotropic in sky. Then, for each BBHs, we introduce the effect of the observer velocity by introducing an aberration following Eq.~\eqref{eq:domegabar}: $\theta'=\theta-(v_o/c)\sin \theta$, where $\theta$ is the angle between the source position and the observer velocity~\footnote{Note that in the AGN case, the dipole may not be due solely to the observer velocity, but rather to a large anisotropy in the distribution of sources, as discussed above. However, when assessing the detectability of such a dipole, it does not matter if it is of kinematic origin or not. Hence, we can simply simulate it as if it were due to a large velocity, 5 times larger than the one measured from the CMB.}. The detector-frame mass and luminosity distance change as described in Eq.~\eqref{eq:pert}. For each BBH, we then calculate the SNR using Eq.~\eqref{Finn1}. Then an ``observed'' SNR is drawn from a $\chi^2$ distribution with 2 times number of detectors d.o.f. If the ``observed'' SNR exceeds a SNR threshold of $9$, the binary is labelled as detected. Finally, to mimic the sky localization uncertainty given by the GW detection, we scatter the position of the GW sources using a gaussian distribution with $3$ deg.
Once the total list of BBHs detections is obtained with their sky position, we calculate the estimator of the GW dipole defined in Eq.~\eqref{eq:estimator}. Fig.~\ref{fig:maps} shows the skymap of $\hat{v}_{\mathbf{n}}$ calculated for the CMB and AGN fiducial values using $10^6$ BBHs detections. As we can see from the figure, in the AGN case, $\hat{v}_{\mathbf{n}}$ is maximized in the direction of the observer velocity and it displays an amplitude of $\hat{v}_{\mathbf{n}}=5.73 \cdot 10^{-3}$, similar to the one injected. For the CMB case on the other hand, the estimator is maximized in a direction different from the one of the observer velocity and it displays a maximum value of $\hat{v}_{\mathbf{n}}=1.3 \cdot 10^{-3}$, higher than the one injected. The reason for this is that, in the CMB case, the GW dipole will only be marginally detectable with $10^6$ BBHs.
\begin{figure}
\centering
\includegraphics{estimator_vs_angle.pdf}
\caption{Distribution of the estimator $\hat{v}_{\mathbf{n}}$ (red points) for the CMB fiducial value (top plot) and the AGN fiducial value (bottom plot), plotted as a function of the angle between ${\bf{n}}$ and the injected velocity direction $\hat{{\bf{v}}}_o$. The shaded areas mark the background $1,2,3 \sigma$ confidence intervals generated by randomly shuffling the detections over the sky. The horizontal dashed lines mark the standard deviations generated with Eq.~(\ref{eq:variance}).
The figure is generated by dividing the sky equal size pixels of 53 deg$^2$.}
\label{fig:estimator_vs_angle}
\end{figure}
Fig.~\ref{fig:estimator_vs_angle} shows another view of the values of the estimator reported in Fig.~\ref{fig:maps}. More precisely, we plot the value of the estimator as a function of the angle between the direction of the observer velocity, $\hat{{\bf{v}}}_o$, and the direction $\mathbf{n}$ at which the estimator is calculated. The figure also displays the $1, 2$ and $3$ $\sigma$ values of $\hat{v}_{\mathbf{n}}$ due to the stochastic fluctuations of events around the monopole given by Eq.~\eqref{eq:variance}. A $3\sigma$ detection of the GW dipole can be claimed if the $\hat{v}_{\mathbf{n}}$ estimator exceeds the $3 \sigma$ threshold from Poisson noise. From the figure we can see that, with $10^6$ detections, a GW dipole with amplitude compatible with that of the CMB dipole is marginally detectable with $1\sigma$ confidence (the probability that it is generated from a fluctuation of the monopole is not negligible), while a GW dipole with amplitude compatible with that of the AGN dipole is clearly detectable.
\begin{figure}
\centering
\includegraphics{detectability.pdf}
\caption{Distribution of the estimator $\hat{v}_{\mathbf{n}}$ evaluated in the injected velocity direction ${\bf{n}}=\hat{{\bf{v}}}_o$ with $10^4,10^5,10^6$ and $10^7$ BBH detections. The histograms are obtained by simulating 1000 populations of BBHs. The vertical dashed lines indicate the fiducial values of the GW dipole in the CMB case (blue) and the AGN case (orange). The dashed and dotted gray lines indicate the $1,2,3 \sigma$ contribution from Poisson noise, generated using Eq.~(\ref{eq:variance}).}
\label{fig:detectability}
\end{figure}
In order to better understand the detectability of the GW dipole, we repeat 1000 times the previous simulations for $10^4,10^5,10^6$ and $10^7$ detected BBHs and we calculate the fraction of cases for which the dipole will be detected. Fig.~\ref{fig:detectability} shows the distribution of the estimator $\hat{v}_{\mathbf{n}}$, evaluated in the velocity direction ${\bf{n}}=\hat{{\bf{v}}}_o$, that we obtain for the CMB and AGN cases. From the figure we can see that the distributions are centered around the injected value of the dipole. This is a confirmation that, for BBHs, threshold effects are negligible and $\alpha=2$ (such that $\alpha/2=1$ in Eq.~\eqref{eq:exp}). From the figure, we can see that when we only have $10^4$ detections, the distribution of the dipole estimator for the AGN and the CMB cases almost coincide. Moreover, the $1,2,3 \sigma$ intervals of these distributions almost coincide with the $1,2,3 \sigma$ variance from Poisson noise. In other words, with only $10^4$ detections the dipole is not detectable. As we collect more and more detections, the dipole estimator distributions tend to become separate from the Poisson noise.
\begin{figure}
\centering
\includegraphics{fappdet.pdf}
\caption{FAP versus detection probability plots for the injected CMB and AGN fiducial values (blue and orange lines). The vertical dashed lines indicate the FAP at the standard $1,2,3 \sigma$ credible intervals. The black dashed line is the reference to indicate when the detection probability coincides with the FAP (they are the same distribution).}
\label{fig:fappdet}
\end{figure}
In order to better quantify the detectability of the dipole, we plot False Alarm probability (FAP) vs detection probability in Fig.~\ref{fig:fappdet}. The FAP identifies a threshold for the dipole detection and it is defined as the probability that a random fluctuation of the number of GW detections in absence of dipole, would result in a false positive. The detection probability is defined as the probability that, in presence of a dipole, the estimator for the dipole detection would exceed the FAP threshold. Decreasing the FAP allows us to be more sure on the dipole nature of our detection but it decreases our sensitivity for the dipole detection. FAP versus detection probabilities plots can be used to: \textit{(i)} estimate what are the detection prospects for a given threshold (significance of the detection) and \textit{(ii)} clearly check in what regime the dipole is detectable. In fact, in the case where we are not able to detect the dipole, we expect the detection probability to be equal to the FAP, as the detection estimator would follow the same distribution.
As we can see from the top left panel of Fig.~\ref{fig:detectability}, with $10^4$ GWs detections, the distribution of the estimator for the CMB case agrees with the confidence intervals traced by the detection thresholds. This is reflected in the top left panel of Fig.~\ref{fig:fappdet}, where we see that the FAP and detection probability follow the same distribution, thus indicating that the only dipole that can be detected in this case, is the one from random fluctuations around the isotropic background (false positive). For the AGN case, the detection probability is slightly larger than the FAP, but still no robust detection could be claimed in this case.
With $10^5$ GWs detections ($\sim$ 2-3 year of observation), we can see from Fig.~~\ref{fig:fappdet} that there is a 75\% probability of detecting a dipole with AGN amplitude using a FAP of $2\sigma$. With $10^6$ detections (achievable in 10 years for an optimistic scenario), a dipole as high as the AGN one would be detectable at 100\%. On the other hand, if the GW dipole is compatible with the one observed in the CMB, then there is a 50\% probability that we would detect it with a FAP of $2 \sigma$.
\subsection{A Bayesian study}
\label{sec:cbayes}
Bayesian statistic can also be used to provide evidence for the GW dipole and estimate its parameters. As discussed in Section~\ref{sec:2.1}, the number of observed BBHs in a pixel situated in direction ${\bf{n}}_i$ is given by
\begin{align}
N_{\rm det}^{i}=\bar{N}_{\rm det}\left(1+\alpha\,{\bf{n}}_i\cdot \frac{{\bf{v}}_o}{c} \right) +\Delta N^i\, , \label{eq:Ni2}
\end{align}
where $\bar{N}_{\rm det}$ is the number of detections per pixel due to the monopole, $\Delta N^i$ a statistical fluctuation , and $\alpha \approx 2$ since for BBHs threshold effects are negligible.
The likelihood of obtaining $k_i$ detections in one pixel ${\bf{n}}_i$ is then given by
\begin{equation}
\mathcal{L}\left(k_i|\bar{N}_{\rm det},\frac{{\bf{v}}_o \alpha}{c}, {\bf{n}}_i \right) \propto e^{-N^i_{\rm det}}\cdot \left(N^i_{\rm det}\right)^{k_i}\, .
\end{equation}
The overall likelihood of obtaining $\{k\}=\{k_1,\ldots,k_{N_{\rm sky}}\}$ detections when dividing the sky in $N_{\rm sky}$ equal area pixels is
\begin{equation}
\mathcal{L}\left(\{k\}|\bar{N}_{\rm det},\frac{{\bf{v}}_o \alpha}{c}\right) = \prod_{i}^{N_{\rm sky}} \mathcal{L}\left(k_i|\bar{N}_{\rm det},\frac{{\bf{v}}_o \alpha}{c}, {\bf{n}}_i\right)\,.
\label{eq:likelihoodtotal}
\end{equation}
Finally, by applying the Bayes theorem, we can obtain posterior distributions on $\bar{N}_{\rm det},\alpha v_o/c$ and $\hat{{\bf{v}}}_o$ by calculating
\begin{align}
p\left(\bar{N}_{\rm det},\frac{{\bf{v}}_o \alpha}{c}| \{k\} \right) &\propto \mathcal{L}\left(\{k\}|\bar{N}_{\rm det},\frac{{\bf{v}}_o \alpha}{c}, \right)\times \pi \left(\bar{N}_{\rm det},\frac{{\bf{v}}_o \alpha}{c}, \right)\,,
\end{align}
where $\pi(\cdot)$ is a prior term.
We implement the likelihood in Eq.~\eqref{eq:likelihoodtotal} in a nested sampling code to obtain posterior distributions for $N_{\rm tot}=\bar{N}_{\rm det} N_{\rm sky}$ and $\alpha v_o/c$, in the case studies of the CMB and AGN dipole amplitudes, estimated with $10^6$ GWs. We use the python code \textsc{bilby} \citep{2019ApJS..241...27A,2020MNRAS.499.3295R} and its implementations of the nested sampling algorithm \textsc{dynesty} \citep{2019S&C....29..891H}. We use an isotropic prior for the dipole direction, a flat in log prior for $\alpha v_o/c$ between $10^{-6}$ and $10^{-1}$ and a flat in log prior for the total number of events $N_{\rm tot}$ between $10^5$ and $10^8$.
\begin{figure}
\centering
\includegraphics[scale=0.35]{AGN_dipoleloc.pdf}
\caption{Sky area at 90\% credible intervals for the dipole sky direction for the AGN test case with $10^6$ detections. The red dot marks the direction of the dipole. The sky localization has a radius of $\sim$ 13 deg.}
\label{fig:AGN_dipoleloc}
\end{figure}
For the case of a GW dipole with amplitude compatible with the AGN one, we find that with $10^6$ GWs detections, we are able to estimate the direction of the observer velocity with an uncertainty of $\sim 13$ deg ($\sim 2200$ deg$^2$ of area) at 90\% credible intervals. Fig.~\ref{fig:AGN_dipoleloc} shows the 90\% credible intervals in the sky identified by the posterior.
\begin{figure}
\centering
\includegraphics[scale=0.5]{amplitude_AGN.pdf}
\caption{Plots of the posterior distribution (and its marginals) for the total number of detections $N_{\rm tot}$ and the dipole amplitude $\alpha v_o/c$ for the AGN test case with $10^6$ detections. The solid black lines mark the injected values (assuming $\alpha=2$).}
\label{fig:amplitude_AGN}
\end{figure}
Fig.~\ref{fig:amplitude_AGN} shows instead the estimate of the total number of detections and the value of the GW dipole amplitude, $\alpha v_o/c$. The value of the recovered dipole amplitude is $\alpha v_o/c=1.15^{+0.18}_{-0.18} \cdot 10^{-2}$ at 68.3\% symmetric credible intervals. The amplitude of a dipole consistent with the AGN one can therefore be measured with a precision of 16\%.
We can also calculate the Bayes factor to asses the detection of the dipole, i.e.
\begin{equation}
\mathcal{B}^{\rm dip}_{\rm mono}= \frac{p(\rm dip|\{k\})}{p(\rm mono|\{k\})},
\end{equation}
where $p(\rm dip|\{k\})$ and $p(\rm mono|\{k\})$ are the evidences for the dipole model (with $\alpha v_o /c \neq 0$) and the monopole model (with $\alpha v_o/c =0$). For the AGN case, with $10^6$ GW detections we obtain a $\log_{10}(\mathcal{B}^{\rm dip}_{\rm mono})=6.0$, so strong preference for the presence of a dipole.
On the other hand, if the GW dipole has an amplitude consistent with that of the CMB dipole, the situation is different. First, we find that with $10^6$ detections it will not be possible to constrain the sky location. This is expected, since in Sec.~\ref{sec:cdet} we showed that the GW dipole would be hardly detectable in this case with $10^6$ events.
\begin{figure}
\centering
\includegraphics[scale=0.5]{amplitude_CMB.pdf}
\caption{Plots of the posterior distribution (and its marginals) for the total number of detections $N_{\rm tot}$ and the dipole amplitude $\alpha v_o/c$ for the CMB test case with $10^6$ detections. The solid black lines mark the injected values (assuming $\alpha=2$).}
\label{fig:amplitude_CMB}
\end{figure}
Moreover, if we look at the posteriors on the total number of events, $N_{\rm tot}$, and on the dipole amplitude, $\alpha v_o/c$, plotted in Fig.~\ref{fig:amplitude_CMB}, we can see that the total number of detections is clearly constrained, while the dipole amplitude cannot be constrained. However, it is possible to define an upper-limit in this case that results in $\alpha v_o/c <3.0 \cdot 10^{-3}$ at 95\% credible intervals (the injected value was $2.4 \cdot 10^{-3}$). We also compute the Bayes factor for this case finding that $\log_{10}(\mathcal{B}^{\rm dip}_{\rm mono})=-0.17$, indicating that there is no clear preference for the presence of a dipole or not.
\section{Discussion}
\label{sec:4}
In the previous section we have discussed several aspects related to the detection and estimation of the cosmic dipole from GW counting using both frequentists and bayesian techniques. We have shown that BBHs detected with the ET+2CE might be a ``clean'' probe for estimation of the dipole, since threshold effects are negligible in this case and the dipole is consequently only due to aberration ($\alpha \approx 2$).
The first crucial task to consider when searching for a dipole with GW number counts is the assessment of the detection significance.
For the frequentists techniques, we have shown that the significance can be evaluated using standard p-value techniques for Poissonian statistic and by reshuffling GW events in the sky to build a distribution of the noise due to the stochastic distribution of sources. For Bayesian statistic instead, we have shown that the detection can be evaluated with Bayes factors between the dipole and the monopole models. In both cases, we have obtained that in order to detect a dipole with amplitude $\alpha v_o/c$, one would need at least $\sim (c/\alpha v_o)^2$ detections, in order to be significantly confident that the detected dipole is not a fluctuation of the monopole distribution. We stress that, even with a few detections, it would not be surprising to find that the distribution of events is not perfectly isotropic in the sky. However, a low number of detections does not have the statistical significance to fit a dipole distribution. For instance, in Fig.~\ref{fig:maps}, we have shown that for a simulated GW dipole consistent with the one of the CMB, our estimator shows a peak over the sky which is not directly aligned with the simulated dipole. However in this case, its value is not statistically significant and a detection cannot be claimed.
This could explain the puzzling results of \citet{2022arXiv220407472K}, where using LIGO/Virgo data the authors reconstruct a dipole direction orthogonal to the CMB one. In that work the authors indeed do find an asymmetry in the distribution of GW events over the sky, however they do not assign any statistical significance to their results. If a significance was worked out, it would very probably indicate that the dipole reconstruction is statistically not significant due to the very low statistics of events used in the analysis.
\begin{figure}
\centering
\includegraphics{ULtrend.pdf}
\caption{95\% confidence level upper-limits (blue dots) on the dipole amplitude generated from populations of BBHs isotropically distributed over the sky (no cosmic dipole present). The black dashed line indicates the $1/\sqrt{N_{\rm tot}}$ scaling.}
\label{fig:ULtrend}
\end{figure}
To better demonstrate that the dipole detection capabilities scales as $1/\sqrt{N_{\rm tot}}$, in Fig.~\ref{fig:ULtrend}, we provide upper limits on the dipole amplitude obtained from BBHs distributions that are generated isotropically over the sky. From the figure, we can see that the 95\% upper limit on the dipole amplitude scales as expected with the number of detections.
In this study, we have focused on the dipole in GW number counting, showing in particular that BBHs provide a very clean way of measuring the observer velocity, since the theoretical prediction for the signal is very simple in this case. This is in contrast with the dipole from radio galaxies and quasars, for which the amplitude of the dipole depends directly on the properties of the sources, namely their spectral index and flux distribution. As shown in~\cite{Dalang:2021ruy}, if the populations evolve with redshift, then the theoretical prediction may change significantly and the tension with the CMB dipole may disappear. In this context, BBHs will provide a robust way to determine if the tension between the CMB dipole and the AGN dipole is due to our imperfect knowledge of quasars and radio galaxies properties, to systematic (that will necessarily be different with GW observations) or to a breaking of the Copernician principle.
This being said, it is still interesting to investigate if other GW estimators could help explaining the tension between the CMB and AGN dipole. In \citet{2022arXiv220407472K}, the authors propose to measure the dipole in the distribution over the sky of the detector-frame mass. Similarly to the case of number counting, detecting a cosmic dipole with this method, would require the detection of a relative discrepancy in detector-frame mass distribution of the order of $v_o/c$. To determine if this method can work, it is therefore crucial to assess the noise expected on such a measurement. Indeed, in addition to Poisson noise that affects number counting, the width of the mass distribution would induce additional contributions to the variance of the estimator. For BBHs, for which the width is expected to be large, the variance may be important~\footnote{Note that the mass distribution of BBHs is directly estimated from observed data, with complex hierarchical analysis tools such as the ones of \cite{2019ApJ...882L..24A,2021ApJ...913L...7A,2021arXiv211103634T}}. For BNSs, on the other hand, the variance may be significantly smaller since the mass distribution is much more peaked. It may also be interesting to combine mass and number counting, to determine if threshold effects can be mitigated with a specific estimator. We will explore all these aspects in a future work.
Finally, let us comment on some limitations of our prospect studies, that can be tackled in future works. In our study, we used the 0PN approximation to calculate the SNR of GW events. This is an approximation that underestimates the SNR of the detections as it neglects the merging part, which is important for BBHs. Therefore, we might expect threshold effects to be even less important than what we discussed. Another crucial assumption that we made was to neglect the sky direction dependence of the Antenna patterns of GW detectors. Due to their geometrical configuration, GW detectors are not equally sensitive to all the directions, and even for a detector network, the antenna patterns averaged during an year of observation would not be isotropic. For instance, for the simulated ET+2CE network, the antenna patterns averaged over one year of observation is
\begin{equation}
\langle F^2\rangle\simeq -0.097 \cos^4\delta-0.040 \cos^2\delta+0.878\,,
\end{equation}
where $\delta$ is the source declination. As ${\rm SNR}^2 \propto \langle F^2\rangle$, and $\langle F^2\rangle$ varies by about $10\%$ over the sky, we expect the SNR to vary by about $3\%$ over the sky. This means that there might be more or less sources detected over the sky due to this sensitivity variation, that may contaminate the measurement of a dipole of the order of $10^{-3}$. However, if threshold effects are negligible, as for the case of ET+2CE and BBHs, this sky-dependent selection bias will be negligible. In other cases, this effect can be calculated form first principles and used to adjust the estimator. Finally, one should consider also the sky-localization uncertainties associated to the GW detection, which could be higher than $3$ deg (although the results reported in \citet{2022arXiv220702771I} suggest this is a fair approximation for most BBH events); this would require an analysis done as in \citet{2022arXiv220705792E}.
\section{Conclusions}
\label{sec:5}
In this paper we have discussed aspects and prospects for detecting and estimating the cosmic dipole due to the observer velocity using GWs detected with XG detectors.
In Sec.~\ref{sec:2} we have introduced the theoretical framework to evaluate the effect of the observer velocity on GW detections. We have shown that this velocity introduces an aberration on the GW localization, and that it modifies the number of detections above a given SNR threshold through the redshifted chirp mass and the luminosity distance. For BBHs, we have demonstrated that these threshold effects are negligible and that the amplitude of the GW dipole is directly given by $2 v_o/c$.
In Sec.~\ref{sec:3} we have discussed several frequentists and bayesian techniques to detect and estimate the presence of the cosmic dipole from GW counting. We have shown that, with $10^6$ BBH detections, which would be observable in a few years of observations with a ET+2CE network, it will be possible to detect a cosmic dipole with amplitude similar to the one estimated from AGN, with a precision of $\sim 16\%$. On the other hand, with $10^6$ detections, a GW dipole with amplitude compatible with that of the CMB would only be marginally detectable. With $10^7$ detections however, we would be able to significantly detect a GW dipole with amplitude compatible with both the AGN and the CMB one. If we include BNSs and model the threshold contributions to the dipole, $10^7$ detections could be reachable in $\sim 10$ years of observations of ET+2CE.
Finally, in Sec.~\ref{sec:4} we have discussed critical aspects related to the detection of the cosmic dipole using number counting and the detector-frame mass distribution. Moreover, we have discussed the impact of some of the assumptions that we made in this exploratory study.
next generation GW detectors, that will give us access to at least hundreds of thousands of GW detections per year, have therefore the potential to solve a growing tension associated to the standard cosmological model.
\section*{Acknowledgements}
We thank Nicola Tamanini and Archisman Ghosh for discussions and exchanges.
S.M. is supported by the ANR COSMERGE project, grant ANR-20-CE31-001 of the French Agence Nationale de la Recherche.
S.F. is supported by the Fonds National Suisse, grant $200020\_191957$, and by the SwissMap National Center for Competence in Research. C.B. acknowledges funding from the Swiss National Science Foundation and from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation program (Grant agreement No.~863929; project title ``Testing the law of gravity with novel large-scale structure observables"). The work of G.C. is supported by the CNRS and by Swiss National Science Foundation (Ambizione grant, ``Gravitational wave propagation in the clustered universe").
\section*{Data Availability}
The simulations and numerical code underlying this paper are available on GitHub \href{https://github.com/simone-mastrogiovanni/cosmic_dipole_GW_3G}{\faGithub}.
\bibliographystyle{mnras}
|
\section{Introduction}
\begin{wrapfigure}{r}{0.47\textwidth}
\centering
\begingroup
\setlength{\tabcolsep}{1pt}
\begin{tabular}{cc}
\includegraphics[width=0.46\linewidth]{imgs/header/header_spima.png}%
& \includegraphics[width=0.46\linewidth]{imgs/header/header_spimb.png}\\
\includegraphics[width=0.46\linewidth]{imgs/header/header_recon.png}%
& \includegraphics[width=0.46\linewidth]{imgs/header/header_original.png}
\end{tabular}
\endgroup
\caption{Dual-Cycle reconstructs a 3D image with isotropic resolution given two views, A and B, of the same sample. \label{fig:header}}
\end{wrapfigure}
\label{sec:intro}
Three-dimensional fluorescence imaging, such as light-sheet fluorescence microscopy (LSFM)~\cite{stelzer2021light,liu2022recovery} is an essential tool for revealing important structural information in biological samples. However, it is common for 3D fluorescence microscopy to suffer from spatial-resolution anisotropy, where the axial direction is more blurry than the lateral imaging plane. Such anisotropy is due to several factors, including the diffraction of light and axial undersampling.
The spatial-resolution anisotropy is often addressed using image deconvolution methods, such as Richardson-Lucy algorithm~\cite{richardson1972bayesian,lucy1974iterative}. However, achieving isotropic resolution from a single 3D volume is an ill-posed inverse problem. The problem can be simplified by using multiview microscopy systems, such as dual-view inverted selective plane illumination microscope (diSPIM)~\cite{wu2013spatially,kumar2014dual}, equipped with classical joint multi-view deconvolution and fusion methods~\cite{wu2013spatially,preibisch2014efficient, temerinac2011multiview}.
Deep learning (DL) has emerged as an alternative to the classical deconvolution algorithms~\cite{guo2020rapid,park2022deep,wu2021multiview}. Neuroclear~\cite{park2022deep} is a recent self-supervised DL framework that uses cycle-consistent generative adversarial network (CycleGAN)~\cite{zhu2017unpaired} to improve the axial resolution from a single 3D input image without any knowledge of the point spread function (PSF). However, in many cases, the experimental PSF can be readily measured using either fluorescent beads~\cite{huygens, temerinac2011multiview} or small structures within samples~\cite{de2003image}, or derived theoretically~\cite{becker2019deconvolution}.
In this paper, we present Dual-Cycle as an improvement to Neuroclear that extends it into a dual-view self-supervised model-based framework. The inclusion of an additional view as input improves the reconstruction capability, while the additional prior on estimated PSFs allows our model to account for the expected degradation process. We experimentally validate Dual-Cycle on synthetic and real data showing that it can outperform Neuroclear as well as traditional dual view reconstruction algorithms.
\begin{figure*}[!t]
\includegraphics[width=\textwidth]{imgs/dualcycle.pdf}
\caption{Schematic illustration of the Dual-Cycle framework. a) Scheme of dual-view inverted selective plane illumination microscope (diSPIM). b)~CycleGAN approach: for two domains $Y$ and $X$, CycleGAN learns two mutually-inverse generator mappings $Gen$ and $Deg$ with the assistance of corresponding discriminators. c)~Dual-Cycle network architecture. d)~Schematic of the generator based on U-Net. e)~Degradation forms two paths each consisting of blurring with known PSF followed by the deep linear generator. f)~PatchGAN-based \cite{isola2017image} discriminators work on 2D slices of input 3D volumes.} \label{fig:framework}
\end{figure*}
\section{Forward Problem}
\label{sec:backg}
We focus on images recorded with single-plane illumination
microscopes (SPIMs) \cite{huisken2004optical} in a dual-view setup (diSPIM, Fig.~\ref{fig:framework}a). Data is acquired by two cameras, A and B, with an ideal relative rotation of 90~degrees. The image formation process (forward model) can be represented as the following linear observation model:
\begin{align}
\begin{split}
{\bm{g}}_A &= {\bm{\mathcal{A}}}_A {\bm{\mathcal{H}}}_A {\bm{u}} + {\bm{n}},\\
{\bm{g}}_B &= {\bm{\mathcal{R}}}_{\perp} {\bm{\mathcal{A}}}_B {\bm{\mathcal{H}}}_B {\bm{u}} + {\bm{n}}.
\end{split}
\label{eq:model2}
\end{align}
where ${\bm{g}}_A$, ${\bm{g}}_B$, and ${\bm{u}}$ correspond to the vectorized forms of deskewed 3D volumes, measured by camera A (View~A), camera B (View~B), and the original high-resolution 3D volume (Fig.~\ref{fig:header}). ${\bm{\mathcal{H}}}_A$ (resp.\ ${\bm{\mathcal{H}}}_B$) denote 3D convolution along the axial direction $z$ (resp.\ $x$) with some known PSF $h_A$ (resp.\ $h_B$). To model the mismatch from an ideal dual view setup, we include operators ${\bm{\mathcal{A}}}_{A/B}$, representing 3D affine transformation. We assume a coordinate system of unknown image ${\bm{u}}$ to be the same as ${\bm{g}}_A$ and that the ideal rotation of View B with respect to View A is 90 degrees around axes $y$, denoted as ${\bm{\mathcal{R}}}_{\perp}$. We omit subsampling in the axial directions by interpolating measurements to have voxels of equal size. In the general case, we consider additive noise ${\bm{n}}$.
Problem \ref{eq:model2} leads to an inherently ill-posed inverse problem. To solve it, we adopt and extend the approach proposed in \cite{park2022deep}.
\section{Inverse Problem}
\label{sec:propo}
Our proposed framework is illustrated in Fig.~\ref{fig:framework}c. In our setup, View A has a higher resolution in the $xy$ plane and is blurred in the axial direction $z$, while View B has a higher resolution in the $yz$ plane and is blurred in the axial direction $x$. Our goal is to reconstruct the original 3D volume with an isotropic resolution. We focus mainly on joint deconvolution and fusion with additional fine registration.
Our framework is based on a CycleGAN approach illustrated in Fig.~\ref{fig:framework}b and consists of two cycle-consistency paths, hence the name Dual-Cycle. It is worth mentioning that Dual-Cycle does not require any external training data beside the test object to be reconstructed.
The two views of the 3D volume are used as input for the 3D U-net-based generator (Fig.~\ref{fig:framework}d). The result of the generator is one 3D image representing the original 3D volume with isotropic resolution. To achieve this, we employ two sets of discriminators A1 and B1 (Fig.~\ref{fig:framework}f). Discriminators A1 distinguish between $xy$ planes of View A and $xy$ and $xz$ planes of the reconstructed volume.
Discriminators B1 distinguish between $yz$ planes of View B and $yz$ and $xz$ planes of the reconstructed volume.
To regularize and stabilize learning, the dual-cycle consistency is imposed. Therefore, the reconstructed image is degraded along two paths to imitate the forward problem \eqref{eq:model2}.
Consequently, \textit{Degradation} A and B, Fig.~\ref{fig:framework}e, consist of 3D convolution with given PSFs $h_A$ and $h_B$ followed by a deep linear generator (DLG) to address ideal model mismatch caused by affine operators ${\bm{\mathcal{A}}}$. For the blind case, when PSFs are unknown, degradation can be performed by DLGs only. Eventually, two other sets of discriminators A2 and B2 are added to map the distribution of corresponding planes of input View A/B onto generated View A/B. All discriminators are PatchGAN-based \cite{isola2017image} and work on 2D slices of analyzed 3D volumes (Fig.~\ref{fig:framework}f). Pixel-wise L1 loss between View A/B and generated View A/B is added to the GAN objective function to enforce cycle consistency.
\section{Experimental Validation}
\label{sec:exper}
We now present the numerical evaluation of Dual-Cycle on synthetic and real light-sheet data.
\begin{figure}[t!]
\centering
\begingroup
\setlength{\tabcolsep}{1pt}
\begin{tabular}{c}
\includegraphics[width=0.5\linewidth]{imgs/synSet.png
\end{tabular}
\endgroup
\caption{The set of six generated 3D volumes used in experiments. \label{fig:synSet}}
\centering
\begingroup
\setlength{\tabcolsep}{1pt}
\begin{tabular}{c}
\includegraphics[width=0.5\linewidth]{imgs/final.png
\end{tabular}
\endgroup
\caption{Comparison of MIPAV-generatefusion \cite{kumar2014dual}, diSPIMFusion \cite{guo2020rapid}, Neuroclear \cite{park2022deep}, and Dual-Cycle applied on the views A and B generated from the first 3D volume in the synthetic dataset in Fig.\ \ref{fig:synSet}. Visualized XY, XZ, and YZ images represent central cross-sections of the corresponding cubes in $xy$, $xz$, and $yz$ planes. Each reconstruction is labeled with its SSIM value with respect to the original volume. \label{fig:synresult}}
\end{figure}
\subsection{Synthetic data}
\label{ssec:syn}
We first illustrate possible improvements due to our dual-view framework over the single-view Neuroclear~\cite{park2022deep}. Additionally, we compare our network with other commonly used multi-view reconstruction techniques diSPIMFusion~\cite{guo2020rapid} and MIPAV-generatefusion~\cite{kumar2014dual}. The performance was measured using the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM).
We consider a dataset of six generated 3D volumes (120 × 120 × 120 voxels), shown in Fig.\ \ref{fig:synSet}. We drew 30-50 lines randomly in space and applied 3D elastic grid-based deformation. These volumes were treated as original ground truth volumes. All images were scaled to have values in the range 0-1. To obtain degraded volumes View A/B, we used the degradation process \eqref{eq:model2}, without noise and 90-degree rotation. The original volume was blurred in the $z$ direction for View A and in the $x$ direction for View B (blurring by Gaussian kernel with a standard deviation in range 2-4). Further, we applied random affine transformations to simulate the imperfection of the registration method. Relatively small mismatch (representing by ${\bm{\mathcal{A}}}$ in eq. \eqref{eq:model2}) is implemented as transformation of 3D points $\bm{p}$ as follows: $\bm{p}^{\prime} = (\mathbb{I} + \mathbb{N})\bm{p} + \bm{t}$, where $\mathbb{I}$ is identity matrix and $\mathbb{N}$ is random matrix with elements from a uniform distribution over $[-0.0025, 0.0025]$, and $\bm{t}$ is random translation vector sampled from a uniform distribution over $[-0.05, 0.05]^3$.
Except for Neuroclear, all methods use prior knowledge about the PSFs and both views as input. Visual comparison of reconstructed volumes corresponding to the first 3D volume of the synthetic dataset is in Fig.\ \ref{fig:synresult}. All methods can effectively perform the reconstruction, yet the improvement of Dual-Cycle compared to single view baseline is visually noticeable and corroborated by an increase in SSIM. Table \ref{table:results} summarizes the average PSNR/SSIM results of the tested methods. Overall, Dual-Cycle improves over the second best methods by 1.49 db (PSNR) and 0.017 (SSIM).
Implementation of Dual-Cycle was based on the Neuroclear and CycleGAN PyTorch framework; we used Adam optimizer and learning rate set to 0.0001. The network was initialized with weights pre-trained on the first volume. The training of the first (resp.\ following volumes) lasted approximately 12 hours (resp.\ 3-6 hours) using NVIDIA RTX A5000.
\subsection{Real data}
\label{ssec:subhead}
\begin{figure*}[!ht]
\includegraphics[width=\textwidth]{imgs/final_real.png}
\caption{Image reconstruction from real diSPIM data from \cite{guo2020rapid} using reconstruction methods MIPAV-generatefusion \cite{kumar2014dual}, diSPIMFusion \cite{guo2020rapid}, Neuroclear \cite{park2022deep}, and the proposed Dual-Cycle framework. \label{fig:realresult}}
\end{figure*}
We also tested reconstruction on diSPIM data from \cite{guo2020rapid}. Data was preprocessed using the Fiji software \cite{schindelin2012fiji}. The preprocessing involved denoising of both views and performing initial coarse registration of View B on View A. For both views: the minimum brightness value was truncated at value 78, volumes were normalized to 0-1 range, and were interpolated to have voxel sizes equal to $(0.1625~\mathrm{\upmu m})^3$. For the registration of view B on view A, we used Fiji plugin Fijiyama \cite{fernandez2021fijiyama}.
Images were cropped to 120 × 120 × 120 voxels and tested with the same methods as in Sec.\ \ref{ssec:syn}.
Visual comparison of reconstructed volumes is presented in Fig.\ \ref{fig:realresult}. The improvement of Dual-Cycle reconstruction over the Neuroclear is indicated via cross sections. Overall, Dual-Cycle achieves comparable or better performance relative to the state-of-the-art methods.
\begin{table}[t!]
\caption{The average PSNR/SSIM results of the blurred view~A/B, MIPAV-generatefusion, diSPIMFusion, Neuroclear and Dual-Cycle on the testing 3D volumes.\vspace{5px}}
\centering
\resizebox{0.5\columnwidth}{!}{\begin{tabular}{lrr}
\hline\hline
\multicolumn{1}{c}{Method} & \multicolumn{1}{c}{PSNR [dB]} & \multicolumn{1}{c}{SSIM}\\
\hline
View A & 29.32 & 0.929 \\
View B & 29.13 & 0.927 \\
MIPAV-generatefusion \cite{kumar2014dual} & 29.13 & 0.931 \\
diSPIMFusion \cite{guo2020rapid} & 28.55 & 0.943 \\
Neuroclear \cite{park2022deep} & 29.79 & 0.942 \\
Dual-Cycle (our) & \textbf{31.28} & \textbf{0.960}\\
[1ex] \hline
\end{tabular}}
\label{table:results}
\end{table}
\newpage
\section{Conclusion}
We presented Dual-Cycle, a self-supervised framework for dual-view fluorescence image reconstruction. The proposed method extends the recent Neuroclear method based on the CycleGAN framework. Compared to Neuroclear, Dual-Cycle includes two perpendicular views of the sample as input and uses prior knowledge on the estimated PSFs as a part of the degradation process within the framework. We have experimentally shown that Dual-Cycle achieves the state-of-the-art performance on synthetic and real data. While we only explored the dual-view setup in this work, our framework can be readily expanded into the multiple-view regime.
\section*{Acknowledgements}
This work was supported in part by the Czech Science Foundation grant GA21-03921S, the NSF
CAREER award CCF-2043134, the Fulbright commission under the Fulbright-Masaryk award, and by the Beckman Center for Advanced Light-Sheet Microscopy at Washington University in St.~Louis.
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:intro}
We consider $N$ bosons in the $d$-dimensional torus $\L_T = [-T/2, T/2]^d$, $d=2,3$, interacting via a two body non negative, radial and compactly supported potential $V$ with scattering length $\mathfrak{a}$ (see \cite[App.C]{LSSY} for a definition of the scattering length). Note that with a slight abuse of notation we denote by $\mathfrak{a}$ both the three and two dimensional scattering lengths. In the units where the particle mass is set to $m=1/2$ and $\hbar=1$, the Hamilton operator has the form
\begin{equation} \label{eq:HT}
H_T = -\sum_{i=1}^N \D_{x_i} + \sum_{1\leq i<j \leq N} V(x_i-x_j)
\end{equation}
and acts on the Hilbert space $L^2_s(\L_T^N)$, the subspace of $L^2(\L_T^N)$ consisting of functions which are symmetric with respect to permutations of the $N$ particles. Here the letter $T$ for the side length of $\L$ stands for `thermodynamic box', since we are going soon to consider the limit $N,T\to \io$ with $\r=N/|\L_T|$ fixed, known as {\it thermodynamic limit}.
In absence of interaction, and in the thermodynamic limit, the system described by \eqref{eq:HT} exhibit a striking feature: at sufficiently low temperature in three dimensions and at zero temperature in two dimensions a macroscopic fraction of the particles (depending on the temperature) occupies the ground state of the kinetic energy operator, namely the zero momentum mode (see \hbox{\it e.g.\ } \cite[Sec.1]{LSSY}).
This macroscopic behaviour, also known as {\it Bose-Einstein condensation} (BEC), is expected to survive for interacting systems (at least for $\r \mathfrak{a}^d$ small enough),
at low temperature in three dimensions and at zero temperature in two dimensions (condensation at positive temperature for interacting bosons in two dimensions is ruled out by the Hohenber-Mermin-Wagner theorem \cite{H,MW}).
Still a proof of the occurrence of BEC for generic bosonic systems is out of reach of rigorous analysis (see \cite{DLS, KLS} for the only available result to date), a fact which is not surprising, being the emergence of BEC an example of
occurrence of long-range order in systems with continuous symmetries, see e.g. \cite[Sect. 2]{SP-book}, \cite{LSY-BEC-SSB} and \cite[Ch. 5]{Tasaki}.
In these notes we are interested in a different - though related - aspect: providing asymptotic expressions for the thermodynamic functions of dilute Bose gases in the thermodynamic limit. As an instance of this goal, we will consider the ground state energy per unit volume in the thermodynamic limit.
Let us denote by $E_d(N,T)$ the ground state energy of the system described by \eqref{eq:HT} in $d=2,3$ dimensions. Then the specific ground state energy, defined by
\begin{equation}\label{eq:e_d}
e_d(\r) = \lim_{\substack{N, T \to \io\\ \r=N/T^d}} \frac{E_d(N,T)}{T^d}
\end{equation}
admits the following expansions in the dilute limit $\r \mathfrak{a}^d \ll 1$:
\begin{align}
e_{3}(\rho) &\;= 4\pi \rho^2 \mathfrak{a} \Big[1+\frac{128}{15\sqrt{\pi}}\sqrt{\rho \mathfrak{a}^3}+o\big(\sqrt{\rho \mathfrak{a}^3}\big)\Big]\,, \label{eq:LHY}
\\
e_2(\r) &\;= 4 \pi \r^2 \, \mathfrak{b} \Big[ 1 - \mathfrak{b} | \log \mathfrak{b} | + \Big( \frac 12 + 2\g + \log \pi \Big) \mathfrak{b} + o(\mathfrak{b})\Big] \label{eq:LHY2d}\,,
\end{align}
with $\mathfrak{b}= |\log (\r \mathfrak{a}^2)|^{-1}$.
The expansions \eqref{eq:LHY} and \eqref{eq:LHY2d}, first predicted in \cite{B,LHY} for the $3d$ case, and in \cite{Bog2D-5,A2d,MC03} for the $2d$ case, have been recently proven to be correct for non negative interactions with finite scattering length, see \cite{FS, FS2, YY, BCS} and \cite{FGJMO} respectively.
More precisely, lower and upper bounds compatible with \eqref{eq:LHY2d} have been proved in \cite{FGJMO} for generic non negative potentials, including the hard core case. Differently, the conditions under which the correct three dimensional upper bound can be proven are much more restrictive ($V \in L^3$) than those necessary for the lower bound, where the hard core case is included. We will come back to the difference among the $2d$ and $3d$ upper bounds in Section \ref{sec:UB}.
A striking feature of \eqref{eq:LHY} is its universal behaviour: in the dilute regime the ground state energy up to second order does not depend on the detail of the system in consideration, but only on the particle density and the scattering length of the interaction.
These physical quantities also define the relevant length scales for the problem, which are: the scattering length $\mathfrak{a}$ of the interaction; the mean inter-particle distance $\r^{-1/d}$; the inverse square root of the ground state energy per particle, given by $(\r \mathfrak{a})^{-1/2}$ in three dimensions and $(\r \mathfrak{b})^{-1/2}$ in two dimensions. The last length scale, usually referred to as {\it de Broglie wavelength}, {\it uncertainty principle length}, or {\it healing length},
defines the length below which a particle cannot be localized without seriously altering its energy (see the remark after Eq.\eqref{eq:localiz-UB} in Sec. \ref{sec:UB}).
In the dilute regime, we have the relations:
\[ \begin{split}
\mathfrak{a} & \ll \r^{-1/3} \ll (\r \mathfrak{a})^{-1/2} \qquad d=3 \\
\mathfrak{a} & \ll \r^{-1/2} \ll (\r \mathfrak{b})^{-1/2} \qquad d=2 \,.
\end{split}\]
Finally, let us mention a last important length scale, which is the one identified by the range of the interaction $R \geq \mathfrak{a}$. In particular it enters in the definitions of {\it high density} and {\it low density} regimes, corresponding to situations where $\r R^d \gg 1$ and $\r R^d \ll 1$ respectively.
Note that a system can be both dilute $\r \mathfrak{a}^d \ll 1$ and high density, provided that $\mathfrak{a}/R \ll 1$.
We will see in the next sections that the length scales introduced above play a key role in the methods allowing for a rigorous derivation of the expansions \eqref{eq:LHY} and \eqref{eq:LHY2d}. Note that in these notes we will not discuss the dilute limit for bosonic systems in one dimension, which is very different from the two and three dimensional cases considered here. We refer the reader to \cite{ARS}, where the latest result in this setting has been recently obtained.
\medskip
{\bf Bogoliubov's theory.} The first understanding of the expansions in \eqref{eq:LHY}
and \eqref{eq:LHY2d}
is based on an approximate exactly solvable model due to Bogoliubov \cite{B} (see also \cite[Appendix A]{LSSY} for a review). Bogoliubov started from the assumption (so far not yet justified) that low-energy states of \eqref{eq:HT} exhibit Bose-Einstein condensation for sufficiently weak interaction.
Guided by this idea, he rewrote the Hamilton operator \eqref{eq:HT} in momentum space, using the formalism of second quantization, and replaced all creation and annihilation operators associated with the zero-momentum mode by factors of $N^{1/2}$. The resulting Hamiltonian contains a constant term (describing the interaction among particles in the condensate), plus terms containing creation and annihilation operators associated with modes with momentum $p \not = 0$, usually referred to as {\it excitations}. In particular we have:
i) terms that are quadratic in creation and annihilation operators associated with $p \not = 0$ modes, describing the kinetic energy of the excitations as well as the interaction among pairs of excitations and the condensate; ii) terms that are cubic and quartic, describing interactions involving more than two excitations. Neglecting all cubic and quartic contributions, Bogoliubov obtained a quadratic Hamiltonian that he could diagonalize explicitly, obtaining the following expression for the ground state energy of the system of \eqref{eq:HT}:
\begin{equation} \label{eq:BogEnergy-anyd}
\begin{split}
E_d(N,T) =\; & \frac{N }{2} \rho \widehat V(0) - \frac 1 2 \sum_{\substack{p \in \frac{2\pi}{T} \mathbb{Z}^d,\\ p \neq 0}} \Big[\; p^2 + \rho \widehat V(p) - \sqrt{p^4 \hskip -0.05cm+ \hskip -0.05cm2 \rho \widehat V(p) p^2 }\; \Big] \,,
\end{split}\end{equation}
where $\widehat V(p)$ denotes the Fourier transform of $V$, defined as
\begin{equation} \label{eq:Fourier}
\widehat V(p)= \int_{\L_T} e^{ip\cdot x} V(x) d^d x\,.
\end{equation}
Let us now focus on the three dimensional case, which is the one directly analysed in \cite{B}. We rewrite \eqref{eq:BogEnergy-anyd} as follows
\begin{equation} \label{eq:BogEnergy} \begin{split}
E_3(N,T) =\; & \frac{N }{2} \rho \widehat V(0) - \frac 1 4 \hskip -0.1cm\sum_{\substack{p \in \frac{2\pi}{T} \mathbb{Z}^3 \\ p \neq 0}} \frac{ (\rho \widehat V(p) )^2 }{p^2} \\
&- \frac 1 2 \sum_{\substack{p \in \frac{2\pi}{T} \mathbb{Z}^3 \\ p \neq 0}} \Big[ p^2 + \rho \widehat V(p) - \sqrt{p^4 \hskip -0.05cm+ \hskip -0.05cm2 \rho \widehat V(p) p^2 } - \frac 1 2 \frac{ (\rho \widehat V(p) )^2 }{p^2} \Big] \,.
\end{split}\end{equation}
and then we take the thermodynamic limit $N,T \to \io$, $\r=N/T^3$.
One easily recognizes, in the expressions appearing on the r.h.s. of \eqref{eq:BogEnergy}, the first and second term of the Born approximations of the scattering length $\mathfrak{a}$ of $V$, which reads:
\[
\mathfrak{a} = a_0 + a_1 + \mathcal{O}\big((a_0/R)^2\big) \,,
\]
with
\begin{equation} \label{eq:Born}
8 \pi a_0= \widehat V(0)\,, \qquad 8 \pi a_1= - \int \frac{d^3 p}{(2\pi)^3} \frac{\widehat V(p)^2}{2p^2} \,. \end{equation}
Let us recall that $\mathfrak{a}$ is defined through the solution $f$ of the zero energy scattering equation, which in three dimension solves
\begin{equation}\label{eq:scatt_3d}
\Big(-\D + \frac12 V\Big)f =0\,, \qquad f(x) \xlongrightarrow[x \to \io]{} 1\,.
\end{equation}
With this definition we have $\int V(x)f(x)\,d^3x=8\pi\mathfrak{a}$.
Eq. \eqref{eq:LHY} was then obtained by Bogoliubov by replacing the sum $a_0 + a_1$ by $\mathfrak{a} $ in first line of Eq. \eqref{eq:BogEnergy}, and $\widehat V(p)$ by $\widehat V(0) \simeq 8 \pi \mathfrak{a} $ in the integral obtained from the sum on the second line of the same equation. Note that the replacement of $\widehat V(p)$ with $\widehat V(0)$ in the integral is possible thanks to the fact that we have added to the sum on \eqref{eq:BogEnergy-anyd} the term $ \sum_{ p \neq 0} (\r\widehat V(p))^2/ (4 p^{2})$, which in the thermodynamic limits reproduces $- 4 \pi a_1 N \r $. Let us now denote by $\r_+$ the density of particles outside the condensate in the ground state (the so called {\it condensate depletion}). Then Bogoliubov's theory predicts (see \hbox{\it e.g.\ } \cite[Eq.(4.48)]{SP-book}):
\begin{equation} \label{eq:depletion}
\frac{\rho_+}{\r}= \frac{8}{3 \sqrt \pi} \sqrt{\r a_0^3} \simeq \frac{8}{3 \sqrt \pi} \sqrt{\r \mathfrak{a}^3}\,,
\end{equation}
with the usual replacement $a_0 \to \mathfrak{a}$.
\medskip
Let us now analyse Eq.\eqref{eq:BogEnergy-anyd} in the two dimensional case. We recall that for a fixed $\ell>R$, the zero energy scattering function in two dimensions solves
\begin{equation} \label{eq:scatt_2d}
\Big( -\D + \frac 12 V \Big) f_\ell =0\,,\qquad f(x)\big|_{|x|=\ell} =1 \,.
\end{equation}
In particular
\[
f_\ell (x) = \frac{\log(|x|/\mathfrak{a})}{\log(\ell/\mathfrak{a})}\,, \qquad \forall x : R <|x| \leq \ell
\]
with $\mathfrak{a}$ denoting here the two dimensional scattering length.
For small interactions $V(x)=\l \widetilde V(x)$, with $0<\l \ll 1$, the Born series for the two dimensional scattering length reads:
\[ \label{eq:bell}
|\log (\ell/\mathfrak{a})|^{-1} = (4 \pi)^{-1} \widehat V(0) + \sum_{k=1}^{\io} (4\pi)^{-(k+1)} \int_{\mathbb{R}^2} \big(\mathcal{L}_V\big)^{k}(V)(x) dx = \sum_{k=0}^\io 2 b_k^{(\ell)}\,,
\]
where $\mathcal{L}_V$ is the operator given by $\mathcal{L}_V(g)(x)= V(x)\int_{\mathbb{R}^2} \log(|x-y|/\ell) g(y)dy$. Clearly $b_k^{(\ell)}= \mathcal{O}(\l^{k+1})$. Hence, setting $\ell= \r^{-1/2}$ we obtain the following expansion for the effective parameter $\mathfrak{b}=|\log(\r \mathfrak{a}^2)|^{-1}$ appearing in \eqref{eq:LHY2d}:
\begin{equation} \label{eq:bb-Born}
\mathfrak{b} = b_0 + b_1 + b_2 + \mathcal{O}(\l^4)\,, \qquad \text{where } b_k := b_k^{(\r^{-1/2})}\,.
\end{equation}
In particular $ 8 \pi b_0 = \widehat V(0)$ and
\[
8 \pi b_1 = \frac{1}{4\pi} \int V(x) V(y) \log(|x-y|/\ell) \, dx dy \,.
\]
Note that with the choice of $\ell$ above the following is satisfied: $R \ll \ell \ll (\r \mathfrak{b})^{-1/2}= \mathcal{O}\big((\l \r )^{-1/2}\big)$, \hbox{\it i.e.\ } $\ell$ is smaller than the $2d$ healing length.
Similarly to what is done in the three dimensional case we can add and subtract on the r.h.s. of \eqref{eq:BogEnergy-anyd} a term of the form $ \frac {\,\r^2} 2 \sum_{p \neq 0} g_p$, with $g_p$ chosen so that $\frac {\,\r^2} 2 \sum_{p \neq 0} g_p$ converges to $8 \pi b_1 N \rho$ in the limit $T\to \io$ with $\r=N/T^2$ fixed\footnote{Note that due to the logarithmic divergence of the two dimensional Green function, the expression for $g_p$ is slightly more complicated than in the three dimensional case (see for example \cite[App. A]{FNRS-2d} or \cite[Eq. after (4.17)]{CCS2}), and in particular $g_p \to 0$ in the limit $p \to 0$.}.
Eq. \eqref{eq:LHY2d} is then obtained by taking the thermodynamic limit, and by substituting in the final result: $ 4 \pi N \r (b_0+b_1)$ with $ 4 \pi N \r \mathfrak{b}$; $b_0$ with $\mathfrak{b}$ in the terms of the order $b_0^2 \log b_0$ and $b_0^2$ coming from the limit of the sum of the second term on the r.h.s. of \eqref{eq:BogEnergy-anyd} and $- \frac {\, \r^2} 2 \sum_{p \neq 0} g_p$. One can also evaluate the condensate depletion, which Bogoliubov approach predicts to be proportional to $\r \mathfrak{b}$.
We refer to \cite{MC09} and references therein for the use of Bogoliubov method in two dimensions in the physics literature. In the next paragraphs we will rather discuss the validity of Bogoliubov approach, both in three and two dimensions, from a mathematical point of view.
\bigskip
{\bf Validity of Bogoliubov's approach in $3d$.} The discussion around \eqref{eq:Born} makes apparent that there exists a regime of parameters $\{R,\mathfrak{a},\r \}$ where the approximations made by Bogoliubov can be proven to be correct.
This is in fact the case when $a_0 = (8\pi)^{-1}\widehat V(0)\ll R$ and moreover
\begin{equation} \label{eq:Bog-exact}
\frac{a_0}{R} \gg \sqrt{\rho a_0^3} \gg \frac{a_0^2}{R^2}\,.
\end{equation}
The above condition guarantees that the final replacements made by Bogoliubov to obtain \eqref{eq:LHY} only produce subleading contributions (see also the remarks at the end of \cite[Section 1]{BriS}). In other words, whenever \eqref{eq:Bog-exact} holds, the contribution to the ground state energy coming from the cubic and quartic terms neglected in Bogoliubov theory is negligible at the order of precision we are looking at.
To make the conditions \eqref{eq:Bog-exact} more explicit,
let us introduce the parameter $\g>0$ defined by $a_0/R = (\r \mathfrak{a}^3)^\g$ (recall that $\r \mathfrak{a}^3 \ll 1$). Then the conditions in \eqref{eq:Bog-exact} are satisfied for $\g \in (1/4, 1/2)$, see the shadowed area in Fig.\ref{Fig:Bog}. On the other hand for $\g \in (1/3,1/2]$ we have
\[ \label{eq:highdensity}
\r R^3 \sim (\r \mathfrak{a}^3)^{1-3\g} \gg 1 \,,
\]
namely the system is in a high density regime.
Indeed, mathematically, the validity of Bogoliubov's approach for three-dimensional Bose gases has been at first established in weak coupling and high density regimes: in \cite{LS-jellium} where the ground state energy of bosonic jellium was obtained, in \cite{LS-charged,S-charged} where a similar result was achieved in the context of the two-component charged Bose gas, and later by Giuliani and Seiringer for the computation of \eqref{eq:LHY}, as discussed in the Sect. \ref{sec:LB} below.
\smallskip
\begin{figure}[t]
\centering
\begin{tikzpicture}[scale =0.7,
every path/.style = {},
]
\begin{scope}
\draw[|-, thick] (0,0)--(3,0);
\draw[|-, thick] (3,0)--(4,0);
\draw[|-|, color=black, thick] (4,0)--(6,0);
\shade[top color=lightgray, bottom color=white, opacity=0.8] (3.01,-0.25) rectangle (5.99,0.25);
%
\draw[<->, thick, black] (4,0.7)--(6,0.7);
\node[below, text width=8.5em, align=center] at (5,1.5) {\textcolor{black}{\footnotesize high density regime}};
%
\node[below] at (0, -0.2) {$0$};
\node[below] at (3, -0.2) {$\frac 1 4$};
\node[below] at (4, -0.2) {\textcolor{black}{$\frac 1 3$}};
\node[below] at (6, -0.2) {$\frac 1 2$};
\node[above] at (0.5, 0.2) {$\g$};
\draw[->] (18/5,-1) --(18/5,-0.4);
\node[below] at (18/5, -1){\cite{BriS}};
\draw[->] (134/23,-1) --(134/23,-0.4);
\node[below] at (134/23, -1){\cite{GiuS}};
\end{scope}
\end{tikzpicture}
\caption{For $\r \mathfrak{a}^3 \ll 1$ (dilute regime), the figure above depicts the range of validity of Bogoliubov approximation in $3d$ (shadowed area) with respect to the value of the parameter $\g>0$ defining the ratio $a_0/R=(\r \mathfrak{a}^3)^\g$. Note that for $\g>1/3$ the systems is in a high density regime, namely $\r R^3 \gg 1$. The case $\g=0$ corresponds to the case where no Born series for the scattering length is available. The threshold reached by \cite{GiuS} ($\g>1/2 - 1/69$) and \cite{BriS} ($\g>3/10$) are also indicated.}
\label{Fig:Bog}
\end{figure}
\medskip
The first paper where the validity of Bogoliubov preditions were proven in a region of parameters where \eqref{eq:Bog-exact} fails is due to Erd\H os, Schlein and Yau in \cite{ESY}.
There, an upper bound for the specific ground state energy was proven considering an interaction of the form $\lambda V$ for $\lambda>0$ small. The result coincides with the asymptotic in Eq.\eqref{eq:LHY}, except for the fact that the second order correction is multiplied by a constant $S_\lambda$ satisfying $1\leq S_\lambda\leq 1+C\lambda$. Note that, under the assumption $\mathfrak{a}/R=(\r \mathfrak{a}^3)^\d$ for any $\d>0,$ \eqref{eq:LHY} is recovered from \cite{ESY}.
The key idea in \cite{ESY} is that a good trial state capturing \eqref{eq:LHY} for smooth potentials is given by a quasi-free state on the top of the condensate, with the quasi-free state describing the creation of couples of excited particles with momenta $(p,-p)$. In \cite{ESY} the creation of couples of excitations is mediated by a kernel $c_p$ which has quite a different behaviour for momenta smaller or higher than the inverse healing length $(\r \mathfrak{a})^{1/2}$. Let $f$ be the solution of the zero energy scattering equation defined in \eqref{eq:scatt_3d}. Then for high momenta $c_p$ is the Fourier transform of $\r (1-f)$ and for low momenta is a quantity related to the diagonalization of a quadratic Hamiltonian, which can be thought as obtained by conjugating the Bogoliubov pair Hamiltonian (as in \cite[A.17]{LSSY}) through a Bogoliubov transformation with kernel $\widehat{\rho(1-f)}_p$.
An important remark is that the result from \cite{ESY} do not improve with different choices of the coefficients $c_p$. Indeed, as showed in \cite{NRS1,NRS2} quasi-free states can only approximate the specific ground
state energy of a $3d$ dilute Bose gas up to errors which are of order $\r^2\mathfrak{a} \sqrt{\r \mathfrak{a}^3}$. To show \eqref{eq:LHY} for general dilute systems, it is needed to go beyond quasi-free states, in order to be able to extract some missing contribution hidden in the cubic and quartic terms \cite{YY, BCS, FS,FS2}.
Finally, let us mention that a systematic study of the perturbation theory around Bogoliubov model in three dimension, and the proof of its order by order convergence after proper resummations, was obtained by Benfatto back in 1997 \cite{Benfatto} (see \cite[Ch. 1]{PhD-Cenatiempo} for an outline of this approach). However the resulting bounds are not enough for constructing the theory in a mathematically complete
form: they allow to derive finite bounds at all orders in renormalized perturbation
theory, growing like $n!$ at the $n$-th order, but the possible Borel summability of the series remains an outstanding open problem. A program addressing this issue has been
started by Balaban, Feldman,
Kn\"orrer and Trubowitz, see \cite{Balaban} for an overview.
\bigskip
{\bf Validity of Bogoliubov approach in $2d$.} As commented around \eqref{eq:bb-Born}, in two dimensions Bogoliubov's approach is expected to give the correct asymptotics for sufficiently weak interactions, so that
\begin{equation}\label{eq:Bog-exact2d}
b_2 \ll \max\{ b_0, b_1, b_0^2, b_0^2 \log b_0\}\,.
\end{equation}
Note that, differently from the three dimensional case, the range of the interaction does not enter in the identification of the regimes in \eqref{eq:Bog-exact2d}. Still we can expect Bogoliubov theory to hold both for low density and high density regimes since for interactions of intensity $\l \ll 1$ we have $R \ll ( \l \r)^{-1/2}$, that is $\rho R^2 \ll 1/\l$.
The validity of Bogoliubov two term asymptotics for the ground state energy for sufficiently weak interactions has been indeed shown in \cite{FNRS-2d}, by considering quasi-free states.
Very recently, a proof establishing \eqref{eq:LHY2d} for generic non negative potentials and dilute $2d$ systems has been obtained in \cite{FGJMO} (see Sec. \ref{sec:LB} and Sec. \ref{sec:UB} below for more details).
Finally, in the case of weak interactions, a systematic study of the corrections to Bogoliubov theory
via renormalization group methods (allowing in principle also to obtain information on correlation functions) can be obtained up to length scales of the order $\l^{-1} (\l \r)^{-1/2} $, which is the largest scale at which the problem has a perturbative nature \cite{CG}.
\medskip
\medskip
{\bf Summarizing.} The fact that the expansion \eqref{eq:LHY} (resp.\eqref{eq:LHY2d}) has been shown to be correct in more general regimes than those identified by \eqref{eq:Bog-exact} (resp. \eqref{eq:Bog-exact2d}) shows that Bogoliubov's predictions remain corrects beyond the regimes of validity of Bogoliubov's approximations. We will see in the next sections, that this is linked to two main facts:
\begin{enumerate}
\item[i)] to make Bogoliubov theory rigorous it is sufficient to show the occurrence of condensation on a sufficiently large, but finite box; \smallskip
\item[ii)] in situations where \eqref{eq:Bog-exact} (resp. \eqref{eq:Bog-exact2d}) are not satisfied, it is possible to extract from the cubic and quartic terms neglected in Bogoliubov theory, some contributions to the ground state energy which renormalize the interaction and make the scattering length to appear.
\end{enumerate}
Finally, the problem shows a clear separation of energies: momenta of the order $\mathfrak{a}^{-1}$ are responsible for the appearance of the full scattering length, while momenta of the order of $(\rho\mathfrak{a})^{1/2}$ in $3d$ and $(\rho\mathfrak{b})^{1/2}$ in $2d$ are responsible for the second term in \eqref{eq:LHY} and \eqref{eq:LHY2d}.
\section{Occurrence of local condensation: scaling regimes } \label{sec:scalings}
A natural question that arises from the overview given in the previous section is what is the largest length scale at which Bose-Einstein condensation can be shown to occur, and whether the results predicted by Bogoliubov, which are in fact based on the occurrence of BEC, can be validated up to the same scale.
The first result in this direction is due to Lieb and Seiringer who, in the celebrated paper \cite{LS}, exhibited a proof of Bose-Einstein condensation for $3d$ interacting bosons in the so-called Gross-Pitaevskii regime, famously introduced in \cite{LS1,LSY}.
Let us consider $N$ bosons trapped in a region of order $L$, and interacting through a potential whose scattering length $\mathfrak{a}$ is of the same order of the range as the interaction $R$. Then the $3d$ Gross-Pitaevskii regime is defined as the limit where we send $L, N$ to infinity by keeping the ratio $\mathfrak{a} N /L$ fixed.
Note that, if we consider $\mathfrak{a}$ fixed as in the previous section, the Gross-Pitaevskii scaling corresponds to an ultradilute regime where the particle density goes to zero as $N^{-2}$. On the other hand, if we rescale lengths and rather consider $L$ of order one, we have that the Gross-Pitaevskii regime corresponds to consider potentials with a scattering length of order $N^{-1}$ (which are nicely describing the intense and short range interactions typical of BEC in cold atomic gases). In the simplest case where the bosons are trapped on a three dimensional torus the Gross-Pitaevskii Hamiltonian is an operator on $L^2_s([0,1]^3)$ of the form:
\begin{equation} \label{eq:HN-GP}
H_N^{(GP)}= \sum_{i=1}^N -\D_{x_i} + \sum_{1\leq 1 < j \leq N} N^{2} V(N (x_i- x_j))\,.
\end{equation}
In \cite{LSY} Lieb, Yngvason and Seiringer were able to provide an expression for the ground state energy of $H_N^{(GP)}$ compatible with \eqref{eq:LHY} at leading order, but also valid in the more general case where the bosons are confined in a region of order $L$ by a generic trapping potential (in the latter case the ground state energy is obtained through the minimization of a non linear energy functional on $L^2(\mathbb{R}^3)$, see \cite[Theorem 6.1]{LSSY}).
In the same years the same results were extended to two dimensions \cite{LS1,LSY2}. There the Gross-Pitaevskii Hamiltonian on $L^2_s([0,1]^2)$ is defined to be:
\begin{equation} \label{eq:HN-GP2d}
H_{N,2d}^{(GP)}= \sum_{i=1}^N -\D_{x_i} + \sum_{1\leq 1 < j \leq N} e^{2N} V( e^N (x_i- x_j))\,.
\end{equation}
The exponential scaling of the interaction potential in \eqref{eq:HN-GP2d} is due to the fact that in $2d$ the Gross-Pitaevskii regime is defined as the limit where $N \to \io$ with $\mathfrak{b} N$ fixed, and $\mathfrak{b}$ depends logarithmically on the scattering length.
Note that, rescaling lengths, the two dimensional Gross-Pitaevskii regime can be interpreted as describing an extended Bose gas (of particles interacting through the unscaled
potential $V$) at a density that is exponentially small in $N$. While the exponential smallness of the density (or, equivalently, of the scattering length) makes it difficult to
directly apply the results obtained in this regime to physically relevant situations, the $2d$ Gross-Pitaevskii provides the simplest scaling limit where the strong correlations typical of two dimensional systems drastically reduce the effective coupling $\mathfrak{b}$.
\medskip
The key fact making possible to show the occurrence of condensation in the Gross-Pitaevskii regime is that in this regime the energy gap between the ground state of the kinetic energy and the first excited state is of the same order of magnitude of the energy per particle (hence one can for example apply the Temple inequality, as discussed in Sec. \ref{sec:LB}). Note also that proving BEC in the Gross-Pitaevskii regimes corresponds to showing condensation on a length scale of order $L \sim (\r \mathfrak{a})^{-1/2}$ in three dimensions and $L\sim (\r \mathfrak{b})^{-1/2}$ in two dimensions.
On the other hand, the Born series expansion for the scattering length fails to be true in the Gross-Pitaevskii regime, both in three dimensions (since the ratio among the scattering length of the interaction and its radious is of order one) and in two dimensions (since the integral of the Gross-Pitaevskii potential is independent of $N$, while $\mathfrak{b} \sim N^{-1}$). This is the reason why showing the validity of Bogoliubov theory in the Gross-Pitaevskii regimes has required twenty years more \cite{BBCS1,BBCS3,BBCS4, CCS1, CCS2}, with respect to the pioneering works by Lieb, Seiringer and Yngvason (later revised in \cite{NRS}).
Note that thanks to the presence of a spectral gap, in the Gross-Pitaevskii regime it also possible to analyse the low energy excitation spectrum \cite{BBCS4, CCS2}, which is particularly important for understanding superfluidity.
We refer to \cite{Schl22} and \cite{Caraci} for a discussion of the methods allowing for a rigorous implementation of Bogoliubov theory in the $3d$ and $2d$ Gross-Pitaevskii regimes respectively. We remark that the validity of Bogoliubov theory can also be established in presence of a generic trapping potentials \cite{BSS,BSS1,NNRT,NT}, and that more recently a simplified approach to verify Bogoliubov theory in the Gross-Pitaevskii regime has been developed in \cite{Hainzl,HST}. Note that the current method used to recover Bogoliubov theory in the Gross-Pitaevskii regime requires the potential to be at least $L^1$; a second order upper bound valid for hard core potentials in the Gross-Pitaevskii regime has been recently showed in \cite{BCOPS}.
\medskip
In the next two paragraphs we see how one can consider less (or more) singular regimes with respect to the Gross-Pitaevskii one. We will analyse the $3d$ and $2d$ cases separately.
\bigskip \medskip
{\bf Three dimensions.} As before let us consider $N$ bosons on the unitary torus. Let $R$ be the range of the interaction. Scaling regimes less singular than the Gross-Pitaevskii regime can be obtained by choosing $\mathfrak{a} \sim N^{-1}$ and $R \sim N^{-\b}$ with $\b \in [0,1)$ so that $\mathfrak{a} \ll R$, namely we consider
\begin{equation} \label{eq:HN-beta}
H_{N,3d}^{(\b)}= \sum_{i=1}^N -\D_{x_i} + \sum_{1\leq 1 < j \leq N} N^{3\b-1} V(N^\b (x_i- x_j)) \qquad \text{on} \qquad L^2_s([0,1]^3)\,.
\end{equation}
Note that the regime corresponding to $\b=0$ is also known as {\it mean-field} scaling, and for $\b=1$ we recover the Gross-Pitaevskii regime. We remark that: i) for $\b\in[0,1/3)$ we have $\r R^3 \gg 1$ hence the system is in a high density regime; ii) for $\b\in [0,1/2)$ the conditions \eqref{eq:Bog-exact} holds, and one can show that Bogoliubov approximation is correct.
For $\b \in [1/2,1)$, despite the fact that the approximations made by Bogoliubov fail, it is possible to recover Bogoliubov's predictions \cite{BBCS2} considering a quasi-free state which has a different behaviour on low and high energy scales \cite[Sec. 2]{Olgiati}, in the same spirit of what is done in \cite{ESY} for weakly interacting bosons in the thermodynamic limit. For a comparison among the regimes described by \eqref{eq:HN-beta} and those discussed around \eqref{eq:Bog-exact}, we refer the reader to \cite[Eqs. (6)-(7)]{BriS}.
\medskip
{\it Remark.} The first rigorous proof of the validity of Bogoliubov predictions for the ground state energy and low energy spectrum of Hamiltonians as in \eqref{eq:HN-beta}, was given by Seiringer for $\b=0$ \cite{Sei}. This result was later extended in various directions in \cite{GS,LNSS,DN,PizzoI, PizzoII, PizzoIII,BPS,Mitrouskas,LNR2, BroS}. Note that the mean field regime can also be regarded as a semiclassical limit, see \cite{AN1,AN2,AN3,AN4,AFP} and references therein. As for the regimes corresponding to $\b \in (0,1)$ the leading order results for the ground state energy and a proof of condensation without optimal rate were first achieved in \cite{LNR}, which also covers the case of attractive interactions in the
regime of stability. Differently, the validity of Bogoliubov predictions (for all $\b \in (0,1)$) have been established so far only for non negative interaction \cite{BBCS2}.
\medskip
Let us now consider regimes more singular than the $3d$ Gross-Pitaevskii, by taking $\mathfrak{a} \sim R$ and letting the product $\mathfrak{a} N$ grow with $N$.
To describe these regimes we introduce a parameter $\kappa \in [0,2/3)$ and consider the following Hamilton operator on $L^2([0,1]^3)$:
\begin{equation} \label{eq:HN-kappa}
H_{N,3d}^{(\kappa)}= \sum_{i=1}^N -\D_{x_i} + \sum_{1\leq 1 < j \leq N} N^{2(1-\kappa)} V(N^{1-\kappa} (x_i- x_j))\,,
\end{equation}
so that $\mathfrak{a} N \sim N^\kappa$. We note that the choice $\kappa=0$ corresponds to the Gross-Pitaevskii regime. On the other hand, by rescaling lengths so that the scattering length is of order one, we see that \eqref{eq:HN-kappa} is equivalent to studying a Hamiltonian as in \eqref{eq:HT} on the torus $[0,L_\kappa]^3$ with $L_\kappa = N^{\kappa-1}$. This corresponds to systems with density $\r_\kappa = N^{3\kappa-2}$, hence less and less dilute as $\kappa \to 2/3$. In particular the choice $\kappa=2/3$ would correspond to the thermodynamic limit. From the point of view of length scales, proving BEC for Hamiltonians described by \eqref{eq:HN-kappa} is equivalent to showing condensation up to length scales of the order
\begin{equation} \label{eq:L3d}
L \sim (\r \mathfrak{a})^{-1/2} (\r \mathfrak{a}^3)^{-\a}\,, \qquad \a=\kappa /(4-6\kappa)\,.
\end{equation}
Indeed, the analysis carried out in \cite{LS,LS1} allows to show condensation also beyond the Gross-Pitaevskii regime, up to $\kappa<1/10$. On the other hand, in the recent paper \cite{F}, this result was extended to all $\kappa<2/5$.
As for the validity of Bogoliubov theory, this was established in \cite{BCaS} for sufficiently small $\kappa>0$, by exploiting key estimates obtained in \cite{ABS}. Finally, we mention \cite{BCS} and \cite{Ba} where a second order upper bound on the ground state energy compatible with \eqref{eq:LHY} is proved for all $\kappa<7/12$.
We will discuss in the next two sections how the analysis of Hamiltonians of the form \eqref{eq:HN-kappa} is closely related to proving lower and upper bounds compatible with \eqref{eq:LHY}. In particular, current methods (which will be discussed below) allow to obtain lower and upper bounds compatible with \eqref{eq:LHY} starting from systems defined on a torus of side length $L\sim N^{\kappa-1}$, at the expenses of errors which are smaller than $\r^2 \mathfrak{a} \sqrt{\r \mathfrak{a}^3}$ if $\kappa>0$ in the case of lower bounds, and $\kappa>1/2$ for upper bounds.
\bigskip
{\it Thomas-Fermi regime.} A scaling regime which is closely related to the Gross-Pitaevskii one, but will not discussed in these notes is the Thomas-Fermi regime \cite[Ch. 6]{LSSY}. In this regime
one considers a system of interacting bosons in a dilute regime with scattering length $\mathfrak{a} \gg R$ and so that the product $N \mathfrak{a}$ (resp. $N\mathfrak{b}$ in $2d$) grows with $N$. In the large $N$ limit the ground state energy per particle is thus described by a Gross-Pitaevskii energy functional where
the gradient term becomes negligible compared to the other terms, see \cite[Thms.6.3 and 6.6]{LSSY}.
The Thomas-Fermi limit is particularly relevant to derive the effective behaviour of rapidly rotating Bose gases in anharmonic traps, see \cite{BCPY}.
Recently, a simultaneous Thomas-Fermi and mean field scaling was considered in \cite{DG}.
\bigskip \bigskip
{\bf Two dimensions.}
Scaling limits interpolating between the mean field and Gross-Pitaevskii scalings (hence corresponding to those described by \eqref{eq:HN-beta} in the three dimensional case) are given in two dimensions by the family of Hamiltonians:
\begin{equation} \label{eq:HN-beta2d}
H_{N,2d}^{(\b)}= \sum_{i=1}^N -\D_{x_i} + \sum_{1\leq 1 < j \leq N} N^{2\b-1} V(N^\b (x_i- x_j)) \qquad \text{on} \qquad L^2_s([0,1]^2)
\end{equation}
for any $\b\geq 0$ independent on $N$. Note in particular that the choice $\b \in [0,1/2)$ corresponds to high density regimes, while setting $\b>1/2$ we are describing low density regimes. Eventually the parameter $\b \geq 0$ might be let to depend on $N$, provided that
\begin{equation} \label{eq:betamax}
\lim_{N \to \io} (\log N^\b)/N=0\,,
\end{equation}
so that \eqref{eq:HN-beta2d} does not include the Gross-Pitaevskii regime. The Hamiltonian \eqref{eq:HN-beta2d} describes a weakly interacting system with $\l \sim N^{-1}$, hence in particular, using the notation
$V_N(x)=N^{2\b-1} V(N^\b x)$, we have $\mathfrak{b} \sim \widehat V_N(0)\sim N^{-1}$. This has to be compared with the Gross-Pitaevskii regime, where $\mathfrak{b} \sim N^{-1}$ is much smaller than the integral of the Gross-Pitaevskii potential, which is of order one.
Systems described by \eqref{eq:HN-beta2d} were first analysed in \cite{LNR,LNR1,NR}. The method used in these papers (see \cite{Rougerie} for a review) allows to obtain a proof of Bose-Einstein condensation and a derivation of the leading order term in the asymptotics for the ground state energy
(which is also valid for non positive potential) provided $\b<1$. These results were later extended in \cite{Caraci}, where an optimal bound for Bose-Einstein condensation has been obtained for all $\b>0$ (up to the threshold defined by \eqref{eq:betamax}), in the case of non negative potentials. The latter result is at the core of a proof of the validity of Bogoliubov theory for systems described by \eqref{eq:HN-beta2d}, see \cite{Caraci2}.
\medskip
Regimes analogous to those described in \eqref{eq:HN-kappa}, less dilute than the Gross-Pitaevskii regime, can be obtained in two dimensions by considering Hamiltonians of the form
\begin{equation} \label{eq:HN-kappa2d}
H_{N,2d}^{(\kappa)}= \sum_{i=1}^N -\D_{x_i} + \sum_{1\leq 1 < j \leq N} e^{2N^{(1-\kappa)}} V( e^{N^{(1-\kappa)}} (x_i- x_j)) \qquad \text{on} \qquad L^2_s([0,1]^2)\,,
\end{equation}
for $\kappa \in [0,1)$. Indeed with this choice of the interaction we have $\mathfrak{b} \sim N^{-1+\kappa}$. Note that regimes as in \eqref{eq:HN-kappa2d} interpolates between the Gross-Pitaevskii regime $\kappa=0$ and the thermodynamic limit (where the interaction is not rescaled and $\mathfrak{b}\sim 1$). From the point of view of length scales, the condition $\mathfrak{b} \sim N^{-1+\kappa}$ is equivalent to fixing
\begin{equation} \label{eq:L2d}
L \sim (\r \mathfrak{b})^{-1/2} \,\mathfrak{b}^{-\a}\,, \qquad \a = \kappa/2(1-\kappa)
\end{equation}
A proof of BEC for regimes described by \eqref{eq:HN-kappa2d} for all $\kappa<1/6$ (corresponding to length scales as in \eqref{eq:L2d} with $\a<1/10$) can be obtained using the methods \cite{LSY2}. For weak interactions, order by orders bounds in perturbation theory compatible with the occurrence of condensation were proved in \cite{CG} up to length scales as in \eqref{eq:L2d} with $\a<1$ (corresponding to $\kappa<2/3$). An upper bound for the ground state energy compatible with \eqref{eq:LHY2d} and valid for all $\a>0$ (namely for all $\kappa<1$) has been recently shown in \cite{FGJMO}.
\bigskip
{\bf Summarising.} In this section we reviewed recent results where the occurrence of condensation and the validity of Bogoliubov prediction for the ground state energy were proved for systems defined on finite length scales.
In the next two sections we will discuss how the progress in proving Bose-Einstein condensation and the validity of Bogoliubov theory on finite, but increasingly larger, boxes has been linked to
the advances in the understanding of the behaviour of interacting bosons in the thermodynamic limit. The mathematical tools allowing to obtain information about thermodynamic functions, starting from the properties of finite-size systems, are usually referred to as {\it localization methods}. Since different localization methods are required to obtain lower and upper bounds compatible with \eqref{eq:LHY} and \eqref{eq:LHY2d}, we will discuss the two problems separately, trying to emphasise the major difficulties to be overcome in the two situations.
\section{Thermodynamic limit: lower bounds} \label{sec:LB}
The aim of this section is to review the methods allowing to reduce the problem of obtaining a lower bound on the specific ground state energy of \eqref{eq:HT}, to a proof of Bose-Einstein condensation on a sufficiently large box.
We begin with the pioneering paper by Lieb-Yngvason \cite{LY} back in 1998, to arrive to the very recent localization technique developed by Briezke, Fournais and Solovej \cite{BrFS}, finally allowing to capture
the second order correction in \eqref{eq:LHY} and \eqref{eq:LHY2d}, as a lower bound, for general non negative interactions, including the hard core case, both in three \cite{FS,FS2} and in two dimensions \cite{FGJMO}.
\medskip
{\bf The Neumann bracketing.} It may seem a long time ago, but it was only in 1998
that a lower bound compatible with the leading order in \eqref{eq:LHY} was first obtained \cite{LY}.
To achieve this result Lieb and Yngvason implemented a localization method usually referred to as {\it the box method} or {\it Neumann bracketing} (see \cite[Eq.(4)~--~ (12)]{LY}). Namely, they decomposed the system described by \eqref{eq:HT} on a box of side $T$ (with Neumann boundary conditions) into boxes of side $L$ with Neumann boundary conditions as well, getting the lower bound
\begin{equation}\label{eq:box_method}
\frac{E_3(N,T)}{N} \geq \frac{1}{\r L^3} \inf_{c_n}\sum_{n\geq 0}c_n E_3(n,L)\,,
\end{equation}
where the coefficients $c_n$ (relative number of boxes containing $n$ particles) have to satisfies the constraints $\sum_{n\geq 0}c_n=1$ and $\sum_{n\geq 0}nc_n=\r L^3.$ The difficulty then relies in proving that to minimize the total energy the number of particles has to be approximately uniform in all boxes. To do so the authors exploit the sub-additivity property of the ground state energy, which follows from the positivity of the interaction.
Eq. \eqref{eq:box_method} reduces the problem of getting a lower bound for $E_3(N,T)$ to the study of the Bose gas in a box of side length $L=\r^{-1/3}(\r \mathfrak{a}^3)^{-\d}$ for some small $\d>0$, with Neumann boundary conditions. Here the condition $L\gg \r^{-1/3}$, ensures that each box contains a large number of particles $n$ so that $n(n-1) \simeq n^2$. The strategy used in \cite{LY} to study the last problem is based on two by now very famous steps. The first one is a generalization of a lemma due to Dyson (see \cite{Dy}), allowing to replace the original interaction by a more regular potential living on a longer scale, at the expense of part of the kinetic energy (we denote by $\varepsilon>0$ the fraction of kinetic energy left after this step). The second step consists in regarding the regularized interaction as a (non negative) perturbation of the remaining kinetic energy, using Temple's inequality. Here a restriction on $L$ related to the gap of the Laplacian comes into play. Indeed, for Temple's inequality to make sense the kinetic energy of the first excited state in the small box, proportional to $ \varepsilon L^{-2}$, has to control the expectation of the interaction energy on the ground state of the kinetic energy, which is of order $\mathfrak{a} \r^2 L^3$ (see \cite[Eq. (28)]{LY}). This constrains $L$ to be so that $ L \ll L_{\text{T}}:=\varepsilon^{1/5} (\r \mathfrak{a})^{-1/2} (\r \mathfrak{a}^3)^{1/10}$.
As a matter of fact, more conditions are required in the course of the proof to control error terms, which are met for
$L \sim \r^{-1/3} (\r \mathfrak{a}^3)^{-1/51}
\ll L_T$.
The same strategy also works in two dimensions \cite{LY2d}, provided a suitable modification of the Dyson Lemma, to account for the logarithmic behaviour of the two dimensional scattering function.
The proof of the leading order term in \eqref{eq:LHY2d} is thus achieved by studying the problem on a box with side length of the order $ \r^{-1/2} \,|\log(\r \mathfrak{a}^2)|^{1/10}$.
The construction in \cite{LY, LY2d} allows also to show the occurrence of Bose-Einstein condensation in $3d$ and $2d$ over lengths scale of order
$(\r \mathfrak{a})^{-1/2}(\r\mathfrak{a}^3)^{-1/34}$
\cite{LS}, and $\r^{-1/2} |\log(\r \mathfrak{a}^2)|^{1/10}$ \cite{LSY2} respectively (see the two paragraphs after \eqref{eq:L3d} and \eqref{eq:L2d} above for a summary about the length scales over which condensation can be currently shown).
\medskip
{\bf The sliding method.} For almost ten years it was not clear how to go beyond the methods of \cite{LY} to rigorously derive the second term in \eqref{eq:LHY}. In \cite{GiuS} -- following a strategy applied first to the bosonic jellium in \cite{LS-jellium, LS-charged} -- Giuliani and Seiringer finally provided a rigorous derivation of \eqref{eq:LHY} in a weak coupling and high density regime satisfying \eqref{eq:Bog-exact} with $a_0/R = (\r \mathfrak{a}^3)^\g$ and $\g$ close to $1/2$. As previously discussed, under the conditions \eqref{eq:Bog-exact} Bogoliubov's heuristic strategy can be proven to be correct. In particular it is possible to show that in the ground state of \eqref{eq:HT} the ratio among the number $n_+$ of particles orthogonal to the condensate and the total number $n$ of particles in a box of side length $L$ is bounded by \cite[Lemma III.4]{GiuS}
\[ \label{eq:BEC-GiuS}
\frac{n_+}{n} \leq C a_0 L^2 R^{-3}\,.
\]
In other words condensation occurs up to $L \ll (\r \mathfrak{a})^{-1/2}(\r \mathfrak{a}^3)^{(1-3\g)/2}$. We remark that with this side length the usual localization error $L^{-2}$ would be of order $\r \mathfrak{a} (\r \mathfrak{a})^{1/2 +3(\g-1/2)}$, which is larger than the Lee-Huang-Yang second order correction for any $\g<1/2$. In fact a key role in the paper by Giuliani and Seiringer is played by the sliding method introduced in \cite{CLY}, where a lower bound is achieved by decomposing the thermodynamic box into small boxes, and then averaging over all possible locations of the boxes.
Let us state the result of the sliding localization, as used in \cite{GiuS}. Let $\L_T=[-T/2,T/2]^3$ be a three dimensional torus, and let us consider on $L^2_s(\L_T^{N})$ the Hamiltonian
\begin{equation} \label{eq:Hprime}
H'_{T,N} = - \sum_{i=1}^N \D_{x_i} + \frac{a_0}{R_0^3} \sum_{1\leq i < j \leq N} v_{R_0}(x_i-x_j) - 4 \pi N \r a_0 \,,
\end{equation}
with $v_{R_0}(x)= \sum_{n \in \mathbb{Z}^3} e^{-|x+n T|/R_0} $. Note that $H'_{T,N}$ is obtained by considering in \eqref{eq:HT} an interaction of the form $V(x)= a_0 R_0^{-3} v_{R_0}(x)$ and shifting the ground state by the leading order term in \eqref{eq:LHY}, after having used the Born series to replace $\mathfrak{a}$ with $a_0$, see \eqref{eq:Born}. Hence the ground state of $H'_{T,N}$ will be negative.
We now introduce a function localizing the particles on lengths smaller than $T$. We fix a small parameter $t=(\r a_0^3)^{\t}$, and define $\chi \in C_0^\io(\mathbb{R}^3)$, spherically symmetric, $0 \leq \chi \leq 1$, such that $\chi(x)$ is supported in $[(-1+t)/2, (1-t)/2]^3$ and is identically one in the smaller box $[(-1+2t)/2, (1-2t)/2]^3$. Then we denote by $\chi_L(x)$ the function on the torus $\L_T$ defined by $\chi_{L}(x)= \sum_{n \in \mathbb{Z}^3} \chi(L^{-1}(x+n T)) $ and with
\[
w_{R_<}(x,y)= \chi(x/L) e^{-|x-y|/R_<}\chi(y/L)
\]
a localized interaction with range $R_<$. On $L^2([-L/2, L/2]^{3n})$ we define the small box Hamiltonian
\begin{equation} \begin{split} \label{eq:HLn}
& H_{L,R}^n = - \sum_{j=1}^n \D_L^{(j)} \\
& + \frac{ a_0 I R }{R_0^4} \bigg[ \sum_{1\leq i < j \leq n} w_R(x_i,x_j) - \r \sum_{j=1}^n \int_{\mathbb{R}^3} w_R(x_j,y)dy +\frac{\r^2}{2} \int_{\mathbb{R}^3\times \mathbb{R}^3} w_R(x,y)dx dy\bigg]\,,
\end{split}\end{equation}
where $\D_L^{(j)}$ is the Neumann Laplacian for the $j$-th particle in the cube $[-L/2,L/2]$ and $I= (\int \chi^2(y) dy)^{-1}$. The following lemma holds:
\begin{lemma} Let $E'_{N,T}$ and $E_{n,L}^{(R)}$ be the ground state energy of the Hamiltonian defined in \eqref{eq:Hprime} and \eqref{eq:HLn} respectively. Then, for $t L/R_0$ large enough, there exists a function of the form $\o(t)= \text{const.} (t L / R_0)^{-1} $ such that if we set $R^{-1}= R_0^{-1} + \o(t)/L$ we have
\begin{equation} \label{eq:GSloc}
\lim_{N \to \io} \frac{E'_{N,T}}{N}\geq \frac{1}{\r L^3} \inf_n E_{n,L}^{(R)} - \frac{a_0 \o(t) R}{2L R_0^3}\,.
\end{equation}
\end{lemma}
The last term on the r.h.s. of \eqref{eq:GSloc} is the error due to the localization to the smaller box. It is proportional to $ \frac{1}{L^2} \frac{a_0}{t R_0}$. Hence, w.r.t. the `standard' localization error $L^{-2}$
we gain some extra smallness from the high density condition $a_0/R_0 = (\r a_0^3)^{\g}$, for $\g \in (1/3, 1/2)$.
In particular, by choosing
\[
t = (\r a_0^3)^{\t}\,, \quad L= (\r a_0)^{-1/2} (\r a_0^3)^{-b}\,, \quad a_0/R_0 = (\r a_0^3)^{1/2 -\nu}
\]
we have that the localization error is smaller than the second-order correction to the ground state energy if $2b >\nu +\t$. Hence, the high density condition in \cite{GiuS} has two key advantages: first it makes the analysis on the small box `simpler' since one only needs to follow Bogoliubov's strategy; moreover it allows to improve the localization error. Note, however, that the conditions in \cite{GiuS} correspond to a very high density regime. In particular the authors fix $\nu <1/69$ (see Fig.~\ref{Fig:Bog}).
\bigskip
{\bf The small box localization.}
Here we discuss the localization method applied in \cite{BrFS} to obtain a lower bound which captures the correct second order with the wrong constant. The main difference w.r.t. \cite{GiuS} is that here the Hamiltonian in the small box exhibits a modified kinetic energy. Consider the Hamiltonian $H_T''$ on the bosonic Fock space $\mathcal{F}_S(L^2(\Lambda_T))$ built on the box $\Lambda_T=[-T/2,T/2]^3$ with Dirichlet boundary conditions and whose restriction on the $N$-particle subspace is given by
\begin{equation*}\label{eq:Hprimeprime}
H_{T,N}''=-\sum_{i=1}^N\Delta_{x_i}+\sum_{i<j}v(x_i-x_j)-8\pi N \rho \mathfrak{a}\,.
\end{equation*}
Note that the Hamiltonian $H_{T,N}''$ is formulated in the grand canonical setting and the term $-8 \pi N \r \mathfrak{a} $ in the above definition can be interpreted as the addition of a term $-\m N$ fixing the total number of particles to $N$, with $\mu= 8 \pi \r \mathfrak{a} $ the expected chemical potential on the basis of Bogoliubov theory.
Clearly, an estimate from below on $H_{T,N}''$ provides a lower bound on the specific ground state energy $e_3(\rho)$ defined in \eqref{eq:e_d} since the canonical ground state can be considered as a trial state for the grand canonical Hamiltonian. More precisely,
\[
e_3(\rho)\geq \lim_{|\Lambda_T|\to\infty}\frac{1}{|\Lambda_T|}\inf_{\psi\in \mathcal{F}_S(L^2(\Lambda_T))}\frac{\langle\psi,H_{T,N}''\psi\rangle}{\|\psi\|^2}+8\pi\mathfrak{a}\rho^2\,.
\]
We are going to localize on a small box of side length $L=K( \rho \mathfrak{a})^{-1/2}$ for $K>0$ to be chosen sufficiently small but independent of $\rho.$ To write the Hamiltonian in the small box we first introduce the localization function $\chi\in C^\infty_0(\mathbb{R}^3)$ even s.t. $0\leq \chi\leq 1$, $\int \chi^2=1$ and $\supp\chi\subset[-1/2,1/2]^ 3.$ For a given $u\in \mathbb{R}^3$ we set $\chi_u(x)=\chi(x/L-u),$ then the potential in the small box $uL+[-L/u,L/u]^3$ centered in $uL$ is given by
\[
w_u(x,y)=\chi_u(x)\frac{v(x-y)}{(\chi\ast\chi)((x-y)/L)}\chi_u(y)\,.
\]
We also introduce the notation
\[
\mathcal{W}_u(x_1,\dots,x_N)=\sum_{i<j}w_u(x_i,x_j)+\sum_{i=1}^N-\rho\int w_u(x_i,y)f(x_i-y)dy\,,
\]
where $f$ is the scattering solution defined in \eqref{eq:scatt_3d}. A direct calculation shows that
\begin{equation}\label{eq:pot}
\sum_{i<j}v(x_i-x_j)-8\pi N\rho\mathfrak{a}= \int_{L^{-1}(\Lambda_T+B(0,L))}\mathcal{W}_u(x_1,\dots,x_N)du\,.
\end{equation}
On the other hand, the usual kinetic energy is replaced in the small box by the operator
\[
\mathcal{T}_u=Q_u\Big[\chi_u\big(-\Delta-\frac14(sL)^{-2}\big)_+\chi_u+b/L^2\Big]Q_u\,,
\]
where $s,b>0$, $(\cdot)_+$ denotes the positive part and $Q_u$ is the orthogonal projection defined by $Q_u\phi=\theta_u\phi-L^{-3}(\theta_u,\phi)\theta_u$ with $\theta_u$ the indicator function of the box centered in $uL.$ Let us comment on the operator $\mathcal{T}_u$. First we note that it vanishes on constant functions, moreover the last term in the square bracket plays the role of a Neumann gap, namely $\mathcal{T}_u$ on functions orthogonal to constants is bounded below by at least $bL^{-2}.$ Finally, $(-\Delta- \frac14(sL)^{-2})_+$ is to be understood as the multiplication operator $(p^2-\frac14(sL)^{-2})_+$ in Fourier space,
hence momenta $p$ of order $(\rho\mathfrak{a})^{1/2},$ which are relevant to capture the second order in the energy (as discussed in Sec. \ref{sec:intro}) are correctly recovered.\\
One can show that the kinetic energy in the thermodynamic box can be bounded by an integral over the kinetic operators $\mathcal{T}_{u}$ on the small boxes, namely there exists a $b$ such that, if $s$ is small enough then
\begin{equation}\label{eq:kin}
-\Delta\geq \int_{L^{-1}(\Lambda_T+B(0,L))}\mathcal{T}_{u} \,du\,.
\end{equation}
The proof of the above inequality relies on the general fact that given $\mathcal{K}:\mathbb{R}^3\to[0,\infty)$ symmetric and polynomially bounded, then the operator
\[
T=\int_{\mathbb{R}^3}Q_u\chi_u\mathcal{K}(-iL\nabla)\chi_uQ_u \,du
\]
is such that $T=F(-iL\nabla)$, \hbox{\it i.e.\ } in Fourier space it is the multiplication operator times $F(p)$ where
\[
F(p)=(2\pi)^{-3}\mathcal{K}\ast|\widehat{\chi}|^2(p)-2(2\pi)^{-3}\widehat{\theta}(p)\widehat{\chi}\ast(\mathcal{K}\widehat{\chi})(p)+(2\pi)^{-3}\Big(\int \mathcal{K}|\widehat{\chi}|^2\Big)\widehat{\theta}(p)^2\,,
\]
where we recall the definition \eqref{eq:Fourier} of Fourier transform $\widehat{f}(p)$ of $f$.
While the proof of \eqref{eq:kin} is quite technical, it can be regarded as an extension of the IMS localization formula. Indeed, the latter can be stated as follows: the operator $\int_{\mathbb{R}^3} \chi_u\mathcal{K}(-i\nabla)\chi_u$ with $\mathcal{K}(p)=p^2$ acts in Fourier space as the multiplication by
\[
(2\pi)^{-3}\mathcal{K}\ast|\widehat{\chi}|^2=p^2+\int |\nabla\chi|^2\,.
\]
We define the Hamiltonian $H_{L,u}$ in Fock space acting on the $N$ particle sector as
\[
H_{L,u,N}=\sum_{i=1}^N\mathcal{T}_{u,i}+\mathcal{W}_u(x_1,\dots,x_N)\,,
\]
and noting that $H_{L,u}$ and $H_{L,u'}$ are unitarily equivalent one concludes from \eqref{eq:pot}, \eqref{eq:kin}
\begin{equation*}\label{eq:bound_small_box}
\lim_{|\Lambda_T|\to\infty}\frac{1}{|\Lambda_T|}\inf_{\psi\in \mathcal{F}_s(L^2(\Lambda_T))}\frac{\langle \psi,H_{T,N}''\psi\rangle}{\|\psi\|^2}\geq\frac1{ L^{3}}\inf\mathrm{spec}H_{L,0}\,.
\end{equation*}
The main result in \cite{BrFS} is the following estimate on the Hamiltonian in the small box:
\begin{equation*}\label{eq:en_small_box}
H_{L,0}\geq -4\pi \mathfrak{a}\rho^2L^3-C\rho^2\mathfrak{a}\sqrt{\rho \mathfrak{a}^3}L^3-C\rho^3\mathfrak{a}^2R^2L^3\,,
\end{equation*}
where $R$ is the range of the interaction. Hence, assuming $\rho$ to be small enough
\[
e_3(\rho)\geq 4\pi\mathfrak{a}\r^2\big(1-C(\sqrt{\r\mathfrak{a}^3}+R^2\mathfrak{a}\r)\big)\,.
\]
Note that the localization method described above allows to prove BEC on length scales $L=C_L(\r\mathfrak{a})^{-1/2} (\r\mathfrak{a}^3)^{-\a}$ for $\a < 1/4$ as shown in \cite{F}.
\bigskip
{\bf The double box localization.}
This is the method introduced in \cite{BriS} to extend the result obtained in \cite{GiuS} to scaled potentials of the form
\[
v_R(x)=\frac1{R^3}v(x/R)
\]
for a given $v$ non-negative, spherically symmetric and compactly supported under the assumptions (see \hbox{\it e.g.\ } \cite[Thm 1.1]{BriS}) that
\begin{equation}\label{eq:scaling_BFS}
\lim_{\r\to 0}\frac{R}{\mathfrak{a}}(\r\mathfrak{a}^3)^{1/2}=0, \qquad \text{ and }\qquad \lim_{\r\to 0}R\r^{1/3}(\r\mathfrak{a}^3)^{-1/30}=\infty\,.
\end{equation}
In particular, assuming $\mathfrak{a}/R=(\r\mathfrak{a}^3)^\gamma$, the condition \eqref{eq:scaling_BFS} implies $3/10<\g<1/2.$ Hence, in contrast to \cite{GiuS} where $1/2-1/69<\gamma<1/2$ was needed, here also {\it low density} regimes are handed. In fact, choosing $3/10<\gamma<1/3$, we get $R\r^{1/3}=(\r\mathfrak{a}^3)^{1/3-\g}\ll1$ (see Fig. \ref{Fig:Bog}).
The novelty in the localization used in \cite{BriS} is to proceed in two steps: the system is first localized to ``large" boxes of side length $L\gg(\r\mathfrak{a})^{-1/2}$ and then to ``small'' boxes smaller than $(\r\mathfrak{a})^{-1/2}$.
The method, which is an ingenious extension of the sliding method described above, comes with a subtle modification of the kinetic energy, see \cite[Eq.(57)]{BriS} for the large box kinetic operator, and \cite[Eq.(64)]{BriS} for the small box one. On the small box, slightly smaller than the Gross-Pitaevskii scale: i) the number of excitations is much smaller than the total number of particles; ii) due to conditions \eqref{eq:scaling_BFS}, which imply the chain of inequalities in \eqref{eq:Bog-exact}, the heuristics proposed by Bogoliubov can be proven to be correct; in particular error terms can be absorbed in the gap.
In a second stage, the estimates on the energy and number of excitation obtained on the small boxes yield a priori bounds on the energy, number of particles (which will be shown to be close to $\r L^3$) and excited particles in the large box
\cite[Sec. 6]{BriS}.
The localization technique from \cite{BriS} was exploited by Fournais and Solovej in the papers \cite{FS, FS2} allowing to get a lower bound compatible with \eqref{eq:LHY}, for general non negative interactions, thus closing a longstanding problem in mathematical physics. This result was achieved thanks to a more refined analysis of the large box Hamiltonian, allowing the authors to extract the missing contributions to the ground state energy hidden in the cubic and quartic terms neglected in Bogoliubov theory, and which could not be extracted by restricting to quasi-free states as in \cite{NRS1,NRS2}. We refer to \cite[Sec. 2]{FS2} for a description of the general strategy pursued in \cite{FS,FS2} and the additional complications to be overcome to treat potentials with large $L^1$ norm.
The same localization method is at the core of the recent paper \cite{FGJMO}, where a lower (and upper) bound matching \eqref{eq:LHY2d} has been recently obtained. There, similarly as in three dimensions, the small and large boxes are taken to be slightly smaller or larger than the Gross-Pitaevskii scale $(\r \mathfrak{b})^{-1/2}$. In two dimensions, however, the authors had to face similar challanges as in the $3d$ hard core case, even for regular potentials, being the effective parameter $\mathfrak{b}$ much smaller than the integral of the potential.
\bigskip
{\bf Summarizing.}
A major challenge for obtaining a second order lower bound compatible with \eqref{eq:LHY} has been the development of a localization method allowing to obtain a Neumann gap both on the small and large boxes (to be able to absorb error terms), still keeping essentially the original kinetic energy at the relevant Bogolubov scales (see \cite[Remark 5.5]{FS} for a discussion of the localized kinetic energy in \cite{BriS, FS}). Thanks to this localization bound, the problem is reduced to showing the
validity of Bogoliubov theory on a box of side length slightly larger than the Gross-Pitaevskii length scale. On this scale, it is essential to extract some missing energy from cubic and quartic terms neglected in Bogoliubov theory, which cannot be captured by restricting to quasi-free states. Indeed one needs to take into account correlations among triples of excitations. This idea, which appeared first in the trial state by Yau-Yin \cite{YY}, is also at the core of the works \cite{BBCS4} and \cite{FS} (where it appeared formulated with very different languages though), and has been later exploited in all works aimed at applying Bogoliubov approach at the Gross-Pitaevskii scale and beyond (see references in Sec. \ref{sec:scalings}.)
An alternative route to the double box localization would require studying directly the Neumann problem rather than the periodic one, on a sufficiently large box.
Very recently, Boccato and Seiringer \cite{BoSei} managed to study the properties of three dimensional Bose gases with Neumann boundary conditions directly on boxes of side length $L \sim (\r \mathfrak{a})^{-1/2}$. Via Neumann bracketing this implies a
lower bound for the leading order of the ground state energy per particle of a Bose gas in the thermodynamic limit, with almost optimal error (the error in \cite{BoSei} is $\r \mathfrak{a} (\r \mathfrak{a}^3)^{1/2}|\log(\r \mathfrak{a}^3)|$ due to a finite size effect related to the Neumann boundary conditions, see \cite[Eq.(1.12)]{BoSei}).
The extension of this result to
boxes of side length larger than $(\r \mathfrak{a})^{-1/2}$ would allow for an alternative proof of a lower bound compatible with \eqref{eq:LHY},
without need of refined localization techniques.
\section{Thermodynamic limit: upper bounds} \label{sec:UB}
In this section we review the methods developed in the literature to get an upper bound on the specific ground state energy $e_d(\rho)$ defined in \eqref{eq:e_d}.
An upper bound correct at leading order was first obtained in \cite{Dy} for the three dimensional hard-core Bose gas and later extended in \cite{LSY} to cover all non-negative interactions and in \cite{LY2d} to treat the two dimensional case. To get an upper bound on the ground state energy it is enough to evaluate the energy on an appropriate state. In the aforementioned papers the authors considered a trial state of the form
\begin{equation}\label{eq:dyson}
\Psi(x_1,\dots,x_N)=F_1( x_1)F_2( x_2)\dots F_N(x_1,\dots,x_N)
\end{equation}
where
\[
F_i( x_1,\dots,x_i)= f_\ell(t_i),\quad t_i=\min_{j=1,\dots,i-1}|x_i-x_j|\,,
\]
and the function $f_\ell$ is a modified version (normalized to 1 at length $\ell$) of the scattering function $f$ defined in \eqref{eq:scatt_3d} in three dimensions, and it is defined by \eqref{eq:scatt_2d} in two dimensions. In \eqref{eq:dyson} the cutoff is put at a length scale of the order $\r^{-1/d}$ (see \cite[Eq. (2.19)]{LSSY} for the precise definition).
The state \eqref{eq:dyson} has to be thought of as a modification of the so-called Jastrow factor, namely the wave function
\begin{equation} \label{eq:jastrow}
\prod_{i<j}f_\ell(x_i-x_j)
\end{equation}
considered in \cite{J}. The idea behind the Jastrow trial state \eqref{eq:jastrow} is the following: in contrast with the non-interacting case where the ground state was just the constant wave function, in the interacting case, correlations have to be considered when two particles come closer. In particular, in the dilute regime, it seems reasonable to expect that only two-body correlations matter, and that two-body correlations are well described by the zero energy scattering function.
The trial state \eqref{eq:jastrow} is relatively manageable for $\ell \ll (\r \mathfrak{a})^{-1/2}$ in $3d$ and $\ell \ll (\r \mathfrak{b})^{-1/2}$ in $2d$; we refer the reader to \cite[Sect. 2.1]{LSSY} and \cite[Sect. 2]{BCOPS-rev} for a computation of the energy of \eqref{eq:dyson} and \eqref{eq:jastrow} respectively. However, it seems difficult to compute the energy of a state as in \eqref{eq:jastrow} for $\ell$ larger than the healing length, a choice which would be needed to capture the second order term in \eqref{eq:LHY}, as discussed in Sec.\ref{sec:intro}.
In fact, the next order in the energy was first tackled in three dimensions in \cite{ESY} with a different method, closer to Bogoliubov heuristics, as described in Sec.\ref{sec:intro}. Finally, it was in \cite{YY} that the Lee-Huang-Yang correction was obtained as an upper bound, by extending the trial state considered in \cite{ESY}, including correlation among triples of excited particles ({\it soft-pair creation} in the language of the paper) and exploiting a localization method discussed in the next subsection. The approach of \cite{YY} has been reviewed and adapted to a grand canonical setting in \cite{Aaen}. The same localization technique, with a simpler trial state, was then used in \cite{BCS} to cover all potentials $V \in L^3.$ Note that a second order upper bound also covering the hard sphere case in three dimensions is still missing.
On the other hand, in the two dimensional case an upper bound in agreement with \eqref{eq:LHY2d} was recently obtained in \cite{FGJMO} for a very general class of potentials including hard-core interactions. This result was achieved by combining the localization method of \cite{YY} (applied to the two dimensional case) with a trial state obtained multiplying the Jastrow function by a quasi-free state. Note that a similar trial state has been used in \cite{BCOPS} to get a second-order upper bound on the energy for hard core bosons in the three-dimensional Gross-Pitaevskii regime. The reason why in three dimension the same idea cannot be applied to the thermodynamic limit will be discussed at the end of this section.
\begin{figure}[t]
\centering a) \hskip 0.3cm
\begin{tikzpicture}[scale =1.1,
every path/.style = {},
]
\begin{scope}
\draw[thick, color=myred!60!white, fill=myred!60!white] (0.05,0.05) rectangle (0.95,0.95);
\draw[color=black,<->, thick] (0.05,-0.3) -- (0.95,-0.3);
\node[below, color=black] at (0.5,-0.4){$L_\a$};
\end{scope}
\end{tikzpicture} \hskip 0.5cm
\begin{tikzpicture}[scale =1.1,
every path/.style = {},
]
\begin{scope}
\draw[thick, color=myred!60!white, fill=myred!60!white] (0.05,0.05) rectangle (0.95,0.95);
\draw[dashed, color=black] (-0.26,-0.26) rectangle (1.26,1.26);
\draw[color=myred!5!white] (-0.25,-0.25) rectangle (1.25,1.25);
\draw[color=myred!8!white, fill=myred!8!white] (-0.2,-0.2) rectangle (1.2,1.2);
\draw[color=myred!15!white, fill=myred!15!white] (-0.15,-0.15) rectangle (1.15,1.15);
\draw[color=myred!30!white, fill=myred!30!white] (-0.1,-0.1) rectangle (1.1,1.1);
\draw[color=myred!40!white, fill=myred!40!white] (-0.05,-0.05) rectangle (1.05,1.05);
\draw[thick, color=myred!50!white, fill=myred!50!white] (0,0) rectangle (1,1);
\draw[thick, color=myred!60!white, fill=myred!60!white] (0.05,0.05) rectangle (0.95,0.95);
\draw[color=black,<->, thick] (-0.25,-0.6) -- (1.25,-0.6);
\node[below, color=black] at (0.5,-0.7){$L_\a + 2 \d$};
\end{scope}
\end{tikzpicture} \hskip 0.8cm b)
\begin{tikzpicture}[scale =0.6,
every path/.style = {},
]
\begin{scope}
\draw[thick, color=myred!60!white, fill=myred!60!white] (0.05,0.05) rectangle (0.95,0.95);
\draw[color=myred!5!white] (-0.25,-0.25) rectangle (1.25,1.25);
\draw[color=myred!8!white, fill=myred!8!white] (-0.2,-0.2) rectangle (1.2,1.2);
\draw[color=myred!15!white, fill=myred!15!white] (-0.15,-0.15) rectangle (1.15,1.15);
\draw[color=myred!30!white, fill=myred!30!white] (-0.1,-0.1) rectangle (1.1,1.1);
\draw[color=myred!40!white, fill=myred!40!white] (-0.05,-0.05) rectangle (1.05,1.05);
\draw[thick, color=myred!50!white, fill=myred!50!white] (0,0) rectangle (1,1);
\draw[thick, color=myred!60!white, fill=myred!60!white] (0.05,0.05) rectangle (0.95,0.95);
\draw[color=black](-0.6, -0.2)--(0.4, -0.2);
\draw[color=black](-0.6, -0.4)--(0.4, -0.4);
\node[above, color=black] at (-0.9,-0.6) {\footnotesize $R$};
\draw[color=black, dashed] (-0.4,1.4) rectangle (4.8,-3.8);
\draw[color=black,<->, thick] (5.2,1.4) -- (5.2,-3.8);
\node[below, color=black] at (6,0) { $T=$};
\node[below, color=black] at (7.2,-0.5) { $t \big(L_\a + 2 \d + R \big)$};
\draw[color=myred!5!white] (1.45,-0.25) rectangle (2.95,1.25);
\draw[color=myred!8!white, fill=myred!8!white] (1.5,-0.2) rectangle (2.9,1.2);
\draw[color=myred!15!white, fill=myred!15!white] (1.55,-0.15) rectangle (2.85,1.15);
\draw[color=myred!30!white, fill=myred!30!white] (1.6,-0.1) rectangle (2.8,1.1);
\draw[color=myred!40!white, fill=myred!40!white] (1.65,-0.05) rectangle (2.75,1.05);
\draw[thick, color=myred!50!white, fill=myred!50!white] (1.7,0) rectangle (2.7,1);
\draw[thick, color=myred!60!white, fill=myred!60!white] (1.75,0.05) rectangle (2.65,0.95);
\draw[color=myred!5!white] (3.15,-0.25) rectangle (4.65,1.25);
\draw[color=myred!8!white, fill=myred!8!white] (3.2,-0.2) rectangle (4.6,1.2);
\draw[color=myred!15!white, fill=myred!15!white] (3.25,-0.15) rectangle (4.55,1.15);
\draw[color=myred!30!white, fill=myred!30!white] (3.3,-0.1) rectangle (4.50,1.1);
\draw[color=myred!40!white, fill=myred!40!white] (3.35,-0.05) rectangle (4.45,1.05);
\draw[thick, color=myred!50!white, fill=myred!50!white] (3.4,0) rectangle (4.4,1);
\draw[thick, color=myred!60!white, fill=myred!60!white] (3.45,0.05) rectangle (4.35,0.95);
\draw[color=myred!5!white] (-0.25,-1.95) rectangle (1.25,-0.45);
\draw[color=myred!8!white, fill=myred!8!white] (-0.2,-1.9) rectangle (1.2,-0.5);
\draw[color=myred!15!white, fill=myred!15!white] (-0.15,-1.85) rectangle (1.15,-0.55);
\draw[color=myred!30!white, fill=myred!30!white] (-0.1,-1.8) rectangle (1.1,-0.6);
\draw[color=myred!40!white, fill=myred!40!white] (-0.05,-1.75) rectangle (1.05,-0.65);
\draw[thick, color=myred!50!white, fill=myred!50!white] (0,-1.7) rectangle (1,-0.7);
\draw[thick, color=myred!60!white, fill=myred!60!white] (0.05,-1.65) rectangle (0.95,-0.75);
\draw[color=myred!5!white] (1.45,-1.95) rectangle (2.95,-0.45);
\draw[color=myred!8!white, fill=myred!8!white] (1.5,-1.9) rectangle (2.9,-0.5);
\draw[color=myred!15!white, fill=myred!15!white] (1.55,-1.85) rectangle (2.85,-0.55);
\draw[color=myred!30!white, fill=myred!30!white] (1.6,-1.8) rectangle (2.8,-0.6);
\draw[color=myred!40!white, fill=myred!40!white] (1.65,-1.75) rectangle (2.75,-0.65);
\draw[thick, color=myred!50!white, fill=myred!50!white] (1.7,-1.7) rectangle (2.7,-0.7);
\draw[thick, color=myred!60!white, fill=myred!60!white] (1.75,-1.65) rectangle (2.65,-0.75);
\draw[color=myred!5!white] (3.15,-1.95) rectangle (4.65,-0.45);
\draw[color=myred!8!white, fill=myred!8!white] (3.2,-1.9) rectangle (4.6,-0.5);
\draw[color=myred!15!white, fill=myred!15!white] (3.25,-1.85) rectangle (4.55,-0.55);
\draw[color=myred!30!white, fill=myred!30!white] (3.3,-1.8) rectangle (4.5,-0.6);
\draw[color=myred!40!white, fill=myred!40!white] (3.35,-1.75) rectangle (4.45,-0.65);
\draw[thick, color=myred!50!white, fill=myred!50!white] (3.4,-1.7) rectangle (4.4,-0.7);
\draw[thick, color=myred!60!white, fill=myred!60!white] (3.45,-1.65) rectangle (4.35,-0.75);
\draw[color=myred!5!white] (-0.25,-3.65) rectangle (1.25,-2.15);
\draw[color=myred!8!white, fill=myred!8!white] (-0.2,-3.6) rectangle (1.2,-2.2);
\draw[color=myred!15!white, fill=myred!15!white] (-0.15,-3.55) rectangle (1.15,-2.25);
\draw[color=myred!30!white, fill=myred!30!white] (-0.1,-3.5) rectangle (1.1,-2.3);
\draw[color=myred!40!white, fill=myred!40!white] (-0.05,-3.45) rectangle (1.05,-2.35);
\draw[thick, color=myred!50!white, fill=myred!50!white] (0,-3.4) rectangle (1,-2.4);
\draw[thick, color=myred!60!white, fill=myred!60!white] (0.05,-3.35) rectangle (0.95,-2.45);
\draw[color=myred!5!white] (1.45,-3.65) rectangle (2.95,-2.15);
\draw[color=myred!8!white, fill=myred!8!white] (1.5,-3.6) rectangle (2.9,-2.2);
\draw[color=myred!15!white, fill=myred!15!white] (1.55,-3.55) rectangle (2.85,-2.25);
\draw[color=myred!30!white, fill=myred!30!white] (1.6,-3.5) rectangle (2.8,-2.3);
\draw[color=myred!40!white, fill=myred!40!white] (1.65,-3.45) rectangle (2.75,-2.35);
\draw[thick, color=myred!50!white, fill=myred!50!white] (1.7,-3.4) rectangle (2.7,-2.4);
\draw[thick, color=myred!60!white, fill=myred!60!white] (1.75,-3.35) rectangle (2.65,-2.45);
\draw[color=myred!5!white] (3.15,-3.65) rectangle (4.65,-2.15);
\draw[color=myred!8!white, fill=myred!8!white] (3.2,-3.6) rectangle (4.6,-2.2);
\draw[color=myred!15!white, fill=myred!15!white] (3.25,-3.55) rectangle (4.55,-2.25);
\draw[color=myred!30!white, fill=myred!30!white] (3.3,-3.5) rectangle (4.5,-2.3);
\draw[color=myred!40!white, fill=myred!40!white] (3.35,-3.45) rectangle (4.45,-2.35);
\draw[thick, color=myred!50!white, fill=myred!50!white] (3.4,-3.4) rectangle (4.4,-2.4);
\draw[thick, color=myred!60!white, fill=myred!60!white] (3.45,-3.35) rectangle (4.35,-2.45);
\end{scope}
\end{tikzpicture}
\caption{Schematic representation of the localization for the upper bound exploited in \cite{YY, Aaen, BCS, FGJMO}. The trial state on the thermodynamic box $T$ is obtained by gluing together $t^3$ small boxes with Dirichlet boundary conditions, leaving corridors of width $R$ among different boxes, see Fig. b). Each small box is obtained by enlarging a bit a box with side length $L_\a$ and periodic boundary condition, to obtain a trial state satisfying Dirichlet boundary conditions, see Fig. a). While on each small box the side length and the density are related, the thermodynamic limit is then taken by sending $t\to \io$ while keeping the density of the small boxes constant. }
\label{Fig:localization}
\end{figure}
\bigskip
{\bf Localization for second order upper bounds.} The proofs of the upper bounds matching \eqref{eq:LHY} (resp. \eqref{eq:LHY2d}) at second order obtained in \cite{YY, Aaen, BCS} (resp. \cite{FGJMO}) are all based on the construction of appropriate trial states for the Hamiltonian \eqref{eq:HT}. However, none of these trial states are constructed directly in $L^2_s (\L_T^N)$. Instead, it turns convenient (for reasons which will be clearer in a while) to build a trial state on smaller boxes, with side length $L_\a$ depending on $\r$, as in \eqref{eq:L3d} and \eqref{eq:L2d}.
It is an easy observation that a thermodynamic trial state can be then built by gluing together trial states obtained on smaller boxes, if we consider Dirichlet rather than periodic boundary conditions, and if we impose that particles in different boxes do not interact (by forcing them to be at distance larger than the range of the potential), see Fig.\ref{Fig:localization}. The core of the localization for the upper bound, whose standard proof \cite{R, YY, Aaen} is discussed in detail in \cite[App.~A]{BCS} and generalized to non compact potentials in \cite[App.~A]{FGJMO}, is identifying the correct length scale for which the error coming from working in with periodic rather than Dirichlet boundary conditions is of smaller order with respect to the Lee-Huang-Yang second order correction.
In the following we will keep the discussion at the level of ideas, while we refer to \cite[Sec. 2.1]{YY} and \cite[Prop.1.2]{BCS}, \cite[App. A]{FGJMO} for a precise statement of the localization result, in the canonical and in the grand canonical setting respectively. To this end, it is sufficient to compare the leading order contribution to the energy in the case of Dirichlet and periodic boundary conditions. This is indeed the only source of error coming from the localization method described above.
\bigskip
{\bf Three dimensional case.} Let us consider $N$ bosons in the three dimensional torus $\Lambda_L$
with density $\r$. Then the ground state energy of the system satisfies \cite[Thm. 2.1]{LSSY}
\begin{equation*}\label{eq:Eperiodic}
E_3(N,L) = 4 \pi \mathfrak{a} N \rho \big(1 + o(1) \big)
\end{equation*}
for $\r \mathfrak{a}^3\ll1$. On the other hand, the energy of $N$ bosons in a box of side length $L$ with Dirichlet boundary condition is given, at leading order, by
\begin{equation*}\label{eq:EDirichlet}
E_3^{(D)}(N,L) \simeq N \min_{\varphi \in \mathcal{C}^\infty_c (\Lambda_L)} \Big[ \int |\nabla\varphi|^2 + 4\pi \mathfrak{a} N \int |\varphi|^4\Big] \,.
\end{equation*}
We now require that
\begin{equation}\label{eq:difference}
\begin{split}
\frac{E_{3}^{(D)}(N,L)-E_3(N,L)}{N\rho} &\simeq L^3 \min_{\varphi \in \mathcal{C}^\infty_c (\Lambda_L)} \Big[ \frac 1N \int |\nabla\varphi|^2 + 4\pi \mathfrak{a} \int |\varphi|^4\Big] - 4\pi\mathfrak{a} \\[0.2cm]
&\ll C \mathfrak{a} (\rho \mathfrak{a}^3)^{1/2}\,.
\end{split}
\end{equation}
Even though we do not know the explicit expression of the minimizer on the first line of \eqref{eq:difference}, we know that it will be constant far from the boundary hence we can approximate it as
\begin{equation*}
\label{eq:phapprox}
\varphi(x) = \begin{cases} \;\tfrac1{(L- 2\d)^{3/2}}\,, \quad & d(x, \dpr \L_L) < \d\,, \\
\; 0\,, & \text{otherwise}\,.\end{cases}
\end{equation*}
Then we have
\[
\begin{split}
\int |\varphi|^4 &\sim \frac{1}{(L - 2\d)^6}(L-2\d)^3 = \frac{1}{(L - 2\d)^3}\sim \frac{1}{L^3} \Big(1+ o\big(\d / L \big) \Big)\\
\int |\nabla\varphi|^2 & \sim \frac{1}{\d^2(L-2\d)^3}[L^3 - (L-2\d)^3] \sim \frac{1}{\d L}\Big(1 + o\big(\d/L\big)\Big)\,.
\end{split}\]
Indeed, being $N =\rho L^3$ and substituting in the expression in Eq. \eqref{eq:difference} we conclude
\begin{equation} \label{eq:Dir-Per}
\begin{split}
\frac{E_3^{D}(N,L)}{N\rho} - \frac{E_3(N,L)}{N\rho}
\simeq \frac{1}{L} \Big(\mathfrak{a} \d + \frac{1}{\rho\d}\Big)\,.
\end{split}\end{equation}
Optimization over $\d$ leads to the choice $\d = (\rho \mathfrak{a})^{-1/2}$. Hence the condition in Eq. \eqref{eq:difference} is satisfied when
\begin{equation} \label{eq:localiz-UB}
L \gg (\rho \mathfrak{a})^{-1/2} (\r \mathfrak{a}^3)^{-1/2}\,.
\end{equation}
In particular, for $L_\a =(\rho \mathfrak{a})^{-1/2} (\r \mathfrak{a}^3)^{-\a}$ being defined as in \eqref{eq:L3d}, this implies $\a > 1$ (or equivalently to study Hamiltonian of the form \eqref{eq:HN-kappa}, for $\kappa >1/2$).
\medskip
{\it Remark.} Note that if one is interested in computing the leading order in the energy on the r.h.s. of \eqref{eq:LHY}, starting from a system were we imposed Dirichlet boundary conditions, the requirement $E_3^{(D)}(N,L) - E_3(N,L) \ll C N \r \mathfrak{a}$, together with \eqref{eq:Dir-Per}, leads to the condition $L \gg (\r \mathfrak{a})^{-1/2}$. This confirms, as mentioned in the introduction, that the energy is spoiled at leading order if one localizes particles at length scales smaller than the healing length.
\bigskip
To conclude, let us comment on the requirement $\kappa>1/2$ corresponding to the condition \eqref{eq:localiz-UB}. One of the reasons why providing second order upper bounds for the ground state energy of Hamiltonians $\eqref{eq:HN-kappa}$ becomes increasingly difficult for increasing values of $\kappa$ is that the number of excitations needed to obtain the correct energy is of the order $N^{3\kappa/2}$ (see \eqref{eq:depletion}). In particular it is independent of $N$ in the Gross-Pitaevskii regime, and then increases up to order $N$ (for $\kappa=2/3$). Moreover, the number of triples of excitations which have to be put on top of the quasi-free state and the condensate, to reproduce the Lee-Huang-Yang formula is of the order $N^{9\kappa/2 -2}$, see for example \cite[Prop. 2.2]{BCS}. While for $\kappa< 4/9$ this number goes to zero with $N$ (and one can control its effect by means of Gronwall type estimates, as done in the Gross-Pitaevskii regime, see \hbox{\it e.g.\ } \cite[Prop. 4.2]{BBCS4}), for $\kappa>4/9$ the control of cubic excitations requires substantially new ideas.
\bigskip
{\bf Two-dimensional case.} As before, we consider $N$ bosons on the two dimensional torus $\L_L$ and we compare their ground state energy, denoted by $E_2(N,L)$ with the ground state energy of the same system in a box of side length $L$ and Dirichlet boundary condition, below denoted by $E_2^{(D)}(N,L)$. Reasoning as above we require
\[ \begin{split}
\frac{E_2^{(D)}(N,L)}{N\rho} - \frac{E_2(N,L)}{N\rho} \simeq \; & L^2 \min \Big\{ \frac 1 N \int |\nabla \varphi|^2 + 4 \pi \mathfrak{b} \int |\varphi|^4 \Big\} - 4 \pi \mathfrak{b} \ll C \,\mathfrak{b}^2 \;,
\end{split}\]
being $N \r \mathfrak{b}^2$ the order at which we want to match \eqref{eq:LHY2d}. As in the $3d$ case we approximate the minimizer $\varphi$ by
\[
\varphi(x) = \begin{cases} \;1/(L- 2\d), \quad & d(x, \dpr \L_L) < \d\,, \\
\; 0\,, & \text{otherwise}\,.\end{cases} \,
\]
Proceeding as before we get
\begin{equation} \begin{split} \label{eq:difference2d}
\frac{E_2^{(D)}(N,L)}{N\rho} - \frac{E_2(N,L)}{N\rho} \sim\; \frac 1 L \Big( \mathfrak{b} \d + \frac 1 {\r \d} \Big) \sim\frac 1 L \sqrt{\mathfrak{b}/\rho} \ll \mathfrak{b}^2 \,,
\end{split}\end{equation}
where in the last step we optimized over the choice of $\d$ by setting $\d^2 = 1/(\mathfrak{b}\rho)$. Then \eqref{eq:difference2d} is satisfied if
\[ L \gg (\rho \mathfrak{b}^3)^{-1/2} = (\r \mathfrak{b})^{-1/2} |\log(\r \mathfrak{a}^2)|\,.
\]
For $L_\a= (\r \mathfrak{b})^{-1/2} \mathfrak{b}^{-\a}$ as in \eqref{eq:L2d}, this implies $\a> 1$, or equivalently to study Hamiltonians of the form \eqref{eq:HN-kappa2d} for $\kappa>2/3$. Note that, analogously to the three dimensional case, to recover the correct leading order term in \eqref{eq:LHY2d} from $E_2^{(D)}(N,L)$, one has to choose $L$ larger than the healing length, which is $(\r \mathfrak{b})^{-1/2}$ in two dimensions.
\medskip
Note that in the regimes described by \eqref{eq:HN-kappa2d} the number of excitations is of the order $N \mathfrak{b} \sim N^{\kappa} \ll N$ for all $\kappa\in [0,1)$. On the other hand, the number of triples needed to recover \eqref{eq:LHY2d} (thinking to a trial state similar to the one in \cite{BCS}) is $N^{-2+2\kappa}$ which is of smaller order with respect to $N$ for all $\kappa \in [0,1)$. This is a key difference with respect to the three dimensional case, where the number of triples increases with $N$ if $\kappa \in (4/9, 2/3)$, and one needs $\kappa>1/2$ to obtain a second order upper bound in the thermodynamic limit.
A similar difference between the $3d$ and $2d$ cases also shows up if one considers trial states of the form used in \cite{BCOPS, FGJMO}, where the Jastrow factor \eqref{eq:jastrow} is multiplied by a quasi free state. Indeed, as discussed in \cite[Paragraph below (2.7)]{FGJMO}, in two dimensions the contribution coming from the Jastrow factor can be expanded for all $\kappa<1$. As a matter of fact, to match the correct energy, correlations described by the Jastrow factor have to be put on a scale $\ell$ so that $R \ll \ell \ll (\r \mathfrak{b})^{-1/2}$. Since for any $\kappa \in [0,1)$ we have that $R$ is exponentially small in $N$, while $(\r \mathfrak{b})^{-1/2} \sim N^\kappa$ there is plenty of space to choose $\ell$ small enough, and hence to make the expansion for the Jastrow factor possible. This is not the case in three dimensions, where correlations described by the Jastrow factor have to be put at a distance $\ell$ satisfying $N^{-1+\kappa} \ll \ell \ll N^{\kappa/2}$.
Indeed, proving a second order upper bound matching \eqref{eq:LHY} in the thermodynamic limit and valid for the hard sphere potential is a current challenge to be overcome.
\bigskip
{\it Acknowledgements.} We acknowledge the support of Istituto Nazionale di Alta Matematica ``F. Severi", through the Intensive Period ``INdAM Quantum Meetings (IQM22)". G.B. acknowledges support through the project ``Progetto Giovani GNFM 2020: Emergent Features in Quantum Bosonic Theories and Semiclassical Analysis". C.C. gratefully acknowledges the support from the European Research Council through the ERC-AdG CLaQS.
\input{references}
\end{document}
|
\section{Introduction}
In the past several years, with the rapid development of the deep artificial neural network, near-field speaker verification (SV) performance has significantly improved\cite{snyder2017deep,vpr2014}. However, for the far-field speaker verification task, the quality of the speech signals is affected by energy decaying, reverberation, and environmental noise. All these factors make the far-field speaker verification problem more challenging. With intelligent speech assistants being used in smartphones and smart-home devices, more and more attention has been paid to the performance of far-field speaker recognition. The Far-Field Speaker Verification Challenge 2020 \cite{FFSVC2020} (FFSVC2020), which focused on both far-field text-dependent and text-independent speaker verification problems under noisy conditions in real scenarios, was launched to derive the development of far-field speaker recognition. FFSVC2020 successfully encouraged more researchers to pay attention to the far-field SV task. The Far-Field Speaker Verification Challenge 2022 (FFSVC2022) \cite{FFSVC2022_Eval_Plan}, heading in the same direction, still focuses on far-field speaker verification. The FFSVC2022 has two tasks, namely Task 1 and Task2.
Task 1 is a fully supervised far-field speaker verification task that can only use the VoxCeleb dataset and FFSVC2020 dataset.
Task 2 focuses on cross-language self-supervised/semi-supervised learning, which is a novelty of this challenge.
In contrast to Task 1, the speaker label of the FFSVC2020 dataset cannot be used in Task 2. In addition to providing a supplementary set of the FFSVC2020 dataset, FFSVC2022 also provided a new far-field development and evaluation set collected in complex environments with multiple scenarios.
The main difficulty of this challenge lies in cross-language and cross-domain. In response to this, we propose a new staged transfer training strategy that can greatly improve the performance of the systems on FFSVC2022 Task 1. We only use the close-talking English-based VoxCeleb data in the pre-training stage and preserve the speaker weights that need to be learned in the fine-tuning stage. In the fine-tuning stage, we add the far-field Mandarin-based FFSVC2020 data. The traditional transfer learning method does not reserve additional speaker weights in the pre-training stage, and randomly initializes these speaker weights in the fine-tuning stage.
All of our systems are deep-learning-based. Two different networks are used as encoders, namely ResNet and RepVGG. Based on the two-dimensional (2D) convolution layer, the networks get state-of-the-art performance for near-field speaker recognition in the reverberant and noisy environment. In this challenge, ResNet-based and RepVGG-based networks are used as encoders to generate frame-level representation. Each encoder is followed by a pooling layer to aggregate the frame-level features across time to obtain utterance-level representation. Global statistic pooling structure \cite{snyder2017deep} and multi-query multi-head attention (MQMHA) pooling structure \cite{zhao2021speakin} are used in our work. We use AM-Softmax and AAM-Softmax loss functions to increase the distance of inter-speakers and decrease the distance of the intra-speakers. We introduce the Sub-Center method \cite{deng2020sub} to reduce the influence of possible noisy samples.
We use cosine similarity for scoring in both tasks. Adaptive score normalization (AS-Norm) \cite{asnorm} is used to increase the robustness against different channels and domains. For Task 2, we found that the Sub-Mean backend method drastically reduces the EER and minDCF. Finally, the scores from different single systems are weighted average fused. Our proposed system significantly outperforms the official baseline with 56.83\% and 27.33\% relative minDCF reductions on the Eval set of Task 1 and Task 2 respectively.
The paper is structured as follows: Section 2 describes the datasets used to train and test, feature extraction procedure, and augmentation methods. Section 3 introduces our network structure, including backbone, pooling structure, and loss function. Section 4 details the training methods of our systems. Section 5 focuses on the backend methods and fusion method. Experimental results are presented and analyzed in Section 6. Section 7 concludes this paper.
\section{Datasets}
\subsection{Training dataset}
For Task 1, VoxCeleb and FFSVC2020 datasets(train, dev, and supplementary set) were used to perform system development. VoxCeleb dataset contains 5994 speakers and the FFSVC2020 dataset has 155 speakers. We here adopted a 3-fold speed augmentation at first to generate extra twice speakers. Each speech segment in this dataset was perturbed by 0.9 and 1.1 factors based on the SoX speed function. Then we obtained total 18447 speakers, which is triple amount of the original speakers.
For Task 2, only VoxCeleb data with speed perturbation was used to perform system development. There are 17982 speakers in this dataset. FFSVC2020 dataset without speaker label information was used in the backend stage.
We applied the following techniques to augment each utterance:
\begin{itemize}
\item Reverberation: artificially reverberation using a convolution with simulated RIRs\cite{ko2017study} from the AIR dataset
\item Music: taking a music file (without vocals) randomly selected from MUSAN\cite{snyder2015musan}, trimmed or repeated as necessary to match duration, and added to the original signal (5-15dB SNR).
\item Noise: MUSAN noises were added at one-second intervals throughout the recording (0-15dB SNR).
\item Babble: MUSAN speech was added to the original signal (13-20dB SNR).
\end{itemize}
We extracted 81-dimensional log Mel filter bank with energy based on Kaldi. The window size is 25 ms, and the frame shift is 10 ms. 200 frames of each feature were extracted without extra Voice Activation Detection (VAD). All features were cepstral mean normalized (CMN) in our training modes.
\subsection{Development \& Evaluation dataset}
The development (Dev) data has the same data distribution as evaluation (Eval) data. All trial pairs of the development set and the evaluation set are single-channel speech segments. All enrollment utterances are close-talking speech segments recorded by telephone, while the test segments are close-talking or far-field audio recorded by tablet or telephone\cite{FFSVC2022_Eval_Plan}.
\section{Systems}
\subsection{ResNet}
As one of the most classical ConvNets, ResNet\cite{he2016deep} has proved its power in speaker verification. In our systems, bottleneck-block-based ResNet (deeper structures:ResNet-74, ResNet-101, ResNet-152) are adopted. Base channels of all these ResNets are 64. We also implemented a deep and thin ResNet-221 structure which used ResNet\_v2\cite{DBLP:journals/corr/HeZR016} BottleNeck with 32 base channels.
\subsubsection{ResNet-D}
ResNet-D\cite{DBLP:journals/corr/abs-1812-01187} is a modification of the ResNet architecture that utilizes an average pooling tweak for downsampling. The motivation is that in the unmodified ResNet, the 1 × 1 convolution for the downsampling block ignores 3/4 of input feature maps. Such modification will not lead to the omission of information.
\subsubsection{fwSE}
A frequency-wise Squeeze-Excitation(fwSE) block\cite{https://doi.org/10.48550/arxiv.2109.04070}, which injects global frequency information across all feature maps, is used in our system.
\subsection{RepVGG}
In our previous work, we have proved that the RepVGG, as one of the re-parameterized models, shows competitive performance in speaker recognition\cite{zhao2021speakin, DBLP:journals/corr/abs-2110-09720}. We select RepVGG-A1, RepVGG-A2, and RepVGG-B1 as our backbones in this challenge. All RepVGG models adopt 64 base channels.
\subsection{Pooling Method}
The pooling layer aims to aggregate the variable sequence to an utterance level embedding. In addition to the global statistics pooling layer (GSP), we also used the multi-query multi-head attention pooling mechanism layer (MQMHA). RepVGG-A1 and RepVGG-A2 are followed by GSP, and other backbones are followed by MQMHA.
\subsection{Loss Function}
Recently, margin-based softmax methods have been widely used in speaker recognition works. To make a much better performance, we strengthen the AM-Softmax\cite{wang2018additive, wang2018cosface} and AAM-Softmax\cite{deng2019arcface} loss functions by the Sub-Center method.
The Sub-Center method \cite{deng2020sub} was introduced to reduce the influence of possible noisy samples. The formulation is given by:
\begin{equation}
cos(\theta_{i,j})=\max_{1\leq k\leq K}(||x_i||\cdot||W_{j,k}||)
\label{eq3}
\end{equation}
where the $\max$ function means that the nearest center is selected and it inhibits possible noisy samples interfering the dominant class center. $K$ means the number of sub-centers for each speaker class, and $k$ is the index of the sub-center.
\section{Training Protocol}
We used Pytorch\cite{paszke2019pytorch} to conduct our experiments. For Task 1, some of our models were trained through two stages, and the others were trained through an additional stage. And on Task 2, all of our systems were trained through the first stage.
\subsection{Stage 1: Pre-Training }
In the first stage, we used all VoxCeleb data with speed perturbation, consisting of 17982 speakers. It should be noted that we do not use any FFSVC2020 data at this stage. The number of classes is set to 17982 or 18447, depending on whether the weights of speakers from the FFSVC2020 dataset are reserved. The SGD optimizer with a momentum of 0.9 and weight decay of 1e-3 was used. We used 8 GPUs with 128 mini-batch and an initial learning rate of 0.08 to train all of our models. 200 frames of each sample in one batch were adopted. We adopted the ReduceLROnPlateau scheduler with a frequency of validating every 2,000 iterations, and the patience is 2. The minimum learning rate is 1.0e-6, and the decay factor is 0.1. All the models were trained with AM-Softmax in the first stage. Furthermore, the margin gradually increases from 0 to 0.2 \cite{liu2019large}.
We propose a novel training method, which preserves the speaker weights in the pre-training stage that need to be learned for the fine-tuning stage, even though there are no positive samples to train these preserved weights in the pre-training phase.
For example, the number of speakers in VoxCeleb dataset with speed perturbation is 17982, and we set the number of classes in loss function to 18447. In the pre-training stage, there are no positive samples but only negative samples to train these preserved weights of 465 classes. The preserved 465 classes correspond to 465 speakers in the FFSVC2020 dataset with speed perturbation, respectively.
When the number of classes is set to 17982, the weights of speakers of the FFSVC2020 dataset will be generated by random initialization in the second-stage training. However, preserving the speaker weights of the FFSVC2020 dataset in stage 1, we could use the trained weights to initialize in the fine-tuning stage. The experimental results confirm that this training method greatly improves the performance of the model, compared with the method of randomly initializing the weights of the last layer in traditional transfer learning. For specific experimental results, we can refer to the results of whether fwSE-ResNet34-D reserves weights in Table 1.
\subsection{Stage 2: Fine-Tuning }
We used VoxCeleb dataset without speed perturbation and the FFSVC2020 dataset with speed perturbation to fine-tune all systems in the second stage. The training dataset consists of 6459 speakers. All the configurations and hyper-parameters were the same as in the first stage except the initial learning rate, which was 2e-5. We removed the speed augmented part from the VoxCeleb dataset, and discarded the corresponding weights at the same time. For this reason, the number of classes is changed to 6459.
\subsection{Stage 3: Large-Margin Fine-Tuning }
Large-Margin Fine-Tuning (LM-FT) \cite{thienpondt2020idlab} helps to further improve model performance for some of our models. We chose the second-stage model to fine-tune for an additional epoch. In the LM-FT stage, settings are slightly different from the second stage. Firstly, we only used the FFSVC2020 dataset set as the training data, removing the speed augmented part from the training set to avoid domain mismatch. Secondly, we changed the chunk size from 200 to 400 and increased the margin exponentially from 0.2 to 0.5. The AM-Softmax loss was replaced by AAM-Softmax loss. We found that the large-margin-based fine-tuning in the third stage is not stable. For some large models, extra large-margin-based fine-tuning after the second stage may make the model performance worse. As a result, we only do the third stage fine-tuning on some models, such as RepVGG-A1 and RepVGG-A2.
\section{Backend}
We used cosine distance for scoring in both Task 1 and Task 2. In addition, adaptive symmetric score normalization (AS-norm) \cite{asnorm} was used for Task 1, and Sub-Mean was used for Task 2.
\subsection{AS-Norm}
For Task 1, AS-Norm was used for all of the models. For AS-Norm, we selected the original VoxCeleb and FFSVC2020 dataset without any augmentation. The cohort was created by using the speaker's random one utterance embedding vector as a speaker center and consisted of 6149 speaker centers. Only part of the cohorts are selected to compute mean and standard deviation for normalization, and top-300 highest scores are selected for Task 1.
\subsection{Sub-Mean }
Sub-Mean was used for models trained for Task 2. We randomly chose 40000 utterances from the FFSVC2022 dataset, then extracted the embedding vectors to compute the global mean embedding vector. The enrollment and test embedding vectors both subtract the mean embedding vector before scoring:
\begin{equation}
s(\bm{x_{e}} ,\bm{x_{t}}) = cos(\bm{x_{e}}-\bm{\bar{x}} ,\bm{x_{t}}-\bm{\bar{x}})
\end{equation}
where $\bm{x_{e}}$, $\bm{x_{t}}$ are enrollment and test speaker embedding vectors respectively, and $\bm{\bar{x}}$ is the mean embedding vector of 40000 utterances randomly chosen from the FFSVC2022 dataset.
\subsection{Fusion }
The results of all systems were fused using Logistic Regression on FFSVC2022 Dev set. We got the weight of each system and then selected the dominant systems to assign weights artificially. In the end, fusion was performed by computing the weighted average of the scores of selected individual systems.
\section{Results}
\begin{table*}[t]
\centering
\caption{Performance on FFSVC2022 Dev\&Eval set in Task 1. All systems used the AS-Norm backend method. Reserved Weight is T means that the corresponding model reserved the speaker weights of the FFSVC2020 dataset in the first stage, and F means that the corresponding model does not. S7 and S9 systems were trained in the third stage, the others were only trained in the second stage.}
\label{tab:tabel_two}
\setlength{\tabcolsep}{7.5mm}{
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccccc}
\hline
System Index & System & Reserved Weight & \multicolumn{2}{c}{Dev} & \multicolumn{2}{c}{Eval} \\ \cline{4-7}
& & & EER(\%) & minDCF & EER(\%) & minDCF \\ \hline
S1 & fwSE-ResNet34-D & F & 3.4278 & 0.3830 & - & - \\
S2 & fwSE-ResNet34-D & T & 5.8806 & 0.5379 & - & - \\
S3 & ResNet74 & T & 3.3361 & 0.3621 & - & - \\
S4 & fwSE-ResNet101-D & T & 3.0694 & 0.3319 & - & - \\
S5 & ResNet152 & T & 2.6667 & 0.2939 & 3.1897 & 0.3108 \\
S6 & ResNet221\_v2(thin) & T & 2.9861 & 0.3237 & 3.3333 & 0.3307 \\
S7 & RepVGG-A1 & T & 3.8417 & 0.3910 & 4.1109 & 0.3933 \\
S8 & RepVGG-A2 & T & 3.4472 & 0.3554 & 3.7269 & 0.3699 \\
S9 & RepVGG-A2 & T & 3.5389 & 0.3536 & 3.7178 & 0.3651 \\
S10 & RepVGG-B1 & T & 3.2861 & 0.3247 & - & - \\ \hline
Fusion & & & & & & \\ \hline
S5,S6,S9 & & & \textbf{2.5000} & \textbf{0.2735} & \textbf{3.0049} & \textbf{0.2938} \\ \hline
\end{tabular}
}
}
\end{table*}
\begin{table*}[t]
\centering
\caption{Performance on FFSVC2022 Dev\&Eval set in Task 2. All models were trained only using the VoxCeleb dataset in stage 1. S2-S8 systems used the Sub-Mean backend method while the S1 system did not.}
\label{tab:tabel_two}
\setlength{\tabcolsep}{7.5mm}{
\resizebox{\textwidth}{!}{
\begin{tabular}{ccccccc}
\hline
System Index & System & Sub-Mean & \multicolumn{2}{c}{Dev} & \multicolumn{2}{c}{Eval} \\ \cline{4-7}
& & & EER(\%) & minDCF & EER(\%) & minDCF \\ \hline
S1 & fwSE-ResNet34-D & F & 8.6861 & 0.6901 & - & - \\
S2 & fwSE-ResNet34-D & T & 7.2444 & 0.5650 & - & - \\
S3 & fwSE-ResNet101-D & T & 7.1472 & 0.5617 & - & - \\
S4 & ResNet152 & T & 6.4639 & 0.5132 & 6.6924 & 0.5374 \\
S5 & ResNet221\_v2(thin) & T & 6.4417 & 0.5399 & - & - \\
S6 & RepVGG-A1 & T & 7.4306 & 0.5929 & - & - \\
S7 & RepVGG-A2 & T & 6.9583 & 0.5600 & - & - \\
S8 & RepVGG-B1 & T & 7.0861 & 0.5556 & - & - \\ \hline
Fusion & & & & & & \\ \hline
S4,S5 & & & \textbf{5.9833} & \textbf{0.5004} & \textbf{6.2060} & \textbf{0.5232} \\ \hline
\end{tabular}
}
}
\end{table*}
Results of experiments on all our systems developed for the Task 1 and Task 2 are displayed in Table 1 and Table 2 respectively. The performance is measured on the FFSVC2022 development and evaluation set in terms of Equal Error Rate (EER) and Minimum Detection Cost(minDCF) with a prior target probability, $P_{tar}$ of 0.01. All systems in Table 1 are the results of models trained in stage 2 or stage 3 for Task 1, while all systems in Table 2 are the results of models trained in stage 1 only using the VoxCeleb dataset for Task 2.
All systems in Table 1 are the results of the cosine score after AS-Norm calibration for Task 1. Among all the single systems, ResNet152 gets the best performance by both EER and minDCF, which has a 6.6924\% EER and 0.5374 minDCF after AS-Norm. The final fusion result used three single systems and get a 3.0049\% EER and 0.2938 minDCF.
The S1 system that reserves the speaker classification weights of the FFSVC2020 dataset in the first training stage gets a much lower minDCF value than the S2 system that does not. According to the results of the S1 system and S2 system, the reserving weight training strategy results in a relative minDCF reduction of 28.8\%. The other systems (S3-S10) all reserve weights in the first stage.
The S8 system was trained in the second stage, and the S9 system was trained in the third stage. Comparing the S8 system and S9 system, we conclude that an additional Large-Margin Fine-Tuning can lower the minDCF slightly. However, we found that for large models (such as ResNet152 and RepVGG-B1) an extra LM-FT probably degraded model performance. Therefore we did not perform LM-FT on all individual systems.
For Task 2, all the models were trained in stage 1. S2-S8 systems used the Sub-Mean method described in Section 5.2 while the S1 system did not. Comparing the S1 system and S2 system, the Sub-Mean method brings 18.13\% relative minDCF reductions. Same as Task 1, the best individual system is ResNet152, which has a 6.6924\% EER and 0.5374 minDCF. The EER and minDCF of the fused result are 6.2060\% and 0.5232 respectively. In our experiments, we found that AS-Norm calibration after Sub-Mean does not further reduce the minDCF value. As a result, we only use the Sub-Mean backend method for Task 2. Finally, our system won the first place in both challenge tasks.
\section{Conclusions}
We experimented with multiple models on the SV, and ResNet achieved the best results on both tasks. Data augmentation, hyper-parameter changes in the fine-tuning stage, and score normalization in the backend have all brought improvement. Experimental results show that the larger model outperforms the small model on both tasks of the FFSVC2022.
Training in stages with using different train dataset and strategies in different stages can greatly improve model performance. The model trained using only VoxCeleb data will perform slightly worse than the model trained using VoxCeleb and FFSVC2020 data in the first stage. But the former will perform far better than the latter after fine-tuning in the second stage.
From the experimental results, we conclude that reserving the weights of the speakers from the FFSVC2020 dataset with only using the VoxCeleb dataset to train in the first stage makes the model perform much better after fine-tuning in the second stage. We speculated that the randomly initialized weights of speakers of the FFSVC2020 dataset are hard to converge possibly due to the small learning rate in the second stage.
We found that an additional large-margin-based fine-tuning after the second-stage fine-tuning can further improve the model performance, such as RepVGG-A1 and RepVGG-A2. However, our experimental results show that this method is unstable. For some large models, such as RepVGG-B1 and ResNet152, the refinement may make the model performance worse.
For Task 2, the speaker labels of the far-field FFSVC2020 dataset cannot be used in training. We only use the close-talking VoxCeleb dataset to train our models. Instead, the development set and evaluation set are far-field audio datasets. Therefore, there is a domain mismatch between training data and evaluation data. We focused on the solution of domain mismatch and used the Sub-Mean to solve the problem. Our experimental results show that Sub-Mean is pretty useful for domain mismatch.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The weakly interacting massive particles (WIMPs) as dark matter (DM) candidates come in various types of interactions with the normal matter \cite{STEIGMAN1985375,Arcadi:2017kky,Bergstrom:2000pn}.
Recent improvements on the direct detection (DD) experiments have been able to exclude a large portion of the parameter space in minimal extensions of the Standard Model (SM).
The minimal extensions are used to be applied in both simplified DM models and in effective field theory (EFT) DM models.
The disadvantage of the latter approach is that the viable parameter space
is shrunk by the constraints form the applicability criterion of the EFT models \cite{Arcadi:2021mag}.
A generic question is that how small the WIMP mass might be when assuming a thermal mechanism for the production of dark matter.
We address this issue in the present work within a $U(1)_X$ vector DM model.
The model consists of a dark gauge boson acting as a DM candidate and a complex scalar which
is responsible for the spontaneous breaking of the gauge symmetry and
thereby giving mass to the dark matter candidate. DM particles can communicate with the normal matter via scalar-Higgs portal. The abelian vector DM model is studied earlier \cite{Lebedev:2011iq,Baek:2012se,Farzan:2012hh,Glaus:2019itb,Arcadi:2016qoz,Arcadi:2020jqf,Ghorbani:2021yiw,Duch:2015jta,YaserAyazi:2022tbn, Foguel:2022unm}.
In all the previous works, points in the parameter space with quite small singlet scalar mass and at the same time very small mixing angle are not considered.
Moreover, we study the parameter space respecting the observed relic density and
upper bounds on the annihilation cross section from Cosmic Microwave Background (CMB).
To address this issue, one has to consider the forbidden and near pole annihilation channels carefully.
There are also various constraints from collider searches, beam-dump experiments, and
astrophysical observations, which are incorporated in this work.
The paper has the following structure. The DM model is described in the next section along
with theoretical bounds on the couplings from vacuum stability.
Focusing on the low mass DM we present the relevant Feynman diagrams for the DM annihilation and study the behavior of the DM relic density in section \ref{Rdensity} making use of MicrOMEGAs code. Bounds from CMB and observed relic density on the DM annihilation are studied simultaneously in \ref{CMBbounds}. The elastic scattering of DM off the proton is obtained in section \ref{DDB} and the viable parameter is found after imposing all the constraints.
We summarize all the relevant experimental and astrophysical bounds in
section \ref{Experimental-bounds}. In \ref{Secresults} our results are discussed.
The conclusion is provided by section \ref{con}. In the Appendix we provide the DM annihilation cross sections.
\section{Model}
\label{model}
We consider a model as a minimal extension to the SM with a new $U(1)_X$ gauge symmetry introducing a complex scalar field, $S$, in the following Lagrangian
\begin{equation}
{\cal L}_{\text{DM}} = ({\cal D}_\mu S)({\cal D}^\mu S)^*
- \frac{1}{4} X^{\mu\nu} X_{\mu\nu} - {\cal V } (S,H) \,,
\end{equation}
where the covariant derivative is ${\cal D}_\mu = \partial_\mu - i \alpha_X q_X X_\mu$, in which $\alpha_X$ is the gauge coupling and $q_X$ is the $U(1)_X$ charge of the corresponding field which is taken unity. The strength field tensor for the gauge field $X_\mu$ is given by $X_{\mu \nu} = \partial_{\mu} X_{\nu} -\partial_{\nu} X_{\mu}$.
The scalar potential entailing the Higgs portal is
\begin{equation}
{\cal V} (S, H) = \mu_H |H|^2 + \lambda_H |H|^4 + m^2 |S|^2 + \lambda_S |S|^4 + \lambda_{HS} |S|^2 |H|^2 \,.
\end{equation}
The Lagrangian enjoys an additional {\it charge conjugation} symmetry
\begin{equation}
X^\mu \to -X^{\mu}, ~~~~ S \to S^* \,.
\end{equation}
When the complex scalar gets a nonzero expectation value this breaks the gauge symmetry spontaneously, after which the remaining residual symmetry for the vector gauge boson
is $X^\mu \to -X^{\mu}$. The parametrization of the complex scalar field around its vacuum is
\begin{equation}
S = \left( \frac{v_s + \phi_1}{2}\right) e^{-i \phi_2 /v_{s}} \,,
\end{equation}
here $\phi_1$ and $\phi_2$ are real scalar fields. It is possible to gauge away
the Goldstone field $\phi_2$ and rendering the massless gauge boson massive, such that $m_X = v_{s} \alpha_X$. On the other hand, the Higgs doublet takes a vacuum expectation value, $v_h$, and in unitary gauge we write down the Higgs field as $H = (0~~v_h + h_1)^T/ \sqrt{2}$. Plugging the Higgs doublet and
singlet scalar in the Lagrangian and after some simplification we arrive at the following interaction terms,
\begin{equation}
\begin{split}
{\cal L}_{int} = & \alpha^2_X v_s X_\mu X^\mu \phi_1
+\frac{1}{2} \alpha^2_X X_\mu X^\mu \phi_1^2
- \frac{1}{4} \lambda_S \phi_1^4 - \lambda_S v_s \phi_1^3
- \frac{1}{4} \lambda_H h_1^4 - \lambda_H v_h h_1^3 \\
&\quad
- \frac{1}{2} \lambda_{HS} v_h h_1 \phi_1^2
- \frac{1}{2} \lambda_{HS} v_s h_1^2 \phi_1
- \frac{1}{4} \lambda_{HS} h_1^2 \phi_1^2 \,.
\end{split}
\end{equation}
The scalar mass matrix is also obtained as
\begin{equation}
{\cal M}^2 = \Big( \begin{matrix}
2\lambda_S v_s^2 & \lambda_{HS} v_h v_s \\
\lambda_{HS} v_h v_s & 2 \lambda_H v_h^2
\end{matrix} \Big) \,.
\end{equation}
A scalar field transformation is needed in order to diagonalize the mass matrix.
Therefore, the physical or mass eigenstates, $h$ are $\phi$, are defined in terms of
the mixing angle $\epsilon$,
\begin{equation}
h = h_1~\cos(\epsilon) + \phi_1~ \sin(\epsilon), ~~~~ \phi = -h_1 \sin(\epsilon) + \phi_1 \cos(\epsilon) \,,
\end{equation}
where the mixing angle is given by the relation,
\begin{equation}
\tan(2\epsilon) = \frac{2v_h v_s}{m_h^2-m_{\phi}^2 } \lambda_{HS} \,,
\end{equation}
with $m_h$ and $m_\phi$ the physical masses of the SM Higgs and the singlet scalar, respectively.
The quartic couplings are obtained in terms of the mixing angle and the physical scalar masses
\begin{equation}
\begin{split}
\lambda_H =& \frac{1}{2v_h^2} (m_h^2 \cos^2 \epsilon + m_{\phi}^2 \sin^2 \epsilon), \\
\lambda_S =& \frac{1}{2 v_s^2} (m_h^2 \sin^2 \epsilon + m_{\phi}^2 \cos^2 \epsilon), \\
\lambda_{HS} =& \frac{\sin 2\epsilon}{4v_h v_s} (m_{\phi}^2- m_h^2) \,.
\end{split}
\end{equation}
We will impose the conditions by which the Higgs potential remains bounded from below. These conditions read
\begin{equation}
\label{BFM}
\lambda_H > 0,~~ \lambda_S > 0,~~ \lambda_{HS} > -2 \sqrt{\lambda_H \lambda_S} \,.
\end{equation}
Also, satisfying the relations in Eq.~\ref{BFM} warrants the electroweak vacuum to be a global minimum \cite{Baek:2012se}.
The free parameters in this model are $m_X, m_\phi, \epsilon,$ and $\alpha_X$.
\section{Relic Density from Thermal Freeze-out}
\label{Rdensity}
Assuming DM is in thermal equilibrium in the early times, dark matter production
is then happening at temperatures very smaller than
the DM mass when it goes out of equilibrium, the so-called freeze-out mechanism of dark matter particles.
The freeze-out temperature, $T_f$, depends on the DM annihilation cross section and DM mass.
The Boltzmann equation describes how the DM number density, $n_X$, evolves in an
expanding Universe,
\begin{equation}
\dot n_{X} +3 {\cal H} n_{X} = - \langle \sigma_{\text{ann}}v_{\text{rel}} \rangle [n^{2}_{X}-(n^{\text{eq}}_{X})^2 ] \,,
\end{equation}
where ${\cal H}$ is the Hubble parameter, $n^{\text{eq}}_{X} \sim (m_X T)^{3/2} e^{-m_X/T}$ is the particle density before particles get out of equilibrium, and the
thermal average annihilation cross section times the relative velocity is given by
\begin{equation}
\langle \sigma_{\text{ann}} v_{\text{rel}} \rangle = \frac{1}{8 m_{X}^4TK^{2}_{2}(\frac{m_{X}}{T})}
\int^{\infty}_{4m^{2}_{X}} ds~(s-4m^{2}_{X})\sqrt{s}~K_{1} \left(\frac{\sqrt{s}}{T} \right)~\sigma_{\text{ann}}(s)\,.
\end{equation}
In the present work DM with mass smaller than $\sim 10$ GeV annihilates through three different ways diagrammatically as depicted in Fig.~\ref{FeynAnni}:
1) A $s$-channel annihilation process into singlet scalars and the SM fermions
as $XX \to \phi \phi, f \bar f$ with mediators being the SM Higgs or the singlet scalar.
2) A contact interaction of DM with the singlet scalars.
3) DM annihilation to singlet scalars via $t$- and $u$-channels, with the vector DM acting as the mediator.
\begin{figure}
\begin{center}
\includegraphics[width=.7\textwidth,angle =0]{feyn-anni-eps-converted-to.pdf}
\caption{The relevant Feynman diagrams for the vector dark matter annihilation with mass smaller
than $\sim 10$ GeV.
The diagram in the left is a $s$-channel annihilation process and the process shown in the right side takes place in both $t$- and $u$-channel.}
\label{FeynAnni}
\end{center}
\end{figure}
It is worth mentioning that the lower limit for a thermal WIMP with $2 \to 2$ s-wave annihilation to visible final states is found to be about 20 GeV \cite{Leane:2018kjk}.
However, since in the present work we consider the forbidden channels and the resonance effect this lower bound is relaxed.
This model is first implemented in the package LanHEP \cite{Semenov:2008jy}
in order to obtain
the possible vertices and then the code micrOMEGAs \cite{Belanger:2013oya} is applied to compute the relic density numerically. First, it is interesting to
see the variation of the relic density as a function of DM mass for three different singlet scalar
masses $m_\phi = 1, 2, 5$ GeV. To do that we fix the gauge coupling at a reasonable
value $\alpha_X = 0.05$. The results are shown for two distinct mixing angles
$\sin \epsilon = 0.001$ and $\sin \epsilon = 0.01$ in Fig.~\ref{relicdensity}.
It is evident from the two plots with different mixing angles that the relic density
behaves almost similarly near the observed relic density, although the bigger $\epsilon$ the bigger is the annihilation cross section, and in turn lower relic abundance.
\begin{figure}
\begin{minipage}{0.5\textwidth}
\includegraphics[width=\textwidth,angle =0]{relic-low-a-eps-converted-to.pdf}
\end{minipage}
\begin{minipage}{0.5\textwidth}
\includegraphics[width=\textwidth,angle =0]{relic-low-b-eps-converted-to.pdf}
\end{minipage}
\caption{Shown are relic density as a function of DM mass for the gauge coupling $\alpha_X = 0.05$.
In the left panel we take the mixing angle such that $\sin \epsilon = 0.001$
and in the right panel we fix it as $\sin \epsilon = 0.01$. The observed DM relic density is shown as a horizontal line at $\Omega h^2 \sim 0.12$.}
\label{relicdensity}
\end{figure}
It is also apparent from the figures that the relic density is
decreasing when the DM mass approaches the singlet scalar mass.
The reason for this behavior relies on the fact that when $m_\text{DM} \sim m_\phi$, the
annihilation channel $XX \to \phi \phi$ opens up and its cross section increases leading to
a reduction of the relic density. When $m_\text{DM} < m_\phi$ the annihilation
channel $XX \to f {\bar f}$ is open, and in case $m_\text{DM}$ is just slightly smaller than $m_\phi$ then the forbidden channel is also open \cite{DAgnolo:2015ujb}.
In summary, we can conclude that in the low mass region and away from the resonance,
the computed relic density complies with the observed value when the singlet
scalar mass is $m_{\phi} \sim m_{\text{DM}}$.
The analytical formulas for the annihilation cross sections are available in the Appendix.
In the analysis in the next sections we constrain the model parameter space by imposing
the observed DM relic density, $\Omega_{\text{DM}} = 0.1198 \pm 0.0012$ \cite{Planck:2018vyg,Hinshaw:2012aka}.
\section{CMB Bounds}
\label{CMBbounds}
Light dark matter annihilation being a $s$-wave process in our scenario, can modify the recombination history by injecting energy into the universe. The anisotropies of the CMB is measured very accurately by Planck \cite{Planck:2018vyg}, and this will put strong indirect constraint on the DM annihilation. The upper limit on the DM annihilation at the recombination is given by
\begin{equation}\label{cmb}
\frac{f_{\text{eff}}}{m_{\text{DM}}} \langle \sigma v \rangle_{\text{rec}} < 3.2 \times 10^{-28} \text{cm}^3 \text{s}^{-1} \text{GeV}^{-1} \,,
\end{equation}
where $f_{\text{eff}}$ is an efficiency factor depending on the energy of
injected electrons and photons. The efficiency factor for models with s-wave annihilation is computed in \cite{Slatyer:2015jla} as a function of DM mass, and
can be applied to any DM model by weighting their results considering the annihilation product. In our analysis for DM mass below 10 GeV,
we take $f_{\text{eff}} = 0.1$.
\begin{figure}
\includegraphics[width=\textwidth,angle =0]{reospnance_0005.png}
\caption{In the {\it left plot} we show the relic density as a function of the singlet scalar, and in the {\it right plot} the $s$-wave annihilation cross section times the velocity as a function of the singlet scalar mass. In both plots we set the DM mass at 0.2 GeV and $\tan \epsilon = 0.005$, with $\alpha_X = 0.1, 0.01$ shown in red and blue, respectively. The observed DM relic density is shown as a horizontal green line at $\Omega h^2 \sim 0.12$ in the left plot, and in the plot in the right we show the CMB bounds as the yellow colored region. The vertical lines in each plot are explained in the text.}
\label{reospnance}
\end{figure}
At first sight, it might look impossible to reconcile the thermal relic
density of about $3\times 10^{-26} \text{cm}^3 \text{s}^{-1}$ to the CMB limits
for DM particles with $s$-wave annihilation and masses below 10 GeV.
As discussed in \cite{Griest:1990kh, DAgnolo:2015ujb}, there are kinematical regions in
the annihilation processes where the $s$-wave annihilation
is not dominating. This happens in two ways.
When Boltzmann distributed DM is slightly lighter than
the mediator mass, it can dominantly annihilates to the heavier states, the so-called
forbidden channels. Another avenue in which higher partial waves dominate is DM annihilation near pole. We proceed by computing the annihilation cross section times the velocity at the recombination era, $\langle \sigma v \rangle_{s-wave}$.
As shown in Fig.~\ref{reospnance} for some benchmark points, it is possible to respect the observed relic density and CMB bounds via forbidden channel as $m_\phi/m_{\text{DM}} = 1.20$, and near the pole as $m_\phi/m_{\text{DM}} = 2.04$, for $\tan \epsilon = 0.005$, $m_{\text{DM}} = 0.2$ GeV and $\alpha_X = 0.1$ (we show the blue curve as another point just for comparison, in which it does not fulfill the relic abundance constraint, but the CMB bounds near the pole does). In particular, in the plot in the left shows the red curve intersecting the green horizontal line (correct relic abundance), and both intersection points in the plot in the right are well below the yellow region, which represents the CMB bounds imposed by Eq.~\ref{cmb}.
\section{Direct Detection Bounds}
\label{DDB}
We consider direct detection bounds on the elastic scattering of DM off the proton from three experiments, Xenon1T \cite{XENON:2019gfn}, DarkSide50-S2 \cite{DarkSide:2018ppu} and CRESST-III \cite{CRESST:2019jnq}.
In the present model the spin-independent (SI) elastic scattering of DM off the normal matter is a $t$-channel process with the SM Higgs or the singlet scalar as the mediator between the DM and the nucleon.
The final formula for the cross sections reads
\begin{equation}
\sigma^{\text{p}}_{\text{SI}} = \frac{\mu_{Xp}^2 m^2_{p}}{4\pi v_h^2} \alpha_X^2 \sin^2 (2\epsilon) \left(\frac{1}{m_h^2}-\frac{1}{m_\phi^2}\right)^2 f_p^2 \,,
\end{equation}
where $m_p$ is the proton mass, the reduced mass of the proton and DM
is $\mu_{Xp} = m_X m_p /(m_X+m_p)$, and the scalar form factor for proton is given by
$f_p = 2/9 + 7/9 \sum_{q= u,d,s} f^{p}_{q}$, in which $f^{p}_{u} = 0.0153$
$f^{p}_{d} = 0.0191$ and $f^{p}_{s} = 0.0447$ \cite{Belanger:2013oya}.
We probe the parameter space while taking into account bounds from invisible Higgs decay and observed relic density. The ranges of the relevant parameters picked out in our scan
are the following: $0.1 < m_{\text{DM}} < 10~$GeV, $0.1 < m_{\phi} < 10~$GeV, $0.001 < \alpha_X < 1$,
and $ 0 < \sin \epsilon < 0.1$.
The model points in the plane $\sigma^{\text{p}}_{\text{SI}}-m_{\text{DM}}$ of Fig.~\ref{DDbounds} indicate that there are DM masses from 0.1 GeV up
to 10 GeV in the viable parameter space. The allowed values for the mixing angle in
the respected regions are smaller that $\sim 10^{-4}$ as shown in the right panel of Fig.~\ref{DDbounds}. In the final section we project these same constraints once again.
\begin{figure}
\hspace{-.45cm}
\begin{minipage}{0.37\textwidth}
\includegraphics[width=\textwidth,angle =-90]{direct-0.1-10GeV-eps-converted-to.pdf}
\end{minipage}
\hspace{2.2cm}
\begin{minipage}{0.37\textwidth}
\includegraphics[width=\textwidth,angle =-90]{direct-0.1-10GeV-sin-eps-converted-to.pdf}
\end{minipage}
\caption{Viable regions are shown after imposing the observed relic density and invisible Higgs decays bounds. The vertical color spectrum shows the singlet scalar mass in the {\it left} panel and the mixing angle in the {\it right} panel. Direct detection bounds from Xenon1T, DarkSide50 and CRESST-III are applied.}
\label{DDbounds}
\end{figure}
\section{Collider and Beam-dump Bounds}
\label{Experimental-bounds}
In this section we provide bounds from various experiments when
dark particles are engaged in the decay or production at collider or beam-dump experiments.
\subsection{Invisible Higgs Decay}
\label{Inv}
In the mass range of interest in this work there are two channels for the SM Higgs invisible decay: $\text{Higgs} \to \phi \phi$ and $\text{Higgs} \to X X$.
The invisible Higgs decay width for decay a pair of dark matter is
\begin{equation}
\Gamma_{h \to XX} = \frac{\alpha_X^2 m_h^3 \sin^2 \epsilon}{16\pi m_X^2} \sqrt{1-4(m_X/m_h)^2} [1-4(m_X/m_h)^2+12(m_X/m_h)^4].
\end{equation}
The invisible Higgs decay width for the process $H \to \phi \phi$ is
\begin{equation}
\Gamma_{h \to \phi \phi} = \frac{c^2}{8\pi m_h} \sqrt{1-4(m_\phi/m_h)^2} \,,
\end{equation}
where the couplings $c$ is given by the relation below
\begin{equation}
\begin{split}
c =& \lambda_{HS} v_h \cos \epsilon
+ 2 \lambda_{HS} v_s \sin \epsilon
-3 v_s \lambda_{HS} \sin^3(\epsilon) + 6 \lambda_H v_h \sin^2(\epsilon) \cos(\epsilon) \\
&\quad -3 \lambda_{HS} v_h \sin^2(\epsilon) \cos(\epsilon)
-6 v_s \lambda_S \cos^2(\epsilon) \sin(\epsilon) \,.
\end{split}
\end{equation}
The measured Higgs mass is $m_h \sim 125.25$ and the total Higgs decay width reads $\Gamma_{\text{Higgs}} = 3.2^{+2.8}_{-2.2}$ MeV \cite{ParticleDataGroup:2020ssz}.
The experimental upper bound on the branching ratio of the invisible Higgs decay
at $95\% $ confidence level is Br($h\to$ invisible) $\lesssim 0.19$
\cite{CMS:2018yfx}.
Thus, the following constraint is applied when the parameter space is scanned
\begin{equation}
\frac{\Gamma_{h \to XX} + \Gamma_{h \to \phi \phi}}{\Gamma_{\text{Higgs}}} \lesssim 0.19 \,.
\end{equation}
\subsection{Singlet Scalar Decay}
\label{Sdecay}
Since the singlet scalar mass varies in the range $(\sim 0.1-10)$ GeV, when $m_\phi > 2 m_f$ it can decay to the SM fermions by the following decay width
\begin{equation}
\Gamma (\phi \to f^+ f^- ) = \frac{N_c^2 m_\phi m_f^2 \sin^2(\epsilon)}{8\pi v_h^2}
\Big(1- \frac{4m_f^2}{m_\phi^2} \Big)^{3/2} \,,
\end{equation}
The total decay width is $\Gamma = \sum_{f} \Gamma (\phi \to f^+ f^- )$, and the corresponding life time of the scalar is $\tau = \Gamma^{-1}$.
Note that the decay $\phi \to XX$ is kinematically open when $m_\phi > 2m_X$.
The decay width is obtained as
\begin{equation}
\Gamma (\phi \to XX) = \frac{\alpha_X^2 m_\phi^3 \cos^2 \epsilon}{16\pi m_X^2} \sqrt{1-4(m_X/m_\phi)^2} [1-4(m_X/m_\phi)^2+12(m_X/m_\phi)^4] \,.
\end{equation}
The hadronic decay of the singlet scalar with final states, $\pi \pi$, $K K$,...,
is also important here and has to be taken into account.
The big bang nucleosynthesis (BBN) predicts the light element abundances in agreement with the observation. This puts upper limit on the singlet scalar life time which
varies between $\sim 0.01$ sec to 1 sec, in the scalar mass range $\sim 0.1-4$ GeV \cite{Fradette:2017sdd}. This upper bound on the scalar life time will be converted to lower bounds on the mixing angle.
\subsection{Mediator Production in $e^+ e^-$ and LHC Colliders}
Direct production of the dark mediator, $\phi$, mixing with the SM Higgs
is possible at colliders like LEP through the process $e^+ e^- \to Z \to \phi Z^*$ \cite{L3:1996ome}. Since the rate of the process is proportional to $\sin^2 \epsilon$, then strong constraint on the mixing angle, $\epsilon$, is expected.
However, for mediator mass smaller than 10 GeV, it is found that regions
with $\sin \epsilon \gtrsim 0.1$ are excluded \cite{Winkler:2018qyg}.
On the other hand, CMS and LHCb in search for a new scalar mediator
put constraints on $\sigma_{pp-\phi} \times \text{Br}(\phi \to \mu^+ \mu^-)$
at $\sqrt{s} = 7$ and 8 TeV, respectively \cite{CMS:2012fgd,LHCb:2018cjc}.
It turns out that the LHC bounds are stronger than that of LEP for scalar mass smaller than about 5 GeV.
\subsection{Rare Decays}
The relevant rare decays are $\Upsilon$ decay and flavor changing decay of $B$ and $K$ mesons. In these processes the singlet scalar can act as a mediator.
The most important decay channel of $\Upsilon$ studied by BaBar is $\Upsilon \to \gamma \phi^*$, where $\phi$ decays hadronically to jets \cite{BaBar:2011kau}.
LHCb provides us constraints on two $B$ meson decays, $B^0 \to K^{*0} \mu^+ \mu^-$ and $B^+ \to K^+ + \mu^+ \mu^-$, where in both decays the scalar mediator $\phi$ is decaying to a dilepton \cite{LHCb:2015nkv,LHCb:2016awg}.
When the scalar mass is below a few hundred of MeV, rare kaon decays become important. Constraints on the mixing angle in terms of the scalar mass can be obtained from the decay $K_L \to \pi^0 \mu^+ \mu^-$ performed by KTeV experiment \cite{KTEV:2000ngj} and the search for the decay $K^+ \to \pi^+ \bar \nu \nu$ by E949 experiment \cite{BNL-E949:2009dza}.
\subsection{Beam Dump Experiments}
We apply bounds from three beam dump experiments at CERN SPS using a 400 GeV proton beam, see \cite{Winkler:2018qyg} and references therein.
1) CHARAM experiments being sensitive to leptonic final states was
operating in 1980s. 2) NA62 in dump mode
which is an upcoming run and as CHARAM is sensitive to higher scalar mass. 3) SHiP is a planned experiment which will cover a large region
of the parameter space not accessible in all earlier experiments.
\begin{comment}
\section{Self-scattering of Dark Matter}
\label{SS}
One of the strongest constraints on the DM self-interaction is the observation from the Bullet Cluster giving rise to $\sigma/m_{\text{DM}} \lesssim 1.3~\text{cm}^2/\text{g}$ ($5.95\times10^3~\text{GeV}^{-3}$) \cite{Randall:2008ppe,Robertson:2016xjh}.
In the limit where the mixing angle is quite small such that $\sin \epsilon \ll 1$ and $\cos \epsilon \sim 1$, we obtain for the self-scattering process \footnote{The code CalcHEP \cite{Belyaev:2012qa} is exploited to compute the amplitude.},$X(p) X(p') \to X(k) X(k')$,
\begin{equation}
\begin{split}
\overline {{\cal M}^2} \simeq& \frac{w^4 \alpha_{X}^8}{9} [ \frac{128}{(s-m_\phi^2)^2}
+ \frac{256}{(t-m_\phi^2)^2} + \frac{256}{(u-m_\phi^2)^2}
+ \frac{128}{(s-m_\phi^2)(t-m_\phi^2)} \\
&\quad
+ \frac{128}{(s-m_\phi^2)(u-m_\phi^2)}
+\frac{64}{(t-m_\phi^2)(u-m_\phi^2)}
] \,,
\end{split}
\end{equation}
where the Mandelstam variables are $s= (p+p')^2$, $t=(p-k)^2$ and $u=(p-k')^2$.
The scattering cross section is
\begin{equation}
\sigma_{XX \to XX} = \int \int \frac{1}{2E_p 2E_{p'} |v_p -v_{p'}|}
\frac{|\vec{k}|}{16\pi^2 E_{CM}} \overline {{\cal M}^2} ~ d\Omega \,,
\end{equation}
where, working in the center-of-mass (CM) frame, $E_p$ and $E_{p'}$ are the energies of the incoming dark matter particles, $v_p$ and $v_{p'}$ are the velocities of dark particles, $|\vec{k}|$ is the magnitude of the tree-momentum of an outgoing dark particle and ${\cal M}$ is the scattering amplitude.
In the CM frame, it follows $E_p = E_{p'} = E = E_{CM}/2$, and the Mandelstam variables
can be written in terms of the scattering angle, $\theta^*$, as $s = 4 E^2$, $t = -2(E^2-m_X^2) (1-\cos \theta^*)$ and
$u = -2(E^2-m_X^2) (1+\cos \theta^*)$. The DM momentum is $E^2-m_X^2 \simeq m_X v_0$ where $v_0 \sim {\cal O}(10^{-3})$ is the DM velocity.
Now, it is possible to obtain a simple expression for the cross section in the non-relativistic limit. In the region of parameter space we do the scan, it follows $m_X v_0 \ll m_\phi$ and
we take $E \simeq m_X$. The final result for the self-scattering cross section in the non-relativistic limit is
\begin{equation}
\sigma_{XX \to XX} \simeq \frac{w^4 \alpha_{X}^8}{9 \pi m_X^2}
\Big[ \frac{15 m_\phi^4 + 144 m_X^4 - 88 m_\phi^2 m_X^2}{m_\phi^4(4m_X^2-m_\phi^2)^2} \Big] \,.
\end{equation}
Our numerical exploration in the viable parameter space (regions respecting bounds from invisible Higgs decay, observed relic density and direct detection) indicates that the observed bound on the
DM self-scattering, $\sigma/m_{\text{DM}} \lesssim 1.3~\text{cm}^2/\text{g}$, is fully satisfied.
\end{comment}
\section{Results}
\label{Secresults}
For the resonance case, collider constraints constraints results to be too strong, in such a way that most of the region is excluded by them, even for the case in which the mixing angle $\epsilon$ is too small. This can be seen on the left plot of Fig.~\ref{results}, where we have overlaid the predictions of the model with the tick black curves for different mass shift of the singlet scalar and the DM mass. In particular, provided the mass shift is too small (solid black tick curve), the mixing angle tend to be reduced, but even in this case $B$ meson decay measurements at LHCb results to be strong. Masses for the new particles above 4 - 5 GeV are not constrained yet by the latter experiment, although future direct detection experiments such as SuperCDMS or DarkSide-50 will be able to test this region. The very low mass region, $\sim 0.1$ GeV is also viable in some discrete range of masses, and experiments such as SHiP or CHARM could explore these parameter space.
\begin{figure}
\includegraphics[width=0.57\textwidth]{full.png}
\includegraphics[width=0.43\textwidth]{2D_relic.png}
\caption{(left) Overlaid of predictions in the resonance region (black tick curves) with constraints by experiments and future projections. (right) Points fulfilling the correct relic abundance by the forbidden mechanism, for $\alpha_X$ values specified in the plot.}
\label{results}
\end{figure}
The forbidden DM relic abundance is dependent only on three parameters: $m_X, m_{\phi}$ and $\alpha_X$. In order to exemplify the relation of them fulfilling the observed relic abundance, in Fig.~\ref{results}(right) we show some results obtained with MicrOMEGAS in the mass plane for different values of $\alpha_X$. This behavior is understood by the interplay between the parameters such that the relic is keep constant \cite{DAgnolo:2015ujb}. Note that too small $\alpha_X$ requires a too high degeneracy between the new components of the new sector, and a very limited range of masses. These results are in agreement with the scalar version of this model \cite{Hara:2021lrj}. Since the relic abundance produced by the forbidden channel does not require any particular value of $\epsilon$, one can arbitrarily take too small values of it in such a way to avoid all the constraints and projections from above shown in Fig.~\ref{results}, then entering into the white region of Fig.~\ref{results}(left). However, two constraints from too small values of $\epsilon$ must be taken into account: BBN and the thermalization of both sectors.
\section{Conclusion}
\label{con}
In this work we reconsidered the Higgs portal $U(1)_X$ vector dark matter
model regarding the low mass of the DM candidate and the singlet scalar.
This region of the parameter space with $m_{DM}$ below 10 GeV
is not fully investigated in earlier works.
In this region of the parameter space we must consider DM annihilation in forbidden
channels and near the pole in order to satisfy the observed relic density and
CMB bounds. The invisible Higgs decay bounds becomes strong and to respect the
corresponding limits, quite small mixing angle is demanded. Other collider searches
in $B$ and $K$ meson are considered along with the limits from beam-dump experiments. Since the singlet scalar can decay to the SM leptons we ensured that
it will not spoil the BBN.
Our main results given in Fig.~\ref{results} concern the viable regions respecting all the theoretical and experimental constraints. We found that when the DM annihilations are near the pole resonance, various parameter space regions result to be excluded, and in the future direct detection constraints will be able to probe most of the region of DM masses $\mathcal{O}(1)$ GeV, whereas lower masses will be tested by experiments such as SHiP and CHARM. On the other hand, regarding the forbidden dark matter production, it shows that as it is independent on the mixing angle $\epsilon$, it is less constrained than the parameter space originated by the resonance production, therefore being more challenging to test regions with such small $\epsilon$.
\textbf{Note added}: while our work was being completed, the article \cite{Yang:2022zlh} was published on the
ArXiv, presenting some overlapping.
\section{Acknowledgments}
B.D.S has been founded by ANID (ex CONICYT) Grant No. 74200120.
\section{Appendix: DM Annihilation Cross Sections}
\label{Apen}
Dark matter annihilation cross sections times the relative velocity are provided in this section.
The first formula is the annihilation of DM to a fermion pair being a $s$-channel process:
\begin{equation}
\sigma_{\text{anni}} v_{rel} (X X \to f^+ f^-) = \frac{N_c m_f^2 w^2 \alpha_{X}^4 \sin^2(2\epsilon) }{36\pi^2 v_h^2} (1-4m_f^2/s)^{3/2}
\Big[\frac{1}{(s-m^2_\phi)^2}-\frac{1}{(s-m^2_h)^2} \Big]^2 \,,
\end{equation}
where $N_c$ is number of colors.
The next formula belongs to the DM annihilation to a pair of singlet scalars in $s$-, $t$-
and $u$-channel:
\begin{equation}
\begin{aligned}
\sigma_{\text{anni}} v_{rel} (X X \to s s) & = \frac{\sqrt{1-4m^2_s/s}}{32\pi^2s}
\int d\Omega \Big[ \frac{8}{9} \cos^4(\epsilon) \alpha_{X}^4 + \frac{64}{9} \cos^4(\epsilon) w^4 \alpha_{X}^8 \Big(\frac{1}{t-m_X^2} + \frac{1}{u-m_X^2}\Big)^2
\\ &
+ 8 w^2 \alpha_X^4 \Big(\frac{c_1 \cos(\epsilon)}{s-m_\phi^2} -
\frac{c_2 \sin(\epsilon)}{s-m_h^2} \Big)^2
-\frac{64}{9} \cos^4(\epsilon) w^2 \alpha_{X}^6 \Big(\frac{1}{t-m_X^2}+\frac{1}{u-m_X^2}\Big)
\\&
- \frac{16c_1 \cos^3(\epsilon) w \alpha_X^4}{3(s-m_X^2)}
+ \frac{16c_2 \cos^2(\epsilon) \sin(\epsilon) w \alpha_X^4}{9(s-m_h^2)}
\\&
+\frac{64c_1 \cos^3(\epsilon)w^3 \alpha_X^6}{3(s-m_\phi^2)}\Big(\frac{1}{t-m_X^2}
+\frac{1}{u-m_X^2}\Big)
\\&
-\frac{64c_2 \cos^2(\epsilon) \sin(\epsilon) w^3 \alpha_X^6}{9(s-m_h^2)}\Big(\frac{1}{t-m_X^2} + \frac{1}{u-m_X^2}\Big)
\Big] \,,
\end{aligned}
\end{equation}
where, the couplings $c_1$ is
\begin{equation}
c_1 = 2 v_h \lambda_H \sin^3(\epsilon) + w \lambda_{HS} \cos(\epsilon) \sin^2(\epsilon)
+2 w \lambda_S \cos^3(\epsilon) \,,
\end{equation}
and the couplings $c_2$ is
\begin{equation}
\begin{aligned}
c_2 &= v_h \lambda_{HS} \cos(\epsilon) + 2 w \lambda_{HS} \sin(\epsilon)
-3 w \lambda_{HS} \sin^3(\epsilon) + 6 v_h \lambda_{H} \cos(\epsilon) \sin^2(\epsilon)
\\&
- 3 v_h \lambda_{HS} \cos(\epsilon) \sin^2(\epsilon)
- 6 w \lambda_{S} \cos^2(\epsilon) \sin(\epsilon) \,.
\end{aligned}
\end{equation}
|
\section{Introduction}
Fractional operators arise in the context of preconditioning of coupled multiphysics systems and, in particular, in the problem formulations where the coupling constraint is enforced by a Lagrange multiplier defined on the interface. Examples include the so-called EMI equations in modeling of excitable tissue \cite{tveito2017cell}, reduced order models of microcirculation in 2$d$-1$d$ \cite{kuchta2016preconditioners,lamichhane2004mortar} and 3$d$-1$d$ \cite{kuchta2021analysis,kuchta2019preconditioning} setting or Darcy/Biot--Stokes models \cite{ambartsumyan2018lagrange,layton2002coupling}. We remark that fractional operators have been recently utilized also in monolithic solvers for formulations of Darcy/Biot--Stokes models without the Lagrange multipliers, see \cite{boon2022parameter,boon2022robust}.
The coupling in the multiplier formulations is naturally posed in Sobolev spaces of fractional order. However, for parameter robustness of iterative methods, a more precise setting must be considered, where the interface problem is posed in the intersection (space) of parameter-weighted fractional order Sobolev spaces \cite{holter2020robust}. Here the sums of fractional operators induce the natural inner product. In the examples mentioned earlier, however, the interface preconditioners were realized by eigenvalue decomposition, and thus, while being parameter robust, the resulting solvers do not scale with mesh size. Using Darcy--Stokes system as the canonical example, we aim to show that rational approximations are crucial for designing efficient preconditioners for multiphysics systems.
There have been numerous works on approximating/preconditioning problems such as $\mathcal{D}^s u=f$, e.g. \cite{2015BonitoPasciak-a,2016ChenNochettoOtarolaSalgado-a,2016NochettoOtarolaSalgado-a}. The rational approximation (RA) approach has been advocated, and several techniques based on it were proposed by Hofreither in~\cite{hofreitherUnifiedViewNumerical2020,2021Hofreither-a} (where also a \verb|Python| implementation of Best Uniform Rational Approximation (BURA) for $x^{\lambda}$ is found). Further rational approximations used in preconditioning can be found in~\cite{2020HarizanovLazarovMargenovMarinov-a,2020HarizanovLazarovMargenovMarinovPasciak-a}. An interesting approach, which always leads to real-valued poles and uses Pad\'{e} approximations of suitably constructed time-dependent problems, is found in~\cite{lazarov2020_time}. In our numerical tests, we use Adaptive Antoulas-Anderson (AAA) algorithm~\cite{nakatsukasaAAAAlgorithmRational2018} which is a greedy strategy for locating the interpolation points and then using the barycentric representation of the rational interpolation to define a function that is close to the target. In short, for a given continuous function $f$, the AAA algorithm returns a rational function that approximates the best uniform rational approximation to $f$.
The rest of the paper is organized as follows. In Section~\ref{sec:darcy-stokes-model}, we introduce a model problem that leads to the sum of fractional powers of a differential operator on an interface acting on weighted Sobolev spaces. In Section~\ref{sec:mixed-FEM}, we introduce the finite element discretizations that we employ for the numerical solution of such problems. Next, Section~\ref{sec:ra} presents some details on the rational approximation and the scaling of the discrete operators. In Section~\ref{sec:results}, we test several relevant scenarios and show the robustness of the rational approximation as well as the efficacy of the preconditioners. Conclusions are drawn in Section~\ref{sec:conclusions}.
\section{Darcy--Stokes model}\label{sec:darcy-stokes-model}
We study interaction between a porous medium occupying $\Omega_D\subset\mathbb{R}^d$, $d=2, 3$ surrounded by a free flow domain $\Omega_S\supset\Omega_D$ given by a Darcy--Stokes model \cite{layton2002coupling} as: For given volumetric source terms $f_S:\Omega_S\rightarrow\mathbb{R}^d$ and $f_D:\Omega_D\rightarrow\mathbb{R}$ find the Stokes velocity and pressure $u_S:\Omega_S\rightarrow\mathbb{R}^d$, $p_S:\Omega_S\rightarrow\mathbb{R}$ and the Darcy flux and pressure $u_D:\Omega_D\rightarrow\mathbb{R}^d$, $p_D:\Omega_D\rightarrow\mathbb{R}$ such that
\begin{equation}\label{eq:darcy_stokes}
\begin{aligned}
-\nabla\cdot\sigma(u_S, p_S) = f_S \text{ and } \nabla\cdot u_S &= 0 &\text{ in }\Omega_S, \\
u_D + K\mu^{-1}\nabla p_D = 0 \text{ and } \nabla\cdot u_D &= f_D &\text{ in }\Omega_D, \\
u_S\cdot \nu_s + u_D\cdot \nu_D &= 0 &\text{ on }\Gamma, \\
-\nu_S\cdot\sigma(u_S, p_s)\cdot\nu_S - p_D &= 0 &\text{ on }\Gamma, \\
-P_{\nu_S}\left(\sigma(u_S, p_S)\cdot\nu_S\right) - \alpha\mu K^{-1/2}P_{\nu_S}u_S &= 0 &\text{ on }\Gamma.
\end{aligned}
\end{equation}
Here, $\sigma(u, p):=2\mu\epsilon(u)-pI$ with $\epsilon(u):=\tfrac{1}{2}(\nabla u + \nabla u^{T})$.
Moreover, $\Gamma:=\partial\Omega_D\cap\partial\Omega_S$ is the common interface,
$\nu_S$ and $\nu_D=-\nu_S$ represent the outward unit normal vectors on $\partial\Omega_S$ and $\partial\Omega_D$. Given a surface with normal vector $\nu$,
$P_{\nu}:=I-\nu\otimes\nu$ denotes a projection to the tangential plane with normal $\nu$. The final three
equations in \eqref{eq:darcy_stokes} represent the coupling conditions on $\Gamma$. Parameters of the model
(which we shall assume to be constant) are viscosity $\mu>0$, permeability $K>0$ and Beavers-Joseph-Saffman
parameter $\alpha>0$.
Finally, let us decompose the outer boundary $\partial\Omega_S\setminus\Gamma=\Gamma_u\cup\Gamma_\sigma$,
$\lvert\Gamma_i\rvert>0$, $i=u, \sigma$ and introduce the boundary conditions to
close the system \eqref{eq:darcy_stokes}
\begin{equation*
\begin{aligned}
u_S\cdot\nu_S &= 0 \text{ and }
P_{\nu_S}\left(\sigma(u_S, p_S)\cdot\nu_S\right) = 0&\text{ on }\Gamma_u, \\
%
\sigma(u_S, p_S)\cdot\nu_S &= 0&\text{ on }\Gamma_{\sigma}.
\end{aligned}
\end{equation*}
Thus, $\Gamma_u$ is an impermeable free-slip boundary. We remark
that other conditions, in particular no-slip on $\Gamma_u$, could be considered without introducing additional challenges. However, unlike the tangential component, constraints for the normal component of velocity are easy\footnote{ Conditions on the normal component can be implemented as Dirichlet boundary conditions and enforced by the constructions of the finite element trial and test spaces. The tangential component can be controlled e.g., by the Nitsche method \cite{stenberg1995some} which modifies the discrete problem operator.} to implement in $H(\div)$-conforming discretization schemes considered below.
Letting $V_S\subset H_1(\Omega_S)$, $V_D\subset H(\div, \Omega_D)$,
$Q_S\subset L^2(\Omega_S)$, $Q_D\subset L^2(\Omega_D)$ and $Q\subset H^{1/2}(\Gamma)$ the variational formulation of \eqref{eq:darcy_stokes} seeks to find $w:=(u, p, \lambda)\in W$, $W:=V\times Q\times \Lambda$, $V:=V_S\times V_D$, $Q:=Q_S \times Q_D$ and $u:=(u_S, u_D)$, $p:=(p_S, p_D)$ such that $\mathcal{A}w = L$ in $W'$, the dual space of $W$, where $L$ is the linear functional of the right-hand sides in \eqref{eq:darcy_stokes} and the problem operator $\mathcal{A}$ satisfies
\begin{equation}\label{eq:darcy_stokes_operator}
\langle \mathcal{A}w, \delta w\rangle = a_S(u_S, v_S) + a_D(u_D, v_D) + b(u, q) + b_{\Gamma}(v, \lambda)+
b(v, p) + b_{\Gamma}(u, \delta\lambda),
\end{equation}
where $\delta w:=(v, q, \delta\lambda)$, $v:=(v_S, v_D)$, $q:=(q_S, q_D)$ and $\langle\cdot,\cdot\rangle$ denotes a duality pairing between $W$ and $W'$. The bilinear forms in \eqref{eq:darcy_stokes_operator} are defined as
\begin{equation}\label{eq:darcy_stokes_components}
\begin{aligned}
a_S(u_S, v_S) &:= \int_{\Omega_S}2\mu\epsilon(u_S):\epsilon(v_S)\,\mathrm{d}x + \int_{\Gamma}{\alpha\mu}{K^{-1/2}}P_{\nu_S}u_S\cdot P_{\nu_S}v_S\,\mathrm{d}x,\\
%
a_D(u_D, v_D) &:= \int_{\Omega_D}{\mu}{K}^{-1}u_D\cdot v_D\,\mathrm{d}x,\\
%
b(v, p) &:= -\int_{\Omega_S}p_S\nabla\cdot v_S\,\mathrm{d}x -\int_{\Omega_D}p_D\nabla\cdot v_D\,\mathrm{d}x,\\
%
b_{\Gamma}(v, \lambda) &:= \int_{\Gamma}(v_S\cdot \nu_S + v_D\cdot \nu_D)\lambda\,\mathrm{d}s.
%
\end{aligned}
\end{equation}
Here, $\lambda$ represents a Lagrange multiplier whose physical meaning is related to the normal component of the traction vector on $\Gamma$, $\lambda:=-\nu_S\cdot\sigma(u_S, p_S)\cdot\nu_S$, see~\cite{layton2002coupling} where also well-posedness of the problem in the space $W$ above is established.
Following \cite{holter2020robust}, parameter-robust preconditioners for Darcy--Stokes operator $\mathcal{A}$
utilize weighted sums of fractional operators on the interface. Specifically, we shall consider the following operator
\begin{equation}\label{eq:darcy_stokes_iface}
S := {\mu}^{-1}(-\Delta_{\Gamma} + I_{\Gamma})^{-1/2} + {K}{\mu}^{-1}(-\Delta_{\Gamma} + I_{\Gamma})^{1/2},
\end{equation}
where we have used the subscript $\Gamma$ to emphasize that the operators are considered on the interface.
We note that $-\Delta_{\Gamma}$ is singular in our setting as $\Gamma$ is a closed surface and
adding lower order term $I_{\Gamma}$ thus ensures positivity.
Letting $A_S$ be the operator induced by the bilinear form $a_S$ in \eqref{eq:darcy_stokes_components} we define the Darcy--Stokes preconditioner as follows,
\begin{equation}\label{eq:darcy_stokes_precond}
\mathcal{B} := {\operatorname{diag}}\left(A_S,\mu {K}^{-1}(I-\nabla\nabla\cdot), {\mu}^{-1}I, {K}{\mu}^{-1}I, S \right)^{-1}.
\end{equation}
Note that the operators $I$ in the pressure blocks of the preconditioner act on different spaces/spatial domains, i.e., $Q_S$ and $Q_D$. We remark that in the context of Darcy--Stokes preconditioning, \cite{harizanov2022rational} consider BURA approximation for a simpler interfacial operator, namely, ${K}{\mu}^{-1}(-\Delta_{\Gamma} + I_{\Gamma})^{1/2}$. However, the preconditioner cannot yield parameter robustness, cf. \cite{holter2020robust}.
\section{Mixed Finite Element Discretization}\label{sec:mixed-FEM}
In order to assess numerically the performance of rational approximation of $S^{-1}$ in \eqref{eq:darcy_stokes_precond}, stable discretization of the Darcy--Stokes system is needed. In addition to parameter variations, here we also wish to show the algorithm's robustness to discretization and, in particular, the construction of the discrete multiplier space. To this end, we require a family of stable finite element discretizations.
Let $k\geq 1$ denote the polynomial degree. For simplicity, to make sure that the Lagrange multiplier fits well with the discretization of both the Stokes and Darcy domain, we employ the same $H(\mbox{div})$ based discretization in both domains. That is, we discretize the Stokes velocity space $V_S$ by Brezzi-Douglas-Marini $\mathbb{BDM}_{k}$ elements \cite{brezzi1985two} over simplicial triangulations $\Omega^h_{S}$ of $\Omega_S$ and likewise, approximations to $V_D$ are constructed with $\mathbb{BDM}_{k}$ elements on $\Omega^h_{D}$. Here, $h$ denotes the characteristic mesh size. Note that by construction, the velocities and fluxes have continuous normal components across the \emph{interior} facets of the respective triangulations. However, on the interface $\Gamma$, we do not impose any continuity between the vector fields. The pressure spaces $Q_S$, $Q_D$ shall be approximated in terms of discontinuous piecewise polynomials of degree $k-1$, $\mathbb{P}^{\text{disc}}_{k-1}$. Finally, the Lagrange multiplier space is constructed by $\mathbb{P}^{\text{disc}}_{k}$ elements on the triangulation $\Gamma^h:=\Gamma^h_S$, with $\Gamma^h_S$ being the trace mesh of $\Omega^h_{S}$ on $\Gamma$. For simplicity, we assume $\Gamma^h_D=\Gamma^h_S$.
Approximation properties of the proposed discretization are demonstrated in \Cref{fig:darcy_stokes_cvrg}. It can be seen that all the quantities converge with order $k$ (or better) in their respective norms. This is particularly the case for the Stokes velocity, where the error is measured in the $H^1$ norm.
Let us make a few remarks about our discretization. First, observe that by using $\mathbb{BDM}$ elements on a global mesh $\Omega^h_{S}\cup\Omega^h_{D}$ the Darcy--Stokes problem \eqref{eq:darcy_stokes} can also be discretized such that the mass conservation condition $u_S\cdot\nu_S+u_D\cdot\nu_D=0$ on $\Gamma$ is enforced by construction, i.e. no Lagrange multiplier is required. Here, $u$ is the global $H(\div)$-conforming vector field, with $u_i:=u|_{\Omega_i}$, $i=S, D$. Second, we note that the chosen discretization of $V_S$ is only $H(\div, \Omega_S)$-conforming. In turn, stabilization of the tangential component of the Stokes velocity is needed, see, e.g., \cite{2014AyusodeDiosBrezziMariniXuZikatanov}, which translates to modification of the bilinear form $a_S$ in \eqref{eq:darcy_stokes_components} as
\begin{equation}\label{eq:sym_grad_BDM}
\begin{split}
a^h_S(u, v) &:= a_S(u, v) - \sum_{e\in F^h_{S}}\int_{e}2\mu \avg{\epsilon(u)}\cdot\jump{P_{\nu_e}v}\,\mathrm{d}s\\
&- \sum_{e\in F^h_{S}}\int_{e}2\mu\avg{\epsilon(v)}\cdot\jump{P_{\nu_e}u}\,\mathrm{d}s
+ \sum_{e\in F^h_{S}}\int_{e}\frac{2\mu\gamma}{h_e}\jump{P_{\nu_e}u}\cdot \jump{P_{\nu_e}v}\,\mathrm{d}s,
\end{split}
\end{equation}
see \cite{hong2016robust}. Here, $F^{h}_S=F(\Omega^h_S)$ is the collection of interior facets $e$ of triangulation $\Omega^h_S$, while $\nu_e$, $h_e$ denote respectively the facet normal and facet diameter. The stabilization parameter $\gamma>0$ has to be chosen large enough to ensure the coercivity of $a^h_S$. However, the value depends on the polynomial degree. Finally, for interior facet $e$ shared by elements $T^{+}$, $T^{-}$ we define the (facet) jump of a vector $v$ as $\jump{v}:=v|_{T^{+}\cap e}-v|_{T^{-}\cap e}$ and the (facet) average of a tensor $\epsilon$ as $\avg{\epsilon}:=\tfrac{1}{2}\left(v|_{T^{+}\cap e}\cdot\nu_e+v|_{T^{-}\cap e}\cdot\nu_e\right)$.
We conclude this section by discussing discretization of the operator $(-\Delta+I)$ needed for realizing the interface preconditioner \eqref{eq:darcy_stokes_iface}. Since, in our case, the multiplier space $\Lambda_h$ is only $L^2$-conforming we adopt the symmetric interior penalty approach \cite{di2011mathematical} so that in turn for $u, v\in \Lambda_h$,
\begin{equation}\label{eq:delta_dg}
\begin{split}
\langle (-\Delta+I) u, v\rangle := &\int_{\Gamma}(\nabla u\cdot \nabla v + u v)\,\mathrm{d}x
- \sum_{e\in F(\Gamma_h)}\int_{e}\avg{\nabla u} \jump{v}\,\mathrm{d}s\\
&- \sum_{e\in F(\Gamma_h)}\int_{e}\avg{\nabla v} \jump{u}\,\mathrm{d}s
+ \sum_{e\in F(\Gamma_h)}\int_{e}\frac{\gamma}{h_e}\jump{u}\jump{v}\,\mathrm{d}s.
\end{split}
\end{equation}
Here, the jump of a scalar $f$ is computed as $\jump{f}:=f|_{T^{+}\cap e}-f|_{T^{-}\cap e}$ and the average of a vector $v$ reads $\avg{v}:=\tfrac{1}{2}\left(v|_{T^{+}\cap e}\cdot\nu_e+v|_{T^{-}\cap e}\cdot\nu_e\right)$. As before, $\gamma>0$ is a suitable stabilization parameter.
\begin{figure}
\centering
\includegraphics[width=0.325\textwidth]{./figures/ds_cvrg_p1.pdf}
\includegraphics[width=0.325\textwidth]{./figures/ds_cvrg_p2.pdf}
\includegraphics[width=0.325\textwidth]{./figures/ds_cvrg_p3.pdf}
\vspace{-10pt}
\caption{Approximation properties of $\mathbb{BDM}_k$-$\mathbb{BDM}_k$-$\mathbb{P}^{\text{disc}}_{k-1}$-$\mathbb{P}^{\text{disc}}_{k-1}$-$\mathbb{P}^{\text{disc}}_k$ discretization of the Darcy--Stokes problem. Two-dimensional setting is considered with $\Omega_S=(0, 1)^2$ and $\Omega_D=(\tfrac{1}{4}, \tfrac{3}{4})^2$. Parameter values are set as $K=2$, $\mu=3$ and $\alpha=\tfrac{1}{2}$. Left figure is for $k=1$ while $k=2,3$ is shown in the middle and right figures, respectively.
}
\label{fig:darcy_stokes_cvrg}
\end{figure}
\section{Rational approximation for the general problems of sums of fractional operators} \label{sec:ra}
Let $s, t \in [-1, 1]$ and $\alpha, \beta \geq 0$ where at least one of $\alpha, \beta$ is not zero. For interval $I \subset \mathbb{R}^+$, consider a function
\(
f(x) = (\alpha x^s + \beta x^t)^{-1}, \quad x \in I.
\).
The basic idea is to find a rational function $R(x)$ approximating $f$ on $I$, that is,
\
R(x) = \frac{P_{k'}(x)}{Q_{k}(x)}\approx f(x)
\),
where $P_{k'}$ and $Q_{k}$ are polynomials of degree $k'$ and $k$, respectively. Assuming $ k' \leq k $, the rational function can be given in the following partial fraction form
\begin{equation*}
R(x) = c_0 + \sum_{i=1}^N \frac{c_i}{x - p_i},
\end{equation*}
for $ c_0 \in {\mathbb{R}} $, $ c_i, p_i \in {\mathbb{C}} $, $ i = 1, 2, \dots, N$. The coefficients $p_i$ and $c_i$ are called \emph{poles} and \emph{residues} of the rational approximation, respectively.
We note that the rational approximation has been predominantly explored to approximate functions with only one fractional power, that is $x^{-\bar{s}}$ for $\bar{s} \in (0, 1) $ and $ x > 0 $. Additionally, the choice of the rational approximation method that computes the poles and residues is not unique. One possibility is the BURA method which first computes the best uniform rational approximation $\bar{r}_{\beta}(x)$ of $x^{\beta -\bar{s}}$ for a positive integer $\beta > \bar{s}$ and then uses $\bar{r}(x) = \frac{\bar{r}_{\beta}(x)}{x}$ to approximate $x^{-\bar{s}}$. Another possible choice is to use the rational interpolation of $z^{-\bar{s}}$ to obtain $\bar{r}(x)$. The AAA algorithm proposed in \cite{nakatsukasaAAAAlgorithmRational2018} is a good candidate. The AAA method is based on the representation of the rational approximation in barycentric form and greedy selection of the interpolation points. Both approaches lead to the poles $ p_i \in {\mathbb{R}}, \, p_i \leq 0 $ for the case of one fractional power. An overview of rational approximation methods can be found in \cite{hofreitherUnifiedViewNumerical2020}.
The location of the poles is crucial in rational approximation preconditioning. For $\bar{f}(x) = x^{\bar{s}}$, $\bar{s} \in (0,1)$ the poles of the rational approximation for $\bar{f}$ are all real and negative. Hence, in the case of a positive definite operator $\mathcal{D}$, the approximation of $\mathcal{D}^{\bar{s}}$ requires only inversion of positive definite operators of the form $\mathcal{D}+|p_i| I$, for $i = 1, 2, \dots, N$, $p_i \neq 0 $. Such a result for rational approximation of $\bar{f}$ with $\bar{s}\in(-1,1)$ is found in a paper by H.~Stahl~\cite{2003Stahl-a}. In the following, we present an extensive set of numerical tests for the class of \emph{sum of fractional operators}, which gives a wide class of efficient preconditioners for the multiphysics problems coupled through an interface. The numerical tests show that the poles remain real and nonpositive in most combinations of fractional exponents $s$ and $t$.
Let $ V $ be a Hilbert space and $ V'$ be its dual. Consider a symmetric positive definite (SPD) operator $ A : V \to V' $. Then, the rational function $ R(\cdot) $ can be used to approximate $ f(A) $ as follows,
\begin{equation*}
z = f(A) r \approx c_0 r + \sum_{i=1}^N c_i \left( A - p_i I \right)^{-1}r
\end{equation*}
with $ z \in V $ and $r \in V'$. The overall algorithm is shown in \Cref{alg:rational-approx}.
\begin{algorithm}
\caption{Compute $ z = f(A) r $ using rational approximation.} \label{alg:rational-approx}
\begin{algorithmic}[1]
\STATE Solve for $ w_i $:
$ \left( A - p_i I \right) w_i = r, \quad i = 1, 2, \dots, N. $
\STATE Compute: $ z = c_0 r + \sum\limits_{i=1}^N c_i w_i $
\end{algorithmic}
\end{algorithm}
Without loss of generality we let the operator $ A $ be a discretization of the Laplacian operator $ -\Delta $,
and $ I $ is the identity defined using the standard $L^2$ inner product. In particular, unlike in \eqref{eq:darcy_stokes_iface}
we assume that $ -\Delta $ is SPD (e.g. by imposing boundary conditions eliminating the constant nullspace).
Therefore, the equations in Step 1 of \Cref{alg:rational-approx} can be viewed as discretizations
of the shifted Laplacian problems $ -\Delta \, w_i - p_i \, w_i = r $, $p_i<0$, and we can use
efficient numerical methods, such as Algebraic MultiGrid (AMG) methods \cite{1stAMG,2017XuZikatanov}, for their solution.
We would like to point out that in the implementation, the operators involved in \Cref{alg:rational-approx} are replaced by their matrix representations on a concrete basis and are properly scaled. We address this in more detail in the following section.
\subsection{Preconditioning} \label{subsec:ra_preconditioning}
Let $ {\mathsf{A}} $ be the stiffness matrix associated with $ -\Delta$ and ${\mathsf{M}}$ a corresponding mass matrix of the $L^2$ inner product. Also, denote with $ n_c $ the number of columns of $ {\mathsf{A}} $. The problem we are interested in is constructing an efficient preconditioner for the solution of the linear system $F({\mathsf{A}}) \mathsf{x} = \mathsf{b}$.
Thus, we would like to approximate $ f({\mathsf{A}}) = F({\mathsf{A}})^{-1} $ using the rational approximation $ R(x) $ of $ f(x)=\frac{1}{F(x)} $.
Let $ {\mathsf{I}} $ be a $ n_c \times n_c $ identity matrix and let $ {\mathsf{U}} $ be an $ {\mathsf{M}} $-orthogonal matrix of the eigenvectors of the generalized eigenvalue problem $ {\mathsf{A}} \mathsf{u}_j = \lambda_j {\mathsf{M}} \mathsf{u}_j $, $ j = 1, 2, \dots, n_c $, namely,
\begin{equation}\label{eq:gep_ra}
{\mathsf{A}} {\mathsf{U}} = {\mathsf{M}} {\mathsf{U}} {\mathsf{\Lambda}}, \quad {\mathsf{U}}^T {\mathsf{M}} {\mathsf{U}} = {\mathsf{I}} \quad \Longrightarrow\quad {\mathsf{U}}^T {\mathsf{A}} {\mathsf{U}} = {\mathsf{\Lambda}}.
\end{equation}
For any continuous function $ G : [0, \rho] \rightarrow \mathbb{R}$ we define
\begin{equation}\label{eq:function_gep}
G({\mathsf{A}}) \coloneqq {\mathsf{M}} {\mathsf{U}} G({\mathsf{\Lambda}}) {\mathsf{U}}^T {\mathsf{M}}, \quad\mbox{i.e.}, \quad
f({\mathsf{A}})={\mathsf{M}} {\mathsf{U}} f({\mathsf{\Lambda}}) {\mathsf{U}}^T {\mathsf{M}}.
\end{equation}
where $ \rho \coloneqq \rho \left({\mathsf{M}}^{-1} {\mathsf{A}} \right) $ is the spectral radius of the matrix $ {\mathsf{M}}^{-1} {\mathsf{A}} $.
A simple consequence from the Chebyshev Alternation Theorem is that the residues/poles for the rational approximation of $f(x)$ for $x\in [0,\rho]$ are obtained by scaling the residues/poles of the rational approximation of $ g(y) = f\left( \rho y \right) $ for $ y \in [0,1] $. Indeed, we have, $c_i(f)=\rho c_i(g)$, and $p_i(f)=\rho p_i(g)$. Therefore, in the implementation, we need an upper bound on $ \rho \left( {\mathsf{M}}^{-1}{\mathsf{A}} \right) $. For $\mathbb{P}_1 $ finite elements, such a bound can be obtained following the arguments from \cite{wathen1987realistic},
\begin{equation}\label{diag}
\rho\left({\mathsf{M}}^{-1} {\mathsf{A}}\right)
\le \frac{1}{\lambda_{\min{}}({\mathsf{M}})} \|{\mathsf{A}}\|_{\infty}
=d(d+1)\left\|\operatorname{diag}({\mathsf{M}})^{-1}\right\|_{\infty} \| {\mathsf{A}} \|_{\infty}
\end{equation}
with $d$ the spatial dimension\footnote{Such estimates can be carried out for Lagrange finite elements of any polynomial degree because the local mass matrices are of a special type: a constant matrix, which depends only on the dimension and the polynomial degree, times the volume of the element.}.
\begin{proposition} \label{prop:ra_for_matrix}
Let $ R_f(\cdot) $ be the rational approximation for the function $ f(\cdot) $ on $ [0, \rho] $. Then for the stiffness matrix $ {\mathsf{A}} $ and mass matrix $ {\mathsf{M}} $ satisfying \eqref{eq:gep_ra}, we have
\begin{equation} \label{eq:ra_for_matrix}
f({\mathsf{A}}) \approx R_f({\mathsf{A}}) = c_0 {\mathsf{M}}^{-1} + \sum_{i=1}^N c_i \left({\mathsf{A}} - p_i {\mathsf{M}} \right)^{-1}.
\end{equation}
\end{proposition}
\begin{proof}
The relations \eqref{eq:gep_ra} imply that $ {\mathsf{U}} {\mathsf{U}}^T = {\mathsf{M}}^{-1}$, and therefore,
\begin{equation*}
{\mathsf{A}} {\mathsf{U}} - p_i {\mathsf{M}} {\mathsf{U}} = {\mathsf{M}} {\mathsf{U}} ({\mathsf{\Lambda}} - p_i {\mathsf{I}})
\Longleftrightarrow
({\mathsf{A}} - p_i {\mathsf{M}}){\mathsf{U}} = {\mathsf{M}} {\mathsf{U}}({\mathsf{\Lambda}} - p_i {\mathsf{I}}).
\end{equation*}
This is equivalent to \(
({\mathsf{\Lambda}} - p_i {\mathsf{I}})^{-1}\underbrace{{\mathsf{U}}^{-1} {\mathsf{M}}^{-1}}_{{\mathsf{U}}^T} = {\mathsf{U}}^{-1}({\mathsf{A}} - p_i {\mathsf{M}})^{-1}
\). Hence,
\begin{equation*}
({\mathsf{\Lambda}} - p_i {\mathsf{I}})^{-1} = {\mathsf{U}}^{-1}({\mathsf{A}} - p_i {\mathsf{M}})^{-1} {\mathsf{U}}^{-T}.
\end{equation*}
A straightforward substitution in~\eqref{eq:function_gep} then shows \eqref{eq:ra_for_matrix}.
\qed
\end{proof}
In addition, to apply the rational approximation preconditioner, we need to compute the actions of ${\mathsf{M}}^{-1}$ and each $\left({\mathsf{A}} - \rho p_i {\mathsf{M}} \right)^{-1}$. If $ p_i \in {\mathbb{R}}, \, p_i \leq 0$, this leads to solving a series of elliptic problems where the AMG methods are very efficient.
\section{Numerical results} \label{sec:results}
In this section, we present two sets of experiments: (1) on the robustness of the rational approximation with respect to the scaling parameters and the fractional exponents; and (2) on the efficacy of the preconditioned minimal residual (MinRes) method as a solver for Darcy--Stokes coupled model. We use the AAA algorithm \cite{nakatsukasaAAAAlgorithmRational2018} to construct a rational approximation. The discretization and solver tools are \verb|Python| modules provided by FEniCS\textunderscore{ii} \cite{kuchta2021assembly}, cbc.block \cite{mardal2012block}, and interfaced with the HAZmath library \cite{hazmath}.
\subsection{Approximating the sum of two fractional exponents} \label{subsec:example_1}
In this example, we test the approximation power of the rational approximation computed by AAA algorithm regarding different fractional exponents $s, t$ and parameters $\alpha, \beta$. That is, we study the number of poles $N $ required to achieve
\begin{equation*
\| f - R \|_{\infty} = \max\limits_{x \in [0, 1]} \left| (\alpha x^s + \beta x^t)^{-1} - \left(c_0 + \sum_{i=1}^N \frac{c_i}{x - p_i} \right) \right| \leq \epsilon_{\text{RA}},
\end{equation*}
for a fixed tolerance $ \epsilon_{\text{RA}} = 10^{-12}$. In this case, we consider the fractional function $ f $ to be defined on the unit interval $I = (0, 1]$. As we noted earlier, however, the approximation can be straightforwardly extended to any interval. We also consider the scaling regarding the magnitude of parameters $ \alpha $ and $ \beta $. Specifically, in case when $ \alpha > \beta $, we rescale the problem with $ \gamma_\alpha = \frac{\beta}{\alpha} < 1 $ and approximate
\begin{equation*}
\widetilde{f}(x) = (x^s + \gamma_\alpha x^t)^{-1} \approx R_{\widetilde{f}}(x).
\end{equation*}
Then the rational approximation for the original function $ f $ is given as
\begin{equation*}
f(x) \approx R(x) = \frac{1}{\alpha} R_{\widetilde{f}}(x).
\end{equation*}
Similar can be done in case when $ \beta > \alpha $.
\begin{figure}[t!]
\centering
\includegraphics[width=\textwidth]{./figures/poles_alpha_beta_s_t.pdf}
\vspace{-25pt}
\caption{Visualization of the number of poles in the rational approximations of the function $f(x) = (\alpha x^s + \beta x^t)^{-1}$ for $ x \in(0, 1] $ with regards to varying fractional exponents $s, t$ and coefficients $\alpha, \beta$.}
\label{fig:number_of_poles}
\end{figure}
The results are summarized in \Cref{fig:number_of_poles}. To obtain different parameter ratios $\gamma_\alpha$, we take $\alpha \in \{10^{-9}, 10^{-6}, 10^{-3}, 1 \} $ and $ \beta \in \{10^{-10}, 10^{-6}, 10^{-2}, 10^2 \} $. Furthermore, we vary the fractional exponents $s, t \in [-1, 1] $ with the step 0.2. We observe that the number of poles $N$ remains relatively uniform with varying the exponents, except in generic cases when $ s, t = \{-1, 0, 1 \}$. For example, for the combination $(s, t) = (1, 1)$, the function we are approximating is $ f(x) = \tfrac{1}{2x} $, thus the rational approximation should return only one pole $ p_1 = 0 $ and residues $ c_0 = 0, c_1 = {\frac{1}{2}} $. We also observe that for the fixed tolerance of $ \epsilon_{\text{RA}} = 10^{-12}$, we obtain a maximum of 22 poles in all cases.
Additionally, we remark that in most test cases, we retain real and negative poles, which is a desirable property to apply the rational approximation as a positive definite preconditioner. However, depending on the choice of fractional exponents $s, t$ and tolerance $\epsilon_{\text{RA}}$, the algorithm can produce positive or a pair of complex conjugate poles. Nevertheless, these cases are rare, and the number and the values of those poles are small. Therefore numerically, we do not observe any significant influence on the rational approximation preconditioner. More concrete analytical results on the location of poles for sums of fractionalities are part of our future research.
\subsection{The Darcy--Stokes problem}\label{subsec:example_ds}
In this section, we discuss the performance of RA approximation of $S^{-1}$ in the Darcy--Stokes preconditioner \eqref{eq:darcy_stokes_precond}. To this end, we consider $\Omega_S=(0, 1)^2$, $\Omega_D=(\tfrac{1}{4}, \tfrac{3}{4})^2$ and fix the value of Beavers-Joseph-Saffman parameter $\alpha=1$ while permeability and viscosity are varied\footnote{These ranges are identified as relevant for many applications in biomechanics \cite{boon2022robust}.} $10^{-6}\leq K\leq 1$ and $10^{-6}\leq\mu\leq 10^{2}$. The system is discretized by $\mathbb{BDM}_k$-$\mathbb{BDM}_k$-$\mathbb{P}^{\text{disc}}_{k-1}$-$\mathbb{P}^{\text{disc}}_{k-1}$-$\mathbb{P}^{\text{disc}}_k$ elements, $k=1, 2, 3$, which were shown to provide convergent approximations in \Cref{fig:darcy_stokes_cvrg}. A hierarchy of meshes $\Omega^h_S$, $\Omega^h_D$ is obtained by uniform refinement. We remark that the stabilization constants $\gamma$ in \eqref{eq:sym_grad_BDM} and \eqref{eq:delta_dg} are chosen as $\gamma=20k$ and $\gamma=10k$ respectively. The resulting linear systems are then solved by preconditioned MinRes method. To focus on the RA algorithm in the preconditioner, for all but the multiplier block, an exact LU decomposition is used. Finally, the iterative solver is always started from a zero initial guess and terminates when the preconditioned residual norm drops below $10^{-10}$.
We present two sets of experiments. First, we fix the tolerance in the RA algorithm at $\epsilon_{\text{RA}}=2^{-40}\approx 10^{-12}$ and demonstrate that using RA in the preconditioner \eqref{eq:darcy_stokes_precond} leads to stable MinRes iterations for any practical values of the material parameters, the mesh resolution $h$ and the polynomial degree $k$ in the finite element discretization (see \Cref{fig:darcy_stokes_param_robust}). Here it can be seen that the number of iterations required for convergence is bounded in the above-listed quantities. Results for different polynomial degrees are largely similar and appear to be mostly controlled by material parameters. However, despite the values of $K$, $\mu$ spanning several orders of magnitude, the iterations vary only between 30 to 100.
\begin{figure}[ht!]
\centering
\includegraphics[width=\textwidth]{./figures/ds_robust_p1.pdf}
\includegraphics[width=\textwidth]{./figures/ds_robust_p2.pdf}
\includegraphics[width=\textwidth]{./figures/ds_robust_p3.pdf}
\vspace{-25pt}
\caption{
Number of MinRes iterations required for convergence with preconditioner \eqref{eq:darcy_stokes_precond} using the RA with tolerance $\epsilon_{\text{RA}}=2^{-40}$. Setup from \Cref{fig:darcy_stokes_cvrg} is considered with the system discretized by $\mathbb{BDM}_k$-$\mathbb{BDM}_k$-$\mathbb{P}^{\text{disc}}_{k-1}$-$\mathbb{P}^{\text{disc}}_{k-1}$-$\mathbb{P}^{\text{disc}}_k$
elements. (Top) $k=1$, (middle) $k=2$, (bottom) $k=3$.
}
\label{fig:darcy_stokes_param_robust}
\end{figure}
Next, we assess the effects of the accuracy in RA on the performance of preconditioned MinRes solver. Let us fix $k=1$ and vary the material parameters as well as the RA tolerance $\epsilon_{\text{RA}}$. In \Cref{fig:darcy_stokes_ra_robust}, we observe that the effect of $\epsilon_{\text{RA}}$ varies with material properties (which enter the RA algorithm through scaling). In particular, it can be seen that for $K=1$, the number of MinRes iterations is practically constant for any $\epsilon_{\text{RA}}\leq 10^{-1}$. On the other hand, when $K=10^{-6}$, the counts vary with $\epsilon_{\text{RA}}$ and to a lesser extent with $\mu$. Here, lower accuracy typically leads to a larger number of MinRes iterations. However, for $\epsilon_{\text{RA}}\leq 10^{-4}$ the iterations behave similarly. We remark that with $K=10^{-6}$ (and any $10^{-6}\leq\mu\leq 10^{2}$), the tolerance $\epsilon_{\text{RA}}=10^{-1}$ leads to 2 poles, cf. \Cref{fig:darcy_stokes_ra_perf}, while for $K=1$ there are at least 5 poles needed in the RA approximation.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{./figures/ds_ra_K1.0.pdf}
\includegraphics[width=\textwidth]{./figures/ds_ra_K1e-06.pdf}
\vspace{-25pt}
\caption{
Number of MinRes iterations required for convergence with preconditioner
\eqref{eq:darcy_stokes_precond} using the RA with varying tolerance $\epsilon_{\text{RA}}$.
(Top) $K=1$. (Bottom) $K=10^{-6}$. Setup as in \Cref{fig:darcy_stokes_param_robust} is used with discretization by
$\mathbb{BDM}_1$-$\mathbb{BDM}_1$-$\mathbb{P}^{\text{disc}}_{0}$-$\mathbb{P}^{\text{disc}}_{0}$-$\mathbb{P}^{\text{disc}}_1$ elements.
}
\label{fig:darcy_stokes_ra_robust}
\end{figure}
Our results demonstrate that RA approximates $S^{-1}$ in \eqref{eq:darcy_stokes_iface}, which leads to a robust, mesh, and parameter-independent Darcy--Stokes solver. We remark that though the algorithm complexity is expected to scale with the number of degrees of freedom on the interface, $n_{h}=\text{dim}\Lambda_h$, which is often considerably smaller than the total problem size, the setup cost may become prohibitive. This is particularly true for spectral realization, which often results in $\mathcal{O}(n^3_h)$ complexity. To address such issues, we consider how the setup time of RA and the solution time of the MinRes solver depend on the problem size.
In \Cref{fig:darcy_stokes_ra_perf} we show the setup time of RA (for fixed material parameters) as function of mesh size and $\epsilon_{\text{RA}}$. It can be seen that the times are $<0.1\,\text{s}$ and practically constant with $h$ (and $n_{h}$). As with the number of poles, the small variations in the timings with $h$ are likely due to different scaling of the matrices ${\mathsf{A}}$ and ${\mathsf{M}}$. We note that in our experiments $32\leq n_h\leq 1024$. Moreover, since $\Lambda_h$ is in our experiments constructed from $\mathbb{P}^{\text{disc}}_1$ we apply the estimate \eqref{diag}.
In \Cref{fig:darcy_stokes_ra_perf}, we finally plot the dependence of the solution time of the preconditioned MinRes solver on the problem size. Indeed we observe that the solver is of linear complexity. In particular, application of rational approximation of $S^{-1}$ in our implementation requires $\mathcal{O}(n_h)$ operations. We remark that here the solvers for the shifted Laplacian problems are realized by the conjugate gradient method with AMG as a preconditioner. Let us also recall that the remaining blocks of the preconditioner are realized by LU, where the setup cost is not included in our timings. However, LU does not define efficient preconditioners for the respective blocks. Here instead, multilevel methods could provide order optimality, and this is a topic of current and future research.
\begin{figure}
\centering
\includegraphics[width=0.325\textwidth, height=0.15\textheight]{./figures/ds_ra_poles.pdf}
\includegraphics[width=0.325\textwidth, height=0.15\textheight]{./figures/ds_ra_ra.pdf}
\includegraphics[width=0.325\textwidth, height=0.15\textheight]{./figures/ds_ra_ksp.pdf}
\vspace{-10pt}
\caption{
Dependence of the number of poles (left), the setup time of RA (center) and runtime of the MinRes solver (right) on mesh size $h$ and RA tolerance $\epsilon_{\text{RA}}$. Parameters in the Darcy--Stokes problem are fixed at $K=10^{-6}$, $\mu=10^{-2}$ and $\alpha=1$. Setup as in \Cref{fig:darcy_stokes_param_robust} is used with discretization by $\mathbb{BDM}_1$-$\mathbb{BDM}_1$-$\mathbb{P}^{\text{disc}}_{0}$-$\mathbb{P}^{\text{disc}}_{0}$-$\mathbb{P}^{\text{disc}}_1$ elements.
}
\label{fig:darcy_stokes_ra_perf}
\end{figure}
\section{Conclusions}\label{sec:conclusions}
We have demonstrated that RA provides order optimal preconditioners for sums of fractional powers of SPD operators and can thus be utilized to construct parameter robust and order optimal preconditioners for multiphysics problems. The results are of practical interest for constructing efficient preconditioning on interfaces in models for which fractional weighted Sobolev spaces are the natural setting for the resulting differential operators, for example, when flow interacts with porous media. The techniques presented here could aid the numerical simulations in a wide range of biology, medicine, and engineering applications.
\section{Acknowledgments}
A. Budi\v{s}a, M. Kuchta, K.~A. Mardal, and L.~Zikatanov acknowledge the financial support from the SciML project funded by the Norwegian Research Council grant 102155. The work of X. Hu is partially supported by the National Science Foundation under grant DMS-2208267. M.~Kuchta acknowledges support from the Norwegian Research Council grant 303362. The work of L.~Zikatanov is supported in part by the U.S.--Norway Fulbright Foundation and the U.S. National Science Foundation grant DMS-2208249.
\bibliographystyle{splncs04}
|
\section{Introduction}
The emerging field of graph signal processing (GSP) deals with processing data indexed by general graphs with concepts and techniques inspired by traditional digital signal processing (DSP). These techniques include graph Fourier transforms (GFT), graph filter designs \cite{8347162,Shuman_Ortega_2013,Isufi_Leus2017,6319640}, and the sampling and recovery of graph signals \cite{9244650,6854325,7352352}. Most of these techniques have been used for various tasks under a linear measurement model. However, modern networks are often large and complex, contain heterogeneous large datasets, and are characterized by nonlinear models \cite{9343697,8347160}, and as a result, graph signals are often difficult to recover in these networks.
Examples of applications with such networks include brain network connectivity \cite{shen2016nonlinear}, environmental monitoring \cite{8496842}, and power flow equations in power systems \cite{drayer2018detection,Grotas2019,shaked2021identification}.
Hence, the development of GSP methods for the estimation of graph signals in nonlinear models has considerable practical significance.
The mean-squared-error (MSE) is one of the most commonly-used criteria
of accuracy for estimation and reconstruction purposes. In Bayesian estimation,
the optimal minimum MSE (MMSE) estimator
usually lacks a closed-form expression in nonlinear models and is often computationally intractable.
Therefore, the linear MMSE (LMMSE) estimator and other
low-complexity estimators (e.g. \cite{Edfors_Borjesson_1996,dd,Berman_letter}) are widely used in practice.
The linear
GSP-LMMSE estimator, which minimizes the MSE among estimators that are represented as an output of a graph filter, has been suggested in \cite{kroizer2021bayesian} for the recovery of graph signals, and its properties are discussed. In particular, it has been shown that
the GSP-LMMSE estimator has low computational complexity, and the ability to adapt to changes in the graph topology. In addition, it has been extended to a widely-linear estimation of complex-valued signals in \cite{Amar_Routtenberg}.
However, linear estimators
fail to take into account the nonlinearity of the measurement model, which
may lead to degraded performance compared with nonlinear methods.
Consequently, developing nonlinear estimation methods that take advantage of the graph structure and GSP theory have the potential to significantly improve the estimation performance.
Nonlinear methods can significantly outperforms linear estimators in terms of MSE.
For example,
the superiority of the nonlinear maximum
{\em a-posteriori} probability (MAP) estimator compared with the LMMSE estimator in nonlinear filtering is well known \cite{Fatemi_Svensson_Morelande2012}.
In order to implement
nonlinear methods, such as
the MAP estimator, in nonlinear settings, iterative methods are used such as the widely-employed Gauss-Newton method.
However, the Gauss-Newton method is sensitive to initialization, and may converge to a local minima or even diverge \cite{Blaschke1997,convergence_GN}.
Thus, integrating the graph information, e.g. by using graph filters, has the potential to enhance the robustness of iterative implementations of the MAP estimator.
Graph filters have been used for many signal processing tasks, such as denoising \cite{7032244,ZHANG20083328}, classification \cite{6778068}, and anomaly detection \cite{drayer2018detection}.
Model-based recovery of graph signals by GSP filters for linear models was treated in \cite{7117446,7032244,Isufi_Leus2017,7891646}.
Nonlinear graph filters were considered in \cite{8496842}, but they require higher-order statistics that are not completely specified in the general nonlinear case. Recently,
graph neural network approaches were considered in \cite{Isufi_Ribeiro,gama2020graphs}.
In addition,
a Gauss-Newton unrolled neural network method was developed in \cite{Yang_Giannakis_Sun2020} for the application of power system state estimation (PSSE).
However, data-based methods necessitate extensive stationary training sets, and do not necessarily utilize the model information. Using the model at hand, one can design estimators with improved performance in terms of MSE, interpretability,
robustness, complexity, and flexibility.
Fitting graph-based models to given data was considered in \cite{7763882,Hua_Sayed_2020,confPaper}. However, model-fitting approaches aim to minimize the modeling error, and in general have significantly lower performance than estimators that minimize the estimation error directly \cite{lecture_notes}.
In this paper, we consider the nonlinear estimation of random graph signals with a nonlinear observation model and a Gaussian distribution.
First, we present the MAP estimator, as well as its implementations in the vertex and in the graph-frequency domains by using the Gauss-Newton method.
Then, we propose two new GSP estimators: 1) the elementwise graph-frequency-domain MAP (eGFD-MAP) estimator; and 2) the GSP-MAP estimator. The eGFD-MAP estimator updates the coordinates of the estimator in the graph-frequency domain separately,
and has a significantly lower computational complexity than the MAP estimator. The GSP-MAP estimator is based on optimizing the graph filters at each iteration of the Gauss-Newton algorithm.
We discuss the conditions under which the eGFD-MAP and the GSP-MAP estimators coincide with the MAP estimator. In particular, we show that for
models with measurement functions that have orthogonal graph frequencies, i.e. separable in the graph-frequency domain,
the eGFD-MAP and the GSP-MAP estimators coincide with the MAP estimator.
We perform numerical simulations for: 1) synthetic data with orthogonal graph frequencies; and 2) power system state estimation (PSSE).
For the first case, it is shown that the eGFD-MAP and the GSP-MAP estimators achieve the same MSE as the MAP estimator, while the eGFD-MAP estimator significantly reduces the computational complexity, especially in large networks where the MAP and the GSP-MAP estimators are intractable.
For the PSSE simulations,
it is shown that the eGFD-MAP and the GSP-MAP estimators almost achieve the performance of the MAP estimator, and significantly outperform
the linear estimator.
Moreover, the proposed eGFD-MAP and GSP-MAP estimators are more robust to perturbed initialization than the MAP estimator.
The rest of this paper is organized as follows. In Section \ref{background_sec} we introduce the basics of GSP required for this paper.
In Section \ref{MAP_objectives_sec}, we formulate the estimation problem, present the MAP
estimation approaches (in the vertex and graph-frequency domains), and describe the Gauss-Newton implementation of the MAP
estimators.
In Section \ref{eGFD_MAP_and_GSP_MAP}, we develop the eGFD-MAP and GSP-MAP estimators and discuss their properties.
Simulations are presented in Section \ref{simulation}.
Finally,
conclusions are outlined in Section \ref{conclusion}.
In the following, we denote vectors by boldface lowercase letters and matrices by boldface uppercase letters.
The operators $(\cdot)^T$ and $(\cdot)^{-1}$ denote the transpose and inverse, respectively. The vector ${\bf{1}}$ denotes a vector of all ones, and $\circ$ denotes the Hadamard product. The $m$th element of the vector ${\bf{a}}$ is denoted by $a_m$ or $[{\bf{a}}]_m$. The $(m,q)$th element of the matrix ${\bf{A}}$ is written as $A_{m,q}$ or $[{\bf{A}}]_{m,q}$.
For a vector ${\bf{a}}$, ${\text{diag}}({\bf{a}})$ is a diagonal matrix whose $i$th diagonal entry is $a_{i}$; when applied to a matrix, ${\text{diag}}({\bf{A}})$ is a vector collecting the diagonal elements of ${\bf{A}}$. In addition, ${\text{ddiag}}({\bf{A}})={\text{diag}}({\text{diag}}({\bf{A}}))$ is the diagonal matrix whose entries on the diagonal are those
of ${\bf{A}}$.
The identity matrix and the zero vector are written as ${\bf{I}}$ and ${\bf{0}}$, respectively.
The cross-covariance matrix of the vectors ${\bf{a}}$ and ${\bf{b}}$ is denoted by ${\bf{C}}_{{\bf{a}}{\bf{b}}}\triangleq \EX[({\bf{a}}-\EX[{\bf{a}}]) ({\bf{b}}-\EX[{\bf{b}}])^T]$.
The Jacobian matrix of a vector function ${\bf{g}}({\bf{x}})$, $\nabla_{{\bf{x}}}{\bf{g}}({\bf{x}})$, is a matrix in $\mathbb{R}^{K\times M}$, with the $(k,m)$th element equal to $\frac{\partial g_k}{\partial x_m}$, where ${\bf{g}}=\left[g_1,\ldots,g_K\right]^T$ and ${\bf{x}}=\left[x_1,\ldots,x_M\right]^T$.
We use the notation $\nabla_{\bf{x}}^T {\bf{g}}({\bf{x}}) \stackrel{\triangle}{=} (\nabla_{\bf{x}} {\bf{g}}({\bf{x}}))^T$.
For a scalar function, $g({\bf{x}})$, we denote
$\nabla_{\bf{x}}^2 g({\bf{x}}) \stackrel{\triangle}{=}\nabla_{\bf{x}} \nabla_{\bf{x}}^T g({\bf{x}})$. Finally,
$||\cdot||$ denotes the Euclidean norm.
\section{Background: Graph Signal Processing (GSP)}
\label{background_sec}
Consider an undirected, connected, weighted graph ${\pazocal{G}}({\pazocal{V}},\xi,{\bf{W}})$, where $\pazocal{V}$ and $\xi$ are sets of vertices and edges, respectively.
The matrix ${\bf{W}} \in \mathbb{R}^{N \times N}$ is the nonnegative weighted adjacency matrix of the graph, where $N \stackrel{\triangle}{=} |\pazocal{V}|$ is the number of vertices in the graph. If there is an edge $(i, j) \in \xi$ connecting vertices $i$ and $j$, the entry ${\bf{W}}_{i,j}$ represents the weight of the edge; otherwise, ${\bf{W}}_{i,j} = 0$. The Laplacian matrix of this graph is defined by
$
{\bf{L}} \stackrel{\triangle}{=} \text{diag}\left({\bf{W}}{\bf{1}}\right) - {\bf{W}}$.
The Laplacian matrix, ${\bf{L}}$, is a real and positive semidefinite matrix; thus, its associated eigenvalue decomposition is given by
\begin{equation}
{\bf{L}} = {\bf{V}}{\bf{\Lambda}} {\bf{V}}^{T}, \label{SVD_new_eq}
\end{equation}
where ${\bf{\Lambda}}$ is a diagonal matrix consisting of the eigenvalues of
${\bf{L}}$, $0= \lambda_1 < \lambda_2 \leq \ldots \leq \lambda_N $, ${\bf{V}}$ is a matrix whose $n$th column, ${\bf{v}}_N$, is the eigenvector of ${\bf{L}}$ that is associated with $\lambda_n$, and ${\bf{V}}^{T}={\bf{V}}^{-1}$.
In this paper, a {\em{graph signal}} is
an $N$-dimensional vector, ${\bf{a}}$, that assigns a scalar value to each vertex, i.e. each entry $a_n$ denotes the signal value at vertex $n$, for $n=1,\ldots, N$.
The GFT of the graph signal ${\bf{a}}$ is defined as \cite{Shuman_Ortega_2013}
\begin{equation}
\label{GFT}
\tilde{{\bf{a}}} \triangleq {\bf{V}}^{T}{\bf{a}}.
\end{equation}
Similarly, the inverse GFT (IGFT) of $\tilde{{\bf{a}}}$ is given by ${\bf{V}}\tilde{{\bf{a}}}$.
A graph signal is a graph-bandlimited signal with cutoff graph frequency $N_s$ if it satisfies \cite{8347162}
\begin{equation} \label{bandlimited_def}
\tilde{a}_n =0,~ n =N_s+1,\dots,N.
\end{equation}
Graph filters are useful tools for various GSP tasks.
Linear and shift-invariant graph filters with respect to (w.r.t.) the graph shift operator (GSO) play essential roles in GSP.
A graph filter is a function $f(\cdot)$ applied to a GSO, where here we use the Laplacian ${\bf{L}}$ as the GSO, which allows the following eigendecomposition \cite{8347162}:
\begin{equation} \label{laplacian_graph_filter}
f({\bf{L}})= {\bf{V}} f({\bf{\Lambda}}){\bf{V}}^T,
\end{equation}
where $f({\bf{\Lambda}})$ is a diagonal matrix.
That is, $f(\lambda_n)$ is the graph frequency response of the filter at graph frequency $\lambda_n$, $n=1,\dots,N$,
and $f({\bf{L}})$ is diagonalized by the eigenvector matrix of ${\bf{L}}$, ${\bf{V}}$.
We assume that the graph filter, $f(\cdot)$, is a well-defined function on the spectrum of ${\bf{L}}$, $\{\lambda_1,\ldots,\lambda_N\}$.
\section{MAP estimator of graph signals}
\label{MAP_objectives_sec}
In this section we formulate the MAP estimator for the problem of estimating a random graph signal by observing its noisy nonlinear function. First, the measurement model and assumption are introduced in Subsection \ref{model_sub_sec}.
Then, we derive the MAP estimator in both the vertex and the graph-frequency domains in Subsection \ref{MAP_subsec}. In Subsection \ref{iterative_subsection}, we show the implementation of the MAP estimator by the Gauss-Newton method.
\subsection{Model}
\label{model_sub_sec}
Consider the problem of recovering a random input graph signal, ${\bf{x}}\in\mathbb{R}^N$, based on the following nonlinear measurement model:
\begin{equation} \label{Model}
{\bf{y}} = {\bf{g}}({\bf{L}},{\bf{x}}) + {\bf{w}},
\end{equation}
where the measurement function, ${\bf{g}}:\mathbb{R}^{N \times N} \times{\mathbb{R}}^N \rightarrow {\mathbb{R}}^N$, and the Laplacian matrix, ${\bf{L}}$, which represents the influence of the graph topology, are assumed to be known.
In addition, it is assumed that
$ {\bf{g}}({\bf{L}},{\bf{x}})$ is
continuously differentiable w.r.t. ${\bf{x}}$.
We assume that
the graph input signal, ${\bf{x}}$, is a Gaussian vector with mean ${\mbox{\boldmath $\mu$}}_{\bf{x}}$ and covariance matrix ${{\bf{C}}}_{{\bf{x}}\xvec}$. The noise, ${\bf{w}}$, is assumed to be a zero-mean Gaussian vector with covariance matrix ${\bf{C}}_{{\bf{w}}\wvec}$. Finally, we assume that ${\bf{w}}$ and ${\bf{x}}$ are independent.
The nonlinear measurement model in \eqref{Model} is adequate for many applications within networked data and the Internet-of Things (IoT)
\cite{sahu2018communication,Sahu_Kar_Moura_Poor2016,kroizer2021bayesian}. In particular, this
model arises in PSSE, which is discussed in Section \ref{simulation}.
\subsection{MAP estimator}
\label{MAP_subsec}
In the following, we develop the MAP estimator for the measurement model in \eqref{Model}.
The MAP estimator of ${\bf{x}}$ from ${\bf{y}}$ is given by
\begin{eqnarray}
\label{MAP_estimator1}
\hat{{\bf{x}}}= \argmax_{{{\bf{x}}} \in \mathbb{R}^N}
f({\bf{x}}|{\bf{y}})
=\argmax_{{{\bf{x}}} \in \mathbb{R}^N}
\log f({\bf{y}}|{\bf{x}})+\log f({\bf{x}}),
\end{eqnarray}
where the second equality is obtained by
using Bayes’s rule,
applying the monotonically-increasing logarithm function and removing constant terms w.r.t. ${\bf{x}}$.
By substituting the considered model, in which ${\bf{y}}$ given ${\bf{x}}$ is also a Gaussian vector with mean ${\bf{g}}({\bf{L}},{\bf{x}})$ and covariance ${{\bf{C}}}_{{\bf{w}}\wvec}$, in
\eqref{MAP_estimator1},
we obtain
\begin{eqnarray}
\label{MAP_estimator2}
\hat{{\bf{x}}}=\argmin_{{{\bf{x}}} \in \mathbb{R}^N} Q({\bf{x}}),
\end{eqnarray}
where, after removing constant terms w.r.t. ${\bf{x}}$, we have
\begin{eqnarray}
\label{Q_def}
Q({\bf{x}})\stackrel{\triangle}{=}
\frac{1}{2}({\bf{x}}-{\mbox{\boldmath $\mu$}}_{\bf{x}})^{T}{{\bf{C}}}_{{\bf{x}}\xvec}^{-1}({\bf{x}}-{\mbox{\boldmath $\mu$}}_{\bf{x}})\hspace{1.75cm}\nonumber\\ + \frac{1}{2}({\bf{y}}-{\bf{g}}({\bf{L}},{\bf{x}}))^{T}{{\bf{C}}}_{{\bf{w}}\wvec}^{-1}({\bf{y}}-{\bf{g}}({\bf{L}},{\bf{x}})).
\end{eqnarray}
The left term in the objective function on the r.h.s. of \eqref{Q_def} corresponds to
the prior, and the right term corresponds to the noisy measurement model.
This objective function can be minimized by iterative algorithms to approximate the MAP estimator, as described in Section \ref{iterative_subsection}.
The objective function in \eqref{Q_def} discards information about the relationship between the graph signal and its underlying graph structure. As a result, it is less robust to perturbations of the initialization that are due to changes in the graph topology.
In addition, the MAP estimator only uses the measurement function and does not exploit additional GSP information on the graph signal, such as smoothness or graph-bandlimitness, that can be utilized to improve estimation performance.
In general, representing data in the graph-frequency domain can yield substantial data reduction, and minimize the computational requirements and memory use.
Thus, as a first step, we suggest to transform $Q({\bf{x}})$ into its graph-frequency domain representation (as a function of $\tilde{{\bf{x}}}$):
\begin{eqnarray}
\label{Q_def_GSP}
Q_{freq}(\tilde{{\bf{x}}})\stackrel{\triangle}{=}
\frac{1}{2}(\tilde{{\bf{x}}}-\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})^{T}{{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1}(\tilde{{\bf{x}}}-\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\hspace{1.5cm}\nonumber\\ + \frac{1}{2}(\tilde{{\bf{y}}}-\tilde{{\bf{g}}}({\bf{L}},{\bf{V}}\tilde{\bf{x}}))^{T}{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})),
\end{eqnarray}
where $\tilde{\bf{y}}$, $\tilde{\bf{x}}$, $\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}}$, and $\tilde{\bf{g}}$ are the GFT representations of ${\bf{y}}$, ${\bf{x}}$, ${{\mbox{\boldmath $\mu$}}}_{\bf{x}}$, and ${\bf{g}}$, respectively, as defined in \eqref{GFT}, such that
$\tilde{{\bf{g}}}({\bf{L}},{\bf{V}}\tilde{{\bf{x}}})={\bf{V}}^T{\bf{g}}({\bf{L}},{\bf{x}})$. In addition, we use the fact that
\begin{equation*}
{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}\stackrel{\triangle}{=} {\rm{E}}\left[{\bf{V}}^{T}({\bf{x}}-{\mbox{\boldmath $\mu$}}_{{\bf{x}}})({\bf{x}}-{\mbox{\boldmath $\mu$}}_{{\bf{x}}})^{T}{\bf{V}}\right]={\bf{V}}^{T}{{\bf{C}}}_{{\bf{x}}\xvec}{\bf{V}}
\end{equation*}
and similarly ${{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}={\bf{V}}^{T}{{\bf{C}}}_{{\bf{w}}\wvec}{\bf{V}}$. It can be verified that the r.h.s. of \eqref{Q_def} and the r.h.s. of \eqref{Q_def_GSP} are identical. Thus, these two objective functions will lead to the same final estimator,
where in \eqref{Q_def} the update is in the vertex domain (${\bf{x}}$), and in \eqref{Q_def_GSP} the update is in the graph-frequency domain ($\tilde{{\bf{x}}}$).
It should be noted that the efficiency and the convergence rate of the specific implementation in each domain may be different.
For example, in our simulations (see Subsection \ref{simulation}), we realized that implementing the MAP estimator in the graph-frequency domain is much faster.
\subsection{Implementation by Gauss-Newton method}
\label{iterative_subsection}
Since ${\bf{g}}({\bf{L}},{\bf{x}})$ is a nonlinear and nonconvex function, direct minimization of the objective functions in
\eqref{Q_def}, \eqref{Q_def_GSP}, or \eqref{Q_def_GSP_incremental} is intractable.
Numerous algorithms have been proposed to minimize nonconvex objectives. Here, we implement the
Gauss-Newton method, which is widely
employed to solve
nonlinear weighted least squares (WLS) problems and to find the MAP estimator.
The Gauss-Newton method has the desirable property of a quadratic rate of convergence under suitable assumptions \cite{convergence_GN}.
{\bf{Initialization:}} For all the algorithms described in this and in the following section, the estimators can be initialized by
the prior mean, i.e. $\hat{{\bf{x}}}^{(0)}={\mbox{\boldmath $\mu$}}_{\bf{x}}$ and $\hat{\tilde{{\bf{x}}}}^{(0)}=\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}}$, where $\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}}\stackrel{\triangle}{=} {\bf{V}}^{T}{\mbox{\boldmath $\mu$}}_{\bf{x}}$.
Alternatively, if training data is available or can be generated, it is possible to initialize the estimators by the LMMSE or the GSP-LMMSE \cite{kroizer2021bayesian} estimators.
In addition,
the stopping condition is attained when two successive MAP estimates, $\hat{{\bf{x}}}^{(t+1)}$
and $\hat{{\bf{x}}}^{(t)}$, are sufficiently close, where $t$ is the iteration index.
\subsubsection{Conventional MAP estimator}
\label{subsection_MAP}
In order to locally minimize the MAP cost function in \eqref{Q_def},
we use the
Gauss-Newton method, which relies on Taylor's expansion to
linearize the measurement function ${\bf{g}}({\bf{L}},{\bf{x}})$ and iteratively updates the estimators until convergence (see Sec. 1.5.1 in \cite{BERTSEKAS_1999}). Specifically,
under the assumption that the given estimator $\hat{{{\bf{x}}}}^{(t)}$ is close enough to ${{\bf{x}}}$,
the linear first-order approximation of ${\bf{g}}({\bf{L}},{\bf{x}})$ is
\begin{equation}
\label{g_approx}
{\bf{g}}({\bf{L}},{\bf{x}})\approx
{\bf{g}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)}) + {\bf{G}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)})({\bf{x}}-\hat{{{\bf{x}}}}^{(t)}),
\end{equation}
where
\begin{equation}
\label{GGG}
{\bf{G}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)})\stackrel{\triangle}{=} \left.\nabla_{\bf{x}} {\bf{g}}({\bf{L}},{\bf{x}})\right|_{{\bf{x}}=\hat{{{\bf{x}}}}^{(t)}}
\end{equation}
is the $N\times N$ Jacobian matrix of the measurement function ${\bf{g}}({\bf{L}},{\bf{x}})$ evaluated at
$\hat{{{\bf{x}}}}^{(t)}$.
By substituting the approximation from \eqref{g_approx} in \eqref{Q_def},
we obtain the approximated linearized objective function:
\begin{eqnarray}
\label{Q_def_approx}
Q({\bf{x}})\approx Q_{lin}({\bf{x}},\hat{{{\bf{x}}}}^{(t)})
\hspace{4.75cm}\nonumber\\\stackrel{\triangle}{=}
\frac{1}{2}({\bf{x}}-{\mbox{\boldmath $\mu$}}_{\bf{x}})^{T}{{\bf{C}}}_{{\bf{x}}\xvec}^{-1}({\bf{x}}-{\mbox{\boldmath $\mu$}}_{\bf{x}})\hspace{3.5cm}\nonumber\\ + \frac{1}{2}({\bf{y}}-{\bf{g}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)}) - {\bf{G}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)})({\bf{x}}-\hat{{{\bf{x}}}}^{(t)}))^{T}{{\bf{C}}}_{{\bf{w}}\wvec}^{-1}
\hspace{0.275cm}\nonumber\\\times
({\bf{y}}-{\bf{g}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)}) - {\bf{G}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)})({\bf{x}}-\hat{{{\bf{x}}}}^{(t)})).\hspace{1.2cm}
\end{eqnarray}
The
Gauss-Newton method finds the next iterate $t+1$ by minimizing $ Q_{lin}({\bf{x}},\hat{{{\bf{x}}}}^{(t)})$ w.r.t. ${\bf{x}}$, which results in
\begin{eqnarray}\label{iteration_Newton2}
\hat{{\bf{x}}}^{(t+1)}=\arg\min_{{\bf{x}}\in\mathbb{R}^N} Q_{lin}({\bf{x}},\hat{{{\bf{x}}}}^{(t)})\hspace{3.65cm}\nonumber\\=
\hat{{\bf{x}}}^{(t)}-\alpha^{(t)}\left ({{\bf{C}}}_{{\bf{x}}\xvec}^{-1} + {\bf{G}}^{T}({\bf{L}},\hat{{\bf{x}}}^{(t)}){{\bf{C}}}_{{\bf{w}}\wvec}^{-1}{\bf{G}}({\bf{L}},\hat{{\bf{x}}}^{(t)})\right)^{-1}
\nonumber\\
\times\Big({{\bf{C}}}_{{\bf{x}}\xvec}^{-1} (\hat{{\bf{x}}}^{(t)}- {\mbox{\boldmath $\mu$}}_{\bf{x}})\hspace{4.5cm}\nonumber\\
- {\bf{G}}^{T}({\bf{L}},\hat{{\bf{x}}}^{(t)}){{\bf{C}}}_{{\bf{w}}\wvec}^{-1}({\bf{y}}-{\bf{g}}({\bf{L}},\hat{{\bf{x}}}^{(t)}))\Big),\hspace{1.85cm}
\end{eqnarray}
with $\alpha^{(t)}=1$.
In practice, wisely setting the step size,
$\alpha^{(t)}\in(0,1]$, can improve the convergence rate \cite{Bell_Cathey_1993}.
In this paper, we compute the step size by a backtracking
line search (Algorithm \ref{line_search_stages}). In this strategy, we iteratively reduce the step size, $\alpha^{(t)}$, until $\hat{{{\bf{x}}}}^{(t+1)}$ from \eqref{iteration_Newton2} with the tested step size that satisfies
\begin{equation}\label{line_search1}
Q(\hat{{{\bf{x}}}}^{(t)})-Q(\hat{{{\bf{x}}}}^{(t+1)})>\Delta |Q(\hat{{{\bf{x}}}}^{(t)})|,
\end{equation}
where $\Delta\in\mathbb{R}$ satisfies $\Delta<<1$.
The MAP-estimator algorithm is summarized in Algorithm \ref{algorithm_MAP}.
\begin{algorithm}[hbt]
\SetAlgoLined
\KwInput{
\begin{itemize}
\item The function ${\bf{g}}({\bf{L}},{\bf{x}})$
\item The mean and covariance matrices ${\mbox{\boldmath $\mu$}}_{{{\bf{x}}}}$, ${{\bf{C}}}_{{{\bf{x}}}{{\bf{x}}}}$, and ${{\bf{C}}}_{{\bf{w}}\wvec}$
\item Initial step size $\alpha_{0}$, $\gamma$ and $\Delta$
\item The tolerance $\delta$
\end{itemize}}
{\textbf{Algorithm Steps:}}
\begin{enumerate}
\setlength\itemsep{0.2em}
\item Initialization: ${\hat{{\bf{x}}}}^{(0)}$
\item Compute: ${\bf{g}}({\bf{L}},\hat{{\bf{x}}}^{(t)})$ and ${\bf{G}}({\bf{L}},\hat{{\bf{x}}}^{(t)})$ from \eqref{GGG}
\item
Choose step size $\alpha^{(t)}$ such that \eqref{line_search1} holds using Algorithm \ref{line_search_stages} with the cost $Q$ from \eqref{Q_def}
\item Compute
${{{\bf{x}}}}^{(t+1)}$ from \eqref{iteration_Newton2} with step size $\alpha^{(t)}$
\item Stopping condition: $||\hat{{\bf{x}}}^{(t+1)}-\hat{{\bf{x}}}^{(t)}||<\delta$
\end{enumerate}
\KwOutput{MAP estimator
$\hat{{\bf{x}}} = {\tilde{{\bf{x}}}}^{(t+1)}$}
\caption{MAP estimator by Gauss-Newton}
\label{algorithm_MAP}
\end{algorithm}
\begin{algorithm}[hbt]
\SetAlgoLined
\KwInput{
\begin{itemize}
\item Current estimator $\hat{{\bf{x}}}^{(t)}$
\item Initial step size $\alpha_{0}\in (0,1]$, tuning parameter $\gamma\in(0,1)$
\item Cost function $Q$
\end{itemize}}
{\textbf{Algorithm Steps:}}
\begin{enumerate}
\item Compute: $Q(\hat{{{\bf{x}}}}^{(t)})$
\item
\For{$k=0,1,\ldots K_{\max}$}{
\begin{itemize} \item
Compute
${\hat{{\bf{x}}}}^{(t+1)}$
with $\alpha^{(t)}=\alpha_{k}$
\item Compute $Q(\hat{{{\bf{x}}}}^{(t+1)})$
\end{itemize}
\If{\eqref{line_search1} does not hold}
{\If{ $k=K_{\max}$}{$\alpha_{K_{\max}}=0$}
\Else{$\alpha_{k+1}=\gamma\alpha_{k}$}
}\Else{\text{break}}
}
\end{enumerate}
\KwOutput{ Step size: $\alpha^{(t)} =\alpha_k$ }
\caption{Backtracking line search}
\label{line_search_stages}
\end{algorithm}
\subsubsection{MAP estimator in the graph-frequency domain} \label{GSP_estimator_section}
In this subsection,
we evaluate the graph-frequency-domain update of the MAP
estimator.
Similar to the derivation of \eqref{iteration_Newton2}, the minimization of
$Q_{freq}(\tilde{{\bf{x}}})$ from \eqref{Q_def_GSP} w.r.t. $\tilde{{\bf{x}}}$ by the Gauss-Newton method results in the following update equation:
\begin{eqnarray}\label{iteration_Newton2_GSP}
\hat{\tilde{{\bf{x}}}}^{(t+1)}=
\hat{\tilde{{\bf{x}}}}^{(t)}\hspace{6.25cm}\nonumber\\-\alpha^{(t)}
\left ({{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1} + \tilde{{\bf{G}}}^{T}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}){{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})\right)^{-1}
\nonumber\\\times
\left({{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1} (\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\hspace{4.5cm} \right.
\nonumber\\
\left.- \tilde{\bf{G}}^{T}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}){{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}))\right),\hspace{1.45cm}
\end{eqnarray}
where
\begin{equation}
\label{GGG_tilde}
\tilde{\bf{G}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})\stackrel{\triangle}{=} \nabla_{\tilde{{\bf{x}}}} \tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{{\bf{x}}}) = {\bf{V}}^{T}{\bf{G}}({\bf{L}},{\bf{x}}){\bf{V}}.
\end{equation}
It can be verified that by multiplying \eqref{iteration_Newton2} by ${\bf{V}}^T$ from the left, we obtain the iteration in the graph-frequency domain in \eqref{iteration_Newton2_GSP}.
In addition, it is known that the convergence of the Gauss-Newton method is invariant under affine transformations of the domain \cite{deuflhard1979affine}.
An advantage of the update in \eqref{iteration_Newton2_GSP} compared with \eqref{iteration_Newton2} is for cases where ${\bf{x}}$ is a
graph bandlimited signal with a cutoff graph frequency $N_s$. In this case, we substitute $[\hat{\tilde{{\bf{x}}}}^{(t)}]_n=0$ $\forall n>N_s$ and update only
the first $N_s$ elements of $\hat{\tilde{{\bf{x}}}}^{(t+1)}$ in
\eqref{iteration_Newton2_GSP} at each step.
The MAP estimator in the graph frequency domain can be implemented using Algorithm \ref{algorithm_MAP}, with the cost $Q$ from \eqref{Q_def_GSP} in step 3) and with \eqref{iteration_Newton2_GSP} instead of \eqref{iteration_Newton2} in step 4) of the algorithm.
In terms of computational complexity, implementing the MAP estimator by \eqref{iteration_Newton2} and \eqref{iteration_Newton2_GSP} requires the computation of the inverse of $N\times N$ matrices $ {{\bf{C}}}_{{\bf{x}}\xvec}^{-1} + {\bf{G}}^{T}({\bf{L}},\hat{{\bf{x}}}^{(t)}){{\bf{C}}}_{{\bf{w}}\wvec}^{-1}{\bf{G}}({\bf{L}},\hat{{\bf{x}}}^{(t)})$ in \eqref{iteration_Newton2},
or ${{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1} + \tilde{{\bf{G}}}^{T}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}){{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})$ in \eqref{iteration_Newton2_GSP} in addition to multiplications of $N\times N$ matrices
at each iteration, which leads to high computational complexity.
Furthermore, due to the nonconvexity of ${\bf{g}}({\bf{L}},{\bf{x}})$ and the quadratic loss function, the Gauss-Newton method is sensitive to initialization and may diverge.
These challenges inhibit its use for real-time estimation in large-scale networks.
Moreover, these methods do not utilize the information about the underlying graph structure, e.g. for the initialization approach.
Thus, they are less robust to changes and misspecification in the graph topology, e.g. in the graph connectivity.
Finally, the estimators in \eqref{iteration_Newton2} and \eqref{iteration_Newton2_GSP} ignore the GSP properties and do not exploit additional information on the graph signal, such as smoothness or graph-bandlimitness, that can improve estimation performance.
\section{eGFD-MAP and GSP-MAP estimators}\label{eGFD_MAP_and_GSP_MAP}
In this section, we propose two new estimators that integrate the graph structure: the eGFD-MAP in Subsection \ref{eGFD-MAP_EST} and the GSP-MAP in Subsection \ref{MAP_GSP_section}.
In Subsection \ref{separately_model_section}, we present special cases and discuss the conditions under which the proposed eGFD-MAP and GSP-MAP estimators coincide with the MAP estimator. Finally, in Subsection \ref{computational_complexity} we compare the computational complexity of the different estimators.
\subsection{eGFD-MAP estimator}\label{eGFD-MAP_EST}
In the nonlinear case, numerical optimization methods are usually used to minimize \eqref{Q_def} or \eqref{Q_def_GSP}.
However, when $N$ is large, the minimization problem is high-dimensional, with high computational complexity and memory demands.
In order to reduce the complexity and
accelerate the convergence rate of the iterative optimization algorithms, we
propose here the eGFD-MAP estimator, which is based on two steps.
In the
first step, we replace the MAP objective function from \eqref{Q_def_GSP} by
the following objective function
in the graph-frequency domain:
\begin{eqnarray}
\label{Q_def_GSP_incremental}
Q_{freq}^{(d)}(\tilde{{\bf{x}}})\stackrel{\triangle}{=}
\frac{1}{2}(\tilde{{\bf{x}}}-\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})^{T}{{\bf{D}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}(\tilde{{\bf{x}}}-\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\hspace{2.25cm}\nonumber\\ + \frac{1}{2}(\tilde{{\bf{y}}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}}))^{T}{{\bf{D}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{(\text{inv})}(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}}))\nonumber\\
=\frac{1}{2}\sum_{n=1}^N
(\tilde{x}_n-[\tilde{\mu}_{\bf{x}}]_n)^2[{{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1}]_{n,n}\hspace{2.1cm}\nonumber\\ + \frac{1}{2}\sum_{n=1}^N(\tilde{y}_n-
[\tilde{{\bf{g}}}({\bf{L}},{\bf{V}}\tilde{{\bf{x}}})]_n
)^2[{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}]_{n,n},\hspace{0.5cm}
\end{eqnarray}
where
\begin{equation}
\label{inv_Dx_define}
{\bf{D}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}\stackrel{\triangle}{=} {\text{ddiag}}({{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1})
\end{equation}
and
\begin{equation}
\label{inv_Dw_define}
{\bf{D}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}\stackrel{\triangle}{=} {\text{ddiag}}({{\bf{C}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{-1}).
\end{equation}
Similar to the derivations of
\eqref{iteration_Newton2} and \eqref{iteration_Newton2_GSP}, the minimization of $Q_{freq}^{(d)}(\tilde{\bf{x}})$ from (10) w.r.t. $\tilde{\bf{x}}$ by the Gauss-Newton method results in the following update equation:
\begin{eqnarray}\label{27_old25_before_approx}
\hat{\tilde{{\bf{x}}}}^{(t+1)}
=\hat{\tilde{{\bf{x}}}}^{(t)}\hspace{6.25cm}\nonumber\\-\alpha^{(t)}\bigg({{\bf{D}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})} + \tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})^{T}{{\bf{D}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})\bigg)^{-1}\nonumber\\\times\bigg({\bf{D}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}(\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\hspace{4.75cm}
\nonumber\\
-\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})^{T}{\bf{D}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{(\text{inv})}(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}))\bigg).\hspace{1.5cm}
\end{eqnarray}
That is, \eqref{Q_def_GSP_incremental} is obtained by replacing the inverse of the full covariance matrices ${{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}$ and ${{\bf{C}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{-1}$ in \eqref{Q_def_GSP} with
diagonal versions in \eqref{inv_Dx_define} and \eqref{inv_Dw_define}.
In Subsection \ref{separately_model_section}, we present the orthogonal-graph-frequencies case, in which the estimator from \eqref{27_old25_before_approx} that minimizes the objective in \eqref{Q_def_GSP_incremental} is separable in the graph-frequency domain and can be implemented with per-coordinate iterations. Thus, in this case, the estimator from \eqref{27_old25_before_approx} has a lower computational complexity than the MAP estimator.
However,
in the general case, a main problem is that the matrix $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)})$ in \eqref{27_old25_before_approx} is a full matrix that changes at each iteration.
To bypass this hurdle, we use the Gauss-Newton iteration in \eqref{27_old25_before_approx} with an additional step of neglecting the
off-diagonal elements of
$\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)})$.
This approach results in a separable form of the estimator in the graph-frequency domain for the general non-orthogonal case.
In the second step, at each iteration of \eqref{27_old25_before_approx} we neglect the non-diagonal elements of the Jacobian matrix, $\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})$ that involves the mixed-derivatives of
$\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$. This results in the following iteration:
\begin{eqnarray}\label{27_old25}
\hat{\tilde{{\bf{x}}}}^{(t+1)}
=\hat{\tilde{{\bf{x}}}}^{(t)}-\alpha^{(t)}\bigg({{\bf{D}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})} + {{\bf{D}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}(\hat{\tilde{{\bf{x}}}}^{(t)})^2\bigg)^{-1}\nonumber\\\times\bigg({\bf{D}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}(\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})
\hspace{3.25cm}
\nonumber\\
-{\bf{D}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{(\text{inv})}\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}(\hat{\tilde{{\bf{x}}}}^{(t)})(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}))\bigg),\hspace{0.6cm} \end{eqnarray}
where \[\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}(\tilde{\bf{x}})=\text{ddiag}(\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})).\]
The estimator in \eqref{27_old25} is denoted the eGFD-MAP estimator.
The advantages of the eGFD-MAP estimator
compared with the conventional MAP iterative methods from Subsection \ref{iterative_subsection} are that: 1) there is no need to calculate the off-diagonal elements of the Jacobian matrix $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$ at each iteration; and 2) there is no need for performing matrix inversion per iteration.
Indeed, in order to compute ${\bf{D}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}$ and ${\bf{D}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}$ there is a need to compute the inverse of the prior covariance matrices ${{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}$ and ${{\bf{C}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}$. However, this calculation does not change at each iteration, and thus can be done offline where ${{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}$ and ${{\bf{C}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}$ are assumed to be known.
In particular, assuming the covariance matrices and the Jacobian matrix are given, preforming a single iteration of the MAP update equation according to \eqref{iteration_Newton2} or \eqref{iteration_Newton2_GSP} requires $\mathcal{O}(N^{3})$ calculations since, in general, there is a need for performing matrix inversion, while the update step of the eGFD-MAP from \eqref{27_old25} requires only $\mathcal{O}(N)$ calculations.
Due to these two advantages, and the fact that \eqref{27_old25} can be implemented in a component-wise fashion,
the eGFD-MAP estimator in \eqref{27_old25}
can result in a significant overall speedup and can be used even when the size of the network is large, in contrast with the MAP estimator, which becomes intractable in large networks.
In addition, in cases were the matrices $ {{\bf{C}}}_{{\bf{x}}\xvec}^{-1} + {\bf{G}}^{T}({\bf{L}},\hat{{\bf{x}}}^{(t)}){{\bf{C}}}_{{\bf{w}}\wvec}^{-1}{\bf{G}}({\bf{L}},\hat{{\bf{x}}}^{(t)})$ in \eqref{iteration_Newton2}
or ${{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1} + \tilde{{\bf{G}}}^{T}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}){{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})$ in \eqref{iteration_Newton2_GSP}, are ill-conditioned (as may happen when the sample covariance matrices are used instead of the true covariances), the calculation of their inverse is prone to large numerical errors. This can badly affect the performance of the MAP estimator, in contrast to the performance of the eGFD-MAP estimator that does not require matrix inversion. Finally,
the eGFD-MAP-estimator algorithm is summarized in Algorithm \ref{algorithm_general}.
\begin{algorithm}[hbt]
\SetAlgoLined
\textbf{Input:}
\begin{itemize}
\item The function $\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$
\item The Laplacian matrix ${\bf{L}}$
\item Initial step size $\alpha^{(0)}$, $\gamma$ and $\Delta$
\item The mean and the diagonal entries of the inverse covariance matrices: $\tilde{\mbox{\boldmath $\mu$}}_{{\bf{x}}}$, ${{\bf{D}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}$, and ${{\bf{D}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}$
\item The tolerance $\delta$
\end{itemize}
{\textbf{Algorithm Steps:}}
\begin{enumerate}
\setlength\itemsep{0.2em}
\item Initialization: ${\hat{\tilde{{\bf{x}}}}}^{(0)}$
\item Compute
$\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)})$ and
$\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}(\hat{\tilde{{\bf{x}}}}^{(t)})$ from \eqref{GGG_tilde}
\item Choose step size $\alpha^{(t)}$ according to Algorithm \ref{line_search_stages} with the cost $Q_{freq}^{(d)}$ from \eqref{Q_def_GSP_incremental}
\item\label{Step4} Apply iteration step according to
\eqref{27_old25}
\item Stopping condition: $||{\hat{\tilde{\bf{x}}}}^{(t+1)}-{\hat{\tilde{\bf{x}}}}^{(t)}||<\delta$
\end{enumerate}
\vspace{1mm}
\KwOutput{eGFD-MAP estimator
$\hat{{\bf{x}}} = {\bf{V}}{\hat{\tilde{{\bf{x}}}}}^{(t+1)}$}
\caption{eGFD-MAP estimator by Gauss-Newton}
\label{algorithm_general}
\end{algorithm}
\subsection{GSP-MAP estimator}
\label{MAP_GSP_section}
While the eGFD-MAP estimator has a lower computational cost, it is an ad-hoc estimator.
In this subsection, we aim to find an optimal approach in the GSP sense.
As shown in Subsection \ref{iterative_subsection}, the update equation of the MAP estimator is obtained by solving a
linearized WLS problem at each iteration.
As a result, the update equations under the three objective functions in \eqref{iteration_Newton2}, \eqref{iteration_Newton2_GSP}, and \eqref{27_old25} are all linear functions of ${\bf{y}}-{\bf{g}}({\bf{L}},\hat{{\bf{x}}}^{(t)})$ and of $\hat{{\bf{x}}}^{(t)}- {\mbox{\boldmath $\mu$}}_{\bf{x}}$.
Based on this representation,
in the following we remain
with the linearized WLS problem
but constrain the estimator at each iteration $t$ to be the output of two graph filters.
In particular,
in the following,
we consider that at the $t$th iteration, the estimator has the form:
\begin{eqnarray} \label{opt_estimator}
\hat{{\bf{x}}}^{(t+1)} =\hat{{\bf{x}}}^{(t)}
+ f_1({\bf{L}},\hat{{\bf{x}}}^{(t)})(\hat{{{\bf{x}}}}^{(t)}- {{\mbox{\boldmath $\mu$}}}_{\bf{x}})\nonumber\\
+f_2({\bf{L}},\hat{{\bf{x}}}^{(t)}) ({\bf{y}}-{\bf{g}}({\bf{L}},\hat{{\bf{x}}}^{(t)})) ,
\end{eqnarray}
where
$f_i(\cdot,\cdot)$, $i=1,2$, are graph filters as defined in \eqref{laplacian_graph_filter}.
By left-multiplying \eqref{opt_estimator}
by ${\bf{V}}^T$, we obtain that
in the graph domain representation the estimator from \eqref{opt_estimator} can be written as
\begin{eqnarray} \label{opt_estimator_graph}
\hat{\tilde{{\bf{x}}}}^{(t+1)}
= \hat{\tilde{{\bf{x}}}}^{(t)}+ f_1({\bf{\Lambda}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)} )(\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\nonumber\\
+f_2({\bf{\Lambda}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)} )(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)})),
\end{eqnarray}
where ${\bf{V}}$ and ${\bf{\Lambda}}$ are the eigenvector-eigenvalue matrices of the Laplacian matrix,
${\bf{L}}$, as defined in \eqref{SVD_new_eq}.
In this form of estimators, the terms $\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}}$ and $\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)})$ are multiplied by {\em{diagonal}} matrices that represent the graph filters.
It can be seen that in the general case,
the estimators in \eqref{iteration_Newton2} and \eqref{iteration_Newton2_GSP} cannot be written in the form of \eqref{opt_estimator} or \eqref{opt_estimator_graph}.
In contrast,
the iteration of the eGFD-MAP estimator from \eqref{27_old25} can be written as the output of graph filters, as described in
\eqref{opt_estimator_graph}, where $f_1(\cdot,\cdot)$ and $f_2(\cdot,\cdot)$ are the following graph filters:
\begin{eqnarray}
\label{f1_def}
f_1^{\text{eGFD-MAP}}({\bf{\Lambda}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})\hspace{4cm}\nonumber\\=-
\alpha^{(t)}\bigg({{\bf{D}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})} + {{\bf{D}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}
\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}^2(\hat{\tilde{{\bf{x}}}}^{(t)})\bigg)^{-1}
{\bf{D}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}
\end{eqnarray}
and
\begin{eqnarray}
\label{f2_def}
f_2^{\text{eGFD-MAP}}({\bf{\Lambda}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})=
\alpha^{(t)}\hspace{3.5cm}\nonumber\\
\times\bigg({{\bf{D}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})} + {{\bf{D}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}
\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}^2(\hat{\tilde{{\bf{x}}}}^{(t)})\bigg)^{-1}
{\bf{D}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{(\text{inv})}
\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}(\hat{\tilde{{\bf{x}}}}^{(t)}).
\end{eqnarray}
In the following, our goal is to choose the graph filters $f_1(\cdot,\cdot)$ and $f_2(\cdot,\cdot)$ in the updated equation in \eqref{opt_estimator} (or, equivalently, in \eqref{opt_estimator_graph}) in an optimal way in the sense that the expected objective function
$ Q_{lin}$ from \eqref{Q_def_approx} is minimized for the general case.
It should be noted that the graph filters can be a function of the previous-iteration estimator, $\hat{{{\bf{x}}}}^{(t)}$.
The following theorem describes the MAP-optimal graph filters.
\begin{Theorem}
\label{claim1}
The graph filters that minimize the expected objective function, \begin{equation}
\label{obj1}
{\rm{E}}\left[\left.Q_{lin}(\hat{{\bf{x}}}^{(t+1)},\hat{{{\bf{x}}}}^{(t)})\right|{\hat{{{\bf{x}}}}^{(t)}={\bf{x}}}\right],
\end{equation}
over the subset of GSP estimators defined in \eqref{opt_estimator},
under
the approximation that $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})$ is a deterministic matrix\footnote{The rationale behind this approximation is similar to that behind the first-order approximation in the Gauss-Newton method \cite{Bell_Cathey1993}.},
are
\begin{eqnarray}
\label{filter1}
f_1^*({\bf{\Lambda}},\hat{{\bf{x}}}^{(t)} )=-
{\emph{diag}}\Big(\Big({{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}
\circ {{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}
\hspace{3cm}
\nonumber\\
\left.
+{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}
\circ (\tilde{\bf{G}}^T({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})
{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}
\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}))\right)^{-1}
{\bf{1}}\Big)
\end{eqnarray}
and
\begin{eqnarray}
\label{filter2}
f_2^*({\bf{\Lambda}},\hat{{\bf{x}}}^{(t)} )=
{\emph{diag}}\Big(\Big({{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}} \circ {{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}
\hspace{3cm}
\nonumber\\
\left.+ {{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}} \circ (\tilde{\bf{G}}^T({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})
{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}
\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}))\right)^{-1}
\nonumber\\
\times{\emph{diag}}(\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}))\Big)
\end{eqnarray}
at the $t+1$ iteration.
\end{Theorem}
\begin{IEEEproof}
The full proof appears in Appendix \ref{appendix_derivation_graph_filters}.
The main course of the proof is composed of two steps; in the first step, it is shown that under the theorem conditions,
we can replace the minimization of \eqref{obj1} by the minimization of
\begin{eqnarray}
\label{LMMSE_GSP_o3_try2}
{\rm{E}}[ Q_{lin}^{approx}(\tilde{\bf{x}},\hat{{\tilde{\bf{x}}}}^{(t)})] \hspace{4.75cm}
\nonumber\\=
\frac{1}{2} {\text{trace}}\left({{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}({\bf{I}}+
f_1({\bf{\Lambda}})){{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}(({\bf{I}}
+ f_1({\bf{\Lambda}}))\right) \nonumber\\
+\frac{1}{2} {\text{trace}}\left(
{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}
f_2({\bf{\Lambda}}) {{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}f_2({\bf{\Lambda}})\right)
\hspace{1.5cm}
\nonumber\\
+\frac{1}{2} {\text{trace}}\left({{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}f_1({\bf{\Lambda}})
{\bf{A}} f_1({\bf{\Lambda}})\right)\hspace{2.1cm}
\nonumber\\
+ {\text{trace}}\left(\frac{1}{2}{\bf{I}}-\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})f_2({\bf{\Lambda}})\right) \hspace{1.25cm}
\nonumber\\
+\frac{1}{2} {\text{trace}}\left({{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}f_2({\bf{\Lambda}}){\bf{A}} f_2({\bf{\Lambda}})\right), \hspace{1.85cm}
\end{eqnarray}
where ${\mathcal{D}}_N$ is the set of diagonal matrices of size $N \times N$ and
\begin{equation}
\label{Adef}
{\bf{A}}\stackrel{\triangle}{=}\tilde{\bf{G}}^T({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})
{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}
\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}).
\end{equation}
Then, in the second step, it is shown that this minimization w.r.t. the graph filters results in \eqref{filter1} and \eqref{filter2}.
\end{IEEEproof}
Similar to the iterative algorithms from Section \ref{iterative_subsection}, in practice, we multiply these graph filters by the step size
$\alpha^{(t)}\in(0,1]$ that is computed by a backtracking line search. The GSP-MAP algorithm is summarized in Algorithm \ref{new_alg}.
\begin{algorithm}[hbt]
\SetAlgoLined
\textbf{Input:}
\begin{itemize}
\item The function $\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$
\item The Laplacian matrix ${\bf{L}}$
\item Initial step size $\alpha^{(0)}$, $\gamma$ and $\Delta$
\item The mean and covariance matrices $\tilde{\mbox{\boldmath $\mu$}}_{{{\bf{x}}}}$, ${{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}$, and ${{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}$
\item The tolerance $\delta$
\end{itemize}
{\textbf{Algorithm Steps:}}
\begin{enumerate}
\setlength\itemsep{0.2em}
\item Initialization: ${\hat{\tilde{{\bf{x}}}}}^{(0)}$
\item Compute
$\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)})$ and $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)})$ from \eqref{GGG_tilde}
\item Compute the graph filters $f_1^*({\bf{\Lambda}},\hat{{\bf{x}}}^{(t)} )$
and
$f_2^*({\bf{\Lambda}},\hat{{\bf{x}}}^{(t)} )$
from \eqref{filter2} and \eqref{filter1}
\item Choose step size $\alpha^{(t)}$ according to Algorithm \ref{line_search_stages} with the cost $Q_{freq}$.
\item Apply iteration step based on \eqref{opt_estimator_graph} with the computed graph filters:
\begin{eqnarray} \label{opt_estimator_graph_filter} \hspace{-0.5cm}\hat{\tilde{{\bf{x}}}}^{(t+1)}
= \hat{\tilde{{\bf{x}}}}^{(t)}+ \alpha^{(t)} f_1^*({\bf{\Lambda}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)} )(\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\nonumber\\
+\alpha^{(t)} f_2^*({\bf{\Lambda}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)} )(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}))
\end{eqnarray}
\item Stopping condition: $||{\hat{\tilde{\bf{x}}}}^{(t+1)}-{\hat{\tilde{\bf{x}}}}^{(t)}||<\delta$
\end{enumerate}
\vspace{1mm}
\KwOutput{Graph-filtered GSP-MAP estimator
$\hat{{\bf{x}}} \hspace{-0.12cm} =\hspace{-0.13cm} {\bf{V}}{\hat{\tilde{{\bf{x}}}}}^{(t+1)}\hspace{-0.2cm}$}
\caption{GSP-MAP estimator by Gauss-Newton}
\label{new_alg}
\end{algorithm}
\subsection{Orthogonal-graph-frequencies models}
\label{separately_model_section}
In this subsection,
we present the special case of orthogonal graph frequencies, and its special case in which the measured graph signal is the output of a GSP filter.
We show that in the case of orthogonal graph frequencies,
the proposed eGFD-MAP and GSP-MAP estimators coincide with the MAP estimator.
The orthogonal-graph-frequencies model is defined as follows.
\begin{definition}
\label{orthogonal_frequencies}
The nonlinear measurements function, ${\bf{g}}({\bf{L}},{\bf{x}})\in {\mathbb{R}}^N$, is separable in the graph-frequency domain (``orthogonal graph frequencies") if it satisfies
\begin{equation} \label{separately_Model}
[\tilde{{\bf{g}}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})]_n = [\tilde{{\bf{g}}}({\bf{L}},\tilde{x}_n{\bf{v}}_n)]_n,~n=1,\ldots,N,
\end{equation}
where $\tilde{x}_n$ is the $n$th element of $\tilde{{\bf{x}}}$ and $\tilde{{\bf{g}}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})={\bf{V}}^T{\bf{g}}({\bf{L}},{\bf{x}})$.
\end{definition}
Definition \ref{orthogonal_frequencies} is satisfied, for example, if the associated Jacobian matrix, ${\bf{G}}({\bf{L}},{\bf{x}})$, is diagonalized by the eigenvector matrix of ${\bf{L}}$, ${\bf{V}}$.
By substituting \eqref{separately_Model} in \eqref{Q_def_GSP_incremental}, we obtain that in this case
\begin{eqnarray}
\label{Q_def_GSP_incremental2}
Q_{freq}^{(d)}(\tilde{{\bf{x}}})
=\frac{1}{2}\sum_{n=1}^N
(\tilde{x}_n-[\tilde{\mu}_{\bf{x}}]_n)^2[{{\bf{C}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1}]_{n,n}\hspace{1.25cm}\nonumber\\ + \frac{1}{2}\sum_{n=1}^N(\tilde{y}_n-
[\tilde{{\bf{g}}}({\bf{L}},\tilde{x}_n{\bf{v}}_n)]_n
)^2[{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}]_{n,n}.
\end{eqnarray}
Thus, in the case described by Definition \ref{orthogonal_frequencies}, the objective $Q_{freq}^{(d)}(\tilde{{\bf{x}}})$ is separable in the graph-frequency domain. The component-wise formulation of \eqref{Q_def_GSP_incremental2} results in a separable nonlinear WLS problem \cite{golub2003separable}, which
simplifies the task of designing the MAP estimator in the graph-frequency domain without the need for neglecting off-diagonal elements (as performed in the eGFD-MAP estimator).
The minimization of \eqref{Q_def_GSP_incremental2} instead of \eqref{Q_def_GSP}
results in $N$ independent optimization problems that
can result in an overall significant speedup.
Minimization of \eqref{Q_def_GSP_incremental2} can be simplified even further if
the graph input signal is a graph-bandlimited signal as defined in \eqref{bandlimited_def}.
In this case, by substituting $\tilde{x}_n=0$, $\forall n>N_s$, the sums in \eqref{Q_def_GSP_incremental2}
can be computed only over $n=1,\ldots,N_s$.
The following theorem states sufficient conditions for the proposed eGFD-MAP and GSP-MAP estimators to coincide with the MAP estimator for the case of an orthogonal-graph frequencies.
\begin{Theorem}\label{claim_coincides}
If the measurement function satisfies Definition \ref{orthogonal_frequencies} and, in addition, the following conditions hold:
\renewcommand{\theenumi}{C.\arabic{enumi}}
\begin{enumerate}
\setcounter{enumi}{0}
\item\label{cond1} The elements of the input graph signal, ${\bf{x}}$, are statistically independent in the graph-frequency domain, i.e. ${\bf{C}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}$ is a diagonal matrix;
\item\label{cond2} The noise vector, ${\bf{w}}$, is uncorrelated in the graph-frequency domain, i.e. ${\bf{C}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}$ is a diagonal matrix;
\end{enumerate}
then, the proposed eGFD-MAP and GSP-MAP estimators coincide with the MAP estimator.
\end{Theorem}
\begin{proof}
By substituting \eqref{separately_Model}, together with ${\bf{C}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{-1}={\bf{D}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}$ and
${\bf{C}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1}={\bf{D}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}$ from Conditions \ref{cond1} and \ref{cond2}, respectively, in the MAP objective function in the graph-frequency domain in \eqref{Q_def_GSP}, we obtain that in this case
\begin{eqnarray}
\label{Q_orth}
Q_{freq}(\tilde{{\bf{x}}})=
\frac{1}{2}(\tilde{{\bf{x}}}-\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})^{T}{{\bf{D}}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}(\tilde{{\bf{x}}}-\tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\hspace{1.67cm}\nonumber\\ + \frac{1}{2}(\tilde{{\bf{y}}}-\tilde{{\bf{g}}}({\bf{L}},{\bf{x}}))^{T}{{\bf{D}}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{(\text{inv})}(\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}}))
\nonumber\\=
\frac{1}{2}\sum_{n=1}^N
(\tilde{x}_n-[\tilde{\mu}_{\bf{x}}]_n)^2[{\bf{C}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1}]_{n,n}\hspace{1.42cm}\nonumber\\ + \frac{1}{2}\sum_{n=1}^N(\tilde{y}_n-
[\tilde{{\bf{g}}}({\bf{L}},\tilde{x}_n{\bf{v}}_n)]_n
)^2[{\bf{C}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{-1}]_{n,n}.
\end{eqnarray}
Thus, the eGFD-MAP objective from \eqref{Q_def_GSP_incremental2} coincides with the MAP objective from \eqref{Q_orth}.
From Definition \ref{orthogonal_frequencies}, we have that $\tilde{{\bf{G}}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})=\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}(\tilde{\bf{x}})$. Therefore, we can conclude that the eGFD-MAP estimator coincides with the MAP estimator.
By substituting, ${\bf{C}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{-1}={\bf{D}}_{\tilde{{\bf{x}}}\tilde{{\bf{x}}}}^{(\text{inv})}$,
${\bf{C}}_{\tilde{{\bf{w}}}\tilde{{\bf{w}}}}^{-1}={{\bf{D}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{(\text{inv})}$,
and $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})=\bar{\bf{D}}_{\tilde{{\bf{G}}}\tilde{{\bf{G}}}}(\hat{\tilde{\bf{x}}}^{(t)})$ in
the optimal graph filters from \eqref{filter1}
and \eqref{filter2}, and using the fact that ${\bf{I}}\circ{\bf{A}}={\text{ddiag}}({\bf{A}})$, ${\text{diag}}({\bf{D}} {\bf{a}})={\bf{D}} {\text{diag}}({\bf{a}})$, and ${\text{ddiag}}({\bf{D}})={\bf{D}}$ for any
matrix ${\bf{A}}$,
diagonal matrix, ${\bf{D}}$, and vector ${\bf{a}}$, we obtain the graph filters of the eGFD-MAP estimator from \eqref{f1_def} and \eqref{f2_def}. Thus,
the GSP-MAP estimator coincides with the eGDF-MAP estimator, which is also the MAP estimator.
\end{proof}
The following corollary presents a special case of Theorem \ref{claim_coincides}.
\begin{corollary}\label{claim_graphical_Model}
The proposed eGFD-MAP and GSP-MAP estimators coincide with the MAP estimator
if Conditions \ref{cond1} and \ref{cond2} hold, and
\renewcommand{\theenumi}{C.\arabic{enumi}}
\begin{enumerate}
\setcounter{enumi}{2}
\item\label{cond3} The measurement function, ${\bf{g}}({\bf{L}},{\bf{x}})$, is the output of a linear graph filter as defined in \eqref{laplacian_graph_filter}, i.e.
\begin{equation}
\label{g_filter1}
{\bf{g}}({\bf{L}},{\bf{x}}) = {\bf{V}} f({\bf{\Lambda}}){\bf{V}}^T {\bf{x}}.
\end{equation}
\end{enumerate}
\end{corollary}
\begin{proof}
By multiplying \eqref{g_filter1} by ${\bf{V}}^{T}$ and using the fact that ${\bf{V}}$ is unitary matrix, i.e. ${\bf{V}}\Vmat^T = {\bf{I}}$, we obtain
\begin{equation}
\label{g_filter2}
\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}}) = f({\bf{\Lambda}}) \tilde{\bf{x}}.
\end{equation}
Since $f({\bf{\Lambda}})$ is a diagonal matrix, we obtain that the measurement function satisfies \eqref{separately_Model} in Definition \ref{orthogonal_frequencies}. Since, in addition, we assume that Conditions \ref{cond1} and \ref{cond2} of Theorem \ref{claim_coincides} hold, we obtain that this is a special case of Theorem \ref{claim_coincides}. Thus, the eGFD-MAP and GSP-MAP estimators coincide with the MAP estimator in this case.
\end{proof}
The special case in Corollary \ref{claim_graphical_Model} fits the model behind
the graphical Wiener filter \cite{7891646}, as explained in \cite{kroizer2021bayesian}, where
under Condition \ref{cond1}, the signal $ {\bf{x}} - \EX[{\bf{x}}]$ is a Graph Wide-Sense Stationary signal (see Definition 3 and Theorem 1 in \cite{7891646}).
Therefore, in the case where the conditions of Corollary \ref{claim_graphical_Model} hold, an
estimator that is obtained by minimizing \eqref{Q_def_GSP_incremental2} coincides with the graphical Wiener filter (Eq. (13) in \cite{7891646}), the MAP estimator, the LMMSE estimator, and the GSP-LMMSE estimator from \cite{kroizer2021bayesian}.
In this case, the Gauss-Newton iterative approach converges in single iteration with a step size equal to 1, since ${\bf{g}}$ is a linear function (see Sec. 1.5.1 in \cite{BERTSEKAS_1999}).
\subsection{Computational complexity}\label{computational_complexity}
The computational complexity and the running time of the proposed iterative estimators mainly depend on: 1) the total number of iterations until convergence of the estimator; 2) the total subiterations in the backtracking line search algorithm (Algorithm \ref{line_search_stages}); and 3) the matrix multiplications in the update step. If we assume that 1) and 2) are roughly similar among the different estimators (given that the parameters $\alpha^{(0)}$, $\gamma$, $\Delta$, and $\delta$ are the same among the estimators), then the differences in the complexity and running time are due to the update steps of the different estimators. In addition, the computations of the inverse covariance matrices of ${\bf{x}}$ and ${\bf{w}}$ in the graph and graph-frequency domains, and the computation of the eigenvalue decomposition of the Laplacian matrix, which is of order $\mathcal{O}(N^3)$, can be done offline. Therefore, we do not consider them in the computational complexity.
The update rule of the MAP estimator in \eqref{iteration_Newton2} consists of full matrix multiplications with a computational complexity of $\mathcal{O}(N^3)$, and inversion of an $N\times N$ full matrix, which has a complexity of $\mathcal{O}(N^3)$. This also holds for the implementation of the MAP estimator in the graph-frequency domain; however, as mentioned, one implementation of the MAP (in the graph or graph-frequency domain) may be more efficient than the second one. The computational complexity of the update step of the GSP-MAP estimator is similar to that of the MAP estimator, since the reconstruction of each filter in \eqref{filter1} and \eqref{filter2} demands full matrix multiplications in the order of $\mathcal{O}(N^3)$, and inversion of a $N\times N$ matrix. The update step of the eGFD-MAP estimator in \eqref{27_old25} can be implemented in a vectorized form, and thus requires only $\mathcal{O}(N)$ multiplications, without the need to invert a matrix. In addition, in order to perform the update steps of the MAP and GSP-MAP estimators, it is necessary to calculate $N^2$ elements of the Jacobian matrix in \eqref{GGG} (or in \eqref{GGG_tilde}), compared to $N$ diagonal elements for the eGFD-MAP estimator in \eqref{27_old25}.
As the size of the network increases, the differences in the computational complexities of the estimators become more significant, so that in very large networks the MAP and GSP-MAP estimators may become intractable.
The total floating point operations order of the update steps of the different estimators is summarized in Table \ref{computational_complexity_tab}.
\begin{table}[hbt!]\centering
\begin{adjustbox}{max width=0.489\textwidth}
\begin{tabular}{|c|c|c|c|}
\hline
Estimator & MAP & eGFD-MAP & GSP-MAP \\ \hline
FLOPs & $11N^{3}+2.5N^{2}+1.5N$ & $11N$ & $10N^3+6.5N^{2}+5.5N$ \\ \hline
Order & $\mathcal{O}(N^3)$ & $\mathcal{O}(N)$ & $\mathcal{O}(N^3)$ \\ \hline
\end{tabular}
\end{adjustbox}
\caption{The total floating point operations (FLOPs) and order of complexity required for the update rules of the different estimators.}
\label{computational_complexity_tab}
\end{table}
\section{Simulation} \label{simulation}
In this section, we evaluate the performance of the different estimators for
a synthetic example in Subsection \ref{toy} and for the PSSE problem in Subsection \ref{seeting_sec}.
The estimators that are used in this section are:
\begin{itemize}
\item MAP estimator, implemented by Algorithm \ref{algorithm_MAP};
\item MAP estimator in the graph-frequency domain (MAP-FD), implemented by Algorithm \ref{algorithm_MAP} while using \eqref{iteration_Newton2_GSP};
\item eGFD-MAP estimator, implemented by Algorithm
\ref{algorithm_general};
\item GSP-MAP estimator implemented by Algorithm \ref{new_alg};
\item LMMSE estimator and GSP-LMMSE estimator from \cite{kroizer2021bayesian}; In Subsection \ref{toy}, the analytic-version of LMMSE and GSP-LMMSE estimators is implemented, while
in Subsection
\ref{seeting_sec}
these estimators are implemented by their sample-mean version (i.e. where the intractable covariance matrices are replaced by the sample covariance matrices), as discussed in detail in \cite{kroizer2021bayesian}. Here,
we used $P=500$ training samples to compute the sample covariance matrices.
\end{itemize}
All iterative estimators (MAP, eGFD-MAP, and GSP-MAP) were initialized by the GSP-LMMSE estimator,
unless written otherwise.
The performance of the different estimators is calculated by performing 10,000 Monte Carlo simulations for each scenario.
\subsection{Example A: Synthetic data - orthogonal graph frequencies}
\label{toy}
In this example we use random graphs that were generated by the Watts-Strogatz small-world graph
model \cite{Watts_Strogatz}, with different number of vertices, $N$, and a mean degree of $K=5$.
We evaluate the different estimators under the model from \eqref{Model} with the following
nonlinear measurement function in the graph-frequency domain:
\begin{equation}\label{sep_func}
[\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})]_{n} = \tilde{\bf{x}}_{n}^{3} \quad n=1,\ldots,N.
\end{equation}
We assume that the {\em{a-priori}} pdf of ${\bf{x}}$ is given by $\tilde{{\bf{x}}} \sim \pazocal{N}({\bf{0}},\sigma_{{\bf{x}}}^{2}{\bf{I}})$. The noise ${\bf{w}}$ from \eqref{Model} is white Gaussian noise with ${\bf{C}}_{{\bf{w}}\wvec}=\sigma_{{\bf{w}}}^{2}{\bf{I}}$. In the simulations, we use $\sigma_{{\bf{x}}}^{2}=0.5$ and $\sigma_{{\bf{w}}}^{2}=0.05$.
It can be verified that this nonlinear model has ``orthogonal graph frequencies", as defined in Definition \ref{orthogonal_frequencies}, and that the conditions of Theorem \ref{claim_coincides} are satisfied since $\tilde{\bf{x}}$ and $\tilde{\bf{w}}$ are white Gaussian noise signals.
Thus, according to Theorem \ref{claim_coincides} the MAP estimator coincides with the eGFD-MAP and the GSP-MAP estimators.
Under this setting, the LMMSE and GSP-LMMSE estimators coincide and can be computed analytically by using ${\bf{C}}_{{\bf{y}}\yvec}=(\sigma_{{\bf{w}}}^{2}+15\sigma_{{\bf{x}}}^{6}){\bf{I}}$ and ${\bf{C}}_{{\bf{x}}{\bf{y}}}=3\sigma_{{\bf{x}}}^{4}{\bf{I}}$.
In the simulations in this subsection, we also add the MAP-FD estimator in order to have a fair comparison in terms of run-time.
In Fig. \ref{fig:MSE_SYN} we present the normalized MSE (NMSE), i.e. the MSE divided by the number of vertices, $N$, of the different estimators versus $N$.
In this case,
the nonlinear estimators (MAP, MAP-FD, eGFD-MAP, and GSP-MAP estimators) significantly outperform the linear estimators (LMMSE and GSP-LMMSE), which have an almost constant NMSE for any $N$ around the value $0.208$ with a standard deviation of $0.001$. Therefore, and due to resolution reasons, the linear estimators are omitted from this figure.
Since in this case the conditions of Theorem \ref{claim_coincides} are satisfied, the MSEs of the nonlinear estimators (MAP, MAP-FD, eGFD-MAP,
and GSP-MAP estimators) are all equal as expected from Theorem \ref{claim_coincides}.
It can be seen that the MSE of the nonlinear estimators increases as $N$ increases.
\begin{figure}[hbt]
\centering
\includegraphics[width=0.8\linewidth]{figures/Synthetic_Example7.eps}
\caption{Example A: The NMSE of the different estimators versus $N$. }
\label{fig:MSE_SYN}
\end{figure}
In Fig. \ref{fig:TIME_SYN} we present the averaged runtime of any estimator till convergence versus $N$, which is evaluated using MATLAB on an Intel Core(TM) i7-7500U CPU computer, 2.7 GHz.
It can be seen that the eGFD-MAP estimator, which does not require matrix inversion per iteration, has the smallest runtime.
This complexity reduction (in terms of run-time) becomes more evident as the dimension of
the network increases.
Hence the eGFD-MAP
estimator is a good alternative to the MAP estimator in the cases where the measurement function is close to separable in the graph-frequency domain, i.e. has almost ``orthogonal frequencies". In addition, it can be seen that in this case, implementation in the graph frequency domain has lower computational complexity; indeed, the run-time of the MAP-FD estimator is lower than the run-time of the MAP estimator.
The averaged run-time of the linear estimators (not shown in this figure due to resolution reasons) is significantly smaller than the iterative estimators (around $1 \cdot 10^{-6}$ [sec]). This run-time would significantly increase in complex models where there is a need to compute the sample covariance matrices, as in the following subsection.
\begin{figure}[hbt]
\centering
\includegraphics[width=0.8\linewidth]{figures/Synthetic_Example7_time.eps}
\caption{Example A: The time complexity of the different estimators versus $N$.}
\label{fig:TIME_SYN}
\end{figure}
\subsection{Example B: PSSE in electrical networks}
\label{seeting_sec}
A power system can be represented as an undirected weighted graph, ${\pazocal{G}}({\pazocal{V}},\xi)$, where the set of vertices, $\pazocal{V}$, is the set of buses (generators or loads) and the edge set, $\xi$, is the set of transmission lines between these buses.
The measurement vector of the active powers at the buses, ${\bf{y}}$,
can be described by the model in \eqref{Model} with the measurement function \cite{Abur_book}:
\begin{eqnarray} \label{g_AC}
\left[{\bf{g}}({\bf{L}},{\bf{x}})\right]_n\
\stackrel{\triangle}{=} \sum_{m=1}^N |v_n||v_m|(G_{n,m}\cos(x_n -x_{m})
\nonumber\\+B_{n,m}\sin(x_n -x_{m})),\hspace{1.5cm}
\end{eqnarray}
$n=1,\ldots,N$. Here, $x_n$ and $|v_n|$ are the voltage phase and amplitude at the $n$th bus,
and $G_{n,m}$ and $B_{n,m}$ are the conductance and susceptance of the transmission line between the buses $n$ and $m$ \cite{Abur_book}, where $(n,m)\in\xi$.
We assume that $|v_n|=1$, which is a common assumption in normalized power systems \cite{Abur_book}, and that $G_{n,m}$ and $B_{n,m}$ are all known. In the graph modeling of the electrical network, usually
the Laplacian matrix, ${\bf{L}}$, is constructed as follows (Subsection II-C in \cite{Grotas2019}):
\begin{eqnarray}
\label{L_B}
{\bf{L}}=
\left\{
\begin{array}{lr}
-B_{n,k} , & k \neq n\\
\mathlarger{\sum}_{\underset{m\neq n}{m=1}}^{N} B_{n,m} , & k = n.
\end{array}\right.
\end{eqnarray}
The goal of PSSE is to recover the state vector, ${\bf{x}}$, from the power measurements,
${\bf{y}}$, described by \eqref{Model} with the nonlinear measurement function
${\bf{g}}({\bf{L}},{\bf{x}})$ in \eqref{g_AC}.
This estimation task is known to be a NP-hard problem \cite{bienstock2019strong},
and is essential for various monitoring purposes
\cite{Abur_book}.
PSSE is traditionally solved by iterative methods, such as the Gauss-Newton method \cite{Abur_book}.
By substituting \eqref{g_AC} in \eqref{GGG}, we obtain that
the associated Jacobian matrix is
\begin{eqnarray}
\label{Jacobian_g}
[{\bf{G}}({\bf{L}},{\bf{x}})]_{n,k}=
\frac{\partial {\bf{g}}_{n}({\bf{L}},{\bf{x}})}{\partial x_{k}}=
\hspace{4cm}
\nonumber\\\left\{
\begin{array}{lr}
G_{n,k} \sin(x_{n}-x_{k})-B_{n,k} \cos(x_{n}-x_{k}), & k \neq n\\
\mathlarger{\sum}_{\underset{m\neq n}{m=1}}^{N} -G_{n,m} \sin(x_{n}-x_{m})+B_{n,m} \cos(x_{n}-x_{m}), & k = n
\end{array}\right..
\end{eqnarray}
In particular, since $G_{n,k}=0$ and $B_{n,k}=0$ for any $(n,k) \notin \xi$, it can be seen that $[{\bf{G}}({\bf{L}},{\bf{x}})]_{n,k}= 0$ for any $(n,k) \notin \xi$.
It should be noted that in the general case, \eqref{Jacobian_g} implies that the conditions of Theorem \ref{claim_coincides} are not satisfied in this case.
The input graph signal, ${\bf{x}}$, has been shown to be smooth w.r.t. the graph \cite{drayer2018detection,dabush2021state}.
Therefore, we model the distribution of ${\bf{x}}$ in the graph-frequency domain as a smooth, normal distribution \cite{Dong_Vandergheynst_2016,ramezani2019graph}, as defined in \eqref{x_distribution} in Appendix \ref{WLS_Interpretation}.
Since in this case study, ${\bf{x}}$ represents phases, the value of $\beta$ in \eqref{x_distribution} is taken such that the probability that the elements of $|{\bf{x}}|$ are larger than $\pi$ is smaller than 0.01.
We assume that the covariance matrix of the noise, ${\bf{w}}$, from the model in \eqref{Model} is
${\bf{C}}_{{\bf{w}}\wvec}=\sigma_{{\bf{w}}}^2{\bf{I}}$.
The values of the different physical parameters in \eqref{g_AC} are taken from the IEEE 118-bus test case
\cite{iEEEdata}, in which $N=118$
In Table \ref{hyperparameters} we present the hyperparameters used to evaluate the different estimators in this and in the following subsections.
\begin{table}[hbt!]\centering
\begin{tabular}{|c|c|c|c|}
\hline
Estimator & MAP & eGFD-MAP & GSP-MAP \\ \hline
$\alpha_{0}$ & 0.5 & 0.5 & 0.5 \\ \hline
$\gamma$ & 0.83 & 0.83 & 0.83 \\ \hline
$\delta$ & 0.1 & 0.1 & 0.1 \\ \hline
$\Delta$ & 0.01 & 0.01 & 0.01 \\ \hline
\end{tabular}
\caption{The hyperparameters used for the simulations in Subsections \ref{seeting_sec} and \ref{sensativity2initialization}
}
\label{hyperparameters}
\end{table}
Finally, it can be seen that in the model in \eqref{g_AC} there is an inherent ambiguity, and one can recover the phases in ${\bf{x}}$ only up to modulo $2\pi$ errors. Therefore, in the following simulations, the error is presented in terms of mean-squared-periodic-error (MSPE) \cite{Routtenberg_Tabrikian_Bayesian}:
\begin{equation}
\label{eq:rmspe}
{\text{MSPE}}({\bf{x}}, \hat{{\bf{x}}}) = {\rm{E}}\left[ \left({\text{mod}}_{2\pi}({\bf{x}}-\hat{{\bf{x}}})\right)^2\right],
\end{equation}
where $\text{mod}_{\pi}$ denotes the element-wise modulo operator.
The units of the MSPE are $[\text{rad}^{2}]$. The normalized MSPE (NMSPE) is obtained by dividing the MSPE by $N$.
In Fig. \ref{fig:Diff N}
we present the NMSPE of the different estimators
versus the inverse of the noise variance, $\frac{1}{\sigma_{{\bf{w}}}^2}$, for the 118-bus case.
It can be seen that the MAP, eGFD-MAP, and GSP-MAP estimators significantly outperform the linear estimators (LMMSE and GSP-LMMSE estimators) in terms of MSPE for $\sigma_w^2<1$.
As the noise variance, $\sigma_{\tilde{\bf{w}}}^2$, increases, the MSPE of the LMMSE and the GSP-LMMSE achieved the MSPE of the nonlinear estimators, since in the presence of significant measurement noise,
all estimators are reduced to the prior-mean estimator, ${\mbox{\boldmath $\mu$}}_{\bf{x}}$, which is a linear estimator.
Therefore, this figure shows that we can achieve good performance with the low-complexity eGFD-MAP estimator
and with the GSP-MAP estimator
even when the measurement function does not have orthogonal graph frequencies. In addition, the proposed estimators has a lower MSPE than the MAP estimator for large noise variance, since they have fewer parameters to determine. Thus, in this case, the proposed methods are more robust to noise than the MAP estimator.
\begin{figure}[hbt]
\centering
{ \includegraphics[width=0.8\linewidth]{figures/MSE_SIGMA_IEEE118.eps}}
\caption{The NMSPE of the different estimators versus $\frac{1}{\sigma_{{\bf{w}}}^2}$ for 118-bus test case with $\beta = 3$.\label{fig:Diff N}}
\end{figure}
In order to examine the robustness of the iterative estimators under different phase distributions, and in particular under different levels of separability of the measurement function, we changed the value of $\beta$ from \eqref{x_distribution}, which directly affects the separability of the measurement function in the sense of Definition \ref{orthogonal_frequencies}.
As $\beta$ increases, the total variation of the signal ${\bf{x}}$, which is defined as ${\bf{x}}^{T}{\bf{L}}{\bf{x}}$, increases.
From the distribution in \eqref{x_distribution}, and recalling that ${\bf{x}}={\bf{V}}\tilde{\bf{x}}$, the variance of $x_{i}$ can be expressed as
\begin{equation}\label{phase_var}
\text{var}\big({x_{i}}\big)=\beta\ \sum_{j=1}^{N} \frac{V^{2}_{i,j}}{\lambda_{j}}.
\end{equation}
Thus, as $\beta$ decreases, the phases are more concentrated around their mean, which is assumed to be the same for any $n=1,\ldots,N$.
Hence, as $\beta$ decreases the difference $x_n-x_m$ is smaller and when it is sufficiently small it is possible to use the first order Taylor approximation:
\begin{equation}
\label{approx}
\sin(x_n-x_m)\approx x_n-x_m,~
\cos(x_n-x_m)\approx 1
\end{equation}
Therefore, by substituting \eqref{approx} in \eqref{g_AC} and taking the derivative w.r.t ${\bf{x}}$, we obtain that the Jacobian of ${\bf{g}}({\bf{L}},{\bf{x}})$ from \eqref{g_AC} can be approximated by
\begin{equation}
\label{G_L}
{\bf{G}}({\bf{L}},{\bf{x}}) \approx -{\bf{L}} = {\bf{V}}{\bf{\Lambda}}{\bf{V}}^{T},
\end{equation}
where we used \eqref{L_B} and the eigenvalue decomposition of ${\bf{L}}$.
By multiplying \eqref{G_L} by ${\bf{V}}^T$ and ${\bf{V}}$ from the left and right, respectively, and by using \eqref{GGG_tilde}, one obtains
\begin{equation}\label{G_tilde_small_beta}
\tilde{\bf{G}}({\bf{L}},{\bf{V}}\tilde{\bf{x}}) \approx -{\bf{\Lambda}},
\end{equation}
i.e. $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$ is a diagonal matrix.
Equation \eqref{G_tilde_small_beta} implies that as $\beta$ decreases, $\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$ becomes more separable in the sense of Definition \ref{orthogonal_frequencies}.
However, for a general $\beta$, $\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$ may not be separable as required by Definition \ref{orthogonal_frequencies} (or, equivalently $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$ may not be a diagonal matrix).
In Fig. \ref{fig:beta_118} we present the NMSPE versus the parameter $\beta$. It can be seen that
the MSPE of the different estimators increases as $\beta$ increases, since $\beta$ is proportional to the variance of the unknown parameters. Moreover,
the iterative estimators have significantly lower MSPE than the linear estimators for any $\beta$. In addition it can be seen that the eGFD-MAP and GSP-MAP estimator have similar MSPE values to those of the MAP estimator for any value of $\beta$ that is large enough. Thus, in this case the proposed methods perform well even when the conditions of Theorem \ref{claim_coincides} are not satisfied.
\begin{figure}[hbt!]
\centering
{ \includegraphics[width=0.8\linewidth]{figures/MSE_BETA_IEEE118.eps}}
\caption{The NMSPE of the different estimators versus $\frac{1}{\beta}$ for 118-bus test case, where ${\sigma}_{{\bf{w}}}^2=0.05$.\label{fig:beta_118}}
\end{figure}
\subsection{Sensitivity to initialization}\label{sensativity2initialization}
The implementation of the MAP estimator by the Gauss-Newton method is known to be sensitive to the initialization of the algorithm \cite{Fatemi_Svensson_Morelande2012}. In this subsection, we examine the robustness of the MAP, eGFD-MAP, and GSP-MAP estimators to perturbed initialization under the same setting of the PSSE problem from Subsection \ref{seeting_sec}.
\subsubsection{Scenario I - noisy initialization}
In the first scenario, we use perturbed initialization, in which the estimators are initialized with
\begin{equation}
\hat{{\bf{x}}}^{(0)} = {\bf{V}}(\tilde{\bf{x}}_{0} + \tilde{\bf{p}}_{0}),
\end{equation}
where $\tilde{\bf{x}}_{0}$ is the original initialization (i.e. the prior mean in the graph-frequency domain, $\tilde{\mbox{\boldmath $\mu$}}_{{\bf{x}}}$, or the GSP-LMMSE estimator in the graph frequency domain), and $\tilde{\bf{p}}_{0}$ is zero-mean Gaussian noise with variance $\sigma_{p}^{2}$.
The NMSPE of the estimators versus $\frac{1}{\sigma_{p}^{2}}$ is presented in Fig.
\ref{fig:noise_INIT_118} for the
118-bus test case with this perturbed initialization.
\begin{figure}[hbt!]
\centering
{ \includegraphics[width=0.8\linewidth]{figures/MSE_INIT_ALL_IEEE118.eps}}
\caption{The NMSPE of the different estimators versus $\frac{1}{\sigma_{p}^{2}}$, where the noisy GSP-LMMSE and the noisy prior mean were used to initialize each of the estimators, where ${\sigma}_{{\bf{w}}}^2=0.05$ and $\beta=3$.}\label{fig:noise_INIT_118}
\end{figure}
\subsubsection{Scenario II - Perturbed topology}
In the second scenario, the GSP-LMMSE and the LMMSE estimators were calculated under a change in the topology. In particular, the sample-mean versions of these estimators were calculated under a misspecified model. While the true dataset was generated with a given topology, the linear estimators assume a different topology obtained from the original topology after removing $M$ edges. Then, the misspecified sample-mean GSP-LMMSE estimator was used as the initial estimator for the iterative estimators (similar results were obtained by using the misspecified sample-mean LMMSE estimator). Thus, this scenario describes a perturbed initialization, where the topology is perturbed and affects the initialization.
The results are presented in Fig. \ref{fig:MSE_PER_LIN_INIT_118} versus the number of removed edges $M$.
\begin{figure}[hbt!]
\centering
{ \includegraphics[width=0.8\linewidth]{figures/MSE_INIT1_IEEE118_50.eps}}
\caption{The NMSPE of the different estimators versus the number of removed edges $M$, where the linear estimators (that were used to initialize the iterative estimators) were perturbed due to changes in the graph topology, where $M$ edges were removed, where ${\sigma}_{{\bf{w}}}^2=0.05$ and $\beta=3$.}\label{fig:MSE_PER_LIN_INIT_118}
\end{figure}
It can be seen in Figs. \ref{fig:noise_INIT_118} and \ref{fig:MSE_PER_LIN_INIT_118}, associated with Scenarios I and II, respectively, that the MAP estimator is more sensitive to initialization than the proposed estimators. Specifically, when the noise variance is small or the number of removed edges is low, the MSPE of the MAP estimator is significantly larger than the MSPE of the proposed eGFD-MAP and GSP-MAP estimators.
When comparing the initialization method, it can be seen that the eGFD-MAP and GSP-MAP estimators are more robust to the wrong initialization for both initialization methods (GSP-LMMSE and prior mean). Thus, we can conclude that when the initialization tends to be a problem, the proposed methods that use GSP information are preferable.
\section{Conclusion}
\label{conclusion}
In this paper, we discuss the recovery of random graph signals from nonlinear measurements using a GSP-based MAP approach. We formulate the MAP estimator in both the vertex and the graph-frequency domains, and show that although they lead to the same estimator, the efficiency and convergence rate of each implementation may be different.
In order to accommodate the complexity and sensitivity to initialization of the MAP estimator, we developed the eGFD-MAP estimator that does not require a matrix inversion per iteration and can update the different elements of the graph signal in the graph-frequency domain
independently.
In addition,
we derived the GSP-MAP estimator that is computed iteratively with an update equation that is composed of the output of two graph filters that are optimal in the sense of the expected objective function.
We showed that when the measurement function is separable in the graph-frequency domain and the input graph signal and the noise vector are uncorrelated, the proposed eGFD-MAP and GSP-MAP estimators coincide with the conventional MAP estimator.
We showed that
the eGFD-MAP estimator has the lowest computational complexity, making its calculation tractable in large networks.
Our numerical simulations show that the proposed eGFD-MAP and the GSP-MAP estimators achieve similar MSEs, compared to the MAP estimator, while the eGFD-MAP estimator is more efficient. Finally, We examined the sensitivity to initialization of the MAP estimator and showed that the eGFD-MAP and the GSP-MAP estimators are significantly more robust to perturbed and noisy initialization.
\appendices
\renewcommand{\thesectiondis}[2]{\Alph{section}:}
\section{Interpretation of the MAP Estimation Problem as a Regularized WLS Problem}\label{WLS_Interpretation}
For the special case where
the distribution of the input graph signal, ${\bf{x}}$, in the graph-frequency domain is a smooth zero-mean Gaussian distribution \cite{Dong_Vandergheynst_2016,ramezani2019graph}:
\begin{equation} \label{x_distribution}
\tilde{{\bf{x}}} \sim \pazocal{N}({\bf{0}},\beta {\bf{\Lambda}}^{\dagger}),
\end{equation}
where $\beta$ is the smoothness level.
In this case, ${\mbox{\boldmath $\mu$}}_{\bf{x}}={\bf{0}}$ and ${{\bf{C}}}_{{\bf{x}}\xvec}=\beta {\bf{L}}^\dagger$, where $(\cdot)^\dagger$ denotes the pseudo-inverse operator. By substituting this prior in \eqref{Q_def}, we obtain
\begin{eqnarray}
\label{Q_def_smooth_prior}
Q({\bf{x}})=
\frac{1}{2}\beta {\bf{x}}^{T}{\bf{L}}{\bf{x}} + \frac{1}{2}({\bf{y}}-{\bf{g}}({\bf{L}},{\bf{x}}))^{T}{{\bf{C}}}_{{\bf{w}}\wvec}^{-1}({\bf{y}}-{\bf{g}}({\bf{L}},{\bf{x}})).
\end{eqnarray}
The left term on the r.h.s. of \eqref{Q_def_smooth_prior}, $\frac{1}{2}\beta {\bf{x}}^{T}{\bf{L}}{\bf{x}}$, can be interpreted as a
regularization term, which corresponds to assuming that the graph signal
${\bf{x}}$ is smooth on the graph.
If we take the model in \eqref{Model} where the prior information can be neglected (e.g. when ${{\bf{C}}}_{{\bf{x}}\xvec}$ is significantly larger than ${{\bf{C}}}_{{\bf{w}}\wvec}$ in the positive-definite matrix sense), then we can develop the WLS estimator instead of the MAP estimator, as considered, for example, in \cite{Sahu_Kar_Moura_Poor2016}. In this case, we do not have the first term in the objective functions (e.g.
$\frac{1}{2}({\bf{x}}-{\mbox{\boldmath $\mu$}}_{\bf{x}})^{T}{{\bf{C}}}_{{\bf{x}}\xvec}^{-1}({\bf{x}}-{\mbox{\boldmath $\mu$}}_{\bf{x}})$ in \eqref{Q_def}). However, we can add a Laplacian regularization term, $\mu {\bf{x}}^{T}{\bf{L}}{\bf{x}}$, to obtain the regularized WLS problem in \eqref{Q_def_smooth_prior}.
Signal recovery with a
Laplacian regularization term
has been used in various applications,
such as image processing \cite{Zheng_Cai2011,Elmoataz_2008} and state estimation in power systems \cite{dabush2021state}.
\section{Proof of Theorem \ref{claim1}}
\label{appendix_derivation_graph_filters}
For the sake of simplicity, in this appendix we use the notation $||{\bf{a}}||_{{\bf{B}}}={\bf{a}}^T{\bf{B}}{\bf{a}}$ for any vector ${\bf{a}}$ and a positive semi-definite matrix ${\bf{B}}$. In addition, we use the short notations $f_i({\bf{L}})$, $f_i({\bf{\Lambda}})$, $i=1,2$, without writing explicitly the dependency on $\hat{{{\bf{x}}}}^{(t)}$.
By substituting \eqref{opt_estimator}
in \eqref{Q_def_approx}, we obtain
\begin{eqnarray}
\label{LMMSE_GSP_o2}
Q_{lin}\Big(\hat{{\bf{x}}}^{(t+1)},\hat{{\bf{x}}}^{(t)}\Big)=
Q_{lin}\Big(\hat{{\bf{x}}}^{(t)}
+f_1({\bf{L}})(\hat{{{\bf{x}}}}^{(t)}- {{\mbox{\boldmath $\mu$}}}_{\bf{x}})
\nonumber\\
+f_2({\bf{L}}) ({\bf{y}}-{\bf{g}}({\bf{L}},\hat{{\bf{x}}}^{(t)})),\hat{{{\bf{x}}}}^{(t)}\Big)\hspace{2.4cm}\nonumber\\=
\frac{1}{2}\left|\left|({\bf{I}}+
f_1({\bf{L}}))(\hat{{{\bf{x}}}}^{(t)}- {{\mbox{\boldmath $\mu$}}}_{\bf{x}})\right.\right.\hspace{2.7cm}
\nonumber\\\left.\left.
+f_2({\bf{L}}) ({\bf{y}}-{\bf{g}}({\bf{L}},\hat{{\bf{x}}}^{(t)}))\right|\right|_{{{\bf{C}}}_{{\bf{x}}\xvec}^{-1}}\hspace{2.5cm}\nonumber\\ + \frac{1}{2}\left|\left|-{\bf{G}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)}) f_1({\bf{L}})(\hat{{{\bf{x}}}}^{(t)}- {{\mbox{\boldmath $\mu$}}}_{\bf{x}})\right.\right.\hspace{1.6cm}
\nonumber\\\left.\left.
+({\bf{I}}-{\bf{G}}({\bf{L}},\hat{{{\bf{x}}}}^{(t)})f_2({\bf{L}})) ({\bf{y}}-{\bf{g}}({\bf{L}},\hat{{\bf{x}}}^{(t)})))\right|\right|_{{{\bf{C}}}_{{\bf{w}}\wvec}^{-1}}
\nonumber\\
=\frac{1}{2}\left|\left|({\bf{I}}+
f_1({\bf{\Lambda}}))(\hat{\tilde{{\bf{x}}}}^{(t)}- {\tilde{\mbox{\boldmath $\mu$}}}_{\bf{x}})\right.\right.\hspace{2.75cm}
\nonumber\\\left.\left.
+f_2({\bf{\Lambda}}) (\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}))\right|\right|_{{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}}\hspace{2.25cm}\nonumber\\ + \frac{1}{2}\left|\left| -\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}) f_1({\bf{\Lambda}})(\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\right.\right.\hspace{1.3cm}
\nonumber\\\left.\left.
+({\bf{I}}-\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})f_2({\bf{\Lambda}})) (\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}))\right|\right|_{{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}},
\end{eqnarray}
where the last equality is in the graph-frequency domain, which holds
since ${\bf{V}}$ is a unitary matrix, i.e. ${\bf{V}}\Vmat^T = {\bf{I}}$.
Based on the assumption that $\hat{{\bf{x}}}^{(t)}$ is close enough to ${\bf{x}}$, we replace $\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}^{(t)}))$ by $\tilde{\bf{g}}({\bf{L}},{\bf{V}}\tilde{\bf{x}})$, and
$\hat{\tilde{{\bf{x}}}}^{(t)}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}}$ by
$\hat{\tilde{{\bf{x}}}}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}}$ in \eqref{LMMSE_GSP_o2}, but keep $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})$ evaluated at $\hat{\tilde{{\bf{x}}}}^{(t)}$ (similar to the rationale behind the first-order approximation in the Gauss-Newton method \cite{Bell_Cathey1993}), to obtain the approximation
\begin{eqnarray}
\label{LMMSE_GSP_o2_approx}
Q_{lin}^{approx}(\tilde{{\bf{x}}},\hat{\tilde{\bf{x}}}^{(t)})=\hspace{4.5cm}\nonumber\\
\frac{1}{2}\left|\left|({\bf{I}}+
f_1({\bf{\Lambda}})({\tilde{{\bf{x}}}}- {\tilde{\mbox{\boldmath $\mu$}}}_{\bf{x}})
+f_2({\bf{\Lambda}}) (\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}\hat{\tilde{\bf{x}}}))\right|\right|_{{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}}\nonumber\\ + \frac{1}{2}\left|\left| -\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)}) f_1({\bf{\Lambda}})({\tilde{{\bf{x}}}}- \tilde{{\mbox{\boldmath $\mu$}}}_{\bf{x}})\right.\right.\hspace{3cm}
\nonumber\\\left.\left.
+({\bf{I}}-\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})f_2({\bf{\Lambda}})) (\tilde{\bf{y}}-\tilde{\bf{g}}({\bf{L}},{\bf{V}}{\tilde{\bf{x}}}))\right|\right|_{{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}^{-1}}.
\end{eqnarray}
Our goal is to choose the filters $f_1(\cdot)$ and $f_2(\cdot)$ such that the objective
$ Q_{lin}^{approx}$ is minimized on average.
Thus,
we minimize the expected cost from \eqref{LMMSE_GSP_o2_approx}, under the assumption that $\tilde{{\bf{x}}}$ and $\tilde{{\bf{w}}}$ are independent, and treating $\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})$ as a deterministic matrix:
\begin{eqnarray}
\label{LMMSE_GSP_o3_try}
\{f_1^*({\bf{\Lambda}} ),f_2^*({\bf{\Lambda}} )\}\hspace{4cm}\nonumber\\=
\arg\min_{f_1({\bf{\Lambda}} ) \in {\mathcal{D}}_N,f_2({\bf{\Lambda}} ) \in {\mathcal{D}}_N} {\rm{E}}[ Q_{lin}^{approx}(\tilde{{\bf{x}}},\hat{\tilde{\bf{x}}}^{(t)})],
\end{eqnarray}
where ${\rm{E}}[ Q_{lin}^{approx}(\tilde{{\bf{x}}},\hat{\tilde{\bf{x}}}^{(t)})]$ is defined in \eqref{LMMSE_GSP_o3_try2}.
Since the minimization is separable w.r.t. $f_1$ and $f_2$, we can solve it independently.
Thus,
\begin{eqnarray}
\label{f1_term}
f_1^*({\bf{\Lambda}} )=
\arg\min_{f_1({\bf{\Lambda}} ) \in {\mathcal{D}}_N}
\sum_{n=1}^N [f_1({\bf{\Lambda}})]_{n,n}
\nonumber\\
+
\frac{1}{2} \sum_{n=1}^N \sum_{k=1}^N
[{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}]_{n,k} [{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}]_{k,n} [f_1({\bf{\Lambda}})]_{k,k}
[f_1({\bf{\Lambda}}))]_{n,n}
\nonumber\\
+
\frac{1}{2} \sum_{n=1}^N\sum_{k=1}^N
[{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}]_{n,k}[{\bf{A}}]_{k,n}
[f_1({\bf{\Lambda}})]_{k,k}
[ f_1({\bf{\Lambda}})]_{n,n}.
\end{eqnarray}
By equating the derivative of \eqref{f1_term} w.r.t. $[f_1({\bf{\Lambda}})]_{l,l}$ to zero, one obtains
\begin{eqnarray}
\label{derivative1}
0= 1+\sum_{n=1}^N
[{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}]_{n,l}
[{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}]_{l,n}
[f_1^*({\bf{\Lambda}})]_{n,n}
\nonumber\\
+ \sum_{n=1}^N
[{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}]_{n,l}[{\bf{A}}]_{l,n}
[ f_1^*({\bf{\Lambda}})]_{n,n}, ~\forall l=1,\ldots,N,
\end{eqnarray}
which results in
\begin{eqnarray}
\label{f1opt}
{\text{diag}}(f_1^*({\bf{\Lambda}}))=-
({{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}
\circ {{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}
+{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}
\circ {\bf{A}})^{-1}
{\bf{1}}.
\end{eqnarray}
By applying the diag operator on both sides of \eqref{f1opt} and substituting \eqref{Adef}, we obtain
the graph filter in \eqref{filter1}.
Similarly, the optimization w.r.t. the filter $f_2({\bf{\Lambda}} )$ is
\begin{eqnarray}
\label{optf2}
f_2^*({\bf{\Lambda}} )=
\arg\min_{f_2({\bf{\Lambda}} ) \in {\mathcal{D}}_N}\nonumber\\
\frac{1}{2} \sum_{n=1}^N\sum_{k=1}^N
[{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}]_{n,k}[{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}]_{k,n}
[f_2({\bf{\Lambda}})]_{k,k}
[ f_2({\bf{\Lambda}})]_{n,n}
\nonumber\\
-\sum_{n=1}^N [\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})]_{n,n}[f_2({\bf{\Lambda}})]_{n,n}
\nonumber\\
+
\frac{1}{2} \sum_{n=1}^N\sum_{k=1}^N
[{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}]_{n,k}[{\bf{A}}]_{k,n}
[f_2({\bf{\Lambda}})]_{k,k}
[ f_2({\bf{\Lambda}})]_{n,n}.
\end{eqnarray}
By equating the derivative of \eqref{optf2} w.r.t. $[f_2({\bf{\Lambda}})]_{l,l}$ to zero, one obtains
\begin{eqnarray}
\label{derivative2}
0=
\sum_{n=1}^N
[{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}]_{n,l}[{{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1}]_{l,n}
[ f_2^*({\bf{\Lambda}})]_m\hspace{2.5cm}
\nonumber\\
- [\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})]_{l,l}
+ \sum_{n=1}^N
[{{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}}]_{n,l}[{\bf{A}}]_{l,n}
[ f_2^*({\bf{\Lambda}})]_m,
\end{eqnarray}
$\forall l=1,\ldots,N$,
which results in
\begin{eqnarray}
\label{f2opt}
{\text{diag}}(f_2^*({\bf{\Lambda}}))\hspace{5.75cm}
\nonumber\\=({{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}} \circ {{\bf{C}}}_{\tilde{\bf{x}}\tilde{\bf{x}}}^{-1} + {{\bf{C}}}_{\tilde{\bf{w}}\tilde{\bf{w}}} \circ{\bf{A}})^{-1}{\text{diag}}(\tilde{\bf{G}}({\bf{L}},{\bf{V}}\hat{\tilde{{\bf{x}}}}^{(t)})).
\end{eqnarray}
By applying the diag operator on both sides of \eqref{f2opt} and substituting \eqref{Adef}, we obtain
the graph filter in \eqref{filter2}.
\bibliographystyle{IEEEtran}
|
\section{\label{s:force_derivation} Derivation of the forces acting on the nanoparticle}
In this section we give further details of the derivation of the electromagnetic forces acting on the nanoparticle that are discussed in the main text.
\newcommand{\gpv}{\vb r}
We consider a point dipole at position $\dpv$ with electric dipole moment given by $\vb d(t)$. The dipole is placed in vacuum and possibly in the vicinity of dielectric material (e.g.\ a surface). The total electromagnetic field in the space is given by $\vb E(\gpv,t)$. The electric dipole moment has a contribution $\vb d_\mathrm{th}(t)$ from intrinsic thermal fluctuations at temperature $T$ (e.g.\ due to the internal temperature of a nanoparticle) and a contribution $\vb d_\mathrm{ind}(t)$ due to the dipole moment induced by the external electric field. The electromagnetic field $\vb E(\gpv,t)$ has a contribution $\vb E_\mathrm{th}(\gpv, t)$ from environmental thermal fluctuations at temperature $T_\mathrm{env}$ (i.e.\ due to the temperature of the vacuum and the surrounding dielectric material), a contribution $\vb E_\mathrm{tw}(\gpv,t)$ from the laser beam used for trapping, and a contribution $\vb E_\mathrm{ind}(\gpv, t)$ from the field emitted by the point dipole. In total, we have
\begin{align}
\vb d(t) & = \vb d_\mathrm{th}(t) + \vb d_\mathrm{ind}(t) \label{seq:total_fluc_dipole}, \\
\vb E(\gpv, t) & = \vb E_\mathrm{th}(\gpv, t) + \vb E_\mathrm{tw}(\gpv, t) + \vb E_\mathrm{ind}(\gpv, t). \label{seq:total_fluc_field}
\end{align}
In order to describe how the induced electric dipole moment depends on the electromagnetic field and vice versa, it is convenient to transform to frequency domain, defined by
\begin{align}
\vb d(\omega) & = \int_{-\infty}^\infty \mathrm d t \, \vb d(t) \mathrm e^{\mathrm i \omega t}, \label{seq:freq_dipole_def} \\
\vb E(\vb r, \omega) & = \int_{-\infty}^\infty \mathrm d t \, \vb E(\vb r, t) \mathrm e^{\mathrm i \omega t}\, . \label{seq:freq_field_def}
\end{align}
The electromagnetic field emitted by the point dipole can be expressed as
\begin{equation}\label{seq:induced_field}
\vb E_\mathrm{ind}(\gpv, \omega) = \tensor{G}(\gpv, \dpv, \omega)\vb d(\omega) = \tensor{G}_0(\gpv, \dpv, \omega)\vb d(\omega) + \tensor G_1(\gpv, \dpv, \omega)\vb d(\omega).
\end{equation}
Here $\tensor{G}(\gpv, \dpv, \omega)$ is Green's tensor for the electromagnetic field that describes the field at position $\gpv$ emitted by the point dipole at position $\dpv$ in the presence of the surrounding material. In the second equality we used $\tensor{G}(\gpv, \dpv, \omega) = \tensor{G}_0(\gpv, \dpv, \omega) + \tensor{G}_1(\gpv, \dpv, \omega)$, where $\tensor{G}_0(\gpv, \dpv, \omega)$ is the Green's tensor in vacuum, namely in absence of any material. The Green's tensor $\tensor{G}_1(\gpv, \dpv, \omega) $ is often referred to as the scattering part of the total Green's tensor.
The induced electric dipole moment can be expressed as
\begin{equation}\label{seq:induced_dipole}
\begin{aligned}[b]
\vb d_\mathrm{ind}(\omega) & = \alpha(\omega)\vb E_\mathrm{ex}(\dpv,\omega) \\
& = \alpha(\omega) \left[ \vb E(\dpv, \omega) - \tensor{G}_0(\dpv, \dpv,\omega)\vb d(\omega) \right] \\
& = \alpha(\omega) \left[ \vb E_\mathrm{th}(\dpv, \omega) + \vb E_\mathrm{tw}(\dpv, \omega) + \tensor{G}_1(\dpv, \dpv,\omega)\vb d(\omega) \right].
\end{aligned}
\end{equation}
Here $\alpha(\omega)$ is the scalar polarizability of the electric point dipole. The electromagnetic field inducing the dipole is given by the so called exciting field defined as $\vb E_\mathrm{ex}(\gpv,\omega) \equiv \vb E(\gpv, \omega) - \tensor{G}_0(\gpv, \dpv,\omega)\vb d(\omega)$. In this way, the dipole can be induced either by the external thermal field, the laser beam, or by the field emitted by the dipole that is reflected by the nearby material.
The instantaneous electromagnetic force acting on the point dipole can be evaluated as~\cite{novotny_PrinciplesNanoOptics_2006}
\begin{equation}
\vb F(t) = \sum_{j\in\{x,y,z\}} d_j(t) \left.\nabla_{\gpv} E_{\mathrm{ex},j}(\gpv, t)\right|_{\gpv=\dpv} + \dfdx{}{t}\left[\vb d(t) \cdot \vb B(\dpv, t)\right] = \nabla_{\gpv} \left[\vb d(t) \cdot \vb E_{\mathrm{ex}}(\gpv, t)\right]_{\gpv=\dpv} + \dfdx{}{t}\left[\vb d(t) \cdot \vb B(\dpv, t)\right].
\end{equation}
This is a fluctuating force due to the thermal fluctuations of the dipole $\vb d_\mathrm{th}(t)$ and the electric field $\vb E_\mathrm{th}(\gpv, t)$. We are interested in the time-averaged value of the force $\vb F = \expval{\vb F(t)}$. The thermally fluctuating quantities are assumed to be both stationary and ergodic, hence the time average is equivalent to an ensemble average. Stationarity implies that the total time derivative of the second term in $\vb F(t)$ is zero under time average. Thus, one has
\begin{equation} \label{seq:definition_force}
\vb F = \expval{\nabla_{\gpv} \left [ \vb d(t) \cdot \vb E_{\mathrm{ex}}(\gpv, t)\right ]_{\gpv=\dpv}} = \frac{1}{4 \pi^2} \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} \mathrm d \omega \mathrm d \omega' \left.\nabla_{\gpv} \expval{\vb d(\omega) \cdot \vb E^*_{\mathrm{ex}}(\gpv, \omega')}\right|_{\gpv=\dpv},
\end{equation}
where the quantity $\expval{\vb d(\omega) \cdot \vb E^*_{\mathrm{ex}}(\gpv, \omega')}$ is the ensemble average over the realisations of the fluctuations in frequency space.
In order to evaluate the frequency correlators in the integrand of \eqnref{seq:definition_force}, we insert \eqnref{seq:induced_field} and \eqnref{seq:induced_dipole} into the frequency-domain versions of \eqnref{seq:total_fluc_dipole} and \eqnref{seq:total_fluc_field} and solve for $\vb d(\omega)$ and $\vb E(\gpv, \omega)$. The result is
\begin{align}
\label{seq:total_dipole_from_fluc}
\vb d(\omega) & = \tensor T(\dpv,\omega)\left\{ \vb d_\mathrm{th}(\omega) + \alpha(\omega) [\vb E_\mathrm{th}(\dpv, \omega) + \vb E_\mathrm{tw}(\dpv, \omega)]\right\} ,\\
\label{seq:ex_field_from_fluc}
\vb E_\mathrm{ex}(\gpv,\omega) & = \vb E_\mathrm{th}(\gpv,\omega) + \vb E_\mathrm{tw}(\gpv, \omega) + \tensor G_1(\gpv,\dpv,\omega)\tensor T(\dpv, \omega)\left\{\vb d_\mathrm{th}(\omega) + \alpha(\omega)[\vb E_\mathrm{th}(\dpv,\omega) + \vb E_\mathrm{tw}(\dpv,\omega)]\right\},
\end{align}
where
\begin{equation}
\tensor T (\dpv,\omega) \equiv \left [ \ut - \alpha (\omega) \tensor G_1(\dpv, \dpv, \omega)\right ]^{-1}.
\label{seq:dresstensor_def}
\end{equation}
The tensor $\tensor T$ accounts for the ability of the dipole to scatter the radiation that it previously has emitted and that returns to the dipole after having been reflected by the environment (e.g.\ surface).
To understand the action of $\tensor T$ in more detail, we can introduce the ``bare'' dipole moment $\vb d^{(0)}$ as
\begin{equation}
\vb d^{(0)}(\omega) = \vb d_\mathrm{th}(\omega) + \alpha(\omega)[\vb E_\mathrm{th}(\dpv, \omega) + \vb E_\mathrm{tw}(\dpv, \omega)].
\end{equation}
Next, recall that $\tensor G \vb d^{(0)}$ is the field radiated by the bare dipole, and in particular, that $\tensor G_1 \vb d^{(0)}$ is the radiation from the bare dipole that is subsequently reflected by the environment. The dipole moment that this field induces when it returns to the dipole is $\alpha \tensor G_1 \vb d^{(0)}$. Therefore, we introduce the $n\text{:th}$ induced dipole moment as
\begin{equation}
\vb d^{(n)}(\omega) = \alpha(\omega)\tensor G_1(\dpv,\dpv,\omega)\vb d^{(n-1)}(\omega).
\label{seq:nth_induced_dipole}
\end{equation}
In words, the dipole $\vb d^{(n)}$ is the dipole moment that gets induced when radiation from the dipole $\vb d^{(n-1)}$ gets reflected by the environment and returns to the location of the point dipole. Then we see that \eqnref{seq:total_dipole_from_fluc} can be formally written as
\begin{equation}
\vb d(\omega) = \tensor T(\dpv,\omega)\vb d^{(0)}(\omega) = \vb d^{(0)}(\omega) + \sum_{n=1}^\infty \vb d^{(n)}(\omega).
\end{equation}
In words, the total dipole moment is the bare dipole moment plus all the dipole moments that get induced in sequence as multiple reflections between the dipole and the environment happen.
We then see that \eqnref{seq:ex_field_from_fluc} states that the total field has independent contributions from the thermal field, laser field, and the field radiated by the dipole including all the multiple scattering events of the dipole.
For sufficiently small polarizability, namely for $||\alpha \tensor G_1|| \ll 1$, one can neglect the contributions of the induced dipole moments to the total dipole moment, i.e.~the dipole's rescattering of its own radiation can be neglected. In this case, expressions of the forces can be approximated by using $\tensor T \approx \ut$.
Under this approximation, the field emitted by the bare dipole only gets reflected once -- by the environment. Roughly speaking, the dipole feels the force from the radiation reflected by the environment, but it does not emit additional radiation due to the interaction. We will use this approximation later.
The frequency correlators in \eqref{seq:definition_force} can now be evaluated in terms of the frequency correlators of the fluctuating and driving fields using \eqnref{seq:total_dipole_from_fluc} and \eqnref{seq:ex_field_from_fluc}. We assume that the cross-correlations of $\vb E_\mathrm{th}$, $\vb E_\mathrm{tw}$, and $\vb d_\mathrm{th}$ are zero, so that any expectation value with a combination of these quantities vanishes. Therefore, the only terms that are nonzero are those containing any of the following three correlators:
\begin{gather}
\begin{aligned}[b]
\expval{d_{\mathrm{th},j}(\omega) d_{\mathrm{th},k}^*(\omega')} = 2\pi\hbar\delta(\omega-\omega') \coth{\left(\frac{\hbar\omega}{2k_\mathrm B T}\right)} \delta_{jk} \Im{\alpha(\omega)},
\label{seq:fdrel_dipole}
\end{aligned}
\\
\begin{aligned}[b]
\expval{E_{\mathrm{th},j}(\dpv,\omega) E^*_{\mathrm{th},k}(\gpv,\omega')} = 2\pi\hbar\delta(\omega-\omega')\coth{\left(\frac{\hbar\omega}{2k_\mathrm B T_\mathrm{env}}\right)}\Im{G_{jk}(\dpv, \gpv, \omega)},
\label{seq:fdrel_field}
\end{aligned}
\\
\expval{E_{\mathrm{tw},j}(\dpv, \omega) E_{\mathrm{tw},k}^*(\gpv, \omega')} = \delta(\omega-\omega') \lim_{\Delta\omega\to0}\int_{\omega - \Delta\omega}^{\omega + \Delta\omega} \mathrm d\omega''\,E_{\mathrm{tw},j}(\dpv, \omega) E_{\mathrm{tw},k}^*(\gpv, \omega'').
\label{seq:cfield_psd}
\end{gather}
The thermal correlators \eqnref{seq:fdrel_dipole} and \eqnref{seq:fdrel_field} are obtained using the fluctuation-dissipation theorem at the temperature of the point dipole and the electromagnetic field environment respectively. In the nonequilibrium situation $T \neq T_\mathrm{env}$ (i.e.\ when the point dipole has a different temperature than its environment), this assumes local equilibrium. The relation \eqnref{seq:cfield_psd} holds for any field which is stationary after a time-average is performed. In the formulas that follow, we make the integral in \eqnref{seq:cfield_psd} implicit by defining:
\begin{equation}
\lim_{\Delta\omega\to0}\int_{\omega - \Delta\omega}^{\omega + \Delta\omega} \mathrm d\omega''\,E_{\mathrm{tw},j}(\dpv, \omega) E_{\mathrm{tw},k}^*(\gpv, \omega'') \equiv E_{\mathrm{tw},j}(\dpv, \omega) \bar E_{\mathrm{tw},k}^*(\gpv, \omega).
\end{equation}
Putting everything together, we identify five contributions to the mean force:
\begin{equation}
\vb F = \vb F_\mathrm{trap} + \vb F_\mathrm{cs} + \vb F_\mathrm{zpf} + \vb F_\mathrm{env} + \vb F_\mathrm{rad}.
\label{seq:total_force}
\end{equation}
For each term on the right-hand side, we now give the general expression of the term, the expression within the $\tensor T\to \ut$ approximation, and a physical explanation of the term's appearance.
(1) The first term $\vb F_\mathrm{trap}$ is given by
\begin{equation}
\begin{aligned}[b]
\vb F_\mathrm{trap} & = \begin{aligned}[t]
\int_{0}^\infty\frac{\mathrm d\omega}{2\pi^2} \nabla_{\gpv}
\Bigl[ & \Re{\alpha(\omega)}\Re{\bar{\vb E}_\mathrm{tw}^*(\vb r,\omega) \cdot \tensor T(\vb r_0, \omega) \vb E_\mathrm{tw}(\dpv, \omega)}_{\gpv=\dpv} \\
& + \Im{\alpha(\omega)}\Im{\bar{\vb E}_\mathrm{tw}^*(\vb r,\omega) \cdot \tensor T(\vb r_0, \omega) \vb E_\mathrm{tw}(\dpv, \omega)}_{\gpv=\dpv}\Bigr]
\end{aligned} \\
& \approx \int_{0}^\infty\frac{\mathrm d\omega}{2\pi^2} \left[\frac{1}{2} \Re{\alpha(\omega)} \nabla_{\dpv} \left[\bar{\vb E}^*_\mathrm{tw}(\dpv,\omega) \cdot \vb E_\mathrm{tw}(\dpv,\omega)\right] + \Im{\alpha(\omega)} \nabla_\gpv \Im{\bar{\vb E}^*_\mathrm{tw}(\dpv,\omega) \cdot \vb E_\mathrm{tw}(\gpv,\omega)}_{\gpv=\dpv}\right] \quad\quad ({\tensor T \to \ut}).
\end{aligned}
\label{seq:trap_force}
\end{equation}
This term describes the optical gradient and scattering force exerted by the laser beam. In our proposal, this term will be used to trap the particle (e.g. optical tweezer). We remark that the field $\vb E_\mathrm{tw}$ describes the laser beam in the presence of the surrounding material but in the absence of the dipole.
(2) The second term $\vb F_\mathrm{cs}$ is given by
\begin{equation}
\begin{aligned}[b]
\vb F_\mathrm{cs} & = \int_{0}^\infty\frac{\mathrm d\omega}{2\pi^2}\,|\alpha(\omega)|^2\nabla_\gpv\Re{\bar{\vb E}_\mathrm{tw}^*(\dpv,\omega) \cdot \tensor T^\dagger(\dpv,\omega)\tensor G_1^\dagger(\gpv,\dpv,\omega) \tensor T(\dpv,\omega) \vb E_\mathrm{tw}(\dpv,\omega)}_{\gpv=\dpv} \\
& \approx \int_{0}^\infty\frac{\mathrm d\omega}{2\pi^2}\, |\alpha(\omega)|^2 \nabla_{\gpv}\Re{\bar{\vb E}_\mathrm{tw}^*(\dpv,\omega) \cdot \tensor G_1^\dagger(\gpv, \dpv, \omega) \vb E_\mathrm{tw}(\dpv,\omega)}_{\gpv=\dpv} \quad\quad (\tensor T \to \ut).
\label{seq:cs_force}
\end{aligned}
\end{equation}
This force arises due to the nanoparticle scattering light from the trapping laser beam, which is subsequently reflected from the environment, and returns to interact with the nanoparticle's dipole moment. This is very similar to the force observed in~\cite{rieser_Science_377:6609_2022}, but here the interaction is between the nanoparticle and its optical mirror image rather than between two distinct nanoparticles. Note that the force depends on the orientation of the electric field at the position of the dipole, that is, on the polarization of the laser beam.
(3) The third term $\vb F_\mathrm{zpf}$ is given by
\begin{equation}
\begin{aligned}[b]
\vb F_\mathrm{zpf} & = \begin{aligned}[t]
\int_0^\infty \frac{\hbar \mathrm d\omega}{\pi} \nabla_\gpv \mathrm{Tr}\biggl\{ & \Re{\alpha(\omega)\tensor T(\dpv,\omega)} \Im{\tensor G(\dpv, \gpv,\omega)} \\
& + \Re{\tensor T^\dagger(\dpv,\omega) \tensor G_1(\gpv, \dpv,\omega)\tensor T(\dpv,\omega)}\left(\Im{\alpha(\omega)} + |\alpha(\omega)|^2\Im{\tensor G(\dpv, \dpv,\omega)}\right)\biggr\}_{\gpv=\dpv}
\end{aligned}\\
& \approx \int_0^\infty \frac{\hbar \mathrm d\omega}{2\pi} \left\{\Im{\alpha(\omega) \nabla_{\dpv} \operatorname{Tr}\left\lbrace\tensor G_1(\vb r_0, \vb r_0, \omega)\right\rbrace} + 2|\alpha(\omega)|^2\nabla_\gpv\operatorname{Tr}\left[\Re{\tensor G_1(\gpv, \dpv,\omega)}\Im{\tensor G(\dpv, \dpv,\omega)}\right]_{\gpv=\dpv}\right\}.
\end{aligned}
\label{seq:zpf_force}
\end{equation}
This force arises due to the quantum zero-point fluctuations of the dipole moment of the nanoparticle correlating with the zero-point fluctuations of the environmental electromagnetic field.
In the $\tensor T \to \ut$ approximation, the first and second terms combine into the $\Im{\alpha \operatorname{Tr}\left\{\tensor G_1\right\}}$ term. It is typically this contribution to the force that is referred to by the name zero-temperature Casimir--Polder force~\cite{wylie_Phys.Rev.A_30:3_1984}.
(4) The fourth term $\vb F_\mathrm{env}$ contains all the dependence on the environmental temperature $T_\mathrm{env}$ and is given by
\begin{equation}
\begin{aligned}[b]
\vb F_\mathrm{env} & =
\begin{aligned}[t]
\int_0^\infty \frac{2\hbar \mathrm d\omega}{\pi} n(T_\mathrm{env},\omega)\nabla_\gpv \mathrm{Tr}\biggl\{ & \Re{\alpha(\omega)\tensor T(\dpv,\omega)} \Im{\tensor G(\dpv, \gpv,\omega)} \\
& + |\alpha(\omega)|^2\Re{\tensor T^\dagger(\dpv,\omega) \tensor G_1(\gpv, \dpv,\omega)\tensor T(\dpv,\omega)}\Im{\tensor G(\dpv, \dpv,\omega)}\biggr\}_{\gpv=\dpv}
\end{aligned} \\
& \begin{aligned}[b]
\approx \int_0^\infty \frac{\hbar \mathrm d\omega}{\pi} n(T_\mathrm{env}, \omega) \Bigl\{
& \Re{\alpha(\omega)}\Im{ \nabla_{\dpv} \operatorname{Tr}\left\{\tensor G(\vb r_0, \vb r_0, \omega)\right\}} \\
& + 2|\alpha(\omega)|^2\nabla_\gpv \operatorname{Tr}\Bigl[\Re{\tensor G_1(\gpv, \dpv,\omega)}\Im{\tensor G(\dpv, \dpv,\omega)}\Bigr]_{\gpv=\dpv}\Bigr\} \quad\quad (\tensor T\to\ut).
\end{aligned}
\end{aligned}
\label{seq:env_force}
\end{equation}
This force arises due to environmental thermal radiation being scattered by the dipole. The first term in the force describes a gradient-type force on the total dipole due to the gradient of the thermal field intensity, and the second term is the thermal version of $\vb F_\mathrm{cs}$, where the dipole scatters light from the thermal field which is then reflected and returns to interact with the dipole.
(5) Finally, the fifth term $\vb F_\mathrm{rad}$ contains all dependence on the particle temperature $T$ and is given by
\begin{equation}
\begin{aligned}[b]
\vb F_\mathrm{rad} & = \int_0^\infty\frac{2\hbar\mathrm d\omega}{\pi}\, n(T,\omega) \Im{\alpha(\omega)}\nabla_{\gpv}\Re{\operatorname{Tr}\left[\tensor T^\dagger(\dpv,\omega) \tensor G_1^\dagger(\gpv,\dpv,\omega) \tensor T(\dpv,\omega)\right]}_{\gpv=\dpv} \\
& \approx \int_0^\infty\frac{\hbar\mathrm d\omega}{\pi}\, n(T,\omega)\Im{\alpha(\omega)}\nabla_{\dpv}\Re{\operatorname{Tr}\left [\tensor G_1^\dagger(\dpv,\dpv,\omega)\right]} \quad\quad (\tensor T\to\ut).
\end{aligned}
\label{seq:bb_force}
\end{equation}
where $n(T,\omega) = [\exp{(\hbar\omega/k_\mathrm B T)} - 1]^{-1}$ and in the second line we used $\nabla_\gpv \operatorname{Tr}[\tensor G(\vb r,\vb r_0)] = \nabla_{\dpv} \operatorname{Tr}[ \tensor G(\vb r_0,\vb r_0)]/2$.
This force arises due to thermal radiation emitted by the dipole which is reflected by the environment and returns to interact with the dipole.
In order for this force to be non-vanishing, the thermal radiation that returns to the dipole must retain some coherence with the dipole. This is possible for \emph{colored} thermal emitters, meaning that $\Im{\alpha(\omega)}$ is peaked in the infrared part of the spectrum. We emphasize that the gradient and scattering force that the dipole would experience due to its reflected field are not included in the $\tensor T \to \ut$ expression, but are part of negligible higher-order terms. To see this, it is useful to use the $n$:th induced thermal dipole moment and field akin to \eqnref{seq:nth_induced_dipole}: $\vb d_\mathrm{th}^{(n)} = \alpha \vb E_\mathrm{th}^{(n-1)}$ and $\vb E^{(n)}_\mathrm{th} = \tensor G_1 \vb d_\mathrm{th}^{(n)}$. Then,
\begin{equation}
\vb F_\mathrm{rad} \propto \expval{\vb d_\mathrm{th}^* \cdot \tensor G_1 \vb d_\mathrm{th}} = \expval{\left(\tensor T \vb d^{(0)}_\mathrm{th}\right)^* \cdot \tensor G_1 \tensor T \vb d^{(0)}_\mathrm{th}}
= \expval{{\vb d^{(0)}_\mathrm{th}}^* \cdot \vb E^{(0)}_\mathrm{th}} + \expval{{\vb d^{(0)}_\mathrm{th}}^*\cdot \vb E^{(1)}_\mathrm{th}} + \alpha\expval{{\vb E^{(0)}_\mathrm{th}}^* \cdot \vb E^{(0)}_\mathrm{th}} + \dots
\end{equation}
where in the last expression we included terms up to first order in $\vb d^{(1)}_\mathrm{th}$. The term ${\vb d^{(0)}_\mathrm{th}}^*\cdot \vb E^{(0)}_\mathrm{th}$ is the coherent interaction included in the $\tensor T \to \ut$ expression, the term ${\vb d^{(0)}_\mathrm{th}}^*\cdot \vb E^{(1)}_\mathrm{th}$ is the next-order coherent interaction term, while the term $\alpha{\vb E^{(0)}_\mathrm{th}}^*\cdot \vb E^{(0)}_\mathrm{th}$ is the lowest-order gradient and scattering force on the nanoparticle due to the thermal emission. As we will see later, in Fig.~\ref{sfig:small_limit_check}, the last two terms are negligible compared to the first term.
\section{Evaluation of forces for a nanoparticle in a plane geometry}
We explain how we evaluate the forces in the particular scenario of an optically trapped nanoparticle in front of a surface. The scattering Green's function above a planar interface is given by~\cite{buhmann_DispersionForces_2012}
\begin{equation}
\tensor G_1(\vb r, \vb r',\omega) = \frac{\mathrm i}{8\pi^2 \epsilon_0} \int_0^\infty \mathrm d k_\parallel \frac{k_\parallel}{k_z} \int_0^{2\pi}\mathrm d\theta\,\exp{\left\{\mathrm i k_\parallel (\cos{\theta} \uv x + \sin{\theta}\uv y) \cdot (\vb r - \vb r') + \mathrm i k_z(z+z')\right\}}\tensor M(k_\parallel,\theta,\omega),
\label{seq:plane_scattering_gf}
\end{equation}
with $k_z = \left[(\omega/c)^2 - k_\parallel^2\right]^{1/2}\ (\Im{k_z}>0)$, and
\begin{equation}
\tensor M(k_\parallel,\theta,\omega) = \left(\begin{array}{ccc}
k^2r_s \sin^2{\theta} -k_z^2 r_p\cos^2\theta & -k^2r_s \sin\theta\cos\theta - k_z^2 r_p\sin \theta \cos \theta &
-k_p k_z r_p\cos\theta \\
-k^2 r_s\sin\theta\cos\theta - k_z^2 r_p\sin\theta\cos\theta & k^2 r_s \cos^2\theta - k_z^2 r_p\sin^2\theta &
-k_p k_z r_p \sin\theta \\
k_p k_z r_p\cos\theta & k_p k_z r_p\sin\theta & k_p^2 r_p \\
\end{array}\right),
\label{seq:plane_scattering_M}
\end{equation}
where $r_{s,p}=r_{s,p}(k_\parallel,\omega)$ are the Fresnel reflection coefficients for $s$ and $p$ polarizations. $\tensor G_1(\vb r_0, \vb r_0, \omega)$ becomes a diagonal matrix after evaluation of the $\theta$-integral, since $\int_0^{2\pi}\tensor M(k_\parallel,\theta,\omega) \,\mathrm d\theta \equiv \overline{\tensor M}(k_\parallel,\omega)$ is a diagonal matrix. Consequently, $\tensor T(\dpv,\omega) = [1 - \alpha \tensor G_1(\dpv, \dpv, \omega)]^{-1}$ is also a diagonal matrix, and the trace in \eqnref{seq:bb_force} can be evaluated as the trace of $\tensor G_1$ weighted with the components of the diagonal matrices:
\begin{equation}
\operatorname{Tr}\left\{\tensor T^\dagger(\dpv,\omega) \tensor G_1(\gpv, \dpv, \omega)^\dagger \tensor T(\dpv,\omega)\right\} = \sum_{j\in\{x,y,z\}} G^*_{1,jj}(\gpv,\dpv,\omega) |T_{jj}(\dpv,\omega)|^2.
\end{equation}
Now,
\begin{equation}
\begin{aligned}[b]
\lim_{\gpv \to \dpv}\nabla_{\gpv}G_{1,jj}(\gpv, \dpv, \omega) & = \frac{\mathrm i}{8\pi^2 \epsilon_0} \int_0^\infty \mathrm d k_\parallel\frac{k_\parallel}{k_z(k_\parallel)} \mathrm e^{2 \mathrm i k_z(k_\parallel) z_0} \int_0^{2\pi}\mathrm d\theta \left(\begin{matrix} \mathrm i k_\parallel \cos{\theta} \\ \mathrm i k_\parallel \sin{\theta} \\ \mathrm i k_z(k_\parallel)
\end{matrix}\right) M_{jj}(k_\parallel,\theta,\omega) \\
& = \frac{-\uv z}{8\pi^2 \epsilon_0} \int_0^\infty \mathrm d k_\parallel\, k_\parallel \mathrm e^{2 \mathrm i k_z(k_\parallel) z_0} \overline{M}_{jj}(k_\parallel,\omega),
\end{aligned}
\label{seq:plane_diagonal_derivative_evaluation}
\end{equation}
since the $\uv x$ and $\uv y$ components in the middle equation integrate to zero. We remark that, as the right-hand-side of \eqnref{seq:plane_diagonal_derivative_evaluation} is equal to $\nabla_{\dpv} G_{1,jj}(\vb r_0,\vb r_0)/2$, this is an explicit demonstration of the general relationship $\lim_{\vb r \to \vb r_0}\nabla_{\vb r} G_{1,jj}(\vb r, \vb r_0, \omega) = \nabla_{\dpv} G_{1,jj}(\vb r_0, \vb r_0, \omega)/2$ that we used to simplify the $\tensor T=1$ expressions in Section~\ref{s:force_derivation}. Consequently,
\begin{multline}
\lim_{\gpv \to \dpv}\nabla_{\gpv} \operatorname{Tr}\left\{\Re{\tensor T^\dagger(\dpv,\omega) \tensor G_1(\gpv, \dpv, \omega)^\dagger \tensor T(\dpv,\omega)}\right\} \\ = \frac{-\uv z}{8\pi^2 \epsilon_0} \sum_{j\in\{x,y,z\}} \Re{\int_0^\infty \mathrm d k_\parallel\, k_\parallel \mathrm e^{2\mathrm i k_z(k_\parallel) z_0} \overline{M}_{jj}(k_\parallel,\omega)} |T_{jj}(\dpv,\omega)|^2,
\label{seq:diagonal_trace_derivative}
\end{multline}
with
\begin{gather}
\overline M_{xx}(k_\parallel,\omega) = \overline M_{yy}(k_\parallel,\omega) = \pi((\omega/c)^2 r_s(k_\parallel,\omega) - k_z^2(k_\parallel) r_p(k_\parallel,\omega)),\quad \overline M_{zz}(k_\parallel,\omega) = 2\pi k_\parallel^2 r_p(k_\parallel,\omega), \label{seq:barM_general} \\
\left(T_{jj}\right)^{-1} = 1-\frac{\mathrm i\alpha(\omega)}{8\pi^2\epsilon_0}\int_0^\infty \mathrm d k_\parallel\,\frac{k_\parallel}{k_z(k_\parallel)}\, \mathrm e^{2\mathrm i k_z(k_\parallel) z_0} \overline{M}_{jj}(k_\parallel,\omega).
\label{seq:Tjj_general}
\end{gather}
Insertion of Eqs.~(\ref{seq:diagonal_trace_derivative}-\ref{seq:Tjj_general}) into \eqnref{seq:bb_force} would now allow numerical evaluation of the integrals. However, we proceed to the special case of a perfectly reflecting plane which has $r_s(k_\parallel,\omega)=-1,\,r_p(k_\parallel,\omega)=1$. In this case the integrals over $k_\parallel$ can be performed analytically, and we obtain, with $k=\omega/c$
\begin{gather}
\lim_{\vb r\to\vb r_0}\nabla_\vb r G_{1,jj}(\vb r,\vb r_0,\omega) = -\frac{\uv z \mathrm e^{2\mathrm i k z_0}}{64\pi \epsilon_0 z_0^4}\left\{
\begin{array}{ll}
3 - 6 \mathrm i k z_0 - 8 k^2 z_0^2 + 8 \mathrm i k^3 z_0^3 \quad \ & j=x,y \\
6 - 12 \mathrm i k z_0 - 8 k^2 z_0^2 \quad \ & j=z
\end{array}
\right. \label{seq:nablaG_jj_mirror}\\
\left(T_{jj}(\dpv,\omega)\right)^{-1} = \left\{
\begin{array}{ll}
1+\alpha(\omega)\mathrm e^{2 \mathrm i k z_0} \left(-1 + 2 \mathrm i k z_0 + 4 k^2 z_0^2\right)/(32 \pi \epsilon_0 z_0^3) \quad \ & j=x,y \\
1+\mathrm i\alpha(\omega) \mathrm e^{2 \mathrm i k z_0} (\mathrm i + 2 k z_0)/(16 \pi \epsilon_0 z_0^3) \quad \ & j=z
\end{array}
\right. \label{seq:T_jj_mirror}
\end{gather}
Usage of these relations with \eqnref{seq:diagonal_trace_derivative} leaves only the $\omega$ integral to be evaluated in \eqnref{seq:bb_force}. We use these relations to investigate the accuracy of the $\tensor T=1$ version of \eqref{seq:bb_force} to the expression including $\tensor T$. The $\tensor T=1$ expansions of Eqs.~(\ref{seq:cs_force}-\ref{seq:zpf_force}) are conveniently evaluated for a perfectly reflecting plane by using
\begin{equation}
\nabla_{\dpv} \operatorname{Tr}\left\{\tensor G_1(\vb r_0, \vb r_0, \omega)\right\} = \frac{\uv z}{{8 \pi \epsilon_0 z_0^4}}\mathrm e^{2 \mathrm i k z_0} \left(-3 + 6 \mathrm i k z_0 + 6 k^2 z_0^2 - 4 \mathrm i k^3 z_0^3\right)
\label{seq:plane_tr_nabla_G}
\end{equation}
which can be obtained from \eqnref{seq:nablaG_jj_mirror} combined with $\lim_{\vb r \to \vb r_0}\nabla_{\vb r} G_{1,jj}(\vb r, \vb r_0, \omega) = \nabla_{\dpv} G_{1,jj}(\vb r_0, \vb r_0, \omega)/2$, or from the method of images. We compare the $\tensor T\neq\ut$ and $\tensor T=\ut$ results for $\vb F_\mathrm{rad}$ in Figure~\ref{sfig:small_limit_check}. We find agreement between the full and expanded results, as expected for a subwavelength particle since it fulfills $||\alpha \tensor G_1||\ll 1$. Therefore we use the $\tensor T=\ut$ versions of all forces in the rest of this work.
\begin{figure}
\centering
\includegraphics{suppfig/Frad_exact_vs_single.pdf}
\caption{Comparison of $\vb F_\mathrm{rad}$ calculated using the $\tensor T \neq \ut$ expression ($\vb F_\mathrm{rad}^{(\infty)}$) and the $\tensor T = \ut$ expansion ($\vb F_\mathrm{rad}^{(1)}$) of \eqnref{seq:bb_force}. Top panel: Darker solid lines are the exact result and lighter dotted lines are the $\tensor T = \ut$ results. Bottom panel: the difference of these results as a fraction of $\vb F_\mathrm{rad}^{(\infty)}$ is displayed with colors matching temperatures in the top panel.}
\label{sfig:small_limit_check}
\end{figure}
We evaluate the total force on the nanoparticle $\vb F$ displayed in Figure~\ref{fig2} in the main text as follows: Starting from \eqref{seq:total_force}, we neglect $\vb F_\mathrm{trap}$ since its later effect on the PSD is captured by the trapping frequency $\Omega$ and equilibrium position $\vb r_0$. $\vb F_\mathrm{rad}$ and $\vb F_\mathrm{env}$ were calculated using \eqref{seq:bb_force} and \eqref{seq:env_force} with \eqref{seq:plane_tr_nabla_G} and $\alpha(\omega)$ as displayed in \subfigref{1}{c} in the main text. In \eqref{seq:env_force}, we neglect the term $\propto |\alpha(\omega)|^2$. We performed the integrals numerically using the Gauss--Kronrod 21-point rule as implemented in the \texttt{scipy.integrate.quad\_vec} method of SciPy 1.8.0~\cite{scipy1.0, piessens_QuadpackSubroutine_2012}. As integration bounds we used the upper and lower bounds $\lambda_\mathrm{max}=\SI{50}{\micro m}$ and $\lambda_\mathrm{min}=\SI{7}{\micro m}$ of the $\epsilon(\lambda)$ dataset~\cite{kitamura_Appl.Opt._46:33_2007}, translated into frequency space as $\omega_\mathrm{min,max} = 2\pi c/\lambda_\mathrm{max,min}$. We evaluate $\vb F_\mathrm{zpf}$ with the far-field approximation of the Casimir force
\begin{equation}
\vb F_\mathrm{zpf}(z \gg \nprad) = -\frac{\hbar c \alpha(0)}{8\pi^2 z^5}\uv{z}
\label{seq:far_field_casimir_force}
\end{equation}
where $\nprad$ is the nanoparticle radius and with $\alpha(0)$ derived from the electrostatic value $\epsilon(0) = 3.8$ of the permittivity of \ch{SiO2}. The well-known formula \eqnref{seq:far_field_casimir_force} can be derived analytically from \eqnref{seq:zpf_force} (neglecting the integrand term $\propto |\alpha(\omega)|^2$) by performing a Wick rotation of the integral in \eqref{seq:zpf_force} onto the imaginary axis to obtain a non-oscillating, rapidly decaying integrand and then expanding the result for $z > \lambda$ where $\lambda$ is the minimum wavelength appearing in the integral~\cite{casimirInfluenceRetardationLondonvan1948, wylie_Phys.Rev.A_30:3_1984}. As a consistency check, we compared a full numerical evaluation of the Wick-rotated \eqref{seq:zpf_force} (translating $\alpha(\omega)$ to the imaginary axis using the Schwarz formula $\alpha(\mathrm i\omega) = \frac{2}{\pi}\int_0^\infty \phi\Im{\alpha(\phi)}/(\phi^2 + \omega^2) \, \mathrm d\phi, \ \omega\in\mathbb{R}$) to the approximation \eqnref{seq:far_field_casimir_force} and found excellent agreement, confirming that the Casimir force is dominated by the electrostatic value of the polarizability also in the present case.
To evaluate $\vb F_\mathrm{cs}$, we approximate the optical tweezer by the time-harmonic function $\vb E_\mathrm{tw}(\vb r, t) = E_\mathrm{tw}(\vb r)\uv x\cos{(\omega_0 t)}$, where we assume the tweezer beam to be polarized parallel to the plane, consistent with the geometry of our proposed setup. This enables us to remove the integral in \eqref{seq:cs_force} to find
\begin{equation}
\vb F_\mathrm{cs} = \frac{|\alpha(\omega_0)|^2}{2}|E_\mathrm{tw}(\vb r_0)|^2\Re{\lim_{\gpv\to\dpv}\nabla_\gpv G_{1,xx}(\vb r, \vb r_0, \omega_0)}.
\end{equation}
The field intensity at the center of the tweezer is given in terms of the experimental parameters as $|E_\mathrm{tw}(\dpv)|^2 = 2\pi N^2 P_\mathrm{tw} / \epsilon_0 c \lambda_0^2$.
We calculate the $\vb F_\mathrm{cs}$ that would result if the surface was perfectly reflecting at $\lambda_0 = 2\pi c/\omega_0$. The force would then become
\begin{equation}
\vb F_\mathrm{cs} = -\uv z\left|\frac{\alpha(\lambda_0)}{\epsilon_0}\right|^2\frac{N^2 P_\mathrm{tw}}{64 c \lambda_0^2 z_0^4}j_x(2\pi z_0/\lambda_0)
\end{equation}
with $j_x(x) = \Re{\exp{(2\mathrm i x)}(3-6\mathrm i x-8x^2+8\mathrm i x^3)}$. We plot this force in Figure~\ref{sfig:fcs}. However, in our scenario the surface is transparent at $\omega_0$ and the optical interaction force will be weaker. We approximate the magnitude of the optical interaction force above our non-perfectly reflecting surface as $|r_\perp(\omega_0)\vb F_\mathrm{cs}|$, where $r_\perp(\omega_0)$ is the reflection coefficient of the surface at $\omega_0$ and normal incidence, and the symbol $\vb F_\mathrm{cs}$ retains its meaning as the coherent scattering force above a perfectly reflecting plane.
\begin{figure}
\includegraphics{suppfig/coherent_scattering_force_x_polarization.pdf}
\caption{$\vb F_\mathrm{cs}$ for a nanoparticle above a surface which reflects the tweezer wavelength $\lambda_0$ perfectly. Solid/dashed lines indicate that the force is positive/negative in the normal direction of the surface.}
\label{sfig:fcs}
\end{figure}
We model the modulation of the surface reflectivity in time as follows. We assume that the presence of the surface gives rise to a Green's tensor which is effectively the Green's tensor above a perfectly reflecting surface in the range $(\omega_\mathrm{peak} - \Gamma, \omega_\mathrm{peak} + \Gamma)$ (where $\omega_\mathrm{peak} = 2\pi c/\lambda_\mathrm{peak}$ is the frequency of the dominant peak of $n(\omega,T)\Im{\alpha(\omega)}$ in the temperatures of interest and $\Gamma$ is the full width at half maximum (FWHM) of the peak), but where the Fresnel coefficients obtain a sinusoidal oscillation about their perfect reflector values with amplitude $\eta$ and frequency $\Omega_\mathrm{d}$: for $\omega \in (\omega_\mathrm{peak} - \Gamma, \omega_\mathrm{peak} + \Gamma)$,
\begin{equation}
\left\{\begin{array}{l}
r_s(\omega, k_\parallel) = -1 \to -(1 - \frac{\eta}{2}[1 - \cos{(\Omega_\mathrm{d} t)}]), \\
r_p(\omega, k_\parallel) = 1 \to 1 - \frac{\eta}{2}[1 - \cos{(\Omega_\mathrm{d} t)}].
\end{array}\right.
\end{equation}
Since the integral giving $\vb F_\mathrm{rad}$ is dominated by the value of the integrand in this range, this effectively gives
\begin{equation}
\vb F_\mathrm{rad} \to (1-\frac{\eta}{2})\vb F_\mathrm{rad} + \frac{\eta}{2}\cos{(\Omega_\mathrm{d} t)}\vb F_\mathrm{rad}
\label{seq:Frad_td}
\end{equation}
in the expression \eqref{seq:total_force}. Note that the symbol $\vb F_\mathrm{rad}$ retains its meaning as the strength of the force above a perfectly reflecting surface. $\vb F_\mathrm{env}$ also obtains a time-dependence in this way, but we neglect investigating it since $|\vb F_\mathrm{rad}| \gg |\vb F_\mathrm{env}|$ when $T > T_\mathrm{env}$. We assume that $\vb F_\mathrm{zpf}$ does not obtain a time-dependence, since we have seen that it is sensitive only to the electrostatic limit of the integrand. Ideally, $\vb F_\mathrm{cs}$ would not obtain a time-dependence since the surface would be perfectly transparent at $\omega_0$ at all times. However, the modulation of the reflection coefficients at $\omega_\mathrm{peak}$ will realistically also lead to a residual modulation of the reflection coefficients at $\omega_0$, which in turn would give rise to a time-dependent force $|r_\perp(t,\omega_0)\vb F_\mathrm{cs}|$ due to the interaction with the nanoparticle's optical image. As we will see later in Section~\ref{s:sensing}, to measure $\vb F_\mathrm{rad}$ we require it to dominate the $\Omega_\mathrm{d}$ frequency component of the total time-dependent force on the nanoparticle. We make this condition quantitative by defining the amplitude of the $\Omega_\mathrm{d}$ frequency component of the driving of $|r_\perp(\omega_0)|$ as $\eta_0 \equiv \lim_{T\to\infty} T^{-1} \int_{-T}^T \mathrm d t \, \cos{(\Omega_\mathrm{d} t)}|r_\perp(t,\omega_0)|$. This gives the condition $|\vb F_\mathrm{cs}|\eta_0 \ll |\vb F_\mathrm{rad}|\eta$ where roughly $|\vb F_\mathrm{rad}|/|\vb F_\mathrm{cs}| \sim 10^{-4}$. If the driving of the reflection coefficients at thermal frequencies is not perfectly sinusoidal, one should consider the $\eta$ appearing in this condition to be defined in an equivalent manner to $\eta_0$.
Assuming the condition on $\eta$ and $\eta_0$ to hold, the relevant effect of the time-dependent reflection coefficients on the total force $\vb F$ is
\begin{equation}
\vb F \to \vb F - \frac{\eta}{2}\vb F_\mathrm{rad} + \frac{\eta}{2}\cos{(\Omega_\mathrm{d} t)}\vb F_\mathrm{rad}.
\label{seq:time_dependent_total_force}
\end{equation}
To simplify the presentation in the main text, we defined the reflection coefficient $R(t,\lambda)$ to be the magnitude of the Fresnel reflection coefficient at normal incidence: $R(t,\lambda) = |r_s(2\pi c/\lambda,0;t)| = |r_p(2\pi c/\lambda,0;t)| = |r_\perp(2\pi c/\lambda,0;t)|$. We remark that the reflectivity of the surface at normal incidence is given by $R^2(t,\lambda)$.
\section{Nanoparticle polarizability and its Lorentzian approximation}
We obtain $\alpha(\lambda)$ from measurements of the relative permittivity $\epsilon(\lambda)$ of fused silica glass in the bulk summarized in~\cite{kitamura_Appl.Opt._46:33_2007}. $\alpha(\lambda)$ is calculated from these data via the quasistatic polarizability for a dielectric sphere
\begin{equation}
\alpha(\lambda) = 3\epsilon_0 V \frac{\epsilon(\lambda) - 1}{\epsilon(\lambda) + 2}.
\label{seq:polarizability}
\end{equation}
The model in~\cite{kitamura_Appl.Opt._46:33_2007} gives values of $\Im{\epsilon(\lambda)}$ in the optical range $\lambda \in [1,7]\ \si{\micro m}$ which are lower than the measurements reported in the optical range in the same work. However, we have checked that the values of $\epsilon(\lambda)$ in the optical range are so small that including them in our model for $\alpha(\lambda)$ does not affect the results for the forces. Rather, we observe that the integrands giving the values of the forces are dominated by $\alpha(\lambda)$ at its peak values; the value of $\vb F_\mathrm{rad}$ is especially dominated by the peak of $\epsilon(\lambda)$ around $\lambda_\mathrm{peak} = \SI{8.9}{\micro m}$. To confirm this, we fit a frequency-domain Lorentzian model
\begin{equation}
\alpha_L(\omega) = 3 \epsilon_0 V \frac{\omega_p^2}{\omega_\mathrm{peak}^2 - \omega^2 - \mathrm i \Gamma \omega},
\label{seq:lorentzian_pol}
\end{equation}
to this peak in the frequency-domain version of \eqnref{seq:polarizability}. Here $\omega_\mathrm{peak}, \omega_p$ and $\Gamma$ are all real parameters. Specifically, we fit $\Im{\alpha_L(\omega)}$ to $\Im{\alpha(\omega)}$ using the \texttt{scipy.optimize.curve\_fit} method of SciPy 1.8.0~\cite{scipy1.0}, obtaining the peak frequency $\omega_0 = 2\pi \times \SI{3.4e13}{\hertz}$, the linewidth $\Gamma = 2\pi\times\SI{1.5e12}{\hertz}$ and $\omega_p = 2\pi\times\SI{8.8e12}{\hertz}$. $\Gamma$ corresponds to a linewidth $2 \pi c \Gamma / \omega_\mathrm{peak}^2 = \SI{0.4}{\micro m}$ of $\alpha_L(\lambda)$. We then use $\alpha_\mathrm{L}$ to compute $\vb F_\mathrm{rad}$ approximately and compare to the results using $\alpha$, finding good agreement, which confirms that the peak at $\omega_\mathrm{peak}$ dominates the integral for $\vb F_\mathrm{rad}$.
Using $\alpha_L$, we also investigate the effect on $\vb F_\mathrm{rad}$ of broadening the linewidth of the peak, which gives insight into the role of colored emission on the strength of $\vb F_\mathrm{rad}$. In Fig.~\ref{sfig:broadening}(a) we plot $\Im{\alpha_L(\lambda;n)}$ where $\alpha_L(\lambda;n)$ is given by \eqnref{seq:lorentzian_pol} with $\Gamma \to n\Gamma$ and renormalized to keep the total radiated power constant (see \eqnref{seq:free_space_power}), and in Fig.~\ref{sfig:broadening}(b) the $\vb F_\mathrm{rad}$ calculated using these polarizabilities. We find that as the linewidth is increased, the scaling of the force decreases progressively to $z^{-6}$, i.e.~a weaker scaling than the $z^{-5}$ far-field Casimir force, and the oscillations in sign vanish.
\begin{figure}
\includegraphics{suppfig/broadening_lorentzians.pdf}
\caption{The effect of broadening the linewidth of the Lorentzian fit on $\vb F_\mathrm{rad}$. (a) The broadened lorentzians. The factor in the legend indicates the linewidth with respect to the original linewidth $\Gamma$. The broadened Lorentzians are renormalized in order to keep the total power radiated by the nanoparticle constant. (b) $\vb F_\mathrm{rad}$ calculated using the Lorentzian of the corresponding color in panel (a) and $T=\SI{600}{K}$.}
\label{sfig:broadening}
\end{figure}
\section{\label{s:sensing}Sensing of $\vb F_\mathrm{rad}$}
\newcommand{\stof}{K}
For low temperatures of the center of mass motion, an optically trapped nanoparticle is well-modelled as a damped harmonic oscillator. The equation of motion of a driven damped harmonic oscillator with position $x(t)$ is
\begin{equation}
\ddot x + \gamma \dot x + \Omega^2 x = \frac{1}{m}\left[F(t) + \stof(t)\right].
\end{equation}
where $\gamma$ is the total damping rate of the nanoparticle's motion, $F(t)$ is a deterministic driving force and $\stof(t)$ is the sum of all the noise forces. By Fourier transform, we can solve this equation to be
\begin{align}
x(\omega) &= \chi(\omega)\left[F(\omega) + \stof(\omega)\right] ,\\
\chi(\omega) &= \frac{1}{m (\Omega^2 - \omega^2 - \mathrm i\gamma\omega)}.
\end{align}
Introducing the motional and force PSDs
\begin{align}
S_{xx}(\omega) & = \frac{1}{2\pi}\int_{-\infty}^{\infty} \expval{x(t)x(t+\tau)} \mathrm e^{\mathrm i\omega\tau}\,\mathrm d\tau,
\label{motional_psd_def}
\\
S_{F}(\omega) & = \frac{1}{2\pi}\int_{-\infty}^{\infty} \langle F(t)F(t+\tau)\rangle \mathrm e^{\mathrm i\omega\tau}\,\mathrm d\tau,
\label{F_force_psd_def}
\\
S_{\stof}(\omega) & = \frac{1}{2\pi}\int_{-\infty}^{\infty} \expval{\stof(t)\stof(t+\tau)} \mathrm e^{\mathrm i\omega\tau}\,\mathrm d\tau,
\label{R_force_psd_def}
\end{align}
one can show that the motional PSD can be expressed in terms of the force PSDs as
\begin{equation}
S_{xx}(\omega) = |\chi(\omega)|^2\left[S_F(\omega) + S_\stof(\omega)\right].
\end{equation}
From this, we can define the thermally limited signal-to-noise ratio (SNR) as
\begin{equation}
\mathrm{SNR} = \frac{S_F(\omega)}{S_\stof(\omega)}.
\label{seq:SNR}
\end{equation}
The square root of $S_F(\omega)$ at $\mathrm{SNR}=1$ is called the force sensitivity $\fs$ and is given by
\begin{equation}
\fs(\omega) = \sqrt{S_\stof(\omega)},
\label{seq:force_sensitivity}
\end{equation}
and has units of \si{N/\sqrt{Hz}}.
To find the force PSD, we assume a harmonic driving force
\begin{equation}
F(t) = A(z,T) \cos{(\Omega_\mathrm{d} t + \phi)},
\end{equation}
with $\Omega_\mathrm{d}$ the driving frequency and $\phi$ the initial phase of the driving, and where we find from \eqnref{seq:time_dependent_total_force} the amplitude $A(z,T)$ to be
\begin{equation}
A(z,t) = \frac{\eta}{2} [\uv z \cdot \vb F_\mathrm{rad}(z,T)].
\end{equation}
From this, one calculates the force PSD
\begin{equation}
S_F(\omega) = \frac{A^2(z,T)}{4}\left[\delta(\omega+\Omega_\mathrm{d}) + \delta(\omega - \Omega_\mathrm{d})\right].
\label{seq:harmonic_force_PSD}
\end{equation}
A perfectly harmonic driving force is an idealization that does not exist in experiments. Any realistic approximately harmonic driving force will have some drift or error in its driving frequency $\Omega_\mathrm{d}$, which smoothens the $\delta$-functions appearing in \eqref{seq:harmonic_force_PSD}.
We can account for the uncertainty in $\Omega_\mathrm{d}$ by broadening the $\delta$-functions in \eqref{seq:harmonic_force_PSD} into Lorentzians with linewidth $\Gamma_\mathrm{d}$ which represent the broadening:
\begin{equation}
\delta(\omega \pm \Omega_\mathrm{d}) \to \frac{1}{\pi}\frac{\Gamma_\mathrm{d}}{(\omega \pm \Omega_\mathrm{d})^2 + \Gamma_\mathrm{d}^2}.
\end{equation}
We can check the validity of this broadening by taking the $\Gamma_\mathrm{d}\to 0$ limit of the RHS using the Plemejl formula, and seeing that it recovers the LHS. This results in the PSD
\begin{equation}
S_F(\omega) = \frac{A^2(z,T)}{4\pi}\left[\frac{\Gamma_\mathrm{d}}{(\omega + \Omega_\mathrm{d})^2 + \Gamma_\mathrm{d}^2} + \frac{\Gamma_\mathrm{d}}{(\omega - \Omega_\mathrm{d})^2 + \Gamma_\mathrm{d}^2}\right].
\label{seq:pseudoharmonic_force_PSD}
\end{equation}
The contributions to the noise force $\stof(t)$ that we model are random collisions with the gas and photon shot noise~\cite{gonzalez-ballestero_Phys.Rev.A_100:1_2019, jain_Phys.Rev.Lett._116:24_2016}. Thus,
\begin{gather}
S_\stof(\omega) = S_\mathrm{photon}(\omega) + S_\mathrm{gas}(\omega), \label{seq:noise_PSD} \\
S_\mathrm{photon}(\omega) = \frac{2}{5} \frac{\hbar\omega_0}{2\pi c^2} P_\mathrm{scatt}, \label{seq:photon_PSD} \\
S_\mathrm{gas}(\omega) = \frac{m\gamma_\mathrm{gas}k_\mathrm B T_\mathrm{gas}}{\pi}. \label{seq:gas_PSD}
\end{gather}
Here $\omega_0$ is the frequency of the trapping laser, $P_\mathrm{scatt}$ is the power scattered from the tweezer laser, $\gamma_\mathrm{gas}$ is the damping rate due to gas collisions, and $T_\mathrm{gas}$ is the temperature of the gas. The power scattered from the tweezer laser is given by $P_\mathrm{scatt} = \sigma_\mathrm{scatt}I_0$, where $\sigma_\mathrm{scatt}$ is the electromagnetic scattering cross-section of the nanoparticle, and $I_0$ is the intensity at the trapping laser focus. These two quantities are given by~\cite{jain_Phys.Rev.Lett._116:24_2016}
\begin{gather}
I_0 = \frac{P_\mathrm{tw} \omega_0^2 N^2}{2\pi c^2}, \\
\sigma_\mathrm{scatt} = \frac{\omega_0^4|\alpha(\omega_0)|^2}{6\pi c^4 \epsilon_0^2}.
\end{gather}
Here $N$ is the numerical aperture of the trap. The gas damping rate is given by~\cite{beresnev_MotionSpherical_1990}
\begin{equation}
\gamma_\mathrm{gas} = 0.619\frac{6\pi \nprad^2}{m} p \sqrt{\frac{2 m_\mathrm{gas}}{\pi k_\mathrm B T_\mathrm{gas}}}.
\end{equation}
Here $\nprad$ is the nanoparticle radius, $p$ is the gas pressure, $m_\mathrm{gas}$ is the gas molecular mass.
Inserting our expressions for the noise PSDs into the expression for $\fs$, we obtain
\begin{equation}
\fs = \sqrt{\frac{\hbar \omega_0^7 |\alpha(\omega_0)|^2 N^2 P_\mathrm{tw}}{60\pi^3\epsilon_0^2 c^8} + 6\cdot0.619 \nprad^2 p \sqrt{\frac{2 m_\mathrm{gas} k_\mathrm B T_\mathrm{gas}}{\pi}}}.
\end{equation}
Putting everything together, we arrive at the expression for the signal-to-noise ratio evaluated at the driving frequency $\Omega_\mathrm{d}$
\begin{equation}
\begin{aligned}[b]
\mathrm{SNR}(\Omega_\mathrm{d}) & = \frac{A^2(z,T)}{4\pi\Gamma_\mathrm{d}}\left(\frac{\hbar \omega_0^7 |\alpha(\omega_0)|^2 N^2 P_\mathrm{tw}}{60\pi^3\epsilon_0^2 c^8} + 6\cdot0.619 \nprad^2 p \sqrt{\frac{2 m_\mathrm{gas} \pi k_\mathrm B T_\mathrm{gas}}{\pi}}\right)^{-1} \\
& \approx \frac{\eta^2 [\uv z \cdot \vb F_\mathrm{rad}(z,T)]^2}{16\pi\Gamma_\mathrm{d}}\left(\frac{\hbar \omega_0^5 |\alpha(\omega_0)|^2 I_0}{30\pi^2\epsilon_0^2 c^6}\right)^{-1}
\end{aligned}
\end{equation}
where in the second row we assumed that we are below the pressure at which $\fs$ becomes photon shot noise-dominated.
This expression lets us identify the scaling of the $\mathrm{SNR}$ with the experimental parameters. In this analysis, we assume a constant $T$ independently controllable from the other parameters. Since $|\vb F_\mathrm{rad}| \propto |\alpha| \propto V$, the $\mathrm{SNR}$ is independent of the particle size. Increasing the laser intensity $I_0$ at the nanoparticle decreases the SNR since the photon shot noise increases.
\section{\label{s:heat_balance} Heat balance equation for the nanoparticle}
\begin{figure}
\centering
\includegraphics{suppfig/power_fit_and_temp_dyns.pdf}
\caption{(a) The power radiated per volume by an \ch{SiO2} nanoparticle in free space in the temperature range relevant for experiments, as well as a quadratic fit from 300 to \SI{2000}{K} which we use to qualitatively understand the nanoparticle thermodynamics. (b) The power emitted above a perfectly reflecting surface as a fraction of the power emitted in free space. (c) The same function as in panel (b), but with a log-scaled horizontal axis. (d) Time evolution of the temperature in the radiative cooling regime for the parameters in the main text. Colors denote different initial temperatures. Solid lines are numerical solutions of \eqnref{seq:heat_balance_eq}, dashed lines are the approximation \eqnref{seq:approx_T_dyns}. The numerical steady state temperature is \SI{543}{K} while the approximation gives $T_\infty = \SI{515}{K}$.}
\label{sfig:power}
\end{figure}
Our goal in this section is to capture the behavior of the temperature dynamics predicted by fluctuational electrodynamics in a simple, qualitative model. We predict the temperature dynamics of the nanoparticle for our experimental parameters using the heat balance equation
\begin{equation}
c_m \rho V\dfdx{T}{t} = P_\mathrm{abs} - P_\mathrm{rad}(T),
\label{seq:heat_balance_eq}
\end{equation}
where we use $c_m = \SI{700}{J/kg}$ for \ch{SiO2}, and
\begin{equation}
P_\mathrm{abs} = P_\mathrm{abs,tw} + P_\mathrm{abs,env}(T_\mathrm{env}),
\end{equation}
Here $P_\mathrm{abs,env}(T_\mathrm{env})$ is the power that the nanoparticle absorbs from environmental radiation, and $P_\mathrm{abs,tw}$ is the power that the nanoparticle absorbs from the laser beam. The latter is
\begin{equation}
P_\mathrm{abs,tw} = \frac{\omega_0\Im{\alpha(\omega_0)}}{c} I_0.
\label{seq:absorbed_laser_power}
\end{equation}
As can be calculated using $P_\mathrm{rad}(T) - P_\mathrm{abs,env}(T_\mathrm{env}) = \expval{[\partial_t\vb d(t)] \cdot \vb E(t)}$ (see e.g.~\cite{messina_Phys.Rev.B_88:10_2013}), in free space
\begin{equation}
P_\mathrm{rad}(T) = \frac{\hbar c^2}{\pi^2}\int_0^\infty\mathrm d k\, \frac{k^4}{e^{\hbar c k/k_\mathrm{B} T} - 1}\Im{\alpha(ck)},
\label{seq:free_space_power}
\end{equation}
and $P_\mathrm{abs,env}(T_\mathrm{env}) = P_\mathrm{rad}(T_\mathrm{env})$. We have not included the correction to \eqnref{seq:free_space_power} from radiation reaction (see \cite{rubiolopez_Phys.Rev.B_98:15_2018} and references therein) because it gives negligible corrections for the material and particle size that we consider. With our parameters, $P_\mathrm{abs,env}(T_\mathrm{env}) \ll P_\mathrm{abs,tw}$, so we neglect $P_\mathrm{abs,env}(T_\mathrm{env})$. We obtain a simplified model of $P_\mathrm{rad}(T)$ for the silica nanoparticle by approximating it as a quadratic function of $T$:
\begin{equation}
P_\mathrm{rad}(T) \approx \sigma_2 V T^2,
\label{seq:quadratic_power}
\end{equation}
where we find $\sigma_2 = \SI{2e4}{W/m^3K^2}$ by numerical fit to~\eqref{seq:free_space_power}. We display~\eqnref{seq:free_space_power} and compare it to the quadratic approximation~\eqnref{seq:quadratic_power} in Fig.~\ref{sfig:power}(a). $P_\mathrm{rad}(T)$ is in general affected by the presence of dielectric material like the surface. As an example, we plot $P_\mathrm{rad}(T)$ for our silica nanoparticle above a perfectly reflecting surface in Fig.~\ref{sfig:power}(b) and (c). The emitted power oscillates with $z$ with a maximum change of \SI{20}{\percent} for $z>\SI{1}{\micro m}$. The qualitative behavior of the temperature dynamics will not be affected by this change, so we find that \eqnref{seq:quadratic_power} still qualitatively captures the emitted power for this example. To make a qualitative model of the temperature dynamics, we insert \eqnref{seq:quadratic_power} into \eqref{seq:heat_balance_eq}. This gives a simple instance of Ricatti's equation for $T$ with solution
\begin{equation}
T(t) = \left\{\begin{matrix} T_\infty\tanh{[t/\tau + \mathrm{artanh}\,(T_0/T_\infty)]} & (T_\infty > T_0) \\
T_\infty\coth{[t/\tau + \mathrm{artanh}\,(T_\infty/T_0)]} & (T_\infty < T_0) \end{matrix}\right.
\label{seq:approx_T_dyns}
\end{equation}
where the time constant is
\begin{equation}
\tau = \frac{\rho c_m}{\sqrt{\frac{P_\mathrm{abs}}{V}\sigma_2}},
\end{equation}
and $T_\infty = \sqrt{P_\mathrm{abs}/V \sigma_2}$. Note that $P_\mathrm{abs}/V$ is independent of $V$. In Fig.~\ref{sfig:power}(d) we compare time trajectories of the temperature obtained with numerical solution of \eqnref{seq:heat_balance_eq} and with the approximation \eqnref{seq:approx_T_dyns}. We find that the approximation reproduces the qualitative behavior of the trajectories and that the final temperature agrees within \SI{30}{K}.
\end{document}
|
\section{Results}
\subsection{Model and Formalism}\label{model}
Consider a set of 1D quantum wires subjected to Rashba~\cite{Rashba} and the Dresselhaus SOI~\cite{Dresselhaus}. The quantum wires form a QN by connecting their ends to different nodes. In two-dimensional electron gases, the dominant contributions to Rashba and Dresselhaus SOI are linear in momentum $p$~\footnote{Here, we neglect cubic corrections to Dresselhaus SOI ~\cite{Dresselhaus,Bercioux_2015}.}. The quantum wire Hamiltonian then reads:
\begin{equation}\label{Hamiltonian1}
\hat{\mathcal{H}}=\frac{p^{2}}{2m^*}+\frac{\hbar k_{\mathrm{R}}}{m^*}p~(\hat{\boldsymbol{\gamma}}\times\hat{\boldsymbol{z}})\cdot\boldsymbol{\sigma}+\frac{\hbar k_{\mathrm{D}}}{m^*}p~\hat{\overline{\boldsymbol{\gamma}}}\cdot\boldsymbol{\sigma},
\end{equation}
where $k_{\mathrm{R}}$ and $k_{\mathrm{D}}$ are the Rashba and the Dresselhaus SOI strengths (in inverse-length units), respectively, $\hat{\boldsymbol{\gamma}}=(\cos\gamma,\sin\gamma,0)$ is the unit vector specifying the direction of the quantum wire, and $\hat{\overline{\boldsymbol{\gamma}}}$ is the $y$-reflected $\hat{\boldsymbol{\gamma}}$. In Eq.~\eqref{Hamiltonian1}, $\boldsymbol{\sigma}$ is the vector of the Pauli matrices associated with the electron spin.
The Rashba and Dresselhaus SOI terms in Hamiltonian~\eqref{Hamiltonian1} can be unified as
\begin{equation}\label{Hamiltonian2}
\hat{\mathcal{H}}=\frac{p^{2}}{2m^*}+\frac{\hbar \kappa}{m^*}p~ \hat{\boldsymbol{\theta}}\cdot\boldsymbol{\sigma},
\end{equation}
where
\begin{subequations}\label{kappa_theta}
\begin{align}
\kappa&=\sqrt{k_{\mathrm{R}}^2+k_{\mathrm{D}}^2+2k_{\mathrm{R}}k_{\mathrm{D}}\sin{(2\gamma)}}, \label{kappa}\\
\theta&=\arg\left[(k_{\mathrm{R}}\sin{\gamma}+k_{\mathrm{D}}\cos{\gamma}) \right.\nonumber \\
& \left. \hspace{1cm}+\mathrm{i}(-k_{\mathrm{R}}\cos{\gamma}-k_{\mathrm{D}}\sin{\gamma}) \right],\label{phaseRD}
\end{align}
\end{subequations}
Hamiltonian~\eqref{Hamiltonian2} describes an equivalent system where an electron moves along a quantum wire perpendicular to $\hat{\boldsymbol{\theta}}=(\cos\theta,\sin\theta,0)$ and is subject to an \emph{effective} SOI with strength $\kappa$. Said SOI term can be interpreted as effective magnetic field $\boldsymbol{B}_\mathrm{SO}=2\hbar \kappa p/(g\mu m^*)\hat{\boldsymbol{\theta}}$. In Fig.~\ref{magneticfieldtexture} we show the effective magnetic field texture experienced by spin carriers following circular trajectories for various values of the Rashba and Dresselhaus SOIs. The arrows indicate the evolution of $\boldsymbol{B}_\mathrm{SO}$ when an electron moves counterclockwise in a circular trajectory. The field texture can be characterised topologically in terms of the winding number around the $z$ axis, see Eq.~\eqref{Eq:winding number}. The winding number changes depending on the relative strength between the Rashba and Dresselhaus SOIs; this is reflected in the spin texture of the polygon eigenstates as shown in Sec.~\ref{Winding number}.
The QN problem is solved by obtaining a wave function for each 1D quantum wire satisfying the Dirichlet boundary conditions. The overall solution is obtained by imposing the general boundary condition on the set of wires composing the QN | see below.
The wave function of a quantum wire can be written as~\cite{Bercioux_2004,Bercioux_2005A}:
\begin{align}\label{wavefunction1}
\boldsymbol{\Psi}(r)=\frac{\mathrm{e}^{-\mathrm{i}\kappa r \hat{\boldsymbol{\theta}}\cdot\boldsymbol{\sigma}}}{\sin(k\ell)}&[\sin{k(\ell-r)}\boldsymbol{\Psi}_\alpha \nonumber\\
&+\sin{(kr)}\mathrm{e}^{\mathrm{i}\kappa \ell \hat{\boldsymbol{\theta}}\cdot\boldsymbol{\sigma}}\boldsymbol{\Psi}_\beta],
\end{align}
where $\boldsymbol{\Psi}_\alpha$ and $\boldsymbol{\Psi}_\beta$ are the spinors evaluated at quantum wire the boundaries and $k=\sqrt{2m\epsilon/\hbar^2+\kappa^2}$. The exponent in the prefactor of Eq.~\eqref{wavefunction1} accounts for the spin precession due to the effective magnetic field created by the two SOIs. When the Rashba and Dresselhaus SOI strengths are equal, and $\gamma=3\pi/4$, both SOI terms cancel in the Hamiltonian, the effective magnetic field vanishes, and the energy dispersion turns spin degenerate~\cite{Schliemann_2003A,Schliemann_2003B}. In this case, $\kappa$ vanishes, so the SU(2) rotation disappears from the wave function. This means that the spinor evolves along the wire as a free particle.
The spin-carrier dynamics in a QN can be solved by considering two general boundary conditions: first, the continuity of the wave function to each node of the QN, and second by requiring the conservation of the probability current at the same points~\cite{Gnutzmann_2006}. Notice that the presence of SOI modifies the definition of probability current~\cite{Hodge_2014}. This is accounted by the extended derivative:
\begin{equation}
D=\frac{\partial}{\partial r}+\mathrm{i}k_{\mathrm{R}}(\hat{\boldsymbol{\gamma}}\times\hat{\boldsymbol{z}})\cdot\boldsymbol{\sigma}+\mathrm{i}k_{\mathrm{D}}\hat{\overline{\boldsymbol{\gamma}}}\cdot\boldsymbol{\sigma}=\frac{\partial}{\partial r}+\mathrm{i}\kappa\hat{\boldsymbol{\theta}}\cdot\boldsymbol{\sigma}.
\end{equation}
Equation~\eqref{wavefunction1} together with the conservation of the current at the vertices provides the values of the wave function at the vertices $\boldsymbol{\Psi}_{\alpha}$, and $\boldsymbol{\Psi}(r)$ by extension. We can evaluate the transport properties by supplementing the QN with an additional connection to external leads. The extension of the method is explained in the Methods section.
\subsection{Conductance pattern}\label{conductance}
In this section, we study the transport properties of square and ring loops by applying the QN formalism. Rings are modelled as regular polygons of perimeter $P$ with a large number of edges ($N \gg 1$)~\cite{Bercioux_2005B} such that the Fermi wavelength and the spin precession length of the carriers are much larger than the edges' length $L=P/N$. This overcomes the problem of dealing with approximate solutions for rings subject to Rashba and Dresselhaus SOI \cite{Lia_2021, Lia_2022}. These restrictions do not apply to square loops. Moreover, mesoscopic experiments are typically carried out in the so-called semiclassical regime where the Fermi wavelength of the carriers is much smaller than the loops' perimeter, such that $k \gg 2\pi/P$ \cite{Nagasawa_2012,Nagasawa_2013,Wang_2019,Frustaglia_2020}.
In Fig.~\ref{conductance1} we show the conductance maps as a function of the dimensionless SOI intensities
$k_\text{R}P/2\pi$ and $k_\text{D}P/2\pi$ for different interferometric loops. Figures~\ref{conductance1}(a) and~\ref{conductance1}(b) show the conductance for the case of square-shaped loops with different orientations, with $\eta$ the rotation angle measured from the ``diamond" configuration depicted in Fig.~\ref{conductance1}(a) (corresponding to square's sides forming an angle of $\pi/4$ with respect to the crystallographic axes). In Fig.~\ref{conductance1}(c), instead, we show the conductance corresponding to a ring-shaped loop. For all configurations, the conductance shows a symmetric behaviour with respect to the critical line $k_\text{R}=k_\text{D}$ along which the Rashba and Dresselhaus SOIs have the same strength.
This can be understood in terms of the Hamiltonian in Eq.~\eqref{Hamiltonian2}. The interaction between the two SOIs can be described by an equivalent QN where each edge is replaced by one with a different orientation $\theta+\pi/2$ with an \emph{effective} SOI strength $\kappa$ given by Eq.~\eqref{kappa}. If the strengths of the interactions are interchanged, $k_{\mathrm{R}} \leftrightarrow k_{\mathrm{D}}$, the value of the equivalent effective SOI strength $\kappa$ for a given edge remains unchanged, but the orientation of the equivalent edge becomes $-\theta$ instead of $\theta+\pi/2$. This means that the edges, and the polygon by extension, are mirrored with respect to the $\hat{r}_+=(\hat{x}+\hat{y})/\sqrt{2}$ direction, so the conductance pattern remains unchanged under the $k_{\mathrm{R}} \leftrightarrow k_{\mathrm{D}}$ transformation.
The most interesting case is presented in the case of Fig.~\ref{conductance1}(a): the conductance presents a checkerboard pattern. In absence of Dresselhaus SOI, the minima of the conductance are presented every $(2n+1)k_\text{R}P/2\pi$ with $n\in\mathbb{N}_0$~\cite{Bercioux_2005B,Hijano_2021}. Similar behaviour is observed in the absence of Dresselhaus SOI with $k_\text{R}$ replaced by~$k_\text{D}$. The combined presence of the two SOIs adds an overall shift of the conductance minima by a factor of $2\pi$. The resulting checkerboard pattern was first reported in Ref.~\cite{Ramaglia_2006}. In the next section, we elaborate on this by studying the geometric properties of spinors in terms of winding numbers. In the Methods section, we show how the conductance and the winding number for the case of the square evolves
from the checkerboard pattern in Fig.~\ref{Winding1}(a) and~\ref{Winding1}(b) to the simpler structure in Fig.~\ref{Winding1}(c) and~\ref{Winding1}(d) upon changing the orientation of the square.
By rotating the square with respect to the crystallographic axes | Fig.~\ref{conductance1}(b) | we find something remarkable: the checkerboard conductance pattern disappears. Instead, we find a conductance map that looks similar to that of ring-shaped loop | Fig.~\ref{conductance1}(c) | except for a period-doubling due to strongly non-adiabatic processes at the square vertices hindering spin-phase development ~\cite{Frustaglia_2004,Bercioux_2005B}.
For the case of a ring, Fig.~\ref{conductance1}(c), the conductance presents a fishbone structure with minima as a function of the Rashba SOI following the sequence predicted theoretically in Ref.~\cite{Frustaglia_2004}. Similar behaviour is observed as a function of the Dresselhaus SOI term. However, in the presence of both SOI terms, the conductance behaviour is more intricate.
Interestingly, the conductance remains constant along the critical line regardless of the orientation or number of edges of the polygon [see Supplemental Material~\cite{supplemental} for the additional case of a hexagonal and octagonal loop]. When the Rashba and Dresselhaus SOIs have the same strength, the effective SOI field decouples from the momentum and points always in the same direction~\cite{Schliemann_2003A,Schliemann_2003B}. Moreover, the spin precession angle only depends on the distance travelled along the $\hat{r}_+$ direction.
This effect is known as \emph{persistent spin helix}~\cite{Bernevig}, and in recent years several experiments claimed to have achieved this effect~\cite{Koralek_2009,Kohda_2012,Walser_2012,Kunihashi_2016,Dettwiler_2017}.
Since all paths contributing to the transmission amplitude begin and end at the same points, the spin precession along each path is the same and the resulting interference is always constructive.
\subsection{Topological characterization} \label{Winding number}
We can characterise spin and field textures topologically in terms of (integer) winding numbers around the $z$-axis. This quantity is defined as
\begin{align}\label{Eq:winding number}
\omega=\frac{1}{2 \pi} \int_{0}^{P} d \ell\left(\widehat{\boldsymbol{n}} \times \frac{d \hat{\boldsymbol{n}}}{d \ell}\right) \cdot \hat{\boldsymbol{z}}
\end{align}
with $\hat{\boldsymbol{n}}(\ell)$ a in-plane unit vector and $0\le \ell\le P$ a linear parametrization of the circuit's perimeter. For the field texture, we identify $\hat{\boldsymbol{n}}(\ell)$ with $\hat{\boldsymbol{\theta}}(\ell)$ in Eq.~\eqref{Hamiltonian2}. This means that $\omega=1$ for dominating Rashba SOI and $\omega=-1$ for dominating Dresselhaus SOI (see Fig.~\ref{magneticfieldtexture}). The transition occurs at the critical line $k_{\text R}=k_{\text D}$. As for the spin texture, this is given by $\hat{s}(\ell)=\langle \Psi(\ell)|\boldsymbol{\sigma}|\Psi(\ell)\rangle$. In this case, $\hat{\boldsymbol{n}}$ is identified with the normalized projection of $\hat{s}(\ell)$ on the $xy$-plane.
Spin textures developed in Rashba and Dresselhaus squares present a periodic, checkerboard-like pattern alternating positive and negative windings | see Fig.~\ref{Winding1}(a). This pattern contrasts with the simplicity of the field texture driving the spin dynamics, demonstrating the possibility of producing topological transitions in the spin texture by slightly tuning the SOI fields. Figure~\ref{Winding1}(b) shows that the winding pattern is fully correlated with the conductance up to a period-doubling.
\begin{figure}[!t]
\includegraphics[width=0.9\columnwidth]{Fig_3.png}
\caption{\label{Winding1} Winding number ($\omega$) for a square (a), a square rotated $\pi/4$ from the initial configuration (c) and a ring (e) subject to Rasba and Dresselhaus SOI. Winding number ($\omega$) overlapped with the conductance in units of $2e^2/h$ for a square (b), a square rotated $\pi/4$ from the initial configuration (d) and a ring (f) subject to Rasba and Dresselhaus SOI.}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=0.75\columnwidth]{Fig_4.png}
\caption{\label{Textures}The spin texture of a propagating mode in the Bloch sphere (up) and its azimuthal projection (down) for different SOI strengths. The SOI strengths are given by lines A (square) and B, C, and D (ring) in Fig.~\ref{Winding1}. In each panel, the colour indicates the circulation of the local spin states as the carrier propagates along the perimeter, from red to violet.}
\end{figure}
In Fig.~\ref{Textures}(A), we present a series of spin textures undergoing a topological transition along segment A in Fig.~\ref{Winding1}(a), corresponding to a square loop. We find that a positive winding texture becomes negative by collapsing into a flat texture (subtending no solid angle and no geometric phase) at the critical line~\cite{Rodriguez_2021}.
In the case of ring loops, Fig.~\ref{Winding1}(e), we find that the spin textures have a dominant tendency to follow the driving field texture by sharing its topological characteristics. Still, we find a fishbone pattern of anomalous winding | see Fig.~\ref{Winding1}(f). To shed some light on it, it is useful to assume a dominating Rashba SOI with $k_{\text R} \gg k_{\text D}$ and work in the rotating frame where the radial Rashba texture results uniform (with an oscillating Dresselhaus perturbation). We notice that the fishbone pattern meets the Rashba axis at points $k_{\text D}P=\pi\sqrt{4j^2-1}$ with $j$ integer. This coincides with the Rabi condition for spin resonance in the rotating frame. As the magnitude of Dresselhaus perturbation increases, the resonance condition changes by undergoing a so-called Bloch-Siegert shift~\cite{Bloch_1940,Reynoso_2017}. Something similar happens along the Dresselhaus axis. Close to the resonance condition, complex spin textures emerge with anomalous winding. Figures~\ref{Textures}(B),~\ref{Textures}(C), and~\ref{Textures}(D) illustrate the winding transitions taking place in the these textures. We notice that, in contrast to the case of square loops, a winding transition does not require a full collapse of the spin texture with vanishing geometric phases.
Still, in both square and ring geometries, the spin winding is antisymmetric with respect to the critical line along which the driving field changes topology.
Additionally, it is possible to disrupt the spin textures' response in a square circuit by reorienting it with respect to the crystallographic axes. In Fig.~\ref{Winding1}(c) and~\ref{Winding1}(d), we show the results for a rotated square (angle $\eta=\pi/4$). In this case we find a winding pattern fully correlated with that of the field texture | similarly to the ring's one shown in Fig.~\ref{Winding1}(e), except for the anomalous fishbone winding.
\section{Conclusions and Outlook}\label{ConclusionsandOutlook}
We demonstrate how the circuit's geometry in SOI interferometers can be used as a resource for manipulating the spin state of the carriers. Effective SOI field textures are built by introducing circuit sections of different curvatures steering the carriers' momentum. In this way, highly curved vertices in polygon circuits act as spin scattering centers for the carriers.
For square circuits subject to Rashba and Dresselhaus SOI, we find that the topological characteristics of the spin textures can be manipulated with relative ease by electric control of the SOIs in semiconducting nanostructures \cite{Nitta_1997,Schapers_1998,Grundler_2000,Dettwiler_2017,Nagasawa_2018}. This contrasts with the case of ring circuits where, as a general rule, a topological change in the field texture is required to induce a similar change in the spin textures. Still, this restraint can be overcome by tuning the SOI to satisfy the spin resonance conditions where complex spin textures develop. Moreover, additional possibilities for spin control appear by in-plane rotation of square circuits. Remarkably, these topological features leave an imprint on the quantum conductance of the circuits, which can be addressed experimentally.
\section*{acknowledgments}
D.B. acknowledges the support from the Spanish MICINN-AEI through Project No. PID2020-120614GB-I00, the Transnational Common Laboratory $Quantum-ChemPhys$, and the Programa Red guipuzcoana de Ciencia, Tecnolog\'ia e Innovaci\'on 2021, Grant nr. 2021-CIEN-000070-01 Gipuzkoa Next, and the funding from the Basque Government's IKUR initiative on Quantum technologies (Department of Education). A.H. acknowledges funding from the University of the Basque Country (Project PIF20/05). D.F. acknowledges support from the Spanish MICINN-AEI through Project No. PID2021-127250NB-I00 and from the Andalusian Government through PAIDI 2020 Project No. P20-00548 and FEDER Project No. US-1380932.
\section*{Methods}
\subsection*{Formalism for quantum transport}\label{transportformalism}
Here we present the quantum network (QN) formalism used to study the transport properties of polygonal QNs, Fig.~\ref{polygons}. Semi-infinite input and output leads are attached to the network's vertices for the transport measurements. Each lead consists of a quantum wire with two spin channels. The leads are not subjected to any interaction, so they are characterised at zero temperature by the Fermi energy and a wave vector $k$. We assume that the channels behave like incoherent sources, so there is no phase relationship between the different input channels~\cite{landauer_formulation}.
In a system with $N_{\mathrm{in}}$ ($N_{\mathrm{out}}$) input (output) channels, if an electron is injected through input channel $\sigma$ with wavenumber $k$, the wave function alongside the channels can be written as
\begin{subequations}
\begin{equation}\label{input}
\Psi_{\mathrm{in},\sigma'}(r)=\mathrm{e}^{\mathrm{i}k r}\delta_{\sigma'\sigma}+r_{\sigma'\sigma}\mathrm{e}^{-\mathrm{i}k r},
\end{equation}
\begin{equation}\label{output}
\Psi_{\mathrm{out},\sigma'}(r)=t_{\sigma'\sigma}\mathrm{e}^{\mathrm{i}k r},
\end{equation}
\end{subequations}
where $r$ is the position measured from the edge, and it is negative for input leads and positive for output leads. Here $r_{\sigma'\sigma}$ and $t_{\sigma'\sigma}$ are the channel-resolved reflection and transmission coefficients respectively, so that $\sum_{\sigma'}^{N_{\mathrm{in}}}\left|r_{\sigma'\sigma}\right|^2+\sum_{\sigma'}^{N_{\mathrm{out}}}\left|t_{\sigma'\sigma}\right|^2=1$. The indices $\sigma$ and $\sigma'$ specify the lead and the spin state of the channel. We define the total transmission and reflection coefficients of a channel $\sigma$ as
\begin{figure}[!t]
\centering
\includegraphics[width=0.9\columnwidth]{SM_Fig_1.png}
\caption{\label{polygons} Sketch of the polygonal structures we will consider for the quantum transport: (a) square, (b) hexagon, (c) octagon, (d) ring.}
\end{figure}
\begin{equation}
T_{\sigma}=\sum_{\sigma'}\left|t_{\sigma\sigma'}\right|^2\; , \qquad R_{\sigma}=\sum_{\sigma'}\left|r_{\sigma\sigma'}\right|^2\; .
\end{equation}
where the sum runs over the input channels. The total transmission (reflection) is given by the sum of the transmission (reflection) coefficients of the output (input) channels,
\begin{subequations}
\begin{align}
T&=\sum_{\sigma}T_{\sigma}=\sum_{\sigma\sigma'}\left|t_{\sigma\sigma'}\right|^2\\
R&=\sum_{\sigma}R_{\sigma}=\sum_{\sigma\sigma'}\left|r_{\sigma\sigma'}\right|^2
\end{align}
\end{subequations}
The zero-temperature conductance $G$ based on the Landauer formula reads~\cite{datta1997electronic}:
\begin{equation}\label{landauer_conductance}
G=\frac{e^2}{h}\mathrm{Tr}\: \left[tt^{\dagger}\right]=\frac{e^2}{h} T.
\end{equation}
Equation~\eqref{landauer_conductance} sets an upper limit for the conductance, which is bounded by the number input channels, such that $G \leq N_{\mathrm{in}} e^2/h$.
The wavefunction of the quantum network satisfies boundary conditions at the vertices, which ensure the continuity (uniqueness) of the wavefunction and the conservation of the probability current. In an isolated quantum network, imposing the continuity of the wave function and conserving the probability current yields a set of linear homogeneous equations where the variables are the values of the wave function at the vertices. This allows us to study the spectral properties of the quantum network. When adding the external leads, the system's energy is fixed by the Fermi energy of the leads. Due to the first term in Eq.~\eqref{input}, the set of equations becomes inhomogeneous, with a unique solution for $T$ and $R$.
In a system with Rashba and Dresselhaus spin-orbit interactions (SOIs), the wave function of a wire is described by the values it takes at the nodes $\boldsymbol{\Psi}_\alpha$ [see Eq.~\eqref{wavefunction1}]. The single-valuedness of the wave function at the nodes is automatically satisfied by this equation. In addition, imposing the continuity of the wave function at the vertices connected to external leads allows writing the reflection and transmission coefficients of the leads in terms of $\boldsymbol{\Psi}_\alpha$. Therefore, the number of unknown variables equals the number of vertices $V$. The conservation of probability current at the nodes allows one to write $V$ equations, which fix the values of $\boldsymbol{\Psi}_\alpha$, and consequently the reflection/transmission coefficients. The conservation of probability current at a node is given by the sum of the outgoing extended derivatives of the wave function, which must be equal to zero. For a generic node $\alpha$, the continuity of probability current reads
\begin{equation}\label{probability current}
\sum_{\langle\alpha,\beta\rangle} \left. D\boldsymbol{\Psi}_{\alpha,\beta}(r) \right|_{r=0}=0\; ,
\end{equation}
where the sum $\sum_{\langle\alpha,\beta\rangle}$ runs over all nodes $\beta$ which are connected to $\alpha$. This equation can be rewritten in terms of $\boldsymbol{\Psi}_{\alpha}$ and $\boldsymbol{\Psi}_{\beta}$. For internal nodes, it reads
\begin{equation}\label{current conservation 1}
\boldsymbol{M}_{\alpha,\alpha}\boldsymbol{\Psi}_{\alpha}+\sum_{\langle\alpha,\beta\rangle}\boldsymbol{M}_{\alpha,\beta}\boldsymbol{\Psi}_{\beta}=0\; ,
\end{equation}
where
\begin{subequations}
\begin{align}
\boldsymbol{M}_{\alpha,\alpha}&=\sum_{\langle\alpha,\beta\rangle}\frac{k_{\beta,\alpha}}{\tan{k_{\beta,\alpha}L}}\\
\boldsymbol{M}_{\alpha,\beta}&=-\frac{k_{\beta,\alpha}}{\sin{k_{\beta,\alpha}L}}\mathrm{e}^{\mathrm{i}\kappa_{\beta,\alpha} L \hat{\boldsymbol{\theta}}_{\beta,\alpha}\cdot\boldsymbol{\sigma}}\; .
\end{align}
\end{subequations}
Here $\kappa_{\beta,\alpha}$ and $\hat{\boldsymbol{\theta}}_{\beta,\alpha}$ indicate the strength and direction of $\boldsymbol{B}_\mathrm{SO}$ for an electron travelling from vertex $\alpha$ towards vertex $\beta$, see Eq.~\eqref{kappa_theta}.
\subsection*{Computation of the winding number}\label{Windingmodel}
The model used to compute the winding number was built upon the one used in \cite{Rodriguez_2021} where we consider a regular polygon of $N$ conducting sides of length $L = P/N$ with $P$ being the perimeter, which lies on the $xy$-plane. Each side connects the vertices $\alpha$ and $\beta$ and it is oriented along directions $\hat{\gamma}_{\beta,\alpha}$ (from $\alpha$ to $\beta$). The spin-carrier dynamics along each side are determined by Hamiltonian~\eqref{Hamiltonian2}. The SOI terms can be interpreted as effective in-plane magnetic field $\boldsymbol{B}_\mathrm{SO}=2\hbar \kappa p/(g\mu m^*)\hat{\boldsymbol{\theta}}$ [see Eq.~\eqref{kappa_theta}] coupled to the itinerant spins, with $g$ the g-factor and $\mu$ the Bohr magneton. \\
\begin{figure}[!ht]
\centering
\includegraphics[width=0.75\columnwidth]{SM_Fig_4.png}
\caption{One-dimensional Rashba and Dresselhaus polygon model showing the effective fields corresponding to CCW propagating spin carriers.}\label{phaseSOI}
\end{figure}
The solutions of the Schr\"{o}dinger equation are plane waves propagating along each side from vertex $\alpha$ towards vertex $\beta$ as
\begin{equation}
|\psi(r)\rangle=\mathrm{e}^{-\mathrm{i} k_\text{F} r} \mathrm{e}^{-\mathrm{i} \kappa r\hat{\boldsymbol{\theta}}_{\beta,\alpha} \cdot \boldsymbol{\sigma}}|\psi(0)\rangle,
\end{equation}
with $k_\text{F}$ the Fermi wavenumber. The first prefactor corresponds to the kinetic phase of the carrier associated with the dynamics of charged particles, while the second prefactor represents the spin phase due to spin precession. The propagation of a spin carrier from $\alpha$ to $\beta$ is then fully determined by the phases $ k_\text{F} L+\kappa L\hat{\boldsymbol{\theta}}_{\beta,\alpha} \cdot \boldsymbol{\sigma}$, and the spin evolution along one side is determined by the momentum-independent spin rotation operator:
\begin{equation}
R_{\beta,\alpha}=\exp \left[-\mathrm{i} \kappa L\hat{\boldsymbol{\theta}}_{\beta,\alpha} \cdot \boldsymbol{\sigma}\right]
\end{equation}
with $R_{\beta,\alpha}^\dagger=R_{\beta,\alpha}$ due to time-reversal symmetry.
The full spin evolution along counterclockwise (CCW) and clockwise (CW) propagating paths from vertex 1 to vertex $N$ is then given by the unitary operators
\begin{equation}
U_{+}(N)=R_{1 N} \ldots R_{32} R_{21}
\end{equation}
and
\begin{equation}
U_{-}(N)=R_{12} \ldots R_{N-1, N} R_{N 1}
\end{equation}
with $U_{-}(N)=U_{+}^\dagger (N)$, see Fig~\ref{phaseSOI}.
The spin rotation operator allows us to obtain the $xy$-projection of the spin texture as $\hat{\mathbf{s}}_{x,y}(r)=\left\langle\chi_{s}|\boldsymbol{\sigma_{x,y}}| \chi_{s}\right\rangle$, then it is possible to compute the angle accumulated around the $z$-axis by the itinerant spin state as the carrier propagates along each segment as the phase of the complex number $s_{x}(r)+\mathrm{i}s_{y}(r)$.
Completing a CCW round trip, we obtain the accumulated angle around the z-axis and, therefore the winding number.
\section*{Author Contributions}
A.H. and E.J.R. contributed equally to this work. All the authors contributed to write the manuscript.
\section*{Code availability}
The codes employed in this study are available from the authors on reasonable request.
\section*{Competing interests}
The authors declare no competing interests.
|
\section{Semantics}
\subsection{Abstract Syntax}
\paragraph{Features, types and constraints.}
Let \textsf{Id}\ be a finite set of features, let \textsf{Kinds}\xspace be a set of domain-specific feature kinds and let \textsf{Flavors}\xspace\ be a set of types that further define a feature's possible values. More precisely, $\textsf{Kinds}\xspace=\{$\textsf{package, component, option, interface}$\}$ and $\textsf{Flavors}\xspace=\{\textsf{none}\xspace, \textsf{bool}\xspace, \textsf{booldata}\xspace, \textsf{data}\xspace\}$. Furthermore, we introduce two types of expressions allowed in CDL: \textit{Goal expressions} and \textit{list expressions}.
Concerning the first one, we define $\text{\textit{Exp}}(\textsf{Id})$ to be a set of goal
expressions over \textsf{Id}, generated by the following grammar:
\begin{equation}
e ::= \textsf{id}
\mid \textsf{const}
\mid e \otimes e
\mid !e
\mid \text\textasciitilde e
\mid e \oplus e
\mid e \oslash e
\mid \textsf{Func}(e,e,...)
\mid e ? e : e
\end{equation}
Here, $\otimes\in\{\ensuremath{\mid\mid,\&\&,\textit{implies},\textit{eqv},\textit{xor}}\}$, $\oplus\in\{\ensuremath{+,-,*,/,\%,<<,>>,\text{\^{}},\&,\mid}\}$, $\oslash\in\{\ensuremath{==,\text{!=},<,>,<=,>=}\}$, $\textsf{Func}\in$\{\textit{get\_data, is\_active, is\_enabled, is\_loaded, is\_substr, is\_xsubstr, version\_cmp}\}\label{cdlfunctions}, $\textsf{id}\in\textsf{Id}$ and $\textsf{const}\in\textsf{Data}$, whereas \textsf{Data}\ is a set of untyped data (say all character strings).
The second type of expressions, so-called list expressions represent an enumeration of values or ranges, which can be computed by goal expressions. Thus, we define $\text{\textit{LExp}}(\textsf{Id})$ to be a set of list
expressions over goal expressions, generated by the following grammar ($e\in\text{\textit{Exp}}(\textsf{Id})$):
\begin{equation}
l ::= \left( e \mid e \textsf{ to } e \right) [\ \text{\textvisiblespace} l\ ]
\end{equation}
\paragraph{CDL models.}
\textsf{Cdl}\ is the set of all possible models
in CDL. Each CDL model $m\in\textsf{Cdl}$ is a
set of nodes, so $\textsf{Cdl} = \Powerset\textsf{Nodes}$, where
\begin{equation}
\textsf{Nodes} = \textsf{Id} \times \lceil\textsf{Id}\rceil \times \textsf{Flavors}\xspace \times \Powerset{\text{\textit{Exp}}(\textsf{Id})} \times \Powerset{\text{\textit{Exp}}(\textsf{Id})} \times \lfloor\text{\textit{Exp}}(\textsf{Id})\rfloor \times \lfloor\text{\textit{LExp}}(\textsf{Id})\rfloor \times\textsf{Kinds}\xspace \times \Powerset{\textsf{Id}}
\end{equation}
If $(n,p,fl,ai,req,cl,lv,knd,imp) \in \textsf{Nodes}$, then $n$ is the name, $p$ is the parent of the node ($p=\top$ for nodes at the top level), $ai$
is a set of \texttt{active\_if} visibility goal expressions, and $req$ is a set of
\texttt{requires} goal expressions. Further, $cl$ denotes a \texttt{calculated}\ goal expression that prescribes the feature's values and $lv$ is a \texttt{legal\_values} list expression restricting its values. Finally, $knd$ specifies the node's domain-specific kind and $imp$ specifies whether the node implements one or more interfaces. There is no further restriction on both values, that is, an interface can even implement other interfaces.
We write $\textsf{Id}(m)$ to denote names of nodes in the model $m$, so $\textsf{Id}(m) = \{
n \mid (n,\_,\_,\_,\_,\_,\_,\_,\_)\in m \}$.
\paragraph{Well-formedness.}
\label{par:wellformedness}
CDL introduces some more constraints on the syntax of the model. If
$(n,p,fl,ai,req,cl,lv,knd,imp) \in \textsf{Nodes}$, it has to fulfill the following invariants:
\begin{itemize}
\item $fl=none \rightarrow cl=\bot$ (\texttt{calculated} has no effect if flavor is \textsf{none}\xspace)
\item $cl\neq\bot \rightarrow lv=\bot$ (\texttt{calculated} and \texttt{legal\_values} exclude each other)
\item $fl \in \{bool\} \rightarrow lv=\bot$ (\texttt{legal\_values} applies to nodes with non-fixed data value only, see Eq.\ref{eq:domain})
\item $knd=\textit{interface} \rightarrow ( fl\neq\textsf{none}\xspace \wedge cl=\bot )$ (Interfaces must neither have the \textsf{none}\xspace\ flavor nor a \texttt{calculated} property)
\item The parent relationship $p$ should define a tree, with the virtual $\top$ as the root. Furthermore, nodes of kind \textit{option} must not be parents of other nodes.
\end{itemize}
\paragraph{Pre-processing notes.}
For convenience and conciseness reasons, the abstract syntax given in this section depends on the following preprocessing steps from the concrete syntax:
\begin{enumerate}
\item Similar to the configuration tool, we introduce a synthetic root element $\top$, which is a parent to every top-level package.
\item In case no flavor is specified for a node, we set the flavor ($fl$) property (according to the CDL documentation) to \textsf{booldata}\xspace for packages, to \textsf{bool}\xspace for components and options, and to \textsf{data}\xspace for interfaces.
\item The \texttt{requires}, \texttt{active\_if} and \texttt{calculated} properties can contain an enumeration of goal expressions separated by whitespace. We convert such enumerations to a disjunction of their goal expressions.
\end{enumerate}
\subsection{Semantic Domain}
A configuration is an assignment of triples of values to nodes. The set of
all possible configurations is:
\begin{equation}
\label{eq:domain}
\textsf{Confs} = \lceil\textsf{Id}\rceil \rightarrow (\{0,1\} \times \{0,1\} \times \textsf{Data})
\end{equation}
If $c\in\textsf{Confs}$ and $x\in\textsf{Id}$, we write $c(x)_1$ for the first component of the
valuation (the \emph{enabled state}), $c(x)_2$ for the second one (the \emph{enabled value}), and $c(x)_3$ for the third component of the valuation (the \emph{data value}). The first component specifies whether the node is actually in the configuration, that is, whether it influences the build of eCos\xspace in some sense. The latter two components refer to values the user can give to a node. We predefine the valuation of the $\top$ element as follows: $c(\top)_1=1, c(\top)_2=1, c(\top)_3=1$.
Now, a semantics of a CDL model is given in
terms of sets of configurations, so \Powerset{\textsf{Confs}} is our semantic domain and the semantic function has the signature:
\begin{equation}
\cdlDenotation\cdot : \textsf{Cdl} \rightarrow \Powerset\textsf{Confs}\\
\end{equation}
\begin{equation}
\denotation\cdot: \textsf{Model} \rightarrow \Powerset\textsf{Confs}\\
\end{equation}
\subsection{Semantics}
\paragraph{Helper functions.}
Let $\textit{access} : \textsf{Id} \times \textsf{Confs} \rightarrow \textsf{Data}$ denote a function that returns the value of a feature under a certain configuration while taking its enabled state into account.
\begin{equation} access(x,c) =
\begin{cases}
0 & \text{iff~}c(x)_1=0\\
c(x)_3 & \text{iff~}c(x)_1=1
\end{cases}
\end{equation}
Since arbitrary values can be returned for a feature's occurrence in an expression, and since they can be direct input to Boolean operators (e.g., ``feature A requires B \&\& C'' and C could have flavor data or booldata), we define a cast of arbitrary values to Boolean values in the TCL/TK style. More precisely, $\textit{bool}: \textsf{Data} \rightarrow \{0,1\}$. Please note that $\textit{bool}$ is also defined for plain Boolean values ($\{0,1\}\subset\textsf{Data}$), which are the return type if nodes are inactive, disabled, or bool.
\begin{equation} \textit{bool}( v ) =
\begin{cases}
0 & \text{iff~} v=\text{0+} \vee v=\text{""+} \\
1 & \text{otherwise}
\end{cases}
\end{equation}
For the evaluation of goal expressions, we define the function $\textit{eval}:\text{\textit{Exp}}(\textsf{Id})\times\textsf{Confs}\rightarrow \textsf{Data}$ recursively as follows, with $x\in \textsf{Id}$, $e_1,e_2,e_3\in\text{\textit{Exp}}(\textsf{Id})$ and $\textsf{const}\in \textsf{Data}$:
\begin{align}
\begin{split}
\textit{eval}(x,c)= &\textit{access}(x,c) \\
\textit{eval}(\textsf{const},c)=&\textsf{const} \\
\textit{eval}(!e_1,c)=&\textsf{non~}\textit{bool}(\textit{eval}(e_1,c)) \\
\textit{eval}(e_1 \otimes e_2,c)=& \phi_0(\textit{bool}(\textit{eval}(e_1,c)),\textit{bool}(\textit{eval}(e_2,c))) \\
&\text{ with } \phi_0=\textsf{vel,et,seq,eq,aut} \text{ for }\otimes=\ensuremath{\mid\mid,\&\&,\textit{implies},\textit{eqv},\textit{xor}} \\
\textit{eval}(e_1 \oplus e_2,c)=& \phi_1(\textit{eval}(e_1,c),\textit{eval}(e_2,c)) \\
&\text{ with }\phi_1\text{ TCL's arithmetic for }\oplus=\ensuremath{+,-,*,/,\%,<<,>>,\text{\^{}},\&,\mid} \\
\textit{eval}(e_1 \oslash e_2,c)=& \phi_2(\textit{eval}(e_1,c),\textit{eval}(e_2,c)) \\
&\text{ with }\phi_2\text{ TCL's comparison operators for }\oslash=\ensuremath{==,\text{!=},<,>,<=,>=} \\
\textit{eval}(e_1?e_2:e_3,c)=& \begin{cases}
\textit{eval}(e_2,c) & \text{iff~} \textit{bool}(\textit{eval}(e_1,c))\\
\textit{eval}(e_3,c) & \text{otherwise}
\end{cases}
\end{split}
\end{align}
We left out CDL's built-in functions (see \textsf{Func}\ in \ref{cdlfunctions}) in the definition of \textit{eval}\ and refer to the CDL online documentation instead.
For the evaluation of values against the \texttt{legal\_values} property, we introduce a satisfaction relation $ \models : \textsf{Data}\times\textsf{Confs}\times\text{\textit{LExp}}(\textsf{Id}) \rightarrow \{0,1\}$. For $d\in\textsf{Data}; c\in\textsf{Confs}; e_1,e_2\in\text{\textit{Exp}}(\textsf{Id}), l_1,l_2\in\text{\textit{LExp}}(\textsf{Id})$, we define the relation:
\begin{align}
\begin{split}
d &\models_c e_1 \hspace{1.5cm} \text{ iff } d = \textit{eval}(e_1,c) \\
d &\models_c e_1 \textsf{ to } e_2 \hspace{0.62cm} \text{ iff } \textit{bool}(\textit{eval}( e_1 <= d\ \&\&\ d <= e_2, c )) \\
d &\models_c l_1\ \text{\textvisiblespace}\ l_2 \hspace{0.85cm} \text{ iff } (d,c) \models l_1 \vee (d,c) \models l_2 \\
\end{split}
\label{eq:listevaluation}
\end{align}
\paragraph{Semantic function.}
The semantics of a model is just an intersection of denotations of all the
nodes except that we need to ensure that all the unloaded packages (that is, their nodes) are
enforced to be false. Furthermore, we adhere to the semantics of the nodes' \texttt{flavor} as well as we take their \texttt{calculated}/\texttt{legal\_values} property and the interface concept into account.
\begin{align}
\begin{split}
\cdlDenotation m = &
\left(\bigcap_{n\in m} \nodeDenotation n \right)
\cap
\left(\bigcap_{n\in m} \flavorDenotation n \right)
\cap
\left(\bigcap_{n\in m} \calculatedDenotation n \right)
\cap
\left(\bigcap_{n\in m} \lvDenotation n \right) \\
& \cap
\left(\bigcap_{n\in m} \interfaceDenotation n \right)
\cap
\left\{~ c\! \in\! \textsf{Confs} \mid c(x)_1=0 \text{ for all } x\in \textsf{Id}\setminus\textsf{Id}(m)
~\right\}
\end{split}
\end{align}
The semantics of a node is a set of all configurations that (1) contain the node's parent, (2) have the node's enabled value set, and (3) can satisfy the node's
constraints. Notably, we ignore the unloaded packages here, as we treated them
all globally above (here 'unloaded' = 'not mentioned in the model'). For brevity, we also introduce a macro for cross-tree constraints: $\textsf{CTC}=\forall e\in ai\cup req.\textit{bool}(\textit{eval}(e,c))$.
\begin{equation}
\nodeDenotation{(n,p,\_,ai,req,\_,_,\_,\_)} = \{ c \in \textsf{Confs} \mid c(n)_1 \leftrightarrow (c(p)_1 \wedge c(n)_2 \wedge \textsf{CTC} ) \}
\label{eq:nodeDenotation}
\end{equation}
Next, we continue with the denotation of nodes according to their flavor property. The values \textsf{none}\xspace and \textsf{data}\xspace are the equivalent to mandatory features in Feature Modeling (FM) with one big difference: In CDL, such nodes can be made optional with cross-tree constraints, whereas in FM, cross-tree constraints of mandatory features also constrain the parent. In the flavor denotation, we set the \textit{enabled value} for \textsf{none}\xspace and \textsf{data}\xspace features. With regard to Eq. \ref{eq:nodeDenotation}, such mandatory nodes still require the parent node present and cross-tree constraints satisfied.
\begin{align}
\begin{split}
\flavorDenotation{(n,p,(\textsf{none}\xspace|\textsf{data}\xspace),ai,req,\_,\_,\_,\_)}&=\{c\in\textsf{Confs} \mid c(n)_2\} \\
\flavorDenotation{(\_,\_,(\textsf{bool}\xspace|\textsf{booldata}\xspace),\_,\_,\_,\_,\_,\_)}&=\textsf{Confs}
\end{split}
\end{align}
The \texttt{calculated}\ property forces a node's \emph{data} and \emph{enabled value} depending on the flavor. Please note that CDL excludes the \textsf{none}\xspace flavor for calculated nodes (cf. Sec. \ref{par:wellformedness}, well-formedness). For $cl\neq\bot$, we define:
\begin{align}
\begin{split}
\calculatedDenotation{(n,\_,\textsf{bool}\xspace,\_,\_,cl,\_,\_,\_)}&=\{c\in\textsf{Confs}\mid c(n)_2=\textit{bool}(\textit{eval}(cl,c))\} \\
\calculatedDenotation{(n,\_,\textsf{booldata}\xspace,\_,\_,cl,\_,\_,\_)}&=\{c\in\textsf{Confs}\mid c(n)_3=\textit{eval}(cl,c) \wedge c(n)_2=\textit{bool}(c(n)_3) \} \\
\calculatedDenotation{(n,\_,\textsf{data}\xspace,\_,\_,cl,\_,\_,\_)}&=\{c\in\textsf{Confs}\mid c(n)_3=\textit{eval}(cl,c)\}
\end{split}
\end{align}
The \texttt{legal\_values} property restricts the data value of a node with a list expression. We define its denotation by using our satisfaction relation from Eq. \ref{eq:listevaluation}. Interestingly, the \texttt{legal\_values} property only excludes the flavor \textsf{bool}\xspace by well-formedness rules (cf. Sec. \ref{par:wellformedness}), whereas it does not have any effect on \textsf{none}\xspace -flavored nodes. For $lv\neq\bot$, we define:
\begin{align}
\begin{split}
\lvDenotation{(n,\_,(\textsf{booldata}\xspace\mid\textsf{data}\xspace),\_,\_,\_,lv,\_,\_)} &= \{c\in\textsf{Confs}\mid c(n)_3 \models_c lv \}\\
\lvDenotation{(n,\_,\textsf{none}\xspace,\_,\_,\_,lv,\_,\_)} &= \textsf{Confs}
\end{split}
\end{align}
Finally, we specify the denotation of interfaces, which represent derived features in CDL. However, please note that the \textsf{none}\xspace flavor is excluded by well-formedness rules (cf. Sec. \ref{par:wellformedness}).
\begin{align}
\begin{split}
\interfaceDenotation{(n,\_,\textsf{booldata}\xspace,\_,\_,\_,\_,\textit{interface},\_)} &= \{c\in\textsf{Confs}\mid c(n)_3 = \mid impls(n,c)\mid \wedge \\
& \hspace{2.34cm} c(n)_2=\textit{bool}(c(n)_3)\}\\
\interfaceDenotation{(n,\_,\textsf{data}\xspace,\_,\_,\_,\_,\textit{interface},\_)}&=\left\{c\in\textsf{Confs}\mid c(n)_3 = \mid impls(n,c)\mid \right\} \\
\interfaceDenotation{(n,\_,\textsf{bool}\xspace,\_,\_,\_,\_,\textit{interface},\_)}&=\left\{c\in\textsf{Confs}\mid c(n)_2 = \textit{bool}(\mid impls(n,c)\mid) \right\}
\end{split}
\end{align}
where $impls:\textsf{Id}\times\textsf{Confs}\rightarrow\Powerset\textsf{Nodes}$ is defined as follows:
\begin{equation}
impls(n,c)=\left\{x\in\textsf{Nodes} \mid n\in x_{impl} \wedge c(x)_1=1 \right\}
\end{equation}
\section{Propositional Semantics}
We now describe a Boolean interpretation of the abstract syntax. Given CDL's expressiveness, there is no precise translation from a CDL model into propositional logic. While it is relatively easy to translate the hierarchy and flavor constraints into a propositional formula, this task becomes more complicated for cross-tree constraints and the interface concept. Our strategy is to approximate constraints as much as possible by loosening the original constraints, that is, the propositional semantics under-approximate the full ones.
We tailor the full semantics down to denotations that can be expressed in propositional logic. Based on these semantics, we implemented\footnote{https://code.google.com/p/variability/wiki/CDLTools} rewriting rules that take a full CDL model and convert it into a Boolean formula, enabling analysis based on SAT solvers. The latter comprises, for instance, satisfiability and dead feature checks, or building implication graphs.
\subsection{Propositional Semantic Domain}
We define
\begin{equation}
\textsf{Confs}_p = \lceil\textsf{Id}\rceil \rightarrow \{0,1\}
\end{equation}
If $c_p\in\textsf{Confs}_p$ and $x\in\textsf{Id}$, we write $c_p(x)$ for the valuation of the node under a configuration. We also predefine $c_p(\top)=1$ for the $\top$ element. Similar to the full semantics, the propositional semantics of a CDL model is given in terms of sets of configurations, so \Powerset{\textsf{Confs}_p} is our semantic domain. Thus, our semantic function has the following signature:
\begin{equation}
\pCdlDenotation\cdot : \textsf{Cdl} \rightarrow \Powerset{\textsf{Confs}_p}\\
\end{equation}
Furthermore, we define some invariants between the full (\textsf{Confs}) and propositional ($\textsf{Confs}_p$) configuration spaces, basically answering the question: What does it mean if a feature under a Boolean configuration $c_p\in\textsf{Confs}_p$ is true or false with regard to the full semantics $c\in\textsf{Confs}$? Table \ref{tab:invariantsBetweenConfigurationSpaces} shows the invariants according to a node's flavor.
\begin{table*}[h]\centering
\begin{tabular}{ll}
Flavor & Invariant \\
\midrule
\textsf{bool}\xspace & $c_p(n) = c(n)_1$ \\
\textsf{none}\xspace & $c_p(n) = c(n)_1$ \\
\textsf{booldata}\xspace & $c_p(n) = c(n)_1 \wedge c(n)_3\neq0$ \\
\textsf{data}\xspace & $c_p(n) = c(n)_1 \wedge c(n)_3\neq0$ \\
\end{tabular}
\caption{Invariants between configuration spaces}
\label{tab:invariantsBetweenConfigurationSpaces}
\end{table*}
\subsection{Propositional Semantics}
\paragraph{Helper functions.}
A function such as $\textit{access}_p : \textsf{Id} \times \textsf{Confs}_p \rightarrow \{0,1\}$ is not necessary anymore, since a node does only have one value left (that is, $access_p(\textsf{id},c) = c_p(\textsf{id})$). However, a slightly different function $impls':\textsf{Id}\times\textsf{Cdl}\rightarrow\Powerset\textsf{Id}$ will be helpful later in this section:
\begin{equation}
impls'(i,m) = \{n\in m \mid i \in n_{impl}\}
\end{equation}
\paragraph{Boolean expressions.}
We just have to consider goal expressions since list expressions only appear in \textit{legal\_values} constraints, which cannot be approximated\footnote{One could argue that it is possible to approximate special cases, such as \texttt{legal\_values} 0 and so on. However, it would spoil our translation with too many exceptions and we have not seen comparable examples in the real models.} in the propositional semantics since our semantic domain contains no \textit{data value} any more.
Let $\textit{BExp}(\textsf{Id})\subset\text{\textit{Exp}}(\textsf{Id})$ be a subset of Boolean expressions over \textsf{Id}, which is defined by the following grammar, with $\otimes=\{\ensuremath{\mid\mid,\&\&,\textit{implies},\textit{eqv}}\}$ and $\textsf{const}\in\{0,1\}$:
\begin{equation}
e ::= \textsf{id}
\mid \textsf{const}
\mid e \otimes e
\mid !id
\end{equation}
\paragraph{Boolean expression evaluation.} The evaluation of \textit{BExp}(\textsf{Id}) now follows ordinary propositional semantics. Thus, the definition of a function $\textit{$eval_p$}:\textit{BExp}(\textsf{Id})\times\textsf{Confs}_p\rightarrow\{0,1\}$ is pretty straightforward, with $x\in \textsf{Id}$, $e_1,e_2\in\textit{BExp}(\textsf{Id})$ and $\textsf{const}\in\{0,1\}$:
\begin{align}
\begin{split}
\textit{$eval_p$}(x,c_p)= &c_p(x) \\
\textit{$eval_p$}(\textsf{const},c_p)=&\textsf{const} \\
\textit{$eval_p$}(!e_1,c_p)=&\textsf{non~}\textit{$eval_p$}(e_1,c_p) \\
\textit{$eval_p$}(e_1 \otimes e_2,c_p)=& \phi_0(\textit{$eval_p$}(e_1,c_p)),\textit{$eval_p$}(e_2,c_p)) \\
&\text{ with } \phi_0=\textsf{vel,et,seq,eq} \text{ for }\otimes=\ensuremath{\mid\mid,\&\&,\textit{implies},\textit{eqv}} \\
\end{split}
\end{align}
\paragraph{Expression rewriting rules.}
Next, we define a partial function \textit{rewrite}\xspace:$\text{\textit{Exp}}(\textsf{Id})\times\textsf{Cdl}\leadsto\textit{BExp}(\textsf{Id})$, which translates goal expressions from the full semantics to reduced Boolean ones. For $x\in\textsf{Id}$, $m\in\textsf{Cdl}$ and $e_1,e_2,e_3\in \text{\textit{Exp}}(\textsf{Id})$:
\begin{align}
\begin{split}
rewrite(x,m) &= \begin{cases}
x &\text{iff~} x\in\textsf{Id}(m)\\
0 &\text{otherwise}\\
\end{cases}\\
rewrite( !x, m ) &= \neg rewrite(x,m)\\
rewrite(\textsf{const},m) &= \textit{bool}( \textsf{const} )\\
rewrite(x=\textsf{const},m) &= \begin{cases}
rewrite(x,m) &\text{iff~} \textit{bool}(\textsf{const})\neq0\\
rewrite(\neg x,m) &\text{otherwise}\\
\end{cases}\\
rewrite(x>\textsf{const},m) &= \begin{cases}
rewrite(x,m) &\text{iff~} \textsf{const}\in INT \wedge \textsf{const}\geq 0\\
1 &\text{otherwise} \text{ (drop it)}\\
\end{cases}\\
rewrite(\textit{is\_substr}(x,\textsf{const}),m) &= rewrite(x,m)\\
rewrite(e_1 \otimes e_2,m) &= rewrite(e_1,m) \otimes rewrite(e_2,m)\\
rewrite(e_1 ? e_2 : e_3,m) &= (rewrite(e_1,m) \rightarrow rewrite(e_2,m))\\
& \hspace{0.5cm} \wedge (\neg rewrite(e_1,m) \rightarrow rewrite(e_3,m))
\end{split}
\end{align}
For the interpretation of interfaces, we need to define a helper function $choose:\Powerset\textsf{Id}\times\mathbb{N}\times\mathbb{N}\rightarrow\textit{BExp}(\textsf{Id})$. More precisely, $choose(ids,min,max)$ converts a set of \textsf{Id} s into a Boolean expression, where at least \textit{min} and at most \textit{max} \textsf{Id} s can be satisfied simultaneously.
For $x\in\textsf{Id}(m)$ and if $x$ denotes an interface, we continue the definition of \textit{rewrite}\xspace as follows:
\begin{align}
\begin{split}
rewrite(x=0,m) &=\neg x \wedge \bigwedge_{i\in impls'(x,m)}{\neg i} \\
rewrite(x>0,m) &=x \wedge \bigvee_{i\in impls'(x,m)}{i} \\
rewrite(x=1,m) &=x \wedge \bigxor_{i\in impls'(x,m)}{i} \\
rewrite(x>=\textsf{const},m) &= x \wedge \textit{choose}(impls'(x,m),\textsf{const},\mid impls'(x,m) \mid) \\
rewrite(x>\textsf{const},m) &= x \wedge \textit{choose}(impls'(x,m),\textsf{const} + 1,\mid impls'(x,m) \mid)
\end{split}
\end{align}
\paragraph{Semantic function}
The propositional semantics of a model $m\in\textsf{Cdl}$ is just an intersection of the propositional denotations of all the nodes, similar to the full semantics. However, we (have to) leave out \texttt{legal\_values} as already pointed out. Furthermore, we need some more context (the current model $m\in\textsf{Cdl}$) for the \textit{rewrite} function, since our propositional semantic domain is not capable of carrying enough information any more.
\begin{equation}
\begin{split}
\pCdlDenotation m = &
\left(\bigcap_{n\in m} \pNodeDenotation{ n, m } \right)
\cap
\left(\bigcap_{n\in m} \pFlavorDenotation{ n, m } \right)
\cap
\left(\bigcap_{n\in m} \pCalculatedDenotation{ n, m } \right)
\cap
\left(\bigcap_{n\in m} \pInterfaceDenotation{ n, m } \right)\\
& \cap
\left\{~ c\! \in\! \textsf{Confs}_p \mid c_p(x)=0 \text{ for all } x\in \textsf{Id}\setminus\textsf{Id}(m)
~\right\}
\end{split}
\end{equation}
The semantics of a node is a set of all configurations that can satisfy its
constraints. For the propositional version, we introduce the macro $\textsf{CTC}_p=\forall e\in ai\cup req.(e\in dom(rewrite) \rightarrow \textit{eval}(rewrite(e,m),c_p)))$.
\begin{equation}
\pNodeDenotation{(n,p,\_,ai,req,\_,\_,\_,\_),m} = \{ c_p \in \textsf{Confs}_p \mid c_p(n) \rightarrow c_p(p) \wedge CTC_p \}
\end{equation}
We continue with the denotation of nodes according to their flavor property:
\begin{align}
\begin{split}
\pFlavorDenotation{(n,p,(\textsf{none}\xspace|\textsf{data}\xspace),\_,\_,\_,\_,\_,\_),m}&=\{c_p\in\textsf{Confs}_p \mid c_p(p) \wedge CTC_p \rightarrow c_p(n)\} \\
\pFlavorDenotation{(n,p,(\textsf{bool}\xspace|\textsf{booldata}\xspace),\_,\_,\_,\_,\_,\_),m}&=\textsf{Confs}_p
\end{split}
\end{align}
Similarly, we define the denotation of the \texttt{calculated}\ property. For $cl\neq\bot$:
\begin{align}
\begin{split}
\pCalculatedDenotation{(n,p,\_,\_,\_,cl,\_,\_,\_),m}=\{c_p\in\textsf{Confs}\mid c_p(p)& \wedge CTC_p \rightarrow \\
&c_p(n)=\textit{eval}(\textit{rewrite}(cl,m),c_p)\}
\end{split}
\end{align}
Finally, the propositional denotation of interfaces can be declared as follows:
\begin{align}
\begin{split}
\pInterfaceDenotation{(n,\_,\_,\_,\_,\_,\_,\textit{interface},\_),m}=\{c_p\in\textsf{Confs}_p \mid & c_p(p) \wedge CTC_p \rightarrow\\
& c_p(n)=\textit{eval}(\bigvee_{i\in impls'(x,m)}{i},c_p) \}
\end{split}
\end{align}
\begin{comment}
Where $\textit{impls}_p:\textsf{Id}\times\textsf{Confs}\rightarrow\{0,1\}$ is (in the propositional version) defined as follows:
\begin{equation}
\textit{impls}_p(x,c_p) = \exists (n,\_,\_,\_,\_,\_,\_,\_,impl)\in\textsf{Nodes}.\left(x\in impl \wedge c_p(n) \right)
\end{equation}
\section{Some Notes on Syntactic Translation}
\paragraph{Invariants.} First of all, let's define an interpretation of $\textsf{Confs}_p$ with regard to the original full semantics under \textsf{Confs}, basically answering the question: What does it mean if a feature under a Boolean configuration $c_p\in\textsf{Confs}_p$ is true or false with regard to the full semantics? Thus, depending on the nodes' structure, we define the following invariants, with $c'\in\textsf{Confs}_p$, $c\in\textsf{Confs}$.
\\
\\
For packages, options and components:
\begin{table*}[h]\centering
\begin{tabular}{ll}
flavor & invariant \\
\midrule
\textsf{bool}\xspace & $c'(n) = c(n)_1$ \\
\textsf{none}\xspace & $c'(n) = c(n)_1$ \\
\textsf{booldata}\xspace & $c'(n) = c(n)_1 \wedge c(n)_3\neq0$ \\
\textsf{data}\xspace & $c'(n) = c(n)_1 \wedge c(n)_3\neq0$ \\
\end{tabular}
\caption{invariants between configuration spaces}
\label{tab:invariantsBetweenConfigurationSpaces2}
\end{table*}
\begin{align}
test
\end{align}For interfaces, the situation is a bit more subtle.
\paragraph{Hierarchy, node types and flavor.}
\begin{itemize}
\item For every node $X$ we add a conjunct $X\rightarrow parent(X)$
\item For every node (except interfaces) $X$ of flavor \textsf{none}\xspace\ or \textsf{data}\xspace\, we add a conjunct $parent(X)\rightarrow X$
\item For every node $X$ of flavor \textsf{bool}\xspace\ and calculated="1", we add a conjunct $X\rightarrow parent(X)$
\end{itemize}
\paragraph{Requires and Active\_If Expressions.}
\begin{itemize}
\item For every node $X$ we add a conjunct $X\rightarrow parent(X)$
\end{itemize}
\newpage
|
\section{Introduction}
\label{intro}
Prime factorization is a problem in the complexity class NP of problems that can be solved in
polynomial time by a nondeterministic machine. Indeed, the prime factors can be verified
efficiently by multiplication.
At present, it is not known if the problem has polynomial computational complexity
and, thus, is in the complexity class P. Nonetheless, the most common cryptographic algorithms
rely on the assumption of hardness of factorization. Website certificates
and bitcoin wallets are examples of resources depending on that
assumption. Since no strict lower bound on the computational complexity is actually
known, many critical services are potentially subject to
future security breaches. Consequently, cryptographic keys have gradually increased
their length to adapt to new findings. For example, the general number-field sieve~\cite{nfsieve}
can break keys that would have been considered secure against previous factoring methods.
Prime factorization is also important for its relation with quantum computing, since
an efficient quantum algorithm for factorization is known. This algorithm is considered
a main argument supporting the supremacy of quantum over classical computing. Thus,
the search for faster classical algorithms is relevant for better understanding the
actual gap between classical and quantum realm.
Many of the known factoring methods use the ring of integers modulo $c$ as a common feature, where
$c$ is the number to be factorized. Examples are Pollard's $\rho$ and $p-1$
algorithms~\cite{pollard1,pollard2}, Williams' $p+1$ algorithm~\cite{williams},
their generalization with cyclotomic polynomials~\cite{bach}, Lenstra elliptic curve
factorization~\cite{lenstra}, and quadratic sieve~\cite{qsieve}.
These methods end up to generate a number, say $m$, having a
factor of $c$. Once $m$ is obtained, the common factor can be efficiently
computed by the Euclidean algorithm. Some of these methods use only operations defined in the ring. Others,
such as the elliptic-curve method, perform also division operations by pretending that $c$ is prime.
If this operation fails at some point, the divisor is taken as outcome $m$.
In other words, the purpose of these methods is to compute a zero over the field of integers modulo
some prime factor of $c$ by possibly starting from some random initial state. Thus, the general scheme
is summarized by a map $X\mapsto m$ from a random state $X$ in some set $\Omega$ to an integer $m$
modulo $c$. Different states may be tried until $m$ is equal to zero modulo some prime of $c$.
The complexity of the algorithm depends on the computational complexity of generating $X$ in $\Omega$,
the computational complexity of evaluating the map, and the average number of trials required to find a zero.
In this paper, we employ this general scheme by focusing on a class $\Theta$
of maps defined as multivariate rational
functions over prime fields $\mathbb{Z}_p$ of order $p$ and, more generally, over a finite
field $\text{GF}(q)$ of order $q=p^k$ and degree $k$. The set $\Omega$ of inputs is taken equal to
the domain of definition of the maps. More precisely, the maps are first defined over some algebraic
number field $\mathbb{Q}(\alpha)$ of degree $k_0$, where $\alpha$ is an algebraic number, that is,
solution of some irreducible polynomial $P_I$ of degree $k_0$. Then, the maps
are reinterpreted over a finite field.
Using the general scheme of the other methods, the class $\Theta$
takes to a factoring algorithm with polynomial complexity if
\begin{enumerate}
\item[(a)] the number of distinct zeros, say $N_P$, of the maps in $\Theta$ is arbitrarily large over $\mathbb{Q}(\alpha)$;
\item[(b)] a large fraction of the zeros remain distinct when reinterpreted over a finite field
whose order is greater than about $N_P^{1/M}$;
\item[(c)] \label{cond_c}
the product of the number of parameters by the field degree is upper-bounded by a
sublinear power function of $\log N_P$;
\item[(d)] the computational complexity of evaluating the map given any input is upper-bounded
by a polynomial function in $\log N_P$.
\end{enumerate}
A subexponential factoring complexity is achieved with weaker scaling conditions on
the map complexity, as discussed later in Sec.~\ref{sec_complex}.
Later, this approach to factorization will be reduced to the search of rational points of a
variety having an arbitrarily large number of rational intersection points with a hypersurface.
The scheme employing rational functions resembles some existing methods, such as Pollard's $p-1$
algorithm. The main
difference is that these algorithms generally rely on algebraic properties over finite
fields, whereas the present scheme relies on algebraic properties over the
field $\mathbb{Q}(\alpha)$. For example,
Pollard's method ends up to build a polynomial $x^n-1$ with $p-1$ roots over a finite
$\mathbb{Z}_p$, where $p$ is some prime factor of $c$. This feature of the polynomial
comes from Fermat's little theorem and is satisfied if the integer $n$ has $p-1$ as factor.
Thus, the existence of a large number of zeros of $x^n-1$ strictly
depends on the field. Indeed, the polynomial does not have more than $2$ roots over the
rationals. In our scheme, the main task is to find rational functions having a
sufficiently large number of zeros over an algebraic number field. This feature is then
inherited by the functions over finite fields.
Some specific properties of finite fields can eventually be useful, such as the reducibility
of $P_I$ over $\mathbb{Z}_p$. This will be mentioned later.
Let us illustrate the general idea with an example.
Suppose that the input of the factorization problem is $c=p p'$, with $p$ and $p'$
prime numbers and $p<p'$. Let the map be a univariate polynomial of the form
\be\label{Eq1}
P(x)=\prod_{i=1}^{N_P}(x-x_i),
\ee
where $x_i$ are integer numbers somehow randomly distributed
in an interval between $1$ and $i_{max}\gg p'$. More generally, $x_i$ can be rational
numbers $n_i/m_i$ with $n_i$ and/or $m_i$ in $\{1,\dots,i_{max}\}$.
When reinterpreted modulo $p$ or modulo $p'$, the numbers
$x_i$ take random values over the finite fields. If $N_P<p$, we expect that the polynomial
has about $N_P$ distinct roots over the finite fields. Thus, the probability that
$P(x)\mod p=0$ or $P(x)\mod p'=0$ is about $N_P/p$ or $N_P/p'$, respectively, which are the
ratio between the number of zeros and the size of the input space $\Omega$ over the finite fields.
The probability that $P(x)\mod c$ contains a nontrivial factor of $c$ is about
$\frac{N_P}{p}\left(1-\frac{N_P}{p'}\right)+\frac{N_P}{p'}\left(1-\frac{N_P}{p}\right)$.
Thus, if $N_P$ is of the order of $p$, we can get a nontrivial factor
by the Euclidean algorithm in few trials. More specifically, if $p\simeq p'$ and
$N_P\simeq \sqrt{c}/2$, then the probability of getting a nontrivial factor is roughly
$1/2$. It is clear that a computational complexity of the map scaling subexponentially or
polynomially in $\log N_P$ leads to a subexponential or polynomial
complexity of the factoring algorithm. Thus, the central problem is to build a polynomial
$P(x)$ or a rational function with friendly computational properties with respect to
the number of zeros.
The scheme can be generalized by taking multivariate
maps with $M$ input parameters. In this case, the number of zeros needs to be of the order of
$p^M$, which is the size of the input space over the field $\mathbb{Z}_p$. As a further
generalization, the rational field can be replaced by an algebraic number field
$\mathbb{Q}(\alpha)$ of degree $k_0$. A number in this field is represented as a
$k_0$-dimensional vector over the rationals. Reinterpreting the components of
the vector over a finite field $\mathbb{Z}_p$, the size of the sampling space
is $p^{k_0 M}$, so that we should have $N_P\sim p^{k_0 M}$ in order to get nontrivial
factors in few trials. Actually, this is the worst-case scenario since the reinterpretation
of $\mathbb{Q}(\alpha)$ modulo $p$ can lead to a degree of the finite field much smaller
than $k_0$. For example, if $\alpha$ is the root $e^{2\pi i/n}$ of the polynomial $x^n-1$,
the degree of the corresponding finite field with characteristic $p$ collapses to $1$ if
$n$ is a divisor of $p-1$.
On one hand, it is trivial to build polynomials with an arbitrarily large number $N_P$
of roots over the rationals as long as the computational cost grows linearly in $N_P$.
On the other hand, it is also simple to build polynomials with friendly computational
complexity with respect to $\log N_P$ if the roots are taken
over algebraically closed fields. The simplest example is the previously mentioned
polynomial $P(x)=x^n-1$, which has $n$ distinct complex roots and a computational complexity
scaling as $\log n$. However, over the rationals, this polynomial has at most
$2$ roots. We can include other roots by extending the rational field to an algebraic number
field, but the extension would have a degree proportional to the number of roots, so that
the computational complexity of evaluating $P(x)$ would grow polynomially in the number
of roots over the extension.
\subsection{Algebraic-geometry rephrasing of the problem}
It is clear that an explicit definition of each root of polynomial~(\ref{Eq1}) leads
to an amount of memory allocation growing exponentially in $\log c$, so that the resulting
factoring algorithm is exponential in time and space. Thus, the roots has to be defined
implicitly by some simple rules. Considering a purely algebraic definition, we associates
the roots to rational solutions of a set of $n$ non-linear polynomial equations in
$n$ variables ${\bf x}=(x_1,\dots,x_n)$,
\be
P_k({\bf x})=0,\;\;\;\; k\in\{0,\dots,n-1\}.
\ee
The solutions are intersection points of $n$ hypersurfaces.
The roots of $P(x)$ are defined as the values of some coordinate, say $x_n$, at the
intersection points. By eliminating the $n-1$ variables $x_1,\dots,x_{n-1}$, we end
up with a polynomial $P(x_n)$ with a number of roots generally growing exponentially in $n$.
This solves the problem of space complexity in the
definition of $P(x)$. There are two remaining problems. First, we have to choose
the polynomials $P_0,\dots,P_{n-1}$ such that an exponentially large fraction of
the intersection points are rational. Second, the variable elimination, given a value
of $x_n$, has to be performed as efficiently as possible over a finite field. If
the elimination has polynomial complexity, then factorization turns out to have
polynomial complexity. Note that the elimination of $n-1$ variables given $x_n$
is equivalent to a consistency test of the $n$ polynomials. The first problem can be
solved in a simple way by defining the polynomials as elements of the ideal generated
by products of linear polynomials. Let us denote a linear polynomial with a symbol
with a hat. For example, the quadratic polynomials
\be\label{quadr_polys}
G_i=\hat a_i\hat b_i, \;\;\;\;\ i\in\{1,\dots,n\}
\ee
have generally $2^n$ rational common zeros, provided that the coefficients of $\hat a_i$
and $\hat b_i$ are rational.
Identifying the polynomials $P_0,\dots,P_{n-1}$ with elements of the ideal generated
$G_1,\dots,G_n$, we have
\begin{equation}
\label{poly_eqs_simple}
P_k=\sum_i c_{k,i}({\bf x})\hat a_i \hat b_i,\;\;\; i\in\{0,\dots,n-1\},
\end{equation}
whose set of common zeros contains the $2^n$ rational points of the generators $G_i$.
In particular, if the polynomials $c_{k,i}$ are set equal to constants, then
the system $P_0=\dots P_{n-1}=0$ is equivalent to the system $G_1=\dots G_n=0$.
At this point, the variable elimination is the final problem. A working method is to
compute a Gr\"obner basis. For the purpose of factorizing $c=p p'$, the task is
to evaluate a Gr\"obner basis to check if the $n$ polynomials given $x_n$ are
consistent modulo $c$. If they are consistent modulo some non-trivial factor $p$ of $c$,
we end up at some point with some integer equal to zero modulo $p$.
However, the complexity of this computation is doubly exponential in the worst
case. Thus, we have to search for a suitable set of polynomials with a large set of
rational
zeros such that there is an efficient algorithm for eliminating $n-1$ variables.
The variable elimination is efficient if $n-1$ out of the $n$ polynomial equations
$P_k=0$ form a suitable triangular system for some set of low-degree polynomials $c_{k,i}$.
Let us assume that the last $n-1$ polynomials have
the triangular form
$$
\left.
\begin{array}{r}
P_{n-1}(x_{n-1},x_n) \\
P_{n-2}(x_{n-2},x_{n-1},x_n) \\
\dots \\
P_1(x_1,\dots,x_{n-2},x_{n-1},x_n)
\end{array}
\right\},
$$
such that the $k$-th polynomial is linear in $x_k$. Thus, the corresponding polynomial
equations can be sequentially solved in the first $n-1$ variables
through the system
\be
\label{rational_system}
\begin{array}{l}
x_{n-1}=\frac{{\cal N}_{n-1}(x_n)}{{\cal D}_{n-1}(x_n)} \\
x_{n-2}=\frac{{\cal N}_{n-2}(x_n,x_{n-1})}{{\cal D}_{n-2}(x_n,x_{n-1})} \\
\dots \\
x_1=\frac{{\cal N}_1(x_n,x_{n-1},\dots,x_2)}{{\cal D}_1(x_n,x_{n-1},\dots,x_2)},
\end{array}
\ee
where ${\cal D}_k\equiv \partial P_k/\partial x_k$ and ${\cal N}_k\equiv P_k|_{x_k=0}$.
This system defines a parametrizable curve,
say $\cal V$, in the algebraic set defined by the polynomials $P_1,\dots,P_{n-1}$, the variable
$x_n$ being the parameter.
Let us remind that a curve is parametrizable if and only if its geometric genus is equal to
zero.
The overall set of variables can be efficiently computed over a finite field, provided that
the polynomial coefficients $c_{k,i}$ are not too complex. Once determined the variables,
the remaining polynomial $P_0$ turns out to be equal to zero if ${\bf x}$
is an intersection point. Provided that the rational intersection points have distinct values
of $x_n$ (which is essentially equivalent to state that the points are distinct and
are in the variety $\cal V$),
then the procedure generates a value $P_0({\bf x})$ which is zero modulo $p$ with
high probability if $p$ is of the order of the number of rational intersection
points. For this inference, it is pivotal to assume that a large fraction of rational points
remain distinct when reinterpreted over the finite field.
This algebraic-geometry rephrasing of the problem can be stated in a more general form.
Let $\cal V$ and $\cal H$ be some irreducible curve in an $n$-dimensional space and a hypersurface,
respectively. The curve $\cal V$ is not necessarily parametrizable, thus its genus may take
strictly positive values. The points in $\cal H$ are the zero locus
of some polynomial $P_0$. Let $N_P$ be the
number of distinct intersection points between $\cal V$ and $\cal H$ over the rational field $\mathbb{Q}$.
Over a finite field $\text{GF}(q)$, Weil's theorem states that the number of rational
points, says $N_1$, of a smooth curve is bounded by the inequalities
\be\label{weil_bounds}
-2 g \sqrt{q} \le N_1-(q+1)\le 2 g\sqrt{q},
\ee
where $g$ is the geometric genus of the curve.
Generalizing to singular curves~\cite{aubry}, we have
\be\label{aubry_eq}
-2 g \sqrt{q}-\delta \le N_1-(q+1)\le 2 g\sqrt{q}+\delta,
\ee
where $\delta$ is the number of singularities, properly counted. These inequalities have the
following geometric interpretation. For the sake of simplicity, let us assume that the
singularities are ordinary double points.
A singular curve, says $\cal S$, with genus $g$ is birationally
equivalent to a smooth curve, say $\cal R$, with same genus, for which Wiel's theorem holds.
That is, the rational points of $\cal R$ are bijectively mapped to the non-singular
rational points of $\cal S$,
apart from a possible finite set $\Omega$ of $2m$ points mapping to $m$ singular points
of $\cal S$.
The cardinality of $\Omega$ is at most $2\delta$ (attained when the $\delta$ singularities
have tangent vectors over the finite field). We have two extremal cases.
In one case, $\#\Omega=2\delta$, so that, $\cal S$ has $\delta$ points less than $\cal R$
(two points in $\Omega$ are merged into a singularity of $\cal S$). This gives the lower
bound in~(\ref{aubry_eq}). In the second case, $\Omega$ is empty and the singular points of
$\cal S$ are rational points. Thus, $\cal S$ has $\delta$ rational points more.
Given this interpretation, Weil's upper bound still holds for the number of
non-singular rational points, say $N_1'$,
\be\label{weil2}
N_1'\le (q+1)+2 g \sqrt{q}.
\ee
Thus, if the genus is much smaller than $\sqrt{q}$, $N_1'$ is upper-bounded by a number close
to the order of the field.
Now, let us assume that most of the $N_P$ rational points in ${\cal V}\cap{\cal H}$ over $\mathbb{Q}$
remain distinct when reinterpreted over $\mathbb{Z}_p$, with $p\simeq a N_P$, where $a$ is a number
slightly greater than $1$, say $a=2$. We also assume that
these points are not singularities of $\cal V$. Weil's inequality~(\ref{weil2}) implies that
the curve does not have more than about $p$ points over $\mathbb{Z}_p$. Since
$p\gtrsim N_1'\gtrsim N_P\sim p/2$, we have that the number of non-singular rational points of
the curve is about the number of intersection points over $\mathbb{Z}_p$. This implies that a
large fraction of points $\bf x\in \cal V$ over the finite field are also points of $\cal H$. We
have the following.
\begin{claim}
\label{claim_rat_curve}
Let ${\cal V}$ and ${\cal H}$ be an algebraic curve with genus $g$ and a hypersurface, respectively.
The hypersurface is the zero locus of the polynomial $P_0$.
Their intersection has $N_P$ distinct points over the rationals, which are not singularities
of $\cal V$. Let us also assume that $g\ll \sqrt{N_P}$ and that most of the $N_P$ rational points remain
distinct over $\mathbb{Z}_p$ with $p\gtrsim 2 N_P$.
If we pick up at random a point in $\cal V$, then
\be
P_0({\bf x})=0 \mod p
\ee
with probability equal to about ratio $N_P/p$.
\end{claim}
If there are pairs $({\cal V},{\cal H})$ for every $N_P$ that satisfy the premises of this claim,
then prime factorization is reduced to the search of rational points of a curve. Actually,
these pairs always exist, as shown later in the section. Assuming that
$c=p p'$ with $p\sim p'$, the procedure for factorizing $c$ is as follows.
\begin{enumerate}
\item[(1)] Take a pair $({\cal V},{\cal H})$ with $N_P\sim c^{1/2}$ such that the premises of
Claim~\ref{claim_rat_curve} hold.
\item[(2)] Search for a rational point ${\bf x}\in\cal V$ over $Z_p$.
\item[(3)] Compute $\text{GCD}[P_0({\bf x}),c]$, the greatest common divisor of $P_0({\bf x})$ and $c$.
\end{enumerate}
The last step gives $\text{GCD}[P({\bf x}),c]$ equal to
$1$, $c$ or one of the factors of $c$. The probability of getting
a nontrivial factor can be made close to $1/2$ with a suitable tuning of $N_P$ (as shown later in
Sec.~\ref{sec_algo}).
Finding rational points of a general curve with genus greater than $2$ is an exceptionally complex
problem. For example, just to prove that the plane curve $x^h+y^h=1$ with $h>2$ does not have zeros over
the rationals took more than three centuries since Fermat stated it.
Curves with genus $1$ are elliptic curves, which have an important role
in prime factorization (See Lenstra algorithm~\cite{lenstra}). Here, we will focus on
parametrizable curves, which have genus $0$. In particular, we will consider parametrizations
generated by the sequential equations~(\ref{rational_system}). It is interesting to note that
it is always possible to find a curve $\cal V$ with parametrization~(\ref{rational_system})
and a hypersurface $\cal H$ such that their intersection contains a given set of rational points.
In particular, there is a set of polynomials $P_0,\dots,P_{n-1}$ of the form~(\ref{poly_eqs_simple}),
such that the zero locus of the last $n-1$ polynomials contains a parametrizable curve with
parametrization~(\ref{rational_system}), whose intersection with the hypersurface $P_0=0$ contains
$2^n$ distinct rational points. Let the intersection points be defined by the
polynomials~(\ref{quadr_polys}). Provided that $x_n$ is a separating variable, the set of
intersection points admits the rational univariate representation~\cite{rouillier}
\be
\left\{
\begin{array}{l}
x_{n-1}=\frac{\bar{\cal N}_{n-1}(x_n)}{{\bar{\cal D}}_{n-1}(x_n)} \\
x_{n-2}=\frac{{\bar{\cal N}}_{n-2}(x_n)}{\bar{\cal D}_{n-2}(x_n)} \\
\dots \\
x_1=\frac{\bar{\cal N}_1(x_n)}{\bar{\cal D}_1(x_n)} \\
{\bar{\cal N}}_0(x_n)=0
\end{array}
\right.
\ee
The first $n-1$ equations are a particular form of equation~(\ref{rational_system}) and
define a parametrizable curve with $x_n$ as parameter. The last equation can be replaced
by some linear combination of the polynomials $G_i$. It is also interesting
to note that the rational univariate representation is unique once the separating
variable is chosen. This means that the parametrizable curve is uniquely determined by
the set of intersection points and the variable that is chosen as parameter.
It is clear that the curve and hypersurface obtained through this construction with a
general set of polynomials $G_i$ satisfy
the premises of Claim~\ref{claim_rat_curve}. Indeed, a large part of the common zeros
of the polynomials $G_i$ are generally distinct over a finite field $\mathbb{Z}_p$
with $p\simeq N_P$. For example, the point $\hat a_1=\dots=\hat a_n=0$ is distinct from the other
points if and only if $\hat b_i\ne 0$ at that point for every $i\in\{1,\dots,n\}$.
Thus, the probability that a given point is not distinct over $\mathbb{Z}_p$ is of
the order of $p^{-1}\sim N_P^{-1}$, hence a large part of the points are distinct over the finite
field. There is an apparent paradox. With a suitable choice of the linear functions $\hat a_i$
and $\hat b_i$, the intersection points can be made distinct over a field $\mathbb{Z}_p$ with
$p\ll N_P$, which contradicts Weil's inequality~(\ref{weil2}). The contradiction is explained
by the fact that the curve is broken into the union of reducible curves over the finite field.
In other words, some denominator ${\cal D}_k$ turns out to be equal to zero modulo $p$
at some intersection points. This may happen also with $p\sim N_P$, which is not a concern.
Indeed, possible zero denominators can be used to find factors of $c$.
\subsection{Contents}
In Section~\ref{sec_algo}, we introduce the general scheme of the factoring algorithm
based on rational maps and discuss
its computational complexity in terms of the complexity of the maps, the number
of parameters and the field degree. In Section~\ref{sec_arg_geo}, the factorization
problem is reduced to the search of rational points of parametrizable algebraic varieties
$\cal V$ having an arbitrarily large number $N_P$ of rational intersection points
with a hypersurface $\cal H$. Provided that the $N_P$ grows exponentially in the
space dimension, the factorization algorithm has polynomial complexity if the
number of parameters and the complexity of evaluating a point in $\cal V$ over a
finite field grow sublinearly and polynomially in the space dimension, respectively.
Thus,
the varieties $\cal V$ and $\cal H$ have to satisfy two requirements.
On one side, their intersection has to contain a large set of rational points.
On the other side, $\cal V$ has to be parametrizable and its points
have to be computed efficiently given
the values of the parameters. The first requirement is fulfilled with a generalization
of the construction given by Eq.~(\ref{poly_eqs_simple}). First, we define an ideal
$I$ generated by products of linear polynomials such that the associated algebraic
set contains $N_P$ rational points. The relevant information on this ideal
is encoded in a satisfiability formula (SAT) in conjunctive normal form (CNF) and a
linear matroid. Then, we define $\cal V$ and $\cal H$ as elements of the ideal.
By construction, ${\cal V}\cap{\cal H}$ contains the $N_P$ rational points.
The ideal $I$ and the polynomials defining the varieties contain some coefficients.
The second requirement is tackled in Sec.~\ref{build_up}. By imposing the
parametrization of $\cal V$, we get a set of polynomial equations for the
coefficients. These equations always admit a solution, provided that the
only constraint on $\cal V$ and $\cal H$ is being an element of $I$.
The task is to find
an ideal $I$ and a set of coefficients such that the computation of points
in $\cal V$ is as efficient as possible, given a number of parameters scaling
sublinearly in the space dimension.
In this general form, the problem of building the varieties $\cal V$ and
$\cal H$ is quite intricate. A good strategy is to start with simple ideals
and search for varieties in a subset of these ideals, so that
the polynomial constraints on the unknown coefficients can be handled with
little efforts. With these restrictions, it is not guaranteed that
the required varieties exist, but we can have hints on how to proceed.
This strategy is employed in Sec.~\ref{sec_quadr_poly}, where we consider an
ideal generated by the polynomials~(\ref{quadr_polys}). The varieties
are defined by linear combinations of these generators with constant
coefficients, that is, $\cal H$ and $\cal V$ are in the zero locus of
$P_0$ and $P_1,\dots,P_{n-1}$, respectively, defined
by Eq.~(\ref{poly_eqs_simple}). The $2^n$ rational points associated
with the ideal are taken distinct in $\mathbb{Q}$. First, we prove
that there is no solution with one parameter ($M=1$), for a dimension
greater than $4$. We give an explicit numerical example of a curve
and hypersurface in dimension $4$. The intersection has $16$ rational
points. We also give a solution with about $n/2$ parameters. Suggestively,
this solution resembles a kind of retro-causal model. Retro-causality
is considered one possible explanation of some strange aspects of
quantum theory, such as non-locality and wave-function collapse after
a measurement. Finally, we close the section by proving that there is
a solution with $2\le M \le (n-1)/3$. This is shown by explicitly
building a variety $\cal V$ with $(n-1)/3$ parameters. Whether
it is possible to drop the number of parameters below this upper
bound is left as an open problem. If $M$ grows
sublinearly in $n$, then there is automatically a factoring
algorithm with polynomial complexity, provided that the coefficients
defining the polynomials $P_k$ are in $\mathbb{Q}$ and can be
computed efficiently over a finite field. The conclusions and
perspectives are drawn in Sec.~\ref{conclusion}.
\section{General scheme and complexity analysis}
\label{sec_algo}
At a low level, the central object of the factoring algorithm under study is a
class $\Theta$ of maps ${\vec\tau}\mapsto {\cal R}(\vec\tau)$
from a set $\vec\tau\equiv(\tau_1,\dots,\tau_M)$ of $M$ parameters over the field $\mathbb{Q}(\alpha)$ to a
number in the same field, where $\cal R$ is a rational function, that is, the algebraic fraction of
two polynomials. Let us write it as
$$
{\cal R}(\vec\tau)\equiv \frac{{\cal N}(\vec\tau)}{{\cal D}(\vec\tau)}.
$$
This function may be indirectly defined by applying consecutively simpler rational functions, as
done in Sec.~\ref{sec_arg_geo}. Note that the computational complexity of evaluating ${\cal R}(\vec\tau)$
can be lower than the complexity of evaluating the numerator ${\cal N}(\vec\tau)$. For this reason we
consider more general rational functions rather than polynomials.
Both $M$ and $\alpha$ are not necessarily fixed in the class $\Theta$.
We denote by $N_P$ the number of zeros of the polynomial $\cal N$ over $\mathbb{Q}(\alpha)$.
The number $N_P$ is supposed to be finite, we will come back to this assumption later in
Sec.~\ref{sec_infinite_points}.
For the sake of simplicity, first we introduce the general scheme of the algorithm over the
rational field. Then, we outline its extension to algebraic number fields.
We mainly consider the case of semiprime input, that is, $c$ is
taken equal to the product of two prime numbers $p$ and $p'$. This case is the most
relevant in cryptography. If the rational points are somehow randomly distributed
when reinterpreted over $\mathbb{Z}_p$, then the polynomial ${\cal N}$ has at least about
$N_P$ distinct zeros over the finite field, provided that $N_P$ is sufficiently smaller than
the size $p^{M}$ of the input space $\Omega$.
We could have additional zeros in the finite field, but we conservatively assume that
$N_P$ is a good estimate for the total number.
For $N_P$ close to $p^M$, two different roots in the rational field may collapse to the same
number in the finite field. We will account for that later in Sec.~\ref{sec_complex}.
Given the class $\Theta$, the factorization procedure has the same general scheme as other methods
using finite fields. Again, the value $m={\cal R}(\tau_1,\tau_2,\dots)$ is computed by pretending
that $c$ is prime and $\mathbb{Z}/c\mathbb{Z}$ is a field. If an algebraic division takes to a
contradiction during the computation of $\cal R$, the divisor is taken as outcome $m$. For the sake of
simplicity, we neglect the zeros of $\cal D$ and consider only the zeros of $\cal N$.
In Section~\ref{sec_arg_geo}, we will see that this
simplification is irrelevant for a complexity analysis. It is clear that the outcome $m$ is zero
in $\mathbb{Z}/c\mathbb{Z}$ with high probability for some divisor $p$ of $c$ if the number of
zeros is about or greater than the number of inputs $p^{M}$. Furthermore, if $p'>p$ and
$N_P$ is sufficiently smaller than $(p')^{M}$, then the outcome $m$
contains the nontrivial factor $p$ of $c$ with high probability. This is guaranteed if
$N_P$ is taken equal to about $c^{M/2}$, which is almost optimal if $p\simeq p'$, as we will see later
in Sec.~\ref{sec_complex}. Thus, we have the following.
\begin{algorithm}
\label{gen_algo0}
Factoring algorithm with input $c=p p'$, $p$ and $p'$ being prime numbers.
\item[(1)] \label{algo0_1}
Choose a map in $\Theta$ with $M$ input parameters and $N_P$ zeros over the rationals
such that $N_P\simeq c^{M/2}$ (see Sec.~\ref{sec_complex} for an optimal choice of $N_P$);
\item[(2)] generate a set of $M$ random numbers $\tau_1,\dots,\tau_M$ over $\mathbb{Z}/c\mathbb{Z}$.
\item[(3)] \label{algo0_3}
compute the value $m={\cal R}(\tau_1,\dots,\tau_M)$ over $\mathbb{Z}/c\mathbb{Z}$
(by pretending that $c$ is prime).
\item[(4)] \label{algo0_4}
compute the greatest common divisor between $m$ and $c$.
\item[(5)] if a nontrivial factor of $c$ is not obtained, repeat from point (2).
\end{algorithm}
The number $M$ of parameters may depend on the map picked
up in $\Theta$. Let $M_{min}(N_P)$ be the minimum of $M$ in $\Theta$ for given $N_P$.
The setting at point~(\ref{algo0_1}) is possible only if $M_{min}$ grows less than linearly in
$\log N_P$, which is condition~(c) enumerated in the introduction. A tighter condition is
necessary if the computational complexity of evaluating the map scales subexponentially, but
not polynomially. This will be discussed with more details in Sec.~\ref{sec_complex}.
If $c$ has more than two prime factors,
$N_P$ must be chosen about equal to about $p^{M}$, where $p$ is an estimate of one
prime factor.
If there is no knowledge about the factors, the algorithm can be executed by trying different
orders of magnitude of $N_P$ from $2$ to $c^{1/2}$. For example, we can
increase the guessed $N_P$ by a factor $2$, so that the overall number of executions grows
polynomially in $\log_2 p$. However, better strategies are available.
A map with a too great $N_P$ ends up to produce zero modulo $p$ for every factor $p$ of $c$
and, thus, the algorithm always generates the trivial factor $c$. Conversely, a too small $N_P$
gives a too small probability of getting a factor. Thus, we can employ a kind of bisection
search. A sketch of the search algorithm is as follows.
\begin{enumerate}
\item set $a_d=1$ and $a_u=c^M$;
\item set $N_P=\sqrt{a_d a_u}$ and choose a map in $\Theta$ with $N_P$ zeros;
\item execute Algorithm~\ref{gen_algo0} from point (2) and break the loop after a certain number
of iterations;
\item if a nontrivial factor is found, return it as outcome;
\item if the algorithm found only the trivial divisor $c$, set $a_u=N_P$, otherwise set $a_d=N_P$;
\item go back to point (2).
\end{enumerate}
This kind of search can
reduce the number of executions of Algorithm~\ref{gen_algo0}. In the following, we will not
discuss these optimizations for multiple prime factors, we will consider mainly semiprime integer
numbers $c=p p'$.
\subsection{Extension to algebraic number fields}
Before outlining how the algorithm can be extended to algebraic number fields,
let us briefly remind what a number field is.
The number field $\mathbb{Q}(\alpha)$ is a rational field extension obtained by adding
an algebraic number $\alpha$ to the field $\mathbb{Q}$.
The number $\alpha$ is solution of some irreducible polynomial $P_I$ of degree $k_0$, which is also
called the degree of $\mathbb{Q}(\alpha)$. The extension
field includes all the elements of the form $\sum_{i=0}^{k_0-1} r_i \alpha^i$, where $r_i$ are rational
numbers. Every power $\alpha^h$ with $h\ge k_0$ can be reduced to that form through the equation
$P_I(\alpha)=0$.
Thus, an element of $\mathbb{Q}(\alpha)$ can be represented as a $k_0$-dimensional vector over
$\mathbb{Q}$. Formally, the extension field is defined as the quotient ring $\mathbb{Q}[X]/P_I$,
the polynomial ring over $\mathbb{Q}$ modulo $P_I$. The quotient ring is also a field as long as
$P_I$ is irreducible.
Reinterpreting the rational function $\cal R$ over a finite field $\text{GF}(p^k)$ means to reinterpret
$r_i$ and the coefficients of $P_I$ as integers modulo a prime number $p$.
Since the polynomial $P_I$ may be reducible over $\mathbb{Z}_p$, the degree $k$ of
the finite field is some value between $1$ and $k_0$ and equal to the degree of one the
irreducible factors of $P_I$. Let $D_1,\dots,D_f$ be the factors of $P_I$. Each $D_i$ is
associated with a finite field $\mathbb{Z}_p[X]/D_i\cong \text{GF}(p^{k_i})$, where $k_i$ is
the degree of $D_i$. Smaller values of $k$ take to a computational advantage, as the
size $p^{k M}$ of the input space $\Omega$ is smaller and the probability, about $N_P/p^{k M}$, of
getting the factor $p$ is higher. For example, the cyclotomic number field with $\alpha=e^{2\pi i/n}$
has a degree equal to $\phi(n)$, where $\phi$ is the Euler totient function, which is asymptotically
lower-bounded by $K n/\log\log n$, for every constant $K<e^{-\gamma}$, $\gamma$ being the Euler
constant. In other words, the highest degree of the polynomial prime factors of $x^n-1$ is equal to
$\phi(n)$. Let $P_I$ be equal to the factor with $e^{2\pi i/n}$ as root.
If $n$ is a divisor of $p-1$ for some prime number $p$, then $P_I$ turns out to have linear
factors over $\mathbb{Z}_p$. Thus, the degree of the number field collapses to $1$
when mapped to a finite field with characteristic $p$. Thus, the bound $k_0$ sets a worst case.
For general number fields, the equality $k=k_0$ is more an exception than a rule, apart from the
case of the rational field, for which $k=k_0=1$. For the sake of simplicity, let us assume
for the moment that $k=k_0$ for one of the two factors of $c$, say $p'$. Algorithm~\ref{gen_algo0}
is modified as follows. The map is chosen at point~(1) of Algorithm~\ref{gen_algo0} such that
$N_P\simeq c^{k_0 M/2}$;
the value $m$ computed at point~(3) is a polynomial over $\mathbb{Z}/c\mathbb{Z}$ of
degree $k_0-1$; the greatest common divisor at point~(4) is computed
between one of the coefficients of the polynomial $m$ and $c$. If the degree $k$ of the finite
field of characteristic $p$ turns out to be smaller than $k_0$, we have to compute the
polynomial greatest common divisor between $m$ and $P_I$ by pretending again that
$\mathbb{Z}/c\mathbb{Z}$ is a field. If $m$ is zero over $\text{GF}(p^{k})$,
then the Euclidean algorithm generates at some point
a residual polynomial with the leading coefficient having $p$ as a factor (generally, all the
coefficients turn out to have $p$ as a factor).
If $k\ne k_0$ for both factors and most of the maps, then the algorithm ends up to generate
the trivial factor $c$, so that we need to decrease $N_P$ until a non-trivial factor is
found.
\subsection{Complexity analysis}
\label{sec_complex}
The computational cost of the algorithm grows linearly with the product between the computational
cost of the map, say ${\bf C}_0({\cal R})$, and the average number of trials, which is
roughly $p^{k_0 M}/N_P$ provided that $N_P\ll p^{k_0 M}$ and $P_I$ is irreducible over
$\mathbb{Z}_p$. The class $\Theta$ may contain many maps with a given number $N_P$ of
zeros over some number field. We can choose the optimal map for each $N_P$, so that we express
$k_0$, $M$ and $\cal R$ as functions of $\log N_P\equiv \xi$. The computational cost
${\bf C}_0({\cal R})$ is written as a function of $\xi$, ${\bf C}_0(\xi)$.
Let us evaluate the computational complexity of the algorithm in terms of the scaling
properties of $k_0(\xi)$, $M(\xi)$ and ${\bf C}_0({\xi})$ as functions of $\xi=\log N_P$.
The complexity ${\bf C}_0(\xi)$ is expected to be a monotonically increasing
function. If the functions $k_0(\xi)$ and $M(\xi)$ were decreasing,
then they would asymptotically tend to a constant, since they are not less than $1$.
Thus, we assume that these two functions are monotonically increasing or constant.
As previously said, the polynomial $\cal N$ has typically about $N_P$ distinct roots
over $\text{GF}(p^{k_0})$, provided that $N_P$ is sufficiently smaller than $p^{k_0 M}$.
If $N_P$ is greater than $p^{k_0 M}$, then almost every
value of $\vec\tau$ is a zero of the polynomial. Assuming that the zeros are somehow
randomly distributed, the probability that a number picked up at random is different from
any zero over $\text{GF}(p^{k_0})$ is equal to $(1-p^{-k_0 M})^{N_P}$.
Thus, the number of roots over $\text{GF}(p^{k_0})$ is expected to be of the order of
$p^{k_0 M} [1-(1-p^{-k_0 M})^{N_P}]$, which is about $N_P$ for $N_P\ll p^{k_0 M}$. Thus,
the average number of trials required for getting a zero is
\be
N_{trials}\equiv \frac{1}{1-(1-p^{-k_0 M})^{N_P}},
\ee
A trial is successful if it gives
zero modulo some nontrivial factor of $c$, thus the number of required trials can
be greater than $N_{trials}$ if some factors are close each other. Let us consider the
worst case with $c=p p'$, where $p$ and $p'$ are two primes with $p'\simeq p$ such
that $(p')^{k_0 M}\simeq p^{k_0 M}$.
Assuming again that the roots are randomly distributed, the probability of a successful
trial is $\text{Pr}_\text{succ}\equiv 2 [1-(1-p^{-k_0 M})^{N_P}](1-p^{-{k_0} M})^{N_P}$.
The probability has a maximum equal to $1/2$ for $\xi$ equal to the value
\be
\xi_0\equiv\log\left[-\frac{\log 2}{\log(1-p^{-k_0 M})}\right].
\ee
Evaluating the Taylor series at $p=\infty$, we have that
\be
\xi_0=k_0 M\log p+\log\log 2-\frac{1}{2 p^{k_0 M}}+O(p^{-2 k_0 M}).
\ee
The first two terms give a very good approximation of $\xi_0$.
At the maximum, the ratio between the number of zeros and the number of states $p^{k_0 M}$
of the sampling space is about $\log 2$. It is worth to note that,
for the same value of $\xi$, the probability of getting an isolated factor
with $p\ll p'$ is again exactly $1/2$. Thus, we have in general
\be
N_P\simeq 0.69 p^{k_0 M} \Rightarrow \text{Pr}_\text{succ}=1/2.
\ee
Since the maximal probability is independent of $k_0$ and $M$, this
value is also maximal if $k_0$ and $M$ are taken as functions of $\xi$.
The maximal value $\xi_0$ is solution of the equation
\be
\label{eq_xi0}
\xi_0=\log\left[-\frac{\log 2}{\log(1-p^{-f(\xi_0)})}\right].
\ee
where $f(\xi)\equiv k_0(\xi) M(\xi)$. If the equation has no positive solution,
then the probability is maximal for $\xi=0$. That is, the optimal map in the
considered class is the one with $N_P=1$. This means that the number of states
$p^{k_0 M}$ of the sampling space grows faster than the number of zeros.
In particular, there is no solution for $\log p$ sufficiently large if
$f(\xi)$ grows at least linearly (keep in mind that $f(\xi)\ge1$). Thus,
the function $f(\xi)$ has to be a sublinear power function, as previously
said.
The computational cost of the algorithm for a given map ${\cal R}(\xi)$ is
\be
{\bf C}(p,\xi)\equiv
\frac{{\bf C}_0(\xi)}{2 [1-(1-p^{-f(\xi)})^{\exp\xi}](1-p^{-f(\xi)})^{\exp\xi }}.
\ee
The optimal map for given $p$ is obtained by minimizing ${\bf C}(p,\xi)$ with
respect to $\xi$. The computational complexity of the algorithm is
\be\label{comp_complexity}
{\bf C}(p)=\min_{\xi>0} {\bf C}(p,\xi)\equiv {\bf C}(p,\xi_m),
\ee
which satisfies the bounds
\begin{equation}
\label{bounds}
{\bf C}_0(\xi_m)\le {\bf C}(p)\le 2{\bf C}_0(\xi_0).
\end{equation}
The upper bound in Eq.~(\ref{bounds}) is the value of ${\bf C}(p,\xi)$ at $\xi=\xi_0$,
whereas the lower bound is the computational complexity of the map at the minimum
$\xi_m$.
It is intuitive that the complexity ${\bf C}_0(\xi)$ must be subexponential in order to
have ${\bf C}(p)$ subexponential in $\log p$. This can be
shown by contradiction. Suppose that the complexity ${\bf C}(p)$ is subexponential
in $\log p$ and ${\bf C}_0(\xi)=\exp(a \xi)$ for some positive $a$. The lower bound
in Eq.~(\ref{bounds}) implies that the optimal $\xi_m$ grows less than $\log p$.
Asymptotically,
\be
\left. \frac{p^{k_0 M}}{N_P}\right|_{\xi=\xi_m}\sim e^{f(\xi_m)\log p-\xi_m}\ge K p^{1/2},
\ee
for some constant $K$.
Thus, the average number of trials grows exponentially in $\log p$, implying
that the computational complexity is exponential, in contradiction
with the premise.
Since $f(\xi)$ and $\log {\bf C}_0(\xi)$ must grow less than linearly, we may assume
that they are concave.
\begin{property}
\label{concave}
The functions $f(\xi)$ and $\log {\bf C}_0(\xi)$ are concave, that is,
\be
\frac{d^2}{d\xi^2}f(\xi)\le 0, \;\;\; \frac{d^2}{d\xi^2}\log {\bf C}_0(\xi)\le 0.
\ee
\end{property}
The lower bound in Eq.~(\ref{bounds}) depends on $\xi_m$, which depends on the
function $C_0(\xi)$. A tighter bound which is also simpler to evaluate can be derived
from Property~\ref{concave} and the inequality
\be
{\bf C}(p,\xi)\ge \frac{1}{2}e^{f(\xi)\log p-\xi}{\bf C}_0(\xi).
\ee
\begin{lemma}
If Property~\ref{concave} holds and ${\bf C}(p)$ is asymptotically sublinear in $p$, then there
is an integer $\bar p$ such that the complexity
${\bf C}(p)$ is bounded from below by $\frac{{\bf C}_0(\xi_0)}{2\log2}$ for $p>\bar p$.
\end{lemma}
{\it Proof}.
The minimum $\xi_m$ is smaller than $\xi_0$, since the function ${\bf C}_0(\xi)$ is monotonically increasing.
Thus, we have
\be
{\bf C}(p)=\min_{\xi\in\{0,\xi_0\}}{\bf C}(p,\xi)\ge \min_{\xi\in\{0,\xi_0\}}
e^{f(\xi)\log p-\xi+\log C_0(\xi)}/2.
\ee
Since the exponential is monotonic and the exponent is concave, the objective function
has a maximum and two local minima at the $\xi=0$ and $\xi=\xi_0$. Keeping in mind that $f(\xi)\ge 1$,
The first local minimum is not less than $p {\bf C}_0(0)/2$. The second minimum is
$e^{f(\xi_0)\log p-\xi_0} {\bf C}_0(\xi_0)/2$,
which is greater than or equal to ${\bf C}_0(\xi_0)/(2 \log2)$. This can be proved by eliminating $p$ through
Eq.~(\ref{eq_xi0}) and minimizing in $\xi_0$. Since ${\bf C}(p)$ is sublinear in $p$,
there is an integer $\bar p$ such that the second minimum is global for $p>\bar p$. $\square$
Summarizing, we have
\begin{equation}
\label{bounds2}
0.72{\bf C}_0(\xi_0)\le {\bf C}(p)\le 2{\bf C}_0(\xi_0)
\end{equation}
for $p$ greater than some integer.
Thus, the complexity analysis of the algorithm is reduced to study the asymptotic behavior
of ${\bf C}_0(\xi_0)$.
The upper bound is asymptotically tight, that is, $\xi=\xi_0$ is asymptotically optimal. Taking
$$
f(\xi)=b \xi^\beta \text{ with } \beta\in[0:1),
$$
the optimal value of $\xi$ is
$$
\xi_0=(b \log p)^\frac{1}{1-\beta}+O(1).
$$
The function $f(\xi)$ cannot be linear, but we can take it very close to a linear function,
\be
f(\xi)=b \frac{\xi}{(\log\xi)^\beta}, \;\;\;\; \gamma>1.
\ee
In this case, the optimal $\xi$ is
$$
\xi_0=e^{(b\log p)^{1/\beta}}+O\left[(\log p)^{1/\beta} \right].
$$
There are three scenarios taking to subexponential or polynomial complexity.
\begin{enumerate}
\item[(a)] The functions ${\bf C}_0(\xi)$ and $f(\xi)$ scale polynomially as $\xi^\alpha$ and
$\xi^\beta$, respectively, with $\beta\in[0:1)$. Then, the computational complexity ${\bf C}(p)$
scales polynomially in $\log p$ as $(\log p)^\frac{\alpha}{1-\beta}$.
\item[(b)] The function ${\bf C}_0(\xi)$ is polynomial and $f(\xi)\sim\xi/(\log\xi)^\beta$ with $\beta>1$.
Then the computational complexity ${\bf C}(p)$ scales subexponentially in $\log p$ as
$\exp\left[b (\log p)^{1/\beta}\right]$.
\item[(c)] The function ${\bf C}_0(\xi)$ and $f(\xi)$ are superpolynomial and polynomial respectively,
with ${\bf C}_0(\xi)\sim\exp\left[b \xi^\alpha\right]$ and $f(\xi)\sim\xi^\beta$. If
$\alpha+\beta<1$, then the complexity ${\bf C}(p)$ is subexponential in $\log p$ and scales
as $\exp\left[b (\log p)^\frac{\alpha}{1-\beta}\right]$.
\end{enumerate}
The algorithm has polynomial complexity in the first scenario. The other cases are subexponential.
This is also implied by the following.
\begin{lemma}
\label{litmus}
The computational complexity ${\bf C}(p)$ is subexponential or polynomial in $\log p$ if
the function ${\bf C}_0(\xi)^{f(\xi)}$ grows less than exponentially, that is, if
$$
\lim_{\xi\rightarrow\infty}\frac{f(\xi)\log {\bf C}_0(\xi)}{\xi}=0.
$$
In particular, the complexity is polynomial if ${\bf C}_0(\xi)$ is polynomial
and $f(\xi)$ scales sublinearly.
\end{lemma}
This lemma can be easily proved directly from Eq.~(\ref{eq_xi0}) and the upper bound in
Eq.~(\ref{bounds}), the former implying the inequality $\xi_0\le f(\xi_0)\log p+\log\log 2$.
Let us prove the first statement.
$$
\lim_{p\rightarrow\infty}\frac{\log{\bf C}(p)}{\log p}\le
\lim_{\xi\rightarrow\infty}\frac{f(\xi)\log 2 {\bf C}_0(\xi)}{\xi-\log\log 2}=
\lim_{\xi\rightarrow\infty}\frac{f(\xi)\log{\bf C}_0(\xi)}{\xi}=0.
$$
Using the lower bound in Eq.~(\ref{bounds2}),
the lemma can be strengthened by adding the inferences in the other directions
(\emph{if} replaced by {\emph{if and only if}).
Summarizing, we have the following.
\begin{claim}
\label{claim1}
The factoring algorithm~\ref{gen_algo0} has subexponential (\emph{polynomial}) complexity if,
for every $\xi=\log N_P>0$ with $N_P$ positive integer, there are rational univariate functions
${\cal R}(\vec\tau)=\frac{{\cal N}(\vec\tau)}{{\cal D}(\vec\tau)}$ of
the parameters $\vec\tau=(\tau_1,\dots,\tau_{M(\xi)})$ over
an algebraic number field $\mathbb{Q}(\alpha)$ of degree $k_0(\xi)$
with polynomials $\cal N$ and $\cal D$ coprime, such that
\begin{enumerate}
\item the number of distinct roots of $\cal N$ in $\mathbb{Q}(\alpha)$ is equal to
about $N_P$. Most of the roots remain distinct when interpreted over finite fields
of order equal to about $N_P^{1/M}$;
\item given any value $\vec\tau$, the computation of ${\cal R}(\vec\tau)$
takes a number ${\bf C}_0(\xi)$ of arithmetic operations growing less than
exponentially (\emph{polynomially}) in $\xi$;
\item the function ${\bf C}_0(\xi)^{k_0(\xi) M(\xi)}$ is subexponential (\emph{the function
$k_0(\xi) M(\xi)$ scales sublinearly}).
\end{enumerate}
\end{claim}
Let us stress that the asymptotic complexity is less than exponential if and only if
${\bf C}_0(\xi)^{f(\xi)}$ is less than exponential. Thus, the latter
condition is a litmus test for a given class of rational functions. However,
the function ${\bf C}_0(\xi)^{f(\xi)}$ does not provide sufficient information on
the asymptotic computational complexity of the factoring algorithm.
The general number-field sieve is the algorithm
with the best asymptotic complexity, which scales as $e^{a (\log p)^{1/3}}$.
Thus, algorithm~\ref{gen_algo0} is asymptotically more efficient than the general number-field
sieve if ${\bf C}_0(\xi)$ and $f(\xi)$ are asymptotically upper-bounded by a subexponential function
$e^{b(\log p)^\alpha}$ and a power function $c \xi^\beta$, respectively, such that $\alpha<(1-\beta)/3$.
In the limit case of $\beta\rightarrow 1$ and polynomial complexity of the map, the
function $f(\xi)$ must be asymptotically upper-bounded by $b \xi/(\log\xi)^3$.
\subsection{Number of rational zeros versus polynomial degree}
Previously we have set upper bounds on the required computational complexity of the
rational function $\cal R={\cal N}/{\cal D}$ in terms of the number of its rational zeros.
For a polynomial (subexponential) complexity of prime factorization, the computational complexity
${\bf C}_0$ of $\cal R$ must scale polynomially (subexponentially) in the logarithm in the number
of rational zeros. Thus, for a univariate rational function, it is clear that ${\bf C}_0$
has to scale polynomially (subexponentially) in the logarithm of the degree $d$ of $\cal N$,
since the number of rational zeros is upper-bounded by the degree
(fundamental theorem of algebra). An extension of this inference to multivariate
functions is more elaborate, as upper bounds on the number of rational zeros
are unknown. However, we are interested more properly to a set of $N_P$ rational zeros
that remain in great part distinct when reinterpreted over a finite field whose order
is greater than about
$N_P^{1/M}$.
Under this restriction, let us show that the number of rational
zeros of a polynomial of degree $d$ and with $M$ variables is upper-bounded by
$K d^{2 M}$ with some constant $K>0$.
This bound allows us to extend the previous inference on ${\bf C}_0$ to the case of
multivariate functions.
Assuming that the $N_P$ rational zeros over $\mathbb{Q}$ are randomly distributed
when reinterpreted over $\text{GF}(q)$, their number over the finite field is
about $q^{M}\left[1-(1-q^{-M})^{N_P}\right]$, as shown previously. Since
an upper bound on the number of zeros $N(q)$ of a smooth hypersurface
over a finite field of order $q$ is known, we can evaluate an upper bound on $N_P$.
Given the inequality~\cite{katz}
\be\label{gen_weil}
N(q)\le
\frac{q^M-1}{q-1} +\left[(d-1)^M-(-1)^M\right]\left(1-d^{-1}\right)q^{(M-1)/2}
\ee
and
\be
\label{bound_ff}
q^M\left[1-(1-q^{-M})^{N_P}\right]\le N(q),
\ee
we get an upper bound on $N_P$ for each $q$.
Requiring that Eq.~(\ref{bound_ff})
is satisfied for every $q>N_P^{1/M}$, we get
\be\label{up_bound}
N_P< K d^\frac{2 M^2}{M+1}< K d^{2 M}
\ee
for some constant $K$ (the same result is obtained by assuming that Eq.~(\ref{bound_ff}) holds
for every $q$).
Note that a slight break of bound~(\ref{up_bound}) with $N_P$ growing as $d_0^{M^a}$
in $M$ for some particular $d=d_0$ and $a>1$ would make the complexity of prime factorization
polynomial, provided the computational complexity of evaluating the function ${\cal R}$
is polynomial in $M$. This latter condition can be actually fulfilled, as shown with an
example later.
Ineq.~(\ref{gen_weil}) holds for smooth irreducible hypersurfaces. However, dropping
these conditions are not expected to affect the bound~(\ref{up_bound}). For example, if
$M=2$, then Ineq.~(\ref{gen_weil}) gives
\be\label{weil_plane}
N_P\le q+1+(d-1)(d-2)\sqrt{q}
\ee
which is the Weil's upper bound~(\ref{weil_bounds}) for
a smooth plane curve, whose geometric genus $g$ is equal to $(d-1)(d-2)/2$.
This inequality holds also for singular curves~\cite{aubry}. Indeed, this comes
from the upper bound~(\ref{aubry_eq}) and the equality $g=(d-1)(d-2)/2-\delta$.
Also reducibility does not affect Ineq~(\ref{up_bound}).
It is simple to find examples of multivariate functions with a number of rational
zero quite close to the bound $K d^{2 M}$.
Trivially, there are polynomials ${\cal N}(\tau_1,\dots,\tau_M)$ of
degree $d$ with a number of rational zeros at least equal to the number of
coefficients minus $1$, that is, equal to
$\bar N_P\equiv M!^{-1}\prod_{k=1}^M(d+k)-1\sim d^M/M!+O(d^{M-1})$.
For $M=1$, this corresponds to take the univariate polynomial
\be\label{univ_poly}
{\cal N}(\tau)=(\tau-x_1)(\tau-x_2)\dots(\tau-x_d).
\ee
A better construction of a multivariate polynomial is a generalization
of the univariate polynomial in Eq.~(\ref{univ_poly}).
Given linear functions $L_{i,s}(\vec\tau)$, the polynomial
$$
\tilde P=\sum_{i=1}^M\prod_{s=1}^d L_{i,s}(\vec\tau)
$$
has generally a number of rational points $N_P$ at least equal to $d^M$, which
is the square root of the upper bound, up to a constant.
For $d<4$ and $M=2$, the number of rational zeros turns
out to be infinite, since the genus is smaller than $2$
(see Sec.~\ref{sec_infinite_points} for the case of
infinite rational points).
A naive computation of $\tilde P(\vec\tau)$ takes $d M^2$
arithmetic operations, that is, its complexity is polynomial in $M$.
This example provides an illustration of the complexity test described
previously in Claim~\ref{claim1}. Expressing
$d$ in terms of $M$ and $\xi=\log N_P$ and assuming that
${\bf C}_0\sim d M^2$, we have that
$$
{\bf C}_0(\xi)=M^2 e^{M^{-1}\xi},
$$
which is subexponential in $\xi$ (provided that $M$ is a growing function of $\xi$),
which
is a necessary condition for a subexponential algorithm. However, the polynomial
does not pass the litmus test, as ${\bf C}_0(\xi)^M$ grows exponentially.
\subsubsection{The case of infinite rational zeros}
\label{sec_infinite_points}
Until now, we have assumed that the rational function has a finite number of rational
zeros over the rationals. However, in the multivariate case, it is possible to
have non-zero functions
with an infinite number of zeros. For example, this is the case of a bivariate
polynomials with genus equal to zero and one, which correspond to parametrizable
curves and elliptic curves, respectively. We can also have functions whose
zero locus contains linear subspaces with positive dimension, which can have
infinite rational points. Since the
probability of having ${\cal R}$ equal to zero modulo $p$ increases with the number
of zeros over the rationals, this would imply that the probability is equal
to $1$ if the number of zeros is infinite. This is not evidently the case. For
example, if ${\cal R}$ is zero for $x_1=0$ and $M>1$, evidently the function
has infinite rational points over $\mathbb{Q}$, but the number of points with
$x_1=0$ over $\mathbb{Z}_p$ is $p^{M-1}$, which is $p$ times less than the
number of points in the space.
Once again,
we are interested more properly to sets of $N_P$ rational zeros over $\mathbb{Q}$
such that a large fraction of them remain distinct over finite fields whose order
is greater than about $N_P^{1/M}$. Under this condition, $N_P$ cannot be infinite
and is constrained by Ineq.~(\ref{up_bound}). If there are linear subspaces with
dimension $h>0$ in the zero locus of $\cal R$, we may fix some of the parameters
$\vec\tau$, so that
these spaces become points. In the next sections, we will build rational functions
having isolated rational points and possible linear subspaces in the zero locus.
If there are subspaces with dimension $k>0$ giving a dominant contribution to
factorization, we can transform them to isolated rational
points by fixing some parameters without changing the asymptotic complexity of
the algorithm.
Isolated rational points are the only relevant points for an asymptotic study of
the complexity of the factoring algorithm, up to a dimension reduction. Thus,
we will consider only them and will not care of the other linear subspaces.
\section{Setting the problem in the framework of algebraic geometry}
\label{sec_arg_geo}
Since the number of zeros $N_P$ is constrained by Ineq.~(\ref{up_bound}),
the complexity of computing the rational function ${\cal R}(\vec\tau)$ must be
subexponential or polynomial in $\log d$ in order to have ${\bf C}_0(\xi)$
subexponential or polynomial. This complexity scaling is attained if, for
example, $\cal R$ is a polynomial with few monomials.
The univariate polynomial $P=\tau^d-1$, which is pivotal in Pollard's $p-1$
algorithm, can be evaluated with a number of arithmetic operations scaling
polynomially in $\log d$. This is achieved by consecutively applying
polynomial maps. For example, if $d=2^g$, then $\tau^d$ is computed through
$g$ applications of the map $x\rightarrow x^2$ by starting with $x=\tau$.
However, polynomials with few terms have generally few zeros over
$\mathbb{Q}$. More general polynomials and rational functions with friendly
computational complexity are obviously available and are obtained by consecutive
applications of simple functions, as done for $\tau^d-1$. This leads us to formulate
the factorization problem in the framework of algebraic geometry as
an intersection problem.
\subsection{Intersection points between a parametrizable variety and
a hypersurface}
\label{sec_intersection}
Considering only the operations defined in the field, the most general rational functions
${\cal R}(\vec\tau)$ with low complexity can be evaluated through the consecutive
application of a small set of simple rational equations of the form
\be\label{ratio_eqs}
\begin{array}{c}
x_{n-M}=\frac{{\cal N}_{n-M}(x_{n-M+1},\dots,x_n)}{{\cal D}_{n-M}(x_{n-M+1},\dots,x_n)} \\
x_{n-M-1}=\frac{{\cal N}_{n-M-1}(x_{n-M},\dots,x_n)}{{\cal D}_{n-M-1}(x_{n-M},\dots,x_n)} \\
\vdots \\
x_1=\frac{{\cal N}_1(x_2,\dots,x_n)}{{\cal D}_1(x_2,\dots,x_n)} \\
{\cal R}=P_0(x_1,\dots,x_n),
\end{array}
\ee
where $P_0$ is a polynomial. If the numerators and denominators ${\cal N}_k$ and ${\cal D}_k$
do not contain too many monomials, then the computation of ${\cal N}_k/{\cal D}_k$ can
be performed efficiently. Assuming that the computational complexity
of these rational functions is polynomial in $n$, the complexity of $\cal R$ is
polynomial in $n$. The computation of ${\cal R}(\vec\tau)$ is performed by setting the
last $M$ components $x_{n-M+1},\dots,x_n$ equal to $\tau_1,\dots,\tau_M$
and generating the sequence $x_{n-M},x_{n-M-1},\dots,x_1,{\cal R}$ according
to Eqs.~(\ref{ratio_eqs}), which ends up with the value of ${\cal R}$.
The procedure may fail to compute the
right value of ${\cal R}(\vec\tau)$ if some denominator
${\cal D}_k(\vec\tau)\equiv {\cal D}_k[x_{k+1}(\vec\tau),\dots,x_n(\vec\tau)]$
turns out to be equal to zero during the computation. However, since
our only purpose is to generate the zero of the field, we can take a zero divisor
as outcome and stop the computation of the sequence. In this way, the
algorithm generates a modified function ${\cal R}'(\vec\tau)$.
Defining $\bar{\cal N}_k(\vec\tau)$ as the numerator of the rational function
${\cal D}_k(\vec\tau)$, we have
\be
{\cal R}'(\vec\tau)=\left\{
\begin{array}{lr}
{\cal R}(\vec\tau) & \;\; \text{if}\;\; \bar{\cal N}_1(\vec\tau)\dots \bar{\cal N}_{n-M}(\vec\tau)\ne0 \\
0 & \text{otherwise}
\end{array}
\right.
\ee
The function ${\cal R}'(\vec\tau)$ has the zeros of
${\cal N}(\vec\tau)\bar{\cal N}_1(\vec\tau)\dots \bar{\cal N}_{n-M}(\vec\tau)$.
For later reference, let us define the following.
\begin{algorithm}
\label{algo1}
Computation of ${\cal R}'(\vec\tau)$.
\begin{enumerate}
\item set $(x_{n-M+1},\dots,x_n)=(\tau_1,\dots,\tau_M)$;
\item set $k=n-M>0$;
\item\label{attr} set
$x_k=\frac{{\cal N}_k(x_{k+1},\dots,x_n)}{{\cal D}_k(x_{k+1},\dots,x_n)}$.
If the division fails, return the denominator as outcome;
\item set $k=k-1$;
\item\label{last_step}
if $k=0$, return $P_0(x_1,\dots,x_n)$, otherwise go back to \ref{attr}.
\end{enumerate}
\end{algorithm}
The zeros of the denominators are not expected to give an effective
contribution on the asymptotic complexity of the factoring algorithm, otherwise it
would be more convenient to reduce the number of steps of the sequence by one and
replace the last function $P_0$ with the denominator ${\cal D}_1$. Let us show that.
Let us denote by $N_1$ the number of rational zeros of ${\cal R}'$ over some
finite field with all the denominators different from zeros. They are the zeros
returned at step~\ref{last_step} of Algorithm~\ref{algo1}.
Let $N_T$ be the total number of zeros. We remind that the factoring complexity
is about $p^{k M}$ times the ratio between the complexity of $\cal R$ and the number
of zeros. If the algorithm is more
effective than the one with one step less and $P_0$ replaced with
${\cal D}_1$, then $\frac{C_T}{N_T}<\frac{C_T-C_1}{N_T-N_1}$.
where $C_1$ and $C_T$ are the number of arithmetic operations of the last
step and of the whole algorithm, respectively. Since $C_1\ge1$, we have
$$
N_1>N_T C_T^{-1}
$$
In order to have a subexponential
factoring algorithm, $C_T$ must scale subexponentially
in $\log N_T$. Thus,
$$
N_1>N_T e^{-\alpha (\log N_T)^\beta}
$$
for some positive $\alpha$ and $0<\beta<1$. That is,
$$
\log N_1>\log N_T -\alpha (\log N_T)^\beta.
$$
If we assume polynomial complexity, we get the tighter bound
$$
\log N_1>\log N_T -\alpha \log\log N_T.
$$
These inequalities imply that the asymptotic complexity of the factoring
algorithm does not change if we discard the zero divisors at Step~\ref{attr}
in Algorithm~\ref{algo1}. Thus, for a complexity analysis, we can consider only
the zeros of ${\cal R}'(\vec\tau)$ with all the denominators ${\cal D}_k(\vec\tau)$
different from zero. This will simplify the subsequent discussion.
Each of these zeros are associated with an $n$-tuple $(x_1,\dots,x_n)$,
generated by Algorithm~\ref{algo1} and solutions of Eqs.~(\ref{ratio_eqs}).
Let us denote by ${\cal Z}_P$ the set of these $n$-tuples.
By definition, an element in ${\cal Z}_P$ is a zero of the set of $M-n+1$ polynomials
\be
\label{poly_affine}
\begin{array}{l}
P_0(x_1,\dots,x_n), \\
P_k(x_1,\dots,x_n)=x_k {\cal D}_k(x_{k+1},\dots,x_n)-{\cal N}_k(x_{k+1},\dots,x_n), \;\;\;
k\in\{1,\dots,n-M\}.
\end{array}
\ee
The last $n-M$ polynomials define an algebraic set of points, say $\cal A$,
having one irreducible branch parametrizable by Eqs.~(\ref{ratio_eqs}).
This branch defines an algebraic variety which we denote by $\cal V$.
The algebraic set may have other irreducible components which do not care about.
The polynomial $P_0$ defines a hypersurface, say $\cal H$. Thus, the
set ${\cal Z}_P$ is contained in the intersection between $\cal V$ and $\cal H$. This
intersection may contain singular points of $\cal V$ with ${\cal D}_k(x_{k+1},\dots,x_n)=0$
for some $k$, which are not relevant for a complexity analysis, as shown previously.
Thus, the factorization problem is reduced to search for non-singular rational points
of a parametrizable variety $\cal V$, whose intersection with a hypersurface
$\cal H$ contains an arbitrarily large number $N_P$ of rational points.
If $N_P$ and ${\bf C}_0$ scale exponentially and polynomially
in the space dimension $n$, respectively, then the complexity of factorization is polynomial,
provided that the number of parameters $M$ scales sublinearly as a power of $n$.
In the limit case of
\be\label{subexp_cond}
M\sim n/(\log n)^\beta
\ee
with $\beta>1$,
the complexity scales subexponentially as $e^{b (\log p)^{1/\beta}}$. Thus, if
$M$ has the scaling property~(\ref{subexp_cond}) with $\beta>3$, then there
is an algorithm outperforming asymptotically the general number field sieve.
A subexponential computational complexity is also obtained if the complexity
of evaluating a point in $\cal V$ is subexponential.
The parametrization of $\cal V$ is a particular case of rational parametrization
of a variety. We call it \emph{Gaussian parametrization} since the triangular form of
the polynomials $P_1,\dots,P_{n-M}$ resembles Gaussian elimination.
Note that this form is invariant under the transformation
\be\label{poly_replacement}
P_k\rightarrow P_k+\sum_{k'=k+1}^{n-M}\omega_{k,k'} P_{k'}.
\ee
The form is also invariant under the variable
transformation
\be\label{invar_trans}
x_k\rightarrow x_k+\sum_{k'=k+1}^{n+1} \eta_{k,k'} x_{k'}
\ee
with $x_{n+1}=1$.
It is interesting to note that if $N_P'$ out of the $N_P$ points in ${\cal Z}_P$
are collinear, then it is possible to build another variety with Gaussian
parametrization and a hypersurface over a $(n-1)$-dimensional subspace such that
their intersection contains the $N_P'$ points. For later reference, let us
state the following.
\begin{lemma}
\label{lemma_dim_red}
Let ${\cal Z}_P$ be the set of common zeros of the polynomials~(\ref{poly_affine})
with ${\cal D}_k(x_{k+1},\dots,x_n)\ne 0$ over ${\cal Z}_P$ for $k\in\{1,\dots,n-M\}$.
If $N_P'$ points in ${\cal Z}_P$ are solutions of the linear equation $L(x_1,\dots,x_n)=0$,
then there is a variety with Gaussian parametrization and a hypersurface over
the $(n-1)$-dimensional subspace defined by $L(x_1,\dots,x_n)=0$ such that their
intersection contains the $N_P'$ points.
\end{lemma}
{\it Proof.}
Given the linear function $L(x_1,\dots,x_n)\equiv l_{n+1}+\sum_{k=1}^n l_k x_k$,
let us first consider the case with $l_k=0$ for $k\in\{1,\dots,n-M\}$. Using
the constraint $L=0$, we can set one of the $M$ variables $x_{n-M+1},\dots,x_n$
as a linear function of the remaining $M-1$ variables. Thus, we get a new
set of polynomials retaining the original triangular form. The new parametrizable
variety, say ${\cal V}'$, has $M-1$ parameters. The intersection of ${\cal V}'$ with
the new hypersurface over the $(n-1)$-dimensional space contains the $N_P'$ points.
Let us now consider the case with $l_k=0$ for $k\in\{1,\dots,\bar k\}$, where
$\bar k$ is some integer between $0$ and $n-M-1$, such that $l_{\bar k+1}\ne 0$.
We can use the constraint $L=0$ to set $x_{\bar k+1}$ as a linear function
of $x_{\bar k+2},\dots,x_n$.
We discard the polynomial
$P_{\bar k+1}$ and eliminate the $(\bar k+1)$-th variable from the remaining polynomials.
We get $n-1$ polynomials retaining the original triangular form in $n-1$ variables
$x_1,\dots,x_{\bar k},x_{{\bar k}+2},\dots,x_n$. The intersection between the
new parametrizable variety and the new hypersurface contain the $N_P'$ points
$\square$.
\newline
This simple lemma will turn out to be a useful tool in different parts of the
paper.
In Section~\ref{sec_common_zeros}, we show how to build a set of polynomials $P_k$
with a given number $N_P$ of common rational zeros by using some tools of algebraic geometry
described in Appendix~\ref{alg_geom}.
In Sec.~\ref{build_up}, we close the circle by
imposing the form~(\ref{poly_affine}) for the polynomials $P_k$ with
the constraint that ${\cal D}_k(x_{k+1},\dots,x_n)\ne 0$ for $k\in\{1,\dots,n-M\}$ over
the set of $N_P$ points.
\subsection{Sets of polynomials with a given number of zeros over a number field}
\label{sec_common_zeros}
In this subsection, we build polynomials with given $N_P$ common rational zeros
as elements of an ideal $I$ generated by products of linear functions. This construction
is the most general.
The relevant information on the ideal $I$ is summarized by a satisfiability
formula in conjunctive normal form without negations and a linear matroid.
The formula and the matroid uniquely determine the number $N_P$ of rational common zeros
of the ideal. Incidentally, we also show that the information can be encoded in a more
general formula with negations by a suitable choice of the matroid.
Every finite set of points in an $n$-dimensional space is an algebraic set, that
is, they are all the zeros of some set of polynomials. More generally, the union
of every finite set of linear subspaces is an algebraic set.
In the following, we
will denote linear polynomials by a symbol with a hat; namely, $\hat a$ is meant
as $a_{n+1}+\sum_{i=1}^n a_i x_i$. Let us denote by $\vec x$ the $(n+1)$-dimensional
vector $(x_1,\dots,x_n,x_{n+1})$, where $x_{n+1}$ is an extra-component that is
set equal to $1$. A linear polynomial $\hat a$ is written in the form
$\vec a\cdot\vec x$.
Let $V_1,\dots, V_{L}$ be a set of linear subspaces and $I_1,\dots,I_L$ their
associated radical ideals. The codimension of the $k$-th subspace is denoted by $n_k$.
The minimal set of generators of the $k$-th ideal contains $n_k$ independent
linear polynomials, say $\hat a_{k,1},\dots,\hat a_{k,n_k}$, so that
\be
\vec x\in V_k \Leftrightarrow\vec a_{k,i}\cdot\vec x=0\;\; \forall i\in\{1,\dots,n_k\}.
\ee
If the codimension $n_k$ is equal to $n$, then $V_k$ contains one
point. We are mainly interested to these points, whose number is
taken equal to $N_P$. The contribution of higher dimensional subspaces to
the asymptotic complexity of the factoring algorithm is irrelevant up to a dimension
reduction (see also Sec.~\ref{sec_infinite_points} and the remark in the
end of the section). Since only isolated points are relevant, we could just consider
ideals whose zero loci contain only isolated points. However,
we allow for the possible presence of subspaces with positive dimension
since they may simplify the set of the generators or the form of the polynomials
$P_k$ that eventually we want to build.
Let $\cal Z$ be the union of the subspaces $V_k$. The product
$I_1\cdot I_2\cdot\dots I_L\equiv \tilde I$ is associated with $\cal Z$, that is,
${\cal Z}={\bf V}(\tilde I)$. A set of generators of the ideal $\tilde I$ is
\be
\prod_{k=1}^{L}\hat a_{k,i_k}\equiv G_{i_1,\dots,i_L}(\vec x) \;\;\;
i_r\in\{1,\dots,n_r\}, r\in\{1,\dots,L\}.
\ee
Thus, we have that
\be
\vec x\in {\cal Z} \Leftrightarrow G_{i_1,\dots,i_L}(\vec x)=0
\;\;\;\; i_r\in\{1,\dots,n_r\}, r\in\{1,\dots,L\}.
\ee
Polynomials in the ideal $\tilde I$ are zero in the set $\cal Z$.
This construction is not the most general, as $\tilde I$ is not radical.
Thus, there are polynomials that are not in $\tilde I$, but
their zero locus contains $\cal Z$.
Furthermore, the number of generators and the number of their
factors grow polynomially and linearly in $N_P$, respectively.
This makes it hard to build polynomials in $\tilde I$ whose
complexity is polynomial in $\log N_P$.
The radicalization of the ideal and the assumption of
special arrangements of the subspaces in $\cal Z$
can reduce drastically both the degree of the generators and
their number. For example, let us assume that $V_1$ and $V_2$
are two isolated points in the $n$-dimensional space and, thus,
$n_1=n_2=n$. The overall number of generators $a_{1,1},\dots,a_{1,n}$ and
$a_{2,1},\dots,a_{2,n}$ is equal to $2 n$. Thus, there are
$n-1$ linear constraints among the generators. Using linear
transformations, we can write these constraints as
$$
\hat a_{1,i}=\hat a_{2,i}\equiv\hat a_i \;\;\;\forall i\in\{2,\dots,n\}.
$$
Every generator $G_{i,i,i_3,\dots,i_L}$ with $i\ne 1$ is equal
to $\bar G_{i,i_3,\dots,i_L}=a_i^2 \prod_{k=3}^L \hat a_{k,i_k}$. The polynomial
$\bar G'_{i,i_3,\dots,i_L}\equiv a_i \prod_{k=3}^L \hat a_{k,i_k}$ is not an element of the
ideal $\tilde I$, but it is an element of its radical. Indeed,
it is zero in the algebraic set $\cal Z$. Thus, we extend
the ideal by adding these new elements. This extension allows
us to eliminate all the generators $G_{i_1,i_2,\dots,i_L}$ with
$i_1=1$ or $i_2=1$, since they are generated by $\bar G'_{i,i_3,\dots,i_L}$.
Thus, the new ideal has the generators,
\begin{equation}
\left.
\begin{array}{l}
\hat a_{1,1}\hat a_{2,1}\prod_{k=3}^L a_{k,i_k} \\
\hat a_i\prod_{k=3}^L a_{k,i_k},\;\;\; i\in\{2,\dots,n\}
\end{array}\right\}
i_r\in\{1,\dots,n_r\}, r\in\{3,\dots,L\}.
\end{equation}
Initially, we had $n^2\prod_{k=3}^{L}n_k$ generators. Now, their
number is $n\prod_{k=3}^{L}n_k$. A large fraction of them has the
degree reduced by one. We can proceed with the other points and
further reduce
both degrees and number of generators. Evidently,
this procedure cannot take to a drastic simplification of the
generators if the points in $\cal Z$ are in general position, since
the generators must contain the information about these positions.
A simplification is possible if the points have special arrangements
taking to contraction of a large number of factors in the generators.
Namely, coplanarity of points is the key feature that can take to a
drastic simplification of the generators.
In a $n$-dimensional space, there are at most $n$ coplanar points
in general position. Let us consider algebraic sets containing
larger groups of coplanar points. For example, let us assume that
the first $m$ sets $V_1,\dots,V_m$ are distinct coplanar points, with
$m\gg n$. Then, there is a vector $\vec a_1$ such that $\vec a_1\cdot\vec x=0$
for every $\vec x$ in the union of the first $m$ linear spaces. It is convenient
to choose the linear polynomial $\vec a_1\cdot\vec x$ as common generator of
the first $m$ ideals $I_1,\dots,I_m$. Let us set $\hat a_{k,1}=\hat a_1$
for $k\in\{1,\dots,m\}$. Every generator $G_{i_1,\dots,i_L}$ with $i_k=1$
for some $k\in\{1,\dots,m\}$ is contracted to a generator of the form
$\hat a\prod_{k=m+1}^L \hat a_{k,i_k}$. If there are other groups of
coplanar points, we can perform other contractions.
\begin{definition}
Given an integer ${\bar n}>n$, we define $\Gamma_s$ with $s\in\{1,\dots,{\bar n}\}$ as
a set of $s$-tuples $(i_1,\dots,i_s)\in\{1,\dots,{\bar n}\}^s$
with $i_k < i_{k'}$ for $k < k'$. That is,
\begin{equation}
\forall s\in\{1,\dots,{\bar n}\},\;\;\;
\Gamma_s\subseteq \{(i_1,\dots,i_s)\in\{1,\dots,{\bar n}\}^s| i_k < i_{k'},\forall k, k' \text{ s.t. }
k<k' \}.
\end{equation}
\end{definition}
The final result of the inclusion of elements of the radical ideal is another
ideal, say $I$, with generators of the form
\be\label{generators}
\begin{array}{l}
\hat a_{i_1} \;\;\;\forall i_1\in\Gamma_1 \\
\hat a_{i_1}\hat a_{i_2} \;\;\;\forall (i_1,i_2)\in\Gamma_2 \\
\dots \\
\hat a_{i_1}\hat a_{i_2}\dots\hat a_{i_{{\bar n}}} \;\;\;\forall (i_1,i_2,\dots,i_{{\bar n}})\in\Gamma_{{\bar n}},
\end{array}
\ee
where $\{\hat a_1,\dots,\hat a_{{\bar n}}\}\equiv\Phi$ is a set of ${\bar n}$ linear polynomials.
Polynomials in this form generate the most general ideals whose zero loci contain a given finite set
of points. This is formalized by the following.
\begin{lemma}
Every radical ideal associated with a finite set ${\cal Z}_P$ of points is generated by a set of
polynomials of form~(\ref{generators}) for some $\bar n$.
\end{lemma}
{\it Proof}. This can be shown with a naive construction. Given a set $\bar S$ of $N_P$ points
associated with the ideals $I_1,\dots,I_{N_P}$, the product
$I_1\cdot \dots I_{N_P}$ is an ideal associated with the set $\bar S$, which can be radicalized by
adding a certain number of univariate square-free polynomials as generators~\cite{seidenberg}.
The resulting ideal is generated by a set of polynomials of form~(\ref{generators}). $\square$
\newline
With the construction used in the proof, $\bar n$ ends up to be equal to the number of points
in ${\cal Z}_P$, which is not optimal for our purposes. We are interested to keep $\bar n$
sufficiently small, possibly scaling polynomially in the dimension $n$.
This is possible only if the points in the zero locus
have a `high degree' of collinearity.
Thus, a bound on $\bar n$ sets a restriction on ${\cal Z}$.
The minimal information on $\Phi$ that is relevant for determining the number $N_P$ of points in
$\cal Z$ is encoded in a linear \emph{matroid}, of which $\Phi$ is one linear representation.
Thus, the sets $\Gamma_s$ and the matroid determine $N_P$.
Note that the last set $\Gamma_{{\bar n}}$ has at most one element. The linear generators
can be eliminated by reducing the dimension of the affine space, see Lemma~\ref{lemma_dim_red}.
Thus, we can set $\Gamma_1=\emptyset$. Every subset $\Phi_{sub}$ of $\Phi$ is associated with a
linear space $V_{sub}$ whose points are the common zeros of the linear functions in $\Phi_{sub}$.
That is, $V_{sub}={\bf V}(I_{sub})$, where $I_{sub}$ is the ideal generated by $\Phi_{sub}$.
Let us denote briefly by ${\bf V}(\Phi_{sub})$ the linear space ${\bf V}(I_{sub})$.
This mapping from subsets of $\Phi$ to subspaces is not generally injective. Let
$\hat a'\in\Phi\setminus \Phi_{sub}$ be a linear superposition of the functions in
$\Phi_{sub}$, then $\Phi_{sub}$ and $\Phi_{sub}\cup\{\hat a'\}$ represent the same linear space.
An injective mapping is obtained by considering only the maximal subset associated with a linear
subspace. These maximal subsets are called \emph{flats} in matroid theory.
\begin{definition}
\emph{Flats} of the linear matroid $\Phi$ are defined as subsets
$\Phi_{sub}\subseteq \Phi$ such that
no function in $\Phi\setminus \Phi_{sub}$ is
linearly dependent on the functions in $\Phi_{sub}$.
\end{definition}
Let us also define the closure of a subset of $\Phi$.
\begin{definition}
Given a subset $\Phi_{sub}$ of $\Phi$ associated with subspace $V$, its closure
$\text{cl}(\Phi_{sub})$ is the flat associated with $V$.
\end{definition}
The number of independent functions in a flat is called \emph{rank} of the flat.
The whole set $\Phi$ is a flat of rank $n+1$, which is associated with an empty space. Flats
of rank $n$ define points of the $n$-dimensional affine space (with $x_{n+1}=1$).
More generally,
flats of rank $k$ define linear spaces of dimension $n-k$.
The dimension of a flat $\Phi_{flat}$ is meant as the dimension of the space
${\bf V}(\Phi_{flat})$.
The structure of the generators~(\ref{generators}) resembles
a Boolean satisfiability problem (SAT) in conjunctive normal form without negations.
Let us interpret $\hat a_i$ as a logical variable $a_i$ which is $\true$ or $\false$ if
the function is zero or different from zero, respectively. Every subset
$\Phi_{sub}\subseteq\Phi$ is identified with a string $(a_1,\dots,a_{{\bar n}})$ such
that $a_i=\true$ if and only if $\hat a_i\in\Phi_{sub}$.
The SAT formula associated with the generators~(\ref{generators}) is
\be\label{SAT}
\bigwedge\limits_{k=2}^{{\bar n}} \left( \bigvee\limits_{i\in \Gamma_k } a_i \right).
\ee
Given a flat $\Phi_{flat}$, the linear space ${\bf V}(\Phi_{flat})$
is a subset of $\cal Z$ if and only if $\Phi_{flat}$
is solution of the SAT formula. If a set $\Phi_{sub}\subseteq\Phi$ is solution
of the SAT formula, then the flat $\text{cl}(\Phi_{sub})$ is also solution
of the formula. Thus, satisfiability implies that there are flats as solutions
of the formula. This does not mean that
satisfiability implies that
$\cal Z$ is non-empty. Indeed, if the dimension of ${\bf V}(\Phi_{sub})$ is
negative for every solution $\Phi_{sub}$, then the set $\Phi$ is the only
flat solution of the formula.
We are interested to the isolated points in $\cal Z$.
A point $p\in{\cal Z}$
is \emph{isolated} if there is a SAT solution $\Phi_{flat}$ with zero dimension
such that $p\in{\bf V}(\Phi_{flat})$ and no flat $\Phi_{flat}'\subset\Phi_{flat}$ is
solution of the Boolean formula. We denote by ${\cal Z}_P$ the subset in $\cal Z$
containing the isolated points. Since the number $N_P$ of isolated points is completely
determined by the SAT formula and the linear matroid, the information on these
two latter objects is the most relevant. Given them, the linear functions $\hat a_i$
have some free coefficients.
{\bf Remark}. In general, we do not rule out sets $\cal Z$
containing subspaces with positive dimension, however these subspaces are irrelevant
for the complexity analysis of the factoring algorithm. For example, if subspaces of
dimension $d_s<M$
give a dominant contribution to factorization, then we can generally eliminate $d_s$ out of
the $M$ parameters by setting them equal to constants,
so that the subspaces are reduced to points. Furthermore,
subspaces with dimension greater than $M-1$ are not in the parametrizable variety $\cal V$,
whose dimension is $M$. Neither the overall contribution of all the subspaces with positive
dimension can provide a significant change in the asymptotic complexity up to
parameter deletions. Thus, only isolated points of
$\cal Z$ are counted without loss of generality.
\subsection{Boolean satisfiability and algebraic-set membership}
As we said previously,
the Boolean formula does not encode all the information about the number of isolated
points in ${\cal Z}$, which also depends on the independence relations among the vectors
$\vec a_i$, specified by the matroid. A better link between the SAT problem
and the membership to ${\cal Z}$ can be obtained if we consider sets $\Phi$
with cardinality equal to $2 n$ and interpret half of the
functions in $\Phi$ as negations of the others. Let us
denote by $\hat a_{0,1},\dots,\hat a_{0,n}$ and $\hat a_{1,1},\dots,\hat a_{1,n}$
the $2n$ linear functions of $\Phi$. For general functions, we have the following.
\begin{property}
\label{indep_functions}
The set of vectors $\{\vec a_{s_1,1},\dots,\vec a_{s_n,n},\vec a_{1-s_k,k}\}$
is independent for every string $\vec s=(s_1,\dots,s_n)\in\{0,1\}^n$ and every
$k\in\{1,\dots,n\}$.
\end{property}
This property generally holds if the functions are picked up at random.
Let us assume that $\Phi$ satisfies Property~\ref{indep_functions}. This
implies that $\{\hat a_{s_1,1},\dots,\hat a_{s_n,n}\}$ are linearly independent
and equal to zero at one point $\vec x_{\vec s}$. Furthermore,
Property~\ref{indep_functions} also implies that different strings $\vec s$ are
associated with different points $\vec x_{\vec s}$.
\begin{lemma}
\label{lemma_indep}
Let $\{\vec a_{0,1},\dots,\vec a_{0,n},\vec a_{1,1},\dots,\vec a_{1,n}\}$ be a set
of $2n$ vectors satisfying Property~\ref{indep_functions}. Let $\vec x_{\vec s}$
be the solution of the equations $\hat a_{s_1,1}=\dots=\hat a_{s_n,n}=0$.
If $\vec s\ne\vec r$, then $\vec x_{\vec s}\ne\vec x_{\vec r}$.
\end{lemma}
{\it Proof}. Let us assume that $\vec x_{\vec s}=\vec x_{\vec r}$ with $\vec s\ne\vec r$.
There is
an integer $k\in\{1,\dots,n\}$ such that $s_k\ne r_k$. Thus, the set of
vectors $\{\vec a_{s_1,1},\dots,\vec a_{s_n,n},\vec a_{1-s_k,k}\}$
are orthogonal to $\vec x_{\vec s}$. Since the dimension of the vector space
is $n+1$, the set of $n+1$ vectors are linearly dependent, in contradiction
with the hypotheses. $\square$
Now, let us define the set ${\cal Z}$ as the zero locus of the ideal generators
\be\label{generators2}
\begin{array}{l}
\hat a_{0,i} \hat a_{1,i} \;\;\;\forall i\in\{1,\dots,n\} \\
\hat a_{s_1,i_1}\hat a_{s_2,i_2} \;\;\;\forall (s_1,i_1;s_2,i_2)\in\Gamma_2 \\
\dots \\
\hat a_{s_1,i_1}\hat a_{s_2,i_2}\dots\hat a_{s_n,i_n} \;\;\;\forall (s_1,i_1;s_2,i_2,\dots,s_n,i_n)\in\Gamma_n.
\end{array}
\ee
The first $n$ generators provide an interpretation of $\hat a_{1,i}$ as negation
of $\hat a_{0,i}$, as consequence of Property~\ref{indep_functions}. The $i$-th generator
implies that $(a_{0,i},a_{1,i})$ is equal to $(\true,\false)$, $(\false,\true)$
or $(\true,\true)$. However, the last case is forbidden by Property~\ref{indep_functions}.
Assume that $(a_{0,i},a_{1,i})$ is equal to $(\true,\true)$ for some $i$.
Then, there would be $n+1$ functions $\hat a_{s_1,1},\dots,\hat a_{s_n,n},\hat a_{1-s_i,i}$
equal to zero, which is impossible since they are independent. Thus, the algebraic
set defined by the first $n$ generators contains $2^n$ distinct points, as implied
by Lemma~\ref{lemma_indep}, which are associated with all
the possible states taken by the logical variables. The remaining
generators set further constraints on these variables and define a Boolean formula
in conjunctive normal form. With this construction there is a one-to-one correspondence
between the points of the algebraic set ${\cal Z}$ and the solutions of the Boolean
formula.
There is a generalization of the generators~(\ref{generators2}) that allows us
to weaken Property~\ref{indep_functions} while retaining the one-to-one correspondence.
Let $R_1,\dots,R_m$ be $m$ disjoint non-empty sets such that
$\cup_{k=1}^m R_k=\{1,\dots,n\}$.
\begin{property}
\label{indep_functions2}
The set of vectors $\cup_{k=1}^m\{\vec a_{s_k,i}|i\in R_k\}\equiv A_{\vec s}$ is
independent for every $\vec s=(s_1,\dots,s_m)\in\{0,1\}^m$. Furthermore,
every vector $\vec a_{s,i}\notin A_{\vec s}$ is not in $\text{span}(A_{\vec s})$,
with $s\in\{0,1\}$ and $i\in\{1,\dots,n\}$.
\end{property}
\begin{lemma}
\label{lemma_indep2}
Let $\{\vec a_{0,1},\dots,\vec a_{0,n},\vec a_{1,1},\dots,\vec a_{1,n}\}$ be a set
of $2n$ vectors satisfying Property~\ref{indep_functions2}. Let $\vec x_{\vec s}$
be the solution of the equations
$$
\hat a_{s_k,i}=0 \;\;\; i\in R_k, k\in\{1,\dots,m\}
$$
for every $\vec s\in\{0,1\}^m$.
If $\vec s\ne\vec r$, then $\vec x_{\vec s}\ne\vec x_{\vec r}$.
\end{lemma}
The generators~(\ref{generators2}) are generalized by replacing the first line
with
\be\label{block_gen}
\hat a_{0,i} \hat a_{1,j}, \;\;\; (i,j)\in \cup_{k=1}^m (R_k \times R_k).
\ee
Provided that Property~\ref{indep_functions2} holds there is a one-to-one correspondence
between the points in the algebraic set and the solutions of a SAT formula built according
to the following interpretation.
Each set of functions $\{\hat a_{0,i}|i\in R_k\}\equiv a_k$ is interpreted as a Boolean variable,
which is true if the functions in there are equal to zero. The set $\{\hat a_{1,i}|i\in R_k\}$
is interpreted as negation of $\{\hat a_{0,i}|i\in R_k\}$. The SAT formula is built in obvious
way from the set of generators. For example, the generator $\hat a_{0,i}\hat a_{0,j}$ with
$i\in R_{1}$ and $j\in R_{2}$ induces the clause $a_{1} a_{2}$. Different generators
can induce the same clause. Since the total number of solutions depends only on the
SAT formula, it is convenient to take the maximal set of generators compatible with
a given formula. That is, if $a_{1} a_{2}$ is a clause, then $\hat a_{0,i}\hat a_{0,j}$
is taken as a generator for every $i\in R_1$ and $j\in R_2$.
\subsection{$3$SAT-like generators}
\label{sec_3SAT}
SAT problems have clauses with an arbitrarily large number of literals. Special cases
are $2$SAT and $3$SAT, in which clauses have at most $2$ or $3$ literals. It is known
that every SAT problem can be converted to a $3$SAT one by increasing the number of
variables and replacing a clause with a certain number of smaller clauses containing
the new variables. For example, the clause $a\lor b\lor c\lor d$ can be replaced by
$a\lor b\lor x$ and $c\lor d\lor (\lnot x)$. An assignment satisfies the first clause
if and only if the other two clauses are satisfied for some $x$. An identical reduction
can be performed also on the generators~(\ref{generators}). For example, a generator in $I$
of the form $\hat a_1\hat a_2\hat a_3\hat a_4\equiv G_0$ can be replaced by
$\hat a_1\hat a_2 y\equiv G_1$, $\hat a_1\hat a_2(1-y)\equiv G_2$
and $y(1-y)\equiv G_3$, where $y$ is an additional variable. Also in this
case, $G_0$ is equal to zero if and only if $G_1$ and $G_2$
are zero for $y=0,1$. Furthermore, the new extended ideal contains the
old one. Indeed, we have that $G_0=\hat a_3\hat a_4 G_1+\hat a_1\hat a_2 G_2$.
Note that all the polynomials in the ideal $I$ are independent of the
additional variables used in the reduction. Thus, if we build the
polynomials~(\ref{poly_affine}) by using $3$SAT-like generators, then
all these polynomials may be independent of some variables.
Thus, we can consider generators in a $3$SAT form,
\be\label{generators_3SAT}
\begin{array}{l}
\hat a_{i_1}\hat a_{i_2} \;\;\;\forall (i_1,i_2)\in\Gamma_2 \\
\hat a_{i_1}\hat a_{i_2}\hat a_{i_{3}} \;\;\;\forall (i_1,i_2,i_3)\in\Gamma_{3}.
\end{array}
\ee
There is no loss of generality, provided that all the polynomials $P_k$ are
possibly independent from $n_I$ variables.
The number of isolated points satisfies the inequality
\be
\label{bound_N0_n}
N_P\le 3^n.
\ee
The actual number can be considerably smaller, depending on the
matroid and the number of clauses defining the Boolean formula.
The bound is attained if $n_c=3 n$, the generators have the
form $a_i b_i c_i$ with $i\in\{1,\dots,n\}$, and the independent
sets of the matroid contain $n+1$ elements.
If there are only clauses with $2$ literals, then the bound is
\be
N_P\le 2^n,
\ee
which is strict if the generators have the form $a_i b_i$ with
$i\in\{1,\cdots,n\}$.
A consequence of these constraints is that the number $M$ of parameters
must scale sublinearly in $n$,
\be
\label{bound_M_n}
M\le K n^\beta, \;\;\;\; 0\le \beta<1
\ee
for some $K>0$.
\section{Building up the parametrizable variety and the hyperplane}
\label{build_up}
In this section, we put together the tools introduced previously to tackle our
problem of building the rational function ${\cal R}$ with the desired properties
of being computationally simple and having a sufficiently large set of zeros.
This problem has being reduced to the search of computationally simple
polynomials $P_k$ of the form~(\ref{poly_affine}) with a number of
common rational zeros growing sufficiently fast with the space dimension.
To build these polynomials, we first choose a set of generators of the
form~(\ref{generators}) such that the associated algebraic set $\cal Z$ has a set
of $N_P$ points. Then, we write the polynomials $P_k$ as elements of the ideal
associated with $\cal Z$. Finally, we impose that the polynomials $P_k$ have the
form of Eqs.~(\ref{poly_affine}).
\begin{procedure}
\label{procedure}
Building up of a parametrizable variety $\cal V$ with $M$ parameters
and $N_P$ intersection points.
\begin{enumerate}
\item Take a set of ${\bar n}$ unknown non-homogeneous linear functions in $n$ variables
with ${\bar n}>n$, say $\hat a_1,\dots,\hat a_{{\bar n}}$. Additionally,
specify which set of vectors are linearly independent. In other words,
a linear matroid with ${\bar n}$ elements is defined.
\item Choose and ideal $I$ with generators of the form~(\ref{generators_3SAT}) such that
the associated algebraic set $\cal Z$ contains a subset ${\cal Z}_P$ of
$N_P$ isolated points over some given number field.
\item Set the polynomials $P_s$ equal to elements of the ideal $I$
with $s\in\{0,\dots,n-M\}$. That is,
\begin{equation}
\label{poly_in_ideal}
P_s(\vec x)=\sum_{(i,j)\in \Gamma_2} C_{s,i,j}(\vec x) \hat a_i \hat a_j+
\sum_{(i,j,k)\in \Gamma_3} D_{s,i,j,k}(\vec x) \hat a_i \hat a_j\hat a_k,
\end{equation}
The polynomials $P_s$ with $s\in\{1,\dots,n-M\}$ define an algebraic set $\cal A$.
The polynomial $P_0$ defines a hyperplane $\cal H$. The number of parameters $M$
and the polynomial coefficients $C_{s,i,j}(\vec x)$ and $D_{s,i,j,k}(\vec x)$ are also unknown.
\item Search for values of the coefficients such that there is a parametrizable branch $\cal V$
in $\cal A$
with a number of parameters as small as possible. All the polynomials $P_s$ with
$s\in\{0,\dots,n-M\}$ are possibly independent of some subset of variables (see Sec.~\ref{sec_3SAT}).
The polynomials ${\cal D}_k$, as defined in Eq.~(\ref{poly_affine}) must be different
from zero in the set ${\cal Z}_P$.
\end{enumerate}
\end{procedure}
More explicitly, the last step leads us to the following.
\begin{problem}
\label{problem1}
Given the sets $\Gamma_2$ and $\Gamma_3$, and polynomials of the
form~(\ref{poly_in_ideal}), find linear functions
$\hat a_1,\dots,\hat a_{{\bar n}}$ and coefficients $C_{s,i,j}(\vec x)$,
$D_{s,i,j,k}(\vec x)$ such that
\be
\label{gauss_constrs}
\begin{array}{l}
\frac{\partial P_s}{\partial x_k}=0, \;\;\;\; 1\le k < s\le n-M,
\vspace{2mm}
\\
\frac{\partial^2 P_s}{\partial x_s^2}=0, \;\;\;\; 1\le s\le n-M,
\vspace{2mm}
\\
\vec x\in{\cal Z}_P \Rightarrow \frac{\partial P_s}{\partial x_s}\equiv
{\cal D}_s(x_{s+1},\dots,x_n)\ne 0,
\end{array}
\ee
under the constraint that $(\hat a_1,\dots,\hat a_{{\bar n}})$ is the representation
of a given matroid.
\end{problem}
{\bf Remark}.
If the algebraic set associated with the ideal $I$ is zero-dimensional,
this problem has always a solution for any $M$, since a rational univariate representation
always exists (see introduction). Essentially, the task is to find ideals such
that there is a solution with the coefficients $C_{s,i,j}(\vec x)$ and
$D_{s,i,j,k}(\vec x)$ as simple as possible, so that their computation is
efficient, given $\vec x$.
Let us remind that the constraints~(\ref{gauss_constrs}) are
invariant under transformations~(\ref{poly_replacement},\ref{invar_trans}).
All the polynomials are possibly independent of a subset of $n_I$
variables, say $\{x_{n-n_I+1},\dots,x_{n}\}$,
\begin{equation}
\frac{\partial P_s}{\partial x_k}=0, \;\;\;\;
\left\{
\begin{array}{l}
s\in\{0,\dots,n-M\} \\
k\in\{n-n_I+1,\dots,n\}
\end{array}\right.
\end{equation}
These $n_I$ variables can be set equal to constants,
so that the actual number of significant parameters is $M-n_I$.
The input of Problem~\ref{problem1} is given by a 3SAT formula
of form~(\ref{generators_3SAT}) and a linear matroid.
\begin{definition}
A 3SAT formula of form~(\ref{generators_3SAT}) and a linear matroid with
$\bar n$ elements is called a \emph{model}.
\end{definition}
\noindent
In literature, the term `model' is occasionally used with a different
meaning and refers to a solution of a SAT formula.
Problem~\ref{problem1} in its general form is quite intricate. First,
it requires the definition of a linear matroid and a SAT formula with an
exponentially large number of solutions associated with isolated points.
Whereas it is easy to find examples of matroids and Boolean formulas
with this feature, it is not generally simple to characterize models
with an exponentially large number of isolated points.
Second, Eqs.~(\ref{gauss_constrs}) take to a large number of polynomial
equations in the unknown coefficients. Lemma~\ref{lemma_dim_red} can help
to reduce the search space by dimension reduction.
This will be shown in Sec.~\ref{sec_reduc_2SAT} with a simple example.
A good strategy is to start with simple models and low-degree coefficients
in Eq.~(\ref{poly_in_ideal}). In particular, we can take the coefficients
constant, as done later in Sec.~\ref{sec_quadr_poly}. This restriction
does not guarantees that Problem~\ref{problem1} has a solution for
a sufficiently small number of parameters $M$, but we can have some
hints on how to proceed.
\subsection{Required number of rational points vs space dimension}
Let assume that the computational complexity ${\bf C}_0$ of $\cal R$
is polynomial in the space dimension $n$, that is,
\begin{equation}
\label{ident_dim}
{\bf C}_0\sim n^{\alpha_0}.
\end{equation}
The factoring algorithm has polynomial complexity if
\be
\left.
\begin{array}{l}
K_1 n^\alpha\le \log N_P\le K_2 n \;\;\; 0<\alpha\le 1 \\
M\le (\log N_P)^\beta \;\;\; \beta<1
\end{array}
\right\} \;\;\; (\text{polynomial complexity})
\ee
for $n$ sufficiently great, where $K_1$ is some positive
constant and $K_2=\log 3$. The upper bound is given
by Eq.~(\ref{bound_N0_n}). The algorithm has subexponential
complexity
${\bf C}\sim e^{b (\log N_P)^{\alpha}}$ with $0<\alpha<1$ if
\be
\left.
\begin{array}{r}
\log N_P\sim (\log n)^{1/\alpha}\;\;\; 0<\alpha<1, \\
M\sim (\log n)^\frac{\beta}{\alpha}\;\;\; 0\le\beta<1-\alpha.
\end{array}
\right\} \;\;\;\; (\text{subexponential complexity})
\ee
The upper bound on $\beta$ comes from Lemma~\ref{litmus}.
Thus, the number of rational points is required to
scale much less than exponentially for getting polynomial
or subexponential factoring complexity. Note that a slower
increase of $N_P$ induces stricter bounds on $M$ in terms
of $n$.
\subsection{Reduction of models}
\label{sec_reduc_2SAT}
In this subsection, we describe an example of model reduction. The model
reduction is based on Lemma~\ref{lemma_dim_red} and can be
useful for simplifying Problem~\ref{problem1}.
The task is to reduce a class of models associated with an efficient
factoring algorithm to a class of simpler models taking to another
efficient algorithm, so that it is sufficient to search for solutions
of Problem~\ref{problem1} over the latter smaller class.
In our example, the matroid contains $2n$ elements and is
represented by the functions
$\hat a_{0,1},\dots,\hat a_{0,n},\hat a_{1,1},\dots,\hat a_{1,n}$
satisfying Property~\ref{indep_functions}.
\newline
{\bf Model A.}
\newline
Matroid with representation
$(\hat a_{0,1},\dots,\hat a_{0,n},\hat a_{1,1},\dots,\hat a_{1,n})$
satisfying Property~\ref{indep_functions}.
\newline
Generators:
\be\label{gen_Scheme_A}
\begin{array}{l}
\hat a_{0,i} \hat a_{1,i} \;\;\; i\in \{1,\dots n\} \\
\hat a_{0,i} \hat a_{0,j} \;\;\; (i,j)\in \Gamma.
\end{array}
\ee
\begin{definition}
A diagonal model is defined as Model A with $\Gamma=\emptyset$.
\end{definition}
Clearly, an diagonal model defines an algebraic set with $2^n$ isolated
points. Each point satisfies the linear equations
\be
\hat a_{s_i,i}=0, \;\;\; i\in \{1,\dots,n\}
\ee
for some $(s_1,\dots,s_n)\in\{0,1\}^n$.
If there is an algorithm with polynomial complexity
and associated with Model~A, then it is possible to prove
that there is
another algorithm with polynomial complexity and
associated with a diagonal model. More generally,
this formula reduction takes to a subexponential factoring
algorithm, provided that the parent algorithm outperforms
the quadratic sieve algorithm. If the parent algorithm outperforms
the general number field sieve, then the reduced algorithm outperforms
the quadratic sieve. Thus, if we are interested to find a competitive
algorithm from Model~A, we need to search only the space of reduced formulas.
If there is no algorithm outperforming the quadratic sieve with $\Gamma=\emptyset$,
then there is no algorithm outperforming the general number field
for $\Gamma\ne \emptyset$.
\begin{theorem}
If there is a factoring algorithm with subexponential asymptotic complexity
$e^{a (\log p)^\gamma}$ and associated with Model~A, then there is another algorithm
associated with the diagonal model with computational complexity upper-bounded by
the function $e^{\bar a (\log p)^\frac{\gamma}{1-\gamma}}$ for some $\bar a>0$.
In particular, if the first algorithm has polynomial complexity,
also the latter has polynomial complexity.
\end{theorem}
{\it Proof}.
Let us assume that the asymptotic computational complexity of the
parent algorithm is $e^{a (\log p)^\gamma}$. For every $N_P$,
there is a Model~A with $N_P$ isolated
points and generating a rational function ${\cal R}$ with complexity
${\bf C}_0(\xi)$ scaling as $e^{a (\log N_P)^\alpha}$ and a number of
parameters $M$ scaling as $(\log N_P)^\beta$, where $\gamma=\alpha/(1-\beta)$
and $0\le\beta<1$ (See~\ref{sec_complex}).
We denote by $\cal Z$ the set of isolated points.
Since the complexity ${\bf C}_0$ is lower-bounded
by a linear function of the dimension $n$, we have
\be\label{ineq_n_N0}
\log n\le a (\log N_P)^\alpha+O(1).
\ee
Let $\hat a_{0,1},\dots,\hat a_{0,n},\hat a_{1,1},\dots,\hat a_{1,n}$
be the set of linear functions representing the matroid and satisfying
Property~\ref{indep_functions}. The ideal generators are given by
Eq.~(\ref{gen_Scheme_A}).
Let $m$ be the maximum number of functions in $\{\hat a_{0,i},\dots,\hat a_{0,n}\}$
which are simultaneously different from zero for $\vec x\in \cal Z$. Thus,
we have that
\be\label{bound_N0}
N_P\le \sum_{j=0}^m\frac{n!}{(n-j)! j!}\le \left(1+n^{-1}\right) {n}^{m} .
\ee
There is a point $\vec x_2$ in $\cal Z$ such that
$\hat a_{0,1},\dots,\hat a_{0,m}$ are different from zero and
$\hat a_{0,m+1}=\hat a_{0,m+1}=\dots=\hat a_{0,n}=0$,
up to permutations of the indices.
Let us set these last $n-m$ functions equal to zero by dimension reduction.
The new set of generators is associated with another factoring algorithm
(Lemma~\ref{lemma_dim_red}) and contains the clauses of the form
\be
\begin{array}{l}
\hat a_{0,i} \hat a_{1,i} \;\;\; i\in \{1,\dots,m\} \\
\hat a_{0,i} \hat a_{0,j} \;\;\; (i,j)\in \bar\Gamma\subseteq
\{1,\dots,m\} \times \{1,\dots,m\}.
\end{array}
\ee
Since there is a point $\vec x_2$ such that $\hat a_{0,i}\ne$ for $i\in\{1,\dots,m\}$,
the set $\bar\Gamma$ turns out to be empty, so that the reduced model is diagonal.
The number of common zeros of the generators, say $N_1$, is equal to $2^m$. Using
Ineq.~(\ref{bound_N0}), we have that
\be
(\log_2 N_1)(\log n)+\log\left(1+n^{-1}\right) \ge \log N_P.
\ee
Ineq.~(\ref{ineq_n_N0}) and this last inequality implies that
\be
\log N_P \le K (\log N_1)^{\frac{1}{1-\alpha}}
\ee
for some constant $K$. Since the computational complexity, say $\bar {\bf C}_0$ of the rational
function $\cal R$ associated with the reduced model is not greater than ${\bf C}_0$,
which scales as $e^{a (\log N_P)^\alpha}$,
we have that
\be
\bar {\bf C}_0\le e^{\bar a (\log N_1)^\frac{\alpha}{1-\alpha}},
\ee
for some constant $\bar a$.
Similarly, since the number of parameters, say $\bar M$, of the reduced rational function
is not greater than $M$, we have that
\be
\bar M\le \bar K (\log \bar N_1)^\frac{\beta}{1-\alpha}
\ee
for some constant $\bar K$. Thus, the resulting factoring algorithm has a computational
complexity upper-bounded by
$$
e^{\bar a(\log p)^\frac{\alpha}{1-\alpha-\beta} }=
e^{\bar a(\log p)^\frac{\gamma}{1-\gamma} }
$$
up to a constant factor.
The last statement of the theorem is proved in a similar fashion.
$\square$
The diagonal model with generators
\be\label{simple_gene}
G_i=\hat a_{0,i} \hat a_{1,i}, \;\;\;\; \forall i\in\{1,\dots,n\}
\ee
provides the simplest example of polynomials with an exponentially large number of
common zeros. The algebraic set ${\cal Z}={\cal Z}_P$ contains $2^n$ points,
which are distinct because of Property~\ref{indep_functions}. This guarantees that
the generated ideal is radical. Thus, Hilbert's Nullstellensatz
implies that every polynomial which is zero in ${\cal Z}$
can be written as $\sum_i F_i(\vec x) G_i(\vec x)$, where $F_1,\dots,F_n$
are polynomials (let us remind that $x_{n+1}=1$).
We impose that the polynomials $P_0,\dots,P_{n-M}$ are in the ideal
generated by $G_1,\dots,G_n$, that is,
\be
P_k(\vec x)=\sum_i C_{k,i}(\vec x) \hat a_{0,i} \hat a_{1,i} \;\;\;
\forall k\in\{0,\dots,n-M\}.
\ee
As there is no particular requirement on $P_0$, we can just set $C_{0,i}(\vec x)$ equal
to constants. In particular, we can take $P_0=\hat a_{0,1}\hat a_{1,1}$.
In this case, the unknown variables of Problem~\ref{problem1} are the
polynomials $C_{k,i}(\vec x)$ and
the linear equations $\hat a_{s,k}$ under the constraints of Property~\ref{indep_functions}.
In the following section we tackle this problem with $C_{k,i}(\vec x)$
constant.
\section{Quadratic polynomials}
\label{sec_quadr_poly}
In this section, we illustrate the procedure described previously by
considering the special case of $n-M+1$ quadratic polynomials
in the ideal $I$ generated by the polynomials~(\ref{simple_gene}).
Namely, we take the polynomials $P_l$ of the form
\begin{equation}
\label{quad_poly}
P_l(\vec x)=\sum_{i=1}^n c_{l,i}\hat a_{0,i}\hat a_{1,i}, \;\;\;\;
l\in\{0,\dots,n-M\},
\end{equation}
where $c_{l,i}$ are rational numbers and the linear functions $\hat a_{s,i}$
satisfy Property~\ref{indep_functions}. Thus, there are $2^n$ common rational
zeros of the $n-M+1$ polynomials, which are also the zeros of the
generators~(\ref{simple_gene}).
Each rational point is associated
with a vector $\vec s\in\{0,1\}^n$ so that the linear equations
$\vec a_{s_1,1}\cdot\vec x=0,\dots,\vec a_{s_n,n}\cdot\vec x=0$ are
satisfied.
First, we consider the case with one parameter ($M=1$). We also assume that
all the $2^n$ rational points are in the parametrizable variety.
Starting
from these assumptions, we end up to build a variety $\cal V$ with a number
$M$ of parameters equal to $n/2-1$
for $n$ even and $n\ge4$. Furthermore, we prove that there is no solution
with $M=1$ if $n>4$. We give a numerical example for $n=4$, which
takes to a rational function $\cal R$ with $16$ zeros.
Then we build a parametrizable variety with a number of parameters equal to
$(n-1)/3$. Thus, the minimal number of parameters is some value between $2$
and $(n-1)/3$ for the considered model with the polynomials of the
form~(\ref{quad_poly}).
\subsection{One parameter? ($M=1$)}
Given polynomials~(\ref{quad_poly}) and vectors $\vec a_{s,i}$ satisfying
Property~(\ref{indep_functions}), we search for a solution of Problem~\ref{problem1}
under the assumption $M=1$.
Let us first introduce some notations and definitions.
We define the $(n-1)\times n$ matrices
\begin{equation}
\label{def_matr_M}
{\bf M}^{\vec s}\equiv
\left(
\begin{array}{ccc}
A_{1,1}^{(s_1)} & \dots & A_{1,n}^{(s_n)} \\
\vdots & \ddots & \vdots \\
A_{n-1,1}^{(s_1)} & \dots & A_{n-1,n}^{(s_n)}
\end{array}
\right),
\end{equation}
where
\begin{equation}
A_{k,i}^{(s)}\equiv\frac{\partial\hat a_{s,i}}{\partial x_k},
\end{equation}
The square submatrix of ${\bf M}^{\vec s}$ obtained by deleting the $j$-th column
is denoted by ${\bf M}_j^{\vec s}$, that is,
\begin{equation}
{\bf M}_j^{\vec s}=
\left(
\begin{array}{cccccc}
A_{1,1}^{(s_1)} & \dots & A_{1,j-1}^{(s_{j-1})} & A_{1,j+1}^{(s_{j+1})} & \dots &A_{1,n}^{(s_n)} \\
\vdots & \ddots & \vdots & \vdots & \ddots & \vdots \\
A_{n-1,1}^{(s_1)} & \dots & A_{n-1,j-1}^{(s_{j-1})} & A_{n-1,j+1}^{(s_{j+1})} & \dots &A_{n-1,n}^{(s_n)}
\end{array}
\right),
\end{equation}
The vectors $\vec a_{0,i}$ and $\vec a_{1,i}$ are also briefly denoted by $\vec a_i$ and
$\vec b_i$, respectively. Similarly, we also use the symbols $A_{k,i}$ and $B_{k,i}$
for the derivatives $A_{k,i}^{(0)}$ and $A_{k,i}^{(1)}$.
Problem~\ref{problem1} takes the specific form
\begin{problem}
\label{problem2}
Find coefficients $c_{l,i}$ and vectors $\vec a_{s,i}$ satisfying
Property~(\ref{indep_functions}) such that
\bey
\label{prob2_1}
\sum_{i=1}^n c_{l,i} \left( A_{k,i}\vec a_i+B_{k,i}\vec b_i\right)=0
\;\;\;\; 1\le k < l\le n-1,
\vspace{2mm}
\\
\label{prob2_2}
\sum_{i=1}^n c_{l,i} A_{l,i} B_{l,i}=0 \;\;\;\; 1\le l\le n-1,
\vspace{2mm}
\\
\label{prob2_3}
\vec x\in{\cal Z}_P \Rightarrow
\sum_{i=1}^n c_{l,i} \left( A_{l,i}\hat a_i+B_{l,i}\hat b_i\right)\ne 0
\;\;\;\; 1\le l\le n-1.
\eey
\end{problem}
Let us stress again that the problem is invariant with respect to
the transformations~(\ref{poly_replacement},\ref{invar_trans}), the latter
taking to the transformation
\be
A_{k,i}^{(s)}\rightarrow A_{k,i}^{(s)}+\sum_{l=1}^{k-1}\bar\eta_{k,l}A_{l,i}^{(s)}
\ee
of the derivatives.
We have the following.
\begin{lemma}
\label{lemma_inde_A}
For every $\vec s\in\{0,1\}^n$ and $j\in\{1,\dots,n\}$, the matrix ${\bf M}_j^{\vec s}$
has maximal rank, that is,
\begin{equation}
\det {\bf M}_j^{\vec s}\ne 0.
\end{equation}
\end{lemma}
{\it Proof}.
Let us prove the lemma by contradiction. There is a $j\in\{1,\dots,n\}$,
$l\in\{1,\dots,n-1\}$, and an $\vec s\in\{0,1\}^n$
such that the $l$-th row of ${\bf M}_j^{{\vec s}}$
is linearly dependent on the first $l-1$ rows. Thus,
there are coefficients $\lambda_1,\dots,\lambda_{l-1}$ such that
\begin{equation}
A_{l,i}^{(s_l)}+\sum_{k=1}^{l-1}\lambda_k A_{k,i}^{(s_k)}=0 \;\;\; \forall i\ne j
\end{equation}
With a change of variables of the form of Eq.~(\ref{invar_trans}), this equation
can be rewritten in the form
\begin{equation}
A_{l,i}^{(s_l)}=0 \;\;\; \forall i\ne j.
\end{equation}
Up to permutations $\hat a_i\leftrightarrow \hat b_i$, we have
\begin{equation}
\label{lin_dep_lemma}
B_{l,i}=0 \;\;\; \forall i\ne j.
\end{equation}
From Eq.~(\ref{prob2_2}), we have
$$
\sum_{i=1}^n c_{l,i} A_{l,i} B_{l,i}=0.
$$
From this equation and Eq.~(\ref{lin_dep_lemma}), we get the
equation $c_{l,j} A_{l,j} B_{l,j}=0$, implying that
$c_{l,j} A_{l,j}=0$ or $c_{l,j} B_{l,j}=0$. Without loss of generality,
let us take
\be
\label{clne0}
c_{l,j} B_{l,j}=0.
\ee
Let $\vec x_0\in{\cal Z}_P$ be the vector orthogonal to $\vec b_1,\dots,\vec b_n$.
From Eq.~(\ref{prob2_3}) we have that
\be
c_{l,j} B_{l,j}(\vec a_j\cdot\vec x_0)\ne 0,
\ee
which is in contradiction with Ineq.~(\ref{clne0}).
$\square$
\begin{corollary}
\label{corol_c}
The coefficients $c_{n-1,i}$ are different from zero for every
$i\in\{1,\dots,n\}$.
\end{corollary}
{\it Proof}.
Let us assume that the statement is false. Up to permutations, we have that
$c_{n-1,1}=0$. Lemma~\ref{lemma_inde_A} implies that there is an
integer $i_0\in\{2,\dots,n\}$ such that
$B_{n-1,i}=0$ for $i\notin\{1,i_0\}$, up to a transformation of the form of Eq.~(\ref{invar_trans}).
Thus,
\be
0=\sum_{i=1}^n c_{n-1,i}A_{n-1,i}B_{n-1,i}= c_{n-1,i_0} A_{n-1,i_0}B_{n-1,i_0},
\ee
the first equality coming from Eq.~(\ref{prob2_2}). Lemma~\ref{lemma_inde_A} also
implies that $A_{n-1,i_0}B_{n-1,i_0}\ne 0$
Thus, on one hand, we have that $c_{n-1,i_0}=0$.
On the other hand, we have
\be
c_{n-1,i_0} B_{n-1,i_0}(\vec a_{i_0}\cdot\vec x_0)\ne 0
\ee
from Eqs.~(\ref{prob2_3}),
where $\vec x_0$ is the vector orthogonal to $\vec b_1,\dots,\vec b_n$.
Thus, we have a contradiction. $\square$
Let us denote by ${\bf M}^{\vec s}_{j_1,\dots,j_m}$
the submatrix of ${\bf M}^{\vec s}$ obtained by deleting the last $m-1$ rows and
the columns $j_1,\dots,j_m$.
Given the coefficient matrix
\be
{\bf c}\equiv
\left(
\begin{array}{ccc}
c_{0,1} & \dots & c_{0,n} \\
\vdots & \ddots & \vdots \\
c_{n-1,1} & \dots & c_{n-1,n}
\end{array}
\right),
\ee
let us define ${\bf c}_{j_1,\dots,j_m}$ as the $m\times m$ submatrix of
${\bf c}$ obtained by keeping the last $m$ rows and the columns
$j_1,\dots,j_m$.
Lemma~\ref{lemma_inde_A} and Corollary~\ref{corol_c} are generalized by
the following.
\begin{theorem}
\label{theorem_det}
For every $m\in\{1,\dots,n-1\}$, $\vec s\in\{0,1\}^n$, and $m$ distinct
integers $j_1,\dots,j_m\in\{1,\dots,n\}$,
the matrices ${\bf M}^{\vec s}_{j_1,\dots,j_m}$ and ${\bf c}_{j_1,\dots,j_m}$
have maximal rank, that is,
\bey
\label{det_reduc}
\det {\bf M}^{\vec s}_{j_1,\dots,j_m}\ne 0, \\
\label{det_reduc2}
\det {\bf c}_{j_1,\dots,j_m}\ne 0.
\eey
\end{theorem}
{Proof.} The proof is by recursion. For $m=1$, the theorem comes from
Lemma~\ref{lemma_inde_A} and Corollary~\ref{corol_c}.
Thus, we just need to prove Eqs.~(\ref{det_reduc},\ref{det_reduc2}) by assuming that
\bey
\label{recur0}
\det {\bf M}^{\vec s}_{j_1,\dots,j_{m-1}}\ne 0, \\
\label{recur1}
\det {\bf c}_{j_1,\dots,j_{m-1}}\ne 0.
\eey
Let us first prove Eq.~(\ref{det_reduc}) by contradiction. If
the equation is false, then there is an $\vec s_0\in\{0,1\}^n$ and
$m$ distinct integers $i_1,\dots,i_m$ in $\{1,\dots,n\}$ such that
$\det {\bf M}^{\vec s_0}_{i_1,\dots,i_m}=0$.
By permutations, we can
set $i_h=h$. By suitable exchanges of $\hat a_i$ and $\hat b_i$, we can
set $s_i=1$ for every $i\in\{1,\dots,n\}$.
There is an integer $l\in\{1,\dots,n-m\}$ such that $B_{l,i}=0$ for
$i\in\{m+1,\dots,n\}$ up to a transformation of the form of Eq.~(\ref{invar_trans}).
From Eqs.~(\ref{prob2_1},\ref{prob2_2}), we have the $m$ equations
\be
\begin{array}{r}
\sum_{i=1}^m c_{l,i}A_{l,i}B_{l,i}=0 \\
\sum_{i=1}^m c_{n+1-m,i}A_{l,i}B_{l,i}=0 \\
\sum_{i=1}^m c_{n+2-m,i}A_{l,i}B_{l,i}=0 \\
\dots \\
\sum_{i=1}^m c_{n-2,i}A_{l,i}B_{l,i}=0 \\
\sum_{i=1}^m c_{n-1,i}A_{l,i}B_{l,i}=0.
\end{array}
\ee
From Eq.~(\ref{recur0}), we have that $A_{l,i}\ne0$ and $B_{l,i}\ne0$ for some
$i\in\{1,\dots,m\}$, so that
\be
\text{rank}
\left(
\begin{array}{ccc}
c_{l,1} & \dots & c_{l,m} \\
c_{n+1-m,1} & \dots & c_{n+1-s,m} \\
c_{n+2-m,1} & \dots & c_{n+2-s,m} \\
\vdots & \ddots & \vdots \\
c_{n-2,1} & \dots & c_{n-1,m} \\
c_{n-1,1} & \dots & c_{n-1,m}
\end{array}
\right)< m.
\ee
Up to a transformation of the form of Eq.~(\ref{poly_replacement}),
there is an integer $l_0\in\{n+1-m,\dots,n-1\}\cup\{l\}$ such
that $c_{l_0,i}=0$ for $i\in\{1,\dots,m\}$. Eq.~(\ref{recur1}) implies
that $l_0=l$. Thus, $c_{l,1}=\dots c_{l,m}=0$, but this contradicts
Eq.~(\ref{prob2_3}) with $\vec x\in{\cal Z}_P$ orthogonal to $\vec b_1,\dots,\vec b_n$.
Let us now prove Eq.~(\ref{det_reduc2}) by contradiction. If the equation is false,
then there are $m$ distinct integers $i_1,\dots,i_m$ in $\{1,\dots,n\}$ such that
$\det {\bf c}_{i_1,\dots,i_m}=0$. Without loss of generality, let us take
$i_h=h$. Up to the transformation~(\ref{poly_replacement}),
there is an integer $l\in\{n-m,\dots,n-1\}$ such that $c_{l,i}=0$ for
$i\in\{1,\dots,m\}$. Eq.~(\ref{recur1}) implies that $l=n-m$. Thus,
\be
c_{n-m,1}=\dots=c_{n-m,m}=0.
\ee
Eq.~(\ref{det_reduc}) implies that there is an integer
$i_0\in\{m+1,\dots,n\}$ such that $A_{n-m,i}=0$ for $i\in\{m+1,\dots,n\}\bs\{i_0\}$
up to transformation~(\ref{invar_trans}). Thus, we have from Eq.~(\ref{prob2_2})
that
\be
0=\sum_{i=1}^n c_{n-m,i}A_{n-m,i}B_{n-m,i}=
c_{n-m,i_0}A_{n-m,i_0}B_{n-m,i_0}.
\ee
Eq.~(\ref{det_reduc}) also implies that $A_{n-m,i_0}B_{n-m,i_0}\ne 0$, so that
$c_{n-m,i_0}=0$, which is in contradiction with Eq.~(\ref{prob2_3}) for $\vec x$
orthogonal to $\vec a_1,\dots,\vec a_n$. $\square$
\newline
In the following, this theorem will be used with $m\in\{1,2\}$.
Since all the coefficients $c_{n-1,i}$ are different from zero, we can set
them equal to $1$ by rescaling the vectors $\vec a_i$ or $\vec b_i$.
Let us denote by $c_i$ the coefficients $c_{n-2,i}$. Theorem~\ref{theorem_det}
with $m=2$ implies that $c_i\ne c_j$ for $i\ne j$.
Eq.~(\ref{prob2_1}) with $l=n-1$ and $l=n-2$ takes the form
\bey
\label{nm2}
\frac{\partial }{\partial x_k}P_{n-1}=\sum_{i=1}^n \left( A_{k,i}\vec a_i+B_{k,i}\vec b_i\right)=0
\;\;\;\; 1\le k \le n-2, \\
\label{nm3}
\frac{\partial }{\partial x_k}P_{n-2}=
\sum_{i=1}^n c_i\left( A_{k,i}\vec a_i+B_{k,i}\vec b_i\right)=0
\;\;\;\; 1\le k \le n-3.
\eey
These equations impose the form~(\ref{poly_affine}) to the last two
polynomials, $P_{n-1}$ and $P_{n-2}$, which must be independent from $n-2$ and
$n-3$ variables, respectively.
The first $n-2$ vector equations are linearly independent. Let us assume
the opposite. Then, there is a set of coefficients $\lambda_1,\dots,\lambda_{n-2}$
such that
$\sum_{k=1}^{n-2}\lambda_k (A_{k,i},B_{k,i})=0$.
But this is impossible because of Property~\ref{indep_functions}. It also
contradicts Theorem~\ref{theorem_det}. The theorem also implies that
Eqs.~(\ref{nm3}) are linearly
independent. Since the vector space is $n+1$-dimensional, the vectors
$\vec a_i$ and $\vec b_i$ must have $n-1$ vector constraints. Thus,
at least $n-4$ out of Eqs.~(\ref{nm3}) are linearly dependent on
Eqs.~(\ref{nm2}).
First, let us show that $n-4$ is the maximal number of dependent equations.
Assuming the converse, we have
\be
c_i(A_{k,i},B_{k,i})=\sum_{l=1}^{n-2}\lambda_{k,l} (A_{l,i},B_{l,i}) \;\;\;\;\;
\forall k\in\{1,\dots,n-3\}.
\ee
for suitable coefficients $\lambda_{k,l}$. Let us define the linear
superposition
\be
(A_i,B_i)\equiv \sum_{k=1}^{n-3} v_k (A_{k,i},B_{k,i})
\ee
with the coefficients $v_k$. Let $\bf \Lambda$ be the
$(n-2)\times(n-2)$ matrix with ${\bf \Lambda}_{k,n-2}=0$
and ${\bf \Lambda}_{k,l}=\lambda_{l,k}$
for $k\in\{1,\dots,n-2\}$ and $l\in\{1,\dots,n-3\}$.
The coefficients
$(v_1,\dots,v_{n-3})\equiv\vec v$ are defined by imposing
the $n-4$ constraints
\be
({\bf\Lambda}^s\vec v)_{n-2}=0 \;\;\;\;\; s\in\{1,\dots,n-4\}.
\ee
By construction, the pairs
\be\label{expo_form}
c_i^{k-1}(A_i,B_i) \;\;\;\; k\in\{1,\dots,n-2\}
\ee
are linear superpositions of the derivatives $(A_{k,i},B_{k,i})$ with
$k\in\{1,\dots,n-2\}$. Furthermore, the first $n-3$ pairs are
linear superpositions of $(A_{k,i},B_{k,i})$ with
$k\in\{1,\dots,n-3\}$. That is,
\be
\label{expo_deps}
\begin{array}{l}
c_i^{k-1} (A_i,B_i)=\sum_{l=1}^{n-3}\bar\lambda_{k,l}(A_{l,i},B_{l,i})
\;\;\; k\in\{1,\dots,n-3\} \\
c_i^{n-3} (A_i,B_i)=\sum_{l=1}^{n-2}\bar\lambda_{n-2,l}(A_{l,i},B_{l,i})
\end{array}
\ee
for some coefficients $\bar\lambda_{k,l}$. From Lemma~\ref{lemma_inde_A}
and Corollary~\ref{corol_c} we have that the $n-2$ pairs~(\ref{expo_form})
are linearly independent. Indeed, Corollary~\ref{corol_c} implies that
$A_i\ne 0$ and $B_i\ne 0$ for every $i\in\{1,\dots,n\}$. Lemma~\ref{lemma_inde_A}
implies that $c_i^{k-1}$ are linearly independent for $k\in\{1,\dots,n-2\}$.
Equations~(\ref{expo_form},\ref{expo_deps}) can be also derived from Jordan's
theorem, Lemma~\ref{lemma_inde_A} and Corollary~\ref{corol_c}. See
Appendix~\ref{lin_alg_tools}.
Thus, by a variable transformation,
Eqs.~(\ref{nm2},\ref{nm3}) take the form
\be
\sum_{i=1}^n c_i^{k-1}\left( A_{i}\vec a_i+B_{i}\vec b_i\right)=0
\;\;\;\;\;\; k\in\{1,\dots,n-2\}.
\ee
and
\be
\frac{\partial (\hat a_i,\hat b_i)}{\partial x_k}=c_i^{k-1}(A_i, B_i)
\;\;\;\;\;\; k\in\{1,\dots,n-2\}.
\ee
These equations imply that
$\sum_{i=1}^n d_i^{k-1} A_i B_i=0$ for $k\in\{1,\dots,2n-5\}$.
For $n>4$, we have in particular that
\be
\label{kill_AB}
\sum_{i=1}^n c_i^{k-1} A_i B_i=0 \;\;\;\; k\in\{1,\dots,n\}.
\ee
Since
\be
\det\left(
\begin{array}{ccc}
1 & \dots & 1 \\
c_1 & \dots & c_n \\
\vdots & \ddots & \vdots \\
c_1^{n-1} & \dots & c_n^{n-1}
\end{array}
\right)=\prod_{j>i} (c_j-c_i)
\ee
and $c_i\ne c_j$ for $i\ne j$, Eq.~(\ref{kill_AB}) implies that $A_i B_i=0$ for every
$i\in\{1,\dots,n\}$. But this is in contradiction with Theorem~\ref{theorem_det}.
Thus, let us take exactly $n-4$ out of Eqs.~(\ref{nm3}) linearly dependent on
Eqs~(\ref{nm2}). Let $\bar k$ be an integer in $\{1,\dots,n-3\}$ such that
Eq.~(\ref{nm3}) with $k=\bar k$ is linearly independent of Eqs.~(\ref{nm2}). Thus,
$$
c_i(A_{k,i},B_{k,i})=\bar\lambda_k c_i (A_{\bar k,i},B_{\bar k,i})+
\sum_{l=1}^{n-2}\lambda_{k,l}(A_{l,i},B_{l,i}) \;\;\;\;
k\in\{1,\dots,n-3\}\bs\{\bar k\}.
$$
By a transformation of the first $n-3$ variables, we can rewrite this
equation in the form.
\be
c_i(A_{k,i},B_{k,i})=\sum_{l=1}^{n-2}\lambda_{k,l}(A_{l,i},B_{l,i}) \;\;\;\;
k\in\{1,\dots,n-4\}.
\ee
By a suitable transformation of the first $n-2$ variables,
the $n-2$ pairs $(A_{k,i},B_{k,i})$ can be split in two groups (see
Appendix~\ref{lin_alg_tools}), say,
\be
\left.
\begin{array}{l}
\frac{\partial}{\partial x_k'} (\hat a_i,\hat b_i) \equiv
(A_{k,i}', B_{k,i}')=c_i^{k-1} (A_i',B_i') \;\;\;\; k\in\{1,\dots,n_1\} \\
\frac{\partial}{\partial x_k''}(\hat a_i,\hat b_i) \equiv
(A_{k,i}'', B_{k,i}'')=c_i^{k-1} (A_i'',B_i'') \;\;\;\; k\in\{1,\dots,n_2\}
\end{array} \right\}
\;\;\; n_1+n_2=n-2.
\ee
Equations~(\ref{nm2}) become
\begin{equation}
\label{nm2_vector_equation}
\begin{array}{l}
\sum_{i=1}^n c_i^{k-1}\left( A_i' \vec b_i+B_i' \vec a_i\right)=0 \;\;\; k\in\{1,\dots,n_1\} \\
\sum_{i=1}^n c_i^{k-1}\left(\bar A_i'' \vec b_i+\bar B_i'' \vec a_i\right)=0 \;\;\; k\in\{1,\dots,n_2\}.
\end{array}
\end{equation}
Given these $n-2$ vector constraints, all $n-2$ the derivatives
$\partial P_{n-1}/\partial x_1',\dots,\partial P_{n-1}/\partial x_{n_1}'$,
$\partial P_{n-1}/\partial x_1'',\dots,\partial P_{n-1}/\partial x_{n_2}''$ are equal to
zero. Furthermore, we also have that
$$
\begin{array}{l}
\frac{\partial}{\partial x_k'}P_{n-2}=0 \;\;\; k\in\{1,\dots,n_1-1\} \\
\frac{\partial}{\partial x_k''}P_{n-2}=0 \;\;\; k\in\{1,\dots,n_2-1\},
\end{array}
$$
so that $P_{n-2}$ is independent of $n-4$ out of the $n-2$ variables
$x_1',\dots,x_{n_1}$, $x_1'',\dots,x_{n_2}''$. Thus, we need to add
another vector equation such that
$\left(w_1 \frac{\partial}{\partial x_{n_1}'}+w_2 \frac{\partial}{\partial x_{n_2}''}\right)P_{n-2}=0$
for some $(w_1,w_2)\ne (0,0)$. Up to a variable transformation, we can set $(w_1,w_2)=(1,0)$
so that the additional vector equation is
\be
\label{additional_equation}
\sum_{i=1}^n c_i^{n_1}\left( A_i' \vec b_i+B_i' \vec a_i\right)
=0.
\ee
Equations~(\ref{prob2_2},\ref{nm2_vector_equation},\ref{additional_equation}) imply that
\bey
\label{eqs_coef_AB}
\sum_{i=1}^n c_i^{k-1} A_i' B_i'=0 \;\;\; k\in\{1,\dots,2 n_1\}, \\
\label{eqs_coef_AbBb}
\sum_{i=1}^n c_i^{k-1} A_i'' B_i''=0 \;\;\; k\in\{1,\dots,2 n_2\}, \\
\label{eqs_coef_AbB}
\sum_{i=1}^n c_i^{k-1} ( A_i' B_i''+A_i'' B_i')=0 \;\;\; k\in\{1,\dots,n_1+n_2\}.
\eey
Since $A_i' B_i'$ and $A_i' B_i'$ are not identically equal to zero (as consequence of
Theorem~\ref{theorem_det}), the number of Eqs.~(\ref{eqs_coef_AB}) and Eqs.~(\ref{eqs_coef_AbBb})
is smaller than $n$, so that
$$
n_1\le \frac{n-1}{2},\;\;\; n_2\le \frac{n-1}{2}.
$$
Without loss of generality, we can assume that $n$ is even. Indeed, if Problem~\ref{problem1}
can be solved for $n$ odd, then Lemma~\ref{lemma_dim_red} implies that it can be solve
for $n$ even, and {\it viceversa}. Since $n_1+n_2=n-2$, we have that
\be
n_1=n_2=\frac{n-2}{2}.
\ee
Let $W_1,\dots, W_n$ be $n$ numbers defined by the equations
\be
\label{def_W}
\sum_{i=1}^n c_i^{k-1} W_i=0 \;\;\;\; k\in\{1,\dots,n-1\}
\ee
up to a constant factor. Equations~(\ref{eqs_coef_AB},\ref{eqs_coef_AbBb},\ref{eqs_coef_AbB})
are equivalent to the equations
\bey
\label{AB}
A_i' B_i'=(k_0+k_1 c_i) W_i, \\
\label{AbBb}
A_i'' B_i''=(r_0+r_1 c_i) W_i, \\
\label{AbB}
A_i' B_i''+A_i'' B_i'=(s_0+s_1 c_i) W_i.
\eey
These equations can be solved over the rationals for the coefficients $c_i$, $B_i'$ and
$B_i''$ in terms of $A_i'$ and $A_i''$. The coefficients $c_i$ take a form which is
independent of $W_i$,
\be
c_i=\frac{r_0 A_i'^{\, 2}+k_0 A_i''^{\,2}-s_0 A_i' A_i''}{r_1 A_i'^{\, 2}+k_1 A_i''^{\,2}-s_1 A_i' A_i''},
\ee
so that we first evaluate $c_i$, then $W_i$ by Eq.~(\ref{def_W}) and, finally, $B_i'$ and
$B_i''$ by Eqs.~(\ref{AB},\ref{AbBb}).
It is possible to show that condition~(\ref{prob2_3}) for $l=n-1$ implies that $(k_1,r_1)\ne(0,0)$.
Indeed, if $(k_1,r_1)=(0,0)$, then only half of the points in ${\cal Z}_P$ satisfies
the inequality in the condition.
Up to a variable change, we have
$$
k_1\ne 0,\;\;\; s_1=0.
$$
Up to now we have been able to solve all the conditions of Problem~\ref{problem2} which
refer to the last two polynomials, that is, for $l=n-2,n-1$. The equations that need to
be satisfied are Eqs.~(\ref{nm2_vector_equation},\ref{additional_equation},
\ref{def_W},\ref{AB},\ref{AbBb},\ref{AbB}). Let us rewrite them all together.
\be
\boxed{
\begin{array}{c}
A_i' B_i'=(k_0+k_1 c_i) W_i, \;\;\;
A_i'' B_i''=(r_0+r_1 c_i) W_i \\
A_i' B_i''+A_i'' B_i'=s_0 W_i, \;\;\; k_1\ne 0
\vspace{1mm} \\
\sum_{i=1}^n c_i^{k-1} W_i=0 \;\;\;\; k\in\{1,\dots,n-1\}
\vspace{1mm} \\
\sum_{i=1}^n c_i^{k-1}\left( A_i' \vec b_i+B_i' \vec a_i\right)=0 \;\;\; k\in\{1,\dots,\frac{n}{2}\} \\
\sum_{i=1}^n c_i^{k-1}\left(\bar A_i'' \vec b_i+\bar B_i'' \vec a_i\right)=0 \;\;\; k\in\{1,\dots,
\frac{n}{2}-1\}
\end{array}
}
\ee
Given $2n$ vectors $\vec a_1,\dots,\vec a_n,\vec b_1,\dots,\vec b_n$ satisfying these equations,
there are $n-1$ directions $\vec u_1,\dots,\vec u_{n-1}$ such that
\be
\begin{array}{l}
\vec u_{2k-1}\cdot \frac{\partial}{\partial\vec x} (\hat a_i,\hat b_i)=c_i^{k-1}(A_i',B_i')
\;\;\; k\in\{1,\dots,\frac{n}{2}-1\} \\
\vec u_{2k}\cdot \frac{\partial}{\partial\vec x} (\hat a_i,\hat b_i)=c_i^{k-1}(A_i'',B_i'')
\;\;\; k\in\{1,\dots,\frac{n}{2}\}.
\end{array}
\ee
This can be easily verified by substitution.
Let us define the coordinate system $(y_1,\dots,y_{n+1})\equiv \vec y$ such that
\be
\vec u_k\cdot \frac{\partial}{\partial\vec x} =\frac{\partial}{\partial y_k} \;\;\;
k\in\{1,\dots,n-1\}.
\ee
Given the polynomials
\begin{equation}
\begin{array}{l}
P_{n-1}=\sum_{i=1}^n \hat a_i\hat b_i \\
P_{n-2}=\sum_{i=1}^n c_i \hat a_i\hat b_i
\end{array}
\end{equation}
with $\hat a_i=\vec a_i\cdot \vec y$ and $\hat b_i=\vec b_i\cdot \vec y$, it is
easy to verify that
\begin{equation}
\begin{array}{l}
\frac{\partial P_{n-1}}{\partial y_k}=0, \;\;\;\; k\in\{1,\dots,n-2\}, \\
\frac{\partial P_{n-2}}{\partial y_k}=0, \;\;\;\; k\in\{1,\dots,n-3\}, \\
\frac{\partial^2 P_{n-2}}{\partial y_{n-2}^2}=0.
\end{array}
\end{equation}
The polynomial $P_{n-1}$ depends on $2$ variables (in the affine space) and
the polynomial $P_{n-2}$ depends linearly on an additional variable $y_{n-2}$.
Thus, the algebraic set of the two polynomials admits a Gaussian parametrization,
that is, the equations $P_{n-1}=0$ and $P_{n-2}=0$ can be solved with \emph{a la} Gauss
elimination of two variables. Note that the polynomial $P_{n-1}$ has rational
roots by construction.
The next step is to satisfy the conditions of Problem~\ref{problem2} for the other
polynomials
$P_{1},\dots,P_{n-3}$ by setting $c_{k,i}$ and the other remaining free coefficients.
It is interesting to note that it is sufficient to take $c_{2s,i}=c_i^{n/2-s}$ with
$s\in\{1,\dots,(n-4)/2\}$ for satisfying every condition of Problem~\ref{problem2}
for $l$ even. The polynomials $P_2,P_4,\dots,P_{n-2}$ take the form
\be\label{poly_even}
P_{2s}=\sum_{i=1}^n c_i^{n/2-s} \hat a_i\hat b_i, \;\;\; s\in\{1,\dots,(n-4)/2\}.
\ee
Furthermore, we can choose $c_{1,i}$ such that $\partial^2 P_1/\partial x_1^2=0$.
With this choice, we have that
\be
\left.
\begin{array}{l}
\frac{\partial P_l}{\partial y_k}=0, \;\;\;\; k\in\{1,\dots,l-1\} \\
\frac{\partial^2 P_l}{\partial y_l^2}=0
\end{array}
\right\} \;\;\; l\in\{2,4,\dots,n-4,n-2\}\cup\{1,n-1\}.
\ee
Thus, we are halfway to solve Problem~\ref{problem2}, about half of the conditions
are satisfied. The hard core of the problem is to solve the conditions for
$P_1,P_3,\dots,P_{n-3}$.
The form of Polynomials~(\ref{poly_even}) is not necessarily the most general. Thus, let
us take a step backward and handle Problem~\ref{problem2} for the polynomial $P_{n-3}$
with the equations derived so far. We will find that this polynomial cannot
satisfy the required conditions if $n>4$, so that the number of parameters has
to be greater than $1$.
Let us denote by $d_i$ the coefficients $c_{n-3,i}$. Eqs.~(\ref{prob2_1},\ref{prob2_2})
with $l=n-3$ give the equations
$$
\sum_{i=1}^n e_i \left(A_{k,i} B_{k',i}+A_{k',i} B_{k,i}\right)=0 \;\;\;
k,k'\in\{1,\dots,n-3\},
$$
which imply that
$$
\begin{array}{l}
\sum_{i=1}^n e_i c_i^{k+k'-2} A_i' B_i'=0\;\;\; k,l\in\{1,\dots,\frac{n}{2}-1\} \\
\sum_{i=1}^n e_i c_i^{k+k'-2} A_i'' B_i''=0\;\;\; k,l\in\{1,\dots,\frac{n}{2}-2\} \\
\sum_{i=1}^n e_i c_i^{k+k'-2} \left( A_i' B_i''+ A_i'' B_i'\right) =0\;\;\;
\left\{
\begin{array}{l}
k\in\{1,\dots,\frac{n}{2}-1\} \\
k'\in\{1,\dots,\frac{n}{2}-2\}
\end{array}
\right.
\end{array}
$$
that is,
\be
\label{eqs_for_e}
\begin{array}{l}
\sum_{i=1}^n e_i c_i^{k-1} A_i' B_i'=0\;\;\; k\in\{1,\dots,n-3\} \\
\sum_{i=1}^n e_i c_i^{k-1} A_i'' B_i''=0\;\;\; k\in\{1,\dots,n-5\} \\
\sum_{i=1}^n e_i c_i^{k-1} \left( A_i' B_i''+ A_i'' B_i'\right) =0\;\;\;
k\in\{1,\dots,n-4\}.
\end{array}
\ee
These equations imply that
\be
\begin{array}{l}
e_i A_i' B_i'=F_{11}(c_i) W_i \\
e_i A_i'' B_i''=F_{22}(c_i) W_i \\
e_i \left(A_i' B_i''+ A_i'' B_i'\right) =F_{12}(c_i) W_i,
\end{array}.
\ee
where $F_{11}(x)$, $F_{22}(x)$ and $F_{12}(x)$ are polynomials of degree lower than $3$, $5$ and $4$,
respectively. Thus,
\be
e_i=\frac{F_{11}(c_i)}{k_0+k_1 c_i}=\frac{F_{22}(c_i)}{r_0+r_1 c_i}=
\frac{F_{12}(c_i)}{s_0}.
\ee
The second and third equalities give polynomials of degree lower than $6$ and $5$,
respectively. Since $c_i\ne c_j$ for $i\ne j$ and $n$ is even, the coefficients of
these polynomials are equal to zero for $n>4$. In particular, $k_0+k_1 c_i$ divides
$F_{11}(c_i)$ and, thus, $e_i$ is equal to a linear function of $c_i$. We have that
$P_{n-3}=q_1 P_{n-2}+q_2 P_{n-1}$ for some constants $q_1$ and $q_2$, so that
there is no independent polynomial $P_{n-3}$ satisfying the required conditions
for $n>4$.
In conclusion, we searched for a solution of Problem~\ref{problem2} with one parameter ($M=1$), but
we ended up to find a solution with $n/2-1$ parameters. Let us stress that we have not
proved that $M$ cannot be less than $n/2-1$, we have only proved that $P_{n-3}$
cannot satisfy the required conditions, so that solutions with $M>1$ may exist.
Furthermore, we employed the condition $M=1$ in some
intermediate inferences. Thus, to check the existence of better solutions, we need to consider
the case $M\ne 1$ from scratch.
For the sake of completeness, let us write down the solution for $n=4$. Eqs.~(\ref{eqs_for_e})
reduce to
\be
\sum_{i=1}^n e_i A_i' B_i'=0,
\ee
Up to a replacement $P_{1}\rightarrow \lambda_1 P_1+\lambda_2 P_2+\lambda_3 P_3$ for some constants
$\lambda_i$ with $\lambda_1\ne 0$, we have that
\be
e_i=\frac{1}{k_0+k_1 c_i}.
\ee
Thus, the $4$ polynomials take the form
\be
\begin{array}{ll}
P_0=\hat a_1\hat b_1, \;\;\;\;
& P_1=\sum_{i=1}^{4}\frac{\hat a_i \hat b_i}{k_0+k_1 c_i} \\
P_2=\sum_{i=1}^{4}c_i \hat a_i \hat b_i \;\;\;\;
& P_3=\sum_{i=1}^{4} \hat a_i \hat b_i.
\end{array}
\ee
Let us give a numerical example with $4$ polynomial, built by using the derived equations.
\subsubsection{Numerical example with $n=$4}
Let us set $A_i'=i$, $A_i''=1$, $k_0=k_1=r_0=1$, $r_1=2$, and $s_0=3$. Up to
a linear transformation of $x_3$ and $x_4$, this setting gives the polynomials
\be
\begin{array}{l}
P_3(x_3,x_4)=\\ 5 x_3 \left(8427 x_4+9430\right)-209 \left(3 x_4 \left(393 x_4+880\right)+1478\right)
\vspace{1mm} \\
P_2(x_2,x_3,x_4)= \\
5538425 x_3^2+18810 \left(1445 x_2+5718 x_4+6421\right) x_3-786258 \left(3 x_4 \left(267 x_4+598\right)+1004\right)
\vspace{1mm} \\
P_1(x_1,x_2,x_3,x_4)= \\
2299 [205346285 x_3-38 (63526809 x_4+35594957)]- 5 [-2045057058 x_2^2+ \\
1630827 (1813 x_3+1254 x_4) x_2+2891872832 x_3^2+495958966272 x_4^2+ \\
4892481 x_1 (1254 x_2-1429 x_3-418)-87093628743 x_3 x_4] \vspace{1mm}\\
P_0(x_1,x_2,x_3,x_4)= \\
\left(627 x_1+627 x_2-46 x_3+1881 \left(x_4+1\right)\right) \left(5016 x_1+6270 x_2+2555 x_3-3762 \left(4 x_4+5\right)\right)
\end{array}
\ee
Taking $x_4$ as the parameter $\tau$ and solving the equations $P_3=P_2=P_1=0$ with respect to $x_3$, $x_2$ and $x_1$,
we replace the result in $P_0$ and obtain, up to a constant factor,
\be
{\cal R}(\tau)=\frac{\prod_{k=1}^{16}(\tau-\tau_k)}{Q_1^2(\tau)Q_2^2Q_3^2(\tau)},
\ee
where
\be
\begin{array}{l}
Q_1(\tau)=8427 \tau +9430, \\
Q_2(\tau)=3 \tau (393 \tau +880)+1478, \\
Q_3(\tau)=3 \tau (9 \tau (7 \tau
(5367293625 \tau +24273841402)+288165964484)+1954792734568)+1657527934720, \\
(\tau_1,\dots\tau_{16})=
-\left(\frac{86}{69},\frac{800}{681},\frac{122}{105},\frac{3166}{2775},\frac{
140}{123},\frac{718}{633},\frac{2452}{2163},\frac{5558}{4929},\frac{2578}{2
289},
\frac{152}{135},\frac{1070}{951},\frac{3932}{3507},\frac{158}{141},
\frac{2072}{1851},\frac{1142}{1023},\frac{218}{201}\right)
\end{array}
\ee
Over a finite field $\mathbb{Z}_p$, one can check that the numerator has about $16$ distinct roots for $p\gg 16$.
For $p\simeq 16$, the roots are lower because of collision or because the denominator of some rational root $\tau_k$
is divided by $p$.
\subsubsection{Brief excursus on retro-causality and time loops}
Previously, we have built the polynomials~(\ref{poly_even}). Setting them equal to zero, we have
a triangular system of about $n/2$ polynomial equations that can be efficiently solved in $n/2$ variables,
say ${\bf x}_1$, given the value of the other variables, say ${\bf x}_2$. This system is more or less
symmetric, that is, the variables ${\bf x}_2$ can be efficiently computed given the first block
${\bf x}_1$ (up to few variables). To determine the overall set of variables, we need the missing
$n/2$ polynomials in the ideal $I$.
It is possible to choose the coefficients $c_{l,i}$ of these polynomials in a
such a way that the associated equations have again a triangular form with respect to one of the
two blocks ${\bf x}_1$ and ${\bf x}_2$, up to few variables. Thus, we end up with two independent
equations and a boundary condition,
\be
\begin{array}{l}
{\bf x}_2={\cal R}_1({\bf x}_1), \\
{\bf x}_3={\cal R}_2({\bf x}_2), \\
{\bf x}_3={\bf x}_1,
\end{array}
\ee
where ${\cal R}_1$ and ${\cal R}_2$ vectorial rational functions.
The first two equations can be interpreted as time-forward and time-backward processes. The last
equation identifies the initial state of the forward process with the final state of the backward
process. The overall process can be seen also as a deterministic process in a time loop.
This analogy is suggestive, since retro-causality is considered one possible explanation
of quantum weirdness. Can a suitable break of causality allow
for a description of quantum processes in a classical framework? To be physically interesting,
this break should not lead to a computational power beyond the power of quantum computers,
otherwise a fine tuning of the theory would be necessary to conceal, in a physical process,
much of the power allowed by the causality break. A similar fine tuning is necessary
if, for example, quantum non-locality is explained with superluminar interactions. These classical
non-local theories need an artificial fine tuning to account for non-signaling of quantum theory.
\subsection{$(n-1)/3$ parameters at most}
In the previous subsection, we have built a class of curves defined by systems of $n-1$ polynomial
equations such that about half of the variables can be efficiently solved over a finite field as
functions of the remaining variables. These curves and the polynomial $P_0$ have $2^n$ rational
intersection points. From a different perspective (discarding about $n/2$ polynomials), we have
found a parametrizable variety with about
$n/2$ parameters such that its intersection with some hypersurface has $2^n$ rational points.
In this subsection, we show that the number of parameters can be dropped to about $n/3$ so
that about $2n/3$ variables can be efficiently eliminated, at least. In the following, we consider
space dimensions $n$ such that $n-1$ is a multiple of $3$. Let us define the integer
\be
n_1\equiv\frac{n-1}{3}.
\ee
Let us define the rational numbers $A_i,B_i,\bar A_i,\bar B_i$, $W_i$, and $c_i$ with $i\in\{1,\dots,n\}$
as a solution of the equations
\be
\begin{array}{c}
A_i B_i=W_i,\;\;\; \bar A_i \bar B_i= W_i, \\
A_i \bar B_i+\bar A_i B_i=2 c_i W_i, \\
\sum_{i=1}^n c_i^{k-1} W_i=0 \;\;\;\; k\in\{1,\dots,n-1\}, \\
i\ne j \Rightarrow c_i\ne c_j.
\end{array}
\ee
The procedure for finding a solution has been given previously.
We define the polynomials
\be
P_s=\sum_i^n c_i^{s-1} \hat a_i \hat b_i, \;\;\;\; s\in\{1,\dots,n\}.
\ee
The linear functions $\hat a_i$ and $\hat b_i$ are defined by the $n-1$ linear equations
\be
\begin{array}{l}
\sum_{i=1}^{n_1} c_i^{k-1}(A_i \hat b_i+ B_i \hat a_i)=0, \;\;\; k\in\{1,\dots,n_1\} \\
\sum_{i=1}^{n_1} c_i^{k-1}(\bar A_i \hat b_i+\bar B_i \hat a_i)=0, \;\;\; k\in\{1,\dots,2 n_1\}.
\end{array}
\ee
These equations uniquely determine $\hat a_i$ and $\hat b_i$, up to a linear transformation
of the variables $x_i,\dots,x_{n+1}$.
Up to a linear transformation, we have
\be
\begin{array}{l}
\frac{\partial (\hat a_i,\hat b_i)}{\partial x_k}=c_i^{k-1} (\bar A_i,\bar B_i), \;\;\; k\in\{1,\dots,n_1\} \\
\frac{\partial (\hat a_i,\hat b_i)}{\partial x_{k+n_1}}=c_i^{k-1} (A_i,B_i), \;\;\; k\in\{1,\dots,n_1\}.
\end{array}
\ee
Since there are rational points in the curve, there is another variable, say $x_{2n_1+1}$, such that
the second derivative $\partial^2 P_1/\partial x_{2n_1+1}^2$ is equal to zero.
Using the above equations, we have
\be
\left.
\begin{array}{l}
\frac{\partial P_s}{\partial x_k}=0, \;\;\; k\in\{1,\dots,2 n_1-s+1\}, \\
\frac{\partial^2 P_s}{\partial x_k^2}=0, \;\;\; k=2 n_1-s+2.
\end{array}
\right\} \;\;\; s\in\{1,\dots,2 n_1+1 \}.
\ee
Thus, the first $2n_1+1=\frac{2n+1}{3}$ polynomials take the triangular form~(\ref{poly_affine}),
up to a reorder of the indices. These polynomials define a parametrizable variety with $(n-1)/3$
parameters. Stated in a different way, there is a curve and a hypersurface such that their
intersection contains $2^n$ points and at least $(2n+1)/3$ coordinates of the points in the curve
can be evaluated efficiently given the value of the other coordinates. It is possible
to show that all the intersection points are in the parametrizable variety, that is,
they satisfy the third of Conditions~(\ref{gauss_constrs}).
\section{Conclusion and perspectives}
\label{conclusion}
In this paper, we have reduced prime factorization to the search of rational
points of a parametrizable variety $\cal V$ having an arbitrarily large number
$N_P$ of rational points in the intersection with a hypersurface $\cal H$.
To reach a subexponential factoring complexity, the number of parameters $M$
has to grow sublinearly in the space dimension $n$. In particular,
If $N_P$ grows exponentially in $n$ and $M$ scales as a sublinear power
of $n$, then the factoring complexity is polynomial (subexponential) if
the computation of a rational point in $\cal V$, given the parameters, requires
a number of arithmetic operations growing polynomially (subexponentially)
in the space dimension. Here, we have considered a particular kind of
rational parametrization. A set of $M$ coordinates, say $x_{n-M+1},\dots,x_n$,
of the points in $\cal V$ are identified with the $M$ parameters,
so that the first $n-M$ coordinates are taken equal to rational functions of
the last $M$ coordinates. In particular, the parametrization is expressed in a
triangular form. The $k$-th variable is taken equal to a rational function
${\cal R}_k={\cal N}_k/{\cal D}_k$ of the variables $x_{k+1},\dots,x_{n}$,
with $k\in\{1,\dots,n-M\}$. That is,
\be\label{triang_par}
\begin{array}{l}
x_k={\cal R}_k(x_{k+1},\dots,x_n), \;\;\; k\in\{1,\dots,n-M\},
\end{array}
\ee
which parametrize a variety in the zero locus of the $n-M$ polynomials,
\be\label{triang_poly_form}
P_k={\cal D}_k x_k-{\cal N}_k, \;\;\;\; k\in\{1,\dots,n-M\}.
\ee
To reach polynomial complexity,
there are two requirements on these polynomials. First, they have to contain
a number of monomials scaling polynomially in $n$, so that the computation of
${\cal R}_k$ is efficient. For example, we could require that the degree is
upper-bounded by some constant. Second, their zero locus has to share an
exponentially large number of rational points with some hypersurface $\cal H$
(a superpolynomial scaling $N_P\sim e^{b\,n^\beta}$ with $0<\beta<1$ is actually
sufficient, provided
that the growth of $M$ is sufficiently slow).
The hypersurface is the zero locus of some polynomial $P_0$. Also the
computation of $P_0$ at a point has to be efficient.
We have proposed a procedure for building pairs $\{{\cal V},{\cal H}\}$
satisfying the two requirements. First, we define the set of $N_P$ rational
points. This set can depend on some coefficients. Since $N_P$ has to grow
exponentially
in the dimension, we need to define them implicitly as common zeros of
a set of polynomials, say $G_1,G_2,\dots$.
The simplest way is to take $G_k$ as
products of linear functions, like the polynomials~(\ref{quadr_polys}).
These polynomials generate an ideal $I$. The relevant
information on the generators
is encoded in a satisfiability formula in conjunctive normal form without
negations and a linear matroid. We have called these two objects a model.
Second, we search for $n-M$ polynomials
in $I$ with the triangular form~(\ref{triang_poly_form}). These polynomials
always exist. Thus, the task is to find a solution such that the polynomials
contain as few monomials as possible.
This procedure is illustrated with the simplest example. The generators
are taken equal to reducible quadratic polynomials of the
form~(\ref{quadr_polys}), whose associated algebraic set
contains $2^n$ rational points. We search for polynomials $P_k$
of the form $\sum_i c_i G_i$ with $c_i$ constant. First, we prove
that there is no solution for $M=1$ and space dimension greater than $4$.
Then, we find a solution for $M=(n-1)/3$. If there are solutions with
$M$ scaling sublinearly in $n$, then a factoring algorithm with polynomial
complexity automatically exists, since the computational complexity of
the rational functions ${\cal R}_k$ is polynomial by construction.
The existence of such solutions is left as an open problem.
This work can proceed in different directions. First, it is necessary
to investigate whether the studied model admits solutions with
a much smaller set of parameters. The search has been performed in
a subset of the ideal. Thus, if these solutions do not exist,
we can possibly expand this subset (if it is sufficiently large, there
is for sure a solution, but the polynomial complexity of ${\cal R}_k$ is
not guaranteed anymore). We could also
relax other hypotheses such as the distinguishibility
of each of the $2^n$ rational points and their membership of
the parametrizable variety. More general ideals are another option.
In this context, we have shown that classes of models can be reduced
to smaller classes by preserving the computational complexity
of the associated factoring algorithms.
This reduction makes the search space smaller. It is interesting to
determine what is the minimal class of models obtained by this
reduction. This is another problem left open.
Apart from the search of better inputs of the procedure, there is a
generalization of the procedure itself. The variety $\cal V$ has
the parametrization~(\ref{triang_par}).
However, there are more general parametrizable
varieties which can be taken in consideration.
It is also interesting to investigate if there is some
deeper relation with retro-causality, time loops and, possibly, a
connection with Shor algorithm. Indeed, in the
attempt of lowering the geometric genus of one of the non-parametrizable
curves derived
in the previous section, we found a set of solutions for the coefficients
over the cyclotomic number field, so that the resulting polynomials have
terms taking the form of a Fourier transform. Quantum Fourier transform
is a key tool in Shor's algorithm. This solution ends up to break
the curve into the union of an exponential large number of parametrizable
curves, thus it is not useful for our purpose. Nonetheless, the Fourier-like
forms in the polynomials remains suggestive.
Finally, the overall framework has some interesting relation with the
satisfiability problem.
Using a particular matroid, we have seen that there is a one-to-one correspondence
between the points of an algebraic set and the solutions of a satisfiability
formula (including also negations). To prove that a formula is satisfiable is
equivalent to prove that a certain algebraic set is not empty. This mapping
of SAT problems to an algebraic-geometry problem turns out to be a generalization
of previous works using the finite field $\mathbb{Z}_2$, see for example
Ref.~\cite{hung}. It can be interesting to investigate whether part of the machinery
introduced here can be used for solving efficiently some classes of SAT formulae.
\section{Acknowledgments}
This work was supported by Swiss National Science Foundation (SNF) and Hasler foundation
under the project n. 16057.
|
\section{Conceptual Model}\label{sec:funcx}
We first describe the conceptual model behind \emph{FSys}{}
to provide context to the implementation architecture.
\emph{FSys}{} allows users to register and then execute \emph{functions}
on arbitrary \emph{endpoints}.
All user interactions with \emph{FSys}{} are performed via a
REST API implemented by a cloud-hosted \emph{FSys}{} service.
\textbf{Functions:}
\emph{FSys}{} is designed to execute \emph{functions}: snippets of Python code
that perform some activity. A \emph{FSys}{} function explicitly defines
a Python function and input signature.
The function body must specify all imported modules. While \emph{FSys}{} supports only Python functions, users can easily write Python functions to invoke tools written in other languages.
Listing~\ref{lst:ssx} shows several functions used in the SSX pipeline mentioned in \S\ref{sec:requirements}. The \texttt{process\_stills} function takes a single input dictionary as input, which includes the locations of the images and the
\textit{phil} file describing the analysis configuration. The function then makes use of the \textit{DIALS}~\cite{winter2018dials} tool to analyze the image.
\textbf{Function registration:}
A function must be registered with \emph{FSys}{} before it can be executed.
The registration includes a name and the serialized function body.
Optionally, it
may also specify users, or groups of users, who may be authorized to invoke the function, and a container image to be used for execution. Containers
allow the construction of
environments with the
dependencies
(system packages and Python libraries) required to execute the function.
\emph{FSys}{} assigns a universally
unique identifier (UUID) for management and invocation. Users may update
functions they own.
\textbf{Endpoints:}
A \emph{FSys}{} endpoint is a logical entity that represents a compute resource.
The corresponding \emph{FSys}{} agent allows the \emph{FSys}{} service to dispatch
functions to that resource for execution. The agent handles
authentication and authorization, provisioning of nodes on the
compute resource, and monitoring and management. Administrators
or users can deploy a \emph{FSys}{} agent and register an endpoint for themselves and others,
providing descriptive (e.g., name, description) metadata.
Each endpoint is assigned a UUID for subsequent use.
\textbf{Function execution:}
Authorized users may invoke a registered function on a selected endpoint.
To do so, they issue a request via \emph{FSys}{} that
identifies
the function and endpoint to be used as well as inputs
to be passed to the function.
Functions are executed asynchronously: each invocation returns an
identifier via which progress may be monitored and results retrieved.
\emph{In this paper, we refer to an invocation of a function as a ``task.''} Importantly, following the FaaS model, while users must specify the specific endpoint for use, they do not manage the resources on which the function is executed (e.g., nodes, containers, or modules)
\textbf{\emph{FSys}{} service:}
Users interact with \emph{FSys}{} via a cloud-hosted service that
exposes a REST API for registering functions and
endpoints, and for executing functions, monitoring their execution, and retrieving results.
The REST API provides a uniform interface via which users can
make asynchronous and stateless calls to manage endpoints and function executions. REST APIs are the the most common interface for FaaS platforms (e.g., AWS Lambda~\cite{amazonlambda} and Google Cloud Functions~\cite{googlecloudfunctions}).
The service is connected to
accessible endpoints via the endpoint
registration process.
\textbf{User interface:}
\emph{FSys}{} provides a Python SDK that wraps the REST API.
Listing~\ref{lst:ssx} shows an example of how the SDK can be used
to register and invoke a function on a specific endpoint.
The example first constructs a \emph{client} and registers the \texttt{process\_stills} function.
It then invokes the registered function using the \texttt{run}
command, passing the unique function identifier,
the endpoint id on which to execute the function, and inputs
(in this case \texttt{data}).
Finally, the example shows that the results can be asynchronously retrieved
using \texttt{get\_result}.
\section{Architecture and Implementation}\label{sec:arch}
\emph{FSys}{} combines a cloud-hosted management service with software agents
deployed on remote resources: see~\figurename{~\ref{fig:arch}}.
\begin{figure}[h]
\vspace{-0in}
\includegraphics[width=\columnwidth,trim=0.2in 0.1in 1.6in 0in,clip]{Figures/funcx_arch_4.png}
\caption{\emph{FSys}{} architecture, showing the \emph{FSys}{} service (top) consisting of a REST interface, Redis store, and Forwarders. \emph{FSys}{} endpoints (bottom) provision resources and coordinate the execution of functions.
}
\label{fig:arch}
\vspace{-0in}
\end{figure}
\subsection{The \emph{FSys}{} Service}
The \emph{FSys}{} service maintains a registry of \emph{FSys}{} endpoints,
functions, and users in a persistent AWS Relational Database Service (RDS) database.
To facilitate rapid function dispatch, \emph{FSys}{} stores serialized function codes
and tasks (including inputs and task metadata) in an AWS ElastiCache Redis hashset.
The service also manages a Redis queue for each endpoint that stores task IDs for
tasks to be dispatched to that endpoint.
The service exposes a REST API to register and manage endpoints,
register functions, execute and monitor functions, and retrieve the output from tasks.
The \emph{FSys}{} service is secured using Globus Auth~\cite{GlobusAuth},
which allows users, programs and applications, and \emph{FSys}{} endpoints
to securely make API calls.
When an endpoint registers with the \emph{FSys}{} service,
a unique \emph{forwarder} process is created for each
endpoint.
Endpoints establish secured ZeroMQ connections with their
forwarder to receive tasks, return results, and perform heartbeats.
\emph{FSys}{} implements a hierarchical task queuing architecture consisting
of queues at the \emph{FSys}{} service, endpoint, and worker. These queues support
reliable fire-and-forget function execution that is resilient to failure
and intermittent endpoint connectivity.
At the first level,
each registered endpoint is allocated a unique Redis \emph{task queue}
and \emph{result queue} to store and track tasks, which are implemented using Redis \texttt{Lists} structure.
We use Redis as it provides a simple yet performant system for brokering tasks. Redis is offered as a hosted Amazon service and can be elastically scaled as workload increases. \emph{FSys}{} serves primarily as a broker to manage and distribute tasks. Redis provides high throughput queuing via an in-memory store with little overhead on the tasks and results passed through the queue---an important requirement for providing low latency execution. One limitation of this approach is that we must implement message acknowledgments to ensure that tasks and results are communicated reliably between clients, endpoints, and the \emph{FSys}{} service. We note that as use cases expand, we may need to consider other message queues, such as Kafka~\cite{kafka}, Pulsar~\cite{pulsar}, or AMQP-based systems (e.g., RabbitMQ~\cite{rabbitmq}).
\figurename~\ref{fig:taskpath} shows the \emph{FSys}{} task lifecycle.
At function submission, the \emph{FSys}{} service routes the task to the specified endpoint's task queue.
The forwarder listens to the queue for tasks and then
dispatches the task to the corresponding agent.
\emph{FSys}{} agents internally queue tasks at both the manager and worker.
These queues ensure that tasks are not lost once they
have been delivered to the endpoint.
Results are returned to the \emph{FSys}{} service and stored in the endpoint's
result queue until they are retrieved by the user.
\begin{figure}[h]
\centering
\includegraphics[width=0.88\columnwidth]{Figures/funcx_submit_path.png}
\caption{\emph{FSys}{} task execution path. A task submitted to \emph{FSys}{} (1) is stored in Redis (2), queued for execution (3), and dispatched via a Forwarder to an endpoint (4); results are returned (5), then stored in Redis for users to retrieve (6).
}
\label{fig:taskpath}
\end{figure}
\emph{FSys}{} relies on AWS-hosted databases, caches, and Web serving infrastructure to reduce operational overhead,
elastically scale resources, and provide high availability.
While these services provide significant benefits to \emph{FSys}{}, they have associated costs.
To minimize these costs we apply several techniques, such as
using small cloud instances with responsive
scaling to minimize the steady state
cost
and restricting the size of input and output data
passed through the
\emph{FSys}{} service to reduce storage (e.g., in Redis store) and data egress costs.
Further, we periodically purge results from the Redis store
once they have been retrieved by the client or after a period
of time.
The \emph{FSys}{} service is designed to provide robustness and fault tolerance using several techniques. First, the \emph{FSys}{} service implements health checks, including a liveness check, CPU utilization, and response time monitoring, etc. The service is automatically restarted when these health checks indicate failures. Second, the RDS database and Redis task queue are replicated to ensure that any data (e.g., users, functions, endpoints, and tasks) are not lost. Third, the forwarder uses configurable periodic heartbeats (30 seconds by default) to detect if an agent is disconnected. A task is sent to an agent only when it is connected. When an agent is disconnected, all the tasks dispatched to the agent are returned back into the task queue and the tasks are forwarded to that agent when it reconnects. Fourth, tasks are cached at each layer and only removed when downstream layers have acknowledged receipt.
Finally, to serve the distributed endpoints in \emph{FSys}{}, the \emph{FSys}{} service is deployed on cloud-hosted services, which internally provide high reliability and robustness.
\subsection{Function Containers}
\emph{FSys}{} uses containers to package function code and dependencies that are to be deployed
on a compute resource.
Our review of container technologies, including Docker~\cite{merkel2014docker}, LXC~\cite{LXC}, Singularity~\cite{kurtzer2017singularity}, Shifter~\cite{jacobsen2015contain}, and CharlieCloud~\cite{priedhorsky2017charliecloud},
led us to adopt Docker, Singularity, and Shifter.
Docker works well for local and cloud deployments, whereas
Singularity and Shifter are designed for use in HPC environments
and are supported at large-scale computing facilities (e.g., Singularity at ALCF and Shifter at NERSC).
Singularity and Shifter implement similar models and thus
it is easy to convert from a common representation (e.g., a Dockerfile)
to both formats.
\emph{FSys}{} requires
that each container includes a base set of software, including Python~3
and \emph{FSys}{} worker software. Other system libraries or Python
modules needed for function execution must also be included.
When registering a function, users may optionally specify a container to be used for execution;
if no container is specified, \emph{FSys}{} executes functions using the worker's Python environment.
In future work, we intend to make this process dynamic, using repo2docker~\cite{repo2docker}
to build Docker images and convert them to site-specific container formats
as needed.
\subsection{The \emph{FSys}{} Endpoint}\label{sec:executor}
The \emph{FSys}{} endpoint represents a remote resource and
delivers high-performance remote execution of
functions in a secure, scalable, and reliable manner.
The endpoint architecture, depicted in the lower portion of \figurename{~\ref{fig:arch}},
is comprised of three components, which are discussed below:
\begin{itemize}
\item \emph{\emph{FSys}{} agent}: a persistent process that queues and forwards tasks and results,
interacts with resource schedulers, and load balances tasks.
\item \emph{Manager}: manages the resources for a single compute node on an endpoint by deploying and managing a set of workers.
\item \emph{Worker}: executes tasks (optionally within a container).
\end{itemize}
The \emph{\emph{FSys}{} agent} is a software agent that is deployed by a user on a
compute resource (e.g., an HPC login node, cloud instance, or a laptop).
It registers with the \emph{FSys}{} service and acts
as a conduit for routing tasks and results between the service and workers.
The \emph{FSys}{} agent manages resources on its system
by working with the local scheduler or cloud API to deploy
\emph{managers} on compute nodes.
The \emph{FSys}{} agent uses a pilot job model~\cite{turilli2018comprehensive} to provision
and communicate with resources in a uniform manner,
irrespective of the resource type (cloud or cluster) or local resource manager (e.g., Slurm, PBS, Cobalt).
As each manager is launched on a compute node, it connects to and registers with the \emph{FSys}{} agent.
The \emph{FSys}{} agent then uses ZeroMQ sockets to communicate
with its managers.
To minimize blocking, all communication is performed by threads
using asynchronous communication patterns.
The \emph{FSys}{} agent uses a randomized scheduling algorithm to allocate tasks
to suitable managers with available capacity.
The \emph{FSys}{} agent can be configured to provide
access to specialized hardware or accelerators.
When deploying the agent, users can specify how worker
containers should be launched, enabling them to mount
specialized hardware and execute functions on that hardware.
In future work, we will
extend the agent configuration to specify custom hardware and
software capabilities and report
this information to the \emph{FSys}{} agent and service for scheduling.
To provide fault tolerance and robustness, for example with respect to
node failures, the \emph{FSys}{} agent relies on periodic heartbeat messages and a
process to detect lost managers. The \emph{FSys}{} agent tracks
tasks that have been distributed to managers so that when failures
do occur, lost tasks can be re-executed (if permitted).
\emph{FSys}{} agents communicate with the \emph{FSys}{} service's forwarder
via a ZeroMQ channel.
Loss of a \emph{FSys}{} agent is detected by the forwarder and when
the \emph{FSys}{} agent recovers, it repeats the registration
process to acquire a new forwarder and continue receiving tasks.
To reduce overheads, the \emph{FSys}{} agent can shut down managers to release resources when they are not
needed, suspend managers to prevent further tasks from being scheduled to them,
and monitor resource capacity to aid scaling decisions.
\emph{Managers} represent, and communicate on behalf of, the
collective capacity of the workers on \emph{a single node}, using just two sockets per node.
Managers determine
the available CPU and memory resources on a node, and partition the
node among the workers.
Managers advertise deployed container
types and available capacity to the endpoint.
\emph{Workers} persist on a node (optionally within containers) and each executes one task
at a time. Since workers have a single responsibility,
they use blocking communication to wait for tasks from the manager.
Once a task is received, it is deserialized, executed,
and the serialized results are returned via the manager.
\subsection{Managing Compute Infrastructure} \label{sec:provider}
\emph{FSys}{} is designed to support a range of computational resources, from
embedded computers to clusters, clouds, and supercomputers,
each with distinct access modes.
As \emph{FSys}{} workloads are often sporadic, resources must be provisioned and deprovisioned
as needed to reduce costs due to idle resources.
\emph{FSys}{} uses Parsl{}'s provider interface~\cite{babuji19parsl}
to interact with various resources, specify resource-specific
requirements (e.g., allocations, queues, limits, cloud instance types),
and define rules for automatic scaling (i.e., limits and scaling aggressiveness).
This interface allows \emph{FSys}{} to be deployed on batch schedulers
such as Slurm, PBS, Cobalt, SGE, and Condor;
major cloud systems
such as AWS, Azure, and Google Cloud;
and Kubernetes.
\subsection{Serialization}
\label{sec:serialization}
\emph{FSys}{} supports the registration of arbitrary Python functions and the
passing of data (e.g., primitive types and complex objects) to/from those functions.
\emph{FSys}{} uses a Facade interface with
several serialization libraries (including pickle, dill, and JSON) as some Python object types cannot be serialized with some serialization libraries, and no single serialization library can serialize all objects.
The \emph{FSys}{} serializer sorts the serialization libraries by speed and applies
them in order successively until the object is successfully serialized.
This approach combines the strengths of various libraries, including
support for complex objects (e.g., machine learning models) and traceback objects in a fast and transparent fashion.
Once objects are serialized, they are packed into buffers with headers that include routing tags and the serialization method,
such that only the buffers need to be unpacked and deserialized at the destination.
\subsection{Batching}
\emph{FSys}{} supports two batching to amortize costs
across many function requests:
internal batching enables managers to request many tasks on behalf of their workers, minimizing network communication costs;
and, user-facing \emph{batching} that
enables users to define batches of function inputs,
allowing users to trade off
efficient execution and increased
per-function latency by creating fewer, larger
requests. The SDK includes a matching batch
interface for retrieving the results of many
tasks concurrently.
\subsection{Security Model}
\emph{FSys}{} requires a security model to ensure that functions are executed on endpoints by authenticated and authorized users and that one function cannot interfere with another.
\textbf{Authentication and authorization:}
Since \emph{FSys}{} endpoints may be deployed across arbitrary resources, we first summarize authentication and authorization requirements.
\begin{itemize}
\item Different research CI may rely on diverse identity management systems and authentication models (e.g., two-factor authentication). To ease the deployment of \emph{FSys}{} agent on any resources, \emph{FSys}{} needs a general model that provides a uniform API, rather than maintaining a set of APIs for the diverse identity providers.
\item Users may have to use different accounts (e.g., institution accounts, national CI credentials, or national laboratory accounts) to access different resources. Users would like to use one account to authenticate \emph{FSys}{} endpoints infrequently.
\item One frequent use case in scientific computing is that resources are shared among a group of scientists. Ideally, the authorization model should enable users to grant access to others while enforcing secure delegation.
\end{itemize}
\emph{FSys}{} uses Globus Auth~\cite{GlobusAuth} for identity and access management (IAM), and protection of all APIs.
We use Globus Auth as it satisfies the above requirements, is widely adopted in scientific community, implements standard protocols (e.g., OAuth~2), enables simple delegation (e.g., such that a user may allow the \emph{FSys}{} service or another user to access their endpoint),
and offers a flexible OAuth client model
for developing the \emph{FSys}{} SDK.
Although Globus Auth is used as the primary implementation, other IAM services that provide similar capabilities and interfaces could be integrated with \emph{FSys}{}.
The \emph{FSys}{} service is registered as a Globus Auth \emph{resource server}, allowing users to
authenticate using a supported identity (e.g., institution, Google, ORCID)
and enabling various OAuth-based authentication flows (e.g., native client)
for different scenarios.
\emph{FSys}{} has associated Globus Auth scopes
(e.g., ``urn:globus:auth:scope:funcx:register\_function'')
via which other clients (e.g., applications and services)
may obtain authorizations for programmatic access.
\emph{FSys}{} endpoints are themselves Globus Auth native clients, each dependent on
the \emph{FSys}{} scopes, which are used to securely connect to the \emph{FSys}{} service.
Endpoints require the administrator to authenticate prior to registration
in order to acquire access tokens used for constructing API requests.
The connection between the \emph{FSys}{}
service and endpoints is established using ZeroMQ. Communication addresses
are sent
as part of the registration process. Inbound
traffic from endpoints to the cloud-hosted service is limited to known IP addresses.
\textbf{Isolation:}
\emph{FSys}{} function execution can be isolated in containers to ensure that functions
cannot access data or devices outside their context.
To enable fine-grained tracking of execution,
we store execution request histories in the \emph{FSys}{} service
and in logs on \emph{FSys}{} endpoints.
\section{Data Management}\label{sec:data_management}
Data management is essential for many
applications:
functions may interact with large and/or remote datasets, and tasks may use the outputs of other tasks as inputs.
This section describes how data can be staged and managed between different \emph{FSys}{} endpoints (\emph{inter-endpoint}) and between different functions within an endpoint (\emph{intra-endpoint}).
\subsection{Inter-endpoint Data Transfers}
To minimize operational costs and performance overheads we limit the size of
data that can be passed through the \emph{FSys}{} service to 10~MB. To enable functions to be seamlessly invoked with large data that may be located on remote
computers, we require an out-of-band data transfer mechanism.
We summarize the primary requirements as follows.
\begin{itemize}
\item Transfers can be managed programmatically by \emph{FSys}{}.
\item The transfer mechanism should be natively supported and approved by the administrations of research CI.
\item Transfers should be optimized and provide high performance, endpoint-to-endpoint movement.
\item The transfer mechanism should be interoperable with \emph{FSys}{}'s authentication and authorization model (i.e., Globus Auth)
to secure data transfers on behalf of users.
\item The transfer mechanism should allow a user's functions to fetch data that is shared among a group.
\end{itemize}
\begin{lstlisting}[style=PythonStyle, caption={Inter-endpoint data transfer with Globus. \label{lst:globus}}]
from funcx import GlobusFile
data = GlobusFile(globus_endpoint_id='65e...',
file_path='/~/file.txt')
task_id = fc.run(func_id, endpoint_id,
remote_data=data)
\end{lstlisting}
We focus on wide area data management, rather than cloud storage, as data may be stored or generated in different locations (e.g., instruments, campus clusters, supercomputers) in many scientific use cases.
Based on the requirements above, we integrate Globus transfer~\cite{chard14efficient} to streamline inter-endpoint data transfers.
Globus has several advantages that lead us to this choice: i) it is a research data management service that provides high-performance data transfers between arbitrary storage resources, such as supercomputers, laptops, and clouds; ii) it is widely deployed on research CI and used in scientific research; iii) data are transferred directly between the source and destination systems via the GridFTP~\cite{allcock2005globus} protocol;
iv) it provides a Python SDK that allows a user's functions to fetch shared data.
To use Globus, the Globus Connect software
must be installed on the storage system, this is often done by administrators installing Globus
Connect Server on large clusters or it can be done individually in user-space using Globus Connect Personal.
Storage systems are registered as a Globus endpoint with associated authentication mechanism
in the Globus service. Each endpoint is given a unique endpoint identifier that is used when transferring data.
In this paper, we extend \emph{FSys}{} to allow for references to Globus-accessible files to be passed as input/output
to/from a function. Specifically, users must specify the Globus endpoint and the path
to the file on that endpoint.
When Globus-accessible files are passed
to/from a \emph{FSys}{} function,
\emph{FSys}{} can automatically stage data either prior to, or after invocation of the function.
An example
of using Globus for inter-endpoint data transfer is shown in Listing~\ref{lst:globus}.
We have found that Globus is well suited for our current use cases; however, other mechanisms (e.g., HTTP, FTP, and rsync) could also be used for inter-endpoint data transfers by augmenting functions to make direct data downloads or uploads. In future work we will extend the inter-endpoint transfer model in \emph{FSys}{} to transparently support these mechanisms as we have done in Parsl.
\subsection{Intra-endpoint Data Transfers}\label{sec:intra}
Modern applications may involve frequent fine-grained communications
among functions executed on an individual endpoint (i.e., intra-endpoint data transfers).
For example, distributed machine learning (ML) training may require that state be coordinated
among all worker nodes; and MapReduce-style applications often
involve a shuffle phase where every map task sends data to every reduce task.
Here we describe the advantages and disadvantages of potential intra-endpoint data management approaches.
\begin{itemize}
\item \textbf{A shared file system} that can be accessed by every worker on an endpoint. The effort to attach such storage to a \emph{FSys}{} endpoint is minimal, as many clusters, clouds, and supercomputers provide built-in shared file system (sharedFS) or object storage. However, they often have high access cost, limited IO performance, and high latency when writing and reading many files.
\item \textbf{MPI} is a message passing fabric that is highly scalable and optimized for data communications on supercomputers with specialized interconnects;
however, MPI libraries are not natively available or optimized for many computers (e.g., clouds and private clusters). More importantly, the synchronous nature of MPI's collective communication is not well-suited for the asynchronous task-based model in \emph{FSys}{}, as it blocks tasks from making progress even when partial results are ready, which is important for many performance-driven asynchronous applications (e.g., distributed machine learning training); HPC containers often must be adapted to make use of local MPI libraries; and a failure of one MPI process may cause other MPI processes to block, which stops other tasks from continuing. We note that fault tolerance has improved in the recent release of MPI 4.0; however, this is not commonly deployed at the time of writing.
\item \textbf{Socket and socket-like connections} (e.g., ZeroMQ) between workers can
provide high throughput and low latency direct data transfers. However, creating pair-wise connections
between workers is expensive and in some cases workers (e.g., in containers) may not be network addressable or may not have sufficient open ports to support connections between all workers.
\item \textbf{In-memory data stores} (e.g., MemCache~\cite{nishtala2013scaling} and Redis~\cite{redis}) provide
higher IOPS and lower latency than shared file systems and support more data types than socket connections (e.g., serialized data). However, they require that storage be provisioned explicitly, that additional services be hosted, and they cannot match the raw throughput or latency of direct socket connection~\cite{li2019socksdirect}.
\end{itemize}
The aforementioned advantages and disadvantages lead us to select the shared file system and in-memory data store (Redis) approaches to support intra-endpoint data transfers in \emph{FSys}{}, as these approaches are both general and are readily available (or can be deployed with minimal effort) on most target resources.
We present a preliminary performance study of these four approaches in \S\ref{sec:data-eval} and the results show that the performance of shared file system and Redis is similar to the other approaches, especially when transferring large volumes of data.
We have extended the \emph{FSys}{} agent such that users may specify a requirement for a Redis cluster to be deployed alongside their endpoint. The \emph{FSys}{} SDK provides a general
interface to retrieve the Redis client which users can interact with, as shown in Listing~\ref{lst:redis}.
\begin{lstlisting}[style=PythonStyle, caption={Intra-endpoint data transfer with Redis. \label{lst:redis}}]
def example(key, data):
from funcx_endpoint import get_redis_client
rc = get_redis_client()
rc.set(key, data)
rc.get(key)
\end{lstlisting}
\section{Container Management}\label{sec:routing}
\emph{FSys}{} uses containers to provide customized execution environments
for functions irrespective of the endpoint's host environment.
In this section, we discuss how the \emph{FSys}{}
agent spawns containers to serve functions, retains warm containers, routes functions to containers, and scales resources based on function requirements.
\subsection{Container Warming}\label{sec:warm}
Commercial FaaS platforms~\cite{wang2018peeking}
keep function containers \emph{warm} by leaving them running for a short period of time (e.g., 5-15 minutes)
following the execution of a function.
Warm containers remove the need to instantiate a new container to execute a function,
significantly reducing latency.
We argue that this need is especially important in HPC environments for several reasons.
First, containers and Python environments (e.g., conda) are generally stored on shared file systems of HPC systems.
Therefore, starting many containers and Python environments concurrently for the workers at the HPC scale may impose significant stress on the shared file systems. Second, many HPC centers implement their own methods for instantiating containers that place
limitations on the number of concurrent requests.
Third, individual cores are often
slower in many-core architectures like Xeon Phis.
As a result, the start time for containers
can be much larger than what would be seen on a PC, as shown in \tablename~\ref{extractor-tab} in \S\ref{sec:eval-routing}.
In \emph{FSys}{}, container warming is implemented by the \emph{FSys}{} agent. To reduce the number of container cold starts, the \emph{FSys}{} agent keeps a container warm until there are insufficient resources available to process pending workloads or the container has been idle for a configurable period of time (e.g., 10 minutes).
The \emph{FSys}{} agent is extensible to support other container-warming strategies, such as releasing the least recently used container and application-agnostic container warming~\cite{shahrad2020serverless} if necessary.
\subsection{Warming-aware Function Routing}
Ideally we aim to minimize the number of container cold starts due to the
cost of starting a container in HPC environments.
To do so, the \emph{FSys}{} agent needs to know which computing nodes have warm containers and what types of warm containers, so that it can route the function tasks to the appropriate warm containers.
The \emph{FSys}{} agent employs a hierarchical, warming-aware scheduling algorithm to route function tasks to workers to optimize throughput. The \emph{FSys}{} agent determines which functions to route to a given manager, and each \emph{manager} determines how to launch and spawn containers to satisfy the arriving workload.
Thus, warming-aware routing involves coordination between managers and \emph{FSys}{} agent.
Each manager advertises its deployed container types and its available resources to the \emph{FSys}{} agent.
Based on the advertised information of each manager,
the \emph{FSys}{} agent implements a warming-aware scheduling algorithm to route tasks
to managers. Specifically, when receiving a task with requirement for a specific container type, the scheduler attempts to send the task to a manager that has a suitable warm container.
When there are multiple available managers with the required container type warm, priority is given to the one with the most available container workers to balance load across managers.
If there are not any warmed containers on any connected managers, the \emph{FSys}{} agent chooses one manager at random to execute the task.
While we use random scheduling in our implementation, other scheduling policies, such as bin-packing and round-robin, could also be used.
To amortize network latency during manager advertising and task dispatching, the \emph{FSys}{} agent also supports \emph{prefetching}, which allows a manager to prefetch a configurable number of additional tasks beyond its current availability.
Upon receiving a set of tasks, the manager determines the required container types and dynamically starts (and stops) containers to serve tasks
in a fair manner: we set the number of deployed containers for a function type to be proportional to the number of received tasks of this function type. For instance, if 30\% of the tasks a manager receives are of type A and the manager can spawn at most 10 containers, the manager will spawn 3 containers of type A.
It is worth mentioning that the function routing is different when an endpoint is deployed on a Kubernetes cluster. Both the manager and its workers are deployed within a container pod that can only serve one type of function. Hence, in this case, each manager is deployed with a
specific container image and the agent simply needs to route
tasks to corresponding managers.
We apply relatively simple scheduling algorithms here to
demonstrate the benefits of warming aware routing; however, the \emph{FSys}{} agent implements modular scheduling interfaces for function routing (at \emph{FSys}{} agents) and container deployment (by managers) which enabling
different algorithms (e.g., priority-aware or deadline-driven scheduling)
to be implemented by users.
We note that when task duration is much larger than the container cold start time, the benefits of warming-aware routing are limited.
\subsection{Elastic Resource Provisioning}
One of the main benefits of the FaaS computing model is elasticity.
To provide elasticity on a \emph{FSys}{} endpoint, a \emph{FSys}{} agent dynamically provisions
resources via an extensible provisioning \emph{strategy} interface.
The strategy interface consists of a monitoring and a scaling component within the \emph{FSys}{} agent. The monitoring component interacts periodically (e.g., every second) with the provider interface (introduced in~\S\ref{sec:provider}) and the \emph{FSys}{} agent to fetch the current endpoint load, including the active and idle resources (i.e., the number of container workers) and the number of pending function requests.
Based on the monitoring information, the scaling component automatically provisions more resources when the number of function requests is greater than the number of idle resources, and releases resources that have been idle for some period of time, via the provider interface. The maximum idle time is set to two minutes by default, but is user-configurable for each endpoint.
Similar to commercial FaaS platforms such as AWS Lambda and Azure Functions, the \emph{FSys}{} strategy allows users to configure the minimum and maximum resources to be used, as well as how aggressively a \emph{FSys}{} agent scales those resources (e.g., request one more resource when there are ten waiting requests).
However, elasticity may be subject to resource request delays, such as the time to request a new instance on a cloud or
to provision a resource via an HPC scheduler.
\section{Conclusion}\label{sec:conclusion}
\emph{FSys}{} is a distributed FaaS platform that is designed to support the unique
needs of research
computing. Unlike existing centralized FaaS platforms,
\emph{FSys}{} combines a reliable and easy-to-use cloud-hosted
interface with the ability to securely execute functions on user-deployed \emph{FSys}{}
endpoints deployed on various remote computing resources.
\emph{FSys}{} supports many HPC systems and cloud platforms, can use three container
technologies, and can expose access to heterogeneous and specialized computing resources.
In this paper we extend \emph{FSys}{} to support inter-endpoint and intra-endpoint data transfers between functions, and optimize function execution performance with advanced container management and warming-aware function routing mechanisms.
We showed that \emph{FSys}{} agents can
scale to execute 1M tasks over \num{130000} concurrent workers
when deployed on the Cori supercomputer.
We also showed that \emph{FSys}{}'s data transfer mechanisms are comparable to alternative methods, and that they can significantly improve application performance.
Finally, we showed that \emph{FSys}{} can dynamically route functions to workers to reduce container warming overhead
and that batching can significantly reduce overheads.
\emph{FSys}{} demonstrates the advantages of adapting the FaaS model to create
a federated computing ecosystem. Based on early experiences using \emph{FSys}{} in
scientific case studies~\cite{chard2020funcx},
we have found that the approach
provides several advantages,
including abstraction, code simplification, portability, scalability, and sharing; however,
we also identified several limitations including suitability for some applications,
conflict with current allocation models, and challenges decomposing
applications into functions.
We hope that \emph{FSys}{} will serve as a flexible platform for research
computing while also enabling new studies in
function scheduling,
dynamic container management, and data management.
In future work,
we will continue our work to explore new scheduling approaches that can select appropriate endpoints for function execution and manage data dependencies between functions. We also plan to provide APIs that allow users to manage and discover functions and endpoints.
We will extend \emph{FSys}{}'s container management capabilities to
create containers dynamically based on function requirements,
and to stage containers to endpoints on-demand.
We will also explore techniques for optimizing performance, for example by
sharing containers among functions with similar dependencies and developing
resource-aware scheduling algorithms.
\section{Experiences with \emph{FSys}{}} \label{sec:discussion}
As of August, 2022 \emph{FSys}{} has been used by 413 users to perform over 19.8 million function invocations, \num{338105} functions have been registered, and 4027 endpoints have been created.
Here we describe our experiences applying \emph{FSys}{} to various scientific case studies.
\textbf{AI-enabled steering of computational campaigns:}
Colmena~\cite{ward2021colmena} is an open-source library that enables researchers to build complex, AI-directed HPC campaigns.
Researchers can implement flexible decision-making policies to steer different tasks (e.g., simulation, model update, and model inferences) of computational campaigns. When tasks are generated, \emph{FSys}{} serves as an execution backend to distribute and execute tasks. The FaaS model of \emph{FSys}{} and the implementation of container management allows Colmena to flexibly dispatch tasks to arbitrary computing resources, enabling ML-enhanced tasks to be sent to GPU-accelerated devices and high throughput simulations to HPC clusters. The integration of data management mechanisms (e.g., Globus and Redis) in \emph{FSys}{} enables data to move between Colmena entities transparently without requiring the user to manage movement; further, it can improve performance and simplify distributed, data-intensive campaigns.
\textbf{Linking instruments and HPC:}
\emph{FSys}{} has been used to combine several experimental instruments with HPC infrastructure~\cite{vescovi2022linking}.
This approach allows scientists to offload computationally-intensive analysis tasks to HPC resources, simplifies large-scale parallel processing for large data rates, and enables online analysis. Such experimental techniques, including serial synchrotron crystallography~\cite{sherrellfixed2022}, X-ray photon correlation spectroscopy~\cite{perakis2020towards}, ptychography~\cite{bicer2021high}, and scientific machine learning~\cite{liu2022braggnn},
depend on orchestration of various activities in various locations. For this purpose, these examples use Globus flows~\cite{chard2022globus} to create complex sequences of actions. For example, when data are acquired from an experiment, run quality control at the edge, move data to an HPC center, run analysis and reconstruction algorithms, and index resulting images in a data catalog. \emph{FSys}{} provides the compute substrate enabling many of these actions to be executed in various locations. Integration of Globus for data management simplifies dispatching tasks to different resources without requiring changes to broader workflows to transfer and retrieve inputs and results. Further, scientific analysis toolkits are often containerized to promote portability and exploit available resources.
The container warming features presented in this paper
enable these workloads to reduce cold starts, which can be costly on large, shared file systems and facilitate rapid computation---a necessary feature to support real-time computation and experiment steering. We report function execution and data transfer characteristics in prior work~\cite{vescovi2022linking}.
\textbf{AlphaFold as a Service:}
AlphaFold~\cite{jumper2021highly} is a cutting edge machine learning model that can predict a protein's 3D structure from its amino acid sequence. AlphaFold has garnered significant interest in the bioinformatics community, with applications in the development of therapeutics and accelerating the practice of deriving crystal structures at light sources. However, AlphaFold relies on powerful GPUs and
large reference datasets, restricting access for many researchers. To address these challenges, the Argonne Leadership Computing Facility deployed AlphaFold as a service using \emph{FSys}{} to dynamically provision GPU resources on-demand. In this work, containers are dispatched to GPU nodes and managed by the \emph{FSys}{} endpoint to serve inference requests. As AlphaFold tasks can take over an hour to process, the Globus integration with \emph{FSys}{} provides the ability to asynchronously transfer results to users.
\textbf{Distributed ML:}
Flox~\cite{kotsehub22flox} is a federated learning (FL) framework that decouples model training and inference from infrastructure management.
Flox uses \emph{FSys}{} to enable users to
train and deploy FL models on one or more remote computers, and in particular on edge devices. We are applying these techniques to Rural AI applications~\cite{ruralai2022}, using \emph{FSys}{} to
facilitate training and deployment of models in remote locations.
Rural AI requires reliable task and result transmission
as devices are deployed in
rural settings where device and network outages are common,
and the quality of wireless networks varies depending on location.
\emph{FSys}{}'s hierarchically
designed queues support the necessary robustness to dispatch (and queue) tasks across rural devices. Containers enable execution of tasks on heterogeneous
edge devices for training and on centralized cloud instances for model
aggregation.
\section{Evaluation}\label{sec:evaluation}
We evaluate the performance of \emph{FSys}{} in terms of latency, scalability, and throughput. We also study the effects of batching, function routing, and data transfer approaches.
\subsection{Latency}
We explore \emph{FSys}{} latency by instrumenting the system.
Figure~\ref{fig:lat_breakdown} shows latencies for a warm
container as follows:
$t_s$: Web service latency to authenticate, store the task in Redis,
and append the task to an endpoint's queue;
$t_f$: forwarder latency to read task from the Redis store, forward
the task to an endpoint, and write the result to the Redis store;
$t_e$: endpoint latency to receive tasks and send results to the forwarder,
and to send tasks and receive results from the worker; and
$t_w$: function execution time.
The endpoint was deployed on ANL's Cooley cluster for this test and had an 18 ms latency on average to the forwarder.
We observe that $t_w$ is fast relative to the overall system latency.
The network latency between service and forwarder includes minimal
communication time due to internal AWS networks (measured at $<$1 ms).
Most \emph{FSys}{} overhead is found
in $t_s$ due to
authentication, and in $t_e$ due to internal queuing and
dispatching.
We note here that the aim of \emph{FSys}{} is not to build yet another low-latency FaaS platform, but instead to provide a new federated model in which functions can be executed on arbitrary remote machines. Nevertheless, in our previous work we showed that the latency of \emph{FSys}{} is comparable to commercial FaaS platforms, such as AWS Lambda, Google Cloud Functions, and Azure Functions~\cite{chard2020funcx}.
\begin{figure}[h]
\includegraphics[width=0.97\columnwidth]{Figures/latency_breakdown.png}
\vspace{-0.1in}
\caption{\emph{FSys}{} latency breakdown for a container.}
\label{fig:lat_breakdown}
\vspace{-0.1in}
\end{figure}
\begin{figure*}[h]
\vspace{-0in}
\centering
\includegraphics[width=0.8\textwidth,trim=0.07in 0.08in 0.07in 0.08in,clip]{Figures/strong-weak.pdf}
\vspace{-0.1in}
\caption{Strong and weak scaling of the \emph{FSys}{} agent.}
\label{fig:scalability}
\vspace{-0.1in}
\end{figure*}
\subsection{Scalability and Throughput}
We study the strong and weak scaling of the \emph{FSys}{} agent on
ANL's Theta and NERSC's Cori supercomputers.
Theta is a 11.69-petaflop system based on the second-generation
Intel Xeon Phi ``Knights Landing" (KNL) processor.
Its \num{4392} nodes each have a 64-core processor
with 16 GB MCDRAM, 192 GB of DDR4 RAM, and are interconnected with high speed InfiniBand.
Cori is a 30-petaflop system with an Intel Xeon ``Haswell" partition
and an Intel Xeon Phi KNL partition.
We ran our tests on the KNL partition, which
has \num{9688} nodes,
each with a 68-core processor (with 272 hardware threads)
with six 16-GB DIMMs, 96 GB DDR4 RAM,
interconnected in a
Dragonfly topology.
We perform experiments using 64 Singularity containers on each Theta node
and 256 Shifter containers on each Cori node.
Due to a limited allocation on Cori we use the four hardware threads
per core to deploy more containers than cores.
Strong scaling evaluates performance when the total number of function invocations is fixed;
weak scaling evaluates performance when the average number of functions executed on each container is fixed.
To measure scalability we created functions of various durations:
a 0-second ``no-op'' function that exits immediately, a 1-second ``sleep'' function, and a 1-minute CPU ``stress'' function that keeps a CPU core at 100\% utilization.
For each case, we measured completion time of a batch of functions as we increased the total number of containers.
Notice that the completion time of running $M$ ``no-op'' functions on $N$ workers indicates the overhead of \emph{FSys}{}
to distribute the $M$ functions to $N$ containers.
Due to limited allocations
we did not execute sleep or stress
functions on Cori, nor did we
execute stress functions for strong scaling on Theta.
We pre-warmed all containers in these experiments.
\subsubsection{Strong scaling}
Figure~\ref{fig:scalability}(a) shows the completion time of \num{100000}
\emph{concurrent} function requests with an increasing number of containers.
On both Theta and Cori, the completion time decreases as the number of containers
increases, until we reach 256 containers for ``no-op'' and 2048 containers for ``sleep'' on Theta.
As reported by Wang et al.~\cite{wang2018peeking} and Microsoft~\cite{azureFunctionsDocs},
Amazon Lambda achieves good scalability for a single function to more
than 200 containers, Microsoft Azure Functions
can scale up to 200 containers, and Google Cloud Functions does not scale well beyond 100 containers.
While these results do not indicate the maximum number of containers that can be used
for a single function, and likely include per-user limits imposed by the platform,
our results show that \emph{FSys}{} scales similarly to commercial platforms.
\subsubsection{Weak scaling}
We performed \emph{concurrent} function
requests such that each container receives, on average, 10 requests.
Figure~\ref{fig:scalability}(b) shows weak scaling for ``no-op,'' ``sleep,'' and ``stress.''
For ``no-op," the completion time increases with more containers on both Theta and Cori.
This reflects the time required to distribute requests to all of the containers.
On Cori, \emph{FSys}{} scales to \num{131072} concurrent containers and
executes more than 1.3 million ``no-op'' functions.
Again, we see that the completion time for ``sleep'' remains close to constant
up to 2048 containers, and the completion time for ``stress'' remains
close to constant up to \num{16384} containers.
Thus, we expect a function with a several-minute duration would scale well to many more containers.
\subsubsection{Throughput}
We observe a maximum throughput for a \emph{FSys}{} agent (computed as number of function requests divided
by completion time) of \num{1694} and \num{1466} requests per second on Theta and Cori, respectively.
\subsubsection{Summary}
Our results show that \emph{FSys}{} agents
i) scale to \num{130000}+ containers for a single function;
ii) exhibit good scaling performance up to approximately \num{2048} containers for a
1-second function and \num{16384} containers for a 1-minute function;
and iii) provide similar scalability and throughput using both Singularity and
Shifter containers on Theta and Cori. It is important to note that these experiments study the
\emph{FSys}{} agent, and not the end-to-end throughput of \emph{FSys}{}.
While the \emph{FSys}{} web service can elastically scale
to meet demand,
communication overhead may limit throughput. To address this challenge and
amortize communication overheads, we enable batch submission of tasks.
These optimizations are discussed in \S\ref{sec:opt}.
\subsection{Data Management}\label{sec:data-eval}
We evaluate four potential approaches for intra-endpoint data transfers (described in \S\ref{sec:intra}) on Theta. We use mpi4py for MPI, ZeroMQ for direct socket connections, Redis for the in-memory store, and Theta's Lustre shared file system. We note that we use mpi4py because it supports direct Python object transfers and previous work~\cite{saxena2018evaluation} has shown that mpi4py does not add significant overhead when compared with OpenMPI in terms of throughput and latency for data transfers.
We emulate different communication patterns (i.e., point-to-point, broadcast, and all-to-all) and vary data transfer size.
\figurename~\ref{fig:redis-broadcast} shows the performance of these four approaches with different communication patterns. As expected, MPI performs the best, and sharedFS the worst.
However, ZeroMQ and Redis achieve similar performance to MPI. As
data volume increases, the performance difference between the four approaches diminishes, as transfer time is mainly determined by available network bandwidth (which is the same for all approaches).
While sharedFS and Redis perform slightly worse than MPI for small data sizes, we adopt them in \emph{FSys}{} because of their generality, ease-of-use, and the challenges of using mpi4py (as well as MPI compiled in C) and ZeroMQ described in \S\ref{sec:intra}.
\begin{figure}[h]
\vspace{-0in}
\includegraphics[width=0.9\columnwidth,trim=0.07in 0.08in 0.07in 0.08in,clip]{Figures/redis-broadcast.pdf}
\vspace{-0.1in}
\caption{Performance of the four intra-endpoint transfer approaches. Top: point-to-point; middle: broadcast to 20 nodes; bottom: all-to-all on 20 nodes.}
\label{fig:redis-broadcast}
\vspace{-0.1in}
\end{figure}
We now evaluate intra-endpoint data management in the context
of MapReduce applications and a real-world science application
deployed on funcX.
\subsubsection{MapReduce}
To demonstrate how Redis and sharedFS can facilitate intra-endpoint data transfers for real applications, we deployed a \emph{FSys}{} endpoint with a three-node Redis cluster. We also used the shared Lustre file system on Theta. We deployed two MapReduce applications: WordCount and Sort. These applications involve an all-to-all communication pattern between map and reduce tasks (i.e., data shuffling).
Each application processes 30~GB of Wikipedia text data, and has 300 map and 300 reduce tasks, requiring communication of 90000 data chunks in total.
\tablename~\ref{tab:mr-redis} shows the average completion time of each task spent in each phase of the MapReduce application: input read, map process, intermediate write, intermediate read, reduce process, output write, when using Redis and sharedFS approaches for data shuffling. WordCount benefits less from Redis than Sort as WordCount shuffles just one tenth of the data. The table shows that Redis can speed up the data shuffling phase of the workload (i.e., intermediate write and read) by up to 3x.
\begin{table}[h]
\caption{Average completion time of the transfer phases in WordCount and Sort when using Redis and shared file system for intra-endpoint data management.}
\label{tab:mr-redis}
\begin{tabular}{lll|ll}
& \multicolumn{2}{c|}{\textbf{WordCount (s)}} & \multicolumn{2}{c}{\textbf{Sort (s)}} \\
\multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{Redis}} & \multicolumn{1}{c|}{\textbf{SharedFS}} & \multicolumn{1}{c}{\textbf{Redis}} & \multicolumn{1}{c}{\textbf{SharedFS}} \\ \hline
Intermediate write & 3.55 & 8.15 & 3.27 & 5.32 \\
Intermediate read & 33.39 & 43.40 & 11.37 & 41.77 \\
\end{tabular}
\end{table}
Note that \tablename~\ref{tab:mr-redis} shows the average task completion time.
The benefits of Redis over sharedFS on the total completion time of a MapReduce application may depend on the amount of parallelism and the portion of the shuffle phase over the other phases.
For example, the total completion time of Sort with Redis and sharedFS are
220 and 520 seconds, respectively, yielding a 55.7\% improvement. The WordCount application runs in 1800 seconds and 2200 seconds, respectively, yielding a 18.2\% improvement. This is because Sort has a heavier shuffle phase than WordCount.
\subsubsection{Colmena}
Finally, we evaluate intra-endpoint data management in the context of a real-world scientific application to demonstrate the benefits of Redis over sharedFS. Colmena~\cite{ward2021colmena} is a framework that manages large-scale, AI-directed steering of computational campaigns (e.g., to efficiently explore large molecular spaces when designing new materials). A Colmena application consists of a \emph{Thinker} that implements the decision-making policy to generate new tasks (e.g., new simulation, new model training, or model inference), a \emph{Task Server} that dispatches task requests to resources and manages task results, and \emph{Workers} that are deployed on compute resources to execute tasks.
These components exchange data (e.g., task requests and results)
with Redis used to facilitate transfers.
We implement a Colmena benchmark with 1000 tasks, each with 1 MB input and 1 MB output data. Table~\ref{tab:colmena} shows the average completion time of the communication stages in Colmena.
Redis yields a lower completion time for all communication stages compared to sharedFS. Such a benefit has been demonstrated to be particularly important when running Colmena at scale with thousands of tasks.
\begin{table}[h]
\centering
\caption{Average completion time of the transfer phases in the Colmena benchmark when using Redis and shared file system for intra-endpoint data management.}
\label{tab:colmena}
\begin{tabular}{c c c}
\textbf{Stage} & \textbf{Redis (ms)} & \textbf{SharedFS (ms)}\\ \hline
Input data write from Thinker & 7.15 & 32.31 \\
Input read on Workers & 0.70 & 11.36 \\
Result write from Workers & 18.04 & 244.72 \\
Result read from Task Server & 0.11 & 3.50 \\
\end{tabular}
\end{table}
\begin{figure*}[h]
\includegraphics[width=0.97\textwidth]{Figures/routing-throughput.pdf}
\vspace{-0.1in}
\caption{Completion time of warming-aware and non-warming-aware routing.}
\label{fig:routing-throughput}
\vspace{-0.2in}
\end{figure*}
\begin{figure*}[h]
\includegraphics[width=0.97\textwidth]{Figures/routing-switches.pdf}
\vspace{-0.1in}
\caption{Number of container cold starts of warming-aware and non-warming-aware routing.}
\label{fig:routing-switch}
\vspace{-0.1in}
\end{figure*}
\subsection{Function Routing}\label{sec:eval-routing}
Before exploring function routing performance, we first quantify the instantiation cost of various container technologies on different resources.
Specifically, we measure the time taken to start a container and execute a Python command to import \emph{FSys}{}'s worker module---a
requirement prior to executing a \emph{FSys}{} function.
We deploy the containers on an AWS EC2 \texttt{m5.large} instance and on compute nodes on Theta and Cori following the facility's documented
best practices.
\tablename~\ref{table:cold_start_cost}
shows the results.
We speculate that the significant performance deterioration of container instantiation on HPC systems can be attributed
to a combination of slower clock speed on KNL nodes and shared file system contention when fetching images.
These results highlight the need to apply function warming approaches to reduce overheads on HPC systems.
\begin{table}[h]
\caption{Cold container instantiation time for different container technologies on different resources.}
\label{extractor-tab}
\vspace{-0.1in}
\begin{center}
\begin{tabular}{l l c c c}
\textbf{System} & \textbf{Container} & \textbf{Min (s)} & \textbf{Max (s)} & \textbf{Mean (s)} \\
\hline
Theta & Singularity & 9.83 & 14.06 & 10.40 \\
Cori & Shifter & 7.25 & 31.26 & 8.49 \\
EC2 & Docker & 1.74 & 1.88 & 1.79 \\
EC2 & Singularity & 1.19 & 1.26 & 1.22 \\
\end{tabular}
\end{center}
\vspace{-0.1in}
\label{table:cold_start_cost}
\end{table}
We evaluate \emph{FSys}{}'s function routing strategy and show that it improves
overall throughput as well as reducing the number of container cold starts.
We deployed an endpoint on Theta and compared the performance of warming-aware routing and randomized (non-warming-aware) routing.
The endpoint is allocated 10 nodes and each node can host 10 workers, each with its own container.
We registered 10 functions, where each function requires a specific container (i.e., 10 different containers.)
We submitted a batch of requests, each of which is chosen from one of the ten functions uniformly at random.
\figurename~\ref{fig:routing-throughput} and \figurename~\ref{fig:routing-switch} show the overall function completion time and the number of container cold starts for different batch sizes and for different function durations (0, 1, 5, and 20 seconds).
We note that the number of requests in a batch is much higher than the available resources (100 container workers) in this experiment, and thus a container worker is more likely to be killed to serve other request when using non-warming-aware routing.
Thus, the warming-aware routing reduces completion time by up to 61\% for a batch of requests (i.e., higher throughput) and reduces the number of container cold starts significantly (e.g., 22 cold starts for 3000 functions), compared to the randomized routing strategy. This is because the warming-aware algorithm attempts to reuse the warm containers as much as possible to reduce the overhead of container instantiation.
As expected, the benefit of warming-aware routing gradually diminishes as the function duration increases, because the function runtime, rather than the cold container instantiation time, becomes dominant.
\subsection{Batching}
\label{sec:opt}
To evaluate the effect of executor-side batching, we submit \num{10000} concurrent ``no-op'' function requests
and measure the completion time when executors can request one function at a time (batching disabled) vs
when they can request many functions at a time based on the number of idle containers (batching enabled).
We use 4 nodes (64 containers each) on Theta.
We observe that the completion time with batching enabled is 6.7~s (compared to 118s when disabled).
\section{Motivations and Requirements} \label{sec:requirements}
Over the last two years the scientific community has been working to understand
SARS-CoV-2 and develop effective tests, therapeutics, and vaccines.
However, progress in these areas is dependent on our ability
to understand SARS-CoV-2 protein structures.
At Argonne's Advanced Photon Source~\cite{kim2020crystal}, scientists use an emerging method called fixed-target serial synchrotron crystallography (SSX) to collect physiological temperature data from thousands of protein crystals.
Data are generated at unprecedented rates with tens of thousands of images captured each hour. Keeping pace with the experiment requires rapid data processing across multiple, heterogeneous
computing resources to efficiently analyze, refine, solve, and
curate structures.
To meet these
data processing and publication needs, SSX scientists have adopted an automated
data management framework~\cite{Wilamowski2020} that can manage data acquisition,
analysis, curation, and visualization.
Throughout this workflow, there are needs for computation both
at the edge to detect and pre-process data rapidly, as well as on
HPC resources to perform computationally expensive analysis tasks and produce structures.
\emph{Each of these steps relies on different packages and functions, has different processing
durations, occurs at different times, and requires different types and amounts of computing resources.
Thus, it is essential that the scientists be able to decompose the entire processing pipeline
into a series of individual functions to perform on data as they are moved and transformed.}
These functions, shown in Listing~\ref{lst:ssx}, analyze
individual images, refine and solve the crystal structure,
and extract metadata and create plots before publishing results.
\begin{lstlisting}[style=PythonStyle, caption={Three functions used in the SSX pipeline and an example of how the funcX SDK is used to register and invoke the \texttt{process\_stills} function. \label{lst:ssx}}]
def process_stills(data):
inputs = data['inputs']
phil = data['phil']
cmd = f'dials.stills_process {phil} {inputs}'
res = subprocess.run(cmd)
return res.stdout
def solve(data):
from gladier.tools import template_prime
pdata = template_prime.substitute(data['template'])
cmd = f"prime.run {pdata} > prime.log"
res = subprocess.run(cmd)
return res.stdout
def extract_metadata(data):
from gladier.tools import (get_dims,
get_lattice_counts, plot_lattice_counts)
lc = get_lattice_counts(xdim, ydim, int_files)
plot_name = f'1int-sinc.png'
plot_lattice_counts(xdim, ydim, lc, plot_name)
return plot_name
fc = FuncXClient()
func_id = fc.register_function(process_stills)
endpoint_id = '863d-...-d820d'
input_data = {'inputs': '...', 'phil': '...'}
task_id = fc.run(func_id, endpoint_id,
data=input_data)
res = fc.get_result(task_id)
\end{lstlisting}
This typical science use case, with parallels in many
other domains described in our previous work~\cite{chard2020funcx},
highlights the benefits of
FaaS approaches (e.g., decomposition, abstraction,
flexibility, scalability, reliability),
and also elucidates several requirements for FaaS approaches.
\begin{itemize}
\item \textbf{Research CI}:
functions may require HPC-scale and/or specialized and heterogeneous resources. Many resources expose batch scheduler interfaces (with long delays, periodic downtimes, proprietary interfaces) and specialized container technology (e.g., Singularity, Shifter) that make it challenging to provide common execution interfaces, on-demand and elastic capacity, and fault tolerance.
\item \textbf{Distribution}: different parts of an application may be most efficiently executed on different, often distributed, resources (e.g., near data, on a specialized computer).
\item \textbf{Data}: functions analyze both small and large data, stored in various locations and formats, and accessible via different methods (e.g., Globus~\cite{chard14efficient}).
\item \textbf{Authentication}: institutional identities and specialized security models are used to access data, compute resources, and other cyberinfrastructure.
\item \textbf{State}: functions may be connected and share state (e.g., files or database connections) to decrease overheads.
\end{itemize}
Existing FaaS solutions may satisfy these requirements partially, but not completely.
For example, some FaaS systems (e.g., OpenWhisk~\cite{openwhisk}, KNIX~\cite{knix}) support on-premise deployments on specialized hardware (e.g, GPU), but not on distributed and federated computing resources.
Some FaaS systems (e.g., DFaaS~\cite{ciavotta2021dfaas}, ChainFaaS~\cite{ghaemi2020chainfaas}) support function execution in distributed environments, but not on research CI.
Here we present \emph{FSys}{}, a federated and scalable FaaS platform that enables researchers to decompose applications into functions and execute them on arbitrary remote computers via the FaaS paradigm.
\section{Introduction}}
\IEEEPARstart{T}{he} exponential growth of data and increasing hardware diversity is driving the need for computation to occur wherever it makes the most sense, for example, on a suitable computer, where particular software is available, or near data.
Prior research, in grid~\cite{Foster2001} and peer-to-peer~\cite{milojicic2002peer} computing,
has studied and explored the foundations for remote computing.
However, with the exception of cloud platforms, general-purpose remote computation has remained elusive due to, for example, slow and unreliable network communications, security challenges, and dependencies between software and heterogeneous computer architectures.
Commercial cloud providers have been at the forefront of
recent advances in networks, hardware, and distributed computing, leveraging
widespread virtualization, universal trust fabrics,
and high-speed networks to deliver serverless computing
services such as function-as-a-service (FaaS)~\cite{baldini2017serverless,fox2017status,varhese19cloud}.
FaaS enables developers to register a high-level programming
function and to then invoke that function many times
by passing input arguments. The user needs not concern
themselves with provisioning infrastructure or
configuring execution environments.
FaaS systems have quickly become integral to a wide
range of applications, particularly for event-based
and dev-ops applications.
The FaaS model is particularly attractive in science
as a way of decomposing monolithic science applications into
a collection of modular, performant, and extensible functions~\cite{foster2017cloud,spillner2017faaster,malawski2016towards,fox2017conceptualizing,kiar2019serverless}.
However, existing FaaS systems are typically centralized and specific to a particular cloud, rather than being designed to be deployed
on heterogeneous research cyberinfrastructure (CI) or to use federated resources.
Typically research CI uses batch scheduling interfaces and inflexible authentication and authorization models,
which does not lend itself to the fine-grain and sporadic function workloads. In response, we propose a federated FaaS model for general-purpose remote computing at scale across diverse CIs, both centrally and at the edge.
In this paper, we present \emph{FSys}{}, a federated, scalable,
and high-performance function execution platform.
\emph{FSys}{} leverages a \emph{distributed endpoint model} to support remote function execution across distributed and heterogeneous research CI.
Users can transform many computing resources, such as laptops, clouds, clusters, supercomputers, or Raspberry Pis they are authorized to access, into function serving systems by deploying \emph{FSys}{}'s endpoint software.
Users then use the cloud-hosted \emph{FSys}{} service to register Python functions and invoke those functions on their deployed endpoints.
\emph{FSys}{} manages the reliable and secure execution of functions, staging function code and inputs, provisioning resources, managing safe and secure
execution (optionally in containers), monitoring execution, and returning outputs to users. Thus, users benefit from the convenience
and reliability of a cloud-hosted service combined with the
flexibility and performance of a federated ecosystem of endpoints.
We extend our previous work~\cite{chard2020funcx}
to support complex data dependencies between scientific functions.
Specifically, we focus on enabling
data transfer between functions that are executing
on the same (\textit{intra-endpoint}) or different
(\textit{inter-endpoint}) endpoints.
For intra-endpoint communication we use an in-memory data store, for inter-endpoint communication we use Globus~\cite{foster2011globus}.
We also present new heuristic-based container management and function routing schemes that reduce container warming overhead and
efficiently route functions to appropriately configured containers.
The primary novelty of our work is in the adaptation of the FaaS paradigm to a federated research ecosystem, combining a distributed endpoint model with a hosted FaaS platform to support remote function execution across distributed and heterogeneous research CI.
We demonstrate the viability of our approach with a highly modularized and extensible design as well as a scalable and performant implementation. We also show that it is beneficial to decompose scientific applications into monolithic functions that may be executed on different remote resources. The contributions of our work are as follows:
\begin{itemize}
\item \emph{FSys}{}, a distributed and federated FaaS platform that can:
be deployed on research CI, dynamically provision and manage
resources, leverage various container technologies, and facilitate
secure, scalable, and distributed function execution.
\item Automated data movement between functions using widely-used in-memory data stores and high-performance data transfer technology to transparently support data dependencies between functions.
\item Design and evaluation of performance enhancements for function
serving on distributed research CI, including function warming,
batching, and function routing.
\item Experimental studies showing that \emph{FSys}{} delivers execution
latencies comparable to those of commercial FaaS platforms
and scales to 1M+ functions across 130K active workers on supercomputers.
\end{itemize}
The rest of this paper is as follows.
\S\ref{sec:requirements} describes an example use case and presents general requirements for FaaS in science.
\S\ref{sec:funcx} presents a conceptual model of \emph{FSys}{}.
\S\ref{sec:arch} describes the \emph{FSys}{} system architecture.
\S\ref{sec:data_management} discusses how data is managed in \emph{FSys}{}.
\S\ref{sec:routing} presents \emph{FSys}{}'s container management model.
\S\ref{sec:evaluation} evaluates \emph{FSys}{} performance.
\S\ref{sec:discussion} reviews \emph{FSys}{}'s use in scientific case studies.
\S\ref{sec:survey} discusses related work.
Finally, \S\ref{sec:conclusion} summarizes our contributions.
\section*{Acknowledgment}
This work was supported in part by NSF 2004894/2004932 and Laboratory Directed Research and
Development funding from Argonne National Laboratory under U.S. Department of
Energy under Contract DE-AC02-06CH11357. This work also used resources of the
Argonne Leadership Computing Facility and Center for Computational Science and Engineering at Southern University of Science and Technology.
\bibliographystyle{abbrv}
\section{Requirements} \label{sec:requirements}
Our work is guided by the unique requirements of FaaS in science.
To illustrate these requirements we present six representative case studies:
scalable metadata extraction, machine learning inference as a service, synchrotron serial crystallography, neuroscience, correlation spectroscopy, and high energy physics.
\figurename~\ref{fig:use-cases} shows function execution time distributions for each case study.
These short duration tasks exemplify opportunities for FaaS in science.
We summarize by highlighting requirements for FaaS in science.
\begin{figure}[ht]
\includegraphics[width=1\columnwidth,trim=0.08in 0.1in 0.08in 0.1in,clip]{Figures/use-cases}
\vspace{-0.15in}
\caption{Distribution of latencies for 100 function calls, for each of the six case studies described in the text.
\label{fig:use-cases}
\vspace{-0.1in}
}
\end{figure}
\textbf{Metadata extraction:}
The effects of high-velocity data expansion are making it increasingly
difficult to organize, discover, and manage data.
Edge file systems and data
repositories now store petabytes of data
which are created and modified at an alarming rate~\cite{paul17monitoring}.
Xtract{}~\cite{skluzacek19serverless}
is a distributed metadata extraction system
that applies a set of
general and specialized metadata extractors, such as those for identifying topics in text, computing aggregate values from tables, and recognizing locations in maps.
To reduce data transfer costs, Xtract{} executes extractors ``near'' to
data, by pushing extraction tasks to the edge. Extractors are implemented
as Python functions, with various dependencies, and each extractor typically executes for between 3 milliseconds and 15 seconds.
\textbf{Machine learning inference:}
DLHub{}~\cite{chard19dlhub}
is a service that supports ML model publication
and on-demand inference.
Users deposit ML models, implemented as functions with a set of
dependencies, in the DLHub{} catalog
by uploading the raw model (e.g., PyTorch, TensorFlow) and model
state (e.g., training data, hyperparameters).
DLHub{} uses this information to dynamically create a container for the model using repo2docker~\cite{repo2docker} that contains all model dependencies and necessary model
state.
Other users may then invoke the model through DLHub{} on arbitrary input
data. DLHub{} currently includes more than one hundred published models,
many of which have requirements for specific ML toolkits and execute
more efficiently on GPUs and accelerators.
\figurename~\ref{fig:use-cases} shows the execution time when invoking the
MNIST digit identification model. Other DLHub{} models execute for between seconds and
several minutes.
\textbf{Synchrotron Serial Crystallography (SSX)}
is an emerging method for imaging small crystal samples 1--2 orders of magnitude faster than other methods.
To keep pace with the increased data production, SSX researchers require automated
approaches that can process the resulting data with great rapidity:
for example, to count the \emph{bright spots} in an image
(``stills processing'') within seconds, both for quality control
and as a first step in structure determination. The DIALS~\cite{waterman2013dials} crystallography processing tools are implemented
as Python functions that execute for
1--2 seconds per sample.
Analyzing large datasets requires HPC resources to derive crystal structures
in a timely manner.
\textbf{Quantitative Neurocartography} and connectomics map the
neurological connections in the brain---a compute- and data-intensive process
that requires processing \textasciitilde 20GB every minute during experiments.
Researchers apply automated workflows
to perform quality control on raw images (to validate that the instrument
and sample are correctly configured), apply ML models
to detect image centers for subsequent reconstruction, and generate
preview images to guide positioning. Each of these steps is implemented
as a function that can be executed sequentially with some data exchange
between steps. However, given the significant data sizes, researchers typically
rely on HPC resources and are subject to scheduling delays.
\textbf{Real-time data analysis in High Energy Physics (HEP)}:
The traditional HEP analysis model uses successive processing steps to reduce the initial dataset (typically, 100s of PB) to a size that permits real-time analysis. This iterative approach requires significant computation time and storage of large intermediate datasets, and may take weeks or months to complete. Low-latency, query-based analysis strategies~\cite{pivarski2017} are being developed to enable real-time analysis
using native operations on hierarchically nested, columnar data.
Such queries are well-suited to FaaS.
To enable interactive analysis, for example as a physicist engages in real-time analysis of several billion particle collision events, successive compiled functions, each running for seconds, need to be dispatched to the data. Analysis needs require sporadic (and primarily remote) invocation, and compute needs increase as new data are collected.
\textbf{X-ray Photon Correlation Spectroscopy (XPCS)}
is an experimental technique used
to study the dynamics in materials at nanoscale by identifying correlations in
time series of area detector images.
This process involves analyzing the pixel-by-pixel correlations
for different time intervals.
The detector acquires megapixel frames at 60 Hz (\textasciitilde 120 MB/sec).
Computing correlations at these data rates is a challenge that requires HPC resources but also rapid response time. Image processing functions, such as XPCS-eigen's corr function,
execute for approximately 50 seconds, and images can be processed in parallel.
\textbf{Requirements for FaaS in science:}
The case studies illuminate benefits of
FaaS approaches (e.g., decomposition, abstraction,
flexibility, scalability, reliability), but also
elucidate requirements unmet by existing FaaS solutions:
\begin{itemize}
\item \textbf{Specialized compute:} functions may require HPC-scale and/or specialized and heterogeneous resources (e.g., GPUs).
\item \textbf{Distribution:} functions may need to execute near to data and/or on a specialized computer.
\item \textbf{Dependencies:} functions often require specific libraries and user-specified dependencies.
\item \textbf{Data:} functions analyze both small and large data, stored in various locations and formats, and accessible via different methods (e.g., Globus~\cite{chard14efficient}).
\item \textbf{Authentication:} institutional identities and specialized security models are used to access data and compute resources.
\item \textbf{State:} functions may be connected and share state (e.g., files or database connections) to decrease overheads.
\item \textbf{Latency:} functions may be used in online (e.g., experiment steering) and interactive environments (e.g., Jupyter notebooks) that require rapid response.
\item \textbf{Research CI:} resources offer batch scheduler interfaces (with long delays, periodic downtimes, proprietary interfaces) and specialized container technology (e.g., Singularity, Shifter) that make it challenging to provide common execution interfaces, elasticity, and fault tolerance.
\item \textbf{Billing:} research CI use allocation-based usage models.
\end{itemize}
\section{A Brief Survey of FaaS}\label{sec:survey}
\kyle{Need to focus more on differences: distributed/multi-site execution, reliable execution on remote EPs, different resource providers, function targeting specific resource (e.g., endpoint X), we can use accelerators, data, scale (I expect the others do not scale that far).}
\kyle{Existing systems designed for kubernetes where you get auto responses, here we had
to do a lot to work with batch systems, provision nodes, scale nodes, deal with unreliability and unknown SLAs}
FaaS platforms have proved wildly successful in industry as a way
to reduce costs and the need to manage infrastructure.
Here we present a brief survey of FaaS platforms, summarized in \tablename{~\ref{tab:survey}}.
We broadly categorize platforms as \emph{commercial}, \emph{open source}, or \emph{academic}.
\definecolor{Gray}{gray}{0.9}
\newcolumntype{a}{>{\columncolor{Gray}}c}
\begin{table*}[]
\caption{Taxonomic survey of common FaaS platforms.} \label{tab:survey}
\vspace{-1ex}
\footnotesize
\begin{tabular}{|p{2.7cm}|p{4.5cm}|p{2.7cm}|p{2.3cm}|}
\hline
\textbf{} & \textbf{Intended Infrastructure} & \textbf{Virtualization} & \textbf{Max Walltime (s)} \\ \hline
\textbf{Amazon Lambda} & Public cloud, Edge (Greengrass) & Firecracker (KVM) & 900 \\ \hline
\textbf{Google Cloud Functions} & Public cloud & Undefined & 540 \\ \hline
\textbf{Azure Functions} & Public cloud, local & OS images & 600 \\ \hline
\textbf{OpenWhisk} & Kubernetes, Private cloud, Public cloud & Docker & 300 \\ \hline
\textbf{Kubeless} & Kubernetes & Docker & Undefined \\ \hline
\textbf{SAND} & Public cloud, Private cloud & Docker & Undefined \\ \hline
\textbf{Fn} & Public cloud, Kubernetes & Docker & 300 \\ \hline
\textbf{Abaco} & TACC clusters & Docker & Undefined \\ \hline
\rowcolor[HTML]{D0D0D0}
\textbf{\emph{FSys}{}} & Local, clouds (AWS, Azure, GCP), clusters (e.g., Cobalt, Slurm, Torque/PBS), supercomputers & Singularity, Shifter, Docker & No limit \\ \hline
\end{tabular}
\end{table*}
\subsection{Commercial Platforms}
Most commercial cloud providers offer FaaS capabilities.
Here we compare three
platforms offered by Amazon, Microsoft, and Google.
\emph{Amazon Lambda}~\cite{AmazonLambda} pioneered the FaaS paradigm in 2014 and has since
be used in many
industry~\cite{netflixlambda} and academic~\cite{chard17ripple} use cases. Lambda is a hosted service
that supports a multitude of function languages and trigger sources (Web interface, CLI, SDK, and other AWS services).
Tight integration with the wider AWS ecosystem means Lambda functions can be associated with triggers from other AWS services, such as CloudWatch, S3, API gateways, SQS queues, and Step Functions. Functions are billed based on their memory allocation and
for every 100ms execution time. Once defined, Lambda uses a custom virtualization technology built on KVM, called Firecracker to create lightweight micro-virtual machines. These microVMs then persist in a \textit{warmed} state for five minutes and continue to serve requests.
While Lambda is provided as a hosted service, functions can be deployed locally or to edge devices
via the Greengrass~\cite{greengrass} IoT platform.
\emph{Google Cloud Functions}~\cite{googlecloudfunctions} is differentiated by its tight coupling to Google Cloud Storage, Firebase mobile backends, and custom IoT configurations via Google's globally distributed message bus (Cloud Pub/Sub).
Like Lambda, Google Cloud Functions also support triggers from arbitrary HTTP webhooks.
Further, users can trigger functions through a number of third party systems including
GitHub, Slack, and Stripe.
While Google Cloud functions apply a similar pricing model to Lambda, the model is slightly more expensive
for high-volume, less computationally intensive tasks as Lambda has lower per-request costs after the first two million invocations (with similar compute duration costs).
\emph{Azure Functions}~\cite{azureFunctions} allow users to create functions in a native language through either the Web interface or the CLI. Functions are packaged and may be tested locally using a local web service before being uploaded to the Azure platform.
Azure functions integrate with other Azure products through triggers.
Triggers are provided from CosmosDB, Blob storage, and Azure storage queues, in addition to custom HTTP and time-based triggers. Azure price-matches AWS for compute and storage (as of November 2018).
\subsection{Open Source Platforms}
Open FaaS platforms resolve two of the key challenges to using FaaS for scientific workloads: they
can be deployed on-premise and can be customized to meet the requirements of data-intensive
workloads without any pricing models.
\emph{Apache OpenWhisk}~\cite{openwhisk} is the most well-known open source FaaS platform. OpenWhisk is the basis of IBM Cloud Functions~\cite{IBMCloudFunctions}. OpenWhisk clearly defines an event-based programming model, consisting of \emph{Actions} which are stateless, runnable functions, \emph{Triggers} which are the types of events OpenWhisk may track, and \emph{Rules} which associate one trigger with one action. OpenWhisk can be deployed locally as a service using a Kubernetes cluster. However, deploying OpenWhisk is non-trivial, requiring installation of dependencies and administrator access to the cluster.
\emph{Fn}~\cite{Fn} is a powerful open-source software from Oracle that can be deployed on any Linux-based compute resource having administrator access to run Docker containers. Applications---or groups of functions---allow users to logically group functions to build runtime utilities (e.g.,
dependency downloads in custom Docker containers) and other resources (e.g.,
a trained machine learning model file) to support functions in the group. Moreover, Fn supports fine-grained logging and metrics, and is one of few open source FaaS platforms deployable on Windows. Fn can be deployed locally or on a Kubernetes cluster. In our experience, one can deploy a fully-functional Fn server in minutes.
\emph{Kubeless}~\cite{Kubeless} is a native
Kubernetes FaaS platform that takes advantage of built-in
Kubernetes primitives. Kubeless uses Apache Kafka for messaging, provides a CLI that mirrors that of AWS Lambda, and supports fine-grained monitoring. Users can invoke functions via the CLI, HTTP, and via a Pub/Sub mechanism. Like Fn, Kubeless allows users to define function groups that share resources. Like OpenWhisk, Kubeless is reliant on Kubernetes and cannot be deployed on other resources.
\subsection{Academic Platforms}
The success of FaaS in industry has spurred academic exploration of FaaS. Two
systems that have resulted from that work are SAND~\cite{akkus2018sand} and Actor Based Co(mputing)ntainers (Abaco)~\cite{stubbs2017containers}.
\emph{SAND}~\cite{akkus2018sand} is a lightweight, low-latency FaaS platform from Nokia Labs that provides application-level sandboxing and a hierarchical message bus. The authors state that they achieve a 43\% speedup and a 22x latency reduction over Apache OpenWhisk
in commonly-used image processing applications.
Further, SAND provides support for function or \textit{grain} chaining via user-submitted workflows. At the time of their writing, it appears that SAND does not support multi-tenancy, only having isolation at the application level.
SAND is closed source and as far as we know cannot be downloaded and installed locally.
\emph{Abaco}~\cite{stubbs2017containers} supports functions written in a wide range of programming languages and supports automatic scaling. Abaco implements the Actor model in which an \textit{actor} is an Abaco runtime mapped to a specific Docker image. Each actor executes in response to messages posted to its \textit{inbox}. Moreover, Abaco provides fine-grained monitoring of container, state, and execution events and statistics. Abaco is deployable via Docker Compose.
\subsection{Summary of FaaS}
Commercial cloud providers implement high performance and reliable FaaS models
that are used by huge numbers of users. However, for science use cases they
are unable to make use of existing infrastructure, they do not integrate
with the science ecosystem (e.g., in terms of data and authentication models),
and they can be costly.
Open source and academic frameworks support on-premise deployments and can be
configured to address a range of use cases. However, each of the systems surveyed
is Docker-based and therefore requires administrator privileges to be deployed
on external systems. Furthermore, the reliance on Docker prohibits use in
most computing centers which instead support user space containers.
In most cases, these systems have been implemented to rely on Kubernetes (or other
container orchestration models such as Mesos and Openshift) which means
they cannot be adapted to existing HPC and HTC environments.
\emph{FSys}{} provides a scalable, low-latency
FaaS platform that can be applied to existing HPC resources with minimal effort. It employs user-space
containers to isolate and execute functions, avoiding the security concerns prohibiting other FaaS
platforms from being used. Finally, it provides an intuitive interface for executing scientific
workloads and includes a number of performance optimizations to support broad scientific use cases.
\subsection{Other Related Approaches}
FaaS builds upon a large amount of related work including
in Grid and cloud computing, container orchestration, and analysis systems.
Grid computing~\cite{Foster2001} laid the foundation for
remote, federated computations, most often applying
federated batch submission~\cite{krauter02gridmanagement}.
GridRPC~\cite{seymour02gridrpc} defines an API for executing
functions on remote servers requiring that developers implement
the client and the server code.
\emph{FSys}{} extends these ideas to allow interpreted functions
to be registered and subsequently to be dynamically
executed within sandboxed containers via a standard endpoint API.
Container orchestration systems~\cite{rodriguez19containers, hightower17kubernetes, hindman11mesos} allow
users to scale deployment of containers while
managing scheduling, fault tolerance, resource provisioning, and addressing
other user requirements.
These systems primarily rely on dedicated, cloud-like
infrastructure and cannot be directly applied to HPC resources.
However, these tools provide a basis for other serverless platforms, such as Knative~\cite{knative} and Kubeless~\cite{kubeless}. Knative is a middleware that simplifies deploying serverless applications on a variety of resources using Kubernetes and public clouds.
\emph{FSys}{}
provides similar functionality,
however it focuses at the level of scheduling and managing functions, that are deployed
across a pool of containers, rather than deploying services. In addition, it does not rely on container orchestration systems and can interface with almost any computing resource. We apply approaches from container orchestration
systems (e.g., warming) to improve performance.
Data-parallel systems such as Hadoop~\cite{hadoop} and Spark~\cite{spark}
enable map-reduce style analyses.
Unlike \emph{FSys}{}, these systems dictate a particular programming
model on dedicated clusters.
Parallel computing libraries such as Dask~\cite{dask}, Parsl~\cite{babuji19parsl},
and Ray~\cite{moritz2018ray} support parallel execution of scripts,
and selected functions within those scripts,
on clusters and clouds. \emph{FSys}{} uses Parsl to manage
function execution in containers.
\section{Related Work}\label{sec:survey}
Both commercial and open-source FaaS platforms have proved extremely successful in industry as a way
to reduce costs and remove the need to manage infrastructure.
\textbf{Hosted FaaS platforms:}
\emph{Amazon Lambda}~\cite{amazonlambda},
\emph{Google Cloud Functions}~\cite{googlecloudfunctions}, and \emph{Azure Functions}~\cite{azureFunctions} are the most well-known FaaS platforms. They support various function languages
and trigger sources, connect directly to other cloud services, and apply fine-grain billing models. Lambda uses Firecracker~\cite{agache2020firecracker}, a custom virtualization technology built on KVM, to create lightweight micro-virtual machines.
To meet the needs of IoT use cases,
some cloud-hosted platforms support local deployment (e.g., AWS Greengrass~\cite{greengrass});
however, they support only single machines and require that functions be exported from the cloud platform.
\textbf{Open source platforms:}
Open FaaS platforms resolve two of the key challenges to using FaaS for scientific workloads: they
can be deployed on-premise and can be customized to meet the requirements of data-intensive
workloads without set pricing models.
\emph{Apache OpenWhisk}~\cite{openwhisk}, the basis of IBM Cloud Functions~\cite{IBMCloudFunctions},
defines an event-based programming model, consisting of \emph{Actions} which are stateless, runnable functions, \emph{Triggers} which are the types of events OpenWhisk may track, and \emph{Rules} which associate one trigger with one action. OpenWhisk can be deployed locally as a service using a Kubernetes cluster.
\emph{Fn}~\cite{Fn} is an event-driven FaaS system that executes functions in Docker containers. Fn allows users to logically group functions into applications.
Fn can be deployed locally (on Windows, MacOS, or Linux) or on Kubernetes.
The \emph{Kubeless}~\cite{Kubeless} FaaS platform builds upon Kubernetes.
It uses Apache Kafka for messaging, provides a CLI that mirrors Amazon Lambda, and supports comprehensive monitoring.
Like Fn, Kubeless allows users to define function groups that share resources.
\emph{SAND}~\cite{akkus2018sand}, which has been recently open-sourced as KNIX MicroFunctions~\cite{knix}, is a lightweight, low-latency FaaS platform from Nokia Bell Labs that provides application-level sandboxing and a light-weight process-based execution model. KNIX provides support for function chaining via user-submitted workflows.
Recently, KNIX has been further extended to support GPU sharing among functions~\cite{satzke2021efficient}.
However, KNIX requires privileged access to nodes, which is generally not possible in research CI.
\emph{Abaco}~\cite{stubbs2017containers} implements the Actor model, where an \textit{actor} is an Abaco runtime mapped to a specific Docker image. Each actor executes in response to messages posted to its \textit{inbox}. It supports functions written in several programming languages and automatic scaling. Abaco also provides fine-grained monitoring of container, state, and execution events and statistics. Abaco is deployable via Docker Compose.
\emph{ChainFaaS}~\cite{ghaemi2020chainfaas} is a blockchain-based FaaS platform that makes use of idle personal computers. The platform allows users to submit functions that utilize contributed computing power, or to be a provider who contributes the idle computing resources for potential profits. While ChainFaaS shares some similar goals with funcX, it focuses on deployment on personal computers, rather than large-scale research CI.
\emph{DFaaS}~\cite{ciavotta2021dfaas} is a federated and decentralized FaaS platform for edge computing. It relies on a peer-to-peer network to share the states of edge nodes to balance loads among all the nodes.
\textbf{Comparison with \emph{FSys}{}:}
Hosted cloud providers implement high performance and reliable FaaS models
that are used by an enormous number of users. However, they often have vendor lock-in, are not designed
to support heterogeneous resources or research CI (e.g., schedulers, containers),
do not integrate with the science ecosystem
(e.g., in terms of data and authentication models), and can be costly.
Open source and academic frameworks support on-premise deployments and can be
configured to address a range of use cases. However, most systems we surveyed
are Docker-based and rely on Kubernetes (or other
container orchestration platforms) for deployment.
Some systems such as ChainFaaS and DFaaS support distributed function execution on personal computers and edge nodes.
However, to the best of our knowledge, there are no systems that support remote execution over a federated ecosystem of endpoints on diverse research CI (from edge to HPC environments).
\textbf{Other Related Approaches:}
FaaS has many predecessors, notably
grid and cloud computing, container orchestration, and analysis systems.
Grid computing~\cite{Foster2001} laid the foundation for
remote, federated computations, often through
federated batch submission~\cite{krauter02gridmanagement}.
GridRPC~\cite{seymour02gridrpc} defines an API for executing
functions on remote servers requiring that developers implement
the client and the server code.
\emph{FSys}{} extends these ideas to allow interpreted functions
to be registered and then executed within sandboxed
containers via standard cloud and endpoint APIs.
Container orchestration systems, such as Mesos~\cite{hindman11mesos}, Kubernetes~\cite{hightower17kubernetes}, KubeFed~\cite{kubefed}, MicroK8s~\cite{microk8s}, and K3s~\cite{k3s}, allow
users to scale deployment of containers while
managing scheduling, fault tolerance, resource provisioning, and addressing
other user requirements.
Mesos and Kubernetes primarily rely on dedicated, cloud-native
infrastructure. KubeFed extends Kubernetes to support multi-cluster deployments.
MicroK8s and K3s are lightweight versions of Kubernetes and are designed for Edge and IoT use cases.
These systems cannot be directly used with diverse research CI (e.g., HPC resources); however, these container orchestration systems serve as a basis for developing serverless platforms, such as Kubeless, and indeed play an increasingly important
role in research CI.
\emph{FSys}{} focuses at the level of scheduling and managing functions,
that are deployed across a pool of containers.
We leverage both container orchestration
systems (e.g., Kubernetes) as well as techniques from
orchestration systems (e.g., warming) in funcX.
Data-parallel systems such as Hadoop~\cite{hadoop} and Spark~\cite{spark}
enable map-reduce style analyses.
Unlike \emph{FSys}{}, these systems dictate a particular programming
model on dedicated clusters.
Python parallel computing libraries such as Parsl and Dask~\cite{dask}
support development of parallel programs,
and parallel execution of selected functions within those scripts,
on clusters and clouds. These systems could be extended to use \emph{FSys}{}
for remote execution of tasks.
LFM~\cite{shaffer2021lfm} provides advanced dependency management for Python functions by using transparent dependency detection and distribution, and dynamic provisioning and resource management at the granularity of a Python function.
Azure Functions~\cite{shahrad2020serverless} proposed a policy that dynamically controls the pre-warming window for application containers to reduce the number of container cold starts, based on the characterization of applications.
Researchers have proposed various methods to mitigate container cold start latency by leveraging various workflow-specific information, such as cascading starts and dependency graphs~\cite{daw2020xanadu,shen2021defuse,bermbach2020using}.
Anna~\cite{sreekanti2020cloudburst} is an autoscaling key-value store that can be used to support stateful serverless computing.
Delta~\cite{kumar2021delta} adds a shim layer on top of \emph{FSys}{} that profiles the function performance on different endpoints and automatically schedules functions to appropriate endpoints.
Several recent papers have aimed to model application performance and optimize performance on FaaS platforms~\cite{HoseinyFarahabady2018model,bao2019performance,kim2020automated,lin2021modeling}.
While \emph{FSys}{} implements its own function routing, container management, data management schemes, and performance metrics,
these systems are orthogonal to this paper and could be integrated with \emph{FSys}{}.
Several frameworks have been implemented on top of \emph{FSys}{} to create workflows for different scientific use cases. For instance, Xtract~\cite{skluzacek2021serverless} uses \emph{FSys}{} to enable workflow compositions for distributed bulk metadata extraction. Globus Automate~\cite{vescovi2022linking} uses \emph{FSys}{} to run arbitrary computations as part of automated and event-based workflows, it uses \emph{FSys}{}'s APIs to automatically monitor the status of a \emph{FSys}{} function and trigger the next step when it completes.
|
\section{Introduction}
In the context of financial risk management, financial risk models are of utmost importance in order to quantify and manage financial risk. Their outputs, risk measurements, can either help in the process of decision making or ensure that the regulatory requirements are met \citep{BISFRTB}. Financial management thus heavily relies on financial risk models and the interpretation of their outputs. The data usually consist of time series
representing a wide variety of risk factors. A major issue of such data is the presence of anomalies. A time series is considered to be abnormal whenever its behaviour is significantly different from the behaviour of the rest of the time series \citep{hawkins1980identification}. In this work, we focus on the detection of abnormal observations in a risk factor time series used to calibrate financial risk models. Indeed, financial risk models may be sensitive to anomalies, when erroneous input data wrongly impact the calibrated model parameters. A typical example could be the estimation of the covariance matrix of a bank risk factors. The covariance matrix is involved, for instance, in the computation of Value-at Risks (VaR) or Expected Shortfalls, that is expected losses above the VaR level. Since the true covariance is unknown, it has to be estimated from the data. However, the presence of anomalies in the data might have an impact on the estimation. For this specific case, robust methods, which are less sensitive to anomalies, can be used. Yet, existing robust estimators are computationally expensive, with a polynomial or even exponential time complexity in terms of the number of risk factors. A faster approach was suggested by \citep{cheng2019faster}. Nonetheless, this algorithm only applies to the estimation of the covariance matrix in the case of a high-dimensional Gaussian distribution. Financial risk models used by banks are widespread and various. Therefore, instead of seeking for a robust version for each of them, we propose to detect anomalies directly on the time series. This way we are able to bypass the problem of sensitivity of all the models to anomalies.
\paragraph{Baseline Algorithms}
Anomaly detection aims at finding an \textit{``observation that deviates so much from other observations as to arouse suspicion that it was generated by a different mechanism"} \citep{hawkins1980identification}. The baseline anomaly detection algorithms, described in \citep{chandola2009anomaly}, struggle to identify anomalies in time series, mainly because their assumptions are invalidated. Indeed, first, if we consider models built for spatial data, a major assumption of these models is that observations are independent, whereas for time series, high dependency exists between different time stamps. Clustering-based approaches, like the density based spatial clustering with noise (DBSCAN) method, are particularly impacted by this aspect: if an anomaly occurs at a given time stamp and is followed by incorrect
values, clustering-based approaches consider that the observations of the time series belong to two different clusters and thus fail to identify the anomaly. Another limitation when considering this type of techniques is the choice of the similarity metric used for data clustering. This task, although being a crucial pillar of these approaches, is not trivial and becomes even more complex for high dimensional problems.
\paragraph{Statistical Approach to Anomaly Detection}
The statistical techniques for anomaly detection can be split in two families: statistical tests and predictive models. Both suffer from the curse of dimensionality and model/data mismatch. When anomaly detection relies on hypothesis tests, usually they test whether the observations are drawn from a known distribution \citep{zhang2017statistical}, which supposes that the user knows the probability distribution of the normal observations. This parametric framework narrows down the scope of applicability of hypothesis tests, as the data does not always coincide with the assumed distribution. Moreover, the tests provided in the literature are not suitable in high dimensional settings since they were originally built for univariate data as explained in \citep{kurt2020real}.
The statistical technique relying on fitting a predictive model to each time series also requires strong assumptions on the data. Predictive models are usually autoregressive (AR), moving average (MA), or ARMA models. Anomalies are then detected relatively to the forecasts suggested by the model \citep{chandola2009anomaly}. However, in this parametric approach, some parameters have to be specified, starting with the order of the models. Selecting the optimal model parameters with respect to an information criterion is not always possible \citep{sedman2018online}. Additionally, these models assume that the time series are homogeneous, i.e. drawn from the same distribution \citep{laptev2015generic}. This is not always satisfied in the financial risk management case where several types of risk factors are treated simultaneously.
\paragraph{Score based Anomaly Detection Models}
Additional anomaly detection challenges are of general concern. Most of anomaly detection algorithms are score-based, in the sense that these approaches return an anomaly score reflecting to which extent the observation is considered to be abnormal by the model. In order to decide whether an observation is abnormal or not, a cut-off value of the score has to be selected. Empirical approaches are often used, consisting in either setting the cut-off value as a quantile of the distribution of the anomaly score, or in considering the cut-off value as the elbow point of this distribution. However, the selected cut-off value according to such methods remains arbitrary. An advanced approach is proposed by \citep{gaoconverting} who propose to rely on the cost of misclassification. In most cases, the latter is taken as a weighted classification accuracy. Approaches to calculate ``optimal'' weights are described in \citep{lu2019learning}, but they involve a grid search, technique that may draw near the optimum but may also diverge. Another alternative is to determine the cut-off value by cross validation on the training data as in \citep{saha2009snake}. Finally, some methods do not select any cut-off value, but are based instead on a contamination rate. However, in a sense, fixing a cut-off value or deciding on a contamination rate is equivalent.
\paragraph{Scarcity of Anomalies and Data Augmentation}
The scarcity of anomalies within the data sets is another typical problem in anomaly detection. Anomalies are, by definition, rare events, therefore they are underrepresented in the data set used to fit the models. This under representation is not helping in the design of a reliable model able to identify anomalies. Classical methods to overcome this issue consider data augmentation. These techniques aims at producing new synthetic samples that will ultimately enhance model performance since a better representation of the feature space is allowed. As reported in \citep{wen2020time}, time series can be augmented by using a simple transformation (in time domain \citep{cui2016multi} or frequency domain \citep{gao2020robusttad}) or more advanced approaches as generative models, which involve deep learning techniques (such as recurrent generative adversarial networks \citep{esteban2017real}). In practice, the use of generative models for data augmentation of time series with anomalies presents two limitations. First, training such models requires a large number of samples to guarantee a satisfactory performance. While restricted Boltzman machines do not exhibit this problem, \citep{kondratyev2020data} have shown that they fail into fitting multivariate complex distributions with non linear dependence structure. A more fundamental limitation affects the very idea behind generative models. Such networks are trained to learn a given distribution. However, by definition, anomalies are different from each other and therefore there is not a distribution that characterises them.
\paragraph{Supervised vs. Unsupervised learning }
Since anomalies are the realisations of atypical events for which the distribution is unknown, it seems quite natural to use unsupervised algorithms. However, these approaches are deemed more suitable for learning complex patterns and are task specific. Moreover, \citep{gornitz2013toward} pointed out that they often do not present a high prediction performance, in particular in high dimensional settings \citep{ruff2019deep}. Indeed, the performance of unsupervised shallow anomaly detection algorithms depend upon a feature engineering step. \citep{akyildirim2022applications} proposed to use signatures as feature extractors which are fed to algorithms such as isolation forest. This combination of techniques is shown to over-perform benchmark approaches. However, the designed model is task specific (detection of pump and dumps attacks) and the feature extraction step is only efficient when at least one one explanatory variable is considered in the analysis.
Moreover, the unavailability of labelled data makes the model building and its evaluation even more complex. As for the supervised methods the only limit on which the literature tends to agree is their incapacity to generalize the learned patterns to new samples, which is the consequence of misrepresentation of anomalies among the training samples \citep{zhao2018xgbod}. However, even a small fraction of labelled data represents an essential information, that should be integrated during the model elaboration. By contrast, under the unsupervised learning framework, the use of this scarce but precious information is not possible \citep{ruff2019deep}. The lack of labelled data can be sidestepped through the use of data augmentation techniques on the fraction of available labelled data. Hence, for these reasons the supervised learning framework is to be preferred even when only a small set of labelled data is available.
\paragraph{Anomaly detection on Time Series}
Usually, anomaly detection models on time series have two main components. The first component aims at extracting a parsimonious yet expressive representation of the time series. Several approaches are suggested in the literature to deal with such feature extraction. Recently, deep neural networks have been shown to suffer from overparametrization and to be often computationally expensive \citep{dempster2020rocket, akyildirim2022applications}. Path signatures are also computationally demanding. This could perhaps be alleviated by the random signatures \citep{compagnoni2022randomized}. However, the information extracted with signatures is of most interest when the considered paths are characterized by several variables. The resulting representation is then transformed into an anomaly score. The second component aims at converting the anomaly score, which is usually a continuous variable, into a binary label \citep{braei2020anomaly}.
\paragraph{PCA NN Anomaly Detection}
In this paper, we propose a methodology to identify anomalies on time series using a principal component analysis (PCA) as underlying model, challenging current methods by providing a competitive alternative overcoming the above described pitfalls.
PCA is involved in our approach as feature extractor. With a relatively low number of features given by PCA, we are able to accurately describe the dynamics of risk factors represented by times series. The reason standing behind that is the high correlation structure displayed by financial risk factors. The particular power-fullness of auto-encoder, a non-linear PCA, as data compressor is not desirable herein, since auto-encoders compress all patterns including abnormal ones. In the literature, PCA is usually used in anomaly detection for its dimension reduction properties. Anomalies are identified on the latent space, either by applying some anomaly detection algorithm or by assuming a given distribution on the principal component and identifying the anomalies relatively to a quantile as in \citep{shyu2006principal}. Once PCA is involved in anomaly detection, the user assumes that a normal subspace representation of the data set can be constructed with the first $k$-components \citep{ringberg2007sensitivity}. Anomaly detection is then achieved by looking at the observations that cannot be expressed in terms of the first $k$-components \citep{bin2016abnormal}. While \citep{ding2016pca} claimed that the PCA-based models are stable with respect to their parameters, such as the number of principal components $k$ spanning the subspace or the cut-off level, \citep{ringberg2007sensitivity} found instead that PCA-based anomaly detection is sensitive to these parameters and to the amplitude of the anomalies. Indeed, the latter may undermine the construction of the normal subspace representation, in turn leading to mis-identification of anomalies. In light of that, we take an extra care regarding these aspects, and appropriate tests are conducted to show that the suggested approach is not subject to such issues. In addition, we introduce a neural network approach to calibrate the anomaly score threshold.
\paragraph{Outline}
The paper is organised as follows. In Section \ref{section: model}, we present the PCA NN approach with a particular focus on each step of the technique. A brief description of PCA is also provided to motivate its use as an efficient feature extractor tool. In Section \ref{section: data}, we present the numerical framework adopted to generate time series in Section \ref{subsection:datasim}, then contaminate them (Section \ref{subsec:DataContamination}) and the process of data augmentation used to overcome the scaricity of anomalies. Our main numerical results are provided in Section \ref{section: model_eval}, with the main performance metrics that were considered to assess the suggested anomaly detection model (Section \ref{subsection:perfmetrics}). The calibration of the latent space dimension is described in Section \ref{subsec:caliK} followed by the model evaluation in Sections \ref{subsection:evalstep1} and \ref{subsection:evalstep2}. We numerically demonstrate the efficiency of our approach over baseline anomaly detection models in Section \ref{subsection: benchmarking}. In Section \ref{subsection:imputation}, we tackle the issue of anomalies imputation once detected, comparing the imputation approach suggested by our approach to comparable imputation techniques (in terms of computational complexity). The calibration of the anomaly score cut-off value is a main contribution of the approach, therefore its robustness was tested in Section \ref{subsection:ThresholdRobustness}. Finally, we show on a practical case of value-at-risk estimation in Section \ref{section:VaR}, that the estimation errors are reduced when the proposed anomaly detection model is used, together with a naive imputation approach to correct the anomaly. A conclusion of the paper with further discussions are presented in Section \ref{section: conclusion}.
\bigskip
\section{Method: PCA NN Anomaly Detection}\label{section: model}
After setting the key notations that are used in the remainder of the paper we proceed with the a general description of our two step method for anomaly detection on time series while briefly recalling tenets of principle components analysis. Then the first step of the approach i.e. the contaminated time series identification is introduced whilst explaining the motivation behind the model suggested for this step. The second step i.e. anomaly localization step is described in further details at the end of this section.
\subsection{Notations}\label{subsec:MethodNotation}
In this section, we introduce the notations that we use in the following sections to describe our two-step PCA methodology for anomaly detection on financial time series.
Let
\begin{equation}\label{def:X}
\bm{X} =
\left(
X^{1}, X^{2}, \ldots , X^{i}, \ldots, X^{n}
\right)^\top,
\end{equation}
where the column-vector $X^i= \left(x_{t_1}^i,x_{t_2}^i, \ldots, x_{t_j}^i\ldots, x_{t_{p}}^i\right) \in \mathbb{R}^{p}$ corresponds to the $i$-th observed time series, and $x_{t_j}^i$ to the value observed at time $t_j$ in the $i$-th time series.
Our method fits into the supervised framework. We thus assume that the data matrix $\bm{X}$ comes along with two label matrices. The first label matrix $\bm{A} = \left(A^1,\ldots,A^n\right) \in \{0,1\}^n$ identifies the time series containing anomalies referred to as contaminated time series. For $i=1,\ldots, n$, we define the identification labels as
\begin{align}
A^{i} = \begin{cases}
1 & \text{if there exists $j$ such that $x^i_{t_j}$ is an anomaly}. \\
0 & \text{otherwise.}
\end{cases}
\label{eq:identLabel}
\end{align}
The second label matrix $\bm{L}$ concerns solely the contaminated time series. Its coefficients correspond to the location labels, that is the time stamps at which an anomaly occurs. For $j=1,\ldots, p$ and $i \in I_c = \{i : A^i = 1\}$ (the set of contaminated time series),
\[
L^{i} = \jmath \quad \text{ when the anomaly occurs at time $t_{\jmath}$ for the $i$-th contaminated time series},
\]
$\bm{X},\bm{A}$ and $\bm{L}$ are general notations. As the model involves a learning phase, we denote by $\bm{X}^{Train}$, $\bm{A}^{Train}$ and $\bm{L}^{Train}$ the data used for calibration and by $\bm{X}^{Test}$, $\bm{A}^{Test}$ and $\bm{L}^{Test}$ the independent data set on which the model performance is evaluated.
\subsection{Model Description}
\label{subsection:modeldescription}
The anomaly detection model we propose is a two step supervised-learning approach. The first step, namely \textit{Contaminated Time Series Identification Step}, aims at identifying among the observed time series the ones with potential anomalies. The second step, called \textit{Anomaly Localization Step}, consists in finding the location of the anomaly in each contaminated time series identified during the first step.
The model used in the first step falls under the scope of binary classification models, as it will assign to each time series a predicted label $\widehat{A}^{i}$ in the following way:
\[
\widehat{A}^{i} = \begin{cases}
1 & \text{if $X^{i}$ is considered as contaminated by the identification model.} \\
0 & \text{otherwise.}
\end{cases}
\]
The second step uses a multi-classification model. For each contaminated time series identified in the first step, the model predicts a unique time stamp $\widehat{L}^{i}$ at which the anomaly has occurred. Formally, the observed value at the time stamp $t_{\hat{\jmath}^{i}}$ of the $i$-th contaminated time series is considered abnormal, meaning that the observation $x^i_{t_{\hat{\jmath}^i}}$ is abnormal. Thus,
\[
\widehat{L}^{i} = \hat{\jmath}^{i} \quad \text{ with } \quad \hat{\jmath}^{i} \in \{ 1,\dots,p \}.
\]
Our two-step anomaly detection model locates only one anomaly per run, if any. Hence, several iterations allow removing all anomalies. Indeed, as long as the time series is identified as contaminated by the first step of the model, the time series can go through the second step of the model. Once, all anomalies have been located the final step of the approach would be to remove the anomalous values and suggest imputation values.
\subsection{Theoretical Basics of Principal Component Analysis}
\label{subsec:PCA}
In this section, we recall the theoretical tenets of the principal component analysis (PCA) useful for our purpose. The idea of PCA is to project the data from the observation space of dimension $p$ into a latent space of dimension $k$, with $k <p$. This latent space is generated by the $k$ directions for which the variance retained under projection is maximal.
The first principal component $U$ is given by the equation
\[
U = w^\top\bm{X} \quad \text{with} \quad \mathrm{var}(U) = w^\top\Sigma w,
\]
with $w$ a vector of weights to be determined. Since the PCA aims at maximizing the variance along its principal components, $w$ is sought for as
\[
\arg \max_{\| w\| = 1} w^\top\Sigma w \,.
\]
The optimal solution by the Lagrangian method is given by
\[
\Sigma w = \lambda w \, ,
\]
where $\lambda>0$ is a Lagrangian multiplier. Multiplying both sides by $w^\top$ results in
\[
w^\top\Sigma w = \lambda.
\]
Following the same process with an additional constraint regarding the orthogonality between the principal components, one can show that the $k$ first principal components correspond to the eigenvectors associated with the $k$ dominant eigenvalues of $\Sigma$.
The transfer matrix $\Omega_{\bm{X}} \in \mathbb{R}^{k\times p}$ is defined by the eigenvectors of $\Sigma$ associated with its $k$ largest eigenvalues. This transfer matrix $\Omega_{\bm{X}}$ applied to any observation in the original observation space projects it into the latent space. Since the transformation is linear, reconstructing the initial observations from their equivalent in the latent space is straightforward. Using PCA on the data matrix $\bm{X}$, the transfer matrix $\Omega_{\bm{X}}$ can be inferred. The projection of the observations into the latent space $\bm{Z} \in \mathbb{R}^{n\times k}$ is then given by $\bm{Z} = \bm{X} \Omega_{\bm{X}}^\top $, and their reconstructed values by $\widehat{\bm{X}} = \bm{Z}\Omega_{\bm{X}}$.
The reconstruction errors $\varepsilon^i$ for each time series defined by
\begin{align}
\varepsilon^{i} =\widehat{X}^{i} - X^{i}, \quad i=1,\ldots, n,
\label{eq:recerror}
\end{align}
are our new representation of data, that are given as inputs to our models, referred in the sequel as the new features. It is worth stressing that the two steps of the model use the same inputs, namely these reconstruction errors. Nevertheless, the processing these inputs undergo in each step differ.
A fundamental intuition behind the anomaly detection algorithm we suggest is the existence of a lower dimensional subspace where normal and abnormal observations are easily distinguishable. Indeed, since the selected principal components explain a given level of the variance of the data, they represent the common and essential characteristics to all observations. As one might reasonably expect, these characteristics mostly represent the normal observations. Therefore, when the inverse transformation is applied, the model successfully reconstructs the normal observations with the help of the extracted characteristics, while it fails into reconstructing the anomalies.
Following this intuition, we expect normal observations to have low reconstruction errors, and higher errors for anomalies. Since this error reflects the degree to which the instance is considered as abnormal, it can be used to propose an anomaly score according to which each observation will be labelled as either normal or abnormal.
A decision rule thus needs to be defined on the anomaly score to convert it into a binary label: classically in anomaly detection, a time series is labelled as abnormal if its anomaly score exceeds a previously determined cut-off value $s$. This cut-off value is, in most of anomaly detection algorithms, a hyper-parameter that needs to be tuned. In this paper, we propose to calibrate $s$ during the learning phase as described in Section \ref{subsec:Step1}. Algorithm \ref{algo: FeaturesExtraction} summarizes how to compute the reconstruction errors.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Features extraction with PCA }
\label{algo: FeaturesExtraction}
\SetKwInOut{Require}{inputs}
\Require{number of principal components $k$\\out of sample data $\bm{X}^{Test}$\\training data if not trained yet $\bm{X}^{Train}$ }
\If{not trained yet}{
$\Omega_{\bm{X}^{Train}} \leftarrow \mathrm{PCA}(\bm{X}^{Train},k)$\;
}
$\bm{\varepsilon}^{Train} = \bm{X}^{Train} \left(\Omega_{\bm{X}^{Train}}^\top\Omega_{\bm{X}^{Train}} - I_{p}\right)$
$\bm{\varepsilon}^{Test} = \bm{X}^{Test} \left(\Omega_{\bm{X}^{Train}}^\top\Omega_{\bm{X}^{Train}} - I_{p}\right)$\\
\SetKwInOut{Output}{outputs}
\Output{$\bm{\varepsilon}^{Train}$ Reconstruction errors for the train set \\
$\bm{\varepsilon}^{Test}$ Reconstruction errors for the test set}
\end{algorithm}
In the next sections, we introduce the two steps of our approach, while the first step identifies the contaminated time series, the second step focuses on the localization of the anomaly in the time series. Both steps use the reconstruction errors as inputs. The main difference is the processing undergone by these inputs, while for the first step we use a neural network to compute an anomaly score which reflects the propensity of the time series of being contaminated (cf. Section \ref{subsec:Step1}), a more shallow approach is applied on the reconstruction errors to localise the abnormal observation in the second step (cf. Section \ref{subsec:Anomalylocalization}).
\subsection{Contaminated Time Series Identification}\label{subsec:Step1}
Recall that the first step of the approach, that is the \textit{Contaminated time series identification} step, aims at identifying the time series with anomalies based on the reconstruction errors that we obtain using PCA and Algorithm \ref{algo: FeaturesExtraction}. In this section, we detail the process these new features undergo. In fact, we assign to each reconstruction error an anomaly score which computation is handled by a neural network (NN). We also numerically demonstrate that our NN have an advantage over the naive approach in terms of being able to assess more accurately the propensity of a time series of being contaminated.
\paragraph{Naive approach}
A naive and natural approach to define the anomaly score is to consider the $\ell^2$-norm of the reconstruction errors $\widetilde{\varepsilon}_i = \lVert \varepsilon^{i}\rVert_2$, $i=1,\ldots, n$. This anomaly score represents the propensity of a time series to be contaminated or not.
The scores are first split into two sets depending on whether the corresponding time series is identified as contaminated or not. We denote by $\widetilde{\bm{\varepsilon}}^c$ and $\widetilde{\bm{\varepsilon}}^u$ the set of contaminated and uncontaminated time series, i.e.
\begin{align*}
\widetilde{\bm{\varepsilon}}^c &:= \{ \lVert \varepsilon^{i}\rVert_2; \;A^{i} = 1\, \} = \{ \widetilde{\varepsilon}_{i}, \, i\in I_c \} \\
\widetilde{\bm{\varepsilon}}^u &:= \{ \lVert \varepsilon^{i}\rVert_2; \; A^{i} = 0\, \} = \{ \widetilde{\varepsilon}_{i}, \, i\notin I_c \} .
\end{align*}
The density distribution function of each class of time series $f^u$ and $f^c$, i.e. for both $\widetilde{\bm{\varepsilon}}^c$ and $\widetilde{\bm{\varepsilon}}^u$, is then estimated using kernel density estimation \citep{wkeglarczyk2018kernel}. For $l=\{c,u\}$, let $\bm{\varepsilon}^l$ be an i.i.d sample of $n^l$ observations from a population with unknown density $f^l$. The corresponding kernel estimator is given by
\begin{align}
\hat{f}^l(s) = \frac{1}{n^l\mathfrak{h}} \sum_{k=1}^{n^l} \mathcal{K}\left( \frac{s - \widetilde{\varepsilon}_k^l}{\mathfrak{h}} \right),
\label{kde}
\end{align}
where $\mathcal{K}$ is a kernel function and $\mathfrak{h}$ is a smoothing parameter.
The cut-off value $s$ is then chosen as the intersection between the two empirical density functions, i.e.
\[
\hat s= \arg\min_s\left\{\lvert \hat{f}^u(s) - \hat{f}^c(s) \rvert < \eta \right\} \, ,
\]
with $\eta$ a precision level to be tuned. The selected cut-off value $\hat s$ represents the value of the score for which the area under the curve of the density of uncontaminated time series above $\hat s$ and the area under the curve of contaminated time series below $\hat s$ are as small as possible. We expect it to correspond to a value exceeded by a relative low amount of $\widetilde{\varepsilon}_i$ with $i \notin I_c$ as $\hat{f}^u(s)$ is expected to be flat around the cut-off value exceedance region, and exceeded only by few $\widetilde{\varepsilon}_i$ with $i \in I_c$.
\begin{figure}[H]
\centering{
\begin{tabular}{cc}
\includegraphics[width=0.45\linewidth]{Graphs/Density/Naive_denstTrain.pdf} &
\includegraphics[width=0.45\textwidth]{Graphs/Density/Naive_denstTest.pdf}\\
a) & b)
\end{tabular}}
\caption{Empirical densities of anomaly scores given by the naive approach for uncontaminated time series in black and contaminated time series in red, a) on the train set and b) on a test set. The dotted dark red line represents the cut-off value.}
\label{fig:DensityPCAReconsErrorNaive}
\end{figure}
As shown in Figure \ref{fig:DensityPCAReconsErrorNaive}, the naive approach results in a non-negligible overlapping region between the two densities, both on the train set (see Figure \ref{fig:DensityPCAReconsErrorNaive} a)) and for the test set (see Figure \ref{fig:DensityPCAReconsErrorNaive} b)). This region represents the anomaly score associated with time series that could be either contaminated or uncontaminated. The uncertainty regarding the nature of the observation when its anomaly score belongs to this region is relatively high in comparison with observations whose anomaly scores lie on the extreme left-hand or right-hand side of the calibrated cut-off value. Moreover, because we are not able to provide a clear separation between the scores of uncontaminated and contaminated time series, we may expect the model to mislabel future observations, resulting a high rate of false positives and true negatives.
\paragraph{Neural network approach}
In view of getting a clearer separation of the densities, we propose an alternative approach for the computation of the anomaly scores built upon the reconstructions errors. Let $F$ denote the function that associates with each reconstruction error $\varepsilon^{i}$ its anomaly score. In the naive approach, $F$ corresponds to the $\ell^2$-norm. Hereafter, $F$ instead represents the outputs of a feed-forward Neural Network (NN) \citep[see e.g.][]{goodfellow2016deep}, whose architecture is depicted in Figure \ref{fig:NNarchi}. The training of the corresponding NN aims at minimising a loss function that reflects our ambition to construct a function $F$ that gives accurate anomaly scores. The network used for computing the anomaly scores is defined by
\begin{align*}
&F(\varepsilon) = \left (h^H \circ h^{H-1} \circ \dots h^2 \circ h^1\right )(\varepsilon),
\end{align*}
where $h^i(\varepsilon) =\left(W^i \cdot \varepsilon + b^i \right)^+$
represents the computation carried in the $i$-th layer of the neural network with $n_h^i$ hidden units. $H$ stands for the number of layers of the NN with ReLU activation applied element-wise. We finally estimate the labels from the outputs following
\begin{align}
\bm{\hat{A}} = \mathbb{1}_{ \left\{(F(\bm{\varepsilon}) - s)^+ > 0\right\}}.
\label{eq:label}
\end{align}
The idea of the learning is to calibrate the weights $\bm{W} := \left \{W^i \in \mathbb{R}^{n_h^i}, i =1, \dots, H \right \}$ and the biases $\bm{b}:= \left \{b^i \in \mathbb{R}, i= 1, \dots, H\right \}$ such that the NN is able to accurately assess to which extent a time series is contaminated, with a clear distinction between the anomaly scores assigned to uncontaminated time series and those assigned to the contaminated ones, while integrating the calibration of the cut-off value $s$ as part of the learning. We denote by $\Theta:=\{\bm{W},\bm{b},s\}$ the set of parameters to be calibrated through the NN training. To meet these needs, the loss we minimise during the learning is given by
\begin{equation}
\mathcal{L}_{\Theta}\left(\bm{A}, \hat{\bm{A}}\right) = \mathrm{BCE}(\bm{A},\bm{\hat{A}}) + \mathrm{AUC Density}^u_{\bm{A},F(\bm{\varepsilon})} + \mathrm{AUC Density}^c_{\bm{A},F(\bm{\varepsilon})}
\label{eq:myLoss}
\end{equation}
where
\begin{align*}
BCE(\bm{A},\bm{\hat{A}}) = - \frac{1}{n}\sum_{i}\left(A^{i} \log(\hat{A}^{i}) + (1-A^{i}) \log(1-\hat{A}^{i})\right) \,
\end{align*}
is the binary cross-entropy, a well-known loss function classically used for classification problems. To this first component of our loss function, we add two components that aim at downsizing the overlapping region between the density of anomaly score of both types of observations. In order to have a control on the latter, we consider $\mathrm{AUC Density}^u$ and $\mathrm{AUC Dentsity}^c$, which correspond to the area under the curve of the probability density function of anomaly scores the model assigns to contaminated and uncontaminated observations, i.e
\begin{align}
& \mathrm{AUC Density}^u(s) = \int_s^{\infty} \hat{f}^u_{\bm{A},F(\bm{\varepsilon})}(\omega) d\omega \quad \quad \quad & \mathrm{AUC Density}^c(s) = \int_{-\infty}^s \hat{f}^c_{\bm{A},F(\bm{\varepsilon})}(\omega) d\omega \,.
\label{eq:AUCs}
\end{align}
Note that the bounds of these integrals depend on the cut-off value $s$ and define the region for which we want the probability density function to be as small as possible, which allows us to estimate $\hat{s}$. The estimated probability density functions $\hat{f}^u_{A,F(\bm{\varepsilon})}$ and $\hat{f}^c_{A,F(\bm{\varepsilon})}$ depend on $F(\bm{\varepsilon})$, the scores assigned by contaminated time series identification model, and of the identification labels $\bm{A}$. We describe with Algorithm \ref{algo: identificationLearning}, the scoring and cut-off value calibration that we achieve throughout the calibration of a feed-forward network with the customized loss \eqref{eq:myLoss}. To this end, the update of $\Theta$, AdamStep in Algorithm \ref{algo: identificationLearning}, is carried following the Adam optimization algorithm of \citep{kingma2014adam}.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Scoring and cut-off calibration }
\label{algo: identificationLearning}
\SetKwInOut{Require}{inputs}
\Require{learning rate $lr$\\
number of maximum iterations $K$\\
kernel density estimator parameter $\mathcal{K}$,$\mathfrak{h}$\\
training data $\bm{\varepsilon}^{Train},\bm{A}^{Train}$\\}
Initialize parameter $\Theta$, $\widehat{\mathcal{L}} = \infty$ and count index $k = 0$\;
\While(){$k< K$}{
$\\ scores \leftarrow F_{\Theta} \left(\bm{\varepsilon}^{Train}\right)$ \;\\
$scores^{c} \leftarrow scores \mathds{1}_{\bm{A}^{Train}=1}$\;\\
$scores^{u} \leftarrow scores \mathds{1}_{\bm{A}^{Train}=0}$\; \tcp*[f]{Density estimation following \eqref{kde}} \; \\
$\hat{f}^u_{\bm{A}^{Train},scores^u} = \mathrm{KernelDensityEstimator}(\mathcal{K},\mathfrak{h},scores^u)$ \; \\
$\hat{f}^c_{\bm{A}^{Train},scores^c} = \mathrm{KernelDensityEstimator}(\mathcal{K},\mathfrak{h},scores^c)$\\
\tcp*[f]{Loss evaluation following \eqref{eq:myLoss} and \eqref{eq:AUCs}} \; \\
$\mathrm{AUCDensity}^u \leftarrow \mathrm{NumericalIntegration}(\hat{f}^u_{\bm{A}^{Train},scores^u},s)$ \\
$\mathrm{AUCDensity}^c \leftarrow \mathrm{NumericalIntegration}(\hat{f}^c_{\bm{A}^{Train},scores^c},s)$\\
$\bm{\hat{A}} \leftarrow \mathds{1}_{scores>s}$\\
$\mathcal{L} \leftarrow \mathrm{BCE}\left(\bm{A}^{Train},\bm{\hat{A}}\right) +\mathrm{AUCDensity}^u + \mathrm{AUCDensity}^c $\\
\If(){ $\widehat{\mathcal{L}} > \mathcal{L}$ }{
$\\\widehat{\mathcal{L}} \leftarrow \mathcal{L}$\; \\
$\widehat{\Theta} \leftarrow \Theta$\;
}
$\\ \Theta \leftarrow \mathrm{AdamStep}(\mathcal{L},\Theta,lr)$\; \\
$k \leftarrow k +1$\;
}
\SetKwInOut{Output}{outputs}
\Output{best calibrated parameter $\widehat{\Theta}=\{\widehat{\bm{W}},\hat{\bm{b}}, \hat{s}\}$.} \end{algorithm}
\begin{figure}[H]
\centering{
\begin{tabular}{cc}
\includegraphics[width=0.45\linewidth]{Graphs/Density/NN_denstTrain.pdf} &
\includegraphics[width=0.45\textwidth]{Graphs/Density/NN_denstTest.pdf}\\
a) & b)
\end{tabular}}
\caption{Empirical densities of anomaly scores given by the NN approach for uncontaminated time series in black and contaminated time series in red, on the train set a) and on a test set b). The dotted dark red line represents the calibrated cut-off value $\hat{s}$.}
\label{fig:DensityPCAReconsErrorNN}
\end{figure}
\begin{table}[H]
\centering
\begin{tabular}{c}
\begin{tabular}{c|c|c}
Approach & $\mathrm{AUC Density}^u$ & $\mathrm{AUC Density}^c$\\
\hline
Naive & 0.1725 & 0.7898 \\
\hline
NN &0.05153 & 0.1550
\end{tabular}\\~\\
a)\\~\\
\begin{tabular}{c|c|c}
Approach & $\mathrm{AUC Density}^u$ & $\mathrm{AUC Density}^c$\\
\hline
Naive & 0.1897 & 0.6354 \\ \hline
NN & 0.1290 & 0.1367
\end{tabular}\\~\\
b)
\end{tabular}
\caption{AUC obtained with the naive and the NN approaches for a) the train set and b) the test set. }
\label{tab:AUC}
\end{table}
Table \ref{tab:AUC} shows that with the NN approach the densities of scores assigned to each type of time series display the expected behaviours on the left-hand (right-hand) side of the cut-off value for contaminated (uncontaminated) time series, on the train and test sets. Actually, the lower $\mathrm{AUC Density}^u$ ($\mathrm{AUC Density}^c$) is, the lower the number of uncontaminated (contaminated) time series to which are assigned anomaly scores above (below) the cut-off value, which prevents mislabelling.
Once the features $\bm{X}$ and the calibrated NN are provided, the contaminated times series identification model is ready for use. This step is described by Algorithm \ref{algo: IdentificationModel}.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Contaminated time series identification model }
\label{algo: IdentificationModel}
\SetKwInOut{Require}{inputs}
\Require{Time series to analyze $X^i$,\\ Calibrated model parameters $\widehat{\Theta}$, \\}
$\varepsilon^{i} \leftarrow \mathrm{PCAFeaturesExtraction}(X^i)$\; \tcp*{cf. Algorithm \ref{algo: FeaturesExtraction} in Section \ref{subsec:PCA}}
$score^{i} \leftarrow F_{\widehat{\Theta}}(\varepsilon^{i})$\;
$\hat{A}^i \leftarrow \mathds{1}_{score^{i}>\hat{s}} $\;
\SetKwInOut{Output}{outputs}
\Output{identification label $\hat{A}^i$.} \end{algorithm}
\tikzset{%
every neuron/.style={
circle,
draw,
minimum size=1cm
},
neuron missing/.style={
draw=none,
scale=2,
text height=0.333cm,
execute at begin node=\color{black}$\vdots$
},
}
\begin{figure}
\centering
\begin{tikzpicture}[x=1.5cm, y=1.5cm, >=stealth]
\foreach \m/\l [count=\y] in {1,2,3,missing,4}
\node [every neuron/.try, neuron \m/.try] (input-\m) at (0,2.5-\y) {};
\foreach \m [count=\y] in {1,missing,2}
\node [every neuron/.try, neuron \m/.try ] (hidden-\m) at (2,2-\y*1.25) {};
\foreach \m [count=\y] in {1}
\node [every neuron/.try, neuron \m/.try ] (output-\m) at (4,.5-\y) {};
\foreach \l [count=\i] in {1,2,3,p}
\draw [<-] (input-\i) -- ++(-1,0)
node [above, midway] {$\varepsilon^{i}_{t_\l}$};
\foreach \l [count=\i] in {1,n^{H}_h}
\node [above] at (hidden-\i.north) {$h_{\l}$};
\foreach \l [count=\i] in {1}
\draw [->] (output-\i) -- ++(1,0)
node [above, midway] {$a^{i}$};
\foreach \i in {1,...,4}
\foreach \j in {1,...,2}
\draw [->] (input-\i) -- (hidden-\j);
\foreach \i in {1,...,2}
\foreach \j in {1}
\draw [->] (hidden-\i) -- (output-\j);
\foreach \l [count=\x from 0] in {Input, Hidden, Ouput}
\node [align=center, above] at (\x*2,2) {\l \\ layer};
\end{tikzpicture}
\caption{Architecture of the network computing the anomaly score $a^{i}$ assuming $H=1$, with the reconstruction error $\bm{\varepsilon}^{i}$ as input.}
\label{fig:NNarchi}
\end{figure}
\subsection{Anomaly localization Step}
\label{subsec:Anomalylocalization}
Once the time series containing an anomaly have been identified, the second step of our approach aims at locating an abnormal observation (see Section \ref{subsection:modeldescription}) among each contaminated time series. Again we use the reconstruction errors defined in \eqref{eq:recerror} as inputs of this second step. The difference lies on the transformation these reconstruction errors undergo before labelling the different time stamp observations.
Namely, we now consider the following element-wise transformation of the reconstruction errors:
\begin{equation*}
F(\varepsilon^{i}) = \lvert \varepsilon^i \rvert = \lvert X^{i} - \widehat{X}^{i} \rvert.
\end{equation*}
The model inputs $F(\varepsilon^{i}) \in \mathbb{R}^{p}_{+}$ is thus assigned to each time series $X^{i} \in \mathbb{R}^{p}$. The time stamp of occurrence of the anomaly is given by
\begin{align*}
\widehat{L}^{i} := \arg\max_{j} \left\{F(\varepsilon^{i}_{t_j}) : j \in \{ 1,\ldots, p \}\right\}.
\end{align*}
Algorithm \ref{algo: localizationModel} recaps the anomaly localization model.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Anomaly Localization Model}
\label{algo: localizationModel}
\SetKwInOut{Require}{inputs}
\Require{time series $X^i$ to analyze.}
$\varepsilon^{i} \leftarrow \mathrm{PCAFeaturesExtraction}(X^i)$
$\widehat{L}^i \leftarrow \arg\max_{\jmath} \left\{\lvert \varepsilon^{i}_{t_\jmath}\rvert : \jmath \in \{ 1,\ldots, p \}\right\}\;$
\SetKwInOut{Output}{outputs}
\Output{anomaly location $\widehat{L}^i$.}
\end{algorithm}
Note that anomalies are not necessarily extrema. For this reason the above PCA feature extraction step is necessary. Indeed, building on the reconstruction errors, the model identifies these extrema-anomalies, but also abnormal observations which are not necessarily extrema. This subtlety of the nature of the observations underlines the importance of going further than taking the index of the highest observed value of the contaminated time series $X^i$.
\tikzstyle{decision} = [diamond, draw, fill=gray!20,
text width=4.5em, text badly centered, node distance=4cm, inner sep=0pt]
\tikzstyle{corner} = [rectangle, draw, fill=white!20,
text width=2em, text centered, rounded corners, minimum height=0.1em, minimum width=0.1em,node distance=3cm]
\tikzstyle{block} = [rectangle, draw, fill=gray!20,
text width=6em, text centered, rounded corners, minimum height=4em,node distance=3cm]
\tikzstyle{blocktransp} = [rectangle, draw,
text width=6em, rounded corners,minimum width=5cm, minimum height=1cm,node distance=3cm]
\tikzstyle{blockEnd} = [rectangle, draw, fill=gray!20,
text width=2em, text centered, rounded corners, minimum height=0.5em, minimum width=0.25em,node distance=3cm]
\tikzstyle{line} = [draw, -latex']
\tikzstyle{linecust} = [draw]
\tikzstyle{cloud} = [draw, ellipse,fill=red!40, node distance=3cm,
minimum height=2em]
\begin{figure}[H]
\centering
\begin{tikzpicture}[node distance = 2cm,auto]
\node [block] (step1) {Features extraction};
\node [cloud, left of=step1] (input) {$X$};
\node [cloud, right of= step1] (output) {$\varepsilon$};
\node [block, below of= step1] (stepad1) {Identification};
\node [cloud, left of= stepad1] (inputad1) {$\varepsilon$};
\node [cloud, right of= stepad1] (outputad1) {$\widehat{A}$};
\node [decision, below of= stepad1] (suspornot) {Is $X$ contaminated?};
\node [cloud, left of= suspornot](inputdec) {$\widehat{A}$};
\node [blockEnd, right of= suspornot] (outputdecno) {End};
\node [block, below of = suspornot] (stepad2) {Localization};
\node [cloud, left of = stepad2] (inputad2) {$\varepsilon$};
\node [cloud, right of = stepad2] (outputad2) {$\hat{L}$};
\node [block, below of = stepad2] (stepad3) {Imputation};
\node [cloud, left of = stepad3] (inputad3) {$X,\hat{L}$};
\node [cloud, right of = stepad3] (outputad3) {$\widetilde{X}$};
\node [corner, below of= outputad3] (endModel) {};
\node [corner, below of= inputad3] (endModel2) {};
\node [corner, left of= endModel2] (endModel3) {};
\node [corner, left of= input] (endModel4) {};
\path [line,dashed] (input) -- (step1);
\path [line,dashed] (step1) -- (output) ;
\path [line](step1)--(stepad1);
\path [line,dashed] (inputad1) -- (stepad1);
\path [line,dashed] (stepad1) -- (outputad1);
\path [line] (stepad1) -- (suspornot);
\path [line,dashed] (inputdec) --(suspornot);
\path [line, dashed] (suspornot) -- node {no} (outputdecno);
\path [line,dashed] (suspornot) -- node {yes} (stepad2);
\path [line,dashed] (inputad2) -- (stepad2);
\path [line, dashed] (stepad2) -- (outputad2);
\path [line] (stepad2) -- (stepad3);
\path [line,dashed] (inputad3) -- (stepad3);
\path [line, dashed] (stepad3) -- (outputad3);
\draw (outputad3) -- (endModel);
\draw (endModel) -- (endModel2);
\draw (endModel2) -- (endModel3);
\draw (endModel3) -- (endModel4);
\path [line] (endModel4) -- (input);
\end{tikzpicture}
\caption{Flow chart of our two step anomaly detection model (PCA NN), depicting the process a time series $X$ goes through. }
\label{fig:FlowChartModel}
\end{figure}
The flow chart of Figure \ref{fig:FlowChartModel} along with Algorithm \ref{algo: AnomalyDetectionModel} summarize our approach. When a time series $X$ is given to our model, we suggest a new representation of $X$, namely $\varepsilon$, through a features engineering step involving PCA. The resulting representation $\varepsilon^i$ feeds the first component of the model, i.e. the identification step, which evaluates the time series likelihood of being contaminated. The optimal parameter $\widehat{\Theta}$ of the identification model solves
\[
\widehat{\Theta} = \arg \min_{\Theta=\left(\bm{W},\bm{b},s\right)} \sum_{X\in\bm{X}} \mathcal{L}_{\Theta} \left( A, \hat{A}\right),
\]
where $\hat{A} = \mathds{1}_{F\left((\Omega\Omega^\top - I_{p})X\right)>s}$ and $\mathcal{L}_{\Theta}$ is the loss function defined in \eqref{eq:myLoss}. Then, if the model considers the time series as contaminated, the localization model takes over to localise the abnormal value in $X$. For this second step of the model the time stamp of occurrence of the anomaly is given by
\[
\arg\max_{j} \quad \left\{\left( X \left(\Omega^\top\Omega - I_{p}\right) \right)_j : j \in \{ 1,\ldots, p \}\right\}
\]
Finally, the anomaly is imputed. As our approach integrates the computation of a reconstruction of the time series, we could replace the anomaly with the corresponding reconstructed value. The model will be then able to detect the anomaly and suggest an imputation value. However, numerical tests described in Section \ref{subsection:imputation} show that imputations with naive approaches perform better.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Anomaly Detection Model }
\label{algo: AnomalyDetectionModel}
\SetKwInOut{Require}{inputs}
\Require{time series $X^i$ to analyze, \\calibrated identification model parameters}
$\widehat{A}^i \leftarrow \mathrm{IdentificationModel}(X^i,\widehat{\Theta})$ \tcp*[f]{Algorithm \ref{algo: IdentificationModel}} \\
\If{$\widehat{A}^i=1$}{
$\widehat{L}^i \leftarrow \mathrm{LocalizationModel}(X^i)$ \tcp*[f]{Algorithm \ref{algo: localizationModel}}
}
\SetKwInOut{Output}{outputs}
\Output{identification label $\widehat{A}^i$,}
\hspace{42pt} anomaly localization $\widehat{L}^i$.
\end{algorithm}
\section{Data}\label{section: data}
Since anomalies are rare by definition, real world data sets are very imbalanced, in the sense that the proportion of anomalies compare to normal observations is very small, making the learning phase of the model difficult. This has led us to consider synthetic data for the model calibration. Our data set is obtained through a three-step process including time series simulations, contamination and data augmentation. We point out that, in the data simulation, care was taken to ensure that the generated data sets stay realistic. Particularly, only few anomalies were added to time series as described in Section \ref{subsec:DataContamination}. To this extent, we still face the problem of scarcity of anomalies in this synthetic framework, and we provide some preprocessing steps to sidestep this issue as well.
\subsection{Data Simulation}
\label{subsection:datasim}
In this section, we describe the model used to simulate the data. Recall that our primary motivation is to detect anomalies in financial time series. For that purpose, we consider share price sample paths generated through the Black and Scholes model i.e. geometric Brownian motions. Under this framework, the share price $(S_t)_t$ is solution to the stochastic differential equation
\begin{align}
dS_t = S_t \mu dt + S_t \sigma dW_t,
\label{eq:bs}
\end{align}
where $W$ represents a standard Brownian motion, $\mu$ the drift and $\sigma$ the volatility of the stock. Applying Itô's lemma to the process defined by $\log(S)$, yields
\begin{align}
S_t = S_0 \exp\left(\left(\mu - \frac{1}{2}\sigma^2\right)t + \sigma W_t\right).
\label{eq:diffeq}
\end{align}
Let $N$ stocks $S^{1}, \ldots, S^{N}$ simulated simultaneously from this model, with $S^{i}:= \left\{S_{t_0}^{j},S_{t_1}^{j},\dots,S_{t_T}^{j}\right\}$ the time series of length $T$ representing the (time discretized) path diffusion of the $j$-th stock. Each stock $S^i$ has its own drift $\mu^i$, volatility $\sigma^i$ and initial value $S^i_{t_0}$. The paths parameters are selected randomly according to
\begin{align}
S_{t_0}^i \sim \mathcal{N}\left(100,1\right), \quad \quad \mu^i \sim \mathcal{U}\left([0.01,0.2]\right), \quad \quad \sigma^i \sim \mathcal{U}\left([0.01,0.1]\right), \quad \, i=1,\ldots, N .
\label{eq:genparams}
\end{align}
For the sake of realism the Brownian motions driving the $N$-stocks are correlated.
We assume that the time series obtained thanks to the paths diffusion do not contain any type of anomalies. The contamination of these time series by anomalies is described in the next section.
\subsection{Time series Contamination}
\label{subsec:DataContamination}
We apply a quite naive approach to introduce anomalies into our time series. We introduce the same fixed number of anomalies $n^{anomaly}$ to each time series by applying a shock on some original values of the observed time series. Formally, the $\jmath$-th added anomaly is characterized by its location $t_{\jmath}$ corresponding to the time stamp at which the anomaly has occurred, its shock $\delta_{\jmath}$, the amplitude of the shock is given by $|\delta_{\jmath}|$ and its sign by $\mathrm{sgn}\left(\delta_{\jmath}\right)$. We denote by $S^{a,i}$ the time series resulting from the contamination of the $i$-th clean time series $S^i$. For $i \in \{1,\dots ,N \}$, let $\mathcal{J}^i$ be the set of indices of the time stamps at which an anomaly occurs for the $i$-th time series. For $\jmath \in \mathcal{J}^i$, the abnormal values are
\begin{align*}
S_{t_{\jmath}}^{a,i} = S_{t_{\jmath}}^{i} \left( 1 + \delta_{\jmath}\right).
\end{align*}
The location, sign and amplitude of the shocks are generated randomly according to uniform distributions:
\begin{align*}
\mathcal{J} \sim \mathcal{U}_{N,n^{anomaly}}\left(\{1,\ldots,p\}\right), \quad \quad \mathrm{sign}\left(\delta\right) \sim \mathcal{U}_{N,n^{anomaly}}\left(\{-1,1\}\right), \quad \quad \lvert \delta\rvert \sim \mathcal{U}_{N,n^{anomaly}}\left([0,\rho]\right),
\label{eq:contamParams}
\end{align*}
where $\rho$ is an upper bound on the shock amplitude.
We define the anomaly mask matrix $\mathbb{R}^{N \times T}$ by setting, for $i=1,\ldots,20$, and $j=1,\ldots, T$,
\begin{equation}
\label{eq:anomalymask}
\mathcal{T}_{i,j} = \begin{cases}
1+ \delta_j & \text{if $j \in \mathcal{J}^i$.} \\
1 & \text{otherwise.}
\end{cases}
\end{equation}
Under this framework, where we incorporate the anomalies to the clean time series driven by the geometric Brownian motion, we assign labels to the time series observations according to whether the values correspond to anomalies or normal observations. We thus provide the labels $Y^{i}_{t_j} $ associated with each value $S_{t_j}^{a,i}$. Hence, for $i=1,\ldots,N$,
\begin{align}
Y^i_{t_j} = \begin{cases}
1 & \text{if $j \in \mathcal{J}^i$.} \\
0 & \text{otherwise.}
\end{cases}
\label{eq:getlabels}
\end{align}
Algorithm \ref{algo: DataContamination} recaps the time series contamination procedure, where AnomalyMask and GetLabels are names for the operators defined by \eqref{eq:anomalymask} and \eqref{eq:getlabels}.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Data Contamination }
\label{algo: DataContamination}
\SetKwInOut{Require}{inputs}
\Require{mplitude range $\rho$,\\ number of anomalies to add in time series $n^{anomaly}$,\\ sata $\bm{S}$}
$\mathrm{sgn}\left(\delta\right)
\leftarrow \mathcal{U}_{N,n^{anomaly}}\left(\{-1,1\}\right)$\\
$ \lvert\delta\rvert
\leftarrow \mathcal{U}_{N,n^{anomaly}}\left(\{0,\rho\}\right)$\\
$ \mathcal{J} \leftarrow \mathcal{U}_{N,n^{anomaly}}\left(\{1,\ldots,p\}\right)$\\
$\mathbb{R}^{N \times T} \ni\mathcal{T} \leftarrow \mathrm{AnomalyMask}(\mathcal{J},\delta)$\\
$\bm{S}^{a} \leftarrow \bm{S} \circ \mathcal{T}$\\
$Y \leftarrow \mathrm{GetLabels}\left(\mathcal{J}\right)$
\SetKwInOut{Output}{outputs}
\Output{time series with anomalies $\bm{S}^{a}$,}
\hspace{42pt} labels associated with each value of the time series $\bm{L}$.
\end{algorithm}
\subsection{Data Augmentation}
By definition, anomalies are rare events and thus represent only a low fraction of the data set. Yet, the suggested approach needs an important training set for an efficient learning. To overcome this issue, we apply a classical data augmentation technique called the sliding window method. This method not only extends the number of anomalies within the data set, but also allows the model to learn that anomalies could be located anywhere in the time series. The sliding window technique as described by \citep{le2016data} consists in extracting $N_{p} = T- p + 1$ sub-time series of length $p$ of the initial observed time series of length $T$.
Note that we have to split the data into train and test sets before augmentation to guarantee that time series considered in the training set do not share any observation with the ones we use for the model evaluation. In view of simplification, we introduce the data augmentation process, without loss of generality, for $\bm{S}$ and $\bm{Y}$, but one should keep in mind that this process has to be applied to $\bm{S^{Train}},\bm{Y^{Train}}$ and $\bm{S^{Test}},\bm{Y^{Test}}$ separately.
For $i=1,\ldots, N$ and $q \in \{1,\ldots,N_{p}\}$, the sub-time series $S^{i,q}$ and the associated labels $Y^{i,q}$ are defined by (see Algorithm \ref{algo: Slidding})
\begin{align*}
S^{i,q} &= \left\{S_{t_j}^{i} \, , \ j =q,\dots,q + p - 1 \right\} \\
Y^{i,q} &= \left\{Y_{t_j}^{i} \, , \ j =q,\dots,q + p - 1 \right\} .
\end{align*}
Each sub-time series $S^{i,q+1}$ thus results from the shift forward in time of one observation of the previous sub-time series $S^{i,q}$. The final data set $\bm{X}$ and the labels $\bm{Y^s}$ are then defined as the matrices which rows correspond to the sub-time series $S^{i,q}_{q=1,\ldots, N_p; i= 1,\ldots,N}$, and $Y^{i,q}_{q=1,\ldots, N_p; i= 1,\ldots,N}$, respectively, i.e.
\[
\bm{X} =
\left(
S^{1,1}, S^{1,2}, \ldots , S^{1,N_{p}}, S^{2,1} , \ldots , S^{i,q}, \ldots, S^{N,N_{p}}
\right)^{\top},
\]
\[
\bm{Y_s} =
\left(
Y^{1,1}, Y^{1,2}, \ldots , Y^{1,N_{p}}, Y^{2,1} , \ldots , Y^{i,q}, \ldots, Y^{20,N_{p}}
\right)^{\top}.
\]
With this data configuration, an observation refers to a time series $S^{i,q}$ obtained through the sliding window technique. Two observations may represent the same stock but on different time intervals.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Data Augmentation with sliding window technique}
\label{algo: Slidding}
\SetKwInOut{Require}{inputs}
\Require{window size $p$,}
\hspace{42pt} data and labels $\bm{S}$,$\bm{Y}$; \\
Initialise empty slided time series and labels matrices $\bm{X}$, $\bm{Y_s}$;\\
\For{i :=1 to N}{
\For{q :=1 to $N_{p}$}{
$S^{i,q} \leftarrow \left\{S_{t_j}^i | j \in \{q,\dots,q+p-1\}\right\}$\\
$Y^{i,q} \leftarrow \left\{Y_{t_j}^i | j \in \{q,\dots,q+p-1\}\right\}$\\
}
$\bm{X} \leftarrow \mathrm{Concatenate}\left(\bm{X}, S^{i,q}\right)$
$\bm{Y_s}\leftarrow \mathrm{Concatenate}\left(\bm{Y_s}, Y^{i,q}\right)$\\
}
\SetKwInOut{Output}{outputs}
\Output{resulting slided time series and labels $\bm{X}, \bm{Y_s}$}
\end{algorithm}
While the fact that the observations share the same values may be argued to wrongly impact the learning process, we point out that a real benefit can be drawn from this situation. Indeed, thanks to this sliding window technique, the number of anomalies is considerably increased. This technique also allows the model to learn that anomalies could be located anywhere in the time series, reducing the dependency on the event location \citep{um2017data}.
However, once we apply the sliding window technique, we do not only extend the number of contaminated time series: the number of uncontaminated time series is also increased. But the minority class (contaminated time series) has, this time, a significant number of instances denoted by $N^c$. Therefore, in order to get a balanced data set for the training set, we perform an undersampling, selecting randomly $N^c$ observations from the $N^u$ uncontaminated time series without any anomalies (RandomSampling in Algorithm \ref{algo: Pruning}). It is important to point that the resulting retained number of observations $2N^c$ is more than enough to train the model. The test set, in turn is umbalanced. To sharpen the imbalanced characteristic of the test set we specify a contamination rate $r_c$ which corresponds to the rate of contaminated time series in the data set. The construction of the test set is described in Algorithm \ref{algo: Pruning}.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Time series selection }
\label{algo: Pruning}
\SetKwInOut{Require}{Inputs}
\Require{slided data and labels $\bm{X}$,$\bm{Y_s}$,}
\hspace{42pt} contamination rate $r_c$ (for test set) \\
\medskip
$\bm{X},\bm{Y_s} \leftarrow \bm{X}\mathds{1}_{\mathrm{sum}(\bm{Y_s})\leq 1},\bm{Y_s}\mathds{1}_{\mathrm{sum}(\bm{Y_s})\leq 1}$ \tcp*{Keep time series with at most one anomaly.}
\If{Train set}{
$\\ N^c = \mathrm{card}\left(\bm{Y_s}\mathds{1}_{\mathrm{sum}(\bm{Y_s})= 1}\right)$\\
\tcp*[f]{Randomly select the indexes of $N^c$ uncontaminated time series} \; \\
$\mathrm{index}^u = \mathrm{RandomSampling}\left(\left\{i; \; \mathrm{sum}(Y_s^i)=0, i \in \{1,\ldots,NN_p\}\right\}, N^c\right)$\\}
\If{Test set}{
$\\ N^c = \mathrm{card}\left(\bm{Y_s}\mathds{1}_{\mathrm{sum}(\bm{Y_s})= 1}\right)$\\ \;
$N^u = \left \lceil \frac{N^c(1-r_c)}{r_c}\right\rceil$\\
\tcp*[f]{Randomly select the indexes of $N^c$ uncontaminated time series} \; \\
$\mathrm{index}^u = \mathrm{RandomSampling}\left(\left\{i; \; \mathrm{sum}(Y_s^i)=0, i \in \{1,\ldots,NN_p\}\right\}, N^u\right)$\\
}
$\\ \mathrm{index}^c =\left\{i; \; \mathrm{sum}(Y_s^i)=1, i \in \{1,\ldots,NN_p\}\right\}$\\
$\bm{X} \leftarrow \left(X^i, \mathrm{for} \; i \; \in \; \mathrm{index}^u \cup \mathrm{index}^c\right)$\\
$\bm{Y_s} \leftarrow \left(Y_s^i, \mathrm{for} \; i \; \in \; \mathrm{index}^u \cup \mathrm{index}^c\right)$\\
\SetKwInOut{Output}{outputs}
\Output{time series $\bm{X}$ with at most one anomaly, \\
corresponding labels for Identification task $\bm{A}$,\\
corresponding labels for localization task $\bm{L}$.}
\end{algorithm}
As mentioned in Section \ref{subsection:modeldescription}, the model is designed to predict the localization of only one anomaly. If there is more than one anomaly in the time series, this will wrongly impact the evaluation of the anomaly localization model. Therefore at this stage we only consider time series with at most one anomaly.
Hence before introducing the labels transformation, it is worth stressing that the following assumption is being made:
\begin{assumption}
A time series $S^{i,q}$ contains at most one anomaly among all its observed values.
\end{assumption}
We assign the identification label $A^{i,q}$ (see Section \ref{subsec:MethodNotation}) to each $S^{i,q}$ following the rule
\begin{equation*}
A^{i,q} = \sum_{j=q}^{q+p -1} Y^{i}_{t_j} = \begin{cases}
1 & \text{if there is an anomaly among the observed values of $S^{i,q}$} \\
0 & \text{otherwise.}
\end{cases}
\end{equation*}
Regarding the location labels, we recall that they only concern the time series with an anomaly, therefore $L^{i,q}$ is defined following
\begin{equation}
L^{i,q} = \arg\max_{j} Y^{i,q} =\arg\max_{j} \left\{Y_{t_j}^{i} \, , \ j =q,\dots,q + p - 1 \right\}
\label{LabelFromY}
\end{equation}
With Algorithm \ref{algo: Labeling}, we give a rundown of the construction process of the identification and localization labels namely $\bm{A}$ and $\bm{L}$ departing from $\bm{Y_s}$.
The supervised learning framework is adopted herein, since we have at our disposal labelled data.
The resulting matrices with the observed values of the time series in $\bm{X}$ the associated identification labels in $\bm{A}$ and location labels in $\bm{L}$ constitute the data set used for our model calibration and evaluation.
\begin{algorithm}[H]
\SetAlgoLined
\caption{Time series labelling}
\label{algo: Labeling}
\SetKwInOut{Require}{Inputs}
\Require{slided labels $\bm{Y_s}$;}
$\bm{A} \leftarrow \mathrm{sum}(\bm{Y_s})$ \; \\
$\bm{L} \leftarrow \arg\max \left\{\bm{Y_s}\mathds{1}_{\mathrm{sum}(\bm{Y_s}) = 1}\right\}$\\
\SetKwInOut{Output}{outputs}
\Output{corresponding labels for identification task $\bm{A}$,\\
corresponding labels for localization task $\bm{L}$.}
\end{algorithm}
\section{Model Evaluation}\label{section: model_eval}
We briefly describe the process of the latent space dimension calibration.
Then we evaluate the performance of the identification and the localization models on synthetic data, using appropriate performance indicators. And we demonstrate numerically the efficiency of the PCA NN over baseline anomaly detection algorithms. To take our approach a leap beyond classical anomaly detection algorithms, we evaluate the anomalies imputation suggest by the PCA NN model. Lastly, we test the robustness of the calibrated cut-off value.
\subsection{Performance Metrics}
\label{subsection:perfmetrics}
Common methods to assess the performance of binary classifiers include true positive and true negative rates, and ROC (Receiver Operating Characteristics) curves, which display the true positive rate against the false positive rate. These methods, however, are uninformative when the classes are severely imbalanced. In this context, $F_1-$score and Precision-Recall curves (PRC) have been shown to be more informative \citep{brownlee2020imbalanced,prc}. They are both based on the values of
\[
\mathrm{Precision} (s)= \frac{\text{true positives}}{\text{true positives} + \text{false positives}}
\]
against the values of
\[
\mathrm{Recall} (s) = \frac{\text{true positives}}{\text{true positives} + \text{false negatives}} \,
\]
where $s$ is a cut-off probability varying between $0$ and $1$.
Precision quantifies the number of correct positive predictions out all positive predictions made; and Recall (often also called Sensitivity) quantifies the number of correct positive predictions out of all positive predictions that could have been made. Both focus on the Positives class (the minority class, anomalies) and are unconcerned with the Negatives (the majority class, normal observations).
The $F_1-$score combines these two measures in a single index by taking the harmonic mean of those two values. It is derived from the F-Measure introduced in \citep{fmes1, fmes2} and defined as
\begin{equation}
F_1 = 2\cdot \frac{ \mathrm{Precision}\cdot \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}.
\end{equation}
To closer the $F_1-$score is to 1, the better the prediction model is.
PRC display the values of Precision and Recall as the cut-off $s$ varies from $0$ to $1$. The PR curve of a skill-ful model bows towards the point with coordinates $(1,1)$. The curve of a no-skill classifier will be a horizontal line on the plot with a y-coordinate proportional to the number of Positives in the data set. For a balanced data set this will be 0.5 \cite{brownlee2020imbalanced}.
PRC and $F_1-$score are complementary in our approach. The PRC is used on the anomaly scores outcomes of the models, it gives the best configuration and the best model, whereas the $F_1-$score is used to select the best cut-off value used in the prediction of the two classes, for each model.
\subsection{Synthetic Data Set}
The anomaly detection task is performed on $N=20$ stocks simultaneously. The stock prices are diffused according to the Black and Scholes model, each stock has its own drift and volatility and the $20$ stocks are correlated, as described in Section \ref{section: data}. Each time series represents $T=1,500$ daily stock prices, split into two sets: $1,000$ observations, corresponding to the train set, are used to learn the model parameters, that is the PCA transfer matrix and the NN weights and biases. The last $500$ observations, corresponding to the test set, are used to assess the quality of the estimated parameters when applied to unseen samples. Hence, for the application of the sliding window technique, we consider that the length of the resulting time series $p$ is 206. Finally, we obtained $\bm{X}^{Train} \in \mathbb{R}^{12,000\times 206}$ and $\bm{X}^{Test} \in \mathbb{R}^{2,500\times 206}$, with $6,000$ contaminated time series in the train set and $400$ in the test set.
\Table{\ref{tab:compo}} sums up the composition of each data set before and after data augmentation. Thanks to data augmentation, we are able to extend the number of time series in both train and test sets, going from 20 time series of length 1,000 with 4 anomalies each to 12,000 time series of length 206, among which 6,000 time series are contaminated for the train set for example.
\begin{table}[h!]
\begin{center}
\begin{tabular}{c}
\begin{tabular}{l|c|c|c}
& Nb of Time Series & Nb of Observed values per time series & Nb of anomalies \\
\hline
Train set & 20 & 1,000& $4 \times 20$\\
Test set & 20& 500 & $2 \times 20$
\end{tabular}\\~\\
a)
\\~\\
\begin{tabular}{l|c|c|c}
& Nb of Time Series & Nb of Observed values per time series & Nb of anomalies \\
\hline
Train set & 12,000 & 206& 6,000\\
Test set & 2,500& 206 & 400\\
\end{tabular}\\~\\
b)
\end{tabular}
\end{center}
\caption{Data set composition a) before and b) after data augmentation}
\label{tab:compo}
\end{table}
We recall that both steps of the model are preceeded by a feature extraction step, for which the latent space dimension calibration is described in Section \ref{subsec:caliK}. The features extraction guarantees the stationarity of the time series used for the anomaly detection task, namely $\bm{\varepsilon}$, as shown by the numerical results provided in Appendix \ref{apx:AppendixB}.
\subsection{Calibration of the Latent Space Dimension}
\label{subsec:caliK}
The dimension $k$ of the latent space in the PCA algorithm needs to be specified. When performing PCA, $k$ is determined through a scree plot, which is the representation of the proportion of variance explained by each component. The optimal $k$ corresponds to the number of principal components explaining a given level of the variance of the original data. However, this method has its limitation as stated in \citep{linting2007nonlinear}, and more importantly it is not suitable for our approach. Indeed, in our case, the number of selected principal components must achieve a given trade-off between information and noise in the latent space. If we consider a too low number of principal components, we may lose information regarding the normal observations, leading to false alarms. If a too high number of principal component is retained, we may include components representing noise, which prevents the model from detecting some anomalies.
To select the optimal dimension of the latent space, we consider the distribution of anomaly score obtained through the application of naive approach Section \ref{subsec:Step1}, empirically calibrating a cut-off $\hat{s}$ through the non parametric estimation of the distribution of the anomaly scores of uncontaminated and contaminated time series. We chose to calibrate the dimension of the latent space with the naive approach, because using the NN to this end would be very costly. Hence, for each $k = 5,10,\dots,200$, we construct a PCA model from which we infer reconstruction errors which are then converted into anomaly scores. Based on these anomaly scores, we tune the cut-off value $\hat{s}$ thanks to the distributions and finally convert the scores into labels. We evaluate the predictions of the naive approach for each value of $k$. The results on the evaluation metrics on the train set, as legitimate to choose $k$, are represented on Figure \ref{fig:kselection}. The highest values are reached for $k \in \{40,\dots,145\}$. The performance seems to be stable in terms of F1-score and accuracy. Therefore, for computational reasons we choose the optimal $k$ to be 40.
\begin{figure}[H]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.4\textwidth]{Graphs/latentS/latentspaceTrain.png} &
\includegraphics[width=0.4\textwidth]{Graphs/latentS/latentspaceTest.png}\\
a) & b)
\end{tabular}
\caption{Performance metrics on train a) and test b) sets with respect to the number of principal component $k$.}
\label{fig:kselection}
\end{figure}
Figure \ref{fig:kselection} also shows, as expected, a downward trend of the scores when represented against the highest values of $k$. This demonstrates that when a high level of variance is explained, it become much harder to perform anomaly detection based on the reconstruction errors.
\subsection{Contaminated Time Series Identification Step}
\label{subsection:evalstep1}
For the feature extraction step, we considered a latent space dimension $k=40$. The NN built to compute the anomaly scores and convert them into labels was calibrated on the train set. The result of this calibration is shown in Table \ref{tab:ResCaliIdentification}.
\begin{table}[H]
\centering
\begin{tabular}{l|l|l|l|l}
Data set & Accuracy &Precision & Recall &F1-score \\
\hline
Train set & 90.97 \% & 97.36\% & 84.21\% & 90.31 \% \\
Test set & 88.58\% & 61.26\% & 85.27\% & 71.30\% \\
\end{tabular}
\caption{Performance evaluation of suggested model on synthetic data set for identification step}
\label{tab:ResCaliIdentification}
\end{table}
Figure \ref{fig:detecsusp} shows two contaminated time series identified as such by the model. Figure \ref{fig:detecnon} displays two examples of time series without anomalies accurately identified by the model. Figure \ref{fig:detecmis} displays two time series misidentified by the model.
\begin{figure}
\centering
\begin{tabular}{c}
\includegraphics[scale=0.3]{Graphs/Detection/iden_visu_ano1.pdf} \\
a) \\
\includegraphics[scale=0.3]{Graphs/Detection/iden_visu_ano2.pdf}\\
b)
\end{tabular}
\caption{Two examples a) and b) of contaminated time series accurately identified by the model. The stock path and the reconstruction errors are represented in black and brown. The red cross shows the anomaly localization.}
\label{fig:detecsusp}
\end{figure}
When an observation deviates significantly from the rest of the time series values, the model is able to recognise that the concerned time series contains an abnormal observation.
\begin{figure}[H]
\centering
\begin{tabular}{c}
\includegraphics[scale=0.3]{Graphs/Detection/iden_visu_norm1.pdf} \\
a) \\
\includegraphics[scale=0.3]{Graphs/Detection/iden_visu_norm2.pdf} \\
b)
\end{tabular}
\caption{Two examples a) and b) of uncontaminated time series accurately identified by the model. The stock path and the reconstruction errors are represented in black and brown.} \label{fig:detecnon}
\end{figure}
In Figure \ref{fig:detecnon} with uncontaminated time series, we see that even when there is a local upward trend in the time series values, the model is able to make the distinction between this market move and the occurrence of an anomaly.
\begin{figure}[H]
\centering
\begin{tabular}{c}
\includegraphics[scale=0.3]{Graphs/Detection/iden_visu_mispred2.pdf} \\
a)\\
\includegraphics[scale=0.3]{Graphs/Detection/iden_visu_mispred1.pdf} \\
b)
\end{tabular}
\caption{Two examples a) and b) of time series misidentified by the model. The stock path and the reconstruction errors are represented in black and brown. The red cross shows the anomaly localization.} \label{fig:detecmis}
\end{figure}
The last set of time series displays some limits of the model. The graphs in Figure \ref{fig:detecmis}a) represent the situation where the model did not manage to identify the contaminated time series. One explanation could be that the size of the anomaly is not significantly large enough to be spotted by the model. Indeed looking at the graph the time series does not seem to contain any anomaly. In contrast, the graphs in Figure \ref{fig:detecmis}b) shows a stock price path predicted to be contaminated whereas it is not.
We illustrate the robustness of the approach by assessing the model predictions on $100$ distinct data sets. \Table{\ref{tab:multiRunsIdent}} shows the mean and standard deviation over the multiple runs.
\begin{table}[H]
\centering
\begin{tabular}{l|l|l|l|l}
Data set & Accuracy &Precision &Recall &F1-score \\
\hline
Train set & 79.02\% (2.4\%)& 78.69\% (2.7\%) & 79.74\% (4.6\%)& 79.13\% (2.7\%) \\
Test set & 77.79\% (3.9\%) & 41.87\% (5.2\%) & 80.16\% (6.4\%) & 54.82\% (5.2\%) \\
\end{tabular}
\caption{Mean (standard deviation) of performance metrics of the suggested model over multiple runs for the identification step.}
\label{tab:multiRunsIdent}
\end{table}
\subsection{Anomaly Localization Step}
\label{subsection:evalstep2}
Regarding the localization step, the dummy approach consists in taking the argmax of time series observations as the anomaly location. We distinguish between two cases, anomalies which are extrema and anomalies which are not. The quality of the detection of our model is assessed for both types of anomalies. The results are displayed in Tables {\ref{tab:ResCaliLocAll}} and \ref{tab:ResCaliLocNE}.
\begin{table}[H]
\centering
\begin{tabular}{l|l|l|l|l}
Data set &Accuracy &Precision &Recall &F1-score \\
\hline
Train set & 89.65\% (34.87\%)& 89.81\% (40.53\%) & 89.65\% (34.88\%)& 89.68\% ( 36.56\%) \\
Test set & 94.39\% (27.78\%) & 94.49\% (31.52\%) & 94.39\% (27.78\%) & 94.38\% (28.94\%) \\
\end{tabular}
\caption{Performance evaluation of the suggested model and the dummy approach on synthetic data set for the location step on all type of anomalies.}
\label{tab:ResCaliLocAll}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{l|l|l|l|l}
Data set &Accuracy &Precision & Recall &F1-score \\
\hline
Train set & 84.22\% (0\%)& 84.55\% (0\%) & 84.28\% (0\%)& 89.68\% (0\%) \\
Test set & 91.92\% (0\%) & 92.10\% (0\%) & 91.92\% (0\%) & 91.90\% (0\%) \\
\end{tabular}
\caption{Performance evaluation of the suggested model and the dummy approach on synthetic data set for the location step on non extrema anomalies.}
\label{tab:ResCaliLocNE}
\end{table}
Numerical tests show the necessity of the feature extraction step for the anomaly location task, as applying the dummy approach alone is not enough when the anomaly is not an extreme value. Figure \ref{fig:locavisu} represents a stock price time series with the true and predicted anomaly location. In these examples, the locations are accurately predicted.
\begin{figure}[H]
\centering
\begin{tabular}{cc}
\includegraphics[scale=0.3]{Graphs/LocalisationVisu/location_visu1.pdf} &
\includegraphics[scale=0.3]{Graphs/LocalisationVisu/location_visu2.pdf}\\
a) & b)
\end{tabular}
\caption{Anomaly localization prediction on two distinct time series a) and b).} \label{fig:locavisu}
\end{figure}
To guarantee the robustness of the model on the anomaly location, we evaluate the model prediction on $100$ data sets. \Table{\ref{tab:multiRunsLocALL}} shows the mean and standard deviation for the localization of all types of anomaly.
\begin{table}[H]
\centering
\begin{tabular}{c|c|c|c|c}
Data set & Accuracy & Precision & Recall &F1-score \\
\hline
Train set & 89.58\% (4.3\%)& 89.58\% (4.3\%) & 89.96\% (4.1\%)& 89.67\% (4.3\%) \\
Test set & 89.49\% (4.8\%) & 89.49\% (4.8\%) & 90.00\% (4.5\%) & 89.58\% (4.7\%)
\end{tabular}
\caption{Mean (standard deviation) of performance metrics of suggested model over multiple runs for location step on all type of anomalies.}
\label{tab:multiRunsLocALL}
\end{table}
Table \ref{tab:multiRunsLocNE} displays the results on the localization of non extrema anomalies. These results show the importance of the feature extraction step. If we only consider the maximal observed value of the time series to be the anomaly, we would not be able to localise any non-extrema anomaly. Applying Algorithm \ref{algo: localizationModel} instead leads to satisfying anomaly detection rates.
\begin{table}[H]
\centering
\begin{tabular}{l|l|l|l|l}
Data set & Accuracy & Precision & Recall &F1-score \\
\hline
Train set & 85.23\% (6.0\%)& 85.23\% (6.0\%) & 85.87\% (5.7\%)& 85.36\% (6.0\%) \\
Test set & 85.17\% (7.1\%) & 85.17\% (7.1\%) & 86.04\% (6.8\%) & 85.30\% (7.1\%) \\
\end{tabular}
\caption{Mean (standard deviation) of performance metrics of suggested model over multiple runs for location step on non-extreme anomalies.}
\label{tab:multiRunsLocNE}
\end{table}
\subsection{Numerical Results Against Benchmark Models}\label{subsection: benchmarking}
To assess the performance, the suggested approach is compared with well-known machine learning algorithms for anomaly detection, that is isolation forest (IF), local outlier factor (LOF), density based clustering of applications with noise (DBSCAN), k-nearest neighbors (KNN), support vector machine (SVM), reviewed in Section \ref{apx:AppendixA}. We assess their performance on the data set described in Section \ref{section: data}. Their results are given in this section.
Tables \ref{tab:perfmethodsIdentification} and \ref{tab:perfmethodsLocation} summarize the performance on train and test sets of each model for both the contaminated time series identification and anomaly localization steps.
\begin{table}[H]
\centering
\begin{tabular}{l|cc|cc}
& \multicolumn{2}{c|}{Train} & \multicolumn{2}{c}{Test} \\
\hline
Model& Accuracy &F1-score& Accuracy &F1-score \\
\hline
IF &42.31\% & 42.31\% & 69.64\% & 7.022\% \\
LOF &59.95\% & 59.95\% & 90.00\% & 62.41\% \\
DBSCAN & 50.00\% & 66.67\% & 16.65\% & 28.53\% \\ \hline
KNN &94.68\% & 94.39\% &64.82\% &26.69\% \\
SVM & 81.96\% & 82.33\% & 44.39\% &27.44\% \\
PCA NN &90.97\% & 90.31\% & 88.58\% & 71.30\%
\end{tabular}
\caption{Performance evaluation of supervised (bottom) and unsupervised (top) models for contaminated time series identification step. }
\label{tab:perfmethodsIdentification}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{l|cccccc}
Algorithm & IF & LOF & DBSCAN & KNN & SVM &PCA NN \\
\hline
Exec. Time & 0.6915 & 0.2015& 0.1275 & 1.725 & 4.572 & 0.003523 \\
\end{tabular}
\caption{Execution time in seconds for identification of contaminated time series step.}
\label{tab:extimeidentification}
\end{table}
The following paragraphs provide similar conclusions drawn from the results on the identification and location steps (see Tables \ref{tab:perfmethodsIdentification} and \ref{tab:perfmethodsLocation}).
For the unsupervised learning methods, since there is not a proper learning step, even though Tables \ref{tab:perfmethodsIdentification} and \ref{tab:perfmethodsLocation} show the scores on the train and test set, these scores should be seen as the ones obtained by testing the models on independent data sets. For IF, LOF and DBSCAN models, the performance across the sets is not stable, as a significant difference could be observed between the scores on the train and test sets. The poor performance of unsupervised learning algorithms could be explained by the difficulty to estimate the contamination rate for IF and the LOF algorithms. For DBSCAN the poor performance is rather due to the high dimensionality of the data.
Regarding the supervised approaches, although KNN is outperforming the suggested approach on the training set, there is a non-negligible decrease of these scores on the test set. This may suggest an over-fitting on the training data, which makes KNN unable to generalize what it has learnt to unseen samples. The same trend is seen on the scores with our approach, however the loss is no as harsh as in the KNN case.
Hence, according to the results, our PCA based approach seem to be the most suitable method for the problem at stake of anomaly detection on time series. Besides the relatively satisfying performance it shows in terms of accuracy and F1-score, its low computational cost (Tables \ref{tab:extimeidentification} and \ref{tab:extimelocalization}) for both steps, makes the approach stand out.
\begin{table}[H]
\centering
\begin{tabular}{l|cc|cc}
& \multicolumn{2}{c|}{Train} & \multicolumn{2}{c}{Test} \\
\hline
Model & Accuracy & F1-score & Accuracy & F1-score\\
\hline
IF & 89.79\% & 2.296\% & 70.94\% &1.611\% \\
LOF & 99.51\% & 0\% & 2.066\% &0.9816\% \\
DBSCAN & N/A& NA & NA & NA \\\hline
KNN & 99.99\% & 99.99\% & 95.56\% & 2.794\% \\
SVM & NA & NA & NA & NA \\
PCA NN & 89.65\% & 89.68\% & 94.39\% & 94.38\%
\end{tabular}
\caption{Performance evaluation of supervised (bottom) and unsupervised (top) models for anomaly localization step. Results for DBSCAN and SVM are not provided due to high computational cost.}
\label{tab:perfmethodsLocation}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{l|cccccc}
Algorithm & IF & LOF & DBSCAN & KNN & SVM &PCA NN \\
\hline
Exec. Time & 14.50 & 2.287& N/A & 14.19 & N/A& 0.002004 \\
\end{tabular}
\caption{Execution time in seconds for the anomaly localization step.Results for DBSCAN and SVM are not provided due to high computational cost.}
\label{tab:extimelocalization}
\end{table}
\subsection{Anomalies Imputation}\label{subsection:imputation}
To assess the imputation values suggested by our approach, we start from time series simulated without anomalies, then we randomly select a time stamp of the time series and add a noise to the corresponding value following the methodology described in Section \ref{section: data}.
The imputation value the PCA NN suggests is the reconstructed observation. This imputation technique is compared to naive methods of missing values imputation, namely backward fill (BF), consisting in replacing the anomaly by the previous value, and linear interpolation (LI). The choice of these imputation methods is motivated by their low computational cost. In fact, the PCA NN imputation comes at no additional cost as stated before. Therefore, we only challenge its performance with methods with similar complexity. To assess the quality of the imputation value suggested by each approach, we consider the following metrics. The imputation errors $\mathrm{ImputationError}^{i}$ are computed on each time series as
\begin{align*}
\mathrm{ImputationError}^{i} = \sqrt{\sum_j \frac{\left(S^{i}_{t_j} -\widetilde{S}^i_{t_j}\right)^2}{n^{anomaly}}}\; ,
\end{align*}
where $\widetilde{S}^i$ refers to as the $i$-th path price with imputed values. The error on the covariance matrix is computed as
\begin{align*}
\mathrm{ErrCov} = \|\Sigma - \widetilde{\Sigma}\|_{\mathrm{Frob}},
\end{align*}
where $\|\cdot\|_{\mathrm{Frob}}$ is the Frobenius norm, $\Sigma$ the sample covariance matrix and $\widetilde{\Sigma}$ the covariance matrix estimated on the data after the anomalies have been replaced by their respective imputation values.
Each stock path was diffused and contaminated 100 times. The anomalies were then imputed following two baseline approach (BF and LI). Figure \ref{fig:boxPCAvsNone} clearly shows that imputation using the reconstructed values with PCA reduces the error on the estimation of the covariance matrix compared to the situation where the covariance matrix is estimated on the data with anomalies. However, Table \ref{table:CovErrorImputation} shows that the baseline imputation approaches achieve even lower errors on the estimation of the covariance matrix. Figure \ref{fig:boximpcov} shows a higher variance of the errors for the PCA-based imputation approach. In conclusion, it would be rather recommended to replace the flagged anomalies with approaches such as backward fill or linear interpolation.
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{Graphs/Imputation/BoxplotStocksImput.pdf}
\caption{Boxplot representation of the distribution of the errors on time series imputation before the imputation of anomalies (Ano) and after replacing the anomalies; with the reconstructed values suggested by PCA (PCA), using linear interpolation (LI) and using backward fill (BF).}
\label{fig:boxPCAvsAll}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width=\linewidth]{Graphs/Imputation/BoxplotStocksImputPCavsAno.pdf}
\caption{Boxplot representation of the distribution of the errors on time series imputation before the imputation of anomalies (Ano) and after replacing the anomalies with the reconstructed values suggested by PCA NN (PCA).}
\label{fig:boxPCAvsNone}
\end{figure}
\begin{figure}[H]
\centering
\begin{tabular}{cc}
\includegraphics[scale=0.3]{Graphs/Imputation/CovBoxwAno.pdf} & \includegraphics[scale=0.3]{Graphs/Imputation/CovBoxwoAno.pdf} \\
a) & b)
\end{tabular}
\caption{Boxplot representation of the distribution of the errors on covariance matrix (on several stocks paths samples). Ano, PCA NN, LI, BF and Clean refer to the errors on the covariance when estimated on time series with anomalies, time series after anomalies imputation following three approaches: imputation by the reconstructed values suggested by PCA (PCA), using linear interpolation (LI), with backward fill (BF) and time series without anomalies (Clean). The Ano, PCA NN, LI, BF and Clean errors on the covariance estimation are all represented in a). For better visualization we remove the Ano errors in b). } \label{fig:boximpcov}
\end{figure}
\begin{table}[H]
\centering
\begin{tabular}{c|ccccc}
Method & Ano & PCA & LI & BF & Clean \\
\hline
Mean & 0.004208& 0.001575& 0.001554& 0.001554& 0.001552\\
Standard deviation & 0.000272& 0.000165& 0.000170& 0.000172& 0.000173
\end{tabular}
\caption{Mean and standard deviation error on covariance matrix after imputation of anomalies with baseline methods.}
\label{table:CovErrorImputation}
\end{table}
A natural explanation to the relative inefficiency of the reconstruction value as imputation value comes from the fact that by definition the reconstructed value integrated in its computation the abnormal value, which is not the case when using other imputation techniques. Therefore even if the reconstructed value is closer to the true value than it is to the abnormal value, the spread between the imputation and the true value is still significant. Figures \ref{fig:ImpuPCAPath} and \ref{fig:ImpuPCAPathZoomIn} illustrate the latter with a plot of a reconstructed and original price path.
\begin{figure}[H]
\centering
\includegraphics[scale=0.5]{Graphs/Imputation/ImpuPCAPath.pdf}
\caption{Original, contaminated and reconstructed stock price path.}
\label{fig:ImpuPCAPath}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.6]{Graphs/Imputation/ImputationZoominPaths.pdf}
\caption{Original, contaminated and imputed stock prices path in black, red and orange. The two additional graphs represent the region around anomalies, where the cross and circle respectively show the true value of the stock and its value after imputation of the anomalies using the reconstructed value suggested by the PCA NN model.}
\label{fig:ImpuPCAPathZoomIn}
\end{figure}
\subsection{Cut-off Value Robustness}
\label{subsection:ThresholdRobustness}
In most anomaly detection models, the cut-off value is a hand-set parameter. In our approach, the cut-off value is a model parameter and as such it is calibrated through the learning. Testing the robustness of the cut-off given by the model is therefore a must. The identification step is the unique step of the approach being concerned by the robustness, since it is the only step involving the cut-off calibration. Robustness is checked by shocking the suggested cut-off with different level of noise and observing the impact of these shocks on the model performance. We consider several shocks amplitude $\gamma \in \pm \{10^{-4},10^{-3},10^{-2},10^{-1},1,2 \}$. Table \ref{tab:RobustCheckThreshold} reports the mean and standard deviation of the scores of the model. The cut-off calibrated on the synthetic training data sample is shocked. Scores on both train and test sets are computed using the shocked cut-off values.
\begin{table}[H]
\centering
\begin{tabular}{c}
\begin{tabular}{c|cccc}
$\gamma$ & Accuracy & Precision & Recall& F1 \\
\hline
$10^{-4}$ & 0.9097 & 0.9736 & 0.8421 & 0.9031 \\
-$10^{-4}$ & 0.9097 & 0.9736 & 0.8421 & 0.9031 \\
\hline
$10^{-3}$ & 0.9097 & 0.9738 & 0.8419 & 0.9031 \\
-$10^{-3}$ & 0.9098 & 0.9737 & 0.8424 & 0.9033 \\
\hline
$10^{-2}$ & 0.9092 & 0.9747 & 0.8403 & 0.9025 \\
-$10^{-2}$ & 0.9098 & 0.9720 & 0.8439 & 0.9035 \\
\hline
$10^{-1}$ & 0.9042 & 0.9860 & 0.8201 & 0.8954 \\
-$10^{-1}$ & 0.9103 & 0.9563 & 0.8599 & 0.9056 \\
\hline
0 & 0.9097 & 0.9736 & 0.8421 & 0.9031 \\
\hline
1 & 0.7771 & 0.9994 & 0.5546 & 0.7133 \\
-1 & 0.5183 & 0.5093 & 0.9998 & 0.6749 \\
\hline
2 & 0.6284 & 1.0000 & 0.2567 & 0.4086 \\
-2 & 0.5001 & 0.5000 & 1.0000 & 0.6667 \\
\end{tabular} \\
a)\\~\\
\begin{tabular}{c|cccc}
$\gamma$ & Accuracy & Precision & Recall & F1 \\
\hline
$10^{-4}$ & 0.8858 & 0.6126 & 0.8527 & 0.7130 \\
$-10^{-4}$ & 0.8858 & 0.6126 & 0.8527 & 0.7130 \\
\hline
$10^{-3}$ &0.8858 & 0.6126 & 0.8527 & 0.7130 \\
$-10^{-3}$ &0.8850 & 0.6105 & 0.8527 & 0.7116 \\
\hline
$10^{-2}$ &0.8877 & 0.6179 & 0.8527 & 0.7166 \\
$-10^{-2}$ & 0.8838 & 0.6074 & 0.8527 & 0.7095 \\
\hline
$10^{-1}$ & 0.8984 & 0.6502 & 0.8432 & 0.7342 \\
$-10^{-1}$ & 0.8672 & 0.5653 & 0.8741 & 0.6866 \\
\hline
0 & 0.8858 & 0.6126 & 0.8527 & 0.7130 \\
\hline
1 & 0.9391 & 0.9435 & 0.6746 & 0.7867 \\
-1 & 0.2660 & 0.1848 & 1.0000 & 0.3120 \\
\hline
2 & 0.8949 & 0.9814 & 0.3753 & 0.5430 \\
-2 & 0.1700 & 0.1670 & 1.0000 & 0.2862 \\
\end{tabular}\\
b)
\end{tabular}
\caption{Performance evaluation after shocking the calibrated cut-off value for a) the training set and b) for the test set.}
\label{tab:RobustCheckThreshold}
\end{table}
Excluding the extreme cases where the shock amplitude is $\pm\{1,2\}$, one could see that the accuracy is barely impacted by the shocked cut-off values both on the train and test sets. The interpretation of the remaining results is split in two and is applicable for both the training and test sets.
When negative shocks are applied, the model predicts more anomalies and less normal observations (compared to the predicted numbers with the calibrated cut-off value). Therefore, the model is able to identify anomalies which were missed initially. Hence, applying negative shocks increases the recall. As for the precision, i.e. the rate of correctly identified contaminated time series, since on left hand side of the cut-off value we are in the density region where contaminated time series represent the dominant class, the new position of the cut-off value leads to a misidentification of this type of observations. This entails a deterioration of the precision.
Opposite behaviours of the precision and recall are observed when positive shocks are applied since some abnormal time series are missed by the model.
\begin{figure}
\centering
\begin{tabular}{cc}
\includegraphics[scale=0.3]{Graphs/ThresholdRob/PrecisionRecallCurveSyntheticTrain.png}&
\includegraphics[scale=0.3]{Graphs/ThresholdRob/PrecisionRecallCurveSyntheticTest.png} \\
a) & b)
\end{tabular}
\caption{Precision-recall curve of the suggested model (in blue), scores with the calibrated threshold (red plus), no skill model scores (green cross) for the training set a) and a test set b).} \label{fig:precisionrecallcurve}
\end{figure}
Figure \ref{fig:precisionrecallcurve} shows the precision and recall when cut-off values other than the one we calibrated are used in our model, on the train and test sets. These scores are also compared with the performance of the no skill model\footnote{Here, no skill model being the model assigning to all observations the positive label}. We note that our approach edges out the no skill model on both sets. The area under the curve (AUC) score on the train set and the test set are respectively 0.97 and 0.87 (as a reminder, AUC score ranges from 0 to 1, 1 being the score associated with a perfect model). This shows that we are better performing on the train set, which was expected, but the performance on test set is just as satisfying. These results reinforce our conclusions regarding the robustness of our approach. One can see that the calibrated cut-off value represents, for the train set, the point where an equilibrium is being reached between precision and recall. The calibrated cut-off allows to achieve high precision and recall scores simultaneously. Thus we conclude that the cut-off given by the model is suitable for the training samples and for unseen samples as well.
To briefly sum up this section, although we mentioned that some shocks on the cut-off induce higher scores, the improvement over the scores with the calibrated cut-off is not significant (unless high amplitude shocks are considered). The numerical tests and the precision-recall curves are consistent with the robustness of the cut-off value suggested by the approach.
\section{Application to a Downstream Task: Value-at-Risk Computations }
\label{section:VaR}
In this section, we illustrate the benefit that could be drawn when applying the PCA NN approach as pre-processing step for value-at-risk computations.
Given a random variable $\Delta X$ representing the potential loss in portfolio position over a time horizon $\Delta t$, the value-at-risk quantile at level $1 - \alpha$, $\mathrm{VaR}_{1-\alpha}$, is defined by
\begin{align*}
\mathrm{VaR}_{1-\alpha}\left( \Delta X\right) := - \inf \{x \colon \mathbb{P} \left(\Delta X \leq x\right)\geq 1 - \alpha \}.
\end{align*}
The $\mathrm{VaR}_{1-\alpha}$ over a time horizon $\Delta t$ corresponds to the greatest loss that may be exceeded with probability $\alpha$ in the period $\Delta t$. Let $(S_t)_{t=t_0, \ldots, t_T}$ be a path price diffusion distributed according to the Black-Scholes model \eqref{eq:diffeq}. The logarithmic returns to maturity are distributed according to a Gaussian distribution, i.e.
\begin{align}
\ln\left(\frac{S_{t+dt}}{S_t}\right) \sim \mathcal{N} \left( \left(\mu - \frac{\sigma^2}{2}\right)dt, \sigma^2dt\right),
\end{align}
with $\mu \in \mathbb R$ and $\sigma >0$.
We assume the vector $\bm{R}$ of log-returns on our stocks to be joint-normal,
\[
\bm{R} =
\left(
R^{1}, R^{2}, \ldots , R^{i}, \ldots, R^N
\right)^{\top} \sim \mathcal{N}_N\left(\mu_R,\Sigma_R\right),
\]
where $R^i = \log\left(\frac{S^i_{t+dt}}{S^i_t}\right) \sim \mathcal{N}\left(\mu_{i,R},\sigma_{i,R}^2\right)$ and $\Sigma_R$ is the covariance matrix.
We consider a portfolio on $N$ stocks, which return is given by $R^{Portfolio} = \mathcal{Q}^\top \bm{R}$. Hence, $R^{Portfolio} \sim \mathcal{N}\left(\mathcal{Q}^\top \mu_R,\mathcal{Q}^\top \Sigma_R \mathcal{Q}\right)$.
The VaR quantile for the time horizon $\Delta t$ and at level $\alpha$ is
\begin{align}
VaR_{1-\alpha}\left(\mu_R,\Sigma_R\right) = q_{1-\alpha} \sqrt{\mathcal{Q}^\top \Sigma_R \mathcal{Q}} + \mathcal{Q}^\top \mu_R,
\label{eq:VaR}
\end{align}
where $q_{1-\alpha}$ is the $(1-\alpha)$-quantile of a standard normal distribution and the parameters $ \mu_R$ and $ \Sigma_R$ are estimated from different types of time series to evaluate the impact of identifying and removing anomalies following our approach.
Under the adopted framework, the true $\mathrm{VaR}_{1-\alpha}\left(\mu_R,\Sigma_R\right) $, $\mathrm{VaR}^{Theoretical}$ is thus known and can be computed using the diffusion parameters. An estimation $\mathrm{VaR}_{1-\alpha}\left(\widehat{\mu_R},\widehat{\Sigma_R}\right)$ can be obtained by replacing the parameters in \eqref{eq:VaR} by their estimates $\widehat{\mu_R}$ and $\widehat{\sigma_R}$ computed from the time series with anomalies, or after imputation of anomalies. Then, absolute errors
and relative errors are computed as
\begin{align*}
\text{AbsoluteError}_{\text{VaR}} = \left \lvert \mathrm{VaR}_{1-\alpha}\left(\mu_R,\Sigma_R\right) - \mathrm{VaR}_{1-\alpha}\left(\widehat{\mu_R},\widehat{\Sigma_R}\right) \right \rvert
\end{align*}
\begin{align*}
\text{RelativeError}_{\text{VaR}} = \frac{\left \lvert \mathrm{VaR}_{1-\alpha}\left(\mu_R,\Sigma_R\right) - \mathrm{VaR}_{1-\alpha}\left(\widehat{\mu_R},\widehat{\Sigma_R}\right) \right \rvert}{\mathrm{VaR}_{1-\alpha}\left(\mu_R,\Sigma_R\right)}
\end{align*}
Table \ref{tab:VaRnames} summarizes the four VaR estimations we consider in the sequel:
\begin{table}[H]
\begin{tabular}{c|l}
VaR estimation Name & $\widehat{\mu_R}, \widehat{\Sigma_R}$ estimated on \\ \hline
$\mathrm{VaR}^{Clean}$ & Time series without anomalies \\
$\mathrm{VaR}^{Ano}$ & Times series with anomalies \\
$\mathrm{VaR}^{Loc,True}$ & Time series after anomalies imputation knowing their true localization \\
$\mathrm{VaR}^{Loc,Pred}$ & Time series after anomalies imputation based on predicted localization \\
\end{tabular}
\caption{Notations of VaR estimation based on the time series the VaR parameters have been estimated from.}
\label{tab:VaRnames}
\end{table}
To conduct this analysis we generated new stocks paths samples that we assume to be clean, fixing the diffusion parameters $\mu_R$ and $\Sigma_R$, then we added the anomalies following the procedure described in Section \ref{section: data}. We apply our model to locate the anomalies and replace the identified anomalies using the backward fill (BF) approach (shown to be the more efficient imputation technique in Section \ref{subsection:imputation}. For each stock and each run of simulation, we obtain four estimates of the distribution parameters of the associated log-returns.
\begin{figure}[H]
\centering
\includegraphics[scale=0.3]{Graphs/VaRImpact/boxVaRPortfolio99.png}
\caption{Boxplot representation of parametric VaR estimations for $\bm{R}^{Portfolio}$. The green square represents the mean of VaR estimation.}
\label{fig:VaRonPortfolioBox}
\end{figure}
\begin{table}[H]
\begin{tabular}{c|c|c|c|c|c}
VaR & $\mathrm{VaR}^{Theoretical}$ & $\mathrm{VaR}^{Clean}$ &$\mathrm{VaR}^{Loc,True}$ &$\mathrm{VaR}^{Loc,Pred}$&$\mathrm{VaR}^{Ano}$ \\ \hline
Mean & 0.546851& 0.546300& 0.548392& 0.548270& 0.569015\\
Standard Deviation &0.0& 0.010105& 0.010739& 0.010832 & 0.012268 \\
\end{tabular}
\caption{Summary of VaR estimations for $\bm{R}^{Portfolio}$.}
\label{table:VaRonPortfolioBox}
\end{table}
Figure \ref{fig:VaRonPortfolioBox} and Table \ref{table:VaRonPortfolioBox} summarize the distribution of the VaR estimates for $\alpha = 0.99$ and $\Delta t= 1(day)$, over several simulation runs. The boxplots show the dispersion of the portfolio VaR estimates on several diffusions. The green square represents the mean of VaR estimation. For the four first boxplots, the means are approximately on the same level, which is confirmed by the results of Table \ref{table:VaRonPortfolioBox}. The anomalies present among the time series observed values have a non-negligible impact on the distribution parameters estimation, which ultimately causes a wrong estimation of the VaR. Thanks to the localization of the anomalies by the suggested model and their imputation, we are able to get a more accurate estimation of the VaR. $\mathrm{VaR}^{Loc,True}$ and $\mathrm{VaR}^{Loc,Pred}$ are quite similar, which shows that the model accurately localizes the anomalies.
\begin{table}[H]
\centering
\begin{tabular}{c|c|c|c|c}
VaR & $\mathrm{VaR}^{Clean}$ &$\mathrm{VaR}^{Loc,True}$ &$\mathrm{VaR}^{Loc,Pred}$ &$\mathrm{VaR}^{Ano}$\\
\hline
Absolute Error & 0.007995& 0.008596& 0.008622 & 0.02235 \\
Relative Error & 0.014620& 0.015720& 0.015767 & 0.04087 \\
\end{tabular}
\caption{Mean Absolute and relative error on VaR estimations for $\bm{R}^{Portfolio}$.}
\label{table:MAEMREVaRPortfolio}
\end{table}
We also evaluate the error on the VaR estimation using the mean absolute error and the mean relative error, taking the $\mathrm{VaR}^{Theoretical}$ as our benchmark VaR. As one can tell from Table \ref{table:MAEMREVaRPortfolio}, even when the distribution parameters are estimated from the clean time series, the VaR computed with these parameters is not exactly the one computed with the theoretical parameters. This can be explained by the historical size of the observed values used to estimate the parameters. This table shows that by removing anomalies we can reduce by a factor two the error on VaR estimation.
Additionally, we assess the impact on $\mathrm{VaR}^{Theoretical},\mathrm{VaR}^{Clean},\mathrm{VaR}^{Ano},\mathrm{VaR}^{Loc,True}$ and $\mathrm{VaR}^{Loc,Pred}$ of increasing $n^{anomaly}$. To this end, we perform 50 simulation runs of stocks paths for $n^{anomaly}$ and for each of those scenarios we estimate the VaR on the portfolio. We summarize the results on Figure \ref{fig:boxVaRwrtContamRate}, where each curve represents the mean VaR estimation with respect to $n^{anomaly}$ along with a representation of the uncertainty around each evaluated point through a confidence interval.
\begin{figure}
\centering
\includegraphics[scale=0.35]{Graphs/VaRImpact/VaRboxwrtContamRate.pdf}
\caption{VaR estimation based on parameter estimation from time series with and without anomalies (purple and light blue curves), time series after identification and imputation of anomalies with the suggested approach (red curve) and time series imputed knowing the true location of the anomalies (brown curve), with respect to $n^{anomaly}$.}
\label{fig:boxVaRwrtContamRate}
\end{figure}
When we compute the VaR using the abnormal time series, we notice that the difference between $\mathrm{VaR}^{Ano}$ and $\mathrm{VaR}^{Theoretical}$ increases with $n^{anomaly}$, which is natural to expect. However, when the time series are cleaned prior to VaR estimation, the curve representing the VaR estimates are much closer to the ones representing $\mathrm{VaR}^{Theoretical}$ and $\mathrm{VaR}^{Clean}$, showing an undeniable improvement in the accuracy of the VaR estimation over the estimation based on abnormal time series. Furthermore, VaR estimation after the imputation following the model prediction or knowing the true location of the anomalies seem to be quite similar for low $n^{anomaly}$, while some discrepancies between the two become more significant as $n^{anomaly}$ increases. A natural explanation could be that when the number of anomalies increases and the model suggests wrong anomalies location, normal values are being replaced while true anomalies remain among the observed values, which wrongly impacts the VaR estimation. However, the results of Tables \ref{tab:meanRErrorVaR} and \ref{tab:stdRErrorVaR} indicate that the anomalies localization suggested by the model are overall correct and allow removing most of the anomalies, as the relative error of $\mathrm{VaR}^{Loc,Pred}$ is, regardless of $n^{anomaly}$, always lower than the relative error of $\mathrm{VaR}^{Ano}$ (e.g a relative error of 0.0248 for $\mathrm{VaR}^{Loc,Pred}$ against 0.0658 for $\mathrm{VaR}^{Ano}$ when there is $76$ anomalies among the $1,500$ observed values of the time series).
\begin{table}[H]
\centering
\begin{tabular}{c|cccc}
$n^{anomaly}$ &$\mathrm{VaR}^{Clean}$ &$\mathrm{VaR}^{Loc,True}$ &$\mathrm{VaR}^{Loc,Pred}$&$\mathrm{VaR}^{Ano}$ \\
\hline
5 & 0.012194 &0.013796 &0.013341&0.037392 \\
15 & 0.012194 &0.012623 &0.016676 &0.059604 \\
76 & 0.012194 &0.014831 &0.024807 &0.065791 \\
153 & 0.012194 &0.014644 &0.034361 &0.073750 \\
307 & 0.012194 &0.020537 &0.052244 &0.074091 \\
\end{tabular}
\caption{Mean relative error of parametric VaR estimations with respect to $n^{anomaly}$, for parameters estimated from clean time series, time series with anomalies, imputed time series following predicted location ($\mathrm{VaR}^{Loc,Pred}$) and true anomalies location ($\mathrm{VaR}^{Loc,True}$). }
\label{tab:meanRErrorVaR}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{c|cccc}
$n^{anomaly}$ & $\mathrm{VaR}^{Clean}$ &$\mathrm{VaR}^{Loc,True}$ &$\mathrm{VaR}^{Loc,Pred}$ &$\mathrm{VaR}^{Ano}$ \\
\hline
5 & 0.010356 &0.011511 &0.009678 &0.020674 \\
15 & 0.010356 &0.010514 &0.012723 &0.020172 \\
76 & 0.010356 &0.011842 &0.018696 &0.019930 \\
153 &0.010356 &0.013683 &0.026117 &0.024243 \\
307 & 0.010356 &0.022526 &0.057903 &0.019207 \\
\end{tabular}
\caption{Standard deviation of relative error of parametric VaR estimations with respect to $n^{anomaly}$, for parameters estimated from clean time series, time series with anomalies, imputed time series following predicted location ($\mathrm{VaR}^{Loc,Pred}$) and true anomalies location ($\mathrm{VaR}^{Loc,True}$). }
\label{tab:stdRErrorVaR}
\end{table}
\section{Conclusion}\label{section: conclusion}
The suggested approach targets anomaly detection on panel of time series possibly reflecting a wide variety of risk factors. We suggest a model able to detect anomalies in risk factor time series suitable for all risk factors. Anomaly detection is achieved in two steps. The first step aims at identifying the contaminated time series (time series with anomalies). The second step focuses on the localization of the anomaly among the observed values of the identified contaminated time series.
In addition, our methodology integrates feature extraction from the time series with PCA. This part of the method is proved to be an essential part of the model, as it provides the models with inputs on which the distinction between abnormal/contaminated and normal instances is eased, and also ensures the stationnarity of the model (time series) inputs. Another main focus of the approach is the calibration of the cut-off, the key parameter in the identification of contaminated time series, by means of feedforward network with a customized loss function.
The proposed approach suggests an imputation value, however this value is strongly impacted by the anomaly value. As a result, naive approaches with similar complexity are preferred.
Furthermore, our numerical experiments do not only show that our approach outperforms baseline anomaly detection models, it also shows the true benefit one could draw from cleaning time series with the suggested approach, as illustrated on the VaR computation task.
\newpage
|
\section{Introduction}
Consider in the upper half $\mathbb{C}_+ = \{ z = x+iy \in \mathbb{C} : y > 0 \}$ of the complex plane, the logarithmic function
$$
\ln z = \ln |z| + i \arg z, \qquad \mbox{Im} \, z > 0
$$
with
$$
\arg z = \frac{\pi}{2} - \arctan \frac{x}{y}, \quad y>0
$$
It is quite an interesting function; let us have a closer look at its properties.
\begin{itemize}
\item[(i)] The function $\ln z$ is holomorphic in $\mathbb{C}_+$, i.e.\ it is a null solution of the Cauchy--Riemann operator
$$
D = \frac{1}{2} \left ( \partial_x + i \partial_y \right )
$$
\item[(ii)] Its real and imaginary parts are conjugate harmonic functions in $\mathbb{C}_+$ and the real part $\ln |z| = \ln r = \ln \sqrt{x^2+y^2}$ is, up to a constant, the fundamental solution of the two--dimensional Laplace operator
$$
\Delta_2 = \partial_{xx}^2 + \partial_{yy}^2 = 4 D \overline{D}
$$
where $\overline{D} = \frac{1}{2} \left ( \partial_x - i \partial_y \right )$ is the complex conjugate Cauchy--Riemann operator; in fact, we have, in distributional sense,
$$
\Delta_2 \left ( \frac{1}{2\pi} \ln |z| \right ) = \delta(z)
$$
with $\delta(z)$ the Dirac or delta distribution in $\mathbb{C}$.
\item[(iii)] As a holomorphic function, $\ln z$ has a complex derivative in $\mathbb{C}_+$, given by
$$
\frac{d}{dz} \ln z = \frac{1}{z}
$$
meaning that $\ln z$ is a holomorphic primitive (or potential) in $\mathbb{C}_+$, with respect to the complex derivative $\frac{d}{dz}$, of the function $\frac{1}{z}$ which, in its turn, is, up to a constant, the fundamental solution of the Cauchy--Riemann operator; in fact we have, in distributional sense,
$$
D \left (\frac{1}{\pi} \frac{1}{z} \right ) = \delta(z)
$$
Note that the complex derivative operator $\frac{d}{dz}$ is nothing else but the conjugate Cauchy--Riemann operator $\overline{D}$, and there also holds in $\mathbb{C}_+$
$$
\frac{d}{dz} \ln z = \overline{D} \ln z = \partial_x \ln z = (-i \partial_y) \ln z = \frac{1}{z}
$$
\item[(iv)] The conjugate harmonic real and imaginary parts $\ln r$ and $\arg z$ satisfy the Cauchy--Riemann system
$$
\left \{ \begin{array}{lllllll}
\partial_x \ln r & = & \phantom{-} \partial_y \arg z & = & \frac{x}{x^2+y^2} & = & \phantom{-} \mbox{Re} \left ( \frac{1}{z} \right ) \\[3mm]
\partial_y \ln r & = & - \partial_x \arg z & = & \frac{y}{x^2+y^2} & = & - \mbox{Im} \left ( \frac{1}{z} \right ) \\
\end{array} \right .
$$
where at the right hand sides one recognizes the Poisson kernel $P(x,y) = \frac{y}{x^2+y^2}$ and its harmonic conjugate
$Q(x,y) = \frac{x}{x^2+y^2}$ in $\mathbb{C}_+$; it follows that
$$
\overline{D} ( 2 \ln r) = \frac{1}{z}
$$
and
$$
\overline{D} ( 2 i \arg z ) = \frac{1}{z}
$$
meaning that the functions $2 \ln r$ and $2i \arg z$ are conjugate harmonic potentials in $\mathbb{C}_+$, with respect to the operator $\overline{D} = \frac{d}{dz}$, of the Cauchy kernel $\frac{1}{z}$.
\item[(v)] The distributional limits for $y \rightarrow 0+$ of the Cauchy kernel $\frac{1}{z}$ and its holomorphic potential $\ln z$, are given by
$$
\lim_{y \rightarrow 0+} \frac{1}{z} = \lim_{y \rightarrow 0+} \frac{x}{x^2+y^2} - i \lim_{y \rightarrow 0+} \frac{y}{x^2+y^2} = \mbox{Pv}\, \frac{1}{x} - i \pi \delta(x)
$$
with Pv$\frac{1}{x}$ the ''principal value'' distribution on the real axis,
and
$$
\lim_{y \rightarrow 0+} \ln z = \lim_{y \rightarrow 0+} \ln |z| + i \lim_{y \rightarrow 0+} \arg z = \ln |x| + i \pi Y(-x)
$$
with $Y(x)$ the Heaviside step function. These distributional boundary values fit into the following two commutative schemes
$$
\begin{array}{ccc}
\ln r & \xrightarrow{\hspace*{2mm} \partial_x \hspace*{2mm}} & \frac{x}{x^2+y^2} \\[2mm]
\hspace*{-7mm} ^{\hspace*{0.1mm}_{y \rightarrow 0+}} \downarrow & & \downarrow \\
\ln |x| & \xrightarrow{\hspace*{2mm} \partial_x \hspace*{2mm}} & \mbox{Pv} \, \frac{1}{x}
\end{array}
\qquad \mbox{and} \qquad
\begin{array}{ccc}
- \arg z & \xrightarrow{\hspace*{2mm} \partial_x \hspace*{2mm}} & \frac{y}{x^2+y^2} \\[2mm]
\hspace*{-5mm} ^{\hspace*{0.1mm}_{y \rightarrow 0+}} \downarrow & & \downarrow \\
- \pi Y(-x) & \xrightarrow{\hspace*{2mm} \partial_x \hspace*{2mm}} & \pi \, \delta(x)
\end{array}
$$
and moreover they form Hilbert pairs, the Hilbert transform on the real axis being given by
$$
\mathcal{H}[T] = H \ast T = \frac{1}{\pi} \mbox{Pv} \, \frac{1}{x} \ast T,
$$
since we have indeed
$$
\mathcal{H} \left[ \pi \delta(x)\right] = \mbox{Pv} \, \frac{1}{x}, \qquad \mathcal{H} \left [ \mbox{Pv} \, \frac{1}{x} \right ] = \pi \delta(x)
$$
and
$$
\mathcal{H} \left[ \ln |x| \right] = \pi \, Y(-x), \qquad \mathcal{H} \left[ \pi \, Y(-x) \right] = \ln |x|
$$
\end{itemize}
The aim of this paper is to construct a generalization of this logarithmic potential function in higher dimension, more specifically in the framework of Clifford analysis, where the functions under consideration take their values in the universal Clifford algebra $\mathbb{R}_{0,m+1}$ constructed over Euclidean space $\mathbb{R}^{m+1}$ equipped with a quadratic form of signature $(0,m+1)$. The concept of a higher dimensional holomorphic function, mostly called monogenic function, is expressed by means of a generalized Cauchy--Riemann operator, which is a combination of the derivative with respect to one of the variables, say $x_0$, and the so--called Dirac operator $\underline{\partial}$ in the remaining variables $(x_1, x_2, \ldots, x_m)$. The generalized Cauchy--Riemann operator and its Clifford algebra conjugate linearize the Laplace operator, whence Clifford analysis is entitled to be qualified as a refinement of harmonic analysis.\\[-2mm]
It is a remarkable fact that the thus constructed monogenic logarithmic function in upper half--space $\mathbb{R}^{m+1}_+$ shows the same, above mentioned, five properties as in the complex plane. Starting point of our construction is the fundamental solution of the generalized Cauchy--Riemann operator, also called Cauchy kernel, and its relation to the Poisson kernel and its harmonic conjugate in $\mathbb{R}^{m+1}_+$. We then proceed by induction in two directions, {\em downstream} by differentiation and {\em upstream} by primitivation, yielding an doubly infinite chain of monogenic, and thus harmonic, potentials. This chain mimics the well--known sequence of holomorphic potentials in $\mathbb{C}_+$ (see e.g. \cite{slang}):
$$
\frac{1}{k!} z^k \left[ \ln z - ( 1 + \frac{1}{2} + \ldots + \frac{1}{k}) \right] \rightarrow \ldots \rightarrow z ( \ln z - 1) \rightarrow \ln z
\stackrel{\frac{d}{dz}}{\longrightarrow} \frac{1}{z} \rightarrow - \frac{1}{z^2} \rightarrow \ldots \rightarrow (-1)^{k-1} \frac{(k-1)!}{z^k}
$$
Identifying the boundary of upper half--space with $\mathbb{R}^m \cong \{(x_0,\underline{x}) \in \mathbb{R}^{m+1} : x_0 = 0\}$, the distributional limits for $x_0 \rightarrow 0+$ of those potentials are computed; they divide into two classes which are linked by the Hilbert transform and encompass well--known distributions in $\mathbb{R}^m$ such as the Dirac or delta distribution, the Hilbert kernel, the fundamental solutions of the Dirac and the Laplace operators, the square root of the negative Laplacian, and the like. It is also shown how each of the monogenic potentials may be recovered from an adjacent kernel in the chain by an appropriate convolution with such a boundary distribution.\\[-2mm]
The organization of the paper is as follows. To make the paper self--contained we recall in Section 2 the basics of Clifford algebra and Clifford analysis. In Section 3 we construct a conjugate harmonic in upper half--space $\mathbb{R}^{m+1}_+$ to the fundamental solution of the $(m+1)$--dimensional Laplace operator, which is essential to obtaining the desired monogenic logarithmic function in $\mathbb{R}^{m+1}_+$. In Section 4 we study the so--called downstream potentials obtained under the action of the Clifford algebra conjugate of the generalized Cauchy--Riemann operator. Finally, in Section 5, we study the monogenic logarithmic function in $\mathbb{R}^{m+1}_+$ and we construct, by an appropriate form of primitivation, the sequence of upstream potentials. Section 6 is concluding.
\section{Basics of Clifford analysis}
Clifford analysis (see e.g. \cite{red}) is a function theory which offers a natural and elegant generalization to higher dimension of holomorphic functions in the complex plane and refines harmonic analysis. Let $(e_0, e_1,\ldots,e_m)$ be the canonical orthonormal basis of Euclidean space $\mathbb{R}^{m+1}$ equipped with a quadratic form of signature $(0,m+1)$. Then the non--commutative multiplication in the universal real Clifford algebra $\mathbb{R}_{0,m+1}$ is governed by the rule
$$
e_{\alpha} e_{\beta} + e_{\beta} e_{\alpha} = -2 \delta_{\alpha \beta}, \qquad \alpha,\beta = 0, 1,\ldots,m
$$
whence $\mathbb{R}_{0,m+1}$ is generated additively by the elements $e_A = e_{j_1} \ldots e_{j_h}$, where $A=\lbrace j_1,\ldots,j_h \rbrace \subset \lbrace 0,\ldots,m \rbrace$, with $0\leq j_1<j_2<\cdots < j_h \leq m$, and $e_{\emptyset}=1$.
For an account on Clifford algebra we refer to e.g. \cite{porteous}.\\[-2mm]
We identify the point $(x_0, x_1, \ldots, x_m) \in \mathbb{R}^{m+1}$ with the Clifford--vector variable
$$
x = x_0 e_0 + x_1 e_1 + \cdots x_m e_m = x_0 e_0 + \underline{x}
$$
and the point $(x_1, \ldots, x_m) \in \mathbb{R}^{m}$ with the Clifford--vector variable $\underline{x}$.
Introducing spherical co--ordinates $\underline{x} = r \underline{\omega}$, $r = |\underline{x}|$, $\underline{\omega} \in S^{m-1}$, gives rise to the Clifford--vector valued locally integrable function $\underline{\omega}$, which is to be seen as the higher dimensional analogue of the {\em signum}--distribution on the real line; we will encounter $\underline{\omega}$ as one of the distributions discussed below.\\[-2mm]
At the heart of Clifford analysis lies the so--called Dirac operator
$$
\partial = \partial_{x_0} e_0 + \partial_{x_1} e_1 + \cdots \partial_{x_m} e_m = \partial_{x_0} e_0 + \underline{\partial}
$$
which squares to the negative Laplace operator: $\partial^2 = - \Delta_{m+1}$, while also $\underline{\partial}^2 = - \Delta_{m}$. Due to the non--commutative character of the multiplication in the Clifford algebra, the Dirac operator may act from the left or from the right on a Clifford algebra valued function with, in general, different results. The (left and right) fundamental solution of the Dirac operator $\partial$ is given by
$$
E_{m+1} (x) = - \frac{1}{\sigma_{m+1}} \ \frac{x}{|x|^{m+1}}
$$
where $\sigma_{m+1} = \frac{2\pi^{\frac{m+1}{2}}}{\Gamma(\frac{m+1}{2})}$ stands for the area of the unit sphere $S^{m}$ in $\mathbb{R}^{m+1}$.
We also introduce the generalized Cauchy--Riemann operator
$$
D = \frac{1}{2} \overline{e_0} \partial = \frac{1}{2} (\partial_{x_0} + \overline{e_0} \underline{\partial})
$$
and its Clifford algebra conjugate $\overline{D} = \frac{1}{2}(\partial_{x_0} - \overline{e_0} \underline{\partial})$. As is the case in the complex plane, both operators decompose the Laplace operator in $\mathbb{R}^{m+1}$: $D \overline{D} = \overline{D} D = \frac{1}{4} \Delta_{m+1}$. \\[-2mm]
A continuously differentiable function $F(x)$, defined in an open region $\Omega \subset \mathbb{R}^{m+1}$ and taking its values in the Clifford algebra $\mathbb{R}_{0,m+1}$, is called (left--)monogenic if it satisfies the equation $D F = 0$ in $\Omega$, which is equivalent with $\partial F = 0$. \\
Singling out the basis vector $e_0$, we can decompose the real Clifford algebra $\mathbb{R}_{0,m+1}$ in terms of the Clifford algebra
$\mathbb{R}_{0,m}$ as $\mathbb{R}_{0,m+1} = \mathbb{R}_{0,m} \oplus \overline{e_0} \mathbb{R}_{0,m}$. Similarly we decompose the functions considered as
$$
F(x_0,\underline{x}) = F_1(x_0,\underline{x}) + \overline{e_0} F_2(x_0,\underline{x})
$$
where $F_1$ and $F_2$ take their values in the Clifford algebra $\mathbb{R}_{0,m}$; mimicking functions of a complex variable, we will call $F_1$ the {\em real} part and $F_2$ the {\em imaginary} part of the function $F$.\\[-2mm]
We will extensively use two families of distributions in $\mathbb{R}^m$, which have been thoroughly studied in \cite{fb1,fb2,distrib}. The first family $\mathcal{T} = \{ T_\lambda : \lambda \in \mathbb{C}\}$ is very classical. It consists of the radial distributions
$$
T_\lambda = {\rm Fp} \ r^{\lambda} = {\rm Fp} \ (x_1^2 + \ldots + x_m^2)^{\frac{\lambda}{2}}
$$
their action on a test function $\phi \in \mathcal{S}(\mathbb{R}^m)$ being given by
$$
\langle T_\lambda, \phi \rangle = \sigma_m \langle {\rm Fp} \; r^\mu_+, \Sigma^{(0)}[\phi] \rangle
$$
with $\mu = \lambda +m-1$. In the above expressions ${\rm Fp}\; r^\mu_+$ is the classical {\em finite part} distribution on the real $r$-axis and $\Sigma^{(0)}$ is the scalar valued generalized spherical mean, defined on scalar valued test functions $\phi(\underline{x})$ by
$$
\Sigma^{(0)}[\phi] = \frac{1}{\sigma_m} \int_{S^{m-1}} \phi(\underline{x}) \, dS(\underline{\omega})
$$
This family $\mathcal{T}$ contains a.o. the fundamental solution of the Laplace operator. As convolution operators they give rise to the traditional Riesz potentials (see e.g. \cite{helgason}). The second family $\mathcal{U} = \{ U_\lambda : \lambda \in \mathbb{C}\}$ of distributions arises in a natural way by the action of the Dirac operator $\underline{\partial}$ on $\mathcal{T}$. The $U_{\lambda}$--distributions thus are typical Clifford analysis objects: they are Clifford--vector valued, and they also arise as products of $T_{\lambda}$--distributions with the distribution $\underline{\omega} = \frac{\underline{x}}{|\underline{x}|}$, mentioned above. The action of $U_\lambda$ on a test function $\phi \in \mathcal{S}(\mathbb{R}^m)$ is given by
$$
\langle U_\lambda, \phi \rangle = \sigma_m \langle {\rm Fp} \; r^\mu_+, \Sigma^{(1)}[\phi] \rangle
$$
with $\mu = \lambda +m-1$, and where the Clifford--vector valued generalized spherical mean $\Sigma^{(1)}$ is defined on scalar valued test functions $\phi(\underline{x})$ by
$$
\Sigma^{(1)}[\phi] = \frac{1}{\sigma_m} \int_{S^{m-1}} \underline{\omega} \ \phi(\underline{x}) \, dS(\underline{\omega})
$$
Typical example in the $\mathcal{U}$--family is the fundamental solution of the Dirac operator.\\[-2mm]
The normalized distributions $T^{*}_\lambda$ and $U^{*}_\lambda$ arise when the singularities of $T_\lambda$ and $U_\lambda$ are removed by dividing them by an appropriate Gamma-function, showing the same simple poles. The $T^{*}_\lambda$--distributions are defined by
\begin{eqnarray*}
\left \{
\begin{array}{ll}
\displaystyle{T_\lambda^* = \pi^{\frac{\lambda+m}{2}} \frac{T_\lambda}{\Gamma \left ( \frac{\lambda+m}{2} \right )}}, & \lambda \ne -m-2l\\[5mm]
\displaystyle{T_{-m-2l}^* = \frac{\pi^{\frac{m}{2}-l}}{2^{2l} \Gamma \left ( \frac{m}{2} + l \right )} (-\Delta)^l \delta (\underline{x})}, & l \in \mathbb{N}_0
\end{array}
\right .
\end{eqnarray*}
while the Clifford--vector valued distributions $U^{*}_\lambda$ are defined by
\begin{eqnarray*}
\left \{
\begin{array}{ll}
\displaystyle{U_\lambda^* = \pi^{\frac{\lambda+m+1}{2}} \, \frac{U_\lambda}{\Gamma \left ( \frac{\lambda + m + 1}{2} \right )}}, & \lambda \ne -m-2l-1\\[5mm]
\displaystyle{U_{-m-2l-1}^* = - \frac{\pi^{\frac{m}{2}-l}}{2^{2l+1} \, \Gamma \left ( \frac{m}{2} + l + 1 \right )} \; \underline{\partial}^{2l+1} \delta(\underline{x})}, & l \in \mathbb{N}_0
\end{array}
\right .
\end{eqnarray*}
The normalized distributions $T_\lambda^*$ and $U_\lambda^*$ are holomorphic mappings from $\lambda \in \mathbb{C}$ to the space $\mathcal{S}'(\mathbb{R}^m)$ of tempered distributions. As already mentioned they are intertwined by the action of the Dirac operator. They enjoy the following properties: for all $\lambda \in \mathbb{C}$ one has
\begin{itemize}
\item[(i)] $\underline{x} \; T_\lambda^* = \frac{\lambda+m}{2\pi} \; U_{\lambda+1}^*$; \quad
$\underline{x} \; U_\lambda^* = U_\lambda^* \; \underline{x} = - T_{\lambda+1}^*$
\item[(ii)] $\underline{\partial} \; T_\lambda^* = \lambda \; U_{\lambda-1}^*$; \quad
$\underline{\partial} \; U_\lambda^* = U_\lambda^* \; \underline{\partial} = - 2\pi \; T_{\lambda-1}^*$
\item [(iii)] $\Delta_m T_\lambda^* = 2 \pi \lambda T_{\lambda-2}^*$ ; \quad
$\Delta_m U_\lambda^* = 2 \pi (\lambda-1) U_{\lambda-2}^*$
\item[(iv)] $r^2 T_\lambda^* = \frac{\lambda + m}{2\pi} \ T_{\lambda+2}^*$; \quad
$r^2 U_\lambda^* = \frac{\lambda + m + 1}{2\pi} \ U_{\lambda+2}^*$
\end{itemize}
Of particular importance for the sequel are the convolution formulae for the $T_\lambda^*$-- and $U_\lambda^*$--distributions; we list them in the following proposition and refer the reader to \cite{distrib} for more details.
\begin{proposition}
\label{prop1}
\rule{0mm}{0mm}
\begin{itemize}
\item[(i)] For all $(\alpha,\beta) \in \mathbb{C} \times \mathbb{C}$ such that $ \alpha \neq 2j, j \in \mathbb{N}_0$, $ \beta \neq 2k, k \in \mathbb{N}_0$ and $\alpha+\beta+m \neq 2l, l \in \mathbb{N}_0$ the convolution $T_{\alpha}^* \ast T_{\beta}^*$ is the tempered distribution given by
$$
T_{\alpha}^* \ast T_{\beta}^* = c_m(\alpha,\beta)\; T_{\alpha+\beta+m}^*
$$
with
\begin{eqnarray*}
c_m(\alpha,\beta) & = & \pi^{\frac{m}{2}}\; \frac{\Gamma \left( -\frac{\alpha + \beta + m}{2} \right)}{\Gamma \left( -\frac{\alpha}{2} \right) \Gamma \left( -\frac{\beta}{2} \right)}
\end{eqnarray*}
\item[(ii)] For $(\alpha,\beta) \in \mathbb{C} \times \mathbb{C}$ such that $ \alpha \neq 2j+1, \ \beta \neq 2k, \ \alpha+\beta \neq -m+2l+1, \ j, k, l \in \mathbb{N}_0$ one has
$$
U_{\alpha}^* \ast T_{\beta}^* = T_{\beta}^* \ast U_{\alpha}^* = c_m(\alpha - 1,\beta)\; U_{\alpha+\beta+m}^*
$$
\item[(iii)] For $(\alpha,\beta) \in \mathbb{C} \times \mathbb{C}$ such that $ \alpha \neq 2j+1, \ \beta \neq 2k+1, \ \alpha+\beta \neq -m+2l, \ j, k, l \in \mathbb{N}_0$ one has
$$
U_{\alpha}^* \ast U_{\beta}^* = U_{\beta}^* \ast U_{\alpha}^* = \pi^{\frac{m}{2}+1} \displaystyle{ \frac{\Gamma(- \frac{\alpha+\beta+m}{2})}{\Gamma(\frac{-\alpha+1}{2}) \Gamma(\frac{-\beta+1}{2})}} \; T_{\alpha+\beta+m}^*
$$
\end{itemize}
\end{proposition}
\begin{remark}
The action of a Clifford algebra valued distribution on a ditto test function is assumed to be carried out componentwise, the respective basis vectors being multiplied in the Clifford algebra.
\end{remark}
\begin{remark}
In general the convolution of Clifford algebra valued distributions is not commutative. However, as is seen from formula (iii) in Proposition \ref{prop1}, the convolution of two distributions from the $\mathcal{U}$--family is indeed commutative. We will frequently use this property in the sequel. Convolution by distributions from the $\mathcal{T}$--family is intrinsically commutative since they are scalar valued.
\end{remark}
\begin{remark}
In general the convolution of distributions is not associative. However, as is seen from the formulae in Proposition \ref{prop1}, the convolution of distributions from the $\mathcal{T}$-- and $\mathcal{U}$--families is associative. Also this property will be frequently used this in the sequel.
\end{remark}
\section{A conjugate harmonic to Green's function}
\label{conjharm}
The fundamental solution of the Laplace operator $\Delta_{m+1}$ in $\mathbb{R}^{m+1}$, sometimes called Green's function, and here denoted, for reasons which will become clear afterwards, by $\frac{1}{2}A_0(x_0,\underline{x})$, is given by
\begin{equation}
\frac{1}{2}A_0(x_0,\underline{x}) = - \frac{1}{m-1} \frac{1}{\sigma_{m+1}} \frac{1}{|x|^{m-1}}
\label{A0}
\end{equation}
Considering the function $A_0(x_0,\underline{x})$ as a harmonic function in the upper half--space $\mathbb{R}^{m+1}_+$, our aim now is to construct its conjugate harmonic in $\mathbb{R}^{m+1}_+$ in the sense of \cite{red}, in this way elaborating further on an earlier result of \cite{Xu}. This means that we have to look for a harmonic function $B_0(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$ such that
$$
C_0(x_0,\underline{x}) = \frac{1}{2} A_0(x_0,\underline{x}) + \frac{1}{2} \overline{e_0} \, B_0(x_0,\underline{x})
$$
is monogenic in $\mathbb{R}^{m+1}_+$ w.r.t.\ the generalized Cauchy--Riemann operator $D$. Expressing the monogenicity of $C_0$ in $\mathbb{R}^{m+1}_+$ leads to the system
\begin{equation}
\left \{ \begin{array}{rcl}
\partial_{x_0} A_0 + \underline{\partial} B_0 & = & 0 \\[2mm]
\partial_{x_0} B_0 + \underline{\partial} A_0 & = & 0
\end{array} \right .
\label{CR12}
\end{equation}
which clearly mimics the Cauchy--Riemann system in the complex plane. Taking into account the explicit expression (\ref{A0}) of $A_0(x_0,\underline{x})$, the system (\ref{CR12}) reduces to
\begin{equation}
\left \{ \begin{array}{rcl}
\underline{\partial} B_0(x_0,\underline{x}) &=& - \displaystyle\frac{2}{\sigma_{m+1}} \, \displaystyle\frac{x_0}{|x|^{m+1}} \ = \ - P(x_0,\underline{x}) \\[3mm]
\partial_{x_0} B_0(x_0,\underline{x}) &=& - \displaystyle\frac{2}{\sigma_{m+1}} \, \displaystyle\frac{\underline{x}}{|x|^{m+1}} \ = \ \phantom{-} Q(x_0,\underline{x})
\end{array} \right .
\label{CR34}
\end{equation}
where $P$ and $Q$ stand for the Poisson kernel and its conjugate in $\mathbb{R}^{m+1}_+$ (see also Section \ref{downstream}). From the second condition in (\ref{CR34}) its follows that, for an arbitrary, but fixed, $x_0^\ast$,
$$
B(x_0,\underline{x}) = \frac{2}{\sigma_{m+1}} \, \frac{\underline{x}}{|\underline{x}|^m} \, F_m \left ( \frac{|\underline{x}|}{x_0} \right ) - \frac{2}{\sigma_{m+1}} \, \frac{\underline{x}}{|\underline{x}|^m} \, F_m \left ( \frac{|\underline{x}|}{x_0^\ast} \right ) + W(\underline{x})
$$
where we have put
$$
F_m(v) = \int_0^v \frac{\eta^{m-1}}{(1+\eta^2)^\frac{m+1}{2}} \, d\eta = \frac{v^m}{m} \, _2F_1 \left ( \frac{m}{2},\frac{m+1}{2};\frac{m}{2}+1;-v^2 \right )
$$
with $_2F_1$ a standard hypergeometric function (see e.g. \cite{grad}).\\[-2mm]
From the first condition in (\ref{CR34}) it then follows that the function $W(\underline{x})$ should satisfy the equation
$$
\underline{\partial} W(\underline{x}) = - \left ( \partial_{x_0} A_0 \right )_{x_0^\ast} = - \frac{2}{\sigma_{m+1}} \, \frac{x_0^\ast}{|x_0^\ast e_0 + \underline{x}|^{m+1}}
$$
and a straightforward calculation shows that the function
$$
W(\underline{x}) = \frac{2}{\sigma_{m+1}} \, \frac{\underline{x}}{|\underline{x}|^m} \, F_m \left ( \frac{|\underline{x}|}{x_0^\ast} \right )
$$
does the job. A conjugate harmonic to $A_0$ in $\mathbb{R}_+^{m+1}$ is thus given by
\begin{equation}
B_0(x_0,\underline{x}) = \frac{2}{\sigma_{m+1}} \, \frac{\underline{x}}{|\underline{x}|^m} \, F_m \left ( \frac{|\underline{x}|}{x_0} \right )
\label{B0}
\end{equation}
or
\begin{equation}
B_0(x_0,\underline{x}) = \frac{2}{m} \, \frac{1}{\sigma_{m+1}} \, \frac{\underline{x}}{x_0^m} \, _2F_1 \left ( \frac{m}{2},\frac{m+1}{2};\frac{m}{2}+1;-\frac{|\underline{x}|^2}{x_0^2} \right )
\label{B02}
\end{equation}
Expression (\ref{B02}) clearly shows that $B_0(x_0,\underline{x})$ is well--defined for $\underline{x}=0$, with
$$
\lim_{\underline{x} \rightarrow 0} B_0(x_0,\underline{x}) = 0, \qquad x_0 > 0
$$
Taking into account that
$$
F_m(+\infty) = \int_0^{+\infty} \frac{\eta^{m-1}}{(1+\eta^2)^\frac{m+1}{2}} \, d\eta = \frac{\sqrt{\pi}}{2} \frac{\Gamma \left ( \frac{m}{2} \right )}{\Gamma \left ( \frac{m+1}{2} \right )}
$$
expression (\ref{B0}) leads to the following distributional limit
\begin{equation}
b_0(\underline{x}) = \lim_{x_0 \rightarrow 0+} B_0(x_0,\underline{x}) = \frac{1}{\sigma_m} \frac{\underline{x}}{|\underline{x}|^m} = \frac{1}{\pi} \, \frac{1}{\sigma_m} \, U^\ast_{-m+1}
\label{b0}
\end{equation}
in which one recognizes, up to a minus sign, the fundamental solution $E_m(\underline{x})$ of the Dirac operator $\underline{\partial}$ in $\mathbb{R}^m$:
$$
- b_0(\underline{x}) = \frac{1}{\sigma_m} \, \frac{\overline{\underline{x}}}{|\underline{x}|^m} = - \frac{1}{\pi} \, \frac{1}{\sigma_m} \, U^\ast_{-m+1} = E_m(\underline{x})
$$
This distribution $E_m(\underline{x})$ may act as a convolution kernel for the so--called $T$--operator, which is a convolution operator acting on Clifford algebra valued Schwartz--functions $f \in \mathcal{S}(\mathbb{R}^m)$ or on ditto tempered distributions as
$$
T[f] = E_m \ast f = - b_0 \ast f
$$
Seen the fact that $E_m(\underline{x})$ is the fundamental solution of the Dirac operator $\underline{\partial}$, this $T$--operator is an inverse to this Dirac operator:
$$
\underline{\partial} \, T[f] = f, \qquad f \in \mathcal{S}(\mathbb{R}^m)
$$
The Green function $A_0(x_0,\underline{x})$ itself shows the following distributional limit:
\begin{equation}
a_0(\underline{x}) = \lim_{x_0 \rightarrow 0+} A_0(x_0,\underline{x}) = - \frac{2}{m-1} \, \frac{1}{\sigma_{m+1}} \, {\rm Fp} \frac{1}{|\underline{x}|^{m-1}} =
- \frac{2}{m-1} \, \frac{1}{\sigma_{m+1}} \, T_{-m+1}^\ast
\label{a0}
\end{equation}
Using this distribution, up to a minus sign, as a convolution kernel, gives rise to the convolution operator $(-\Delta)^{-\frac{1}{2}}$, acting on Schwartz--functions or tempered distributions by, see e.g.\ \cite{helgason},
$$
\left ( - \Delta \right )^{- \frac{1}{2}} [f] = \frac{2}{m-1} \, \frac{1}{\sigma_{m+1}} \, T^\ast_{-m+1} \ast f = - a_0 \ast f
$$
The functions $A_0(x_0,\underline{x})$ and $B(x_0,\underline{x})$ being conjugate harmonic in $\mathbb{R}^{m+1}_{+}$, we expect their distributional boundary values $a_0(\underline{x})$ and $b_0(\underline{x})$ to be intimately related. This is indeed the case, as will be shown in Section \ref{upstream}.
\section{Downstream potentials}
\label{downstream}
\subsection{The Cauchy kernel as a potential}
As is well--known, the Cauchy kernel of Clifford analysis, i.e. the fundamental solution of the generalized Cauchy--Riemann operator $D$,
$$
C_{-1}(x_0,\underline{x}) = \frac{1}{\sigma_{m+1}} \, \frac{x \overline{e_0}}{|x|^{m+1}} = \frac{1}{\sigma_{m+1}} \, \frac{x_0 -
\overline{e_0} \underline{x}}{|x|^{m+1}}
$$
may be decomposed in terms of the Poisson kernels in $\mathbb{R}^{m+1}_+$:
$$
C_{-1}(x_0,\underline{x}) = \frac{1}{2} A_{-1}(x_0,\underline{x}) + \frac{1}{2} \overline{e_0} \, B_{-1}(x_0,\underline{x})
$$
where, also mentioning the traditional notations, for $x_0 >0$,
\begin{equation}
\left \{ \begin{array}{rcl}
A_{-1}(x_0,\underline{x}) & = & P(x_0,\underline{x}) \ = \ \phantom{-} \frac{2}{\sigma_{m+1}} \, \frac{x_0}{|x|^{m+1}}\\[4mm]
B_{-1}(x_0,\underline{x}) & = & Q(x_0,\underline{x}) \ = \ - \frac{2}{\sigma_{m+1}} \, \frac{\underline{x}}{|x|^{m+1}}
\end{array} \right .
\label{A-1B-1}
\end{equation}
Note that the Poisson kernel $A_{-1}$ is real--valued, while its conjugate harmonic kernel $B_{-1}$ is Clifford vector--valued. Their distributional limits for $x_0 \rightarrow 0+$ are given by
\begin{eqnarray*}
a_{-1}(\underline{x}) & = & \lim_{x_0 \rightarrow 0+} A_{-1}(x_0,\underline{x}) \ = \ \delta(\underline{x}) \ \ = \ \phantom{-} \frac{2}{\sigma_m} \, T^\ast_{-m} \\
b_{-1}(\underline{x}) & = & \lim_{x_0 \rightarrow 0+} B_{-1}(x_0,\underline{x}) \ = \ H(\underline{x}) \ = \ - \frac{2}{\sigma_{m+1}} \, U^\ast_{-m}
\end{eqnarray*}
and also
$$
c_{-1}(\underline{x}) = \lim_{x_0 \rightarrow 0+} C_{-1}(x_0,\underline{x}) \ = \ \frac{1}{2} \delta(\underline{x}) + \frac{1}{2} \overline{e_0} \, H(\underline{x})
$$
Note that the distribution
$$
H(\underline{x}) = - \frac{2}{\sigma_{m+1}} \, U^\ast_{-m} = - \frac{2}{\sigma_{m+1}} \, \mbox{Pv} \frac{\underline{x}}{|\underline{x}|^{m+1}}
$$
where $\mbox{Pv}$ stands for the {\em principal value} distribution in $\mathbb{R}^m$, is the convolution kernel of the Hilbert transform $\mathcal{H}$ in $\mathbb{R}^m$ (see e.g.\ \cite{gilmur}). Note also that both distributional boundary values are linked by this Hilbert transform:
\begin{eqnarray*}
\mathcal{H} \left [ a_{-1} \right ] & = & \mathcal{H} \left [ \delta \right ] \ = \ H \ast \delta \ = \ H \ = \ b_{-1} \\
\mathcal{H} \left [ b_{-1} \right ] & = & \mathcal{H} \left [ H \right ] \ = \ H \ast H \ = \ \delta \ = \ a_{-1}
\end{eqnarray*}
since $\mathcal{H}^2 = \mathbf{1}$, while$$\overline{e_0} \, \mathcal{H} \left [ c_{-1} \right ] = c_{-1}$$
Conversely, the Poisson kernels are the Poisson transforms of these distributional limits:
\begin{eqnarray*}
\mathcal{P} \left [ a_{-1} \right ] & = & P(x_0,\cdot) \ast a_{-1}(\cdot)(\underline{x}) \ = \ P(x_0,\cdot) \ast \delta(\cdot)(\underline{x}) \ = \ P(x_0,\underline{x})\\
\mathcal{P} \left [ b_{-1} \right ] & = & P(x_0,\cdot) \ast b_{-1}(\cdot)(\underline{x}) \ = \ P(x_0,\cdot) \ast H(\cdot)(\underline{x}) \ = \ Q(x_0,\underline{x})
\end{eqnarray*}
It follows that also the Poisson kernels themselves are linked by the Hilbert transform in the variable $\underline{x} \in \mathbb{R}^m$:
\begin{eqnarray*}
\mathcal{H} \left [ A_{-1} \right ] & = & H(\cdot) \ast A_{-1}(x_0,\cdot)(\underline{x}) \ = \ H(\cdot) \ast P(x_0,\cdot)(\underline{x}) \\
& = & P(x_0,\cdot) \ast H(\cdot)(\underline{x}) \ = \ Q(x_0,\underline{x}) \ = \ B_{-1}(x_0,\underline{x}) \\[2mm]
\mathcal{H} \left [ B_{-1} \right ] & = & \mathcal{H}^2 \left [ A_{-1} \right ] \ = \ A_{-1}
\end{eqnarray*}
For a function $f \in L_2(\mathbb{R}^m)$, its Poisson transforms
\begin{eqnarray*}
\mathcal{P}[f] & = & P(x_0,\cdot) \ast f(\cdot)(\underline{x}) \ =\ A_{-1}(x_0,\cdot) \ast f(\cdot)(\underline{x}) \\
\mathcal{Q}[f] & = & Q(x_0,\cdot) \ast f(\cdot)(\underline{x}) \ = \ B_{-1}(x_0,\cdot) \ast f(\cdot)(\underline{x})
\end{eqnarray*}
belong to the Clifford--Hardy space $\mbox{Harm}^2 \left ( \mathbb{R}^{m+1}_+ \right )$ of Clifford algebra valued harmonic functions in $\mathbb{R}_+^{m+1}$:
$$
\mbox{Harm}^2 \left( \mathbb{R}^{m+1}_+ \right) = \left \{ F(x_0,\underline{x}) \, : \, F \mbox{\ is harmonic in $\mathbb{R}_+^{m+1}$ and\ } \sup_{x_0 > 0} \int_{\mathbb{R}^m} |F(x_0,\underline{x})|^2 \, d\underline{x} \; < \; + \infty \right \}
$$
and show the non--tangential $L_2$--boundary values $$\lim_{x_0 \rightarrow 0+} \mathcal{P}[f] = f \quad {\rm and} \quad \lim_{x_0 \rightarrow 0+} \mathcal{Q}[f] = \mathcal{H}[f]$$ with
$$
\mathcal{H}[f] = H \ast f = \frac{2}{\sigma_{m+1}} \, \mbox{Pv} \int_{\mathbb{R}^m} \frac{\overline{\underline{u}}}{|\underline{u}|^{m+1}} \, f(\underline{x} - \underline{u}) \, d\underline{u}
$$
the explicit expression for the Hilbert transform of $f$. In its turn the Cauchy transform of $f \in L_2(\mathbb{R}^m)$, given by
$$
\mathcal{C}[f] = C_{-1}(x_0,\cdot) \ast f(\cdot)(\underline{x}) = \frac{1}{2} \mathcal{P}[f] + \frac{1}{2} \overline{e_0} \mathcal{Q}[f]
$$
belongs to the Clifford--Hardy space $H^2(\mathbb{R}^{m+1}_+)$ of monogenic functions in $\mathbb{R}^{m+1}_+$:
$$
H^2 \left ( \mathbb{R}^{m+1}_+ \right ) = \left \{ F(x_0,\underline{x}) \, : \, F \mbox{\ is monogenic in $\mathbb{R}_+^{m+1}$ and\ } \sup_{x_0 > 0} \int_{\mathbb{R}^m} |F(x_0,\underline{x})|^2 \, d\underline{x} \; < \; + \infty \right \}
$$
and shows the following non--tangential $L_2$--boundary value:
$$
\lim_{x_0 \rightarrow 0+} \mathcal{C}[f] = \frac{1}{2} f + \frac{1}{2} \overline{e_0} \mathcal{H}[f] = \left ( \frac{1}{2} \delta + \frac{1}{2} \overline{e_0} H \right ) \ast f = \mathcal{A}\mathcal{S}[f]
$$
which belongs to the Clifford--Hardy space $H^2(\mathbb{R}^m)$, see \cite{gilmur}. In signal analysis the functions in $H^2(\mathbb{R}^m)$ are called {\em analytic signals} ; they show no negative-frequency components (see e.g. \cite{hahn}). Whence the notation $\mathcal{A}\mathcal{S}$ for the boundary value of the Cauchy transform. Note that for this Cauchy transform we have several equivalent expressions:
$$
\mathcal{C} [f] = \mathcal{C} \left [\overline{e_0} \, \mathcal{H}[f] \right ] = \mathcal{C} \left [ \mathcal{A}\mathcal{S}[f] \right ] = \mathcal{P} \left [ \mathcal{A}\mathcal{S}[f] \right ]
$$
From the monogenicity of the Cauchy kernel $C_{-1}(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$, i.e.
$$
DC_{-1} = \frac{1}{2} \left ( \partial_{x_0} + \overline{e_0} \underline{\partial} \right ) C_{-1}=0
$$
it follows that the Poisson kernels $A_{-1}(x_0,\underline{x})$ and $B_{-1}(x_0,\underline{x})$ satisfy the generalized Cauchy--Riemann system
\begin{equation}
\left \{ \begin{array}{rcl}
\partial_{x_0} A_{-1} + \underline{\partial} B_{-1} & = & 0 \\[2mm]
\partial_{x_0} B_{-1} + \underline{\partial} A_{-1} & = & 0 \\
\end{array} \right .
\label{GCR}
\end{equation}
and that
\begin{equation}
\overline{D} C_{-1} = \frac{1}{2} \partial_{x_0} C_{-1} - \frac{1}{2} \overline{e_0} \underline{\partial} C_{-1} = \partial_{x_0} C_{-1} = - \overline{e_0} \underline{\partial} C_{-1}
\label{DC-1}
\end{equation}
and also that
\begin{equation}
\left \{ \begin{array}{l}
\overline{D} A_{-1} = \frac{1}{2} \partial_{x_0} A_{-1} - \frac{1}{2} \overline{e_0} \underline{\partial} A_{-1} = \frac{1}{2} \partial_{x_0} A_{-1} + \frac{1}{2} \overline{e_0} \partial_{x_0} B_{-1} = \partial_{x_0} C_{-1} = \overline{D} C_{-1} \\[2mm]
\overline{D}(\overline{e_0} B_{-1}) = \frac{1}{2} \overline{e_0} \partial_{x_0} B_{-1} - \frac{1}{2} \overline{e_0} \underline{\partial} \overline{e_0} B_{-1} = \frac{1}{2} \overline{e_0} \partial_{x_0} B_{-1} + \frac{1}{2} \partial_{x_0} A_{-1} = \partial_{x_0} C_{-1} = \overline{D} C_{-1}
\end{array} \right .
\label{DAB-1}
\end{equation}
Now we put
$$
\overline{D} C_{-1} = C_{-2} = \frac{1}{2} A_{-2} + \frac{1}{2} \overline{e_0} B_{-2}
$$
clearly a monogenic function in $\mathbb{R}^{m+1}_+$, since $DC_{-2} = D \overline{D} C_{-1} = \frac{1}{4} \Delta_{m+1} C_{-1} = 0$. From this definition it follows that
$$
\left \{ \begin{array}{l}
A_{-2} = \partial_{x_0} A_{-1} = - \underline{\partial} B_{-1} \\[2mm]
B_{-2} = \partial_{x_0} B_{-1} = - \underline{\partial} A_{-1}
\end{array} \right .
$$
leading to the explicit expressions for the conjugate harmonic components of $C_{-2}$:
\begin{equation}
\left \{ \begin{array}{rcl}
A_{-2} & = & \displaystyle\frac{2}{\sigma_{m+1}} \, \displaystyle\frac{1}{|x|^{m+3}} \left ( |x|^2 - (m+1) x_0^2 \right ) = \displaystyle\frac{2}{\sigma_{m+1}} \, \displaystyle\frac{- m x_0^2 + |\underline{x}|^2}{|x|^{m+3}} \\[4mm]
B_{-2} & = & (m+1) \, \displaystyle\frac{2}{\sigma_{m+1}} \, \displaystyle\frac{x_0 \underline{x}}{|x|^{m+3}}
\end{array} \right .
\label{A-2B-2}
\end{equation}
Note that $A_{-2}(x_0, \underline{x})$ is real--valued, while $B_{-2}(x_0, \underline{x})$ is Clifford vector--valued. Moreover it is readily confirmed that they satisfy the generalized CR--system
$$
\left \{ \begin{array}{l}
\partial_{x_0} A_{-2} + \underline{\partial} B_{-2} = 0 \\[2mm]
\partial_{x_0} B_{-2} + \underline{\partial} A_{-2} = 0
\end{array} \right .
$$
The above relations (\ref{DC-1})--(\ref{DAB-1}) imply that the monogenic function $C_{-2}(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$ shows the monogenic potential (or primitive) $C_{-1}(x_0,\underline{x})$ and the conjugate harmonic potentials $A_{-2}(x_0,\underline{x})$ and $\overline{e_0} B_{-2}(x_0,\underline{x})$. The distributional limits for $x_0 \rightarrow 0+$ of these harmonic potentials are given by
$$
\left \{ \begin{array}{rcl}
a_{-2}(\underline{x}) = \lim_{x_0 \rightarrow 0+} A_{-2}(x_0,\underline{x}) & = & \displaystyle\frac{2}{\sigma_{m+1}} \, {\rm Fp} \displaystyle\frac{1}{|\underline{x}|^{m+1}} \ = \ - \displaystyle\frac{4 \pi}{\sigma_{m+1}} T^\ast_{-m-1}\\[4mm]
b_{-2}(\underline{x}) = \lim_{x_0 \rightarrow 0+} B_{-2}(x_0,\underline{x}) & = & - \underline{\partial} \delta \ = \ \displaystyle\frac{2m}{\sigma_m} \, U^\ast_{-m-1}
\end{array} \right .
$$
Conversely, the harmonic potentials $A_{-2}(x_0,\underline{x})$ and $B_{-2}(x_0,\underline{x})$ are recovered from these distributional boundary values by the Poisson transform
\begin{eqnarray*}
A_{-2}(x_0,\underline{x}) & = & \mathcal{P} \left [ a_{-2}(\underline{x}) \right ] \ = \ P(x_0,\cdot) \ast a_{-2}(\cdot)(\underline{x}) \ = \ A_{-1}(x_0,\cdot) \ast a_{-2}(\cdot)(\underline{x}) \ = \ a_{-2}\ast A_{-1} \\
& = & \mathcal{Q} \left [ b_{-2}(\underline{x}) \right ] \ = \ Q(x_0,\cdot) \ast b_{-2}(\cdot)(\underline{x}) \ = \ B_{-1}(x_0,\cdot) \ast b_{-2}(\cdot)(\underline{x}) \ = \ b_{-2} \ast B_{-1}
\end{eqnarray*}
and
\begin{eqnarray*}
B_{-2}(x_0,\underline{x}) & = & \mathcal{P} \left [ b_{-2}(\underline{x}) \right ] \ = \ P(x_0,\cdot) \ast b_{-2}(\cdot)(\underline{x}) \ = \ A_{-1}(x_0,\cdot) \ast b_{-2}(\cdot)(\underline{x}) \ = \ b_{-2} \ast A_{-1}\\
& = & \mathcal{Q} \left [ a_{-2}(\underline{x}) \right ] \ = \ Q(x_0,\cdot) \ast a_{-2}(\cdot)(\underline{x}) \ = \ B_{-1}(x_0,\cdot) \ast a_{-2}(\cdot)(\underline{x}) \ = \ a_{-2} \ast B_{-1}
\end{eqnarray*}
In the distribution $a_{-2}$ one recognizes the convolution kernel $-\underline{\partial} H = - H \underline{\partial}$, known as the Hilbert--Dirac kernel, see \cite{fbhds}, or perhaps better known as the convolution kernel for the pseudodifferential operator $(-\Delta)^\frac{1}{2}$ (see \cite{helgason}). The distribution $b_{-2}$ is, up to a minus sign, the Dirac derivative of the delta-distribution. Both distributional boundary values are linked by the Hilbert transform, as shown a.o.\ in the following lemma.
\begin{lemma}
One has
\begin{itemize}
\item[(i)] $- \underline{\partial} a_{-1} = b_{-2}$, $-\underline{\partial} b_{-1} = a_{-2}$, $- \overline{e_0} \underline{\partial} c_{-1} = c_{-2}$
\item[(ii)] $\mathcal{H} \left [ a_{-2} \right ] = b_{-2}$, $\mathcal{H} \left [ b_{-2} \right ] = a_{-2}$, $\overline{e_0} \mathcal{H} \left [ c_{-2} \right ] = c_{-2}$
\item[(iii)] $c_{-1} \ast a_{-2} = c_{-2}$, $c_{-1} \ast \overline{e_0} b_{-2} = c_{-2}$, $c_{-1} \ast c_{-2} = c_{-2}$
\end{itemize}
\end{lemma}
\par\noindent {\bf Proof}~\par\noindent
\noindent (i) Follows by direct calculation.\\[-2mm]
\noindent (ii) Making use of the convolution calculation rules, recalled in Proposition \ref{prop1}, we have
\begin{eqnarray*}
\mathcal{H} \left [ a_{-2} \right ] & = & - \frac{4 \pi}{\sigma_{m+1}} \, H \ast T^\ast_{-m-1} \ = \ \frac{8 \pi}{\sigma_{m+1}^2} U^\ast_{-m} \ast T^\ast_{-m-1} \\
& = & \frac{8 \pi}{\sigma_{m+1}^2} \, \pi^\frac{m}{2} \, \frac{\Gamma \left ( \frac{m+2}{2} \right )}{\left ( \Gamma \left ( \frac{m+1}{2} \right ) \right )^2} \, U^\ast_{-m-1} \ = \ \frac{2m}{\sigma_m} \, U^\ast_{-m-1} \ = \ - \underline{\partial} \delta \ = \ b_{-2}
\end{eqnarray*}
and
\begin{eqnarray*}
\mathcal{H} \left [b_{-2} \right ] & = & \frac{2m}{\sigma_m} \, H \ast U^\ast_{-m-1} \ = \ - \frac{4m}{\sigma_m \sigma_{m+1}} U^\ast_{-m} \ast U^\ast_{-m-1} \\
& = & - \frac{4m}{\sigma_m \sigma_{m+1}} \, \pi^\frac{m}{2} \, \frac{\Gamma \left ( \frac{m+1}{2} \right )}{\Gamma \left ( \frac{m+1}{2} \right ) \Gamma \left ( \frac{m+2}{2} \right )} \, T^\ast_{-m-1} \ = \ - \frac{4}{\sigma_{m+1}} \, T^\ast_{-m-1} \ = \ a_{-2}
\end{eqnarray*}
\noindent (iii) We subsequently find
$$
c_{-1} \ast a_{-2} = \left ( \frac{1}{2} \delta + \frac{1}{2} \overline{e_0} H \right ) \ast a_{-2} = \frac{1}{2} a_{-2} + \frac{1}{2} \overline{e_0} \mathcal{H} \left [ a_{-2} \right ] = \frac{1}{2} a_{-2} + \frac{1}{2} \overline{e_0} b_{-2} = c_{-2}
$$
$$
c_{-1} \ast \overline{e_0} b_{-2} = \left ( \frac{1}{2} \delta + \frac{1}{2} \overline{e_0} H \right ) \ast \overline{e_0} b_{-2} = \frac{1}{2} \overline{e_0} b_{-2} + \frac{1}{2} \mathcal{H} \left [ b_{-2} \right ] = \frac{1}{2} \overline{e_0} b_{-2} + \frac{1}{2} a_{-2} = c_{-2}
$$
and
$$
c_{-1} \ast c_{-2} = c_{-1} \ast \left ( \frac{1}{2} a_{-2} + \frac{1}{2} \overline{e_0} b_{-2} \right ) = \frac{1}{2} c_{-2} + \frac{1}{2} c_{-2} = c_{-2}
$$
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
Through the Poisson transform, the Hilbert--link between the distributional boundary values $a_{-2}$ and $b_{-2}$ is reflected in a similar relationship between the harmonic potentials $A_{-2}$ and $B_{-2}$, as it was also the case for $A_{-1}$ and $B_{-1}$. Indeed, one has
\begin{eqnarray*}
\mathcal{H} \left [ A_{-2} \right ] & = & H(\cdot) \ast A_{-2}(x_0,\cdot)(\underline{x}) \ = \ H(\cdot) \ast P(x_0,\cdot) \ast a_{-2}(\cdot)(\underline{x}) \\
& = & P(x_0,\cdot) \ast H(\cdot) \ast a_{-2}(\cdot)(\underline{x}) \ = \ P(x_0,\cdot) \ast b_{-2}(\cdot)(\underline{x}) \ = \ B_{-2}(x_0,\underline{x})\\
\mathcal{H} \left [ B_{-2} \right ] & = & \mathcal{H}^2 \left [ A_{-2} \right ] = A_{-2}
\end{eqnarray*}
These relations may also be rewritten as
\begin{eqnarray*}
b_{-1}(\cdot) \ast A_{-2}(x_0,\cdot)(\underline{x}) = B_{-2}(x_0,\underline{x}) \\
b_{-1}(\cdot) \ast B_{-2}(x_0,\cdot)(\underline{x}) = A_{-2}(x_0,\underline{x})
\end{eqnarray*}
while, quite trivially,
\begin{eqnarray*}
a_{-1}(\cdot) \ast A_{-2}(x_0,\cdot)(\underline{x}) = A_{-2}(x_0,\underline{x}) \\
a_{-1}(\cdot) \ast B_{-2}(x_0,\cdot)(\underline{x}) = B_{-2}(x_0,\underline{x})
\end{eqnarray*}
Note the following two commutative schemes, which are each others Hilbert image:
\begin{equation}
\begin{array}{ccc}
A_{-1} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & B_{-2} \\[2mm]
\hspace*{-8mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
\delta = a_{-1} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & b_{-2} = - \underline{\partial} \delta
\end{array}
\qquad \qquad \mbox{and} \qquad \qquad
\begin{array}{ccc}
B_{-1} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & A_{-2} \\[2mm]
\hspace*{-7mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
H = b_{-1} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & a_{-2} = - \underline{\partial} H
\end{array}
\label{cscheme1}
\end{equation}
By means of the distributional limits $a_{-2}$ and $b_{-2}$, we are now able to prove some remarkable relations between the conjugate harmonic components of $C_{-1}$ and $C_{-2}$; in fact they are shown to be linked by the distributional limits $a_0$ and $b_0$ of the Green function and its conjugate (see Section \ref{conjharm}).
\begin{proposition}
\label{propdownstream}
One has, convolutions being taken in the variable $\underline{x} \in \mathbb{R}^m$:
\begin{itemize}
\item[(i)] $a_0(\cdot) \ast A_{-2}(x_0,\cdot)(\underline{x}) = A_{-1}(x_0,\underline{x}) = b_0(\cdot) \ast B_{-2}(x_0,\underline{x})$
\item[(ii)] $a_0(\cdot) \ast B_{-2}(x_0,\cdot)(\underline{x}) = B_{-1}(x_0,\underline{x}) = b_0(\cdot) \ast A_{-2}(x_0,\underline{x})$
\item[(iii)] $a_0(\cdot) \ast C_{-2}(x_0,\cdot)(\underline{x}) = C_{-1}(x_0,\underline{x}) = b_0(\cdot) \ast C_{-2}(x_0,\underline{x})$
\item[(iv)] $c_0(\cdot) \ast A_{-2}(x_0,\cdot)(\underline{x}) = c_0(\cdot) \ast B_{-2}(x_0,\underline{x}) = c_0(\cdot) \ast C_{-2}(x_0,\underline{x}) = C_{-1}(x_0,\underline{x})$
\end{itemize}
\end{proposition}
\par\noindent {\bf Proof}~\par\noindent
\noindent (i)(ii) Put $b_0 \ast B_{-2} = A'_{-1}$. Then
\begin{eqnarray*}
\partial_{x_0} A'_{-1} & = & b_0 \ast \partial_{x_0} B_{-2} = b_0 \ast \left ( - \underline{\partial} A_{-2} \right ) = - b_0 \underline{\partial} \ast A_{-2} = \delta \ast A_{-2} = A_{-2} \\
\underline{\partial} A'_{-1} & = & \underline{\partial} b_0 \ast B_{-2} = - \delta \ast B_{-2} = B_{-2}
\end{eqnarray*}
while moreover
$$
\lim_{x_0 \rightarrow 0+} A'_{-1} = b_0 \ast b_{-2} = b_0 \ast (-\underline{\partial} \delta) = - b_0 \underline{\partial} \ast \delta = \delta \ast \delta = \delta = a_{-1}
$$
Similarly, by putting $b_0 \ast A_{-2} = B'_{-1}$, we have
\begin{eqnarray*}
\partial_{x_0} B'_{-1} & = & b_0 \ast \partial_{x_0} A_{-2} = b_0 \ast \left ( - \underline{\partial} B_{-2} \right ) = - b_0 \underline{\partial} \ast B_{-2} = \delta \ast B_{-2} = B_{-2} \\
\underline{\partial} B'_{-1} & = & \underline{\partial} b_0 \ast A_{-2} = - \delta \ast A_{-2} = A_{-2}
\end{eqnarray*}
while moreover
$$
\lim_{x_0 \rightarrow 0+} B'_{-1} = b_0 \ast a_{-2} = b_0 \ast (-\underline{\partial} H) = - b_0 \underline{\partial} \ast H = \delta \ast H = H = b_{-1}
$$
So $A'_{-1}$ and $B'_{-1}$ satisfy the CR--system (\ref{GCR}) and show the same distributional limits for $x_0 \rightarrow 0+$ as $A_{-1}$ and $B_{-1}$, respectively, from which it follows that they have to coincide: $A'_{-1} = A_{-1}$ and $B'_{-1} = B_{-1}$. Now note that
\begin{eqnarray*}
a_0 \ast A_{-2} & = & a_0 \ast \mathcal{H} \left [ B_{-2} \right ] = a_0 \ast H \ast B_{-2} = H \ast a_0 \ast B_{-2} = b_0 \ast B_{-2} = A_{-1} \\
a_0 \ast B_{-2} & = & a_0 \ast \mathcal{H} \left [ A_{-2} \right ] = a_0 \ast H \ast A_{-2} = H \ast a_0 \ast A_{-2} = b_0 \ast A_{-2} = B_{-1}
\end{eqnarray*}
to complete the proof of (i) and (ii).\\[-2mm]
\noindent (iii)(iv) It suffices to make the appropriate combinations of the results in (i) and (ii).
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
For a function $f \in L_2(\mathbb{R}^m)$ we can define in $\mathbb{R}^{m+1}_+$ the conjugate harmonic functions
$$
\mathcal{A}_{-2} [f] = A_{-2}(x_0,\cdot) \ast f(\cdot) (\underline{x}) \qquad \mbox{and} \qquad \mathcal{B}_{-2}[f] = B_{-2}(x_0,\cdot) \ast f(\cdot) (\underline{x})
$$
and the monogenic function
$$
\mathcal{C}_{-2} [f] = C_{-2}(x_0,\cdot) \ast f(\cdot) (\underline{x})
$$
They show non--tangential $L_2$--boundary values on condition that $f$ belongs to the Clifford--Sobolev space
$$
W_2^1(\mathbb{R}^m) = \left \{ f \in L_2(\mathbb{R}^m) \, : \, \underline{\partial} f \in L_2(\mathbb{R}^m) \right \}
$$
Under these assumptions there holds
\begin{eqnarray*}
\mathcal{A}_{-2}^+[f] = \lim_{x_0 \rightarrow 0+} \mathcal{A}_{-2}[f] &=& a_{-2} \ast f \ = \ - H \underline{\partial} \ast f \ = \ - \mathcal{H} \left [ \underline{\partial} f \right ] \ = \ - \underline{\partial} \mathcal{H}[f] \\
\mathcal{B}_{-2}^+[f] = \lim_{x_0 \rightarrow 0+} \mathcal{B}_{-2}[f] &=& b_{-2} \ast f \ = \ - \underline{\partial} \delta \ast f \ = \ - \underline{\partial} f
\end{eqnarray*}
and also
$$
\mathcal{C}_{-2}^+[f] = \lim_{x_0 \rightarrow 0+} \mathcal{C}_{-2}[f] = - \frac{1}{2} \underline{\partial} \mathcal{H} [f] - \frac{1}{2} \overline{e_0} \underline{\partial} f = (- \overline{e_0} \underline{\partial}) \left ( \frac{1}{2} f + \frac{1}{2} \overline{e_0} \mathcal{H}[f] \right ) = (- \overline{e_0} \underline{\partial}) \left ( \mathcal{A} \mathcal{S} [f] \right )
$$
Note that the convolution operators $\mathcal{A}_{-2}^+$, $\mathcal{B}_{-2}^+$ and $\mathcal{C}_{-2}^+$ are bounded operators from $W_2^1(\mathbb{R}^m)$ into $L_2(\mathbb{R}^m)$, and that, for $f \in W_2^1(\mathbb{R}^m)$, the $L_2$--boundary value
$\mathcal{C}_{-2}^+[f]$ belongs to the Clifford--Hardy space $H^2(\mathbb{R}^m)$. Also note the following commutative scheme for a function $f \in W_2^1(\mathbb{R}^m)$:
$$
\begin{array}{rcl}
\mathcal{C}_{-1}[f] & \xrightarrow{\hspace*{2mm} -\overline{e_0} \underline{\partial} \hspace*{2mm}} & \mathcal{C}_{-2}[f] \\[2mm]
^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow \hspace*{4mm} & & \hspace*{4mm} \downarrow ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}}\\
\frac{1}{2} f + \frac{1}{2} \overline{e_0} \mathcal{H}[f] = \mathcal{A}\mathcal{S}[f] & \xrightarrow{\hspace*{2mm} -\overline{e_0} \underline{\partial} \hspace*{2mm}} & - \overline{e_0} \underline{\partial} \mathcal{A} \mathcal{S}[f] = - \frac{1}{2} \underline{\partial} \mathcal{H}[f] - \frac{1}{2} \overline{e_0} \underline{\partial} f
\end{array}
$$
which reflects at the level of the operators, the commutative schemes (\ref{cscheme1}) at the level of the convolution kernels.
\subsection{Further derived potentials}
Proceeding in the same manner as in Subsection 4.1, we can define a sequence of monogenic potentials in $\mathbb{R}_+^{m+1}$:
$$
C_{-k-1} = \overline{D} C_{-k} = \overline{D}^2 C_{-k+1} = \ldots = \overline{D}^k C_{-1}, \qquad k=1,2,\ldots
$$
where each monogenic potential decomposes into two conjugate harmonic potentials:
$$
C_{-k-1} = \frac{1}{2} A_{-k-1} + \frac{1}{2} \overline{e_0} B_{-k-1}, \qquad k=1,2,\ldots
$$
with, for $k$ odd, say $k=2\ell-1$,
$$
\left \{ \begin{array}{rcl}
A_{-2\ell} & = & \partial_{x_0}^{2\ell-1} A_{-1} \ = \ - \partial_{x_0}^{2\ell-2} \underline{\partial} B_{-1} \ = \ \ldots \ = \ - \underline{\partial}^{2\ell-1} B_{-1} \\[2mm]
B_{-2\ell} & = & \partial_{x_0}^{2\ell-1} B_{-1} \ = \ - \partial_{x_0}^{2\ell-2} \underline{\partial} A_{-1} \ = \ \ldots \ = \ - \underline{\partial}^{2\ell-1} A_{-1}
\end{array} \right .
$$
while for $k$ even, say $k=2\ell$,
$$
\left \{ \begin{array}{rcl}
A_{-2\ell-1} & = & \partial_{x_0}^{2\ell} A_{-1} \ = \ - \partial_{x_0}^{2\ell-1} \underline{\partial} B_{-1} \ = \ \ldots \ = \ \underline{\partial}^{2\ell} A_{-1} \\[2mm]
B_{-2\ell-1} & = & \partial_{x_0}^{2\ell} B_{-1} \ = \ - \partial_{x_0}^{2\ell-1} \underline{\partial} A_{-1} \ = \ \ldots \ = \ \underline{\partial}^{2\ell} B_{-1}
\end{array} \right .
$$
Note that also holds
$$
C_{-k-1} = \partial_{x_0} C_{-k} = (-\overline{e_0} \underline{\partial}) C_{-k} = \partial_{x_0}^2 C_{-k+1} = (-\overline{e_0} \underline{\partial})^2 C_{-k+1} = \ldots = \partial_{x_0}^k C_{-1} = (-\overline{e_0} \underline{\partial})^k C_{-1}
$$
while the conjugate harmonic components satisfy the recurrence relations
\begin{equation}
\left \{ \begin{array}{rcl}
A_{-k-1} & = & \partial_{x_0} A_{-k} = - \underline{\partial} B_{-k} \\[2mm]
B_{-k-1} & = & \partial_{x_0} B_{-k} = - \underline{\partial} A_{-k}
\end{array} \right .
\label{AB-k-1}
\end{equation}
whence
$$
\left \{ \begin{array}{rcl}
\overline{D} A_{-k} & = & \displaystyle\frac{1}{2} \left ( \partial_{x_0} - \overline{e_0} \underline{\partial} \right ) A_{-k} = \displaystyle\frac{1}{2} A_{-k-1} + \displaystyle\frac{1}{2} \overline{e_0} B_{-k-1} = C_{-k-1} \\[4mm]
\overline{D} ( \overline{e_0} B_{-k}) & = & \displaystyle\frac{1}{2} \left ( \partial_{x_0} \overline{e_0} - \underline{\partial} \right ) B_{-k} = \displaystyle\frac{1}{2} A_{-k-1} + \displaystyle\frac{1}{2} \overline{e_0} B_{-k-1} = C_{-k-1}
\end{array} \right .
$$
which expresses the fact that $A_{-k}$ and $\overline{e_0} B_{-k}$ are indeed potentials (or primitives) of $C_{-k-1}$. Their distributional limits for $x_0 \rightarrow 0+$ are given by
$$
\left \{ \begin{array}{rcl}
a_{-2\ell} & = & (- \underline{\partial})^{2\ell-1} H = - 2^{2\ell-1} \displaystyle\frac{\Gamma \left ( \frac{m+2\ell-1}{2} \right )} {\pi^{\frac{m-2\ell+1}{2}}} \, T^{\ast}_{-m-2\ell+1} \\[5mm]
& = & (-1)^{\ell-1} 2^{\ell-1} (2\ell-1)!! \displaystyle\frac{\Gamma \left ( \frac{m+2\ell-1}{2} \right )} {\pi^{\frac{m+1}{2}}} \, {\rm Fp} \displaystyle\frac{1}{r^{m+2\ell-1}} \\[7mm]
b_{-2\ell} & = & (- \underline{\partial})^{2\ell-1} \delta = 2^{2\ell-1} \displaystyle\frac{\Gamma \left (\frac{m+2\ell}{2} \right )} {\pi^{\frac{m-2\ell+2}{2}}} \, U^{\ast}_{-m-2\ell+1} \end{array} \right .
$$
and
$$
\left \{ \begin{array}{rcl}
a_{-2\ell-1} & = & \underline{\partial}^{2\ell} \delta = 2^{2\ell} \displaystyle\frac{\Gamma \left (\frac{m+2\ell}{2} \right )} {\pi^\frac{m-2\ell}{2}} \, T^{\ast}_{-m-2\ell}
\\[7mm]
b_{-2\ell-1} & = & \underline{\partial}^{2\ell} H \ = \ - 2^{2\ell} \displaystyle\frac{\Gamma \left ( \frac{m+2\ell+1}{2} \right )} {\pi^\frac{m-2\ell+1}{2}} \, U^{\ast}_{-m-2\ell} \\[5mm]
& = &(-1)^{\ell-1} 2^{\ell} (2\ell-1)!! \displaystyle\frac{\Gamma \left ( \frac{m+2\ell+1}{2} \right )}{\pi^\frac{m+1}{2}} \, {\rm Fp} \displaystyle\frac{1}{r^{m+2\ell}} \, \omega \end{array} \right .
$$
They show the following properties, which can be verified by direct calculation.
\begin{lemma}
\label{lem2}
One has for $j,k=1,2,\ldots$
\begin{itemize}
\item[(i)] $a_{-k} \xrightarrow{\hspace*{1mm} -\underline{\partial} \hspace*{1mm}} b_{-k-1} \xrightarrow{\hspace*{1mm} -\underline{\partial} \hspace*{1mm}} a_{-k-2}$
\item[(ii)] $\mathcal{H} \left [ a_{-k} \right ] = b_{-k}$, $\mathcal{H} \left [ b_{-k} \right ] = a_{-k}$
\item[(iii)] $a_{-j} \ast a_{-k} = a_{-j-k+1}$ \\
$a_{-j} \ast b_{-k} = b_{-j} \ast a_{-k} = b_{-j-k+1}$ \\
$b_{-j} \ast b_{-k} = a_{-j-k+1}$.
\end{itemize}
\end{lemma}
Through the Poisson transform, the above Hilbert--link (Lemma \ref{lem2}(ii)) between the distributional boundary values $a_{-k}$ and $b_{-k}$ is reflected into a similar relationship between the harmonic potentials $A_{-k}$ and $B_{-k}$, as was already shown for $k=1$ and $k=2$. Indeed, we have, the Hilbert transform being taken in the variable $\underline{x} \in \mathbb{R}^m$:
$\mathcal{H} \left [ A_{-k} \right ] = B_{-k}$ and $\mathcal{H} \left [ B_{-k} \right ] = A_{-k}$, which may also be written as
\begin{equation}
\label{ter}
\left \{ \begin{array}{rcl}
b_{-1} (\cdot) \ast A_{-k}(x_0,\cdot)(\underline{x}) & = & B_{-k}(x_0,\underline{x}) \\[2mm]
b_{-1} (\cdot) \ast B_{-k}(x_0,\cdot)(\underline{x}) & = & A_{-k}(x_0,\underline{x})
\end{array} \right .
\end{equation}
while, trivially,
\begin{equation}
\label{quater}
\left \{ \begin{array}{rcl}
a_{-1} (\cdot) \ast A_{-k}(x_0,\cdot)(\underline{x}) & = & A_{-k}(x_0,\underline{x}) \\[2mm]
a_{-1} (\cdot) \ast B_{-k}(x_0,\cdot)(\underline{x}) & = & B_{-k}(x_0,\underline{x})
\end{array} \right .
\end{equation}
Note also the following commutative scheme:
\begin{equation}
\begin{array}{ccccc}
A_{-k} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & B_{-k-1} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & A_{-k-2} \\[2mm]
\hspace*{-8mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow & & \downarrow \\
a_{-k} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & b_{-k-1} & \xrightarrow{\hspace*{2mm} -\underline{\partial} \hspace*{2mm}} & a_{-k-2}
\end{array}
\label{cscheme2}
\end{equation}
The formulae (\ref{ter}) and (\ref{quater}) are special cases of the more general, and remarkable, result that the distributional boundary values $a_{-k}$ and $b_{-k}$ may act as convolution operators to convert the harmonic potentials into harmonic potentials of a lower order.
\begin{proposition}
One has for $k=1,2,\ldots$ and $j=0,1,\ldots$
\begin{itemize}
\item[(i)] $b_{-j-1} \ast A_{-k} = B_{-k-j}$
\item[(ii)] $b_{-j-1} \ast B_{-k} = A_{-k-j}$
\item[(iii)] $a_{-j-1} \ast A_{-k} = A_{-k-j}$
\item[(iv)] $a_{-j-1} \ast B_{-k} = B_{-k-j}$
\end{itemize}
\end{proposition}
\par\noindent {\bf Proof}~\par\noindent
\noindent (i) First assume that $k$ is even, say $k=2\ell$, and put $b_{-j-1} \ast A_{-2\ell} = B'_{-2\ell-j}$. Then we have
$$
\partial_{x_0} B'_{-2\ell-j} = b_{-j-1} \ast \partial_{x_0} A_{-2\ell} = b_{-j-1} \ast \left ( -\underline{\partial} B_{-2\ell} \right ) = - \left ( b_{-j-1} \underline{\partial} \right ) \ast B_{-2 \ell} = a_{-j-2} \ast B_{-2\ell}
$$
Assuming $j$ to be even, say $j=2i$, there holds
$$
a_{-2i-2} \ast B_{-2\ell} = (-\underline{\partial})^{2i+1} H \ast B_{-2\ell} = (-\underline{\partial})^{2i+1} A_{-2\ell} = B_{-2\ell-2i-1} = B_{-2\ell-j-1}
$$
while for $j$ odd, say $j=2i-1$, we have
$$
a_{-2i-2} \ast B_{-2\ell} = \underline{\partial}^{2i} \delta \ast B_{-2\ell} = (-\underline{\partial})^{2i} B_{-2\ell} = B_{-2\ell-2i} = B_{-2\ell-j-1}
$$
and so $\partial_{x_0} B'_{-2\ell-j} = B_{-2\ell - j-1}$. For the action of the Dirac operator $\underline{\partial}$ on $B'_{-2\ell-j}$ we obtain
$$
\underline{\partial} B'_{-2\ell-j} = \underline{\partial} b_{-j-1} \ast A_{-2\ell} = - a_{-j-2} \ast A_{-2\ell}
$$
where now for $j$ even, say $j=2i$,
$$
-a_{-2i-2} \ast A_{-2\ell} = - (-\underline{\partial})^{2i+1} H \ast A_{-2 \ell} = - (-\underline{\partial})^{2i+1} B_{-2 \ell} = - A_{-2 \ell -2i -1} = - A_{-2\ell -j -1}
$$
while for $j$ odd, say $j=2i-1$,
$$
-a_{-2i-1} \ast A_{-2\ell} = - \underline{\partial}^{2i} \delta \ast A_{-2\ell} = - (-\underline{\partial})^{2i} A_{-2\ell} = A_{-2\ell-2i} = - A_{-2\ell-j-1}
$$
and hence $\underline{\partial} B'_{-2\ell-j} = -A_{-2\ell-j-1}$. Moreover
$$
\lim_{x_0 \rightarrow 0+} B'_{-2\ell-j} = b_{-j-1} \ast a_{-2\ell} = b_{-j-1} \ast (- \underline{\partial})^{2\ell-1} H = b_{-j-1} (- \underline{\partial})^{2\ell-1} \ast H = a_{-2\ell - j} \ast H = b_{-2\ell-j}
$$
This means that $B'_{-2\ell-j}$ and $B_{-2\ell-j}$ satisfy the same system (\ref{AB-k-1}) and show the same distributional boundary value for $x_0 \rightarrow 0+$, so they have to coincide, whence $b_{-j-1} \ast A_{-2\ell} = B_{-2\ell-j}$.\\
Next assume that $k$ is odd, say $k=2\ell-1$, and put $b_{-j-1} \ast A_{-2\ell+1} = B''_{-2\ell-j+1}$. Then we have, with a similar calculation as above: $\partial_{x_0} B''_{-2\ell-j+1} = B_{-2\ell -j}$ and $\underline{\partial} B''_{-2\ell-j+1} = - A_{-2\ell-j}$, while $\lim_{x_0 \rightarrow 0+} B''_{-2\ell-j+1} = b_{-2\ell-j+1}$, from which it follows that indeed $b_{-j-1} \ast A_{-2\ell+1} = B_{-2\ell - j+1}$, which completes the proof of (i). \\[-2mm]
\noindent (ii) The proof of (ii) is similar to that of (i).\\[-2mm]
\noindent (iii) Using the Hilbert--link between the distributional boundary values $a_{-j-1}$ and $b_{-j-1}$, we have indeed
$$
a_{-j-1} \ast A_{-k} = \mathcal{H} \left [ b_{-j-1} \right ] \ast A_{-k} = H \ast b_{-j-1} \ast A_{-k} = b_{-j-1} \ast H \ast A_{-k} = b_{-j-1} \ast B_{-k} = A_{-k-j}
$$
Note however that an alternative proof of (iii) is already contained in the calculations in the proofs of (i) and (ii).\\[-2mm]
\noindent (iv) The proof of (iv) is similar to that of (iii), now making use of (ii).
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
For a function $f \in L_2(\mathbb{R}^m)$ we can define in $\mathbb{R}^{m+1}_+$ the conjugate harmonic functions
$$
\mathcal{A}_{-k}[f] = A_{-k}(x_0,\cdot) \ast f(\cdot) (\underline{x}) \qquad \mbox{and} \qquad \mathcal{B}_{-k}[f] = B_{-k}(x_0,\underline{x}) \ast f(\cdot) (\underline{x}), \quad k=1,2,\ldots
$$
and the monogenic function
$$
\mathcal{C}_{-k}[f] = C_{-k}(x_0,\cdot) \ast f(\cdot) (\underline{x}), \qquad k=1,2,\ldots
$$
By definition of the potential kernels $A_{-k}$, $B_{-k}$ and $C_{-k}$, it is readily obtained that $\mathcal{A}_{-k}[f]$ and $\mathcal{B}_{-k}[f]$ are conjugate harmonic potentials of $\mathcal{C}_{-k-1}[f]$, while $\mathcal{C}_{-k}[f]$ is a monogenic potential (or primitive) of $\mathcal{C}_{-k-1}[f]$ in $\mathbb{R}^{m+1}_+$. These potentials will show non--tangential $L_2$--boundary values for $x_0 \rightarrow 0+$ on condition that $f$ belongs to the Clifford--Sobolev space
$$
W_2^{k-1}(\mathbb{R}^m) = \left \{ f \in L_2(\mathbb{R}^m) \, : \, \underline{\partial} f, \; \underline{\partial}^2 f, \; \ldots, \underline{\partial}^{k-1} f \in L_2(\mathbb{R}^m) \right \}
$$
Under this assumption we have
\begin{eqnarray*}
\mathcal{A}_{-2\ell}^+ [f] = \lim_{x_0 \rightarrow 0+} \mathcal{A}_{-2\ell} [f] & = & a_{-2\ell} \ast f = - \underline{\partial}^{2\ell-1} \mathcal{H}[f] = - \mathcal{H} \left [ \underline{\partial}^{2\ell-1} f \right ] \\[2mm]
\mathcal{B}_{-2\ell} ^+ [f] = \lim_{x_0 \rightarrow 0+} \mathcal{B}_{-2\ell} [f] & = & b_{-2\ell} \ast f = - \underline{\partial}^{2\ell-1} f \\[2mm]
\mathcal{A}_{-2\ell-1}^+ [f] = \lim_{x_0 \rightarrow 0+} \mathcal{A}_{-2\ell-1} [f] & = & a_{-2\ell-1} \ast f = \underline{\partial}^{2\ell} f \\[2mm]
\mathcal{B}_{-2\ell-1}^+ [f] = \lim_{x_0 \rightarrow 0+} \mathcal{B}_{-2\ell-1} [f] & = & b_{-2\ell-1} \ast f = \underline{\partial}^{2\ell} \mathcal{H}[f] = \mathcal{H} \left [ \underline{\partial}^{2\ell} f \right ]
\end{eqnarray*}
and the convolution operators $\mathcal{A}_{-k}^+$ and $\mathcal{B}_{-k}^+$ are bounded operators from $W_2^{k-1}(\mathbb{R}^m)$ into $L_2(\mathbb{R}^m)$. For a function $f \in W_2^{k-1}(\mathbb{R}^m)$ we also obtain the following expressions of the non--tangential $L_2$--boundary values of the monogenic potentials:
\begin{eqnarray*}
\mathcal{C}_{-2\ell}^+ [f] = \lim_{x_0 \rightarrow 0+} \mathcal{C}_{-2\ell} [f] & = & - \frac{1}{2} \underline{\partial}^{2\ell-1} \mathcal{H}[f] - \frac{1}{2} \overline{e_0} \underline{\partial}^{2\ell-1} f \\
& = & (- \overline{e_0} \underline{\partial}^{2\ell-1}) \left ( \frac{1}{2} f + \frac{1}{2} \overline{e_0} \mathcal{H}[f] \right ) \ = \ (- \overline{e_0} \underline{\partial})^{2\ell-1} \left ( \mathcal{A} \mathcal{S} [f] \right ) \\[4mm]
\mathcal{C}_{-2\ell-1}^+ [f] = \lim_{x_0 \rightarrow 0+} \mathcal{C}_{-2\ell-1} [f] & = & \frac{1}{2} \underline{\partial}^{2\ell} f + \frac{1}{2} \overline{e_0} \underline{\partial}^{2\ell} \mathcal{H} [f] \\
& = & \underline{\partial}^{2\ell} \left ( \frac{1}{2} f + \frac{1}{2} \overline{e_0} \mathcal{H}[f] \right ) \ = \ \underline{\partial}^{2\ell} \left ( \mathcal{A} \mathcal{S} [f] \right ) = (- \overline{e_0} \underline{\partial})^{2\ell} \left ( \mathcal{A} \mathcal{S} [f] \right )
\end{eqnarray*}
which belong to the Clifford--Hardy space $H^2(\mathbb{R}^m)$. This
leads to the following commutative schemes for a function $f \in W_2^k(\mathbb{R}^m)$:
$$
\begin{array}{ccccc}
\mathcal{C}_{-2\ell-1}[f] & \xrightarrow{\hspace*{2mm} - \overline{e_0} \underline{\partial} \hspace*{2mm}} & \mathcal{C}_{-2\ell-2}[f] & \xrightarrow{\hspace*{2mm} - \overline{e_0} \underline{\partial} \hspace*{2mm}} & \mathcal{C}_{-2\ell-3}[f] \\[2mm]
\hspace*{-8mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow & & \downarrow \\
(- \overline{e_0} \underline{\partial})^{2\ell} \left ( \mathcal{A} \mathcal{S} [f] \right ) & \xrightarrow{\hspace*{2mm} - \overline{e_0} \underline{\partial} \hspace*{2mm}} & (- \overline{e_0} \underline{\partial})^{2\ell+1} \left ( \mathcal{A} \mathcal{S} [f] \right ) & \xrightarrow{\hspace*{2mm} - \overline{e_0} \underline{\partial} \hspace*{2mm}} & (- \overline{e_0} \underline{\partial})^{2\ell+2} \left ( \mathcal{A} \mathcal{S} [f] \right )
\end{array}
$$
The above scheme reflects at the level of the convolution operators, the commutative schemes (\ref{cscheme2}) at the level of the convolution kernels.
\subsection{Explicit expression of the downstream potentials}
In Subsection 4.1 we have already obtained the explicit expressions of the harmonic potentials $A_{-1}, B_{-1}, A_{-2}$ and $B_{-2}$ (see \ref{A-1B-1}, \ref{A-2B-2}). Putting forward for the harmonic potentials $A_{-k}$ and $B_{-k}$ the following form:
$$
A_{-k} = \frac{2}{\sigma_{m+1}} \ \frac{1}{|x|^{m+2k-1}} \ P_k(x_0,|\underline{x}|^2) \quad {\rm and} \quad
B_{-k} = \frac{2}{\sigma_{m+1}} \ \frac{\underline{x}}{|x|^{m+2k-1}} \ Q_{k-1}(x_0,|\underline{x}|^2)
$$
where $P_k$ and $Q_{k-1}$ are scalar--valued homogeneous polynomials of degree $k$ and $k-1$ respectively, it is shown, by a direct calculation, that these polynomials satisfy the following recurrence relations:
$$
P_{k+1}(t,u^{2}) = (t^2+u^{2}) \, \partial_t P_k - (m+2k-1) \, t \, P_k \quad , \quad P_1(t,u^{2}) = t
$$
and
$$
Q_{k}(t,u^{2}) = (t^2+u^{2}) \, \partial_t Q_{k-1} - (m+2k-1) \, t \, Q_{k-1} \quad , \quad Q_0(t,u^{2}) = -1
$$
The fact that $A_{-k}$ and $B_{-k}$ are related by harmonic conjugacy leads to the following intertwined relations for those polynomials:
$$
Q_{k}(x_0,|\underline{x}|^2) = (m+2k-1) \, P_k + \frac{x_0^2 + |\underline{x}|^2}{|\underline{x}|^2} \, \underline{x} \underline{\partial} P_k
$$
and
$$
P_{k+1}(x_0,|\underline{x}|^2) = \left(m x_0^2 - (2k-1)|\underline{x}|^2\right) Q_{k-1} - (x_0^2 + |\underline{x}|^2) \, \underline{\partial} Q_{k-1} \, \underline{x}
$$
It is possible to obtain an explicit expression for $P_{k}(t,u^{2})$ and $Q_{k}(t,u^{2})$ in terms of well-known orthogonal polynomials. This is achieved in the following way. First rewrite these polynomials as
\begin{align*}
P_{k}(t,u^{2}) &= u^{k} \widetilde{P}_{k}\left( \frac{t}{u} \right)\\
Q_{k}(t,u^{2}) &= u^{k} \widetilde{Q}_{k}\left( \frac{t}{u} \right)
\end{align*}
with $\widetilde{P}_{k}(w)$ and $\widetilde{Q}_{k}(w)$ polynomials of degree $k$ in the variable $w=t/u$. The recursion relations for $P_{k}$ and $Q_{k}$ may now be rewritten as
\begin{align*}
\widetilde{P}_{k}(w) & = (1+w^{2}) \partial_{w} \widetilde{P}_{k-1}(w) - (m+2k-3) w \widetilde{P}_{k-1}(w), \qquad \widetilde{P}_{0}(w) = -1/(m-1)\\
\widetilde{Q}_{k}(w) & = (1+w^{2}) \partial_{w} \widetilde{Q}_{k-1}(w) - (m+2k-1) w \widetilde{Q}_{k-1}(w), \qquad \widetilde{Q}_{0}(w) =1
\end{align*}
Using the operator identity
\[
(1+ w^{2}) \partial_{w} +2(\alpha+1)w = (1+w^{2})^{-\alpha} \partial_{w} (1+w^{2})^{\alpha+1}
\]
we subsequently find
\begin{align*}
\widetilde{P}_{k}(w) & = - \frac{1}{m-1}(1+w^{2})^{k + \frac{m-1}{2}} (\partial_{w})^{k} (1+w^{2})^{- \frac{m-1}{2}}\\[1mm]
\widetilde{Q}_{k}(w) & = - (1+w^{2})^{k + \frac{m+1}{2}} (\partial_{w})^{k} (1+w^{2})^{- \frac{m+1}{2}}
\end{align*}
Comparing this result with the Rodrigues' formula for the Gegenbauer polynomials, we obtain:
$$
\widetilde{P}_{k}(w) = (-1)^{k+1} 2^k i^k k! \frac{1}{m-1} \frac{\Gamma(-m-2k+2)}{\Gamma(-m-k+2)} \frac{\Gamma(-\frac{m}{2} + \frac{3}{2})}{\Gamma(-\frac{m}{2} + \frac{3}{2}-k)} C^{-k+1-\frac{m}{2}}_{k}(i w)
$$
and
$$
\widetilde{Q}_{k}(w) = (-1)^{k+1} 2^k i^k k! \frac{\Gamma(-m-2k)}{\Gamma(-m-k)} \frac{\Gamma(-\frac{m}{2} + \frac{1}{2})}{\Gamma(-\frac{m}{2} + \frac{1}{2}-k)} C^{-k-\frac{m}{2}}_{k}(i w)
$$
where $i$ is the imaginary unit, eventually leading to
$$
A_{-k} = (-1)^{k+1} \frac{2^{k+1}}{\sigma_{m+1}} \ k! \ \frac{1}{m-1} \frac{\Gamma(-m-2k+2)}{\Gamma(-m-k+2)} \frac{\Gamma(-\frac{m}{2} + \frac{3}{2})}{\Gamma(-\frac{m}{2} + \frac{3}{2}-k)} \ \frac{|\underline{x}|^k}{|x|^{m+2k-1}} \ i^k C^{-k+1-\frac{m}{2}}_{k}\left( i \frac{ x_0}{|\underline{x}|} \right)
$$
$$
= - \frac{2^{2k+1}}{\sigma_{m+1}} \ \frac{1}{m-1} \frac{\Gamma(-m-2k+2)}{\Gamma(-m-k+2)} \frac{\Gamma(-\frac{m}{2} + \frac{3}{2})}{\Gamma(-\frac{m}{2} + \frac{3}{2}-k)} \frac{\Gamma(-\frac{m}{2}+1)}{\Gamma(-\frac{m}{2}-k+1)} \ \frac{x_0^k}{|x|^{m+2k-1}} \ _2 F_1\left(-\frac{k}{2}, \frac{1-k}{2}; \frac{m}{2}; - \frac{|\underline{x}|^2}{x_0^2}\right)
$$
and
$$
B_{-k} = (-1)^{k} \frac{2^{k}}{\sigma_{m+1}} \ (k-1)! \ \frac{\Gamma(-m-2k+2)}{\Gamma(-m-k+1)} \frac{\Gamma(-\frac{m}{2} + \frac{1}{2})}{\Gamma(-\frac{m}{2} + \frac{3}{2}-k)} \ \frac{|\underline{x}|^{k-1} \underline{x}}{|x|^{m+2k-1}} \ i^{k-1} C^{-k+1-\frac{m}{2}}_{k-1}\left( i \frac{ x_0}{|\underline{x}|} \right)
$$
$$
= - \frac{2^{2k+1}}{\sigma_{m+1}} \ \frac{\Gamma(-m-2k+2)}{\Gamma(-m-k+1)} \frac{\Gamma(-\frac{m}{2} + \frac{1}{2})}{\Gamma(-\frac{m}{2} + \frac{3}{2}-k)} \frac{\Gamma(-\frac{m}{2}+1)}{\Gamma(-\frac{m}{2}-k+2)} \ \frac{x_0^{k-1} \underline{x}}{|x|^{m+2k-1}} \ _2 F_1\left(-\frac{k+1}{2}, 1-\frac{k}{2}; \frac{m}{2}; - \frac{|\underline{x}|^2}{x_0^2}\right)
$$
\\[2mm]
Finally, to give an idea, let us state the explicit expressions of the potentials $A_{-k}$ and $B_{-k}$ for a couple of low values of $k$:
\begin{eqnarray*}
A_{-3} (x_0,|\underline{x}|) & = & \frac{2}{\sigma_{m+1}} \ \frac{1}{|x|^{m+5}} \left( m(m+1) x_0^3 - 3(m+1) x_0 |\underline{x}|^2 \right)\\
A_{-4} (x_0,|\underline{x}|) & = & \frac{2}{\sigma_{m+1}} \ \frac{1}{|x|^{m+7}} \left( - m(m+1)(m+2) x_0^4 + 6(m+1)(m+2) x_0^2 |\underline{x}|^2 - 3(m+1) |\underline{x}|^4 \right)
\end{eqnarray*}
and
\begin{eqnarray*}
B_{-3} (x_0,|\underline{x}|) & = & \frac{2}{\sigma_{m+1}} \ \frac{\underline{x}}{|x|^{m+5}} \ (m+1) \left( - (m+2) x_0^2 + |\underline{x}|^2 \right)\\
B_{-4} (x_0,|\underline{x}|) & = & \frac{2}{\sigma_{m+1}} \ \frac{\underline{x}}{|x|^{m+7}} \ (m+1)(m+3) \left( (m+2) x_0^3 - 3 x_0 |\underline{x}|^2 \right)
\end{eqnarray*}
\section{Upstream potentials}
\label{upstream}
\subsection{The monogenic logarithmic function}
Recall that Green's function $A_0(x_0,\underline{x}) = - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \frac{1}{|x|^{m-1}}$, (\ref{A0}), and its conjugate harmonic $B_0(x_0,\underline{x}) = \frac{2}{\sigma_{m+1}} \frac{\underline{x}}{|\underline{x}|^m} F_m ( \frac{|\underline{x}|}{x_0} )$, (\ref{B0}), satisfy in $\mathbb{R}^{m+1}_+$ the system
$$
\left \{ \begin{array}{rcl}
\partial_{x_0} A_0 & = & - \underline{\partial} B_0 \ = \ P \ = \ A_{-1} \\[2mm]
\partial_{x_0} B_0 & = & - \underline{\partial} A_0 \ = \ Q \ = \ B_{-1} \\[2mm]
\end{array} \right .
$$
(see also (\ref{CR12}) in Section 3) from which it follows that
\begin{equation}
\overline{D} A_0 = \frac{1}{2} \left ( \partial_{x_0} - \overline{e_0} \underline{\partial} \right ) A_0 \ = \ \frac{1}{2} P + \frac{1}{2} \overline{e_0} Q = C_{-1}
\label{rel1}
\end{equation}
and
\begin{equation}
\overline{D} \overline{e_0} B_0 = \frac{1}{2} \left ( \overline{e_0} \partial_{x_0} - \underline{\partial} \right ) B_0 \ = \ \frac{1}{2} \overline{e_0} Q + \frac{1}{2} P = C_{-1}
\label{rel2}
\end{equation}
Relations (\ref{rel1}) and (\ref{rel2}) express the fact that $A_0(x_0,\underline{x})$ and $\overline{e_0} B_0(x_0,\underline{x})$ are conjugate harmonic potentials (or primitives), with respect to the operator $\overline{D}$, of the Cauchy kernel $C_{-1}(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$. Putting, as in Section 3, $C_0(x_0,\underline{x}) = \frac{1}{2} A_0(x_0,\underline{x}) + \frac{1}{2} \overline{e_0} B_0 (x_0,\underline{x})$, it is readily seen that
$$
\overline{D} C_0 (x_0,\underline{x}) = \frac{1}{2} C_{-1} (x_0,\underline{x}) + \frac{1}{2} C_{-1} (x_0,\underline{x}) = C_{-1}(x_0,\underline{x})
$$
which implies that $C_0(x_0,\underline{x})$ is a monogenic potential (or primitive), with respect to te operator $\overline{D}$, of the Cauchy kernel $C_{-1}(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$. Moreover there holds, in view of $DC_0 = \frac{1}{2} (\partial_{x_0} +\overline{e_0} \underline{\partial}) C_0 = 0$, that
$$
C_{-1} = \overline{D} C_0 = \partial_{x_0} C_0 = \left ( - \overline{e_0} \underline{\partial} \right ) C_0
$$
Recall the distributional limits for $x_0 \rightarrow 0+$ of $A_0(x_0,\underline{x})$ and $B_0(x_0,\underline{x})$:
\begin{eqnarray*}
a_0(\underline{x}) & = & - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \mbox{Fp} \frac{1}{|\underline{x}|^{m-1}} \ = \ - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} T^\ast_{-m+1} \\
b_0(\underline{x}) & = & \frac{1}{\sigma_m} \frac{\underline{x}}{|\underline{x}|^m} \ = \ \frac{1}{\pi} \frac{1}{\sigma_m} U^\ast_{-m+1}
\end{eqnarray*}
(see also Section 3, (\ref{a0}) and (\ref{b0})), yielding
$$
c_0(\underline{x}) = \lim_{x_0 \rightarrow 0+} C_0(x_0,\underline{x}) = - \frac{1}{m-1} \frac{1}{\sigma_{m+1}} \mbox{Fp} \frac{1}{|\underline{x}|^{m-1}} + \frac{1}{2} \overline{e_0} \frac{1}{\sigma_m} \frac{\underline{x}}{|\underline{x}|^m}
$$
As was expected these distributional boundary values are intimately related, as is shown in the following lemma.
\begin{lemma}
\label{lemintiem}
One has
\begin{itemize}
\item[(i)] $-\underline{\partial} a_0 = b_{-1} = H$
\item[(ii)] $-\underline{\partial} b_0 = a_{-1} = \delta$
\item[(iii)] $\mathcal{H} \left [a_0 \right ] = b_0$
\item[(iv)] $\mathcal{H} \left [b_0 \right ] = a_0$
\item[(v)] $c_{-1} \ast a_0 = c_{-1} \ast \overline{e_0} b_0 = c_{-1} \ast c_0 = c_0$
\item[(vi)] $-\overline{e_0} \underline{\partial} c_0 = c_{-1}$
\item[(vii)] $ \overline{e_0} \mathcal{H} \left [c_0 \right ] = c_0$
\end{itemize}
\end{lemma}
\par\noindent {\bf Proof}~\par\noindent
We make use of the calculation rules for the $T^\ast$-- and $U^\ast$--distributions, recalled in Proposition \ref{prop1}. For (i) we have
$$
- \underline{\partial} a_0 = \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \underline{\partial} T^\ast_{-m+1} = \frac{2}{m-1} \frac{1}{\sigma_{m+1}} (-m+1) U^\ast_{-m} = -\frac{2}{\sigma_{m+1}} U^\ast_{-m} = b_{-1} = H
$$
while for (ii)
$$
- \underline{\partial} b_0 = -\frac{1}{\pi} \frac{1}{\sigma_m} \underline{\partial} U^\ast_{-m+1} = -\frac{1}{\pi} \frac{1}{\sigma_m} (-2\pi) T^\ast_{-m} = -\frac{2}{\sigma_m} T^\ast_{-m} = a_{-1} = \delta
$$
Then (iii) is obtained by
\begin{eqnarray*}
\mathcal{H} \left [ a_0 \right ] &=& H \ast a_0 = \left ( - \frac{2}{\sigma_{m+1}} U^\ast_{-m} \right ) \ast \left ( - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} T^\ast_{-m+1} \right ) \\
& = & \frac{4}{m-1} \frac{1}{\left ( \sigma_{m+1} \right )^2} U^\ast_{-m} \ast T^\ast_{-m+1} \ = \ \frac{4}{m-1} \frac{1}{
\left ( \sigma_{m+1} \right )^2} \pi^\frac{m}{2} \frac{\Gamma \left ( \frac{m}{2} \right )}{\Gamma \left ( \frac{m+1}{2} \right ) \Gamma \left ( \frac{m-1}{2}\right )} U^\ast_{-m+1} \\
& = & \frac{\Gamma \left ( \frac{m}{2} \right ) }{2 \pi^{\frac{m}{2}+1} } U^\ast_{-m+1} \ = \ \frac{1}{\pi} \frac{1}{\sigma_m} U^\ast_{-m+1} \ = \ b_0
\end{eqnarray*}
from which also (iv) follows: $\mathcal{H} \left [ b_0 \right ] = \mathcal{H}^2 \left [ a_0 \right ] = a_0$.
To obtain (v) it suffices to observe that
\begin{eqnarray*}
c_{-1} \ast a_0 & = & \left ( \frac{1}{2} a_{-1} + \frac{1}{2} \overline{e_0} b_{-1} \right ) \ast a_0 \ = \ \left ( \frac{1}{2} \delta + \frac{1}{2} \overline{e_0} H \right ) \ast a_0 \ = \ \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \ = \ c_0 \\
c_{-1} \ast \overline{e_0} b_0 & = & \left ( \frac{1}{2} \delta + \frac{1}{2} \overline{e_0} H \right ) \ast \overline{e_0} b_0 \ = \ \frac{1}{2} \overline{e_0} b_0 + \frac{1}{2} \mathcal{H} \left [ b_0 \right ] \ = \ \frac{1}{2} \overline{e_0} b_0 + \frac{1}{2} a_0 \ = \ c_0 \\
c_{-1} \ast c_0 & = & c_{-1} \ast \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) \ = \ \frac{1}{2} c_0 + \frac{1}{2} c_0 \ = \ c_0
\end{eqnarray*}
while (vi) is directly obtained by
$$
- \overline{e_0} \underline{\partial} c_0 = \overline{e_0} \left ( -\underline{\partial} \frac{1}{2} a_0 - \underline{\partial} \frac{1}{2} \overline{e_0} b_0 \right ) \ = \ \frac{1}{2} \overline{e_0} b_{-1} + \frac{1}{2} a_{-1} = c_{-1}
$$
Finally, we have
$$
\mathcal{H} \left [ c_0 \right ] = \frac{1}{2} \mathcal{H} \left [ a_0 \right ] + \frac{1}{2} \mathcal{H} \left [ \overline{e_0} b_0 \right ] = \frac{1}{2} b_0 + \frac{1}{2} e_0 a_0
$$
from which (vii) follows:
$$
\overline{e_0} \mathcal{H} \left [ c_0 \right ] = \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 = c_0
$$
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
Note the following commutative schemes which are each others Hilbert image:
\begin{equation}
\begin{array}{ccc}
A_0(x_0,\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & B_{-1}(x_0,\underline{x}) = Q(x_0,\underline{x}) \\[2mm]
\hspace*{-8mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
a_{0}(\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & b_{-1}(\underline{x}) = H(\underline{x})
\end{array}
\quad \mbox{and} \qquad
\begin{array}{ccc}
B_{0}(x_0,\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & A_{-1}(x_0,\underline{x}) = P(x_0,\underline{x}) \\[2mm]
\hspace*{-7mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
b_{0}(\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & a_{-1}(\underline{x}) = \delta(\underline{x})
\end{array}
\label{cschemes4}
\end{equation}
The following lemma, also in terms of distributions, makes the relationships between the harmonic potentials $A_0$, $B_0$, $A_{-1} = P$ and $B_{-1}=Q$ more transparent.
\begin{lemma}
\label{lem52}
In distributional sense one has, convolutions being taken in the variable $\underline{x} \in \mathbb{R}^m$:
\begin{itemize}
\item[(i)] $A_0 = a_0 \ast A_{-1} = b_0 \ast B_{-1}$
\item[(ii)] $B_0 = b_0 \ast A_{-1} = a_0 \ast B_{-1}$
\item[(iii)] $H \ast A_0 = B_0 = A_0 \ast H$
\item[(iv)] $H \ast B_0 = A_0 = B_0 \ast H$
\end{itemize}
\end{lemma}
\par\noindent {\bf Proof}~\par\noindent
\noindent (i)(ii) The technique is the same as the one used in the proof of Proposition \ref{propdownstream}.\\[-2mm]
\noindent (iii)(iv) It suffices to observe that
$$
A_0 \ast H = A_{-1} \ast a_0 \ast H = A_{-1} \ast b_0 = B_0
$$
and
$$
B_0 \ast H = A_0 \ast H \ast H = A_0
$$
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
Similar properties hold for the monogenic potentials $C_0$ and $C_{-1}$.
\begin{lemma}
\label{lem53}
In distributional sense one has, convolutions being taken in the variable $\underline{x} \in \mathbb{R}^m$:
\begin{itemize}
\item[(i)] $C_0 = C_{-1} \ast a_0 = C_{-1} \ast \overline{e_0} b_0 = C_{-1} \ast c_0$
\item[(ii)] $C_0 = A_{-1} \ast c_0 = \overline{e_0} B_{-1} \ast c_0$
\end{itemize}
\end{lemma}
\par\noindent {\bf Proof}~\par\noindent
Making use of the results of Lemma \ref{lem52}, we have for (i)
\begin{eqnarray*}
C_{-1} \ast a_0 & = & \left ( \frac{1}{2} A_{-1} + \frac{1}{2} \overline{e_0} B_{-1} \right ) \ast a_0 \ = \ \frac{1}{2} A_0 + \frac{1}{2} \overline{e_0} B_0 \ = \ C_0 \\
C_{-1} \ast \overline{e_0} b_0 & = & \left ( \frac{1}{2} A_{-1} + \frac{1}{2} \overline{e_0} B_{-1} \right ) \ast \overline{e_0} b_0 \ = \ \frac{1}{2} \overline{e_0} B_0 + \frac{1}{2} A_0 \ = \ C_0 \\
C_{-1} \ast c_0 & = & C_{-1} \ast \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) = \frac{1}{2} C_0 + \frac{1}{2} C_0 \ = \ C_0
\end{eqnarray*}
while for (ii)
\begin{eqnarray*}
A_{-1} \ast c_0 & = & A_{-1} \ast \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) \ = \ \frac{1}{2} A_0 + \frac{1}{2} \overline{e_0} B_0 \ = \ C_0 \\
\overline{e_0} B_{-1} \ast c_0 & = & \overline{e_0} B_{-1} \ast \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) \ = \ \frac{1}{2} \overline{e_0} B_0 + \frac{1}{2} B_{-1} \ast b_0 \ = \ C_0
\end{eqnarray*}
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
\begin{remark}
The results (i)--(ii) of Lemma \ref{lem52} are the analogues of the results of Proposition \ref{propdownstream} for the case where $j=-1$, $k=1$.
\end{remark}
The potential kernels $A_0(x_0,\underline{x})$, $B_0(x_0,\underline{x})$ and $C_0(x_0,\underline{x})$ may now be used in their corresponding convolution operators defining the conjugate harmonic potentials in $\mathbb{R}^{m+1}_+$
$$
\mathcal{A}_0[f](x_0,\underline{x}) = A_0(x_0,\cdot) \ast f(\cdot)(\underline{x}) \quad \mbox{and} \quad \mathcal{B}_0[f](x_0,\underline{x}) = B_0(x_0,\cdot) \ast f(\cdot)(\underline{x})
$$
and the monogenic potential
$$
\mathcal{C}_0[f](x_0,\underline{x}) = C_0(x_0,\cdot) \ast f(\cdot) (\underline{x}) = \frac{1}{2} \mathcal{A}_0[f](x_0,\underline{x}) + \frac{1}{2} \overline{e_0} \mathcal{B}_0[f](x_0,\underline{x})
$$
The properties they enjoy, summarized in the next proposition, reflect the corresponding properties of the potential kernels.
\begin{proposition}
\label{prop51}
For a Schwartz function or a distribution $f$ one has
\begin{itemize}
\item[(i)] $\overline{D} \mathcal{A}_0[f] = \overline{D} \mathcal{B}_0[f] = \overline{D} \mathcal{C}_0[f] = C_{-1}[f]$
\item[(ii)] $\lim_{x_0 \rightarrow 0+} \mathcal{A}_0[f] = a_0 \ast f = - (- \Delta)^{-\frac{1}{2}} [f]$\\
$\lim_{x_0 \rightarrow 0+} \mathcal{B}_0[f] = b_0 \ast f = - E \ast f = T[f]$ \\
$\lim_{x_0 \rightarrow 0+} \mathcal{C}_0[f] = c_0 \ast f = a_0 \ast \mathcal{A} \mathcal{S}[f] = \overline{e_0} b_0 \ast f$
\item[(iii)] $\overline{e_0} \mathcal{H} \left [c_0 \ast f \right ] = c_0 \ast f$
\item[(iv)] $\mathcal{C}_0[f] = \mathcal{C}_{-1} \left [a_0 \ast f \right ] = \mathcal{C}_{-1} \left [ b_0 \ast \mathcal{H}[f] \right ] = \mathcal{C}_{-1} \left [\overline{e_0} b_0 \ast f \right ] = \mathcal{C}_{-1} \left [ a_0 \ast \overline{e_0} \mathcal{H}[f] \right ]$ \\
$\phantom{\mathcal{C}_0[f]} = \mathcal{C}_{-1} \left [ c_0 \ast f \right ] = \mathcal{P} \left [ c_0 \ast f \right ] = \mathcal{A}_0 \left [ \mathcal{A}\mathcal{S}[f] \right ] = \overline{e_0} \mathcal{B}_0 \left [ \mathcal{A} \mathcal{S}[f] \right ]$
\end{itemize}
\end{proposition}
Note also the following commutative scheme, which may be derived from the commutative schemes (\ref{cschemes4}):
$$
\begin{array}{rcl}
\mathcal{C}_{0}[f] & \xrightarrow{\hspace*{2mm} -\overline{e_0} \underline{\partial} \hspace*{2mm}} & \mathcal{C}_{-1}[f] \\[2mm]
^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow \hspace*{4mm} & & \hspace*{4mm} \downarrow ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}}\\
c_0 \ast f & \xrightarrow{\hspace*{2mm} -\overline{e_0} \underline{\partial} \hspace*{2mm}} & c_{-1} \ast f = \mathcal{A} \mathcal{S}[f]
\end{array}
$$
\begin{remark}
As already explained in the introduction, in the upper half of the complex plane the function $\ln(z)$ is a holomorphic potential (or primitive) of the Cauchy kernel $\frac{1}{z}$ and its real and imaginary components are the fundamental solution $\ln |z|$ of the Laplace operator, and its conjugate harmonic $i\, {\rm arg}(z)$ respectively. By similarity we could say that $C_0(x_0,\underline{x}) = \frac{1}{2} A_0(x_0,\underline{x}) + \frac{1}{2} \overline{e_0} B_0(x_0,\underline{x})$, being a monogenic potential of the Cauchy kernel $C_{-1}(x_0,\underline{x})$ and the sum of the fundamental solution $A_0(x_0,\underline{x})$ of the Laplace operator and its conjugate harmonic $\overline{e_0} B_0(x_0,\underline{x})$, is a {\em monogenic logarithmic function} in the upper half--space $\mathbb{R}^{m+1}_+$.
\end{remark}
\subsection{The potentials of the logarithmic monogenic function}
\noindent
Inspired by the properties contained in Proposition \ref{propdownstream} and Lemma \ref{lem52} we proceed as follows for the construction of harmonic and monogenic potentials of $C_0(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$. We put
$$
\left \{ \begin{array}{rclcl}
A_1(x_0,\underline{x}) & = & a_0(\cdot) \ast A_0(x_0,\cdot)(\underline{x}) & = & b_0(\cdot) \ast B_0(x_0,\cdot) \\[2mm]
B_1(x_0,\underline{x}) & = & a_0(\cdot) \ast B_0(x_0,\cdot)(\underline{x}) & = & b_0(\cdot) \ast A_0(x_0,\cdot)
\end{array} \right .
$$
and we verify at once that
\begin{equation}
\begin{array}{rcl}
\partial_{x_0} A_1 & = & a_0 \ast \partial_{x_0} A_0 \ = \ a_0 \ast A_{-1} \ = \ A_{-1} \ast a_0 \ = \ \mathcal{P}[a_0] \ = \ A_0\\[2mm]
\partial_{x_0} A_1 & = & b_0 \ast \partial_{x_0} B_0 \ = \ b_0 \ast (-\underline{\partial} A_{0}) \ = \ (-b_0 \underline{\partial}) \ast A_0 \ = \ \delta \ast A_0 \ = \ A_0
\end{array}
\label{verify1}
\end{equation}
and
\begin{equation}
\begin{array}{rcl}
\partial_{x_0} B_1 & = & a_0 \ast \partial_{x_0} B_0 \ = \ (-a_0 \underline{\partial}) \ast A_{0} \ = \ \mathcal{H}[A_0] \ = \ B_0\\[2mm]
\partial_{x_0} B_1 & = & b_0 \ast \partial_{x_0} A_0 \ = \ b_0 \ast (-\underline{\partial} B_{0}) \ = \ (-b_0 \underline{\partial}) \ast B_0 \ = \ \delta \ast B_0 \ = \ B_0
\end{array}
\label{verify2}
\end{equation}
while also
\begin{equation}
\begin{array}{rcl}
- \underline{\partial} A_1 & = & - \underline{\partial} a_0 \ast A_0 \ = \ H \ast A_0 \ = \ \mathcal{H}[A_0] \ = \ B_0 \\[2mm]
- \underline{\partial} A_1 & = & - \underline{\partial} b_0 \ast B_0 \ = \ \delta \ast B_0 \ = \ B_0
\end{array}
\label{verify3}
\end{equation}
and
\begin{equation}
\begin{array}{rcl}
- \underline{\partial} B_1 & = & - \underline{\partial} a_0 \ast B_0 \ = \ \mathcal{H}[B_0] \ = \ A_0 \\[2mm]
- \underline{\partial} B_1 & = & - \underline{\partial} b_0 \ast A_0 \ = \ \delta \ast A_0 \ = \ A_0
\end{array}
\label{verify4}
\end{equation}
The relations (\ref{verify1})--(\ref{verify4}) precisely are the relations needed for $A_1(x_0,\underline{x})$ and $B_1(x_0,\underline{x})$ to be conjugate harmonic potentials in $\mathbb{R}^{m+1}_+$ of the function $C_0(x_0,\underline{x})$. It then follows at once that
$$
C_1(x_0,\underline{x}) = \frac{1}{2} A_1(x_0,\underline{x}) + \frac{1}{2} \overline{e_0} B_1(x_0,\underline{x})
$$
is a monogenic potential in $\mathbb{R}^{m+1}_+$ of $C_0$ and there holds that $\overline{D} C_1 = \partial_{x_0} C_1 = (-\overline{e_0} \underline{\partial}) C_1 = C_0$. Also note that the conjugate harmonic potentials $A_1(x_0,\underline{x})$ and $B_1(x_0,\underline{x})$ form a Hilbert pair, the Hilbert transform being taken in the variable $\underline{x} \in \mathbb{R}^m$. Indeed, we have
\begin{eqnarray*}
\mathcal{H} \left [ A_1 \right ] &=& H(\cdot) \ast A_1(x_0,\cdot) (\underline{x}) \ = \ H(\cdot) \ast a_0(\cdot) \ast A_0(x_0,\cdot)(\underline{x}) \ = \ b_0(\cdot) \ast A_0(x_0,\cdot) \ = \ B_1(x_0,\underline{x}) \\
\mathcal{H} \left [ B_1 \right ] & = & \mathcal{H}^2 \left [ A_1 \right ] \ = \ A_1
\end{eqnarray*}
The distributional limits for $x_0 \rightarrow 0+$ of the conjugate harmonic potentials $A_1$ and $B_1$ are given by
$$
\left \{ \begin{array}{rcl}
a_1(\underline{x}) & = & \lim_{x_0 \rightarrow 0+} A_1(x_0,\underline{x}) \ = \ a_0(\cdot) \ast a_0(\cdot)(\underline{x}) \ = \ b_0(\cdot) \ast b_0(\cdot) (\underline{x}) \\[2mm]
b_1(\underline{x}) & = & \lim_{x_0 \rightarrow 0+} B_1(x_0,\underline{x}) \ = \ a_0(\cdot) \ast b_0(\cdot)(\underline{x}) \ = \ b_0(\cdot) \ast a_0(\cdot) (\underline{x})
\end{array} \right .
$$
Making use of the calculation rules for the convolution of the $T^\ast$-- and $U^\ast$--distributions (see Section 2, Proposition 2.1), these distributional boundary values are explicitly given by
\begin{equation}
\label{a1b1}
\left \{ \begin{array}{rcl}
a_1(\underline{x}) & = & \phantom{-} \displaystyle\frac{1}{\pi} \displaystyle\frac{1}{\sigma_m} \displaystyle\frac{1}{m-2} \, T^\ast_{-m+2} \ = \ \displaystyle\frac{1}{\sigma_m} \displaystyle\frac{1}{m-2} \displaystyle\frac{1}{|\underline{x}|^{m-2}}
\\[5mm]
b_1(\underline{x}) & = & - \displaystyle\frac{1}{\pi} \displaystyle\frac{1}{\sigma_{m+1}} \displaystyle\frac{1}{m-1} \, U^\ast_{-m+2} \ = \ - \displaystyle\frac{1}{\sigma_{m+1}} \displaystyle\frac{2}{m-1} \displaystyle\frac{\underline{x}}{|\underline{x}|^{m-1}}
\end{array} \right .
\end{equation}
They show the following properties, where we have put, quite naturally, $c_1(\underline{x}) = \frac{1}{2} a_1(\underline{x}) + \frac{1}{2} \overline{e_0} b_1(\underline{x})$.
\begin{lemma}
\label{lem54}
\rule{0mm}{0mm}
\begin{itemize}
\item[(i)] $- \underline{\partial} a_1 = b_0$, $-\underline{\partial} b_1 = a_0$, $- \overline{e_0} \underline{\partial} c_1 = c_0$
\item[(ii)] $\mathcal{H} \left [ a_1 \right ] = b_1$, $\mathcal{H} \left [ b_1 \right ] = a_1$, $\overline{e_0} \mathcal{H} \left [ c_1 \right ] = c_1$
\item[(iii)] $c_{-1} \ast a_1 = c_{-1} \ast \overline{e_0} b_1 = c_{-1} \ast c_1 = c_1$
\item[(iv)] $a_0 \ast c_0 = c_0 \ast a_0 = c_1$, $\overline{e_0} b_0 \ast c_0 = c_0 \ast \overline{e_0} b_0 = c_1$
\end{itemize}
\end{lemma}
\par\noindent {\bf Proof}~\par\noindent
For (i) we consecutively have
\begin{eqnarray*}
- \underline{\partial} a_1 & = & - \frac{1}{\pi} \frac{1}{\sigma_m} \frac{1}{m-2} \, \underline{\partial} T^\ast_{-m+2} = - \frac{1}{\pi} \frac{1}{\sigma_m} \frac{1}{m-2} (-m+2) \, U^\ast_{-m+1} = \frac{1}{\pi} \frac{1}{\sigma_m} \, U^\ast_{-m+1} = b_0 \\
- \underline{\partial} b_1 & = & \frac{1}{\pi} \frac{1}{\sigma_{m+1}} \frac{1}{m-1} \, \underline{\partial} U^\ast_{-m+2} = \frac{1}{\pi} \frac{1}{\sigma_{m+1}} \frac{1}{m-1} (-2\pi) \, T^\ast_{-m+1} = - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \, T^\ast_{-m+1} = a_0 \\
\end{eqnarray*}
and
$$
- \overline{e_0} \underline{\partial} c_1 = - \overline{e_0} \underline{\partial} \left ( \frac{1}{2} a_1 + \frac{1}{2} \overline{e_0} b_1 \right ) = \overline{e_0} \frac{1}{2} b_0 + \frac{1}{2} a_0 = c_0
$$
For (ii) we obtain
\begin{eqnarray*}
\mathcal{H} \left [ a_1 \right ] & = & \mathcal{H} \left [ a_0 \ast a_0 \right ] = b_0 \ast a_0 = b_1 \\
\mathcal{H} \left [ b_1 \right ] & = & \mathcal{H}^2 \left [ a_1 \right ] = a_1
\end{eqnarray*}
and
$$
\overline{e_0} \mathcal{H} \left [ c_1 \right ] = \overline{e_0} \left ( \frac{1}{2} \mathcal{H} \left [ a_1 \right ] + \frac{1}{2} \mathcal{H} \left [ \overline{e_0} b_1 \right ] \right ) = \frac{1}{2} \overline{e_0} b_1 + \frac{1}{2} a_1 = c_1
$$
For (iii) it holds that
\begin{eqnarray*}
c_{-1} \ast a_1 & = & \left ( \frac{1}{2} \delta + \frac{1}{2} \overline{e_0} H \right ) \ast a_1 = \frac{1}{2} a_1 + \frac{1}{2} \overline{e_0} b_1 = c_1 \\
c_{-1} \ast \overline{e_0} b_1 & = & \left ( \frac{1}{2} \delta + \frac{1}{2} \overline{e_0} H \right ) \ast (\overline{e_0} b_1 ) = \frac{1}{2} \overline{e_0} b_1 + \frac{1}{2} a_1 = c_1
\end{eqnarray*}
and
$$
c_{-1} \ast c_1 = c_{-1} \ast \left ( \frac{1}{2} a_1 + \frac{1}{2} \overline{e_0} b_1 \right ) = \frac{1}{2} c_1 + \frac{1}{2} c_1 = c_1
$$
Finally, (iv) follows from the following calculations:
\begin{eqnarray*}
a_0 \ast c_0 & = & a_0 \ast \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) = \frac{1}{2} a_1 + \frac{1}{2} \overline{e_0} b_1 = c_1 \\
c_0 \ast a_0 & = & \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) \ast a_0 = \frac{1}{2} a_1 + \frac{1}{2} \overline{e_0} b_1 = c_1
\end{eqnarray*}
and
\begin{eqnarray*}
\overline{e_0} b_0 \ast c_0 & = & \overline{e_0} b_0 \ast \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) = \frac{1}{2} \overline{e_0} b_1 + \frac{1}{2} a_1 = c_1 \\
c_0 \ast \overline{e_0} b_0 & = & \left ( \frac{1}{2} a_0 + \frac{1}{2} \overline{e_0} b_0 \right ) \ast \overline{e_0} b_0 = \frac{1}{2} \overline{e_0} b_1 + \frac{1}{2} a_1 = c_1
\end{eqnarray*}
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
Now that we have the distributional boundary values $a_1(\underline{x})$ and $b_1(\underline{x})$ at our disposal, the following relations between the harmonic potentials $(A_1,B_1)$ and $(A_{-1},B_{-1})$ may be readily shown.
\begin{lemma}
\label{lem55}
One has, convolutions being taken in the variable $\underline{x} \in \mathbb{R}^m$:
\begin{itemize}
\item[(i)] $A_1(x_0,\underline{x}) = a_1(\cdot) \ast A_{-1}(x_0,\cdot)(\underline{x}) = b_1(\cdot) \ast B_{-1} (x_0,\cdot)(\underline{x})$
\item[(ii)] $B_1(x_0,\underline{x}) = a_1(\cdot) \ast B_{-1}(x_0,\cdot)(\underline{x}) = b_1(\cdot) \ast A_{-1} (x_0,\cdot)(\underline{x})$
\end{itemize}
\end{lemma}
Note that $( - \Delta_m) a_1 = ( - \underline{\partial})^2 a_1 = ( - \underline{\partial}) b_0 = \delta$, and indeed, in $-a_1 = - \frac{1}{\sigma_m} \frac{1}{m-2} \frac{1}{|\underline{x}|^{m-2}}$ we recognize the fundamental solution of the Laplace operator $\Delta_m$ in $\mathbb{R}^m$. Also note the commutative schemes which are each others Hilbert image:
\begin{equation}
\begin{array}{ccc}
A_1(x_0,\underline{x}) & \xrightarrow{\hspace*{2mm} - \underline{\partial} \hspace*{2mm}} & B_{0}(x_0,\underline{x}) \\[2mm]
\hspace*{-8mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
a_{1}(\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & b_{0}(\underline{x})
\end{array}
\quad \mbox{and} \qquad
\begin{array}{ccc}
B_{1}(x_0,\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & A_{0}(x_0,\underline{x}) \\[2mm]
\hspace*{-7mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
b_{1}(\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & a_{0}(\underline{x})
\end{array}
\label{cschemes6}
\end{equation}
As before, the potential kernels $A_1(x_0,\underline{x})$, $B_1(x_0,\underline{x})$ and $C_1(x_0,\underline{x})$ may be used as convolution kernels to define conjugate harmonic functions and their monogenic sum in $\mathbb{R}^{m+1}_+$, by putting, for a function $f \in L_2(\mathbb{R}^m)$:
$$
\left \{ \begin{array}{rcl}
\mathcal{A}_1[f] & = & A_1(x_0,\cdot) \ast f(\cdot) (\underline{x}) \ = \ \mathcal{A}_0 \left [ a_0 \ast f \right ] \ = \ \mathcal{B}_0 \left [ b_0 \ast f \right ] \\[2mm]
\mathcal{B}_1[f] & = & B_1(x_0,\cdot) \ast f(\cdot) (\underline{x}) \ = \ \mathcal{A}_0 \left [ b_0 \ast f \right ] \ = \ \mathcal{B}_0 \left [ a_0 \ast f \right ]
\end{array} \right .
$$
and
$$
\mathcal{C}_1[f] = C_1(x_0,\cdot) \ast f(\cdot) (\underline{x}) = \mathcal{C}_0 \left [ a_0 \ast f \right ] = \mathcal{C}_0 \left [ \overline{e_0} b_0 \ast f \right ]
$$
The corresponding non--tangential $L_2$--boundary values for $x_0 \rightarrow 0+$ are given by $\lim_{x_0 \rightarrow 0+} \mathcal{A}_1 [f] = a_1 \ast f$, $\lim_{x_0 \rightarrow 0+} \mathcal{B}_1 [f] = b_1 \ast f$ and $\lim_{x_0 \rightarrow 0+} \mathcal{C}_1 [f] = c_1 \ast f$, and the commutative schemes (\ref{cschemes6}) eventually lead to the following one:
$$
\begin{array}{rcl}
\mathcal{C}_{1}[f] & \xrightarrow{\hspace*{2mm} -\overline{e_0} \underline{\partial} \hspace*{2mm}} & \mathcal{C}_{0}[f] \\[2mm]
^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow \hspace*{4mm} & & \hspace*{4mm} \downarrow ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}}\\
c_1 \ast f & \xrightarrow{\hspace*{2mm} -\overline{e_0} \underline{\partial} \hspace*{2mm}} & c_{0} \ast f
\end{array}
$$
The conjugate harmonic potentials $A_1(x_0,\underline{x})$ and $B_1(x_0,\underline{x})$ may now be determined explicitly by a computation similar to the one used in Section 3 to determine the conjugate harmonic of Green's function. For this and subsequent calculations the dimension $m$ is assumed to be great enough in order that the expressions obtained should remain valid. Starting from the equation (\ref{verify1})
$$
\partial_{x_0} A_1(x_0,\underline{x}) = A_0(x_0,\underline{x}) = - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \frac{1}{|x_0 e_0 + \underline{x}|^{m-1}}
$$
we find in $\mathbb{R}^{m+1}_+$
\begin{equation}
A_1(x_0,\underline{x}) = a_1(\underline{x}) - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \frac{1}{|\underline{x}|^{m-2}} \, \widetilde{F}_{m-2} \left ( \frac{x_0}{|\underline{x}|} \right )
\label{A1}
\end{equation}
where we have put
$$
\widetilde{F}_{m-2} \left ( u \right ) = \int_0^u \frac{d\zeta}{\left ( 1+ \zeta^2 \right )^\frac{m-1}{2}}
$$
A priori it is not clear that $A_1(x_0,\underline{x})$ is well--defined for $\underline{x}=0$. However, in virtue of the relation
$$
\widetilde{F}_{m-2}(u) = F_{m-2}(+\infty) - F_{m-2} \left ( \frac{1}{u} \right ) = \frac{\sqrt{\pi}}{2} \frac{2}{m-2} \frac{
\Gamma \left ( \frac{m}{2} \right )}{\Gamma \left ( \frac{m-1}{2} \right )} - F_{m-2} \left ( \frac{1}{u} \right )
$$
expression (\ref{A1}) for $A_1(x_0,\underline{x})$ is turned into, with $m>2$:
$$
A_1(x_0,\underline{x}) = \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \frac{1}{|\underline{x}|^{m-2}} \, F_{m-2} \left ( \frac{|\underline{x}|}{x_0} \right )
$$
or, introducing again the hypergeometric function $_2F_1$,
$$
A_1(x_0,\underline{x}) = \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \frac{1}{m-2} \frac{1}{x_0^{m-2}} \; _2F_1 \left ( \frac{m}{2} -1; \frac{m-1}{2}; \frac{m}{2} ; - \frac{|\underline{x}|^2}{x_0^2} \right )
$$
showing that $A_1(x_0,\underline{x})$ indeed is well--defined for $|\underline{x}|=0$ with
$$
A_1(x_0,0) = \frac{2}{(m-1)(m-2)} \frac{1}{x_0^{m-2}}, \qquad x_0 > 0
$$
By some lengthy calculations it may be verified that the above function $A_1(x_0,\underline{x})$ also satisfies the equation
$$
- \underline{\partial} A_1(x_0,\underline{x}) = B_0(x_0,\underline{x}) = \frac{2}{\sigma_{m+1}} \frac{\underline{x}}{|\underline{x}|^m} \, F_m \left ( \frac{|\underline{x}|}{x_0} \right )
$$
and also shows the distributional limit (\ref{a1b1}) given by
$$
\lim_{x_0 \rightarrow 0+} \ A_1(x_0, \underline{x}) = \frac{1}{\sigma_m} \frac{1}{m-2} \frac{1}{|\underline{x}|^{m-2}} = a_1(\underline{x})
$$
It is perhaps interesting to mention that in the course of these calculations, use has been made of the following recurrence relation for the function $F_m$:
$$
F_m(v) = \frac{m-2}{m-1} \, F_{m-2}(v) - \frac{1}{m-1} \frac{v^{m-2}}{\left ( 1+v^2 \right )^\frac{m-1}{2}}
$$
For the harmonic potential $B_1(x_0,\underline{x})$ we start the computation from equation (\ref{verify2}):
$$
\partial_{x_0} B_1 (x_0,\underline{x}) = B_0(x_0,\underline{x}) = \frac{2}{\sigma_{m+1}} \frac{\underline{x}}{|\underline{x}|^m} \, F_m \left ( \frac{|\underline{x}|}{x_0} \right )
$$
leading to the expression, with $m>1$:
$$
B_1(x_0,\underline{x}) = \frac{2}{\sigma_{m+1}} \frac{x_0 \underline{x}}{|\underline{x}|^m} \, F_m \left ( \frac{|\underline{x}|}{x_0} \right ) - \frac{2}{\sigma_{m+1}} \frac{1}{m-1} \frac{\underline{x}}{|x|^{m-1}}
$$
or
$$
B_1(x_0,\underline{x}) = \frac{2}{m} \frac{1}{\sigma_{m+1}} \frac{\underline{x}}{x_0^{m-1}} \, _2F_1 \left ( \frac{m}{2}; \frac{m+1}{2} ; \frac{m}{2}+1 ; - \frac{|\underline{x}|^2}{x_0^2} \right ) - \frac{2}{\sigma_{m+1}} \frac{1}{m-1} \frac{\underline{x}}{|x|^{m-1}}
$$
showing that $B_1(x_0,\underline{x})$ is well--defined for $\underline{x}=0$ with
$$
B_1(x_0,0) =0 ,\qquad x_0 > 0
$$
Note that the distributional limit $b_1(\underline{x})$ is indeed recovered:
$$
\lim_{x_0 \rightarrow 0+} B_1(x_0,\underline{x}) = - \frac{2}{\sigma_{m+1}} \frac{1}{m-1} \frac{\underline{x}}{|x|^{m-1}} = b_1((\underline{x})
$$
It may now readily be verified that the above function $B_1(x_0,\underline{x})$ also satisfies equation (\ref{verify4}):
$$
- \underline{\partial} B_1(x_0,\underline{x}) = A_0(x_0,\underline{x}) = - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \frac{1}{|x|^{m-1}}
$$
\subsection{The next step}
\noindent
Proceeding in a similar way we may define in the next step
$$
\left \{
\begin{array}{rcl}
A_2(x_0,\underline{x}) & = & a_0(\cdot) \ast A_1(x_0,\cdot)(\underline{x}) = b_0(\cdot) \ast B_1(x_0,\cdot)(\underline{x}) \\[2mm]
B_2(x_0,\underline{x}) & = & a_0(\cdot) \ast B_1(x_0,\cdot)(\underline{x}) = b_0(\cdot) \ast A_1(x_0,\cdot)(\underline{x})
\end{array}
\right .
$$
and verify that in $\mathbb{R}^{m+1}_+$ it holds that
\begin{eqnarray*}
\partial_{x_0} A_2 &=& a_0 \ast \partial_{x_0} A_1 \ = \ a_0 \ast A_0 \ = \ A_1 \\
\partial_{x_0} A_2 &=& b_0 \ast \partial_{x_0} B_1 \ = \ b_0 \ast B_0 \ = \ A_1
\end{eqnarray*}
and
\begin{eqnarray*}
\partial_{x_0} B_2 &=& a_0 \ast \partial_{x_0} B_1 \ = \ a_0 \ast B_0 \ = \ B_1 \\
\partial_{x_0} B_2 &=& b_0 \ast \partial_{x_0} A_1 \ = \ b_0 \ast A_0 \ = \ B_1
\end{eqnarray*}
while also
\begin{eqnarray*}
-\underline{\partial} A_2 &=& -\underline{\partial} a_0 \ast A_1 \ = \ H \ast A_1 \ = \ \mathcal{H} \left [ A_1 \right ] = B_1\\
-\underline{\partial} A_2 &=& -\underline{\partial} b_0 \ast B_1 \ = \ \delta \ast B_1 \ = \ B_1
\end{eqnarray*}
and
\begin{eqnarray*}
-\underline{\partial} B_2 &=& -\underline{\partial} a_0 \ast B_1 \ = \ H \ast B_1 \ = \ \mathcal{H} \left [ B_1 \right ] = A_1\\
-\underline{\partial} B_2 &=& -\underline{\partial} b_0 \ast A_1 \ = \ \delta \ast A_1 \ = \ A_1
\end{eqnarray*}
These relations justify $A_2(x_0,\underline{x})$ and $B_2(x_0,\underline{x})$ to be called conjugate harmonic potentials in $\mathbb{R}^{m+1}_+$ of the function $C_1(x_0,\underline{x})$. It follows that
$$
C_2(x_0,\underline{x}) = \frac{1}{2} A_2(x_0,\underline{x}) + \frac{1}{2} \overline{e_0} B_2(x_0,\underline{x})
$$
is a monogenic potential in $\mathbb{R}^{m+1}_+$ of $C_1$ and there also holds $\overline{D} C_2 = \partial_{x_0} C_2 = (-\overline{e_0} \underline{\partial}) C_2 = C_1$. As before, the conjugate harmonic potentials $A_2(x_0,\underline{x})$ and $B_2(x_0,\underline{x})$ form a Hilbert pair in the variable $\underline{x} \in \mathbb{R}^m$:
$$
\left \{ \begin{array}{rcl}
\mathcal{H} \left [ A_2(x_0,\underline{x}) \right ] & = & H(\cdot) \ast A_2(x_0,\cdot)(\underline{x}) \ = \ b_{-1}(\cdot) \ast A_2(x_0,\cdot)(\underline{x}) \ = \ B_2(x_0,\underline{x}) \\[2mm]
\mathcal{H} \left [ B_2(x_0,\underline{x}) \right ] & = & \mathcal{H}^2 \left [ A_2(x_0,\underline{x}) \right ] \ = \ b_{-1}(\cdot) \ast B_2(x_0,\cdot)(\underline{x}) \ = \ A_2(x_0,\underline{x})
\end{array} \right .
$$
while, trivially,
$$
\left \{ \begin{array}{rcl}
a_{-1}(\cdot) \ast A_2(x_0,\cdot)(\underline{x}) & = & A_2(x_0,\underline{x}) \\[2mm]
a_{-1}(\cdot) \ast B_2(x_0,\cdot)(\underline{x}) & = & B_2(x_0,\underline{x})
\end{array} \right .
$$
Their distributional limits for $x_0 \rightarrow 0+$ are given by
$$
\left \{ \begin{array}{rcl}
a_2(\underline{x}) & = & \lim_{x_0 \rightarrow 0+} A_2(x_0,\underline{x}) \ = \ a_0 \ast a_1(\underline{x}) \ = \ b_0 \ast b_1(\underline{x}) \\[2mm]
b_2(\underline{x}) & = & \lim_{x_0 \rightarrow 0+} B_2(x_0,\underline{x}) \ = \ a_0 \ast b_1(\underline{x}) \ = \ b_0 \ast a_1(\underline{x})
\end{array} \right .
$$
which may be calculated explicitly to be
$$
a_2(\underline{x}) = \left ( - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} T^\ast_{-m+1} \right ) \ast \left ( \frac{1}{\pi} \frac{1}{\sigma_m} \frac{1}{m-2} T^\ast_{-m+2} \right ) = - \frac{2}{(m-1)(m-3)} \frac{1}{\sigma_{m+1}} \frac{1}{|\underline{x}|^{m-3}}
$$
or
$$
a_2(\underline{x}) = \left ( \frac{1}{\pi} \frac{1}{\sigma_{m}} U^\ast_{-m+1} \right ) \ast \left ( - \frac{1}{\pi} \frac{1}{\sigma_{m+1}} \frac{1}{m-1} U^\ast_{-m+2} \right ) = - \frac{1}{\pi} \frac{1}{(m-1)(m-3)} \frac{1}{\sigma_{m+1}} \, T^\ast_{-m+3}
$$
and
$$
b_2(\underline{x}) = \left ( - \frac{2}{m-1} \frac{1}{\sigma_{m+1}} T^\ast_{-m+1} \right ) \ast \left ( - \frac{1}{\pi} \frac{1}{\sigma_{m+1}} \frac{1}{m-1} U^\ast_{-m+2} \right ) = \frac{1}{2} \frac{1}{\sigma_{m}} \frac{1}{m-2} \frac{\underline{x}}{|\underline{x}|^{m-2}}
$$
or
$$
b_2(\underline{x}) = \left ( \frac{1}{\pi} \frac{1}{\sigma_{m}} U^\ast_{-m+1} \right ) \ast \left ( \frac{1}{\pi} \frac{1}{\sigma_{m}} \frac{1}{m-2} T^\ast_{-m+2} \right ) = \frac{1}{2 \pi^2} \frac{1}{\sigma_{m}} \frac{1}{m-2} U^\ast_{-m+3}
$$
Putting $c_2(\underline{x}) = \frac{1}{2} a_2(\underline{x}) + \frac{1}{2} \overline{e_0} b_2(\underline{x})$ we can prove the following properties of those distributional boundary values.
\begin{lemma}
\label{lem56}
\rule{0mm}{0mm}
\begin{itemize}
\item[(i)] $-\underline{\partial} a_2 = b_1$, $- \underline{\partial} b_2 = a_1$, $- \overline{e_0} \underline{\partial} c_2 = c_1$
\item[(ii)] $\mathcal{H} \left [ a_2 \right ] = b_2$, $\mathcal{H} \left [ b_2 \right ] = a_2$, $\overline{e_0} \mathcal{H} \left [ c_2 \right ] = c_2$
\item[(iii)] $c_{-1} \ast a_2 = c_{-1} \ast b_2 = c_{-1} \ast c_2 = c_2$
\item[(iv)] $c_0 \ast a_1 = a_1 \ast c_0 = c_2$, $c_0 \ast \overline{e_0} b_1 = \overline{e_0} b_1 \ast c_0 = c_2$
\end{itemize}
\end{lemma}
\par\noindent {\bf Proof}~\par\noindent
For (i) we obtain
\begin{eqnarray*}
- \underline{\partial} a_2(\underline{x}) & = & - \underline{\partial} \left ( - \frac{1}{\pi} \frac{1}{\sigma_{m+1}} \frac{1}{m-1} \frac{1}{m-3} \, T^\ast_{-m+3} \right ) \ = \ \frac{1}{\pi} \frac{1}{\sigma_{m+1}} \frac{1}{m-1} \frac{1}{m-3} (-m+3) \, U^\ast_{-m+2} \\
& = & - \frac{1}{\pi} \frac{1}{m-1} \frac{1}{\sigma_{m+1}} \, U^\ast_{-m+2} \ = \ b_1(\underline{x}) \\
- \underline{\partial} b_2(\underline{x}) & = & - \underline{\partial} \left ( - \frac{1}{2\pi^2} \frac{1}{\sigma_{m}} \frac{1}{m-2} \, U^\ast_{-m+3} \right ) \ = \ - \frac{1}{2\pi^2} \frac{1}{\sigma_{m}} \frac{1}{m-2} (-2\pi) \, T^\ast_{-m+2} \\
& = & \frac{1}{\pi} \frac{1}{m-2} \frac{1}{\sigma_{m}} \, T^\ast_{-m+2} \ = \ a_1(\underline{x}) \\
\left ( \overline{e_0} \underline{\partial} \right ) c_2(\underline{x}) & = & \left ( \overline{e_0} \underline{\partial} \right ) \left ( \frac{1}{2} a_2 + \frac{1}{2} \overline{e_0} b_2 \right ) = \frac{1}{2} \overline{e_0} b_1 + \frac{1}{2} a_1 \ = \ c_1(\underline{x})
\end{eqnarray*}
while for (ii)
\begin{eqnarray*}
\mathcal{H} \left [ a_2 \right ] & = & \mathcal{H} \left [ a_0 \ast a_1 \right ] \ = \ \mathcal{H} \left [ a_0 \right ] \ast a_1 \ = \ b_0 \ast a_1 \ = \ b_2 \\[1mm]
\mathcal{H} \left [ b_2 \right ] & = & \mathcal{H}^2 \left [ a_2 \right ] \ = \ a_2 \\
\overline{e_0} \mathcal{H} \left [ c_2 \right ] & = & \overline{e_0} \mathcal{H} \left [ \frac{1}{2} a_2 + \frac{1}{2} \overline{e_0} b_2 \right ] \ = \ \frac{1}{2} \overline{e_0} b_2 + \frac{1}{2} a_2 \ = \ c_2
\end{eqnarray*}
Statements (iii) and (iv) follow by direct computation.
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
Making use of the distributional boundary values $a_1(\underline{x})$, $b_1(\underline{x})$, $a_2(\underline{x})$ and $b_2(\underline{x})$ we may now prove by direct computation the following equivalent expressions for the conjugate harmonic potentials $A_2(x_0,\underline{x})$ and $B_2(x_0,\underline{x})$.
\begin{lemma}
\label{lem57}
One has, convolutions being taken in the variable $\underline{x} \in \mathbb{R}^{m}$:
\begin{itemize}
\item[(i)] $A_2 = a_1 \ast A_0 = b_1 \ast B_0 = a_2 \ast A_{-1} = b_2 \ast B_{-1}$
\item[(ii)] $B_2 = a_1 \ast B_0 = b_1 \ast A_0 = a_2 \ast B_{-1} = b_2 \ast A_{-1}$
\end{itemize}
\end{lemma}
Also note the commutative schemes
$$
\begin{array}{ccc}
A_2(x_0,\underline{x}) & \xrightarrow{\hspace*{2mm} - \underline{\partial} \hspace*{2mm}} & B_{1}(x_0,\underline{x}) \\[2mm]
\hspace*{-8mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
a_{2}(\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & b_{1}(\underline{x})
\end{array}
\quad \mbox{and} \qquad
\begin{array}{ccc}
B_{2}(x_0,\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & A_{1}(x_0,\underline{x}) \\[2mm]
\hspace*{-7mm} ^{\hspace*{0.1mm}_{x_0 \rightarrow 0+}} \downarrow & & \downarrow \\
b_{2}(\underline{x}) & \xrightarrow{\hspace*{1mm} - \underline{\partial} \hspace*{1mm}} & a_{1}(\underline{x})
\end{array}
$$
We also have explicitly determined the conjugate harmonic potentials $A_2(x_0,\underline{x})$ (for $m>3$) and $B_2(x_0,\underline{x})$ (for $m>2)$:
\begin{eqnarray*}
A_2(x_0,\underline{x}) & = & \frac{2}{m-1} \frac{1}{\sigma_{m+1}} \frac{x_0}{|\underline{x}|^{m-2}} \, F_{m-2} \left ( \frac{|\underline{x}|}{x_0} \right ) - \frac{2}{m-1} \frac{1}{m-3} \frac{1}{\sigma_{m+1}} \frac{1}{|x|^{m-3}} \\
B_2(x_0,\underline{x}) & = & \frac{1}{\sigma_{m+1}} \frac{\underline{x} |x|^2}{|\underline{x}|^{m}} \, F_{m} \left ( \frac{|\underline{x}|}{x_0} \right ) - \frac{m-3}{m-1} \frac{1}{\sigma_{m+1}} \frac{\underline{x}}{|\underline{x}|^{m-2}} \, F_{m-2} \left ( \frac{|\underline{x}|}{x_0} \right )
\end{eqnarray*}
\subsection{The general case}
Inspired by the properties of the harmonic potentials $A_1(x_0,\underline{x})$, $B_1(x_0,\underline{x})$, $A_2(x_0,\underline{x})$ and $B_2(x_0,\underline{x})$, we define recursively, for general $k=1,2,3,\ldots$ , the following functions in $\mathbb{R}^{m+1}_+$, the convolutions being taken in the variable $\underline{x} \in \mathbb{R}^{m}$:
\begin{eqnarray*}
A_k(x_0,\underline{x}) & = & a_0 \ast A_{k-1} \ = \ a_1 \ast A_{k-2} \ = \ \ldots \ = \ a_{k-1} \ast A_0 \\
& = & b_0 \ast B_{k-1} \ = \ b_1 \ast B_{k-2} \ = \ \ldots \ = \ b_{k-1} \ast B_0 \\
B_k(x_0,\underline{x}) & = & a_0 \ast B_{k-1} \ = \ a_1 \ast B_{k-2} \ = \ \ldots \ = \ a_{k-1} \ast B_0 \\
& = & b_0 \ast A_{k-1} \ = \ b_1 \ast A_{k-2} \ = \ \ldots \ = \ b_{k-1} \ast A_0 \\
\end{eqnarray*}
and
$$
C_k(x_0,\underline{x}) = \frac{1}{2} A_k(x_0,\underline{x}) + \frac{1}{2} \overline{e_0} B_k(x_0,\underline{x})
$$
Note that for $k=1,2$ we indeed recover the harmonic potentials studied in the previous subsections. In a similar way as above, it is now shown that $A_k(x_0,\underline{x})$, $B_k(x_0,\underline{x})$ and $C_k(x_0,\underline{x})$ satisfy the following equations:
\begin{itemize}
\item[(i)] $\partial_{x_0} A_k = A_{k-1}$ \\[1mm]
$- \underline{\partial} A_k = B_{k-1}$ \\[1mm]
$\overline{D} A_k = \frac{1}{2} \left ( \partial_{x_0} - \overline{e_0} \underline{\partial} \right ) A_k = \frac{1}{2} A_{k-1} + \frac{1}{2} \overline{e_0} B_{k-1} = C_{k-1}$
\item[(ii)] $\partial_{x_0} B_k = B_{k-1}$ \\[1mm]
$- \underline{\partial} B_k = A_{k-1}$ \\[1mm]
$\overline{D} \left ( \overline{e_0} B_k \right ) = \frac{1}{2} \left ( \partial_{x_0} - \overline{e_0} \underline{\partial} \right ) \overline{e_0} B_k = \frac{1}{2} \overline{e_0} B_{k-1} + \frac{1}{2} A_{k-1} = C_{k-1}$
\item[(iii)] $D C_k = \frac{1}{2} \left ( \partial_{x_0} + \overline{e_0} \underline{\partial} \right ) \left ( \frac{1}{2} A_k + \frac{1}{2} \overline{e_0} B_k \right ) = 0$
\item[(iv)] $\overline{D} C_k = \overline{D} \left ( \frac{1}{2} A_k + \frac{1}{2} \overline{e_0} B_k \right ) = C_{k-1}$
\end{itemize}
which clearly show that $A_k(x_0,\underline{x})$ and $B_k(x_0,\underline{x})$ are conjugate harmonic potentials of $C_{k-1}(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$, while $C_k(x_0,\underline{x})$ is a monogenic potential of the same $C_{k-1}(x_0,\underline{x})$ in $\mathbb{R}^{m+1}_+$. Their distributional boundary values for $x_0 \rightarrow 0+$ are given by the recurrence relations
\begin{eqnarray*}
a_k(\underline{x}) & = & a_0 \ast a_{k-1}\ = \ a_1 \ast a_{k-2} \ = \ \ldots \ = \ a_{k-1} \ast a_0 \\
& = & b_0 \ast b_{k-1}\ = \ b_1 \ast b_{k-2} \ = \ \ldots \ = \ b_{k-1} \ast b_0 \\
b_k(\underline{x}) & = & a_0 \ast b_{k-1}\ = \ a_1 \ast b_{k-2} \ = \ \ldots \ = \ a_{k-1} \ast b_0 \\
& = & b_0 \ast a_{k-1}\ = \ b_1 \ast a_{k-2} \ = \ \ldots \ = \ b_{k-1} \ast a_0
\end{eqnarray*}
for which the following explicit formulae may be deduced:
$$
\left \{ \begin{array}{rcl}
a_{2j} & = & -\displaystyle\frac{1}{2^{j-1}} \displaystyle\frac{1}{\pi^j} \displaystyle\frac{1}{(m-1)(m-3) \ldots (m-2j-1)} \displaystyle\frac{1}{\sigma_{m+1}} \, T^\ast_{-m+2j+1} \\[5mm]
a_{2j-1} & = & \phantom{-} \displaystyle\frac{1}{2^{j-1}} \displaystyle\frac{1}{\pi^j} \displaystyle\frac{1}{(m-2)(m-4) \ldots (m-2j)} \displaystyle\frac{1}{\sigma_{m}} \, T^\ast_{-m+2j}
\end{array} \right .
$$
$$
\left \{ \begin{array}{rcl}
b_{2j} & = & \phantom{-} \displaystyle\frac{1}{2^{j}} \displaystyle\frac{1}{\pi^{j+1}} \displaystyle\frac{1}{(m-2)(m-4) \ldots (m-2j)} \displaystyle\frac{1}{\sigma_{m}} \, U^\ast_{-m+2j+1} \\[5mm]
b_{2j-1} & = & - \displaystyle\frac{1}{2^{j-1}} \displaystyle\frac{1}{\pi^j} \displaystyle\frac{1}{(m-1)(m-3) \ldots (m-2j+1)} \displaystyle\frac{1}{\sigma_{m+1}} \, U^\ast_{-m+2j}
\end{array} \right .
$$
These distributional limits show the following, by now traditional, properties.
\begin{lemma}
\label{lem58}
One has for $k=1,2,\ldots$:
\begin{itemize}
\item[(i)] $- \underline{\partial} a_k = b_{k-1}$
\item[(ii)] $- \underline{\partial} b_k = a_{k-1}$
\item[(iii)] $\mathcal{H} \left [ a_k \right ] = b_{-1} \ast a_k = b_k$
\item[(iv)] $\mathcal{H} \left [ b_k \right ] = b_{-1} \ast b_k = a_k$
\end{itemize}
\end{lemma}
\par\noindent {\bf Proof}~\par\noindent
Follows by direct computation using the derivation and convolution formulae for the $T^\ast$-- an $U^\ast$--distributions.
~\hfill{$\square$}\pagebreak[1]\par\medskip\par
\section{Conclusion}
While constructing a higher dimensional analogue in upper half--space $\mathbb{R}^{m+1}_+$ of the function $\ln{z}$ in the upper half of the complex plane, preserving its fundamental property of being a holomorphic potential of the Cauchy kernel $\frac{1}{z}$, it became clear that this monogenic logarithmic function is but one of a double sequence of such kind of potentials, just as $\ln{z}$ is the central element in the double sequence of holomorphic primitives:
$$
\frac{1}{k!} z^k \left[ \ln z - ( 1 + \frac{1}{2} + \ldots + \frac{1}{k}) \right] \rightarrow \ldots \rightarrow z ( \ln z - 1) \rightarrow \ln z
\stackrel{\frac{d}{dz}}{\longrightarrow} \frac{1}{z} \rightarrow - \frac{1}{z^2} \rightarrow \ldots \rightarrow (-1)^{k-1} \frac{(k-1)!}{z^k}
$$
The sequence of monogenic potentials corresponding to the negative integer powers of $z$, which we called {\em downstream potentials}, were rather easily constructed via differentiation with the conjugate generalized Cauchy--Riemann operator $\overline{D}$. The explicit construction of the monogenic potentials corresponding to the logarithmic functions in $\mathbb{C}_+$, which we termed {\em upstream potentials}, requires tedious calculations involving primitivation with respect to $\overline{D}$, and up to now we have executed three inductive steps. A general expression for these upstream potentials is lacking, but their properties are known since they arise as convolutions of adjacent potentials with their distributional boundary values in $\mathbb{R}^m$. Also with an eye on possible applications, the upstream potentials will be further calculated in the lower dimensional cases where $m=2,3$, and it is hoped for that a general formula, mimicking the one in the complex plane, will appear.\\[-2mm]
The above mentioned distributional boundary values are really fundamental, since not only they are used in the definition of the potentials, but also uniquely determine the conjugate harmonic potentials obtained by primitivation, thanks to the simple, but crucial, fact that a monogenic function in $\mathbb{R}^{m+1}_+$ vanishing at the boundary $\mathbb{R}^{m}$ indeed is zero. For those distributional boundary values we have established a general formula, showing that they all fit into two families of distributions in $\mathbb{R}^{m}$, one scalar--valued, the second one Clifford vector--valued. In some particular cases they have been identified as fundamental solutions of the Dirac operator or the Laplace operator, or as convolution kernels for some pseudodifferential operators related to both these operators. The forthcoming paper \cite{bdbds} will treat this remarkable relationship between the distributional boundary values of the harmonic potentials and specific integer and half--integer powers of the Dirac and Laplace operators.
|
\section{Introduction}
Since its inception, the phenomenon of Anderson localization
in a disordered lattice~\cite{pwa1} has been found to be ubiquitous in
diverse fields of condensed matter physics and materials science. While the
popular domain of interest is related to electronic systems, where the
quantum interference plays a pivotal role in localizing electronic
eigenstates in presence of
disorder~\cite{angus,thouless,schreiber,alberto1,alberto2,zilly}, the effect
is by no means, confined to electrons, and extend over a variety of
phenomena ranging from spin freezing in one dimensional
semiconductors~\cite{echeverria}, localization in optical
lattices~\cite{sankar1,edwards,roati}, or the localization of matter waves
(cold atoms forming Bose-Einstein condensates)~\cite{gavish,lye}, to name a
few. Incidentally, the latter has recently been observed experimentally
in one dimensional matter waveguides, where the random potential is
generated by laser speckles~\cite{billy}.
In 1984 Sajeev John pointed out that the idea of localization goes far
beyond the electronic systems, and is actually a general phenomenon common
to any wave propagation in systems with disorder~\cite{john}. Anderson
followed with a seminal paper considering the idea of localization of
classical waves, in an attempt to work out the theory of white
paints~\cite{pwa2}. The field gathered momentum in the last couple of
decades and a considerable volume of literature related to the localization
of classical waves, particularly light, ultrasound and microwave is now in
existence~\cite{jovic1,bliokh,jovic2,sutherland,asatryan,hodges,he,yablonovich1,shi}.
The study of localization of light in disordered media has been patronized
by the discovery of the photonic band gap (PBG)
materials~\cite{soukoulis,john2,yablonovitch2}. These systems exhibit gaps
in the frequency spectrum in which propagation of electromagnetic waves is
forbidden. This has important implications in both fundamental science and
technological applications.
Photonic gaps, apart from materials with a large dielectric constant, can
also be observed in waveguide networks, as proposed by several groups over
the past years~\cite{zhang,dobrzynski,vasseur,pradhan,li,sheelan1,sheelan2,lu1,lu2}.
Anderson localized eigenmodes are observed inside the photonic gaps and
excellent agreement between theory and experiments has been
obtained~\cite{zhang}. The network models are able to localize a propagating
wave by virtue of the geometrical arrangement of the waveguide segments. Of
particular interest are a wide variety of models based on waveguide networks
designed following a deterministic fractal
geometry~\cite{li,sheelan1,sheelan2,lu1,lu2}, where the gaps result from the
typical topology exhibited the hierarchical arrangement of the
waveguide segments. The present communication also deals with a
hierarchically designed (fractal geometry) waveguide network, but addresses
a deeper fundamental question regarding wave localization in such systems,
as explained below.
Fractals or hierarchical geometries in general cause an excitation to localize~\cite{kadanoff,rammal,kappertz,schwalm1}. The energy spectrum
turns out to be singular continuous~\cite{kadanoff}, with a gap in the
vicinity of every energy. Nevertheless, there can be a countable
infinity of {\it extended} eigenfunctions with high (or even, perfect)
transmittivity, even though there is no translational order in such systems.
Once again, this is true for electrons~\cite{bibhas,anirban,xiang,arun1,macia,moritz,arun2},
and classical waves as well~\cite{sheelan1,sheelan2}. A curious point,
apparently gone unnoticed or un-appreciated so far is that, while a precise
determination of the eigenvalues corresponding to the extended eigenmodes is
possible in the above cases of hierarchically grown fractal networks, the
task seems to be practically impossible when it comes to an exact evaluation
of eigenvalues of the localized modes in such hierarchical systems in their
{\it thermodynamic limits}. Direct diagonalization of the Hamiltonian (in
the electronic case), or an exact numerical solution of the wave equation
doesn't help, as the overall character of the spectrum in all the cases is
highly fragmented, and the eigenvalues obtained from a finite sized network
are likely to {\it slip away} from the spectrum once we go over to a higher
generation. This problem has recently been addressed in the context of
electron localization in fractal space~\cite{biplab}, and we carry forward
the central idea floated in Ref.~\cite{biplab} to evaluate the {\it exact}
wavelengths (wave vectors) of electromagnetic waves that can be localized
{\it at will} in a properly designed hierarchical single channel network. To
the best of our knowledge, this issue remains unaddressed so
far in the literature in the context of localization of wave, light in particular.
We design a Vicsek fractal network~\cite{vicsek} consisting of {\it diamond}
shaped loops each arm of which mimics a single mode linear waveguide~[see
Fig.~\ref{system}]. While examination of the localized mode eigenvalues and
the nature of localization are indeed the major factors motivating this work,
other interests in such a study are related to the general spectral
character and classical wave transport in these systems.
\begin{figure*}[ht]
\includegraphics[clip,width=16cm]{lattice.eps}
\caption{ (Color online) (a) Schematic view of the second generation of an
infinite diamond-Vicsek waveguide network. The vertices at the junction of
two waveguides are named as `$A$' in the text, while those in the bulk (at
the crossing of four waveguides) are termed `$B$'. (b) Renormalized version
of (a) with the dotted lines indicating the diagonal `hopping' which is
generated due to renormalization.}
\label{system}
\end{figure*}
As one can easily appreciate, the geometry of a diamond-Vicsek network
provides an interesting culmination of the `open' character of a typical
Vicsek pattern and closed loops at shorter scales of length. This is in
marked contrast to the much studied Sierpinski gasket waveguide
network~\cite{li}, which is a closed structure, or to the other
deterministic waveguide networks~\cite{sheelan2,lu1}. The presence of the
loops generates a possibility of an effectively long ranged propagation of waves
between the various vertices, and its effect on the localization or de-
localization of waves is worth studying.
We find interesting results. For an infinite hierarchical
geometry, such as presented above, a countable infinity of eigenmodes with a
multitude of localization lengths can be {\it precisely detected}. One can
work out an exact mathematical prescription to specify the length scale at
which the onset of localization begins. The localization can in principle,
be {\it delayed} (staggered) in position space and the corresponding wave
vectors (or, wavelengths) can be exactly evaluated following the same
prescription based on a real space renormalization group (RSRG) method~\cite{southern}.
In addition, it is shown that for a given set of parameters, the center of the
spectrum corresponds to a perfectly extended eigenmode, with the parameters
describing the system exhibiting a fixed point behavior. This central
extended mode is flanked on either side by localized wave functions with a
hierarchy of localization lengths.
In what follows we describe the results. In Sec.~\ref{sec2}, the model and
the mathematical method of handling the problem are presented.
Section~\ref{sec3} discusses the results and their analyses, and in
Sec.~\ref{sec4} we draw our conclusions .
\section{The model and the method}
\label{sec2}
\subsection{The wave equation and its discretization}
We have considered a waveguide network formed by waveguide segments having
the same lengths arranged in a Vicsek fractal geometry~\cite{vicsek}. Each
segment has a single channel for wave propagation. The wave function
$\psi_{ij}$ between any two nodal points $i$ and $j$ satisfies the wave
equation:
\begin{equation}
\label{waveeqn}
\frac{\partial^2 \psi_{ij}(x)}{\partial x^2}
+\frac{\omega^2}{c^2}\psi_{ij}(x) = 0
\end{equation}
where $\omega$ is the frequency of the wave, $c$ is the speed of wave
propagation and $x$ is the distance measured from the $i$th node. The above
equation has the solution of the form~\cite{alexander,ping}
\begin{equation}
\label{soln}
\psi_{ij}(x)= \psi_{i}\frac{\sin[k(\ell_{ij}-x)]}{\sin(k\ell_{ij})}
+ \psi_{j}\frac{\sin(kx)}{\sin(k\ell_{ij})}
\end{equation}
where $k=\omega/c$, $\ell_{ij}$ is the length of the segment between the
node $i$ and $j$, and $\psi_{i}$ and $\psi_{j}$ are the values of the wave
function at the $i$th and $j$th nodes respectively. The flux conservation
condition:
\begin{equation}
\label{condition}
\sum_{j}\left[\frac{\partial}{\partial x}\psi_{ij}(x)\right]_{x=0} = 0
\end{equation}
where the summation $j$ is over all the nodes linked directly to $i$, leads to a discretized version of Eq.~\eqref{waveeqn}~\cite{ping}, viz.,
\begin{equation}
\label{difference1}
-\psi_{i}\sum_{j}\cot\theta_{ij}+\sum_{j}\psi_{j}/\sin\theta_{ij}
= 0
\end{equation}
where $\theta_{ij}= k\ell_{ij} = ka$, $a$ being the constant length of a
waveguide, as considered in all the calculations which follow.
Eq.~\eqref{difference1} resembles a tight binding difference equation
depicting the propagation of non-interacting electrons in a lattice, viz.,
\begin{equation}
\label{difference2}
(E-\epsilon_{i})\psi_{i}=\sum_{j}t_{ij}\psi_{j}
\end{equation}
with the `electron energy' $E$ can be put equal to $2\cos ka$, the `on-site
potential' $\epsilon_{i}=2\cos ka+\sum\limits_{j}\cot \theta_{ij}$ and
`hopping matrix element' $t_{ij}=1/\sin\theta_{ij}$. It should be
appreciated that the choice of $E = 2 \cos ka$ is completely arbitrary, and
this does not affect the final results in any way. As seen in
Fig.~\ref{system}, the Vicsek waveguide network has two types of nodal
points, viz., type $A$ (having two neighboring nodal points) and type $B$
(having four neighboring nodal points). Accordingly, the equivalent `on-site
potentials' are assigned two values, viz.,
$\epsilon_{A}= 2\cos ka+2\cot ka$ and
$\epsilon_{B}= 2\cos ka+4\cot ka$ respectively. The `overlap integral' along
an arm of the waveguide is $t = 1/\sin ka$. There is no second neighbor
tunneling of the wave to begin with. But it will grow with RSRG.
We now proceed to describe the physics of wave propagation in such a fractal
geometry by exploiting this exact analogy with the corresponding electronic
problem.
\subsection{The RSRG scheme}
A renormalized version of the fractal network is easily obtained by
decimating a subset of vertices from the original geometry. This implies one
has to eliminate a subset of the wave amplitudes from the difference
equations~\eqref{difference2} in terms of the surviving vertices [see
Fig.~\ref{system}(b)]. This results in the following set of recursion
relations for the system parameters.
\begin{widetext}
\begin{eqnarray}
\epsilon_{A}^{\prime} &=& \epsilon_{A} +\left[\frac{2t^2}{E-\epsilon_{A}- \lambda} +
\frac{\left(\lambda + \dfrac{2t^2}{E-\epsilon_{A}- \lambda}\right)^{2}
\left\{(E-\epsilon_{B})-\dfrac{2t^2}{E-\epsilon_{A}- \lambda}-\xi\right\}}
{\left\{(E-\epsilon_{B})-\dfrac{2t^2}{E-\epsilon_{A}- \lambda}-\xi\right\}^{2}
-\left(\lambda + \xi\right)^{2}}\right],\nonumber\\
\epsilon_{B}^{\prime} &=& \epsilon_{B} +2\left[\frac{2t^2}{E-\epsilon_{A}- \lambda} +
\frac{\left(\lambda + \dfrac{2t^2}{E-\epsilon_{A}- \lambda}\right)^{2}
\left\{(E-\epsilon_{B})-\dfrac{2t^2}{E-\epsilon_{A}- \lambda}-\xi\right\}}
{\left\{(E-\epsilon_{B})-\dfrac{2t^2}{E-\epsilon_{A}- \lambda}-\xi\right\}^{2}
-\left(\lambda + \xi\right)^{2}}\right],\nonumber\\
t^{\prime} &=& \frac{t\left\{\lambda(E-\epsilon_{A}- \lambda)+2t^2\right\}^{2}}
{\left\{E^2-(\epsilon_{A}+\epsilon_{B})(E-\lambda)
-2E\lambda +\lambda^{2}-2t^2+\epsilon_{A}\epsilon_{B}\right\}^{2}
-4t^2(E-\epsilon_{A}-\lambda)^{2}},\nonumber\\
\lambda^{\prime} &=& \frac{\left(\lambda + \dfrac{2t^2}{E-\epsilon_{A}- \lambda}\right)^{2}
\left(\lambda + \xi\right)}
{\left\{(E-\epsilon_{B})-\dfrac{2t^2}{E-\epsilon_{A}- \lambda}-\xi\right\}^{2}
-\left(\lambda + \xi\right)^{2}}.
\end{eqnarray}
\end{widetext}
where $\xi=2t^2(E-\epsilon_{A}-\lambda)/\Delta$ with
$\Delta=(E-\epsilon_{A}-\lambda)(E-\epsilon_{B}-\lambda)-2t^{2}$,
and $\lambda=0$ at the beginning.
The scaling generates an effective second neighbor `hopping' (overlap)
$\lambda^{\prime}$, as is obvious from the above set of recursion relations
and the diagram~\ref{system}(b). It is important to appreciate that, the
growth of a $\lambda^{\prime}$, unlike the
electronic case~\cite{biplab}, does not have
any physical significance here, as there is no real `overlap' of the wave amplitudes.
The light is essentially confined within the waveguide.
These recursion relations will now be used to obtain information about the
local density of eigenmodes at specific sites of the system, and the
localized or extended character of the modes, as discussed below.
\section{Results and discussion}
\label{sec3}
\subsection{Local density of eigenmodes}
As already stated, the exact mapping of the wave equation on to a discrete
Schr\"{o}dinger type equation allows us to extract information about the
density of electromagnetic modes through a Green's function
analysis~\cite{southern}. We present in Fig.~\ref{ldos1} the density of
modes at a $B$-vertex, which is given by,
\begin{equation}
\rho^{(B)}(ka)= \lim_{\eta \rightarrow 0}\,\left[-\dfrac{1}{\pi}\,
\text{Im}\,\{G^{(B)}(ka + i\eta)\}\right]
\label{dos}
\end{equation}
The distribution of eigenmodes, plotted within $ka < 0 < 2\pi$, shows
clusters of non-zero values over a finite range of the wave vector $k$, and
is found to be symmetric around $ka = \pi/2$. The fragmented Cantor-like
character, typical signature of a fractal spectrum, is apparent. An idea
about the character of the eigenmodes can be obtained by observing the flow
of $t$ under successive RSRG iterations. In general, for an arbitrary value
of $ka$, for which the density of modes is non-zero, $t^{(n)} \rightarrow 0$
as the number of iterations $n$ increases, implying a localized character of
the corresponding eigenmode~\cite{southern,bibhas}. However, for $ka = (2m+1) \pi/2$,
both the first and the second neighbor `hopping integrals' $t$ and $\lambda$
remain non-zero for an indefinite number of iterations. In fact, we observe
a one cycle {\it fixed point} of the entire parameter space, viz.,
$\{\epsilon_{A}^\prime, \epsilon_{B}^\prime, t^\prime, \lambda^\prime\} =
\{\epsilon_{A}, \epsilon_{B}, t, \lambda\}$.
The fact that $t$ and $\lambda$ remain finite at all stages of RSRG implies
that there is a non-zero overlap between the wave amplitudes at all scales
of length, and the corresponding mode is an extended one.
The neighborhood of $ka= (2m+1)\pi/2$ has been scanned minutely. The self
similarity of the spectrum is always seen with dense patches of eigenvalues
clustered throughout the intervals. For many of these eigenvalues $t$ and
$\lambda$ remain finite under successive decimation for a large number of steps.
This indicates that every local {\it band center} at $ka = (2m+1)
\pi/2$ is flanked either by extended modes or at least, eigenmodes with very
large localization lengths.
\begin{figure}[ht]
\includegraphics[clip,width=7cm,angle=-90]{dom.eps}
\caption{ (Color online) Local density of eigenmodes at a bulk $B$-type
vertex of an infinite diamond-Vicsek waveguide network.}
\label{ldos1}
\end{figure}
\subsection{Explicit construction of localized modes}
The deterministic Vicsek fractal waveguide is self-similar at all scales of
length. This feature allows us to explicitly construct a special
distribution of wave functions by suitably exploiting the difference
equations Eq.~\eqref{difference2}. These `special' eigenmodes are localized
and extend over clusters of single channel waveguide segments of various
sizes. The planar extent of such clusters depends on the eigenvalue
corresponding to the localized mode, and can be small or enormous, depending
on the wavelength (or, wave vector). The construction is similar to our
recent work~\cite{biplab} on electronic states.
To elaborate, let us set
\begin{equation}
E = \epsilon_B(n) - 2 \lambda(n)
\label{roots}
\end{equation}
where, $n$ refers to the stage of renormalization. This is in general, a
polynomial equation in $E$ (and hence, in $k$). The zeros of this equation
will be the allowed wave vectors for the infinite system if, and only if,
with them, one can satisfy Eq.~\eqref{difference2} locally at every vertex
of the network. This task can be accomplished by trying to draw a non-
trivial distribution of amplitudes for a value of $E$ ($ka$) obtained from
Eq.~\eqref{roots} on the undecimated vertices of an $n$-step renormalized
network, and then trying to figure out the amplitude distribution on the
original waveguide structure at the bare length scale. This can indeed be
done, as we demonstrate in Fig.~\ref{ampdistribution1}(a) for $n=1$.
\begin{figure*}[ht]
\includegraphics[clip,width=16cm]{waveguide1.eps}
\caption{ (Color online) (a) Distribution of amplitudes of the wave
function for $ka = \pi/6$ (obtained by solving Eq.~\eqref{roots} for $n=1$)
on a second generation network. The dark shaded plaquettes with black lines
at the boundary embrace network vertices with zero amplitude.
The deep red waveguide segments (covering a dark red shaded region) connect vertices with
non-zero amplitude, and `glow' with maximum intensity. The lighter red lines
represent waveguide segments which have an intermediate intensity profile.
The amplitudes of the wave are marked by the numbers $\pm 1$, $0$, $\pm \alpha$ and
$\pm \beta$ respectively, where $\alpha=-\sqrt{3}/2$ and $\beta=1/2$.
(b) The distribution of wave amplitudes for $ka = \pi/6$ on a third generation
network. The thick blue highlighted lines represent one- and two-step
renormalized lattices in (a) and (b) respectively.}
\label{ampdistribution1}
\end{figure*}
For $n=1$ Eq.~\eqref{roots} reduces to
\begin{equation}
\cos ka(2\cos 2ka-1)=0
\end{equation}
Roots of the above equation are $ka = \pm \pi/2$ and $ka = \pm \pi/6$. $ka = \pi/2$ (or,
equivalently, $-\pi/2$) is of course, the extended mode. The root $ka = \pi/6$
leads to the construction of wave amplitudes as shown in Fig.~\ref{ampdistribution1}.
It is not difficult to extend the construction depicted in
Fig.~\ref{ampdistribution1}(a) even to a network of an arbitrarily large
size, where the {\it end} vertices are not actually visible. We are still
able to satisfy Eq.~\eqref{difference2} locally at every vertex while
drawing this distribution and thus, $k =\pi/6a$ definitely belongs to
the spectrum of the infinite system, a fact that has been cross-checked by
evaluation the LDOS at the $A$- and the $B$-sites at this special value of
$k$. We get a stable, finite value of the LDOS which supports our argument
above.
In Fig.~\ref{ampdistribution1}(a) we show the distribution of amplitudes on
the central cluster of an infinite diamond-Vicsek hierarchical network for
$ka =\pi/6$. The deep red arms connect network vertices where the wave
amplitudes are non-zero, and thus these arms are the brightest looking ones
as far as the distribution of light intensity is concerned. The black lines represent
waveguides which will appear completely dark, as the wave amplitudes at
their vertices will have to be zero in order to satisfy Eq.
\eqref{difference2}. There will be arms connecting one vertex with zero
amplitude and another with a non-zero one. These are depicted by a lighter shed of
red, and will `glow' with less intensity compared to the deep red ones. The
distribution of intensity in any arm (apart from the black ones) is by no
means uniform. The colors just represent the fact that the intensity is non-
zero. The significant observation is that, clusters of non-zero amplitude
span over a finite distance, but ultimately get `decoupled' from each other
on a larger scale of length. This can be appreciated if we look at
Fig.~\ref{ampdistribution1}(b) which is a larger version of the previous
figure. The red shaded clusters are distributed along the principal $X$- and
$Y$-axes, but are {\it separated} from each other beyond a certain extent by
light red boxes. The black clusters representing amplitude-voids are now
seen to span larger spatial distances. A similar construction is possible
for $ka =-\pi/6$ which is another solution of Eq.~\eqref{roots} for $n=1$,
but this does not have any additional significance.
In terms of light, the entire hierarchical geometry will have an appearance
where light will be {\it localized} with higher intensity at certain
clusters of waveguides, decoupled from each other by completely dark
patches.
It is apparent from the above discussion that the eigenfunction
corresponding to $ka = \pm \pi/6$ will be localized in the fractal space.
This is easily re-confirmed by studying the evolution of the hopping
integrals under successive RSRG steps. The hopping integrals $t$ and
$\lambda$ (zero initially, but grows later) remain non-zero at the first
stage of RSRG (that is, $n=1$), indicating that the nearest neighboring
sites on a one step renormalized lattice will have a non-zero overlap of the
wave functions. They start decaying for $n>2$ with the decay in $\lambda(n)$
taking place at a much slower rate compared to $t(n)$. This indicates that
over larger scale of length the corresponding states are {\it localized},
but the effect is a weak one.
\subsection{The staggering effect}
The previous observation immediately leads to an innovative way of exactly
determining the wave vectors (wavelengths) corresponding to localized wave
functions on such a deterministic geometry at an arbitrary scale of length.
We do it using the following method.
We can solve Eq.~\eqref{roots} in principle, to get the desired $k$-values
for any $n$. For example, we have done it explicitly for $n=2$. The roots
are obtained from the equation,
\begin{equation}
\cos ka(2\cos 2ka-1)(\cos 2ka-2\cos 8ka)=0
\end{equation}
and are given in Table~\ref{table}.
\begin{table}
\begin{tabular}{|>{\centering\arraybackslash}m{0.4in}|>{\centering\arraybackslash}m{1.4in}|}
\hline
{\boldmath\(n\)} & {\bf Values of \boldmath\(ka\)}\\
\hline
$1$ & $\pm 1.570796$, $\pm 0.523599$\\
\hline
& $\pm 1.570796$, $\pm 0.523599$, \\
& $\pm 2.617994$, $\pm 1.823731$, \\
$2$ & $\pm 1.317862$, $\pm 2.132603$, \\
& $\pm 1.008990$, $\pm 2.531027$, \\
& $\pm 0.610566$, $\pm 3.008153$, $\pm 0.133440$ \\
\hline
\end{tabular}
\caption{Values of $ka$ obtained from the Eq.~\eqref{roots} for $n=1$ and
$n=2$ respectively.}
\label{table}
\end{table}
In every case, on beginning the RSRG iteration with the wave vector $k$
chosen arbitrarily from the above set, the nearest neighbor overlap integral
$t$, and the diagonal one, viz., $\lambda$ remain non zero {\it at least up
to} that specific $n$-th stage of renormalization. After that, as the RSRG
progress, the hoppings flow to zero with $\lambda$ dominating over $t$ at
every step of renormalization. This implies that for any such $k$-value one
can draw a non-trivial distribution of wave-amplitudes on the renormalized
fractal network. When mapped back on to the original lattice the amplitudes
will be found to span clusters of increasing size. The exact size of the
spanning clusters will be determined by the value of $n$. The size, for
example, with $n=2$ exceeds that for $n=1$.
The spanning clusters finally get decoupled from similar clusters when one
looks at the distribution over a large enough network. Speaking in terms of
the red and grey shaded zones, it should be appreciated that the size of the
red zone is much bigger for $n=2$ compared to the $n=1$ case. We refer the
reader to Ref.~\cite{biplab} for understanding the result.
It is now obvious that higher the value of $n$, more will be the number of roots
of the polynomial equation Eq.~\eqref{roots}. The roots have a nice nesting
property. The roots obtained from any $(n-1)$-th stage
are found included in the solution for the $n$-th stage (see Table~\ref{table}).
The additional
roots obtained at the $n$-th level over the existing roots from the $(n-1)$-
th level keep the overlap integrals non-zero up to that specific $n$-th RSRG
step. Beyond this step, the overlap finally starts to weaken in magnitude.
This immediately implies that for larger values of $n$, the clusters of non-
zero amplitudes span wider fractal space, and the localization effect begins
much later. That is, on-set of localization can be {\it delayed in space}, at
one's will, by choosing to solve Eq.~\eqref{roots} for larger value of $n$.
We can thus bring a staggering effect on the localization of light in such a
fractal waveguide network.
\subsection{Transmission of electromagnetic waves}
To get the two terminal conductance for a finite size diamond-Vicsek
fractal, we attach the system between two semi-infinite one-dimensional
single channel waveguides. The wave equation obeyed by the incident wave in
these waveguides is discretized, and the leads are artificially converted
in to arrays of effective {\it nodes} characterized by a constant on-site
potential $\epsilon_{l} = 2 \cos{ka} + 2 \cot{ka}$ as before, and a nearest
neighbor overlap integral $t_{l} = 1/\sin{ka}$. We then successively renormalize
the finite network to reduce it into an effective two-vertex system, with
renormalized effective on-site term equal to $\mathcal{U}$ and with
an effective hopping integral $\mathcal{T}$ between them. The transmission
coefficient across the effective dimer is given by the well known
formula~\cite{stone},
\begin{equation}
T=\dfrac{4\sin^{2}ka}{\mathcal{D}_{1}^{2}+\mathcal{D}_{2}^{2}}
\end{equation}
where $\mathcal{D}_{1}=\left[(M_{12}-M_{21})+(M_{11}-M_{22})\cos ka
\right]$ and
$\mathcal{D}_{2}=\left[(M_{11}+M_{22})\sin ka \right]$.
The matrix elements $M_{ij}$ are given by,
$M_{11} =\dfrac{(E-\mathcal{U})^{2}}{\mathcal{T}t_{l}}
-\dfrac{\mathcal{T}}{t_{l}},\
M_{12} =-\dfrac{(E-\mathcal{U})}{\mathcal{T}},\
M_{21} =-M_{12},\
M_{22} =-\dfrac{t_{l}}{\mathcal{T}}$
and $\cos ka = (E-\epsilon_{l})/2t_{l}$, `$a$' being the lattice constant
and is taken to be equal to unity throughout the calculation.
\begin{figure}[ht]
\centering
\includegraphics[clip,width=7cm,angle=-90]{trans.eps}
\caption{ (Color online) Transmission characteristics for a $3$rd generation
Vicsek waveguide network. The length of each monomode segment $a$ is chosen
as unity.}
\label{trans}
\end{figure}
In Fig.~\ref{trans}, we have shown the two-terminal transmission
characteristics for a 3rd generation waveguide network system. The
transmission spectrum is, as expected, full of gaps, with resonant
transmission exhibited at certain $k$-values. With increasing generation,
the resonances become rarer and the spectrum become much more fragmented.
\section{Concluding remarks}
\label{sec4}
In conclusion, we have examined the distribution of intensity of light (or
any electromagnetic wave) on a Vicsek geometry consisting of diamond shaped
single channel waveguides. The major result is that we have been able to
identify a countable infinity of localized eigenmodes displaying a multitude
of localization lengths. A prescription is given for an exact determination
of the wave vectors corresponding to all such modes, a problem that is far
from trivial in the case of a deterministically disordered system. The
localized wave functions span the fractal space in clusters of
increasing sizes, the size being precisely controlled by the length scale at
which the wave vector ($k$) is evaluated. The onset of localization can be
exactly predicted from the stage of RSRG and can be delayed (staggered) in
space. The study provides a unique
opportunity to experimentally observe the localization of light triggered by
the lattice topology without bothering about the high dielectric constant
materials. It may also be useful in developing novel photonic band gap
structures, where the wavelengths to be screened or allowed to go through
can be controlled over arbitrarily small domains exploiting the fractal
character of the network.
\begin{acknowledgments}
B. Pal acknowledges the financial support through an INSPIRE Fellowship
from the Department of Science and Technology, India.
\end{acknowledgments}
|
\section{Introduction}
Biharmonic mappings, as a generalization of harmonic ones, are among the
most important mappings in physics; initially appearing from problems of
elasticity theory and fluid mechanics, \cite{Selvadurai}, in the latter
decades, they proved to be useful also in computer graphics, geometry
processing, \cite{Lipman} and radar imaging, \cite{Andersson}. Mathematical
arguments, \cite{Montaldo}, for the use of biharmonic maps include the fact
that harmonic maps do not always exist - and biharmonic maps can "succeed
where harmonic maps have failed"- together with stability issues. On the
other side, Finslerian models seem to gain more and more ground wherever
anisotropy of some kind is involved (and not only), from domains such as:
kinematics, elasticity theory, \cite{Bucataru}, seismic ray theory, \cit
{Anto-Slawinski}, \cite{Yajima}, \cite{Yajima1}, gravity theories, \cite{Li
, \cite{Munteanu}, \cite{Vacaru}, geometrical optics, \cite{Anto},
thermodynamics, statistical mechanics, \cite{Anto}, \cite{Ootsuka}, and up
to biology, \cite{Anto}, \cite{Anto-biology}.
While in Riemannian geometry, biharmonic mappings have been quite
intensively studied (see, for instance, \cite{Balmus}, \cite{Jiang}, \cit
{Montaldo}, \cite{Oniciuc}, \cite{Oniciuc2}), to our knowledge, in Finsler
geometry, only harmonic maps have been considered, \cite{Mo}, \cite{Mo2},
\cite{Mo-book}, \cite{Mosel}, \cite{Nishikawa}, \cite{Shen-Zhang}. Still,
the rich potential of Finslerian geometric models makes us think that such a
study is at least necessary.
As a first step in this direction, we study in this paper biharmonic
mappings having as domain real Finsler spaces $(M,g)$ and as codomain,
Riemannian ones $(\tilde{M},\tilde{g})$ or, briefly, Finsler-to-Riemann
mappings. Our study continues the one made by Xiaohuan Mo and collaborators
\cite{Mo}, \cite{Mo2}, \cite{Mo-book}) for Finsler-to-Riemann harmonic
mappings.
First of all, we extend the concept of bienergy functional for
Finsler-to-Riemann mappings and determine its Euler-Lagrange equations,
i.e., the equations of Finsler-to-Riemann biharmonic maps. This process
points out a generalization of the rough Laplacian from Riemannian geometry.
Any Finsler-to Riemann harmonic map is biharmonic. Just as in Riemannian
geometry, there exist several cases in which the converse is also true; two
notable results in Riemannian geometry, due to Guoying Jiang, \cite{Jiang},
and C. Oniciuc, \cite{Oniciuc}, respectively, can be generalized without
difficulty to our situation:
1) Any biharmonic mapping whose domain is compact and boundaryless and whose
codomain has nonpositive sectional curvature, is harmonic.
2)\ Any biharmonic mapping whose codomain has strictly negative sectional
curvature, obeying the conditions: a) the norm of its tension is constant
and b)\ its rank is greater or equal to 2 at least at one point of its
domain, is harmonic.
Further, we study the biharmonicity of the identity map $id:(M,g)\rightarrow
(M,\tilde{g})$ in two cases of Finsler-to-Riemann transformations of metrics
$g\mapsto \tilde{g},$ thus pointing out examples of nonharmonic biharmonic
maps. The second case, that of linearized perturbations, is inspired from
general relativity; even though we only consider here positive definite
metrics, in our opinion, it is illustrative.
In the last section, we determine the second variation of the bienergy
functional. Except for the facts that each of the expressions of the tension
and of the rough Laplacian gains an extra term and of the use of nonlinear
connections on $TM$, the first and second variation of the bienergy remain
formally similar to their Riemannian counterparts.
In the study of a Finsler space $(M,g)$, there are two major - and
equivalent - approaches: the one based on the tangent bundle $(TM,\pi ,M)$,
via horizontal lifts, and the one based on the pullback bundle $\pi ^{\ast
}TM.$ As noticed in \cite{Nishikawa}, the study of harmonic maps between
real Finsler manifolds is usually carried out on $\pi ^{\ast }TM$ (as in
\cite{Mo}, \cite{Shen-Zhang}) while in the case of complex Finsler
manifolds, it relies on the geometry of $TM.$ In order to obtain a more
unified method, we preferred to work, also in the real case, on $TM$; the
geometric structures we used are the $TM$-correspondents of those in \cit
{Mo}, \cite{Mo2}, \cite{Mo-book}.
\section{Biharmonic maps in Riemannian geometry}
In this section, we present in brief some results in \cite{Montaldo}, \cit
{Jiang}, \cite{Balmus}.
Let $(M,g)$ and $(\tilde{M},\tilde{g})$ be two $\mathcal{C}^{\infty }
-smooth, connected Riemannian manifolds without boundary, of dimensions $n$
and $\tilde{n}$; unless elsewhere specified, we will assume, as in \cit
{Jiang}, that $M$ is compact and orientable. On the two manifolds, we denote
the local coordinates by $(x^{i})_{i=\overline{1,n}},$ $(\tilde{x}^{\alpha
})_{\alpha =\overline{1,\tilde{n}}},$ the Levi-Civita connections by $\nabla
,$ $\tilde{\nabla}$ (with coefficients $\Gamma _{~jk}^{i},$ $\tilde{\Gamma
_{~\beta \gamma }^{\alpha }$) and by $\Gamma (E),$ $\Gamma (\tilde{E}),$ the
modules of $\mathcal{C}^{\infty }$-smooth sections of any vector bundles $E
\tilde{E}$ over $M$ and $\tilde{M}.$ Commas $_{,i}$ and $_{,\alpha }$ will
mean partial differentiation with respect to $x^{i}$ and $\tilde{x}^{\alpha
} $ and $\partial _{i},$ $\tilde{\partial}_{\alpha },$ the natural bases of
the modules $\Gamma (TM)$ and $\Gamma (T\tilde{M})$ respectively.
A $\mathcal{C}^{\infty }$-smooth mapping $\phi :M\rightarrow \tilde{M}$ is
called \textit{harmonic, }if it is a critical point of the \textit{energy
functional}
\begin{equation}
E:\mathcal{C}^{\infty }(M,\tilde{M})\rightarrow \mathbb{R},~~E(\phi )=\dfrac
1}{2}\underset{M}{\int }\left\Vert d\phi \right\Vert ^{2}d\mathcal{V}_{g},
\label{energy_r}
\end{equation
where $d\phi $ is regarded as a section of the bundle $T^{\ast }M\otimes
\phi ^{-1}T\tilde{M},$ $\left\Vert d\phi \right\Vert ^{2}=trace_{g}(\phi
^{\ast }\tilde{g})=g^{ij}\tilde{g}_{\alpha \beta }\phi _{~,i}^{\alpha }\phi
_{~,j}^{\beta }$ is the squared Hilbert-Schmidt norm of $d\phi $ and $
\mathcal{V}_{g}$ is the Riemannian volume element on $M.$
Harmonic maps are solutions of the equation $\tau (\phi )=0,$ where, \cit
{Montaldo}
\begin{equation}
\tau (\phi )=g^{ij}\{\nabla _{\partial _{i}}^{\phi }d\phi (\partial
_{j})-d\phi (\nabla _{\partial _{i}}\partial _{j})\}=:g^{ij}(\nabla
_{\partial _{i}}^{\phi }d\phi )\partial _{j}, \label{tension_r_covar}
\end{equation
is a section of the bundle $\phi ^{-1}T\tilde{M},$ called the \textit
tension }of $\phi $ and $\nabla ^{\phi }$ is the connection induced by
\tilde{\nabla}$ in the pullback bundle $\phi ^{-1}T\tilde{M},$ \cite{Balmus
. In local writing
\begin{equation*}
\tau ^{\alpha }(\phi )=g^{ij}\{\phi _{~,ij}^{\alpha }+\tilde{\Gamma}_{~\beta
\gamma }^{\alpha }\phi _{~,i}^{\beta }\phi _{~,j}^{\gamma }-\Gamma
_{~ij}^{k}\phi _{~,k}^{\alpha }\}.
\end{equation*}
The above notion of harmonicity generalizes the usual one for mappings
between Euclidean spaces; notable examples\ include geodesic curves and
minimal Riemannian immersions.
\bigskip
\textit{Biharmonic maps} $\phi \in \mathcal{C}^{\infty }(M,\tilde{M})$ are
defined as critical points of the \textit{bienergy functional:
\begin{equation}
E_{2}(\phi )=\dfrac{1}{2}\underset{M}{\int }\left\langle \tau (\phi ),\tau
(\phi )\right\rangle d\mathcal{V}_{g}; \label{bienergy_r}
\end{equation
here $\left\langle ~~,~~\right\rangle $ denotes the scalar product on the
fibers of $T\tilde{M},$ determined by $\tilde{g}.$ The Euler-Lagrange
equation attached to the bienergy is, \cite{Montaldo}
\begin{equation}
\tau _{2}(\phi )=0, \label{biharmonic_eqn_R}
\end{equation
where the \textit{bitension }$\tau _{2}(\phi )$ of $\phi $ is the section of
the bundle $\phi ^{-1}T\tilde{M}$ given by:
\begin{equation}
\tau _{2}(\phi )=-\Delta ^{\phi }\tau (\phi )-trace_{g}(R^{\tilde{\nabla
}(d\phi ,\tau (\phi ))d\phi ) \label{bitension_r}
\end{equation
and the operator $\Delta ^{\phi }=-g^{ij}(\nabla _{\partial _{i}}^{\phi
}\nabla _{\partial _{j}}^{\phi }-\nabla _{\nabla _{\partial _{i}}\partial
_{j}}^{\phi })$ is the rough Laplacian, acting on sections of $\phi ^{-1}
\tilde{M}$.
\textbf{Remarks: }1) Equation (\ref{biharmonic_eqn_R}) is the Riemannian
generalization of the biharmonic equation in Euclidean spaces, \cit
{Selvadurai}.
2)\ Any harmonic map $\phi :M\rightarrow \tilde{M}$ is biharmonic.
\section{Finsler structures}
In the following, except for the metric structure on $M$ (and related
quantities) we preserve the notations and conventions in Section 2. We
denote by $TM$ and $T\tilde{M}$ the tangent bundles of the manifolds $M$ and
$\tilde{M}$ and their local coordinates, by $(x,y):=(x^{i},y^{i}),$ $(\tilde
x},\tilde{y}):=(\tilde{x}^{\alpha },\tilde{y}^{\alpha });$ dots $_{\cdot i}$
and $_{\cdot \alpha }$ will mean partial differentiation with respect to
y^{i}$ and $\tilde{y}^{\alpha }.$
\bigskip
\textbf{A. Metric structure: }A \textit{Finsler structure}, \cite{Shen},
\cite{Mo-book}, on the manifold $M$ is a function $F:TM\rightarrow \mathbb{R}
$ with the properties:
1) $F(x,y)\ $is $\mathcal{C}^{\infty }$-smooth for $y\not=0$ and continuous
at $y=0.$
2)$\ F(x,\lambda y)=\lambda F(x,y),$ $\forall \lambda >0$;
3)\ The \textit{Finslerian metric tensor}
\begin{equation}
g_{ij}(x,y):=\dfrac{1}{2}(F^{2}(x,y))_{\cdot ij} \label{metric_tensor_F}
\end{equation
is positive definite.
The arc length of a curve $c$ on the Finsler space $(M,g)$ is given, \cit
{Shen}, by
\begin{equation}
l(c)=~\underset{c}{\int }F(x,dx). \label{arc_length}
\end{equation
Condition 2) above insures the independence of $l(c)$ of the chosen
parametrization of $c.$
\bigskip
\textbf{B. Nonlinear connection and adapted frame on }$TM:$ Ehresmann (or
\textit{nonlinear, }\cite{Shen}) connections $TTM=HTM\oplus VTM,$ with
VTM=Span(\dfrac{\partial }{\partial y^{i}})$ help simplify computations in
Finsler geometry and obtain geometric objects with simple transformation
rules. A typical choice is the \textit{Cartan nonlinear connection}, built
as follows, \cite{Shen}.
Geodesics of the Finsler space $(M,g)$ are defined as critical points $c$ of
the arc length (\ref{arc_length}); in the natural parametrization, their
equations are
\begin{equation}
\dfrac{dy^{i}}{ds}+2G^{i}(x,y)=0,~\ y=\dot{x}; \label{spray_coeff}
\end{equation
with $2G^{i}(x,y)=\dfrac{1}{2}g^{ih}\left( (F^{2})_{\cdot
h,k}y^{k}-(F^{2})_{,h}\right) ;$ this defines the local coefficients
G_{~j}^{i}=G_{~j}^{i}(x,y)$ of the Cartan nonlinear connection as:
\begin{equation}
G_{~j}^{i}:=G_{~\cdot j}^{i}. \label{Cartan_conn}
\end{equation
The Cartan nonlinear connection gives rise to the adapted basis:
\begin{equation}
(\delta _{i}=\dfrac{\partial }{\partial x^{i}}-G_{~i}^{j}(x,y)\dfrac
\partial }{\partial y^{j}},~~\ \ \dot{\partial}_{i}=\dfrac{\partial }
\partial y^{i}}) \label{general_adapted_basis}
\end{equation
on $\Gamma (TTM)$ and to its dual $(dx^{i},~\ \delta
y^{i}=dy^{i}+G_{~j}^{i}dx^{j}).$
With respect to coordinate transformations on $TM,$ induced by coordinate
transformations $x^{i^{\prime }}=x^{i^{\prime }}(x)$ on $M,$ the elements of
the adapted basis/cobasis transform by the same rules as vector/covector
fields on $M$, \cite{Lagrange}.
Any vector field $X$ on $TM$ can be decomposed as: $X=hX+vX,$ $\
hX=X^{i}\delta _{i},$ $vX=\hat{X}^{i}\dot{\partial}_{i};$ its \textit
horizontal component} $hX$ and its \textit{vertical component} $vX$ are
vector fields on $TM$. This leads to a simple rule of transformation for
X^{i},\hat{X}^{i}.$ A similar situation holds for 1-forms $\omega =h\omega
+v\omega ,$ $h\omega =\omega _{i}dx^{i},$ $v\omega =\hat{\omega}_{i}\delta
y^{i}$, \cite{Lagrange}) and more generally, for tensors of any rank on $TM$.
Using Cartan nonlinear connection, tangent vector fields to lifts $c^{\prime
}:=(c,\dot{c})$ to $TM$ of unit speed geodesics of $M$ are always
horizontal, \cite{Shen}, \cite{Lagrange}.
\bigskip
The adapted basis $\{\delta _{i},\dot{\partial}_{i}\}$ is generally
nonholonomic, the Lie brackets of its elements are
\begin{equation*}
\lbrack \delta _{j},\delta _{k}]=R_{~jk}^{i}(x,y)\dot{\partial
_{i},~~~~[\delta _{j},\dot{\partial}_{k}]=G_{~jk}^{i}(x,y)\dot{\partial
_{i},~\ \ [\dot{\partial}_{j},\dot{\partial}_{k}]=0;
\end{equation*
where:
\begin{equation}
R_{~jk}^{i}(x,y)=\delta _{k}G_{~j}^{i}-\delta _{j}G_{~k}^{i},~\ \ \
G_{~jk}^{i}(x,y):=G_{~\cdot j\cdot k}^{i}(x,y). \label{Lie_bracket_comps}
\end{equation}
\bigskip
\textbf{C. }As \textbf{covariant differentiation }rule on $TM,$ we will use
the one given by the \textit{Chern-Rund affine connection} $D$ on $TM,$ \cit
{Bucataru}, locally described by
\begin{equation}
D_{\delta _{k}}\delta _{j}=~\Gamma _{~jk}^{i}\delta _{i},~\ D_{\delta _{k}
\dot{\partial}_{j}=\Gamma _{~jk}^{i}\dot{\partial}_{i},~\ D_{\dot{\partial
_{k}}\delta _{j}=~D_{\dot{\partial}_{k}}\dot{\partial}_{j}=0,
\label{chern_conn}
\end{equation
where~\ $\Gamma _{~jk}^{i}=\dfrac{1}{2}g^{ih}(\delta _{k}g_{hj}+\delta
_{j}g_{hk}-\delta _{h}g_{jk})$ are the "adapted" Christoffel symbol
\footnote
In the Finslerian case, the usual Christoffel symbols $\gamma _{~jk}^{i}
\dfrac{1}{2}g^{ih}(g_{hj,k}+g_{hk,j}-g_{jk,h})$ do \textit{not }generally
represent the coefficients of an affine connection on $TM.$} of $g.$ The
Chern-Rund connection preserves by parallelism the horizontal and vertical
distributions on $TTM$, i.e.
\begin{equation*}
D_{X}(hY)=h~D_{X}Y,~~~D_{X}(vY)=v~D_{X}Y;
\end{equation*
it is generally, only $h$-metrical:
\begin{equation*}
D_{hX}g=0,~\forall X\in \Gamma (TTM).
\end{equation*}
The Chern-Rund connection $D$ has nontrivial torsion
\begin{equation}
T=R_{~jk}^{i}\dot{\partial}_{i}\otimes dx^{k}\otimes dx^{j}+P_{~jk}^{i}\dot
\partial}_{i}\otimes \delta y^{k}\otimes dx^{j}, \label{torsion}
\end{equation
with $R_{~jk}^{i}$ as in (\ref{Lie_bracket_comps})\ and
P_{~jk}^{i}=G_{~jk}^{i}-\Gamma _{~jk}^{i};$ the latter defines a horizontal
1-form
\begin{equation}
P=P_{i}dx^{i},~\ P_{i}:=P_{~ij}^{j}, \label{torsion_comps}
\end{equation
which will be used in the following. We notice that the torsion of $D$ has
only vertical components
\begin{equation}
hT(X,Y)=0,~\ \forall X,Y\in \Gamma (TTM). \label{torsion_property}
\end{equation}
The curvature $R$ of $D$ is locally described by
\begin{eqnarray}
R &=&R_{j~kl}^{~i}dx^{l}\otimes dx^{k}\otimes dx^{j}\otimes \delta
_{i}+R_{j~kl}^{~i}dx^{l}\otimes dx^{k}\otimes \delta y^{j}\otimes \dot
\partial}_{i}+; \label{curvature_comps} \\
&&+P_{j~kl}^{~i}\delta y^{l}\otimes dx^{k}\otimes dx^{j}\otimes \delta
_{i}+P_{j~kl}^{~i}\delta y^{l}\otimes dx^{k}\otimes \delta y^{j}\otimes \dot
\partial}_{i}, \notag
\end{eqnarray
where $R_{j~kl}^{~i}=\delta _{l}\Gamma _{~jk}^{i}-\delta _{k}\Gamma
_{~jl}^{i}+\Gamma _{~jk}^{h}\Gamma _{~hl}^{i}-\Gamma _{~jl}^{h}\Gamma
_{~hk}^{i}$ and$\ P_{j~kl}^{~i}=\Gamma _{~jk\cdot l}^{i}.$
\bigskip
We consider the above notions also for the Riemannian manifold $(\tilde{M}
\tilde{g})$ and designate them by tildes. In this case: $\tilde{\Gamma
_{~\beta \gamma }^{\alpha }=\tilde{G}_{~\beta \gamma }^{\alpha }=\tilde
\gamma}_{~\beta \gamma }^{\alpha }$, that is:
\begin{equation}
(\tilde{\nabla}_{X}Y)^{\tilde{h}}=\tilde{D}_{X^{\tilde{h}}}Y^{\tilde{h}},~\
\forall X,Y\in \Gamma (T\tilde{M}), \label{Levi-Civita_lift}
\end{equation
where the superscript $^{\tilde{h}}$ indicates the horizontal of vector
fields from $\tilde{M}$ to $T\tilde{M};$\ also, $\tilde{G}_{~\beta }^{\alpha
}(\tilde{x},\tilde{y})=\tilde{\gamma}_{~\beta \gamma }^{\alpha }(\tilde{x}
\tilde{y}^{\gamma },$ $\tilde{P}_{~\beta \gamma }^{\alpha }=0;$ the
Chern-Rund connection becomes "fully"\ metrical
\begin{equation}
\tilde{D}_{X}\tilde{g}=0,~\forall X\in \Gamma (TT\tilde{M}).
\label{full_metricity}
\end{equation
The only nonzero local component of curvature tensor $\tilde{R}$ is $\tilde{
}_{\beta ~\gamma \delta }^{~\alpha },$ i.e.
\begin{equation}
\tilde{R}(X,Y)Z=\tilde{R}(\tilde{h}X,\tilde{h}Y)Z,~\ \ \ \ \ \forall
X,Y,Z\in \Gamma (TT\tilde{M}); \label{Riem_R}
\end{equation
$\tilde{R}_{\beta ~\gamma \delta }^{~\alpha }$ coincide with the components
of the curvature $R^{\tilde{\nabla}}$ of the Levi-Civita connection of
\tilde{g}$ and are thus subject to the same symmetries; Bianchi identities,
\cite{Shen}, also acquire the same form as those of $R^{\tilde{\nabla}}.$
Ricci identities of $\tilde{D}$ take the local form
\begin{equation}
\begin{array}{l}
\tilde{D}_{\delta _{\rho }}\tilde{D}_{\delta _{\gamma }}Z^{\alpha }-\tilde{D
_{\delta _{\gamma }}\tilde{D}_{\delta _{\rho }}Z^{\alpha }=\tilde{R}_{\beta
~\gamma \rho }^{~\alpha }Z^{\beta }+\tilde{R}_{~\gamma \rho }^{\beta
}Z_{~\cdot \beta }^{\alpha }; \\
\ \tilde{D}_{\dot{\partial}_{\rho }}\tilde{D}_{\delta _{\gamma }}Z^{\alpha }
\tilde{D}_{\delta _{\gamma }}\tilde{D}_{\dot{\partial}_{\rho }}Z^{\alpha
}=0,~\ \ ~\ \ \ \ \forall Z=Z^{\alpha }\delta _{\alpha }\in \Gamma (HT\tilde
M}).\
\end{array}
\label{Ricci_R}
\end{equation}
Another useful property of $\tilde{D}$ is
\begin{equation}
\tilde{D}_{\tilde{\delta}_{\beta }}\tilde{y}^{\alpha }=0. \label{deflection}
\end{equation}
The Riemannian metric $\tilde{g}$ gives rise to a scalar product on the
fibers on $HT\tilde{M},$ which will be denoted by $\left\langle
~~,~~\right\rangle :
\begin{equation}
\left\langle X,Y\right\rangle =\tilde{g}_{\alpha \beta }X^{\alpha }Y^{\beta
},~\ \ ~\ \forall X=X^{\alpha }\tilde{\delta}_{\alpha },Y=Y^{\alpha }\tilde
\delta}_{\alpha }\in \Gamma (HT\tilde{M}). \label{scalar_product}
\end{equation}
\bigskip
\textbf{D. Volume form and integration domain. }Consider the Riemannian
volume element $d\mathcal{V}_{g}=\sqrt{\det G}dx\wedge \delta y=\det
gdx\wedge \delta y$ on $TM,$ determined by the \textit{Sasaki lift, }\cit
{Shen}, $G:=g_{ij}dx^{i}\otimes dx^{j}+g_{ij}\delta y^{i}\otimes \delta
y^{j} $ of $g$ to $TM$. For an $x,y$-dependent function $\alpha
:TM\rightarrow \mathbb{R},$ we will consider its integral on the total space
of the unit ball bundle of $M$ (which is compact, since $g$ is positive
definite), divided by the volume $Vol\mathbb{B}^{n}$ of the unit ball in the
Euclidean space $\mathbb{R}^{n}$, i.e.
\begin{equation}
\underset{BM}{\int }\alpha (x,y)d\mathcal{V}_{g}=\dfrac{1}{Vol\mathbb{B}^{n}
\underset{M}{\int }(\underset{B_{x}}{\int }\alpha (x,y)\det g(x,y)dy)dx,
\end{equation
where $B_{x}=\left\{ y\in T_{x}M~|~g_{ij}y^{i}y^{j}\leq 1\right\} $ (on $M,$
this construction provides a generalization of the Riemannian volume
element, called the \textit{Holmes-Thompson volume element, }\cite{Shen1}).
The divergence $divX=\dfrac{1}{\det g}\delta _{i}(X^{i}\det
g)-G_{~ij}^{j}X^{i},$ of a horizontal vector field $X=X^{i}\delta _{i}$ on
TM,$ \cite{Zhong}, can be expressed in terms of Chern-Rund covariant
derivatives as
\begin{equation*}
divX=~D_{\delta _{i}}X^{i}-P(X);
\end{equation*
we will also use this relation in the form
\begin{equation}
g^{ij}\delta _{i}X_{j}=divX+g^{ij}\Gamma _{~ij}^{k}X_{k}+P_{i}X^{i}.
\label{divergence}
\end{equation}
\section{Some remarks on Finsler-to-Riemann maps}
Let $\phi :M\rightarrow \tilde{M},$ $(x^{i})\mapsto (\phi ^{\alpha }(x^{i}))$
be $\mathcal{C}^{\infty }$-smooth. Between the tangent bundles $TM$ and $
\tilde{M}$, it acts the differential $\Phi :=d\phi $ (regarded as a mapping
between manifolds); throughout this section, we will use alternatively the
two notations $\Phi $ and $d\phi .$ The connection $\tilde{D}$ determines a
connection $D^{d\phi }$ in the pullback bundle $d\phi ^{-1}(TT\tilde{M})$:
\begin{equation}
D_{X}^{d\phi }(\Phi ^{\ast }Y):=\tilde{D}_{\Phi _{\ast }X}Y,~~~~\forall X\in
\Gamma (TTM),~Y\in \Gamma (TT\tilde{M}); \label{pullback_conn_TM}
\end{equation
with $(\Phi ^{\ast }Y)_{(x,y)}=Y_{\Phi (x,y)}$; hence
\begin{equation}
D_{X}^{d\phi }(\tilde{h}Z)=\tilde{h}D_{X}^{d\phi }Z,~\ \ D_{X}^{d\phi }
\tilde{v}Z)=\tilde{v}D_{X}^{d\phi }Z,~\ \ \ \forall Z\in \Gamma (d\phi
^{-1}(TT\tilde{M})). \label{d-connection}
\end{equation}
The mapping $\Phi $ is locally described by:
\begin{equation*}
\Phi :~\ \ \tilde{x}^{\alpha }=\phi ^{\alpha }(x),~\ \tilde{y}^{\alpha
}=\phi _{~,j}^{\alpha }(x)y^{j}.
\end{equation*
With $\phi ^{\alpha ^{\prime }}(x,y):=\phi _{~,j}^{\alpha }(x)y^{j},$ we
have, for $X\in \Gamma (HTM):$ $\Phi _{\ast }X=X(\phi ^{\alpha })\partial
_{\alpha }+X(\phi ^{\alpha ^{\prime }})\dot{\partial}_{\alpha }.$ Taking
X=X^{i}\delta _{i}$ and expressing $\Phi _{\ast }X$ in the adapted bases
\begin{equation}
\Phi _{\ast }X=X^{i}\phi _{~,i}^{\alpha }\tilde{\delta}_{\alpha }+{\large (
X^{i}\delta _{i}\phi ^{\alpha ^{\prime }}+\tilde{N}_{~\beta }^{\alpha
}X(\phi ^{\beta }){\large )}\dot{\partial}_{\alpha }. \label{image_Phi}
\end{equation
The horizontal component $\tilde{h}\Phi _{\ast }X$ will have a peculiar
importance.
\begin{lemma}
For any horizontal vector field $X=X^{i}\delta _{i}$ on $TM:
\begin{equation}
\tilde{h}\Phi _{\ast }X=X^{i}\phi _{~,i}^{\alpha }\delta _{\alpha }=:d\phi ^
\tilde{h}}(X), \label{notation_horiz_comp_differential}
\end{equation
where $d\phi ^{\tilde{h}}:=\phi _{\,,i}^{\alpha }dx^{i}\otimes \delta
_{\alpha }$ is the horizontal lift of the vector-valued 1-form $d\phi $.
\end{lemma}
Consider a 1-parameter variation $f:I_{\varepsilon }\times
M,~~f=f(\varepsilon ,x),$ $f(0,x)=\phi (x)$ of $\phi $ and
\begin{equation*}
F:=df:T(I_{\varepsilon }\times M)\rightarrow T\tilde{M}.
\end{equation*}
On $T(I_{\varepsilon }\times M),$ the local coordinates are $(\varepsilon
,x,\varepsilon ^{\prime },y).$ Taking on the interval $I_{\varepsilon
}\subset \mathbb{R},$ the Euclidean metric and the product Finsler metric on
$I_{\varepsilon }\times M,$ we will obtain a trivial prolongation of the
Cartan nonlinear connection to this new manifold, which produces the adapted
basis $\{\partial _{\varepsilon },\delta _{i},\partial _{\varepsilon
^{\prime }},\dot{\partial}_{i}\}$ and a trivial prolongation of the Chern
connection $D$ (which we will denote again by $D$). i.e., $D_{\partial
_{\varepsilon }}\delta _{i}=D_{\partial _{\varepsilon }}\dot{\partial
_{i}=0,~\ D_{\delta _{i}}\partial _{\varepsilon }=D_{\dot{\partial
_{i}}\partial _{\varepsilon }=0$ etc. We also notice that $[\partial
_{\varepsilon },\delta _{i}]=0$ and $[\partial _{\varepsilon },\dot{\partial
_{i}]=0.$
The connection $D^{df}$ will be prolonged to $df^{-1}(TT\tilde{M}),$ by:
\begin{equation}
D_{\partial _{\varepsilon }}^{df}(F^{\ast }X):=~\tilde{D}_{F_{\ast }\partial
_{\varepsilon }}X,~\ \ X\in \Gamma (TT\tilde{M}). \label{pullback_conn}
\end{equation}
\begin{lemma}
For any $X,Y\in \Gamma (TTM)$, $Z\in \Gamma (df^{-1}(TT\tilde{M})),$
p=(x,y)\in TM$:
\begin{equation}
D_{\partial _{\varepsilon }}^{df}(df^{\tilde{h}}(X))=D_{X}^{df}(df^{\tilde{h
}(\partial _{\varepsilon })). \label{commutation_rule_2}
\end{equation}
\end{lemma}
\begin{proof}
(\ref{torsion_property}) says that $0=$ $\tilde{h}\tilde{T}(F_{\ast
}\partial _{\varepsilon },F_{\ast }X)=$ $h\{D_{\partial _{\varepsilon
}}^{df}(F_{\ast }X)-D_{X}^{df}(F_{\ast }\partial _{\varepsilon })-[F_{\ast
}X,F_{\ast }\partial _{\varepsilon }]\}.$ The result follows then from (\re
{d-connection})\ and $[X,\partial _{\varepsilon }]=0$.
\end{proof}
\section{Bienergy and its first variation}
The energy of a Finsler-to-Riemann mapping $\phi :M\rightarrow \tilde{M},$
\cite{Mo}, is $E(\phi )=\dfrac{1}{2}\underset{BM}{\int }g^{ij}\tilde{g
_{\alpha \beta }\phi _{~,i}^{\alpha }\phi _{~,j}^{\beta }d\mathcal{V}_{g},$
(note that $\phi =\phi (x),$ hence $\phi _{~,i}^{\alpha }=\delta _{i}\phi
^{\alpha }$), or, in our language:
\begin{equation*}
E(\phi )=\dfrac{1}{2}\underset{BM}{\int }g^{ij}\left\langle d\phi ^{\tilde{h
}(\delta _{i}),d\phi ^{\tilde{h}}(\delta _{i})\right\rangle d\mathcal{V}_{g},
\end{equation*
with $\left\langle ~~,~~\right\rangle $ as in (\ref{scalar_product}). The
\textit{tension} of $\phi :M\rightarrow \tilde{M}$, \cite{Mo}, can be
regarded as a section of $(d\phi )^{-1}HT\tilde{M}$
\begin{equation}
\tau (\phi )=g^{ij}\{D_{\delta _{i}}^{d\phi }d\phi ^{\tilde{h}}(\delta
_{j})-d\phi ^{\tilde{h}}(D_{\delta _{i}}\delta _{j})-P_{i}d\phi ^{\tilde{h
}(\delta _{j})\}. \label{tension_FR}
\end{equation
The mapping $\phi $ is \textit{harmonic} iff its tension vanishes
identically.
\bigskip
It appears as natural to define the \textit{bienergy }of a mapping $\phi
:M\rightarrow \tilde{M}$ as
\begin{equation}
E_{2}(\phi )=\dfrac{1}{2}\underset{BM}{\int }\left\langle \tau (\phi ),\tau
(\phi )\right\rangle d\mathcal{V}_{g}. \label{bienergy}
\end{equation}
Accordingly, by a Finsler-to-Riemann \textit{biharmonic map} we will mean a
critical point of the bienergy (\ref{bienergy}).
In order to determine the critical points of $E_{2}$, we take variations
f=f(\varepsilon ,x)$ of $\phi $ as above and denote b
\begin{equation}
\mathbf{V}:=(df(\partial _{\varepsilon }))^{\tilde{h}}=df^{\tilde{h
}(\partial _{\varepsilon }^{~\tilde{h}}),~\ \ \ \ \ \ V:=\mathbf{V
_{|\varepsilon =0.} \label{variation_vector_field}
\end{equation
the horizontal lift of the associated deviation vector field $df(\partial
_{\varepsilon })$.
Since the Chern-Rund connection on the codomain $T\tilde{M}$ is metrical
\begin{equation*}
\dfrac{dE_{2}}{d\varepsilon }(f)=\dfrac{1}{2}\dfrac{d}{d\varepsilon
\underset{BM}{\int }\left\langle \tau (f),\tau (f)\right\rangle d\mathcal{V
_{g}=\underset{BM}{\int }\left\langle D_{\partial _{\varepsilon }}^{df}\tau
(f),\tau (f)\right\rangle d\mathcal{V}_{g}.
\end{equation*}
Let us evaluate the term $D_{\partial _{\varepsilon }}^{df}\tau (f):$
\begin{eqnarray*}
&&D_{\partial _{\varepsilon }}^{df}\tau (f)=D_{\partial _{\varepsilon
}}^{df}\{g^{ij}(D_{\delta _{i}}^{df}(df^{\tilde{h}}(\delta _{j}))-df^{\tilde
h}}(D_{\delta _{i}}\delta _{j})-P_{i}df^{\tilde{h}}(\delta _{j}))\}= \\
&=&g^{ij}\left\{ D_{\partial _{\varepsilon }}^{df}D_{\delta _{i}}^{df}(df^
\tilde{h}}(\delta _{j}))-D_{\partial _{\varepsilon }}^{df}(df^{\tilde{h
}(D_{\delta _{i}}\delta _{j}))-D_{\partial _{\varepsilon }}^{df}(P_{i}df^
\tilde{h}}(\delta _{j}))\right\}
\end{eqnarray*
($g^{ij}$ can be taken in front of the $\partial _{\varepsilon }
-derivative, since in $g^{ij}=g^{ij}(x,y)$ the coordinates $x,y$ do not
depend on $\varepsilon $). Commuting derivatives by means of the curvature
tensor of $\tilde{D},$ taking (\ref{Riem_R}) and $[\delta _{i},\partial
_{\varepsilon }]=0$ into account,
\begin{equation*}
D_{\partial _{\varepsilon }}^{df}D_{\delta _{i}}^{df}(df^{\tilde{h}}(\delta
_{j}))=\tilde{R}(\mathbf{V},df^{\tilde{h}}(\delta _{i}))df^{\tilde{h
}(\delta _{j})+D_{\delta _{i}}^{df}D_{\partial _{\varepsilon }}^{df}(df^
\tilde{h}}(\delta _{j})).
\end{equation*
By (\ref{commutation_rule_2}), the term $D_{\delta _{i}}^{df}D_{\partial
_{\varepsilon }}^{df}(df^{\tilde{h}}(\delta _{j}))$ becomes $D_{\delta
_{i}}^{df}D_{\delta _{j}}^{df}\mathbf{V.}$ Using (\ref{commutation_rule_2})
also in the expression $D_{\partial _{\varepsilon }}^{df}(df^{\tilde{h
}(D_{\delta _{i}}\delta _{j}))-D_{\partial _{\varepsilon }}^{df}(P_{i}df^
\tilde{h}}(\delta _{j}))$ and summing up, we get
\begin{eqnarray}
D_{\partial _{\varepsilon }}^{df}\tau (f) &=&g^{ij}\{\tilde{R}(\mathbf{V
,df^{\tilde{h}}(\delta _{i}))df^{\tilde{h}}(\delta _{j})+D_{\delta
_{i}}^{df}D_{\delta _{j}}^{df}\mathbf{V}\mathcal{-} \label{derivative_tau}
\\
&&-D_{D_{\delta _{i}}\delta _{j}}^{df}\mathbf{V}-P_{i}D_{\delta _{j}}^{df
\mathbf{V}\}. \notag
\end{eqnarray}
We notice the operator
\begin{equation}
g^{ij}(-D_{\delta _{i}}^{df}D_{\delta _{j}}^{df}+D_{D_{\delta _{i}}\delta
_{j}}^{df}+P_{i}D_{\delta _{j}}^{df})=:\Delta ^{df},~\mathcal{J}=-\Delta
^{df}-trace_{g}\tilde{R}(df^{\tilde{h}},~\cdot ~)df^{\tilde{h}},
\label{rough_Laplacian}
\end{equation
acting on sections of the bundle $(df)^{-1}(HT\tilde{M})$. With this, we
have:
\begin{equation}
D_{\partial _{\varepsilon }}^{df}\tau (f)=-\Delta ^{df}\mathbf{V}-g^{ij
\tilde{R}(df^{\tilde{h}}(\delta _{i}),\mathbf{V})df^{\tilde{h}}(\delta _{j})
\mathcal{J}(\mathbf{V}). \label{deriv_tau_1}
\end{equation
Evaluating at $\varepsilon =0$ and substituting into the expression of the
variation,
\begin{equation}
\dfrac{dE_{2}}{d\varepsilon }(f)|_{\varepsilon =0}=\underset{BM}{\int
\left\langle \mathcal{J}(V),\tau (\phi )\right\rangle d\mathcal{V}_{g}.
\label{first_variation_interm}
\end{equation
It remains to transform the above expression so as to have $V$ in the right
hand side of the scalar product. This will be easy using the following lemma.
\begin{lemma}
\label{Lemma1} The operators $\Delta ^{d\phi }$ and $\mathcal{J}$ are
self-adjoint
\begin{equation}
\underset{BM}{\int }\left\langle \Delta ^{d\phi }X,Y\right\rangle d\mathcal{
}_{g}=\underset{BM}{\int }\left\langle X,\Delta ^{d\phi }Y\right\rangle
\mathcal{V}_{g},\ \underset{BM}{\int }\left\langle \mathcal{J
X,Y\right\rangle d\mathcal{V}_{g}=\underset{BM}{\int }\left\langle X
\mathcal{J}Y\right\rangle d\mathcal{V}_{g}. \label{self-adj}
\end{equation
for any $X,Y\in \Gamma (d\phi ^{-1}(HT\tilde{M})).$
\end{lemma}
\begin{proof}
We start from the left hand side of the first relation (\ref{self-adj});
integrating by parts the term $\underset{BM}{\int }\left\langle
-g^{ij}D_{\delta _{i}}^{df}D_{\delta _{j}}^{df}X,Y\right\rangle d\mathcal{V
_{g}$ and applying (\ref{divergence}), we get
\begin{equation}
\underset{BM}{\int }\left\langle \Delta ^{d\phi }X,Y\right\rangle d\mathcal{
}_{g}=-\underset{BM}{\int }g^{ij}\left\langle D_{~\delta _{i}}^{d\phi
}X,~D_{~\delta _{j}}^{d\phi }Y\right\rangle d\mathcal{V}_{g}.
\label{adjoint1}
\end{equation
Integrating once again by parts, we obtain (\ref{self-adj}). The
self-adjointness of $\mathcal{J}$ follows then from the symmetries of
\tilde{R}.$
\end{proof}
The operator $\Delta ^{d\phi }$ is a generalization of the rough Laplacian
from Riemannian geometry, built in the same spirit as the horizontal
Laplacian acting on differential forms in \cite{Zhong}, \cite{Zhong1}.
Using Lemma \ref{Lemma1} in (\ref{first_variation_interm}), we get:
\begin{proposition}
a) The first variation of the bienergy of a mapping $\phi :M\rightarrow
\tilde{M}$ from the Finsler space $(M,g)$ to the Riemann space $(\tilde{M}
\tilde{g})$ is
\begin{equation}
\dfrac{dE_{2}(f)}{d\varepsilon }|_{\varepsilon =0}=\underset{BM}{\int
\left\langle -\Delta ^{d\phi }\tau (\phi )-trace_{g}\tilde{R}(d\phi ^{\tilde
h}},\tau (\phi ))d\phi ^{\tilde{h}},V\right\rangle d\mathcal{V}_{g};
\label{first_var_bienergy}
\end{equation}
b) The mapping $\phi $ is biharmonic iff:
\begin{equation}
\tau _{2}(\phi ):=-\Delta ^{d\phi }\tau (\phi )-trace_{g}\tilde{R}(d\phi ^
\tilde{h}},\tau (\phi ))d\phi ^{\tilde{h}}=0.
\label{biharmonic_Finsler_to_Riemann}
\end{equation}
\end{proposition}
\textbf{Remarks. }1) In the above, we considered, as in \cite{Jiang}, that
M $ is compact and without boundary. Elsewhere, all the discussion can be
made on a compact subset $\mathcal{D}$ of $M$; in this case, we assume that,
on the boundary of $\mathcal{D}$, the vector field $V$ and the covariant
derivatives $D_{\delta _{i}}^{d\phi }V$ vanish.
2) Any harmonic map from a Finsler space to a Riemann one is biharmonic,
namely, a minimum point for the bienergy functional. A biharmonic map which
is not harmonic will be called \textit{proper biharmonic.}
\textbf{Particular cases:}
1)\ If $\tilde{M}=\mathbb{R}^{n}$ with the Euclidean metric, then the
biharmonic equation (\ref{biharmonic_Finsler_to_Riemann}) becomes
\begin{equation*}
\Delta ^{d\phi }\tau (\phi )=0.
\end{equation*}
2)\ If $\tilde{M}=\mathbb{S}^{n}$ is the unit Euclidean sphere, then, using
the expression of the Riemann tensor of a space form, we get that $\phi
:M\rightarrow \mathbb{S}^{n}$ is biharmonic iff
\begin{equation*}
\Delta ^{d\phi }\tau (\phi )+2e(\phi )\tau (\phi )-trace_{g}\left\langle
d\phi ^{\tilde{h}},\tau (\phi )\right\rangle d\phi ^{\tilde{h}}=0,
\end{equation*
where $e(\phi )=\dfrac{1}{2}trace_{g}\left\langle d\phi ^{\tilde{h}},d\phi ^
\tilde{h}}\right\rangle $ is the energy density of $\phi .$ The result is
similar to the one in the Riemannian case, \cite{Balmus}.
3)\ If $M$ is a \textit{weakly Landsberg }manifold, i.e., \cite{Mo}, if
P=0, $ then the expressions of the tension and of the rough Laplacian become
formally similar to the ones in the Riemannian case - just, depending on the
fiber coordinates $y^{i}$: $\tau (\phi )=trace_{g}D^{d\phi }(d\phi ^{\tilde{
}}),~\ \ \Delta ^{df}=g^{ij}(-D_{\delta _{i}}^{df}D_{\delta
_{j}}^{df}+D_{D_{\delta _{i}}\delta _{j}}^{df}).$
\section{Existence of proper biharmonic maps}
The following two results represent generalizations to Finsler-to-Riemann
maps of two theorems in \cite{Jiang} and \cite{Oniciuc} respectively.
\begin{theorem}
If $(M,g)$ is a compact Finslerian manifold without boundary and $(\tilde{M}
\tilde{g})$ is Riemannian with nonpositive sectional curvature, then any
biharmonic map $\phi :M\rightarrow \tilde{M}$ is harmonic.
\end{theorem}
\begin{proof}
The proof follows similar steps to the one in the Riemannian case, \cit
{Jiang}. We apply the horizontal Laplace-Beltrami operator, \cite{Zhong},
\Delta f:=-div(grad_{h}~f),$ where $grad_{h}f:=(g^{ij}\delta _{j}f)\delta
_{i},$ to the scalar function $f:=\left\Vert \tau (\phi )\right\Vert ^{2},$
defined on $TM:
\begin{equation*}
\begin{array}{c}
-\dfrac{1}{2}\Delta \left\Vert \tau (\phi )\right\Vert ^{2}=\dfrac{1}{2
\{D_{\delta _{i}}(g^{ij}\delta _{j}\left\Vert \tau (\phi )\right\Vert
^{2})-g^{ij}P_{i}\delta _{j}\left\Vert \tau (\phi )\right\Vert ^{2}\}= \\
=\dfrac{1}{2}g^{ij}\{\delta _{i}\delta _{j}\left\Vert \tau (\phi
)\right\Vert ^{2}-\Gamma _{~ij}^{k}\delta _{k}\left\Vert \tau (\phi
)\right\Vert ^{2}-P_{i}\delta _{j}\left\Vert \tau (\phi )\right\Vert ^{2}\}
\end{array
\end{equation*}
Taking into account that $\Gamma _{~ij}^{k}\delta _{k}=D_{\delta _{i}}\delta
_{j}$ and expressing the action of the adapted basis vector fields $\delta
_{i},\delta _{j}$ on $\left\Vert \tau (\phi )\right\Vert ^{2}=\left\langle
\tau (\phi ),\tau (\phi )\right\rangle $ in terms of $D^{d\phi }$-covariant
derivatives, we obtain
\begin{equation}
-\dfrac{1}{2}\Delta \left\Vert \tau (\phi )\right\Vert ^{2}=-\left\langle
\Delta ^{d\phi }\tau (\phi ),\tau (\phi )\right\rangle +g^{ij}\left\langle
D_{\delta _{i}}^{d\phi }\tau (\phi ),D_{\delta _{j}}^{d\phi }\tau (\phi
)\right\rangle . \notag
\end{equation}
By means of the biharmonic equation (\ref{biharmonic_Finsler_to_Riemann}),
this becomes
\begin{equation}
-\dfrac{1}{2}\Delta \left\Vert \tau (\phi )\right\Vert ^{2}=\left\langle
trace_{g}\tilde{R}(d\phi ^{\tilde{h}},\tau (\phi ))d\phi ^{\tilde{h}},\tau
(\phi )\right\rangle +g^{ij}\left\langle D_{\delta _{i}}^{d\phi }\tau (\phi
),D_{\delta _{j}}^{d\phi }\tau (\phi )\right\rangle . \label{Weitz1}
\end{equation
According to the hypothesis that the sectional curvature of $(\tilde{M}
\tilde{g})$ is nonpositive, the curvature term above is nonnegative; since
g^{ij}\left\langle D_{\delta _{i}}^{d\phi }\tau (\phi ),D_{\delta
_{j}}^{d\phi }\tau (\phi )\right\rangle $ (as a squared norm) is
nonnegative, too, we get: $-\dfrac{1}{2}\Delta \left\Vert \tau (\phi
)\right\Vert ^{2}\geq 0.$
On the other side, we have, \cite{Zhong}, $\underset{BM}{\int }\Delta
\left\Vert \tau (\phi )\right\Vert ^{2}d\mathcal{V}_{g}=0$, hence, \ $\Delta
\left\Vert \tau (\phi )\right\Vert ^{2}=0;$ thus, by (\ref{Weitz1}),\
g^{ij}\left\langle D_{\delta _{i}}^{d\phi }\tau (\phi ),D_{\delta
_{j}}^{d\phi }\tau (\phi )\right\rangle =0;$ as a consequence,
\begin{equation}
D_{\delta _{j}}^{d\phi }\tau (\phi )=0. \label{*}
\end{equation
Take the horizontal vector field $X:=(g^{ij}\left\langle \phi _{,i},\tau
(\phi )\right\rangle )\delta _{j}$ on $TM;$ by (\ref{*}), we get
\begin{equation*}
0=\underset{BM}{\int }divXd\mathcal{V}_{g}=\underset{BM}{\int }\underset
\geq 0}{\underbrace{\left\langle \tau (\phi ),\tau (\phi )\right\rangle }}
\mathcal{V}_{g}
\end{equation*
and therefore, $\left\langle \tau (\phi ),\tau (\phi )\right\rangle
=0\Rightarrow $ $\tau (\phi )=0,$ i.e., $\phi $ is harmonic.
\end{proof}
Dropping any condition upon the compactness or on the boundary of $M,$ we
have:
\begin{theorem}
Let $(M,g)$ be an arbitrary Finsler space (not necessarily compact), $
\tilde{M},\tilde{g}),$ a Riemannian manifold with strictly negative
sectional curvature and $\phi :M\rightarrow \tilde{M},$ a biharmonic map. If
$\phi $ has the properties: 1)\ $\left\Vert \tau (\phi )\right\Vert =const.$
and 2)\ there exists a point $x_{0}\in M$ at which the rank of $\phi $ is at
least 2, then $\phi $ is harmonic.
\end{theorem}
\begin{proof}
The proof is similar to the one in the Riemannian case, \cite{Oniciuc}. From
the hypothesis $\left\Vert \tau (\phi )\right\Vert =const.,$ in (\ref{Weitz1
), the left hand side is 0; but both terms in the right hand side are
nonnegative, hence: $\left\langle trace_{g}\tilde{R}(d\phi ^{\tilde{h}},\tau
(\phi ))d\phi ^{\tilde{h}},\tau (\phi )\right\rangle =0.$ Since $Riem_
\tilde{g}}<0,$ we must have, for all $i=\overline{1,n}:$ $d\phi ^{\tilde{h
}(\delta _{i})~||~\tau (\phi )$. Taking into account that at $x_{0},$
rank(\phi )\geq 2,$ the only possibility is $\tau (\phi )(x_{0})=0$. Using
\left\Vert \tau (\phi )\right\Vert =const,$ it follows that $\tau (\phi
)\equiv 0,$ i.e., $\phi $ is harmonic.
\end{proof}
\section{Biharmonicity of the identity map}
Throughout this section, we assume that $M=\tilde{M}$ (not necessarily
compact), $\dim M=n,$ and denote the coordinates on $TM$ by $(x^{i},y^{i}).$
Considering on $M$ two metrics: a Riemannian one $\tilde{g}$ and a
Finslerian one $g,$ we will explore the biharmonicity of the
Finsler-to-Riemann mapping
\begin{equation}
id:(M,g)\rightarrow (M,\tilde{g}). \label{identity}
\end{equation}
In this situation, there appear two adapted bases $(\delta _{i},\dot{\partia
}_{i})$ and $(\tilde{\delta}_{i},\dot{\partial}_{i})$ on $TM,$ together with
the covariant differentiations given by $D,$ $\tilde{D}$ and $D^{d(id)}$.
According to \cite{Mo-book}, the tension of the identity map has the local
components
\begin{equation}
\tau ^{i}(id)=g^{jk}(\tilde{\Gamma}_{~jk}^{i}-G_{~jk}^{i})
\label{tension_id_Mo}
\end{equation
(note: our $G^{i}$ is half the one in \cite{Mo-book}).
Let us denote $b:=F^{2}-\tilde{F}^{2},$ i.e.
\begin{equation}
g_{ij}(x,y)=\tilde{g}_{ij}(x)+b_{ij}(x,y), \label{perturbed metric}
\end{equation
where the function $b=b(x,y)$ is homogeneous of degree 2 in $y$ and
b_{ij}=b_{\cdot ij}.$
In the geodesic equations (\ref{spray_coeff}), we express the derivatives
F_{~,k}^{2}$ in terms of $\tilde{D}_{\tilde{\delta}_{i}}$-covariant
derivatives, denoted in the following by double bars $_{||i}~$; we obtain:
\begin{equation}
2G^{i}=2\tilde{G}^{i}+2B^{i}, \label{geodesic_eqns_perturbation}
\end{equation
where
\begin{equation}
2B^{i}:=\dfrac{1}{2}g^{ih}(2y_{h||j}y^{j}-F_{~||h}^{2}) \label{expr_B}
\end{equation
and $y_{h}:=\dfrac{1}{2}F_{~\cdot h}^{2}=g_{hj}y^{j}.$ The tension of $id$
is
\begin{equation}
\tau ^{i}(id)=-g^{jk}B_{\cdot j\cdot k}^{i}. \label{tau_id}
\end{equation}
A direct computation shows that, in (\ref{expr_B}), the covariant derivative
$2y_{h||j}$ can be rewritten as
\begin{equation}
2y_{h||j}=F_{~||j\cdot h}^{2}. \label{commutation_F}
\end{equation}
\textbf{Remarks: 1) }If $b$ is parallel with respect to $\tilde{D},$ then
F_{~||k}^{2}=\tilde{F}_{~||k}^{2}+b_{||k}=0;$ taking into account (\re
{commutation_F}), we get $B^{i}=0\Rightarrow \tau ^{i}=0;$ in this case, the
identity map is harmonic, i.e., also biharmonic.
\textbf{2) }Assuming that $g$ is a Berwald-type metric, i.e.,
G_{~jk}^{i}=G_{~jk}^{i}(x),$ then there exists, \cite{Matveev}, \cite{Szabo
, a Riemannian metric such that $G_{~jk}^{i}=\tilde{G}_{~jk}^{i}$; thus, the
identity map is, again, harmonic, hence, biharmonic.
We will find in the following two examples of Finslerian perturbations $b$
for which the identity of $M$ is proper biharmonic.
With $\tau ^{i}:=\tau ^{i}(id),$ the relation between the $D^{d(id)}$- and
\tilde{D}$-covariant derivatives of $\tau ^{i}$ is
\begin{equation}
D_{~\delta _{j}}^{d(id)}\tau ^{i}=\delta _{j}\tau ^{i}+\tilde{\Gamma
_{~jk}^{i}\tau ^{k}=\tau _{~||j}^{i}-B_{~\cdot j}^{k}\tau _{~\cdot k}^{i}.
\label{rel_derivs}
\end{equation}
\textbf{1. }Suppose that the Finslerian function satisfies
\begin{equation}
F_{~||h}^{2}=~\left\langle a,y\right\rangle _{g}y_{h},
\label{perturbation_x}
\end{equation
where $\left\langle a,y\right\rangle _{g}:=g_{ij}a^{i}y^{j}$ and
a^{i}=a^{i}(x)$ are components of a vector field $A=a^{i}\partial _{i}$ on $M
$ (relations (\ref{perturbation_x}) are equivalent to a first order ODE
system in $g_{ij}$).
A brief calculation using (\ref{commutation_F}) leads to:
2y_{j||h}y^{h}-F_{~||j}^{2}=a_{j}F^{2}$, that is, $2B^{i}=a^{i}(x)F^{2}.$
From (\ref{tau_id}), we obtain
\begin{equation}
\tau ^{i}=-\dfrac{1}{2}na^{i}. \label{tau_x}
\end{equation}
Since $\tau ^{i}=\tau ^{i}(x)$, relation (\ref{rel_derivs})\ becomes simply:
$D_{~\delta _{j}}^{d(id)}\tau ^{i}=\tau _{~||j}^{i}$ and the biharmonic
equation is written as
\begin{equation}
g^{jk}(\tau _{~||j||k}^{i}-\tilde{\Gamma}_{~jk}^{l}\tau _{~||l}^{i}-\tilde{R
_{j~lk}^{~i}\tau ^{l})=0. \label{biharmonic_x}
\end{equation
Here, taking into account that $\tilde{R}_{jilk}=\tilde{R}_{lkij}$ and Ricci
identities (\ref{Ricci_R}) for $\tilde{D}$, the curvature term $\tilde{R
_{j~lk}^{~i}\tau ^{l}$ can be expressed by commuting $\tilde{D}$-covariant
derivatives of $\tau ^{i}.$ It turns out that a sufficient condition for the
biharmonicity of $id$ is
\begin{equation}
\tau _{~||j}^{i}=0. \label{suff_cond_x}
\end{equation}
(\textit{Note:}\ this statement is always true in the Riemannian case, but
generally, not in the Finsler-to-Riemann one, where, as a rule, $\tau
^{i}=\tau ^{i}(x,y)$).
Using (\ref{tau_x}), we deduce that (\ref{suff_cond_x}) is identically
satisfied if the vector field $A^{h}=a^{i}\delta _{i}$ is parallel with
respect to $\tilde{D}.$ But, according to (\ref{Levi-Civita_lift}), this is
nothing but: $\tilde{\nabla}_{\partial _{i}}A=0.$ In other words:
\begin{proposition}
If, in (\ref{perturbation_x}), the nonzero vector field $A=a^{i}(x)\partial
_{i}$ is parallel with respect to $\tilde{g},$ then the identity map $id:(M
\tilde{g})\rightarrow (M,g)$ is proper biharmonic.
\end{proposition}
\bigskip
\textbf{2. Linearized Finslerian perturbations of the Euclidean metric.
Assume that $(M,\tilde{g}_{ij})=(\mathbb{R}^{n},\delta _{ij})$ and the
perturbation $b_{ij}=:\varepsilon _{ij}(x,y)$ is small (linearly
approximable), that is, we may neglect all terms of degree greater than one
in $\varepsilon _{ij}$ and its derivatives, \cite{Landau}. In this case, the
inverse metric is given by: $g^{ik}=\delta ^{ik}-\varepsilon ^{ik}$ and
relation (\ref{expr_B}) becomes
\begin{equation*}
2B^{i}=\dfrac{1}{2}\delta ^{ih}(\varepsilon _{hj,k}+\varepsilon
_{hk,j}-\varepsilon _{jk,h})y^{j}y^{k}.
\end{equation*}
We notice that the tension $\tau $ will be of the same order of smallness as
$\varepsilon ;$ it means that products of $\tau $ with $\varepsilon $ and
its derivatives can be neglected. For instance, we have: $B_{~l}^{h}\tau
_{~\cdot h}^{i}\simeq 0,$ which, substituted into (\ref{rel_derivs}), leads
to
\begin{equation*}
D_{\delta _{j}}^{d(id)}\tau ^{i}=\tau _{~,j}^{i}.
\end{equation*
The biharmonic equation takes the simple form: $\delta ^{lm}\tau
_{~,l,m}^{i}=0.$ Again, a sufficient condition for biharmonicity i
\begin{equation*}
\tau _{~,l}^{i}=0,
\end{equation*
(or: $\tau ^{i}=\tau ^{i}(y)$), that is, $\delta ^{ih}(\varepsilon
_{hj,k,l}+\varepsilon _{hk,j,l}-\varepsilon _{jk,h,l})y^{j}y^{k}=0.$ We
obtain:
\begin{proposition}
Let the Finsler metric $g_{ij}(x,y)=\delta _{ij}+\varepsilon _{ij}(x,y)$ be
a linearized perturbation of the Euclidean metric on $\mathbb{R}^{n}.$ If
the components $\varepsilon _{ij}(x,y)$ are non-constant and affine in $x$,
then $id:(\mathbb{R}^{n},\delta _{ij})\rightarrow (\mathbb{R}^{n},g)$ is
proper biharmonic.
\end{proposition}
\section{Second variation of the bienergy}
Take a biharmonic map $\phi :(M,g)\rightarrow (\tilde{M},\tilde{g})$ and a
smooth 2-parameter variation $f=f(\varepsilon _{1},\varepsilon _{2},x),$
f(0,0,x)=\phi $ of $\phi ,$ wit
\begin{equation*}
\mathbf{V}_{1}=df^{\tilde{h}}(\partial _{\varepsilon _{1}}),~~\mathbf{V
_{2}=df^{\tilde{h}}(\partial _{\varepsilon _{2}}),~\ \ V_{1}:=\mathbf{V
_{1|\varepsilon _{1}=\varepsilon _{2}=0},~\ V_{2}:=\mathbf{V}_{1|\varepsilon
_{1}=\varepsilon _{2}=0}
\end{equation*
(if $M$ has a boundary, then $V_{1},$ $V_{2}$ and their $\delta _{i}
-covariant derivatives are assumed to vanish on $\partial M$).
The deduction of the second variation of $E_{2}$ follows the same steps as
in the Riemannian case, with two differences: in the expressions of $\tau
(f) $ and of $\Delta ^{df},$ there appear extra terms and we have to take
into account that $\Phi _{\ast }(\partial _{\varepsilon _{i}})$ is,
generally, not horizontal. Fortunately, as we will see below, these will
finally not complicate the expression of the variation.
We denote, for simplicity, $\tau :=\tau (f).$ According to (\re
{first_var_bienergy}), (\ref{biharmonic_Finsler_to_Riemann}):
\begin{equation}
\dfrac{\partial E_{2}(f)}{\partial \varepsilon _{1}}=\underset{BM}{\int
\left\langle \tau _{2}(f),\mathbf{V}_{1}\right\rangle d\mathcal{V}_{g};
\end{equation
differentiating with respect to $\varepsilon _{2}:
\begin{equation*}
\dfrac{\partial ^{2}E_{2}(f)}{\partial \varepsilon _{1}\partial \varepsilon
_{2}}=\underset{BM}{\int }\{\left\langle D_{\partial _{\varepsilon
_{2}}}^{df}\tau _{2}(f),\mathbf{V}_{1}\right\rangle +\left\langle \tau
_{2}(f),D_{\partial _{\varepsilon _{2}}}^{df}\mathbf{V}_{1}\right\rangle \}
\mathcal{V}_{g}.
\end{equation*
At $\varepsilon _{1}=\varepsilon _{2}=0,$ since $\phi $ is biharmonic, the
second term in the right hand side term will vanish. It is thus enough to
evaluate the first one; we have:
\begin{equation}
D_{\partial _{\varepsilon _{2}}}^{df}\tau _{2}(f)=-D_{\partial _{\varepsilon
_{2}}}^{df}(\Delta ^{df}\tau )-D_{\partial _{\varepsilon
_{2}}}^{df}(trace_{g}\tilde{R}(df^{\tilde{h}},\tau )df^{\tilde{h}}).
\label{to_evaluate}
\end{equation}
The covariant derivative of the Laplacian $-\Delta ^{df}\tau $ is
\begin{equation}
T_{1}:=-D_{\partial _{\varepsilon _{2}}}^{df}(\Delta ^{df}\tau
)=g^{ij}D_{\partial _{\varepsilon _{2}}}^{df}\left( D_{~\delta
_{i}}^{df}D_{~\delta _{j}}^{df}\tau -D_{~D_{\delta _{i}}\delta
_{j}}^{df}\tau -P_{i}D_{\delta _{j}}^{df}\tau )\right) .
\label{deriv_Laplacian}
\end{equation}
Commuting covariant derivatives by means of the curvature $\tilde{R}$ (twice
for the term $D_{~\delta _{i}}^{df}D_{~\delta _{j}}^{df}\tau $), taking into
account that $[\partial _{\varepsilon },\delta _{i}]=0$ and (\ref{Riem_R}),
we find:
\begin{equation*}
\begin{array}{c}
T_{1}=g^{ij}{\Large \{}\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta
_{i}))D_{~\delta _{j}}^{df}\tau +D_{~\delta _{i}}^{df}\left( \tilde{R}
\mathbf{V}_{2},df^{\tilde{h}}(\delta _{j}))\tau +D_{~\delta
_{j}}^{df}D_{\partial _{\varepsilon _{2}}}^{df}\tau \right) - \\
-\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(D_{\delta _{i}}\delta _{j}))\tau
-D_{~D_{\delta _{i}}\delta _{j}}^{df}D_{\partial _{\varepsilon
_{2}}}^{df}\tau -P_{i}\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta
_{j}))\tau ~-P_{i}D_{\delta _{j}}^{df}D_{\partial _{\varepsilon
_{2}}}^{df}\tau {\Large \}}
\end{array
\end{equation*}
The terms in $D_{\partial _{\varepsilon _{2}}}^{df}\tau $ can be grouped
into $-\Delta ^{df}(D_{\partial _{\varepsilon _{2}}}^{df}\tau ):$
\begin{equation*}
\begin{array}{c}
T_{1}=-\Delta ^{df}(D_{\partial _{\varepsilon _{2}}}^{df}\tau )+g^{ij
{\Large \{}\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta _{i}))D_{~\delta
_{j}}^{df}\tau + \\
D_{~\delta _{i}}^{df}\left( \tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta
_{j}))\tau \right) -\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(D_{\delta
_{i}}\delta _{j}))\tau -P_{i}\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta
_{j}))\tau {\Large \}}
\end{array
\end{equation*}
Splitting $D_{~\delta _{i}}^{df}\left( \tilde{R}(\mathbf{V}_{2},df^{\tilde{h
}(\delta _{j}))\tau \right) $ as a sum of derivatives, we recognize in the
resulting expression $\tilde{R}(\mathbf{V}_{2},\tau )\tau :$
\begin{equation}
\begin{array}{l}
T_{1}=-\Delta ^{df}(D_{\partial _{\varepsilon _{2}}}^{df}\tau )+\tilde{R}
\mathbf{V}_{2},\tau )\tau +g^{ij}{\Large \{}(D_{~\delta _{i}}^{df}\tilde{R})
\mathbf{V}_{2},df^{\tilde{h}}(\delta _{j}))\tau + \\
+2\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta _{i}))D_{~\delta
_{j}}^{df}\tau +\tilde{R}(D_{~\delta _{i}}^{df}\mathbf{V}_{2},df^{\tilde{h
}(\delta _{j}))\tau {\Large \}}
\end{array}
\label{T1}
\end{equation}
The curvature term $T_{2}:=-D_{\partial _{\varepsilon _{2}}}^{df}(trace_{g
\tilde{R}(df^{\tilde{h}},\tau )df^{\tilde{h}})$ in (\ref{to_evaluate}) is
\begin{equation}
\begin{array}{c}
T_{2}=-g^{ij}\{(D_{\partial _{\varepsilon _{2}}}^{df}\tilde{R})(df^{\tilde{h
}(\delta _{i}),\tau )df^{\tilde{h}}(\delta _{j})+\tilde{R}(D_{\partial
_{\varepsilon _{2}}}^{df}df^{\tilde{h}}(\delta _{i}),\tau )df^{\tilde{h
}(\delta _{j})+ \\
+\tilde{R}(df^{\tilde{h}}(\delta _{i}),D_{\partial _{\varepsilon
_{2}}}^{df}\tau )df^{\tilde{h}}(\delta _{j})+\tilde{R}(df^{\tilde{h}}(\delta
_{i}),\tau )D_{\partial _{\varepsilon _{2}}}^{df}(df^{\tilde{h}}(\delta
_{j}))
\end{array
\end{equation}
Taking into account that $\tilde{R}=\tilde{R}(x)$ only, we obtain
D_{\partial _{\varepsilon _{2}}}^{df}\tilde{R}=\tilde{D}_{\mathbf{V}_{2}
\tilde{R}.$ Transforming $D_{\partial _{\varepsilon _{2}}}^{df}df^{\tilde{h
}(\delta _{i}),$ $D_{\partial _{\varepsilon _{2}}}^{df}df^{\tilde{h}}(\delta
_{j})$ by (\ref{commutation_rule_2}) and then using first Bianchi identity
in the second term
\begin{equation}
\begin{array}{c}
T_{2}=-g^{ij}\{(\tilde{D}_{\mathbf{V}_{2}}\tilde{R})(df^{\tilde{h}}(\delta
_{i}),\tau )df^{\tilde{h}}(\delta _{j})+\tilde{R}(D_{\delta _{i}}^{df
\mathbf{V}_{2},\tau )df^{\tilde{h}}(\delta _{j})+ \\
+\tilde{R}(df^{\tilde{h}}(\delta _{i}),D_{\partial _{\varepsilon
_{2}}}^{df}\tau )df^{\tilde{h}}(\delta _{j})+\tilde{R}(df^{\tilde{h}}(\delta
_{i}),\tau )\tilde{D}_{\delta _{j}}\mathbf{V}_{2}\}= \\
=-g^{ij}\{(\tilde{D}_{\mathbf{V}_{2}}\tilde{R})(df^{\tilde{h}}(\delta
_{i}),\tau )df^{\tilde{h}}(\delta _{j})+2\tilde{R}(df^{\tilde{h}}(\delta
_{i}),\tau )\tilde{D}_{\delta _{j}}\mathbf{V}_{2}- \\
-\tilde{R}(df^{\tilde{h}}(\delta _{j}),D_{\delta _{i}}^{df}\mathbf{V
_{2})\tau )+\tilde{R}(df^{\tilde{h}}(\delta _{i}),D_{\partial _{\varepsilon
_{2}}}^{df}\tau )df^{\tilde{h}}(\delta _{j})\}
\end{array
\end{equation}
Second, and then first Bianchi identities for the $(\tilde{D}_{\mathbf{V
_{2}}\tilde{R})$-term tell us that
\begin{equation*}
\begin{array}{c}
-g^{ij}(\tilde{D}_{\mathbf{V}_{2}}\tilde{R})(df^{\tilde{h}}(\delta
_{i}),\tau )df^{\tilde{h}}(\delta _{j})=g^{ij}\{(\tilde{D}_{\tau }\tilde{R})
\mathbf{V}_{2},df^{\tilde{h}}(\delta _{i}))df^{\tilde{h}}(\delta _{j})- \\
-(\tilde{D}_{\delta _{i}}\tilde{R})(df^{\tilde{h}}(\delta _{j}),\tau
\mathbf{V}_{2}-(\tilde{D}_{\delta _{i}}\tilde{R})(\mathbf{V}_{2},df^{\tilde{
}}(\delta _{j}))\tau \}
\end{array
\end{equation*}
Substituting into $T_{2}$ and adding: $T_{1}+T_{2}=D_{\partial _{\varepsilon
_{2}}}^{df}\tau _{2}(f),$ we get
\begin{equation*}
\begin{array}{l}
D_{\partial _{\varepsilon _{2}}}^{df}\tau _{2}(f)=\mathcal{J}(D_{\partial
_{\varepsilon _{2}}}^{df}\tau )+\tilde{R}(\mathbf{V}_{2},\tau )\tau +g^{ij
{\Large \{}(\tilde{D}_{\tau }\tilde{R})(\mathbf{V}_{2},df^{\tilde{h}}(\delta
_{i}))df^{\tilde{h}}(\delta _{j})- \\
-(D_{~\delta _{i}}^{df}\tilde{R})(df^{\tilde{h}}(\delta _{j}),\tau )\mathbf{
}_{2}+2\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta _{i}))D_{~\delta
_{j}}^{df}\tau -2\tilde{R}(df^{\tilde{h}}(\delta _{i}),\tau )D_{~\delta
_{i}}^{df}\mathbf{V}_{2}{\Large \}}
\end{array
\end{equation*
with $\mathcal{J}$ as in (\ref{rough_Laplacian}). Using (\ref{deriv_tau_1})
and evaluating at $\varepsilon _{1}=\varepsilon _{2}=0,$ we get:
\begin{proposition}
The second variation of the bienergy of a Finsler-to-Riemann biharmonic map
\phi :M\rightarrow \tilde{M}$ is
\begin{eqnarray}
&&\dfrac{\partial ^{2}E_{2}(f)}{\partial \varepsilon _{1}\partial
\varepsilon _{2}}|_{\varepsilon _{1}=\varepsilon _{2}=0}=\underset{BM}{\int
\left\langle V_{1},\right. \mathcal{J}^{2}V_{2}+\tilde{R}(V_{2},\tau )\tau +
\label{second variation_E2} \\
&&+g^{ij}{\Large \{}(\tilde{D}_{\tau }\tilde{R})(\mathbf{V}_{2},df^{\tilde{h
}(\delta _{i}))df^{\tilde{h}}(\delta _{j})-(D_{~\delta _{i}}^{df}\tilde{R
)(df^{\tilde{h}}(\delta _{j}),\tau )\mathbf{V}_{2}+ \notag \\
&&+2\tilde{R}(\mathbf{V}_{2},df^{\tilde{h}}(\delta _{i}))D_{~\delta
_{j}}^{df}\tau -2\tilde{R}(df^{\tilde{h}}(\delta _{i}),\tau )D_{~\delta
_{i}}^{df}\mathbf{V}_{2}{\Large \}}\left. {}\right\rangle d\mathcal{V}_{g}.
\notag
\end{eqnarray}
\end{proposition}
In particular cases (for instance, $\tilde{M}=\mathbb{R}^{n}$ or $S^{n}$),
\ref{second variation_E2})\ becomes considerably simpler.
The Hessian $\mathcal{H}:(V_{1},V_{2})\mapsto \mathcal{H}(V_{1},V_{2})
\dfrac{\partial ^{2}E_{2}(f)}{\partial \varepsilon _{1}\partial \varepsilon
_{2}}|_{\varepsilon _{1}=\varepsilon _{2}=0}$ of the bienergy is a symmetric
bilinear form. A solution $\phi $ of the biharmonic equation is \textit
stable} if the quadratic form $\mathcal{H}(V,V)$ is nonnegative for any $V$.
As an example, harmonic maps are stable biharmonic maps.
\textbf{Acknowledgments. }1) The work was supported by the Sectorial
Operational Program Human Resources Development (SOP HRD), financed from the
European Social Fund and by Romanian Government under the Project number
POSDRU/89/1.5/S/59323.
2)\ Special thanks to prof. G. Munteanu for proofreading the text.
|
\section{Introduction}\label{SectionIntroduction}
Renshaw and S\"arkk\"a presented their spatio-temporal growth-interaction (GI) process in \cite{RS1}, and usually this process is described as a spatio-temporal point process (see e.g.\ \cite{VerreJones}) with growing and interacting marks.
The GI-process then has been further studied in a series of papers (e.g.\ \cite{Comas,Cronie,CronieYu,RenshawComas,RenshawComasMateu,RS2}), where, among other things, different inference tools have been developed.
Consider some suitable spatial study region (usually some subset of ${\mathbb R}^2$ or a torus).
The basis of the GI-process is spatio-temporal point process which here will be referred to as a spatial immigration-death (SID) process.
Specifically, the SID-process lets new individuals (points) arrive to a population (the study region) according to the jumps of a Poisson process on ${\mathbb R}_+$, assigns iid locations to them, which are uniformly distributed in the study region, and removes the points after iid exponential times (the temporal dynamics of the SID-process constitute a so-called immigration-death process; see e.g.\ \cite{CronieYu}).
Moreover, once an individual has received its location, centred on its location we place a closed disk/ball with some given radius (the initial mark).
As time evolves, we let the radius of each disk grow according to a given deterministic growth structure, which depends on the radius itself as well as the locations and the radii of the other (neighbouring) marked points.
More precisely, the structure of the simultaneous growth of the radii is given by a system of ordinary differential equations (ODEs).
Furthermore, this system of ODEs is such that when an individual has not yet arrived to the population or if it has been removed (which is governed by the SID-process), its corresponding component of the system of ODEs is set to zero.
Considering its possible application areas,
one important area is the dynamical modelling of a forest stand.
Here, as time passes, new trees arrive, grow and compete with each other until they die.
When we let each growth equation (ODE component) contain an inhibitive part,
such that the growth of a point's radius/mark is reduced when the point is surrounded (spatially) by points with large marks,
this inhibition reflects the natural competition for nutrients and light among trees in a forest.
The description of the GI-process as a spatio-temporal marked point process is rather vague and questions regarding an appropriate representation quickly emerge.
Following \cite{Mateu}, we will call any marked point process where the marks are function-valued a \emph{functional marked point process}, and it has been pointed out by \cite{Mateu} that the GI-process in fact should be represented as a functional marked point process.
More specifically, it can be treated as a marked point process (see e.g.\ \cite{DaleyVereJones,Diggle,SchneiderWeil}) for which the location space is that of a spatial Poisson process and the mark space (representing the radii of the balls/disks) is given by the space of c\`adl\`ag (right continuous with existing left limits) functions (see e.g.\ \cite{Billingsley}).
This can be realised by letting the unmarked part of this functional marked point process be given by the collection of points scattered in the study region by the underlying SID-process during the time interval we are studying the GI-process.
Moreover,
we see that each mark consists of three parts:
1) the size of the mark before the individual has arrived (zero),
2) the size when the individual is present (governed by the ODEs), and
3) the size once the individual is removed (zero).
Since there clearly are jumps between parts 1) and 2), and parts 2) and 3), we see that it makes sense to consider a mark space which is of c\`adl\`ag-type.
Note that the connection between these two representations of the GI-process can be compared to the representations of a one-dimensional Poisson process as either a point process (random measure) on $R_+$ or as a Levy-process (evolutionary process).
Naturally the growth structure of the GI-process includes some parameters that need to be estimated when the model is fitted to data.
For the case where the process is sampled at discrete times (hence creating a time series of marked point patterns), \cite{RS2} suggested a least-squares scheme to estimate the parameters related to the growth and interaction of the marks (the parameters in the ODEs).
This estimation was further considered in \cite{Cronie}, where a spatio-temporal edge correction was added to the estimation procedure.
Regarding the parameters of the underlying (spatial) immigration-death process, which are the arrival and death rates of individuals, these have been estimated separately by means of different maximum likelihood (ML) estimation approaches (see e.g.\ \cite{Cronie,CronieYu,GibsonRenshaw,RS2}).
In \cite{CronieYu} the full ML-estimation of the discretely sampled immigration-death process was treated, and, besides treating some practical aspects of the estimation,
consistency and asymptotic normality of the ML-estimators were proved.
We note that it is unlikely that all marks in a marked point pattern (e.g.\ trees in a forest stand) have the same (deterministic) underlying growth pattern, as is the assumption in the original GI-process.
For instance, when we model a forest stand, in order to reflect phenomena such as (cumulative) measurement errors and individual growth features of each tree, there should be some noise or randomness present in the part of the model which handles the growth of the marks.
Trying to rectify this lack of individuality in the mark growth structures, our main aim here will be to take a first step in the process of adding randomness to the growth of the marks.
The approach chosen here is to add a (scaled continuous time) white noise to each component of the system of mark growth ODEs in the GI-process (see ''part 2)'' above).
This will generate a set of (possibly dependent) Brownian motion driven stochastic differential equations (SDEs) with jumps at the birth and death times (see e.g.\ \cite{KaratzasShreve,Klebaner,Protter,Shreve}).
In other words, we will be considering a functional marked point process with marks given by diffusion processes with jumps.
By considering the temporal evolution of such a process, we may also say that we have defined a
multivariate stochastic jump (diffusion) process which is time shifted and parametrized by an SID-process.
We note that turning the marks into diffusions has some clear advantages.
To begin with, as a consequence of the randomness in the mark growth equations, we may be able to write down a likelihood structure,
which is based on sampling the marks over time,
so that we can treat the simultaneous ML-estimation of the whole GI-process
instead of using the separate spatial and temporal estimators previously considered.
In particular, we may exploit the Markovianity of the diffusions and the SID-process in the derivation of a full likelihood function.
Additionally, given a temporally discrete sampling scheme,
standard tools from likelihood theory can help us derive results about the asymptotic behaviour of the estimators (see e.g.\ \cite{Hoadley,Iacus,Rao,VanDerVaart}), when the process is sampled at discrete times.
Ultimately it may also be possible to use tools from stochastic process theory and stochastic calculus to derive other theoretical results, such as asymptotic distributional properties of functionals of the process, which in the non-stochastic version of the GI-process only have been achievable through simulations.
Here, as an initial extension of the GI-process to the setting where the marks are stochastic, we will assume that all diffusions (mark radius processes) will be independent and of the same type (thereby generated from the same set of parameters).
Note that we hereby remove the interaction between the marks which was present in the GI-process.
Hence, we obtain a functional marked point process with independent diffusion process marks.
We have chosen to study only the Cox-Ingersoll-Ross (CIR) process (see e.g.\ \cite{CIR,Feller,Iacus}) to describe the marks (growth of the radii of the disks).
However, any other strictly positive diffusion which meets the requirements of the modelling setting in question could be chosen.
Note that a nice property of the CIR-process (besides being strictly positive under certain restrictions) is that it is one of the few diffusions which possesses known closed form expressions for its transition densities, and since the transition densities in turn are the building blocks of the likelihood function we can obtain a closed form expression for the likelihood function.
The paper is organised as follows.
In Section \ref{SectionSGI} the stochastic GI-process is defined and in Section \ref{SectionDistribution} we move on to further discuss some of its distributional properties and its building blocks.
Then, in Section \ref{SectionLikelihood}, with the finite dimensional distributions at hand, we define the ML-estimation regime which, in Section \ref{SectionAsymptotics}, in turn allows us to look at large sample properties of ML-estimators.
Since the consistency and the asymptotic normality of the ML-estimators are proved when the mark processes are stationary, in Section \ref{SectionEvaluation} we wish to see how robust these estimators are to the stationarity assumption.
In Section \ref{SectionPines} we evaluate the estimators on the same set of Scots pine data considered in \cite{Cronie} and
in Section \ref{SectionDiscussion} further comments/possible extensions are given as well as a general discussion of the paper.
Finally, in the Appendix proofs of the main results can be found.
\section{The SG(I) process}\label{SectionSGI}
The \emph{stochastic growth-interaction} (SGI) process $\Phi_M$ is a functional marked point process which can be considered to be a stochastic extension of the (non-stochastic) growth-interaction (GI) process (see e.g.\ \cite{Cronie,RS1,RS2,Comas}).
Heuristically $\Phi_M$ is described in the following way.
As time evolves, balls/disks (marked points) appear in a spatial study region $W$ at stochastic times and the radii of these balls/disks change size randomly over time until they disappear after stochastic times.
Due to the usual biological context it will be natural to refer to each point as an \emph{individual}.
As previously noted, we here consider a simplified version of the SGI-process since we do not include any interaction between the marked points, and we therefore will refer to $\Phi_M$ as the \emph{stochastic growth} (SG) process.
It should be pointed out that due to the GI-process' forestry application it is usual that we illustrate the process in such a way that its marks describe the aforementioned growth of disks in ${\mathbb R}^2$ (the space occupied by the tree stocks), however, this need not be the case.
For instance, when modelling the spatio-temporal development of a forest stand, one could instead consider the case where the marks are used to describe, say, the development of the height of the trees.
As was mentioned in Section \ref{SectionIntroduction}, there are essentially two ways to construct the SG-process $\Phi_M$.
In the first representation of the SG-process, which is given in Section \ref{SectionFMPP}, $\Phi_M$ is obtained by treating it as a functional marked point process with c\`adl\`ag function-valued marks.
The second representation (Section \ref{SectionEvolutionary}) is obtained by considering the temporal evolution of the mark processes (which are parametrized by the other relevant information).
Note that the latter representation is of significance since it will be exploited when we develop the statistical inference for the SG-process when the marks are sampled at discrete times.
Throughout we will assume that the spatial study region is given by a subset $W\subseteq{\mathbb R}^d$ of $d$-dimensional Euclidean space, $d\geq1$, with Borel sets ${\mathcal B}(W)\subseteq{\mathcal B}(R^d)$ and Lebesgue measure $\nu(W)<\infty$ (note that this also includes the case of identifying the edges of a rectangle in order to construct a torus).
Moreover, we write ${\mathbb N}=\{0,1,\ldots\}$ and we denote the Euclidean norm and metric by $|x|$ and $d(x,y)$, respectively, for $x,y\in{\mathbb R}^d$, $d\geq1$.
Furthermore, for a given set $A$, ${\mathbf 1}_{A}(a)={\mathbf 1}\{a\in A\}$ will denote the related indicator function and $|A|$ will denote the related cardinality (it will be clear from context whether we consider the norm or the cardinality).
Following \cite{Mateu} and the construction of a functional marked point process given therein, the mark space will be given by the set ${\mathbb F}=D_{[0,T]}({\mathbb R})$, $T\in{\mathbb R}_+=(0,\infty)$, of c\`adl\`ag (right continuous with existing left limits) functions $f:[0,T]\rightarrow{\mathbb R}$ (or ${\mathbb F}=D_{[0,\infty)}({\mathbb R})$ when $f:[0,\infty)\rightarrow{\mathbb R}$).
The underlying probability space will be denoted by $(\mathcal{X},\mathcal{F},{\mathbb P})$.
\subsection{Functional marked point process representation of the SG-process}\label{SectionFMPP}
Assume now that the SG-process under consideration is given by a c\`adl\`ag functional marked point process $\Phi_M = \{[X_i,M_i]:X_i\in\Phi'\}$, with locations $X_i\in W$ and functional marks $M_i\in{\mathbb F}$.
More specifically, we let the unmarked process $\Phi'=\{X_1,\ldots,X_N\}$ be given by a homogeneous Poisson process on $W$, with intensity $\alpha\nu(W)T$, $\alpha\in\Theta_{\alpha}\subseteq{\mathbb R}_+$. We note that we hereby have $N\sim Poi(\alpha\nu(W)T)$ locations $X_1,\ldots,X_N\in W$ which are iid $Uni(W)$-distributed (their indices are assigned to them according to their ''birth times'' defined below).
We now turn to the construction of the ${\mathbb F}$-valued random functional marks $M_1,\ldots,M_N$, which we will require to be almost surely (a.s.)\ positive.
In order to generate the supports $\mathrm{supp}(M_i)=\{t\in[0,T]:M_i(t)\neq0\} = [B_i,D_i)$, $i=1,\ldots,N$, conditionally on $\Phi'$ (or simply $N$), let $B_1,\ldots,B_N$ be iid $Uni(0,T)$-distributed random variables (relabelled according to ascending size) and
let additionally $L_1,\ldots,L_N$ be iid $Exp(\mu)$-distributed, $\mu\in\Theta_{\mu}\subseteq{\mathbb R}_+$.
By now defining $D_i = (B_i+L_i)\wedge T = \min\{B_i+L_i,T\}$, $i=1,\ldots,N$,
we have that $M_i(t)=0$ for all $t\notin[B_i,D_i)$ and $M_i(t)>0$ for all $t\in[B_i,D_i)$ a.s..
We note here that the ''birth/arrival times'' $B_1<\ldots<B_N$ form a Poisson process on $[0,T]$ with intensity $\alpha\nu(W)$. In order to use a terminology which illustrates how $\Phi_M$ can be used to model the dynamics of a population (e.g.\ a forest stand), in connection to the birth times, we additionally call $L_1,\ldots,L_N$ the ''lifetimes'' and $D_1,\ldots,D_N$ the ''death times'' of the individuals.
As previously mentioned, each mark $M_i=\{M_i(t)\}_{t\in[0,T]}$, $i=1,\ldots,N$, can be illustrated by the space which it occupies in ${\mathbb R}^d$ at a given time $t$. This is done by means of the ball $B_{X_i}[M_i(t)]=\{{\bf y}\in{\mathbb R}^d:d(X_i,{\bf y})\leq M_i(t)\}$, with centre $X_i$ and radius $M_i(t)$.
Hereby, for a fixed time $t\in[0,T]$, $\Phi_M$ can be illustrated as a forest stand, or rather a Boolean model (see e.g.\ \cite{StoyanKendallMecke}), by considering the union of the disks or trees $\bigcup_{i=1}^{N} B_{X_i}[M_i(t)]$ or the union of all disks $B_{X_i}[M_i(t)]$, $i=1,\ldots,N$, such that $t\in[B_i,D_i)$ (whereby we only observe ''alive individuals'').
Considering now to the actual structure put on each $M_i=\{M_i(t)\}_{t\in[0,T]}$ when $t\in[B_i,D_i)$ (i.e.\ when the $i$th individual is alive), we will let $M_i(B_i) = M_{i}^{0}$ be the initial size of the $i$th mark process and to illustrate how the construction of $\Phi_M$ originates from the GI-process, we first recall (see e.g.\ \cite{RS2,Cronie,Comas}) that in the GI-process the marks $M_i(t)$, $i=1,\ldots,N$, were set to develop deterministically according to
\begin{eqnarray}
\label{GI}
M_i(t)&=&M_i^0 + \int_{B_i}^{D_i}dM_i(s) \\
&=& M_i^0 + \int_{B_i}^{D_i} f(M_i(s);\theta) + \sum_{i\neq j}h(M_i(s),M_j(s),X_i,X_j;\theta)\;ds,\nonumber
\end{eqnarray}
for $t\in[B_i,D_i)$. Here $f(\cdot)$ controls the growth of radius $i$ in absence of spatial competition (so-called open growth in forestry terminology), $h(\cdot)$ controls the spatial interaction between individual $i$ and the other individuals and $\theta$ is a vector of parameters which controls $f(\cdot)$ and $h(\cdot)$ (see e.g.\ \cite{Cronie,RS2}).
Here, however, in order to initialise the more realistic growth scenario where the marks have random growth patterns, as previously mentioned, we assume that each radius grows stochastically according to an a.s.\ positive
stochastic process (note that an ${\mathbb F}$-valued random variable/element is a stochastic process with c\`adl\`ag sample paths).
By calling $t\in[0,T]$ our \emph{global time} and $t-B_i$ our $i$th \emph{local time},
in order to properly express $M_i(t)$ through the global time scale, we will let the processes $Y_i(t)$, $i=1,\ldots,N$, where
\begin{eqnarray}
\label{Diffusion}
M_i(t) =
\left\{
\begin{array}{ll}
Y_i(t-B_i) & \text{for } t\in[B_i,D_i)\\
0 & \text{for } t\notin[B_i,D_i),
\end{array}
\right.
\end{eqnarray}
be given by a system of independent (time-shifted) CIR-processes (see e.g.\ \cite{CIR,Feller,Iacus})
\begin{eqnarray}
\label{SDE}
dY_i(t) &=& \lambda\left(1 - Y_{i}(t)/K\right)dt + \sigma\sqrt{Y_{i}(t)}dW_i(t),\\
\label{CIR}
Y_i(t) &=& M_{i}^{0} + \int_{0}^{t} \lambda\left(1 - \frac{Y_{i}(s)}{K}\right)ds + \int_{0}^{t}\sigma\sqrt{Y_{i}(s)}dW_i(s),
\end{eqnarray}
with $Y_i(0)=M_{i}^{0}$, $i=1,\ldots,N$, where the $W_i(t)$'s are independent standard Brownian motions.
We note that this is equivalent to setting $f(x)=\lambda(1-x/K)$, $h(\cdot)=0$ and adding a stochastic integral to expression (\ref{GI}).
The parameters $(\lambda,K,\sigma)\in\Theta_{\lambda}\times\Theta_{K}\times\Theta_{\sigma}\subseteq{\mathbb R}_+^{3}$ in expressions (\ref{SDE}) and (\ref{CIR}) control different aspects of the growth of a radius $M_i(t)$ of a ball/disk $B_{X_i}[M_i(t)]$:
The \emph{diffusion coefficient} $\sigma$ controls the magnitude of the random individual fluctuations of the radii.
The interpretation of the remaining two parameters becomes most clear by noticing that $Y_i(t)$ is a so called mean-reverting process, i.e.\ as $Y_i(t)$ starts to move away from its long term equilibrium $K$, the drift term starts pulling it back towards $K$ and the speed at which this occurs is given by $\lambda/K$.
Related to this interpretation we find that if we set $\sigma=0$ in expression (\ref{SDE}), we retrieve
expression (\ref{GI}) with $h(\cdot)=0$ (the GI-process without interaction)
or equivalently
the ODE $dY_{i}(t) = \lambda\left(1 - Y_{i}(t)/K\right)dt$.
This ODE is often referred to as the linear growth function (see e.g.\ \cite{RenshawComasMateu,RS2}) and in this setting the parameter $\lambda$ is referred to as the (individual) \emph{growth rate} while the upper bound $K$ often is referred to as the \emph{carrying capacity}.
In conclusion, $\Phi_M$ is controlled by the parameter vector $\theta = (\lambda,K,\sigma,\alpha,\mu)\in\Theta=\Theta_{\lambda}\times\Theta_{K}\times\Theta_{\sigma}\times\Theta_{\alpha}\times\Theta_{\mu}\subseteq{\mathbb R}_{+}^{5}$,
where the pair $(\alpha,\mu)$ controls the time intervals during which the mark functions are non-zero and the remaining parameters control the growth of the marks.
Regarding the initial size $M_i(B_i) = Y_i(0) = M_{i}^{0}$, a few different options are available.
In \cite{Cronie}, the approach was to use the same constant initial value $M_{i}^{0}\equiv M_0\in{\mathbb R}_{+}$ for all individuals in the GI-process,
and in \cite{RS2} the $M_{i}^{0}$'s were chosen as independent $Uni(0,\epsilon)$-distributed random variables, $\epsilon>0$.
Here, however, we also have the further option to sample $M_{i}^{0}$ from the stationary distribution of $Y_i$ (see Section \ref{SectionDistribution} for details).
\subsection{Temporal evolution representation}\label{SectionEvolutionary}
In order to stress that we here consider the temporal evolution of the SG-process (or rather the temporal evolution of the marks), we often write $\Phi_M(t) = (M_1(t),\ldots,M_N(t))$, $t\geq0$.
Furthermore, in order to treat it properly we let it be adapted to some filtered probability space $(\mathcal{X},\mathcal{F},\{\mathcal{F}_t\}_{t\in[0,T)},{\mathbb P})$.
Specifically, the family of $\sigma$-algebras $\{\mathcal{F}_t\}_{t\in[0,T]}$ is such that, for any $s \leq t$, $\mathcal{F}_s\subseteq\mathcal{F}_t\subseteq\mathcal{F}$ and, for each $t\in[0,T]$, $\Phi_M(t)$ is $\mathcal{F}_t$-measurable.
We here will construct $\Phi_M(t)$ through two building blocks.
The first building block is given by the underlying point process $\Phi(t)$, which can be constructed as spatio-temporal point process on $W\times[0,T]$, and we call it a spatial immigration-death (SID) process.
This process governs the assignment of the spatial locations of the individuals in $W$, as well as their arrival times and their lifetimes.
The second building block, which may be regarded as an extension of $\Phi(t)$, is the set of ${\mathbb F}$-valued functional marks (stochastic processes).
We start by describing the underlying SID-process $\Phi(t)$.
Let us consider the SID-process $\{\Phi(t)\}_{t\in[0,T]}$ which is a spatial birth-death process (see e.g.\ \cite{Moller,Kasper}), taking values in the collection $N^f=\{{\bf x}\subseteq W:|x|<\infty\}$ of finite point configurations. It has birth rate function $b(\cdot,\cdot)=\alpha$, death rate function $d(\cdot,\cdot)=\mu$ and reference probability measure $\upsilon(B)=\nu(B)/\nu(W)$, $B\in{\mathcal B}(W)$, where $(\alpha,\mu)\in\Theta_{\alpha}\times\Theta_{\mu}\subseteq{\mathbb R}_+^2$. Hence, it
can easily be verified that the underlying Markov jump process is given by a so-called immigration-death (ID) process ($M/M/\infty$-queue) $\{N(t)\}_{t\in[0,T]}$ (see e.g.\ \cite{CronieYu,GibsonRenshaw,Grimmett}) with arrival rate $\alpha\nu(W)$ and death rate $\mu$.
Furthermore, we see that the spatial location kernel is such that all locations $X_i\in\Phi(t)\in N^f$, $t\in[0,T]$, are iid $Uni(W)$-distributed.
Looking closer at the ID-process, which is
a time-homogeneous irreducible positive recurrent Markov chain with state space ${\mathbb N}$, we see that it can be used to describe a population where the ''birth/arrival times'' $B_1<\ldots<B_N$ of the individuals occur according to a Poisson process on $[0,T]$ with intensity $\alpha\nu(W)$ (whereby $N\sim Poi(\alpha\nu(W)T)$) and it generates ''lifetimes'' $L_1,\ldots,L_N$ for the individuals which are iid $Exp(\mu)$-distributed.
Hereby, by defining $D_i = (B_i+L_i)\wedge T$, $i=1,\ldots,N$, we finalise the equivalence with the construction of the previously defined supports $\mathrm{supp}(M_i)=\{t\in[0,T]:M_i(t)\neq0\} = [B_i,D_i)$, $i=1,\ldots,N$, of the functional marks.
It is sometimes important to keep track of which individuals are alive/visible and we therefore define the index process $\Omega(t)=\left\{i\in\{1,\ldots,N\}:t\in[B_i,D_i)\right\}$, $\Omega(0)=\emptyset$, which is a Markov process which controls which individuals are alive at time $t$ (note that $N(t) = |\Phi(t)| = |\Omega(t)|$).
We note that we just as well could have defined $\Phi(t)$ as a marked Poisson process on $[0,T]$, with jump times $B_1<\ldots<B_N$ and marks $(L_i,X_i)$, $i=1,\ldots,N$.
We now turn to the second building block of $\Phi_M$.
Similarly to the previous scenario, the idea here is to consider the stochastic processes $M_i(t) = M_i(t;\Phi) ={\mathbf 1}_{[B_i,D_i)}(t)Y_i(t-B_i)$,
$i=1,\ldots,N$, where the $Y_i(t)$'s are defined in expressions (\ref{SDE}) and (\ref{CIR}).
Just as before the parameter vector will be given by $\theta = (\lambda,K,\sigma,\alpha,\mu)\in\Theta=\Theta_{\lambda}\times\Theta_{K}\times\Theta_{\sigma}\times\Theta_{\alpha}\times\Theta_{\mu}\subseteq{\mathbb R}_{+}^{5}$.
We note that under this representation, for each $t\in[0,T]$, we may treat $\Phi_M(t)$ as a (marginal) random vector of a multivariate $d$-dimensional, $d\leq N$, diffusion process with jumps, for which the component processes are independent, stopped and time-shifted CIR-processes with jumps ($d$ is controlled by the supports $[B_i,D_i)$, $i=1,\ldots,N$).
Note that for the conditional process $\Phi_M(t)|\Phi$, the randomness is present only in the $Y_i(t)$'s.
Moreover, we note that since $\nu(W)<\infty$ and $T<\infty$, we have that $N<\infty$ a.s..
It is this representation of $\Phi_M$ which mainly will be exploited in the statistical inference parts in the remainder of this paper.
\section{Distributional properties of the SG-process and its components}\label{SectionDistribution}
\subsection{Properties of the CIR-process}
Given below are some results concerning different properties of the CIR-process and they can all be found in e.g.\ \cite{CIR,Iacus}.
The explicit solution of the CIR-process, which is given by
\begin{eqnarray*}
\label{SolutionCIR}
Y_{i}(t) &=& K - \left(K - M_{i}^{0}\right)\e^{-t\lambda/K} + \sigma\int_{0}^{t}\e^{(s-t)\lambda/K}\sqrt{Y_{i}(s)}dW_i(s),
\end{eqnarray*}
is obtained by applying Ito's formula with $f(x,t) = x\e^{t\lambda/K}$ to the SDE (\ref{SDE}).
Furthermore, when $2\lambda\geq\sigma^2$ the process a.s.\ stays strictly positive whereas it may reach zero otherwise.
This condition, loosely speaking, says that the drift of the SDE must be large enough, in comparison to the diffusion term, to ensure that the mean-reversion is strong enough to keep the process a.s.\ positive.
Hence, we will require that $2\lambda\geq\sigma^2$ so that $M_i(t)>0$ for all $t\in[B_i,D_i)$.
Since $Y_i(t)$ is a Markov process, when we require that $2\lambda\geq\sigma^2$, it is possible to derive explicit statements about the transition distributions, i.e.\ the distributions of the random variables $Y_{i}(t)|Y_{i}(s)$, $s\leq t$.
For instance, when $s<t$ the conditional expectation and variance are given by
\begin{eqnarray}
\label{MeanCIR}
{\mathbb E}\left[Y_{i}(t)|Y_{i}(s) = y_{s}\right] &=& K - (K - y_{s})\e^{-(t-s)\lambda/K},\\
\Var\left(Y_{i}(t)|Y_{i}(s) = y_{s}\right)
&=& y_{s}\frac{\sigma^2 K}{\lambda}\left(\e^{-(t- s)\lambda/K} - \e^{-2(t-s)\lambda/K}\right)\nonumber\\
&&+ \frac{\sigma^2 K^2}{2\lambda} \left(1 - \e^{-(t-s)\lambda/K}\right)^2,\nonumber
\end{eqnarray}
respectively.
More interesting for our purposes, however, is that under the hypothesis that $2\lambda\geq\sigma^2$ and $s\leq t$, conditional on $Y_{i}(s) = y_{s}$, the transition density of $Y_{i}(t)$ is given by the non-central $\chi^2$-distribution density
\begin{eqnarray}
\label{TransitionCIR}
p_{Y_{i}}(t-s, y_{t}|y_{s};\lambda,K,\sigma) = a \e^{-(u+v)} \left(\frac{v}{u}\right)^{q/2} I_{q}\left(2\sqrt{uv}\right),
\end{eqnarray}
where $a = 2\lambda/\left(\sigma^2 K\left(1-\e^{-(t-s)\lambda/K}\right)\right)$,
$u = a y_{s}\e^{-(t-s)\lambda/K}$,
$v = a y_{t}$ and
$q = 2\lambda/\sigma^2 - 1$.
The function $I_{q}(x)=\sum_{k=0}^{\infty}(x/2)^{2k+q}/k!\Gamma(k+q+1)$, $x\in{\mathbb R}$, where $\Gamma(\cdot)$ denotes the gamma function, is the modified Bessel function of the first kind of order $q$.
This ergodic process also has a stationary (invariant) distribution $\pi=\pi_{\lambda,K,\sigma}$ which is given by the Gamma distribution with shape parameter $2\lambda/\sigma^2$ and scale parameter $\sigma^2 K/2\lambda$.
Hereby, the density of the stationary distribution is given by
\begin{eqnarray}
\label{StatDensity}
\pi(x;\lambda,K,\sigma) = \frac{\left(2\lambda/\sigma^2K\right)^{2\lambda/\sigma^2}}{\Gamma(2\lambda/\sigma^2)} x^{2\lambda/\sigma^2-1}\e^{-x(2\lambda/\sigma^2 K)},\quad x\geq0,
\end{eqnarray}
so that $\pi$ has mean $K$ and variance $\sigma^2 K^2/2\lambda$ and, moreover, for $s<t$, the covariance function of $Y_i(t)$ is given by $\Cov(Y_i(s),Y_i(t)) = \frac{\sigma^2K^2}{2\lambda}\e^{-(t-s)}$.
As previously mentioned, $Y_i(t)$ is a Markov process and given that we start a Markov process in its stationary distribution, it is a strictly stationary process.
In the case of $Y_i(t)$ this means that $Y_i(0) = M_{i}^{0}\sim\pi$ and its finite dimensional distributions (fdds) are shift invariant w.r.t.\ time, i.e.\ $(Y_i(T_1),\ldots,Y_i(T_n))=^d(Y_i(T_1+h),\ldots,Y_i(T_n+h))$ for any set of times $T_1 < \ldots < T_n$, any $h\geq0$ and any $n\in{\mathbb N}$.
Hereby the marginal/transition distributions do not change, i.e.\ for any $(s,t)$, $t>s\geq0$, $Y_i(t)\sim\pi$ and $Y_{i}(t)|Y_{i}(s)\sim\pi$.
\subsection{Properties of the ID-process}
Recall from Section \ref{SectionEvolutionary} the underlying SID-process and its temporal component, the ID-process, $\left\{N(t)\right\}_{t\geq 0}$.
The following result, which can be found in \cite{CronieYu}, gives us the transition probabilities and the stationary distribution of $N(t)$.
\begin{lem}\label{ImDeTransProb}
The transition probabilities of the ID-process, $N(t)$, are given as convolutions of Poisson densities and Binomial densities such that, for $h,t\geq0$ and $x,y\in{\mathbb N}$,
\begin{eqnarray*}
\label{ImmigrationDeathTransProb}
p_{N}(t,y|x;\alpha,\mu) &=& \left(P_{Poi(\rho)} \ast P_{Bin(x,\e^{-\mu t})}\right)\big(y\big) \\
&=& \sum_{k=0}^{y} P_{Poi(\rho)}(k) P_{Bin(x,\e^{-\mu t})}(y-k),
\end{eqnarray*}
where $P_{Poi(\rho)}(\cdot)$ is the Poisson density with parameter $\rho=\alpha\left(1-\e^{-\mu t}\right)/\mu$ and $P_{Bin(x,\e^{-\mu t})}(\cdot)$ is the Binomial density with parameters $x$ and $\e^{-\mu t}$.
Furthermore, the stationary distribution of $N(t)$ is given by $$\pi_{N}(\cdot)={\mathbb P}(Poi(\alpha/\mu)\in\cdot)$$ and the expected value and second moment of the $p_{N}(t,y|x;\alpha,\mu)$-distribution are given by ${\mathbb E}[N(h+t)|N(h)=i]=i\e^{-\mu t} + \rho$
and ${\mathbb E}[N^2(h+t)|N(h)=i] = i(i-1)\e^{-2\mu t} + (1+2\rho)i\e^{-\mu t} + \rho^2 + \rho$, respectively.
\end{lem}
This lemma will be further exploited in Proposition \ref{fidi}, where the fdds of $\Phi_M(t)$ are derived.
\subsection{Finite dimensional distributions of the SG-process}
Consider now the SID-process $\Phi(t)$, or alternatively the index process $\Omega(t)$ and the population size process $N(t)=|\Phi(t)|=|\Omega(t)|$.
Recall that these processes as well as the CIR-process are Markov processes, which in turn implies that also $\Phi_M(t)$ is a Markov process.
This observation will be of great importance since in Proposition \ref{fidi} the Markov property we will be exploited in the derivation of the fdds of $\Phi_M(t)$.
In order to set the framework, we consider the (sample) times $0 = T_0 < T_1 < \ldots < T_n \leq T$ and the distribution of $\left(\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n)\right)^T$, when we are concerned with exactly, say, $d\in\{1,\ldots,N\}$ individuals who appear at $T_1,\ldots,T_n$ (recall that $N$ is the total number of individuals observed if we monitor the process continuously).
Furthermore, provided that the joint density of $\left(\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n)\right)^T$ exists,
when evaluated at the size-time matrix
\begin{eqnarray*}
\mathbf{M} =
\begin{pmatrix}
m_{11} & \cdots & m_{1n}\\
\vdots & \ddots & \vdots\\
m_{d1} & \cdots & m_{dn}
\end{pmatrix}
\in{\mathbb R}^{d\times n},
\end{eqnarray*}
we will denote this density by $\mathbf{p}_{T_1,\ldots,T_n}(\mathbf{M};\theta)$.
It should be emphasised that the $i$th row of $\mathbf{M}$ represents the evaluation-sizes of the $i$th individual under consideration, at the respective times $T_1,\ldots,T_n$.
We further also note that if $m_{ik} = 0$, we are considering the case where the $i$th individual is not alive at time $T_k$.
Hence, if $m_{ik} = 0$ for all $k=1,\ldots,l-1<n$, and $m_{il} > 0$, we evaluate a scenario where $B_i\in(T_{l-1},T_{l}]$, and when $m_{il} > 0$ and $m_{ik} = 0$ for all $k=l+1,\ldots,n$ we consider $D_i\in(T_{l},T_{l+1}]$.
Consequently, if a row were to contain only zeros, we would be considering an individual who is not alive at any of $T_1,\ldots,T_n$, whence that individual/row may be removed from consideration.
The exact form of $\mathbf{p}_{T_1,\ldots,T_n}(\mathbf{M};\theta)$ is given in Proposition \ref{fidi} and the main feature exploited in its derivation is the Markovianity of $\Phi_M(t)$.
We note that the distribution of $\left(\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n)\right)^T$ may be expressed through $\Phi_M(t)$'s transition probabilities/densities, which are given by
\begin{eqnarray}
\label{TransSDE}
{\mathbb P}\left(\Phi_{M}(t)\in\mathbf{A}|\mathcal{F}_s;\theta\right) &=& {\mathbb P}\left(\Phi_{M}(t)\in\mathbf{A}|\Phi_{M}(s); \theta\right),
\end{eqnarray}
where $\mathbf{A} = A_1 \times\ldots\times A_d\in\mathcal{B}({\mathbb R}^{d})$ and $0 \leq s < t \leq T$.
The proof of Proposition \ref{fidi} can be found in the Appendix.
\begin{prop}[Fdds of $\Phi_{M}(t)$]\label{fidi}
Given $0 = T_0 < T_1 < \ldots < T_n \leq T$ and $\Phi_{M}(T_0)$, if we let $M_i^0=M_0>0$ for all $i$, then the joint density of $\left(\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n)\right)^T$, evaluated at $\mathbf{M}\in{\mathbb R}^{d\times n}$, $d\geq1$, is given by
\begin{eqnarray}
\label{JointDensity}
\mathbf{p}_{T_1,\ldots,T_n}(\mathbf{M};\theta)
&=& C
\prod_{k=1}^{n} p_{N}\left(\Delta T_{k}, |\omega_k| \Big| |\omega_{k-1}|;\alpha\nu(W),\mu\right)\\
&&\times \prod_{k=1}^{n} \prod_{i\in\omega_{k-1}\cap\omega_{k}} p_{Y_{1}}(\Delta T_{k}, m_{ik}|m_{i(k-1)};\lambda,K,\sigma)\nonumber\\
&&\times
\prod_{i=1}^{d}
\int_{T_{k_{i}-1}}^{T_{k_{i}}} \frac{p_{Y_{1}}(T_{k_{i}}-t, m_{i(k_{i}-1)}|M_0;\lambda,K,\sigma)}{T_{k_{i}}-T_{k_{i}-1}} dt,\nonumber
\end{eqnarray}
where $\Delta T_{k} = T_{k}-T_{k-1}$, $\omega_k = \{i:m_{ik} > 0\}$, $k=1,\ldots,n$, and $k_{i}=\min\{k:i\in\omega_k\}$, $i=1,\ldots,d$.
The constant $C = C(\nu(W),\mathbf{M})>0$ can be found in expression (\ref{NormalizingConstant}), and the densities $p_{Y_{1}}(\cdot)$ and $p_{N}(\cdot)$ are given, respectively, by expression (\ref{TransitionCIR}) and Lemma \ref{ImDeTransProb}.
\end{prop}
Conditioning on $\Phi_{M}(T_0)$ is reasonable since we in most applications already have all the information about the marked points present at the first sample time point.
Note that if we choose all $M_i^0$ fixed but not necessarily equal, expression (\ref{JointDensity}) only changes in that $M_i^0$ replaces $M_0$.
Furthermore, from the proof of Proposition \ref{fidi} we see that the transition probabilities (\ref{TransSDE}) are obtained by finding
\begin{eqnarray*}
{\mathbb P}\left(\Phi_{M}(t)\in\mathbf{A}|\Phi_{M}(s)={\bf y}; \theta\right)
= \int_{\mathbf{A}} \mathbf{p}_{\Phi_{M}(t)|\Phi_{M}(s)}({\mathbf m}|{\bf y};\theta)d{\mathbf m},
\end{eqnarray*}
where $\mathbf{A} = A_1 \times\ldots\times A_d\in\mathcal{B}({\mathbb R}^{d})$, ${\mathbf m}=(m_1,\ldots,m_d)^T\in{\mathbb R}^{d}$, ${\bf y}=(y_1,\ldots,y_d)^T\in{\mathbb R}^{d}$,
\begin{eqnarray*}
\mathbf{p}_{\Phi_{M}(t)|\Phi_{M}(s)}({\mathbf m}|{\bf y};\theta) &=&
C(\nu(W),{\mathbf m},{\bf y})\;
p_{N}\left(t-s, |\omega({\mathbf m})| \Big| |\omega({\bf y})|; \alpha\nu(W),\mu\right)\\
&&\times
\prod_{i\in\omega({\bf y})\cap\omega({\mathbf m})} p_{Y_{1}}(t-s, m_{i}|y_{i};\lambda,K,\sigma)\\
&&\times
\prod_{i\in\omega({\bf y})^c\cap\omega({\mathbf m})} \frac{1}{t-s} \int_{s}^{t} p_{Y_{1}}(t-v, m_{i}|M_0;\lambda,K,\sigma) dv,
\end{eqnarray*}
$\omega({\mathbf m}) = \{i:m_i>0\}$, $\omega({\bf y}) = \{i:y_i>0\}$ and $C(\nu(W),{\mathbf m},{\bf y})$ is a constant.
As mentioned before, when $M_{i}^{0}\sim\pi$ we have that $Y_i(t)$ is a strictly stationary process and this will have a further impact on the joint densities in Proposition \ref{fidi}.
\begin{cor}
Given the preliminaries and notation of Proposition \ref{fidi}, by instead assuming that $M_{i}^{0}\sim\pi$, the joint density (\ref{JointDensity}) becomes
\begin{eqnarray}
\label{StatJointDensity}
\mathbf{p}_{T_1,\ldots,T_n}(\mathbf{M};\theta)
&=&
C
\prod_{k=1}^{n} p_{N}\left(\Delta T_{k},|\omega_k| \Big| |\omega_{k-1}|;\alpha\nu(W),\mu\right)\nonumber\\
&&\times \prod_{k=1}^{n} \prod_{i\in\omega_{k}} \pi(m_{ik};\lambda,K,\sigma).
\end{eqnarray}
\end{cor}
\begin{proof}
We note that all transition densities $p_{Y_{i}}(\Delta T_{k}, \cdot|\cdot;\lambda,K,\sigma)$ in expression (\ref{MarkJointDensity}) may be replaced by the stationary Gamma densities $\pi(\cdot;\lambda,K,\sigma)$ of expression (\ref{StatDensity}).
\end{proof}
\begin{rem}
We may additionally require that also $N(t)$ starts in its stationary distribution $\pi_{N}$ (see Lemma \ref{ImDeTransProb}) so that also $N(t)$ becomes a strictly stationary process.
Hereby the transition probabilities $p_{N}(\cdot)$ in expression (\ref{StatJointDensity}) will be replaced by $\pi_{N}\left(|\omega_k|;\alpha\nu(W),\mu\right)$.
Note that this change will imply that $N(t)=|\Omega(t)|\sim Poi(\alpha/\mu)$ for all $t\geq0$ and under this setup, since all $Y_i$'s are stationary, we have that $M_{i}(0)\sim\pi$ for all individuals $i\in\Omega(0)$.
\end{rem}
\begin{rem}
As we previously noted, conditionally on $\Omega(0)=\emptyset$, the process $\Xi(t) = \bigcup_{i\in\Omega(t)}B_{X_i}[M_i(t)]$ at each fixed time $t$ corresponds to a Boolean model (see e.g.\ \cite{StoyanKendallMecke}). The \emph{germs} $\{X_i\}_{i\in\Omega(t)}$ are generated from a Poisson process with intensity measure $\Lambda_t(B) = \frac{\alpha}{\mu}(1-\e^{-\mu t})\nu(B\cap W)$, $B\in{\mathcal B}({\mathbb R}^2)$, and the \emph{grains} are given by $\{B_{X_i}[M_i(t)]\}_{i\in\Omega(t)}$, where all $M_i(t)$'s are iid $\Gamma(2\lambda/\sigma^2,\sigma^2 K/2\lambda)$-distributed.
Note that this follows since $\Omega(t)$ can be generated as a thinned Poisson process (see \cite{CronieYu}).
\end{rem}
\section{Maximum likelihood estimation}\label{SectionLikelihood}
Conditionally on $\Phi_{M}(T_0)=\Phi_{M}(0)$, we now assume that we sample the SG-process $\Phi_{M}(t)$ as $M=(\phi_1,\ldots,\phi_n)$ at the sample times $T_1,\ldots,T_n$ on some region $W$.
Here $\phi_k = \left(m_{1k},\ldots,m_{dk}\right)^T$,
$k=1,\ldots,n$, where $d=|\bigcup_{k=1}^{n}\omega_k|$ and $\omega_k = \{\text{indices of individuals present at time } T_k\} = \{i:m_{ik}>0\}$
(we may write $\phi_k = \left({\mathbf 1}_{\omega_k}(1)m_{1k},\ldots,{\mathbf 1}_{\omega_k}(N)m_{dk}\right)^T$ to emphasise the individuals' life status).
Now, based on this sampling scheme we want to find the Maximum Likelihood (ML) estimate of the parameter vector $\theta = (\lambda,K,\sigma,\alpha,\mu)\in\Theta$.
We note that when $\Phi_{M}$ is treated as in Section \ref{SectionFMPP}, i.e.\ as a functional marked point process instead of as an evolutionary process, the estimation based on the current sampling is equivalent to estimating a thinned version of the process.
More specifically, this thinned version is such that all marked points $A=\{[X_i,M_i]:[B_i,D_i)\cap\{T_1,\ldots,T_n\}=\emptyset\}$ are removed and only the partial information $\{(X_i,M_i(T_1),\ldots,M_i(T_n)):i\notin A\}$ is available to estimate the actual structure of the non-thinned process $\Phi_{M}$ (think of this as a sample from the previously mentioned Boolean model which was based on solely the ''alive individuals'').
The likelihood function of the parameters of the SG-process, ${\mathcal L}_n(\theta)={\mathcal L}_n(\theta;M)$, is given by the joint density of $\left(\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n)\right)$, evaluated at $M=\left(\phi_1,\ldots,\phi_n\right)$ and treated as a function of $\theta\in\Theta$.
Therefore, depending on whether we choose $M_{i}(0)$ to be fixed or drawn from the stationary distribution, we end up evaluating either expression (\ref{JointDensity}) or expression (\ref{StatJointDensity}) when we evaluate ${\mathcal L}_n(\theta)$.
\subsection{ML-estimation: $M_{i}^{0}=M_0\in{\mathbb R}_+$}\label{SectionNonStationaryMLE}
When we let all $Y_{i}(0)=M_i^0=M_0\in{\mathbb R}_+$ be given by the same fixed value, from expression (\ref{JointDensity}) we obtain
\begin{eqnarray}
\label{FullLikelihood}
{\mathcal L}_n(\theta) =
C
{\mathcal L}_{1,n}(\theta) {\mathcal L}_{2,n}(\theta) {\mathcal L}_{3,n}(\theta)
\propto {\mathcal L}_{1,n}(\theta) {\mathcal L}_{2,n}(\theta) {\mathcal L}_{3,n}(\theta),
\end{eqnarray}
where, for $k_{i}=\min\{k:i\in\omega_k\}$ and $\Delta T_{k} = T_{k} - T_{k-1}$, $k=1,\ldots,n$,
\begin{eqnarray*}
{\mathcal L}_{1,n}(\theta) &=& \prod_{k=1}^{n} \prod_{i\in\omega_{k-1}\cap\omega_{k}} p_{Y_{1}}(\Delta T_{k}, m_{ik}|m_{i(k-1)};\lambda,K,\sigma)\\
{\mathcal L}_{2,n}(\theta) &=& \prod_{i\in\bigcup_{k=1}^{n}\omega_k} \frac{1}{\Delta T_{k_{i}}} \int_{0}^{\Delta T_{k_{i}}} p_{Y_{1}}(t, m_{i(k_{i}-1)}|M_0;\lambda,K,\sigma) dt\\
{\mathcal L}_{3,n}(\theta) &=& \prod_{k=1}^{n} p_{N}\left(\Delta T_{k},|\omega_k| \Big| |\omega_{k-1}|;\alpha\nu(W),\mu\right).
\end{eqnarray*}
The (rescaled) log-likelihood is given by
\begin{eqnarray*}
\label{FullLogLikelihood}
l_n(\theta) &=&
\log\left(C^{-1}{\mathcal L}_n(\theta)\right)
= \log{\mathcal L}_{1,n}(\theta) + \log{\mathcal L}_{2,n}(\theta) + \log{\mathcal L}_{3,n}(\theta)\nonumber\\
&=:& l_{1,n}(\theta) + l_{2,n}(\theta) + l_{3,n}(\theta),
\end{eqnarray*}
whereby the ML-estimator of $\theta\in\Theta$, based on $\left(\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n)\right)$, will be given by
\begin{eqnarray*}
\widetilde{\theta}_n &:=& \widetilde{\theta}_n\left(\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n)\right)\\
&=& \operatornamewithlimits{arg\;max}_{\theta\in\Theta} l_n(\theta;\Phi_{M}(T_1),\ldots,\Phi_{M}(T_n))\\
&=& \operatornamewithlimits{arg\;max}_{\theta\in\Theta} (l_{1,n}(\theta) + l_{2,n}(\theta) + l_{3,n}(\theta)).
\end{eqnarray*}
We now want to express the ML-estimator $\widetilde{\theta}_n = (\widetilde{\lambda}_n, \widetilde{K}_n, \widetilde{\sigma}_n, \widetilde{\alpha}_n, \widetilde{\mu}_n)$ as the sum of two estimators $\widetilde{\theta}_{1,n}$ and $\widetilde{\theta}_{2,n}$ which, respectively, handle the separate estimation of $(\lambda,K,\sigma)$ and $(\alpha,\mu)$.
We note that $l_{1,n}(\theta) + l_{2,n}(\theta)$, which only involves $\lambda$, $K$ and $\sigma$, will be maximized by any $\widetilde{\theta}_{1,n} = (\widetilde{\lambda}_n,\widetilde{K}_n,\widetilde{\sigma}_n,\alpha,\mu)$, $(\alpha,\mu)\in{\mathbb R}^2$.
Similarly we have that $l_{3,n}(\theta)$, which only involves $\alpha$ and $\mu$, will be maximized by $\widetilde{\theta}_{2,n} = (\lambda,K,\sigma,\widetilde{\alpha}_n,\widetilde{\mu}_n)$, for any $(\lambda,K,\sigma)\in{\mathbb R}^3$.
Hence, in order for $\widetilde{\theta}_{n} = \widetilde{\theta}_{1,n} + \widetilde{\theta}_{2,n}$ to hold, we must require that $\widetilde{\theta}_{1,n} = (\widetilde{\lambda}_n,\widetilde{K}_n,\widetilde{\sigma}_n,0,0)$ and $\widetilde{\theta}_{2,n} = (0,0,0,\widetilde{\alpha}_n,\widetilde{\mu}_n)$, i.e.\
\begin{eqnarray}
\label{MLE}
\widetilde{\theta}_n
&=& \widetilde{\theta}_{1,n} + \widetilde{\theta}_{2,n}\\
&=& \operatornamewithlimits{arg\;max}_{\theta\in\Theta_{\lambda}\times\Theta_{K}\times\Theta_{\sigma}\times\{0\}^2} \left\{l_{1,n}(\theta) + l_{2,n}(\theta)\right\}
+ \operatornamewithlimits{arg\;max}_{\theta\in\{0\}^3\times\Theta_{\alpha}\times\Theta_{\mu}} l_{3,n}(\theta), \nonumber
\end{eqnarray}
and consequently we may estimate the parameters of the ID-process and the parameters related to the mark growth separately.
When the amount of data is large or when the $\Delta T_k$'s are small, we may consider instead the approximate ML-estimation where we set $l_{2,n}(\theta)=0$ so that the only information about the diffusions comes from the observed transitions.
This is reasonable since the amount of information about the actual parameter values which is carried by $l_{2,n}(\theta)$ is not really substantial (in comparison to $l_{1,n}(\theta)$).
Moreover, since there is no closed form expression available for the ML-estimator $(\widetilde{\alpha}_n,\widetilde{\mu}_n)$ of the ID-process (see \cite{CronieYu}), there is also no closed form available for $\widetilde{\theta}_n$ in (\ref{MLE}).
Hence, in modelling situations one has to rely on numerical methods to find $\widetilde{\theta}_n$.
\subsection{ML-estimation: $M_{i}^{0}\sim\pi$}\label{SectionStationaryMLE}
Under the assumption that we start the diffusions in their stationary distributions, $M_{i}^{0}\sim\pi$, from expression (\ref{StatJointDensity}) we obtain the likelihood function
\begin{eqnarray}
\label{StatFullLikelihood}
{\mathcal L}_n(\theta) = C {\mathcal L}_{1,n}(\theta) {\mathcal L}_{2,n}(\theta)
\propto
{\mathcal L}_{1,n}(\theta) {\mathcal L}_{2,n}(\theta)
\end{eqnarray}
and the (rescaled) log-likelihood
\begin{eqnarray*}
\label{FullLogLikelihood}
l_n(\theta) &=& \log\left(C^{-1}{\mathcal L}_n(\theta)\right)
= \log{\mathcal L}_{1,n}(\theta) + \log{\mathcal L}_{2,n}(\theta)\nonumber\\
&=:&
l_{1,n}(\theta) + l_{2,n}(\theta),\qquad
\end{eqnarray*}
where, for $\Delta T_{k} = T_{k} - T_{k-1}$, $k=1,\ldots,n$,
\begin{eqnarray*}
l_{1,n}(\theta) &=& \log\left(\prod_{k=1}^{n} \prod_{i\in\omega_{k}} \pi(m_{ik};\lambda,K,\sigma)\right)
=\sum_{k=1}^{n} \sum_{i\in\omega_{k}} \log\pi(m_{ik};\lambda,K,\sigma)\\
l_{2,n}(\theta) &=& \log\left(\prod_{k=1}^{n} p_{N}\left(\Delta T_{k},|\omega_k| \Big| |\omega_{k-1}|;\alpha\nu(W),\mu\right)\right)\\
&=& \sum_{k=1}^{n} \log p_{N}\left(\Delta T_{k},|\omega_k| \Big| |\omega_{k-1}|;\alpha\nu(W),\mu\right).
\end{eqnarray*}
Here, just as in the fixed initial value case of Section \ref{SectionNonStationaryMLE}, we deal with the separate estimators
\begin{eqnarray}
\label{StatMLE}
\hat{\theta}_n &=& \hat{\theta}_{1,n} + \hat{\theta}_{2,n}\\
&=& \operatornamewithlimits{arg\;max}_{\theta\in\Theta_{\lambda}\times\Theta_{K}\times\Theta_{\sigma}\times\{0\}^2} l_{1,n}(\theta)
+ \operatornamewithlimits{arg\;max}_{\theta\in\{0\}^3\times\Theta_{\alpha}\times\Theta_{\mu}} l_{2,n}(\theta)\nonumber
\end{eqnarray}
and, similarly, there is no closed form expression available for $\hat{\theta}_n$.
\section{Asymptotic inference under stationarity}\label{SectionAsymptotics}
When dealing with asymptotic spatial statistics, there are different types of asymptotics which may be considered.
In the case of the SG-process, within the framework of so called increasing domain asymptotics (see e.g.\ \cite{Zhang}), there are essentially two different ways to increase the total number of individuals observed, and consequently also the number of transitions taking place between pairs of consecutive sample times $T_{k-1}$ and $T_{k}$.
The first approach is to increase the number of sample times of the mark processes, i.e.\ we let $n$ grow, whereby $T_n=T$ also will grow.
The second approach is to gradually increase the size of the sampling window $W$ (with the number of sample times fixed).
The two approaches are similar since in both cases we increase the parameter of the Poisson distribution of $N\sim Poi(\alpha T\nu(W))$.
Here, we choose to consider only the first of the two alternatives.
Consider the situation where we, without loss of generality, let $W=[0,1]^2$ and apply the equidistant sampling scheme $T_k = k\Delta$, $k=1,\ldots,n$, $\Delta>0$, where $T = T_n = n\Delta$.
In what follows we denote by $\theta_0=(\lambda_0,K_0,\sigma_0,\alpha_0,\mu_0)\in\Theta$ the true/underlying parameter vector which is responsible for generating $\Phi_M$ and we assume that $\Theta$ is a subset of ${\mathbb R}_{+}^5$ such that
\begin{eqnarray}
\label{ParamSpace}
\Theta\cap\{(\lambda,K,\sigma,\alpha,\mu)\in{\mathbb R}_+^5:2\lambda<\sigma^2\} = \emptyset.
\end{eqnarray}
Recall that this is required to keep the $Y_i(t)$'s positive.
In the theorems and corollaries below we give the strong consistency and the asymptotic normality of the ML-estimator.
The proofs are given in the Appendix.
The consistency proof follows the approach suggested by Wald \cite{Wald} and the asymptotic normality follows the lines of the classical approach of Cram\'er (see e.g.\ \cite{Ferguson}).
\begin{thm}[Consistency]\label{ThmCons}
Let $\Theta$ be a compact subset of ${\mathbb R}_+^5$ such that (\ref{ParamSpace}) holds.
Then, for $\theta_0\in\Theta$, the estimator $\hat{\theta}_n$ in expression (\ref{StatMLE}) is strongly consistent, i.e.\ as $n\rightarrow\infty$,
\begin{center}
$\hat{\theta}_n \stackrel{a.s.}{\longrightarrow} \theta_0$.
\end{center}
\end{thm}
Now, by putting some additional restrictions on the parameters we may also prove the following theorem.
\begin{thm}[Asymptotic normality]\label{ThmAsNorm}
Let $\theta_0$ be an interior point of $\Theta$, where $\Theta$ is a compact subset of ${\mathbb R}_+^5$ such that (\ref{ParamSpace}) holds.
Require further that $\theta_0$ and $\Delta>0$ are such that $(\log(\alpha_0 + \mu_0) - \log(\alpha_0))/\mu_0 \geq 2\Delta$.
Assume that $\lambda_0$ is known, so that $\hat{\theta}_n = (\hat{K}_n, \hat{\sigma}_n, \hat{\alpha}_n, \hat{\mu}_n)$ is the ML-estimator of $\theta_0 = (K_0,\sigma_0,\alpha_0,\mu_0)$.
Then, as $n\rightarrow\infty$, we obtain
\begin{eqnarray*}
\sqrt{n}\big(\hat{\theta}_{n} - \theta_0\big)
\stackrel{d}{\longrightarrow}
\mathbf{Y}\sim
N\left(\mathbf{0}_{4\times1},
\begin{bmatrix}
\frac{\mu_0}{\alpha_0}\frac{K_{0}^2\sigma_{0}^2}{2\lambda_{0}} & 0 & \mathbf{0}_{1\times2}\\
0 & \frac{\mu_0}{\alpha_0}\frac{\sigma_{0}^4}{8\lambda_{0}C(\theta_0)} & \mathbf{0}_{1\times2}\\
\mathbf{0}_{2\times1} & \mathbf{0}_{2\times1} & I_N(\theta_0)^{-1}
\end{bmatrix}
\right),
\end{eqnarray*}
where $C(\theta) = \frac{2\lambda}{\sigma^2}\psi'\left(\frac{2\lambda}{\sigma^2}\right) - 1 > 0$, $\psi(x) = \Gamma'(x)/\Gamma(x)$, $\Gamma(\cdot)$ is the gamma function,
$\mathbf{0}_{i\times j}$ denotes the $i\times j$ zero matrix and the $2\times2$ matrix $I_N(\theta_0)^{-1}$, which can be found in expression (\ref{InvFisherImDea}), is the covariance matrix related to the ID-process.
Similarly, when $\sigma_0$ is known, we estimate $\theta_0 = (\lambda_0,K_0,\alpha_0,\mu_0)$
by means of the ML-estimator $\hat{\theta}_n = (\hat{\lambda}_n, \hat{K}_n, \hat{\alpha}_n, \hat{\mu}_n)$ and, as $n\rightarrow\infty$, we obtain
\begin{eqnarray*}
\sqrt{n}\big(\hat{\theta}_{n} - \theta_0\big)
\stackrel{d}{\longrightarrow}
\mathbf{Y}\sim
N\left(\mathbf{0}_{4\times1},
\begin{bmatrix}
\frac{\mu_0}{\alpha_0}\frac{\lambda_{0}\sigma_{0}^2}{2 C(\theta_0)} & 0 & \mathbf{0}_{1\times2}\\
0 & \frac{\mu_0}{\alpha_0}\frac{K_{0}^2\sigma_{0}^2}{2\lambda_{0}} & \mathbf{0}_{1\times2}\\
\mathbf{0}_{2\times1} & \mathbf{0}_{2\times1} & I_N(\theta_0)^{-1}
\end{bmatrix}
\right).
\end{eqnarray*}
The Fisher information for the discretely sampled ID-process is given by
\begin{eqnarray}
\label{FisherImDea}
I_N(\theta_0) &=&
\begin{pmatrix}
I_N(\theta_0)_{11} & I_N(\theta_0)_{12}\\
I_N(\theta_0)_{12} & I_N(\theta_0)_{22}
\end{pmatrix},
\end{eqnarray}
where
\begin{eqnarray*}
I_N(\theta_0)_{11} &=& (\Xi-1)\frac{\rho_{0}^2}{\alpha_{0}^2},
\qquad
I_N(\theta_0)_{12} = \frac{(\Xi-1)\rho_{0}(\mu_0\Delta - \tau_{0}) - \mu_{0}\Delta}{\mu^2_{0}},\\
I_N(\theta_0)_{22} &=& \frac{\alpha_0^2 \mu_0 \Delta(2\tau_0 -\mu_0t)}{\rho_0 \mu_0^4} + \frac{\alpha_0^2 \Delta^2 \e^{-\mu_0 \Delta}}{\mu_0^2 \rho_0}
+ \frac{(\Xi-1)\alpha_{0}^2\left(\tau_{0} - \mu_{0}\Delta\right)^2}{\mu_{0}^4},
\end{eqnarray*}
$\Xi = \sum_{i, j\in{\mathbb N}} \frac{p_{N}\left(\Delta,j-1|i;\alpha_0,\mu_0\right)^2}{p_{N}\left(\Delta,j|i;\alpha_0,\mu_0\right)} \pi_{N}(i;\alpha_0,\mu_0)$,
$\tau_0 = 1 - \e^{-\mu_0 \Delta} - \mu_0 \Delta\e^{-\mu_0 \Delta}$,
$\rho_{0} = \frac{\alpha_0}{\mu_0}(1 - e^{-\Delta\mu_0})$,
and its inverse is given by
\begin{equation}
\label{InvFisherImDea}
\begin{array}{l}
I_N(\theta_0)^{-1} = \frac{\mu_0}{\Delta\left(\left(1 + \e^{-\mu_0 \Delta}\right)\rho_0(\Xi - 1) - 1\right)}\\
\times
\begin{pmatrix}
\frac{\rho_0\left(2\tau_0 - \mu_0 \Delta\left(1 - \e^{-\mu_0 \Delta}\right)\right) + \frac{\rho_0^2}{\mu_0 \Delta}(\Xi - 1)(\tau_0 - \mu_0 \Delta)^2}{\left(1 - \e^{-\mu_0 \Delta}\right)^2}
& 1 + \frac{\rho_0(\Xi - 1)(\tau_0 - \mu_0 \Delta)}{\mu_0 \Delta} \\
1 + \frac{\rho_0(\Xi - 1)(\tau_0 - \mu_0 \Delta)}{\mu_0 \Delta}
& \frac{(\Xi - 1)\left(1 - \e^{-\mu_0 \Delta}\right)^2}{\mu_0 \Delta}
\end{pmatrix}.
\end{array}
\end{equation}
\end{thm}
The reason that we require knowledge of either $\lambda_0$ or $\sigma_0$ in Theorem \ref{ThmAsNorm} is related to the over parametrization of the $\Gamma(\beta_1,\beta_2)$-distribution, $\beta_1 = 2\lambda/\sigma^2$, $\beta_2 = \sigma^2 K/2\lambda$.
We note to begin with that $\beta_2 = K/\beta_1$ and for a random variable $Z\sim\Gamma(\beta_1,\beta_2)$, by consulting expression (\ref{2orderPartialDeriavtives}) in the Appendix, we obtain the related positive semi-definite singular (non-invertible) Fisher information
\begin{eqnarray*}
\label{FisherGamma}
I_Z(\theta) = -{\mathbb E}_{\theta}\left[\frac{\partial^2\log\pi(Z;\lambda,K,\sigma)}{\partial(\lambda,K,\sigma)^2}\right]
=
\frac{2 C(\theta)}{\sigma^2}
\begin{pmatrix}
\frac{1}{\lambda}
& 0
& \frac{-2}{\sigma}\\
0
& \frac{\lambda}{K^2}C(\theta)^{-1}
& 0 \\
\frac{-2}{\sigma}
& 0
& \frac{4\lambda}{\sigma^2}
\end{pmatrix}.
\end{eqnarray*}
\begin{rem}
From the proofs of Theorem \ref{ThmCons} and Theorem \ref{ThmAsNorm} it may be seen that if we reduce $l_{1,n}(\theta)$ to
\begin{eqnarray*}
\widetilde{l}_{1,n}(\theta) &=& \sum_{k=1}^{n} \sum_{i\in\widetilde{\omega}_{k}} \log\pi(m_{ik};\lambda,K,\sigma),
\end{eqnarray*}
where $\widetilde{\omega}_{k}\subseteq\omega_{k}$ ($\emptyset=\widetilde{\omega}_{k}$ iff $\omega_{k}=\emptyset$), the proofs of the consistency and the asymptotic normality still go through (with obvious modifications).
However, the convergence speed will be different as well as the Fisher information $I(\theta_0)$.
An example of such a reduction is to choose $\widetilde{\omega}_{k} = \{\omega_{k,1}\}$, i.e.\ we choose just one element from $\omega_{k}$.
Another example of a reduction under which the results still hold is to consider the subsequence $k_n=n\wedge A(n)$, $A(n)=\sum_{k=1}^{n}|\omega_{k}|$, and the reduction
\begin{eqnarray*}
\widetilde{l}_{1,n}(\theta) &=& \sum_{i=1}^{k_n} \log\pi(m_{ik};\lambda,K,\sigma).
\end{eqnarray*}
\end{rem}
\section{Evaluation of the estimators}\label{SectionEvaluation}
We now turn to the numerical evaluation of our ML-estimators and precisely we are interested in investigating the asymptotic robustness of the stationarity assumption.
This is carried out by assuming that the data is generated with some fixed $M_i^0=M_0$ and some $\theta_0$, while we instead are employing the estimator $\hat{\theta}_n$ in expression (\ref{StatMLE}) of Section \ref{SectionStationaryMLE}, i.e.\ the estimator based on the assumption that $M_i^0\sim\pi$, to estimate $\theta_0$.
We then compare the behaviours of $|\hat{\theta}_n-\theta_0|$ and $|\widetilde{\theta}_n-\theta_0|$.
We first note that we from expression (\ref{MeanCIR}) may conclude that $\e^{-t\lambda_0/K_0} = |K_0 - {\mathbb E}[Y_{i}(t)|Y_{i}(0) = M_i^0]|/|K_0 - M_i^0|$.
Clearly, if $|K^0 - M_i^0|$ is small then $Y_i(t)$ quickly approaches its steady state $K_0$, whence the distance $|\widetilde{\theta}_n-\hat{\theta}_n|$ between the two estimators should be reduced.
The same should hold if (additionally) $\lambda_0$ is large, since under this condition the mean reversion is strong, which results in small deviations from the long term mean $K_0$.
Similarly, if $\sigma_0$ is small then the random fluctuations do not influence the growth as much as the drift coefficient $\lambda_0(1-Y_i(t)/K_0)$ and hereby the drift becomes the main determining factor of the speed of convergence to $K_0$.
We also note that if $\mu_0$ is small then the expected lifetime of an individual, ${\mathbb E}_{\theta_0}[L_i] = 1/\mu_0$, tends to be longer whereby we obtain more samples of $Y_i(t)$ when it is close to its steady state $K_0$.
We simulate $30$ trajectories of $\Phi_M(t)$ on $W=[0,1]^2$ and sample them discretely according to the sampling scheme $T_k=k$, $k=1,\ldots,100$.
Then, by using the stationary ML-estimator $\hat{\theta}_n$ of Section \ref{SectionStationaryMLE}, we reestimate the parameters and compare the behaviour of $\hat{\theta}_n$ with the (non-stationary) ML-estimator $\widetilde{\theta}_n$ of Section \ref{SectionNonStationaryMLE}.
We use different values for the parameters $\theta_0$ and $M_i^0$ to assess when $|\hat{\theta}_n-\theta_0|$ and $|\widetilde{\theta}_n-\theta_0|$ are small.
We here only consider the estimation of $\lambda_0$, $K_0$ and $\sigma_0$ since the performance of $(\hat{\alpha}_n,\hat{\mu}_n)$ already has been evaluated in \cite{CronieYu}.
As we can see in Table \ref{TableEvaluation}, as expected, in the case of $\hat{\theta}_n$ the main determining factor of the bias is the size of $\lambda_0$, although the size of $|K-M_i^0|$ certainly plays a role.
It should also be noted that a higher $\sigma_0$ seems to imply a lower bias for $\hat{\sigma}_n$.
Furthermore, we see that $\widetilde{\theta}_n$ outperforms $\hat{\theta}_n$ in each case given in Table \ref{TableEvaluation}, which is to be expected since $\widetilde{\theta}_n$ is the estimator which is based on the correct model assumption.
Note, however, that there are parameter choices for which even the performance of $\widetilde{\theta}_n$ is a bit poor.
\begin{table}[htbp]
\centering
\begin{tabular}{lccclccc}
\hline
$M_i^0 = 0.1$ & $\lambda$ & $K$ & $\sigma$ & $M_i^0 = 0.1$ & $\lambda$ & $K$ & $\sigma$\\
\hline
True ($\theta_0$) & 0.5 & 5 & 0.1 & True ($\theta_0$) & 0.5 & 5 & 0.1 \\
Mean $\widetilde{\theta}_n$ & 0.5027 & 5.1698 & 0.1086 & Mean $\hat{\theta}_n$ & 3.2856 & 3.9807 & 0.9761 \\
Bias $\widetilde{\theta}_n$ & 0.5\% & 3.4\% & 8.6\% & Bias $\hat{\theta}_n$ & 557.1\% & -20.4\% & 876.1\% \\
S.e. $\widetilde{\theta}_n$ & 0.0605 & 0.5623 & 0.0139 & S.e. $\hat{\theta}_n$ & 1.3928 & 0.1266 & 0.2480 \\
\hline
$M_i^0 = 5$ & $\lambda$ & $K$ & $\sigma$ & $M_i^0 = 5$ & $\lambda$ & $K$ & $\sigma$\\
\hline
True ($\theta_0$) & 0.5 & 5 & 0.1 & True ($\theta_0$) & 0.5 & 5 & 0.1 \\
Mean $\widetilde{\theta}_n$ & 0.4241 & 5.0385 & 0.1006 & Mean $\hat{\theta}_n$ & 2.9054 & 4.9987 & 0.2185 \\
Bias $\widetilde{\theta}_n$ & -15.2\% & 0.8\% & 0.6\% & Bias $\hat{\theta}_n$ & 481.1\% & -0.03\% & 118.5\% \\
S.e. $\widetilde{\theta}_n$ & 0.1981 & 0.4780 & 0.0063 & S.e. $\hat{\theta}_n$ & 1.2660 & 0.0539 & 0.0540 \\
\hline
$M_i^0 = 0.1$ & $\lambda$ & $K$ & $\sigma$ & $M_i^0 = 0.1$ & $\lambda$ & $K$ & $\sigma$\\
\hline
True ($\theta_0$) & 3 & 5 & 0.1 & True ($\theta_0$) & 3 & 5 & 0.1 \\
Mean $\widetilde{\theta}_n$ & 2.9950 & 4.9926 & 0.1036 & Mean $\hat{\theta}_n$ & 3.1261 & 4.8713 & 0.2425 \\
Bias $\widetilde{\theta}_n$ & -0.2\% & -0.1\% & 3.6\% & Bias $\hat{\theta}_n$ & 4.2\% & -2.6\% & 142.5\% \\
S.e. $\widetilde{\theta}_n$ & 0.2196 & 0.0708 & 0.0121 & S.e. $\hat{\theta}_n$ & 1.2823 & 0.0320 & 0.0569 \\
\hline
$M_i^0 = 0.1$ & $\lambda$ & $K$ & $\sigma$ & $M_i^0 = 0.1$ & $\lambda$ & $K$ & $\sigma$\\
\hline
True ($\theta_0$) & 3 & 5 & 0.5 & True ($\theta_0$) & 3 & 5 & 0.5 \\
Mean $\widetilde{\theta}_n$ & 2.9866 & 5.0513 & 0.4974 & Mean $\hat{\theta}_n$ & 2.7822 & 4.9437 & 0.5126 \\
Bias $\widetilde{\theta}_n$ & -0.4\% & 1.0\% & -0.5\% & Bias $\hat{\theta}_n$ & -7.3\% & -1.1\% & 2.5\% \\
S.e. $\widetilde{\theta}_n$ & 0.2883 & 0.1505 & 0.0226 & S.e. $\hat{\theta}_n$ & 1.2616 & 0.1524 & 0.1288 \\
\hline
\end{tabular}
\caption{Parameter (re)estimation of $\lambda_0$, $K_0$ and $\sigma_0$ using the non-stationary ML-estimator $\widetilde{\theta}_n$
and the stationary ML-estimator $\hat{\theta}_n$.
The estimates are based on 30 realisations of $\{\Phi_M(t)\}_{t\in[0,100]}$ (non-stationary) sampled at $T_k=k$, $k=1,\ldots,100$
(with time discretisation step $dt=0.01$) which are generated from $\alpha_0=0.5$, $\mu_0=0.01$, $W=[0,1]^2$ and the above parameters ($M_i^0$ and $\theta_0$).
}
\label{TableEvaluation}
\end{table}
\section{Modelling Scots pines}\label{SectionPines}
As previously mentioned, the SG-process is constructed as a stochastic extension of the GI-process, under the assumption that the interaction between the marks is negligible.
Hence, when considering the GI-process' main application area, which is the dynamical modelling of forest stands, it makes sense to employ the stationary mark SG-process when we want to model a homogenous forest stand (trees of the same species with similar ages) where e.g.\ the distances between the trees are large (we may ignore the interaction).
One data set which (arguably) may be considered to fulfil these requirements is the set of Swedish Scots pines considered in \cite{Cronie}, which is illustrated in Figure \ref{PinesExample} (all tree radii have been scaled by a factor of 10 for increased visibility).
The spatial region $W$ under consideration here is given by a circular region of radius 10 meters and the actual data set is given by a time series of marked point patterns, recorded at the years 1985, 1990 and 1996, where the approximate age of the forest stand in 1985 was 22 years.
Hereby we may set $T_1 = 22$, $T_2 = 27$ and $T_3 = 33$, and we have $N(T_1)=13$, $N(T_2)=26$ and $N(T_3)=43$.
To be precise, for each $T_k$, $k=1,2,3$, each marked point pattern consists of measurements
of radii (at breast height) $m_{ik}$ and locations (stock centres) $x_i\in W$ of the trees $i\in\{j:m_{jk}>0\}$ which are present at $T_k$, and only trees having reached a radius of 0.005 meter are included in the data set.
\begin{figure}[!htbp]
\begin{center}
\begin{tabular}{ccc}
{\includegraphics[width=0.3\textwidth]{Pines1985Scaled10new.eps}} &
{\includegraphics[width=0.3\textwidth]{Pines1990Scaled10new.eps}} &
{\includegraphics[width=0.3\textwidth]{Pines1996Scaled10new.eps}}
\end{tabular}
\end{center}
\caption{Swedish Scots pines: plots recorded in $1985$ (left), $1990$ (middle) and $1996$ (right). The radii of the pines are scaled by a factor of 10.}
\label{PinesExample}
\end{figure}
The approach used in \cite{Cronie} to model this data set was to employ the so-called logistic growth function $f(Y_{i}(t);\theta) = \lambda Y_{i}(t)\left(1-Y_{i}(t)/K\right)$ as individual/open growth function in (\ref{GI})
and the so-called area-interaction function $h(\cdot)$ (see expression (\ref{GI})) to describe the spatial interaction between the marked points.
We note that both this individual growth function and the (linear growth function) drift coefficient $f(Y_{i}(t);\theta) = \lambda\left(1-Y_{i}(t)/K\right)$ in the CIR process are special cases of the so-called Von Bertalanffy-Chapman-Richards (VBCR) growth function (see e.g.\ \cite{RenshawComasMateu}), whence their behaviours are quite similar.
As previously mentioned, besides $\alpha$ and $\mu$, the parameters under consideration here are the growth rate $\lambda$, the carrying capacity $K$ and the diffusion parameter $\sigma$.
In Table \ref{ResultsPines} we find, together with the results obtained in \cite{Cronie}, the results obtained after having fit the SG-process to the data set in Figure \ref{PinesExample}.
Note that the choice $M_i^0=0.005$ has been made (in the non-stationary SG-process and in the GI-process) since the trees in the data set have been measured only once they have grown to at least a radius (at breast height) of 0.05 meter.
Regarding the estimation of $(\alpha,\mu)$, \cite{Cronie} obtained $\hat{\alpha} = 0.0042$ and $\hat{\mu} = 0$ (based on the estimators given therein).
Here, we obtain $(\hat{\alpha},\hat{\mu})=(0.0613, 0.7020)$ whence, once the forest stand has become old, we would expect $\hat{\alpha}\nu(W)/\hat{\mu} \approx 27$ trees in $W$.
\begin{table}[htbp]
\centering
\begin{tabular}{lccc}
\hline
&$\hat{\lambda}$ & $\hat{K}$ &$\hat{\sigma}$\\
\hline
GI & 0.078 & 0.095 & -- \\
SG & 0.371 & 0.073 & 0.151 \\
Stationary SG & 1.269 & 0.062 & 0.218 \\
\hline
\end{tabular}
\caption{Results obtained after fitting the stationary mark SG-process and the GI-process (results from \cite{Cronie}) to a data set of Swedish Scots pines. }
\label{ResultsPines}
\end{table}
It comes as no surprise that $\hat{K}$ is larger in the GI-process than in the SG-process.
This follows since in the GI-process, the estimation of the open growth ($\lambda$ and $K$ in the logistic growth function) takes into account also that the observed sizes $m_{ik}$ are results of an open growth which has been inhibited by spatial interaction, i.e.\ $f(\cdot)$ is inhibited by $h(\cdot)$.
Since $\max_{i,k}m_{ik}=0.0860$ (see \cite{Cronie}) it is probable that the SG-process underestimates $K$ a bit.
Moreover, by comparing the results for the stationary and the non-stationary SG-process, we conclude that an increased $\hat{\sigma}$ (larger fluctuations) for the stationary case also results in a stronger estimated mean reversion (increased $\hat{\lambda}$).
From the differences in $\hat{\lambda}$ and $\hat{\sigma}$ for the two SG-processes
we have indications that the data set has not (yet) reached stationarity, which is to be expected since the forest stand we are considering is quite young.
In conclusion, mainly due to the difference in $\hat{K}$ between the SG-process and the GI-process as well as the sensibility of having stochastic marks in the GI-process,
this pilot study certainly motivates a further investigation of the applicability of the full SGI-process, where we include an interaction function $h(\cdot)$ in the drift term of each diffusion $M_i(t)$, $t\in[B_i,D_i)$, i.e.\ where we add a stochastic integral term $\int_{B_i}^{D_i}\sigma(M_i(t))dW_i(t)$ to expression (\ref{GI}).
\section{Discussion}\label{SectionDiscussion}
We have here considered the GI-process only in the context of the CIR-mark process, but we may just as well employ any other positive diffusion for the growth of the marks.
As previously noted, the linear growth function, which is the drift function in the CIR-process, is a special case of the Von Bertalanffy-Chapman-Richards (VBCR) growth function (see e.g.\ \cite{RenshawComasMateu}).
Another special case of the VBCR growth function is the
aforementioned logistic growth function $f(Y_{i}(t);\theta) = \lambda Y_{i}(t)\left(1-Y_{i}(t)/K\right)$ which has been used in the GI-process in e.g.\ \cite{Cronie,RS1,RS2}.
A further modification which may be made is to change the diffusion term $\sigma(Y_{i}(t);\theta) = \sigma\sqrt{Y_{i}(t)}$ into any other diffusion term which keeps $Y_{i}(t)$ positive, e.g.\ $\sigma(Y_{i}(t);\theta) = \sigma Y_{i}(t)^{\gamma}$, $\gamma>0$, which is the diffusion coefficient found in the CKLS-model (see e.g.\ \cite{CKLS}).
Note that when applying these changes, we would typically not have known closed form expressions for the transition densities, $p_{Y_i}(t,y_1|y_0;\theta)$.
The transition densities are know only for a few special cases, including the CIR-process. Therefore, we have to use different approximated/pseudo likelihood methods for the estimation of the parameters (see \cite{Iacus} for a good general overview).
Our final goal is to ML-estimate all parameters of the full SGI-process, i.e.\ to include also the spatial interaction function $h(\cdot)$ in expression (\ref{GI}).
Here the lack of closed form expressions for the transition densities remains and, just as for the previous adjustments suggested, the estimation requires that we employ approximated/pseudo likelihood methods.
For instance, \cite{Ait} suggests an approach where the transition densities of multivariate diffusions may be approximated by series expansions based on hermite polynomials.
Note further that within this setting, in order to reduce edge effects (absence of individuals outside the boundary of $W$), it would be sensible to choose $W$ to be a torus.
Also, thus far we have introduced the type of death which occurs when $t\geq D_i$, i.e.\ the life-time of the individual has expired.
Following the terminology of \cite{RS2}, we can refer to this type of death as \emph{natural death}.
It is possible, however, to introduce another type of death, namely so called \emph{competitive death} (or \emph{interactive death}), and its introduction entails a slightly different formulation of the diffusions $M_{i}(t)$, $i=1,\ldots,N$.
By defining the \emph{death-time} of individual $i$ to be (the stopping-time) $\zeta_i = \inf\{t>B_i:M_{i}(t) = 0\}\wedge D_i$, we have that
if $M_i$ reaches the absorbing state $M_{i}(t)=0$ for some $t\in(B_i,D_i)$ it stays 0 and we say that it has suffered a competitive death.
Furthermore, if it does not die from competition in $(B_i,D_i)$ it will still die at time $D_i$, i.e.\ at its natural death time.
As soon as $t>\zeta_i$ the interaction between $M_{i}(t)$ and the other marks will terminate, hence we remove individual $i$ from consideration.
\section*{Acknowledgements}
The author would like to thank Peter Guttorp (University of Washington), Aila S\"arkk\"a (Chalmers university of technology) and Jun Yu (Swedish university of agricultural sciences) for useful suggestions and discussions.
|
\section{Introduction} \label{sec:intro}
Asymptotic-waveform evaluation (AWE) is a long standing challenge in the
computation of waves. Whether for acoustic, electromagnetic,
or gravitational waves the goal is to identify the far-field
or asymptotic signal radiated to future null infinity $\mathscr{I}^+$
using only knowledge of the
solution on a truncated, spatially
finite, computational domain. For the unit-speed ordinary
wave equation, the far-field signal is $f(u,\theta,\phi) =
\lim_{r\rightarrow\infty}r\psi(u,r,\theta,\phi)$, where
$\psi(u,r,\theta,\phi)$ is a solution to the wave equation
written with respect to retarded-time $u=t-r$ and spherical
polar coordinates $(r,\theta,\phi)$. Continuing with this
example, we view computation of
$r_\infty \psi(u,r_\infty,\theta,\phi)$ as AWE, so long as
the {\em evaluation radius} $r_\infty$ can be taken arbitrarily
large, even if ultimately finite. Indeed, for the ordinary
wave equation the asymptotic signal and signal at
$r_\infty = 10^{15}$ are identical to about double precision machine
epsilon (see Appendix~\ref{sec:ErrorEst} for the error estimate).
The asymptotic signal we compute corresponds to
observation on the timelike hyper-cylinder $r = r_\infty$.
In the perturbative gravitational setting considered here, with
$r_\infty = 2M \left(1 \times 10^{15}\right)$ in terms of
the Schwarzschild mass $M$, such observations take place
in the {\em astrophysical zone}~\cite{Leaver,Barack1999,Purrer:2004nq}.
An observation in the astrophysical zone is better approximated
as taking place at $\mathscr{I}^+$ rather than future
timelike infinity $i^+$ \cite{Zenginoglu:2008wc,Leaver,Purrer:2004nq,Barack1999}.
For this reason, and for clarity of exposition, throughout this
paper we refer to such observations as taking place at $\mathscr{I}^+$.
Identifying the gravitational wave signal at $\mathscr{I}^+$ is of
both theoretical and practical importance. Theoretically, in
the context of asymptotically flat spacetimes Sachs \cite{Sachs:1962}
identified the asymptotic metric factors corresponding to
$f(u,\theta,\phi)$ for gravitation, and exploited this
identification in his discussion of the radiative degrees
of freedom for general relativity. Using Geroch's
calculation framework \cite{Geroch}, Ashtekar and Streubel
expanded on the Sachs approach in their fundamental analysis
\cite{AshtekarStreubel} of the symplectic structure of
radiative modes and gravitational flux in general relativity.
Several works \cite{DrayStreubel,Dray,Shaw}
then investigated the general charge integral (where the
integration is over a two-surface ``cut" of $\mathscr{I}^+$) corresponding
to the Ashtekar-Streubel flux. The practical importance of AWE stems
from the upcoming generation of advanced-sensitivity ground-based
gravitational wave interferometer detectors
(i.e., advanced LIGO, advanced Virgo, and KAGRA)
\cite{KAGRA_web,LIGO_web,VIRGO_web,GEO_web}
and anticipated space-based detectors
like LISA \cite{LISA_web,ESA_web}. These
instruments are well-modeled as idealized observers located
at $\mathscr{I}^+$.
While the problem of AWE for
gravitation shares difficulties with its counterparts for
the ordinary wave and Maxwell equations, for example the slow
fall-off of the waves (in our case metric perturbations) in
powers of $1/r$, the gravitational problem is further
complicated by the backscattering of waves, coordinate
(gauge) issues, and non-linearities.
For a perturbed Schwarzschild blackhole, covariant and gauge
invariant approaches exist for the construction of ``master
functions" from the spacetime metric perturbations (see, for
example, \cite{Sarbach:2001qq,Martel_CovariantPert,SopuertaLaguna}).
In the asymptotic limit these master functions specify the
gravitational waveform. Here we consider the wave equations
which directly govern these master functions.
In this perturbative setting,
AWE is precisely the technique (perhaps extrapolation,
for example) used to compute the master functions at
arbitrarily large distances from the central blackhole.
This paper introduces a new technique based on signal
{\em teleportation} between two finite radial values.
A straightforward and longstanding approach to AWE in both
full general relativity \cite{Pollney:2009yz,Boyle2009}
and perturbative settings \cite{Sundararajan:2007jg} has been
to record relevant field quantities at a variety of radii,
perform a numerical fit, and then extrapolate to larger radii.
However, the accuracy of this method ultimately relies on an
{\em Ansatz} for the expected fall-off of the field with
larger $r$, as well as recording field values at multiple and
preferably large values of $r$ \cite{Pollney:2009ut}.
An alternative approach, known as Cauchy characteristic
extraction, is to record geometrical data from a Cauchy evolution
on a world-tube, which is later used as interior boundary data
for a second characteristic evolution whose coordinates have been
compactified to formally include $\mathscr{I}^+$ within the numerical
grid~\cite{Bishop,Reisswig:2009us,Babiuc:2010ze,Winicour_LRR}.
When the background coordinates are fixed, $\mathscr{I}^+$ can
be directly included within a Cauchy evolution by a
geometric prescription using hyperboloidal methods
\cite{Zenginoglu:2008wc,Zenginoglu:2007jw,Zenginoglu:2009hd,Zenginoglu:2009ey,
Zenginoglu:2010cq,Zenginoglu:2010zm,Bernuzzi:2011aj,Zenginoglu:2011zz}.
Another approach due to Abrahams and Evans shows
how one may {\em exactly} evaluate asymptotic waveforms
from gravitational multipoles for general relativity
linearized about flat spacetime \cite{AE1,AE2}.
This paper presents a new analytical and
numerical method to evaluate asymptotic
gravitational waveforms from perturbations of a
non-spinning (Schwarzschild)
blackhole. It also presents new results on the asymptotic signal
evaluation problem for the acoustic (i.e.~ordinary) wave equation.
Our approach is most similar to that of Abrahams and Evans.
Essentially, we reformulate their approach in a way which
subsequently generalizes to blackhole
perturbations. However, while our approach generalizes
the Abrahams-Evans one to a curved background spacetime,
we do not match their careful discussion of gauge issues.
For the Einstein equations linearized about Minkowski spacetime in
the Lorenz gauge the trace-reversed metric perturbation obeys the
flatspace (ordinary d'Alembertian) tensor wave equation. Therefore,
these perturbations are akin to solutions of either the ordinary
wave equation or the Maxwell equations; solutions characterized by
the sharp Huygen's principle and, therefore, which possess secondary
lacunae \cite{PetropavlovskyTsynkov}: given trivial initial
data and an inhomogeneous source which is bounded in space and time,
the solution vanishes on the intersection of all forward light cones
whose vertices sweep over the support of the source. The secondary
lacunae is a region of spacetime which is ``dark" because all
waves have already passed. Similar statements hold for the
homogeneous case with non-trivial initial data of compact support.
Actually, for the Maxwell case with certain sources, the solution may
have a quasi-lacunae featuring a late-time static electric field
\cite{PetropavlovskyTsynkov}.
Wave propagation on a curved spacetime
is more complicated due to the backscattering of waves off of curvature.
Even within the relatively simple setting of perturbations of
Schwarzschild blackholes, backscattering effects are present and
the resulting late-time ``tails" \cite{Price1972,DSS2011}
have been extensively studied both theoretically and numerically.
Backscattering confounds our intuitive sense of ``outgoing" and
``ingoing"; one might reasonably take the viewpoint that a partially
backscattered wave has both outgoing and ingoing pieces. Nevertheless,
for the {\em linear} master equations which describe perturbations of
Schwarzschild blackholes, there is an unambiguous notion of ``outgoing",
provided initial data of compact support. Away from the support of the
initial data, Laplace transformation of a master equation
\eqref{eq:RWZeqn} yields a {\em homogeneous} second-order ODE, which
therefore has two linearly independent solutions,
$\widehat{\Psi}^{(1)}(s,r)$ and $\widehat{\Psi}^{(-1)}(s,r)$,
where $s$ is Laplace frequency. Here we have suppressed
harmonic indices $(\ell,m)$ and assumed that the area radius $r$
is the independent spatial variable. We may assume
that $\widehat{\Psi}^{(\pm 1)}(s,r) \sim \exp(\mp sr_*)$ as
$r, r_* \rightarrow \infty$, where $r_*$ is the Regge-Wheeler
tortoise coordinate defined below.
At a radial location beyond the support
of the initial data, the frequency-domain solution has the
form $\alpha(s)\widehat{\Psi}^{(+1)}(s,r)$, where the details of
the initial data are buried in the coefficient $\alpha(s)$.
Physically, this notion of ``outgoing" would perhaps be better
characterized as ``asymptotically outgoing". Nevertheless,
provided the solution has this form, we can derive {\em at a finite
radius} both time-domain boundary conditions
\cite{Lau1,Lau2,Lau3} and an AWE procedure. The strategy in
both cases is to write down the exact conditions/procedure in the
frequency domain, and then accurately approximate this exact
relationship in a fashion that allows for simple inversion under
the inverse Laplace transform. For the case of boundary conditions,
one approximates the exact Dirichlet-to-Neumann map as a rational
function (in fact a sum of simple poles) along the axis of
imaginary Laplace frequency (the inversion contour). The exact
time-domain boundary condition is a history-dependent convolution,
which maybe approximated to machine precision as a
convolution involving a kernel given by a small sum of
exponentials. As we show, this type of kernel effectively
localizes the history dependence.
Our reformulation of the Abrahams-Evans procedure
(and its generalization to curved spacetimes) features a similar
history-dependent convolution involving a sum-of-exponentials
time-domain kernel. Section \ref{sec:Theory} demonstrates that,
in the (Laplace) frequency domain, an AWE
kernel is exactly expressible as an ``integral over boundary kernels",
thereby allowing us to leverage existing codes and knowledge
for generating and approximating boundary kernels. While
the construction of AWE kernels is
computationally intensive, this is an {\em offline} cost. Once the
kernel has been calculated, efficient and accurate AWE
can be implemented within an existing code in a non-intrusive manner.
Furthermore, AWE can be effected as a post-processing
step on existing data recorded at a fixed radial location.
Kernels used in this paper, as well as others, will be
available at \cite{Kernel_web1}.
This paper is organized as follows. Section \ref{sec:HowTo}
provides a self-contained guide on using
boundary and AWE kernels in either
existing codes or data post-processing. Towards this end,
Section~\ref{subsec:QNRTails} considers the numerical evolution
of late time tails from an approximate asymptotic signal which we find to
decay at the rate predicted for $\ell = 2$ perturbations
at $\mathscr{I}^+$.
Section \ref{sec:Theory} presents the theoretical underpinnings
of both radiation boundary conditions and waveform teleportation,
considering both wave propagation on flat (Minkowski) and
Schwarzschild spacetime.
For $\ell = 2$, $3$, $64$ perturbations, Section~\ref{subsec:PulseT}
considers accurate signal teleportation to a finite (near-field)
radial value. In Section~\ref{subsec:luminosities} we apply our
method to compute gravitational waveforms and luminosities from
extreme mass ratio binary systems, finding excellent
agreement with frequency domain computations. In these studies
we have observed that spurious junk radiation is
problematic for accurate $\mathscr{I}^+$ computations. Finally, we
conclude in Section \ref{sec:conclude} by discussing open issues,
both theoretical and practical.
\section{Implementation how-to guide} \label{sec:HowTo}
Our aim in this section is not to give a
derivation of our AWE method. Rather, adopting the
simplest possible evolution scheme and coordinates, we focus on
how AWE is implemented. By presenting an
implementation of our AWE method for a simple scheme, we hope to convey
the key points
to the reader, who will then grasp how to implement the method
within their own evolution scheme. Since our implementation of AWE
relies on certain radiation boundary conditions (RBC),
themselves essential when working on a spatially finite domain,
we first describe how to implement these within our simple scheme.
Here we do not discuss our RBC and AWE methods for different
background coordinate systems (i.e.~Kerr-Schild or hyperboloidal
foliations), but return to this issue in Appendix~\ref{sec:OtherFol}.
Multipole gravitational perturbations of a Schwarzschild blackhole are
described by the Regge-Wheeler (axial) and Zerilli (polar)
formalisms \cite{RW,ZER}. In geometric units the corresponding ``master"
wave equations have the form
\begin{equation}\label{eq:RWZeqn}
\frac{\partial^2\Psi}{\partial t^2}
-\frac{\partial^2\Psi}{\partial r_*^2} +
V^\mathrm{RW,Z}(r)\Psi = S,
\end{equation}
where $S(t,r)$ is a possible source and in terms of
the blackhole mass $M$ the Regge-Wheeler tortoise coordinate
is $r_* = r + 2M\log(\frac{1}{2}M^{-1}r - 1)$, which we also
denote by $x$. Until Section \ref{subsec:luminosities},
we alway choose $S = 0$ for the source.
Expressions for the Zerilli $V^\mathrm{Z}(r)$
and Regge-Wheeler $V^\mathrm{RW}(r)$ potentials are given
in, for example, Eqs.~(2) and (3) of Ref.~\cite{FHL}.
Both potentials depend on the orbital angular index
$\ell$. We have suppressed this index on $V(r)$, as well as the
orbital and azimuthal indices $(\ell,m)$ on the mode $\Psi$
and source $S$.
\subsection{Simple evolution algorithm with RBC}
\label{subsec:simpleRBC}
Most numerical schemes for evolution of \eqref{eq:RWZeqn} employ
some form of ``auxiliary variables", for example the variables
$\Psi,\Pi \equiv -\partial_t\Psi,\Phi \equiv \partial_x \Psi$.
However, for transparency here we use ``characteristic
variables":
\begin{align}\label{eq:charvars}
\Psi, \qquad W = -\Pi - \Phi, \qquad X = -\Pi +\Phi ,
\end{align}
for which the evolution equations are
\begin{align} \label{eq:UPsiSystem}
\partial_t \Psi = \frac{1}{2}(W + X),\quad
\partial_t W = -\partial_x W - V(r(x))\Psi,\quad
\partial_t X = \partial_x X - V(r(x))\Psi.
\end{align}
These equations show that $W$ propagates left-to-right ($\nearrow$)
and $X$ propagates right-to-left ($\nwarrow$). While
from the theoretical standpoint we prefer to view fields,
for example $\Psi(t,r)$, as depending spatially on $r$, from a
computationally standpoint we discretize Eqs.~\eqref{eq:UPsiSystem}
in $x$. Therefore, suppose the computational domain is the interval
$[a,b]$ in tortoise coordinate $x$, corresponding to the interval
$[r_a, r_b]$ in Schwarzschild radius $r$. Boundary conditions must
be specified for $W$ at $x=a$ and for $X$ at $x=b$.
We discretize \eqref{eq:UPsiSystem} using first-order
upwind stencils in space, and the forward Euler method in time.
Respectively, let $\{t_n\}$ and $\{x_k\}_{k=0}^K$ denote uniformly
spaced temporal and spatial grid points
such that $\Delta x = x_{k+1}-x_k$. Moreover, let $V_k \equiv
V(r(x_k))$ and $\Psi^n_k \simeq \Psi(t_n,r(x_k))$, with the same
notation for the other fields. Then the scheme for updating the
fields at all $x_k$ from time $t_n$ to time $t_{n+1} =
t_n + \Delta t$ is given by Algorithm~\ref{alg:FD_scheme}.
{\scriptsize
\begin{algorithm}[H]
\caption{Finite difference scheme for Eqs.~\eqref{eq:UPsiSystem}}
\label{alg:FD_scheme}
\begin{algorithmic}[1]
\State $\Psi^{n+1}_0 = \Psi^n_0
+ (\Delta t/2)
\big[W^n_0 + X^n_0\big]$
\State $W^{n+1}_0 = W_a^{n+1}$
\Comment{{\bf Boundary condition at} $a$}
\State $X^{n+1}_0 = X^n_0
+ (\Delta t/\Delta x)
\big(X^n_1-X^n_0\big)
- \Delta t V_0\Psi^n_0$
\For{$k = 1 \text{ \bf{to} } K-1$}
\State \hspace{5mm} $\Psi^{n+1}_k = \Psi^n_k
+ (\Delta t/2)
\big[W^n_k + X^n_k\big]$
\State \hspace{5mm} $W^{n+1}_k = W^n_k
- (\Delta t/\Delta x)
\big(W^n_k-W^n_{k-1}\big)
- \Delta t V_k \Psi^n_k$
\State \hspace{5mm} $X^{n+1}_k = X^n_k
+ (\Delta t/\Delta x)
\big(X^n_{k+1}-X^n_k\big)
- \Delta t V_k\Psi^n_k$
\EndFor
\State $\Psi^{n+1}_K = \Psi^n_K
+ (\Delta t/2)
\big(W^n_K + X^n_K\big)$
\State $W^{n+1}_K = W^n_K
- (\Delta t/\Delta x)
\big(W^n_K-W^n_{K-1}\big)
- \Delta t V_K \Psi^n_K$
\State $X^{n+1}_K = X_b^{n+1}$
\Comment{{\bf Boundary condition at} $b$}
\end{algorithmic}
\end{algorithm}
}
To complete the scheme, we must specify both
$W_a^n \simeq W(t_n,r_a)$ and $X_b^n \simeq X(t_n,r_b)$
as functions of (discrete) time.
So long as $a \ll 0$, the inner boundary value of the potential
$V(r)$ near $r = r_a$ is zero to machine precision, and the
Sommerfeld boundary condition $W_a^n = 0, \forall n$
is highly accurate. The RBC at
$x=b$ is determined by a Laplace convolution,
\begin{align}\label{eq:LaplaceConvApprox}
X(t,r_b) = \frac{f(r_b)}{r_b}
\int_0^t \Xi(t-t',r_b)
\Psi(t',r_b)dt',
\end{align}
where $f(r) = 1-2M/r$ and the boundary time-domain {\em kernel} is
\begin{equation}\label{eq:compressedBHkern}
\Xi(t,r_b) = \sum_{q=1}^d \frac{\gamma_q(\rho_b)}{2M} \Xi_q(t,r_b), \qquad
\Xi_q(t,r_b) = \exp\left(\frac{\beta_q(\rho_b)t}{2M}\right).
\end{equation}
The parameters $\{(\gamma_q(\rho_b),\beta_q(\rho_b))\}_{q=1}^d$
depend on the rescaled boundary radius $\rho_b = (2M)^{-1}r_b$
(as well as the orbital index $\ell$ which is suppressed) and
they are listed in numerical tables, such as those given in
the Appendix \ref{App:Tables}.\footnote{These
parameters can be redefined through division by $2M$,
thereby removing $2M$ factors in the formulas which follow. Indeed,
such a redefinition would be made in an actual code. Nevertheless, we
retain the original $\{(\gamma_q(\rho_b),\beta_q(\rho_b))\}_{q=1}^d$
parameters with $2M$ factors, in order to ensure that the
parameters here correspond precisely to those listed in our
tables.} Some of the parameters $\{(\gamma_q,\beta_q)\}_{q=1}^d$
are complex, but the kernel $\Xi(t,r_b)$ is real. We stress that,
insofar as implementation of the convolution
\eqref{eq:LaplaceConvApprox} is concerned, the origin of
these numbers is unimportant. Defining, for example,
$\Psi_b(t) = \Psi(t,r_b)$, we write \eqref{eq:LaplaceConvApprox}
as $X_b(t) = r_b^{-1}f(r_b)(\Xi(\cdot,r_b) * \Psi_b)(t)$. With
a similar notation, the constituent
convolution $(\Xi_q * \Psi_b) \equiv
(\Xi_q(\cdot,r_b) * \Psi_b)(t)$ obeys an ODE
{\em at the boundary},
\begin{equation}\label{eq:boundaryODE}
\frac{d}{dt}
(\Xi_q * \Psi_b) = \big[(2M)^{-1}\beta_q (\Xi_q * \Psi_b) + \Psi_b\big].
\end{equation}
These $d$ ODE can be integrated along side the system
\eqref{eq:UPsiSystem}. Indeed, we define $(\Xi_q * \Psi_b)^n \simeq
(\Xi_q * \Psi_b)(t_n)$ and complete our scheme as follows.
{\scriptsize
\begin{algorithm}[H]
\caption{Finite difference scheme for Eqs.~\eqref{eq:UPsiSystem} with RBC}
\label{alg:RBCs}
\begin{algorithmic}[1]
\For{$q = 1 \text{ \bf{to} } d$} \Comment{{\bf First update constituent convolutions}}
\State \hspace{5mm} $(\Xi_q * \Psi_b)^{n+1} = (\Xi_q * \Psi_b)^n
+ \Delta t \big[(2M)^{-1}\beta_q (\Xi_q * \Psi_b)^n
+ \Psi^n_K\big]$
\EndFor
\State $X_b^{n+1} = (2Mr_b)^{-1}f(r_b)\sum_{q=1}^d
\gamma_q (\Xi_q * \Psi_b)^{n+1}$
\State $W_a^{n+1} = 0$
\vskip 10pt
\State Run Algorithm~\ref{alg:FD_scheme} with
$W_a^{n+1}$ and
$X_b^{n+1}$ given by above values.
\end{algorithmic}
\end{algorithm}
}
\subsection{Evaluation of the asymptotic waveform}
\label{subsec:simpleEXT}
Our approach to AWE is similar to the described implementation
of radiation boundary conditions. We introduce new parameters
$\{(\gamma^E_q(\rho_b,\rho_\infty),
\beta^E_q(\rho_b,\rho_\infty))
\}_{q=1}^{d^E}$ and a new kernel
\begin{equation}\label{eq:XiEkernel}
\Xi^E(t,r_b,r_\infty) =
\sum_{q=1}^{d^E}\frac{\gamma^E_q(\rho_b,\rho_\infty)}{2M}
\Xi^E_q(t,\rho_b,\rho_\infty),\qquad
\Xi^E_q(t,\rho_b,\rho_\infty) =
\exp\left(\frac{\beta^E_q(\rho_b,\rho_\infty)t}{2M}\right).
\end{equation}
The parameters $\{(\gamma^E_q(\rho_b,\rho_\infty),
\beta^E_q(\rho_b,\rho_\infty)\}_{q=1}^{d^E}$
also depend on $\ell$, but this dependence has been suppressed.
The $E$ here stands for ``evaluation"
and differentiates this kernel
from the RBC one. This evaluation kernel enacts {\em teleportation}
(the term is defined in Section~\ref{sec:Theory}) of the waveform
from the boundary radius $r_b = 2M\rho_b$ to the evaluation
radius $r_\infty = 2M\rho_\infty \gg r_b$ (from $b$ to $x_\infty$
in tortoise coordinate). As discussed below, the
choice $r_\infty = \infty$ is formally
possible (see Sec.~\ref{sec:conclude}); however, in this
paper $r_\infty$ is an {\em arbitrarily large}, albeit finite,
radius. To ensure the signals recovered at $r_\infty$ and $\mathscr{I}^+$
are identical to about machine precision,
we choose $r_\infty = 2M(1\times 10^{15})$ for double precision
simulations, and would choose $r_\infty = 2M(1\times 10^{30})$
for quadruple precision simulations. We then
approximate the asymptotic waveform as
\begin{equation}\label{eq:XiEkernelConv}
\Psi_\infty(t) := \Psi(t+(x_\infty-b),r_\infty) \simeq
\int_0^t \Xi^E(t-t',r_b,r_\infty)\Psi(t',r_b)dt' + \Psi(t,r_b).
\end{equation}
The offset by $\Psi(t,r_b)$ in this formula stems from a
technicality explained in Section~\ref{Sec:flatspace}.
As before, this formula can be implemented through integration
of ODE at the boundary, only now these ODE are not coupled to
the numerical evolution. With $(\Xi_q^E * \Psi_b)^n
\simeq (\Xi_q^E(\cdot,r_b) * \Psi_b)(t_n)$ and
$\Psi_\infty^{n} \simeq \Psi_\infty(t_n)$, the algorithm
using forward Euler is as follows.
{\scriptsize
\begin{algorithm}[H]
\caption{Evaluation of asymptotic waveform, placed after field update}
\label{alg:WEs}
\begin{algorithmic}[1]
\For{$q = 1 \text{ \bf{to} } d^E$}
\State \hspace{5mm} $(\Xi_q^E * \Psi_b)^{n+1}
= (\Xi_q^E * \Psi_b)^n
+ \Delta t \big[
(2M)^{-1}\beta_q^E (\Xi_q^E * \Psi_b)^n
+ \Psi^n_K\big]$
\EndFor
\State $\Psi_\infty^{n+1} = (2M)^{-1}\sum_{q=1}^d \gamma_q^E
(\Xi_q^E * \Psi_b)^{n+1}
+ \Psi^{n+1}_K$
\end{algorithmic}
\end{algorithm}
}
\noindent
The following condition would yield particularly efficient AWE:
\begin{equation}\label{eq:WEandRBCpoles}
d^E = d,\quad
\beta^E_q(\rho_b,\rho_\infty) = \beta_q(\rho_b),\quad \forall q,
\text{ (preferred, but perhaps {\em not} possible).}
\end{equation}
Indeed, integration of the same ODE \eqref{eq:boundaryODE} at the
boundary would then determine both the RBC and AWE. In this case,
steps 1 through 3 of Algorithm \ref{alg:WEs} have already been
carried out in Algorithm \ref{alg:RBCs}. However, the assumption
in \eqref{eq:WEandRBCpoles} may not always be possible, and even
when possible appears to yield less accurate teleportation kernels. We have
constructed $\ell = 2$ AWE kernels which satisfy \eqref{eq:WEandRBCpoles}
(an example is given in the Appendix \ref{App:Tables}); however,
relative to our best kernels, they indeed yield less accuracy.
Moreover, we have been unable to achieve \eqref{eq:WEandRBCpoles}
for $\ell=64$ teleportation kernels. Therefore, in this paper
we will {\em not} assume \eqref{eq:WEandRBCpoles}.
\vskip 10pt
\noindent
{\bf Remark.} If $\Psi$ is itself complex (as is the case in
applications), then round-off issues will lead to mixing of the
real and imaginary parts in the simple algorithms above.
In this case, we advocate splitting the complex exponentials which
make up $\Xi$ and $\Xi^E$ into manifestly real expressions
involving sine and cosine terms. Such splitting amounts to extra
bookkeeping, but hardly complicates the above treatment.
\begin{figure}[t]
\includegraphics[clip=true,width=14cm]{DecayTails.eps}
\caption{{\sc Quasinormal ringing and decay tails.}
Each dashed curve corresponds to power law decay, with the
indicated rate determined by a least squares fit of
the field over the time window $[500,700]M$.
The time shift for the asymptotic waveform has not been included.
\label{fig:decaytails}}
\end{figure}
\subsection{Numerical experiment: quasinormal ringing and decay tails}
\label{subsec:QNRTails}
Section \ref{sec:Experiments} documents the results of
several numerical experiments which validate and test
our methods. This subsection also describes a numerical
experiment, although here with the goal of providing
further assistance toward implementation. An interested
reader might first repeat the experiment described below.
We consider the $\ell = 2$ Regge-Wheeler equation and
Gaussian initial data
\begin{equation}\label{eq:GaussianData}
\Psi = e^{-[2-x/(2M)]^2},\quad
\Phi = \frac{4M-x}{2M^2}e^{-[2-x/(2M)]^2},\quad
\Pi = \Phi(0,r(x)).
\end{equation}
For this experiment the computational $x$-domain is $[a,b]$,
where $a = -200M$ and $b = 30M + 2M\log(14)$ corresponds to
$\rho_b = r_b/(2M) = 15$. We evolve the data
\eqref{eq:GaussianData} until time $t = 600M$,
with a Sommerfeld boundary condition at $x=a$ and
the convolution boundary condition \eqref{eq:LaplaceConvApprox}
at $x=b$. Table \ref{tab:appendixRBCtable} in Appendix \ref{App:Tables}
lists the 26 pole RBC kernel $\Xi$ used for this convolution.
Instead of Algorithm~\ref{alg:FD_scheme} we use
a multidomain Chebyshev collocation method with classical
Runge Kutta 4 as the timestepper.
During the simulation we record as a time series both the field
$\Psi(t,20M)$ and $\Psi(t,r_b)$. Through the convolution
\eqref{eq:XiEkernelConv} determined by Table
\ref{tab:appendixEXTtableLong} in Appendix \ref{App:Tables},
the field $\Psi(t,r_b)$ is teleported from $r_b = 30M$
to $r_\infty = 2M(1\times 10^{15})$ providing a time series
which approximates the asymptotic waveform $\Psi_\infty(t)$.
In absolute value $\Psi(t,20M)$ (solid blue line) and
$\Psi_\infty(t)$ (solid black line) are depicted in
Fig.~\ref{fig:decaytails}. The time shift for the
teleported waveform has not been included, and we have chosen
to record $\Psi(t,20M)$ at $r = 20M < r_b$ only to ensure that
the time series in the plot do not lie on top of each other
at early times. These series exhibit the phenomena of
quasinormal ringing and late time decay tails. Each dashed
curve in the figure corresponds to power law decay, with the
indicated rate determined by a least squares fit based on
the numerical decay of the field over dashed curve's time
window. The decay rates $p=-7$ and $p=-4$ are respectively the
theoretical predictions \cite{gundlach1994late,Barack:1998bw,Zenginoglu:2009ey}
for a finite radius and $\mathscr{I}^+$.
Figure \ref{fig:scritail} shows the decay rate computed from
$p = \partial_{\mathrm{ln} t} \mathrm{ln} | \Psi_\infty(t) |$
for the teleported signal with the evolution carried out to
$t=25000M$. As a post-processing procedure,
generation of this figure from existing data would take a few seconds.
The decay for the teleported signal asymptotically approaches $p=-4$ at
later times. The teleported pulse corresponds to a time series recorded
along the wordline $(t_\mathrm{obs}+x_\infty-b,r_\infty)$ by an
observer safely within the astrophysical zone. Here $x_\infty-b$
is the time shift, and the astrophysical zone is defined as the region
where $t_\mathrm{obs} \ll r_\infty$, with $t_\mathrm{obs}$ the time
elapsed after the pulse's leading edge passes our fictitious
observer \cite{Leaver,Purrer:2004nq,Barack1999}. Since
an observation in the astrophysical zone is well approximated as taking
place at $\mathscr{I}^+$, the observed $-4$ decay rate is expected. For
very late times the decay rate should settle towards $-7$,
although extended precision might be necessary to capture the transition.
\begin{figure}[t]
\includegraphics[clip=true,width=13cm,trim=0 7.25cm 0 0cm]{scritail.eps}
\caption{{\sc Tail decay rate for the teleported
signal.} The rate $p$ for $\Psi_\infty(t) \propto t^p$ has been computed using
logarithmic difference quotients based on
$\partial_{\ln t}\ln|\Psi_\infty(t)|$.
\label{fig:scritail}}
\end{figure}
\section{Theoretical discussion} \label{sec:Theory}
To fix ideas and motivate the new method, the next subsection
describes AWE for flatspace multipole solutions
of the ordinary 3+1 wave equation, formally the $M=0$ case of
Eq.~\eqref{eq:RWZeqn}. Formulas derived in the next
subsection motivate similar ones given for blackhole
perturbations in subsection \ref{subsec:BH_perts}.
\subsection{Flatspace waves} \label{Sec:flatspace}
This subsection describes
(i) outgoing multipole solutions to the flatspace radial
wave equation, (ii) the exact RBC obeyed by outgoing
multipoles, and (iii) the relationship between this RBC
and AWE for outgoing multipoles. Throughout this subsection,
we often choose $\ell=2$ as a representative example,
but similar (obvious) results hold for any multipole-order $\ell$.
\subsubsection{Structure of outgoing and ingoing flatspace
multipoles}
General outgoing ($\epsilon = 1$) and ingoing ($\epsilon = -1$)
order-$\ell$ multipole solutions of the 3+1 wave equation
$(-\partial_t^2+\partial_x^2 + \partial_y^2 +
\partial_z^2)\psi = 0$ have the form (see, for example,
Refs.~\cite{Burke,gundlach1994late})
\begin{equation}\label{eq:defpsil}
\psi(t,x,y,z)
= \frac{1}{r}\Psi^{(\epsilon)}_\ell(t,r)
Y_{\ell m}(\theta,\phi),
\quad
\Psi^{(\epsilon)}_\ell(t,r) = \sum_{k=0}^\ell
\frac{\epsilon^k}{r^{k}}c_{\ell k}
f^{(\ell-k)}(t-\epsilon r),
\quad
c_{\ell k} = \frac{1}{2^k k!}
\frac{(\ell+k)!}{(\ell-k)!},
\end{equation}
where we have suppressed the azimuthal index $m$ on the
``mode" $\Psi^{(\epsilon)}_\ell(t,r)$. In
Eq.~\eqref{eq:defpsil}
$f^{(p)}(u)$ is the $p$th derivative of an underlying
function $f(u)$ of retarded time $u = t - r$, and similarly
for $f^{(p)}(v)$, where $v = t + r$ is advanced time. In
(\ref{eq:defpsil}) we view $(x,y,z)$ as place holders for
$(r\sin\theta\cos\phi,r\sin\theta\sin\phi,r\cos\theta)$.
For both $\epsilon = \pm 1$ cases the mode
obeys the {\em flatspace radial wave equation}
\begin{equation}\label{eq:flatradial}
\frac{\partial^2\Psi^{(\epsilon)}_\ell}{\partial t^2}
-\frac{\partial^2\Psi^{(\epsilon)}_\ell}{\partial r^2}
+\frac{\ell(\ell + 1)}{r^2}\Psi^{(\epsilon)}_\ell = 0 ,
\end{equation}
and, specializing to the representative example, the outgoing
quadrupole is
\begin{equation}\label{eq:ell2mult_td}
\Psi^{(1)}_2(t,r) = f''(t-r)
+ \frac{3}{r}f'(t-r)
+ \frac{3}{r^2}f(t-r).
\end{equation}
We are interested in the $(\epsilon = 1)$
outgoing case, and so now write
$\Psi_\ell(t,r)$ to mean $\Psi^{(1)}_\ell(t,r)$.
Given a fixed radius $r_b$ which specifies the outer
boundary, consider the following assumption on the
initial data (one of compact support):
\begin{equation} \label{eq:IDassume}
\Psi_\ell(0,r) = 0 = (\partial_t\Psi_\ell)(0,r),\qquad
r > r_b - \delta,\qquad \text{for any small }\delta > 0.
\end{equation}
Provided that \eqref{eq:IDassume} holds,
in the region $r \geq r_b$ Laplace transformation of
an {\em outgoing} mode $\Psi_{\ell}(t,r)$ yields
\begin{equation}\label{eq:ellmult_fd}
\widehat{\Psi}_\ell(s,r) = a(s)s^\ell e^{-sr} W_\ell(sr),\quad
W_\ell(z) = \sum_{k=0}^\ell \frac{c_{\ell k}}{z^k},
\end{equation}
where $z=sr$ and
\begin{equation}\label{eq:profile_fd}
a(s) \equiv e^{sr}\int_0^\infty e^{-st}f(t-r)dt
= \int_{-r_b}^\infty e^{-su}f(u) du.
\end{equation}
Notice that $a(s)$ is indeed independent of $r$.
For the quadrupole case, we have
\begin{equation}\label{eq:ell2mult_fd}
\widehat{\Psi}_2(s,r) = a(s)s^2 e^{-sr} W_2(sr),\qquad
W_2(z) = 1 + \frac{3}{z}+\frac{3}{z^2}
\end{equation}
[compare this expression with Eq.~\eqref{eq:ell2mult_td}].
To obtain \eqref{eq:ellmult_fd} and \eqref{eq:profile_fd},
we have used \eqref{eq:IDassume} as follows. First
consider the Laplace transform of $f^{(\ell-k)}(t-r)$ which
appears in \eqref{eq:defpsil}. Repeated integration by parts
generates $t=0$ boundary terms of the form $f^{(\ell-k-p)}(-r)$
for $1 \leq p \leq \ell - k$. All such terms vanish, as can be
shown by the following identity:
\begin{equation}\label{eq:ffromPsi}
f(t-r) = (-1)^\ell \frac{2^\ell}{(2\ell)!}
\left[r^2\left(\partial_t
+ \partial_r\right)\right]^\ell\Psi_\ell(t,r).
\end{equation}
Since the initial data $\Psi_\ell$ and $\partial_t\Psi_\ell$
vanishes on an open {\em spatial} neighborhood of the spatial
point with coordinate $r \geq r_b$, in fact $\Psi_\ell$ vanishes in an
open {\em spacetime} neighborhood of the spacetime point
with coordinates $(0,r)$. Therefore, {\em all} derivatives of
$\Psi_\ell$ vanish in the same neighborhood, which implies
$f^{(\ell-k-p)}(-r) = 0$. This implication stems from repeated
differentiation of \eqref{eq:ffromPsi} with $t=0$ enforced
afterward.
The previous argument establishes that
\begin{equation}
\int^\infty_0 e^{-st}f^{(\ell-k)}(t-r)dt
= s^{\ell-k} e^{-sr}\int^\infty_{-r} e^{-su}f(u)du.
\end{equation}
The lower limit $-r$ of integration can now be replaced with $-r_b$.
Indeed, \eqref{eq:IDassume} implies that $\Psi(t,r)=0$
for $0 \leq t < \delta + (r-r_b)$ and $r\geq r_b$. Using
\eqref{eq:ffromPsi}, we conclude that $f(u) = 0$ for $-r \leq
u \leq -r_b$. Therefore, the right-hand side of the last equation
is $a(s) s^{\ell-k} e^{-sr}$.
\subsubsection{Radiation boundary conditions for flatspace multipoles}
We continue to derive expressions for $r \geq r_b$
where the assumption \eqref{eq:IDassume}
of compact support holds. The explicit expression \eqref{eq:ellmult_fd}
for $\widehat{\Psi}_\ell(s,r)$ determines an exact
frequency-domain boundary condition
\begin{equation} \label{eq:SommerfeldReshat}
s\widehat{\Psi}_\ell(s,r) + \partial_r\widehat{\Psi}_\ell(s,r)
= \frac{1}{r} \widehat{\Omega}_\ell(s,r)\widehat{\Psi}_\ell(s,r),
\end{equation}
where the {\em frequency-domain radiation kernel}
$\widehat{\Omega}_\ell(s,r)$
defines the Sommerfeld residual. Indeed, the operator on the left-hand
of \eqref{eq:SommerfeldReshat} corresponds to the Sommerfeld operator
$\partial_t + \partial_r$ in the time-domain.
If $\ell = 0$, then $\widehat{\Omega}_\ell(s,r) = 0$;
otherwise a simple computation based on
Eqs.~(\ref{eq:ellmult_fd},\ref{eq:SommerfeldReshat}) shows
that the frequency-domain kernel is given by (with the prime
indicating differentiation in argument):
\begin{equation} \label{eq:Omegahat}
\widehat{\Omega}_\ell(s,r)
\equiv sr \frac{W'_\ell(sr)}{W_\ell(sr)}
= \sum_{k=1}^\ell \frac{b_{\ell,k}/r}{s-b_{\ell,k}/r},
\end{equation}
where $\{b_{\ell,k} : 1 \leq k \leq \ell \}$ are the roots
of $W_\ell(z)$, all of which are simple.\footnote{\label{fn:MacDonald}
The last equality also follows from the identity
\begin{equation}
W_\ell(z) = \sqrt{\frac{2 z}{\pi}}e^z K_{\ell + 1/2}(z),
\end{equation}
showing that the $b_{\ell,k}$ are also the roots of the
half-integer MacDonald function $K_{\ell+1/2}(z)$, which
are simple and lie in the left-half plane \cite{Olver,Watson}.
The appearance of $K_{\ell+1/2}(z)$
may have been anticipated; indeed, the modified Bessel equation
arises when finding separable solutions to the Laplace
transformed flatspace radial wave equation \eqref{eq:flatradial}.
\label{fn:MacDonald}}
For the quadrupole case
\begin{equation} \label{eq:Omegahat2}
\widehat{\Omega}_2(s,r)
= \frac{z_+/r}{s-z_+/r} + \frac{z_-/r}{s-z_-/r},
\quad
z_\pm = -\frac{3}{2}\pm \mathrm{i}\frac{\sqrt{3}}{2},
\end{equation}
where $z_+ = b_{2,1}$ and $z_- = b_{2,2}$ solve $W_2(z_\pm) = 0$.
The time-domain RBC is the inverse Laplace transform of
\eqref{eq:SommerfeldReshat}, i.e.~the Laplace
convolution \cite{GroteKeller1,GroteKeller2,Sofronov1,Sofronov2,AGH}
\begin{equation} \label{eq:SommerfeldRes}
\partial_t\Psi_\ell + \partial_r\Psi_\ell = \frac{1}{r}\int_0^t
\Omega_\ell(t-t',r)\Psi_\ell(t',r)dt',\quad
\Omega_\ell(t,r)= \sum_{k=1}^\ell \frac{b_{\ell,k}}{r}
\exp\left(\frac{b_{\ell,k}t}{r}\right).
\end{equation}
Subject to our assumption \eqref{eq:IDassume}
of compact support, the outgoing
multipole [$\epsilon = 1$ in Eq.~\eqref{eq:defpsil}] obeys
Eq.~\eqref{eq:SommerfeldRes} exactly, as can also be shown via
direct calculation using repeated integration by parts;
see Appendix \ref{sec:NoLaplace}. If, on the other hand,
assumption \eqref{eq:IDassume} does not hold, then
Eq.~\eqref{eq:SommerfeldRes} is violated, but only by terms
which decay exponentially fast in $t$; again, see Appendix
\ref{sec:NoLaplace}.
\subsubsection{Asymptotic waveform evaluation and teleportation
for flatspace multipoles}
For a generic outgoing solution, it is possible to recover
the profile function $f(t-r)$ and asymptotic waveform
\begin{equation}
\Psi_\ell(t,r) \sim f^{(\ell)}(t-r),\qquad r\rightarrow\infty,
\end{equation}
via data recorded solely at a finite and {\em fixed}
radial location, again taken as $r = r_b$.
Let us consider the $\ell = 2$ case as a concrete example.
Generalization to higher $\ell$ is straightforward. Equation
\eqref{eq:ell2mult_td} suggests that we solve the ODE initial
value problem
\begin{equation}\label{eq:extractionODE}
y'' + \frac{3}{r} y' + \frac{3}{r^2}y
= \Psi_2(t,r), \qquad
y(0) = 0 = y'(0),\qquad
(\ell = 2 \text{ problem})
\end{equation}
in which case $f(u) = y(u+r)$.
In a pioneering series of
papers \cite{AE1,AE2}, Abrahams and Evans showed how
the above procedure carries over to the theory of gravitational
multipoles for general relativity linearized about flat
spacetime. We now re-examine the basic
idea behind Abrahams-Evans AWE from the standpoint of
Laplace convolution, and will consider two kernels:
one $\Theta_\ell$ for evaluation of the underlying
function $f(u)$ and another $\Phi_\ell$ more suited for
evaluation of the waveform $f^{(\ell)}(u)$. Our implementations
have mostly relied on the $\Phi_\ell$ kernel.
Continuing with the $\ell = 2$ example, we introduce a
{\em frequency-domain profile evaluation kernel}
$\widehat{\Theta}_2(s,r)$ tailored to satisfy
\begin{equation}
\widehat{\Theta}_2(s,r)\widehat{\Psi}_2(s,r)
= a(s) e^{-sr} = \widehat{y}(s).
\end{equation}
That is, the product of $\widehat{\Theta}_2(s,r)$ and
$\widehat{\Psi}_2(s,r)$ is $\int_0^\infty e^{-st}f(t-r)dt$
[cf.~Eq.~\eqref{eq:profile_fd}]. Comparison with
\eqref{eq:ell2mult_fd} immediately shows that
\begin{equation}\label{eq:flatspaceTheta}
\widehat{\Theta}_2(s,r) = \frac{1}{s^2 W_2(sr)} =
\frac{\mathrm{i}r}{\sqrt{3}}
\left[
\frac{1}{s-z_-/r} - \frac{1}{s-z_+/r}
\right].
\end{equation}
The corresponding time-domain profile evaluation kernel is
\begin{equation}\label{eq:FLAT_TDK_2}
\Theta_2(t,r) = \frac{\mathrm{i}r}{\sqrt{3}}
\left[
\exp\left(\frac{z_-t}{r}\right)-
\exp\left(\frac{z_+t}{r}\right)
\right],
\end{equation}
and $y(t) = (\Theta_2(\cdot,r) * \Psi_2(\cdot,r))(t)$ solves the
Abrahams-Evans initial value problem \eqref{eq:extractionODE}.
Essentially the same arguments show that the order-$\ell$
profile evaluation kernel is
\begin{equation}
\widehat{\Theta}_\ell(s,r) =
\frac{1}{s^\ell W_\ell(sr)} .
\end{equation}
Despite appearances, the kernel is regular at $s = 0$. For example,
$1/W_2(sr) \sim (sr)^2/3$, and in general
$1/W_\ell(sr) \sim (sr)^\ell/c_{\ell\ell}$, as $s\rightarrow 0$.
Direct evaluation of the asymptotic waveform is also possible.
{\em Teleportation} by a positive shift $r_2-r_1$ means conversion of
$\Psi(t,r_1)$ to $\Psi(t+(r_2-r_1),r_2)$, and it might
correspond to a small finite shift $r_2-r_1$. However, when $r_2$
is suitably large, we write $r_\infty$ for $r_2$ and view
teleportation as an AWE procedure (in which case typically
$r_1 = r_b$, and it is the boundary waveform $\Psi(t,r_b)$ which
is teleported). Teleportation is accomplished with a
{\em frequency-domain teleportation kernel}
\begin{equation}\label{eq:freqPhidef}
\widehat{\Phi}_\ell(s,r_1,r_2)
= -1 + \frac{W_\ell(sr_2)}{W_\ell(sr_1)}
\end{equation}
rigged to satisfy
\begin{equation}
e^{s(r_2-r_1)}\widehat{\Psi}_\ell(s,r_2) =
\widehat{\Phi}_\ell(s,r_1,r_2)\widehat{\Psi}_\ell(s,r_1)
+ \widehat{\Psi}_\ell(s,r_1).
\end{equation}
We have included the $-1$ factor in \eqref{eq:freqPhidef} to
ensure that $\widehat{\Phi}_\ell(s,r_1,r_2)$ has a
well-defined inverse Laplace transform $\Phi_\ell(t,r_1,r_2)$.
In the time domain we recover the desired property
\begin{equation}\label{eq:teleport}
\Psi_\ell(t+(r_2-r_1),r_2) =
(\Phi_\ell(\cdot,r_1,r_2)*\Psi_\ell(\cdot,r_1))(t) +
\Psi_\ell(t,r_1).
\end{equation}
Adjusting for the $(r_2 - r_1)$ time delay, this
formula allows for conversion of the signal at $r_1$
to the signal at $r_2$. Since $r_2 \leq \infty$,
this method can also be used for evaluation
of the asymptotic waveform $f^{(\ell)}(u)$. We refer
to the $r_2 = \infty$ case $\widehat{\Phi}_\ell(s,r_1,\infty)$
as the {\em frequency-domain waveform evaluation kernel}.
The relationship between RBC and
AWE/teleportation kernels
is a key insight of this paper, and the one which is exploited
to numerically construct AWE/teleportation kernels.
For example, the profile evaluation kernel can be written as
\begin{equation}\label{eq:flatspaceThetaOmega}
\widehat{\Theta}_\ell(s,r) =
\frac{1}{s^\ell}
\underbrace{\exp\left[\int_r^\infty
\frac{\widehat{\Omega}_\ell(s,\eta)}{\eta} d\eta \right]}_{1/W_\ell(sr)}.
\end{equation}
That the underbraced quantity is indeed $1/W_\ell(sr)$ follows
easily from the identity $\eta^{-1}\widehat{\Omega}_\ell(s,\eta)
= \partial_\eta \log W_\ell(s\eta)$,
that is essentially the definition \eqref{eq:Omegahat}.
The integration in \eqref{eq:flatspaceThetaOmega} can of course be
carried out, recovering \eqref{eq:flatspaceTheta} for the
$\ell = 2$ case; however, when considering similar expressions
for blackhole perturbations at least some of the integration
will be performed by numerical quadrature.
Similarly, one can express the teleportation kernel through
\begin{equation}\label{eq:flatspacePhi}
\widehat{\Phi}_\ell(s,r_1,r_2)
= -1 + \underbrace{\exp\left[\int_{r_1}^{r_2}
\frac{\widehat{\Omega}_\ell(s,\eta)}{\eta}d\eta\right]}_{
W_\ell(sr_2)/W_\ell(sr_1)}.
\end{equation}
In Section~\ref{subsec:BH_perts} we introduce the analogous
kernels for waveform teleportation in the
Regge-Wheeler and Zerilli formalisms. As mentioned,
we have mostly used the $\Phi_\ell$ kernels.
\subsubsection{Efficiency and storage} \label{sec:Eff}
Here we comment on RBC and AWE for the ordinary 3+1 wave equation
from the standpoint of efficiency and storage
as $\ell \rightarrow \infty$, both summarizing
known results for RBC \cite{AGH} and considering these issues
for AWE. Let $\lambda$ represent a
characteristic wavelength, say determined by the initial
data or inputted boundary conditions.
For numerical evolution to a fixed final time $T$,
an implementation of the {\em exact} flatspace RBC
\eqref{eq:SommerfeldRes} (with a kernel comprised
of $\ell$ exponentials) has the following
work and storage requirements:
\begin{equation}\label{eq:exactRBCscalings}
\mathrm{Work}_\mathrm{exactRBC} = O(\lambda^{-1}\ell T), \qquad
\mathrm{Storage}_\mathrm{exactRBC} = O(\ell).
\end{equation}
These scalings are deduced from the cost of integrating $\ell$
ODE of the form \eqref{eq:boundaryODE} with
approximately $\lambda^{-1} T$ timesteps.
A {\em spatially and temporally resolved} numerical
integration (with arbitrary boundary conditions) of
Eq.~\eqref{eq:flatradial} on a radial domain of fixed
size corresponds to the following work and storage scalings:
$O(\lambda^{-2}T)$ and $O(\lambda^{-1})$. Indeed,
a resolved spatial discretization of Eq.~\eqref{eq:flatradial}
yields a coupled system of approximately $\lambda^{-1}$ ODE.
As more spatial/temporal resolution is typically required for
large $\ell$ solutions, it is reasonable to view
$\lambda^{-1} \simeq \ell$, in which case the scalings
for the interior solver are comparable to
\eqref{eq:exactRBCscalings}.
However, implementation of the exact RBC is still preferable to
choosing the computational domain so large that the outer
boundary is casually disconnected from the wordline of an
interior ``detector". Spatial discretization after
such domain enlargement yields $\lambda^{-1}T$
coupled ODE, whence $O(\lambda^{-2}T^2)$ and $O(\lambda^{-1}T)$ for
the work and storage.
{\em Kernel compression} yields a more efficient implementation
of RBC. As proven in Ref.~\cite{AGH}, the
kernel $\widehat{\Omega}_\ell(s,r)$ admits a rational
approximation\footnote{The
$\gamma_{\ell,n}$ and $\beta_{\ell,n}$ appearing in
the approximate (frequency-domain) flatspace kernel
\eqref{eq:compressedFLTkern} are different
than the similar parameters appearing in the
approximate (time-domain) blackhole kernel
\eqref{eq:compressedBHkern}. Here
$\gamma_{\ell,n}$ and $\beta_{\ell,n}$ {\em do not}
depend on $r$, whereas the parameters in
\eqref{eq:compressedBHkern} {\em do} depend on the
(rescaled) radius. We use similar notations for
the flatspace and blackhole cases, hoping this
practice does not cause confusion.}
\begin{equation}\label{eq:compressedFLTkern}
\widehat{\Xi}_\ell(s,r) =
\sum_{n=1}^d \frac{\gamma_{\ell,n}/r}{s-\beta_{\ell,n}/r},
\qquad
\sum_{s\in \mathrm{i}\mathbb{R}}
\left|\frac{\widehat{\Omega}_\ell(s,r)-\widehat{\Xi}_\ell(s,r)}{
\widehat{\Omega}_\ell(s,r)}\right| < \varepsilon,
\end{equation}
where $\varepsilon$ is a prescribed tolerance and the number
of approximating poles scales like \cite{AGH}
\begin{align}\label{eq:AGHdscaling}
d = O\big(\log\nu\log(1/\varepsilon)
+\log^2\nu+\nu^{-1}\log^2(1/\varepsilon)\big)
\end{align}
as $\nu = \ell+1/2 \rightarrow \infty$ and $\varepsilon
\rightarrow 0^{+}$. The frequency domain bound in
\eqref{eq:compressedFLTkern} implies a long-time
bound on the relative convolution error
in the time-domain, see Appendix \ref{sec:ErrorEst}. Since $d$
grows sublinearly in $\ell$ and $1/\varepsilon$, the approximation
$\widehat{\Xi}_\ell(s,r)$ [likewise its inverse Laplace transform
$\Xi_\ell(t,r)$] is called a {\em compressed kernel}. An
implementation of Laplace convolution RBC based on compressed
kernels $\Xi_\ell(t,r)$ scales like
\begin{equation}\label{eq:RBCscalings}
\mathrm{Work}_\mathrm{compressedRBC} = O(\lambda^{-1}dT),
\qquad \mathrm{Storage}_\mathrm{compressedRBC} = O(d),
\end{equation}
with clear performance in the large-$\ell$ limit.
The proof of \eqref{eq:AGHdscaling} relies on the large-$\ell$
asymptotics \cite{AGH,Olver} of the roots
$\{b_{\ell,k}: k = 1,\dots,\ell\}$ of $K_{\ell + 1/2}(z)$.
Precisely, as $\ell \rightarrow \infty$ the scaled roots
$b_{\ell,k}/(\ell+1/2)$ accumulate on a curve $\mathcal{C}$
given by \cite{AGH,Olver}
\begin{equation}\label{eq:curveC}
z(\lambda) = -\sqrt{\lambda^2 -\lambda \tanh\lambda}
\pm \mathrm{i} \sqrt{\lambda \coth\lambda-\lambda^2},
\qquad
\lambda \in [0,\lambda_0],\qquad
\tanh\lambda_0 = 1/\lambda_0.
\end{equation}
Since the pole locations appearing in {\em both} the exact
flatspace RBC and AWE kernels are
$\{b_{\ell,k}/r: k = 1,\dots,\ell\}$,
we conjecture that an implementation
of AWE based on kernel compression
{\em formally} satisfies the scalings \eqref{eq:RBCscalings}.
However, we are unsure if these scalings hold in practice.
As a nascent investigation, we consider compressed kernels
for $\ell = 64$ flatspace RBC and teleportation.
Figure \ref{fig:PolesFlatspaceRBCandTLPell64} plots
{\em scaled} pole locations for a 20-pole compressed
kernel $\widehat{\Xi}_{64}(s,15)$ which approximates
$\widehat{\Omega}_{64}(s,15)$ and for 20, 28, and
36-pole versions of a compressed kernel
$\widehat{\Xi}{}^E_{64}(s,15,240)$ which approximates
$\widehat{\Phi}_{64}(s,15,240)$. Here we have scaled
all pole locations by a factor $r/(\ell+1/2) = 15/64.5$ in
order to plot them relative to the curve $\mathcal{C}$,
on which the actual scaled zeros $b_{64,k}/64.5$ lie
(at least to the eye). Figure \ref{fig:PolesFlatspaceRBCandTLPell64}
shows that, compared with poles for compressed teleportation
kernels, the poles for the compressed RBC kernel lie much closer
to $\mathcal{C}$. Nevertheless, for both compressed RBC and
teleportation kernels as the number of approximating poles
increases (corresponding to a smaller tolerance $\varepsilon$),
more of the approximating poles ``lock on" to $\mathcal{C}$.
This behavior is evident in the right blow-up plot, where for
20, 28, and 32-pole compressed teleportation kernels, we respectively
find 0(circles), 1(diamond), and 3(squares) ``locked-on" poles.
Moreover, at least to the eye, these correspond to ``locked-on"
poles (crosses) for the compressed RBC kernel. Sec.~\ref{sec:EffBH}
briefly discusses these issues for the gravitational case.
\begin{figure}
\includegraphics[clip=true,width=16cm,trim=0 4.25cm 0 0cm]
{PolesFlatspaceRBCandTLPell64.eps}
\caption{{\sc Scaled pole locations for compressed kernels.}
Here TLP means {\em teleportation}, and the curve $\mathcal{C}$
is described by the parameterization $z(\lambda)$ given in
\eqref{eq:curveC}. In the blow-up plot there is a
cross-diamond-square coalescence on $\mathcal{C}$ (near its
right end). See the text for further explanation.}
\label{fig:PolesFlatspaceRBCandTLPell64}
\end{figure}
\subsection{Blackhole perturbations} \label{subsec:BH_perts}
We consider the following
rescaled versions of the generic master
equation \eqref{eq:RWZeqn} (retaining
the same stem letter $V$ for the potentials):
\begin{equation}\label{eq:RWZeqn_rho}
\frac{\partial^2\Psi_\ell}{\partial \tau^2}
-\frac{\partial^2\Psi_\ell}{\partial \rho_*^2} +
V^\mathrm{RW,Z}_\ell(\rho)\Psi_\ell = 0,
\end{equation}
here in terms of rescaled coordinates
\begin{equation}
\rho = r/(2M),\quad
\tau = t/(2M),\quad
\rho_* = \rho + \log(\rho -1).
\end{equation}
Expressions for the Regge-Wheeler
$V^\mathrm{RW}_\ell(\rho)$ and Zerilli
$V^\mathrm{Z}_\ell(\rho)$
potentials are given in Eqs.~(3) and (4) of
Ref.~\cite{Lau3} (expressions in
terms of $r$ rather than $\rho$ are given in
\cite{FHL}).
The formulas we present here hold
for both formalisms, and have been drawn from
Refs.~\cite{Lau1,Lau2,Lau3}. As before in our analysis
of the flatspace radial wave equation, here we also
suppress the azimuthal index $m$ on $\Psi_{\ell}$.
\subsubsection{Structure of outgoing solutions}
With $\sigma = 2Ms$ the rescaled Laplace frequency,
formal Laplace transformation of \eqref{eq:RWZeqn} yields
\begin{equation}\label{eq:RWZeqn_Laplace}
- \frac{d^2\widehat{\Psi}_\ell}{d \rho_*^2}
+ V^\mathrm{RW,Z}_\ell(\rho)\widehat{\Psi}_\ell +
\sigma^2\widehat{\Psi}_\ell = 0,
\end{equation}
Outgoing solutions of the last equation can be expressed
as an asymptotic series\footnote{\label{fn:asymptotic_series}
The coefficients $d_{\ell,k}(\sigma)$ defining
the asymptotic series $W_\ell(z,\sigma) \sim \sum_{k=0}^\infty
d_{\ell,k}(\sigma)z^{-k}$ are respectively defined by three-term
and five-term recursion relations in the Regge-Wheeler
and Zerilli formalisms~\cite{Lau1,Lau2,Lau3}.} about $\rho=\infty$,
\begin{equation} \label{eq:OutgoingBH}
\widehat{\Psi}_\ell(\sigma,\rho) = a(\sigma)\sigma^\ell
e^{-\sigma\rho_*} W_\ell(\sigma\rho,\sigma),\qquad
W_\ell(\sigma\rho,\sigma)
\mathop{\sim}_{\scriptstyle \rho\rightarrow\infty} 1
\end{equation}
[cf.~Eq.~\eqref{eq:ellmult_fd} for a flatspace
multipole].
Notice that $W_\ell(\sigma\rho,\sigma) = W_\ell(sr,2Ms)$,
and so {\em formally} the flatspace expression $W_\ell(sr) =
W_\ell(sr,0)$.
\subsubsection{Radiation boundary conditions}
We again assume initial data of compact support, namely that
$\Psi_\ell(0,\rho) = 0 = (\partial_\tau\Psi_\ell)(0,\rho)$ for
$\rho > \rho_b - \delta$, where $\rho_b$ specifies the outer
boundary. We now work with any $\rho \geq \rho_b$, in terms
of which exact radiation conditions satisfied by a generic
asymptotically outgoing multipole \eqref{eq:OutgoingBH}
have the following frequency-domain and time-domain
forms~\cite{Lau1,Lau2,Lau3}:
\begin{equation}\label{eq:BH_RBC}
\sigma \widehat{\Psi}_\ell
+ \partial_{\rho_*}\widehat{\Psi}_\ell
= \frac{1}{\rho}\left(1-\frac{1}{\rho}\right)
\widehat{\omega}_\ell\widehat{\Psi}_\ell
\iff
\partial_\tau \Psi_\ell
+ \partial_{\rho_*}\Psi_\ell
= \frac{1}{\rho}\left(1-\frac{1}{\rho}\right)
\omega_\ell * \Psi_\ell,
\end{equation}
where the frequency domain radiation kernel is
\begin{equation}
\widehat{\omega}_\ell(\sigma,\rho) \equiv
\sigma\rho
\frac{W'_\ell(\sigma\rho,\sigma)}{W_\ell(\sigma\rho,\sigma)},
\end{equation}
with the prime denoting differentiation in the first argument.
\begin{figure}
\includegraphics[clip=true,width=14.0cm]{ProfilesEll64Rho15RBC.eps}
\caption{{\sc Profiles for an $\rho = 15$, $\ell = 64$ Regge-Wheeler
frequency domain RBC kernel.}
}
\label{fig:RWell64}
\end{figure}
Refs.~\cite{Lau1,Lau3} have argued that the kernel
$\widehat{\omega}_\ell(\sigma,\rho)$ has the following
``sum of poles" representation:\footnote{This terminology
is suggestive only, since in complex analysis poles are
{\em isolated singularities}. Therefore, the integral term
$\widehat{\omega}_\ell^\mathrm{cut}(\sigma,\rho)$ appearing
in \eqref{eq:omegaSOP} is not, strictly speaking, a
``continuous distribution of poles."}
\begin{equation}\label{eq:omegaSOP}
\widehat{\omega}_\ell(\sigma,\rho)
= \widehat{\omega}_\ell^\mathrm{pole}(\sigma,\rho) +
\widehat{\omega}_\ell^\mathrm{cut}(\sigma,\rho)
\equiv \sum_{k=1}^{N_\ell}
\frac{\sigma_{\ell,k}'(\rho)}{\sigma - \sigma_{\ell,k}(\rho)}
-\frac{1}{\pi}\int_0^\infty
\frac{f_\ell(\chi;\rho)}{\sigma+\chi}d\chi,
\end{equation}
where $f_\ell(\chi;\rho) \equiv
\mathrm{Im}\widehat{\omega}_\ell(\chi e^{\mathrm{i}\pi},\rho)$
and the $\sigma_{\ell,k}(\rho)$ are simple roots of
$W_\ell(\sigma\rho,\sigma)$ (analogous to the roots
$b_{\ell,k}/r$ of $W_\ell(sr)$ in the flatspace case).
At least for $\rho \geq 15$, the integer $N_\ell =
\ell$ or $\ell + 1$, when $\ell$ is respectively even
or odd \cite{Lau1,Lau3}. The origin of the extra root,
relative to the flatspace case, in the odd-$\ell$ case
is discussed in Ref.~\cite{Lau3}.
\begin{figure}
\includegraphics[clip=true,width=14.0cm]{ProfilesEll2Rho15_30_60RBC.eps}
\caption{{\sc Profiles for $\ell = 2$ Regge-Wheeler
frequency domain RBC kernels.}
As $\rho$ increases these profiles shrink towards the origin
(the same phenomena occurs for the flatspace RBC kernels).
}
\label{fig:RWell2_shrink}
\end{figure}
Insofar as numerical implementation is concerned, a key
requirement is the ability to evaluate the profiles
$\mathrm{Re}\widehat{\omega}_\ell(\mathrm{i}y,\rho)$ and
$\mathrm{Im}\widehat{\omega}_\ell(\mathrm{i}y,\rho)$ for
$y\in\mathbb{R}$. These evaluations are along
the imaginary $\sigma$-axis, typically the inversion
contour for the inverse Laplace transform. Accurate
methods for such evaluation have been described in \cite{Lau1}.
In fact, these methods are {\em not} based on the sum-of-poles
representation \eqref{eq:omegaSOP}, but this issue
is of no concern here. For example, the profiles
for an $\rho = 15$, $\ell = 64$ Regge-Wheeler kernel are
shown in Fig.~\ref{fig:RWell64}. As $\rho$ increases
these profiles ``shrink" towards the origin (as do the
corresponding flatspace profiles), and this phenomena is
documented in Fig.~\ref{fig:RWell2_shrink}.
With the ability to numerically
generate the profiles
$\mathrm{Re}\widehat{\omega}_\ell(\mathrm{i}y,\rho)$
and $\mathrm{Im}\widehat{\omega}_\ell(\mathrm{i}y,\rho)$,
we are then able to construct approximate kernels
via Alpert-Greengard-Hagstrom (AGH) compression \cite{AGH}.
Here a compressed kernel is a sum of simple poles,
\begin{equation}\label{eq:compressedK}
\widehat{\xi}_\ell(\sigma,\rho) \equiv
\sum_{q=1}^d
\frac{\gamma_{\ell,q}(\rho)}{\sigma-\beta_{\ell,q}(\rho)}
\simeq
\widehat{\omega}_\ell(\sigma,\rho),
\qquad
\mathrm{Re}\beta_{\ell,q}(\rho) < 0, \forall q,
\end{equation}
where the approximation $\widehat{\xi}_\ell(\sigma,\rho)$
satisfies
\begin{equation}\label{eq:superr}
\left|\widehat{\omega}_\ell(\sigma,\rho)
- \widehat{\xi}_\ell(\sigma,\rho)
\right| <
\varepsilon \left|\widehat{\omega}_\ell(\sigma,\rho)\right|,
\qquad
\sigma \in
\mathrm{i}\mathbb{R},
\end{equation}
with $\varepsilon$ a prescribed tolerance. The number
$d$ clearly depends on $\varepsilon$ and $\ell$,
and the numbers $\beta_{\ell,q}$ and $\gamma_{\ell,q}$
depend both on the boundary radius $\rho$ (as indicated) and
on $\ell$ (the dependence on which we have restored here).
The modifier {\em compressed} in the description
of $\widehat{\xi}_\ell(\sigma,\rho)$ is apt.
Indeed, as described in Sec.~\ref{sec:Eff}, for the ordinary
wave equation
the exact frequency domain kernel admits a similar rational
approximation with $d$ scaling as \eqref{eq:AGHdscaling}.
Similar scaling has been observed empirically for approximations
$\widehat{\xi}_\ell(\sigma,\rho)$ of blackhole
kernels $\widehat{\omega}_\ell(\sigma,\rho)$ \cite{Lau1}.
Algorithm \ref{tab:AGHcomp} summarizes
our implementation of AGH compression (see Ref.~\cite{Lau2}
for a complete description). Let us further comment on
Algorithm \ref{tab:AGHcomp}, with numbers appropriate for
$\ell = 2$. Typically $y_\mathrm{max} = 300/\rho$ for step 1.
For step 2 we have typically chosen 10 to 20 adaptive
levels centered around the origin with 65 points at the
bottom level, about $10^3$ grid points $y_j$ in all.
Fig.~\ref{fig:ygrid} depicts an example $y$-grid.
For step 3 the evaluation at each $y_j$ requires ODE integration
in the complex plane with upwards of $10^{10}$ floating
point operations in double precision (more in quad precision).
Step 5 is a confirmation step meant to verify \eqref{eq:superr}.
Ideally, this confirmation takes place with a
much larger $y$-window than $[-y_\mathrm{max},y_\mathrm{max}]$, and
on a different (dense and uniform) $y$-grid. This step involves
further ODE integration and is therefore as or more expensive than
step 3.
\begin{figure}
\includegraphics[trim=0cm 4cm 0cm 4cm, clip=true,width=12.0cm]{ygrid.eps}
\caption{{\sc $y$-grid to evaluate the profiles
$\mathrm{Re}\widehat{\omega}_2(\mathrm{i}y_j,\rho)$
and $\mathrm{Im}\widehat{\omega}_2(\mathrm{i}y_j,\rho),\forall j$.}
This grid has 21 points, 4 adaptive levels, and 9 points at the bottom level.}
\label{fig:ygrid}
\end{figure}
\begin{algorithm}[H]
\caption{{\sc Steps for compressing an RBC kernel.}\label{alg:compressRBC} }
\label{tab:AGHcomp}
\vspace{5pt}
INPUT: $\ell$, $\rho = \rho_b$, $d$
(orbital index, dimensionless boundary radius,
desired number of poles) \\
OUTPUT: $\{ \beta_{\ell,q}(\rho),
\gamma_{\ell,q}(\rho):
q = 1,\dots,d\}$
(compressed kernel)
\vspace{5pt}
\begin{algorithmic}[1]
\State Choose an approximation window
$[-y_\mathrm{max},y_\mathrm{max}]$
on the $\sigma = \mathrm{i}y$ imaginary axis.
\State Partition $[-y_\mathrm{max},y_\mathrm{max}]$ to
form a $y$-grid, typically with mesh refinement
at the origin.
\State
Numerically evaluate the profiles
$\mathrm{Re}\widehat{\omega}_2(\mathrm{i}y_j,\rho)$
and $\mathrm{Im}\widehat{\omega}_2(\mathrm{i}y_j,\rho)$
on the $y$-grid.
\State
Compute the numbers $\{\beta_{\ell,q}(\rho),
\gamma_{\ell,q}(\rho): q = 1,\dots,d\}$
by AGH compression.
\State
Using $\{\beta_{\ell,q}(\rho),
\gamma_{\ell,q}(\rho): q = 1,\dots,d\}$,
verify \eqref{eq:superr}. If not verified,
repeat with $d \leftarrow d+1$.
\end{algorithmic}
\end{algorithm}
From the standpoint of implementation, the representation
\eqref{eq:compressedK} is crucial, since it implies
that the time-domain convolution can be approximately
evaluated via integration of ODE at the boundary.
For a typical explicit ODE scheme
and a sufficiently small time-step, integration of these
ODE is numerically stable since the relevant poles in
\eqref{eq:compressedK} lie in the left-half plane.
Let $\xi_\ell(\tau,\rho)$ be the inverse Laplace transform
of $\widehat{\xi}_\ell(\sigma,\rho)$ with respect to $\sigma$.
Then the approximate time-domain kernel $\Xi_\ell(t,r)
\equiv (1/2M)\xi_\ell(t/(2M),r/(2M))$ appearing in
\eqref{eq:LaplaceConvApprox} is the inverse Laplace
transform (with respect to $s$) of
\begin{equation}
\widehat{\Xi}_\ell(s,r) = \sum_{q=1}^d
\frac{\gamma_{\ell,q}(\rho)}{2M}
\widehat{\Xi}_{\ell,q}(s,r),
\qquad
\widehat{\Xi}_{\ell,q}(s,r) =
\frac{1}{s-(2M)^{-1}\beta_{\ell,q}(\rho)},
\end{equation}
where, unlike in Section \ref{sec:HowTo},
here $\ell$-dependence has not been suppressed.
\subsubsection{Asymptotic waveform evaluation
and teleportation}\label{subsec:BHWE}
Similar to before, we introduce two kernels: (i)
one $\theta_\ell$ for evaluation of an underlying profile,
and (ii) another $\phi_\ell$ for AWE/teleportation of the
waveform. The first type of kernel is
defined by [cf.~Eq.~\eqref{eq:flatspaceThetaOmega}]
\begin{equation}
\widehat{\theta}_\ell(\sigma,\rho)
= \frac{1}{\sigma^\ell}\underbrace{\exp\left[\int_\rho^\infty
\frac{\widehat{\omega}_\ell(\sigma,\eta)}{\eta}
d\eta\right]}_{1/W_\ell(\sigma\rho,\sigma)} \, ,
\end{equation}
and satisfies $\widehat{\theta}_\ell(\sigma,\rho)
\widehat{\Psi}_{\ell}(\sigma,\rho) = a(\sigma)\exp(-\sigma\rho_*)$,
as can be seen directly from Eq.~\eqref{eq:OutgoingBH}.
We can not analytically perform the integration here.
However, the pole part of the kernel can
be exactly integrated to remove the singularity.
Indeed, we find
\begin{align}
\widehat{\theta}_\ell(\sigma,\rho) =
\frac{
\exp\left[\int_\rho^\infty
\frac{\widehat{\omega}_\ell^\mathrm{cut}(\sigma,\eta)}{\eta}
d\eta\right]}{
\sigma^\ell\exp\left[-\int_\rho^\infty
\frac{\widehat{\omega}_\ell^\mathrm{pole}(\sigma,\eta)}{\eta}
d\eta\right]}
= \frac{\exp\left[\int_\rho^\infty
\widehat{\omega}_\ell^\mathrm{cut}(\sigma,\eta)\eta^{-1}d\eta
\right]}{\prod_{k=1}^{N_\ell}
\big[\sigma-\sigma_{\ell,k}(\rho)\big]}.
\end{align}
Teleportation $(\rho_1 \rightarrow \rho_2 \leq \infty)$ is defined
through the kernel [cf.~Eq.~\eqref{eq:flatspacePhi}]
\begin{align}\label{eq:BHteleportK}
\widehat{\phi}_\ell(\sigma,\rho_1,\rho_2) & = -1 +
\underbrace{\exp\left[
\int_{\rho_1}^{\rho_2}\frac{\widehat{\omega}_\ell(\sigma,\eta)}{\eta}
d\eta\right]}_{W_\ell(\sigma\rho_2,\sigma)/W_\ell(\sigma\rho_1,\sigma)}.
\end{align}
Adjusting for the $(\rho^*_2 - \rho^*_1)$ time delay, this
kernel allows for conversion of the signal at $\rho_1$
to the signal at $\rho_2$, as it satisfies
\begin{equation}
e^{\sigma(\rho^*_2-\rho^*_1)}\widehat{\Psi}_\ell(\sigma,\rho_2) =
\widehat{\phi}_\ell(\sigma,\rho_1,\rho_2)\widehat{\Psi}_\ell(\sigma,\rho_1)
+ \widehat{\Psi}_\ell(\sigma,\rho_1).
\end{equation}
In the time domain we therefore recover the desired property
\begin{equation}\label{eq:teleport}
\Psi_\ell(\tau+(\rho^*_2-\rho^*_1),\rho_2) =
(\phi_\ell(\cdot,\rho_1,\rho_2)*\Psi_\ell(\cdot,\rho_1))(\tau) +
\Psi_\ell(\tau,\rho_1).
\end{equation}
Exact evaluation of the asymptotic waveform corresponds to $\rho_2 = \infty$.
\begin{figure}
\includegraphics[clip=true,width=14.0cm]{ProfilesEll2Rho15to1E15EXT.eps}
\caption{Teleportation kernel $\widehat{\phi}_2(\mathrm{i}y,15,10^{15})$
for $\rho_1 = 15 \rightarrow \rho_2 = 10^{15}$.}
\end{figure}
Let us describe how we numerically approximate
$\widehat{\phi}_\ell(\sigma,\rho_1,\rho_2)$ as a pole
sum $\widehat{\xi}_\ell^E(\sigma,\rho_1,\rho_2)$.
A simple version of the procedure is to follow the steps listed
in Algorithm \ref{tab:AGHcomp}, replacing step 3 with evaluations of
the profiles
$\mathrm{Re}\widehat{\phi}_\ell(\mathrm{i}y_j,\rho_1,\rho_2)$ and
$\mathrm{Im}\widehat{\phi}_\ell(\mathrm{i}y_j,\rho_1,\rho_2)$.
To generate these profiles, we use \eqref{eq:BHteleportK}, and
for each $y_j$ evaluation point perform the $\eta$ integration
using a composite Gauss-Kronrod rule.
Assuming $[A,B]$ is one
subinterval of $[\rho_1,\rho_2]$ in the composite
rule, the details are as follows. Let $\eta = AB/[(B-A)q + A]$ for
$q \in [0,1]$, so that
\begin{equation}
\int_A^B \frac{\widehat{\omega}_\ell(\mathrm{i}y_j,\eta)}{\eta}
d\eta
\simeq \sum_{p=1}^{N_\mathrm{GK}} \frac{\widehat{\omega}_\ell
\big(\mathrm{i}y_j,AB/[(B-A)q_p + A]\big)}{q_p + A/(B-A)}w_p,
\end{equation}
where $(q_p,w_p)_{p=1}^{N_\mathrm{GK}}$ are the 15 nodes and
weights for the Gauss-Kronrod rule relative to $[0,1]$, and by
the above equation we mean the same rule is separately applied
to the real and imaginary parts of
$\widehat{\omega}_\ell(\mathrm{i}y_j,\eta)$. This
numerical integration is accurate, because, in fact, for
each grid-point $y_j$ integration of both
Re$\widehat{\omega}_\ell(\mathrm{i}y_j,\eta)$ and
Im$\widehat{\omega}_\ell(\mathrm{i}y_j,\eta)$ always
involves terms of the same sign. With $N_C$ denoting the number
composite subintervals, profiles for the RBC kernels on the
$y$-grid must be computed $N_\mathrm{GK}\cdot N_C$ times.
If $\rho_2 = \infty$ the last composite interval might
be handled through a semi-infinite quadrature.
Approximation of $\widehat{\theta}_\ell(\sigma,\rho)$
follows a procedure similar (although more complicated) to
the one outlined above for
$\widehat{\phi}_\ell(\sigma,\rho_1,\rho_2)$.
Unfortunately, this simple procedure becomes too costly
when $\rho_2 \gg \rho_1$. The problem is two-fold. First,
$N_C$ must be chosen large. Second, and more serious, the
approximation window $[-y_\mathrm{max},y_\mathrm{max}]$
is fixed by the profiles for $\rho_1$ (the ``widest" profiles
in the integration). However, since, as seen from
Fig.~\ref{fig:RWell2_shrink}, the RBC profiles shrink as $\rho$
increases, the $y$-grid needs many adaptive levels to resolve
the contribution to the $\eta$-integration from the profiles at
and near $\rho_2$. The $y$-grid then must have a large number
of points, on which $N_\mathrm{GK}\cdot N_C$ complex function
evaluations are made (with each such evaluation costing
$\simeq 10^7$ floating point operations). To bypass this issue,
we follow another rather complicated procedure, whereby the
interval $[\rho_1,\rho_2]$ is first broken up into $\mathcal{N}$
chunks $[\rho_\alpha,\rho_{\alpha+1}]$ which are typically
decades like $[10^\alpha,10^{\alpha+1}]$ if $\rho_2$ is very
large. Each chunk has its own approximation window
$[-y^\alpha_\mathrm{max},y^\alpha_\mathrm{max}]$ and
(relatively small) $y^\alpha$-grid, and we choose these
to conform with how shrunken the profiles are for
$\rho \simeq \rho_\alpha$. Next, using the relatively simple
procedure described in the last paragraph, for each of the
$\mathcal{N}$ chunks we construct a compressed kernel (table)
which approximates
$\widehat{\phi}_\ell(\mathrm{i}y,\rho_\alpha,\rho_{\alpha+1})$ as a
sum of poles
$\widehat{\xi}_\ell^E(\mathrm{i}y,\rho_\alpha,\rho_{\alpha+1})$.
The last step is to generate profiles (for the compression algorithm)
on a large $y$-grid associated with a wide approximation window
and sufficient resolution near the origin. However, these
evaluations are now done via combination of all $\mathcal{N}$
tables. Therefore, they are drastically faster,
since they are carried out through auxiliary evaluations
made with the $\mathcal{N}$ distinct pole sums (rather than
ODE integration). Finally, we note that the physical
teleportation kernel used in a numerical simulation is
\begin{equation}
\Xi^E_\ell(t,r_1,r_2) =
\frac{1}{2M}\xi^E_\ell(t/(2M),r_1/(2M),r_2/(2M)),\qquad
\end{equation}
where $\xi_\ell^E(\tau,\rho_1,\rho_2)$ is the
inverse Laplace transform (with respect to $\sigma$) of
$\widehat{\xi}_\ell^E(\sigma,\rho_1,\rho_2)$.
\subsubsection{Efficiency and storage} \label{sec:EffBH}
As either $\ell$ or $\varepsilon^{-1}$ becomes large the scaling
observed in \cite{Lau1} for compressed, blackhole, RBC kernels
appears similar to the flatspace result \eqref{eq:AGHdscaling}
described in Sec.~\ref{sec:Eff}.
However, we stress that these are empirical
observations, and there is no corresponding proof of
\eqref{eq:AGHdscaling} or a similar result in the blackhole case.
Nevertheless, provided that the number of approximating
exponentials in the blackhole case indeed grows sublinearly with
both $\ell$ and the inverse $1/\varepsilon$ of the relative
approximation error [cf.~Eq.~\eqref{eq:superr}], our implementation
of RBC satisfies the same efficient scalings \eqref{eq:RBCscalings}
established for flatspace compressed kernels.
One might similarly ask whether or not our implementation of AWE
for blackholes satisfies these scalings; we do not have
an answer for this question, but here consider kernels from
the $\ell= 64$ teleportation experiment considered later in Section
\ref{subsec:PulseT}. Let us focus on the compressed (frequency domain)
kernels $\widehat{\xi}{}_{64}(\sigma,15)$ and
$\widehat{\xi}{}_{64}(\sigma,15,240)$,
respectively for RBC at $\rho_b = 15$ and teleportation from
$\rho_1 = 15$ to $\rho_2 = 240$. Figure \ref{fig:RWell64} has
already depicted the profiles from which the 25-pole approximation
$\widehat{\xi}{}_{64}(\sigma,15)$ is constructed. Notice that the
approximation $\widehat{\xi}{}_{64}(\sigma,15,240)$ has 32 poles,
whereas the corresponding exact {\em flatspace} teleportation kernel
would have 64 poles. Similar reduction for the $\ell = 64$
occurred for compressed flatspace kernels considered earlier.
As depicted in Fig.~\ref{fig:RBCandEXTell64kernComparison} and
similar to the situation encountered in
Fig.~\ref{fig:PolesFlatspaceRBCandTLPell64},
the pole locations for $\widehat{\xi}{}_{64}(\sigma,15)$ and
$\widehat{\xi}{}_{64}(\sigma,15,240)$ are different.
Finally, we remark that we have tried to enforce the condition that
the teleportation kernel has the same pole {\em locations}
as the RBC kernel [cf.~the discussion
around Eq.~\eqref{eq:WEandRBCpoles}]; however, we are then
unable to achieve a compressed kernel with any accuracy whatsoever.
The previous paragraph has considered the large-$\ell$,
small-$\varepsilon$ limits. However, in this paper we mostly consider
$\varepsilon$ fixed (typically machine precision) and
small $\ell$, in which case, as we have seen, $d > \ell$. We remark
that this situation is similar to the case of ordinary wave propagation
on $2+1$ flat spacetime. In that setting the low-$n$ (Fourier index)
``circle kernels" are also expensive to evaluate, with
scalings similar to \eqref{eq:AGHdscaling}
only exhibited in the large $n$ limit \cite{AGH}.
While the previous discussion has presented scalings for various limits,
in practice our implementation of RBC/teleportation
for $\ell = 2,3,$ and $64$
amounts to adding on the order of 20 to 30 points to the spatial domain,
and modest increase in work and storage costs for the
numerical experiments considered here.
\begin{figure}
\includegraphics[clip=true,width=14.0cm]{RBCandEXTell64kernComparison.eps}
\caption{{\sc Compressed RBC and teleportation kernels for $\ell = 64$.}
The right pane plots the modulus of the kernels for
positive $y$. For the RBC kernel (dotted line) this corresponds to the
modulus of the combined profiles shown in Fig.~\ref{fig:RWell64}, but
here with only half of the domain for the independent variable $y$.
\label{fig:RBCandEXTell64kernComparison}}
\end{figure}
\section{Numerical Experiments}\label{sec:Experiments}
To carry out numerical simulations, we have used both
the nodal Legendre discontinuous Galerkin method described
in Ref.~\cite{FHL} (further details of this method will
not be given here) and a nodal Chebyshev method. Both
methods feature multiple subdomains and upwinding.
\subsection{Pulse teleportation}\label{subsec:PulseT}
First consider the $\ell = 2$ Regge-Wheeler equation
with the same initial data \eqref{eq:GaussianData} given in
Subsection \ref{subsec:QNRTails}. Using our multidomain nodal
Chebyshev method, we perform five separate evolutions on
domains with outer boundaries taken as the $b$ values
corresponding to $r_b = 30M$, $60M$, $120M$, $240M$, and $480M$.
We have respectively used 32, 37, 45, 62, and 95 subintervals
of uniform size, and in each case with 32 Chebyshev-Lobatto
points per subinterval. Therefore, the spatial
resolution for each evolution is comparable to the others.
Evolutions are performed by the classical
4-stage explicit Runge Kutta method with timestep
$\Delta t \simeq M(2.6794\times 10^{-4})$. For each
evolution the inner boundary is $a = -200M$, and therefore the
Sommerfeld boundary condition $-(\Pi+\Phi)= 0$ at the inner
boundary is essentially exact. For all choices of outer boundary
$b$ we adopt the Laplace convolution RBC \eqref{eq:LaplaceConvApprox}.
Tables for $r_b = 30M$, $60M$, $120M$, $240M$, and $480M$ respectively
have 19, 19, 19, 18, and 17 poles, with each table computed in
quadruple precision to satisfy the tolerance $\varepsilon = 10^{-15}$.
These tables are available at \cite{Kernel_web1}.
\begin{figure}
\includegraphics[clip=true,width=14.0cm]{PulseComparisonEll2.eps}
\caption{{\sc Errors in read-off and teleported $\ell = 2$ waveforms
relative to read-off waveform at $r_2 = 480M$.}
\label{fig:PulseComparison2}
}
\end{figure}
In all cases the field $\Psi(t,r_b)$ is recorded as a
time series at the boundary $b$, and in all cases but
the last ($b$ corresponding to $r_b = 480M$)
we ``teleport" the field from $r_1 = r_b$ to $r_2 = 480M$.
Each approximate teleportation kernel
$\Xi^E_2 \simeq \Phi_2$ features the same pole locations
as the corresponding approximate RBC kernel
$\widehat{\Xi}\simeq \Omega_2$. For the last $r_b = 480M$
simulation we simply record the field at the boundary, with
this record then serving as a reference time series.
We account for time delays by starting all recorded times
series (whether read off or teleported) at time $b - 6M$.
The top panel in Figure \ref{fig:PulseComparison2} plots the
errors in the waveforms recorded at the different $b$
boundaries as compared to the reference $r_b = 480M$ waveform;
as expected the systematic errors are large.
The bottom panel plots the errors in the
($r_1 = r_b$ to $r_2 = 480M$) teleported
time series relative to the reference time series.
With the reference time series viewed as the ``asymptotic signal",
this ``AWE" clearly yields 10 or more digits of
accuracy relative to simple read-off. We have found similar results
using other ``pulses" based on polynomial, Lorentzian,
and trigonometric profiles (in all case with the initial
data initially supported away from the boundary, either
exactly or to machine precision).
\begin{figure}
\includegraphics[clip=true,width=14.0cm]{PulseComparisonEll3.eps}
\caption{{\sc Errors in read-off and teleported $\ell = 3$ waveforms
relative to read-off waveform at $r_2 = 480M$.}
\label{fig:PulseComparison3}
}
\end{figure}
\begin{figure}
\includegraphics[clip=true,width=14.0cm]{PulseComparisonEll64.eps}
\caption{{\sc Errors in read-off and teleported $\ell = 64$ waveforms
relative to read-off waveform at $r_2 = 480M$.}
\label{fig:PulseComparison64}
}
\end{figure}
We repeat the experiment for two different $\ell$ values. First, for
$\ell = 3$ we adopt the same initial data and experimental setup,
except for the numerical tables which specify the RBC and
teleportation kernels. For the $\ell = 3$ experiment the
number of poles for the RBC tables is either 15 or 16, and the
number of poles for teleportation tables ranges from 15 to 20.
The results, shown in Fig.~\ref{fig:PulseComparison3},
are comparable to those for $\ell = 2$. Lastly, we repeat the
experiment for $\ell = 64$. For such a high $\ell$ the evolutions
are much more expensive, due to finer oscillations in both space and
time. We now use 42 points per subdomain, with the number of subdomains
typically increased by a factor of 2 or 3 relative to the numbers
given above for $\ell = 2$. Moreover, we adopt the timestep
$\Delta t = M(4.0461 \times 10^{-5})$ and inner boundary $a = -80M$
which is casually disconnected from each waveform read-off/teleportation
at the outer boundary. For $\ell = 64$
our RBC tables have between 23 and 25 poles, and our teleportation
tables between 30 and 32 poles. While these tables are large,
note that even the corresponding {\em exact} flatspace RBC
kernels would have 64 poles. Hence, in this experiment the savings
afforded by kernel compression is already evident. Results are
depicted in Fig.~\ref{fig:PulseComparison64}.
\subsection{Luminosities from extreme-mass-ratio binaries}
\label{subsec:luminosities}
An extreme mass ratio binary (EMRB) is a system comprised of
a small mass-$m_p$ compact object (the ``particle")
orbiting a much larger mass-$M$ blackhole, where the mass
ratio $m_p/M \ll 1$. EMRB systems
are expected to emit gravitational radiation in a low frequency
band ($10^{-5}$ to $10^{-1}$ {\tt Hz}), and therefore offer the
promise of detection by a space-based
gravitational wave observatory like the earlier proposed LISA
project \cite{ScottHughes_LISA}. Located within the solar system,
such an observatory would be well approximated as positioned at
$\mathscr{I}^+$ relative to expected sources.
A standard method for studying EMRBs uses the perturbation
theory of Schwarzschild blackholes in an approximation which
treats the particle as a point-like Dirac delta function.
The particle follows a timelike
geodesic in the background Schwarzschild spacetime and is
responsible for generating small metric perturbations which
radiate away
(see Refs.~\cite{Martel_CovariantPert,SopuertaLaguna,Sarbach:2001qq}
for modern accounts of the subject).
Here we note that the axial metric perturbations
for each $(\ell,m)$ mode may be combined to form a gauge
invariant scalar quantity $\Psi_{\ell m}^{\mathrm{CPM}}$ which
obeys Eq.~\eqref{eq:RWZeqn} with the Regge-Wheeler
potential $V^\mathrm{RW}(r)$
and a distributional forcing term
$S_{\ell m}^{\mathrm{CPM}}(t,r)$.\footnote{
\label{foot:RWCPM}
We use the Cunningham-Price-Moncrief (CPM) master function
\cite{CPM:paper} which yields formulas for the axial sector which are on the
same footing as those for the polar sector.}
Likewise, the polar metric perturbations for each $(\ell,m)$
mode may be combined to form a gauge invariant scalar
quantity $\Psi_{\ell m}^{\mathrm{Z}}$ which obeys
Eq.~\eqref{eq:RWZeqn} with
the Zerilli potential $V^\mathrm{Z}(r)$ and a
distributional forcing term $S_{\ell m}^{\mathrm{Z}}(t,r)$.
Both $S^{\mathrm{Z}}_{\ell m}$ and $S^{\mathrm{CPM}}_{\ell m}$
are built from linear combinations of an $(\ell,m)$ mode
decomposition of the stress-energy tensor and, as such,
depend on the particle trajectory
$(r_p(t),\pi/2,\phi_p(t))$ in the equatorial plane.
Bounded and stable orbits are characterized
by an eccentricity constant $e$ and a semi-latus rectum
constant $p$. Upon specification of $(e,p)$, the resulting
trajectory is found by integrating the relevant system of
ODEs given in Eq.~(5) of Ref.~\cite{FHL}. Appendix C of
\cite{FHL} gives exact expressions
for $S_{\ell m}^{\mathrm{CPM,Z}}$
(see also \cite{SopuertaLaguna,Hopper:2010uv}).
With both $S_{\ell m}^{\mathrm{CPM,Z}}$ specified,
we numerically solve for $\Psi_{\ell m}^{\mathrm{CPM,Z}}$,
starting with trivial initial data and smoothly turning on
the source term $S_{\ell m}(t,r)$ over
a timescale $\tau \simeq 150M$ to $450M$
to prohibit static {\em Jost junk} solutions which may appear
in some formulations when using inconsistent initial data
\cite{Field2010,PhysRevD.83.061503}. Respectively,
Sommerfeld and Laplace convolution RBC are enforced at the left and right
physical boundaries (cf.~Sec.~\ref{subsec:simpleRBC}). The computational domain
is given by the interval
$[-400M,b]$, where the tortoise coordinate value $b$
corresponds to $r_b = 60M$ in Schwarzschild radius.
Notice that as an approximation to the asymptotic
signal at $\mathscr{I}^+$, the waveform read-off at $r_b$ will
have an $O(r_b^{-1})$ systematic error, suggesting relative errors
greater than one percent for $r_b = 60M$.
For our simulations, we have chosen $16$ and $3$
subdomains to the left and right of the delta function
respectively and represent the numerical solution by
an order-$40$ or order-$46$ polynomial on each subdomain.
The distributional source terms
$S_{\ell m}^{\mathrm{CPM,Z}}$ determine jump conditions in
the fields $\Psi_{\ell m}^{\mathrm{CPM,Z}}$ at the particle
location which we impose as junction conditions between
subintervals ~\cite{FHL}, with the motion of the particle
incorporated through a time-dependent coordinate transformation.
Our particular choice of $r_b = 60M$ ensures that the particle
does not come too close to the outer computational boundary
which might lead to over stretching of the coordinates.
Temporal integration is carried out with a classical
fourth-order Runge-Kutta method with timestep
$\Delta t = M(5 \times 10^{-3})$.
Computation of the luminosities for a particular orbit of the perturbing
particle is a standard benchmark test. For each $(\ell, m)$
mode the energy and angular momentum luminosities at $\mathscr{I}^+$
(denoted by $\infty$) and at the event horizon $r = 2M$ (denoted by $H$)
are given by
\begin{subequations} \label{eqn:flux_mode}
\begin{align}
\dot{E}^{\infty,H}_{\ell m} & = \frac{1}{64 \pi}
\frac{\left(\ell + 2\right)!}{\left(\ell - 2\right)!}
\left< | \dot{\Psi}_{\ell m}^\mathrm{Z}|^2
+ |\dot{\Psi}_{\ell m}^\mathrm{CPM}|^2 \right>, \\
\dot{L}^{\infty,H}_{\ell m} & = \frac{\mathrm{i}m}{64 \pi}
\frac{\left(\ell + 2\right)!}{\left(\ell - 2\right)!}
\left< \dot{\Psi}^\mathrm{Z}_{\ell m} \bar{\Psi}^\mathrm{Z}_{\ell m}
+ \dot{\Psi}_{\ell m}^\mathrm{CPM} \bar{\Psi}^\mathrm{CPM}_{\ell m} \right> ,
\end{align}
\end{subequations}
where the average $\left< A(t) \right>$ of a time
series $A(t)$ is computed as
\begin{equation} \label{eqnAveragedQuant}
\langle A \rangle \equiv \frac{1}{T_2-T_1}\int_{T_1}^{T_2} A(t) dt,
\qquad
T_2 - T_1 = T_r.
\end{equation}
Here $T_r$ is the period of radial oscillation for the particle orbit.
Before presenting our numerical results, we remark on the
potential sources of error in an evaluated asymptotic waveform.
At $b$ we record both $\Psi_{\ell m}^{\mathrm{CPM,Z}}(t,b)$
and their first temporal derivatives as a time-series. With
the numerical setup described above, the relative
pointwise error associated with these read-off waveforms
is better than $10^{-10}$. An
additional source of (systematic) error is due to
trivial initial data which, both incorrect and inconsistent,
is known to generate spurious junk. At a finite and fixed
radial location, spurious junk radiation propagates away (the
potential for {\em static} Jost junk is discussed
in Ref.~\cite{Field2010}), although due to backscattering
a ``junk error tail" may develop which decays more slowly. Tail fields
are expected to fall off like $t^{-4}$ at $\mathscr{I}^+$, and $t^{-7}$
at a fixed (much smaller) radial value. Evidently, the situation
is worse at $\mathscr{I}^+$ where junk error tails decay more slowly.
Additionally, we often need to average luminosity quantities over
long periods of time. Taken
together, these facts conspire to make the temporal average
of a $\mathscr{I}^+$ waveform especially prone to contamination by junk error
tails, even at late times and especially when high accuracy is desired.
Unfortunately, simply waiting for junk errors to die out may
not be practical, because ODE and PDE numerical integrators typically
introduce numerical errors which grow linearly with time.
While convolution with an approximate teleportation
kernel will introduce additional error,
we believe that the dominant errors in our
asymptotic waveforms stem from numerical method error and spurious junk.
\begin{table*}[h!]
{\scriptsize
\begin{tabular}{| cc || ll | ll | ll | ll |}
\hline
$m$ & Alg.
& \multicolumn{2}{c|}{$\dot{E}^\infty_{2 m}$}
& \multicolumn{2}{c|}{$\dot{E}^H_{2 m}$}
& \multicolumn{2}{c|}{$\dot{L}^\infty_{2 m}$}
& \multicolumn{2}{c|}{$\dot{L}^H_{2 m}$} \\
\hline
\hline
0 & FR
& 1.27486196317 & $\times 10^{-8}$
& 1.66171571270 & $\times 10^{-8}$
& \multicolumn{2}{c|}{0}
& \multicolumn{2}{c|}{0} \\
& AWE
& 1.27486196187 & $\times 10^{-8}$
& 1.66171571269 & $\times 10^{-8}$
& \multicolumn{2}{c|}{0}
& \multicolumn{2}{c|}{0} \\
\hline
1 & FR
& 1.15338054092 & $\times 10^{-6}$
& 3.08063328605 & $\times 10^{-7}$
& 1.44066000650 & $\times 10^{-5}$
& 2.77518962557 & $\times 10^{-6}$\\
& AWE
& 1.15338054091 & $\times 10^{-6}$
& 3.08063328606 & $\times 10^{-7}$
& 1.44066000619 & $\times 10^{-5}$
& 2.77518962558 & $\times 10^{-6}$\\
\hline
2 & FR
& 1.55967717209 & $\times 10^{-4}$
& 1.84497995136 & $\times 10^{-6}$
& 2.07778922470 & $\times 10^{-3}$
& 1.85014840343 & $\times 10^{-5}$\\
& AWE
& 1.55967717211 & $\times 10^{-4}$
& 1.84497995135 & $\times 10^{-6}$
& 2.07778922439 & $\times 10^{-3}$
& 1.85014840342 & $\times 10^{-5}$ \\
\hline
\end{tabular}}
\caption{Mode-by-mode $\ell = 2$ luminosities for the
eccentric orbit described in the text. For a particle of
mass $m_p$ these values should be scaled by $m_p^2$.
The table compares our asymptotic-waveform
evaluation (AWE) method with the accurate frequency domain (FR) luminosities.
FR results refer to Table III of Ref.~\cite{Hopper:2010uv}
and are quoted to a relative error of $10^{-12}$.
For this experiment the outer boundary is $r_b = 60M$.
\label{highEFluxTable}
}
\end{table*}
Through $r_b \rightarrow r_{\infty} = 2M(1 \times 10^{15})$
teleportation, we approximately obtain the
signals $\Psi_{\ell m}^{\mathrm{CPM,Z}}$ at $\mathscr{I}^+$, and
with these signals compute the energy and angular momentum
luminosities (\ref{eqn:flux_mode}a,b).
The orbital parameters $(e,p) = (0.764124,8.75455)$ and
initial location $(r_p, \phi_p) = ((pM)/(1+e\cos(\pi/2)),0)$
specify the particle's path.
As described above, we slowly turn-on the distributional source
over a timescale of $\tau \simeq 150M$ to $450M$. A physically
meaningful luminosity measurement will not depend on our choice
of $\tau$, and from this consideration we find that by $3000M$
the spurious junk's effect is minimal. Table~\ref{highEFluxTable}
compares the $\ell=2$ luminosity measurements at $\mathscr{I}^+$ with
the accurate frequency-domain results reported in
Ref.~\cite{Hopper:2010uv}. We match their
stated accuracy to better than $9$ digits.
As our final experiment we consider a circular orbit
specified by the orbital parameters $(e,p) = (0,10)$.
Circular-orbit luminosity measurements are time-independent,
thereby allowing us to (i) better understand the influence of
junk error tails on $\mathscr{I}^+$ waveforms and (ii) estimate
errors due to our AWE procedure in a clean setting.
With the same numerical parameters used for our eccentric
orbit simulations, we compute $\ell=2$
luminosities at $\mathscr{I}^+$ and compare them to accurate
frequency-domain results generated with the code
described in Ref.~\cite{Hopper:2010uv}. By $T=6000M$
our results agree with the frequency-domain
results to within a relative difference of less than $10^{-12}$.
Furthermore, we find the same level of agreement when the
outer boundary is moved inward to $r_b = 30M$,
in which case we use the teleportation/RBC
kernel tables given in Appendix~\ref{App:Tables}
(with the longer teleportation table for AWE).
As the final measurement time is taken earlier,
the agreement becomes progressively worse due to
spurious junk radiation. Indeed, the solid black line in
Fig.~\ref{fig:Junktails} (left) plots the relative error
$|\dot{E}^{\infty}_{22}(t)
- \dot{E}^{\infty}_{22,\mathrm{FR}}|\big/
\dot{E}^{\infty}_{22,\mathrm{FR}}$
as a time-series, where $\dot{E}^{\infty}_{2 2,\mathrm{FR}}$ is
the frequency-domain value. For comparison we also compute
a ``luminosity" quantity\footnote{At finite radial values,
especially ones this small, $\dot{E}^{b}_{2 2}$ is certainly not
the energy radiated by the system. However, this value is computable
and, furthermore, is theoretically (although perhaps not
numerically) constant for circular orbits.
Therefore, our intention here is to quantify the effect of junk error tails
on its computation. Approximation of $\dot{E}^{\infty}_{2 2}$,
perhaps by extrapolation, might rely on such measurements.}
$\dot{E}^{b}_{2 2}(t)$ from $\dot{\Psi}_{2 2}^\mathrm{Z}(t,b)$.
The solid black line in Fig.~\ref{fig:Junktails} (left)
shows the relative error
$|\dot{E}^{b}_{2 2}(t)
- \dot{E}^{b}_{2 2}(6500M)|\big/
\dot{E}^{b}_{2 2}(6500M)$,
where $\dot{E}^{b}_{2 2}(6500M)$ is a late-time
computation less contaminated
by spurious junk. Comparing the black
and red lines, we see that the junk error tails
at $\mathscr{I}^+$ persist longer than those at the outer boundary $b$. This
observation suggests that spurious junk radiation is
a stubborn problem for high accuracy studies.
The right panel of Fig.~\ref{fig:Junktails}
indicates that the energy luminosity errors (due to spurious junk)
respectively decay as $t^{-8}$ and $t^{-5}$
for a fixed radial value and $\mathscr{I}^+$.
If we view $\dot{\Psi}_\mathrm{exact}$
as either $\dot{\Psi}_{22}(6500,b)$ or $\dot{\Psi}_{22}^\infty$, then
these rates are consistent with the expected decay rate
for {\em field} error tails $\Psi_\text{junk tail}$
and the relationship for a numerically computed
energy luminosity $\dot{E} \propto | \dot{\Psi}_\mathrm{exact}
+ \dot{\Psi}_\text{junk tail}|^2 \simeq
|\dot{\Psi}_\mathrm{exact}|^2 + 2\mathrm{Re}\big(\dot{\Psi}_\mathrm{exact}
\dot{\Psi}_\text{junk tail}\big)$.
\begin{table*}[h!]
{\scriptsize
\begin{tabular}{| cc || ll | ll | ll | ll |}
\hline
$m$ & Alg.
& \multicolumn{2}{c|}{$\dot{E}^\infty_{2 m}$}
& \multicolumn{2}{c|}{$\dot{E}^H_{2 m}$}
& \multicolumn{2}{c|}{$\dot{L}^\infty_{2 m}$}
& \multicolumn{2}{c|}{$\dot{L}^H_{2 m}$} \\
\hline
\hline
1 & FR
& 1.93160935116 & $\times 10^{-7}$
& 1.22691683145 & $\times 10^{-9}$
& 6.10828509933 & $\times 10^{-6}$
& 3.87985168700 & $\times 10^{-8}$\\
& AWE
& 1.93160935114 & $\times 10^{-7}$
& 1.22691683145 & $\times 10^{-9}$
& 6.10828509953 & $\times 10^{-6}$
& 3.87985168700 & $\times 10^{-8}$\\
\hline
2 & FR
& 5.36879547910 & $\times 10^{-5}$
& 1.13082774691 & $\times 10^{-8}$
& 1.69776220056 & $\times 10^{-3}$
& 3.57599132155 & $\times 10^{-7}$\\
& AWE
& 5.36879547910 & $\times 10^{-5}$
& 1.13082774691 & $\times 10^{-8}$
& 1.69776220057 & $\times 10^{-3}$
& 3.57599132154 & $\times 10^{-7}$ \\
\hline
\end{tabular}}
\caption{Mode-by-mode $\ell = 2$ luminosities for
the circular orbit described in the text.
For a particle of mass $m_p$ these values should be scaled by $m_p^2$.
The table compares our asymptotic-waveform evaluation
(AWE) results with frequency domain (FR) results computed by the code described
in Ref.~\cite{Hopper:2010uv}. We are grateful to S.~Hopper for
generating these previously unpublished FR luminosity values.
\label{highEFluxTable_Circ}
}
\end{table*}
\begin{figure}
\includegraphics[clip=true,width=0.48\linewidth]{JunkTails1.eps}
\includegraphics[clip=true,width=0.48\linewidth]{JunkTails2.eps}
\caption{{\sc Energy luminosity errors due to spurious junk radiation.}
Both panels show
$|\dot{E}^{\infty}_{22}(t) - \dot{E}^{\infty}_{2 2,\mathrm{FR}}|
/\dot{E}^{\infty}_{22,\mathrm{FR}}$ (red line, $r_{\infty}$) and
$|\dot{E}^{b}_{2 2}(t) - \dot{E}^{b}_{2 2}(6500M)|
/\dot{E}^{b}_{2 2}(6500M)$ (black line, $r_b$) as a time-series.
Comparing the black and red lines, we see that the energy
luminosity error (due to spurious junk) is particularly persistent
at $\mathscr{I}^+$. See the text for further explanation.
\label{fig:Junktails}}
\end{figure}
\section{Conclusions} \label{sec:conclude}
In the context of the Regge-Wheeler/Zerilli (describing blackhole
perturbations) and ordinary (describing acoustic phenomena) wave
equations we have developed a procedure
for obtaining the asymptotic far-field signal from a time-series
recorded at a finite radial value located beyond the spatial
compact support of the initial data. Furthermore, we have viewed
asymptotic waveform evaluation as a limiting
case of signal teleportation between two
finite radial values. For each of these wave equations our steps
are to (i) write down the exact relationship for teleportation in
the Laplace frequency domain, (ii) approximate this relationship
along the inversion contour (of the inverse Laplace transform) by
a sum of simple poles, and (iii) then represent, through inversion,
the asymptotic signal as a convolution [cf.~Eq.~\eqref{eq:XiEkernelConv}]
of the solution with time-domain kernel [cf.~Eq.~\eqref{eq:XiEkernel}]
comprised of damped exponentials. A similar recipe might be
used both to impose boundary conditions for and evaluate asymptotic
waveforms from perturbations of a Kerr blackhole. The Teukolsky
equation describing such perturbations is, like the cases treated
in this paper, separable in the frequency domain.
Through pre-computed numerical tables specifying each exponential's
strength and damping rate, we have demonstrated that accurate
asymptotic waveform evaluation through
teleportation can be easily implemented. Our simulations
based on these numerical tables correctly exhibit $t^{-4}$ as the
asymptotic decay rate for $\ell=2$ tails. We have also performed
generic-orbit, extreme-mass-ratio, binary simulations. From the solution
recorded as close as $30M$ and $60M$, we compute far-field
luminosities which agree with accurate frequency domain results to a
relative error of better than $10^{-9}$ ($10^{-12}$ for circular
orbits). Our studies indicate that spurious junk radiation is
particularly problematic for $\mathscr{I}^+$ computations, because far-field
luminosity errors (due to spurious junk) decay at a slow rate.
These results have been achieved without a compactification
scheme to include $\mathscr{I}^+$ in the computational domain. Instead,
they have relied on a Laplace convolution \eqref{eq:XiEkernelConv}
which, decoupled from a numerical evolution, can also be carried
out as a post-processing step on an existing time series.
Finally, we have demonstrated effective signal teleportation between
two finite radial locations, for $\ell = 2,3,64$ and with relative
errors $\simeq 10^{-10}$. These demonstrations are a powerful test of
our AWE/teleportation method's
accuracy as well as a practical sanity check of its implementation.
As discussed in Sec.~\ref{sec:EffBH}, our Laplace-convolution RBC
and AWE methods would seem efficient from work and storage
standpoints. Lower $\ell$ kernels appear similar to the case of ordinary
wave propagation on $2+1$ flat spacetime. In that setting the low-$n$
(Fourier index) ``circle kernels" are also expensive to evaluate to
account for tail-like phenomena. In any case, for the $\ell =
2,3,64$ cases considered in this paper a kernel's overall computational
cost is roughly equivalent to to adding 20 to 30 points to
the spatial domain. Moreover, it is non-intrusive, requiring no grid
stretching or supplemental coordinate transformations, and
may be carried out at any radial value beyond the spatial compact support
of the initial data and sources. Finally, in their
frequency-domain form, our kernels might be used to
implement radiation boundary conditions and AWE
in {\em frequency-domain codes.} Here we envision that
the kernels would first undergo a ``Wick rotation" prior
to use.
While the results of this paper are encouraging, we believe that
further careful study is merited. First, the task of computing
teleportation/AWE tables is daunting for a number of reasons.
The main one is cost. Here we refer to the {\em offline} cost in
generating a table, not the cost incurred by the user implementing
AWE with such a table. As discussed in Sec.~\ref{subsec:BHWE},
generation of AWE kernels costs upwards of $10^{13}$
floating point operations. Moreover, since the cost is offline,
with the resulting numerical table then ``good for all time", we believe
the process should be carried out in quadruple precision in order to achieve
$\varepsilon = 10^{-15}$ error tolerances.\footnote{Indeed, once such
accurate tables have been constructed, smaller tables (say corresponding
to $\varepsilon = 10^{-8}$) can be achieved by compressing the accurate
table, i.e.~using the accurate table for fast evaluation of profiles that
are again subject to AGH compression.} This further adds to the cost.
Due to the difficulties associated with the computation of AWE tables,
we have not adequately isolated all sources of error in their construction.
A systematic and optimized procedure for computing kernels would greatly
reduce the offline costs. One possibility is an application specific
quadrature rule. So far, we have employed the familiar Gauss-Kronrod rule,
which is designed for high-order integration of polynomials. We might instead
design a quadrature rule which is exact for the corresponding
flatspace kernels~\cite{Antil:ROQ}.
We plan to construct a family of RBC and AWE
tables for general use:
Regge-Wheeler and Zerilli tables likely for $2 \leq \ell \leq 64$,
boundary radii $\rho_b = 15,30,60,120$,
and an evaluation radius $\rho_\infty = 1\times 10^{15}$ (or
$\rho_\infty = \infty$ with a semi-infinite quadrature rule,
see Sec.~\ref{subsec:BHWE}). All kernels used in this paper, as well as these
others, will be available at \cite{Kernel_web1}.
\section{Acknowledgements}
This work has been substantially supported by NSF grant No.~PHY
0855678 to the University of New Mexico; AGB and SRL gratefully
acknowledge this support. AGB's contributions to this work
were also supported by NSF Grant No.~0739417 as an Undergraduate
Research Project supervised by SRL. SEF acknowledges support from
the Joint Space Science Institute,
NSF Grants No.~PHY 1208861 and No.~PHY 1005632 to the University of
Maryland, and NSF grant No.~PHY05-51164 to the University of
California at Santa Barbara. SRL is also grateful for support from
the Erwin Schr\"{o}dinger Institute for Mathematical Physics,
Vienna, during the workshop {\em Dynamics of General Relativity
Analytical and Numerical Approaches}, 4 July to 2 September, 2011.
SEF thanks the Kavli Institute for Theoretical Physics, University
of California at Santa Barbara, where this work was completed,
for its hospitality. We are grateful for comments
from M.~Blair, P.~Ca\~nizares, M.~Cao, C.~Evans, T.~Hagstrom, S.~Hopper,
L.~Lehner, M.~Tiglio, B.~Whiting, and A.~Zengino\u{g}lu. We further
thank A.~Zengino\u{g}lu for comments on the manuscript, and SRL
particularly thanks C.~Evans for discussions early on in the
development of RBCs which also influenced the ideas presented here.
|
\section{Introduction}
\label{sec:1}
From the outset, inspired by gas physics, plasma physicists derived kinetic equations to describe microscopic aspects of their physics, in particular the Vlasov equation. This trend has been the dominant one till nowadays. However, for plasmas where transport due to short range interactions (``collisions") is weak, it is possible to work directly with $N$-body classical mechanics. As will be recalled in section \ref{WPDRT}, this approach led to a new description of wave-particle interaction making it more intuitive and unifying particle and wave evolutions, as well as collective and finite-$N$ physics.
This paper develops this approach further by deriving a rigorous fundamental equation for the electrostatic potential (section \ref{FEP}), which is extended to enable a correct description of trapping or chaos due to Langmuir waves (section \ref{WPDRT}). It brings further unifications: Debye shielding (or screening) and Landau theory of linear waves (section \ref{DSLD}), Landau growth or damping and spontaneous emission (section \ref{WPDRT}), Debye shielding and collisional transport (section \ref{DSCT}). These results come with a considerable simplification of the mathematical framework with respect to textbooks (see for instance Refs. \cite{HW,Nic,Chen,BS}), and with new insights into microscopic plasma physics. In particular it solves the following issue: How can each particle be shielded by all other ones, while all the plasma particles are in uninterrupted motion? This turns out to be a mere consequence of the almost independent deflections of particles due to the Coulomb force. Furthermore, the new path to Landau damping goes first through Debye shielding, a totally unexpected fact, since classical textbooks present these concepts in different and unrelated chapters.
As to collisional transport, this paper does not simply recover the classical result presented in textbooks, but it provides the first calculation of this transport without any ad hoc cut-off. This is made possible by a mechanical calculation of the deflection of a particle which covers the scales corresponding to the Debye length, to the inter-particle distance, and to the distance of minimum approach of two electrons in a Rutherford collision. Till now none of the papers about collisional transport, and no textbook, was able to correctly describe the scale of the inter-particle distance.
The derivations are elementary (we refrain from invoking the Liouville equation or the BBGKY hierarchy). After section \ref{FEP}, the successive approximations of the fundamental equation (\ref{phihat}) enable to get rid of the heavy notations related to its generality. These approximations involve linearization in sections \ref{FEP}, \ref{SCP}, and \ref{STSP}, but section \ref{MIIDS}, the beginning of section \ref{WPDRT}, and section \ref{DSCT} do not appeal to linearization. For the sake of brevity, this paper is more oriented toward concepts and intuitive physics, than toward ready-to-use formulas to be presented elsewhere.
\section{Claims}
\label{Cl}
Here are the main results of this paper:
1. By using the Fourier and Laplace transforms, a rigorous equation (Eq.~(\ref{phihat})) is derived for the electrostatic potential of an infinite plasma made up of the periodic replication of $N$ electrons coupled by Coulomb forces in a volume $L^3$ with a neutralizing ionic background (One Component Plasma (OCP) model \cite{Salp,Abe,BH}). This equation is of the type ${\mathcal{E}}\hat \varphi=$ source term, where ${\mathcal{E}}$ is a linear operator, acting on the infinite dimensional array $\hat \varphi$ whose components are all the Fourier-Laplace components of the potential.
2. Considering that particle trajectories stay close to ballistic, and replacing inside ${\mathcal{E}}$ the discrete sums over particles by integrals over a smooth distribution function, the potential turns out to be the sum of the shielded Coulomb potentials of the individual particles (Eq.~(\ref{phi})). Such potentials were first computed by a kinetic approach in section II.A of Ref.\ \cite{Gasio} and later on in \cite{Bal,Rost}.
3. Replacing the discrete sums over particles of their shielded potentials by integrals over a smooth distribution function, yields the classical Vlasovian expression (Eqs~(\ref{phihatL}) and (\ref{phi0hatcg})) including initial conditions in Landau contour calculations of Langmuir wave growth or damping, obtained by linearizing Vlasov equation and using Fourier-Laplace transform, as described in many textbooks (see for instance Refs. \cite{HW,Nic,BS}). Therefore in these calculations, the electrostatic potential turns out to be the smoothed version of the actual shielded potential in the plasma.
4. In the spirit of Refs. \cite{OWM,OLMSS,AEE,EEB}, in order to accommodate a correct description of trapping or chaos due to Langmuir waves, the set of particles is split into bulk and tail, where the bulk is the set of particles which cannot resonate with Langmuir waves. Repeating the analysis leading to Eq.~(\ref{phihat}) for the bulk particles, the same equation is recovered with an additional source term due to the tail particles (Eq.~(\ref{phihatU})).
5. Using the fact that the number of tail particles is small with respect to the bulk one, and a technique introduced in Refs. \cite{OWM,OLMSS}, an amplitude equation is derived for any Fourier component of the potential where tail particles provide a source term (Eq.~(\ref{eqampl})).
6. This equation, together with the equation of motion of the tail particles, enables to show that, in the linear regime, the amplitude of a Langmuir wave is ruled by Landau growth or damping, and by spontaneous emission (Eq.~(\ref{evampfinal})), a generalization to 3 dimensions of a one-dimensional result of Refs. \cite{EZE,EEB}.
7. Using the shielded potential, the collisional diffusion coefficient $D$ of a given particle is computed by a convergent expression including the particle deflections for all impact parameters. These deflections are computed by first order perturbation theory in the total electric field, except for those due to close encounters. The contribution to $D$ of the former ones is matched with that of the latter ones provided by Ref.\ \cite{Ros}. The detailed matching procedure includes the scale of the inter-particle distance, and is reminiscent of that in \cite{Hub}, without invoking the cancellation of three infinite integrals. $D$ has the same expression as that in Ref.\ \cite{Ros}, except for the Coulomb logarithm which is modified by a velocity dependent quantity of order 1.
8. Picard iteration technique is applied to the equation of motion of particle $P$ due to the Coulomb forces of all other ones. It reveals that a part of the effect on particle $P$ of another particle $P'$ is mediated by all other particles (Eq.~(\ref{rsecnAccDev2})). Indeed particle $P'$ modifies the position of all other particles, implying that the action of the latter ones on particle $P$ is modified by particle $P'$.
9. This calculation yields the following interpretation of shielding.
At $t =0$ consider a set of randomly distributed particles, and especially particle $P$. At a later time $t$, the latter has deflected all particles which made a closest approach to it with an impact parameter $b \lesssim v_{\rm{th}} t$ where $v_{\rm{\rm{th}}}$ is the thermal velocity. This part of their global deflection due to particle $P$ reduces the number of particles inside the sphere $S(t)$ of radius $v_{\rm{th}} t$ about it. Therefore the effective charge of particle $P$ as seen out of $S(t)$ is reduced: the charge of particle $P$ is shielded due to these deflections. This shielding effect increases with $t$, and thus with the distance to particle $P$. As a result, the typical time-scale for shielding to set in when starting from random particle positions is the time for a thermal particle to cross a Debye sphere, i.e.\ $\omega_{\rmpp}^{-1}$, where $\omega_{\rmpp}$ is the plasma frequency. Furthermore, shielding is a cooperative dynamical process: it results from the accumulation of almost independent repulsive deflections with the same qualitative impact on the effective electric field of particle $P$ (if point-like ions were present, the attractive deflection of charges with opposite signs would have the same effect). It is a cooperative effect, but not a collective one. So, shielding and collisional transport are two aspects of the same two-body repulsive process.
\section{Fundamental equation for the potential}
\label{FEP}
This paper deals with the One Component Plasma model \cite{Salp,Abe,BH}, which considers the plasma as infinite with spatial periodicity $L$ in three orthogonal directions with coordinates $(x,y,z)$, and as made up of $N$ electrons in each elementary cube with volume $L^3$. Ions are present only as a uniform neutralizing background, which enables periodic boundary conditions. This choice is made to simplify the analysis which focuses on $\varphi(\textbf{r})$, the potential created by the $N$ particles at any point where there is no particle. The discrete Fourier transform of $\varphi(\textbf{r})$, readily obtained from the Poisson equation, is given by $\tilde{\varphi}(\textbf{0}) = 0$ and, for $\textbf{m} \neq \textbf{0}$, by
\begin{equation}
\tilde{\varphi}(\textbf{m})
= - \frac{e}{\varepsilon_0 k_{\textbf{m}}^2} \sum_{j \in S}
\exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_j),
\label{phitildetotM}
\end{equation}
where $-e$ is the electron charge, $\varepsilon_0$ is the vacuum permittivity, $\textbf{r}_j$ is the position of particle $j$, $S$ is the set of integers going from 1 to $N$, $\tilde{\varphi}(\textbf{m}) = \int \varphi(\textbf{r}) \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}) \rmd^3 \textbf{r}$, $\textbf{m} = (m_x,m_y,m_z)$ is a vector with three integer components, $\textbf{k}_{\textbf{m}} = \frac{2 \pi}{L}\textbf{m}$, and $k_{\textbf{m}} = \|\textbf{k}_{\textbf{m}}\|$. Reciprocally,
\begin{equation}
\varphi(\textbf{r}) = \frac{1}{L^3}\sum_{\textbf{m}} \tilde{\varphi} (\textbf{m}) \exp(\rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}),
\label{phiInv}
\end{equation}
where $\sum_{\textbf{m}}$ means the sum over all components of $\textbf{m}$ running from $- \infty$ to $+ \infty$.
The dynamics of particle $l$ is determined by Newton's equation
\begin{equation}
\ddot{\textbf{r}}_l = \frac{e}{m_e} \nabla \varphi_l(\textbf{r}_l),
\label{rsectot}
\end{equation}
where $m_e$ is the electron mass, $\varphi_l$ is the electrostatic potential acting on particle $l$, i.e.\ the one created by all other particles and by
the background charge. Its Fourier transform is given by Eq.~(\ref{phitildetotM}) with the supplementary condition $j \neq l$.
Let
\begin{equation}
\textbf{r}_l^{(0)} = \textbf{r}_{l0} + \textbf{v}_{l} t
\label{rl0}
\end{equation}
be a ballistic approximation of the motion of particle $l$. In the following, we consider two instances of this approximation: the one where $\textbf{r}_{l0}$ and $\textbf{v}_{l}$ are respectively the initial position and velocity of particle $l$, and the one where they are slightly shifted from these values by low amplitude Langmuir waves. Let $\delta \textbf{r}_l = \textbf{r}_l - \textbf{r}_l^{(0)}$, so that Eq.~(\ref{rsectot}) is equivalent to
\begin{equation}
\delta \ddot{\textbf{r}}_l = \frac{\rmi e}{L^3 m_e} \sum_{\textbf{n}} \textbf{k}_{\textbf{n}} \ \tilde{\varphi}_l(\textbf{n}) \exp[\rmi \textbf{k}_{\textbf{n}} \cdot (\textbf{r}_l^{(0)} + \delta \textbf{r}_l)].
\label{delrsec}
\end{equation}
Sections \ref{DSLD} and \ref{WPDRT} consider cases where the $\delta \textbf{r}_l$'s are small in some sense. To this end, we split $\tilde{\varphi}_l(\textbf{m})$ as
\begin{equation}
\tilde{\varphi}_l(\textbf{m}) = \tilde{\phi}_l(\textbf{m}) + \Delta \tilde{\varphi}_l(\textbf{m})
\label{phitildnapp}
\end{equation}
where
\begin{equation}
\tilde{\phi}_l (\textbf{m}) = \sum_{j \in S;j \neq l} \delta \tilde{\phi}_{j} (\textbf{m}),
\label{phitildn}
\end{equation}
with
\begin{equation}
\delta \tilde{\phi}_{j} (\textbf{m}) = -\frac{e}{\varepsilon_0 k_{\textbf{m}}^2} \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j}^{(0)})(1 - \rmi \textbf{k}_{\textbf{m}} \cdot \delta \textbf{r}_j),
\label{phitildnj}
\end{equation}
and
\begin{equation}
\Delta \tilde{\varphi}_l(\textbf{m}) = -\frac{e}{\varepsilon_0 k_{\textbf{m}}^2} \sum_{j \in S;j \neq l}
\exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j}^{(0)}) R_j(\textbf{m}),
\label{Deltaphi}
\end{equation}
with
\begin{equation}
R_j(\textbf{m}) = \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \delta \textbf{r}_j) -1 + \rmi \textbf{k}_{\textbf{m}} \cdot \delta \textbf{r}_j,
\label{Rj}
\end{equation}
which is of order two in $\delta \textbf{r}_j$.
We now introduce the time Laplace-transform which transforms a function $ f(t)$ into $\hat{f}(\omega) = \int_0^{\infty} f(t) \exp(\rmi \omega t) \rmd t$ (with $\omega$ complex). The Laplace transform of Eq.~(\ref{delrsec}) is
\begin{equation}
\omega^2 \delta \hat{\textbf{r}}_l(\omega) = - \frac{\rmi e}{L^3 m_e} \sum_{\textbf{n}} \textbf{k}_{\textbf{n}} \exp(\rmi \textbf{k}_{\textbf{n}} \cdot \textbf{r}_{l0}) \ \Psi_l(\hat{\varphi}_l \ ; \textbf{n},\omega + \omega_{\textbf{n},l}) + \rmi \omega \delta \textbf{r}_l(0) - \delta \dot{\textbf{r}}_l(0).
\label{rLapl}
\end{equation}
where carets indicate the Laplace transformed versions of the quantities in Eq.~(\ref{delrsec}), $\omega_{\textbf{n},l} = \textbf{k}_{\textbf{n}} \cdot \textbf{v}_{l}$ comes from the time dependence of $\textbf{r}_l^{(0)}$ in the exponent of Eq.~(\ref{delrsec}), and the operator $\Psi_l$ acting on a function $g(\textbf{m},\omega)$ is defined by
\begin{equation}
\Psi_l(g\ ; \textbf{n},\cdot) = g(\textbf{n},\cdot) \ast T_{l}(\textbf{n},\cdot),
\label{psil}
\end{equation}
where $\cdot$ stands for the frequencies, $ \ast $ is the convolution product, and $T_{l}(\textbf{n},\omega)$ is the Laplace transform of $\exp(\rmi \textbf{k}_{\textbf{n}} \cdot \delta \textbf{r}_l)$. The Laplace transform of Eqs~(\ref{phitildnapp})-(\ref{Rj}) yields
\begin{equation}
k_{\textbf{m}}^2\hat{\varphi}_l(\textbf{m},\omega) = k_{\textbf{m}}^2 \hat{\phi}_l^{(00)}(\textbf{m},\omega) + \frac{\rmi e}{\varepsilon_0} \sum_{j \in S;j \neq l} \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j0}) \ [\textbf{k}_{\textbf{m}} \cdot \delta \hat{\textbf{r}}_j(\omega - \omega_{\textbf{m},j}) + \rmi \hat{R}_j(\textbf{m},\omega - \omega_{\textbf{m},j})],
\label{phihatn}
\end{equation}
where $\hat{R}_j(\textbf{m},\omega)$ is the Laplace transform of $R_j$, and
$\hat{\phi}_l^{(00)}(\textbf{m},\omega) $ is the Laplace transform of $\tilde{\phi}_l (\textbf{m}) $ computed from Eqs~(\ref{phitildn}) and (\ref{phitildnj}) by setting $\delta \textbf{r}_j =0$ for all $j$'s in the latter. Substituting the $\delta \hat{\textbf{r}}_j$'s with their expression Eq.~(\ref{rLapl}) yields
\begin{eqnarray}
&&
k_{\textbf{m}}^2 \hat{\varphi}_l(\textbf{m},\omega)
- \frac{e^2}{ L^3 m_e \varepsilon_0}
\sum_{\textbf{n}} \textbf{k}_{\textbf{m}} \cdot \textbf{k}_{\textbf{n}}
\ \sum_{j \in S;j \neq l}
\frac{\Psi_j(\hat{\varphi}_j\ ; \textbf{n},\omega + \omega_{\textbf{n},j} - \omega_{\textbf{m},j})}
{(\omega - \omega_{\textbf{m},j})^2}
\exp[\rmi (\textbf{k}_{\textbf{n}}-\textbf{k}_{\textbf{m}})
\cdot \textbf{r}_{j0}]
\nonumber\\
& = &
k_{\textbf{m}}^2 \hat{\phi}_l^{(0)}(\textbf{m},\omega)
- \frac{e}{\varepsilon_0} \sum_{j \in S;j \neq l}
\exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j0})
\hat{R}_j(\omega - \omega_{\textbf{m},j}),
\label{phihatnf}
\end{eqnarray}
where $\hat{\phi}_l^{(0)}(\textbf{m},\omega) $ is the Laplace transform of $\tilde{\phi}_l (\textbf{m}) $ computed from Eqs~(\ref{phitildn}) and (\ref{phitildnj}) by setting now $\delta \textbf{r}_j = \delta \textbf{r}_j(0) + \delta \dot{\textbf{r}}_j(0) t$ for all $j$'s in the latter.
Summing Eq.~(\ref{phihatnf}) over $l = 1,... N$ and dividing by $N-1$, yields
\begin{eqnarray}
k_{\textbf{m}}^2\hat{\varphi}(\textbf{m},\omega)
&-& \frac{e^2}{ L^3 m_e \varepsilon_0}
\sum_{\textbf{n}} \textbf{k}_{\textbf{m}} \cdot \textbf{k}_{\textbf{n}}
\ \sum_{j \in S} \frac{\Psi_j(\hat{\varphi}\ ; \textbf{n},\omega + \omega_{\textbf{n},j} - \omega_{\textbf{m},j})}{(\omega - \omega_{\textbf{m},j})^2} \exp[\rmi (\textbf{k}_{\textbf{n}}-\textbf{k}_{\textbf{m}}) \cdot \textbf{r}_{j0}]
\nonumber\\
= k_{\textbf{m}}^2 \hat{\phi}^{(0)}(\textbf{m},\omega) &-& \frac{e}{\varepsilon_0} \sum_{j \in S} \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j0}) \hat{R}_j(\omega - \omega_{\textbf{m},j}),
\label{phihat}
\end{eqnarray}
where $\hat{\phi}^{(0)}(\textbf{m},\omega) $ is $\hat{\phi}_l^{(0)}(\textbf{m},\omega) $ complemented by the missing $l$-th term.
Equation (\ref{phihat}) is the fundamental equation of this paper, and is a rigorous consequence of Eqs~(\ref{phitildetotM}) and (\ref{rsectot}): no approximation was made. This fundamental equation is of the type ${\mathcal{E}}\hat \varphi=$ source term, where ${\mathcal{E}}$ is a linear operator, acting on the infinite dimensional array whose components are all the $\hat{\varphi}(\textbf{m},\omega)$'s.
\section{Debye shielding}
\label{DSLD}
\subsection{Shielded Coulomb potential}
\label{SCP}
We now specialize Eq.~(\ref{phihat}) by considering the lowest order contribution of the $\delta \textbf{r}_j$'s, which makes the $\hat{R}_j$'s vanish (Approximation 1). We further consider $\varphi$ to be small, and the $\delta \textbf{r}_l$'s to be of the order of $\varphi$ (Approximation 2). These two approximations reduce $\Psi_j(\hat \varphi ; \textbf{n},\omega)$ to $\hat{\varphi}(\textbf{n},\omega)$. This corresponds to substituting the true dynamics in Eq.~(\ref{rsectot}) with an approximate one ruled by
\begin{equation}
\delta \ddot{\textbf{r}}_l = \frac{e}{m_e} \nabla \phi_l(\textbf{r}_l^{(0)} + \delta \textbf{r}_l),
\label{rsec}
\end{equation}
where $\phi_l (\textbf{r}) = \sum_{j \in S;j \neq l} \delta \phi_{j} (\textbf{r})$ is the inverse Fourier transform of Eq.~(\ref{phitildn}), so that
\begin{equation}
\lim_{L \to \infty} \delta \phi_{j} (\textbf{r}) = - \frac{e}{4 \pi \varepsilon_0 \| \textbf{r} - \textbf{r}_j^{(0)}\|} - \frac{e \delta \textbf{r}_j \cdot (\textbf{r} - \textbf{r}_j^{(0)})}{4 \pi \varepsilon_0 \| \textbf{r} - \textbf{r}_j^{(0)}\|^3},
\label{deltaphi}
\end{equation}
The $j$th component of the approximate electric field acting over particle $l$ turns out to be due to a particle located at $\textbf{r}_{j}^{(0)}$ instead of $\textbf{r}_j$, and is made up of a Coulombian part and of a dipolar part with a dipole moment $- e \delta \textbf{r}_j$. The cross-over between the two contributions occurs for $\| \textbf{r}_l - \textbf{r}_j^{(0)}\|$ on the order of $\| \delta \textbf{r}_j \|$, i.e.\ when the distance between particle $l$ and the ballistic particle $j$ is equal to the distance between the latter and the true particle $j$. For larger values of $\| \textbf{r}_l - \textbf{r}_j^{(0)}\|$, the dipolar component is subdominant. For smaller ones, it is dominant, but with a direction which is a priori random with respect to the Coulombian one ($(\textbf{r}_l - \textbf{r}_j^{(0)})$ is almost independent from $\delta \textbf{r}_j$). Since the $\| \delta \textbf{r}_j \|$'s are assumed small, the latter case should be rare since it corresponds to a very close encounter between particle $l$ and the ballistic particle $j$. As a result the approximate electric field stays dominantly of Coulombian nature, but with a small mismatch of the charge positions with respect to the actual ones.
We introduce a smooth function $f(\textbf{r},\textbf{v})$,
the smoothed velocity distribution function at $t=0$ such that
the distribution
\begin{equation}
\sum_{l \in S} \bullet
=
\iint \bullet f(\textbf{r},\textbf{v}) \rmd^3 \textbf{r} \, \rmd^3 \textbf{v} + W(\bullet),
\label{fxv}
\end{equation}
where the distribution $W$ yields a negligible contribution when applied to space dependent function which evolve slowly on the scale of the inter-particle distance; there the spatial integration is performed over the elementary cube with volume $L^3$, and the velocity integration is over all velocities.
Replacing the discrete sums over particles by integrals over the smooth distribution function $f(\textbf{r},\textbf{v})$ (Approximation 3), Eq.~(\ref{phihat}) becomes
\begin{equation}
k_{\textbf{m}}^2 \hat{\Phi}(\textbf{m},\omega)
= k_{\textbf{m}}^2 \hat{\phi}^{(0)}(\textbf{m},\omega)
+ \frac{e^2}{L^3 m_e \varepsilon_0} \sum_{\textbf{n}} \textbf{k}_{\textbf{m}} \cdot \textbf{k}_{\textbf{n}}
\int \frac{\hat{\Phi}(\textbf{n},\omega + (\textbf{k}_{\textbf{n}} - \textbf{k}_{\textbf{m}}) \cdot \textbf{v})}{(\omega -\textbf{k}_{\textbf{m}} \cdot \textbf{v})^2} \tilde{f}({\textbf{n}} - {\textbf{m}},\textbf{v}) \ \rmd^3 \textbf{v},
\label{phihatcg}
\end{equation}
where $\hat{\Phi}$ is the smoothed version of $\hat{\varphi}$ resulting from Approximations 1 to 3, and $\tilde{f}$ is the spatial Fourier transform of $f$. We further assume the initial distribution $f$ to be a spatially uniform distribution function $f_0(\textbf{v})$ plus a small perturbation of the order of $\Phi$ (in agreement with Approximation 2). Then operator $\mathcal{E}$ becomes diagonal with respect to both $\textbf{m}$ and $\omega$ (a complex quantity), and linearizing Eq.~(\ref{phihatcg}) for $\hat{\Phi}$ yields
\begin{equation}
\epsilon(\textbf{m},\omega) \hat{\Phi}(\textbf{m},\omega) = \hat{\phi}^{(0)}(\textbf{m},\omega),
\label{phihatL}
\end{equation}
where
\begin{equation}
\epsilon(\textbf{m},\omega) = 1 - \frac{e^2}{L^3 m_e \varepsilon_0}
\int \frac{f_0(\textbf{v}) }{(\omega -\textbf{k}_{\textbf{m}} \cdot \textbf{v})^2} \ \rmd^3 \textbf{v}.
\label{eps}
\end{equation}
This shows that the smoothed self-consistent potential $\hat{\Phi}$ is determined by the response function $\epsilon(\textbf{m},\omega)$. It is the classical plasma dielectric function. A first check of this can be obtained for a cold plasma: then
$\epsilon(\textbf{m},\omega) = 1 - {\omega_{\rmpp}^2}/{\omega^2}$,
where $\omega_{\rmpp}^2 = [(e^2 n)/(m_e \epsilon_0)]^{1/2}$ is the plasma frequency ($n = N/L^3$ is the plasma density). The classical expression involving the gradient of $f_0$ in $\textbf{v}$ is obtained by a mere integration by parts.
The contribution of particle $j$ to $\tilde{\phi}^{(0)}(\textbf{m})$ is $\delta \tilde{\phi}_j^{(0)}(\textbf{m}) = - \frac{e}{\varepsilon_0 k_{\textbf{m}}^2} \exp[- \rmi \textbf{k}_{\textbf{m}} \cdot (\textbf{r}_{j0} + \textbf{v}_j t)]$. Its Laplace transform is
\begin{equation}
\delta \hat{\phi}_j^{(0)}(\textbf{m},\omega) = - \frac{\rmi e}{\varepsilon_0 k_{\textbf{m}}^2}
\frac{\exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j0})}{\omega -\textbf{k}_{\textbf{m}} \cdot \textbf{v}_j}.
\label{phij0hat}
\end{equation}
The corresponding part of $\hat{\Phi}(\textbf{m},\omega)$ is $\delta \hat{\Phi}_j(\textbf{m},\omega) = \delta \hat{\phi}_j^{(0)}(\textbf{m},\omega)/\epsilon(\textbf{m},\omega)$. This turns out to be the shielded potential of particle $j$ \cite{Gasio,Bal,Rost}.
By inverse Fourier-Laplace transform, after some transient whose duration is estimated at the end of section \ref{MIIDS}, the potential due to particle $j$ becomes the shielded Coulomb potential
\begin{equation}
\delta \Phi_j (\textbf{r}) = \delta \Phi(\textbf{r} - \textbf{r}_{j0} - \textbf{v}_j t,\textbf{v}_j),
\label{phij}
\end{equation}
where
\begin{equation}
\delta \Phi (\textbf{r},\textbf{v}) = - \frac{e}{L^3 \varepsilon_0} \sum_{\textbf{m}} \frac{\exp(\rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r})}{ k_{\textbf{m}}^2 \epsilon(\textbf{m},\textbf{k}_{\textbf{m}} \cdot \textbf{v})}.
\label{phi}
\end{equation}
Therefore, after this transient, the dominant contribution to the full potential in the plasma turns out to be the sum of the shielded Coulomb potentials of individual particles located at their ballistic positions. Let $\lambda_{\rm{D}} = [(\epsilon_0 k_{\rm{B}} T)/(n e^2)]^{1/2}
= [(k_{\rm{B}} T)/m_e]^{1/2} \omega_{\rmpp}^{-1}$ be the Debye length, where $k_{\rm{B}}$ is the Boltzmann constant, $T$ is the temperature, and $n$ is the density. The wavenumbers resolving scale $\| \textbf{r}\|$ are such that $k_{\textbf{m}} \|\textbf{r}\| \gtrsim 1$. If $\|\textbf{r}\| \ll \lambda_{\rm{D}}$, the corresponding wavenumbers are such that $k_{\textbf{m}} \lambda_{\rm{D}} \gg 1$. Therefore there is no shielding for $\|\textbf{r}\| \ll \lambda_{\rm{D}}$, since $\epsilon(\textbf{m},\textbf{k}_{\textbf{m}} \cdot \textbf{v}) - 1 \simeq (k_{\textbf{m}} \lambda_{\rm{D}})^{-2}$.
In the following, Eq.~(\ref{phij}) is used by substituting $\delta \Phi(\textbf{r} - \textbf{r}_{j0} - \textbf{v}_j t,\textbf{v}_j)$ with $\delta \Phi(\textbf{r} - \textbf{r}_{j},\textbf{v}_j)$: the shielded potential of particle $j$ is computed by taking into account its actual position, since it is the original Coulomb one close to $\textbf{r}_{j}$. The error made for $\textbf{r} - \textbf{r}_{j} $ of the order of $\lambda_{\rm{D}}$ is small as long as the mismatch of $\textbf{r}_{j} $ from the ballistic orbit is much smaller than $\lambda_{\rm{D}}$. As was done for the bare potential of Eq.~(\ref{phitildetotM}), the field acting on a given particle $l$ is obtained by removing its own divergent contribution $\delta \Phi_l$ from $\Phi$.
\subsection{Mediated interactions imply Debye shielding}
\label{MIIDS}
In the above introduction of Debye shielding, using the Laplace transform of the particle positions does not provide an intuitive picture of this effect. We now show that such a picture can be obtained by turning back to the mechanical description of microscopic dynamics with the full Coulomb potential of Eq.~(\ref{phitildetotM}). In order to compute the dynamics, we use Picard iteration technique. From Eq.~(\ref{rsectot}), $\textbf{r}_l^{(n)}$, the $n$th iterate of $\textbf{r}_l$, is computed by
\begin{equation}
\ddot{\textbf{r}}^{(n)}_l
= \frac{e}{m_e} \nabla \varphi_l^{(n-1)}(\textbf{r}^{(n-1)}_l),
\label{rsecn}
\end{equation}
where $\varphi_l^{(n-1)}$ is computed by the inverse Fourier transform of Eq.~(\ref{phitildetotM}) with the $\textbf{r}_j$'s substituted with the $\textbf{r}_j^{(n-1)}$'s. The iteration starts with the ballistic approximation of the dynamics defined by Eq.~(\ref{rl0}), and the actual orbit of Eq.~(\ref{rsectot}) corresponds to $n \rightarrow \infty$. Let $\delta \textbf{r}_l^{(n)} = \textbf{r}_l^{(n)} - \textbf{r}_l^{(0)}$ be the mismatch of the position of particle $l$ with respect to the ballistic one at the $n$th iterate. It is convenient to write Eq.~(\ref{rsecn}) as
\begin{equation}
\delta \ddot{\textbf{r}}^{(n)}_l = \sum_{j \in S;j \neq l} \delta \ddot{\textbf{r}}^{(n)}_{lj},
\label{rsecnAccl}
\end{equation}
with
\begin{equation}
\delta \ddot{\textbf{r}}^{(n)}_{lj} = \textbf{a}_{\rm{C}}(\textbf{r}_l^{(n-1)}-\textbf{r}_j^{(n-1)}),
\label{rsecnAcclj}
\end{equation}
and
\begin{equation}
\textbf{a}_{\rm{C}}(\textbf{r}) = - \frac{e}{m_e} \nabla \delta \varphi_{\rm{C}} (\textbf{r}),
\label{AccC}
\end{equation}
with $\delta \varphi_{\rm{C}}$, the $L$-periodized, OCP bare Coulomb potential, given by
\begin{equation}
\delta \varphi_{\rm{C}}(\textbf{r}) =
-\frac{e}{\varepsilon_0 L^3}
\sum_{{\textbf{m}}\neq {\textbf{0}}} k_{\textbf{m}}^{-2} \exp(\rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}),
\label{phiCbPer}
\end{equation}
such that $\varphi(\textbf{r}) = \sum_{j \in S} \delta \varphi_{\rm{C}}(\textbf{r} - \textbf{r}_j)$.
Then $\delta \ddot{\textbf{r}}^{(1)}_l = \sum_{j \in S;j \neq l} \textbf{a}_{\rm{C}}(\textbf{r}_l^{(0)}-\textbf{r}_j^{(0)})$ and for $n \geq 2$
\begin{equation}
\delta \ddot{\textbf{r}}^{(n)}_l = \delta \ddot{\textbf{r}}^{(1)}_l + \sum_{j \in S;j \neq l} \nabla \textbf{a}_{\rm{C}}(\textbf{r}_l^{(0)}-\textbf{r}_j^{(0)}) \cdot (\delta \textbf{r}_l^{(n-1)} - \delta \textbf{r}_j^{(n-1)}) + O(a^3),
\label{rsecnAccDev}
\end{equation}
where $a$ is the order of magnitude of the total Coulombian acceleration. Equation (\ref{rsecnAccDev}) may be written
\begin{equation}
\ddot{\textbf{r}}^{(n)}_l = \sum_{j \in S;j \neq l}
[(\delta \ddot{\textbf{r}}^{(1)}_{lj} + M_{lj}^{(n-1)})
+ 2 \nabla \textbf{a}_{\rm{C}}(\textbf{r}_l^{(0)}-\textbf{r}_j^{(0)}) \cdot \delta \textbf{r}_{lj}^{(n-1)}] + O(a^3),
\label{rsecnAccDev2}
\end{equation}
where $M_{lj}^{(n-1)} = \nabla \textbf{a}_{\rm{C}}(\textbf{r}_l^{(0)}-\textbf{r}_j^{(0)}) \cdot \sum_{i \in S; i \neq l,j} (\delta \textbf{r}_{li}^{(n-1)} - \delta \textbf{r}_{ji}^{(n-1)})$
is the modification of the bare Coulomb acceleration of particle $j$ on particle $l$ due to the following phenomenon: particle $j$ modifies the position of all other particles, which implies the action of the latter ones on particle $l$ is modified by particle $j$. Therefore $M_{lj}^{(n-1)} $ is the acceleration of particle $l$ due to particle $j$ mediated by all other particles. The last term in the bracket in Eq.~(\ref{rsecnAccDev2}), where
$\delta \textbf{r}_{lj}^{(n-1)} = \int_0^t \int_0^{t'} \delta \ddot {\textbf{r}}_{lj}^{(n-1)}(t'') \rmd t'' \rmd t'$, accounts for the fact that both particles $j$ and $l$ are shifted with respect to their ballistic positions.
Since the shielded potential of section \ref{SCP} was found by first order perturbation theory, it is felt in the acceleration of particles computed to second order. This acceleration is provided by Eq.~(\ref{rsecnAccDev2}) for $n=2$. Therefore its term in brackets is the shielded acceleration of particle $l$ due to particle $j$. As a result, though the summation runs over all particles, its effective part is only due to particles $j$ typically inside the Debye sphere about particle $l$. Starting from the third iterate of Picard scheme, the effective part of the summation in Eq.~(\ref{rsecnAccDev2}) ranges inside this Debye sphere, since the $\delta \textbf{r}_{lj}^{(n-1)}$'s are then computed with a shielded acceleration. This justifies the use of the shielded potential to compute collisional transport in section \ref{DSCT}, and is the basis of the intuitive interpretation of Debye shielding in claim 9. This interpretation shows that, when starting form random particle positions, the typical time-scale for shielding to set in is the time for a thermal particle to cross a Debye sphere, i.e.\ $\omega_{\rmpp}^{-1}$, which sets the duration of the transients occurring in the inverse Laplace transform leading to Eq.~(\ref{phij}). This order of magnitude is correct for a plasma close to equilibrium.
\subsection{Debye shielding and Landau damping}
\label{STSP}
We now apply the smoothing using distribution function $f$ to $\tilde{\phi}^{(0)}(\textbf{m},\omega)$ too in Eq.~(\ref{phihatL}). As a result of Eqs~(\ref{phitildn}-\ref{phitildnj}), this yields
\begin{equation}
\tilde{\Phi}^{(0)}(\textbf{m})
= - \frac{e}{\varepsilon_0 k_{\textbf{m}}^2}\iint \exp[- \rmi \textbf{k}_{\textbf{m}} \cdot (\textbf{r} + \textbf{v} t)]f(\textbf{r},\textbf{v}) \rmd^3 \textbf{r} \rmd^3 \textbf{v}
\label{phi0tildetcg}
\end{equation}
whose Laplace transform is
\begin{equation}
\hat{\Phi}^{(0)}(\textbf{m},\omega)
= - \frac{\rmi e}{\varepsilon_0 k_{\textbf{m}}^2} \int
\frac{\tilde{f}(\textbf{m},\textbf{v})}{\omega -\textbf{k}_{\textbf{m}} \cdot \textbf{v}} \ \rmd^3 \textbf{v},
\label{phi0hatcg}
\end{equation}
which shows this second smoothing makes Eq.~(\ref{phihatL}) to become the expression allowing Landau contour calculations, as stated in Claim 3.
We point out that, in this paper, the smoothed velocity distribution is introduced after particle dynamics has been taken into account, and not before, as occurs when kinetic equations are used. This avoids addressing the issues of the exact definition of the smoothed distribution for a given realization of the plasma, and of the uncertainty as to the way the smoothed dynamics departs from the actual $N$-body one.
\section{Wave-particle dynamics}
\label{WPDRT}
In section \ref{STSP}, the existence of Langmuir waves is asserted by connection with Landau theory. To describe Langmuir waves with discrete particles, we consider that the $\textbf{r}_{l0}$'s are random, and we allow for non zero $\delta \textbf{r}_j(0)$'s and $\delta \dot{\textbf{r}}_j(0)$'s for the $\delta \textbf{r}_j$'s in Eq.~(\ref{phitildnj}). Therefore, in the formulas of section \ref{DSLD}, the $\textbf{r}_{j0}$ and $\textbf{v}_j$'s are slightly shifted with respect to the initial $\textbf{r}_j(0)$'s and $\dot{\textbf{r}}_j(0)$'s due to Langmuir waves.
Up to this point we described Langmuir waves by a linear theory. We now generalize the analysis of section \ref{FEP} to afford the description of nonlinear effects in wave-particle dynamics. Indeed, resonant particles may experience trapping or chaotic dynamics, which imply $\textbf{k}_{\textbf{m}} \cdot \delta \textbf{r}_l$'s of the order of $2 \pi$ or larger for wave $\textbf{k}_{\textbf{m}}$'s. To describe such a dynamics, it is not appropriate to expand $\phi$ as was done in Eqs~(\ref{phitildn}-\ref{phitildnj}) for such particles. However, this expansion may still be justified for non resonant particles over times where trapping and chaos show up for resonant ones. In order to keep the capability to describe the latter effects, we now split the set of $N$ particles into bulk and tail, in the spirit of Refs. \cite{OWM,OLMSS,AEE,EZE,EEB}. The bulk is defined as the set of particles which are not resonant with Langmuir waves. We then perform the analysis of section \ref{FEP} for the $N_{\mathrm{bulk}}$ particles, while keeping the exact contribution of the $N_{\mathrm{tail}}$ particles to the electrostatic potential. To this end, we number the tail particles from 1 to $N_{\mathrm{tail}}$, the bulk ones from $N_{\mathrm{tail}}+1$ to $N = N_{\mathrm{bulk}} + N_{\mathrm{tail}}$, and we call these respective sets of integer $S_{\mathrm{tail}}$ and $S_{\mathrm{bulk}}$. For $l \in S_{\mathrm{bulk}}$, we now substitute Eq.~(\ref{phitildn}) with
\begin{equation}
\tilde{\phi}_l (\textbf{m}) = \frac{N_{\mathrm{bulk}} - 1}{N_{\mathrm{bulk}}} U(\textbf{m})
+ \sum_{j \in S_{\mathrm{bulk}};j \neq l} \ \delta \tilde{\phi}_{j} (\textbf{m}) ,
\label{phitildnT}
\end{equation}
where
\begin{equation}
U(\textbf{m}) = -\frac{e N_{\mathrm{bulk}}}{\varepsilon_0 k_{\textbf{m}}^2 (N_{\mathrm{bulk}}-1)} \sum_{j \in S_{\mathrm{tail}}} \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_j).
\label{U}
\end{equation}
In the r.h.s.\ of Eq.~(\ref{phitildnT}), the first term vanishes if $N_{\mathrm{tail}} = 0$. We now perform the calculation of section \ref{FEP} on substituting the previous summations with index running from 1 to $N$ by ones where the index runs over $S_{\mathrm{bulk}}$, while keeping the exclusion of $j=l$ where indicated. The previous division by $N-1$ preceding Eq.~(\ref{phihat}) is now a division by $ N_{\mathrm{bulk}} - 1$. This yields
\begin{eqnarray}
k_{\textbf{m}}^2\hat{\varphi}(\textbf{m},\omega)
&-& \frac{e^2}{ L^3 m_e \varepsilon_0}
\sum_{\textbf{n}} \textbf{k}_{\textbf{m}} \cdot \textbf{k}_{\textbf{n}}
\ \sum_{j \in S} \frac{\Psi_j(\hat{\varphi}\ ; \textbf{n},\omega + \omega_{\textbf{n},j} - \omega_{\textbf{m},j})}{(\omega - \omega_{\textbf{m},j})^2} \exp[\rmi (\textbf{k}_{\textbf{n}}-\textbf{k}_{\textbf{m}}) \cdot \textbf{r}_{j0}]
\nonumber\\
= k_{\textbf{m}}^2 \hat{\phi}^{(0)}(\textbf{m},\omega) &-& \frac{e}{\varepsilon_0} \sum_{j \in S} \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j0}) \hat{R}_j(\omega - \omega_{\textbf{m},j})
+ k_{\textbf{m}}^2 \hat{U}(\textbf{m},\omega),
\label{phihatU}
\end{eqnarray}
where $\hat{U}(\textbf{m},\omega)$ is the Laplace transform of $U(\textbf{m})$. Then Eq.~(\ref{phihatL}) becomes
\begin{equation}
\epsilon(\textbf{m},\omega) \hat{\Phi}(\textbf{m},\omega) = \hat{\phi}^{(0)}(\textbf{m},\omega) + \hat{U}.
\label{phihatLU}
\end{equation}
Let $\tilde{\Phi}(\textbf{m},t)$ be the inverse Laplace transform of $\hat{\Phi}(\textbf{m},\omega)$, $\hat{\Phi}_{\mathrm{bulk}}(\textbf{m},\omega)$ be the solution of Eq.~(\ref{phihatL}) computed for the bulk particles, and $\tilde{\Phi}_{\mathrm{bulk}}(\textbf{m},t)$ be its inverse Laplace transform. We now derive an amplitude equation for $\tilde{\Phi}(\textbf{m},t)$ in a way similar to Refs. \cite{OWM,OLMSS}. Let $\omega_{\textbf{m}}$ be such that $\epsilon(\textbf{m},\omega_{\textbf{m}}) = 0$; because of the definition of the bulk, this frequency is real. Then $\tilde{\Phi}_{\mathrm{bulk}}(\textbf{m},t) = A \exp(-\rmi \omega_{\textbf{m}} t)$, where $A$ is a constant, and
\begin{equation}
\hat{\phi}^{(0)}(\textbf{m},\omega) = \frac{\rmi A}{\omega - \omega_{\textbf{m}}},
\label{phi0A}
\end{equation}
according to Eq.~(\ref{phihatL}).
Let $g(\textbf{m},t) = \tilde{\Phi}( \textbf{m},t) / \tilde{\Phi}_{\mathrm{bulk}}(\textbf{m},t)$. Therefore $\hat{\Phi}(\textbf{m},\omega) = A \hat{g}(\omega - \omega_{\textbf{m}})$, which together with Eq.~(\ref{phihatLU}) and (\ref{phi0A}) yields
\begin{equation}
A \epsilon(\textbf{m},\omega_{\textbf{m}} + \omega') [\hat{g}(\textbf{m},\omega') - \frac{\rmi}{\omega'}] = \hat{U}(\textbf{m},\omega_{\textbf{m}} + \omega'),
\label{eqf}
\end{equation}
where $\omega' = \omega - \omega_{\textbf{m}}$. If $ N_{\mathrm{tail}} \ll N_{\mathrm{bulk}}$, $g(\textbf{m},t)$ is a slowly evolving amplitude, and the support of $\hat{g}(\textbf{m},\omega)$ is narrow about zero. This justifies Taylor-expanding $\epsilon(\textbf{m},\omega_{\textbf{m}} + \omega')$ about $\omega' = 0$ in Eq.~(\ref{eqf}), which yields $\frac{\partial \epsilon(\textbf{m},\omega_{\textbf{m}})}{\partial \omega} \omega'$ to lowest order. Setting this in Eq.~(\ref{eqf}) and performing the inverse Laplace transform finally yields an amplitude equation for $\tilde{\Phi}(\textbf{m},t)$
\begin{equation}
\frac{\partial \tilde{\Phi}(\textbf{m},t)}{\partial t} + \rmi \omega_{\textbf{m}} \tilde{\Phi}(\textbf{m},t) =
\frac{\rmi e N_{\mathrm{bulk}}}{\varepsilon_0 k_{\textbf{m}}^2 (N_{\mathrm{bulk}} - 1) \frac{\partial \epsilon(\textbf{m},\omega_{\textbf{m}})}{\partial \omega}} \sum_{j \in S_{\mathrm{tail}}} \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_j).
\label{eqampl}
\end{equation}
The self-consistent dynamics of the potential and of the tail particles is ruled by this equation and by the equation of motion of these particles
\begin{equation}
\ddot{\textbf{r}}_j = \frac{\rmi e}{L^3 m_e} \sum_{\textbf{n}} \textbf{k}_{\textbf{n}} \ \tilde{\Phi}_j(\textbf{n}) \exp(\rmi \textbf{k}_{\textbf{n}} \cdot \textbf{r}_j).
\label{delrsecwv}
\end{equation}
These two sets of equations generalize to 3 dimensions the self-consistent dynamics defined in Refs. \cite{AEE,EEB}. For the sake of brevity, we do not develop here the full generalization of the analysis in these papers; it is lengthy, but straightforward. However, since this analysis unifies spontaneous emission and Landau growth and damping, we give the result ruling the evolution of the amplitude of a Langmuir wave provided by perturbation calculation where the right hand sides of Eqs~(\ref{eqampl}-\ref{delrsecwv}) are considered as small of order one. This is natural for Eq.~(\ref{eqampl}) since $N_{\mathrm{tail}} \ll N_{\mathrm{bulk}}$, and for Eq.~(\ref{delrsecwv}) if the Langmuir waves have a low amplitude. Let $J(\textbf{m},t) = \langle \tilde{\Phi}(\textbf{m},t)\tilde{\Phi}(- \textbf{m},t) \rangle$, where the average is over the random initial positions of the tail particles (thus their distribution is spatially uniform). Then a second order calculation in $\Phi$ yields \begin{equation}
\frac {\rmd J(\textbf{m},t)} {\rmd t}
= 2 \gamma_{\textbf{m}{\rm L}} J(\textbf{m},t) + S_{\textbf{m} \, \mathrm {\rm{spont}} },
\label{evampfinal}
\end{equation}
where $\gamma_{\textbf{m}{\rm L}}$ is the Landau growth or damping rate given by
\begin{equation}
\gamma_{\textbf{m} {\rm L}} = \alpha_{\textbf{m}}
{\frac {\rmd f_{\rm{red}}} {\rmd v}}(\frac{\omega_{\textbf{m}}}{ k_{\textbf{m}}} ; \textbf{m})
\label{SZ122}
\end{equation}
with
\begin{equation}
\alpha_{\textbf{m}}
= \frac{\pi e^2 }{m_e \varepsilon_0 k_{\textbf{m}}^2 \frac{\partial \epsilon(\textbf{m},\omega_{\textbf{m}})}{\partial \omega}},
\label{alj}
\end{equation}
and $f_{\rm{red}}(v; \textbf{m})$ is the reduced smoothed distribution function $f_{\rm{red}}(v; \textbf{m}) = \iint f(v \hat{\textbf{k}}_{\textbf{m}} + \textbf{v}_{\bot}) \ \rmd^2 \textbf{v}_{\bot}$ where $\hat{\textbf{k}}_{\textbf{m}}$ is the unit vector along $\textbf{k}_{\textbf{m}}$ and $\textbf{v}_{\bot}$ is the component of the velocity perpendicular to $\textbf{k}_{\textbf{m}}$; $S_{\textbf{m} \, \mathrm {\rm{spont}} }$ is given by
\begin{equation}
S_{\textbf{m} \, \mathrm {\rm{spont}} } = \frac{2 \alpha_{\textbf{m}}^2}{\pi e^2 k_{\textbf{m}} n} f_{\rm{red}}(\frac{\omega_{\textbf{m}}}{ k_{\textbf{m}}}),
\label{Spont}
\end{equation}
where $ n=N/L^3$ is the plasma density. $S_{\textbf{m} \, \mathrm {\rm{spont}} }$ corresponds to the spontaneous emission of waves by particles and induces an exponential relaxation of the waves to the thermal level in the case of Landau damping. The second order calculation for the particles yields the diffusion and friction coefficients of the Fokker-Planck equation ruling the tail dynamics. This equation corresponds to the classical quasilinear result, plus a dynamical friction term mirroring the spontaneous emission of waves by particles, as found in the one-dimensional case in Refs. \cite{EZE,EEB}.
Finally, we point out that Eqs~(\ref{eqampl}-\ref{delrsecwv}) enable to tackle the linear theory of Langmuir waves by considering a single mechanical system \cite{EZE,EEB}. This is naturally obtained by considering the unperturbed plasma as made up of a series of monokinetic parallel beams, and each beam as an array of equidistributed particles. This approach finds eigenmodes corresponding to the Landau instability \cite{La}, if the distribution function has a positive slope. If the slope is negative, it recovers the van Kampen modes \cite{vK}, as a series of beam modes providing Landau damping as a result of the phase mixing of the latter. This phase mixing turns out to be also important to recover a correct growth of the wave amplitude in the unstable case (section 3.8.3 of reference \cite{EEB}). We point out that in Refs. \cite{EZE,EEB} the equivalent of Eqs~(\ref{eqampl},\ref{delrsecwv}) was obtained without using any smoothing, but by a mechanical reduction of degrees of freedom, which made even stronger the idea of considering a single mechanical system.
\section{Debye shielding and collisional transport}
\label{DSCT}
We now focus on the case where the particles have random initial positions, i.e.\ where the plasma has a uniform density, and for simplicity we consider the plasma to be in thermal equilibrium. Then the dynamics of particles has no collective aspect, but is ruled by the cumulative effect of two-body deflections. More specifically, we choose random $\textbf{r}_{l0}$'s, and vanishing $\delta \textbf{r}_l(0)$'s and $\delta \dot{\textbf{r}}_l(0)$'s, and we assume that at $t=0$ all particles are in the same cube with volume $L^3$, and we consider the limit $L/\lambda_{\mathrm{D}} \rightarrow \infty$. In contrast, each particle has a well defined velocity, in such a way that the overall initial smoothed velocity distribution is close to some given function. We focus on particle $l$ which is assumed to be close to the center of the cube. In this section, we approximate the true dynamics by that due to the shielded Coulombian interactions, i.e.\ we write
\begin{equation}
\delta \ddot{\textbf{r}}_l = \sum_{j \in S;j \neq l} \textbf{a}(\textbf{r}_l-\textbf{r}_j,\textbf{v}_j),
\label{delrsecscreen}
\end{equation}
with
\begin{equation}
\textbf{a}(\textbf{r},\textbf{v}) = \frac{e}{m_e} \nabla \delta \Phi (\textbf{r},\textbf{v}),
\label{acc}
\end{equation}
where $\delta \Phi (\textbf{r},\textbf{v})$ is given by Eq.~(\ref{phi}).
We compute particle $l$ deflection in a sequence of steps. First, we use first order perturbation theory in $\delta \Phi$, which shows the total deflection to be the sum of the individual deflections due to all other particles. For an impact parameter much smaller than $\lambda_{\rm{D}}$, the deflection due to a particle turns out to be the perturbative value of the Rutherford deflection due to this particle if it were alone. Second, for a close encounter with particle $n$, we show that the deflection of particle $l$ is exactly the one it would undergo if the other $N-2$ particles were absent. Third, the deflection for an impact parameter of order $\lambda_{\rm{D}}$ is given by the Rutherford expression multiplied by some function of the impact parameter reflecting shielding. These three steps yield an analytical expression for deflection whatever the impact parameter.
We first compute $\delta \textbf{r}_l$ by first order perturbation theory in $\delta \Phi$, taking the ballistic motion defined by Eq.~(\ref{rl0}) as zeroth order approximation. This yields
\begin{equation}
\delta \dot{\textbf{r}}_{l1}(t) = \sum_{j \in S;j \neq l} \delta \dot{\textbf{r}}_{lj1}(0,t) ,
\label{delrl1}
\end{equation}
where
\begin{equation}
\delta \dot{\textbf{r}}_{lj1}(t_1,t_2) = \int_{t_1}^{t_2} \textbf{a}[\textbf{r}_l^{(0)}(t')-\textbf{r}_j^{(0)}(t'),\textbf{v}_j] \rmd t'.
\label{delrlj1}
\end{equation}
It is convenient to write
\begin{equation}
\textbf{r}_l^{(0)}(t')-\textbf{r}_j^{(0)}(t') = \textbf{b}_{lj}+ \Delta \textbf{v}_{lj} (t' - t_{lj}),
\label{blj}
\end{equation}
where $t_{lj}$ is the time of closest approach of the two ballistic orbits, and $\textbf{b}_{lj}$ is the vector joining particle $j$ to particle $l$ at this time. Then $b_{lj} = \| \textbf{b}_{lj}\|$ is the impact parameter of these two orbits when singled out. The initial random positions of the particles translate into random values of $\textbf{b}_{lj}$ and of $t_{lj}$. The typical duration of the deflection of particle $l$ given by Eq.~(\ref{delrlj1}) is $\Delta t_{lj} \equiv b_{lj} / \Delta v_{lj}$ where $\Delta v_{lj} = \| \Delta \textbf{v}_{lj} \|$, but a certain number, say $\alpha$, of $\Delta t_{lj}$'s are necessary for the deflection to be mostly completed. For a given $b_{lj} $ and for $t \gg \Delta t_{lj}$, the deflection of particle $l$ given by Eq.~(\ref{delrlj1}) is maximum if $t_{lj}$ is in the interval $[\alpha \Delta t_{lj}, t - \alpha \Delta t_{lj}]$. We notice that $\Delta t_{lj}$ is about the inverse of the plasma frequency for $b_{lj} \sim \lambda_{\rm{D}}$ and $\Delta v_{lj}$ on the order of the thermal velocity.
For the sake of brevity, we compute here just the trace of the diffusion tensor for the particle velocities. To this end, we perform an average over all the $\textbf{r}_{l0}$'s to get
\begin{equation}
\langle \delta \dot{\textbf{r}}_{l1}^2(t)\rangle = \sum_{j \in S;j \neq l}
\langle \delta \dot{\textbf{r}}_{lj1}^2(t)\rangle,
\label{delrlj1^2}
\end{equation}
taking into account Eq.~(\ref{phi}), and the fact that the initial positions are independently random, as well as the $\textbf{r}_i - \textbf{r}_j$'s for $i \neq j$. Therefore, though being due to the simultaneous scattering of particle $l$ with the many particles inside its Debye sphere, $\langle \delta \dot{\textbf{r}}_{l1}^2(t)\rangle$ turns out to be the sum of individual two-body deflections for $b_{lj} $'s such that first order perturbation theory is correct. Hence the contribution to $\langle \delta \dot{\textbf{r}}_{l1}^2(t)\rangle$ of particles with given $b_{lj}$ and $\Delta v_{lj}$ can be computed as if it would result from successive two-body collisions, as was done in Ref.\ \cite{Ros} and in many textbooks.
For an impact parameter much smaller than $\lambda_{\rm{D}}$, the main contribution of $\textbf{a}[\textbf{r}_l^{(0)}(t')-\textbf{r}_j^{(0)}(t'),\textbf{v}_j]$ to the deflection of particle $l$ comes from times $t'$ where $\| \textbf{r}_l^{(0)}(t')-\textbf{r}_j^{(0)}(t') \| \ll \lambda_{\rm{D}}$. Therefore $\textbf{a}(\textbf{r},\textbf{v})$ takes on its bare Coulombian value, and $\langle \delta \dot{\textbf{r}}_{l1}^2(t)\rangle$ is a first order approximation of the effect on particle $l$ of a Rutherford collision with particle $j$. Comparing this approximate value with the exact one shows the perturbative calculation to be correct for $b_{lj} \gg \lambda_{\rm{ma}} = \frac{e^2}{ \pi m_e \varepsilon_0 \Delta v_{lj}^2}$, the distance of minimum approach of two electrons in a Rutherford collision, as given by energy conservation.
Second, we consider the case of the close approach of particle $n$ to particle $l$, i.e.\ $b_{ln} \sim \lambda_{\rm{ma}}$. We write the acceleration of particle $l$ as
\begin{equation}
\ddot{\textbf{r}}_l = \textbf{a}(\textbf{r}_l-\textbf{r}_n) + \sum_{j \in S;j \neq l,n} \textbf{a}(\textbf{r}_l-\textbf{r}_j).
\label{delrsecEx}
\end{equation}
For particle $n$, we write the same equation by exchanging indices $l$ and $n$. Since the two particles are at distances much smaller than the inter-particle distance $d = n^{-1/3} = L/N^{1/3}$, the accelerations they get from all other particles are almost the same. Therefore, when subtracting the two rigorous equations of motion, the two summations over $j$ almost cancel, yielding
\begin{equation}
\frac{\rmd^2 (\textbf{r}_l - \textbf{r}_n)}{\rmd t^2} = 2 \textbf{a}(\textbf{r}_l-\textbf{r}_n),
\label{delrsecEx2}
\end{equation}
which is the equation describing the Rutherford collision of these two particles in their center of mass frame, in the absence of all other particles (at such distances the shielded potential is the bare Coulomb one). Since $b_{ln} \ll d$, $\Delta t_{ln}$ is much smaller than the $\Delta t_{lj}$'s of the other particles. Therefore the latter produce a negligible deflection of the center of mass during the Rutherford two-body collision, and the deflection of particle $l$ during this collision is exactly that of a Rutherford two-body collision. The contribution of such collisions to $\langle \delta \dot{\textbf{r}}_{l}^2(t)\rangle $ was calculated in Ref.\ \cite{Ros}.
Now, since the deflection of particle $l$ due to particle $j$ as computed by the above perturbation theory is an approximation of the Rutherford deflection for the same impact parameter, we may approximate the perturbative deflection one by the full Rutherford one, which provides an obvious matching of the theories for $b_{lj} \sim \lambda_{\rm{ma}}$ and for $\lambda_{\rm{D}} \gg b_{lj} \gg \lambda_{\rm{ma}}$: we may use the estimate of \cite{Ros} in the whole domain $b_{lj} \ll \lambda_{\rm{D}}$.
Third, we deal with impact parameters of the order of $\lambda_{\rm{D}}$. Then the deflection due to particle $j$ must be computed with Eq.~(\ref{delrlj1}). For the sake of simplicity, we make the calculation for the case where $\textbf{v}_j$ is small, which makes $\delta \Phi (\textbf{r},\textbf{v}) \simeq \delta \Phi (\textbf{r},\textbf{0})$ which is the Yukawa potential $\delta \Phi_{\rm{Y}} (\textbf{r}) = - \frac{e}{4 \pi \varepsilon_0 \| \textbf{r} \|} \exp (- \frac{\| \textbf{r} \|}{\lambda_{\rm{D}}})$ (Eq.~(18) of Ref.\ \cite{Gasio}). The first order correction in $\textbf{k}_{\textbf{m}} \cdot \textbf{v}_j$ to this approximation is a dipolar potential with an electric dipole moment proportional to $\textbf{v}_j$. Since a Maxwellian distribution is symmetrical in $\textbf{v}$, these individual dipolar contributions cancel globally. As a result, the first relevant correction to the Yukawa potential is of second order in $\textbf{k}_{\textbf{m}} \cdot \textbf{v}_j$. This should make the Yukawa approximation relevant for a large part of the bulk of the Maxwellian distribution.
In the small deflection limit, a calculation using the fact that the force derives from a central potential shows the full deflection of particle $l$ due to particle $j$ is provided by
\begin{equation}
\delta \dot{\textbf{r}}_{lj1}(- \infty,+ \infty) =
\frac{e^2}{ 4 \pi m_e \varepsilon_0} \textbf{b}_{lj}
\int_{- \infty}^{+ \infty} [\frac{1}{r^3(t)} + \frac{1}{\lambda_{\rm{D}} r^2(t)}] \exp [- \frac{r(t)}{\lambda_{\rm{D}}}]\rmd t,
\label{delrljT}
\end{equation}
where $r(t) = (b_{lj}^2 +\Delta v_{lj}^2 t^2)^{1/2}$ and $\textbf{b}_{lj}$ was defined with Eq.~(\ref{blj}). Defining $\theta = \arcsin [ \Delta v_{lj} t / r(t)]$, this equation becomes
\begin{equation}
\delta \dot{\textbf{r}}_{lj1}(- \infty,+ \infty) =
- \frac{2 e^2}{ 4 \pi m_e \varepsilon_0 \Delta v_{lj}}
\, \frac{h(b_{lj})}{b_{lj}^2 } \, \textbf{b}_{lj},
\label{delrljTfin}
\end{equation}
where
\begin{equation}
h(b) =
\int_{0}^{\pi/2} [ \cos (\theta) + \frac{b}{\lambda_{\rm{D}}}] \exp [- \frac{b}{\lambda_{\rm{D}} \cos (\theta)}] \ \rmd \theta
< [1 + \frac{\pi b}{2 \lambda_{\rm{D}}}] \exp [- \frac{b}{\lambda_{\rm{D}}}].
\label{delrljT2}
\end{equation}
During time $t \gg \Delta t_{lj}$, a volume $2 \pi \Delta v_{lj} t b_{lj} \delta b_{lj}$ of particles with velocity $\textbf{v}_j$ and impact parameters between $b_{lj}$ and $b_{lj} + \delta b_{lj}$ produce the deflection of particle $l$ given by Eq.~(\ref{delrljTfin}), and a contribution scaling like $\frac{h^2(b_{lj})}{b_{lj}} \delta b_{lj}$ to $\langle \delta \dot{\textbf{r}}_{l1}^2(t)\rangle$. Let $b_{\rm{min}}$ be such that $\lambda_{\rm{D}} \gg b_{\rm{min}} \gg \lambda_{\rm{ma}}$. The contribution of all impact parameters between $b_{\rm{min}}$ and some $b_{\rm{max}}$ is thus scaling like the integral $\int_{b_{\rm{min}}}^{b_{\rm{max}}} h^2(b)/b \ \rmd b$. Since for $b$ small $h(0) \simeq 1$, if $b_{\rm{max}} \ll \lambda_{\rm{D}}$, this is the non-shielded contribution of orbits relevant to the above perturbative calculation. Since, by approximating it by the Rutherford-like result of Ref.\ \cite{Ros}, this contribution matches that for impact parameters on the order of $\lambda_{\rm{ma}}$, the contribution of all impact parameters between $\lambda_{\rm{ma}}$ and some $b_{\rm{max}}$ small with respect to $\lambda_{\rm{D}}$ is thus scaling like the integral $\int_{\lambda_{\rm{ma}}}^{b_{\rm{max}}} 1/b \ \rmd b$ as was computed in Ref.\ \cite{Ros}. The matching of this result for $b \sim \lambda_{\rm{D}}$ is simply accomplished by setting a factor $h^2(b)$ in the integrand which makes the integral converge for $b \rightarrow \infty$. Taking this limit, one finds that the Coulomb logarithm $\ln (\lambda_{\rm{D}} / \lambda_{\rm{ma}})$ of the second Eq.~(14) of Ref.\ \cite{Ros} becomes $\ln (\lambda_{\rm{D}} / \lambda_{\rm{ma}}) + C$ where $C$ is of order unity. If the full dependence of the shielding on $\textbf{v}_j$ were taken into account, the modification of the Coulomb logarithm would be velocity dependent.
For the sake of brevity, we do not develop here the calculation of the dynamical friction which will be presented elsewhere. This calculation requires using second order perturbation theory, but follows the same lines as those for the diffusion coefficient.
For an inhomogeneous plasma, the acceleration of particle $l$ may be split into a homogeneous and a wave part. When using the linearized versions of Eq.~(\ref{phihat}) and subsequent ones in section \ref{DSLD}, we can split all $\phi$'s and $\Phi$'s into a homogeneous part and an independent inhomogeneous one. Therefore the diffusion coefficient and the dynamical friction estimated by perturbative calculation of the dynamics up to second order are the sum of the collisional contribution and of a contribution due to waves, the latter as calculated for instance in Refs. \cite{EZE,EEB}. For the sake of brevity, we defer this point to later publication.
\section{Conclusion}
\label{Concl}
In this paper, Debye shielding, collisional transport, Landau damping of Langmuir waves, and spontaneous emission of these waves were introduced by a mechanical approach. This approach brings unification and simplification in basic microscopic plasma physics, and may be useful for pedagogical purposes. One might think about trying to apply the above approach to plasmas with more species, or with a magnetic field, or where particles experience trapping and chaotic dynamics. The first generalization sounds rather trivial, and the third one is under way, at least in one dimension (see a pedestrian introduction in \cite{Houches} and more specific results in \cite{BEEB,BEEBEPS}).
As in many textbooks, linearization was applied in sections \ref{FEP}, \ref{SCP}, and \ref{STSP} without questioning deeply its range of validity. However, the smallness of the perturbation is not a sufficient criterion. Indeed, as reviewed in Ref.\ \cite{HS}, perturbation theory that relies on linearization has to be questioned, as it yields a solution of the linearized set of equations only. Whether it also generates a solution of the full set has to be shown explicitly, and this may be a hard task. For instance the full proof of existence of Landau damping \cite{MV} in a Vlasovian frame was a mathematical tour de force, the equivalent of a Kolmogorov-Arnold-Moser theorem for continuous systems, and led one of its authors to be awarded the 2010 Fields medal.
As to collisional transport, as usual, it was assumed that some kind of decorrelation justifies the calculation of the transport coefficients over the finite times considered in section \ref{DSCT}. However, in reality particle dynamics is chaotic, and one is facing the calculation of transport coefficients for a chaotic motion. One should be cautious as, for the motion of a charged particle in a spectrum of longitudinal waves, a perturbation calculation yields the quasilinear estimate for the diffusion coefficient,
while a super-quasilinear regime, a synergetic effect in chaos, is found to exist in this chaotic dynamics for intermediate resonance overlap \cite{CEV}. There, diffusion becomes quasilinear for strong resonance overlap, but not because the perturbation calculation becomes valid again (\cite{EE2}, section 6.8.2 of \cite{EEB}).
Furthermore, we used only a very specific part of the fundamental equation (\ref{phihat}): the one involving linearization and smoothing. It would be interesting to study the effect of the coupling of Fourier components with both coherent and incoherent effects, in particular, to perform the analysis of section \ref{WPDRT} by substituting $k_{\textbf{m}}^2 \hat{U}(\textbf{m},\omega)$ with $- \frac{e}{\varepsilon_0} \sum_{j \in S} \exp(- \rmi \textbf{k}_{\textbf{m}} \cdot \textbf{r}_{j0}) \hat{R}_j(\omega - \omega_{\textbf{m},j})$. The question arises: is it possible to recover the hole solutions propagating near thermal velocity or slower, which are smooth and nonlinear structures satisfying the full nonlinear Vlasov-Poisson system (Ref.\ \cite{HS} and references therein)?
\vspace{8 mm}
Ph. Choquard, L. Cou\"edel, M.-C. Firpo, W. Horton, J.T. Mendon\c{c}a, F. Pegoraro, Y. Peysson, H. Schamel, and J.-Zh. Zhu are thanked for very useful comments and new references.
|
\section{Introduction}
We consider
Feynman integrals in $D$-dimensional Minkowski space with one time- and $(D-1)$
Euclidean space dimensions, $\varepsilon = D - 4$ and $\varepsilon \in {\mathbb R}$ with
$|\varepsilon| \ll 1$, and with at most one mass.
Here the discrete Mellin parameter $n$ comes from local operator insertions.
As worked out in detail in~\cite{Blumlein:2009ta,BKSF:12} these integrals can be
transformed
to integrals of the form\begin{eqnarray}
\label{Equ:HypInt}
{\cal I}(\ep,n) = C(\ep, n, M) \int_0^1 dx_1 \ldots \int_0^1 dx_m
\frac{\sum_{i=1}^k \prod_{{l}=1}^{r_i}
[P_{i,l}(x_1,\dots,x_m)]^{\alpha_{i,l}(\varepsilon,n)}}{[Q(x_1,\dots,x_m)]^{\beta(\varepsilon)}}~,
\end{eqnarray}
with $k\in\set N$, $r_1,\dots,r_k\in\set N$ and
where $\beta(\ep)$ is given by a rational function in $\ep$, i.e., $\beta(\ep)\in\set Q(\ep)$, and similarly
$\alpha_{i,l}(\ep,n) = n_{i,l} n + \overline{\alpha}_{i,l}$ for some $n_{i,l} \in \{0,1\}$ and $\overline{\alpha}_{i,l}\in\set Q(\ep)$, see also \cite{BOGNER}
in the case no local operator insertions are present.
$C(\ep, n, M)$ is a factor, which depends on the dimensional parameter $\ep$,
the integer parameter $n$ and the mass $M$. $P_i(x_1,\dots,x_m), Q(x_1,\dots,x_m)$
are polynomials in the $x_i$. Integrals of the type (\ref{Equ:HypInt}) emerge in the
calculation of unpolarized and polarized massive operator matrix elements (OMEs)
\cite{Bierenbaum:2007qe,Bierenbaum:2007dm,BBKS:08,Bierenbaum:2009zt,
Bierenbaum:2009mv,Blumlein:2009rg,HYP2,ABHKS:12,BHKC:13} and in
other single scale higher loop calculations. In
\cite{Bierenbaum:2009mv,Blumlein:2009rg} 3-loop moments of the corresponding OMEs
have been calculated.
In addition, such integrals~\eqref{Equ:HypInt} can be transformed to proper hypergeometric multi-sums of the form\footnote{For convenience, we assume that the summand is written terms of the Gamma function $\Gamma(x)$. Later, also Pochhammer symbols or binomial coefficients are used which can (if necessary) be rewritten in terms of Gamma-functions.}
\begin{equation}\label{Eq:GenericMultiSum}
{\cal S}(\ep,n) = \sum_{n_1=1}^\infty ... \sum_{n_r=1}^\infty
\sum_{k_1=1}^{L_1(n)} ... \sum_{k_v=1}^{L_v(n,k_1, ..., k_{v-1})}
\sum_{k=1}^l C_k(\ep, n, M)
\frac{\Gamma(t_{1,k}) \ldots \Gamma(t_{v',k})}
{\Gamma(t_{v'+1,k}) \ldots \Gamma(t_{w',k})}.
\end{equation}
Here the upper bounds $L_1(n),\dots,L_{v}(n,k_1,\dots,k_{v-1})$ are integer linear (i.e., linear combinations of the variables over the integers) in
the dependent parameters or $\infty$, and $t_{l,k}$ are linear combinations of the $n_1,\dots,n_r$, of the $k_1,\dots,k_v$, and of $\ep$ over $\set Q$.
Finally, if the sums~\eqref{Eq:GenericMultiSum} are uniformly convergent, one of the most common tactics is as follows. First one expands the summand of~\eqref{Eq:GenericMultiSum}, say
\begin{equation}\label{Equ:SummandExpand}
F(n,n_1,\dots,n_r,v_1,\dots,v_k)=F_t(n,n_1,\dots,v_k)\ep^t+F_t(n,n_1,\dots,v_k)\ep^{t+1}+\dots
\end{equation}
with $t\in\set Z$
by using formulas such as
\begin{equation}\label{Eq:GammaExpand}
\Gamma(n+1+\bar{\ep}) = \frac{\Gamma(n) \Gamma(1+\bar{\ep})}{B(n,1+\bar{\ep})}\text{ and }B(n, 1 + \bar{\ep}) = \frac{1}{n}\exp\left(\sum_{k=1}^\infty \frac{(-\bar{\ep})^k}{k} S_k(n)\right)
= \frac{1}{n}\sum_{k=0}^\infty (-\bar{\ep})^k S_{\underbrace{\mbox{\scriptsize 1, \ldots
,1}}_{\mbox{\scriptsize
$k$}}}(n)
\end{equation}
with $\bar{\ep} = r \ep$ for some $r\in\set Q$.
Here
$B(x,y)=\Gamma(x)\Gamma(y)/\Gamma(x+y)$ denotes the Beta-function and $S_{\mbox{\scriptsize 1, \ldots
,1}}(n)$ is a special instance of the harmonic sums~\cite{Vermaseren:99,Bluemlein:99} defined by
\begin{equation}\label{Equ:HarmonicSums}
S_{m_1,\dots,m_k}(n)=
\sum_{i_1=1}^n\frac{\text{\small$\text{sign}(m_1)^{i_1}$}}{i_1^{|m_1|}}\dots
\sum_{i_k=1}^{i_{k-1}}\frac{\text{\small$\text{sign}(m_k)^{i_k}$}}{i_k^{|m_k|}}
\end{equation}
with $m_1,\dots,m_k$ being nonzero integers. Then one applies the summation signs to each of the coefficients in~\eqref{Equ:SummandExpand}. I.e., the $i$th coefficient of the $\ep$-expansion of~\eqref{Eq:GenericMultiSum} yields
$$\sum_{n_1=1}^\infty ... \sum_{n_r=1}^\infty
\sum_{k_1=1}^{L_1(n)} ... \sum_{k_v=1}^{L_v(n,k_1, ..., k_{v-1})}
\sum_{k=1}^lF_i(n,n_1,\dots,n_r,v_1,\dots,v_k).$$
Then the essential problem is the simplification of these sums to special functions, like, e.g., harmonic sums, $S$--sums~\cite{Moch:02}
\begin{equation}\label{Equ:SSums}
S_{m_1,\dots,m_k}(x_1,\dots,x_k,n)=
\sum_{i_1=1}^n\frac{x_1^{i_1}}{i_1^{m_1}}\dots
\sum_{i_k=1}^{i_{k-1}}\frac{x_k^{i_k}}{i_k^{m_k}},
\end{equation}
cyclotomic harmonic sums~\cite{ABC:11}, or more generally to indefinite nested sums and products~\cite{Schneider:10b}.
For various special cases, this simplification can be carried out with efficient
methods available, e.g., in {\tt Form}; see in~\cite{Vermaseren:99,Moch:02}.
\noindent More general sums can be handled with the Mathematica package \texttt{EvaluateMultiSum} \cite{SchneiderSummation:10,BHC11} based on the summation package \texttt{Sigma} \cite{Schneider:07a}. With the underlying difference field algorithms~\cite{Karr:81,Schneider:05a,Schneider:05c,Schneider:08c,Schneider:10b,ABPS:12} generalizing the hypergeometric summation paradigms~\cite{AequalB} to multi-sum\-mation we are currently simplifying sums up to nesting depth 7. The compact representation of the output, i.e., the elimination of all algebraic relations among the arising indefinite nested sums and products can be guaranteed by difference field theory~\cite{Schneider:10c}. For harmonic sums, cyclotomic sums, $S$-sums and cyclotomic $S$-sums and their infinite versions (quasi-)shuffle algebras are utilized; see e.g.,~\cite{Bluemlein:04,ANCONT3,MZV:10,ABC:11,ABC:12} and references therein. In this regard, the Mathematica package \texttt{HarmonicSums} is heavily used~\cite{Ablinger:12}. This general machinery has been applied to non-trivial massive 3-loop diagrams arising, e.g., in~\cite{BBKS:08,HYP2,ABHKS:12,BHKC:13}.
\noindent Another possibility is the method of hyperlogarithms~\cite{Brown:09} which can be used to evaluate integrals of the form~\eqref{Equ:HypInt} for specific values $n\in\set N$ if one can set $\ep=0$. An adaption of this method for symbolic $n$ has been described and applied to massive 3-loop ladder graphs in~\cite{ABHKS:12}.
In this article we follow another promising approach for the all $n$ expansion\\
1. Calculate a recurrence in $n$ for the multi-integral~\eqref{Equ:HypInt} or multi-sum~\eqref{Eq:GenericMultiSum}\\
2. Given this recurrence and initial values (using, e.g., the \texttt{EvaluateMultiSum} package),\\
\hspace*{0.4cm}calculate the $\ep$-expansion by a recurrence solver for Laurent expansions.
\smallskip
\noindent In~\cite{BKSF:12} we followed this approach by calculating recurrences of multi-sums using techniques of~\cite{WZtheory} and efficient algorithms developed in~\cite{Wegschaider}.
Subsequently, we present two new techniques to compute such recurrences.
In the first approach we apply an enhanced and optimized version~\cite{Ablinger:12} of the multivariate Almkvist--Zeilberger algorithm~\cite{ZEILB} to calculate recurrences for Feynman integrals in the form~\eqref{Equ:HypInt}. In the second approach we use a common framework~\cite{Schneider:05b} within the summation package Sigma that combines difference field~\cite{Karr:81,Schneider:05a} and holonomic summation techniques~\cite{Zeilberger-holonomic,Chyzak} to compute recurrences for Feynman integrals in the form~\eqref{Eq:GenericMultiSum}.
Two new packages, \texttt{MultiIntegrate} by J. Ablinger and \texttt{RhoSum} by M. Round facilitate these tasks completely automatically for the integral or sum representation, respectively.
The outline of the article is as follows. In Section~\ref{Sec:LaurentSolver} we present the general idea of how the Laurent series representation can be calculated from the given recurrence representation. With this knowledge, we illustrate our integration and summation methods in Sections~\ref{Sec:Integrate} and~\ref{Sec:Sum}, respectively.
\section{Finding Laurent series solutions of linear recurrences}\label{Sec:LaurentSolver}
One of the key ingredients of the summation and integration tools under consideration is a recurrence solver for $\ep$-expansions.
To illustrate the ideas of the solver, we consider the single sum
\begin{equation}\label{Equ:SimpleSumProblem}
\begin{split}
\mathcal{S}(\ep, n)&=\sum_{k=1}^{\infty }\frac{B\left(\frac{\ep}{2}+k,n\right)}{(k+n)^
2}=\sum_{k=1}^{\infty } \frac{\Gamma(n) \Gamma\big(\frac{\ep}{2}+k\big)}{(k+n)^2 \Gamma\big(\frac{\ep}{2}+k+n\big)}
\stackrel{?}{=}F_0(n)+F_1(n)\ep+F_2(n)\ep^2+\dots
\end{split}
\end{equation}
which is related, e.g., to sums arising in~\cite{BBKS:08}.
In order to derive the coefficients $F_i(n)$, we compute in a first step a recurrence relation using the summation package \texttt{Sigma}. Internally, it activates a difference field version of Zeilberger's creative telescoping paradigm~\cite{AequalB}. In our example it turns out that $\mathcal{S}(\ep, n)$ satisfies for all integer $n\geq1$, as an analytic function in $\varepsilon$ throughout an annular region centered by $0$,
the recurrence
\begin{multline}\label{Equ:SingleSumRec}
a_0(\varepsilon,n)\mathcal{S}(\ep, n)+a_1(\varepsilon,n) \mathcal{S}(\ep, n+1)+a_2(\varepsilon,n)\mathcal{S}(\ep, n+2)\\
=\frac{-4 (2 n+3) \left(\ep+4 n^2+12
n+8\right) \Gamma
\left(\frac{\ep}{2}+1\right)
\Gamma (n+1)}{(n+1) (n+2)^2 (\ep+2
n+2) \Gamma
\left(\frac{\ep}{2}+n+1\right)}
\end{multline}
with
$a_0(\varepsilon,n)=-4 n (n+1)$,
$a_1(\varepsilon,n)=4 (n+1) (\ep-2n-3)$, and
$a_2(\varepsilon,n)=(\ep-2 n-4)^2$.
Next, we compute the $\ep$-expansion
$h_0(n)+h_1(n)\ep+h_2(n)\ep^2+\dots$
of the right hand side of~\eqref{Equ:SingleSumRec}
by formulas such as~\eqref{Eq:GammaExpand}; here we get
$h_0(n)=\tfrac{-8 (2 n+3)}{(n+1) (n+2)}$, $h_1(n)=\big(\tfrac{4(2 n+3) S_1(n)}{(n+1) (n+2)}\!+\!\tfrac{2 (2 n+3)^2}{(n+1)^2 (n+2)^2}\big)$,
$h_2(n)=-\big(\tfrac{(2 n+3)^2 S_1(n)}{(n+1)^2 (n+2)^2}\!+\!\frac{(2 n+3) S_1(n){}^2}{(n+1) (n+2)}+\tfrac{(2 n+3) S_2(n)}{(n+1) (n+2)}+\frac{(2 n+3)^2}{(n+1)^3 (n+2)^2}\big)$.
As a consequence, for the Laurent series expansion $\mathcal{S}(\ep,n)=F_0(n)+F_1(n)\ep+\dots$ the following relation holds:
\begin{equation}\label{Equ:SingleSumAnsatz}
\begin{split}
a_0(\ep,n)\Big[F_0(n)+F_1(n)\ep+\dots]
+a_1(\ep,n)\Big[F_0(n+1)+F_1(n+1)\ep+\dots]\\
+a_2(\ep,n)\Big[F_0(n+2)+F_1(n+2)\ep+\dots]
=h_0(n)+h_1(n)\ep+h_2(n)\ep^2+\dots
\end{split}
\end{equation}
Next, we expand the first two initial values\footnote{Here $\zeta_r$ stands for the Riemann Zeta function $\zeta(r)=\sum_{i=1}^{\infty}\frac{1}{i^r}$.} of ${\cal S}(\ep,n)$, $n=1,2$:
\begin{equation}\label{Equ:InitEpSolve}
\begin{split}
\mathcal{S}(\ep, 1)&=2-\zeta_2+(\tfrac{3}{2}-\zeta_2 )\ep+( -\tfrac{3 \zeta_2}{4}+\tfrac{\zeta_3}{4}+1)\ep^2+\dots,\\
\mathcal{S}(\ep,2)&=\tfrac{\zeta_2}{2}-\tfrac{3}{4}+(\frac{3 \zeta_2}{4}-\tfrac{41}{32})\ep+(\tfrac{21 \zeta_2}{32}-\frac{3 \zeta_3}{16}-\frac{53}{64})\ep^2+\dots
\end{split}
\end{equation}
by using the package \texttt{EvaluateMultiSum}~\cite{BHC11} in Mathematica; alternatively, one could use the package \texttt{Summer}~\cite{Vermaseren:99} in Form.
Then given the recurrence~\eqref{Equ:SingleSumAnsatz} and the first initial values~\eqref{Equ:InitEpSolve} (to be more precise the polynomials $a_i(\ep,n)$, the first coefficients $h_i(n)$, and the first values of their expansions), we are ready to calculate the first three coefficients of the all $n$ series expansion~\eqref{Equ:SimpleSumProblem}.
\noindent Namely, by setting $\varepsilon=0$ in~\eqref{Equ:SingleSumAnsatz}, it follows that the constant term $F_0(n)$ satisfies the recurrence
\begin{equation}\label{Equ:SingleSumConstrained}
a_0(0,n)F_0(n)+a_1(0,n)F_0(n+1)+a_2(0,n)F_0(n+2)=h_0(n).
\end{equation}
Note that together with $F_0(1)=2-\zeta_2$ and $F_0(2)=\tfrac{\zeta_2}{2}-\tfrac{3}{4}$ the sequence $F_0(n)$ is completely determined. At this point we exploit algorithms from~\cite{Petkov:92,Abramov:94,Schneider:05a,ABPS:12} which can constructively decide if a solution with certain initial values is expressible in terms of indefinite nested products and sums. More precisely, with~\texttt{Sigma}\ one obtains
\begin{equation}\label{Equ:SingleSumF0}
F_0(n)=\frac{2 (-1)^n S_{-2}(n)}{n}+\frac{(-1)^n \zeta_2}{n}.
\end{equation}
\smallskip
\noindent Now, plugging in the partial solution
$$\mathcal{S}(\ep, n)=\frac{2 (-1)^n S_{-2}(n)}{n}+\frac{(-1)^n \zeta_2}{n}+F_1(n)\varepsilon+F_2(n)\varepsilon^2+\dots$$
into~\eqref{Equ:SingleSumAnsatz} and moving $F_0(n)$ to the right hand side yields
\begin{multline}\label{Equ:RecWithEp}
a_0(\varepsilon,n)\big[F_1(n)\varepsilon+F_2(n)\varepsilon^2+\dots\big]+a_1(\varepsilon,n)\big[F_1(n+1)\varepsilon+F_2(n+1)\varepsilon^2+\dots\big]\\
+a_2(\varepsilon,n)\big[F_1(n+2)\varepsilon+F_2(n+2)\varepsilon^2+\dots\big]
=h'_0(n)\ep+h'_1(n)\ep^2+\dots
\end{multline}
with $h'_0(n)=\frac{4 (2 n+3) S_1(n)}{(n+1) (n+2)}+\frac{2 (4 n+5)}{(n+1)^2
(n+2)^2}$
and
\begin{align*}
h'_1(n)=\tfrac{-\big(2 n^2+9 n+8\big) (2 n+3)}{(n+1)^3 (n+2)^3}-\tfrac{(2 n+3)^2 S_1(n)}{(n+1)^2 (n+2)^2}-\tfrac{(2 n+3) S_1(n){}^2}{(n+1) (n+2)}+\tfrac{(-2 n-3) S_2(n)}{(n+1) (n+2)}+\tfrac{(-1)^n(2S_{-2}(n)+\zeta_2)}{n+2}.
\end{align*}
\normalsize
Observe that the coefficients and the inhomogeneous side of the recurrence~\eqref{Equ:RecWithEp} have all the factor $\ep$. Hence dividing the recurrence through $\ep$, we obtain again a recurrence of the form~\eqref{Equ:SingleSumAnsatz} where the explicitly given $h_i$ are changed to $h'_i$ and $F_1$ is the new constant term. In other words, we can repeat this construction process for $F_1(n)$. Namely, by coefficient comparison $F_1(n)$ is uniquely determined by
\begin{equation*}
a_0(0,n)F_1(n)+a_1(0,n)F_1(n+1)+a_2(0,n)F_1(n+2)=h'_0(n)
\end{equation*}
and the initial values given in~\eqref{Equ:InitEpSolve}.
Solving the recurrence with these initial values leads for all $n\geq1$ to the sum representation
\begin{equation}
F_1(n)=(-1)^n \big(\frac{5 S_{-3}(n)}{2 n}-\frac{3 S_{-2,1}(n)}{n}\big)+\frac{(-1)^n \zeta_2 S_1(n)}{n}+\frac{2 (-1)^n S_1(n) S_{-2}(n)}{n}.
\end{equation}
Similarly, one can loop further and calculates, e.g., the coefficients $F_0(n),\dots,F_6(n)$ (in terms of $\zeta_2$, $\zeta_3$, $\zeta_5$, $\zeta_7$ and harmonic sums up to weight 8) in about 30 minutes.
\medskip
\noindent More generally, let ${\cal S}(\varepsilon,n)$ be a function
\begin{itemize}
\item which is for each integer $n$ with $n\geq\lambda$ analytic in $\varepsilon$ throughout an annular region centered by $0$; let ${\cal S}(\varepsilon,n)=\sum_{i=t}^{\infty}F_i(n)\varepsilon^i$ be its Laurent series expansion for some $t\in\set Z$;
\item which satisfies (as $\ep$-expansion) the recurrence
\begin{equation*}
a_0(\varepsilon,n){\cal S}(\varepsilon,n)+\dots+a_d(\varepsilon,n){\cal S}(\varepsilon,n+d)=h_t(n)\ep^t+h_{t+1}(n)\ep^{t+1}\dots
\end{equation*}
for polynomials $a_i(\ep,n)\in\set K[\ep,t]$ with $a_d(0,n)\neq 0$ for all $n\geq\lambda$ and for functions $h_i(n)$ where the first coefficients $h_t(n),h_{t+1}(n)\dots,h_u(n)$ with $n\geq\lambda$ are expressible in terms of indefinite nested sums and products.
\end{itemize}
\noindent Then there is the following algorithm~\cite[Cor.~1]{BKSF:12} implemented in \texttt{Sigma}:\\
\noindent\textbf{Input:} the polynomials $a_i(\ep,n)$ and the product-sum expressions $h_i(n)$ ($t\leq i\leq u$) as above; the values $c_{ij}$ ($t\leq i\leq u$, $\lambda\leq j\leq \lambda+d)$ such that
${\cal S}(\ep,j)=c_{tj}\ep^t+c_{t+1,j}\ep^{t+1}+\dots+c_{u,j}\ep^{u}+\dots$\\
\textbf{Output:} the maximal $r\in\{-\infty,t,t+1,\dots,u\}$ such that the coefficients $F_t(n),F_{t+1}(n)\dots,F_r(n)$ of the $\ep$-expansion of ${\cal S}(\varepsilon,n)$ can be expressed in terms of indefinite nested sums and products. If $r\neq-\infty$, these representations of the coefficients are computed explicitly.
\medskip
\noindent For rigorous proofs, further details concerning efficiency, generalizations, and the function call within the package Sigma we refer to~\cite{BKSF:12}.
\section{Calculating $\ep$-expansions for multi-integrals}\label{Sec:Integrate}
We aim at computing a recurrence relation for multi-integrals of the form~\eqref{Equ:HypInt} and finding a Laurent-series solution that agrees with the input integral.
Here we exploit an enhanced version~\cite{Ablinger:12} of the multivariate Almkvist--Zeilberger algorithm~\cite{ZEILB} which contains as input class these integrals. More generally, it can handle integrands being hyperexponential in the integration variables $x_i$ (i.e., the logarithmic derivative of the integrand w.r.t.\ $x_i$ is a rational function in the $x_i$ and $n$) and hypergeometric in the discrete parameter $n$ (i.e., the shift quotient w.r.t.\ $n$ of the integrand is a rational function in the $x_i$ and $n$).
In order to illustrate the basic ideas, consider the double integral
\begin{equation}\label{Equ:DoubleInt}
{\cal I}(\ep,n)=\int_0^1\int_0^1\underbrace{\frac{(1+x_1\cdot x_2)^n}{(1+x_1)^\ep}}_{F(n,x_1,x_2):=}dx_1dx_2
\stackrel{?}{=}F_0(n)+F_1(n)\ep+F_2(n)\ep^2+\dots
\end{equation}
First, one applies the multivariate Almkvist--Zeilberger algorithm. To be more precise, given $d\in\set N$ one looks for polynomials $e_i(n)$ and rational functions $R_i(n,x_1,x_2)$ such that
\begin{multline*}
e_0(n)F(n,x_1,x_2)+e_2(n)F(n+1,x_1,x_2)+\dots+e_d(n)F(n+d,x_1,x_2)\\
=D_{x1}(R_1 F(n,x_1,x_2))+D_{x2}(R_2 F(n,x_1,x_2));
\end{multline*}
here $D_{x_i}$ stands for the differentiation w.r.t.\ $x_i$. Internally, a clever ansatz is performed with undetermined coefficients which amounts to solving a linear system of equations. To hunt for a solution, one starts with $d=0$ and increases the recurrence order $d$ step by step until a solution is found. In our particular case, for $d=1$, one gets
\begin{equation}\label{Equ:IntegrandRec}
-(n+1)F(n,x_1,x_2)+(n+2)F(n+1,x_1,x_2)=D_{x_1}0+D_{x_2}\frac{x_2(x_1\cdot x_2+1)^{n+1}}{(1+x_1)^{\ep}}.
\end{equation}
Applying now the two integral signs on both sides of~\eqref{Equ:IntegrandRec} leads to the following recurrence
\begin{equation}\label{mAZExrec1}
-(n+1){\cal I}(\ep,n)+(n+2){\cal I}(\ep,n+1)=\underbrace{\int_0^1(x_1+1)^{n+1-\ep}dx_1}_{I_1(\ep,n)}-\int_0^1 0 dx_1.
\end{equation}
Note that the right hand side of~\eqref{Equ:IntegrandRec} consists again of an integral. However, the nested depth is decreased from two to one. By recursion, we treat now this simpler integral again by the method under consideration. In this case, we find
\begin{equation*}
I_1(\ep,n)=\tfrac{2^{n+2}-1}{n+2}+\tfrac{\left(-2^{n+2} \left(\log(2) (n+2)-1\right)-1\right)}{(n+2)^2}\ep\\
+\tfrac{\left(2^{n+1}\left(\log(2)^2 (n+2)^2-2 \log(2) (n+2)+2\right)-1\right)}{(n+2)^3}\ep^2+\dots
\end{equation*}
Finally, plugging this result into \eqref{mAZExrec1} gives a recurrence that fits into the input class of the recurrence solver presented in Section~\ref{Sec:LaurentSolver}. Together with the expanded initial values (which can be calculated easily)
$${\cal I}(\ep,0)=\tfrac{5}{4}+\left(\tfrac{7}{8}-2 \log(2)\right)\ep+ \left(\tfrac{11}{16}+\log^2(2)-\tfrac{3 \log (2)}{2}\right)\ep^2+\dots$$
we are in the position to calculate the first three coefficients $F_i(n)$ of the expansion~\eqref{Equ:DoubleInt}.
In order to calculate this integral within Mathematica, several packages have to be load in: the \texttt{Sigma}\ package to use the recurrence solver from Section~\ref{Sec:LaurentSolver}, and the {\tt EvaluateMultiSum} package to deal with expansions. Finally, we load in the package {\tt MultiIntegrate} which contains an efficient implementation of the multivariate Almkquist--Zeilberger algorithm with the help of homomorphic image testing; see~\cite{Ablinger:12}. In addition, it combines all the steps described above (together with variations of the presented method) to perform the $\ep$-expansion.
\begin{mma}
\In << Sigma.m \vspace*{-0.06cm}\\
\Print Sigma - A summation package by Carsten Schneider
\copyright\ RISC\\
\In << EvaluateMultiSums.m \vspace*{-0.06cm}\\
\Print EvaluateMultiSums by Carsten Schneider -- \copyright\ RISC\\
\In << MultiIntegrate.m \vspace*{-0.06cm}\\
\Print MultiIntegrate by Jakob Ablinger -- \copyright\ RISC\\
\end{mma}
\noindent Now we are ready to calculate the expansion of the integral above. The coefficients are returned in list form, i.e., $\{F_0(n),F_1(n),F_2(n)\}$:
\begin{mma}
\In {\text{\bf sol=mAZExpandedIntegrate$\Big[\frac{(1 + x_1\; x_2)^n}{(1 + x_1)^\ep}$, n, \{$\ep$, 0, 2\}, \{\{$x_1$, 0, 1\}, \{$x_2$, 0, 1\}\}$\Big]$}}\vspace*{-0.3cm}\\
\Out \big\{\frac{1}{(n+1)}\Big(2 \sum_{i_1=1}^n
\tfrac{2^{i_1}}{1+i_1}-\sum_{i_1=1}^n \tfrac{1}{1+i_1}+1\Big),\frac{1}{(n+1)}\Big(1-2 \text{ln2}\,\Big(\sum_{i_1=1}^n
\tfrac{2^{i_1}}{1+i_1}+1\Big)-\sum_{i_1=1}^n
\tfrac{1}{\big(1+i_1\big){}^2}+2 \sum_{i_1=1}^n
\tfrac{2^{i_1}}{\big(1+i_1\big){}^2}\Big),\newline
\frac{1}{(n+1)}\Big(\text{ln2}^2\,
\Big(\sum_{i_1=1}^n \tfrac{2^{i_1}}{1+i_1}+1\Big)-2 \text{ln2}\,
\Big(\sum_{i_1=1}^n
\tfrac{2^{i_1}}{\big(1+i_1\big){}^2}+1\Big)-\sum_{i_1=1}^n
\tfrac{1}{\big(1+i_1\big){}^3}+2 \sum_{i_1=1}^n
\tfrac{2^{i_1}}{\big(1+i_1\big){}^3}+1\Big)\big\}\\
\end{mma}
\noindent Note that the involved sums can be rewritten in terms of $S$-sums, see~\eqref{Equ:SSums}, using the command \texttt{TransformToSSums} from the package \texttt{HarmonicSums}:
\begin{mma}
\In << HarmonicSums.m \vspace*{-0.06cm}\\
\Print HarmonicSums by Jakob Ablinger -- \copyright\ RISC\\
\In TransformToSSums[sol]\vspace*{-0.3cm}\\
\Out \big\{\frac{S_1(2,n)}{n+1}-\frac{S_1(n)}{n+1}+\frac{2^{n+1}}{(n+1)^2}-\
\frac{1}{(n+1)^2},\text{ln2}\,
\Big(\frac{-S_1(2,n)}{n+1}-\frac{2^{n+1}}{(n+1)^2}\Big)+\frac{S_2(2,n)}{n+1}-\frac{S_2(n)}{n+1}+\frac{(2^{n+1}-1)}{(n+1)^3},\newline
\text{ln2}^2\, \Big(\frac{S_1(2,n)}{2 (n+1)}+\frac{2^n}{(n+1)^2}\Big)+\text{ln2}
\,\Big(-\frac{S_2(2,n)}{n+1}-\frac{2^{n+1}}{(n+1)^3}\Big)+\frac{S_3(2,n)
}{n+1}-\frac{S_3(n)}{n+1}+\frac{2^{n+1}}{(n+1)^4}-\frac{1}{(n+1)^4}\big\}\\
\end{mma}
\noindent The proposed method extends to integrals with higher nesting depth. We conclude this approach by the calculation the coefficients $F_0(n),F_1(n),F_2(n)$ of the $\ep$-expansion of the following triple integral:
$$\int_0^1\int_0^1\int_0^1\frac{x_1^{\ep}(x_1+x_2x_3)^n}{
(1+x_3)^{\ep}}dx_1dx_2dx_3=F_0(n)+F_1(n)\ep+F_2(n)\ep^2+\dots$$
Integrals of this type emerge as partial integrals, e.g., for 3-loop ladder
topologies.
\begin{mma}
\In mAZExpandedIntegrate[\tfrac{x_1^{\ep}(x_1+x_2x_3)^n}{ (1+x_3)^{\ep}},n,\{\ep,0,2\},\vspace*{-0.cm}\newline
\hspace*{1cm}\{\{x_1,0,1\},\{x_2,0,1\},\{x_3,0,1\}\}]//TransformToSSums//ReduceToBasis\vspace*{-0.4cm}\\
\Out \big\{\tfrac{-3 n-4}{(n+1)^2 (n+2)^2}+\tfrac{2^{n+1} (3 n+4)}{(n+1)^2 (n+2)^2}-\tfrac{S_1(n)}{(n+1) (n+2)}+\tfrac{S_1(2,n)}{(n+1) (n+2)},\quad
-\tfrac{2^{n+1} \big(3 n^2+6 n+2\big)}{(n+1)^3 (n+2)^3}+\tfrac{4 n^2+9 n+4}{(n+1)^3 (n+2)^3}+\tfrac{2 S_1(n)}{(n+1) (n+2)^2}-\tfrac{S_2(n)}{(n+1) (n+2)}-\tfrac{(-1)^n S_{-1}(n)+2 S_1(2,n)}{(n+1) (n+2)^2}+\text{ln2} \big(-\tfrac{2^{n+1} (3 n+4)}{(n+1)^2 (n+2)^2}-\tfrac{S_1(2,n)}{(n+1) (n+2)}-\tfrac{(-1)^n}{(n+1) (n+2)^2}+\tfrac{1}{(n+1) (n+2)^2}\big)+\tfrac{S_2(2,n)}{(n+1) (n+2)},
\newline
\big(\tfrac{2^n (3 n+4)}{(n+1)^2 (n+2)^2}+\tfrac{S_1(2,n)}{2 (n+1) (n+2)}+\tfrac{(-1)^n}{2 (n+1) (n+2)^2}-\tfrac{1}{2 (n+1) (n+2)^2}\big) \text{ln2}^2+\big(\tfrac{-3 n-5}{(n+1)^2 (n+2)^3}
\tfrac{2^{n+1} \big(3 n^2+6 n+2\big)}{(n+1)^3 (n+2)^3}-\tfrac{S_1(n)}{(n+1) (n+2)^2}+\vspace*{-0.3cm}\newline
+(-1)^n \big(\tfrac{S_{-1}(n)}{(n+1) (n+2)^2}+\tfrac{1}{(n+1) (n+2)^3}\big)+\tfrac{2 S_1(2,n)}{(n+1) (n+2)^2}-\tfrac{S_2(2,n)}{(n+1) (n+2)}\big) \text{ln2}-\tfrac{S_1(n){}^2}{2 (n+1) (n+2)^2}-\tfrac{(3 n+4) \big(3 n^2+9 n+7\big)}{(n+1)^4 (n+2)^4}\newline
+\tfrac{2^{n+1} \big(7 n^3+30 n^2+44 n+22\big)}{(n+1)^4 (n+2)^4}-\tfrac{2 (2 n+3) S_1(n)}{(n+1)^2 (n+2)^3}+\tfrac{S_2(n)-S_2(2,n)}{2 (n+1) (n+2)^2}-\tfrac{S_3(n)}{(n+1) (n+2)}+\tfrac{2 (2 n+3) S_1(2,n)}{(n+1)^2 (n+2)^3}+\tfrac{S_3(2,n)}{(n+1) (n+2)}\newline
+\tfrac{S_{1,1}(1,2,n)}{(n+1) (n+2)^2}+(-1)^n \big(\tfrac{S_{-1}(n)}{(n+1) (n+2)^3}+\tfrac{2 S_{-2}(n)}{(n+1) (n+2)^2}+\tfrac{S_{-1,1}(n)}{(n+1) (n+2)^2}-\tfrac{S_2(-2,n)}{(n+1) (n+2)^2}-\tfrac{S_{1,1}(-1,2,n)}{(n+1) (n+2)^2}\big)\big\}\\
\end{mma}
\noindent Here the command \texttt{ReduceToBasis} ensures that no algebraic relations exist among the arising $S$-sums; see also~\cite{Bluemlein:04,ANCONT3,ABC:11,Ablinger:12,ABC:12}.
\section{Calculating $\ep$-expansions for multi-sums}\label{Sec:Sum}
In our second strategy we rely on a common summation framework~\cite{Schneider:05b} of the difference field and holonomic approach~\cite{ZEILB,Chyzak}. The input class of the proposed method covers multi-sums of the form~\eqref{Eq:GenericMultiSum}. More generally, the summand itself can be an expression in terms of indefinite nested sums and products.
To illustrate this approach we aim at calculating the first coefficients of the double sum
$$\mathcal{S}(\ep,n)=\sum_{k=0}^{n-3}\underbrace{\sum_{j=1}^{n-k-2} \tfrac{(-1)^j \binom{-k+n-2}{j}\Gamma (j+k+1) \Gamma (n-k)
\big(1-\frac{\ep}{2}\big)_k \big(2-\frac{\ep}{2}\big)_j}{\
\Gamma (-k+n-1) (3-\ep)_{j+k} \big(\frac{\ep}{2}+3\big)_{j+k}}}_{F(n,k)}
\stackrel{?}{=}F_0(n)+F_1(n)\ep+F_2(n)\ep^2+\dots$$
First one considers the inner sum $F(n,k)$ which itself is an analytic function in $\ep$ for each integer $n$ and $k$ with $n\geq3$ and $0\leq k\leq n-3$. First, we hunt for a recurrence relation in $k$. As described in Section~\ref{Sec:LaurentSolver} one can use the package Sigma and calculates
\begin{multline}\label{Equ:PureRec}
(k+1) (\ep-2 k-2)(k-n+2)F(n,k)\\
-(k-n+1) \big(\ep^2-\ep k+\ep n-\ep-4 k^2+2 k n-14 k-14\big)F(n,k+1)\\
-2 (\ep+k+2)(k-n+1) (k-n+2)F(n,k+2)=e_0(n,k)+e_1(n,k)\ep+e_2(n,k)\ep^2+\dots
\end{multline}
with
\scriptsize
\begin{align*}
e_0&(n,k)=-\frac{16 (k-n+1) (k-n+2) \big(k^2 n+4 k^2+2 k n+22 k-4 n+28\big)}{(k+2)^2 (k+3)^2 (k+4)^2},\\
e_1&(n,k)=\frac{8(k-n+1) (k-n+2) \big(2 k^5 n+10 k^5+18 k^4 n+134 k^4+43 k^3 n+698 k^3-32 k^2 n+1760 k^2-220 k n+2140 k-184 n+1000\big)}{(k+2)^3 (k+3)^3 (k+4)^3},\\
e_2&(n,k)=
-\frac{1}{2}(k+1) k! (k-n+1) (k-n+2) \big(\frac{16 \big(k^2 n+4 k^2+2 k n+22 k-4 n+28\big) S_2(k)}{(k+1) (k+2)^2 (k+3)^2 (k+4)^2 k!}\\
&-2 \Big(5 k^{10} n+4 k^{10}+142 k^9 n+230 k^9+1602 k^8 n+3812 k^8+9552 k^7 n+31332 k^7+32861 k^6 n\\
&\quad+150820 k^6+64970 k^5 n+452886 k^5+64724 k^4 n+857364 k^4+11424 k^3 n+985264 k^3-29328 k^2 n\\
&\quad+606320 k^2-11712 k n+121344 k+5952 n-30144\Big)\Big/({(k+1)^3 (k+2)^4 (k+3)^4 (k+4)^4 k!})\big).
\end{align*}
\normalsize
In addition, one computes a mixed recurrence, i.e., besides shifts in $k$ we allow in addition one shift in $n$, but keep $k$ unchanged. Using again Sigma, one obtains
\begin{multline}\label{Equ:OnePureRec}
(k-n) \big(-\ep^2+\ep k-\ep n+2 k^2-2 k n+2 k+2 n^2+2 n+2\big)F(n,k)\\
+(\ep-n-1) (\ep+2 n+2) (k-n+1) F(n+1,k)\\
-2 (\ep+k+1)(k-n) (k-n+1)F(n,k+1)=f_0(n,k)+f_1(n,k)\ep+f_2(n,k)\ep^2+\dots
\end{multline}
with
\scriptsize
\begin{align*}
f_0&(n,k)=\frac{16 (k-n) (k-n+1) \big(k^2-2 k n+2 k-4 n-1\big)}{(k+1) (k+2)^2 (k+3)^2},\\
f_1&(n,k)=-\frac{8(k-n) (k-n+1) \big(2 k^5-4 k^4 n+12 k^4-26 k^3 n+17 k^3-56 k^2 n-17 k^2-42 k n-47 k-4 n-19\big)}{(k+1)^2 (k+2)^3 (k+3)^3},\\
f_2&(n,k)=-\frac{1}{2} (k+1) k! (k-n) (k-n+1) \Big(-\frac{16 \big(k^2-2 k n+2 k-4 n-1\big) S_2(k)}{(k+1)^2 (k+2)^2 (k+3)^2 k!}\\
&\;\;+2 \Big(5 k^8-10 k^7 n+102 k^7-188 k^6 n+745 k^6-1300 k^5 n+2644 k^5-4456 k^4 n+4855 k^4-8162 k^3 n\\
&\;\;+4150 k^3-7644 k^2 n+635 k^2-2816 k n-944 k+96 n-192\Big)\Big/\Big((k+1)^4 (k+2)^4 (k+3)^4 k!\Big)\Big).
\end{align*}
\normalsize
We emphasize
that these two recurrences together with the initial values $F(4,0)=-\frac{13}{12}+\frac{23\ep}{288}-\frac{29\ep^2}{768}\dots$ and $F(4,1)=-\frac{1}{18}+\frac{11\ep}{432}-\frac{11\ep^2}{3456}
+\dots$ enables one to calculate the first three coefficients of the $\ep$-expansion for each $F(n,k)$ with $n\geq4$ and $0\leq k\leq n-3$.
Given the two recurrences above (with this particular shape of shifts), one can apply the algorithm from~\cite{Schneider:05b} to calculate a recurrence for $\mathcal{S}(\ep,n)$. Using again Sigma, one gets the relation
\begin{multline}\label{Equ:HoloFinalRec}
\ep (\ep+2)^2 (2 n+1) (\ep-n-3) (\ep-n-2) (\ep+2 n+2) (\ep-n-1)^3 \mathcal{S}(\ep,n+1)\\
+4 \ep(\ep+2) (n+1)^3 (n+3) (2 n+3) (\ep-n-2) (\ep-n-1)\mathcal{S}(\ep,n)\\
=h_0(n)+h_1(n)\ep+h_2(n)\ep^2+h_3(n)F(n,0)\\
+h_4(n)F(n,1)+h_5(n)F(n,n-3)+h_6(n)F(n,n-2)+\dots
\end{multline}
\vspace*{-0.3cm}
\noindent with
\vspace*{-0.2cm}
\scriptsize
\begin{align*}
h_0(n)=&-\frac{64 \big(15 n^9+5 n^8-188 n^7-194 n^6+655 n^5+1129 n^4-266 n^3-1660 n^2-1224 n-288\big)}{9 (n-1)^2 n^2},\\
h_1(n)=&=\tfrac{32\big(15 n^{13}+87 n^{12}-45 n^{11}-329 n^{10}-215 n^9+249 n^8+2331 n^7+1545 n^6-6396 n^5-5440 n^4+6686 n^3+7776 n^2+648 n-864\big)}{27 (n-1)^3 n^3},\\
h_2(n) =&\frac{16 \big(4 n^{12}+16 n^{11}-23 n^{10}-188 n^9-219 n^8+208 n^7+503 n^6+12 n^5-445 n^4-96 n^3+340 n^2+272 n+64\big) S_2(n)}{(n-1)^2 n^2}\\
&\quad-\Big(4(708 n^{16}+916 n^{15}-9426 n^{14}-15711 n^{13}+30903 n^{12}+61829 n^{11}-49101 n^{10}-115105 n^9+44049 n^8\\
&\quad\quad+121347 n^7-789 n^6-57380 n^5-41832 n^4-14040 n^3+16848 n^2+6048 n-3456\big)\Big)\Big/\big(27 (n-1)^4 n^4\big),\\
h_3(n)=& \big(-\tfrac{4 \ep^2 \big(2 n^6+3 n^5-14 n^4-9 n^3-49 n^2-115 n-62\big) (n+1)^2}{n-1}+\tfrac{8 \ep \big(2 n^4+10 n^3-15 n^2-30 n+8\big) (n+1)^3}{n-1}-\tfrac{32 \big(3 n^2+10 n+8\big) (n+1)^3}{n-1}\big),\\
h_4(n)=&-32 (3 n+4) (n+1)^3+8\big(2 n^3-6 n^2-25 n-12\big) (n+1)^3\ep-4\big(2 n^5-3 n^4+16 n^3+32 n^2-18 n-31\big)(n+1)^2\ep^2,\\
h_5(n)=&16 (n-2)^2 \big(3 n^3+n^2-7 n-4\big) (n+1)^3-4\big(2 n^7-15 n^6+46 n^5-18 n^4-125 n^3+86 n^2+96 n-16\big) (n+1)^2\ep\\
&\quad+2\big(16 n^8-58 n^7-7 n^6+221 n^5-59 n^4-325 n^3-26 n^2+146 n+60\big)\ep^2,\\
h_6(n)=&32 \big(3 n^4-8 n^3-10 n^2+20 n+16\big)(n+1)^3-8\big(2 n^6-13 n^5+27 n^4+56 n^3+14 n^2+66 n+60\big)(n+1)^2\ep\\
&\quad+8\big(6 n^7-23 n^6-43 n^5+46 n^4+68 n^3-23 n^2-43 n-12\big)\ep^2;
\end{align*}
\normalsize
for the different command calls within \texttt{Sigma}\ we refer to~\cite{Schneider:05b,Schneider:07a}.
As indicated above, the algorithm itself only uses the two recurrence relations~\eqref{Equ:PureRec} and~\eqref{Equ:OnePureRec} and thus the occurring expressions $F(n,0),F(n,1),F(n,n-3),F(n,n-2)$ remain unevaluated.
Next, one applies the proposed method recursive on these sums.
Since these objects are simpler than the input sum $\mathcal{S}(\ep,n)$, the termination of our method is guaranteed. E.g., the calculation of the $\ep$-expansion
\begin{align*}
F(n,0)&=\sum_{j=1}^{n-2}\tfrac{(-1)^j \Gamma (j+1) \Gamma (n)
\binom{n-2}{j}
\left(2-\frac{\ep}{2}\right)_j}{\Gamma (n-1) (3-\ep)_j
\left(\frac{\ep}{2}+3\right)_j}\\
&=4 (n-1) \big(\tfrac{5-n}{4 (n-1)}-\tfrac{S_1(n)}{(n-1) n}\big)+ (1-n) \big(\tfrac{S_1(n){}^2}{(n-1) n}-\tfrac{S_1(n)}{(n-1) n}+\tfrac{S_2(n)}{(n-1) n}+\tfrac{1}{1-n}\big)\ep\\
&+\tfrac{n-1}{6}\big(-\tfrac{S_1(n){}^3}{(n-1) n}+\tfrac{3 S_1(n){}^2}{2 (n-1) n}+\big(\tfrac{15}{(n-1) n}
-\tfrac{3 S_2(n)}{(n-1) n}\big) S_1(n)+\tfrac{3 S_2(n)}{2 (n-1) n}-\tfrac{2 S_3(n)}{(n-1) n}-\tfrac{12 S_{2,1}(n)}{(n-1) n}\big)\ep^2+\dots
\end{align*}
boils down to the method described in Section~\ref{Sec:LaurentSolver}.
Similarly one proceeds for $F(n,1)$ and $F(n,n-3),F(n,n-2)$. This finally leads to the following simplified right and side of~\eqref{Equ:HoloFinalRec}:
\small
\begin{align*}
&0\times\ep^0 + \Big(32 n (n+1)^2 (3 n+4) \big(n^2+3 n+4\big)-128 (n+1)^3 (3 n+4) S_1(n)\Big)\ep\\
&+\Big(-32 \big(2 n^2+7 n-2\big) (n+1)^3 S_1(n)+16 \big(4 n^5+12 n^4-27 n^3-129 n^2-130 n-28\big) (n+1)^3 S_2(n)\\
&-8 n \big(12 n^6+58 n^5-27 n^4-566 n^3-1125 n^2-804 n-156\big) (n+1)-32 (3 n+4) (n+1)^3 S_1(n){}^2\Big)\ep^2+\dots
\end{align*}
\normalsize
Note that the left hand side of~\eqref{Equ:HoloFinalRec} and its right hand side (after the simplification) can be divided by $\ep$, i.e., the coefficient of $\mathcal{S}(\ep,n+1)$ evaluated at $\ep=0$ does not vanish. Hence together with the expanded initial values
$\mathcal{S}(\ep,3)= -\frac{4}{9}+\frac{1}{27}\ep+\dots$ and $\mathcal{S}(\ep,4)=-\frac{41}{36} +\frac{91}{864}\ep+\dots$ one can activate our recurrence solver for Laurent series to calculate the first two coefficients of the $\ep$-expansion
\begin{align*}
\mathcal{S}(\ep,n)=&-\tfrac{8}{(n+1) (n+2)}S_1(n)-4 (2 n+1) S_2(n)+\tfrac{4 n (3 n+7)}{n+2}
+\Big(\tfrac{-2 \big(2 n^3+3 n^2+3 n+6\big)}{(n+1)^2 (n+2)^2}S_1(n)\\
&+\tfrac{\big(6 n^3+23 n^2+27 n+12\big)}{(n+1) (n+2)}S_2(n)
-\tfrac{2}{(n+1) (n+2)}S_1(n){}^2-2 (2 n+1) S_3(n)-\tfrac{n (n+1) (5 n+14)}{(n+2)^2}\Big)\ep+\dots
\end{align*}
\noindent Summarizing, in the presented method one constructs step by step suitable inhomogeneous recurrences from the innermost sum to the outermost sum\footnote{
In~\cite{Chyzak} this idea has been considered for homogeneous recurrences with polynomial coefficients. In our approach~\cite{Schneider:05b} we observed that setting up the recurrence system in the special form given above (instead of allowing a general holonomic system) one can derive an efficient algorithm without using Gr\"obner basis. In this way, the holonomic approach could be extended in~\cite{Schneider:05b} to handle also inhomogeneous recurrences formulated in difference fields. In order to take into account the $\ep$-expansion of the inhomogeneous sides, new ideas have been added into \texttt{Sigma}.}.
As one can see already for double sums, this construction is quite involved and is fairly complicated for more nested sums (e.g., taking care of poles, estimating how far one should expand\footnote{E.g., in the illustrated example from above one has to start to calculate three coefficients of the $\ep$-expansion and ends up only with the first two coefficients.}, or exploiting a refined difference field theory~\cite{Schneider:08c}).
The new package \texttt{RhoSum} deals with all these aspects using as backbone the packages \texttt{Sigma}, \texttt{HarmonicSums}, and \texttt{EvaluateMultiSums}. After loading
\begin{mma}
\In << RhoSum.m \vspace*{-0.06cm}\\
\Print RhoSum - Package for Refined Holonomic Summation
\copyright\ RISC\\
\end{mma}
\noindent we can perform the calculation from above with the function call
\begin{mma}
\In FindSum[\tfrac{(-1)^j \binom{-k+n-2}{j}\text{$\Gamma$}(j+k+1) \text{$\Gamma$}(n-k)
\big(1-\frac{\ep}{2}\big)_k \big(2-\frac{\ep}{2}\big)_j}{
\text{$\Gamma$}(-k+n-1) (3-\ep)_{j+k} \big(\frac{\ep}{2}+3\big)_{j+k}}
,\{\{j,1,n-k-2\},\{k,0,n-3\}\},\{n\},\{3\},\{\infty\},\vspace*{-0.2cm}\newline
\hspace*{10cm}ExpandIn\to\{ep,0,1\}\vspace*{-0.5cm}\\
\Out\big\{-\tfrac{8}{(n+1) (n+2)}S_1(n)-4 (2 n+1) S_2(n)+\tfrac{4 n (3 n+7)}{n+2},\newline
\tfrac{-2 \big(2 n^3+3 n^2+3 n+6\big)}{(n+1)^2 (n+2)^2}S_1(n)
+\tfrac{\big(6 n^3+23 n^2+27 n+12\big)}{(n+1) (n+2)}S_2(n)
-\tfrac{2}{(n+1) (n+2)}S_1(n){}^2-2 (2 n+1) S_3(n)-\tfrac{n (n+1) (5 n+14)}{(n+2)^2}\big\}\\
\end{mma}
\noindent A more involved problem is, e.g.,
\begin{multline*}
\sum_{j=0}^{n-2}\sum_{j_1=0}^j\sum_{j_2=1}^{n-4+j_1-j}
\tfrac{(-1)^{-{j_1}-{j_2}+n-4} e^{-\frac{3 \ep \gamma }{2}} \binom{j}{{j_1}} \binom{-j+{j_1}+n-4}{{j_2}}\Gamma\big(-\frac{\ep}{2}-j+{j_1}-{j_2}+n-1\big)\Gamma(-\ep-j+n-1) \Gamma\big(-\frac{\ep}{2}-j+{j_1}+n\big)}{(\ep+1) (\ep+2) \Gamma\big(2-\frac{\ep}{2}\big) \Gamma(\ep+4) \Gamma(j-{j_1}+{j_2}+2) \Gamma(-2 \ep-j+{j_1}+n)}\times\\
\times\tfrac{\Gamma\big(\frac{\ep}{2}+3\big)^2 \Gamma\big(-\frac{3 \ep}{2}\big) \Gamma\big(-\frac{\ep}{2}\big) \Gamma(\ep) \Gamma(j-{j_1}+1) \Gamma(-\ep+{j_1}+1) \Gamma({j_2}+1)}{\Gamma\big(\frac{\ep}{2}-j+{j_1}+n\big) \Gamma\big(-\frac{\ep}{2}-j+{j_1}-{j_2}+n\big)}\stackrel{?}{=}F_{-3}\ep^{-3}+F_{-2}\ep^{-2}+F_{-1}\ep^{-1}+F_{0}+\dots
\end{multline*}
Sums of this type occur, e.g., in case of 3-loop topologies with one massive and one
massless fermion line.
If we insert this sum into Mathematica in the variable $f$, then we get the following expansion. The constant term is too large to present it here.
\begin{mma}
\In FindSum[f
,\{\{j_2,1,n-4+j_1-j\},\{j_1,0,j\},\{j,0,n-2\}\},\{n\},\{2\},\{\infty\},ExpandIn\to\{ep,-3,-1\}\\
\Out \big\{-\tfrac{4 \big(n^3-5 n^2+6 n-4\big)}{9 (n-2)^2 (n-1) n^3}+\tfrac{2 (-1)^n \big(n^4-2 n^3+n^2-12 n+8\big)}{9 (n-2) (n-1) n^3}-\tfrac{8 (-1)^n (n-3) S_ 1(n)}{9 (n-2) n^2},\newline
+\tfrac{16 (-1)^n S_ {-2}(n)}{9 (n-2) n^2}+ \tfrac{(-1)^n (10-n) S_ 1(n){}^2}{9 (n-2) n^2}+\big(\tfrac{(-1)^n \big(17 n^3-179 n^2+240 n-96\big)}{27 (n-2) (n-1) n^3}+\tfrac{4}{9 (n-2)^2 (n-1) n^2}\big) S_ 1(n)\newline
+\tfrac{(-1)^n \big(-2 n^5+19 n^4-5 n^3+111 n^2-190 n+84\big)}{27 (n-2) (n-1) n^4}+\tfrac{2 \big(8 n^5-63 n^4+163 n^3-218 n^2+140 n-36\big)}{27 (n-2)^2 (n-1)^2 n^4}+\tfrac{(-1)^n (14-n) S_ 2(n)}{9 (n-2) n^2},\newline
\tfrac{(-1)^n (22-n) S_ 1(n){}^3}{54 (n-2) n^2}+\big(\tfrac{(-1)^n \big(n^3-299 n^2+412 n-168\big)}{108 (n-2) (n-1) n^3}+\tfrac{1}{9 (n-2)^2 (n-1) n^2}\big) S_ 1(n){}^2\newline
+\big(\tfrac{-6 n^2+5 n-5}{27 (n-2)^2 (n-1)^2 n^2}+\tfrac{(-1)^n \big(-313 n^4+1708 n^3-2262 n^2+1740 n-648\big)}{162 (n-2) (n-1) n^4}+\tfrac{(-1)^n (7 n+46) S_ 2(n)}{18 (n-2) n^2}\big) S_ 1(n)\newline
+\tfrac{-289 n^6+2430 n^5-6803 n^4+10222 n^3-8536 n^2+4176 n-1008}{324 (n-2)^2 (n-1)^2 n^5}\newline
+\tfrac{(-1)^n \big(475 n^7-1768 n^6+2423 n^5-6170 n^4+17912 n^3-21856 n^2+15744 n-5184\big)}{648 (n-2)^2 (n-1) n^5}-\tfrac{2 (-1)^n (11 n-18) S_ {2,1}(n)}{9 (n-2) n^2}\newline
+S_ {-2}(n) \big(\tfrac{16 (-1)^n S_ 1(n)}{9 (n-2) n}-\tfrac{4 (-1)^n \big(9 n^3+22 n^2-43 n+18\big)}{27 (n-2) (n-1) n^3}\big)-\tfrac{2 (-1)^n (3 n-16) S_ {-3}(n)}{9 (n-2) n^2}\newline
+\big(\tfrac{-6 n^3+30 n^2-35 n+24}{9 (n-2)^2 (n-1) n^3}+\tfrac{(-1)^n \big(n^4-641 n^3+1646 n^2-976 n+192\big)}{108 (n-2)^2 (n-1) n^3}\big) S_ 2(n)+\tfrac{(-1)^n (47 n-8) S_ 3(n)}{27 (n-2) n^2}
\newline
-\tfrac{4 (-1)^n S_ {-2,1}(n)}{3 (n-2) n}+\big(\tfrac{-n^3+5 n^2-6 n+4}{6 (n-2)^2 (n-1) n^3}+\tfrac{(-1)^n \big(n^4-2 n^3+n^2-12 n+8\big)}{12 (n-2) (n-1) n^3}+\tfrac{(-1)^n (3-n) S_ 1(n)}{3 (n-2) n^2}\big) \zeta_2\Big\}\\
\end{mma}
\noindent The above expressions can be analytically continued to complex values of the Mellin
variable $n$ using relations given in \cite{Blumlein:2000hw,Blumlein:2005jg,
Blumlein:2009ta,ANCONT3}.
\section{Conclusion}
Massive Feynman integrals with operation insertion can be expressed in terms of multi-in\-tegrals and multi-sums over hypergeometric and hyperexponential functions.
We presented new methods to calculate the first coefficients of the $\ep$-expansion of such multi-sums and multi-integrals. Here the multivariate Almkvist-Zeilberger algorithm and the common framework of the holonomic and difference field algorithms have been enhanced to calculate recurrences. Then a recurrence solver for Laurent series expansion is used to extract the all $n$ coefficients of the $\ep$-expansion. Besides of the usage of the Mathematica packages \texttt{Sigma}, \texttt{HarmonicSums} and \texttt{EvaluateMultiSums}, two new packages \texttt{MultiIntegrate} and \texttt{Rho} have been developed that can carry out these calculations in a completely automatic fashion.
\vspace*{2mm}
\noindent
{\bf Acknowledgment.}
This work has been supported in part by DFG Sonderforschungsbereich Transregio 9,
Computergest\"utzte Theoretische Teilchenphysik, Austrian Science Fund (FWF)
grant P203477-N18, and EU Network {\sf LHCPhenoNet} PITN-GA-2010-264564.
|
\section{Introduction}
\label{sec:introduction}
Computationally hard problems can be found in almost every area of computer
science, hence the quest for general tools and methods that help designing
solutions to such problems
is one of the central research challenges in our field.
One particularly successful
approach
is
Answer-Set Programming~\cite{niem-99,mare-trus-99,BrewkaET11}%
---ASP, for short---%
where
highly sophisticated solvers \cite{dlv,GebserKKOSS11} provide a rich declarative
language
to specify the given problem in an intuitive and succinct manner.
On the other hand, the concept of dynamic programming \cite{Larson,Wagner}
denotes a general recursive
strategy where an optimal solution to a problem is defined in terms of optimal solutions to its
subproblems, thus
constructing a solution ``bottom-up'', from simpler to more complex problems.
One particular application of dynamic programming is the design of algorithms
which proceed along tree decompositions~\cite{DBLP:journals/jct/RobertsonS84}
of the problem at hand, or rather of its graph representation.
The significance of this approach is highlighted
by Courcelle's seminal result (see, e.g., \cite{DBLP:books/el/leeuwen90/Courcelle90}) which
states that every problem definable in monadic second-order logic
can be solved in linear time on structures of bounded treewidth;
formal definitions of these concepts will be provided in
Section~\ref{sub:treedecompositions}.
This suggests a two-phased methodology for problem solving, where first
a tree decomposition of the given problem instance is obtained
and subsequently used in a second phase to solve the problem
by a specifically designed algorithm (usually employing dynamic programming) traversing the tree decomposition;
see, e.g., \cite{BodlaenderK08} for an overview of this approach.
Such tree decomposition based algorithms have
been successful in several applications including
constraint satisfaction problems such as \prob{Max-Sat}~\cite{Koster02},
and also bio-informatics~\cite{Xu05,CaiHLRS08}.
While good heuristics to obtain tree decompositions exist \cite{Dermaku08,Bodlaender10} and
implementations thereof are available,
the actual implementation of dynamic programming algorithms that work on
tree decompositions had, as yet, often to be done from scratch.
In this paper, we present a new
method for declaratively
specifying the dynamic programming part, namely by means of
ASP programs. As mentioned, dynamic programming relies on the evaluation of subproblems which
themselves are often combinatorial in nature. Thanks to the Guess \& Check principle
of the ASP paradigm, using ASP to describe the treatment of subproblems is thus
a natural choice and also separates our approach from existing systems (which
we will discuss in the conclusion section of this paper). Using ASP as a
language in order to describe the constituent elements of a dynamic programming
algorithm obviously suggests to also employ sophisticated
off-the-shelf ASP systems
as internal machinery for evaluating the specified algorithm. Thus, our approach
not only takes full advantage of the rich syntax ASP offers to describe
the dynamic programming algorithm, but also delegates
the burden of local computations to highly efficient ASP systems.
We have implemented our approach in the
novel
system D-FLAT\footnote{%
(D)ynamic Programming (F)ramework with (L)ocal Execution of (A)SP on
(T)ree Decompositions; available at
\url{http://www.dbai.tuwien.ac.at/research/project/dynasp/dflat/}.%
}
which takes care of the computation of tree decompositions
and also
handles the ``bottom-up'' data flow of the dynamic programming algorithm.\footnote
Concerning the decomposition, we employ
the htdecomp library of \cite{Dermaku08}
making our system amenable to hypertree decompositions, as well.
This, in fact, allows to decompose arbitrary finite structures thus
extending the range of applicability for our system.}
All that is required of users is an idea for such an algorithm.
Hence, D-FLAT serves as a tool
for rapid prototyping of dynamic programming algorithms but can also
be considered for educational purposes. As well, ASP users are provided with
an easy-to-use interface to decompose problem instances---an issue which might
allow large instances of practical importance to be solved, which so far could
not be handled by ASP systems. To summarize, D-FLAT provides a new
method for problem solving, combining the
advantages
of
ASP and tree decomposition based dynamic programming.
Some attempts to ease the specification of dynamic programming algorithms
already exist (see Section~\ref{sec:conclusion}), but we are not aware
of any other system that employs ASP for dynamic programming on tree
decompositions.
The structure of this paper is as follows: In Section~\ref{sec:preliminaries}, we first briefly
recall logic programming under the answer-set semantics and
provide
the necessary background for (hyper)tree decompositions and dynamic programming. In Section~\ref{sec:examples}, we introduce the features of D-FLAT step-by-step, providing ASP specifications of several
standard (but also some novel) dynamic programming algorithms.
Section~\ref{sec:systemdescription} gives some
system specifics, and Section~\ref{sec:experiments}
reports on a preliminary experimental evaluation.
In the conclusion of the paper, we address related
work and give a brief summary and an outlook.
\section{Preliminaries}
\label{sec:preliminaries}
\subsection{Logic Programs and Answer-Set Semantics}
\label{sub:logicprogramsandanswersetsemantics}
The proposed system D-FLAT uses ASP to specify a dynamic programming algorithm.
In this section, we will therefore briefly introduce the syntax and semantics of
ASP. The reader is referred to \cite{BrewkaET11} for a more thorough
introduction.
A \emph{logic program} $\Pi$ consists of a set of \emph{rules} of the form
\[a_1 \lor \dots \lor a_k \ensuremath{\longleftarrow} b_1, \dots, b_m, \ensuremath{\operatorname{not}} b_{m+1}, \dots, \ensuremath{\operatorname{not}}
b_n.\]
We call $a_1, \dots, a_k$ and $b_1, \dots, b_n$ \emph{atoms}. A \emph{literal} is either an
atom or its default negation which is obtained by putting $\ensuremath{\operatorname{not}}$ in front.
For a rule $r \in \Pi$, we call $h(r) = \{a_1, \dots, a_k\}$ its \emph{head} and $b(r)
= \{b_1, \dots, b_n\}$ its \emph{body} which is further divided into a \emph{positive body},
$b^+(r) = \{b_1, \dots, b_m\}$, and a \emph{negative body}, $b^-(r) = \{b_{m+1}, \dots,
b_n\}$.
Note that the head may be empty, in which case we call $r$ an \emph{integrity
constraint}.
If the body is empty, $r$ is called a \emph{fact}, and the $\ensuremath{\longleftarrow}$ symbol can be
omitted.
A rule $r$ is satisfied by a set of atoms
$I$ (called an \emph{interpretation})
iff $I \cap
h(r) \neq \emptyset$ or $b^-(r) \cap I \neq \emptyset$ or $b^+(r) \setminus I
\neq \emptyset$.
$I$ is a \emph{model} of a set of rules iff it satisfies each rule.
$I$ is an \emph{answer set} of a program $\Pi$ iff it is a subset-minimal model of
the program $\Pi^I =\{h(r) \ensuremath{\longleftarrow} b^+(r) \mid r \in \Pi, b^-(r) \cap I =
\emptyset\}$, called the
\emph{Gelfond-Lifschitz reduct} \cite{gelf-lifs-91} of $\Pi$ w.r.t.\ $I$.
In this paper, we use the input language of Gringo
\cite{clasp-guide}
for logic
programs. Atoms in this language are predicates whose arguments can be either
variables or ground terms. Such programs can be seen as abbreviations of
variable-free programs, where all the variables are instantiated (i.e., replaced
by ground terms). The process of instantiation is also called \emph{grounding}.
It results in a propositional program which can be represented as a set of rules
which adhere to the above definition.
As an example, an instance for the \prob{3-Col} problem\footnote{%
\prob{3-Col} is defined as the problem of deciding whether, for a given graph
$(V,E)$, there exists a 3-coloring, i.e., a mapping $f: V \to
\{\text{red},\text{green},\text{blue}\}$ s.t.\ for each edge $(a,b) \in E$ it
holds that $f(a) \neq f(b)$.}
(covered in Section~\ref{sec:graph-coloring}) consists of a set of vertices and
a set of edges, so the following set of facts represents a valid instance:
\begin{lstlisting}[numbers=none]
vertex(a). vertex(b). vertex(c). vertex(d). vertex(e).
edge(a,b). edge(a,c). edge(b,c). edge(b,d). edge(c,d). edge(d,e).
\end{lstlisting}
The following logic program solves \prob{3-Col} for instances of this form:
\begin{lstlisting}
color(red;green;blue).
1 { map(X,C) : color(C) } 1 :- vertex(X).
:- edge(X,Y), map(X,C), map(Y,C).
\end{lstlisting}
Solving \prob{3-Col} for an instance amounts to grounding this encoding together
with the instance as input and then solving the resulting ground program.
Line~1 of the encoding is expanded by the grounder to three facts which state
that ``red'', ``green'' and ``blue'' are colors. Line~2 contains a cardinality
constraint in the head and is conceptually expanded to
\begin{lstlisting}[numbers=none]
1 { map(X,red), map(X,green), map(X,blue) } 1 :- vertex(X).
\end{lstlisting}
before the grounder expands this rule by substituting ground terms for
\verb!X!.
Roughly speaking, a cardinality constraint $l \{ L_1, \dots, L_n \} u$ is
satisfied by an interpretation $I$ iff at least $l$ and at most $u$ of the literals $L_1, \dots, L_n$ are
true in $I$. Therefore, the rule in question expresses a choice of exactly one of
\verb!map(X,red)!, \verb!map(X,green)!, \verb!map(X,blue)!, for any vertex
\verb!X!. Finally, the integrity constraint in line~3 of the encoding ensures
that no answer set maps the same color to adjacent vertices.
For space reasons, we refer the reader to \cite{clasp-guide} for more details on
the input language.
\vspace{-2mm}
\subsection{Hypertree Decompositions}
\label{sub:treedecompositions}
Tree decompositions and treewidth, originally defined in
\cite{DBLP:journals/jct/RobertsonS84}, are a well known tool to tackle
computationally hard problems (see, e.g.,
\cite{DBLP:journals/actaC/Bodlaender93,DBLP:conf/sofsem/Bodlaender05} for an
overview). Informally, treewidth is a measure of the cyclicity of a graph and many
\textbf{NP}-hard problems become tractable if the treewidth is bounded.
The intuition behind tree decompositions is obtaining a tree from a (potentially
cyclic) graph by subsuming multiple vertices under one node and thereby
isolating the parts responsible for the cyclicity.
Several problems are better represented as hypergraphs for which the concept of
tree decomposition can be generalized,
see, e.g., \cite{GottlobLS02}. In the following, we therefore define hypertree
decompositions, of which tree decompositions are a special case.
A \emph{hypergraph} is a pair $H=(V,E)$ with a set $V$ of vertices and a set
$E$ of hyperedges. A hyperedge $e \in E$ is itself a set of vertices with $e
\subseteq V$.
A \emph{hypertree decomposition} of a hypergraph $H=(V,E)$ is a pair
$\mathit{HD}=\langle T,\chi\rangle$, where $T=(N,F)$ is a (rooted) tree,
with $N$ being its set of nodes and $F$ its set of edges,
and
$\chi$
is a labeling function such that for each node $n \in
N$ the so-called \emph{bags} $\chi(n)\subseteq V$ of $\mathit{HD}$
meet the following requirements:
\vspace{-1mm}
\begin{enumerate}
\item for every
$v\in V$ there exists a node $n\in N$ such that
$v\in\chi(n)$,
\item for every
$e\in E$ there exists a node $n\in N$ such that
$e\subseteq\chi(n)$,
\item for every
$v\in V$ the set $\lbrace n\in N\mid
v\in\chi(n)\rbrace$ induces a connected subtree of $T$.
\end{enumerate}
\vspace{-1mm}
A hypertree decomposition $\langle T,\chi\rangle$ is called \emph{semi-normalized} if each node
$n$ in $T$ has at most two children---nodes with zero (resp.\ one, two) children are
called leaf (resp.\ exchange, join) nodes---and for a join node $n$ with children $n_1$ and $n_2$,
$\chi(n)=\chi(n_1)=\chi(n_2)$ holds.%
\footnote{A few terminological remarks: The reason we use the term
``\emph{semi}-normalized'' is that the (more restrictive) concept of
\emph{normalized} (also called \emph{nice}) tree decompositions appears in the
literature. Normalized tree decompositions are also semi-normalized. Moreover,
hypertree decompositions generalize the notion of tree decompositions to the
case of hypergraphs. Therefore, we often use both terms interchangeably if we
are dealing with ordinary graphs, even though, strictly speaking, D-FLAT always
produces a hypertree decomposition.}
Figure~\ref{fig:td-example} depicts a semi-normalized tree decomposition for the
example graph given by the problem instance in
Section~\ref{sub:logicprogramsandanswersetsemantics}.
\begin{figure}
\caption{A graph and a corresponding semi-normalized tree decomposition}
\label{fig:td-example}
\subfloat{\input{3col-graph}}
\qquad
\subfloat{\input{3col-td}}
\end{figure}
For defining the \emph{width} of a hypertree decomposition $\langle(N,F),\chi\rangle$ of
a hypergraph $(V,E)$, we need an additional labeling function $\lambda:N\rightarrow 2^E$
such that, for every $n\in N$,
$\chi(n)\subseteq\bigcup_{e\in\lambda(n)}e$.
A hypertree decomposition is called \emph{complete} when for each hyperedge
$e\in E$ there exists a node $n\in N$ such that $e\in\lambda(n)$.
The \emph{width} of a hypertree decomposition is then the maximum $\lambda$-set
size over all its nodes. The minimum width over all possible hypertree
decompositions is called the \emph{generalized hypertree width}.
The idea of this parameter is to capture the inherent difficulty of the
given problem, such that the smaller the generalized hypertree width
(i.e., the level of cyclicity), the easier
the problem is to solve, see, e.g., \cite{GottlobLS02}.
If $G=(V,E)$ is an ordinary graph, then the width of a tree decomposition $\langle T,\chi\rangle$
is defined differently, namely as the maximum bag size $\lvert \chi(n) \rvert$, minus one.
For a given hypergraph, it is \textbf{NP}-hard to compute a hypertree
decomposition of minimum width.
However, efficient heuristics have been
developed that offer good approximations (cf. \cite{Dermaku08,Bodlaender10}).
It should be
noted that in general there exist many possible
hypertree decompositions for a given hypergraph, and there always exists at
least one:
the degenerated hypertree decomposition
consisting of just a single node which covers the entire hypergraph. D-FLAT
expects the dynamic programming algorithms provided by the user to work on any
semi-normalized hypertree decomposition, so users do not have to worry
about (nor can they currently determine)
which decomposition is
generated. This is up to the heuristic methods
of the htdecomp library, which attempt to construct decompositions of small width.
\subsection{Dynamic Programming on Hypertree Decompositions}
\label{sub:dynamicprogramming}
The value of hypertree decompositions is that their size is only linear in the
size of the given graph. Moreover, they provide a suitable structure to design
dynamic programming algorithms for a wide range of problems. These algorithms
generally start at the leaf nodes and traverse the tree to the root, whereby at each node
a set of partial solutions is generated by taking those solutions into account
that have been computed for the child nodes. The most difficult part in
constructing such an algorithm is to identify an appropriate data structure to
represent the partial solutions at each node: On the one hand, this data
structure must contain sufficient information so as to
compute the representation of the partial solutions at each node from the
corresponding representation at the child node(s). On the other hand, the size
of the data structure should only depend on the size of the bag (and not on the
total size of the instance to solve).
\begin{figure}
\caption{The \prob{3-Col} tuple tables for the instance and tree decomposition
given in Figure~\ref{fig:td-example}}
\label{fig:3col-example}
\input{3col-table}
\end{figure}
For this purpose, at each tree decomposition node we maintain a data structure
which we call \emph{table}. A table contains rows which we call \emph{tuples}
(i.e., mappings that assign some value to each current bag element), in which we
store the partial solutions. Additionally, each tuple in non-leaf nodes may be
associated with tuples from child nodes by means of pointers. These pointers
determine which child tuples a tuple has originated from and are used to
construct complete solutions out of the respective partial solutions when the
computation of all tables is finished.
\footnote{If only the decision problem needs to be solved, these pointers are
not necessary; in general, the algorithm and the structure of the tables depend
on the problem type (cf.\ \cite{GrecoS10,GottlobGS09}).}
We illustrate this idea for the \prob{3-Col} problem. Consider a given graph
$G$ and a corresponding tree decomposition. For each node $n$ in the decomposition, we
basically compute the valid colorings of the subgraph of $G$ induced by
$\chi(n)$, i.e., by the current bag, and store these colorings in the rows of
the table associated with $n$. Hence each tuple of node $n$'s table assigns
either ``r'', ``g'' or ``b'' to each vertex that is contained in the current bag
$\chi(n)$. So far, the steps taken by the dynamic programming algorithm in $n$
do not differ from a general Guess \& Check approach to \prob{3-Col}.
However, when computing the colorings for an exchange node $n$, we start from
the colorings for the child node $n'$ of $n$ and adapt them with respect to
the new vertices $\chi(n)\setminus\chi(n')$.
In addition, we also keep track of which tuples of $n'$ give rise to
which newly calculated tuples of $n$. Figure~\ref{fig:3col-example} depicts the respective
tables of each node of the semi-normalized tree decomposition of
Figure~\ref{fig:td-example}. In each table, column $i$ contains an index
for each tuple, that can be used in column $j$ of a potential parent node to
reference it. For exchange nodes, an entry in column $j$
is a set of pointers to such child tuples. For join nodes, an entry in column $j$ is a
set of pairs $(l,r)$, where $l$ (resp.\ $r$) references a tuple of the left
(resp.\ right) child node. Using these pointers, it is possible to enumerate
all complete solutions to the entire problem by a final top-down traversal.
Note that the size of the tables only depends on the width of the tree
decomposition, and the number of tables is linear in the size of the input
graph. Thus, when the width is bounded by a constant, the search space for each
subproblem remains constant as well, and the number of subproblems only grows by
a linear factor for larger instances.
\section{D-FLAT by Example}
\label{sec:examples}
In this section, we introduce the usage of the D-FLAT system by means of
specific example problems. We begin with a relatively simple case to illustrate
the basic functionality of the system and then continue with more complex
applications.
In the course of this, we will gradually introduce the features of D-FLAT and
the special predicates responsible for the communication between D-FLAT and the
user's programs. These predicates are summarized in
Table~\ref{tab:special_predicates} for future reference.
A general description of the
system and its possible applications
is delegated to
Section~\ref{sec:systemdescription}.
\input{table-special-predicates}
\vspace{-3mm}
\subsection{Graph Coloring}
\label{sec:graph-coloring}
As a first example of how to solve an \textbf{NP}-complete problem using D-FLAT, we
consider \prob{3-Col}, the 3-colorability problem for graphs.
We have already given an encoding for \prob{3-Col} in Section~\ref{sub:logicprogramsandanswersetsemantics}.
Since this program, together with the instance as input, solves the whole
problem, we call it a \emph{monolithic encoding}.
As \prob{3-Col} is
fixed-parameter tractable w.r.t.\ the treewidth, we can take advantage of low
treewidths by solving the problem with a dynamic programming algorithm that
operates on a tree decomposition of the original input graph.
We have sketched the dynamic programming algorithm for \prob{3-Col} in
Section~\ref{sub:dynamicprogramming}.
D-FLAT now provides a means to specify this dynamic programming algorithm in the ASP
language. It reads the instance, stores its graph representation, and from this
it constructs a semi-normalized tree decomposition.
In order for D-FLAT to obtain a graph representation from the instance, the user
only needs to specify on the command line which predicates indicate (hyper)edges
in the graph representation (in this case only \verb!edge/2!).
Following the idea of dynamic programming, we wish to compute a table
for each
tree decomposition node, such that a table can be computed using the
tables of the child nodes; i.e., we perform the calculations in a bottom-up
manner.
Since semi-normalized tree decompositions only allow for two kinds of nodes---%
\emph{exchange nodes}, which have one child, and \emph{join nodes},
which have exactly two children with the same contents as the join node---it
suffices to provide D-FLAT with an ASP program that computes the table for
exchange nodes (\emph{exchange program} for short) and with a program that
computes the table for join nodes (\emph{join program} for short).%
\footnote{In fact,
D-FLAT produces tree decompositions where all leaves as well as the root node
have an empty bag. Empty leaf nodes have the advantage that they do not involve
any problem-specific
computation (at least in the use cases we considered so far) but just deliver
the empty tuple (with cost $0$) to its parent node. (This default behavior is
mirrored in D-FLAT in the
sense that it currently only supports user-specified ASP programs for exchange
and join nodes).
On the other hand, having an empty root node guarantees that
final actions of
the dynamic programming algorithm
can always be specified in the program for the exchange nodes.}
These two programs are all that is required of the user to solve a problem.
D-FLAT traverses the tree decomposition in post-order and, for each node,
invokes the ASP solver with the proper program, i.e., with the exchange program
for exchange nodes and with the join program for join nodes.
\begin{figure}
\caption{Workflow for calculating a tree decomposition node's table}
\label{fig:dataflow}
\subfloat[Data flow while processing an exchange node]{
\input{exchange-dataflow}}
\hfill
\subfloat[Data flow while processing a join node]{
\input{join-dataflow}
}
\end{figure}
An illustration of
such single
steps is given in
Figure~\ref{fig:dataflow}.
The input for the exchange or join program
consists of
\begin{itemize}
\item the original problem instance as supplied by the user,
\item the tuples from the child nodes as a set of facts,
\item the current bag, i.e., the current, introduced and
removed vertices, as a set of facts.
\end{itemize}
The answer sets then constitute the current node's table. D-FLAT takes care of
processing them and filling the appropriate data structures. Once all new tuples
have thus been stored, it proceeds to the next node. This procedure continues
until the root has been processed.
We now present an exchange program for \prob{3-Col}.
Regarding the input that is
supplied by D-FLAT, the set of current, introduced and removed vertices is
given by
predicates \verb!current/1!, \verb!introduced/1! and
\verb!removed/1!, respectively. Each child tuple has some identifier $I$ and is
declared by the fact \verb!childTuple(I)!. The corresponding mapping is given
by facts of the form \verb!mapped(I,X,C)!, which signifies that in the tuple $I$
the vertex $X$ is assigned the color $C$. (The predicate name ``mapped'' was
chosen in analogy to the present tense term ``map''.)
\begin{lstlisting}
color(red;green;blue).
1 { chosenChildTuple(I) : childTuple(I) } 1.
map(X,C) :- chosenChildTuple(I), mapped(I,X,C), current(X).
1 { map(X,C) : color(C) } 1 :- introduced(X).
:- edge(X,Y), map(X,C), map(Y,C).
\end{lstlisting}
The program above is based on the intuition that in an exchange node each child tuple
gives rise to a set of new tuples, such that the new tuples coincide on the
coloring of common vertices~(line~3), and the coloring of introduced vertices is
guessed~(line~4), followed by a check~(line~5) which uses the \verb!edge/2!
predicate of the problem instance.
Each answer set here constitutes exactly one new tuple in the current node's
table. The output predicate \verb!map/2! is used to specify the partial
coloring of the new tuple, just like it was used in the monolithic encoding. It
must assign a color to each vertex in the current node. Another output
predicate recognized by D-FLAT is \verb!chosenChildTuple/1!. It indicates which
child tuple a partial coloring (characterized by an answer set) corresponds to
and must, of course, only have one child tuple identifier as its extension. This
predicate is required for the reconstruction of complete solutions after all
tables have been computed and can therefore be omitted if just the decision
problem should be solved.
It still remains to provide D-FLAT with a program for processing the other kind
of nodes, viz., the join nodes. A join program for this purpose could look like
this:
\begin{lstlisting}
1 { chosenChildTupleL(I) : childTupleL(I) } 1.
1 { chosenChildTupleR(I) : childTupleR(I) } 1.
:- chosenChildTupleL(L), chosenChildTupleR(R), mapped(L,X,A), mapped(R,X,B), A != B.
map(X,V) :- chosenChildTupleL(L), mapped(L,X,V).
\end{lstlisting}
Here an answer set obviously must state \emph{for each} of the two child nodes
the chosen preceding tuple (indicated by \verb!chosenChildTupleL! and
\verb!chosenChildTupleR!).
In the case of \prob{3-Col}, two tuples match (i.e., are joined to a new tuple)
iff their partial assignments coincide (line~3). The tuple resulting from such a
join is then also equal to each of the two matching child tuples (line 4).
This particular join behavior is so basic that it also applies to other common
problems. Therefore, D-FLAT offers a (quite efficient) default implementation
of this behavior, which users can resort to (if it suffices) instead of
writing their own join programs.
We would like to
point out here the difference between the
role of exchange and join nodes.
In
particular, one might wonder why for join nodes there is a default
implementation that can often be used, whereas there is none for exchange nodes.
This difference is due to the idea of dynamic programming which, on the one
hand, involves solving partial problems and, on the other hand, requires
combining partial solutions (which is clearly separated, thanks to the
concept of semi-normalized tree decompositions). Obviously, the combining is explicitly done in join
nodes, whereas usually the solving of partial problems is performed in exchange
nodes, which therefore require problem-specific code. Exchange nodes
\emph{always} require problem-specific code since the entire problem can be seen
as a single degenerated exchange node.
Hence, parts of exchange programs resemble monolithic encodings very much,
even though these are technically irrelevant to D-FLAT.
Indeed, exchange programs can be seen as more general than monolithic encodings
since a correct exchange program must in principle also solve the entire
problem.
In some cases however, knowledge of the problem is also needed in join nodes, as the
next subsection will show.
\subsection{Boolean Satisfiability}
\label{sec:sat}
Instances for the
satisfiability problem (\prob{Sat}) are given by the
predicates \verb!pos(C,V)! and \verb!neg(C,V)!, denoting that the propositional
variable $V$ occurs positively resp.\ negatively in the clause $C$. Each clause
and variable is declared using the predicates
\verb!clause/1! and \verb!atom/1!, respectively.
The following monolithic encoding solves such instances:
\begin{lstlisting}
1 { map(A,true), map(A,false) } 1 :- atom(A).
sat(C) :- pos(C,A), map(A,true).
sat(C) :- neg(C,A), map(A,false).
:- clause(C), not sat(C).
\end{lstlisting}
We can obtain a graph representation of a \prob{Sat} instance by constructing
its incidence graph, i.e, the graph obtained by considering the clauses and variables as vertices
and connecting a clause vertex with a variable vertex by an edge iff the
respective variable occurs in the respective clause. Again, in order for D-FLAT
to transform the input into a graph representation, the user is only required to
state that \verb!pos/2! and \verb!neg/2! indicate edges.
A dynamic programming algorithm for the model counting problem working on tree
decompositions of the incidence graph is given in \cite{Samers10}.
We now present a possible ASP encoding for the exchange node that follows that
work's general idea of such an algorithm and
generalizes it for semi-normalized tree decompositions.
It
should be noted that we primarily assign truth values to current propositional
atoms like in the monolithic encoding, but as a consequence of this we also
assign true or false to each current clause depending on whether or not it is
satisfied by the partial interpretation represented by the tuple. We need this
information on the status of a clause because, when a clause is removed, all
tuples not satisfying this clause must be eliminated.
\begin{lstlisting}
1 { chosenChildTuple(I) : childTuple(I) } 1.
:- clause(C), removed(C), chosenChildTuple(I), not mapped(I,C,true).
map(X,true) :- chosenChildTuple(I), mapped(I,X,true), current(X).
1 { map(A,true), map(A,false) } 1 :- atom(A), introduced(A).
map(C,true) :- pos(C,A), current(C), map(A,true).
map(C,true) :- neg(C,A), current(C), map(A,false).
map(X,false) :- current(X), not map(X,true).
\end{lstlisting}
The mentioned elimination of non-satisfying truth assignments is performed by
the check in line~2. The child tuple's partial interpretation is retained
(lines~3 and 7) and extended by a guess on introduced atoms (line~4). If a
clause was satisfied at some point before, it remains so due to line 3, whereas
lines~5 and 6 mark clauses as satisfied by the current partial assignment.
For the \prob{Sat} problem we cannot use D-FLAT's default implementation for
join nodes. The reason is that, in order to be joined, two tuples do not need to
coincide on the values they assign to clauses, but only on the assignments for
atoms. If, for a given truth assignment on the current atoms, a clause is
satisfied by either child tuple, it is also satisfied by the tuple resulting
from joining the two. The following join program follows this idea:
\begin{lstlisting}
1 { chosenChildTupleL(I) : childTupleL(I) } 1.
1 { chosenChildTupleR(I) : childTupleR(I) } 1.
:- mapped(L,A,true), mapped(R,A,false), atom(A), chosenChildTupleL(L), chosenChildTupleR(R).
:- mapped(L,A,false), mapped(R,A,true), atom(A), chosenChildTupleL(L), chosenChildTupleR(R).
map(X,true) :- chosenChildTupleL(I), mapped(I,X,true).
map(X,true) :- chosenChildTupleR(I), mapped(I,X,true).
map(X,false) :- current(X), not map(X,true).
\end{lstlisting}
\subsection{Minimum Vertex Cover}
Encodings as outlined above suffice to instruct D-FLAT to solve the respective
enumeration problems (``What are the solutions?''), counting problems (``How
many solutions exist?'') and decision problems (``Is there a solution?'').
Often it is also desired to solve optimization problems, for which additional
information regarding the cost of a (partial) solution must be supplied in the
encodings. As an example, we briefly introduce how the ``drosophila'' of
fixed-parameter algorithms, the \prob{Minimum Vertex Cover} problem, can be
solved. Instances are again given by \verb!vertex/1! and \verb!edge/2!. As
before, we begin with a monolithic encoding for comparison:
\begin{lstlisting}
1 { map(X,in), map(X,out) } 1 :- vertex(X).
:- edge(X,Y), map(X,out), map(Y,out).
cost(C) :- C = #count{ map(X,in) }.
#minimize[ cost(C) = C ].
\end{lstlisting}
By means of the \verb!#minimize! statement, we leave it to the ASP solver to
filter out suboptimal solutions. However, it is a mistake to use such an
optimization statement when writing the exchange program for D-FLAT because
then one would filter out tuples whose local cost might exceed that of others
but which in the end would yield a better global solution.
An exchange program for \prob{Minimum Vertex Cover} could look like this:
\begin{lstlisting}
1 { chosenChildTuple(I) : childTuple(I) } 1.
map(X,Y) :- current(X), chosenChildTuple(I), mapped(I,X,Y).
1 { in(X), out(X) } 1 :- introduced(X).
map(X,in) :- in(X).
map(X,out) :- out(X).
:- edge(X,Y), map(X,out), map(Y,out).
currentCost(C) :- C = #count{ map(_,in) }.
cost(C) :- chosenChildTuple(I), childCost(I,CC), IC = #count{ in(_) }, C = CC + IC.
\end{lstlisting}
The \verb!cost/1! predicate is recognized by D-FLAT and specifies the cost of
the partial solution (obtained by extending the current tuple with its
predecessors, recursively). This number is computed in line~8 by adding to the
preceding tuple's cost (which is provided by D-FLAT as \verb!childCost/2!) the
cost which is due to introduced vertices---in this case, the number of vertices
guessed as \verb!in!.
A peculiarity when using the default join implementation (as in this example) is
that in each tuple we not only need to store the cost of the corresponding
partial solution but also the cost of the tuple itself, declared by
\verb!currentCost/1!, because upon joining two coinciding tuples and adding the
total costs of their associated partial solutions, the portion of the cost that
is due to the coinciding child tuples is counted twice and must thus be
subtracted from the sum. This is why D-FLAT also recognizes the
\verb!currentCost/1! predicate. When implementing the join program manually,
this predicate is useless, as calculating the cost of a tuple resulting from a
join is then up to the user.
Note that, as soon as the root of the tree decomposition has been processed by
D-FLAT, it possesses all the information required for determining the cost and
number of optimal solutions in constant time (i.e., without an additional tree
traversal), as well as for enumerating all optimal solutions with linear delay
by traversing the tree top-down, following each tuple's pointers to its
predecessors to construct solutions. The processing of an optimization problem
instance is no different from that of problems without optimization, except that
partial solution costs must be specified for each tuple.
\subsection{Cyclic Ordering}
Up until now, we have always used a graph representation of the problem
instance. Some problems, however, can be represented more naturally by a
hypergraph. An example is the \prob{Cyclic Ordering} problem,
which takes
as instance
a set $V$, described by the predicate \verb!vertex/1!,
and a set of triples described by \verb!order(A,B,C)!, where $A, B, C$ are
declared vertices. $f : V \to \{1,\dots,\lvert V \rvert\}$ is called an
ordering of the vertices. A triple \verb!order(A,B,C)! is said to be satisfied
by an ordering $f$ iff $f(A) < f(B) < f(C)$ or $f(B) < f(C) < f(A)$ or $f(C) <
f(A) < f(B)$. The objective is to bring the vertices into an ordering such that
all triples are satisfied.
The following monolithic encoding achieves this:
\begin{lstlisting}
1 { map(V,1..N) } 1 :- vertex(V), N = #count { vertex(_) }.
:- map(V1,K), map(V2,K), V1 < V2.
lt(V1,V2) :- map(V1,K1), map(V2,K2), K1 < K2.
sat(A,B,C) :- order(A,B,C), lt(A,B), lt(B,C).
sat(A,B,C) :- order(A,B,C), lt(B,C), lt(C,A).
sat(A,B,C) :- order(A,B,C), lt(C,A), lt(A,B).
:- order(A,B,C), not sat(A,B,C).
\end{lstlisting}
One can naturally represent the problem by a hypergraph where one considers the
elements of $V$ as vertices and the triples as hyperedges. This ensures that the
vertices in each triple must appear together in at least one bag and each triple
can therefore be checked in an exchange node.
We use this opportunity to introduce a distinction between two ways of how to
write exchange encodings: We can either do this, as we did in all the cases
before, in a ``bottom-up'' way, i.e., by selecting preceding child tuples and
then constructing the new tuples originating from them. When dealing with just a
decision problem, another possibility is to proceed ``top-down'', i.e., by
guessing first an assignment of values to current vertices and then checking if
some child tuple exists that is a valid predecessor of the guess. This has the
advantage that we avoid a guess over a potentially huge number of child tuples
when probably many of the resulting tuples would coincide.
The following exchange encoding for \prob{Cyclic Ordering} proceeds
``top-down'':
\begin{lstlisting}
1 { map(V,1..N) } 1 :- current(V), N = #count { current(_) }.
:- map(V1,K), map(V2,K), V1 < V2.
lt(V1,V2) :- map(V1,K1), map(V2,K2), K1 < K2.
sat(A,B,C) :- order(A,B,C), lt(A,B), lt(B,C).
sat(A,B,C) :- order(A,B,C), lt(B,C), lt(C,A).
sat(A,B,C) :- order(A,B,C), lt(C,A), lt(A,B).
:- order(A,B,C), current(A;B;C), not sat(A,B,C).
gtChild(I,V1,V2) :- mapped(I,V1,K1), mapped(I,V2,K2), current(V1;V2), K1 > K2.
noMatch(I) :- lt(V1,V2), gtChild(I,V1,V2).
match :- childTuple(I), not noMatch(I).
:- not match.
\end{lstlisting}
The first seven lines are very similar to the monolithic encoding. The
remainder of the program makes sure that there is some valid predecessor among
the child tuples. Note that such an approach can only be employed for decision
problems because counting or enumerating solutions require pointers to child
tuples.
In fact, this problem differs from the previously discussed insofar as for those
the set of values that can be mapped to a vertex was fixed (e.g., three colors
or a truth value), whereas here it would be required to assign each vertex a
number from $1$ up to the total number of vertices (just as the monolithic
program does), if one were to really construct a solution to the problem
with this approach. This
would violate the principles of dynamic programming because subproblems would
depend on the whole problem. However, note that in line~1, only a \emph{local}
ordering, i.e., an ordering of the current bag elements, is guessed, so the
value mapped to a vertex is \emph{not} its position in an ordering of all the
vertices.
\subsection{Further Examples and Overview}
\label{sec:further-examples}
In this section, we have illustrated the functioning of D-FLAT via several
examples in the course of which we have also introduced step-by-step the special
predicates D-FLAT provides or recognizes.
D-FLAT
solutions for
further problems
are provided on the system website at
\url{http://www.dbai.tuwien.ac.at/research/project/dynasp/dflat/}.
Let it be noted here that it is also possible to handle problems higher on the
polynomial hierarchy than \textbf{NP}. For instance, one can solve (propositional) ASP itself with
D-FLAT, for instance by implementing the algorithm presented in \cite{JaklPW09}.
For purposes like this, D-FLAT allows not only for assignments to all
vertices but more generally for a hierarchy of assignments. For example, each
top-level assignment (as we have been dealing with before) might have subsidiary
assignments that are used to determine if the respective top-level assignment is
valid. A detailed account of this is beyond the scope of this paper. We refer
the reader to the system website for examples of this.
Let us briefly summarize the main features of D-FLAT.
In many cases,
the user has to implement only the program for exchange nodes, while for
join nodes D-FLAT offers a default implementation. For certain problems, this
default has to be overridden. In their programs, users can exploit the full
language accepted by Gringo. The only restriction is that some predicates are
reserved
by D-FLAT.
Table~\ref{tab:special_predicates} gives a summary of these predicates and also
points out under which circumstances (depending on the program or problem type)
which predicates are used.
Predicates used in the facts that constitute the problem instance
are supplied verbatim
to the exchange and join programs.
\section{The D-FLAT System}
\subsection{System Description}
\label{sec:systemdescription}
D-FLAT is written in C++ and can be compiled for many platforms. The
main libraries used are:
(1)
the SHARP framework\footnote{See \url{http://www.dbai.tuwien.ac.at/research/project/sharp/}.}
which takes care of constructing a tree decomposition of the input graph and
then semi-normalizing it; as well, SHARP provides the skeleton for D-FLAT's data management;
(2)
SHARP itself uses the htdecomp library\footnote{See \url{%
http://www.dbai.tuwien.ac.at/proj/hypertree/downloads.html}.}
which implements several heuristics for (hyper)tree decompositions, see also \cite{Dermaku08};
(3)
the
Gringo/Clasp\footnote{See \url{http://potassco.sourceforge.net/}.}
family of ASP tools (see also \cite{GebserKKOSS11} for a recent survey)
is finally used for grounding and solving the user-specified ASP programs.
Figure~\ref{fig:flowchart} depicts the control flow between these components in
a simplified way:
D-FLAT initially parses the instance and constructs a hypergraph representation,
which is then used by htdecomp to build a hypertree decomposition. SHARP is now
responsible for traversing the tree in post-order. For each node, it calls
D-FLAT which flattens the child tables, i.e., converts them to a set of facts
which is given to the ASP solver to compute the new tuples as answer sets.
These are collected by D-FLAT which populates the current node's table. When
all nodes have been processed like this, D-FLAT reconstructs the solutions from
the tables.\footnote{%
The data flow during the procedure within a node (indicated by the dashed box)
is illustrated in Figure~\ref{fig:dataflow}. Note that D-FLAT also features a
default join implementation which does not use ASP and is not depicted.}
\begin{figure}
\caption{A flowchart illustrating the (simplified) interplay of D-FLAT's
components}
\label{fig:flowchart}
\input{flowchart}
\end{figure}
Since we leave the decomposition part as well as the solving of the ASP programs
to external libraries, D-FLAT immediately takes advantage of improvements in
these libraries. It is thus also possible to switch to another ASP solver
altogether without changing \mbox{D-FLAT} internals except, of course, the parts calling
the solver. Likewise, our approach allows to replace the tree decomposition
part with weaker (but more efficient) concepts like graph cuts, etc.
Since the user provides ASP programs for the exchange and, optionally, join
nodes at runtime, D-FLAT is independent of the particular problem being solved.
This means that there is no need to recompile in order to change the dynamic
programming algorithm. It can be used out of the box as a rapid prototyping tool
for a wide range of problems.
\begin{table}
\caption{D-FLAT's most important command-line options}
\label{tab:command-line-options}
\begin{tabular}{p{4cm}p{8cm}}
\hline
\hline
Argument & Meaning\\
\hline
\verb!-e edge_predicate! \newline \hspace*{1em} \footnotesize (mandatory at
least once) &
\verb!edge_predicate! is used in the problem instance as the predicate declaring
(hyper)edges. Arguments of this predicate implicitly declare the vertices of the
input graph.\\
\noalign{\vspace{1ex}}
\verb!-x exchange_program! \newline \hspace*{1em} \footnotesize (mandatory) &
\verb!exchange_program! is the file name of the ASP program processing exchange
nodes.\\
\noalign{\vspace{1ex}}
\verb!-j join_program! \newline \hspace*{1em} \footnotesize (optional) &
\verb!join_program! is the file name of the ASP program processing join nodes.
If omitted, the default implementation is used.\\
\noalign{\vspace{1ex}}
\verb!-p problem_type! \newline \hspace*{1em} \footnotesize (optional) &
\verb!problem_type! specifies what kind of solution the user is interested in.
It can be either ``enumeration'' (default), ``counting'', ``decision'',
``opt-enum'', ``opt-counting'' or ``opt-value''.\\
\hline
\hline
\end{tabular}
\end{table}
When executing the D-FLAT binary, the user can adjust its behavior using
command-line options. The most important ones are briefly described in
Table~\ref{tab:command-line-options}.
The problem instance, which must be a set of facts, is read from the standard
input. The parser of D-FLAT recognizes the predicates that declare (hyper)edges,
whose names are given as command-line options, and for each such fact introduces
a (hyper)edge into the instance graph. The arguments of the (hyper)edge
predicates implicitly declare the instance graph's set of vertices. Note that
the tree decomposition algorithm used currently does not allow for isolated vertices, but
this is no real limitation because usually solutions of a modified instance
without the isolated vertices can be trivially extended.
Executing the D-FLAT binary \verb!dflat! is illustrated by the following example
call, presupposing a \prob{3-Col} instance
with the file name \verb!instance.lp! and an exchange program
with the file name \verb!exchange.lp! (cf.\ Section~\ref{sec:graph-coloring}),
instructing D-FLAT
to
print the number of solutions and enumerate them:
\begin{verbatim}
dflat -e edge -x exchange.lp < instance.lp
\end{verbatim}
\subsection{Experiments}
\label{sec:experiments}
In this section, we briefly report on first experimental results for the
discussed problems. We compared D-FLAT encodings to monolithic ones using
Gringo and Clasp. Each instance was limited to 15~minutes of CPU time and 6~GB
of memory. Instances were constructed such that they have small treewidth by
starting with a tree-like structure and then introducing random elements until
the desired fixed decomposition width is reached.
Traditional ASP solvers employ clever heuristics to quickly either find some
model or detect unsatisfiability, thereby being able to solve the decision
variant of problems particularly well. In contrast,
the dynamic programming
approach of
D-FLAT currently always calculates the tuple tables in the same way,
whatever the problem variant may be---it is only in the final materialization
stage that solutions are assembled differently, depending on the problem type.
\prob{3-Col} and \prob{Sat} are prime examples of problems where traditional ASP
solvers are very successful in solving the decision variant efficiently.
However, when it no longer suffices to merely find \emph{some} model (e.g., when
dealing with counting or enumeration problems), the decomposition exploited by
D-FLAT pays off for small treewidths, especially when there is a great number of
solutions.
In our experiments, the monolithic encodings indeed
soon hit the time limit. D-FLAT, on the other hand, even sooner ran out of
memory for enumeration due to its materializing all solutions in memory at the
same time, and because the number of solutions increased rapidly with larger
instances. This is a motivation to improve the materialization procedure in the
future with incremental solving techniques.
Where D-FLAT excelled was the counting variant of these problems. It could solve
each instance in a matter of seconds, while the monolithic program again ran out
of time soon. Figure~\ref{fig:sat-plot} illustrates the strengths and weaknesses
of D-FLAT for the counting and decision variants of \prob{Sat}: Although the
monolithic program almost instantaneously solved the decision problem of all of
the test instances, its running time on the counting problem soon exploded
(standard ASP-solvers do not provide a dedicated functionality for counting and thus
have to implicitly enumerate all answer sets)
whereas D-FLAT remained almost unaffected on average.
\begin{figure}
\caption{Comparison of D-FLAT to monolithic encodings}
\subfloat[The decision and counting variant of \prob{Sat} on instances of
treewidth 8]{%
\label{fig:sat-plot}%
\includegraphics[scale=0.4]{sat-plot.pdf}%
}%
\quad
\subfloat[Determining the optimum cost for \prob{Minimum Vertex Cover} on
instances of treewidth 12]{%
\label{fig:vertex-cover-plot}%
\includegraphics[scale=0.4]{vertex-cover-plot.pdf}%
}
\end{figure}
Although most of the time traditional ASP-solvers
perform
very efficiently on decision problems, for some problems they have more difficulties, in particular
when the grounding becomes
huge. Our investigations show that
for
the \prob{Cyclic Ordering} problem, D-FLAT often outperforms the monolithic
program, but it could also be observed that D-FLAT's running time is heavily
dependent on the constructed tree decomposition. For this reason, we averaged
over the performance on multiple tree decompositions for each instance size.
The \prob{Minimum Vertex Cover} problem proved to be a strong suit of D-FLAT
(cf.\ Figure~\ref{fig:vertex-cover-plot}). In optimization problems in general,
stopping after the first solution has been found is not an option for
traditional solvers, since yet undiscovered solutions might have lower cost.
Another advantage of D-FLAT is that traditional solvers (at least in the case of
Clasp) require two runs for counting or enumerating all optimal solutions: The
first run only serves to determine the optimum cost, while the second starts
from scratch and outputs all models having that cost. D-FLAT, in contrast, only
requires one run at the end of which it immediately has all the information
needed to determine the optimum cost.
As a concluding remark, recall that D-FLAT's main purpose is to provide a means
to specify dynamic programming algorithms declaratively and not to compete with
traditional ASP solvers, which is why we refrain from extensive benchmarks in
this paper.
Nonetheless,
it can be concluded that
D-FLAT is particularly successful for optimization and counting
problems (provided the treewidth is small), especially when the number of
solutions or the size of the monolithic grounding explodes.
\section{Conclusion}
\label{sec:conclusion}
\paragraph*{Summary.}
We have introduced D-FLAT, a novel system which allows to specify
dynamic programming algorithms over (hyper)tree decompositions by means of
Answer-Set Programming.
To this end, D-FLAT employs an ASP system as an underlying inference
engine to compute the local solutions of the specified algorithm.
We have provided
case studies illustrating how
the rich syntax of ASP allows for succinct and easy-to-read specifications
of such algorithms.
The system%
---together with the example programs given in Section~\ref{sec:examples}
and the benchmark instances used in Section~\ref{sec:experiments}---%
is free software and
available at
\begin{quote}
\url{http://www.dbai.tuwien.ac.at/research/project/dynasp/dflat/}.
\end{quote}
\paragraph*{Related Work.}
Several forms of support of
dynamic programming exist in the world of PROLOG, where tabling
of intermediate results is a natural concept. We mention here the
systems TLP \cite{GuoG08} and B-Prolog \cite{DBLP:journals/corr/abs-1103-0812}.
In the world of datalog, we refer to
the Dyna system \cite{dyna} which provides a wide range of features
for memoization. On the other side of the spectrum,
the work of \cite{GottlobPW10}
shows that the simple formalism
of monadic datalog can be placed instead of monadic second-order (MSO) logic
in meta-theorems (about fixed-parameter tractability in terms of
treewidth) \`a la Courcelle.
The latter approach is thus more of theoretical interest.
The main difference between D-FLAT and the mentioned systems
is that the user can make use of the full language of ASP and, in particular,
employ the Guess \& Check methodology to subproblems.
However, it has to be mentioned that our system so far supports only dynamic
programming in connection with (hyper)tree decompositions.
\paragraph{Future Work.}
As a next step, we have to compare D-FLAT in more detail to the aforementioned
approaches in terms of both modeling capacities and
performance issues. As well, we plan to provide different decomposition options
within D-FLAT. In particular, we anticipate that even a rather simple (but efficient)
decomposition of the graph, say a simple split into two parts, might in practice lead to
performance gains over monolithic encodings.
Another line of optimization concerns
lazy evaluation strategies, for
which incremental ASP techniques \cite{GebserSS11} have been developed.
Finally, the relation of our approach to
reactive ASP \cite{GebserGKS11} might provide interesting new research directions.
\paragraph*{Acknowledgments} Supported by Vienna University of Technology
special fund ``Innovative Projekte 9006.09/008''.
\bibliographystyle{acmtrans}
|
\section{Acknowledgements}
This work received support from the Brazilian agency, CNPq, under grant 305639/2010-2 (A.D.), and by FAPESP under grant 2012/05085-1 (LM).
\bibliographystyle{aipproc}
|
\section{Introduction}
It is commonly believed that gapped phases of quantum many-body systems exhibit area law: entanglement entropy of a simply connected subsystem increases as the area of the boundary.\cite{Eisert2008} Overwhelming amount of evidences supporting this statement has been suggested, including the explicit proof for a ground state of 1D gapped system\cite{Hastings2007a}, exactly solvable models\cite{Levin2005}, and variational wavefunctions\cite{Verstraete2006}. Constant subcorrection to the entanglement entropy - also known as the topological entanglement entropy - can be extracted by judiciously choosing a set of subsystems that cancel out the boundary contributions.\cite{Kitaev2006,Levin2006} Topological entanglement entropy is believed to be a universal constant characterizing the phase of the quantum many-body system.
Li and Haldane(LH) were the first to realize that the spectrum of the reduced density matrix may reveal an information about the phase that cannot be inferred from the entanglement entropy alone.\cite{Li2008,Thomale2010} While LH studied reduced density matrix in the orbital cuts, one may study its spectrum along a real-space partition and arrive at a similar conclusion.\cite{Sterdyniak2011,Dubail2011,Rodriguez2012} In particular, it has been recently suggested by several authors that entanglement spectrum along a real-space partition has a low-lying part that can be described by a local field theory.\cite{Cirac2011,Dubail2012}
Topological entanglement entropy can be obtained from a real-space entanglement spectrum of variational wavefunctions, similar to the way it is extracted from the entanglement entropy.\cite{Dubail2012} Consequently, the corresponding linear combination of entanglement spectrum is ``topological'', in a sense that i) it does not interact with any local observable ii) it is equal to the topological entanglement entropy.
Here we claim that the existence of such topological operator can be attributed to an approximate conditional independence of these quantum states. A tripartite state $\rho_{ABC}$ is conditionally independent if conditional mutual information $I(A:C|B) = S_{AB} + S_{BC} - S_{B} - S_{ABC}$ is equal to $0$. A state is approximately conditionally independent if $0$ is replaced by a small number $\epsilon>0$. To the best of author's knowledge, Hastings and Poulin were the first to point out that there can be configurations that are conditionally independent even in a quantum many-body system with long range entanglement.\cite{Poulin2010a} To illustrate their idea, suppose entanglement entropy satisfies an area law with a \emph{universal} constant subcorrection term.
\begin{equation}
S_A = a|\partial A| - \gamma, \label{eq:area_law}
\end{equation}
One can show that $I(A:C|B)=0$ for a choice of $A,B,C$ such that i) $AB,BC,B,ABC$ are all simply connected ii) $A$ and $C$ do not share a boundary.
A state that is conditionally independent saturates the equality condition of the strong subadditivity of entropy.\cite{Lieb1973} Such state forms a quantum Markov chain, and the structure of the reduced density matrix is vastly restricted compared to an arbitrary state.\cite{Petz2003,Hayden2004,Leifer2007} It is important to note that one cannot directly use these results for a generic quantum many-body system, since the conditional independence condition is unlikely to hold exactly. Still, one may hope these properties to hold approximately for a sufficiently small conditional mutual information. This is precisely the key idea behind this paper. More specifically, we shall use the recently discovered operator extension of the strong subadditivity of entropy as our main technical tool.\cite{Kim2012a}
Rest of the paper is structured as follows. In Section \ref{section:ACI}, we shall briefly review several information-theoretic inequalities. In Section \ref{section:correlation_bound}, we shall introduce a diagrammatic trick that leads to the main result of this paper. Its physical interpretation shall be given in Section \ref{section:interpretation}.
\section{Approximately conditionally independent states\label{section:ACI}}
Strong subadditivity of entropy is one of the most widely used tools in quantum information theory. Its importance stems from the fact that there exists a variety of nontrivial structure theorems that relate the reduced density matrix of different subsystems if the inequality is saturated with an equality condition.\cite{Petz2003,Hayden2004,Leifer2007} In particular, Petz showed that the following relation holds if and only if the conditional mutual information $I(A:C|B)$ is equal to $0$.\cite{Petz2003}
\begin{equation}
\hat{H}_{AB} + \hat{H}_{BC} - \hat{H}_B - \hat{H}_{ABC}=0,\label{eq:equality_condition}
\end{equation}
where $\hat{H}_A = -I_{A^c} \otimes \log \rho_A$ is a formal definition of the entanglement spectrum. From now on, we denote the left hand side of the equation as $\hat{H}_{A:C|B}$ and refer to it as a \emph{conditional mutual spectrum} of $ABC$. It follows that
\begin{equation}
\mathcal{C}(\hat{H}_{A:C|B} , X) = 0, \label{eq:ccf}
\end{equation}
where $\mathcal{C}(\hat{H}_{A:C|B} , X) = \langle \hat{H}_{A:C|B} X\rangle - \langle \hat{H}_{A:C|B}\rangle \langle X\rangle$ is a connected correlation function between the conditional mutual spectrum and an arbitrary operator $X$. $\langle \cdots \rangle$ denotes ground state expectation value.
While such operator trivially has zero correlation with any local operator, exact conditional independence is rarely satisfied by any realistic physical systems. Motivated by this observation, author has recently obtained a nontrivial statement about the spectrum of the reduced density matrices.\cite{Kim2012a}
\begin{thm}
\begin{equation}
\Tr_{BC}(\rho_{ABC} \hat{H}_{A:C|B}) \geq 0. \label{eq:OSSA}
\end{equation}
\end{thm}
We would like to emphasize two important facts about this inequality. First, Eq.\ref{eq:OSSA} reproduces a statement similar to Eq.\ref{eq:equality_condition} when the conditional mutual information is $0$. This can be seen from the following lemma.
\begin{lem}
\begin{equation}
|\Tr_{ABC}(\rho_{ABC} \hat{H}_{A:C|B} O_A)| \leq \|O_A \| I(A:C|B),
\end{equation}
where $\| \cdots \|$ is $l_{\infty}$ norm. \label{lemma:lemma1}
\end{lem}
\begin{proof}
Let $\mathcal{I}_A=\Tr_{BC}(\rho_{ABC} \hat{H}_{A:C|B})$. $|\Tr_A (\mathcal{I}_AO_A) | \leq \|O_A \| |\mathcal{I}_A|_1$. $|\cdots|_1$ is the $l_1$ norm. Since $\mathcal{I}_A$ is positive, $|\mathcal{I}_A|_1 = \Tr_A \mathcal{I}_A = I(A:C|B)$.
\end{proof}
If the conditional mutual information vanishes, the corresponding conditional mutual spectrum has zero correlation with any operator supported on $A$. Furthermore, since both $\hat{H}_{A:C|B}$ and $I(A:C|B)$ are symmetric under the exchange of $A$ and $C$, the same statement holds for an operator supported on $C$ as well. Secondly, Eq.\ref{eq:OSSA} is satisfied by any quantum states. Therefore, unlike Eq.\ref{eq:equality_condition}, it can be applied to quantum states that \emph{approximately} saturate the strong subadditivity of entropy.
\section{Correlation bound for entanglement spectrum\label{section:correlation_bound}}
The main goal of this section is to obtain a statement that resembles Eq.\ref{eq:ccf} when the global state is a ground state of a gapped quantum many-body system. Such correlation bound can be easily obtained in certain cases using Lemma \ref{lemma:lemma1} alone, but there are also important caveats. For example, there are choices of subsystems that yield a nonzero value of conditional mutual information even at a fixed point of some renormalization-group flow.\cite{Kitaev2006,Levin2006} Furthermore, Lemma \ref{lemma:lemma1} alone cannot produce any bound on the correlation between the conditional mutual spectrum $\hat{H}_{A:C|B}$ and an operator supported on $B$. We shall show that, despite these shortcomings, it is still possible to obtain a bound analogous to Eq.\ref{eq:ccf} under a reasonable set of assumptions.
A brief comment on the notation is in order. For a conditional mutual spectrum $\hat{H}_{A:C|B}$, we shall refer $B$ as the {\it reference party} and $A,C$ as {\it target parties}. Also, we shall diagrammatically represent the operator $\hat{H}_{A:C|B}$ with the following rule. The reference party corresponds to the region with a `R' sign. Each of the target parties corresponds to one of the simply connected regions with a `T' sign. When taking a partial trace, subsystem $X$ is used to denote the nontrivial support of operator $X$. Shaded region in the diagram is a nontrivial support of $X$.
We postulate the following modified formula for the entanglement entropy to account for the deviations from the ideal area law.
\begin{equation}
S_A = a|\partial A| - \gamma + \epsilon_A.
\end{equation}
\begin{equation}
S_A + S_B - S_{AB} = \epsilon_{A:B}.
\end{equation}
For a large enough subsystem size, we expect $\epsilon_A$ to approach $0$. $\epsilon_{A:B}$ denotes a long range correlation of the ground state. Due to the exponential clustering theorem, we expect $\epsilon_{A:B}$ to scale as $\min(|A|,|B|)^2 e^{-\frac{2l}{ \xi}}$, where $\xi$ is the correlation length and $|A|$ is the volume of the subsystem $A$.\footnote{The volume factor was chosen in such a way that the bound on connected correlation function from mutual information in Reference \cite{Wolf2007} yields the exponential clustering theorem in Reference \cite{Nachtergaele2006}.}
To simplify the analysis, we assume that each of the subsystems are sufficiently smooth and their boundary lengths are $O(l)$. We assume that the support of $X$ is sufficiently small compared to the size of the subsystems. We also assume that $X$ is supported on only one of the subsystems that partitions the system.
\subsection{Deformation moves}
The key idea for generalizing Eq.\ref{eq:ccf} is that one can decompose $\hat{H}_{A:C|B}$ into a sum of $\hat{H}_{A_i:C_i|B_i}$ in such a way that either i) $I(A_i:C_i|B_i)$ is small or ii) $A_iB_iC_i$ is sufficiently far away from the support of $X$. Such decomposition can be derived from a simple application of the following \emph{chain rule of conditional mutual spectrum}.
\begin{equation}
\hat{H}_{A_1A_2:C|B} = \hat{H}_{A_2:C|B} + \hat{H}_{A_1:C|A_2B}\label{eq:chain_rule}.
\end{equation}
The chain rule can be easily verified from the definition of the conditional mutual spectrum. While any deformation of the subsystem can be expressed as a linear combination of the chain rule, we define three elementary deformation moves for the clarity of the exposition.
First example is an {\it isolation move}. Goal of the isolation move is to deform the boundary between the target party and the reference party so that $X$ can be sufficiently separated from the reference party, see FIG.\ref{fig:isolation}.
\begin{figure}[h!]
\includegraphics[width=0.48\textwidth]{isolation}
\caption{After applying the isolation move, the conditional mutual spectrum is deformed in such a way that i) for the new conditional mutual spectrum, $X$ is sufficiently far away from the reference party ii) the difference is a conditional mutual spectrum with small conditional mutual information. \label{fig:isolation}}
\end{figure}
We also define a {\it separation move}. The purpose of the separation move is to deform the target party so that $X$ is sufficiently separated from the target party, See FIG.\ref{fig:separation}.
\begin{figure}[h!]
\includegraphics[width=0.48 \textwidth]{separation}
\caption{After applying the separation move, the conditional mutual spectrum is deformed in such a way that i) for the new conditional mutual spectrum, $X$ is sufficiently far away from both the reference and target party ii) the difference is a conditional mutual spectrum with small conditional mutual information. \label{fig:separation}}
\end{figure}
By first applying the isolation move and then the separation move, one can always deform the configuration to be distance $O(l)$ away from $X$. The correction from the deformation procedure is of the form $\Tr(\rho_{A_iB_iC_iX} \hat{H}_{A_i:C_i|B_i}X)$ with $I(A_i:C_i|B_i)=o(1)$. To bound these terms, we introduce an {\it absorption move}, see FIG.\ref{fig:absorption}.
\begin{figure}[h!]
\includegraphics[width=0.48 \textwidth]{absorption}
\caption{Goal of the absorption move is to change the correction terms into conditional mutual spectrum $\hat{H}_{A_i:C_i|B_i}$ such that i) support of $X$ is contained in either $A_i$ or $C_i$ ii) $I(A_i:C_i|B_i)$ is small. \label{fig:absorption}}
\end{figure}
After applying the absorption move, corrections from the deformation move can be expressed as a sum of terms of the form $\Tr(\rho_{A_iB_iC_i} \hat{H}_{A_i:C_i|B_i} X)$ with $X \subset A_i$. These terms can be bounded using Lemma \ref{lemma:lemma1}.
To summarize, given a topologically nontrivial configuration, $\mathcal{C}(\hat{H}_{A:C|B},X)$ can be expressed as $\mathcal{C}(\hat{H}_{A':C'|B'},X)$ with $d(A'B'C', X) = O(l)$ and the correction terms that can be expressed as a sum of $\epsilon_{A_i} \|X \|$ and $\epsilon_{A_i:B_i} \|X \|$. Assuming i) $X$ is localized in one of the original subsystems $A,B,C,(ABC)^c$ ii) each of the subsystems are sufficiently large, the correction terms vanish in the $l \to \infty$ limit. One may be tempted to think that $\mathcal{C}(\hat{H}_{A':C'|B'},X)$ vanishes in the $l \to \infty$ limit as well, since correlation decays exponentially in the ground state of a gapped system.\cite{Hastings2006,Nachtergaele2006} While this speculation turns out to be correct, we emphasize that a slight modification of the exponential clustering theorem is necessary.
\subsection{Modified form of exponential clustering theorem}
Before we explain the details of our analysis, we would like to present a technical background about the subject. Exponential clustering theorem states that
\begin{equation}
|\mathcal{C}(O_A,O_B)| \leq c \|O_A \| \|O_B \| \min(|A|, |B|) e^{-\frac{d(A,B)}{\xi}}
\end{equation}
for two spatially separated operator $O_A$ and $O_B$, provided there is a gapped parent hamiltonian that consists of sum of geometrically local bounded-norm terms.\cite{Hastings2006,Nachtergaele2006} Since the spectrum of $\hat{H}_A$ is formally unbounded, one cannot directly apply exponential clustering theorem. We circumvent this problem by regularizing the entanglement spectrum and bounding the error from the regularization procedure.
\begin{defi}
Regularized entanglement spectrum $\hat{H}_{A}^{\Lambda}$ with a cutoff $\Lambda$ is
\begin{equation}
\hat{H}_A^{\Lambda} = -\sum_{p \geq 1/\Lambda} \log p_i \ket{i} \bra{i}.
\end{equation}
\end{defi}
Simple consequence of this construction is that $l_{\infty}$ norm is bounded, i.e. $\| \hat{H}_A^{\Lambda}\| \leq \log \Lambda$. Correction from the regularization can be bounded using the following lemma.
\begin{lem}
\begin{equation}
\Tr (\rho_{AB}\Delta_A^{\Lambda} O_B ) \leq \|O_B \| \frac{\log \Lambda}{\Lambda} d_A
\end{equation}
for $\Lambda \geq 2$, where $\Delta_{A}^{\Lambda} = \hat{H}_A - \hat{H}_A^{\Lambda}$.
\end{lem}
\begin{proof}
Purify $\rho_{AB}$ to $\ket{\psi}_{ABC}$. Rewrite the formula as $\Tr (\rho_{AB}\Delta_A^{\Lambda} O_B ) = \bra{\psi}_{ABC} \Delta_A^{\Lambda} O_B \ket{\psi}_{ABC}$. Note that $\ket{\psi}_{ABC}$ admits a Schmidt decomposition $\ket{\psi}_{ABC} = \sum_i \sqrt{p_i} \ket{i}_A \ket{i}_{BC}$, where $\rho_A = \sum_i p_i \ket{i}_A \bra{i}_A$. This in turn can be expressed as
\begin{equation}
\sum_{p_i \leq 1/\Lambda} -p_i \log p_i \bra{i}_{BC} O_B \ket{i}_{BC}.
\end{equation}
Using $-p_i \log p_i \leq \frac{1}{\Lambda} \log \Lambda$ and $|\bra{i} O_B \ket{i}| \leq \|O_B \|$, one can complete the proof.
\end{proof}
\section{Physical interpretation\label{section:interpretation}}
Setting $\Lambda= d_{ABC} e^{O(l)/ \xi}$, we arrive at the following conclusion.
\begin{equation}
|\mathcal{C}(\hat{H}_{A:C|B},X)| \leq \| X \|(\epsilon_1(l) + \epsilon_2(l))l^2,\label{eq:correlation_bound}
\end{equation}
where $\epsilon_1$ represents a deviation from the ideal area law, and $\epsilon_2$ represents an error from the long range correlation. As $l \to \infty$, conditional mutual spectrum has vanishing correlation with any local operator, provided that i) $X$ is supported on one of $A,B,C,$ or $(ABC)^c$ ii) both $\epsilon_1$ and $\epsilon_2$ decays sufficiently fast. In $l \to \infty$ limit, we have
\begin{equation}
\langle \hat{H}_{A:C|B} X\rangle = I(A:C|B) \langle X \rangle.\label{eq:main_result}
\end{equation}
We conclude that operator $\hat{H}_{A:C|B}$ is topological, in a sense that i) it has vanishing correlation with any operator that is localized in one of the subsystems ii) its eigenvalues contain information about the phase. A set of assumptions to conclude so was that i) correlation decays exponentially ii) the extensive terms of the entanglement entropy cancel out each other iii) the deformation procedure separating $X$ from $ABC$ does not change the topology of the configuration.
We emphasize that the derivation of our result is not necessarily restricted to a pure state. At finite temperature, entanglement entropy obtains volume contributions, but one may be able to show that those contributions can be canceled out as well. In particular, we expect these conditions to be met for quantum many-body systems at sufficiently high temperature.
In the large volume limit, it seems the local contribution of the reduced density matrices cancel out each other, at least when $I(A:C|B)=o(\frac{1}{l^2})$. We do not have a definitive proof for this statement, but we argue as follows. If $\hat{H}_{A:C|B}$ contains a localized term, one could have chosen $X$ to be an operator supported nearby so as to have a large correlation with the local term. Such terms will violate Eq.\ref{eq:main_result}. Our result suggests a decomposition of the entanglement spectrum into i) terms that can be canceled out by a suitable choice of subsystems ii) terms that cannot be canceled out and have small correlation with almost any local operators. It would be interesting if the terms of the first kind can be shown to be quasilocal.
\section{Conclusion}
We have presented a general argument as to why certain linear combination entanglement spectrum allows cancelation of local degrees of freedom, owing in part to a recently discovered information-theoretic inequality. While our formulation is not as precise as the ones described by the variational wavefunction,\cite{Cirac2011,Sterdyniak2011,Dubail2012} it has an advantage of being applicable to a more general class of quantum states. Indeed, we have only used an approximate form of area law and the exponential clustering theorem, which are strongly believed to be a generic property of a gapped phase.
It would be interesting if the approximate conditional independence can be shown to hold in other systems. There are evidences suggesting that models based on BF theory should satisfy such condition\cite{Castelnovo2008}, yet no studies have been performed for exotic models in 3D such as Haah's code.\cite{Haah2011} As for the finite temperature states, approximate conditional independence is one of the key ideas of quantum belief propagation(QBP) algorithm.\cite{Hastings2007c} Success of the QBP indicates that our result may be applicable to finite temperature quantum states as well.\cite{Bilgin2009}
On the other hand, we wish to find a deeper insight as to why conditional independence arises in these systems. In particular, exactly solvable models which satisfy exact conditional independence can be thought as a fixed point of some renormalization-group procedure.\cite{Koenig2008} Does conditional mutual information of topologically trivial configurations monotonically decrease under such flow?
We conclude with a remark that our correlation bound cannot be applied to operators that are supported on more than one of the subsystems. Ability to bound correlation of such form can be used for showing perturbative stability of topological entanglement entropy, but that shall be published elsewhere.\cite{Kim2012}
{\it Acknowledgements---} This research was supported in part by NSF under Grant No. PHY-0803371, by ARO Grant No. W911NF-
09-1-0442, and DOE Grant No. DE-FG03-92-ER40701. I would like to thank Alexei Kitaev for helpful discussions.
|
\section{Introduction}
Let $X$ be a time-homogeneous Markov process. Hunt's hypothesis
(H) says that ``every semipolar set of $X$ is polar". This
hypothesis plays a crucial role in the potential theory of (dual)
Markov processes. To illustrate its importance, let us recall some
potential-theoretic principles.
Suppose that $E$ is a locally compact space with a countable base.
Let $(X, P^x)$ and $(\hat{X},\hat P^x)$ be a pair of dual standard
Markov processes on $E$ as described in Blumenthal and Getoor \cite[VI]{BG68}. Denote by
${\cal B}^n$ the family of all nearly Borel measurable subsets of
$E$. For $D\subset E$, we define the first hitting time of $D$ by
$$
\sigma_D:=\inf\{t>0:X_t\in D\}.
$$
A set $D\subset E$ is called polar (respectively, essentially
polar) if there exists a set $C\in {\cal B}^n$ such that $D\subset
C$ and $P^x(\sigma_C<\infty)=0$ for every $x\in E$ (respectively,
almost every $x\in E$ with respect to the reference measure). $D$
is called a thin set if there exists a set $C\in {\cal B}^n$ such
that $D\subset C$ and {$P^x(\sigma_C=0)=0$} for every $x\in E$.
$D$ is called semipolar if $D\subset\bigcup_{n=1}^{\infty}D_n$ for
some thin sets $\{D_n\}_{n=1}^{\infty}$.
Denote by $E^x$ the expectation with respect to $P^x$. Let $\alpha>0$. A finite $\alpha$-excessive function $f$ on $E$ is called a regular potential provided that
$E^x\{e^{-\alpha T_n}f(X_{T_n})\}\rightarrow E^x\{e^{-\alpha
T}f(X_{T})\}$ for $x\in E$ whenever $\{T_n\}$ is an increasing
sequence of stopping times with limit $T$. Denote by
$(U^{\alpha})_{\alpha>0}$ the resolvent operators for $X$.
\begin{itemize}
\item {\bf Bounded positivity principle $(P^*_{\alpha})$}: If $\nu$ is a finite signed measure
such that $U^{\alpha}\nu$ is bounded, then $\nu U^{\alpha}\nu\geq 0$, where $\nu U^{\alpha}\nu:=\int_EU^{\alpha}\nu(x)\nu(dx)$.
\item {\bf Bounded energy principle $(E^*_{\alpha})$}: If $\nu$ is a finite measure with compact support
such that $U^{\alpha}\nu$ is bounded, then $\nu$ does not charge
semipolar sets.
\item {\bf Bounded maximum principle $(M^*_{\alpha})$}: If $\nu$ is a finite measure
with compact support $K$ such that $U^{\alpha}\nu$ is bounded, then
$\sup\{U^{\alpha}\nu(x):x\in E\}=\sup\{U^{\alpha}\nu(x):x\in K\}$.
\item {\bf Bounded regularity principle $(R^*_{\alpha})$}: If $\nu$ is a finite measure
with compact support such that $U^{\alpha}\nu$ is bounded, then
$U^{\alpha}\nu$ is regular.
\item {\bf Polarity principle} ({\bf Hunt's hypothesis (H)}): Every semipolar set is polar.
\end{itemize}
\begin{pro}\label{thm1}
Assume that all 1-excessive (equivalently, all $\alpha$-excessive, $\alpha>0$) functions are lower
semicontinuous. Then
$$(P^*_{\alpha})\Leftrightarrow (E^*_{\alpha})
\Leftrightarrow(M^*_{\alpha})\Leftrightarrow(R^*_{\alpha})\Leftrightarrow ({\rm H}).
$$
\end{pro}
{\bf Proof.} $(R^*_{\alpha})\Leftrightarrow ({\rm H})$ is proved in Blumenthal and Getoor \cite{BG68} and $(M^*_{\alpha})\Leftrightarrow ({\rm H})$ is proved in Blumenthal and Getoor \cite{BG70}.
$(P^*_{\alpha})\Rightarrow (M^*_{\alpha})$ is proved in Rao \cite{R77} and
$(M^*_{\alpha})\Rightarrow (P^*_{\alpha})$ is proved in Fitzsimmons \cite{Fi90}.
By \cite[Propsition (2.1)]{BG70}, $(E^*_{\alpha})
\Rightarrow (M^*_{\alpha})$. By \cite[Proposition (5.1)]{BG70} and the equivalence of $(M^*_{\alpha})$ and $({\rm H})$, $(M^*_{\alpha})
\Rightarrow (E^*_{\alpha})$.\hfill\fbox
Hunt's hypothesis (H) is also equivalent to some other important
properties of Markov processes. For example, Blumenthal and Getoor
\cite[Proposition (4.1)]{BG70} and Glover \cite[Theorem
(2.2)]{G83} showed that (H) holds if and only if the fine and
cofine topologies differ by polar sets; Fitzsimmons and Kanda
\cite{FK} showed that (H) is equivalent to the dichotomy of
capacity.
In spite of its importance, (H) has been verified only in some
special situations. Some forty years ago, Getoor conjectured that
essentially all L\'{e}vy processes satisfy (H).
From now on we let $(\Omega,{\cal F},P)$ be a probability space
and $X=(X_t)_{t\ge 0}$ be an $\mathbf{R}^n$-valued L\'{e}vy
process on $(\Omega,{\cal F},P)$ with L\'{e}vy-Khintchine exponent
$\psi$, i.e.,
\begin{eqnarray*}
E[\exp\{i\langle z,X_t\rangle\}]=\exp\{-t\psi(z)\},\ z\in
\mathbf{R}^n,t\ge 0,
\end{eqnarray*}
where $E$ denotes the expectation {with respect to} $P$ and $\langle\cdot,\cdot\rangle$ denotes the Euclidean inner product of $\mathbf{R}^n$. The classical L\'{e}vy-Khintchine formula tells us that
\begin{eqnarray*}
\psi(z)=i\langle a,z\rangle+\frac{1}{2}\langle
z,Qz\rangle+\int_{\mathbf{R}^n} \left(1-e^{i\langle
z,x\rangle}+i\langle z,x\rangle 1_{\{|x|<1\}}\right)\mu(dx),
\end{eqnarray*}
where $a\in \mathbf{R}^n,Q$ is a symmetric nonnegative definite
$n\times n$ matrix, and $\mu$ is a measure (called the L\'evy
measure) on $\mathbf{R}^n\backslash\{0\}$ satisfying
$\int_{\mathbf{R}^n\backslash\{0\}} (1\wedge
|x|^2)\mu(dx)<\infty$. Hereafter, we use Re$(\psi)$ and Im$(\psi)$
to denote the real and imaginary parts of $\psi$, respectively,
and use $(a,Q,\mu)$ to denote $\psi$.
Let us recall some important results obtained so far for Getoor's
conjecture. When $n=1$, Kesten \cite{Ke69} {(cf. also Bretagnolle
\cite{Br71}) showed that if $X$ is not a compound Poisson
process, then every $\{x\}$ is non-polar} if and only if
\begin{eqnarray*}\label{Ke69-a}
\int_0^{\infty}\mbox{Re}([1+\psi(z)]^{-1})dz<\infty.
\end{eqnarray*}
Port and Stone \cite{PS69} proved that for the asymmetric
Cauchy process on the line every $x$ is regular for $\{x\}$, and thus (H) holds in this case. Further, Blumenthal and Getoor
\cite{BG70} showed that all stable processes with index
$\alpha\in (0,2)$ on the line satisfy (H).
Kanda \cite{Ka76} and Forst \cite{F75} proved that (H)
holds if $X$ has bounded continuous transition densities (with respect to the Lebesgue measure $dx$) and
the L\'{e}vy-Khintchine exponent $\psi$ satisfies $|\mbox{Im} (\psi)|\leq
M(1+\mbox{Re}(\psi))$ for some positive constant $M$.
Rao \cite{R77} gave a short proof of the Kanda-Forst theorem under the weaker condition that $X$ has resolvent densities.
In particular, for $n\ge 1$, all
stable processes with index $\alpha\neq 1$ satisfy (H). Kanda
\cite{Ka78} proved that (H) holds for stable processes on
$\mathbf{R}^n$ with index $\alpha= 1$ if we assume that the linear
term vanishes. Silverstein {\cite{Si77} extended the Kanda-Forst
condition to the non-symmetric Dirichlet forms setting,
Fitzsimmons \cite{Fi01} extended it to the semi-Dirichlet forms
setting and Han et al. \cite{HMS11} extended it to the
positivity-preserving forms setting. Glover and Rao \cite{GR86}
proved that $\alpha$-subordinates of general Hunt processes
satisfy (H) (cf. Theorem \ref{GR} below).
{Rao \cite{R88} proved that if all 1-excessive functions of $X$ are
lower semicontinuous and $|{\rm Im}(\psi)|\leq (1+{\rm
Re}(\psi))f(1+{\rm Re}(\psi))$, where $f$ is an increasing function
on $[1,\infty)$ such that $\int_N^{\infty}(\lambda f(\lambda))^{-1}d\lambda=\infty$ for
every $N\geq 1$, then $X$ satisfies (H).}
Let $X$ be a L\'{e}vy process on $\mathbf{R}^n$ with
L\'{e}vy-Khintchine exponent $(a,Q,\mu)$. In \cite{HS11}, we showed that if $Q$ is non-degenerate then $X$ satisfies (H); if $Q$ is degenerate then, under the assumption that $\mu({\mathbf{R}^n\backslash
\sqrt{Q}\mathbf{R}^n})<\infty$, $X$ satisfies (H) if
and only if the equation
$$
\sqrt{Q}y=-a-\int_{\{x\in {\mathbf{R}^n\backslash
\sqrt{Q}\mathbf{R}^n}:\,|x|<1\}}x\mu(dx)
$$
has at least one solution $y\in \mathbf{R}^n$. We also showed that if $X$ is a
subordinator and satisfies (H) then its drift coefficient must be 0.
In this paper, we will continue to explore (H) for L\'{e}vy
processes. The rest of the paper is organized as follows. In
Section 2, we present a comparison result on L\'{e}vy processes
which shows that big jumps have no effect on the validity of (H)
in some sense. Based on this result and the Kanda-Forst-Rao
theorem, in Section 3, we give examples of subordinators
satisfying (H). In Section 4, we give a new necessary and
sufficient condition for (H) and obtain an extended
Kanda-Forst-Rao theorem. By virtue of this theorem, we give a new
class of L\'{e}vy processes satisfying (H). In section 5, we
construct a type of subordinators that does not satisfy Rao's
condition. To the best of our knowledge, no existing criteria can
be applied to this example. It suggests that maybe new ideas and
methods are needed in order to completely solve Getoor's
conjecture even for the case of subordinators.
\section{A comparison result on L\'{e}vy processes}\setcounter{equation}{0}
In this section, we prove a comparison result on L\'{e}vy
processes which implies that big jumps have no effect on the
validity of (H) in some sense.
Let $X$ be a L\'{e}vy process on $\mathbf{R}^n$ with
L\'{e}vy-Khintchine exponent $(a,Q,\mu)$. Suppose that $\mu_1$
is a finite measure on $\mathbf{R}^n\backslash\{0\}$ such that
$\mu_1\leq \mu$. Denote $\mu_2:=\mu-\mu_1$ and let $X'$ be a
L\'{e}vy process on $\mathbf{R}^n$ with L\'{e}vy-Khintchine
exponent $(a',Q,\mu_2)$, where
\begin{eqnarray*}
a':=a+\int_{\{|x|<1\}}x\mu_1(dx).
\end{eqnarray*}
\begin{thm}\label{thm2.1}
Let $X$ and $X'$ be L\'{e}vy processes defined as above.
Then
\noindent (i) they have same semipolar sets.
\noindent (ii) they have same essentially polar sets.
\noindent (iii) if both $X$ and $X'$ have resolvent densities, then $X$
satisfies (H) if and only if $X'$ satisfies (H).
\end{thm}
{\bf Proof.} Denote by $\psi$ and $\psi'$ the L\'{e}vy-Khintchine
exponents of $X$ and $X'$, respectively. Then,
\begin{eqnarray}\label{thm2.1-a}
\psi'(z)&=&i\langle a',z\rangle+\frac{1}{2}\langle
z,Qz\rangle+\int_{\mathbf{R}^n} \left(1-e^{i\langle
z,x\rangle}+i\langle z,x\rangle
1_{\{|x|<1\}}\right)\mu_2(dx),\nonumber\\
\psi(z)&=&i\langle a,z\rangle+\frac{1}{2}\langle
z,Qz\rangle+\int_{\mathbf{R}^n} \left(1-e^{i\langle
z,x\rangle}+i\langle z,x\rangle
1_{\{|x|<1\}}\right)\mu(dx)\nonumber\\
&=&\psi'(z)+\int_{\mathbf{R}^n}\left(1-e^{i\langle
z,x\rangle}\right)\mu_1(dx).
\end{eqnarray}
(i) Suppose that $Y$ is a compound Poisson process with L\'{e}vy
measure $\mu_1$ and is independent of $X'$. By (\ref{thm2.1-a}),
$X$ has the same law as that of $X'+Y$. Let $T_1$ be the first
jumping time of $Y$. Then $T_1$ possesses an exponential
distribution and thus $P(T_1>0)=1$. Hence, for any set $A$ and any
point $x\in \mathbf{R}^n$, $x$ is a regular point of $A$ relative
to $X$ if and only if it is a regular point of $A$ relative to
$X'$. Therefore $X$ and $X'$ have same semipolar sets.
(ii) Set $C:=\mu_1(\mathbf{R}^n\backslash\{0\})$. By
(\ref{thm2.1-a}), we get
\begin{eqnarray}\label{sec2-3}
\mbox{Re}\psi'(z)\leq \mbox{Re}\psi(z)\leq
\mbox{Re}\psi'(z)+C
\end{eqnarray}
and
\begin{eqnarray}\label{sec2-4}
|\mbox{Im}\psi(z)|\leq |\mbox{Im}\psi'(z)|+C,\ \
|\mbox{Im}\psi'(z)|\leq |\mbox{Im}\psi(z)|+C.
\end{eqnarray}
For $\lambda>0$, we have
\begin{eqnarray}
\mbox{Re}\left(\frac{1}{\lambda+\psi(z)}\right)
&=&\frac{\lambda+\mbox{Re}\psi(z)}{(\lambda+\mbox{Re}\psi(z))^2+(\mbox{Im}\psi(z))^2},\label{sec2-5}\\
\mbox{Re}\left(\frac{1}{\lambda+\psi'(z)}\right)
&=&\frac{\lambda+\mbox{Re}\psi'(z)}{(\lambda+\mbox{Re}\psi'(z))^2+
(\mbox{Im}\psi'(z))^2}.\label{sec2-6}
\end{eqnarray}
By (\ref{sec2-3}) and (\ref{sec2-4}), we find that if $\lambda\geq
\sqrt{2}C$ then
\begin{eqnarray}\label{sec2-7}
\frac{\lambda+\mbox{Re}\psi(z)}{(\lambda+\mbox{Re}\psi(z))^2+(\mbox{Im}\psi(z))^2}
&\geq&
\frac{\lambda+\mbox{Re}\psi'(z)}{(\lambda+\mbox{Re}\psi'(z)+C)^2+
(|\mbox{Im}\psi'(z)|+C)^2}\nonumber\\
&\geq&\frac{\lambda+\mbox{Re}\psi'(z)}
{2[(\lambda+\mbox{Re}\psi'(z))^2+2C^2+(\mbox{Im}\psi'(z)^2]}\nonumber\\
&\geq&\frac{1}{4}\frac{\lambda+\mbox{Re}\psi'(z)}
{(\lambda+\mbox{Re}\psi'(z))^2+(\mbox{Im}\psi'(z))^2}.
\end{eqnarray}
Similar to (\ref{sec2-7}), we find that if $\lambda\geq 2C$ then
\begin{eqnarray}\label{sec2-8}
\frac{\lambda+\mbox{Re}\psi'(z)}{(\lambda+\mbox{Re}\psi'(z))^2+(\mbox{Im}\psi'(z))^2}
&\geq&
\frac{\lambda+\mbox{Re}\psi(z)-C}{(\lambda+\mbox{Re}\psi(z))^2+
(|\mbox{Im}\psi(z)|+C)^2}\nonumber\\
&\geq&\frac{\frac{1}{2}\lambda+\mbox{Re}\psi(z)}{(\lambda+\mbox{Re}\psi(z))^2+2C^2+
2(\mbox{Im}\psi(z))^2}\nonumber\\
&\geq&\frac{1}{4}\frac{\lambda+\mbox{Re}\psi(z)}{(\lambda+\mbox{Re}\psi(z))^2+(\mbox{Im}\psi(z))^2}.
\end{eqnarray}
By (\ref{sec2-5})-(\ref{sec2-8}), we obtain that if $\lambda\geq
2C$ then for any $z\in \mathbf{R}^n$,
\begin{eqnarray}\label{sec2-9}
\frac{1}{4}\ \mbox{Re}\left(\frac{1}{\lambda+\psi'(z)}\right)\leq
\mbox{Re}\left(\frac{1}{\lambda+\psi(z)}\right)\leq 4\
\mbox{Re}\left(\frac{1}{\lambda+\psi'(z)}\right).
\end{eqnarray}
By (\ref{sec2-9}) and Hawkes \cite[Theorem 3.3]{Ha79}, we obtain that a
set is essentially polar for $X$ if and only if it is essentially
polar for $X'$.
(iii) This is a direct consequence of (i), (ii) and \cite[Theorem 2.1]{Ha79}.\hfill\fbox
For $\delta>0$, we define
$$
B_{\delta}:=\{x\in\mathbf{R}^n:0<|x|<
\delta\}.
$$
\begin{cor}\label{cor2.2}
Let $X_{\delta}$ be a L\'{e}vy process on $\mathbf{R}^n$
with L\'{e}vy-Khintchine exponent
$(a_{\delta},Q,\mu|_{B_{\delta}})$, where
\begin{eqnarray*}
a_{\delta}:=\left\{
\begin{array}{ll}
a+\int_{\{\delta\leq |x|<1\}}x\mu(dx),&\ \mbox{if}\ \ 0<\delta<1,\\
a,&\ \mbox{if}\ \ \delta\geq 1.
\end{array}
\right.
\end{eqnarray*}
Then, all the assertions of Theorem \ref{thm2.1} hold with $X'$ replaced by
$X_{\delta}$.
\end{cor}
\begin{rem}\label{rem2.3}
If $\int_{|x|\leq 1}|x|\mu(dx)<\infty$, then $\psi$ can be
expressed by
$$\psi(z)=i\langle
d,z\rangle+\frac{1}{2}\langle z,Qz\rangle+\int_{\mathbf{R}^n}
\left(1-e^{i\langle z,x\rangle}\right)\mu(dx),
$$
where $-d$ is called the drift of $X$. In this case, we call $(d,
Q,\mu)$ the L\'{e}vy-Khintchine exponent of $X$. For $\delta>0$,
we define $B_{\delta}$ and $X_{\delta}$ as above. Let
$X_{\delta}'$ be a L\'{e}vy process on $\mathbf{R}^n$ with
L\'{e}vy-Khintchine exponent $(d, Q,\mu|_{B_{\delta}})$. We claim
that $X_{\delta}$ and $X_{\delta}'$ have the same law and then all
the assertions of Theorem \ref{thm2.1} hold with $X'$ replaced by
$X_{\delta}'$. In fact, we have
\begin{eqnarray}\label{rem2.2-a}
d=a+\int_{\{|x|<1\}}x\mu(dx).
\end{eqnarray}
If $0<\delta<1$, then
\begin{eqnarray}\label{rem2.2-b}
a_{\delta}+\int_{\{|x|<1\}}x\mu|_{B_{\delta}}(dx)=\left(a+\int_{\{\delta\leq
|x|<1\}}\mu(dx)\right)+\int_{\{|x|<\delta\}}x\mu(dx)=d;
\end{eqnarray}
if $\delta\geq 1$, then
\begin{eqnarray}\label{rem2.2-c}
a_{\delta}+\int_{\{|x|<1\}}x\mu|_{B_{\delta}}(dx)=a+\int_{\{|x|<1\}}x\mu(dx)=d.
\end{eqnarray}
By (\ref{rem2.2-a})-(\ref{rem2.2-c}), we know that $X_{\delta}$
and $X_{\delta}'$ have the same L\'{e}vy-Khintchine exponent $(d,
Q,\mu|_{B_{\delta}})$ and thus have the same law.
\end{rem}
\section{Examples of subordinators satisfying (H)}\setcounter{equation}{0}
In this section, we will present new examples of subordinators
satisfying (H) by virtue of the comparison result given in Section
2 and the Kanda-Forst-Rao theorem. To the best of our knowledge,
which subordinators satisfy (H) is unknown in general. To
appreciate the importance of the validity of (H) for
subordinators, let us recall the following remarkable result of
Glover and Rao.
\begin{thm}\label{GR} {\rm (Glover and Rao \cite{GR86})} Let $(X_t)_{t\ge 0}$ be a standard process on a locally compact space with a
countable base and $(T_t)_{t\ge 0}$ be an independent subordinator
satisfying Hunt's hypothesis (H). Then $(X_{T_t})_{t\ge 0}$
satisfies (H).
\end{thm}
Let $X$ be a subordinator. Then, its L\'{e}vy-Khintchine exponent
$\psi$ can be expressed by
$$\label{1-a}
\psi(z)=-idz+\int_{(0,\infty)}{\left(1-e^{izx}\right)}\mu(dx),\
z\in \mathbf{R},
$$
where $d\geq 0$ (called the drift coefficient) and $\mu$ satisfies
$\int_{(0,\infty)}(1\wedge x)\mu(dx)<\infty$. In \cite{HS11}, we
have proved the following result.
\begin{pro}\label{pro4.1}
If $X$ is a subordinator and satisfies (H), then $d=0$.
\end{pro}
By Proposition \ref{pro4.1}, when we consider (H) for
subordinators, we may concentrate on the case that $d=0$.
Hereafter we use $c_1, c_2,\dots$ to denote constants whose values
can change from one appearance to another.
\subsection{Special subordinators}
Let $X$ be a subordinator. Recall that the potential measure $U$
of $X$ is defined by
$$
U(A)=E\left[\int_0^{\infty}1_{\{X_t\in A\}}dt\right],\ A\subset
[0,\infty).
$$
For $\alpha>0$, the $\alpha$-potential measure $U^{\alpha}$ of $X$
is defined by
$$
U^{\alpha}(A)=E\left[\int_0^{\infty}e^{-\alpha t}1_{\{X_t\in
A\}}dt\right],\ A\subset [0,\infty).
$$
$X$ is called a {\it special subordinator} if $U|_{(0,\infty)}$
has a decreasing density with respect to the Lebesgue measure.
\begin{thm}\label{thm4.2}
Let $X$ be a special subordinator. Then $X$ satisfies (H) if and
only if $d=0$.
\end{thm}
{\bf Proof.} By Proposition \ref{pro4.1}, we need only prove the
sufficiency. Suppose that $d=0$. If $\mu$ is a finite measure,
then $X$ is a compound Poisson process and thus satisfies (H).
Now we consider the case that $\mu$ is an infinite measure. By
Bretagnolle \cite[Theorem 8]{Br71}, $X$ does not hit points, i.e., any single
point set $\{x\}$ is a polar set of $X$, which together with the
assumption that $U|_{(0,\infty)}$ has a decreasing density with
respect to the Lebesgue measure, implies that $U|_{[0,\infty)}$
has a density with respect to the Lebesgue measure. Since for any
$\alpha>0$, $U^{\alpha}(\cdot)\leq U(\cdot)$, we obtain that for
any $\alpha\geq 0$, $U^{\alpha}$ is absolutely continuous with
respect to the Lebesgue measure. Then by Hawkes \cite[theorem 2.1]{Ha79},
we know that for any $\alpha\geq 0$, all $\alpha$-excessive
functions are lower semicontinuous. Therefore, by the fact that
$X$ does not hit points and Blumenthal and Getoor \cite[Proposition (5.1), Theorem
(5.3)]{BG70}, following the same argument for stable subordinators
\cite[page 140]{BG70}, we obtain that $X$ satisfies
(H).\hfill\fbox
\subsection{Locally quasi-stable subordinators}
Let $S$ be a stable subordinator of index
$\alpha$, $0<\alpha<1$. Then, its L\'{e}vy-Khintchine exponent
$\psi_S$ has the form
$$\label{3.1}
\psi_S(z)=c|z|^{\alpha}(1-i\,\mbox{sgn}(z)\tan(\pi\alpha/2)),\
z\in (-\infty,\infty),
$$
where $c>0$. Its L\'{e}vy measure $\mu_S$ is absolutely continuous
with respect to the Lebesgue measure $dx$ and can be expressed by
\begin{eqnarray}\label{add001}
\mu_S(dx)=\left\{ \begin{array}{ll} c^+x^{-\alpha-1}dx,&\mbox{if}\
x>0,\\
0,&\mbox{if}\ x\le 0,
\end{array}
\right.
\end{eqnarray}
where $c^+>0$.
\begin{defi}\label{defi4.3}
Let $X$ be a subordinator with drift $0$ and L\'{e}vy measure
$\mu$. We call $X$ a {locally quasi-stable subordinator} if there
exist a stable subordinator $S$ with L\'{e}vy measure $\mu_S$,
positive constants $c_1,c_2,\delta$, and finite measures $\mu_1$
and $\mu_2$ on $(0,\delta)$ such that
$$c_1\mu_S-\mu_1\le
\mu\le c_2\mu_S+\mu_2\ \ {\rm on}\ (0,\delta). $$
\end{defi}
\begin{pro}\label{pro4.4}
Any locally quasi-stable subordinator satisfies (H).
\end{pro}
{\bf Proof.} Let $X,S,\mu_1$, $\mu_2$ and $\delta$ be as in
Definition \ref{defi4.3}. By Theorem \ref{thm2.1} and Remark
\ref{rem2.3}, we assume without loss of generality that
$\mu|_{[\delta,\infty)}=0$ and $\mu_1=0$. Denote by $\psi$ and
$\psi_S$ the L\'{e}vy-Khintchine exponents of $X$ and $S$,
respectively. Let $\mu_S$ be as in (\ref{add001}). Then
\begin{eqnarray}\label{pro5.4-a}
{\rm Re}\psi(z)&=&\int_0^{\infty}(1-\cos(zx))\mu(dx)\nonumber\\
&\ge&c_1\int_0^{\delta}(1-\cos(zx))\mu_S(dx)\nonumber\\
&=&c_1\left(\int_0^{\infty}(1-\cos(zx))\mu_S(dx)-\int_{\delta}^{\infty}(1-\cos(zx))\mu_S(dx)\right)\nonumber\\
&=&c_1{\rm Re}\psi_S(z)-K_1\nonumber\\
&=&c'|z|^{\alpha}-K_1,
\end{eqnarray}
where $c_1,c',K_1$ are positive constants.
\begin{eqnarray}\label{pro5.4-b}
|{\rm Im}\psi(z)|&\le&\int_0^{\infty}|\sin(zx)|\mu(dx)\nonumber\\
&\le&c_2\int_0^{\delta}|\sin(zx)|\mu_S(dx)\nonumber\\
&=&c_2\int_0^{\infty}|\sin(zx)|\mu_S(dx)-
c_2\int_{\delta}^{\infty}|\sin(zx)|\mu_S(dx)\nonumber\\
&\le&c_2c\left\{\int_0^{1/|z|}|\sin(zx)|x^{-1-\alpha}dx+
\int_{1/|z|}^{\infty}|\sin(zx)|x^{-1-\alpha}dx\right\}+K_2\nonumber\\
&\le&c_2c\left\{|z|\int_0^{1/|z|}x^{-\alpha}dx+\int_{1/|z|}^{\infty}x^{-1-\alpha}dx\right\}+K_2\nonumber\\
&=&c''|z|^{\alpha}+K_2,
\end{eqnarray}
where $c_2,c'',K_2$ are positive constants. By (\ref{pro5.4-a})
and (\ref{pro5.4-b}) we know that the Kanda-Forst condition holds
for $\psi$. By (\ref{pro5.4-a}) and Hartman and Wintner \cite{HW42}, we know that $X$
has bounded continuous transition densities. Therefore, $X$ satisfies (H) by the Kanda-Forst theorem.\hfill\fbox
\begin{cor}\label{cor4.5}
Let $\varphi$ be a L\'{e}vy-Khintchine exponent and $\mu$ be a
L\'{e}vy measure of some special subordinator with drift 0 or some
locally quasi-stable subordinator. Then, the L\'{e}vy process with
L\'{e}vy-Khintchine exponent
\begin{eqnarray}\label{cor3.5-a}
\Phi(z):=\int_{(0,\infty)}\left(1-e^{-\varphi(z)x}\right) \mu(dx)
\end{eqnarray}
satisfies (H).
\end{cor}
{\bf Proof.} Let $X$ be a L\'{e}vy process with
L\'{e}vy-Khintchine exponent $\varphi$ and $(T_t)_{t\ge 0}$ be a
subordinator with drift 0 and L\'{e}vy measure $\mu$, which is
independent of $X$. Then $Y_t:=X_{T_t}$ has the L\'{e}vy exponent
$\Phi$ defined by (\ref{cor3.5-a}). Therefore, by Theorem
\ref{GR}, Theorem \ref{thm4.2} and Proposition \ref{pro4.4}, we
obtain that $Y$ satisfies (H).\hfill\fbox
\subsection{Further examples}
In this subsection, we give further examples of subordinators
satisfying (H) by virtue of the comparison result given in
Section 2 and the following theorem of Rao.
\begin{thm}\label{R2} {\rm (Rao \cite{R88})} Let $X$ be a L\'evy process such that all 1-excessive functions are lower semicontinuous. Suppose there is an increasing function $f$ on $[1,\infty)$ such that
$ \int_N^{\infty}(\lambda f(\lambda))^{-1}d\lambda=\infty $ for
any $N\ge 1$ and $|1+\psi|\le (1+{\rm Re}(\psi))f(1+{\rm
Re}(\psi))$. Then (H) holds.
\end{thm}
Let $0<\alpha<1$ and $0<\delta<1$. We define
$$
\mu_T(dx):=\frac{1}{-\log(x)x^{1+\alpha}}dx,\ \ 0<x<\delta
$$
and
$$
\mu_V(dx)=\frac{-\log(x)}{x^{1+\alpha}}dx,\ \ 0<x<\delta.
$$
Let $X$ be a subordinator with drift 0 and L\'{e}vy measure
$\mu$.
\noindent (i) If $c_1\mu_T-\mu_1\le \mu\le c_2\mu_S+\mu_2$ on
$(0,\delta)$ for some positive constants $c_1,c_2$ and finite
measures $\mu_1,\mu_2$ on $(0,\delta)$, then $X$ satisfies (H).
In fact, by Theorem \ref{thm2.1} and Remark \ref{rem2.3}, we may
assume without loss of generality that $\mu|_{[\delta,\infty)}=0$
and $\mu_1=0$. For any $z\in \mathbf{R}$ with $|z|>1$, we have
\begin{eqnarray}\label{exm-1-a}
{\rm Re}\psi(z)&=&\int_0^{\infty}(1-\cos(zx))\mu(dx)\nonumber\\
&\ge&c_1\int_0^{\delta}(1-\cos(zx))\mu_T(dx)\nonumber\\
&=&c_1\int_0^{\infty}(1-\cos(zx))\mu_T(dx)-c_1\int_{\delta}^{\infty}(1-\cos(zx))\mu_T(dx)\nonumber\\
&\ge&c_1\int_{1/2|z|}^{1/|z|}(1-\cos(zx))\frac{1}{-\log(x)x^{1+\alpha}}dx-K_3\nonumber\\
&\ge&c_1'z^2\int_{1/2|z|}^{1/|z|}\frac{x^2}{-\log(x)x^{1+\alpha}}dx-K_3\nonumber\\
&\ge&c_1'\frac{z^2}{\log(2|z|)}\int_{1/2|z|}^{1/|z|}\frac{x^2}{x^{1+\alpha}}dx-K_3\nonumber\\
&=&c_1{''}\frac{|z|^{\alpha}}{\log(2|z|)}-K_3,
\end{eqnarray}
where $c_1',c_2'',K_3$ are positive constants. By (\ref{pro5.4-b})
and (\ref{exm-1-a}), we obtain that $|{\rm Im}\psi(z)|\le
c^*(1+{\rm Re}\psi(z))\log(1+{\rm Re}\psi(z))$ for some positive
constant $c^*$. By Hartman and Wintner \cite{HW42} and (\ref{exm-1-a}), we know that
$X$
has bounded continuous transition densities. Therefore, $X$ satisfies (H) by Theorem \ref{R2}.
\bigskip
\noindent (ii) If $c_1\mu_S-\mu_1\le \mu\le c_2\mu_V+\mu_2$ on
$(0,\delta)$ for some positive constants $c_1,c_2$ and finite
measures $\mu_1,\mu_2$ on $(0,\delta)$, then $X$ satisfies (H).
In fact, by Theorem \ref{thm2.1} and Remark \ref{rem2.3}, we may
assume without loss of generality that $\mu|_{[\delta,\infty)}=0$
and $\mu_1=0$. For any $z\in \mathbf{R}$ with $|z|>1/\delta$, we
have
\begin{eqnarray}\label{exm-2-a}
|{\rm
Im}\psi(z)|&\le&\int_0^{\infty}|\sin(zx)|\mu(dx)\nonumber\\
&\le& c_2\int_0^{\delta}|\sin(zx)|\mu_V(dx)
+K_4\nonumber\\
&\le&c_2\left\{\int_0^{1/|z|}|\sin(zx)|\frac{-\log(x)}{x^{1+\alpha}}dx+
\int_{1/|z|}^{\delta}|\sin(zx)|\frac{-\log(x)}{x^{1+\alpha}}dx\right\}+K_4\nonumber\\
&\le&c_2'\left\{|z|\int_0^{1/|z|}{\frac{-\log(x)}{x^{\alpha}}}dx+\log(|z|)
\int_{1/|z|}^{\infty}x^{-1-\alpha}dx\right\}+K_4\nonumber\\
&\le&c_2^{''}\left\{|z|\int_0^{1/|z|}{\frac{-(1-\alpha)\log(x)-1}
{x^{\alpha}}}dx+|z|^{\alpha}\log(|z|)\right\}+K_4\nonumber\\
&=&2c_2^{''}|z|^{\alpha}\log(|z|)+K_4,
\end{eqnarray}
where $c_2',c_2'',K_4$ are positive constants. By (\ref{pro5.4-a})
and (\ref{exm-2-a}), we obtain that
$|{\rm Im}\psi(z)|\le c^{**}{\rm Re}\psi(z)$ $\log({\rm
Re}\psi(z))$ for some positive constant $c^{**}$. By
(\ref{pro5.4-a}) and Hartman and Wintner \cite{HW42}, we know that $X$
has bounded continuous transition densities. Therefore, $X$ satisfies (H) by Theorem \ref{R2}.
\section{A new necessary and sufficient condition for (H) and an extended Kanda-Forst-Rao theorem}\setcounter{equation}{0}
Let $X$ be a L\'{e}vy process on $\mathbf{R}^n$. From now on we
assume that all 1-excessive functions are lower semicontinuous,
equivalently, $X$ has resolvent densities. Define
$$
A:=1+{\rm Re}(\psi),\ \ B:=|1+\psi|.
$$
\begin{thm}\label{thm112} {\rm (Rao \cite{R88})} Let $\nu$ be a finite measure of finite 1-energy, i.e.,
$$\int_{\mathbf{R}^n} B^{-2}(z)A(z)|\hat \nu(z)|^2dz<\infty.$$ Then
\begin{equation}\label{H1}
\lim_{\lambda\rightarrow\infty}\int_{\mathbf{R}^n}|\hat\nu(z)|^2(\lambda+{\rm Re}\psi(z))|\lambda+\psi(z)|^{-2}dz
\end{equation}
exists. The limit is zero if and only if $U^1\nu$ is regular.
\end{thm}
Based on Theorems \ref{thm112} and \ref{R2}, we can prove the
following result.
\begin{lem}\label{Thm33} Let
$\nu$ be a finite measure of finite 1-energy and $f$ be an
increasing function on $[1,\infty)$ such that $
\int_N^{\infty}(\lambda f(\lambda))^{-1}d\lambda=\infty $ for some
$N\ge 1$. Then $U^1\nu$ is regular if and only if
$$
\lim_{\lambda\rightarrow\infty}\sum_{k=1}^{\infty}\int_{\{B(z)>A(z)f(A(z)),\,k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,A(z)\le\lambda<(k+1)|{\rm Im}\psi(z)|\}}\frac{\lambda}{\lambda^2+({\rm Im}\psi(z))^2}|\hat \nu(z)|^2dz=0.
$$
\end{lem}
\noindent {\bf Proof.} Since $f$ is an increasing function on
$[1,\infty)$, $ \int_N^{\infty}(\lambda
f(\lambda))^{-1}d\lambda=\infty $ for some $N\ge 1$ if and only if
$ \int_N^{\infty}(\lambda f(\lambda))^{-1}d\lambda=\infty $ for
any $N\ge 1$. From the proof of Theorem \ref{R2} (see Rao \cite{R88}),
we know that the limit
\begin{equation}\label{gap2}
\lim_{\lambda\rightarrow\infty}\int_{A(z)\le\lambda}\frac{\lambda}{\lambda^2+B^2(z)}|\hat \nu(z)|^2dz
\end{equation}
exists and equals the limit in (\ref{H1}). We now show that the limit in (\ref{gap2}) equals 0 if and only if
\begin{equation}\label{H21}
\lim_{\lambda\rightarrow\infty}\int_{\{A(z)\le\lambda,\,B(z)>A(z)f(A(z))\}}\frac{\lambda}{\lambda^2+B^2(z)}|\hat \nu(z)|^2dz=0.
\end{equation}
To this end, we need only show that (\ref{H21}) implies that
\begin{equation}\label{gap}
\lim_{\lambda\rightarrow\infty}\int_{A(z)\le\lambda}\frac{\lambda}{\lambda^2+B^2(z)}|\hat \nu(z)|^2dz=0.
\end{equation}
Suppose that (\ref{H21}) holds. Then, the limit
$$
\lim_{\lambda\rightarrow\infty}\int_{\{A(z)\le\lambda,\,B(z)\le A(z)f(A(z))\}}\frac{\lambda}{\lambda^2+B^2(z)}|\hat \nu(z)|^2dz
$$
exists since the limit in (\ref{gap2}) always exists. Note that
\begin{eqnarray*}
& &\int_1^{\infty}\lambda^{-1}f(\lambda)^{-1}d\lambda\int_{\{A(z)\le\lambda,\,B(z)\le A(z)f(A(z))\}}\lambda(\lambda^2+B^2(z))^{-1}|\hat \nu(z)|^2dz\\
&=&\int_{\{B(z)\le A(z)f(A(z))\}}|\hat \nu(z)|^2dz\int_{A(z)}^{\infty}[f(\lambda)(\lambda^2+B^2(z))]^{-1}d\lambda\\
&\le&\frac{\pi}{2}\int_{\{B(z)\le A(z)f(A(z))\}}[B(z)f(A(z))]^{-1}|\hat \nu(z)|^2dz\\
&\le&\frac{\pi}{2}\int_{\mathbb{R}^d}B^{-2}(z)A(z)|\hat \nu(z)|^2dz\\
&<&\infty.
\end{eqnarray*}
Since $\int_1^{\infty}\lambda^{-1}f(\lambda)^{-1}d\lambda=\infty$,
$$\lim_{\lambda\rightarrow\infty}\int_{\{A(z)\le\lambda,\,B(z)\le A(z)f(A(z))\}}\frac{\lambda}{\lambda^2+B^2(z)}|\hat \nu(z)|^2dz=0.$$
Therefore, (\ref{gap}) holds by (\ref{H21}).
For each $k\in \mathbf{N}$, we have
\begin{eqnarray}\label{kljh1}
&&1_{\{k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,\lambda\ge (k+1)|{\rm Im}\psi(z)|\}}\frac{\lambda}{\lambda^2+({\rm Im}\psi(z))^2}|\hat \nu(z)|^2\nonumber\\
&\le&1_{\{k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,\lambda\ge (k+1)|{\rm Im}\psi(z)|\}}\frac{1}{\lambda}|\hat \nu(z)|^2\nonumber\\
&\le&\frac{1}{k+1}1_{\{k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1\}}\frac{|\hat \nu(z)|^2}{|{\rm Im}\psi(z)|}.
\end{eqnarray}
We assume without loss of generality that $f(1)=\sqrt{2}$. Note that $B(z)>A(z)f(A(z))$ implies that $B(z)\le\sqrt{2}|{\rm Im}\psi(z)|$. Then, we obtain by $\int_{\mathbf{R}^n} B^{-2}(z)A(z)|\hat \nu(z)|^2dz<\infty$ that
\begin{equation}\label{kljh2}
\sum_{k=1}^{\infty}\frac{1}{2(k+1)}\int_{\{B(z)>A(z)f(A(z)),\,k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1\}} \frac{|\hat \nu(z)|^2}{|{\rm Im}\psi(z)|}dz<\infty.
\end{equation}
By (\ref{kljh1}), (\ref{kljh2}) and the dominated convergence theorem, we get
$$
\lim_{\lambda\rightarrow\infty}\sum_{k=1}^{\infty}\int_{\{B(z)>A(z)f(A(z)),\,k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,\lambda\ge (k+1)|{\rm Im}\psi(z)|\}}\frac{\lambda}{\lambda^2+({\rm Im}\psi(z))^2}|\hat \nu(z)|^2dz=0.
$$
Therefore, the proof is complete by noting (\ref{H21}).\hfill\fbox
Note that if $\nu$ is a finite measure such that $U^1\nu$ is bounded then $\nu$ has finite 1-energy (cf. Rao \cite[page 622]{R88}). By Lemma \ref{Thm33} and Proposition \ref{thm1}, we obtain the following necessary and sufficient condition for (H).
\begin{thm}\label{thmnb} Let $f$ be an increasing
function on $[1,\infty)$ such that $ \int_N^{\infty}(\lambda
f(\lambda))^{-1}d\lambda=\infty $ for some $N\ge 1$. Then (H)
holds if and only if
\begin{eqnarray}\label{thm2.4-1}
\lim_{\lambda\rightarrow\infty}\sum_{k=1}^{\infty}\int_{\{B(z)>A(z)f(A(z)),\,k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,A(z)\le\lambda<(k+1)|{\rm Im}\psi(z)|\}}\frac{\lambda}{\lambda^2+({\rm Im}\psi(z))^2}|\hat \nu(z)|^2dz=0
\end{eqnarray}
for any finite measure $\nu$ with compact support such that $U^1\nu$ is bounded.
\end{thm}
\begin{rem}
Theorem \ref{thmnb} indicates that the validity of (H) is closely
related to the behavior of $\psi(z)$ where ${\rm Im}(\psi(z))$ is
not well controlled by ${\rm Re}(\psi(z))$, which is possible and
can be seen from the uniform motion on $\mathbf{R}$ and the
example given in Section 5.\end{rem}
By virtue of Theorem \ref{thmnb}, we obtain the following result extending the Kanda-Forst-Rao theorem on (H).
\begin{thm}\label{cor11}
(H) holds if the following extended Kanda-Forst-Rao condition ((EKFR) for short) holds:\\
(EKFR) There are two measurable functions $\psi_1$ and $\psi_2$ on $\mathbf{R}^n$ such that $\rm{Im}(\psi)=\psi_1+\psi_2$,
and
\begin{eqnarray}\label{vbn1}
&&\quad\quad\quad |\psi_1|\leq Af(A),\nonumber\\
&&\int_{{\mathbf{R}^n}}\frac{|\psi_2(z)|}{(1+{\rm Re}\psi(z))^2+({\rm Im}\psi(z))^2}dz<\infty,
\end{eqnarray}
where $f$ is an increasing function on $[1,\infty)$ such that $
\int_N^{\infty}(\lambda f(\lambda))^{-1}d\lambda=\infty $ for some
$N\ge 1$.
\end{thm}
\begin{rem}
If $\psi_2=0$, then the (EKFR) condition is just Rao's condition.
In particular, if $f=1$, then it is just the Kanda-Forst
condition. An integrability condition similar to (\ref{vbn1}) has
been used in Glover \cite[Theorem 3.1]{Gl81}.
\end{rem}
\noindent {\bf Proof of Theorem \ref{cor11}.}
By Theorem \ref{thmnb}, we need only show that the limit in (\ref{thm2.4-1}) equals 0. We assume without loss of generality that $f(1)=1/3$. Note that $B(z)>3\sqrt{2}A(z)f(A(z))$ implies that $|{\rm Im}\psi(z)|> A(z)$ and $|{\rm Im}\psi(z)|> B(z)/\sqrt{2}$, and $|\psi_2(z)|>2A(z)f(A(z))$ implies that $|\psi_2(z)|>|{\rm Im}\psi(z)|/2$. Then, by (\ref{vbn1}), the fact that $A(z)\leq c(1+|z|^2)$ for some positive constant $c$ and the dominated convergence theorem, we obtain that
\begin{eqnarray*}
&&\sum_{k=1}^{\infty}\int_{\{B(z)>3\sqrt{2}A(z)f(A(z)),\,k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,A(z)\le\lambda<(k+1)|{\rm Im}\psi(z)|\}}\frac{\lambda}{\lambda^2+({\rm Im}\psi(z))^2}|\hat \nu(z)|^2dz\\
&\le&\sum_{k=1}^{\infty}\int_{\{|{\rm Im}\psi(z)|>3A(z)f(A(z)),\,k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,A(z)\le\lambda<(k+1)|{\rm Im}\psi(z)|\}}\frac{1}{2|{\rm Im}\psi(z)|}|\hat \nu(z)|^2dz\\
&\le&\sum_{k=1}^{\infty}\int_{\{|\psi_2(z)|>2A(z)f(A(z)),\,k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,A(z)\le\lambda<(k+1)|{\rm Im}\psi(z)|\}}\frac{|\psi_2(z)|}{|{\rm Im}\psi(z)|^2}| \hat \nu(z)|^2dz\\
&\le&\sum_{k=1}^{\infty}\int_{\{k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,A(z)\le\lambda<(k+1)|{\rm Im}\psi(z)|\}}\frac{2|\psi_2(z)|}{B^2(z)}|\hat \nu(z)|^2dz\\
&\le&\sum_{k=1}^{\infty}\int_{\{k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,\lambda<(k+1)^2A(z)\}}\frac{2|\psi_2(z)|}{B^2(z)}|\hat \nu(z)|^2dz\\
&\le&\sum_{k=1}^{\infty}\int_{\{k\le \frac{|{\rm Im}\psi(z)|}{A(z)}<k+1,\,\lambda<c(k+1)^2(1+|z|^2)\}}\frac{2|\psi_2(z)|}{B^2(z)}|\hat \nu(z)|^2dz\\
&\rightarrow&0\ \ {\rm as}\ \lambda\rightarrow\infty.
\end{eqnarray*}
The proof is complete.\hfill\fbox
In the following, we give an application of Theorem \ref{cor11}.
\begin{thm}\label{new111} Let $\gamma>0$ and $X$ be
a L\'evy process on $ \mathbf{R}$ satisfying
\begin{equation}\label{new090}
\liminf_{|z|\rightarrow\infty}\frac{{\rm
Re}\psi(z)}{|z|\log^{\gamma}(|z|)}>0.
\end{equation}
Then $X$ satisfies (H).
\end{thm}
{\bf Proof.} By (\ref{new090}), we get
$$
\lim_{|z|\to\infty}\frac{{\rm Re}\psi(z)}{\log(1+|z|)}=\infty.
$$
Hence $X$ has bounded continuous transition densities by
Hartman and Wintner \cite{HW42}. Let $f(\lambda)=\log(\lambda)$ for $\lambda\in
[1,\infty)$ and set $\psi_1(z):=1_{\{|{\rm Im}\psi(z)|\le
A(z)f(A(z))\}}{\rm Im}\psi(z)$, $\psi_2(z):=1_{\{|{\rm
Im}\psi(z)|> A(z)f(A(z))\}}{\rm Im}\psi(z)$ for $z\in \mathbf{R}$.
Condition (\ref{new090}) implies that there exists a constant
$c>0$ such that
$$|\psi_2(z)|\ge c1_{\{|\psi(z)|>
A(z)f(A(z))\}}|z|\log^{1+\gamma}(|z|)$$ when $|z|$ is sufficiently
large. Therefore, (\ref{vbn1}) holds and the proof is complete by
Theorem \ref{cor11}.\hfill\fbox
\begin{exa} By Theorem \ref{new111}, Theorem \ref{thm2.1} and Corollary \ref{cor2.2}, we obtain a new class of 1-dimensional L\'{e}vy processes satisfying (H). Let $X$ be a L\'{e}vy process on $\mathbf{R}$ with
L\'{e}vy-Khintchine exponent $(a,Q,\mu)$. Suppose that there exist
constants $\gamma>0$, $0<\delta<1$, $c>0$, and a finite measure
$\mu'$ on $\{x\in \mathbf{R}^n: 0<|x|<\delta\}$ such that
$$
d\mu\geq \frac{c(-\log(|x|))^{\gamma}}{x^2}dx-d\mu' \ \ {\rm on}\
\{x\in \mathbf{R}: 0<|x|<\delta\}.$$ Similar to (\ref{exm-1-a}),
we can show that (\ref{new090}) holds. Then, $X$ satisfies (H).
Note that in this example it does not matter if $a$ or $Q$ equals
0.
Let $Y$ be another 1-dimensional L\'{e}vy process which is
independent of $X$. Theorem \ref{new111} implies that the
perturbed process $Y+X$ also satisfies (H).
\end{exa}
\begin{rem}
Blumenthal and Getoor introduced in \cite{BG61} the following
index $\beta''$ defined by
\begin{equation}\label{mnb}
\beta''=\sup\left\{\tau\geq 0:\frac{{\rm
Re}\psi(z)}{|z|^{\tau}}\to \infty\ \mbox{as}\ |z|\to
\infty\right\}.
\end{equation}
Let $X$ be a L\'evy process on $ \mathbf{R}$. Then, Theorem
\ref{new111} implies that (H) holds when $\beta''>1$. This result
is also a direct consequence of the following proposition.
\begin{pro}\label{pro-1} Let $X$ be
a L\'evy process on $ \mathbf{R}$. Suppose that
\begin{equation}\label{zxcv}
\liminf_{|z|\to \infty}\frac{|\psi(z)|}{|z|\log^{1+\gamma}|z|}>0
\end{equation}
for some constant $\gamma>0$. Then (H) holds.
\end{pro}
{\bf Proof.} Let $f\equiv 1$ and set $\psi_1(z):=1_{\{|{\rm
Im}\psi(z)|\le A(z)f(A(z))\}}{\rm Im}\psi(z)$,
$\psi_2(z):=1_{\{|{\rm Im}\psi(z)|> A(z)f(A(z))\}}{\rm Im}\psi(z)$
for $z\in \mathbf{R}$. Condition (\ref{zxcv}) implies that
\begin{eqnarray*}
\limsup_{|z|\to \infty}\left\{\frac{|\psi_2(z)|}{(1+{\rm
Re}\psi(z))^2+({\rm
Im}\psi(z))^2}\cdot|z|\log^{1+\gamma}|z|\right\}<\infty.
\end{eqnarray*}
Therefore, (\ref{vbn1}) holds and the proof is complete by Theorem
\ref{cor11}.\hfill\fbox
We remark that Proposition \ref{pro-1} can also be proved by
Theorem \ref{thm112}. In fact, the limit in (\ref{H1}) equals the
limit in (\ref{gap2}) and hence equals 0 by (\ref{zxcv}) and the
dominated convergence theorem.
\end{rem}
\section{A type of subordinators that does not satisfy Rao's condition}\setcounter{equation}{0}
As pointed out in Rao \cite{R88}, from the proof of Theorem \ref{R2}
it seems that the condition $B\le Af(A)$ is not far from being
necessary. In this section, however, we will construct a type of
subordinators that does not satisfy Rao's condition.
\subsection{Construction of the example}
We fix an $\alpha$ such that $\frac{1}{2}<\alpha<1$. In the
sequel, we define a function $\rho$ on $\mathbf{R}$ which will be
used as the density function of a L\'evy measure $\mu$.
First, we set $n_1=2$. Define a function $\rho_1$ on $\mathbf{R}$
as follows.
$$
\rho_1(x)=\frac{1}{x^{1+\alpha}},\ \ {\rm if}\ \frac{1}{2n^2_1}<x<\frac{1}{n^2_1};\ \ 0,\ {\rm otherwise}.
$$
We define $\mu_1(dx)=\rho_1(x)dx$ and denote by $\psi_1$ the
L\'evy-Khintchine exponent of $\mu_1$. Then, for $z\in
[\frac{n_1}{2}, 2n_1]$, we have
\begin{eqnarray}\label{jkl1}
{\rm Re}\psi_1(z)&=&\int_0^1(1-\cos(zx))\mu_1(dx)\nonumber\\
&\le&\frac{1}{2}\int_{1/2n_1^2}^{1/n_1^2}z^2x^2\frac{1}{x^{1+\alpha}}dx\nonumber\\
&\le&\frac{2n_1^{2\alpha-2}}{2-\alpha}\nonumber\\
&\le&2
\end{eqnarray}
and
\begin{eqnarray}\label{jkl2}
{\rm Im}\psi_1(z)&=&\int_0^1\sin(zx)\mu_1(dx)\nonumber\\
&=&\int_{1/2n_1^2}^{1/n_1^2}\sin(zx)\mu_1(dx)\nonumber\\
&\ge&\int_{1/2n_1^2}^{1/n_1^2}\frac{zx}{2x^{1+\alpha}}dx\nonumber\\
&\ge&\frac{1}{8}n_1^{2\alpha-1}.
\end{eqnarray}
We increase $n_1$ so that
$\frac{1}{8}n_1^{2\alpha-1}>\frac{6}{1-\alpha}$.
For any $z\in \mathbf{R}$, we have
\begin{equation}\label{21}
{\rm Re}\psi_1(z)=\int_0^1(1-\cos(zx))\mu_1(dx)\le
\int_{1/2n_1^2}^1\frac{1}{x^{1+\alpha}}dx\le
\frac{2^{\alpha}n_1^{2\alpha}}{\alpha}\le 4n_1^{2\alpha}
\end{equation}
and
\begin{equation}\label{22}
|{\rm Im}\psi_1(z)|\le\int_0^1|\sin(zx)|\mu_1(dx)\le
\int_{1/2n_1^2}^1\frac{1}{x^{1+\alpha}}dx\le
\frac{2^{\alpha}n_1^{2\alpha}}{\alpha}\le 4n_1^{2\alpha}.
\end{equation}
\vskip 0.3cm We choose an $n_2\in \mathbf{N}$ such that
$n^2_2>2n_1^2$. We define a function $\rho_2$ on $\mathbf{R}$ as
follows.
$$
\rho_2(x)=\frac{1}{x^{1+\alpha}},\ \ \mbox{if}\ \frac{1}{2n^2_2}<x<\frac{1}{n^2_2};\ \ 0,\ {\rm otherwise}.
$$
Note that there is no overlap between $\rho_1$ and $\rho_2$. We
define $\mu_2(dx)=\rho_2(x)dx$ and denote by $\psi_2$ the
L\'evy-Khintchine exponent of $\mu_2$. Then, similar to the
above, we can show that for $z\in [\frac{n_2}{2}, 2n_2]$
\begin{equation}\label{1}
{\rm Re}\psi_2(z)\le 2\ \ {\rm and}\ \ {\rm Im}\psi_2(z)\ge
\frac{1}{8}n_2^{2\alpha-1}\left(>\frac{6}{1-\alpha}\right).
\end{equation}
Note that for $z\in [\frac{n_1}{2}, 2n_1]$ we have
\begin{eqnarray}\label{p1}
{\rm Re}\psi_2(z)&=&\int_0^1(1-\cos(zx))\mu_2(dx)\nonumber\\
&\le&\frac{1}{2}\int_{1/2n_2^2}^{1/n_2^2}z^2x^2\frac{1}{x^{1+\alpha}}dx\nonumber\\
&\le&\frac{2n_1^2n_2^{2\alpha-4}}{2-\alpha}
\end{eqnarray}
and
\begin{eqnarray}\label{p2}
|{\rm Im}\psi_2(z)|&\le&\int_0^1|\sin(zx)|\mu_2(dx)\nonumber\\
&\le&\int_{1/2n_2^2}^{1/n_2^2}|\sin(zx)|\frac{1}{x^{1+\alpha}}dx\nonumber\\
&\le&\int_{1/2n_2^2}^{1/n_2^2}2n_1x\frac{1}{x^{1+\alpha}}dx\nonumber\\
&\le&\frac{2n_1n_2^{2\alpha-2}}{1-\alpha}.
\end{eqnarray}
We increase $n_2$ (with $n_1$ fixed) so that $n_2\ge
n_1^{5/(2-2\alpha)}$. By (\ref{p1}) and (\ref{p2}), we get
\begin{equation}\label{az1}
{\rm Re}\psi_2(z)\le\frac{2}{(1-\alpha)n_1^4},\ \ |{\rm
Im}\psi_2(z)|\le\frac{2}{(1-\alpha)n_1^4},\ \ z\in
\left[\frac{n_1}{2}, 2n_1\right].
\end{equation}
Then, by (\ref{jkl1}), (\ref{jkl2}) and (\ref{az1}), we obtain
that for $z\in [\frac{n_1}{2}, 2n_1]$,
\begin{equation}\label{r1}
{\rm Re}\psi_1(z)+{\rm Re}\psi_2(z)\le 2+\frac{2}{(1-\alpha)n_1^4}
\end{equation}
and
\begin{equation}\label{r2}
{\rm Im}\psi_1(z)+{\rm Im}\psi_2(z)\ge
\frac{1}{8}n_1^{2\alpha-1}-\frac{2}{(1-\alpha)n_1^4}.
\end{equation}
We further increase $n_2$ so that $n_2\ge
(96)^{1/(2\alpha-1)}n_1^{(4+2\alpha)/(2\alpha-1)}$ which ensures
that for any $z\in \mathbf{R}$ (cf. (\ref{21}), (\ref{22}) and
(\ref{1})),
\begin{equation}\label{az2}
{\rm Re}\psi_1(z)\le \frac{1}{3n_1^4}{\rm
Im}\psi_2\left(\frac{n_2}{2}\right),\ \ |{\rm Im}\psi_1(z)|\le
\frac{1}{3n_1^4}{\rm Im}\psi_2\left(\frac{n_2}{2}\right).
\end{equation}
By (\ref{1}) and (\ref{az2}), we obtain that for $z\in
[\frac{n_2}{2}, 2n_2]$,
\begin{equation}\label{q1}
{\rm Re}\psi_1(z)+{\rm Re}\psi_2(z)\le \frac{1}{3n_1^4}{\rm
Im}\psi_2\left(\frac{n_2}{2}\right)+2
\end{equation}
and
\begin{equation}\label{q2}
{\rm Im}\psi_1(z)+{\rm Im}\psi_2(z)\ge
\left(1-\frac{1}{3n_1^4}\right){\rm
Im}\psi_2\left(\frac{n_2}{2}\right).
\end{equation}
Define
\begin{equation}\label{q20}\vartheta:=\max\left\{\frac{5}{2-2\alpha},
\frac{4+2\alpha}{2\alpha-1}\right\}.\end{equation} We can set
$n_2$ to be $cn_1^{\vartheta}$, for some positive constant $c$
depending only on $\alpha$, such that (\ref{r1}), (\ref{r2}),
(\ref{q1}) and (\ref{q2}) hold.
For any $z\in \mathbf{R}$, we have
\begin{equation}\label{31}
{\rm Re}\psi_2(z)=\int_0^1(1-\cos(zx))\mu_2(dx)\le
\int_{1/2n_2^2}^1\frac{1}{x^{1+\alpha}}dx\le
\frac{2^{\alpha}n_2^{2\alpha}}{\alpha}\le4n_2^{2\alpha}
\end{equation}
and
\begin{equation}\label{32}
|{\rm Im}\psi_2(z)|\le\int_0^1|\sin(zx)|\mu_2(dx)\le
\int_{1/2n_2^2}^1\frac{1}{x^{1+\alpha}}dx\le
\frac{2^{\alpha}n_2^{2\alpha}}{\alpha}\le4n_2^{2\alpha}.
\end{equation}
\vskip 0.3cm We choose an $n_3\in \mathbf{N}$ such that
$n^3_2>2n_2^2$. We define a function $\rho_3$ on $\mathbf{R}$ as
follows.
$$
\rho_3(x)=\frac{1}{x^{1+\alpha}},\ \ \mbox{if}\ \frac{1}{2n^2_3}<x<\frac{1}{n^2_3};\ \ 0,\ {\rm otherwise}.
$$
Note that there is no overlap among $\rho_1$, $\rho_2$ and
$\rho_3$. We define $\mu_3(dx)=\rho_3(x)dx$ and denote by
$\psi_3$ the L\'evy-Khintchine exponent of $\mu_3$. Then, similar
to the above, we can show that for $z\in [\frac{n_3}{2}, 2n_3]$,
\begin{equation}\label{41}
{\rm Re}\psi_3(z)\le 2\ \ {\rm and}\ \ {\rm Im}\psi_3(z)\ge
\frac{1}{8}n_3^{2\alpha-1}
\end{equation}
and for any $z\in \mathbf{R}$,
$$
{\rm Re}\psi_3(z)\le4n_3^{2\alpha},\ \ |{\rm Im}\psi_3(z)|\le
4n_3^{2\alpha}.
$$
Similar to (\ref{p1}) and (\ref{p2}), we obtain that for $z\in
[\frac{n_1}{2}, 2n_1]$,
\begin{equation}\label{v1}
{\rm Re}\psi_3(z)\le\frac{2n_1^2n_3^{2\alpha-4}}{2-\alpha},\ \
|{\rm Im}\psi_3(z)|\le\frac{2n_1n_3^{2\alpha-2}}{1-\alpha}
\end{equation}
and for $z\in [\frac{n_2}{2}, 2n_2]$,
\begin{equation}\label{v2}
{\rm Re}\psi_3(z)\le\frac{2n_2^2n_3^{2\alpha-4}}{2-\alpha},\ \
|{\rm Im}\psi_3(z)|\le\frac{2n_2n_3^{2\alpha-2}}{1-\alpha}.
\end{equation}
We increase $n_3$ (with $n_1,n_2$ fixed) so that $n_3\ge
n_2^{5/(2-2\alpha)}$. By (\ref{v1}) and (\ref{v2}), we get
\begin{equation}\label{az22}
{\rm Re}\psi_3(z)\le\frac{2}{(1-\alpha)n_2^4},\ \ |{\rm
Im}\psi_3(z)|\le\frac{2}{(1-\alpha)n_2^4},\ \ z\in
\left[\frac{n_1}{2}, 2n_1\right]\bigcup\left[\frac{n_2}{2},
2n_2\right].
\end{equation}
Hence, by (\ref{r1}), (\ref{r2}) and (\ref{az22}), we obtain that
for $z\in [\frac{n_1}{2}, 2n_1]$,
\begin{equation}\label{rr1}
{\rm Re}\psi_1(z)+{\rm Re}\psi_2(z)+{\rm Re}\psi_3(z)\le
2+\frac{2}{(1-\alpha)n_1^4}+\frac{2}{(1-\alpha)n_2^4}
\end{equation}
and
\begin{equation}\label{rr2}
{\rm Im}\psi_1(z)+{\rm Im}\psi_2(z)+{\rm Im}\psi_3(z)\ge
\frac{1}{8}n_1^{2\alpha-1}-\frac{2}{(1-\alpha)n_1^4}-\frac{2}{(1-\alpha)n_2^4}.
\end{equation}
By (\ref{q1}), (\ref{q2}), (\ref{az22}) and (\ref{1}), we obtain
that for $z\in [\frac{n_2}{2}, 2n_2]$,
\begin{equation}\label{qq1}
{\rm Re}\psi_1(z)+{\rm Re}\psi_2(z)+{\rm Re}\psi_3(z)\le
\frac{2}{3n_1^4}{\rm
Im}\psi_2\left(\frac{n_2}{2}\right)+2+\frac{2}{(1-\alpha)n_2^4}
\end{equation}
and
\begin{equation}\label{qq2}
{\rm Im}\psi_1(z)+{\rm Im}\psi_2(z)+{\rm Im}\psi_3(z)\ge
\left(1-\frac{1}{3n_1^4}-\frac{1}{3n_2^4}\right){\rm
Im}\psi_2\left(\frac{n_2}{2}\right).
\end{equation}
We further increase $n_3$ so that $n_3\ge
(192)^{1/(2\alpha-1)}n_2^{(4+2\alpha)/(2\alpha-1)}$ which ensures
that for any $z\in \mathbf{R}$ (cf. (\ref{21}), (\ref{22}),
(\ref{31}), (\ref{32}) and (\ref{41})),
\begin{equation}\label{az2223}
{\rm Re}\psi_1(z),{\rm Re}\psi_2(z),|{\rm Im}\psi_1(z)|,|{\rm
Im}\psi_2(z)|\le \frac{1}{6n_2^4}{\rm
Im}\psi_2\left(\frac{n_3}{2}\right).
\end{equation}
Therefore, we obtain by (\ref{41}) and (\ref{az2223}) that for
$z\in [\frac{n_3}{2}, 2n_3]$,
\begin{equation}\label{cq1}
{\rm Re}\psi_1(z)+{\rm Re}\psi_2(z)+{\rm Re}\psi_3(z)\le
\frac{1}{3n_2^4}{\rm Im}\psi_3\left(\frac{n_3}{2}\right)+2
\end{equation}
and
\begin{equation}\label{cq2}
{\rm Im}\psi_1(z)+{\rm Im}\psi_2(z)+{\rm Im}\psi_3(z)\ge
\left(1-\frac{1}{3n_1^4}-\frac{1}{3n_2^4}\right){\rm
Im}\psi_3\left(\frac{n_3}{2}\right).
\end{equation}
We set $n_3$ to be $ 2^{1/(2\alpha-1)}cn_2^{\vartheta}$, where
$\vartheta$ and $c$ are as the same as above.
\vskip 0.3cm Continue in this way, we define $\rho_4,\rho_5,\dots$
All of these functions have no overlap and we have estimates
similar to (\ref{rr1})-(\ref{qq2}), (\ref{cq1}) and (\ref{cq2}).
Now we define
$$
\rho=\sum_{i=1}^{\infty}\rho_i.
$$
One finds that $\mu(dx)=\rho(x)dx$ is the L\'evy measure of a
subordinator $X$ with the L\'evy-Khintchine exponent
$$
\psi=\sum_{i=1}^{\infty}\psi_i.
$$
Moreover, we have that for $k\ge 2$,
\begin{equation}\label{hgf}
n_k=(k-1)^{1/(2\alpha-1)}cn_{k-1}^{\vartheta},
\end{equation} and for $z\in [\frac{n_k}{2},2n_k]$,
\begin{equation}\label{hgf7}
{\rm Im}\psi_k(z)\ge\frac{1}{8}n_k^{2\alpha-1},
\end{equation}
\begin{equation}\label{general1}
{\rm Re}\psi(z)\le \frac{1}{3n_{k-1}^4}{\rm
Im}\psi_k\left(\frac{n_k}{2}\right)+2+\frac{2}{1-\alpha}\sum_{k=1}^{\infty}\frac{1}{n_k^4},
\end{equation}
and
\begin{equation}\label{general2}
{\rm Im}\psi(z)\ge
\left(1-\frac{1}{3}\sum_{k=1}^{\infty}\frac{1}{n_k^4}\right){\rm
Im}\psi_k\left(\frac{n_k}{2}\right).
\end{equation}
\subsection{Discussions}
In this subsection, we make discussion about the subordinators
constructed in Subsection 5.1. Below we use $c_1, c_2,\dots$ to
denote positive constants depending only on $\alpha$.
\noindent \textbf{1.} By the estimates (\ref{general1}) and
(\ref{general2}), we can show that Rao's condition does not hold
for the subordinators. In fact, by (\ref{hgf}), there exists a
constant $c_1> 1$ such that
\begin{equation}\label{explain}
n_k>c_1^{c_1^k}, \ \ k\in \mathbf{N}. \end{equation}
By (\ref{hgf7}),
(\ref{general1}) and (\ref{general2}), we find that there exist
constants $c_2,c_3,c_4>0$ such that for any $k\ge2$,
\begin{equation}\label{add345}
\frac{{\rm Im}\psi(z)}{1+{\rm Re}\psi(z)}\ge c_2n_{k-1}^4\ge
c_3n_k^{3/\vartheta}\ge
c_3\left(\frac{z}{2}\right)^{3/\vartheta},\ \ \forall z\in[n_k/2,
2n_k].
\end{equation}
\begin{equation}\label{break1}
{\rm Re}\psi(z)\le c_4n_{k-1}^{\alpha\vartheta-3},\ \ \forall
z\in[n_k/2, 2n_k].
\end{equation}
The estimates (\ref{add345}) and (\ref{break1}) imply that there
does not exist an increasing function $f$ on $[1,\infty)$
satisfying $ \int_N^{\infty}(\lambda
f(\lambda))^{-1}d\lambda=\infty $ for some $N\ge 1$ and
$|1+\psi|\le (1+{\rm Re}(\psi))f(1+{\rm Re}(\psi))$. That is,
Rao's condition does not hold for the subordinators constructed in
Subsection 5.1.
By Theorem \ref{thm2.1}, we can modify the L\'evy measure $\mu$
defined in Subsection 5.1 by a finite measure and hence obtain a
subordinator which does not satisfy Rao's condition and whose
L\'evy measure $\mu$ has a smooth density $\rho$ with respect to
the Lebesgue measure on $(0,\infty)$.
\vskip 0.3cm \noindent \textbf{2.} Besides the index $\beta''$
(see (\ref{mnb})), Blumenthal and Getoor introduced also in
\cite{BG61} the indexes $\beta$ and $\sigma$ defined by
$$
\beta=\inf\left\{\tau>0:
\int_{\{|x|<1\}}|x|^{\tau}\mu(dx)<\infty\right\}
$$
and
$$
\sigma=\sup\left\{\tau\le 1:
\int_1^{\infty}\frac{x^{\tau-1}}{\int_0^{\infty}(1-e^{-xy})\mu(dy)}dx<\infty\right\}.
$$
From the construction of the subordinators given in Subsection
5.1, we obtain by \cite[Theorem 6.1]{BG61} that
$$
\sigma=\beta=\alpha.
$$
By (\ref{hgf}) and (\ref{general1}) (cf. (\ref{pro5.4-b})), we
get
$$
\beta''\le \alpha-\frac{4}{\vartheta}.
$$
\vskip 0.1cm \noindent \textbf{3.} Take $\alpha=3/4$. For the
subordinators constructed in Subsection 5.1, we claim that there
exists a finite signed measure $d\nu=g_1dx-g_2dx$ with $g_1,g_2\in
L^1_+(\mathbb{R};dx)$ such that
\begin{equation}\label{pop11}\int_{\mathbf{R}}
B^{-2}(z)A(z)|{\hat\nu}(z)|^2dz<\infty\end{equation}
but
\begin{equation}\label{pop12}
\lim_{\lambda\rightarrow\infty}\int_{\mathbf{R}}|{\hat{\nu}}(z)|^2(\lambda+{\rm
Re}\psi(z))|\lambda+\psi(z)|^{-2}dz=\infty.\end{equation}
Let $\omega$ be a sufficiently large number. We define
$$
\zeta_{\omega}(x):=\left\{1-\frac{1-1/({\omega})^{0.1}}{{\omega}}\cdot|x|\right\},
\ \ {\rm if}\ |x|\le\omega;\ \ \frac{1}{|x|^{0.1}},\ {\rm
otherwise},
$$
and
$$
\eta_{\omega}(x):=\left\{1-\frac{1-1/({\omega})^{0.1}}{{\omega}}\cdot|x|\right\}\vee
0, \ \ x\in \mathbb{R}.
$$
By Polya's theorem (cf. Lukacs \cite[Theorem 4.3.1]{Luka}), both
$\zeta_{\omega}$ and $\eta_{\omega}$ are characteristic functions
of absolutely continuous symmetric distributions. Define
$\varsigma_{\omega}:=\eta_{\omega}-\zeta_{\omega}$. Then, $
\varsigma_{\omega}(x)=0$ if $|x|\le{\omega}$; $
\varsigma_{\omega}(x)=1/|x|^{0.1}$ if $|x|\ge (1.1){\omega}$; and
$0\le \varsigma_{\omega}(x)\le 1/|x|^{0.1}$ otherwise.
Let $k_0\in \mathbf{N}$ be a sufficiently large number. For $k\ge
k_0$, we define
$\xi_k:=\varsigma_{\frac{n_k}{2}}-\varsigma_{\frac{2n_k}{1.1}}$.
We find that $\xi_k$ is a characteristic function of the
difference of two functions $g^k_1,g^k_2\in L^1_+(\mathbb{R};dx)$
with $\|g_1^k\|_{L^1},\|g_2^k\|_{L^1}\le 2$. Define
$g_1:=\sum_{k=1}^{\infty}g^k_1/2^k$,
$g_2:=\sum_{k=1}^{\infty}g^k_2/2^k$ and $d\nu:=g_1dx-g_2dx$. By
applying (\ref{q20}), (\ref{hgf}), (\ref{explain}) and the first
inequality of (\ref{add345}) to $B(z)/A(z)$ and applying
(\ref{hgf7}), (\ref{general2}) to $B(z)$, we find that there
exists a constant $c_5>0$ such that
\begin{eqnarray*}
\int_{\mathbf{R}}
B^{-2}(z)A(z)|{\hat\nu}(z)|^2dz
&=&\int_{\mathbf{R}}
\frac{1}{\frac{B(z)}{A(z)}\cdot B(z)}|{\hat\nu}(z)|^2dz\\
&\le&c_{5}\sum_{k=1}^{\infty}\frac{1}{n_k^{\frac{4}{\vartheta}-\frac{1}{22}}\cdot
n_k^{2\alpha-1}\cdot
2^{2k}}\int_{n_k/2}^{2n_k}\frac{1}{z^{0.2}}dz\\
&=&c_{5}\sum_{k=1}^{\infty}\frac{1}{n_k^{9/11}\cdot
2^{2k}}\int_{n_k/2}^{2n_k}\frac{1}{z^{0.2}}dz\\
&<&\infty.\end{eqnarray*}
However, there exists a constant $c_6>0$
such that (cf. ({\ref{pro5.4-b} and (\ref{explain}))
\begin{eqnarray*}
\int_{\mathbf{R}}
|{\hat\nu}(z)|^2\frac{n_k^{\alpha}}{(n_k^{\alpha})^2+({\rm
Im}\psi(z))^2}dz&\ge&c_{6}\frac{1}{n_k^{\frac{3}{4}}\cdot 2^{2k}}\int_{(0.55)n_k}^{\frac{2n_k}{1.1}}\frac{1}{z^{0.2}}dz\\
&\rightarrow&\infty\ \ {\rm as}\
k\rightarrow\infty,\end{eqnarray*} which implies (\ref{pop12}).
By (\ref{pop11}) and (\ref{pop12}) we can also conclude that Rao's
condition does not hold for the subordinators constructed in
Subsection 5.1. In fact, from the proof of Theorem \ref{R2} (see
Rao \cite{R88}), we can see that under Rao's condition, $$
\lim_{\lambda\rightarrow\infty}\int_{\mathbf{R}}|{\hat{\nu}}(z)|^2(\lambda+{\rm
Re}\psi(z))|\lambda+\psi(z)|^{-2}dz=0
$$
holds for any finite signed measure of finite 1-energy.
It is interesting to
compare (\ref{pop11}) and (\ref{pop12}) with the following result,
which is a consequence of Theorem \ref{thm112}.
\begin{thm}\label{slight} Let $X$ be a L\'evy process on $\mathbf{R}^n$ such that all 1-excessive functions are lower
semicontinuous. Then (H) holds if and only if
\begin{equation}\label{sx}
\lim_{\lambda\rightarrow\infty}\int_{\mathbf{R}^n}|\hat\nu(z)|^2(\lambda+{\rm
Re}\psi(z))|\lambda+\psi(z)|^{-2}dz=0
\end{equation}
for any finite measure $\nu$ of finite 1-energy.
\end{thm}
{\bf Proof.} By Theorem \ref{thm112}, Rao \cite[Remark, page 622]{R88}
and Blumenthal and Getoor \cite[VI. (4.8)]{BG68}, we need only prove the necessity.
Suppose that (H) holds for $X$. Let $\nu$ be a finite measure of
finite 1-energy and $\kappa$ be the standard Gaussian measure on
$\mathbf{R}^n$. Then, $\nu+\kappa$ has finite 1-energy, which
implies that
\begin{equation}\label{pop13}
\int_{\mathbf{R}^n}U^1(\nu+\kappa)d(\nu+\kappa)<\infty.
\end{equation}
By (\ref{pop13}), $\kappa(\{x:U^1(\nu+\kappa)(x)=\infty\})=0$.
Hence $U^1(\nu+\kappa)$ is locally integrable (with respect to the
Lebesgue measure $dx$) by \cite[VI. (2.3)]{BG68}. By (H) and
\cite[VI. (4.9)]{BG68}, we find that $U^1(\nu+\kappa)$ is regular.
Therefore, (\ref{sx}) holds by Theorem \ref{thm112} and the proof
is complete.\hfill\fbox
So far we have not been able to
prove or disprove that (H) holds for the subordinators constructed
in Subsection 5.1. This example suggests that maybe completely new
ideas and methods are needed for resolving Getoor's conjecture.
\bigskip
{ \noindent {\bf\large Acknowledgments} \vskip 0.1cm \noindent We
thank Professor Fengyu Wang for helpful comments that improved a
previous version of the paper. We are grateful to the support of
NNSFC, Jiangsu Province Basic Research Program (Natural Science
Foundation) (Grant No. BK2012720), and NSERC.}
|
\section{Introduction}
More than two decades after pioneering works \cite{coll, shur}, the phenomenology related to the deconfined phase of QCD, \textit{i.e.}~the quark-gluon plasma (QGP) is still a fascinating topic both experimentally and theoretically. On the experimental side, the QCD matter was or is studied in heavy-ion collisions (RHIC, SPS, FAIR, LHC). These experiments seem to show that the QGP behaves like a perfect fluid. On the theoretical side, the study of QCD at finite temperature deserves also much interest because it is a challenging problem in itself and because of the many connections with experiments.
\par The aim of this work is to study the thermodynamic features of QGP by resorting to a $T$-matrix approach. The power of this approach is that the bound states and scattering states of the system can be studied in a whole picture. Such an approach has already proved to give relevant results in the study of hadronic matter above the critical temperature of deconfinement ($T_c$) \cite{cabre06} but has not yet been applied to compute the equation of state (EoS). This observable will be performed here thanks to the Dashen, Ma and Bernstein's formulation of statistical mechanics in terms of the ${S}$-matrix (or ${T}$-matrix)~\cite{dashen}. Such a formulation is particularly well suited for systems whose microscopic constituents behave according to relativistic quantum mechanics. The QGP is indeed identified to a quantum gas of gluons and quarks, which are seen as the effective degrees of freedom propagating in the plasma. This assumption is actually common to all the so-called quasiparticle approaches \cite{quasip}, with the crucial difference that the use of a ${T}$-matrix formulation allows us to investigate the behavior of the QGP in a temperature range where it is strongly interacting. This strong interaction means here that bound states are expected to still survive above $T_c$.
\par Although the above formulation can be applied to the full QGP, this paper is dedicated to the description of the gluon plasma. Dealing with only one particle species simplifies drastically the problem while the main feature of the description, \textit{i.e.}~the explicit inclusion of interactions in a quasiparticle approach, is kept. Moreover, the pure gauge thermodynamic features (in particular, the EoS) are well-known in lattice QCD; This will allow an accurate comparison between our phenomenological approach and the lattice QCD calculations.
\par A particularity of this paper is the generalization of the formalism to any gauge groups, with a particular attention for SU($N$) and the large-$N$ limit, and for G$_2$. This group has originally attracted attention because, the center of G$_2$ being trivial, models relating deconfinement to the breaking of a center of symmetry are no longer valid as for SU($N$). However, it still exhibits a first-order phase transition as SU($N$) does \cite{pepe}. Hence, G$_2$ appears quite attractive from a theoretical point of view.
\par The paper is organized as follows. Sec.~II is dedicated to the presentation of the general quasiparticle approach based on the ${T}$-matrix formalism proposed in~\cite{dashen}. In Sec.~III, the model is particularized to a Yang-Mills plasma with the inclusion of 2-body interactions and, in Sec.~IV, useful analytic comments concerning the thermodynamic observables in the SU($N$) and G$_2$ cases are discussed. The model parameters are fixed in Sec.~V and the existence of the bound states inside the gluon plasma is discussed in Sec.~VI. In Sec.~VII, the computation of the EoS is presented. Finally, Sec.~VIII is devoted to the conclusions and perspectives.
\section{$T$-matrix formalism}\label{Tmatsec}
\subsection{Generalities}
The results of~\cite{dashen} can be summarized as follows: The grand potential $\Omega$, expressed as an energy density, of an interacting particle gas is given by (in units where $\hbar=c=k_B=1$).
\begin{equation}\label{pot0}
\Omega=\Omega_0+\sum_\nu\left[\Omega_\nu-\frac{{\rm e}^{\beta\vec \mu\cdot\vec N}}{2\pi^2\beta^2}\int^\infty_{M_\nu} \frac{d\epsilon}{4\pi i}\, \epsilon^2\, K_2(\beta\epsilon)\, \left. {\rm Tr}_\nu \left({\cal S}S^{-1}\overleftrightarrow{\partial_\epsilon}S \right)\right|_c\right].
\end{equation}
In the above equation, the first term, $\Omega_0$, is the grand potential of the free relativistic particles, \textit{i.e.} the remaining part of the grand potential if the interactions are turned off. The second term accounts for interactions in the plasma and is a sum running on all the species, the number of particles included, and the quantum numbers necessary to fix a channel. The set of all these channels is generically denoted $\nu$. The vectors $\vec \mu=(\mu_1, \mu_2,\dots)$ and $\vec N=(N_1,N_2,\dots)$ contain the chemical potentials and the particle number of each species taking part in a given scattering channel.
The contributions above and below the threshold\footnote{Within this approach, the threshold is the summation on the mass of all the particles included in a given channel $\nu$.} $M_\nu$ are separated. Below the threshold, one has $\Omega_\nu$ the grand potential coming from bound states, seen as free additional species in the plasma and appearing as poles of the ${S}$-matrix. Above the threshold, one has the scattering contribution, where the trace is taken in the center of mass frame of the channel $\nu$ and where $S$ is the ${S}$-matrix, depending in particular on the total energy $\epsilon$. The symmetrizer ${\cal S}$ enforces the Pauli principle when a channel involving identical particles is considered, and the subscript $c$ means that only the connected scattering diagrams are taken into account. Notice that $K_2(x)$ is the modified Bessel function of the second kind, that $\beta$ is linked to the temperature $T$ thanks to $\beta = 1/T$, and that the notation $A\overleftrightarrow{\partial_x} B=A(\partial_xB)-(\partial_xA)B$ is used.
By definition, $S=1-2\pi i\, \delta(\epsilon-H_0)\, {\cal T}$, where ${\cal T}$ is the off-shell $T$-matrix and where $H_0$ is the free Hamiltonian of the system. A convenient way to compute ${\cal T}$ is to solve the Lippmann-Schwinger equation for the off-shell $T$-matrix, schematically given by
\begin{equation}\label{ls}
{\cal T}=V+ V\, G_0\, {\cal T},
\end{equation}
with $G_0$ the free propagator and $V$ the interaction potential.
Once the ${T}$-matrix is known, two problems can be simultaneously addressed:
The existence of bound states in the plasma and its EoS. The $T$-matrix
formalism has the advantage of treating bound and scattering states on the same
footing, and is particularly suited for the present situation where we expect
bound states to become less and less bound when the temperature increases,
eventually crossing over and melting into the continuum. This dissociation
mechanism has been shown to provide considerable threshold enhancement effects
in heavy quark anti-quark correlation functions \cite{cabre06}.
The plasma EoS is obtained from (\ref{pot0}). Then, the pressure is simply given by
\begin{equation}
p=-\Omega ,
\end{equation}
and the other thermodynamic observables can derived from $p$. For example, the trace anomaly ($\Delta=e-3\,p$) and the entropy density ($s$) read
\begin{equation}
\Delta=-\frac{1}{\beta^3} \left[ \partial_\beta\left(\beta^4 p\right)\right] _{{\cal V} \text{,}\, \beta\vec \mu}, \quad s=-\beta^2\left[ \partial_\beta p\right]_{{\cal V} \text{,}\, \vec \mu}
\end{equation}
where ${\cal V}$ is the volume of the system.
For later convenience, the thermodynamic quantities will be normalized to the Stefan-Boltzmann pressure, which is defined as
\begin{equation}\label{psb}
p_{SB}=-\lim_{m_i\rightarrow 0}\Omega_0,
\end{equation}
$m_i$ being the masses of the particles propagating in the medium.
\subsection{Interaction potential}
The explicit computation of $\Omega$ obviously requires the knowledge of the on-shell $T$-matrix that can be derived in particular from (\ref{ls}). A key ingredient of the present approach is thus the potential $V$, encoding the interactions between the particles in the plasma. In the following, $V$ is chosen as pairwise: For a $n$-body channel, $V=\sum_{i<j} V_{ij}$ with $V_{ij}$ the potential between two particle species $i$ and $j$. Having in mind the building of an effective framework describing the deconfined phase of a non-abelian gauge theory, each particle composing the plasma should be in a given representation of the considered gauge (or color) group. It is therefore reasonable to assume that the potential $V$ between two particles in the representations $R_i$ and $R_j$ of the considered gauge group has the color-dependence of a (screened) one-gluon-exchange potential, that is, in momentum space,
\begin{equation}\label{V00}
V_{ij}=\tilde M_{R_i} \cdot \tilde M_{R_j} \alpha_S\, \bar v(\beta,\vec{q},\vec{q}\,'),
\end{equation}
where $\tilde M_{R}$ denotes the generator of the considered gauge algebra in the representation $R$, and where the real function $\bar v$ only depends on the temperature and two momenta (no dependence on the mass or other attributes of the particle). We keep the name gluon for the gauge particle even if the gauge group can formally be arbitrary. In the above definition, it has to be remembered that $\alpha_S=g^2/4\pi$ and that $g^2=\lambda/C_2^{ adj}$, $adj$ being the adjoint representation of the gauge group under study and $C_2^R$ being the value of the quadratic Casimir in the representation $R$. Note that in the case of SU($N$), $\lambda$ is the 't~Hooft coupling (fixed in the large-$N$ limit).
Introducing quadratic Casimirs, one can rewrite (\ref{V00}) as
\begin{equation}\label{V0}
V_{ij}=\frac{C^{{\cal C}}_2-C^{R_i}_2-C^{R_j}_2}{2}\, \alpha_S\, \bar v\equiv \kappa_{{\cal C};ij}\, v,
\end{equation}
with ${\cal C}$ the pair representation and
\begin{equation}\label{kappa0}
\kappa_{{\cal C};ij}=\frac{C^{{\cal C}}_2-C^{R_i}_2-C^{R_j}_2}{2\,C_2^{ adj}}.
\end{equation}
Again, the real function $v=v(\beta,\vec{q},\vec{q} \,')$ only depends on the temperature and on two momenta -- an explicit form for $v$ will be given later. The validity of the form (\ref{V0}) for $V_{ij}$ has partially been checked in pure gauge SU(3) lattice calculations, showing that the static potential between two sources, in different representations and bound in a color singlet, follows the Casimir scaling expected from a process of one-gluon-exchange type~\cite{gupta}. The peculiar scaling (\ref{V0}) also leads to a relevant large-$N$ behavior of the EoS when the gauge group SU($N$) is chosen, as it will be shown in Sec.~\ref{sunT}.
Among the various possible representations, the case where a singlet (denoted $\bullet$) appears in the tensor product $R_i\otimes R_j$ is particularly relevant: Since $C_2^\bullet=0$ and the other quadratic Casimirs are positive, the singlet is the most attractive channel in any two-body scattering process, so the most favorable one for the formation of bound states. Such two-particle bound states should presumably be the lowest-lying ones and, being color-singlet, would give rise to low-lying glueballs or mesons for instance.
\subsection{Born approximation} \label{Born}
The scattering term in (\ref{pot0}), given by
\begin{equation}\label{pot_s}
\Omega_s= - \sum_{\nu}\frac{{\rm e}^{\beta\vec\mu\cdot\vec N}}{2\pi^2\beta^2} \int^\infty_{M_\nu} \frac{d\epsilon}{4\pi i}\, \epsilon^2\, K_2(\beta\epsilon)\, \left. {\rm Tr}_\nu \left({\cal S}S^{-1}\overleftrightarrow{\partial_\epsilon}S \right)\right|_c ,
\end{equation}
can be considerably simplified by using the Born approximation, \textit{i.e.} by noticing that if the interactions are weak enough, ${\cal T}=V+{\rm O}(V^2)$. Such conditions are generally expected to be valid at high enough temperatures, where the typical interaction energy is small with respect to the typical thermal energy of the particles. Note also that, in some cases, this approximation can be relevant when the factor $\kappa_{{\cal C};ij}$ is negligible, irrespective of the temperature: Such cases will be encountered when the gauge group is SU($N$) (see Sec.~\ref{sunT}).
\par To the first order in $V$, (\ref{pot_s}) becomes
\begin{equation}\label{pot_s2}
\Omega_s = \sum_\nu \frac{{\rm e}^{\beta\vec\mu\cdot\vec N}}{2\pi^2\beta^2}\int^\infty_{M_\nu} d\epsilon \, \epsilon^2\, K_2(\beta\epsilon)\, \left. {\rm Tr}_\nu \,\partial_\epsilon (\delta(\epsilon -H_0) V )\right|_c + {\rm O}(V^2).
\end{equation}
Let us write explicitly $\nu=(n,\tilde \nu)$, where $n$ is the total number of particles involved in a given scattering channel, and where $\tilde \nu$ are the remaining quantum numbers. A useful remark to be done at this stage is that the pairwise structure of $V$ causes $ \left.V \right|_c$ to be always vanishing excepted in two-body channels. Here, at the Born approximation, $n$ is always equal to 2. Then, $\left. {\rm Tr}_\nu \, \partial_\epsilon (\delta(\epsilon -H_0) V) \right|_c= {\rm Tr}_{\tilde \nu} \,\kappa_{{\cal C},ij}\,\partial_\epsilon (\delta v)$, with $\delta=\delta(\epsilon-\epsilon_{ij}(q))$ and $\epsilon_{ij}(q)=\sqrt{q^2+m_i^2}+\sqrt{q^2+m_j^2}$. Note that the color channel ${\cal C}$ and the particles species $i$, $j$ are part of $\tilde \nu$. After having extracted from the trace the color and $J^{PC}$ dependences ($J^P$ if the charge conjugation is not relevant), one is led to
\begin{equation}\label{pot_s3}
\Omega_s= \sum_{(i,j)}\frac{{\rm e}^{\beta(\mu_{i} + \mu_j)}}{2\pi^2\beta^2}\sum_{J^{PC}} (2J+1) \sum_{{\cal C}} {\rm dim}\,{\cal C} \, \kappa_{{\cal C},ij} \int^\infty_{M_{\tilde\nu}} d\epsilon \, \epsilon^2\, K_2(\beta\epsilon)\, {\rm Tr}_q \,\partial_\epsilon (\delta \, v_{J^{PC}} ) +{\rm O}(V^2),
\end{equation}
with ${\rm dim}\,{\cal C}$ is the pair representation dimension, ${\rm Tr}_{q}$ the remaining trace on the momentum space and $v_{J^{PC}}$ the potential with the angular symmetry of the considered channel.
\par Among the various summations to be performed in $\sum_{\tilde \nu}$, two are of particular interest: The one over the different interacting species, that can be denoted $\sum_{(i,j)}$, and the one over the representations appearing in $R_i\otimes R_j$, that is $\sum_{{\cal C}}$. Because of $\kappa_{{\cal C},ij}$, (\ref{pot_s3}) is thus proportional to a factor $\sum_{{\cal C}}{\rm dim}\,{\cal C} \ \kappa_{{\cal C},ij}$ for a given pair $i$, $j$ in a given $J^{PC}$ channel. When the combinations of species does not have to respect a symmetry principle, this last sum runs on all the representations appearing in $R_i\otimes R_j$; one can then show that
\begin{equation}\label{group}
\sum_{{\cal C}}{\rm dim}\,{\cal C} \ \kappa_{{\cal C},ij}=0.
\end{equation}
Indeed, it is known in group theory that the second order Dynkin indices $I^ R$ in a tensor product obey a sum rule that can be rewritten using our notations as $I^{R_i}\, {\rm dim}\,R_j+I^{R_j}\, {\rm dim}\,R_i=\sum_{{\cal C}} I^{{\cal C}}$ \cite{group}. Using $C_2^{R}=({\dim }\,{adj} /{\dim }\,R) I^R$ \cite{group}, one straightforwardly shows that (\ref{group}) holds. Note that (\ref{group}) and (\ref{pot_s3}) are thus \textit{a priori} nonzero when a symmetry principle has to be respected: The summation cannot then be performed on all possible color representations.
\section{Yang-Mills plasma}\label{YM}
\subsection{Grand canonical potential}
Let us now particularize the general formalism presented in the previous section to a genuine Yang-Mills plasma, \textit{i.e.}~with no matter fields. The bosonic degrees of freedom propagating in the plasma are then quasigluons, that are transverse spin-1 bosons in the adjoint representation of the gauge group. The baryonic potential can be set equal to 0 and according to standard formulas in statistical mechanics, one has
\begin{equation}\label{pg}
\Omega_0=2 \, {\dim }\, adj\, \omega_0(m_g),
\end{equation}
where the quasigluons are \textit{a priori} supposed to have a mass $m_g$, and where
\begin{equation}\label{oo}
\omega_0(m)=\frac{1}{2\pi^2\beta}\int^\infty_0dk\, k^2 \ln\left(1-{\rm
e}^{-\beta\sqrt{k^2+m^2}}\right)
\end{equation}
is the grand potential per degree of freedom associated to a bosonic species with mass $m$. Equation (\ref{psb}) leads to
\begin{equation}\label{psbg}
p_{SB}=\frac{\pi^2}{45 \beta^4}{\rm dim}\, adj.
\end{equation}
Let us recall that in the following, the term gluon indifferently denotes the gauge field of Yang-Mills theory and the quasigluons.
The sum $\sum_\nu$ appearing in (\ref{pot0}) now explicitly reads $\sum_{n_g}\sum_{{\cal C}}\sum_{J^{PC}}$, where $n_g$ is the number of gluons involved in the interaction process. As soon as $n_g>2$, the determination of the allowed color channels and of the correct symmetrized gluon states generally becomes a painful task, to which the problem of finding the $T$-matrix in many-body scattering must be added. Intuitively, one can nevertheless expect the dominant scattering processes to be two-gluon ones, and thus only consider $n_g=$ 2 in a first approach. After simplification, the grand potential (\ref{pot0}) eventually reads
\begin{eqnarray}\label{omega2}
&&\Omega^{(2)}=2\,{\dim }\, adj\,\omega_0(m_g)+\sum_{{\cal C}}\sum_{J^{P}}{\rm dim}\, {\cal C}\, (2J+1) \Bigg\lbrace\omega(M_{{\cal C},J^{P}})
\\
&& + \frac{1}{2\pi^2\beta^2}\int^\infty_{2m_g} d\epsilon \, \epsilon^2\, K_2(\beta\epsilon)\, {\rm Tr}_{{\cal C},J^{P}}\Big[ ( \delta {\rm Re}{\cal T })'- 2\pi \big( (\delta {\rm Re}{\cal T}) (\delta {\rm Im}{\cal T})'-(\delta {\rm Im} {\cal T}) (\delta {\rm Re}{\cal T})' \big) \Big] \Bigg\rbrace , \nonumber
\end{eqnarray}
where the symbol ``prime" is the derivative respective to the energy and $M_{{\cal C},J^{P}}$ is the mass of the two-gluon bound state with color ${\cal C}$ and quantum numbers $J^{P}$, if it exists. The index $C$ in the $J^{PC}$ channel is dropped since the charge conjugation is always positive for a two-gluon state \cite{Boul}. In the remaining trace, it is understood that the $T$-matrix has been computed in a given two-body channel with color ${\cal C}$ and quantum numbers $J^{P}$, and that the Dirac $\delta$ reads $\delta(\epsilon-2\epsilon(q))$, with the dispersion relation $\epsilon(q)=\sqrt{q^2+m_g^2}$. Note also that, in connection with nuclear many-body approaches, (\ref{omega2}) can be rewritten in terms of a weighted thermal average of scattering phase shifts by means of unitarity of the $S$-matrix. The computation of the two-gluon $T$-matrix is explained in detail in the following section.
\subsection{Helicity states and the Lippman-Schwinger equation}\label{helicity}
\subsubsection{Two gluon states}
Jacob and Wick's helicity formalism~\cite{jaco} can be applied to describe a two-gluon state, where the gluons are seen as transverse spin-1 particles. Let us generically define $\left|\psi(\vec p,\lambda)\right\rangle=a^{\dagger}_{\lambda}(\vec p\,)\left|0\right\rangle$ the quantum state of a particle with momentum $\vec p$, spin $s$, and helicity $\lambda$. If the particle is transverse, only $\lambda=\pm s$ is allowed, while all the projections from $-s$ to $+s$ are allowed if the particle has a usual spin degree of freedom. Then it can be deduced from~\cite{jaco} that the quantum state
\begin{equation}\label{hstate}
\left|J^P,M;\lambda_1,\lambda_2,\eta\right\rangle=\frac{1}{\sqrt 2} \Big[ \left|J,M;\lambda_1,\lambda_2\right\rangle+\eta\left|J,M;-\lambda_1,-\lambda_2\right\rangle\Big],
\end{equation}
with $\eta=\pm1$ and
\begin{eqnarray}
\left|J,M;\lambda_1,\lambda_2\right\rangle&=&\left[\frac{2J+1}{4\pi}\right]^{\frac{1}{2}}\int^{2\pi}_0d\phi\int^\pi_0d\theta\, \sin\theta\ {\cal D}^{J*}_{M,\lambda_1-\lambda_2}(\phi,\theta,-\phi)\, R(\phi,\theta,-\phi)\nonumber\\
&&\qquad\qquad\qquad\qquad\qquad\times \, a^{\dagger}_{\lambda_1}(\vec p\,)a^{\dagger}_{\lambda_2}(-\vec p\,)\left|0\right\rangle,
\end{eqnarray}
is a two-particle helicity state in the rest frame of the system which is also an eigenstate of the total spin $\vec J$ and of the parity, \textit{i.e.} $\vec J^{\, 2}=J(J+1)$, $J_z=M$, and $P=\epsilon\, \eta_1\eta_2(-1)^{J-s_1-s_2}$ with $\eta_i$ and $s_i$ the intrinsic parity and spin of particle $i$. Moreover, $J\geq|\lambda_1-\lambda_2|$. In the above definition, $R(\alpha,\beta,\gamma)$ is the rotation operator of Euler angles $\{\alpha,\beta,\gamma\}$ and ${\cal D}^{J}_{M,\lambda}(\alpha,\beta,\gamma)$ are the Wigner $D$-matrices. The coordinates $\{\theta,\phi\}$ are the polar angles of $\vec p$. When both particles have a spin degree of freedom, the helicity basis, spanned by the helicity states~(\ref{hstate}), is equivalent to a standard $\left|^{2S+1}L_J\right\rangle$ basis up to an orthogonal transformation \cite{gie}. When at least one of the particles is transverse, both basis are no longer equivalent, but the helicity states can still be expressed as particular linear combinations of $\left|^{2S+1}L_J\right\rangle$ states \cite{jaco}. This will be convenient in view of future computations.
When the two particles are identical ($m_1=m_2=m$, $s_1=s_2=s$), it is relevant to study the action of the permutation operator $P_{12}$. One finds~\cite{jaco}
\begin{equation}\label{symdef}
\left[1+(-1)^{2s}P_{12}\right]\left|J^P,M;\lambda_1,\lambda_2,\eta\right\rangle=\left|J^P,M;\lambda_1,\lambda_2,\eta\right\rangle+(-1)^{J}\left|J^P,M;\lambda_2,\lambda_1,\eta\right\rangle,
\end{equation}
where the operator $\left[1+(-1)^{2s}P_{12}\right]={\cal S}$ is nothing else than a projector on the symmetric ($s$ integer) or antisymmetric ($s$ half-integer) part of the helicity state. It is readily seen in (\ref{symdef}) that symmetrizing the state will eventually lead to selection rules for $J$ (this is particularly clear if one sets $\lambda_1=\lambda_2$). When extra degrees of freedom are added, it is also of interest to use the antisymmetrizer $\left[1-(-1)^{2s}P_{12}\right]={\cal A}$ as done in Table~\ref{tabs}.
A general discussion about the two-gluon helicity states can be found in~\cite{heli}, to which we refer the interested reader. For the present work, it is sufficient to recall that four families of helicity states can be found, separated in helicity singlets $\left|S_\pm; J^P\right\rangle$ and doublets $\left|D_\pm; J^P\right\rangle$ following the pioneering work~\cite{barnes}. The corresponding quantum numbers are given in Table~\ref{tabs}, as well as the average value of the squared orbital angular momentum, $\left\langle \vec L^2\right\rangle$, computed with these states.
\begin{table}[t]
\caption{Symmetrized and antisymmetrized two-gluon helicity states, following the notation of~\cite{heli,barnes}, with the corresponding quantum numbers and averaged squared orbital angular momentum.}
\begin{tabular}{c|ccc}
State & Symmetrized & Antisymmetrized & $\left\langle \vec L^2\right\rangle$\\
\hline
$\left|S_+; J^P\right\rangle$ & (even-$J\geq0$)$^+$ & (odd-$J\geq1$)$^-$ & $J(J+1)+2$ \\
$\left|S_-; J^P\right\rangle$ & (even-$J\geq0$)$^-$ & (odd-$J\geq1$)$^+$ & $J(J+1)+2$ \\
$\left|D_+; J^P\right\rangle$ & (even-$J\geq2$)$^+$ & (odd-$J\geq3$)$^-$ & $J(J+1)-2$ \\
$\left|D_-; J^P\right\rangle$ & (odd-$J\geq3$)$^+$ & (even-$J\geq2$)$^-$ & $J(J+1)-2$ \\
\end{tabular}
\label{tabs}
\end{table}
The averaged orbital angular momentum is an interesting quantity since it helps to globally understand the mass hierarchy of the glueball spectrum~\cite{heli}. Moreover, in a naive nonrelativistic picture, it estimates the strength of the orbital barrier in scattering theory. For obvious numerical reasons, all the possible $J^P$ channels contributing to $\Omega$ can not be included, that is why it is of interest to find the channels that will presumably contribute the most, \textit{i.e.} those with the lowest value of $\left\langle \vec L^2\right\rangle$. First, one has the symmetric states
\begin{eqnarray}\label{scaps}
\left|S_+;0^{+}\right\rangle&=&\left[\frac{2}{3}\right]^{1/2}\left|^1 S_0\right\rangle+\left[\frac{1}{3}\right]^{1/2}\left|^5 D_0\right\rangle,\\
\left|S_-;0^{-}\right\rangle&=&-\left|^3P_0\right\rangle,
\end{eqnarray}
expressed in a standard $\left|^{2S+1}L_J\right\rangle$ basis, with $\left\langle \vec L^2\right\rangle=2$. In the singlet channel, they correspond to the $0^{++}$ and $0^{-+}$ glueballs respectively, which are indeed found to be among the lightest ones at zero temperature, see \textit{e.g.} the review~\cite{rev}. Then, with $\left\langle \vec L^2\right\rangle=4$, one has the symmetric state
\begin{eqnarray} \label{tensor}
\left|D_+;2^{+}\right\rangle&=&\left[\frac{2}{5}\right]^{1/2}\left|^5S_2\right\rangle
+\left[\frac{4}{7}\right]^{1/2}\left|^5 D_2\right\rangle+\left[\frac{1}{35}\right]^{1/2}\left|^5 G_2\right\rangle,
\end{eqnarray}
and the antisymmetric states
\begin{eqnarray}
\left|S_+;1^{-}\right\rangle&=&\left[\frac{2}{3}\right]^{1/2}\left|^1 P_1\right\rangle-\left[\frac{2}{15}\right]^{1/2}\left|^5 P_1\right\rangle+\left[\frac{1}{5}\right]^{1/2}\left|^5 F_1\right\rangle,\\
\left|S_-;1^{+}\right\rangle&=&\left[\frac{1}{3}\right]^{1/2}\left|^3 S_1\right\rangle-\left[\frac{2}{3}\right]^{1/2}\left|^3 D_1\right\rangle,\\
\left|D_-;2^{-}\right\rangle&=&-\left[\frac{4}{5}\right]^{1/2}\left|^5 P_2\right\rangle-\left[\frac{1}{5}\right]^{1/2}\left|^5 F_2\right\rangle .
\end{eqnarray}
The above three states do not exist in the singlet channel, but the symmetric $2^+$ state corresponds to the $2^{++}$ glueball in the singlet channel; the $0^{\pm+}$ and $2^{++}$ are indeed the lightest states at zero temperature~\cite{rev}. Only the color-symmetric channels with the lowest value of $\left\langle \vec L^2\right\rangle$ will be kept in the following study, which aims at being a first step toward a description of the Yang-Mills plasma within a $T$-matrix formulation.
\subsubsection{Lippman-Schwinger equation}
Solving (\ref{ls}) is a crucial technical part of this work since it eventually leads to the on-shell $T$-matrix. As it will be discussed in Sec.\ref{param}, the potential to be used is known in position space and has firstly to be Fourier-transformed. For a potential with spherical symmetry in configuration space, we use
\begin{equation}\label{Vqq}
V(q,q',\theta_{q,q'}) = 4 \pi \displaystyle\int_0^\infty dr \, r V(r) \displaystyle\frac{\sin(Qr)}{Q}, \quad {\rm where}\quad Q = \sqrt{q^2 + q'^2 - 2 q q' \cos \theta_{q,q'}}\text{,}
\end{equation}
and where $\theta_{q,q'}$ is the angle between the momenta $\vec q$ and $\vec q\,'$.
Since two-gluon interactions are considered, the basis states are two-gluon helicity states, given in the above section. As we assume $V$ to be spin independent (see Appendix~\ref{appU1}), only the orbital angular momentum containing of the helicity states has to be taken into account. According to a standard integration, the $L$-wave part of potential (\ref{Vqq}) reads
\begin{equation}
V_L(q,q') = 2\pi \displaystyle\int_{-1}^1 dx P_L(x) V(q,q',x),
\end{equation}
where $P_L$ is the Legendre polynomial of order $L$ and $x = \cos \theta_{q,q'}$. Our choice is to focus on the scalar, pseudoscalar and tensor scattering channels, for which one can compute from (\ref{scaps})-(\ref{tensor}) that
\begin{eqnarray}
V_{0^{+}}(q,q') &=& \displaystyle\frac{2}{3} V_0(q,q') + \displaystyle\frac{1}{3} V_2(q,q')\text{,}\\
V_{0^{-}}(q,q') &=& V_1(q,q')\text{,} \\
V_{2^{+}}(q,q') &=& \displaystyle\frac{2}{5} V_0(q,q') + \displaystyle\frac{4}{7} V_2(q,q') + \displaystyle\frac{1}{35} V_4(q,q') \text{.}
\end{eqnarray}
Note that once $V_{J^P}(q,q')$ is known, that is the potential in a given $J^P$ scattering channel, the off-shell $T$-matrix can be computed from (\ref{ls}) as follows in \cite{cabre06}:
\begin{equation}\label{tosolve}
{\cal T}(E; q,q') = V_{J^P}(q,q') + \displaystyle\frac{1}{8\pi^3} \displaystyle\int_0^\infty dk\, k^2\, V_{J^P}(q,k) \,G_0(E;k)\, {\cal T}(E;k,q'),
\end{equation}
where the two-gluon propagator reads
\begin{equation}
G_0(E;k) = \displaystyle\frac{m_g^2}{\epsilon(k)} \displaystyle\frac{1}{E^2/4- \epsilon(k)^2 - 2i \, \epsilon(k)\, \Sigma_I}
\end{equation}
with the gluon dispersion relation $\epsilon(k)= \displaystyle\sqrt{k^2+m_g^2}$. Note that the normalization conventions of the $T$-matrix are not the same as the ones in \cite{cabre06} (see Appendix~\ref{trace}). The parameter $\Sigma_I$ accounts for the imaginary part of the gluon self-interaction, whereas the real part is reabsorbed in the effective gluon mass. A more complete calculation of the gluon self-energy would require summing the $T$-matrix over the gluon thermal distribution self-consistently in a Brueckner-Hartree-Fock scheme (schematically, $\Pi_g=\int f^g {\cal T} D_g$ with $f^g$ the gluon distribution and $D_g$ the gluon single-particle propagator). We leave such determination for a future work. In the present evaluation, we shall approximate the gluon self-interaction by using an effective, in-medium gluon mass (to be discussed in Sec.~\ref{param}) together with a small imaginary part for numerical purposes (we use $\Sigma_I=0.01$~GeV as in \cite{cabre06}).
Once ${\cal T} (E;q,q')$ is known, the on-shell $T$-matrix is readily obtained
as ${\cal T}(E;q_E,q_E)$, with $q_E=\sqrt{E^2/4-m_g^2}$. The
Haftel-Tabakin algorithm is used to solve (\ref{tosolve}) \cite{haftel}. The
momentum integral is discretized within an appropriate quadrature, thus turning
the integral equation into a matrix equation, namely, $\sum {\cal F}_{ik} {\cal
T}_{kj} = V_{ij}$, where, schematically, ${\cal F}=1-wVG$ (and $w$ denotes the
integration weight). The solution follows trivially by matrix inversion. It can
be shown that the determinant of the transition function $\cal F$ (referred to
as the Fredholm determinant) vanishes at the bound state energies, which provides
a numerical criterion for solving the bound state problem. This strategy has
already been successfully used to compute $T$-matrices in the case of
quark-antiquark scattering \cite{cabre06}.
\section{Thermodynamic observables with SU($N$) and G$_2$}\label{sunT}
\subsection{Pure gauge sector}
\subsubsection{SU($N$) case}
The explicit computation of $\Omega^{(2)}$, given by (\ref{omega2}), obviously requires the knowledge of the on-shell $T$-matrix, that can be derived in particular from (\ref{ls}). In this last equation, $G_0$ is the propagator of two gluons, that has been discussed in Sec.~\ref{YM}. It is only worth saying that $G_0= {\rm O}(1)$ with respect to the number of colors since $m_g$ is assumed to be ${\rm O}(1)$ (see Sec.~\ref{gm}). The color dependence of the $T$-matrix actually comes from the two-gluon interaction potential only. More precisely, the color-dependence of the potential is all included in the factor (\ref{kappa0}), reading in the present case
\begin{equation}\label{kappa0b}
\kappa_{{\cal C};gg}=\frac{C^{{\cal C}}_2-2N}{2N}.
\end{equation}
The subscript $gg$ is used to recall that two-gluon interactions are concerned in the above formula, and $C_2^g = C_2^{adj} = N$ in the SU($N$) case.
The adjoint representation of SU($N$), to which gluons belong, can be written as the ($N-1$)-component vector $(1,0,\dots ,0,1)$ in a highest weight representation, corresponding to a Young diagram with $1$ column of length $N-1$ and 1 column of length 1. More generally, $(a_1,\dots,a_k,\dots,a_{N-1})$ corresponds to a Young diagram with $a_k$ columns of length $k$. The tensor product of the adjoint representation by itself gives the allowed two-gluon color channels:
\begin{eqnarray}\label{colo}
(1,0,\dots ,0,1)&\otimes &(1,0,\dots,0,1)=\nonumber\\
&&\bullet^S\oplus \, (1,0,\dots,0,1)^A\oplus (2,0,\dots,0,2)^S\qquad\qquad\ \ \hspace{1.5cm} N\geq 2\nonumber \\
& &\oplus \,(1,0,\dots,0,1)^S \oplus (0,1,0,\dots,0,2)^A \oplus (2,0,\dots,1,0)^A \quad N\geq 3\nonumber \\
& &\oplus \,(0,1,0,\dots,0,1,0)^S \hspace{6.4cm} N\geq 4.
\end{eqnarray}
The superscript $S$/$A$ denotes a symmetric/antisymmetric channel. The first/second/third line exists as soon as $N\geq$ 2/3/4. Note that in the special case $N=2$, the above tensor product reduces to $(2)\otimes(2)=(0)^S\oplus(2)^A\oplus (4)^S$, and one recovers usual spin-coupling rules. The dimensions and color factors of the representations appearing in (\ref{colo}) can be found in Table~\ref{tab1}.
In the singlet channel, one has $\kappa_{\bullet}=-1$ for any $N$. It is such that ${\cal T}={ \rm O} (1)$ since $V={\rm O}(1)$. Consequently, the properties of glueballs in singlet above the deconfinement temperature are not dependent of $N$, in agreement with \cite{hage}, where it is suggested that this argument is even gauge-group independent. The singlet finally brings a contribution ${\rm O} (1)$ to $\Omega^{(2)}$ since its dimension is 1.
Using the same arguments as for the singlet, one finds that the adjoint channels also lead to a $T$-matrix that is $N$-independent. They may lead to bound states since the potential is attractive, though less strongly than for the singlet. The symmetric adjoint channel will presumably be the most favorable for the formation of bound states since it demands a completely symmetric spin-space wave function for the two-gluon state in virtue of the Pauli principle, and the most attractive $J^P$ channels are indeed symmetric. Note that this symmetric color channel is actually absent for $N=2$. In the adjoint channel, ${\cal T}= {\rm O}(1)$ since $V= {\rm O} (1)$ but, unlike the singlet, its contribution to $\Omega^{(2)}$ is ${\rm O}(N^2)$ since ${\rm dim} (1,0,\dots,0,1)=N^2-1$.
The two remaining channels with nonzero potential, namely $(2,0,\dots,0,2)$ (the $\bm{27}$ for SU(3)) and $(0,1,0,\dots,0,1,0)$ (only when $N>3$), have in common that they are symmetric and that their color factor scales in $1/N$, thus vanishes in the large-$N$ limit. The fact that $V= {\rm O}(1/N)$ in both cases leads to the exact large-$N$ result
\begin{equation}
{\cal T}=V+V\, G_0\, V+{\rm O}(N^{-3}),
\end{equation}
or
\begin{equation}
{\cal T}=\pm\frac{1}{N}v+\frac{1}{N^2}v\, G_0\, v+{\rm O}(N^{-3}),
\end{equation}
the $\pm$ coming from one channel or another. Because of the weakness of $V$ at large-$N$, one can reasonably suppose that even the attractive channel $(0,1,0,\dots,0,1,0)$ does not lead to the formation of bound states. For the two channels under consideration,
\begin{equation}
\left. \left({\cal S}S^{-1}\overleftrightarrow{\partial_\epsilon}S \right)\right|_c \propto
\,\partial_\epsilon{\rm Re}\left(\pm\frac{v}{N}+\frac{vG_0v}{N^2}\right)+ {\rm O}(N^{-3}).
\end{equation}
One sees in (\ref{omega2}) that the contributions of both channels have to be summed and, since they are symmetric, the sums on the allowed $J^{P}$ is identical in both cases. This causes the term in $1/N$ to vanish in the trace at large-$N$ limit, the first nontrivial one being in $1/N^2$, leading to an overall contribution to $\Omega^{(2)}$ scaling as $N^2$ because the dimension of both channels scale as $N^4$.
Although the color singlet is relevant in view of studying glueballs, it does not bring any contribution to the EoS at large-$N$. So, the large-$N$ EoS is dominated by free gluons and scattering processes above threshold in colored channels. The more $N$ is large, the more important is the gap between the confined phase and the deconfined one, whose EoS scales as $N^2$. It is indeed known that the large $N$-case corresponds to a strongly first-order phase transition ($N=3$ is already weakly first order)~\cite{TcTh0}.
\subsubsection{G$_2$ case} \label{G2}
Another interesting group under consideration is G$_2$ which is also the best studied gauge group so far beyond SU($N$). The main features of this group are summarized in what follows.
\par The adjoint representation of G$_2$ has dimension 14, and reads $(0,1)$ in a highest weight representation. The two-gluon channels are then given by
\begin{equation}
(0,1)\otimes (0,1)=\bullet^S +(0,1)^A +(0,2)^S+(2,0)^S+(3,0)^A
\end{equation}
or, in terms of the dimensions, $14\otimes 14=1+14+77'+27+77$. Using the same normalisation than in the SU($N$) case, the color factors respectively read \cite{lipi,buiss11} $\kappa_{{\rm C};gg}=-1$, $-1/2$, 1/4, $-5/12$, and 0. The color factors in the singlet and adjoint channels are equal to those of SU($N$), so the glueball properties are unchanged in the singlet and antisymmetric adjoint channels. The symmetric $(2,0)^S$ channel is almost as attractive as the adjoint one: It may lead to bound states.
\subsubsection{Scaling relations for SU($N$) and G$_2$} \label{relations}
Some interesting relations about the scaling of the EoS can be deduced thanks to the $T$-matrix.
Let us write the on-shell $T$-matrix as ${\cal T}= \sum_k a_k\,\kappa_{{\cal C};gg}^k $ where all $a_k$ do not depend on the color but rather on the other quantum numbers involved. The color dependence of the thermodynamic observables is then given by the quantities $\sum_{{\cal C};A/S} \text{dim} \, {\cal C}\,\kappa_{{\cal C};gg}^k$. Using the results of Appendix~\ref{sun} and Sec.~\ref{G2}, one can check that, for SU($N$) and G$_2$,
\begin{eqnarray}
\label{sumdimk}
\sum_{{\cal C;S}} \text{dim} \, {\cal C}_{gg} \kappa_{{\cal C};gg} &=& \displaystyle\frac{1}{2} \, \text{dim} \,{adj} \text{,} \\
\sum_{{\cal C;S}} \text{dim} \, {\cal C}_{gg} \kappa_{{\cal C};gg}^2 &=& \displaystyle\frac{3}{4}\,\text{dim} \,{adj} \text{,} \\
\sum_{{\cal C;S}} \text{dim} \, {\cal C}_{gg} \kappa_{{\cal C};gg}^3 &=& -\displaystyle\frac{1}{8}\, \text{dim}\, {adj}\text{,}\\
\sum_{{\cal C;A}} \text{dim} \, {\cal C}_{gg} \kappa_{{\cal C};gg}^k &=& \left(- \displaystyle\frac{1}{2}\right)^k \, \text{dim} \,{adj} \text{.}
\end{eqnarray}
\par For SU($N$) at large-$N$, the previous relations can be written
\begin{eqnarray} \label{KcR}
\sum_{{\cal C;S}} \text{dim} \, {\cal C}_{gg} \kappa_{{\cal C};gg}^k&=& N^2 \left[ \left(-\frac{1}{2} \right)^k+ \delta_{k,1} + \frac{1}{2}\delta_{k,2} \right] +{\rm O}(1) \text{,} \\
\sum_{{\cal C;A}} \text{dim} \, {\cal C}_{gg} \kappa_{{\cal C};gg}^k &=& N^2\left(-\frac{1}{2} \right)^k +{\rm O}(1)\text{.}
\end{eqnarray}
Again that means that the expected scaling like $N^2$ (actually like dim $adj$) of the EoS is found using the present approach. This can be viewed as a confirmation of the relevance of the chosen color scaling (\ref{V00}).
\subsection{Quarks and antiquarks in the 't~Hooft limit}
Even if the rest of this study will be concerned with a genuine Yang-Mills plasma, it is worth making some comments about the possible inclusion of matter (quarks and antiquarks) in the model. Informations about the color channels appearing in interactions involving at least one (anti)quark are given in Tables \ref{tab3} and \ref{tab4}; remark that quarks (antiquarks) have been considered to be in the fundamental (conjugate) representation of SU($N$), as it is the case in 't~Hooft large-$N$ limit \cite{hoof}. Other interesting large-$N$ limits have been proposed, in which quarks belong to the two-index antisymmetric representation of SU($N$) for example \cite{qcdas}, but they will not be studied here.
First of all, the quark-quark and antiquark-antiquark color factors are of order $1/N$: The Born approximation for the $T$-matrix becomes exact at large-$N$ (at any temperature) and, since the dimension of any of the corresponding representations scales as $N^2$, the quark-quark and antiquark-antiquark interactions bring a term scaling as $N$ to the grand potential. More precisely, this term scales as $N_f \, N$ at the Born approximation once the trace over the different flavors is performed. It is shown in Sec.~\ref{Born} that only the interaction of two identical species can contribute to $\Omega$ in this limit. The number of quark flavors remains finite in the 't~Hooft limit, which is the case under study here.
The quark-antiquark interactions lead to a $T$-matrix which is ${\rm O}(1)$ when the pair is in the singlet, or ${\rm O}(1/N^2)$ when the pair is in the adjoint representation. In both cases however, the contribution to the grand potential scales as $N_f(N_f + 1)/2$ at large-$N$. So the quark-antiquark contributions to the thermodynamic observables is negligible with respect to the quark-quark and antiquark-antiquark ones in the 't~Hooft limit.
Finally, using similar arguments, one can show that the contribution of the quark-gluon and antiquark-gluon interactions to the grand potential scale as $N_f \, N$ at large-$N$. One concludes that, in the 't~Hooft large-$N$ limit, the grand potential is dominated by the gluonic contributions only, scaling as $N^2$.
\section{Parameters of the model}\label{param}
\subsection{Potential and gluon mass}\label{gm}
Two ingredients are now missing to start numerical computations: The interaction
potential between two gluons and the gluon mass. The procedure followed to fix
the potential is similar to the one followed in~\cite{cabre06} in the case of
heavy quark-antiquark bound states. The first step is to take some input from
lattice QCD, from which accurate computations of the static free energy of a
quark-antiquark pair bound in a color singlet, $F_1(r,T)$, are available. In
particular, computations in quenched SU(3) lattice QCD can be found in
\cite{kacz3}; they are especially relevant for our purpose since we focus on the
pure Yang-Mills plasma. There is still debate on the proper potential term to
use in phenomenological approaches, namely $F_1$ or the internal energy
$U_1=F_1-T\partial_T F_1$. An entropic contribution is subtracted from the free
energy in $U_1$, causing the internal energy to be more attractive than the free
energy, eventually leading to larger dissociation temperatures for bound states
in the deconfined medium. Spectral function analysis of heavy quarkonia from
lattice QCD simulations of Euclidean correlation functions typically suggest
that the $\eta_c$ and $J/\psi$ states may survive up to about 2$T_c$. Such
values of the dissociation temperature can be accommodated if the singlet
internal energy is used in potential model calculations.
\cite{Asakawa:2003re,Alberico:2005xw,Wong:2006bx,Riek:2010fk}.
As in~\cite{cabre06}, the internal energy is used as potential term. The explicit expression of the internal energy $U_1$ used in this work can be found in Appendix~\ref{appU1}.
The assumed color scaling (\ref{V0}) allows to derive the two-gluon potential from the lattice quark-antiquark one. Indeed, given $U_1(r, T)$ computed in quenched SU($N_{lat}$) lattice data, the color factor of the singlet quark-antiquark pair reads
\begin{equation}
\kappa_{q\bar q}=-\frac{N^2_{lat}-1}{2N_{lat}^2}.
\end{equation}
According to (\ref{V0}), the potential (in position space) between two quasigluons in the color channel ${\cal C}$ is then given by
\begin{equation}\label{Vg}
V(r,T)=\frac{\kappa_{{\cal C};gg}}{\kappa_{q \bar q}} \left[U_1(r,T)-U_1(\infty , T)\right],
\end{equation}
where the long-distance limit of the potential has be normalized to zero in order to ensure the convergence of the scattering equation and to perform the Fourier transform. This is actually a standard procedure in finite-temperature calculations.
According to the suggestion made in ~\cite{mocsy05}, the nonzero value of $U_1(\infty,T)$ should eventually be responsible of an effective in-medium contribution to the gluon mass. The intuitive argument is that, when both gluons are infinitely separated, they no longer interact. Therefore, the remaining potential energy should be seen as a manifestation of self-energy effects induced by the surrounding medium. These effects are encoded in the model as a mass shift to the ``bare" quasigluon mass, whose value has still to be fixed. Since $U_1(\infty,T) = 2\,m_q(T)$, the adaptation to the gluon must be done by extracting the correct color-dependence. From HTL computations \cite{HTL}, the self-energy color dependence is given by $C_2^ q/C_2^ {adj}$ at the first order when it is added in the propagator as a mass term ($m^ 2$), that means here that
\begin{equation}
\displaystyle\frac{U_1(\infty, T)}{2} = m_q(T) = \displaystyle\sqrt{\displaystyle\frac{C_2^q}{C_2^{adj}}} \Delta(T).
\end{equation}
So, $m_q(T) = 2\Delta(T)/3$ in the SU(3) case.
\par In the same way as done for the two-body color scaling, $\Delta(T)$ is considered as universal and the gluon thermal mass reads
\begin{equation} \label{scalingM}
\delta(T) = \displaystyle\sqrt{\displaystyle\frac{C_2^g}{C_2^{adj}}} \Delta(T) = \Delta(T),
\end{equation}
since $C_2^g = C_2^{adj}$. So, $\delta(T)$ is gauge-group independent. The effective in-medium gluon mass is finally given in our approach as
\begin{equation}\label{mg}
m_g(T)^2=m_0^2+\delta(T)^2.
\end{equation}
where the value $m_0$ has still to be fitted (see the following section). All the contributions are quadratically added as it is the case when one is dealing with bosonic propagators. The gluon mass is thus gauge-group independent.
\par It is obvious that the problem of the gluon mass is far more complicated than the simple prescription (\ref{mg}), that has to be seen as valid in a first approximation only. A more refined gluon mass should probably be momentum-dependent. There is indeed an increasing amount of evidences favoring the existence of a dynamically generated gluon mass due to nonperturbative effects, at least at zero temperature. Such a dynamically generated gluon mass $m_g(p)$, with $m_g(\infty)=0$ and $m_g(0)$ finite, is favored by some lattice results in Landau gauge, see \textit{e.g.} \cite{gluml1,gluml2}. Also nonperturbative field-theoretical calculations, using for example the pinch technique, find a nonzero dynamically generated gluon mass in $3+1$ YM theory \cite{glumass,glumass2}. It is also worth quoting the recent Coulomb gauge study \cite{mgfinit}, which is a first step in view of understanding the behavior of $m_g(p,T)$ at a nonperturbative level. From a different perspective, non-perturbative contributions to the gluon potential and mass are analyzed at finite
temperature in connection with the gluon condensates in \cite{Megias:2007pq,Megias:2009mp}.
Such improvements of the gluon mass are left for future works.
The above discussion gives a more precise meaning to the term ``quasigluon" used in this paper: It denotes transverse particles in the adjoint representation of SU($N$) that gain an effective mass $m_g(T)$ given by (\ref{mg}) and interact through the potential (\ref{Vg}).
\subsection{Zero temperature results with SU(3)} \label{zeroT}
Before performing finite-temperature computations, it is worth checking whether the values retained for the various parameters of our model may give relevant results at zero temperature or not. In particular, is the present $T$-matrix formalism able to reproduce at least qualitatively the features of the low-lying glueball spectrum computed in pure gauge SU(3) lattice QCD at zero temperature ? \cite{glulat}
It is known that, at zero temperature and in quenched SU(3) lattice QCD, the potential between a static quark-antiquark pair is compatible with the funnel form \cite{bali}
\begin{equation}
V_f(r) = \sigma r -\frac{4}{3}\frac{\alpha}{r}.
\end{equation}
In order to stay coherent with the potential above $T_c$, $\alpha=0.141$ (see Appendix \ref{appU1}) and $\sigma = 0.176$~GeV$^2$ (a standard value for the string tension) are used. The Fourier transform of $V_f(r)$ is not defined: This flaw can be cured by making it saturate at some value $V_{sb}$, interpreted as a string-breaking value, that is the energy above which a light quark-antiquark pair can be created from the vacuum and break the QCD string. This scale is then subtracted and the potential effectively taken into account is $V_f(r)-V_{sb}$, while $V_{sb}$ is interpreted as an effective quark mass using the same arguments as those detailed in Sec.~\ref{gm}. According to the color scaling (\ref{kappa0}), the potential that should be used to describe the interactions between two gluons at zero temperature is
\begin{equation}
V_0(r) = \displaystyle\frac{9}{4} V_f(r) - V_{sb}^g
\end{equation}
when the gauge group is SU(3).
In this case, the string breaking scale should rather be interpreted as the energy scale necessary to form two gluelumps, a gluelump being a gluon bound in the color field of a static adjoint source. It is known indeed that adjoint string breaking may be observed, and occurs at twice the lightest gluelump mass ($\sim 2$~GeV) \cite{defor}. So, $V_{sb}^g = 2$~GeV is used here, a value in agreement with lattice data showing that the mass of the lightest gluelump is given by $0.85(17)$~GeV \cite{pineda}.
The only free parameter left to compute the ${T}$-matrix is the bare gluon mass, $m_0$.
Keeping the same structure as in Eq.~ (\ref{mg}) we have
\begin{equation}
m_g(0)^2 = m_0^2 + \left(\displaystyle\frac{V_{sb}^g}{2}\right)^ 2,
\end{equation}
where again we have traded the subtracted potential at infinite separation distance (string breaking energy in this case) into a self-energy-like contribution to the quasi-particle gluon mass.
We fix $m_0 = 0.7$~GeV, which is a typical value for the zero-momentum limit of the gluon propagator at zero temperature. Advancing results, such a value will ensure both a correct agreement with the zero temperature lattice glueball spectrum (see Table \ref{tabglueb}), and an excellent agreement with the pure gauge EoS computed on the lattice (see Sec. \ref{EoS}).
\begin{table}[ht]
\caption{Masses (in GeV) of the lowest-lying glueball states at zero temperature with the gauge group SU(3). Our results (third column), are compared to the lattice data of \cite{glulat} (second column) and to the Coulomb gauge QCD (CGQCD) study \cite{cg0} (last column). }
\begin{tabular}{c|rrr}
State & Lattice \cite{glulat} & $T$-matrix & CGQCD \cite{cg0} \\
\hline
$0^{++}$ & 1.73 (5)(8) & 2.17 & 1.98\\
$0^{-+}$ & 2.59 (4)(13) & 2.39 & 2.22 \\
$2^{++}$ & 2.40 (2.5)(12) & 2.34 & 2.42 \\
\end{tabular}
\label{tabglueb}
\end{table}
The results are given in Table~\ref{tabglueb} for $J^P$ channels discussed in Sec~\ref{helicity}. At least our model is able to reproduce the mass hierarchy of the lightest glueballs as well as the typical mass scale of 2~GeV for those states. The accuracy of the model can be compared to Coulomb gauge QCD \cite{cg0}, sharing formally many similarities with our $T$-matrix approach: The results of this last reference are also given in Table~\ref{tabglueb}. The agreement between lattice QCD and CGQCD is better, but it is worth saying that the parameters used in CGQCD have been chosen to reach an optimal agreement with the zero temperature lattice data, while here the values are mostly designed to give good results above $T_c$. So our model results are satisfactory in that sense. What makes us to find such a high mass for the scalar glueball is the quite small value $\alpha=0.141$ that has been taken (in order to fit static potentials above $T_c$), while values as high as $\alpha=0.4$ have sometimes to be used to reach a good agreement between lattice data and effective approaches, see \textit{e.g.} \cite{gluheli}. The scalar glueball being dominantly a $S$-wave state, it is particularly sensitive to the strength of the Coulomb term and to the running of $\alpha$ with the temperature, that is neglected here.
Finally, the extension of the above calculations to any gauge group is straightforward in our approach: The interested reader will find a discussion of such a generalization in \cite{buiss11}, where it is shown that the lowest-lying glueball masses is gauge-group independent within a constituent framework. In particular, the lowest-lying glueball masses are found independent of $N$ in \cite{buiss11}, in agreement with what is observed on the lattice \cite{gluLNC}. That is why the $T$-matrix masses given in Table~\ref{tabglueb} are considered as valid for any gauge group too.
\subsection{Relevance of the parameters for other gauge groups}
At this stage, it is important to summarize the various parameters introduced and their possible dependence -- or not -- on the gauge group. Comparison with existing results when it is possible can also shed light on that issue.
First of all, the basic ingredient underlying the present study is the static quark-antiquark potential computed in finite-temperature quenched lattice QCD. The assumed one-gluon-exchange nature of the two-particle interactions leads to the universality of the momentum-dependent part of the potential, and to a well-defined prescription for its gauge-group dependence. Similarly, the gluon thermal mass has a peculiar color scaling originating in its interpretation as a self-energy term.
More freedom is apparently left for the numerical parameters at our disposal. Let us comment them briefly. First, the propagator imaginary part $\Sigma_I$ has been introduced for computational convenience. Hence it can be kept constant when changing the gauge group. Second, it can be checked that, by dimensional analysis, our results can all be expressed in terms of the ratios $T/T_c$, $T_c/\sqrt\sigma$ and $m_0/\sqrt\sigma$.
According to glueball gas models with a Hagedorn spectrum describing the high-lying glueball states, the critical temperature is given by \cite{meyer,hage}
\begin{equation}
\label{Tcsig}
\frac{T_c}{\sqrt\sigma}=\sqrt{\frac{3}{2\pi}}=0.69.
\end{equation}
This value is due to the Hagedorn spectrum, not defined above a certain temperature. This temperature is here interpreted as the deconfinement one. It is worth saying that it leads to an EoS in very good agreement with lattice results \cite{meyer,hage,pane2} below $T_c$. In this picture, the ratio $T_c/\sqrt\sigma$ is gauge-group independent: This is only valid in a first approximation since, for example, there are lattice evidences showing that $T_c/\sqrt\sigma$ is only constant up to $1/N^2$ corrections \cite{TcTh}. Nevertheless, such deviation are beyond the scope of this exploratory work. Note that according to \cite{Braun}, the critical temperature is found to be pretty close to 300~MeV up to fluctuation of about 10\% for the gauge groups SU($N$), Sp(2), and E$_7$. So, we fixed $T_c = 300$~MeV in our calculations. This value is in good agreement with (\ref{Tcsig}) for the value $\sigma = 0.176$~GeV$^2$ chosen for $T=0$ calculations (see Sec.~\ref{zeroT}).
Concerning the ratio $m_0/\sqrt{\sigma}$, it is worth mentioning the work \cite{maas}, in which it is shown that the nonperturbative gluon propagator at zero temperature (thus $m_0$ in particular) shows no significant quantitative differences when expressed in units of the string tension for the groups SU($N$) and G$_2$. It is thus tempting to say that the ratio $m_0/\sqrt\sigma$ may be gauge-group independent also: This is assumed in the rest of this paper. The value $m_0/\sqrt\sigma=1.67$ obtained from the zero-temperature glueball spectrum is retained.
Let us finally mention that, at zero temperature, the string breaking scale is found to be two times the gluelump mass for SU($N$) and G$_2$ \cite{lipi}. This means that the extension of the above zero-temperature calculations to any gauge group is straightforward in our approach: The $T$-matrix masses given in Table~\ref{tabglueb} can be considered as valid for any gauge group once divided by $\sqrt\sigma$.
\section{Existence of glueballs above $T_c$}\label{BSsec}
\subsection{Singlet states}
Now that the parameters of the model have been all fixed, $T$-matrix
calculations above $T_c$ can be performed. Technical details will not be given
here since the method is identical to the one used in \cite{cabre06}, which
involves the Haftel-Tabakin algorithm to solve the $T$-matrix Lippmann-Schwinger
equation \cite{haftel}. The bound and resonant states appear as poles in the
on-shell $T$-matrix, or more precisely as zeros of $\det {\cal F}$. The
corresponding masses are given in Table~\ref{tab:bsatc} in the
different considered color-singlet channels. Since in this case, $\kappa_{{\rm
C};\bullet}=-1$ for all gauge groups and since the gluon mass is independent of
$N$, the masses of the color-singlet are the same for SU($N$) for all $N$, and for
$G_2$.
Only a few papers devoted to the existence of glueballs on the lattice are currently known \cite{suga3}, and the interpretation of their results depends mostly on the way the glueball correlators are fitted: Either a single narrow pole, or a Breit-Wigner shape. Let us focus on the narrow pole fit, which identifies bound states in a way similar to ours. The main observation to be made from \cite{suga3} is that the glueball masses decrease above $T_c$ with increasing temperature, with a mass near $T_c$ that is similar to the zero temperature one. This nontrivial behavior is well-checked within our approach. Two competing effects are responsible for the temperature evolution of the spectrum: reduction of the binding energy and downward shift of the threshold energy. Overall, the singlet scalar bound state experiences a mild shift to
lower energies and dissociates at $T_{\textrm{dis}} \approx 1.3~T_c$. This is the value from which
$\det {\cal F}$ does not vanish anymore. Nevertheless, considerable strength remains at
threshold up to about $1.5~T_c$. This is in qualitative agreement with the spectral function
analysis of Euclidean correlators by the CLQCD Collaboration \cite{CLQCD}.
\begin{table}
\caption{Masses (in units of $\sqrt{\sigma}$) of lowest-lying glueballs above $T_c$ ($T_c = 0.3$~GeV). A line mark the temperature at which a bound state is not detected anymore.}
\begin{center}
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c}
\multicolumn{2}{c|}{Channel} & \multicolumn{3}{|c|}{Singlet} & \multicolumn{3}{|c|}{Adjoint$^S$ } & \multicolumn{3}{|c}{(2,0)$^S$ } \\ \hline
\multicolumn{2}{c|}{Group} & \multicolumn{3}{|c|}{All} & \multicolumn{3}{|c|}{SU($N \geq 3$)} & \multicolumn{3}{|c}{$G_2$} \\ \hline
$T/T_c$ & 2\,$m_g$ & $0^{++}$ & $0^{-+} $ & $2^{++}$ & $0^{++}$ & $0^{-+} $ & $2^{++}$ & $0^{++}$ & $0^{-+} $ & $2^{++}$ \\ \hline
1.05 & 6.50 & 4.52 & 5.43 & 5.43 & 6.00 & 6.45 & 6.31 & 6.14 & - & 6.38 \\
& & 6.48\footnote{Radial excitation below the threshold} & & & & & & & & \\
1.10 & 5.24 & 4.57 & 5.21 & 5.00 & 5.14 & - & - & 5.21 & & - \\
1.15 & 4.71 & 4.43 & - & 4.67 & - & & & - & & \\
1.20 & 4.43 & 4.33 & \, & - & \, & \, & \, & \, & \, & \, \\
1.25 & 4.26 & 4.24 & \, & \,& \, & \, & \, & \, & \, & \, \\
1.30 & 4.14 & - & & \, & \, & \, & \, & \, & \, & \\
1.35 & - & & & \, & \, & \, & \, & \, & \, & \\
\end{tabular}
\end{center}
\label{tab:bsatc}
\end{table}
The evolution of the imaginary part of the on-shell $T$-matrix in the singlet
scalar channel versus the temperature is displayed in Fig. \ref{fig3bs}: This
gives an overall picture of the glueball progressive dissolution in the medium.
The peak in the imaginary part, depicting a bound state, becomes broader and
broader before melting into the continuum (and thus $\det {\cal F}$ does not
vanish anymore below threshold) as the temperature is increased. Still, for $T
> T_{\textrm{dis}}$ and above the threshold energy one finds sizable strength from the
bound state relic, the $T-$matrix exhibiting a resonant behavior well beyond the
Born approximation.
\begin{figure}[ht]
\begin{center}
\includegraphics*[width=0.6\textwidth]{plot-ImT-0ppS-SqrtCasimirM07-Tall.eps}
\caption{
$\textrm{Im} T$ for $gg$
scattering in the
$0^{++}$ singlet channel for various $T$ with $T_c = 0.3$~GeV.
}
\label{fig3bs}
\end{center}
\end{figure}
Concerning the pseudoscalar channel, singlet bound states are found up to 1.10 $T_c$. Note that states in the pseudoscalar
channels, which in our approach correspond to pure $P$-wave scattering, are
just mildly bound due to the centrifugal barrier. The tensor states, having an $S$-wave component, lie between the scalar and pseudoscalar channels, regarding
binding and dissociation temperatures.
\subsection{Colored states}
\begin{figure}
\begin{center}
\includegraphics*[width=1.0\textwidth]{T-0pp-Tevol-S-and-O-bigger.eps}
\caption{ $T$-matrix for $gg$ scattering in the scalar singlet and symmetric adjoint channels for SU($N \geq 3$). From left to right the temperatures are
(1.05;1.10;1.15) $T_c$, with $T_c = 0.3$~GeV.
}
\label{fig2bs}
\end{center}
\end{figure}
Bound states in the symmetric adjoint channel of SU($N \geq 3$) are also observed (see Table~\ref{tab:bsatc}), although they are less bound since $\kappa_{{\rm C};\bullet}=-1/2$. The scalar channel disappears above 1.10 $T_c$, whereas in the pseudoscalar and tensor channels, bound states are lying right below the threshold energy at the lowest considered temperature (\textit{i.e.} $1.05~T_c$). The differences between singlet and adjoint channels have to be attributed to the strength of the potential, which is two times smaller in the adjoint channel than in the color singlet.
The evolution of the $T$-matrix in the singlet and symmetric adjoint scalar channel versus the temperature is displayed in Fig.~\ref{fig2bs}. One clearly sees the disappearance of this bound state at 1.15 $T_c$ while the singlet state is still well bound at this temperature.
There are in general other colored channels than the adjoint one. For SU($N$) gauge groups in particular, the only one that could \textit{a priori} lead to bound states is the $(0,1,0,\dots,0,1,0)^S$ channel, which is weakly attractive and exists only for $N>3$. It has been checked that even the scalar state (the most attractive channel) is unbound at $N=4$. Hence this color channel does not admit bound states.
In the case of $G_2$, that is the other group considered in this study, the $(2,0)^S$ channel leads to bound states with the same melting temperatures as in the adjoint channel, up to our current precision of 0.05 $T_c$ (see Table~\ref{tab:bsatc}). \\
\section{Equation of state}\label{EoS}
The pure gauge EoS can now be computed without introducing extra parameters: The
two-body potential and the thermal mass contribution to the gluon mass have been
fitted on lattice data by using respectively the scaling (\ref{Vg}) and
(\ref{scalingM}), and they ensure a correct agreement with zero temperature
results. A crucial point to establish the EoS thanks to (\ref{omega2}) is to
correctly express the summation on the different channels. For 2-gluon
interactions, the channels are cued by the $ J^{P}$ number and the color number.
The summation on the $J^{P}$ channels is formally infinite but in this work,
only the $0^{++}$, $0^ {-+}$ and $2^{++}$ channels are taken into account. This
restriction is supported by the following argument. These three channels are the
most attractive ones and generate the lightest glueballs. The
lighter the mass, the more important
the thermodynamic contribution is in the bound state sector.
Thus, the bound state thermodynamic contribution coming from other
$J^{P}$ should be negligible in comparison of these three channels. In order to be
coherent, this $J^{P}$ restriction is implemented in the scattering sector. It
creates also a restriction in the allowed color channels. Since the $0^{++}$,
$0^ {-+}$ and $2^{++}$ are symmetric $J^{P}$ channels, the color channels must
be symmetric too in order to respect the Pauli's principle.
\begin{figure}[ht]
\begin{center}
\includegraphics*[width=0.8\textwidth]{free.eps}
\caption{Normalized pressure versus temperature in units of $T_c$ (with $T_c$ = 0.3~GeV), computed for the gauge group SU(3) in the free gluon gas case and in the full approach.
}
\label{fig3}
\end{center}
\end{figure}
\par In Fig.~\ref{fig3}, the normalized pressure $P/P_{SB}$ computed in the free gluon gas
case with the thermal mass (\ref{mg}) is compared with the normalized pressure
obtained by our approach in the SU(3) case. At low temperature ($T \leq $ 1.3
$T_c$), the bound state and the scattering parts both give thermodynamic
contributions that modify the free gas pressure. For $T > 1.3 T_c$, only the
scattering part keeps to contribute. As it can be observed in Fig.~\ref{fig3},
the main global effect of the interaction is to decrease the pressure. If each
contribution is analyzed, it is seen that the bound state formation increases
the pressure because bound states are simply added as new species that does not
interact with the other particles inside the plasma. Concerning the two-gluon
scattering part, the sign of the pressure contribution can not be analytically
predicted at each temperature. Only at the Born approximation, one can observe
that attractive (repulsive) channels contribute to increase (decrease) the
pressure. Indeed, in momentum space representation (see Appendix~\ref{trace}),
(\ref{pot_s3}) becomes here
\begin{equation} \label{Born0}
\Omega_s = \displaystyle\frac{1}{64\pi^5 \beta} \displaystyle\sum_{J^P} (2J+1) \displaystyle\sum_{{\cal C}, g} \text{dim} \, {\cal C} \,\kappa_{{\cal C}, gg} \displaystyle\int_{2m_g}^{\infty} d\epsilon \, \epsilon^3 \displaystyle\sqrt{\frac{\epsilon^2}{4} - m_g^2} K_1(\beta \epsilon) \, v_{J^{P}}
\end{equation}
where $K_1(x)$ is the modified Bessel function of the first kind. In the attractive (repulsive) channels, the sign of the potential is negative (positive). Since $\Omega_s$ is the scattering contribution to the grand potential, it can be deduced that attractive (repulsive) channels increase (decrease) the pressure. In the present SU(3) case, the only repulsive channel is the $(2,2)^S$. That means that the decreasing of the pressure in our approach compared with the free gas pressure is only driven by the $(2,2)^S$ channel.
\par It is also worth wondering whether some constraints arise or not from the
high-temperature limit of our framework concerning the behavior of the two-body
interactions. In this limit, the Born approximation should be relevant. Using
(\ref{sumdimk}) and (\ref{Born0}), one can write
\begin{eqnarray}
\Omega_s &\sim &\displaystyle\frac{1}{64\pi^5 \beta} \displaystyle\frac{\text{dim}\,adj}{2}\displaystyle\int_{2m_g}^{\infty} d\epsilon \, \epsilon^3 \displaystyle\sqrt{\frac{\epsilon^2}{4} - m_g^2} K_1(\beta \epsilon) \, v_{0} .
\end{eqnarray}
Only the scalar channel has been taken into account for the sake of clarity, but the following argument can be extended to any spin. According to Hard-Thermal-Loop results, it is relevant to assume a Yukawa form for the potential $v_0$ at high temperature \cite{HTL}. Then,
\begin{eqnarray}
\Omega_s &\sim &\displaystyle\frac{1}{64\pi^5 \beta^4} \displaystyle\frac{\text{dim}\,adj}{2} \displaystyle\int_{2 \beta m_g}^{\infty} dx \, x^3 K_1(x) \, \frac{\sqrt{\frac{x^2}{4} -\beta^2 m_g^2}}{\frac{x^2}{4} -\beta^2 m_g^2-\beta^2 M^2},
\end{eqnarray}
where $M$ is the screening mass of the theory, \textit{a priori} temperature-dependent. Still in HTL theory, it is found that, because of the running of the strong coupling constant,
\begin{equation}\label{limmu}
{\rm lim}_{\beta\rightarrow 0} \beta m_g ={\rm lim}_{\beta\rightarrow 0} \beta M=0.
\end{equation}
More precisely, the quark and gluon thermal masses are found to behave as $\alpha_s(T) \, T$.
Consequently, at high enough temperatures, it is found that
\begin{eqnarray}
\Omega_s &\sim &\ \displaystyle\frac{\text{dim}\,adj}{32 \pi^ 5\beta^4},
\end{eqnarray}
\textit{i.e.} a scattering contribution that has the same behavior with respect to the temperature as the free part, ensuring a well-defined large-temperature limit.
\begin{figure}[ht]
\begin{center}
\includegraphics*[width=0.8\textwidth]{EOS1.eps}
\caption{Normalized pressure versus temperature in units of $T_c$ (with $T_c$ = 0.3~GeV), computed for the gauge groups SU(2,3,$\infty$) and $G_2$ (solid lines). Note that all the curves are nearly indistinguishable. Our results are compared to the lattice data of \cite{su2lat} for SU(2) (dots) and \cite{panero} for SU(3,4,6,8) (dots), and of the minimal $G_2$ model of \cite{dumi} for $G_2$ (dashed line). Note that all Lattice data have been normalized to the lattice Stefan-Boltzmann pressure \cite{su2lat,panero}.
}
\label{fig4}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics*[width=0.8\textwidth]{EOS2.eps}
\caption{Normalized pressure versus temperature in units of $T_c$ (with $T_c$ = 0.3~GeV), computed for the gauge groups SU(2), SU(3), $G_2$, and SU($\infty$). The temperature range is the one where the differences between the curves are the most important.
}
\label{fig5}
\end{center}
\end{figure}
Notice that our fit of the screening mass does not follow the constraints (\ref{limmu}), but it is designed to fit the static potential below 3~$T_c$. A more involved form would be needed to reach the HTL predictions at high temperatures, but it is not the scope of the present work.
\par In Fig.~\ref{fig4}, the normalized pressure $P/P_{SB}$ is presented for different gauge groups: SU(2), SU(3), SU($\infty$) and G$_2$.
Severals remarks can be done. First, the free gluon thermodynamic contribution is gauge-group invariant once normalized to $p_{SB}$. The gauge-group dependence is only present in the bound state and scattering sectors. The number of allowed color channels (\textit{i.e.} the symmetric ones) depends on the gauge group (see (\ref{colo})) and determines the allowed maximum number of bound states and the number of scattering channels. The bound state thermodynamic contribution comes from two effets: The number and the mass of the existing glueballs. Because of the glueball dissociation, this contribution is only taken into account up to the temperature of dissociation (see Tables~\ref{tab:bsatc}). One can observe on Fig.~\ref{fig4} that the produced EoS are not very sensitive to the gauge-group. The most important difference between the curves occurs between 1.05 and 1.35 $T_c$ (see Fig.~\ref{fig5}): In this range, the gluon-gluon interactions are maximal. When the temperature increases, the Born approximation becomes more and more valid and the pressure then scales as $\text{dim}~adj$. Thus the normalized pressure tends to be universal.
\par In Fig~\ref{fig4}, it is also worth noticing that the EoS computed in our
approach favorably compares with QCD lattice data for gauge groups SU(3-8)
\cite{panero} where such universal curves seem to appear (note that lattice data exist also for very high values of $T/T_c$ but only for SU(3) \cite{bors12}). Concerning G$_2$, no lattice data about EoS are currently available but a
new effective matrix model describing pure Yang-Mills thermodynamics has been
proposed in \cite{dumi}. These last results are compared to ours in
Fig~\ref{fig4}.
\section{Conclusions}
The relevance of gluon-gluon interactions beyond the critical temperature in the pure gauge
SU(3) plasma has been addressed in a non-perturbative $T$-matrix many-body framework with the
input of Casimir-scaled potentials from thermal lattice QCD and a model of quasigluon mass independent of the gauge group. Scalar glueball bound states in the
singlet channel survive up to temperatures of about 1.3-1.5~$T_c$, together with sizable threshold effects due to strong correlations beyond the two-particle threshold. With only one free parameter, the gluon mass at $T=0$, the EoS of the gluon-glueball gas is reproduced in good agreement with quenched lattice SU($N$) simulations (the other parameters can be fixed \textit{a priori} by resorting to either lattice results or theoretical arguments). Predictions for the $G_2$ EoS are also given, the main feature being that it should be very close to the SU(N) one once normalized to the Stefan-Boltzmann pressure.
\par The present $T$-matrix formalism can in principle be systematically
improved, for example by including three-gluon scattering. The number of
channels under consideration is quite larger compared with the two-gluon case,
with more elaborated symmetries. Moreover, finding the $T$-matrix would then
become a three-body problem, whose resolution through \textit{e.g.} Faddeev equations can be addressed in future works.
\par Another natural extension to this paper is to study the light meson spectrum at finite temperature and the QCD EoS by including quarks within the model. Computations with baryonic potential can be also considered.
\par Finally, the $T$-matrix formalism can also be applied to calculate bulk thermodynamical properties of the system such as the sheer viscosity, which can be easily computed in relaxation-time approximation within a quasi-particle picture. Such a work is in progress.
\acknowledgments
This work has been partly supported by grant FPA2011-27853-C02-02 (Ministerio de
Econom\'{\i}a y Competitividad, Spain).
D.C. acknowledges financial support from Centro Nacional de F\'{\i}sica de
Part\'{\i}culas, Astropart\'{\i}culas y Nuclear (CPAN, Consolider - Ingenio
2010). G.L. and C.S. thank F.R.S-FNRS for financal supports. The authors thank M. Panero for useful comments.
\begin{appendix}
\section{Some SU($N$) relations}\label{sun}
The dimensions of the color channels appearing in (\ref{colo}) are given in Table~\ref{tab1}, together with the color factors (\ref{kappa0b}). Note that a general method for computing the quadratic Casimir of SU($N$) can be found in \cite{cas}.
\begin{table}[ht]\caption{Symmetry, dimension (${\rm dim}\ {\cal C}$) and color factor ($\kappa_{{\cal C}}$) defined in (\ref{kappa0b}), of the color channels (${\cal C}$) appearing in the tensor product of the SU($N$) adjoint representation by itself. This table actually displays the two-gluons color channels, denoted by $gg$. The SU(3) case is also indicated.}
\begin{tabular}{c|ccccc}
${\cal C}$ for $gg$ & $\bullet$ & $(1,0,\dots,0,1)$ & $(2,0,\dots,0,2)$ & $(2,0,\dots,1,0)$ & $(0,1,0,\dots,0,1,0)$ \\
& & & & $(0,1,\dots,2)$ & \\
SU(3) & $\bullet$ & (1,1) & (2,2) & (0,3), (3,0) & - \\
\hline
Symmetry & S & S, A & S & A & S \\
${\rm dim}\ {\cal C}$ & 1 & $N^ 2-1$ &$\frac{N^2(N+3)(N-1)}{4}$ & $\frac{(N^ 2-4)(N^ 2-1)}{4}$ & $\frac{N^2(N-3)(N+1)}{4}$ \\
$\kappa_{{\cal C}}$ & $-1$ & $-\frac{1}{2}$ & $\frac{1}{N}$ & 0 & $-\frac{1}{N}$ \\
\end{tabular}
\label{tab1}
\end{table}
Similar results can be written in the case where only the fundamental and/or conjugate representations are taken into account. The terms appearing in the tensor product of the fundamental and conjugate representations by themselves are given in Table~\ref{tab3}, as well as those appearing in the tensor product of the fundamental representation by the conjugate one.
\begin{table}[ht]
\caption{Symmetry, dimension (${\rm dim}\ {\cal C}$) and color factor ($\kappa_{{\cal C}}$) defined in (\ref{kappa0}), of the color channels (${\cal C}$) appearing in the tensor product of the SU($N$) fundamental representation by itself (left), of the conjugate representation by itself (middle), and of the fundamental representation by the conjugate one (right). This table actually displays the quark-quark, antiquark-antiquark and quark-antiquark color channels, denoted by $qq$, $\bar q \bar q$, and $q\bar q$ respectively. }
\begin{tabular}{c|cc|cc|cc}
${\cal C}$ for $qq$, $\bar q \bar q$, $q\bar q$ & $(2,0,\dots,0)$ & $(0,1,0,\dots,0)$ & $(0,\dots,0,2)$ & $(0,\dots,0,1,0)$ & $\bullet$ & $(1,0,\dots,0,1)$ \\
\hline
Symmetry & S & A & S & A & & \\
${\rm dim}\ {\cal C}$ & $\frac{N(N+1)}{2}$ & $\frac{N(N-1)}{2}$ & $\frac{N(N+1)}{2}$ & $\frac{N(N-1)}{2}$ & 1 &$N^2-1$ \\
$\kappa_{{\cal C}}$ & $\frac{N-1}{2N^2}$ & $-\frac{N+1}{2N^2}$ & $\frac{N-1}{2N^2}$ & $-\frac{N+1}{2N^2}$ & $-\frac{N^2-1}{2N^2}$ & $\frac{1}{2N^2}$ \\
\end{tabular}
\label{tab3}
\end{table}
Finally, useful results concerning the tensor product of the fundamental (conjugate) representation by the adjoint representation are given in Table~\ref{tab4}.
\begin{table}[ht]
\caption{Dimension (${\rm dim}\ {\cal C}$) and color factor ($\kappa_{{\cal C}}$) defined in (\ref{kappa0}) of the color channels (${\cal C}$) appearing in the tensor product of the SU($N$) fundamental representation by the adjoint one (left), and of the conjugate representation by the adjoint one (right). This table actually displays the quark-gluon and antiquark-gluon color channels, denoted by $qg$ and $\bar q g$ respectively.}
\begin{tabular}{c|ccc|ccc}
${\cal C}$ for $qg$, $\bar q g$ & $(1,0,\dots,0)$ & $(2,0,\dots,0,1)$ & $(0,1,\dots,0,1)$ & $(0,\dots,0,1)$ & $(1,0,\dots,0,2)$ & $(1,0,...1,0) $\\
\hline
${\rm dim}\ {\cal C}$ & $N$ & $\frac{(N+2)N(N-1)}{2}$ & $\frac{(N+1)N(N-2)}{2}$ & $N$ & $\frac{(N+2)N(N-1)}{2}$ & $\frac{(N+1)N(N-2)}{2}$ \\
$\kappa_{{\cal C}}$ & $-\frac{1}{2}$ & $\frac{1}{2N}$ & $-\frac{1}{2N}$ & $-\frac{1}{2}$ & $\frac{1}{2N}$ & $-\frac{1}{2N}$ \\
\end{tabular}
\label{tab4}
\end{table}
\section{Lattice potential}\label{appU1}
The lattice data that are used as a starting point to build our interaction potential are those of \cite{kacz3}, \textit{i.e.} the static free energy between a quark-antiquark pair bound in a color singlet for $N_{lat}=3$. For numerical convenience, it is preferable to deal with a fitted form of these, rather than with interpolations of the available points. To fit the data of~\cite{kacz3}, the analytic form proposed by Satz in~\cite{satz} is used:
\begin{equation}\label{f1lat}
F_1(r,T)=\frac{\sigma}{\mu (T)}\left[ \frac{\Gamma(1/4)}{2^{3/2}\Gamma(3/4)}-\frac{\sqrt{\mu(T)r}}{2^{3/4}\Gamma(3/4)}K_{1/4}\left(\mu(T)^2 r^2\right)\right]-\frac{4}{3}\frac{\alpha}{r}\Big[{\rm e}^{-\mu(T) r}+\mu(T) r\Big].
\end{equation}
The way of obtaining this formula is the following. First, it is known that the static quark-antiquark energy at zero temperature is accurately fitted by a so-called funnel shape
\begin{equation}\label{f1z}
F_1(r,0)=\sigma\, r-\displaystyle\frac{4}{3} \displaystyle\frac{\alpha}{r}=U_1(r,0),
\end{equation}
see \textit{e.g.}~\cite{bali}. When $T>0$, one can imagine that this potential is progressively screened by thermal fluctuations. An effective theory for studying the screening of a given potential is the Debye-H\"uckel theory, in which the thermal fluctuations are all contained in a screening function $\mu(T)$, that modifies the zero-temperature potential and eventually leads to the form (\ref{f1lat}).
The explicit form of $\mu(T)$ is unknown \textit{a priori} and has to be fitted on the lattice data. As it can be seen in Fig.~\ref{fig1a}, the form
\begin{eqnarray}\label{mufit}
\frac{\mu(T)}{\sqrt\sigma}&=&0.537\frac{T}{T_c}+0.644+0.112\ln\left(\frac{T}{T_c}-
0.967\right),
\end{eqnarray}
with
\begin{equation}\label{afit}
\alpha=0.141,
\end{equation}
provides an accurate fit of the lattice data in the range 1-3~$T_c$. A more complete fit should be such that $\mu(0)=0$, but our model is not intended to be able to ``cross" the phase transition in $T_c$. The simple form (\ref{mufit}) is already satisfactory. The corresponding internal energy $U_1=F_1-T\partial_T F_1$ is plotted in Fig.~\ref{fig2a}.
\begin{figure}[ht]
\includegraphics*[width=8.5cm]{f1_3.eps}
\caption{Static free energy $F_1(r,T)$ of a quark-antiquark pair bound in a color singlet, computed in SU(3) quenched lattice QCD and plotted for different temperatures (symbols). Data are taken from~\cite{kacz3} and expressed in units of $\sqrt\sigma$, with $r$ the quark-antiquark separation. The fitted form (\ref{f1lat})-(\ref{afit}) is compared to the lattice data (solid lines). }
\label{fig1a}
\end{figure}
\begin{figure}[ht]
\includegraphics*[width=8.5cm]{u1_3.eps}
\caption{Internal energy $U_1(rT)$ of a quark-antiquark pair bound in a color singlet, computed from the fitted form (\ref{f1lat})-(\ref{afit}) and plotted for different temperatures (solid lines). }
\label{fig2a}
\end{figure}
\section{Dashen, Ma and Bernstein's formalism in momentum space} \label{trace}
To compute (\ref{omega2}), it is necessary to use a given representation. In order to use the calculation of the $T$-matrix proposed in \cite{cabre06}, the scattering part of (\ref{pot_s3}) must be computed in the momentum space representation (the two first term are simply free gas contributions and can be easily computed). Let us focus on
\begin{eqnarray} \label{omS}
\Omega_s &=& \sum_{{\cal C}}\sum_{J^{P}}\frac{{\rm dim}\, {\cal C}}{2\pi^2\beta^2}\, (2J+1)\int^\infty_{2m_g} d\epsilon \, \epsilon^2\, K_2(\beta\epsilon) \nonumber \\
&& \times{\rm Tr}_{{\cal C},J^{P}} \Bigg[ \left( \delta {\rm Re} {\cal T }\right)'- 2\pi \Big( (\delta {\rm Re}{\cal T})(\delta {\rm Im}{\cal T})'-(\delta {\rm Im} {\cal T}) (\delta {\rm Re}{\cal T})' \Big) \Bigg] .
\end{eqnarray}
Using the following definitions concerning the trace of an operator $A$ in momentum space
\begin{equation}
{\rm Tr} A = \displaystyle\frac{1}{(2\pi)^ 3} \displaystyle\int_{-\infty}^{\infty} d\vec{q} \, \langle \vec{q}\, | A | \vec{q} \,\rangle
\end{equation}
and the partial wave expansion
\begin{equation}
\langle \vec{q}\, | A | \vec{q} \,'\rangle = A(q,q',\hat{q}.\hat{q}') = \displaystyle\frac{1}{4\pi} \displaystyle\sum_l (2l + 1) A_l(q,q') P_l(\hat{q}.\hat{q}')
\end{equation}
where $P_l(x)$ is the Legendre polynomial of order $l$, (\ref{omS}) reads
\begin{eqnarray}
\Omega_s &=& \displaystyle\frac{1}{64 \pi^ 5\beta^ 2} \displaystyle\sum_{J^P} (2 J + 1) \displaystyle\sum_{{\cal C}} \text{dim} {\cal C} \left( \beta \displaystyle\int_{2m_g}^\infty d\epsilon \, \epsilon^3 \displaystyle\sqrt{\frac{\epsilon^2}{4} - m_g^2} K_1(\beta\epsilon) \,{\rm Re} {\cal T}_{J^P}(\epsilon; q_\epsilon, q_\epsilon) \right.\nonumber \\ \nonumber
&& -\left. \displaystyle\frac{1}{16 \pi^2} \displaystyle\int_{2m_g}^\infty d\epsilon \, \epsilon^4 \, \left(\frac{\epsilon^2}{4} - m_g^2 \right) K_2(\beta\epsilon) \left[{\rm Re} {\cal T}_{J^P}(\epsilon; q_\epsilon, q_\epsilon) \left({\rm Im} {\cal T}_{J^P}(\epsilon; q_\epsilon, q_\epsilon) \right)' \right] \right. \\
&& + \left. \displaystyle\frac{1}{16 \pi^2} \displaystyle\int_{2m_g}^\infty d\epsilon \, \epsilon^4 \, \left(\frac{\epsilon^2}{4} - m_g^2\right) K_2(\beta\epsilon) \left[\left({\rm Re} {\cal T}_{J^P}(\epsilon; q_\epsilon, q_\epsilon)\right)' {\rm Im} {\cal T}_{J^P}(\epsilon; q_\epsilon, q_\epsilon) \right] \right).
\end{eqnarray}
\end{appendix}
|
Subsets and Splits